1 //===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SelectionDAG::Legalize method.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/CodeGen/Analysis.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineJumpTableInfo.h"
23 #include "llvm/IR/CallingConv.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DebugInfo.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetFrameLowering.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
39 //===----------------------------------------------------------------------===//
40 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
41 /// hacks on it until the target machine can handle it. This involves
42 /// eliminating value sizes the machine cannot handle (promoting small sizes to
43 /// large sizes or splitting up large values into small values) as well as
44 /// eliminating operations the machine cannot handle.
46 /// This code also does a small amount of optimization and recognition of idioms
47 /// as part of its processing. For example, if a target does not support a
48 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
49 /// will attempt merge setcc and brc instructions into brcc's.
52 class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
53 const TargetMachine &TM;
54 const TargetLowering &TLI;
57 /// \brief The iterator being used to walk the DAG. We hold a reference to it
58 /// in order to update it as necessary on node deletion.
59 SelectionDAG::allnodes_iterator &LegalizePosition;
61 /// \brief The set of nodes which have already been legalized. We hold a
62 /// reference to it in order to update as necessary on node deletion.
63 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
65 /// \brief A set of all the nodes updated during legalization.
66 SmallSetVector<SDNode *, 16> *UpdatedNodes;
68 EVT getSetCCResultType(EVT VT) const {
69 return TLI.getSetCCResultType(*DAG.getContext(), VT);
72 // Libcall insertion helpers.
75 SelectionDAGLegalize(SelectionDAG &DAG,
76 SelectionDAG::allnodes_iterator &LegalizePosition,
77 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
78 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
79 : SelectionDAG::DAGUpdateListener(DAG), TM(DAG.getTarget()),
80 TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
81 LegalizePosition(LegalizePosition), LegalizedNodes(LegalizedNodes),
82 UpdatedNodes(UpdatedNodes) {}
84 /// \brief Legalizes the given operation.
85 void LegalizeOp(SDNode *Node);
88 SDValue OptimizeFloatStore(StoreSDNode *ST);
90 void LegalizeLoadOps(SDNode *Node);
91 void LegalizeStoreOps(SDNode *Node);
93 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
94 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
95 /// is necessary to spill the vector being inserted into to memory, perform
96 /// the insert there, and then read the result back.
97 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
98 SDValue Idx, SDLoc dl);
99 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
100 SDValue Idx, SDLoc dl);
102 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
103 /// performs the same shuffe in terms of order or result bytes, but on a type
104 /// whose vector element type is narrower than the original shuffle type.
105 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
106 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
107 SDValue N1, SDValue N2,
108 ArrayRef<int> Mask) const;
110 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
111 bool &NeedInvert, SDLoc dl);
113 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
114 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
115 unsigned NumOps, bool isSigned, SDLoc dl);
117 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
118 SDNode *Node, bool isSigned);
119 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
120 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
121 RTLIB::Libcall Call_F128,
122 RTLIB::Libcall Call_PPCF128);
123 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
124 RTLIB::Libcall Call_I8,
125 RTLIB::Libcall Call_I16,
126 RTLIB::Libcall Call_I32,
127 RTLIB::Libcall Call_I64,
128 RTLIB::Libcall Call_I128);
129 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
130 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
132 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
133 SDValue ExpandBUILD_VECTOR(SDNode *Node);
134 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
135 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
136 SmallVectorImpl<SDValue> &Results);
137 SDValue ExpandFCOPYSIGN(SDNode *Node);
138 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
140 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
142 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
145 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
146 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
148 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
149 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
150 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
152 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
154 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
156 void ExpandNode(SDNode *Node);
157 void PromoteNode(SDNode *Node);
159 void ForgetNode(SDNode *N) {
160 LegalizedNodes.erase(N);
161 if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
164 UpdatedNodes->remove(N);
168 // DAGUpdateListener implementation.
169 void NodeDeleted(SDNode *N, SDNode *E) override {
172 void NodeUpdated(SDNode *N) override {}
174 // Node replacement helpers
175 void ReplacedNode(SDNode *N) {
176 if (N->use_empty()) {
177 DAG.RemoveDeadNode(N);
182 void ReplaceNode(SDNode *Old, SDNode *New) {
183 DAG.ReplaceAllUsesWith(Old, New);
184 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i)
185 DAG.TransferDbgValues(SDValue(Old, i), SDValue(New, i));
187 UpdatedNodes->insert(New);
190 void ReplaceNode(SDValue Old, SDValue New) {
191 DAG.ReplaceAllUsesWith(Old, New);
192 DAG.TransferDbgValues(Old, New);
194 UpdatedNodes->insert(New.getNode());
195 ReplacedNode(Old.getNode());
197 void ReplaceNode(SDNode *Old, const SDValue *New) {
198 DAG.ReplaceAllUsesWith(Old, New);
199 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i)
200 DAG.TransferDbgValues(SDValue(Old, i), New[i]);
202 UpdatedNodes->insert(New->getNode());
208 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
209 /// performs the same shuffe in terms of order or result bytes, but on a type
210 /// whose vector element type is narrower than the original shuffle type.
211 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
213 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
214 SDValue N1, SDValue N2,
215 ArrayRef<int> Mask) const {
216 unsigned NumMaskElts = VT.getVectorNumElements();
217 unsigned NumDestElts = NVT.getVectorNumElements();
218 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
220 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
222 if (NumEltsGrowth == 1)
223 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
225 SmallVector<int, 8> NewMask;
226 for (unsigned i = 0; i != NumMaskElts; ++i) {
228 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
230 NewMask.push_back(-1);
232 NewMask.push_back(Idx * NumEltsGrowth + j);
235 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
236 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
237 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
240 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
241 /// a load from the constant pool.
243 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
247 // If a FP immediate is precise when represented as a float and if the
248 // target can do an extending load from float to double, we put it into
249 // the constant pool as a float, even if it's is statically typed as a
250 // double. This shrinks FP constants and canonicalizes them for targets where
251 // an FP extending load is the same cost as a normal load (such as on the x87
252 // fp stack or PPC FP unit).
253 EVT VT = CFP->getValueType(0);
254 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
256 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
257 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
258 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
263 while (SVT != MVT::f32 && SVT != MVT::f16) {
264 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
265 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
266 // Only do this if the target has a native EXTLOAD instruction from
268 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
269 TLI.ShouldShrinkFPConstant(OrigVT)) {
270 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
271 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
277 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
278 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
281 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
283 CPIdx, MachinePointerInfo::getConstantPool(),
284 VT, false, false, Alignment);
288 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
289 MachinePointerInfo::getConstantPool(), false, false, false,
294 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
295 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
296 const TargetLowering &TLI,
297 SelectionDAGLegalize *DAGLegalize) {
298 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
299 "unaligned indexed stores not implemented!");
300 SDValue Chain = ST->getChain();
301 SDValue Ptr = ST->getBasePtr();
302 SDValue Val = ST->getValue();
303 EVT VT = Val.getValueType();
304 int Alignment = ST->getAlignment();
305 unsigned AS = ST->getAddressSpace();
308 if (ST->getMemoryVT().isFloatingPoint() ||
309 ST->getMemoryVT().isVector()) {
310 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
311 if (TLI.isTypeLegal(intVT)) {
312 // Expand to a bitconvert of the value to the integer type of the
313 // same size, then a (misaligned) int store.
314 // FIXME: Does not handle truncating floating point stores!
315 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
316 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
317 ST->isVolatile(), ST->isNonTemporal(), Alignment);
318 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
321 // Do a (aligned) store to a stack slot, then copy from the stack slot
322 // to the final destination using (unaligned) integer loads and stores.
323 EVT StoredVT = ST->getMemoryVT();
325 TLI.getRegisterType(*DAG.getContext(),
326 EVT::getIntegerVT(*DAG.getContext(),
327 StoredVT.getSizeInBits()));
328 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
329 unsigned RegBytes = RegVT.getSizeInBits() / 8;
330 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
332 // Make sure the stack slot is also aligned for the register type.
333 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
335 // Perform the original store, only redirected to the stack slot.
336 SDValue Store = DAG.getTruncStore(Chain, dl,
337 Val, StackPtr, MachinePointerInfo(),
338 StoredVT, false, false, 0);
339 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy(AS));
340 SmallVector<SDValue, 8> Stores;
343 // Do all but one copies using the full register width.
344 for (unsigned i = 1; i < NumRegs; i++) {
345 // Load one integer register's worth from the stack slot.
346 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
347 MachinePointerInfo(),
348 false, false, false, 0);
349 // Store it to the final location. Remember the store.
350 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
351 ST->getPointerInfo().getWithOffset(Offset),
352 ST->isVolatile(), ST->isNonTemporal(),
353 MinAlign(ST->getAlignment(), Offset)));
354 // Increment the pointers.
356 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
358 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
361 // The last store may be partial. Do a truncating store. On big-endian
362 // machines this requires an extending load from the stack slot to ensure
363 // that the bits are in the right place.
364 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
365 8 * (StoredBytes - Offset));
367 // Load from the stack slot.
368 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
369 MachinePointerInfo(),
370 MemVT, false, false, 0);
372 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
374 .getWithOffset(Offset),
375 MemVT, ST->isVolatile(),
377 MinAlign(ST->getAlignment(), Offset),
379 // The order of the stores doesn't matter - say it with a TokenFactor.
380 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
381 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
384 assert(ST->getMemoryVT().isInteger() &&
385 !ST->getMemoryVT().isVector() &&
386 "Unaligned store of unknown type.");
387 // Get the half-size VT
388 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
389 int NumBits = NewStoredVT.getSizeInBits();
390 int IncrementSize = NumBits / 8;
392 // Divide the stored value in two parts.
393 SDValue ShiftAmount = DAG.getConstant(NumBits,
394 TLI.getShiftAmountTy(Val.getValueType()));
396 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
398 // Store the two parts
399 SDValue Store1, Store2;
400 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
401 ST->getPointerInfo(), NewStoredVT,
402 ST->isVolatile(), ST->isNonTemporal(), Alignment);
404 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
405 DAG.getConstant(IncrementSize, TLI.getPointerTy(AS)));
406 Alignment = MinAlign(Alignment, IncrementSize);
407 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
408 ST->getPointerInfo().getWithOffset(IncrementSize),
409 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
410 Alignment, ST->getAAInfo());
413 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
414 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
417 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
419 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
420 const TargetLowering &TLI,
421 SDValue &ValResult, SDValue &ChainResult) {
422 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
423 "unaligned indexed loads not implemented!");
424 SDValue Chain = LD->getChain();
425 SDValue Ptr = LD->getBasePtr();
426 EVT VT = LD->getValueType(0);
427 EVT LoadedVT = LD->getMemoryVT();
429 if (VT.isFloatingPoint() || VT.isVector()) {
430 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
431 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
432 // Expand to a (misaligned) integer load of the same size,
433 // then bitconvert to floating point or vector.
434 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
435 LD->getMemOperand());
436 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
438 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
439 ISD::ANY_EXTEND, dl, VT, Result);
446 // Copy the value to a (aligned) stack slot using (unaligned) integer
447 // loads and stores, then do a (aligned) load from the stack slot.
448 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
449 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
450 unsigned RegBytes = RegVT.getSizeInBits() / 8;
451 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
453 // Make sure the stack slot is also aligned for the register type.
454 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
456 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
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(),
487 MinAlign(LD->getAlignment(), Offset),
489 // Follow the load with a store to the stack slot. Remember the store.
490 // On big-endian machines this requires a truncating store to ensure
491 // that the bits end up in the right place.
492 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
493 MachinePointerInfo(), MemVT,
496 // The order of the stores doesn't matter - say it with a TokenFactor.
497 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
499 // Finally, perform the original load only redirected to the stack slot.
500 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
501 MachinePointerInfo(), LoadedVT, false, false, 0);
503 // Callers expect a MERGE_VALUES node.
508 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
509 "Unaligned load of unsupported type.");
511 // Compute the new VT that is half the size of the old one. This is an
513 unsigned NumBits = LoadedVT.getSizeInBits();
515 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
518 unsigned Alignment = LD->getAlignment();
519 unsigned IncrementSize = NumBits / 8;
520 ISD::LoadExtType HiExtType = LD->getExtensionType();
522 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
523 if (HiExtType == ISD::NON_EXTLOAD)
524 HiExtType = ISD::ZEXTLOAD;
526 // Load the value in two parts
528 if (TLI.isLittleEndian()) {
529 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
530 NewLoadedVT, LD->isVolatile(),
531 LD->isNonTemporal(), Alignment, LD->getAAInfo());
532 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
533 DAG.getConstant(IncrementSize, Ptr.getValueType()));
534 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
535 LD->getPointerInfo().getWithOffset(IncrementSize),
536 NewLoadedVT, LD->isVolatile(),
537 LD->isNonTemporal(), MinAlign(Alignment, IncrementSize),
540 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
541 NewLoadedVT, LD->isVolatile(),
542 LD->isNonTemporal(), Alignment, LD->getAAInfo());
543 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
544 DAG.getConstant(IncrementSize, Ptr.getValueType()));
545 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
546 LD->getPointerInfo().getWithOffset(IncrementSize),
547 NewLoadedVT, LD->isVolatile(),
548 LD->isNonTemporal(), MinAlign(Alignment, IncrementSize),
552 // aggregate the two parts
553 SDValue ShiftAmount = DAG.getConstant(NumBits,
554 TLI.getShiftAmountTy(Hi.getValueType()));
555 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
556 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
558 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
565 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
566 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
567 /// is necessary to spill the vector being inserted into to memory, perform
568 /// the insert there, and then read the result back.
