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.getDataLayout(), *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(), dl,
253 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
258 while (SVT != MVT::f32 && SVT != MVT::f16) {
259 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
260 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
261 // Only do this if the target has a native EXTLOAD instruction from
263 TLI.isLoadExtLegal(ISD::EXTLOAD, OrigVT, SVT) &&
264 TLI.ShouldShrinkFPConstant(OrigVT)) {
265 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
266 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
273 DAG.getConstantPool(LLVMC, TLI.getPointerTy(DAG.getDataLayout()));
274 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
277 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
279 CPIdx, MachinePointerInfo::getConstantPool(),
280 VT, false, false, false, Alignment);
284 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
285 MachinePointerInfo::getConstantPool(), false, false, false,
290 /// Expands an unaligned store to 2 half-size stores.
291 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
292 const TargetLowering &TLI,
293 SelectionDAGLegalize *DAGLegalize) {
294 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
295 "unaligned indexed stores not implemented!");
296 SDValue Chain = ST->getChain();
297 SDValue Ptr = ST->getBasePtr();
298 SDValue Val = ST->getValue();
299 EVT VT = Val.getValueType();
300 int Alignment = ST->getAlignment();
301 unsigned AS = ST->getAddressSpace();
304 if (ST->getMemoryVT().isFloatingPoint() ||
305 ST->getMemoryVT().isVector()) {
306 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
307 if (TLI.isTypeLegal(intVT)) {
308 // Expand to a bitconvert of the value to the integer type of the
309 // same size, then a (misaligned) int store.
310 // FIXME: Does not handle truncating floating point stores!
311 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
312 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
313 ST->isVolatile(), ST->isNonTemporal(), Alignment);
314 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
317 // Do a (aligned) store to a stack slot, then copy from the stack slot
318 // to the final destination using (unaligned) integer loads and stores.
319 EVT StoredVT = ST->getMemoryVT();
321 TLI.getRegisterType(*DAG.getContext(),
322 EVT::getIntegerVT(*DAG.getContext(),
323 StoredVT.getSizeInBits()));
324 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
325 unsigned RegBytes = RegVT.getSizeInBits() / 8;
326 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
328 // Make sure the stack slot is also aligned for the register type.
329 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
331 // Perform the original store, only redirected to the stack slot.
332 SDValue Store = DAG.getTruncStore(Chain, dl,
333 Val, StackPtr, MachinePointerInfo(),
334 StoredVT, false, false, 0);
335 SDValue Increment = DAG.getConstant(
336 RegBytes, dl, TLI.getPointerTy(DAG.getDataLayout(), AS));
337 SmallVector<SDValue, 8> Stores;
340 // Do all but one copies using the full register width.
341 for (unsigned i = 1; i < NumRegs; i++) {
342 // Load one integer register's worth from the stack slot.
343 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
344 MachinePointerInfo(),
345 false, false, false, 0);
346 // Store it to the final location. Remember the store.
347 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
348 ST->getPointerInfo().getWithOffset(Offset),
349 ST->isVolatile(), ST->isNonTemporal(),
350 MinAlign(ST->getAlignment(), Offset)));
351 // Increment the pointers.
353 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
355 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
358 // The last store may be partial. Do a truncating store. On big-endian
359 // machines this requires an extending load from the stack slot to ensure
360 // that the bits are in the right place.
361 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
362 8 * (StoredBytes - Offset));
364 // Load from the stack slot.
365 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
366 MachinePointerInfo(),
367 MemVT, false, false, false, 0);
369 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
371 .getWithOffset(Offset),
372 MemVT, ST->isVolatile(),
374 MinAlign(ST->getAlignment(), Offset),
376 // The order of the stores doesn't matter - say it with a TokenFactor.
377 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
378 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
381 assert(ST->getMemoryVT().isInteger() &&
382 !ST->getMemoryVT().isVector() &&
383 "Unaligned store of unknown type.");
384 // Get the half-size VT
385 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
386 int NumBits = NewStoredVT.getSizeInBits();
387 int IncrementSize = NumBits / 8;
389 // Divide the stored value in two parts.
390 SDValue ShiftAmount = DAG.getConstant(NumBits, dl,
391 TLI.getShiftAmountTy(Val.getValueType()));
393 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
395 // Store the two parts
396 SDValue Store1, Store2;
397 Store1 = DAG.getTruncStore(Chain, dl,
398 DAG.getDataLayout().isLittleEndian() ? Lo : Hi,
399 Ptr, ST->getPointerInfo(), NewStoredVT,
400 ST->isVolatile(), ST->isNonTemporal(), Alignment);
402 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
403 DAG.getConstant(IncrementSize, dl,
404 TLI.getPointerTy(DAG.getDataLayout(), AS)));
405 Alignment = MinAlign(Alignment, IncrementSize);
406 Store2 = DAG.getTruncStore(
407 Chain, dl, DAG.getDataLayout().isLittleEndian() ? Hi : Lo, Ptr,
408 ST->getPointerInfo().getWithOffset(IncrementSize), NewStoredVT,
409 ST->isVolatile(), ST->isNonTemporal(), Alignment, ST->getAAInfo());
412 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
413 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
416 /// Expands an unaligned load to 2 half-size loads.
418 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
419 const TargetLowering &TLI,
420 SDValue &ValResult, SDValue &ChainResult) {
421 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
422 "unaligned indexed loads not implemented!");
423 SDValue Chain = LD->getChain();
424 SDValue Ptr = LD->getBasePtr();
425 EVT VT = LD->getValueType(0);
426 EVT LoadedVT = LD->getMemoryVT();
428 if (VT.isFloatingPoint() || VT.isVector()) {
429 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
430 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
431 // Expand to a (misaligned) integer load of the same size,
432 // then bitconvert to floating point or vector.
433 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
434 LD->getMemOperand());
435 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
437 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
438 ISD::ANY_EXTEND, dl, VT, Result);
445 // Copy the value to a (aligned) stack slot using (unaligned) integer
446 // loads and stores, then do a (aligned) load from the stack slot.
447 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
448 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
449 unsigned RegBytes = RegVT.getSizeInBits() / 8;
450 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
452 // Make sure the stack slot is also aligned for the register type.
453 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
456 DAG.getConstant(RegBytes, dl, TLI.getPointerTy(DAG.getDataLayout()));
457 SmallVector<SDValue, 8> Stores;
458 SDValue StackPtr = StackBase;
461 // Do all but one copies using the full register width.
462 for (unsigned i = 1; i < NumRegs; i++) {
463 // Load one integer register's worth from the original location.
464 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
465 LD->getPointerInfo().getWithOffset(Offset),
466 LD->isVolatile(), LD->isNonTemporal(),
468 MinAlign(LD->getAlignment(), Offset),
470 // Follow the load with a store to the stack slot. Remember the store.
471 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
472 MachinePointerInfo(), false, false, 0));
473 // Increment the pointers.
475 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
476 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
480 // The last copy may be partial. Do an extending load.
481 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
482 8 * (LoadedBytes - Offset));
483 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
484 LD->getPointerInfo().getWithOffset(Offset),
485 MemVT, LD->isVolatile(),
488 MinAlign(LD->getAlignment(), Offset),
490 // Follow the load with a store to the stack slot. Remember the store.
491 // On big-endian machines this requires a truncating store to ensure
492 // that the bits end up in the right place.
493 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
494 MachinePointerInfo(), MemVT,
497 // The order of the stores doesn't matter - say it with a TokenFactor.
498 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
500 // Finally, perform the original load only redirected to the stack slot.
501 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
502 MachinePointerInfo(), LoadedVT, false,false, false,
505 // Callers expect a MERGE_VALUES node.
510 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
511 "Unaligned load of unsupported type.");
513 // Compute the new VT that is half the size of the old one. This is an
515 unsigned NumBits = LoadedVT.getSizeInBits();
517 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
520 unsigned Alignment = LD->getAlignment();
521 unsigned IncrementSize = NumBits / 8;
522 ISD::LoadExtType HiExtType = LD->getExtensionType();
524 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
525 if (HiExtType == ISD::NON_EXTLOAD)
526 HiExtType = ISD::ZEXTLOAD;
528 // Load the value in two parts
530 if (DAG.getDataLayout().isLittleEndian()) {
531 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
532 NewLoadedVT, LD->isVolatile(),
533 LD->isNonTemporal(), LD->isInvariant(), Alignment,
535 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
536 DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
537 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
538 LD->getPointerInfo().getWithOffset(IncrementSize),
539 NewLoadedVT, LD->isVolatile(),
540 LD->isNonTemporal(),LD->isInvariant(),
541 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
543 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
544 NewLoadedVT, LD->isVolatile(),
545 LD->isNonTemporal(), LD->isInvariant(), Alignment,
547 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
548 DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
549 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
550 LD->getPointerInfo().getWithOffset(IncrementSize),
551 NewLoadedVT, LD->isVolatile(),
552 LD->isNonTemporal(), LD->isInvariant(),
553 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
556 // aggregate the two parts
557 SDValue ShiftAmount = DAG.getConstant(NumBits, dl,
558 TLI.getShiftAmountTy(Hi.getValueType()));
559 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
560 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
562 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
569 /// Some target cannot handle a variable insertion index for the
570 /// INSERT_VECTOR_ELT instruction. In this case, it
571 /// is necessary to spill the vector being inserted into to memory, perform
572 /// the insert there, and then read the result back.
573 SDValue SelectionDAGLegalize::
574 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
580 // If the target doesn't support this, we have to spill the input vector
581 // to a temporary stack slot, update the element, then reload it. This is
582 // badness. We could also load the value into a vector register (either
583 // with a "move to register" or "extload into register" instruction, then
584 // permute it into place, if the idx is a constant and if the idx is
585 // supported by the target.
586 EVT VT = Tmp1.getValueType();
587 EVT EltVT = VT.getVectorElementType();
588 EVT IdxVT = Tmp3.getValueType();
589 EVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
590 SDValue StackPtr = DAG.CreateStackTemporary(VT);
592 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
595 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
596 MachinePointerInfo::getFixedStack(SPFI),
599 // Truncate or zero extend offset to target pointer type.
600 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
601 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
602 // Add the offset to the index.
603 unsigned EltSize = EltVT.getSizeInBits()/8;
604 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,
605 DAG.getConstant(EltSize, dl, IdxVT));
606 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
607 // Store the scalar value.
608 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
610 // Load the updated vector.
611 return DAG.getLoad(VT, dl, Ch, StackPtr,
612 MachinePointerInfo::getFixedStack(SPFI), false, false,
617 SDValue SelectionDAGLegalize::
618 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
619 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
620 // SCALAR_TO_VECTOR requires that the type of the value being inserted
621 // match the element type of the vector being created, except for
622 // integers in which case the inserted value can be over width.
623 EVT EltVT = Vec.getValueType().getVectorElementType();
624 if (Val.getValueType() == EltVT ||
625 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
626 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
627 Vec.getValueType(), Val);
629 unsigned NumElts = Vec.getValueType().getVectorNumElements();
630 // We generate a shuffle of InVec and ScVec, so the shuffle mask
631 // should be 0,1,2,3,4,5... with the appropriate element replaced with
633 SmallVector<int, 8> ShufOps;
634 for (unsigned i = 0; i != NumElts; ++i)
635 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
637 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
641 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
644 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
645 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
646 // FIXME: We shouldn't do this for TargetConstantFP's.
647 // FIXME: move this to the DAG Combiner! Note that we can't regress due
648 // to phase ordering between legalized code and the dag combiner. This
649 // probably means that we need to integrate dag combiner and legalizer
651 // We generally can't do this one for long doubles.
652 SDValue Chain = ST->getChain();
653 SDValue Ptr = ST->getBasePtr();
654 unsigned Alignment = ST->getAlignment();
655 bool isVolatile = ST->isVolatile();
656 bool isNonTemporal = ST->isNonTemporal();
657 AAMDNodes AAInfo = ST->getAAInfo();
659 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
660 if (CFP->getValueType(0) == MVT::f32 &&
661 TLI.isTypeLegal(MVT::i32)) {
662 SDValue Con = DAG.getConstant(CFP->getValueAPF().
663 bitcastToAPInt().zextOrTrunc(32),
664 SDLoc(CFP), MVT::i32);
665 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
666 isVolatile, isNonTemporal, Alignment, AAInfo);
669 if (CFP->getValueType(0) == MVT::f64) {
670 // If this target supports 64-bit registers, do a single 64-bit store.
671 if (TLI.isTypeLegal(MVT::i64)) {
672 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
673 zextOrTrunc(64), SDLoc(CFP), MVT::i64);
674 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
675 isVolatile, isNonTemporal, Alignment, AAInfo);
678 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
679 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
680 // stores. If the target supports neither 32- nor 64-bits, this
681 // xform is certainly not worth it.