569 SDValue SelectionDAGLegalize::
570 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
576 // If the target doesn't support this, we have to spill the input vector
577 // to a temporary stack slot, update the element, then reload it. This is
578 // badness. We could also load the value into a vector register (either
579 // with a "move to register" or "extload into register" instruction, then
580 // permute it into place, if the idx is a constant and if the idx is
581 // supported by the target.
582 EVT VT = Tmp1.getValueType();
583 EVT EltVT = VT.getVectorElementType();
584 EVT IdxVT = Tmp3.getValueType();
585 EVT PtrVT = TLI.getPointerTy();
586 SDValue StackPtr = DAG.CreateStackTemporary(VT);
588 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
591 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
592 MachinePointerInfo::getFixedStack(SPFI),
595 // Truncate or zero extend offset to target pointer type.
596 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
597 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
598 // Add the offset to the index.
599 unsigned EltSize = EltVT.getSizeInBits()/8;
600 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
601 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
602 // Store the scalar value.
603 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
605 // Load the updated vector.
606 return DAG.getLoad(VT, dl, Ch, StackPtr,
607 MachinePointerInfo::getFixedStack(SPFI), false, false,
612 SDValue SelectionDAGLegalize::
613 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
614 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
615 // SCALAR_TO_VECTOR requires that the type of the value being inserted
616 // match the element type of the vector being created, except for
617 // integers in which case the inserted value can be over width.
618 EVT EltVT = Vec.getValueType().getVectorElementType();
619 if (Val.getValueType() == EltVT ||
620 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
621 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
622 Vec.getValueType(), Val);
624 unsigned NumElts = Vec.getValueType().getVectorNumElements();
625 // We generate a shuffle of InVec and ScVec, so the shuffle mask
626 // should be 0,1,2,3,4,5... with the appropriate element replaced with
628 SmallVector<int, 8> ShufOps;
629 for (unsigned i = 0; i != NumElts; ++i)
630 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
632 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
636 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
639 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
640 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
641 // FIXME: We shouldn't do this for TargetConstantFP's.
642 // FIXME: move this to the DAG Combiner! Note that we can't regress due
643 // to phase ordering between legalized code and the dag combiner. This
644 // probably means that we need to integrate dag combiner and legalizer
646 // We generally can't do this one for long doubles.
647 SDValue Chain = ST->getChain();
648 SDValue Ptr = ST->getBasePtr();
649 unsigned Alignment = ST->getAlignment();
650 bool isVolatile = ST->isVolatile();
651 bool isNonTemporal = ST->isNonTemporal();
652 AAMDNodes AAInfo = ST->getAAInfo();
654 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
655 if (CFP->getValueType(0) == MVT::f32 &&
656 TLI.isTypeLegal(MVT::i32)) {
657 SDValue Con = DAG.getConstant(CFP->getValueAPF().
658 bitcastToAPInt().zextOrTrunc(32),
660 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
661 isVolatile, isNonTemporal, Alignment, AAInfo);
664 if (CFP->getValueType(0) == MVT::f64) {
665 // If this target supports 64-bit registers, do a single 64-bit store.
666 if (TLI.isTypeLegal(MVT::i64)) {
667 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
668 zextOrTrunc(64), MVT::i64);
669 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
670 isVolatile, isNonTemporal, Alignment, AAInfo);
673 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
674 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
675 // stores. If the target supports neither 32- nor 64-bits, this
676 // xform is certainly not worth it.
677 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
678 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
679 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
680 if (TLI.isBigEndian()) std::swap(Lo, Hi);
682 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
683 isNonTemporal, Alignment, AAInfo);
684 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
685 DAG.getConstant(4, Ptr.getValueType()));
686 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
687 ST->getPointerInfo().getWithOffset(4),
688 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
691 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
695 return SDValue(nullptr, 0);
698 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
699 StoreSDNode *ST = cast<StoreSDNode>(Node);
700 SDValue Chain = ST->getChain();
701 SDValue Ptr = ST->getBasePtr();
704 unsigned Alignment = ST->getAlignment();
705 bool isVolatile = ST->isVolatile();
706 bool isNonTemporal = ST->isNonTemporal();
707 AAMDNodes AAInfo = ST->getAAInfo();
709 if (!ST->isTruncatingStore()) {
710 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
711 ReplaceNode(ST, OptStore);
716 SDValue Value = ST->getValue();
717 MVT VT = Value.getSimpleValueType();
718 switch (TLI.getOperationAction(ISD::STORE, VT)) {
719 default: llvm_unreachable("This action is not supported yet!");
720 case TargetLowering::Legal: {
721 // If this is an unaligned store and the target doesn't support it,
723 unsigned AS = ST->getAddressSpace();
724 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT(), AS)) {
725 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
726 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
727 if (ST->getAlignment() < ABIAlignment)
728 ExpandUnalignedStore(cast<StoreSDNode>(Node),
733 case TargetLowering::Custom: {
734 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
736 ReplaceNode(SDValue(Node, 0), Res);
739 case TargetLowering::Promote: {
740 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
741 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
742 "Can only promote stores to same size type");
743 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
745 DAG.getStore(Chain, dl, Value, Ptr,
746 ST->getPointerInfo(), isVolatile,
747 isNonTemporal, Alignment, AAInfo);
748 ReplaceNode(SDValue(Node, 0), Result);
755 SDValue Value = ST->getValue();
757 EVT StVT = ST->getMemoryVT();
758 unsigned StWidth = StVT.getSizeInBits();
760 if (StWidth != StVT.getStoreSizeInBits()) {
761 // Promote to a byte-sized store with upper bits zero if not
762 // storing an integral number of bytes. For example, promote
763 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
764 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
765 StVT.getStoreSizeInBits());
766 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
768 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
769 NVT, isVolatile, isNonTemporal, Alignment,
771 ReplaceNode(SDValue(Node, 0), Result);
772 } else if (StWidth & (StWidth - 1)) {
773 // If not storing a power-of-2 number of bits, expand as two stores.
774 assert(!StVT.isVector() && "Unsupported truncstore!");
775 unsigned RoundWidth = 1 << Log2_32(StWidth);
776 assert(RoundWidth < StWidth);
777 unsigned ExtraWidth = StWidth - RoundWidth;
778 assert(ExtraWidth < RoundWidth);
779 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
780 "Store size not an integral number of bytes!");
781 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
782 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
784 unsigned IncrementSize;
786 if (TLI.isLittleEndian()) {
787 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
788 // Store the bottom RoundWidth bits.
789 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
791 isVolatile, isNonTemporal, Alignment,
794 // Store the remaining ExtraWidth bits.
795 IncrementSize = RoundWidth / 8;
796 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
797 DAG.getConstant(IncrementSize, Ptr.getValueType()));
798 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
799 DAG.getConstant(RoundWidth,
800 TLI.getShiftAmountTy(Value.getValueType())));
801 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
802 ST->getPointerInfo().getWithOffset(IncrementSize),
803 ExtraVT, isVolatile, isNonTemporal,
804 MinAlign(Alignment, IncrementSize), AAInfo);
806 // Big endian - avoid unaligned stores.
807 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
808 // Store the top RoundWidth bits.
809 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
810 DAG.getConstant(ExtraWidth,
811 TLI.getShiftAmountTy(Value.getValueType())));
812 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
813 RoundVT, isVolatile, isNonTemporal, Alignment,
816 // Store the remaining ExtraWidth bits.
817 IncrementSize = RoundWidth / 8;
818 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
819 DAG.getConstant(IncrementSize, Ptr.getValueType()));
820 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
821 ST->getPointerInfo().getWithOffset(IncrementSize),
822 ExtraVT, isVolatile, isNonTemporal,
823 MinAlign(Alignment, IncrementSize), AAInfo);
826 // The order of the stores doesn't matter.
827 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
828 ReplaceNode(SDValue(Node, 0), Result);
830 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
831 StVT.getSimpleVT())) {
832 default: llvm_unreachable("This action is not supported yet!");
833 case TargetLowering::Legal: {
834 unsigned AS = ST->getAddressSpace();
835 // If this is an unaligned store and the target doesn't support it,
837 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT(), AS)) {
838 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
839 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
840 if (ST->getAlignment() < ABIAlignment)
841 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
845 case TargetLowering::Custom: {
846 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
848 ReplaceNode(SDValue(Node, 0), Res);
851 case TargetLowering::Expand:
852 assert(!StVT.isVector() &&
853 "Vector Stores are handled in LegalizeVectorOps");
855 // TRUNCSTORE:i16 i32 -> STORE i16
856 assert(TLI.isTypeLegal(StVT) &&
857 "Do not know how to expand this store!");
858 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
860 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
861 isVolatile, isNonTemporal, Alignment, AAInfo);
862 ReplaceNode(SDValue(Node, 0), Result);
869 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
870 LoadSDNode *LD = cast<LoadSDNode>(Node);
871 SDValue Chain = LD->getChain(); // The chain.
872 SDValue Ptr = LD->getBasePtr(); // The base pointer.
873 SDValue Value; // The value returned by the load op.
876 ISD::LoadExtType ExtType = LD->getExtensionType();
877 if (ExtType == ISD::NON_EXTLOAD) {
878 MVT VT = Node->getSimpleValueType(0);
879 SDValue RVal = SDValue(Node, 0);
880 SDValue RChain = SDValue(Node, 1);
882 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
883 default: llvm_unreachable("This action is not supported yet!");
884 case TargetLowering::Legal: {
885 unsigned AS = LD->getAddressSpace();
886 // If this is an unaligned load and the target doesn't support it,
888 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT(), AS)) {
889 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
890 unsigned ABIAlignment =
891 TLI.getDataLayout()->getABITypeAlignment(Ty);
892 if (LD->getAlignment() < ABIAlignment){
893 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
898 case TargetLowering::Custom: {
899 SDValue Res = TLI.LowerOperation(RVal, DAG);
902 RChain = Res.getValue(1);
906 case TargetLowering::Promote: {
907 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
908 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
909 "Can only promote loads to same size type");
911 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
912 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
913 RChain = Res.getValue(1);
917 if (RChain.getNode() != Node) {
918 assert(RVal.getNode() != Node && "Load must be completely replaced");
919 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
920 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
922 UpdatedNodes->insert(RVal.getNode());
923 UpdatedNodes->insert(RChain.getNode());
930 EVT SrcVT = LD->getMemoryVT();
931 unsigned SrcWidth = SrcVT.getSizeInBits();
932 unsigned Alignment = LD->getAlignment();
933 bool isVolatile = LD->isVolatile();
934 bool isNonTemporal = LD->isNonTemporal();
935 AAMDNodes AAInfo = LD->getAAInfo();
937 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
938 // Some targets pretend to have an i1 loading operation, and actually
939 // load an i8. This trick is correct for ZEXTLOAD because the top 7
940 // bits are guaranteed to be zero; it helps the optimizers understand
941 // that these bits are zero. It is also useful for EXTLOAD, since it
942 // tells the optimizers that those bits are undefined. It would be
943 // nice to have an effective generic way of getting these benefits...
944 // Until such a way is found, don't insist on promoting i1 here.
946 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
947 // Promote to a byte-sized load if not loading an integral number of
948 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
949 unsigned NewWidth = SrcVT.getStoreSizeInBits();
950 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
953 // The extra bits are guaranteed to be zero, since we stored them that
954 // way. A zext load from NVT thus automatically gives zext from SrcVT.
956 ISD::LoadExtType NewExtType =
957 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
960 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
961 Chain, Ptr, LD->getPointerInfo(),
962 NVT, isVolatile, isNonTemporal, Alignment, AAInfo);
964 Ch = Result.getValue(1); // The chain.
966 if (ExtType == ISD::SEXTLOAD)
967 // Having the top bits zero doesn't help when sign extending.
968 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
969 Result.getValueType(),
970 Result, DAG.getValueType(SrcVT));
971 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
972 // All the top bits are guaranteed to be zero - inform the optimizers.
973 Result = DAG.getNode(ISD::AssertZext, dl,
974 Result.getValueType(), Result,
975 DAG.getValueType(SrcVT));
979 } else if (SrcWidth & (SrcWidth - 1)) {
980 // If not loading a power-of-2 number of bits, expand as two loads.
981 assert(!SrcVT.isVector() && "Unsupported extload!");
982 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
983 assert(RoundWidth < SrcWidth);
984 unsigned ExtraWidth = SrcWidth - RoundWidth;
985 assert(ExtraWidth < RoundWidth);
986 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
987 "Load size not an integral number of bytes!");
988 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
989 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
991 unsigned IncrementSize;
993 if (TLI.isLittleEndian()) {
994 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
995 // Load the bottom RoundWidth bits.
996 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
998 LD->getPointerInfo(), RoundVT, isVolatile,
999 isNonTemporal, Alignment, AAInfo);
1001 // Load the remaining ExtraWidth bits.
1002 IncrementSize = RoundWidth / 8;
1003 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1004 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1005 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1006 LD->getPointerInfo().getWithOffset(IncrementSize),
1007 ExtraVT, isVolatile, isNonTemporal,
1008 MinAlign(Alignment, IncrementSize), AAInfo);
1010 // Build a factor node to remember that this load is independent of
1012 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1015 // Move the top bits to the right place.
1016 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1017 DAG.getConstant(RoundWidth,
1018 TLI.getShiftAmountTy(Hi.getValueType())));
1020 // Join the hi and lo parts.
1021 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1023 // Big endian - avoid unaligned loads.
1024 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1025 // Load the top RoundWidth bits.
1026 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1027 LD->getPointerInfo(), RoundVT, isVolatile,
1028 isNonTemporal, Alignment, AAInfo);
1030 // Load the remaining ExtraWidth bits.
1031 IncrementSize = RoundWidth / 8;
1032 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1033 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1034 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1035 dl, Node->getValueType(0), Chain, Ptr,
1036 LD->getPointerInfo().getWithOffset(IncrementSize),
1037 ExtraVT, isVolatile, isNonTemporal,
1038 MinAlign(Alignment, IncrementSize), AAInfo);
1040 // Build a factor node to remember that this load is independent of
1042 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1045 // Move the top bits to the right place.
1046 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1047 DAG.getConstant(ExtraWidth,
1048 TLI.getShiftAmountTy(Hi.getValueType())));
1050 // Join the hi and lo parts.
1051 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1056 bool isCustom = false;
1057 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1058 default: llvm_unreachable("This action is not supported yet!");
1059 case TargetLowering::Custom:
1062 case TargetLowering::Legal: {
1063 Value = SDValue(Node, 0);
1064 Chain = SDValue(Node, 1);
1067 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1068 if (Res.getNode()) {
1070 Chain = Res.getValue(1);
1073 // If this is an unaligned load and the target doesn't support
1075 EVT MemVT = LD->getMemoryVT();
1076 unsigned AS = LD->getAddressSpace();
1077 if (!TLI.allowsUnalignedMemoryAccesses(MemVT, AS)) {
1079 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1080 unsigned ABIAlignment =
1081 TLI.getDataLayout()->getABITypeAlignment(Ty);
1082 if (LD->getAlignment() < ABIAlignment){
1083 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1084 DAG, TLI, Value, Chain);
1090 case TargetLowering::Expand:
1091 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) &&
1092 TLI.isTypeLegal(SrcVT)) {
1093 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1094 LD->getMemOperand());
1098 ExtendOp = (SrcVT.isFloatingPoint() ?
1099 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1101 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1102 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1103 default: llvm_unreachable("Unexpected extend load type!");
1105 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1106 Chain = Load.getValue(1);
1110 assert(!SrcVT.isVector() &&
1111 "Vector Loads are handled in LegalizeVectorOps");
1113 // FIXME: This does not work for vectors on most targets. Sign-
1114 // and zero-extend operations are currently folded into extending
1115 // loads, whether they are legal or not, and then we end up here
1116 // without any support for legalizing them.
1117 assert(ExtType != ISD::EXTLOAD &&
1118 "EXTLOAD should always be supported!");
1119 // Turn the unsupported load into an EXTLOAD followed by an
1120 // explicit zero/sign extend inreg.
1121 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1122 Node->getValueType(0),
1124 LD->getMemOperand());
1126 if (ExtType == ISD::SEXTLOAD)
1127 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1128 Result.getValueType(),
1129 Result, DAG.getValueType(SrcVT));
1131 ValRes = DAG.getZeroExtendInReg(Result, dl,
1132 SrcVT.getScalarType());
1134 Chain = Result.getValue(1);
1139 // Since loads produce two values, make sure to remember that we legalized
1141 if (Chain.getNode() != Node) {
1142 assert(Value.getNode() != Node && "Load must be completely replaced");
1143 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1144 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1146 UpdatedNodes->insert(Value.getNode());
1147 UpdatedNodes->insert(Chain.getNode());
1153 /// LegalizeOp - Return a legal replacement for the given operation, with
1154 /// all legal operands.
1155 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1156 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1159 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1160 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1161 TargetLowering::TypeLegal &&
1162 "Unexpected illegal type!");
1164 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1165 assert((TLI.getTypeAction(*DAG.getContext(),
1166 Node->getOperand(i).getValueType()) ==
1167 TargetLowering::TypeLegal ||
1168 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1169 "Unexpected illegal type!");
1171 // Figure out the correct action; the way to query this varies by opcode
1172 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1173 bool SimpleFinishLegalizing = true;
1174 switch (Node->getOpcode()) {
1175 case ISD::INTRINSIC_W_CHAIN:
1176 case ISD::INTRINSIC_WO_CHAIN:
1177 case ISD::INTRINSIC_VOID:
1178 case ISD::STACKSAVE:
1179 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1182 Action = TLI.getOperationAction(Node->getOpcode(),
1183 Node->getValueType(0));
1184 if (Action != TargetLowering::Promote)
1185 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1187 case ISD::FP_TO_FP16:
1188 case ISD::SINT_TO_FP:
1189 case ISD::UINT_TO_FP:
1190 case ISD::EXTRACT_VECTOR_ELT:
1191 Action = TLI.getOperationAction(Node->getOpcode(),
1192 Node->getOperand(0).getValueType());
1194 case ISD::FP_ROUND_INREG:
1195 case ISD::SIGN_EXTEND_INREG: {
1196 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1197 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1200 case ISD::ATOMIC_STORE: {
1201 Action = TLI.getOperationAction(Node->getOpcode(),
1202 Node->getOperand(2).getValueType());
1205 case ISD::SELECT_CC:
1208 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1209 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1210 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1211 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1212 ISD::CondCode CCCode =
1213 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1214 Action = TLI.getCondCodeAction(CCCode, OpVT);
1215 if (Action == TargetLowering::Legal) {
1216 if (Node->getOpcode() == ISD::SELECT_CC)
1217 Action = TLI.getOperationAction(Node->getOpcode(),
1218 Node->getValueType(0));
1220 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1226 // FIXME: Model these properly. LOAD and STORE are complicated, and
1227 // STORE expects the unlegalized operand in some cases.
1228 SimpleFinishLegalizing = false;
1230 case ISD::CALLSEQ_START:
1231 case ISD::CALLSEQ_END:
1232 // FIXME: This shouldn't be necessary. These nodes have special properties
1233 // dealing with the recursive nature of legalization. Removing this
1234 // special case should be done as part of making LegalizeDAG non-recursive.
1235 SimpleFinishLegalizing = false;
1237 case ISD::EXTRACT_ELEMENT:
1238 case ISD::FLT_ROUNDS_:
1246 case ISD::MERGE_VALUES:
1247 case ISD::EH_RETURN:
1248 case ISD::FRAME_TO_ARGS_OFFSET:
1249 case ISD::EH_SJLJ_SETJMP:
1250 case ISD::EH_SJLJ_LONGJMP:
1251 // These operations lie about being legal: when they claim to be legal,
1252 // they should actually be expanded.
1253 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1254 if (Action == TargetLowering::Legal)
1255 Action = TargetLowering::Expand;
1257 case ISD::INIT_TRAMPOLINE:
1258 case ISD::ADJUST_TRAMPOLINE:
1259 case ISD::FRAMEADDR:
1260 case ISD::RETURNADDR:
1261 // These operations lie about being legal: when they claim to be legal,
1262 // they should actually be custom-lowered.
1263 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1264 if (Action == TargetLowering::Legal)
1265 Action = TargetLowering::Custom;
1267 case ISD::READ_REGISTER:
1268 case ISD::WRITE_REGISTER:
1269 // Named register is legal in the DAG, but blocked by register name
1270 // selection if not implemented by target (to chose the correct register)
1271 // They'll be converted to Copy(To/From)Reg.
1272 Action = TargetLowering::Legal;
1274 case ISD::DEBUGTRAP:
1275 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1276 if (Action == TargetLowering::Expand) {
1277 // replace ISD::DEBUGTRAP with ISD::TRAP
1279 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1280 Node->getOperand(0));
1281 ReplaceNode(Node, NewVal.getNode());
1282 LegalizeOp(NewVal.getNode());
1288 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1289 Action = TargetLowering::Legal;
1291 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1296 if (SimpleFinishLegalizing) {
1297 SDNode *NewNode = Node;
1298 switch (Node->getOpcode()) {
1305 // Legalizing shifts/rotates requires adjusting the shift amount
1306 // to the appropriate width.
1307 if (!Node->getOperand(1).getValueType().isVector()) {
1309 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1310 Node->getOperand(1));
1311 HandleSDNode Handle(SAO);
1312 LegalizeOp(SAO.getNode());
1313 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1317 case ISD::SRL_PARTS:
1318 case ISD::SRA_PARTS:
1319 case ISD::SHL_PARTS:
1320 // Legalizing shifts/rotates requires adjusting the shift amount
1321 // to the appropriate width.
1322 if (!Node->getOperand(2).getValueType().isVector()) {
1324 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1325 Node->getOperand(2));
1326 HandleSDNode Handle(SAO);
1327 LegalizeOp(SAO.getNode());
1328 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1329 Node->getOperand(1),
1335 if (NewNode != Node) {
1336 ReplaceNode(Node, NewNode);
1340 case TargetLowering::Legal:
1342 case TargetLowering::Custom: {
1343 // FIXME: The handling for custom lowering with multiple results is
1345 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1346 if (Res.getNode()) {
1347 SmallVector<SDValue, 8> ResultVals;
1348 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
1350 ResultVals.push_back(Res);
1352 ResultVals.push_back(Res.getValue(i));
1354 if (Res.getNode() != Node || Res.getResNo() != 0)
1355 ReplaceNode(Node, ResultVals.data());
1360 case TargetLowering::Expand:
1363 case TargetLowering::Promote:
1369 switch (Node->getOpcode()) {
1376 llvm_unreachable("Do not know how to legalize this operator!");
1378 case ISD::CALLSEQ_START:
1379 case ISD::CALLSEQ_END:
1382 return LegalizeLoadOps(Node);
1385 return LegalizeStoreOps(Node);
1390 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1391 SDValue Vec = Op.getOperand(0);
1392 SDValue Idx = Op.getOperand(1);
1395 // Before we generate a new store to a temporary stack slot, see if there is
1396 // already one that we can use. There often is because when we scalarize
1397 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1398 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1399 // the vector. If all are expanded here, we don't want one store per vector
1401 SDValue StackPtr, Ch;
1402 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1403 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1405 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1406 if (ST->isIndexed() || ST->isTruncatingStore() ||
1407 ST->getValue() != Vec)
1410 // Make sure that nothing else could have stored into the destination of
1412 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1415 StackPtr = ST->getBasePtr();
1416 Ch = SDValue(ST, 0);
1421 if (!Ch.getNode()) {
1422 // Store the value to a temporary stack slot, then LOAD the returned part.
1423 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1424 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1425 MachinePointerInfo(), false, false, 0);
1428 // Add the offset to the index.
1430 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1431 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1432 DAG.getConstant(EltSize, Idx.getValueType()));
1434 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1435 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1437 if (Op.getValueType().isVector())
1438 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1439 false, false, false, 0);
1440 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1441 MachinePointerInfo(),
1442 Vec.getValueType().getVectorElementType(),
1446 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1447 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1449 SDValue Vec = Op.getOperand(0);
1450 SDValue Part = Op.getOperand(1);
1451 SDValue Idx = Op.getOperand(2);
1454 // Store the value to a temporary stack slot, then LOAD the returned part.
1456 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1457 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1458 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1460 // First store the whole vector.
1461 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1464 // Then store the inserted part.
1466 // Add the offset to the index.
1468 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1470 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1471 DAG.getConstant(EltSize, Idx.getValueType()));
1472 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1474 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1477 // Store the subvector.
1478 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1479 MachinePointerInfo(), false, false, 0);
1481 // Finally, load the updated vector.
1482 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1483 false, false, false, 0);
1486 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1487 // We can't handle this case efficiently. Allocate a sufficiently
1488 // aligned object on the stack, store each element into it, then load
1489 // the result as a vector.
1490 // Create the stack frame object.
1491 EVT VT = Node->getValueType(0);
1492 EVT EltVT = VT.getVectorElementType();
1494 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1495 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1496 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1498 // Emit a store of each element to the stack slot.
1499 SmallVector<SDValue, 8> Stores;
1500 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1501 // Store (in the right endianness) the elements to memory.
1502 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1503 // Ignore undef elements.
1504 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1506 unsigned Offset = TypeByteSize*i;
1508 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1509 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1511 // If the destination vector element type is narrower than the source
1512 // element type, only store the bits necessary.
1513 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1514 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1515 Node->getOperand(i), Idx,
1516 PtrInfo.getWithOffset(Offset),
1517 EltVT, false, false, 0));
1519 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1520 Node->getOperand(i), Idx,
1521 PtrInfo.getWithOffset(Offset),
1526 if (!Stores.empty()) // Not all undef elements?