682 const APInt &IntVal = CFP->getValueAPF().bitcastToAPInt();
683 SDValue Lo = DAG.getConstant(IntVal.trunc(32), dl, MVT::i32);
684 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), dl, MVT::i32);
685 if (DAG.getDataLayout().isBigEndian())
688 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
689 isNonTemporal, Alignment, AAInfo);
690 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
691 DAG.getConstant(4, dl, Ptr.getValueType()));
692 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
693 ST->getPointerInfo().getWithOffset(4),
694 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
697 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
701 return SDValue(nullptr, 0);
704 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
705 StoreSDNode *ST = cast<StoreSDNode>(Node);
706 SDValue Chain = ST->getChain();
707 SDValue Ptr = ST->getBasePtr();
710 unsigned Alignment = ST->getAlignment();
711 bool isVolatile = ST->isVolatile();
712 bool isNonTemporal = ST->isNonTemporal();
713 AAMDNodes AAInfo = ST->getAAInfo();
715 if (!ST->isTruncatingStore()) {
716 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
717 ReplaceNode(ST, OptStore);
722 SDValue Value = ST->getValue();
723 MVT VT = Value.getSimpleValueType();
724 switch (TLI.getOperationAction(ISD::STORE, VT)) {
725 default: llvm_unreachable("This action is not supported yet!");
726 case TargetLowering::Legal: {
727 // If this is an unaligned store and the target doesn't support it,
729 unsigned AS = ST->getAddressSpace();
730 unsigned Align = ST->getAlignment();
731 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
732 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
733 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
734 if (Align < ABIAlignment)
735 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
739 case TargetLowering::Custom: {
740 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
741 if (Res && Res != SDValue(Node, 0))
742 ReplaceNode(SDValue(Node, 0), Res);
745 case TargetLowering::Promote: {
746 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
747 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
748 "Can only promote stores to same size type");
749 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
751 DAG.getStore(Chain, dl, Value, Ptr,
752 ST->getPointerInfo(), isVolatile,
753 isNonTemporal, Alignment, AAInfo);
754 ReplaceNode(SDValue(Node, 0), Result);
761 SDValue Value = ST->getValue();
763 EVT StVT = ST->getMemoryVT();
764 unsigned StWidth = StVT.getSizeInBits();
765 auto &DL = DAG.getDataLayout();
767 if (StWidth != StVT.getStoreSizeInBits()) {
768 // Promote to a byte-sized store with upper bits zero if not
769 // storing an integral number of bytes. For example, promote
770 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
771 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
772 StVT.getStoreSizeInBits());
773 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
775 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
776 NVT, isVolatile, isNonTemporal, Alignment, AAInfo);
777 ReplaceNode(SDValue(Node, 0), Result);
778 } else if (StWidth & (StWidth - 1)) {
779 // If not storing a power-of-2 number of bits, expand as two stores.
780 assert(!StVT.isVector() && "Unsupported truncstore!");
781 unsigned RoundWidth = 1 << Log2_32(StWidth);
782 assert(RoundWidth < StWidth);
783 unsigned ExtraWidth = StWidth - RoundWidth;
784 assert(ExtraWidth < RoundWidth);
785 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
786 "Store size not an integral number of bytes!");
787 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
788 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
790 unsigned IncrementSize;
792 if (DL.isLittleEndian()) {
793 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
794 // Store the bottom RoundWidth bits.
795 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
797 isVolatile, isNonTemporal, Alignment,
800 // Store the remaining ExtraWidth bits.
801 IncrementSize = RoundWidth / 8;
802 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
803 DAG.getConstant(IncrementSize, dl,
804 Ptr.getValueType()));
805 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
806 DAG.getConstant(RoundWidth, dl,
807 TLI.getShiftAmountTy(Value.getValueType())));
808 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
809 ST->getPointerInfo().getWithOffset(IncrementSize),
810 ExtraVT, isVolatile, isNonTemporal,
811 MinAlign(Alignment, IncrementSize), AAInfo);
813 // Big endian - avoid unaligned stores.
814 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
815 // Store the top RoundWidth bits.
816 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
817 DAG.getConstant(ExtraWidth, dl,
818 TLI.getShiftAmountTy(Value.getValueType())));
819 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
820 RoundVT, isVolatile, isNonTemporal, Alignment,
823 // Store the remaining ExtraWidth bits.
824 IncrementSize = RoundWidth / 8;
825 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
826 DAG.getConstant(IncrementSize, dl,
827 Ptr.getValueType()));
828 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
829 ST->getPointerInfo().getWithOffset(IncrementSize),
830 ExtraVT, isVolatile, isNonTemporal,
831 MinAlign(Alignment, IncrementSize), AAInfo);
834 // The order of the stores doesn't matter.
835 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
836 ReplaceNode(SDValue(Node, 0), Result);
838 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
839 StVT.getSimpleVT())) {
840 default: llvm_unreachable("This action is not supported yet!");
841 case TargetLowering::Legal: {
842 unsigned AS = ST->getAddressSpace();
843 unsigned Align = ST->getAlignment();
844 // If this is an unaligned store and the target doesn't support it,
846 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
847 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
848 unsigned ABIAlignment = DL.getABITypeAlignment(Ty);
849 if (Align < ABIAlignment)
850 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
854 case TargetLowering::Custom: {
855 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
856 if (Res && Res != SDValue(Node, 0))
857 ReplaceNode(SDValue(Node, 0), Res);
860 case TargetLowering::Expand:
861 assert(!StVT.isVector() &&
862 "Vector Stores are handled in LegalizeVectorOps");
864 // TRUNCSTORE:i16 i32 -> STORE i16
865 assert(TLI.isTypeLegal(StVT) &&
866 "Do not know how to expand this store!");
867 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
869 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
870 isVolatile, isNonTemporal, Alignment, AAInfo);
871 ReplaceNode(SDValue(Node, 0), Result);
878 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
879 LoadSDNode *LD = cast<LoadSDNode>(Node);
880 SDValue Chain = LD->getChain(); // The chain.
881 SDValue Ptr = LD->getBasePtr(); // The base pointer.
882 SDValue Value; // The value returned by the load op.
885 ISD::LoadExtType ExtType = LD->getExtensionType();
886 if (ExtType == ISD::NON_EXTLOAD) {
887 MVT VT = Node->getSimpleValueType(0);
888 SDValue RVal = SDValue(Node, 0);
889 SDValue RChain = SDValue(Node, 1);
891 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
892 default: llvm_unreachable("This action is not supported yet!");
893 case TargetLowering::Legal: {
894 unsigned AS = LD->getAddressSpace();
895 unsigned Align = LD->getAlignment();
896 // If this is an unaligned load and the target doesn't support it,
898 if (!TLI.allowsMisalignedMemoryAccesses(LD->getMemoryVT(), AS, Align)) {
899 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
900 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
901 if (Align < ABIAlignment){
902 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
907 case TargetLowering::Custom: {
908 SDValue Res = TLI.LowerOperation(RVal, DAG);
911 RChain = Res.getValue(1);
915 case TargetLowering::Promote: {
916 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
917 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
918 "Can only promote loads to same size type");
920 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
921 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
922 RChain = Res.getValue(1);
926 if (RChain.getNode() != Node) {
927 assert(RVal.getNode() != Node && "Load must be completely replaced");
928 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
929 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
931 UpdatedNodes->insert(RVal.getNode());
932 UpdatedNodes->insert(RChain.getNode());
939 EVT SrcVT = LD->getMemoryVT();
940 unsigned SrcWidth = SrcVT.getSizeInBits();
941 unsigned Alignment = LD->getAlignment();
942 bool isVolatile = LD->isVolatile();
943 bool isNonTemporal = LD->isNonTemporal();
944 bool isInvariant = LD->isInvariant();
945 AAMDNodes AAInfo = LD->getAAInfo();
947 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
948 // Some targets pretend to have an i1 loading operation, and actually
949 // load an i8. This trick is correct for ZEXTLOAD because the top 7
950 // bits are guaranteed to be zero; it helps the optimizers understand
951 // that these bits are zero. It is also useful for EXTLOAD, since it
952 // tells the optimizers that those bits are undefined. It would be
953 // nice to have an effective generic way of getting these benefits...
954 // Until such a way is found, don't insist on promoting i1 here.
956 TLI.getLoadExtAction(ExtType, Node->getValueType(0), MVT::i1) ==
957 TargetLowering::Promote)) {
958 // Promote to a byte-sized load if not loading an integral number of
959 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
960 unsigned NewWidth = SrcVT.getStoreSizeInBits();
961 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
964 // The extra bits are guaranteed to be zero, since we stored them that
965 // way. A zext load from NVT thus automatically gives zext from SrcVT.
967 ISD::LoadExtType NewExtType =
968 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
971 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
972 Chain, Ptr, LD->getPointerInfo(),
973 NVT, isVolatile, isNonTemporal, isInvariant, Alignment,
976 Ch = Result.getValue(1); // The chain.
978 if (ExtType == ISD::SEXTLOAD)
979 // Having the top bits zero doesn't help when sign extending.
980 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
981 Result.getValueType(),
982 Result, DAG.getValueType(SrcVT));
983 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
984 // All the top bits are guaranteed to be zero - inform the optimizers.
985 Result = DAG.getNode(ISD::AssertZext, dl,
986 Result.getValueType(), Result,
987 DAG.getValueType(SrcVT));
991 } else if (SrcWidth & (SrcWidth - 1)) {
992 // If not loading a power-of-2 number of bits, expand as two loads.
993 assert(!SrcVT.isVector() && "Unsupported extload!");
994 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
995 assert(RoundWidth < SrcWidth);
996 unsigned ExtraWidth = SrcWidth - RoundWidth;
997 assert(ExtraWidth < RoundWidth);
998 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
999 "Load size not an integral number of bytes!");
1000 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
1001 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
1003 unsigned IncrementSize;
1004 auto &DL = DAG.getDataLayout();
1006 if (DL.isLittleEndian()) {
1007 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
1008 // Load the bottom RoundWidth bits.
1009 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1011 LD->getPointerInfo(), RoundVT, isVolatile,
1012 isNonTemporal, isInvariant, Alignment, AAInfo);
1014 // Load the remaining ExtraWidth bits.
1015 IncrementSize = RoundWidth / 8;
1016 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1017 DAG.getConstant(IncrementSize, dl,
1018 Ptr.getValueType()));
1019 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1020 LD->getPointerInfo().getWithOffset(IncrementSize),
1021 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1022 MinAlign(Alignment, IncrementSize), AAInfo);
1024 // Build a factor node to remember that this load is independent of
1026 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1029 // Move the top bits to the right place.
1030 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1031 DAG.getConstant(RoundWidth, dl,
1032 TLI.getShiftAmountTy(Hi.getValueType())));
1034 // Join the hi and lo parts.
1035 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1037 // Big endian - avoid unaligned loads.
1038 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1039 // Load the top RoundWidth bits.
1040 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1041 LD->getPointerInfo(), RoundVT, isVolatile,
1042 isNonTemporal, isInvariant, Alignment, AAInfo);
1044 // Load the remaining ExtraWidth bits.
1045 IncrementSize = RoundWidth / 8;
1046 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1047 DAG.getConstant(IncrementSize, dl,
1048 Ptr.getValueType()));
1049 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1050 dl, Node->getValueType(0), Chain, Ptr,
1051 LD->getPointerInfo().getWithOffset(IncrementSize),
1052 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1053 MinAlign(Alignment, IncrementSize), AAInfo);
1055 // Build a factor node to remember that this load is independent of
1057 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1060 // Move the top bits to the right place.
1061 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1062 DAG.getConstant(ExtraWidth, dl,
1063 TLI.getShiftAmountTy(Hi.getValueType())));
1065 // Join the hi and lo parts.
1066 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1071 bool isCustom = false;
1072 switch (TLI.getLoadExtAction(ExtType, Node->getValueType(0),
1073 SrcVT.getSimpleVT())) {
1074 default: llvm_unreachable("This action is not supported yet!");
1075 case TargetLowering::Custom:
1078 case TargetLowering::Legal: {
1079 Value = SDValue(Node, 0);
1080 Chain = SDValue(Node, 1);
1083 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1084 if (Res.getNode()) {
1086 Chain = Res.getValue(1);
1089 // If this is an unaligned load and the target doesn't support
1091 EVT MemVT = LD->getMemoryVT();
1092 unsigned AS = LD->getAddressSpace();
1093 unsigned Align = LD->getAlignment();
1094 if (!TLI.allowsMisalignedMemoryAccesses(MemVT, AS, Align)) {
1095 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1096 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
1097 if (Align < ABIAlignment){
1098 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, Value, Chain);
1104 case TargetLowering::Expand:
1105 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, Node->getValueType(0), SrcVT)) {
1106 // If the source type is not legal, see if there is a legal extload to
1107 // an intermediate type that we can then extend further.
1108 EVT LoadVT = TLI.getRegisterType(SrcVT.getSimpleVT());
1109 if (TLI.isTypeLegal(SrcVT) || // Same as SrcVT == LoadVT?
1110 TLI.isLoadExtLegal(ExtType, LoadVT, SrcVT)) {
1111 // If we are loading a legal type, this is a non-extload followed by a
1113 ISD::LoadExtType MidExtType =
1114 (LoadVT == SrcVT) ? ISD::NON_EXTLOAD : ExtType;
1116 SDValue Load = DAG.getExtLoad(MidExtType, dl, LoadVT, Chain, Ptr,
1117 SrcVT, LD->getMemOperand());
1119 ISD::getExtForLoadExtType(SrcVT.isFloatingPoint(), ExtType);
1120 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1121 Chain = Load.getValue(1);
1126 assert(!SrcVT.isVector() &&
1127 "Vector Loads are handled in LegalizeVectorOps");
1129 // FIXME: This does not work for vectors on most targets. Sign-
1130 // and zero-extend operations are currently folded into extending
1131 // loads, whether they are legal or not, and then we end up here
1132 // without any support for legalizing them.
1133 assert(ExtType != ISD::EXTLOAD &&
1134 "EXTLOAD should always be supported!");
1135 // Turn the unsupported load into an EXTLOAD followed by an
1136 // explicit zero/sign extend inreg.
1137 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1138 Node->getValueType(0),
1140 LD->getMemOperand());
1142 if (ExtType == ISD::SEXTLOAD)
1143 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1144 Result.getValueType(),
1145 Result, DAG.getValueType(SrcVT));
1147 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1149 Chain = Result.getValue(1);
1154 // Since loads produce two values, make sure to remember that we legalized
1156 if (Chain.getNode() != Node) {
1157 assert(Value.getNode() != Node && "Load must be completely replaced");
1158 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1159 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1161 UpdatedNodes->insert(Value.getNode());
1162 UpdatedNodes->insert(Chain.getNode());
1168 /// Return a legal replacement for the given operation, with all legal operands.