1527 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1529 StoreChain = DAG.getEntryNode();
1531 // Result is a load from the stack slot.
1532 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1533 false, false, false, 0);
1536 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1538 SDValue Tmp1 = Node->getOperand(0);
1539 SDValue Tmp2 = Node->getOperand(1);
1541 // Get the sign bit of the RHS. First obtain a value that has the same
1542 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1544 EVT FloatVT = Tmp2.getValueType();
1545 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1546 if (TLI.isTypeLegal(IVT)) {
1547 // Convert to an integer with the same sign bit.
1548 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1550 // Store the float to memory, then load the sign part out as an integer.
1551 MVT LoadTy = TLI.getPointerTy();
1552 // First create a temporary that is aligned for both the load and store.
1553 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1554 // Then store the float to it.
1556 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1558 if (TLI.isBigEndian()) {
1559 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1560 // Load out a legal integer with the same sign bit as the float.
1561 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1562 false, false, false, 0);
1563 } else { // Little endian
1564 SDValue LoadPtr = StackPtr;
1565 // The float may be wider than the integer we are going to load. Advance
1566 // the pointer so that the loaded integer will contain the sign bit.
1567 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1568 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1569 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1570 DAG.getConstant(ByteOffset, LoadPtr.getValueType()));
1571 // Load a legal integer containing the sign bit.
1572 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1573 false, false, false, 0);
1574 // Move the sign bit to the top bit of the loaded integer.
1575 unsigned BitShift = LoadTy.getSizeInBits() -
1576 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1577 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1579 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1580 DAG.getConstant(BitShift,
1581 TLI.getShiftAmountTy(SignBit.getValueType())));
1584 // Now get the sign bit proper, by seeing whether the value is negative.
1585 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1586 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1588 // Get the absolute value of the result.
1589 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1590 // Select between the nabs and abs value based on the sign bit of
1592 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1593 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1597 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1598 SmallVectorImpl<SDValue> &Results) {
1599 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1600 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1601 " not tell us which reg is the stack pointer!");
1603 EVT VT = Node->getValueType(0);
1604 SDValue Tmp1 = SDValue(Node, 0);
1605 SDValue Tmp2 = SDValue(Node, 1);
1606 SDValue Tmp3 = Node->getOperand(2);
1607 SDValue Chain = Tmp1.getOperand(0);
1609 // Chain the dynamic stack allocation so that it doesn't modify the stack
1610 // pointer when other instructions are using the stack.
1611 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
1614 SDValue Size = Tmp2.getOperand(1);
1615 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1616 Chain = SP.getValue(1);
1617 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1618 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1619 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1620 if (Align > StackAlign)
1621 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1622 DAG.getConstant(-(uint64_t)Align, VT));
1623 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1625 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1626 DAG.getIntPtrConstant(0, true), SDValue(),
1629 Results.push_back(Tmp1);
1630 Results.push_back(Tmp2);
1633 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1634 /// condition code CC on the current target.
1636 /// If the SETCC has been legalized using AND / OR, then the legalized node
1637 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1638 /// will be set to false.
1640 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1641 /// then the values of LHS and RHS will be swapped, CC will be set to the
1642 /// new condition, and NeedInvert will be set to false.
1644 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1645 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1646 /// will be set to true. The caller must invert the result of the SETCC with
1647 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1648 /// of a true/false result.
1650 /// \returns true if the SetCC has been legalized, false if it hasn't.
1651 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1652 SDValue &LHS, SDValue &RHS,
1656 MVT OpVT = LHS.getSimpleValueType();
1657 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1659 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1660 default: llvm_unreachable("Unknown condition code action!");
1661 case TargetLowering::Legal:
1664 case TargetLowering::Expand: {
1665 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1666 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1667 std::swap(LHS, RHS);
1668 CC = DAG.getCondCode(InvCC);
1671 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1674 default: llvm_unreachable("Don't know how to expand this condition!");
1676 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1677 == TargetLowering::Legal
1678 && "If SETO is expanded, SETOEQ must be legal!");
1679 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1681 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1682 == TargetLowering::Legal
1683 && "If SETUO is expanded, SETUNE must be legal!");
1684 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1697 // If we are floating point, assign and break, otherwise fall through.
1698 if (!OpVT.isInteger()) {
1699 // We can use the 4th bit to tell if we are the unordered
1700 // or ordered version of the opcode.
1701 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1702 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1703 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1706 // Fallthrough if we are unsigned integer.
1711 // We only support using the inverted operation, which is computed above
1712 // and not a different manner of supporting expanding these cases.
1713 llvm_unreachable("Don't know how to expand this condition!");
1716 // Try inverting the result of the inverse condition.
1717 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1718 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1719 CC = DAG.getCondCode(InvCC);
1723 // If inverting the condition didn't work then we have no means to expand
1725 llvm_unreachable("Don't know how to expand this condition!");
1728 SDValue SetCC1, SetCC2;
1729 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1730 // If we aren't the ordered or unorder operation,
1731 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1732 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1733 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1735 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1736 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1737 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1739 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1748 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1749 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1750 /// a load from the stack slot to DestVT, extending it if needed.
1751 /// The resultant code need not be legal.
1752 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1756 // Create the stack frame object.
1758 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1759 getTypeForEVT(*DAG.getContext()));
1760 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1762 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1763 int SPFI = StackPtrFI->getIndex();
1764 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1766 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1767 unsigned SlotSize = SlotVT.getSizeInBits();
1768 unsigned DestSize = DestVT.getSizeInBits();
1769 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1770 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1772 // Emit a store to the stack slot. Use a truncstore if the input value is
1773 // later than DestVT.
1776 if (SrcSize > SlotSize)
1777 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1778 PtrInfo, SlotVT, false, false, SrcAlign);
1780 assert(SrcSize == SlotSize && "Invalid store");
1781 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1782 PtrInfo, false, false, SrcAlign);
1785 // Result is a load from the stack slot.
1786 if (SlotSize == DestSize)
1787 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1788 false, false, false, DestAlign);
1790 assert(SlotSize < DestSize && "Unknown extension!");
1791 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1792 PtrInfo, SlotVT, false, false, DestAlign);
1795 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1797 // Create a vector sized/aligned stack slot, store the value to element #0,
1798 // then load the whole vector back out.
1799 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1801 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1802 int SPFI = StackPtrFI->getIndex();
1804 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1806 MachinePointerInfo::getFixedStack(SPFI),
1807 Node->getValueType(0).getVectorElementType(),
1809 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1810 MachinePointerInfo::getFixedStack(SPFI),
1811 false, false, false, 0);
1815 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1816 const TargetLowering &TLI, SDValue &Res) {
1817 unsigned NumElems = Node->getNumOperands();
1819 EVT VT = Node->getValueType(0);
1821 // Try to group the scalars into pairs, shuffle the pairs together, then
1822 // shuffle the pairs of pairs together, etc. until the vector has
1823 // been built. This will work only if all of the necessary shuffle masks
1826 // We do this in two phases; first to check the legality of the shuffles,
1827 // and next, assuming that all shuffles are legal, to create the new nodes.
1828 for (int Phase = 0; Phase < 2; ++Phase) {
1829 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1831 for (unsigned i = 0; i < NumElems; ++i) {
1832 SDValue V = Node->getOperand(i);
1833 if (V.getOpcode() == ISD::UNDEF)
1838 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1839 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1842 while (IntermedVals.size() > 2) {
1843 NewIntermedVals.clear();
1844 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1845 // This vector and the next vector are shuffled together (simply to
1846 // append the one to the other).
1847 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1849 SmallVector<int, 16> FinalIndices;
1850 FinalIndices.reserve(IntermedVals[i].second.size() +
1851 IntermedVals[i+1].second.size());
1854 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1857 FinalIndices.push_back(IntermedVals[i].second[j]);
1859 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1861 ShuffleVec[k] = NumElems + j;
1862 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1867 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1868 IntermedVals[i+1].first,
1870 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1872 NewIntermedVals.push_back(std::make_pair(Shuffle, FinalIndices));
1875 // If we had an odd number of defined values, then append the last
1876 // element to the array of new vectors.
1877 if ((IntermedVals.size() & 1) != 0)
1878 NewIntermedVals.push_back(IntermedVals.back());
1880 IntermedVals.swap(NewIntermedVals);
1883 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1884 "Invalid number of intermediate vectors");
1885 SDValue Vec1 = IntermedVals[0].first;
1887 if (IntermedVals.size() > 1)
1888 Vec2 = IntermedVals[1].first;
1890 Vec2 = DAG.getUNDEF(VT);
1892 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1893 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1894 ShuffleVec[IntermedVals[0].second[i]] = i;
1895 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1896 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1899 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1900 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1907 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1908 /// support the operation, but do support the resultant vector type.
1909 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1910 unsigned NumElems = Node->getNumOperands();
1911 SDValue Value1, Value2;
1913 EVT VT = Node->getValueType(0);
1914 EVT OpVT = Node->getOperand(0).getValueType();
1915 EVT EltVT = VT.getVectorElementType();
1917 // If the only non-undef value is the low element, turn this into a
1918 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1919 bool isOnlyLowElement = true;
1920 bool MoreThanTwoValues = false;
1921 bool isConstant = true;
1922 for (unsigned i = 0; i < NumElems; ++i) {
1923 SDValue V = Node->getOperand(i);
1924 if (V.getOpcode() == ISD::UNDEF)
1927 isOnlyLowElement = false;
1928 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1931 if (!Value1.getNode()) {
1933 } else if (!Value2.getNode()) {
1936 } else if (V != Value1 && V != Value2) {
1937 MoreThanTwoValues = true;
1941 if (!Value1.getNode())
1942 return DAG.getUNDEF(VT);
1944 if (isOnlyLowElement)
1945 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1947 // If all elements are constants, create a load from the constant pool.
1949 SmallVector<Constant*, 16> CV;
1950 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1951 if (ConstantFPSDNode *V =
1952 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1953 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1954 } else if (ConstantSDNode *V =
1955 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1957 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1959 // If OpVT and EltVT don't match, EltVT is not legal and the
1960 // element values have been promoted/truncated earlier. Undo this;
1961 // we don't want a v16i8 to become a v16i32 for example.
1962 const ConstantInt *CI = V->getConstantIntValue();
1963 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1964 CI->getZExtValue()));
1967 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1968 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1969 CV.push_back(UndefValue::get(OpNTy));
1972 Constant *CP = ConstantVector::get(CV);
1973 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1974 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1975 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1976 MachinePointerInfo::getConstantPool(),
1977 false, false, false, Alignment);
1980 SmallSet<SDValue, 16> DefinedValues;
1981 for (unsigned i = 0; i < NumElems; ++i) {
1982 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
1984 DefinedValues.insert(Node->getOperand(i));
1987 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
1988 if (!MoreThanTwoValues) {
1989 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1990 for (unsigned i = 0; i < NumElems; ++i) {
1991 SDValue V = Node->getOperand(i);
1992 if (V.getOpcode() == ISD::UNDEF)
1994 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1996 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1997 // Get the splatted value into the low element of a vector register.
1998 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2000 if (Value2.getNode())
2001 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2003 Vec2 = DAG.getUNDEF(VT);
2005 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2006 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2010 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2015 // Otherwise, we can't handle this case efficiently.
2016 return ExpandVectorBuildThroughStack(Node);
2019 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
2020 // does not fit into a register, return the lo part and set the hi part to the
2021 // by-reg argument. If it does fit into a single register, return the result
2022 // and leave the Hi part unset.
2023 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2025 TargetLowering::ArgListTy Args;
2026 TargetLowering::ArgListEntry Entry;
2027 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2028 EVT ArgVT = Node->getOperand(i).getValueType();
2029 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2030 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2031 Entry.isSExt = isSigned;
2032 Entry.isZExt = !isSigned;
2033 Args.push_back(Entry);
2035 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2036 TLI.getPointerTy());
2038 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2040 // By default, the input chain to this libcall is the entry node of the
2041 // function. If the libcall is going to be emitted as a tail call then
2042 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2043 // node which is being folded has a non-entry input chain.
2044 SDValue InChain = DAG.getEntryNode();
2046 // isTailCall may be true since the callee does not reference caller stack
2047 // frame. Check if it's in the right position.
2048 SDValue TCChain = InChain;
2049 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2053 TargetLowering::CallLoweringInfo CLI(DAG);
2054 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2055 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2056 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2058 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2060 if (!CallInfo.second.getNode())
2061 // It's a tailcall, return the chain (which is the DAG root).
2062 return DAG.getRoot();
2064 return CallInfo.first;
2067 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
2068 /// and returning a result of type RetVT.
2069 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2070 const SDValue *Ops, unsigned NumOps,
2071 bool isSigned, SDLoc dl) {
2072 TargetLowering::ArgListTy Args;
2073 Args.reserve(NumOps);
2075 TargetLowering::ArgListEntry Entry;
2076 for (unsigned i = 0; i != NumOps; ++i) {
2077 Entry.Node = Ops[i];
2078 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2079 Entry.isSExt = isSigned;
2080 Entry.isZExt = !isSigned;
2081 Args.push_back(Entry);
2083 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2084 TLI.getPointerTy());
2086 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2088 TargetLowering::CallLoweringInfo CLI(DAG);
2089 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2090 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2091 .setSExtResult(isSigned).setZExtResult(!isSigned);
2093 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2095 return CallInfo.first;
2098 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
2099 // ExpandLibCall except that the first operand is the in-chain.
2100 std::pair<SDValue, SDValue>
2101 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2104 SDValue InChain = Node->getOperand(0);
2106 TargetLowering::ArgListTy Args;
2107 TargetLowering::ArgListEntry Entry;
2108 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2109 EVT ArgVT = Node->getOperand(i).getValueType();
2110 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2111 Entry.Node = Node->getOperand(i);
2113 Entry.isSExt = isSigned;
2114 Entry.isZExt = !isSigned;
2115 Args.push_back(Entry);
2117 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2118 TLI.getPointerTy());
2120 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2122 TargetLowering::CallLoweringInfo CLI(DAG);
2123 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2124 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2125 .setSExtResult(isSigned).setZExtResult(!isSigned);
2127 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2132 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2133 RTLIB::Libcall Call_F32,
2134 RTLIB::Libcall Call_F64,
2135 RTLIB::Libcall Call_F80,
2136 RTLIB::Libcall Call_F128,
2137 RTLIB::Libcall Call_PPCF128) {
2139 switch (Node->getSimpleValueType(0).SimpleTy) {
2140 default: llvm_unreachable("Unexpected request for libcall!");
2141 case MVT::f32: LC = Call_F32; break;
2142 case MVT::f64: LC = Call_F64; break;
2143 case MVT::f80: LC = Call_F80; break;
2144 case MVT::f128: LC = Call_F128; break;
2145 case MVT::ppcf128: LC = Call_PPCF128; break;
2147 return ExpandLibCall(LC, Node, false);
2150 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2151 RTLIB::Libcall Call_I8,
2152 RTLIB::Libcall Call_I16,
2153 RTLIB::Libcall Call_I32,
2154 RTLIB::Libcall Call_I64,
2155 RTLIB::Libcall Call_I128) {
2157 switch (Node->getSimpleValueType(0).SimpleTy) {
2158 default: llvm_unreachable("Unexpected request for libcall!");
2159 case MVT::i8: LC = Call_I8; break;
2160 case MVT::i16: LC = Call_I16; break;
2161 case MVT::i32: LC = Call_I32; break;
2162 case MVT::i64: LC = Call_I64; break;
2163 case MVT::i128: LC = Call_I128; break;
2165 return ExpandLibCall(LC, Node, isSigned);
2168 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
2169 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2170 const TargetLowering &TLI) {
2172 switch (Node->getSimpleValueType(0).SimpleTy) {
2173 default: llvm_unreachable("Unexpected request for libcall!");
2174 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2175 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2176 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2177 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2178 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2181 return TLI.getLibcallName(LC) != nullptr;
2184 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2186 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2187 // The other use might have been replaced with a divrem already.
2188 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2189 unsigned OtherOpcode = 0;
2191 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2193 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2195 SDValue Op0 = Node->getOperand(0);
2196 SDValue Op1 = Node->getOperand(1);
2197 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2198 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2202 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2203 User->getOperand(0) == Op0 &&
2204 User->getOperand(1) == Op1)
2210 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2213 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2214 SmallVectorImpl<SDValue> &Results) {
2215 unsigned Opcode = Node->getOpcode();
2216 bool isSigned = Opcode == ISD::SDIVREM;
2219 switch (Node->getSimpleValueType(0).SimpleTy) {
2220 default: llvm_unreachable("Unexpected request for libcall!");
2221 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2222 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2223 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2224 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2225 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2228 // The input chain to this libcall is the entry node of the function.
2229 // Legalizing the call will automatically add the previous call to the
2231 SDValue InChain = DAG.getEntryNode();
2233 EVT RetVT = Node->getValueType(0);
2234 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2236 TargetLowering::ArgListTy Args;
2237 TargetLowering::ArgListEntry Entry;
2238 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2239 EVT ArgVT = Node->getOperand(i).getValueType();
2240 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2241 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2242 Entry.isSExt = isSigned;
2243 Entry.isZExt = !isSigned;
2244 Args.push_back(Entry);
2247 // Also pass the return address of the remainder.
2248 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2250 Entry.Ty = RetTy->getPointerTo();
2251 Entry.isSExt = isSigned;
2252 Entry.isZExt = !isSigned;
2253 Args.push_back(Entry);
2255 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2256 TLI.getPointerTy());
2259 TargetLowering::CallLoweringInfo CLI(DAG);
2260 CLI.setDebugLoc(dl).setChain(InChain)
2261 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2262 .setSExtResult(isSigned).setZExtResult(!isSigned);
2264 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2266 // Remainder is loaded back from the stack frame.
2267 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2268 MachinePointerInfo(), false, false, false, 0);
2269 Results.push_back(CallInfo.first);
2270 Results.push_back(Rem);
2273 /// isSinCosLibcallAvailable - Return true if sincos libcall is available.
2274 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2276 switch (Node->getSimpleValueType(0).SimpleTy) {
2277 default: llvm_unreachable("Unexpected request for libcall!");
2278 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2279 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2280 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2281 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2282 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2284 return TLI.getLibcallName(LC) != nullptr;
2287 /// canCombineSinCosLibcall - Return true if sincos libcall is available and
2288 /// can be used to combine sin and cos.
2289 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2290 const TargetMachine &TM) {
2291 if (!isSinCosLibcallAvailable(Node, TLI))
2293 // GNU sin/cos functions set errno while sincos does not. Therefore
2294 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2295 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2296 if (isGNU && !TM.Options.UnsafeFPMath)
2301 /// useSinCos - Only issue sincos libcall if both sin and cos are
2303 static bool useSinCos(SDNode *Node) {
2304 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2305 ? ISD::FCOS : ISD::FSIN;
2307 SDValue Op0 = Node->getOperand(0);
2308 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2309 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2313 // The other user might have been turned into sincos already.
2314 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2320 /// ExpandSinCosLibCall - Issue libcalls to sincos to compute sin / cos
2323 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2324 SmallVectorImpl<SDValue> &Results) {
2326 switch (Node->getSimpleValueType(0).SimpleTy) {
2327 default: llvm_unreachable("Unexpected request for libcall!");
2328 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2329 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2330 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2331 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2332 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2335 // The input chain to this libcall is the entry node of the function.
2336 // Legalizing the call will automatically add the previous call to the
2338 SDValue InChain = DAG.getEntryNode();
2340 EVT RetVT = Node->getValueType(0);
2341 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2343 TargetLowering::ArgListTy Args;
2344 TargetLowering::ArgListEntry Entry;
2346 // Pass the argument.
2347 Entry.Node = Node->getOperand(0);
2349 Entry.isSExt = false;
2350 Entry.isZExt = false;
2351 Args.push_back(Entry);
2353 // Pass the return address of sin.
2354 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2355 Entry.Node = SinPtr;
2356 Entry.Ty = RetTy->getPointerTo();
2357 Entry.isSExt = false;
2358 Entry.isZExt = false;
2359 Args.push_back(Entry);
2361 // Also pass the return address of the cos.
2362 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2363 Entry.Node = CosPtr;
2364 Entry.Ty = RetTy->getPointerTo();
2365 Entry.isSExt = false;
2366 Entry.isZExt = false;
2367 Args.push_back(Entry);
2369 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2370 TLI.getPointerTy());
2373 TargetLowering::CallLoweringInfo CLI(DAG);
2374 CLI.setDebugLoc(dl).setChain(InChain)
2375 .setCallee(TLI.getLibcallCallingConv(LC),
2376 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2378 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2380 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2381 MachinePointerInfo(), false, false, false, 0));
2382 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2383 MachinePointerInfo(), false, false, false, 0));
2386 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2387 /// INT_TO_FP operation of the specified operand when the target requests that
2388 /// we expand it. At this point, we know that the result and operand types are
2389 /// legal for the target.
2390 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2394 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2395 // simple 32-bit [signed|unsigned] integer to float/double expansion
2397 // Get the stack frame index of a 8 byte buffer.
2398 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2400 // word offset constant for Hi/Lo address computation
2401 SDValue WordOff = DAG.getConstant(sizeof(int), StackSlot.getValueType());
2402 // set up Hi and Lo (into buffer) address based on endian
2403 SDValue Hi = StackSlot;
2404 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2405 StackSlot, WordOff);
2406 if (TLI.isLittleEndian())
2409 // if signed map to unsigned space
2412 // constant used to invert sign bit (signed to unsigned mapping)
2413 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2414 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2418 // store the lo of the constructed double - based on integer input
2419 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2420 Op0Mapped, Lo, MachinePointerInfo(),
2422 // initial hi portion of constructed double
2423 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2424 // store the hi of the constructed double - biased exponent
2425 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2426 MachinePointerInfo(),
2428 // load the constructed double
2429 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2430 MachinePointerInfo(), false, false, false, 0);
2431 // FP constant to bias correct the final result
2432 SDValue Bias = DAG.getConstantFP(isSigned ?
2433 BitsToDouble(0x4330000080000000ULL) :
2434 BitsToDouble(0x4330000000000000ULL),
2436 // subtract the bias
2437 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2440 // handle final rounding
2441 if (DestVT == MVT::f64) {
2444 } else if (DestVT.bitsLT(MVT::f64)) {
2445 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2446 DAG.getIntPtrConstant(0));
2447 } else if (DestVT.bitsGT(MVT::f64)) {
2448 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2452 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2453 // Code below here assumes !isSigned without checking again.
2455 // Implementation of unsigned i64 to f64 following the algorithm in
2456 // __floatundidf in compiler_rt. This implementation has the advantage
2457 // of performing rounding correctly, both in the default rounding mode
2458 // and in all alternate rounding modes.
2459 // TODO: Generalize this for use with other types.
2460 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2462 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2463 SDValue TwoP84PlusTwoP52 =
2464 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2466 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2468 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2469 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2470 DAG.getConstant(32, MVT::i64));
2471 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2472 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2473 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2474 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2475 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2477 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2480 // Implementation of unsigned i64 to f32.
2481 // TODO: Generalize this for use with other types.
2482 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2483 // For unsigned conversions, convert them to signed conversions using the
2484 // algorithm from the x86_64 __floatundidf in compiler_rt.
2486 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2488 SDValue ShiftConst =
2489 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2490 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2491 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2492 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2493 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2495 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2496 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2498 // TODO: This really should be implemented using a branch rather than a
2499 // select. We happen to get lucky and machinesink does the right
2500 // thing most of the time. This would be a good candidate for a
2501 //pseudo-op, or, even better, for whole-function isel.
2502 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2503 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2504 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2507 // Otherwise, implement the fully general conversion.
2509 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2510 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2511 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2512 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2513 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2514 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2515 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2516 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2517 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2518 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2519 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2521 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2522 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2524 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2525 DAG.getConstant(32, SHVT));
2526 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2527 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2529 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2530 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2531 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2532 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2533 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2534 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2535 DAG.getIntPtrConstant(0));
2538 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2540 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2541 Op0, DAG.getConstant(0, Op0.getValueType()),
2543 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2544 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2545 SignSet, Four, Zero);
2547 // If the sign bit of the integer is set, the large number will be treated
2548 // as a negative number. To counteract this, the dynamic code adds an
2549 // offset depending on the data type.
2551 switch (Op0.getSimpleValueType().SimpleTy) {
2552 default: llvm_unreachable("Unsupported integer type!");
2553 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2554 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2555 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2556 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2558 if (TLI.isLittleEndian()) FF <<= 32;
2559 Constant *FudgeFactor = ConstantInt::get(
2560 Type::getInt64Ty(*DAG.getContext()), FF);
2562 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2563 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2564 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2565 Alignment = std::min(Alignment, 4u);
2567 if (DestVT == MVT::f32)
2568 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2569 MachinePointerInfo::getConstantPool(),
2570 false, false, false, Alignment);
2572 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2573 DAG.getEntryNode(), CPIdx,
2574 MachinePointerInfo::getConstantPool(),
2575 MVT::f32, false, false, Alignment);
2576 HandleSDNode Handle(Load);
2577 LegalizeOp(Load.getNode());
2578 FudgeInReg = Handle.getValue();
2581 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2584 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2585 /// *INT_TO_FP operation of the specified operand when the target requests that
2586 /// we promote it. At this point, we know that the result and operand types are
2587 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2588 /// operation that takes a larger input.
2589 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2593 // First step, figure out the appropriate *INT_TO_FP operation to use.
2594 EVT NewInTy = LegalOp.getValueType();
2596 unsigned OpToUse = 0;
2598 // Scan for the appropriate larger type to use.
2600 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2601 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2603 // If the target supports SINT_TO_FP of this type, use it.
2604 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2605 OpToUse = ISD::SINT_TO_FP;
2608 if (isSigned) continue;
2610 // If the target supports UINT_TO_FP of this type, use it.