1169 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1170 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
1172 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1176 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1177 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1178 TargetLowering::TypeLegal &&
1179 "Unexpected illegal type!");
1181 for (const SDValue &Op : Node->op_values())
1182 assert((TLI.getTypeAction(*DAG.getContext(),
1183 Op.getValueType()) == TargetLowering::TypeLegal ||
1184 Op.getOpcode() == ISD::TargetConstant) &&
1185 "Unexpected illegal type!");
1188 // Figure out the correct action; the way to query this varies by opcode
1189 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1190 bool SimpleFinishLegalizing = true;
1191 switch (Node->getOpcode()) {
1192 case ISD::INTRINSIC_W_CHAIN:
1193 case ISD::INTRINSIC_WO_CHAIN:
1194 case ISD::INTRINSIC_VOID:
1195 case ISD::STACKSAVE:
1196 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1199 Action = TLI.getOperationAction(Node->getOpcode(),
1200 Node->getValueType(0));
1201 if (Action != TargetLowering::Promote)
1202 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1204 case ISD::FP_TO_FP16:
1205 case ISD::SINT_TO_FP:
1206 case ISD::UINT_TO_FP:
1207 case ISD::EXTRACT_VECTOR_ELT:
1208 Action = TLI.getOperationAction(Node->getOpcode(),
1209 Node->getOperand(0).getValueType());
1211 case ISD::FP_ROUND_INREG:
1212 case ISD::SIGN_EXTEND_INREG: {
1213 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1214 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1217 case ISD::ATOMIC_STORE: {
1218 Action = TLI.getOperationAction(Node->getOpcode(),
1219 Node->getOperand(2).getValueType());
1222 case ISD::SELECT_CC:
1225 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1226 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1227 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1228 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1229 ISD::CondCode CCCode =
1230 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1231 Action = TLI.getCondCodeAction(CCCode, OpVT);
1232 if (Action == TargetLowering::Legal) {
1233 if (Node->getOpcode() == ISD::SELECT_CC)
1234 Action = TLI.getOperationAction(Node->getOpcode(),
1235 Node->getValueType(0));
1237 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1243 // FIXME: Model these properly. LOAD and STORE are complicated, and
1244 // STORE expects the unlegalized operand in some cases.
1245 SimpleFinishLegalizing = false;
1247 case ISD::CALLSEQ_START:
1248 case ISD::CALLSEQ_END:
1249 // FIXME: This shouldn't be necessary. These nodes have special properties
1250 // dealing with the recursive nature of legalization. Removing this
1251 // special case should be done as part of making LegalizeDAG non-recursive.
1252 SimpleFinishLegalizing = false;
1254 case ISD::EXTRACT_ELEMENT:
1255 case ISD::FLT_ROUNDS_:
1257 case ISD::MERGE_VALUES:
1258 case ISD::EH_RETURN:
1259 case ISD::FRAME_TO_ARGS_OFFSET:
1260 case ISD::EH_SJLJ_SETJMP:
1261 case ISD::EH_SJLJ_LONGJMP:
1262 // These operations lie about being legal: when they claim to be legal,
1263 // they should actually be expanded.
1264 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1265 if (Action == TargetLowering::Legal)
1266 Action = TargetLowering::Expand;
1268 case ISD::INIT_TRAMPOLINE:
1269 case ISD::ADJUST_TRAMPOLINE:
1270 case ISD::FRAMEADDR:
1271 case ISD::RETURNADDR:
1272 // These operations lie about being legal: when they claim to be legal,
1273 // they should actually be custom-lowered.
1274 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1275 if (Action == TargetLowering::Legal)
1276 Action = TargetLowering::Custom;
1278 case ISD::READ_REGISTER:
1279 case ISD::WRITE_REGISTER:
1280 // Named register is legal in the DAG, but blocked by register name
1281 // selection if not implemented by target (to chose the correct register)
1282 // They'll be converted to Copy(To/From)Reg.
1283 Action = TargetLowering::Legal;
1285 case ISD::DEBUGTRAP:
1286 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1287 if (Action == TargetLowering::Expand) {
1288 // replace ISD::DEBUGTRAP with ISD::TRAP
1290 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1291 Node->getOperand(0));
1292 ReplaceNode(Node, NewVal.getNode());
1293 LegalizeOp(NewVal.getNode());
1299 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1300 Action = TargetLowering::Legal;
1302 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1307 if (SimpleFinishLegalizing) {
1308 SDNode *NewNode = Node;
1309 switch (Node->getOpcode()) {
1316 // Legalizing shifts/rotates requires adjusting the shift amount
1317 // to the appropriate width.
1318 if (!Node->getOperand(1).getValueType().isVector()) {
1320 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1321 Node->getOperand(1));
1322 HandleSDNode Handle(SAO);
1323 LegalizeOp(SAO.getNode());
1324 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1328 case ISD::SRL_PARTS:
1329 case ISD::SRA_PARTS:
1330 case ISD::SHL_PARTS:
1331 // Legalizing shifts/rotates requires adjusting the shift amount
1332 // to the appropriate width.
1333 if (!Node->getOperand(2).getValueType().isVector()) {
1335 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1336 Node->getOperand(2));
1337 HandleSDNode Handle(SAO);
1338 LegalizeOp(SAO.getNode());
1339 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1340 Node->getOperand(1),
1346 if (NewNode != Node) {
1347 ReplaceNode(Node, NewNode);
1351 case TargetLowering::Legal:
1353 case TargetLowering::Custom: {
1354 // FIXME: The handling for custom lowering with multiple results is
1356 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1357 if (Res.getNode()) {
1358 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1361 if (Node->getNumValues() == 1) {
1362 // We can just directly replace this node with the lowered value.
1363 ReplaceNode(SDValue(Node, 0), Res);
1367 SmallVector<SDValue, 8> ResultVals;
1368 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1369 ResultVals.push_back(Res.getValue(i));
1370 ReplaceNode(Node, ResultVals.data());
1375 case TargetLowering::Expand:
1378 case TargetLowering::Promote:
1384 switch (Node->getOpcode()) {
1391 llvm_unreachable("Do not know how to legalize this operator!");
1393 case ISD::CALLSEQ_START:
1394 case ISD::CALLSEQ_END:
1397 return LegalizeLoadOps(Node);
1400 return LegalizeStoreOps(Node);
1405 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1406 SDValue Vec = Op.getOperand(0);
1407 SDValue Idx = Op.getOperand(1);
1410 // Before we generate a new store to a temporary stack slot, see if there is
1411 // already one that we can use. There often is because when we scalarize
1412 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1413 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1414 // the vector. If all are expanded here, we don't want one store per vector
1416 SDValue StackPtr, Ch;
1417 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1418 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1420 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1421 if (ST->isIndexed() || ST->isTruncatingStore() ||
1422 ST->getValue() != Vec)
1425 // Make sure that nothing else could have stored into the destination of
1427 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1430 StackPtr = ST->getBasePtr();
1431 Ch = SDValue(ST, 0);
1436 if (!Ch.getNode()) {
1437 // Store the value to a temporary stack slot, then LOAD the returned part.
1438 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1439 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1440 MachinePointerInfo(), false, false, 0);
1443 // Add the offset to the index.
1445 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1446 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1447 DAG.getConstant(EltSize, SDLoc(Vec), Idx.getValueType()));
1449 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy(DAG.getDataLayout()));
1450 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1454 if (Op.getValueType().isVector())
1455 NewLoad = DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,
1456 MachinePointerInfo(), false, false, false, 0);
1458 NewLoad = DAG.getExtLoad(
1459 ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr, MachinePointerInfo(),
1460 Vec.getValueType().getVectorElementType(), false, false, false, 0);
1462 // Replace the chain going out of the store, by the one out of the load.
1463 DAG.ReplaceAllUsesOfValueWith(Ch, SDValue(NewLoad.getNode(), 1));
1465 // We introduced a cycle though, so update the loads operands, making sure
1466 // to use the original store's chain as an incoming chain.
1467 SmallVector<SDValue, 6> NewLoadOperands(NewLoad->op_begin(),
1469 NewLoadOperands[0] = Ch;
1471 SDValue(DAG.UpdateNodeOperands(NewLoad.getNode(), NewLoadOperands), 0);
1475 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1476 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1478 SDValue Vec = Op.getOperand(0);
1479 SDValue Part = Op.getOperand(1);
1480 SDValue Idx = Op.getOperand(2);
1483 // Store the value to a temporary stack slot, then LOAD the returned part.
1485 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1486 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1487 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1489 // First store the whole vector.
1490 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1493 // Then store the inserted part.
1495 // Add the offset to the index.
1497 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1499 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1500 DAG.getConstant(EltSize, SDLoc(Vec), Idx.getValueType()));
1501 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy(DAG.getDataLayout()));
1503 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1506 // Store the subvector.
1507 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr,
1508 MachinePointerInfo(), false, false, 0);
1510 // Finally, load the updated vector.
1511 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1512 false, false, false, 0);
1515 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1516 // We can't handle this case efficiently. Allocate a sufficiently
1517 // aligned object on the stack, store each element into it, then load
1518 // the result as a vector.
1519 // Create the stack frame object.
1520 EVT VT = Node->getValueType(0);
1521 EVT EltVT = VT.getVectorElementType();
1523 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1524 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1525 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1527 // Emit a store of each element to the stack slot.
1528 SmallVector<SDValue, 8> Stores;
1529 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1530 // Store (in the right endianness) the elements to memory.
1531 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1532 // Ignore undef elements.
1533 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1535 unsigned Offset = TypeByteSize*i;
1537 SDValue Idx = DAG.getConstant(Offset, dl, FIPtr.getValueType());
1538 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1540 // If the destination vector element type is narrower than the source
1541 // element type, only store the bits necessary.
1542 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1543 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1544 Node->getOperand(i), Idx,
1545 PtrInfo.getWithOffset(Offset),
1546 EltVT, false, false, 0));
1548 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1549 Node->getOperand(i), Idx,
1550 PtrInfo.getWithOffset(Offset),
1555 if (!Stores.empty()) // Not all undef elements?
1556 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1558 StoreChain = DAG.getEntryNode();
1560 // Result is a load from the stack slot.
1561 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1562 false, false, false, 0);
1565 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1567 SDValue Tmp1 = Node->getOperand(0);
1568 SDValue Tmp2 = Node->getOperand(1);
1570 // Get the sign bit of the RHS. First obtain a value that has the same
1571 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1573 EVT FloatVT = Tmp2.getValueType();
1574 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1575 if (TLI.isTypeLegal(IVT)) {
1576 // Convert to an integer with the same sign bit.
1577 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1579 auto &DL = DAG.getDataLayout();
1580 // Store the float to memory, then load the sign part out as an integer.
1581 MVT LoadTy = TLI.getPointerTy(DL);
1582 // First create a temporary that is aligned for both the load and store.
1583 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1584 // Then store the float to it.
1586 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1588 if (DL.isBigEndian()) {
1589 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1590 // Load out a legal integer with the same sign bit as the float.
1591 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1592 false, false, false, 0);
1593 } else { // Little endian
1594 SDValue LoadPtr = StackPtr;
1595 // The float may be wider than the integer we are going to load. Advance
1596 // the pointer so that the loaded integer will contain the sign bit.
1597 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1598 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1599 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1600 DAG.getConstant(ByteOffset, dl,
1601 LoadPtr.getValueType()));
1602 // Load a legal integer containing the sign bit.
1603 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1604 false, false, false, 0);
1605 // Move the sign bit to the top bit of the loaded integer.
1606 unsigned BitShift = LoadTy.getSizeInBits() -
1607 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1608 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1610 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1611 DAG.getConstant(BitShift, dl,
1612 TLI.getShiftAmountTy(SignBit.getValueType())));
1615 // Now get the sign bit proper, by seeing whether the value is negative.
1616 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1618 DAG.getConstant(0, dl, SignBit.getValueType()),
1620 // Get the absolute value of the result.
1621 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1622 // Select between the nabs and abs value based on the sign bit of
1624 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1625 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1629 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1630 SmallVectorImpl<SDValue> &Results) {
1631 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1632 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1633 " not tell us which reg is the stack pointer!");
1635 EVT VT = Node->getValueType(0);
1636 SDValue Tmp1 = SDValue(Node, 0);
1637 SDValue Tmp2 = SDValue(Node, 1);
1638 SDValue Tmp3 = Node->getOperand(2);
1639 SDValue Chain = Tmp1.getOperand(0);
1641 // Chain the dynamic stack allocation so that it doesn't modify the stack
1642 // pointer when other instructions are using the stack.
1643 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
1645 SDValue Size = Tmp2.getOperand(1);
1646 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1647 Chain = SP.getValue(1);
1648 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1649 unsigned StackAlign =
1650 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
1651 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1652 if (Align > StackAlign)
1653 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1654 DAG.getConstant(-(uint64_t)Align, dl, VT));
1655 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1657 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
1658 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
1660 Results.push_back(Tmp1);
1661 Results.push_back(Tmp2);
1664 /// Legalize a SETCC with given LHS and RHS and condition code CC on the current
1667 /// If the SETCC has been legalized using AND / OR, then the legalized node
1668 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1669 /// will be set to false.
1671 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1672 /// then the values of LHS and RHS will be swapped, CC will be set to the
1673 /// new condition, and NeedInvert will be set to false.
1675 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1676 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1677 /// will be set to true. The caller must invert the result of the SETCC with
1678 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1679 /// of a true/false result.
1681 /// \returns true if the SetCC has been legalized, false if it hasn't.
1682 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1683 SDValue &LHS, SDValue &RHS,
1687 MVT OpVT = LHS.getSimpleValueType();
1688 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1690 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1691 default: llvm_unreachable("Unknown condition code action!");
1692 case TargetLowering::Legal:
1695 case TargetLowering::Expand: {
1696 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1697 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1698 std::swap(LHS, RHS);
1699 CC = DAG.getCondCode(InvCC);
1702 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1705 default: llvm_unreachable("Don't know how to expand this condition!");
1707 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1708 == TargetLowering::Legal
1709 && "If SETO is expanded, SETOEQ must be legal!");
1710 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1712 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1713 == TargetLowering::Legal
1714 && "If SETUO is expanded, SETUNE must be legal!");
1715 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1728 // If we are floating point, assign and break, otherwise fall through.
1729 if (!OpVT.isInteger()) {
1730 // We can use the 4th bit to tell if we are the unordered
1731 // or ordered version of the opcode.
1732 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1733 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1734 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1737 // Fallthrough if we are unsigned integer.
1742 // We only support using the inverted operation, which is computed above
1743 // and not a different manner of supporting expanding these cases.
1744 llvm_unreachable("Don't know how to expand this condition!");
1747 // Try inverting the result of the inverse condition.
1748 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1749 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1750 CC = DAG.getCondCode(InvCC);
1754 // If inverting the condition didn't work then we have no means to expand
1756 llvm_unreachable("Don't know how to expand this condition!");
1759 SDValue SetCC1, SetCC2;
1760 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1761 // If we aren't the ordered or unorder operation,
1762 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1763 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1764 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1766 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1767 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1768 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1770 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1779 /// Emit a store/load combination to the stack. This stores
1780 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1781 /// a load from the stack slot to DestVT, extending it if needed.
1782 /// The resultant code need not be legal.
1783 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1787 // Create the stack frame object.
1788 unsigned SrcAlign = DAG.getDataLayout().getPrefTypeAlignment(
1789 SrcOp.getValueType().getTypeForEVT(*DAG.getContext()));
1790 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1792 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1793 int SPFI = StackPtrFI->getIndex();
1794 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1796 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1797 unsigned SlotSize = SlotVT.getSizeInBits();
1798 unsigned DestSize = DestVT.getSizeInBits();
1799 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1800 unsigned DestAlign = DAG.getDataLayout().getPrefTypeAlignment(DestType);
1802 // Emit a store to the stack slot. Use a truncstore if the input value is
1803 // later than DestVT.
1806 if (SrcSize > SlotSize)
1807 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1808 PtrInfo, SlotVT, false, false, SrcAlign);
1810 assert(SrcSize == SlotSize && "Invalid store");
1811 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1812 PtrInfo, false, false, SrcAlign);
1815 // Result is a load from the stack slot.
1816 if (SlotSize == DestSize)
1817 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1818 false, false, false, DestAlign);
1820 assert(SlotSize < DestSize && "Unknown extension!");
1821 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1822 PtrInfo, SlotVT, false, false, false, DestAlign);
1825 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1827 // Create a vector sized/aligned stack slot, store the value to element #0,
1828 // then load the whole vector back out.
1829 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1831 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1832 int SPFI = StackPtrFI->getIndex();
1834 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1836 MachinePointerInfo::getFixedStack(SPFI),
1837 Node->getValueType(0).getVectorElementType(),
1839 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1840 MachinePointerInfo::getFixedStack(SPFI),
1841 false, false, false, 0);
1845 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1846 const TargetLowering &TLI, SDValue &Res) {
1847 unsigned NumElems = Node->getNumOperands();
1849 EVT VT = Node->getValueType(0);
1851 // Try to group the scalars into pairs, shuffle the pairs together, then
1852 // shuffle the pairs of pairs together, etc. until the vector has
1853 // been built. This will work only if all of the necessary shuffle masks
1856 // We do this in two phases; first to check the legality of the shuffles,
1857 // and next, assuming that all shuffles are legal, to create the new nodes.
1858 for (int Phase = 0; Phase < 2; ++Phase) {
1859 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1861 for (unsigned i = 0; i < NumElems; ++i) {
1862 SDValue V = Node->getOperand(i);
1863 if (V.getOpcode() == ISD::UNDEF)
1868 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1869 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1872 while (IntermedVals.size() > 2) {
1873 NewIntermedVals.clear();
1874 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1875 // This vector and the next vector are shuffled together (simply to
1876 // append the one to the other).
1877 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1879 SmallVector<int, 16> FinalIndices;
1880 FinalIndices.reserve(IntermedVals[i].second.size() +
1881 IntermedVals[i+1].second.size());
1884 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1887 FinalIndices.push_back(IntermedVals[i].second[j]);
1889 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1891 ShuffleVec[k] = NumElems + j;
1892 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1897 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1898 IntermedVals[i+1].first,
1900 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1902 NewIntermedVals.push_back(
1903 std::make_pair(Shuffle, std::move(FinalIndices)));
1906 // If we had an odd number of defined values, then append the last
1907 // element to the array of new vectors.
1908 if ((IntermedVals.size() & 1) != 0)
1909 NewIntermedVals.push_back(IntermedVals.back());
1911 IntermedVals.swap(NewIntermedVals);
1914 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1915 "Invalid number of intermediate vectors");
1916 SDValue Vec1 = IntermedVals[0].first;
1918 if (IntermedVals.size() > 1)
1919 Vec2 = IntermedVals[1].first;
1921 Vec2 = DAG.getUNDEF(VT);
1923 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1924 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1925 ShuffleVec[IntermedVals[0].second[i]] = i;
1926 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1927 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1930 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1931 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1938 /// Expand a BUILD_VECTOR node on targets that don't
1939 /// support the operation, but do support the resultant vector type.
1940 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1941 unsigned NumElems = Node->getNumOperands();
1942 SDValue Value1, Value2;
1944 EVT VT = Node->getValueType(0);
1945 EVT OpVT = Node->getOperand(0).getValueType();
1946 EVT EltVT = VT.getVectorElementType();
1948 // If the only non-undef value is the low element, turn this into a
1949 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1950 bool isOnlyLowElement = true;
1951 bool MoreThanTwoValues = false;
1952 bool isConstant = true;
1953 for (unsigned i = 0; i < NumElems; ++i) {
1954 SDValue V = Node->getOperand(i);
1955 if (V.getOpcode() == ISD::UNDEF)
1958 isOnlyLowElement = false;
1959 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1962 if (!Value1.getNode()) {
1964 } else if (!Value2.getNode()) {
1967 } else if (V != Value1 && V != Value2) {
1968 MoreThanTwoValues = true;
1972 if (!Value1.getNode())
1973 return DAG.getUNDEF(VT);
1975 if (isOnlyLowElement)
1976 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1978 // If all elements are constants, create a load from the constant pool.
1980 SmallVector<Constant*, 16> CV;
1981 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1982 if (ConstantFPSDNode *V =
1983 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1984 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1985 } else if (ConstantSDNode *V =
1986 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1988 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1990 // If OpVT and EltVT don't match, EltVT is not legal and the
1991 // element values have been promoted/truncated earlier. Undo this;
1992 // we don't want a v16i8 to become a v16i32 for example.
1993 const ConstantInt *CI = V->getConstantIntValue();
1994 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1995 CI->getZExtValue()));
1998 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1999 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
2000 CV.push_back(UndefValue::get(OpNTy));
2003 Constant *CP = ConstantVector::get(CV);
2005 DAG.getConstantPool(CP, TLI.getPointerTy(DAG.getDataLayout()));
2006 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2007 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
2008 MachinePointerInfo::getConstantPool(),
2009 false, false, false, Alignment);
2012 SmallSet<SDValue, 16> DefinedValues;
2013 for (unsigned i = 0; i < NumElems; ++i) {
2014 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
2016 DefinedValues.insert(Node->getOperand(i));
2019 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
2020 if (!MoreThanTwoValues) {
2021 SmallVector<int, 8> ShuffleVec(NumElems, -1);
2022 for (unsigned i = 0; i < NumElems; ++i) {
2023 SDValue V = Node->getOperand(i);
2024 if (V.getOpcode() == ISD::UNDEF)
2026 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
2028 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
2029 // Get the splatted value into the low element of a vector register.
2030 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2032 if (Value2.getNode())
2033 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2035 Vec2 = DAG.getUNDEF(VT);
2037 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2038 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2042 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2047 // Otherwise, we can't handle this case efficiently.
2048 return ExpandVectorBuildThroughStack(Node);
2051 // Expand a node into a call to a libcall. If the result value
2052 // does not fit into a register, return the lo part and set the hi part to the
2053 // by-reg argument. If it does fit into a single register, return the result
2054 // and leave the Hi part unset.
2055 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2057 TargetLowering::ArgListTy Args;
2058 TargetLowering::ArgListEntry Entry;
2059 for (const SDValue &Op : Node->op_values()) {
2060 EVT ArgVT = Op.getValueType();
2061 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2064 Entry.isSExt = isSigned;
2065 Entry.isZExt = !isSigned;
2066 Args.push_back(Entry);
2068 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2069 TLI.getPointerTy(DAG.getDataLayout()));
2071 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2073 // By default, the input chain to this libcall is the entry node of the
2074 // function. If the libcall is going to be emitted as a tail call then
2075 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2076 // node which is being folded has a non-entry input chain.
2077 SDValue InChain = DAG.getEntryNode();
2079 // isTailCall may be true since the callee does not reference caller stack
2080 // frame. Check if it's in the right position.
2081 SDValue TCChain = InChain;
2082 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2086 TargetLowering::CallLoweringInfo CLI(DAG);
2087 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2088 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2089 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2091 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2093 if (!CallInfo.second.getNode())
2094 // It's a tailcall, return the chain (which is the DAG root).
2095 return DAG.getRoot();
2097 return CallInfo.first;
2100 /// Generate a libcall taking the given operands as arguments
2101 /// and returning a result of type RetVT.
2102 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2103 const SDValue *Ops, unsigned NumOps,
2104 bool isSigned, SDLoc dl) {
2105 TargetLowering::ArgListTy Args;
2106 Args.reserve(NumOps);
2108 TargetLowering::ArgListEntry Entry;
2109 for (unsigned i = 0; i != NumOps; ++i) {
2110 Entry.Node = Ops[i];
2111 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2112 Entry.isSExt = isSigned;
2113 Entry.isZExt = !isSigned;
2114 Args.push_back(Entry);
2116 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2117 TLI.getPointerTy(DAG.getDataLayout()));
2119 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2121 TargetLowering::CallLoweringInfo CLI(DAG);
2122 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2123 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2124 .setSExtResult(isSigned).setZExtResult(!isSigned);
2126 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2128 return CallInfo.first;
2131 // Expand a node into a call to a libcall. Similar to
2132 // ExpandLibCall except that the first operand is the in-chain.
2133 std::pair<SDValue, SDValue>
2134 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2137 SDValue InChain = Node->getOperand(0);
2139 TargetLowering::ArgListTy Args;
2140 TargetLowering::ArgListEntry Entry;
2141 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2142 EVT ArgVT = Node->getOperand(i).getValueType();
2143 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2144 Entry.Node = Node->getOperand(i);
2146 Entry.isSExt = isSigned;
2147 Entry.isZExt = !isSigned;
2148 Args.push_back(Entry);
2150 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2151 TLI.getPointerTy(DAG.getDataLayout()));
2153 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2155 TargetLowering::CallLoweringInfo CLI(DAG);
2156 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2157 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2158 .setSExtResult(isSigned).setZExtResult(!isSigned);
2160 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2165 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2166 RTLIB::Libcall Call_F32,
2167 RTLIB::Libcall Call_F64,
2168 RTLIB::Libcall Call_F80,
2169 RTLIB::Libcall Call_F128,
2170 RTLIB::Libcall Call_PPCF128) {
2172 switch (Node->getSimpleValueType(0).SimpleTy) {
2173 default: llvm_unreachable("Unexpected request for libcall!");
2174 case MVT::f32: LC = Call_F32; break;
2175 case MVT::f64: LC = Call_F64; break;
2176 case MVT::f80: LC = Call_F80; break;
2177 case MVT::f128: LC = Call_F128; break;
2178 case MVT::ppcf128: LC = Call_PPCF128; break;
2180 return ExpandLibCall(LC, Node, false);
2183 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2184 RTLIB::Libcall Call_I8,
2185 RTLIB::Libcall Call_I16,
2186 RTLIB::Libcall Call_I32,
2187 RTLIB::Libcall Call_I64,
2188 RTLIB::Libcall Call_I128) {
2190 switch (Node->getSimpleValueType(0).SimpleTy) {
2191 default: llvm_unreachable("Unexpected request for libcall!");
2192 case MVT::i8: LC = Call_I8; break;
2193 case MVT::i16: LC = Call_I16; break;
2194 case MVT::i32: LC = Call_I32; break;
2195 case MVT::i64: LC = Call_I64; break;
2196 case MVT::i128: LC = Call_I128; break;
2198 return ExpandLibCall(LC, Node, isSigned);
2201 /// Return true if divmod libcall is available.
2202 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2203 const TargetLowering &TLI) {
2205 switch (Node->getSimpleValueType(0).SimpleTy) {
2206 default: llvm_unreachable("Unexpected request for libcall!");
2207 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2208 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2209 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2210 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2211 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2214 return TLI.getLibcallName(LC) != nullptr;
2217 /// Only issue divrem libcall if both quotient and remainder are needed.
2218 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2219 // The other use might have been replaced with a divrem already.
2220 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2221 unsigned OtherOpcode = 0;
2223 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2225 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2227 SDValue Op0 = Node->getOperand(0);
2228 SDValue Op1 = Node->getOperand(1);
2229 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2230 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2234 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2235 User->getOperand(0) == Op0 &&
2236 User->getOperand(1) == Op1)
2242 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
2244 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2245 SmallVectorImpl<SDValue> &Results) {
2246 unsigned Opcode = Node->getOpcode();
2247 bool isSigned = Opcode == ISD::SDIVREM;
2250 switch (Node->getSimpleValueType(0).SimpleTy) {
2251 default: llvm_unreachable("Unexpected request for libcall!");
2252 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2253 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2254 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2255 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2256 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2259 // The input chain to this libcall is the entry node of the function.
2260 // Legalizing the call will automatically add the previous call to the
2262 SDValue InChain = DAG.getEntryNode();
2264 EVT RetVT = Node->getValueType(0);
2265 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2267 TargetLowering::ArgListTy Args;
2268 TargetLowering::ArgListEntry Entry;
2269 for (const SDValue &Op : Node->op_values()) {
2270 EVT ArgVT = Op.getValueType();
2271 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2274 Entry.isSExt = isSigned;
2275 Entry.isZExt = !isSigned;
2276 Args.push_back(Entry);
2279 // Also pass the return address of the remainder.
2280 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2282 Entry.Ty = RetTy->getPointerTo();
2283 Entry.isSExt = isSigned;
2284 Entry.isZExt = !isSigned;
2285 Args.push_back(Entry);
2287 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2288 TLI.getPointerTy(DAG.getDataLayout()));
2291 TargetLowering::CallLoweringInfo CLI(DAG);
2292 CLI.setDebugLoc(dl).setChain(InChain)
2293 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2294 .setSExtResult(isSigned).setZExtResult(!isSigned);
2296 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2298 // Remainder is loaded back from the stack frame.