2611 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2612 OpToUse = ISD::UINT_TO_FP;
2616 // Otherwise, try a larger type.
2619 // Okay, we found the operation and type to use. Zero extend our input to the
2620 // desired type then run the operation on it.
2621 return DAG.getNode(OpToUse, dl, DestVT,
2622 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2623 dl, NewInTy, LegalOp));
2626 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2627 /// FP_TO_*INT operation of the specified operand when the target requests that
2628 /// we promote it. At this point, we know that the result and operand types are
2629 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2630 /// operation that returns a larger result.
2631 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2635 // First step, figure out the appropriate FP_TO*INT operation to use.
2636 EVT NewOutTy = DestVT;
2638 unsigned OpToUse = 0;
2640 // Scan for the appropriate larger type to use.
2642 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2643 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2645 // A larger signed type can hold all unsigned values of the requested type,
2646 // so using FP_TO_SINT is valid
2647 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2648 OpToUse = ISD::FP_TO_SINT;
2652 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2653 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2654 OpToUse = ISD::FP_TO_UINT;
2658 // Otherwise, try a larger type.
2662 // Okay, we found the operation and type to use.
2663 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2665 // Truncate the result of the extended FP_TO_*INT operation to the desired
2667 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2670 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2672 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2673 EVT VT = Op.getValueType();
2674 EVT SHVT = TLI.getShiftAmountTy(VT);
2675 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2676 switch (VT.getSimpleVT().SimpleTy) {
2677 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2679 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2680 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2681 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2683 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2684 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2685 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2686 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2687 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2688 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2689 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2690 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2691 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2693 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2694 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2695 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2696 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2697 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2698 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2699 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2700 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2701 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2702 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2703 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2704 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2705 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2706 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2707 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2708 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2709 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2710 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2711 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2712 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2713 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2717 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2719 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2722 default: llvm_unreachable("Cannot expand this yet!");
2724 EVT VT = Op.getValueType();
2725 EVT ShVT = TLI.getShiftAmountTy(VT);
2726 unsigned Len = VT.getSizeInBits();
2728 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2729 "CTPOP not implemented for this type.");
2731 // This is the "best" algorithm from
2732 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2734 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2735 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2736 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2737 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2739 // v = v - ((v >> 1) & 0x55555555...)
2740 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2741 DAG.getNode(ISD::AND, dl, VT,
2742 DAG.getNode(ISD::SRL, dl, VT, Op,
2743 DAG.getConstant(1, ShVT)),
2745 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2746 Op = DAG.getNode(ISD::ADD, dl, VT,
2747 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2748 DAG.getNode(ISD::AND, dl, VT,
2749 DAG.getNode(ISD::SRL, dl, VT, Op,
2750 DAG.getConstant(2, ShVT)),
2752 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2753 Op = DAG.getNode(ISD::AND, dl, VT,
2754 DAG.getNode(ISD::ADD, dl, VT, Op,
2755 DAG.getNode(ISD::SRL, dl, VT, Op,
2756 DAG.getConstant(4, ShVT))),
2758 // v = (v * 0x01010101...) >> (Len - 8)
2759 Op = DAG.getNode(ISD::SRL, dl, VT,
2760 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2761 DAG.getConstant(Len - 8, ShVT));
2765 case ISD::CTLZ_ZERO_UNDEF:
2766 // This trivially expands to CTLZ.
2767 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2769 // for now, we do this:
2770 // x = x | (x >> 1);
2771 // x = x | (x >> 2);
2773 // x = x | (x >>16);
2774 // x = x | (x >>32); // for 64-bit input
2775 // return popcount(~x);
2777 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2778 EVT VT = Op.getValueType();
2779 EVT ShVT = TLI.getShiftAmountTy(VT);
2780 unsigned len = VT.getSizeInBits();
2781 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2782 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2783 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2784 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2786 Op = DAG.getNOT(dl, Op, VT);
2787 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2789 case ISD::CTTZ_ZERO_UNDEF:
2790 // This trivially expands to CTTZ.
2791 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2793 // for now, we use: { return popcount(~x & (x - 1)); }
2794 // unless the target has ctlz but not ctpop, in which case we use:
2795 // { return 32 - nlz(~x & (x-1)); }
2796 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2797 EVT VT = Op.getValueType();
2798 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2799 DAG.getNOT(dl, Op, VT),
2800 DAG.getNode(ISD::SUB, dl, VT, Op,
2801 DAG.getConstant(1, VT)));
2802 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2803 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2804 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2805 return DAG.getNode(ISD::SUB, dl, VT,
2806 DAG.getConstant(VT.getSizeInBits(), VT),
2807 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2808 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2813 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2814 unsigned Opc = Node->getOpcode();
2815 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2820 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2821 case ISD::ATOMIC_SWAP:
2822 switch (VT.SimpleTy) {
2823 default: llvm_unreachable("Unexpected value type for atomic!");
2824 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2825 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2826 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2827 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2828 case MVT::i128:LC = RTLIB::SYNC_LOCK_TEST_AND_SET_16;break;
2831 case ISD::ATOMIC_CMP_SWAP:
2832 switch (VT.SimpleTy) {
2833 default: llvm_unreachable("Unexpected value type for atomic!");
2834 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2835 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2836 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2837 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2838 case MVT::i128:LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16;break;
2841 case ISD::ATOMIC_LOAD_ADD:
2842 switch (VT.SimpleTy) {
2843 default: llvm_unreachable("Unexpected value type for atomic!");
2844 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2845 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2846 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2847 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2848 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_ADD_16;break;
2851 case ISD::ATOMIC_LOAD_SUB:
2852 switch (VT.SimpleTy) {
2853 default: llvm_unreachable("Unexpected value type for atomic!");
2854 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2855 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2856 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2857 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2858 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_SUB_16;break;
2861 case ISD::ATOMIC_LOAD_AND:
2862 switch (VT.SimpleTy) {
2863 default: llvm_unreachable("Unexpected value type for atomic!");
2864 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2865 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2866 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2867 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2868 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_AND_16;break;
2871 case ISD::ATOMIC_LOAD_OR:
2872 switch (VT.SimpleTy) {
2873 default: llvm_unreachable("Unexpected value type for atomic!");
2874 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2875 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2876 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2877 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2878 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_OR_16;break;
2881 case ISD::ATOMIC_LOAD_XOR:
2882 switch (VT.SimpleTy) {
2883 default: llvm_unreachable("Unexpected value type for atomic!");
2884 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2885 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2886 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2887 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2888 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_XOR_16;break;
2891 case ISD::ATOMIC_LOAD_NAND:
2892 switch (VT.SimpleTy) {
2893 default: llvm_unreachable("Unexpected value type for atomic!");
2894 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2895 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2896 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2897 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2898 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_NAND_16;break;
2901 case ISD::ATOMIC_LOAD_MAX:
2902 switch (VT.SimpleTy) {
2903 default: llvm_unreachable("Unexpected value type for atomic!");
2904 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MAX_1; break;
2905 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MAX_2; break;
2906 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MAX_4; break;
2907 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MAX_8; break;
2908 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MAX_16;break;
2911 case ISD::ATOMIC_LOAD_UMAX:
2912 switch (VT.SimpleTy) {
2913 default: llvm_unreachable("Unexpected value type for atomic!");
2914 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMAX_1; break;
2915 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMAX_2; break;
2916 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMAX_4; break;
2917 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMAX_8; break;
2918 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMAX_16;break;
2921 case ISD::ATOMIC_LOAD_MIN:
2922 switch (VT.SimpleTy) {
2923 default: llvm_unreachable("Unexpected value type for atomic!");
2924 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MIN_1; break;
2925 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MIN_2; break;
2926 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MIN_4; break;
2927 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MIN_8; break;
2928 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MIN_16;break;
2931 case ISD::ATOMIC_LOAD_UMIN:
2932 switch (VT.SimpleTy) {
2933 default: llvm_unreachable("Unexpected value type for atomic!");
2934 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMIN_1; break;
2935 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMIN_2; break;
2936 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMIN_4; break;
2937 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMIN_8; break;
2938 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMIN_16;break;
2943 return ExpandChainLibCall(LC, Node, false);
2946 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2947 SmallVector<SDValue, 8> Results;
2949 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2951 switch (Node->getOpcode()) {
2954 case ISD::CTLZ_ZERO_UNDEF:
2956 case ISD::CTTZ_ZERO_UNDEF:
2957 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2958 Results.push_back(Tmp1);
2961 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2963 case ISD::FRAMEADDR:
2964 case ISD::RETURNADDR:
2965 case ISD::FRAME_TO_ARGS_OFFSET:
2966 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2968 case ISD::FLT_ROUNDS_:
2969 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2971 case ISD::EH_RETURN:
2975 case ISD::EH_SJLJ_LONGJMP:
2976 // If the target didn't expand these, there's nothing to do, so just
2977 // preserve the chain and be done.
2978 Results.push_back(Node->getOperand(0));
2980 case ISD::EH_SJLJ_SETJMP:
2981 // If the target didn't expand this, just return 'zero' and preserve the
2983 Results.push_back(DAG.getConstant(0, MVT::i32));
2984 Results.push_back(Node->getOperand(0));
2986 case ISD::ATOMIC_FENCE: {
2987 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2988 // FIXME: handle "fence singlethread" more efficiently.
2989 TargetLowering::ArgListTy Args;
2991 TargetLowering::CallLoweringInfo CLI(DAG);
2992 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
2993 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
2994 DAG.getExternalSymbol("__sync_synchronize",
2995 TLI.getPointerTy()), std::move(Args), 0);
2997 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2999 Results.push_back(CallResult.second);
3002 case ISD::ATOMIC_LOAD: {
3003 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
3004 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
3005 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3006 SDValue Swap = DAG.getAtomicCmpSwap(
3007 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3008 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
3009 cast<AtomicSDNode>(Node)->getMemOperand(),
3010 cast<AtomicSDNode>(Node)->getOrdering(),
3011 cast<AtomicSDNode>(Node)->getOrdering(),
3012 cast<AtomicSDNode>(Node)->getSynchScope());
3013 Results.push_back(Swap.getValue(0));
3014 Results.push_back(Swap.getValue(1));
3017 case ISD::ATOMIC_STORE: {
3018 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
3019 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
3020 cast<AtomicSDNode>(Node)->getMemoryVT(),
3021 Node->getOperand(0),
3022 Node->getOperand(1), Node->getOperand(2),
3023 cast<AtomicSDNode>(Node)->getMemOperand(),
3024 cast<AtomicSDNode>(Node)->getOrdering(),
3025 cast<AtomicSDNode>(Node)->getSynchScope());
3026 Results.push_back(Swap.getValue(1));
3029 // By default, atomic intrinsics are marked Legal and lowered. Targets
3030 // which don't support them directly, however, may want libcalls, in which
3031 // case they mark them Expand, and we get here.
3032 case ISD::ATOMIC_SWAP:
3033 case ISD::ATOMIC_LOAD_ADD:
3034 case ISD::ATOMIC_LOAD_SUB:
3035 case ISD::ATOMIC_LOAD_AND:
3036 case ISD::ATOMIC_LOAD_OR:
3037 case ISD::ATOMIC_LOAD_XOR:
3038 case ISD::ATOMIC_LOAD_NAND:
3039 case ISD::ATOMIC_LOAD_MIN:
3040 case ISD::ATOMIC_LOAD_MAX:
3041 case ISD::ATOMIC_LOAD_UMIN:
3042 case ISD::ATOMIC_LOAD_UMAX:
3043 case ISD::ATOMIC_CMP_SWAP: {
3044 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
3045 Results.push_back(Tmp.first);
3046 Results.push_back(Tmp.second);
3049 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
3050 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
3051 // splits out the success value as a comparison. Expanding the resulting
3052 // ATOMIC_CMP_SWAP will produce a libcall.
3053 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3054 SDValue Res = DAG.getAtomicCmpSwap(
3055 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3056 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
3057 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
3058 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
3059 cast<AtomicSDNode>(Node)->getFailureOrdering(),
3060 cast<AtomicSDNode>(Node)->getSynchScope());
3062 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
3063 Res, Node->getOperand(2), ISD::SETEQ);
3065 Results.push_back(Res.getValue(0));
3066 Results.push_back(Success);
3067 Results.push_back(Res.getValue(1));
3070 case ISD::DYNAMIC_STACKALLOC:
3071 ExpandDYNAMIC_STACKALLOC(Node, Results);
3073 case ISD::MERGE_VALUES:
3074 for (unsigned i = 0; i < Node->getNumValues(); i++)
3075 Results.push_back(Node->getOperand(i));
3078 EVT VT = Node->getValueType(0);
3080 Results.push_back(DAG.getConstant(0, VT));
3082 assert(VT.isFloatingPoint() && "Unknown value type!");
3083 Results.push_back(DAG.getConstantFP(0, VT));
3088 // If this operation is not supported, lower it to 'abort()' call
3089 TargetLowering::ArgListTy Args;
3090 TargetLowering::CallLoweringInfo CLI(DAG);
3091 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3092 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3093 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
3094 std::move(Args), 0);
3095 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3097 Results.push_back(CallResult.second);
3102 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3103 Node->getValueType(0), dl);
3104 Results.push_back(Tmp1);
3106 case ISD::FP_EXTEND:
3107 Tmp1 = EmitStackConvert(Node->getOperand(0),
3108 Node->getOperand(0).getValueType(),
3109 Node->getValueType(0), dl);
3110 Results.push_back(Tmp1);
3112 case ISD::SIGN_EXTEND_INREG: {
3113 // NOTE: we could fall back on load/store here too for targets without
3114 // SAR. However, it is doubtful that any exist.