2299 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2300 MachinePointerInfo(), false, false, false, 0);
2301 Results.push_back(CallInfo.first);
2302 Results.push_back(Rem);
2305 /// Return true if sincos libcall is available.
2306 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2308 switch (Node->getSimpleValueType(0).SimpleTy) {
2309 default: llvm_unreachable("Unexpected request for libcall!");
2310 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2311 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2312 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2313 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2314 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2316 return TLI.getLibcallName(LC) != nullptr;
2319 /// Return true if sincos libcall is available and can be used to combine sin
2321 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2322 const TargetMachine &TM) {
2323 if (!isSinCosLibcallAvailable(Node, TLI))
2325 // GNU sin/cos functions set errno while sincos does not. Therefore
2326 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2327 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2328 if (isGNU && !TM.Options.UnsafeFPMath)
2333 /// Only issue sincos libcall if both sin and cos are needed.
2334 static bool useSinCos(SDNode *Node) {
2335 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2336 ? ISD::FCOS : ISD::FSIN;
2338 SDValue Op0 = Node->getOperand(0);
2339 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2340 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2344 // The other user might have been turned into sincos already.
2345 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2351 /// Issue libcalls to sincos to compute sin / cos pairs.
2353 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2354 SmallVectorImpl<SDValue> &Results) {
2356 switch (Node->getSimpleValueType(0).SimpleTy) {
2357 default: llvm_unreachable("Unexpected request for libcall!");
2358 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2359 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2360 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2361 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2362 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2365 // The input chain to this libcall is the entry node of the function.
2366 // Legalizing the call will automatically add the previous call to the
2368 SDValue InChain = DAG.getEntryNode();
2370 EVT RetVT = Node->getValueType(0);
2371 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2373 TargetLowering::ArgListTy Args;
2374 TargetLowering::ArgListEntry Entry;
2376 // Pass the argument.
2377 Entry.Node = Node->getOperand(0);
2379 Entry.isSExt = false;
2380 Entry.isZExt = false;
2381 Args.push_back(Entry);
2383 // Pass the return address of sin.
2384 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2385 Entry.Node = SinPtr;
2386 Entry.Ty = RetTy->getPointerTo();
2387 Entry.isSExt = false;
2388 Entry.isZExt = false;
2389 Args.push_back(Entry);
2391 // Also pass the return address of the cos.
2392 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2393 Entry.Node = CosPtr;
2394 Entry.Ty = RetTy->getPointerTo();
2395 Entry.isSExt = false;
2396 Entry.isZExt = false;
2397 Args.push_back(Entry);
2399 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2400 TLI.getPointerTy(DAG.getDataLayout()));
2403 TargetLowering::CallLoweringInfo CLI(DAG);
2404 CLI.setDebugLoc(dl).setChain(InChain)
2405 .setCallee(TLI.getLibcallCallingConv(LC),
2406 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2408 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2410 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2411 MachinePointerInfo(), false, false, false, 0));
2412 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2413 MachinePointerInfo(), false, false, false, 0));
2416 /// This function is responsible for legalizing a
2417 /// INT_TO_FP operation of the specified operand when the target requests that
2418 /// we expand it. At this point, we know that the result and operand types are
2419 /// legal for the target.
2420 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2424 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2425 // simple 32-bit [signed|unsigned] integer to float/double expansion
2427 // Get the stack frame index of a 8 byte buffer.
2428 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2430 // word offset constant for Hi/Lo address computation
2431 SDValue WordOff = DAG.getConstant(sizeof(int), dl,
2432 StackSlot.getValueType());
2433 // set up Hi and Lo (into buffer) address based on endian
2434 SDValue Hi = StackSlot;
2435 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2436 StackSlot, WordOff);
2437 if (DAG.getDataLayout().isLittleEndian())
2440 // if signed map to unsigned space
2443 // constant used to invert sign bit (signed to unsigned mapping)
2444 SDValue SignBit = DAG.getConstant(0x80000000u, dl, MVT::i32);
2445 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2449 // store the lo of the constructed double - based on integer input
2450 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2451 Op0Mapped, Lo, MachinePointerInfo(),
2453 // initial hi portion of constructed double
2454 SDValue InitialHi = DAG.getConstant(0x43300000u, dl, MVT::i32);
2455 // store the hi of the constructed double - biased exponent
2456 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2457 MachinePointerInfo(),
2459 // load the constructed double
2460 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2461 MachinePointerInfo(), false, false, false, 0);
2462 // FP constant to bias correct the final result
2463 SDValue Bias = DAG.getConstantFP(isSigned ?
2464 BitsToDouble(0x4330000080000000ULL) :
2465 BitsToDouble(0x4330000000000000ULL),
2467 // subtract the bias
2468 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2471 // handle final rounding
2472 if (DestVT == MVT::f64) {
2475 } else if (DestVT.bitsLT(MVT::f64)) {
2476 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2477 DAG.getIntPtrConstant(0, dl));
2478 } else if (DestVT.bitsGT(MVT::f64)) {
2479 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2483 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2484 // Code below here assumes !isSigned without checking again.
2486 // Implementation of unsigned i64 to f64 following the algorithm in
2487 // __floatundidf in compiler_rt. This implementation has the advantage
2488 // of performing rounding correctly, both in the default rounding mode
2489 // and in all alternate rounding modes.
2490 // TODO: Generalize this for use with other types.
2491 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2493 DAG.getConstant(UINT64_C(0x4330000000000000), dl, MVT::i64);
2494 SDValue TwoP84PlusTwoP52 =
2495 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), dl,
2498 DAG.getConstant(UINT64_C(0x4530000000000000), dl, MVT::i64);
2500 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2501 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2502 DAG.getConstant(32, dl, MVT::i64));
2503 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2504 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2505 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2506 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2507 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2509 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2512 // Implementation of unsigned i64 to f32.
2513 // TODO: Generalize this for use with other types.
2514 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2515 // For unsigned conversions, convert them to signed conversions using the
2516 // algorithm from the x86_64 __floatundidf in compiler_rt.
2518 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2520 SDValue ShiftConst =
2521 DAG.getConstant(1, dl, TLI.getShiftAmountTy(Op0.getValueType()));
2522 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2523 SDValue AndConst = DAG.getConstant(1, dl, MVT::i64);
2524 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2525 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2527 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2528 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2530 // TODO: This really should be implemented using a branch rather than a
2531 // select. We happen to get lucky and machinesink does the right
2532 // thing most of the time. This would be a good candidate for a
2533 //pseudo-op, or, even better, for whole-function isel.
2534 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2535 Op0, DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
2536 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2539 // Otherwise, implement the fully general conversion.
2541 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2542 DAG.getConstant(UINT64_C(0xfffffffffffff800), dl, MVT::i64));
2543 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2544 DAG.getConstant(UINT64_C(0x800), dl, MVT::i64));
2545 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2546 DAG.getConstant(UINT64_C(0x7ff), dl, MVT::i64));
2547 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), And2,
2548 DAG.getConstant(UINT64_C(0), dl, MVT::i64),
2550 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2551 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), Op0,
2552 DAG.getConstant(UINT64_C(0x0020000000000000), dl,
2555 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2556 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2558 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2559 DAG.getConstant(32, dl, SHVT));
2560 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2561 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2563 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), dl,
2565 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2566 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2567 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2568 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2569 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2570 DAG.getIntPtrConstant(0, dl));
2573 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2575 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2577 DAG.getConstant(0, dl, Op0.getValueType()),
2579 SDValue Zero = DAG.getIntPtrConstant(0, dl),
2580 Four = DAG.getIntPtrConstant(4, dl);
2581 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2582 SignSet, Four, Zero);
2584 // If the sign bit of the integer is set, the large number will be treated
2585 // as a negative number. To counteract this, the dynamic code adds an
2586 // offset depending on the data type.
2588 switch (Op0.getSimpleValueType().SimpleTy) {
2589 default: llvm_unreachable("Unsupported integer type!");
2590 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2591 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2592 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2593 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2595 if (DAG.getDataLayout().isLittleEndian())
2597 Constant *FudgeFactor = ConstantInt::get(
2598 Type::getInt64Ty(*DAG.getContext()), FF);
2601 DAG.getConstantPool(FudgeFactor, TLI.getPointerTy(DAG.getDataLayout()));
2602 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2603 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2604 Alignment = std::min(Alignment, 4u);
2606 if (DestVT == MVT::f32)
2607 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2608 MachinePointerInfo::getConstantPool(),
2609 false, false, false, Alignment);
2611 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2612 DAG.getEntryNode(), CPIdx,
2613 MachinePointerInfo::getConstantPool(),
2614 MVT::f32, false, false, false, Alignment);
2615 HandleSDNode Handle(Load);
2616 LegalizeOp(Load.getNode());
2617 FudgeInReg = Handle.getValue();
2620 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2623 /// This function is responsible for legalizing a
2624 /// *INT_TO_FP operation of the specified operand when the target requests that
2625 /// we promote it. At this point, we know that the result and operand types are
2626 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2627 /// operation that takes a larger input.
2628 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2632 // First step, figure out the appropriate *INT_TO_FP operation to use.
2633 EVT NewInTy = LegalOp.getValueType();
2635 unsigned OpToUse = 0;
2637 // Scan for the appropriate larger type to use.
2639 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2640 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2642 // If the target supports SINT_TO_FP of this type, use it.
2643 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2644 OpToUse = ISD::SINT_TO_FP;
2647 if (isSigned) continue;
2649 // If the target supports UINT_TO_FP of this type, use it.
2650 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2651 OpToUse = ISD::UINT_TO_FP;
2655 // Otherwise, try a larger type.
2658 // Okay, we found the operation and type to use. Zero extend our input to the
2659 // desired type then run the operation on it.
2660 return DAG.getNode(OpToUse, dl, DestVT,
2661 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2662 dl, NewInTy, LegalOp));
2665 /// This function is responsible for legalizing a
2666 /// FP_TO_*INT operation of the specified operand when the target requests that
2667 /// we promote it. At this point, we know that the result and operand types are
2668 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2669 /// operation that returns a larger result.
2670 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2674 // First step, figure out the appropriate FP_TO*INT operation to use.
2675 EVT NewOutTy = DestVT;
2677 unsigned OpToUse = 0;
2679 // Scan for the appropriate larger type to use.
2681 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2682 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2684 // A larger signed type can hold all unsigned values of the requested type,
2685 // so using FP_TO_SINT is valid
2686 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2687 OpToUse = ISD::FP_TO_SINT;
2691 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2692 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2693 OpToUse = ISD::FP_TO_UINT;
2697 // Otherwise, try a larger type.
2701 // Okay, we found the operation and type to use.
2702 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2704 // Truncate the result of the extended FP_TO_*INT operation to the desired
2706 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2709 /// Open code the operations for BSWAP of the specified operation.
2710 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2711 EVT VT = Op.getValueType();
2712 EVT SHVT = TLI.getShiftAmountTy(VT);
2713 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2714 switch (VT.getSimpleVT().SimpleTy) {
2715 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2717 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2718 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2719 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2721 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2722 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2723 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2724 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2725 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2726 DAG.getConstant(0xFF0000, dl, VT));
2727 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT));
2728 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2729 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2730 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2732 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2733 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2734 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2735 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2736 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2737 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2738 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2739 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2740 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7,
2741 DAG.getConstant(255ULL<<48, dl, VT));
2742 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6,
2743 DAG.getConstant(255ULL<<40, dl, VT));
2744 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5,
2745 DAG.getConstant(255ULL<<32, dl, VT));
2746 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4,
2747 DAG.getConstant(255ULL<<24, dl, VT));
2748 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2749 DAG.getConstant(255ULL<<16, dl, VT));
2750 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2,
2751 DAG.getConstant(255ULL<<8 , dl, VT));
2752 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2753 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2754 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2755 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2756 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2757 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2758 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2762 /// Expand the specified bitcount instruction into operations.
2763 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2766 default: llvm_unreachable("Cannot expand this yet!");
2768 EVT VT = Op.getValueType();
2769 EVT ShVT = TLI.getShiftAmountTy(VT);
2770 unsigned Len = VT.getSizeInBits();
2772 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2773 "CTPOP not implemented for this type.");
2775 // This is the "best" algorithm from
2776 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2778 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)),
2780 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)),
2782 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)),
2784 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)),
2787 // v = v - ((v >> 1) & 0x55555555...)
2788 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2789 DAG.getNode(ISD::AND, dl, VT,
2790 DAG.getNode(ISD::SRL, dl, VT, Op,
2791 DAG.getConstant(1, dl, ShVT)),
2793 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2794 Op = DAG.getNode(ISD::ADD, dl, VT,
2795 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2796 DAG.getNode(ISD::AND, dl, VT,
2797 DAG.getNode(ISD::SRL, dl, VT, Op,
2798 DAG.getConstant(2, dl, ShVT)),
2800 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2801 Op = DAG.getNode(ISD::AND, dl, VT,
2802 DAG.getNode(ISD::ADD, dl, VT, Op,
2803 DAG.getNode(ISD::SRL, dl, VT, Op,
2804 DAG.getConstant(4, dl, ShVT))),
2806 // v = (v * 0x01010101...) >> (Len - 8)
2807 Op = DAG.getNode(ISD::SRL, dl, VT,
2808 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2809 DAG.getConstant(Len - 8, dl, ShVT));
2813 case ISD::CTLZ_ZERO_UNDEF:
2814 // This trivially expands to CTLZ.
2815 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2817 // for now, we do this:
2818 // x = x | (x >> 1);
2819 // x = x | (x >> 2);
2821 // x = x | (x >>16);
2822 // x = x | (x >>32); // for 64-bit input
2823 // return popcount(~x);
2825 // Ref: "Hacker's Delight" by Henry Warren
2826 EVT VT = Op.getValueType();
2827 EVT ShVT = TLI.getShiftAmountTy(VT);
2828 unsigned len = VT.getSizeInBits();
2829 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2830 SDValue Tmp3 = DAG.getConstant(1ULL << i, dl, ShVT);
2831 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2832 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2834 Op = DAG.getNOT(dl, Op, VT);
2835 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2837 case ISD::CTTZ_ZERO_UNDEF:
2838 // This trivially expands to CTTZ.
2839 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2841 // for now, we use: { return popcount(~x & (x - 1)); }
2842 // unless the target has ctlz but not ctpop, in which case we use:
2843 // { return 32 - nlz(~x & (x-1)); }
2844 // Ref: "Hacker's Delight" by Henry Warren
2845 EVT VT = Op.getValueType();
2846 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2847 DAG.getNOT(dl, Op, VT),
2848 DAG.getNode(ISD::SUB, dl, VT, Op,
2849 DAG.getConstant(1, dl, VT)));
2850 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2851 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2852 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2853 return DAG.getNode(ISD::SUB, dl, VT,
2854 DAG.getConstant(VT.getSizeInBits(), dl, VT),
2855 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2856 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2861 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2862 unsigned Opc = Node->getOpcode();
2863 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2864 RTLIB::Libcall LC = RTLIB::getATOMIC(Opc, VT);
2865 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!");
2867 return ExpandChainLibCall(LC, Node, false);
2870 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2871 SmallVector<SDValue, 8> Results;
2873 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2875 switch (Node->getOpcode()) {
2878 case ISD::CTLZ_ZERO_UNDEF:
2880 case ISD::CTTZ_ZERO_UNDEF:
2881 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2882 Results.push_back(Tmp1);
2885 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2887 case ISD::FRAMEADDR:
2888 case ISD::RETURNADDR:
2889 case ISD::FRAME_TO_ARGS_OFFSET:
2890 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0)));
2892 case ISD::FLT_ROUNDS_:
2893 Results.push_back(DAG.getConstant(1, dl, Node->getValueType(0)));
2895 case ISD::EH_RETURN:
2899 case ISD::EH_SJLJ_LONGJMP:
2900 // If the target didn't expand these, there's nothing to do, so just
2901 // preserve the chain and be done.
2902 Results.push_back(Node->getOperand(0));
2904 case ISD::EH_SJLJ_SETJMP:
2905 // If the target didn't expand this, just return 'zero' and preserve the
2907 Results.push_back(DAG.getConstant(0, dl, MVT::i32));
2908 Results.push_back(Node->getOperand(0));
2910 case ISD::ATOMIC_FENCE: {
2911 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2912 // FIXME: handle "fence singlethread" more efficiently.
2913 TargetLowering::ArgListTy Args;
2915 TargetLowering::CallLoweringInfo CLI(DAG);
2917 .setChain(Node->getOperand(0))
2918 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
2919 DAG.getExternalSymbol("__sync_synchronize",
2920 TLI.getPointerTy(DAG.getDataLayout())),
2921 std::move(Args), 0);
2923 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2925 Results.push_back(CallResult.second);
2928 case ISD::ATOMIC_LOAD: {
2929 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2930 SDValue Zero = DAG.getConstant(0, dl, Node->getValueType(0));
2931 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
2932 SDValue Swap = DAG.getAtomicCmpSwap(
2933 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
2934 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
2935 cast<AtomicSDNode>(Node)->getMemOperand(),
2936 cast<AtomicSDNode>(Node)->getOrdering(),
2937 cast<AtomicSDNode>(Node)->getOrdering(),
2938 cast<AtomicSDNode>(Node)->getSynchScope());
2939 Results.push_back(Swap.getValue(0));
2940 Results.push_back(Swap.getValue(1));
2943 case ISD::ATOMIC_STORE: {
2944 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2945 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2946 cast<AtomicSDNode>(Node)->getMemoryVT(),
2947 Node->getOperand(0),
2948 Node->getOperand(1), Node->getOperand(2),
2949 cast<AtomicSDNode>(Node)->getMemOperand(),
2950 cast<AtomicSDNode>(Node)->getOrdering(),
2951 cast<AtomicSDNode>(Node)->getSynchScope());
2952 Results.push_back(Swap.getValue(1));
2955 // By default, atomic intrinsics are marked Legal and lowered. Targets
2956 // which don't support them directly, however, may want libcalls, in which
2957 // case they mark them Expand, and we get here.
2958 case ISD::ATOMIC_SWAP:
2959 case ISD::ATOMIC_LOAD_ADD:
2960 case ISD::ATOMIC_LOAD_SUB:
2961 case ISD::ATOMIC_LOAD_AND:
2962 case ISD::ATOMIC_LOAD_OR:
2963 case ISD::ATOMIC_LOAD_XOR:
2964 case ISD::ATOMIC_LOAD_NAND:
2965 case ISD::ATOMIC_LOAD_MIN:
2966 case ISD::ATOMIC_LOAD_MAX:
2967 case ISD::ATOMIC_LOAD_UMIN:
2968 case ISD::ATOMIC_LOAD_UMAX:
2969 case ISD::ATOMIC_CMP_SWAP: {
2970 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
2971 Results.push_back(Tmp.first);
2972 Results.push_back(Tmp.second);
2975 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
2976 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
2977 // splits out the success value as a comparison. Expanding the resulting
2978 // ATOMIC_CMP_SWAP will produce a libcall.
2979 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
2980 SDValue Res = DAG.getAtomicCmpSwap(
2981 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
2982 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
2983 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
2984 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
2985 cast<AtomicSDNode>(Node)->getFailureOrdering(),
2986 cast<AtomicSDNode>(Node)->getSynchScope());
2988 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
2989 Res, Node->getOperand(2), ISD::SETEQ);
2991 Results.push_back(Res.getValue(0));
2992 Results.push_back(Success);
2993 Results.push_back(Res.getValue(1));
2996 case ISD::DYNAMIC_STACKALLOC:
2997 ExpandDYNAMIC_STACKALLOC(Node, Results);
2999 case ISD::MERGE_VALUES:
3000 for (unsigned i = 0; i < Node->getNumValues(); i++)
3001 Results.push_back(Node->getOperand(i));
3004 EVT VT = Node->getValueType(0);
3006 Results.push_back(DAG.getConstant(0, dl, VT));
3008 assert(VT.isFloatingPoint() && "Unknown value type!");
3009 Results.push_back(DAG.getConstantFP(0, dl, VT));
3014 // If this operation is not supported, lower it to 'abort()' call
3015 TargetLowering::ArgListTy Args;
3016 TargetLowering::CallLoweringInfo CLI(DAG);
3018 .setChain(Node->getOperand(0))
3019 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3020 DAG.getExternalSymbol("abort",
3021 TLI.getPointerTy(DAG.getDataLayout())),
3022 std::move(Args), 0);
3023 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3025 Results.push_back(CallResult.second);
3030 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3031 Node->getValueType(0), dl);
3032 Results.push_back(Tmp1);
3034 case ISD::FP_EXTEND:
3035 Tmp1 = EmitStackConvert(Node->getOperand(0),
3036 Node->getOperand(0).getValueType(),
3037 Node->getValueType(0), dl);
3038 Results.push_back(Tmp1);
3040 case ISD::SIGN_EXTEND_INREG: {
3041 // NOTE: we could fall back on load/store here too for targets without
3042 // SAR. However, it is doubtful that any exist.
3043 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3044 EVT VT = Node->getValueType(0);
3045 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3048 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3049 ExtraVT.getScalarType().getSizeInBits();
3050 SDValue ShiftCst = DAG.getConstant(BitsDiff, dl, ShiftAmountTy);
3051 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3052 Node->getOperand(0), ShiftCst);
3053 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3054 Results.push_back(Tmp1);
3057 case ISD::FP_ROUND_INREG: {
3058 // The only way we can lower this is to turn it into a TRUNCSTORE,
3059 // EXTLOAD pair, targeting a temporary location (a stack slot).
3061 // NOTE: there is a choice here between constantly creating new stack
3062 // slots and always reusing the same one. We currently always create
3063 // new ones, as reuse may inhibit scheduling.
3064 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3065 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3066 Node->getValueType(0), dl);
3067 Results.push_back(Tmp1);
3070 case ISD::SINT_TO_FP:
3071 case ISD::UINT_TO_FP:
3072 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3073 Node->getOperand(0), Node->getValueType(0), dl);
3074 Results.push_back(Tmp1);
3076 case ISD::FP_TO_SINT:
3077 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3078 Results.push_back(Tmp1);
3080 case ISD::FP_TO_UINT: {
3081 SDValue True, False;
3082 EVT VT = Node->getOperand(0).getValueType();
3083 EVT NVT = Node->getValueType(0);
3084 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3085 APInt::getNullValue(VT.getSizeInBits()));
3086 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3087 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3088 Tmp1 = DAG.getConstantFP(apf, dl, VT);
3089 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3090 Node->getOperand(0),
3092 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3093 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3094 DAG.getNode(ISD::FSUB, dl, VT,
3095 Node->getOperand(0), Tmp1));
3096 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3097 DAG.getConstant(x, dl, NVT));
3098 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3099 Results.push_back(Tmp1);
3103 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3104 EVT VT = Node->getValueType(0);
3105 Tmp1 = Node->getOperand(0);
3106 Tmp2 = Node->getOperand(1);
3107 unsigned Align = Node->getConstantOperandVal(3);
3109 SDValue VAListLoad =
3110 DAG.getLoad(TLI.getPointerTy(DAG.getDataLayout()), dl, Tmp1, Tmp2,
3111 MachinePointerInfo(V), false, false, false, 0);
3112 SDValue VAList = VAListLoad;
3114 if (Align > TLI.getMinStackArgumentAlignment()) {
3115 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3117 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3118 DAG.getConstant(Align - 1, dl,
3119 VAList.getValueType()));
3121 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3122 DAG.getConstant(-(int64_t)Align, dl,
3123 VAList.getValueType()));
3126 // Increment the pointer, VAList, to the next vaarg
3127 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3128 DAG.getConstant(DAG.getDataLayout().getTypeAllocSize(
3129 VT.getTypeForEVT(*DAG.getContext())),
3130 dl, VAList.getValueType()));
3131 // Store the incremented VAList to the legalized pointer
3132 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3133 MachinePointerInfo(V), false, false, 0);
3134 // Load the actual argument out of the pointer VAList
3135 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3136 false, false, false, 0));
3137 Results.push_back(Results[0].getValue(1));
3141 // This defaults to loading a pointer from the input and storing it to the
3142 // output, returning the chain.
3143 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3144 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3145 Tmp1 = DAG.getLoad(TLI.getPointerTy(DAG.getDataLayout()), dl,
3146 Node->getOperand(0), Node->getOperand(2),
3147 MachinePointerInfo(VS), false, false, false, 0);
3148 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3149 MachinePointerInfo(VD), false, false, 0);
3150 Results.push_back(Tmp1);
3153 case ISD::EXTRACT_VECTOR_ELT:
3154 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3155 // This must be an access of the only element. Return it.
3156 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3157 Node->getOperand(0));
3159 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3160 Results.push_back(Tmp1);
3162 case ISD::EXTRACT_SUBVECTOR:
3163 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3165 case ISD::INSERT_SUBVECTOR:
3166 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3168 case ISD::CONCAT_VECTORS: {
3169 Results.push_back(ExpandVectorBuildThroughStack(Node));
3172 case ISD::SCALAR_TO_VECTOR:
3173 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3175 case ISD::INSERT_VECTOR_ELT:
3176 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3177 Node->getOperand(1),
3178 Node->getOperand(2), dl));
3180 case ISD::VECTOR_SHUFFLE: {
3181 SmallVector<int, 32> NewMask;
3182 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3184 EVT VT = Node->getValueType(0);
3185 EVT EltVT = VT.getVectorElementType();
3186 SDValue Op0 = Node->getOperand(0);
3187 SDValue Op1 = Node->getOperand(1);
3188 if (!TLI.isTypeLegal(EltVT)) {
3190 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3192 // BUILD_VECTOR operands are allowed to be wider than the element type.
3193 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3195 if (NewEltVT.bitsLT(EltVT)) {
3197 // Convert shuffle node.
3198 // If original node was v4i64 and the new EltVT is i32,
3199 // cast operands to v8i32 and re-build the mask.
3201 // Calculate new VT, the size of the new VT should be equal to original.
3203 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3204 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3205 assert(NewVT.bitsEq(VT));
3207 // cast operands to new VT
3208 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3209 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3211 // Convert the shuffle mask
3212 unsigned int factor =
3213 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3215 // EltVT gets smaller
3218 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3220 for (unsigned fi = 0; fi < factor; ++fi)
3221 NewMask.push_back(Mask[i]);
3224 for (unsigned fi = 0; fi < factor; ++fi)
3225 NewMask.push_back(Mask[i]*factor+fi);
3233 unsigned NumElems = VT.getVectorNumElements();
3234 SmallVector<SDValue, 16> Ops;
3235 for (unsigned i = 0; i != NumElems; ++i) {
3237 Ops.push_back(DAG.getUNDEF(EltVT));
3240 unsigned Idx = Mask[i];
3242 Ops.push_back(DAG.getNode(
3243 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op0,
3244 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))));
3246 Ops.push_back(DAG.getNode(
3247 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op1,
3248 DAG.getConstant(Idx - NumElems, dl,
3249 TLI.getVectorIdxTy(DAG.getDataLayout()))));
3252 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3253 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3254 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3255 Results.push_back(Tmp1);
3258 case ISD::EXTRACT_ELEMENT: {
3259 EVT OpTy = Node->getOperand(0).getValueType();
3260 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3262 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3263 DAG.getConstant(OpTy.getSizeInBits()/2, dl,
3264 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3265 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3268 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3269 Node->getOperand(0));
3271 Results.push_back(Tmp1);
3274 case ISD::STACKSAVE:
3275 // Expand to CopyFromReg if the target set
3276 // StackPointerRegisterToSaveRestore.