3115 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3116 EVT VT = Node->getValueType(0);
3117 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3120 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3121 ExtraVT.getScalarType().getSizeInBits();
3122 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
3123 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3124 Node->getOperand(0), ShiftCst);
3125 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3126 Results.push_back(Tmp1);
3129 case ISD::FP_ROUND_INREG: {
3130 // The only way we can lower this is to turn it into a TRUNCSTORE,
3131 // EXTLOAD pair, targeting a temporary location (a stack slot).
3133 // NOTE: there is a choice here between constantly creating new stack
3134 // slots and always reusing the same one. We currently always create
3135 // new ones, as reuse may inhibit scheduling.
3136 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3137 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3138 Node->getValueType(0), dl);
3139 Results.push_back(Tmp1);
3142 case ISD::SINT_TO_FP:
3143 case ISD::UINT_TO_FP:
3144 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3145 Node->getOperand(0), Node->getValueType(0), dl);
3146 Results.push_back(Tmp1);
3148 case ISD::FP_TO_SINT:
3149 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3150 Results.push_back(Tmp1);
3152 case ISD::FP_TO_UINT: {
3153 SDValue True, False;
3154 EVT VT = Node->getOperand(0).getValueType();
3155 EVT NVT = Node->getValueType(0);
3156 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3157 APInt::getNullValue(VT.getSizeInBits()));
3158 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3159 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3160 Tmp1 = DAG.getConstantFP(apf, VT);
3161 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3162 Node->getOperand(0),
3164 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3165 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3166 DAG.getNode(ISD::FSUB, dl, VT,
3167 Node->getOperand(0), Tmp1));
3168 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3169 DAG.getConstant(x, NVT));
3170 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3171 Results.push_back(Tmp1);
3175 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3176 EVT VT = Node->getValueType(0);
3177 Tmp1 = Node->getOperand(0);
3178 Tmp2 = Node->getOperand(1);
3179 unsigned Align = Node->getConstantOperandVal(3);
3181 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
3182 MachinePointerInfo(V),
3183 false, false, false, 0);
3184 SDValue VAList = VAListLoad;
3186 if (Align > TLI.getMinStackArgumentAlignment()) {
3187 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3189 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3190 DAG.getConstant(Align - 1,
3191 VAList.getValueType()));
3193 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3194 DAG.getConstant(-(int64_t)Align,
3195 VAList.getValueType()));
3198 // Increment the pointer, VAList, to the next vaarg
3199 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3200 DAG.getConstant(TLI.getDataLayout()->
3201 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
3202 VAList.getValueType()));
3203 // Store the incremented VAList to the legalized pointer
3204 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3205 MachinePointerInfo(V), false, false, 0);
3206 // Load the actual argument out of the pointer VAList
3207 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3208 false, false, false, 0));
3209 Results.push_back(Results[0].getValue(1));
3213 // This defaults to loading a pointer from the input and storing it to the
3214 // output, returning the chain.
3215 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3216 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3217 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
3218 Node->getOperand(2), MachinePointerInfo(VS),
3219 false, false, false, 0);
3220 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3221 MachinePointerInfo(VD), false, false, 0);
3222 Results.push_back(Tmp1);
3225 case ISD::EXTRACT_VECTOR_ELT:
3226 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3227 // This must be an access of the only element. Return it.
3228 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3229 Node->getOperand(0));
3231 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3232 Results.push_back(Tmp1);
3234 case ISD::EXTRACT_SUBVECTOR:
3235 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3237 case ISD::INSERT_SUBVECTOR:
3238 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3240 case ISD::CONCAT_VECTORS: {
3241 Results.push_back(ExpandVectorBuildThroughStack(Node));
3244 case ISD::SCALAR_TO_VECTOR:
3245 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3247 case ISD::INSERT_VECTOR_ELT:
3248 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3249 Node->getOperand(1),
3250 Node->getOperand(2), dl));
3252 case ISD::VECTOR_SHUFFLE: {
3253 SmallVector<int, 32> NewMask;
3254 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3256 EVT VT = Node->getValueType(0);
3257 EVT EltVT = VT.getVectorElementType();
3258 SDValue Op0 = Node->getOperand(0);
3259 SDValue Op1 = Node->getOperand(1);
3260 if (!TLI.isTypeLegal(EltVT)) {
3262 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3264 // BUILD_VECTOR operands are allowed to be wider than the element type.
3265 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3267 if (NewEltVT.bitsLT(EltVT)) {
3269 // Convert shuffle node.
3270 // If original node was v4i64 and the new EltVT is i32,
3271 // cast operands to v8i32 and re-build the mask.
3273 // Calculate new VT, the size of the new VT should be equal to original.
3275 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3276 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3277 assert(NewVT.bitsEq(VT));
3279 // cast operands to new VT
3280 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3281 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3283 // Convert the shuffle mask
3284 unsigned int factor =
3285 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3287 // EltVT gets smaller
3290 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3292 for (unsigned fi = 0; fi < factor; ++fi)
3293 NewMask.push_back(Mask[i]);
3296 for (unsigned fi = 0; fi < factor; ++fi)
3297 NewMask.push_back(Mask[i]*factor+fi);
3305 unsigned NumElems = VT.getVectorNumElements();
3306 SmallVector<SDValue, 16> Ops;
3307 for (unsigned i = 0; i != NumElems; ++i) {
3309 Ops.push_back(DAG.getUNDEF(EltVT));
3312 unsigned Idx = Mask[i];
3314 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3316 DAG.getConstant(Idx, TLI.getVectorIdxTy())));
3318 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3320 DAG.getConstant(Idx - NumElems,
3321 TLI.getVectorIdxTy())));
3324 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3325 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3326 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3327 Results.push_back(Tmp1);
3330 case ISD::EXTRACT_ELEMENT: {
3331 EVT OpTy = Node->getOperand(0).getValueType();
3332 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3334 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3335 DAG.getConstant(OpTy.getSizeInBits()/2,
3336 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3337 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3340 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3341 Node->getOperand(0));
3343 Results.push_back(Tmp1);
3346 case ISD::STACKSAVE:
3347 // Expand to CopyFromReg if the target set
3348 // StackPointerRegisterToSaveRestore.
3349 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3350 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3351 Node->getValueType(0)));
3352 Results.push_back(Results[0].getValue(1));
3354 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3355 Results.push_back(Node->getOperand(0));
3358 case ISD::STACKRESTORE:
3359 // Expand to CopyToReg if the target set
3360 // StackPointerRegisterToSaveRestore.
3361 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3362 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3363 Node->getOperand(1)));
3365 Results.push_back(Node->getOperand(0));
3368 case ISD::FCOPYSIGN:
3369 Results.push_back(ExpandFCOPYSIGN(Node));
3372 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3373 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3374 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3375 Node->getOperand(0));
3376 Results.push_back(Tmp1);
3379 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3380 EVT VT = Node->getValueType(0);
3381 Tmp1 = Node->getOperand(0);
3382 Tmp2 = DAG.getConstantFP(0.0, VT);
3383 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3384 Tmp1, Tmp2, ISD::SETUGT);
3385 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3386 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3387 Results.push_back(Tmp1);
3391 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3392 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3393 RTLIB::SQRT_PPCF128));
3397 EVT VT = Node->getValueType(0);
3398 bool isSIN = Node->getOpcode() == ISD::FSIN;
3399 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3400 // fcos which share the same operand and both are used.
3401 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3402 canCombineSinCosLibcall(Node, TLI, TM))
3403 && useSinCos(Node)) {
3404 SDVTList VTs = DAG.getVTList(VT, VT);
3405 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3407 Tmp1 = Tmp1.getValue(1);
3408 Results.push_back(Tmp1);
3410 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3411 RTLIB::SIN_F80, RTLIB::SIN_F128,
3412 RTLIB::SIN_PPCF128));
3414 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3415 RTLIB::COS_F80, RTLIB::COS_F128,
3416 RTLIB::COS_PPCF128));
3421 // Expand into sincos libcall.
3422 ExpandSinCosLibCall(Node, Results);
3425 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3426 RTLIB::LOG_F80, RTLIB::LOG_F128,
3427 RTLIB::LOG_PPCF128));
3430 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3431 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3432 RTLIB::LOG2_PPCF128));
3435 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3436 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3437 RTLIB::LOG10_PPCF128));
3440 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3441 RTLIB::EXP_F80, RTLIB::EXP_F128,
3442 RTLIB::EXP_PPCF128));
3445 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3446 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3447 RTLIB::EXP2_PPCF128));
3450 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3451 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3452 RTLIB::TRUNC_PPCF128));
3455 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3456 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3457 RTLIB::FLOOR_PPCF128));
3460 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3461 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3462 RTLIB::CEIL_PPCF128));
3465 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3466 RTLIB::RINT_F80, RTLIB::RINT_F128,
3467 RTLIB::RINT_PPCF128));
3469 case ISD::FNEARBYINT:
3470 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3471 RTLIB::NEARBYINT_F64,
3472 RTLIB::NEARBYINT_F80,
3473 RTLIB::NEARBYINT_F128,
3474 RTLIB::NEARBYINT_PPCF128));
3477 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3481 RTLIB::ROUND_PPCF128));
3484 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3485 RTLIB::POWI_F80, RTLIB::POWI_F128,
3486 RTLIB::POWI_PPCF128));
3489 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3490 RTLIB::POW_F80, RTLIB::POW_F128,
3491 RTLIB::POW_PPCF128));
3494 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3495 RTLIB::DIV_F80, RTLIB::DIV_F128,
3496 RTLIB::DIV_PPCF128));
3499 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3500 RTLIB::REM_F80, RTLIB::REM_F128,
3501 RTLIB::REM_PPCF128));
3504 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3505 RTLIB::FMA_F80, RTLIB::FMA_F128,
3506 RTLIB::FMA_PPCF128));
3508 case ISD::FP16_TO_FP: {
3509 if (Node->getValueType(0) == MVT::f32) {
3510 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3514 // We can extend to types bigger than f32 in two steps without changing the
3515 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3516 // the option of emitting that before resorting to a libcall.
3518 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3520 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3523 case ISD::FP_TO_FP16: {
3525 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3526 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3527 Results.push_back(ExpandLibCall(LC, Node, false));
3530 case ISD::ConstantFP: {
3531 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3532 // Check to see if this FP immediate is already legal.
3533 // If this is a legal constant, turn it into a TargetConstantFP node.
3534 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3535 Results.push_back(ExpandConstantFP(CFP, true));
3539 EVT VT = Node->getValueType(0);
3540 assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3541 TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
3542 "Don't know how to expand this FP subtraction!");
3543 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3544 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3545 Results.push_back(Tmp1);
3549 EVT VT = Node->getValueType(0);
3550 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3551 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3552 "Don't know how to expand this subtraction!");
3553 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3554 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3555 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3556 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3561 EVT VT = Node->getValueType(0);
3562 bool isSigned = Node->getOpcode() == ISD::SREM;
3563 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3564 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3565 Tmp2 = Node->getOperand(0);
3566 Tmp3 = Node->getOperand(1);
3567 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3568 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3569 // If div is legal, it's better to do the normal expansion
3570 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3571 useDivRem(Node, isSigned, false))) {
3572 SDVTList VTs = DAG.getVTList(VT, VT);
3573 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3574 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3576 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3577 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3578 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3579 } else if (isSigned)
3580 Tmp1 = ExpandIntLibCall(Node, true,
3582 RTLIB::SREM_I16, RTLIB::SREM_I32,
3583 RTLIB::SREM_I64, RTLIB::SREM_I128);
3585 Tmp1 = ExpandIntLibCall(Node, false,
3587 RTLIB::UREM_I16, RTLIB::UREM_I32,
3588 RTLIB::UREM_I64, RTLIB::UREM_I128);
3589 Results.push_back(Tmp1);
3594 bool isSigned = Node->getOpcode() == ISD::SDIV;
3595 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3596 EVT VT = Node->getValueType(0);
3597 SDVTList VTs = DAG.getVTList(VT, VT);
3598 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3599 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3600 useDivRem(Node, isSigned, true)))
3601 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3602 Node->getOperand(1));
3604 Tmp1 = ExpandIntLibCall(Node, true,
3606 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3607 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3609 Tmp1 = ExpandIntLibCall(Node, false,
3611 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3612 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3613 Results.push_back(Tmp1);
3618 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3620 EVT VT = Node->getValueType(0);
3621 SDVTList VTs = DAG.getVTList(VT, VT);
3622 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3623 "If this wasn't legal, it shouldn't have been created!");
3624 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3625 Node->getOperand(1));
3626 Results.push_back(Tmp1.getValue(1));
3631 // Expand into divrem libcall
3632 ExpandDivRemLibCall(Node, Results);
3635 EVT VT = Node->getValueType(0);
3636 SDVTList VTs = DAG.getVTList(VT, VT);
3637 // See if multiply or divide can be lowered using two-result operations.