3277 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3278 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3279 Node->getValueType(0)));
3280 Results.push_back(Results[0].getValue(1));
3282 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3283 Results.push_back(Node->getOperand(0));
3286 case ISD::STACKRESTORE:
3287 // Expand to CopyToReg if the target set
3288 // StackPointerRegisterToSaveRestore.
3289 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3290 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3291 Node->getOperand(1)));
3293 Results.push_back(Node->getOperand(0));
3296 case ISD::FCOPYSIGN:
3297 Results.push_back(ExpandFCOPYSIGN(Node));
3300 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3301 Tmp1 = DAG.getConstantFP(-0.0, dl, Node->getValueType(0));
3302 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3303 Node->getOperand(0));
3304 Results.push_back(Tmp1);
3307 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3308 EVT VT = Node->getValueType(0);
3309 Tmp1 = Node->getOperand(0);
3310 Tmp2 = DAG.getConstantFP(0.0, dl, VT);
3311 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3312 Tmp1, Tmp2, ISD::SETUGT);
3313 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3314 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3315 Results.push_back(Tmp1);
3322 // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B
3324 switch (Node->getOpcode()) {
3325 default: llvm_unreachable("How did we get here?");
3326 case ISD::SMAX: Pred = ISD::SETGT; break;
3327 case ISD::SMIN: Pred = ISD::SETLT; break;
3328 case ISD::UMAX: Pred = ISD::SETUGT; break;
3329 case ISD::UMIN: Pred = ISD::SETULT; break;
3331 Tmp1 = Node->getOperand(0);
3332 Tmp2 = Node->getOperand(1);
3333 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp1, Tmp2, Pred);
3334 Results.push_back(Tmp1);
3339 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
3340 RTLIB::FMIN_F80, RTLIB::FMIN_F128,
3341 RTLIB::FMIN_PPCF128));
3344 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
3345 RTLIB::FMAX_F80, RTLIB::FMAX_F128,
3346 RTLIB::FMAX_PPCF128));
3349 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3350 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3351 RTLIB::SQRT_PPCF128));
3355 EVT VT = Node->getValueType(0);
3356 bool isSIN = Node->getOpcode() == ISD::FSIN;
3357 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3358 // fcos which share the same operand and both are used.
3359 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3360 canCombineSinCosLibcall(Node, TLI, TM))
3361 && useSinCos(Node)) {
3362 SDVTList VTs = DAG.getVTList(VT, VT);
3363 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3365 Tmp1 = Tmp1.getValue(1);
3366 Results.push_back(Tmp1);
3368 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3369 RTLIB::SIN_F80, RTLIB::SIN_F128,
3370 RTLIB::SIN_PPCF128));
3372 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3373 RTLIB::COS_F80, RTLIB::COS_F128,
3374 RTLIB::COS_PPCF128));
3379 // Expand into sincos libcall.
3380 ExpandSinCosLibCall(Node, Results);
3383 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3384 RTLIB::LOG_F80, RTLIB::LOG_F128,
3385 RTLIB::LOG_PPCF128));
3388 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3389 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3390 RTLIB::LOG2_PPCF128));
3393 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3394 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3395 RTLIB::LOG10_PPCF128));
3398 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3399 RTLIB::EXP_F80, RTLIB::EXP_F128,
3400 RTLIB::EXP_PPCF128));
3403 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3404 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3405 RTLIB::EXP2_PPCF128));
3408 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3409 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3410 RTLIB::TRUNC_PPCF128));
3413 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3414 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3415 RTLIB::FLOOR_PPCF128));
3418 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3419 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3420 RTLIB::CEIL_PPCF128));
3423 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3424 RTLIB::RINT_F80, RTLIB::RINT_F128,
3425 RTLIB::RINT_PPCF128));
3427 case ISD::FNEARBYINT:
3428 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3429 RTLIB::NEARBYINT_F64,
3430 RTLIB::NEARBYINT_F80,
3431 RTLIB::NEARBYINT_F128,
3432 RTLIB::NEARBYINT_PPCF128));
3435 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3439 RTLIB::ROUND_PPCF128));
3442 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3443 RTLIB::POWI_F80, RTLIB::POWI_F128,
3444 RTLIB::POWI_PPCF128));
3447 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3448 RTLIB::POW_F80, RTLIB::POW_F128,
3449 RTLIB::POW_PPCF128));
3452 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3453 RTLIB::DIV_F80, RTLIB::DIV_F128,
3454 RTLIB::DIV_PPCF128));
3457 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3458 RTLIB::REM_F80, RTLIB::REM_F128,
3459 RTLIB::REM_PPCF128));
3462 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3463 RTLIB::FMA_F80, RTLIB::FMA_F128,
3464 RTLIB::FMA_PPCF128));
3467 llvm_unreachable("Illegal fmad should never be formed");
3470 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
3471 RTLIB::ADD_F80, RTLIB::ADD_F128,
3472 RTLIB::ADD_PPCF128));
3475 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
3476 RTLIB::MUL_F80, RTLIB::MUL_F128,
3477 RTLIB::MUL_PPCF128));
3479 case ISD::FP16_TO_FP: {
3480 if (Node->getValueType(0) == MVT::f32) {
3481 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3485 // We can extend to types bigger than f32 in two steps without changing the
3486 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3487 // the option of emitting that before resorting to a libcall.
3489 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3491 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3494 case ISD::FP_TO_FP16: {
3495 if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) {
3496 SDValue Op = Node->getOperand(0);
3497 MVT SVT = Op.getSimpleValueType();
3498 if ((SVT == MVT::f64 || SVT == MVT::f80) &&
3499 TLI.isOperationLegalOrCustom(ISD::FP_TO_FP16, MVT::f32)) {
3500 // Under fastmath, we can expand this node into a fround followed by
3501 // a float-half conversion.
3502 SDValue FloatVal = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Op,
3503 DAG.getIntPtrConstant(0, dl));
3505 DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, FloatVal));
3511 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3512 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3513 Results.push_back(ExpandLibCall(LC, Node, false));
3516 case ISD::ConstantFP: {
3517 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3518 // Check to see if this FP immediate is already legal.
3519 // If this is a legal constant, turn it into a TargetConstantFP node.
3520 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3521 Results.push_back(ExpandConstantFP(CFP, true));
3525 EVT VT = Node->getValueType(0);
3526 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3527 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3528 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3529 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3530 Results.push_back(Tmp1);
3532 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
3533 RTLIB::SUB_F80, RTLIB::SUB_F128,
3534 RTLIB::SUB_PPCF128));
3539 EVT VT = Node->getValueType(0);
3540 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3541 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3542 "Don't know how to expand this subtraction!");
3543 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3544 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
3546 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, dl, VT));
3547 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3552 EVT VT = Node->getValueType(0);
3553 bool isSigned = Node->getOpcode() == ISD::SREM;
3554 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3555 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3556 Tmp2 = Node->getOperand(0);
3557 Tmp3 = Node->getOperand(1);
3558 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3559 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3560 // If div is legal, it's better to do the normal expansion
3561 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3562 useDivRem(Node, isSigned, false))) {
3563 SDVTList VTs = DAG.getVTList(VT, VT);
3564 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3565 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3567 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3568 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3569 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3570 } else if (isSigned)
3571 Tmp1 = ExpandIntLibCall(Node, true,
3573 RTLIB::SREM_I16, RTLIB::SREM_I32,
3574 RTLIB::SREM_I64, RTLIB::SREM_I128);
3576 Tmp1 = ExpandIntLibCall(Node, false,
3578 RTLIB::UREM_I16, RTLIB::UREM_I32,
3579 RTLIB::UREM_I64, RTLIB::UREM_I128);
3580 Results.push_back(Tmp1);
3585 bool isSigned = Node->getOpcode() == ISD::SDIV;
3586 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3587 EVT VT = Node->getValueType(0);
3588 SDVTList VTs = DAG.getVTList(VT, VT);
3589 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3590 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3591 useDivRem(Node, isSigned, true)))
3592 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3593 Node->getOperand(1));
3595 Tmp1 = ExpandIntLibCall(Node, true,
3597 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3598 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3600 Tmp1 = ExpandIntLibCall(Node, false,
3602 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3603 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3604 Results.push_back(Tmp1);
3609 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3611 EVT VT = Node->getValueType(0);
3612 SDVTList VTs = DAG.getVTList(VT, VT);
3613 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3614 "If this wasn't legal, it shouldn't have been created!");
3615 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3616 Node->getOperand(1));
3617 Results.push_back(Tmp1.getValue(1));
3622 // Expand into divrem libcall
3623 ExpandDivRemLibCall(Node, Results);
3626 EVT VT = Node->getValueType(0);
3627 SDVTList VTs = DAG.getVTList(VT, VT);
3628 // See if multiply or divide can be lowered using two-result operations.
3629 // We just need the low half of the multiply; try both the signed
3630 // and unsigned forms. If the target supports both SMUL_LOHI and
3631 // UMUL_LOHI, form a preference by checking which forms of plain
3632 // MULH it supports.
3633 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3634 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3635 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3636 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3637 unsigned OpToUse = 0;
3638 if (HasSMUL_LOHI && !HasMULHS) {
3639 OpToUse = ISD::SMUL_LOHI;
3640 } else if (HasUMUL_LOHI && !HasMULHU) {
3641 OpToUse = ISD::UMUL_LOHI;
3642 } else if (HasSMUL_LOHI) {
3643 OpToUse = ISD::SMUL_LOHI;
3644 } else if (HasUMUL_LOHI) {
3645 OpToUse = ISD::UMUL_LOHI;
3648 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3649 Node->getOperand(1)));
3654 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3655 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3656 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3657 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3658 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3659 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) {
3660 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3661 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3662 SDValue Shift = DAG.getConstant(HalfType.getSizeInBits(), dl,
3663 TLI.getShiftAmountTy(HalfType));
3664 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3665 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3669 Tmp1 = ExpandIntLibCall(Node, false,
3671 RTLIB::MUL_I16, RTLIB::MUL_I32,
3672 RTLIB::MUL_I64, RTLIB::MUL_I128);
3673 Results.push_back(Tmp1);
3678 SDValue LHS = Node->getOperand(0);
3679 SDValue RHS = Node->getOperand(1);
3680 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3681 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3683 Results.push_back(Sum);
3684 EVT ResultType = Node->getValueType(1);
3685 EVT OType = getSetCCResultType(Node->getValueType(0));
3687 SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType());
3689 // LHSSign -> LHS >= 0
3690 // RHSSign -> RHS >= 0
3691 // SumSign -> Sum >= 0
3694 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3696 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3698 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3699 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3700 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3701 Node->getOpcode() == ISD::SADDO ?
3702 ISD::SETEQ : ISD::SETNE);
3704 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3705 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3707 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3708 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3713 SDValue LHS = Node->getOperand(0);
3714 SDValue RHS = Node->getOperand(1);
3715 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3716 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3718 Results.push_back(Sum);
3720 EVT ResultType = Node->getValueType(1);
3721 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3723 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3724 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3726 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3731 EVT VT = Node->getValueType(0);
3732 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3733 SDValue LHS = Node->getOperand(0);
3734 SDValue RHS = Node->getOperand(1);
3737 static const unsigned Ops[2][3] =
3738 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3739 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3740 bool isSigned = Node->getOpcode() == ISD::SMULO;
3741 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3742 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3743 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3744 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3745 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3747 TopHalf = BottomHalf.getValue(1);
3748 } else if (TLI.isTypeLegal(WideVT)) {
3749 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3750 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3751 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3752 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3753 DAG.getIntPtrConstant(0, dl));
3754 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3755 DAG.getIntPtrConstant(1, dl));
3757 // We can fall back to a libcall with an illegal type for the MUL if we
3758 // have a libcall big enough.
3759 // Also, we can fall back to a division in some cases, but that's a big
3760 // performance hit in the general case.
3761 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3762 if (WideVT == MVT::i16)
3763 LC = RTLIB::MUL_I16;
3764 else if (WideVT == MVT::i32)
3765 LC = RTLIB::MUL_I32;
3766 else if (WideVT == MVT::i64)
3767 LC = RTLIB::MUL_I64;
3768 else if (WideVT == MVT::i128)
3769 LC = RTLIB::MUL_I128;
3770 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3772 // The high part is obtained by SRA'ing all but one of the bits of low
3774 unsigned LoSize = VT.getSizeInBits();
3776 DAG.getNode(ISD::SRA, dl, VT, RHS,
3777 DAG.getConstant(LoSize - 1, dl,
3778 TLI.getPointerTy(DAG.getDataLayout())));
3780 DAG.getNode(ISD::SRA, dl, VT, LHS,
3781 DAG.getConstant(LoSize - 1, dl,
3782 TLI.getPointerTy(DAG.getDataLayout())));
3784 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3785 // pre-lowered to the correct types. This all depends upon WideVT not
3786 // being a legal type for the architecture and thus has to be split to
3788 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3789 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3790 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3791 DAG.getIntPtrConstant(0, dl));
3792 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3793 DAG.getIntPtrConstant(1, dl));
3794 // Ret is a node with an illegal type. Because such things are not
3795 // generally permitted during this phase of legalization, make sure the
3796 // node has no more uses. The above EXTRACT_ELEMENT nodes should have been
3798 assert(Ret->use_empty() &&
3799 "Unexpected uses of illegally type from expanded lib call.");
3803 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1, dl,
3804 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3805 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3806 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3809 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3810 DAG.getConstant(0, dl, VT), ISD::SETNE);
3812 Results.push_back(BottomHalf);
3813 Results.push_back(TopHalf);
3816 case ISD::BUILD_PAIR: {
3817 EVT PairTy = Node->getValueType(0);
3818 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3819 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3820 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3821 DAG.getConstant(PairTy.getSizeInBits()/2, dl,
3822 TLI.getShiftAmountTy(PairTy)));
3823 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3827 Tmp1 = Node->getOperand(0);
3828 Tmp2 = Node->getOperand(1);
3829 Tmp3 = Node->getOperand(2);
3830 if (Tmp1.getOpcode() == ISD::SETCC) {
3831 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3833 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3835 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3836 DAG.getConstant(0, dl, Tmp1.getValueType()),
3837 Tmp2, Tmp3, ISD::SETNE);
3839 Results.push_back(Tmp1);
3842 SDValue Chain = Node->getOperand(0);
3843 SDValue Table = Node->getOperand(1);
3844 SDValue Index = Node->getOperand(2);
3846 EVT PTy = TLI.getPointerTy(DAG.getDataLayout());
3848 const DataLayout &TD = DAG.getDataLayout();
3849 unsigned EntrySize =
3850 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3852 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(), Index,
3853 DAG.getConstant(EntrySize, dl, Index.getValueType()));
3854 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3857 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3858 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3859 MachinePointerInfo::getJumpTable(), MemVT,
3860 false, false, false, 0);
3862 if (TM.getRelocationModel() == Reloc::PIC_) {
3863 // For PIC, the sequence is:
3864 // BRIND(load(Jumptable + index) + RelocBase)
3865 // RelocBase can be JumpTable, GOT or some sort of global base.
3866 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3867 TLI.getPICJumpTableRelocBase(Table, DAG));
3869 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3870 Results.push_back(Tmp1);
3874 // Expand brcond's setcc into its constituent parts and create a BR_CC
3876 Tmp1 = Node->getOperand(0);
3877 Tmp2 = Node->getOperand(1);
3878 if (Tmp2.getOpcode() == ISD::SETCC) {
3879 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3880 Tmp1, Tmp2.getOperand(2),
3881 Tmp2.getOperand(0), Tmp2.getOperand(1),
3882 Node->getOperand(2));
3884 // We test only the i1 bit. Skip the AND if UNDEF.
3885 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3886 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3887 DAG.getConstant(1, dl, Tmp2.getValueType()));
3888 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3889 DAG.getCondCode(ISD::SETNE), Tmp3,
3890 DAG.getConstant(0, dl, Tmp3.getValueType()),
3891 Node->getOperand(2));
3893 Results.push_back(Tmp1);
3896 Tmp1 = Node->getOperand(0);
3897 Tmp2 = Node->getOperand(1);
3898 Tmp3 = Node->getOperand(2);
3899 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3900 Tmp3, NeedInvert, dl);
3903 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3904 // condition code, create a new SETCC node.
3906 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3909 // If we expanded the SETCC by inverting the condition code, then wrap
3910 // the existing SETCC in a NOT to restore the intended condition.
3912 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
3914 Results.push_back(Tmp1);
3918 // Otherwise, SETCC for the given comparison type must be completely
3919 // illegal; expand it into a SELECT_CC.
3920 EVT VT = Node->getValueType(0);
3922 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
3923 case TargetLowering::ZeroOrOneBooleanContent:
3924 case TargetLowering::UndefinedBooleanContent:
3927 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3931 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3932 DAG.getConstant(TrueValue, dl, VT),
3933 DAG.getConstant(0, dl, VT),
3935 Results.push_back(Tmp1);
3938 case ISD::SELECT_CC: {
3939 Tmp1 = Node->getOperand(0); // LHS
3940 Tmp2 = Node->getOperand(1); // RHS
3941 Tmp3 = Node->getOperand(2); // True
3942 Tmp4 = Node->getOperand(3); // False
3943 EVT VT = Node->getValueType(0);
3944 SDValue CC = Node->getOperand(4);
3945 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
3947 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) {
3948 // If the condition code is legal, then we need to expand this
3949 // node using SETCC and SELECT.
3950 EVT CmpVT = Tmp1.getValueType();
3951 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
3952 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
3955 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT);
3956 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
3957 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
3961 // SELECT_CC is legal, so the condition code must not be.
3962 bool Legalized = false;
3963 // Try to legalize by inverting the condition. This is for targets that
3964 // might support an ordered version of a condition, but not the unordered
3965 // version (or vice versa).
3966 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
3967 Tmp1.getValueType().isInteger());
3968 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
3969 // Use the new condition code and swap true and false
3971 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
3973 // If The inverse is not legal, then try to swap the arguments using
3974 // the inverse condition code.
3975 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
3976 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
3977 // The swapped inverse condition is legal, so swap true and false,
3980 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
3985 Legalized = LegalizeSetCCCondCode(
3986 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
3989 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
3991 // If we expanded the SETCC by inverting the condition code, then swap
3992 // the True/False operands to match.
3994 std::swap(Tmp3, Tmp4);
3996 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3997 // condition code, create a new SELECT_CC node.
3999 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
4000 Tmp1, Tmp2, Tmp3, Tmp4, CC);
4002 Tmp2 = DAG.getConstant(0, dl, Tmp1.getValueType());
4003 CC = DAG.getCondCode(ISD::SETNE);
4004 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
4005 Tmp2, Tmp3, Tmp4, CC);
4008 Results.push_back(Tmp1);
4012 Tmp1 = Node->getOperand(0); // Chain
4013 Tmp2 = Node->getOperand(2); // LHS
4014 Tmp3 = Node->getOperand(3); // RHS
4015 Tmp4 = Node->getOperand(1); // CC
4017 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
4018 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
4020 assert(Legalized && "Can't legalize BR_CC with legal condition!");
4022 // If we expanded the SETCC by inverting the condition code, then wrap
4023 // the existing SETCC in a NOT to restore the intended condition.
4025 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
4027 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
4029 if (Tmp4.getNode()) {
4030 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
4031 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
4033 Tmp3 = DAG.getConstant(0, dl, Tmp2.getValueType());
4034 Tmp4 = DAG.getCondCode(ISD::SETNE);
4035 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
4036 Tmp2, Tmp3, Node->getOperand(4));
4038 Results.push_back(Tmp1);
4041 case ISD::BUILD_VECTOR:
4042 Results.push_back(ExpandBUILD_VECTOR(Node));
4047 // Scalarize vector SRA/SRL/SHL.
4048 EVT VT = Node->getValueType(0);
4049 assert(VT.isVector() && "Unable to legalize non-vector shift");
4050 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
4051 unsigned NumElem = VT.getVectorNumElements();
4053 SmallVector<SDValue, 8> Scalars;
4054 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
4055 SDValue Ex = DAG.getNode(
4056 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(0),
4057 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
4058 SDValue Sh = DAG.getNode(
4059 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(1),
4060 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
4061 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
4062 VT.getScalarType(), Ex, Sh));
4065 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Scalars);
4066 ReplaceNode(SDValue(Node, 0), Result);
4069 case ISD::GLOBAL_OFFSET_TABLE:
4070 case ISD::GlobalAddress:
4071 case ISD::GlobalTLSAddress:
4072 case ISD::ExternalSymbol:
4073 case ISD::ConstantPool:
4074 case ISD::JumpTable:
4075 case ISD::INTRINSIC_W_CHAIN:
4076 case ISD::INTRINSIC_WO_CHAIN:
4077 case ISD::INTRINSIC_VOID:
4078 // FIXME: Custom lowering for these operations shouldn't return null!
4082 // Replace the original node with the legalized result.
4083 if (!Results.empty())
4084 ReplaceNode(Node, Results.data());
4087 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4088 SmallVector<SDValue, 8> Results;
4089 MVT OVT = Node->getSimpleValueType(0);
4090 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4091 Node->getOpcode() == ISD::SINT_TO_FP ||
4092 Node->getOpcode() == ISD::SETCC) {
4093 OVT = Node->getOperand(0).getSimpleValueType();
4095 if (Node->getOpcode() == ISD::BR_CC)
4096 OVT = Node->getOperand(2).getSimpleValueType();
4097 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4099 SDValue Tmp1, Tmp2, Tmp3;
4100 switch (Node->getOpcode()) {
4102 case ISD::CTTZ_ZERO_UNDEF:
4104 case ISD::CTLZ_ZERO_UNDEF:
4106 // Zero extend the argument.
4107 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4108 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4109 // already the correct result.
4110 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4111 if (Node->getOpcode() == ISD::CTTZ) {
4112 // FIXME: This should set a bit in the zero extended value instead.
4113 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
4114 Tmp1, DAG.getConstant(NVT.getSizeInBits(), dl, NVT),
4116 Tmp1 = DAG.getSelect(dl, NVT, Tmp2,
4117 DAG.getConstant(OVT.getSizeInBits(), dl, NVT), Tmp1);
4118 } else if (Node->getOpcode() == ISD::CTLZ ||
4119 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4120 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4121 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4122 DAG.getConstant(NVT.getSizeInBits() -
4123 OVT.getSizeInBits(), dl, NVT));
4125 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4128 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4129 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4130 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
4131 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
4132 DAG.getConstant(DiffBits, dl,
4133 TLI.getShiftAmountTy(NVT)));
4134 Results.push_back(Tmp1);
4137 case ISD::FP_TO_UINT:
4138 case ISD::FP_TO_SINT:
4139 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
4140 Node->getOpcode() == ISD::FP_TO_SINT, dl);
4141 Results.push_back(Tmp1);
4143 case ISD::UINT_TO_FP:
4144 case ISD::SINT_TO_FP:
4145 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
4146 Node->getOpcode() == ISD::SINT_TO_FP, dl);
4147 Results.push_back(Tmp1);
4150 SDValue Chain = Node->getOperand(0); // Get the chain.
4151 SDValue Ptr = Node->getOperand(1); // Get the pointer.
4154 if (OVT.isVector()) {
4155 TruncOp = ISD::BITCAST;
4157 assert(OVT.isInteger()
4158 && "VAARG promotion is supported only for vectors or integer types");
4159 TruncOp = ISD::TRUNCATE;
4162 // Perform the larger operation, then convert back
4163 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
4164 Node->getConstantOperandVal(3));
4165 Chain = Tmp1.getValue(1);
4167 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4169 // Modified the chain result - switch anything that used the old chain to
4171 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4172 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4174 UpdatedNodes->insert(Tmp2.getNode());
4175 UpdatedNodes->insert(Chain.getNode());
4183 unsigned ExtOp, TruncOp;
4184 if (OVT.isVector()) {
4185 ExtOp = ISD::BITCAST;
4186 TruncOp = ISD::BITCAST;
4188 assert(OVT.isInteger() && "Cannot promote logic operation");
4189 ExtOp = ISD::ANY_EXTEND;
4190 TruncOp = ISD::TRUNCATE;
4192 // Promote each of the values to the new type.
4193 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4194 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4195 // Perform the larger operation, then convert back
4196 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4197 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4201 unsigned ExtOp, TruncOp;
4202 if (Node->getValueType(0).isVector() ||
4203 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4204 ExtOp = ISD::BITCAST;
4205 TruncOp = ISD::BITCAST;
4206 } else if (Node->getValueType(0).isInteger()) {
4207 ExtOp = ISD::ANY_EXTEND;
4208 TruncOp = ISD::TRUNCATE;
4210 ExtOp = ISD::FP_EXTEND;
4211 TruncOp = ISD::FP_ROUND;
4213 Tmp1 = Node->getOperand(0);
4214 // Promote each of the values to the new type.
4215 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4216 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4217 // Perform the larger operation, then round down.
4218 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4219 if (TruncOp != ISD::FP_ROUND)
4220 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4222 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4223 DAG.getIntPtrConstant(0, dl));
4224 Results.push_back(Tmp1);
4227 case ISD::VECTOR_SHUFFLE: {
4228 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4230 // Cast the two input vectors.
4231 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4232 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4234 // Convert the shuffle mask to the right # elements.
4235 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4236 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4237 Results.push_back(Tmp1);
4241 unsigned ExtOp = ISD::FP_EXTEND;
4242 if (NVT.isInteger()) {
4243 ISD::CondCode CCCode =
4244 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4245 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4247 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4248 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4249 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4250 Tmp1, Tmp2, Node->getOperand(2)));
4254 unsigned ExtOp = ISD::FP_EXTEND;
4255 if (NVT.isInteger()) {
4256 ISD::CondCode CCCode =
4257 cast<CondCodeSDNode>(Node->getOperand(1))->get();
4258 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4260 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4261 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(3));
4262 Results.push_back(DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0),
4263 Node->getOperand(0), Node->getOperand(1),
4264 Tmp1, Tmp2, Node->getOperand(4)));
4274 case ISD::FCOPYSIGN:
4276 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4277 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4278 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4279 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4280 Tmp3, DAG.getIntPtrConstant(0, dl)));
4284 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4285 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4286 Tmp3 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(2));
4288 DAG.getNode(ISD::FP_ROUND, dl, OVT,
4289 DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2, Tmp3),
4290 DAG.getIntPtrConstant(0, dl)));
4294 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4295 Tmp2 = Node->getOperand(1);
4296 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4297 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4298 Tmp3, DAG.getIntPtrConstant(0, dl)));
4304 case ISD::FNEARBYINT:
4317 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4318 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4319 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4320 Tmp2, DAG.getIntPtrConstant(0, dl)));
4325 // Replace the original node with the legalized result.
4326 if (!Results.empty())
4327 ReplaceNode(Node, Results.data());
4330 /// This is the entry point for the file.
4331 void SelectionDAG::Legalize() {
4332 AssignTopologicalOrder();
4334 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4335 SelectionDAGLegalize Legalizer(*this, LegalizedNodes);
4337 // Visit all the nodes. We start in topological order, so that we see
4338 // nodes with their original operands intact. Legalization can produce
4339 // new nodes which may themselves need to be legalized. Iterate until all
4340 // nodes have been legalized.
4342 bool AnyLegalized = false;
4343 for (auto NI = allnodes_end(); NI != allnodes_begin();) {
4347 if (N->use_empty() && N != getRoot().getNode()) {
4353 if (LegalizedNodes.insert(N).second) {
4354 AnyLegalized = true;
4355 Legalizer.LegalizeOp(N);
4357 if (N->use_empty() && N != getRoot().getNode()) {
4368 // Remove dead nodes now.
4372 bool SelectionDAG::LegalizeOp(SDNode *N,
4373 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4374 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4375 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes);
4377 // Directly insert the node in question, and legalize it. This will recurse
4378 // as needed through operands.
4379 LegalizedNodes.insert(N);
4380 Legalizer.LegalizeOp(N);
4382 return LegalizedNodes.count(N);