3638 // We just need the low half of the multiply; try both the signed
3639 // and unsigned forms. If the target supports both SMUL_LOHI and
3640 // UMUL_LOHI, form a preference by checking which forms of plain
3641 // MULH it supports.
3642 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3643 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3644 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3645 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3646 unsigned OpToUse = 0;
3647 if (HasSMUL_LOHI && !HasMULHS) {
3648 OpToUse = ISD::SMUL_LOHI;
3649 } else if (HasUMUL_LOHI && !HasMULHU) {
3650 OpToUse = ISD::UMUL_LOHI;
3651 } else if (HasSMUL_LOHI) {
3652 OpToUse = ISD::SMUL_LOHI;
3653 } else if (HasUMUL_LOHI) {
3654 OpToUse = ISD::UMUL_LOHI;
3657 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3658 Node->getOperand(1)));
3663 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3664 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3665 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3666 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3667 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3668 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) {
3669 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3670 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3671 SDValue Shift = DAG.getConstant(HalfType.getSizeInBits(),
3672 TLI.getShiftAmountTy(HalfType));
3673 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3674 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3678 Tmp1 = ExpandIntLibCall(Node, false,
3680 RTLIB::MUL_I16, RTLIB::MUL_I32,
3681 RTLIB::MUL_I64, RTLIB::MUL_I128);
3682 Results.push_back(Tmp1);
3687 SDValue LHS = Node->getOperand(0);
3688 SDValue RHS = Node->getOperand(1);
3689 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3690 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3692 Results.push_back(Sum);
3693 EVT ResultType = Node->getValueType(1);
3694 EVT OType = getSetCCResultType(Node->getValueType(0));
3696 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3698 // LHSSign -> LHS >= 0
3699 // RHSSign -> RHS >= 0
3700 // SumSign -> Sum >= 0
3703 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3705 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3707 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3708 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3709 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3710 Node->getOpcode() == ISD::SADDO ?
3711 ISD::SETEQ : ISD::SETNE);
3713 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3714 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3716 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3717 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3722 SDValue LHS = Node->getOperand(0);
3723 SDValue RHS = Node->getOperand(1);
3724 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3725 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3727 Results.push_back(Sum);
3729 EVT ResultType = Node->getValueType(1);
3730 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3732 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3733 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3735 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3740 EVT VT = Node->getValueType(0);
3741 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3742 SDValue LHS = Node->getOperand(0);
3743 SDValue RHS = Node->getOperand(1);
3746 static const unsigned Ops[2][3] =
3747 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3748 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3749 bool isSigned = Node->getOpcode() == ISD::SMULO;
3750 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3751 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3752 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3753 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3754 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3756 TopHalf = BottomHalf.getValue(1);
3757 } else if (TLI.isTypeLegal(WideVT)) {
3758 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3759 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3760 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3761 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3762 DAG.getIntPtrConstant(0));
3763 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3764 DAG.getIntPtrConstant(1));
3766 // We can fall back to a libcall with an illegal type for the MUL if we
3767 // have a libcall big enough.
3768 // Also, we can fall back to a division in some cases, but that's a big
3769 // performance hit in the general case.
3770 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3771 if (WideVT == MVT::i16)
3772 LC = RTLIB::MUL_I16;
3773 else if (WideVT == MVT::i32)
3774 LC = RTLIB::MUL_I32;
3775 else if (WideVT == MVT::i64)
3776 LC = RTLIB::MUL_I64;
3777 else if (WideVT == MVT::i128)
3778 LC = RTLIB::MUL_I128;
3779 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3781 // The high part is obtained by SRA'ing all but one of the bits of low
3783 unsigned LoSize = VT.getSizeInBits();
3784 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3785 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3786 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3787 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3789 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3790 // pre-lowered to the correct types. This all depends upon WideVT not
3791 // being a legal type for the architecture and thus has to be split to
3793 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3794 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3795 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3796 DAG.getIntPtrConstant(0));
3797 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3798 DAG.getIntPtrConstant(1));
3799 // Ret is a node with an illegal type. Because such things are not
3800 // generally permitted during this phase of legalization, delete the
3801 // node. The above EXTRACT_ELEMENT nodes should have been folded.
3802 DAG.DeleteNode(Ret.getNode());
3806 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3807 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3808 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3809 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3812 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3813 DAG.getConstant(0, VT), ISD::SETNE);
3815 Results.push_back(BottomHalf);
3816 Results.push_back(TopHalf);
3819 case ISD::BUILD_PAIR: {
3820 EVT PairTy = Node->getValueType(0);
3821 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3822 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3823 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3824 DAG.getConstant(PairTy.getSizeInBits()/2,
3825 TLI.getShiftAmountTy(PairTy)));
3826 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3830 Tmp1 = Node->getOperand(0);
3831 Tmp2 = Node->getOperand(1);
3832 Tmp3 = Node->getOperand(2);
3833 if (Tmp1.getOpcode() == ISD::SETCC) {
3834 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3836 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3838 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3839 DAG.getConstant(0, Tmp1.getValueType()),
3840 Tmp2, Tmp3, ISD::SETNE);
3842 Results.push_back(Tmp1);
3845 SDValue Chain = Node->getOperand(0);
3846 SDValue Table = Node->getOperand(1);
3847 SDValue Index = Node->getOperand(2);
3849 EVT PTy = TLI.getPointerTy();
3851 const DataLayout &TD = *TLI.getDataLayout();
3852 unsigned EntrySize =
3853 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3855 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(),
3856 Index, DAG.getConstant(EntrySize, Index.getValueType()));
3857 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3860 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3861 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3862 MachinePointerInfo::getJumpTable(), MemVT,
3865 if (TM.getRelocationModel() == Reloc::PIC_) {
3866 // For PIC, the sequence is:
3867 // BRIND(load(Jumptable + index) + RelocBase)
3868 // RelocBase can be JumpTable, GOT or some sort of global base.
3869 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3870 TLI.getPICJumpTableRelocBase(Table, DAG));
3872 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3873 Results.push_back(Tmp1);
3877 // Expand brcond's setcc into its constituent parts and create a BR_CC
3879 Tmp1 = Node->getOperand(0);
3880 Tmp2 = Node->getOperand(1);
3881 if (Tmp2.getOpcode() == ISD::SETCC) {
3882 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3883 Tmp1, Tmp2.getOperand(2),
3884 Tmp2.getOperand(0), Tmp2.getOperand(1),
3885 Node->getOperand(2));
3887 // We test only the i1 bit. Skip the AND if UNDEF.
3888 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3889 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3890 DAG.getConstant(1, Tmp2.getValueType()));
3891 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3892 DAG.getCondCode(ISD::SETNE), Tmp3,
3893 DAG.getConstant(0, Tmp3.getValueType()),
3894 Node->getOperand(2));
3896 Results.push_back(Tmp1);
3899 Tmp1 = Node->getOperand(0);
3900 Tmp2 = Node->getOperand(1);
3901 Tmp3 = Node->getOperand(2);
3902 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3903 Tmp3, NeedInvert, dl);
3906 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3907 // condition code, create a new SETCC node.
3909 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3912 // If we expanded the SETCC by inverting the condition code, then wrap
3913 // the existing SETCC in a NOT to restore the intended condition.
3915 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
3917 Results.push_back(Tmp1);
3921 // Otherwise, SETCC for the given comparison type must be completely
3922 // illegal; expand it into a SELECT_CC.
3923 EVT VT = Node->getValueType(0);
3925 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
3926 case TargetLowering::ZeroOrOneBooleanContent:
3927 case TargetLowering::UndefinedBooleanContent:
3930 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3934 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3935 DAG.getConstant(TrueValue, VT), DAG.getConstant(0, VT),
3937 Results.push_back(Tmp1);
3940 case ISD::SELECT_CC: {
3941 Tmp1 = Node->getOperand(0); // LHS
3942 Tmp2 = Node->getOperand(1); // RHS
3943 Tmp3 = Node->getOperand(2); // True
3944 Tmp4 = Node->getOperand(3); // False
3945 EVT VT = Node->getValueType(0);
3946 SDValue CC = Node->getOperand(4);
3947 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
3949 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) {
3950 // If the condition code is legal, then we need to expand this
3951 // node using SETCC and SELECT.
3952 EVT CmpVT = Tmp1.getValueType();
3953 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
3954 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
3956 EVT CCVT = TLI.getSetCCResultType(*DAG.getContext(), CmpVT);
3957 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
3958 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
3962 // SELECT_CC is legal, so the condition code must not be.
3963 bool Legalized = false;
3964 // Try to legalize by inverting the condition. This is for targets that
3965 // might support an ordered version of a condition, but not the unordered
3966 // version (or vice versa).
3967 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
3968 Tmp1.getValueType().isInteger());
3969 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
3970 // Use the new condition code and swap true and false
3972 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
3974 // If The inverse is not legal, then try to swap the arguments using
3975 // the inverse condition code.
3976 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
3977 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
3978 // The swapped inverse condition is legal, so swap true and false,
3981 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
3986 Legalized = LegalizeSetCCCondCode(
3987 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
3990 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
3992 // If we expanded the SETCC by inverting the condition code, then swap
3993 // the True/False operands to match.
3995 std::swap(Tmp3, Tmp4);
3997 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3998 // condition code, create a new SELECT_CC node.
4000 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
4001 Tmp1, Tmp2, Tmp3, Tmp4, CC);
4003 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
4004 CC = DAG.getCondCode(ISD::SETNE);
4005 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
4006 Tmp2, Tmp3, Tmp4, CC);
4009 Results.push_back(Tmp1);
4013 Tmp1 = Node->getOperand(0); // Chain
4014 Tmp2 = Node->getOperand(2); // LHS
4015 Tmp3 = Node->getOperand(3); // RHS
4016 Tmp4 = Node->getOperand(1); // CC
4018 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
4019 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
4021 assert(Legalized && "Can't legalize BR_CC with legal condition!");
4023 // If we expanded the SETCC by inverting the condition code, then wrap
4024 // the existing SETCC in a NOT to restore the intended condition.
4026 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
4028 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
4030 if (Tmp4.getNode()) {
4031 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
4032 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
4034 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
4035 Tmp4 = DAG.getCondCode(ISD::SETNE);
4036 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
4037 Tmp2, Tmp3, Node->getOperand(4));
4039 Results.push_back(Tmp1);
4042 case ISD::BUILD_VECTOR:
4043 Results.push_back(ExpandBUILD_VECTOR(Node));
4048 // Scalarize vector SRA/SRL/SHL.
4049 EVT VT = Node->getValueType(0);
4050 assert(VT.isVector() && "Unable to legalize non-vector shift");
4051 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
4052 unsigned NumElem = VT.getVectorNumElements();
4054 SmallVector<SDValue, 8> Scalars;
4055 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
4056 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
4058 Node->getOperand(0), DAG.getConstant(Idx,
4059 TLI.getVectorIdxTy()));
4060 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
4062 Node->getOperand(1), DAG.getConstant(Idx,
4063 TLI.getVectorIdxTy()));
4064 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
4065 VT.getScalarType(), Ex, Sh));
4068 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Scalars);
4069 ReplaceNode(SDValue(Node, 0), Result);
4072 case ISD::GLOBAL_OFFSET_TABLE:
4073 case ISD::GlobalAddress:
4074 case ISD::GlobalTLSAddress:
4075 case ISD::ExternalSymbol:
4076 case ISD::ConstantPool:
4077 case ISD::JumpTable:
4078 case ISD::INTRINSIC_W_CHAIN:
4079 case ISD::INTRINSIC_WO_CHAIN:
4080 case ISD::INTRINSIC_VOID:
4081 // FIXME: Custom lowering for these operations shouldn't return null!
4085 // Replace the original node with the legalized result.
4086 if (!Results.empty())
4087 ReplaceNode(Node, Results.data());
4090 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4091 SmallVector<SDValue, 8> Results;
4092 MVT OVT = Node->getSimpleValueType(0);
4093 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4094 Node->getOpcode() == ISD::SINT_TO_FP ||
4095 Node->getOpcode() == ISD::SETCC) {
4096 OVT = Node->getOperand(0).getSimpleValueType();
4098 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4100 SDValue Tmp1, Tmp2, Tmp3;
4101 switch (Node->getOpcode()) {
4103 case ISD::CTTZ_ZERO_UNDEF:
4105 case ISD::CTLZ_ZERO_UNDEF:
4107 // Zero extend the argument.
4108 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4109 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4110 // already the correct result.
4111 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4112 if (Node->getOpcode() == ISD::CTTZ) {
4113 // FIXME: This should set a bit in the zero extended value instead.
4114 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
4115 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
4117 Tmp1 = DAG.getSelect(dl, NVT, Tmp2,
4118 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
4119 } else if (Node->getOpcode() == ISD::CTLZ ||
4120 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4121 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4122 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4123 DAG.getConstant(NVT.getSizeInBits() -
4124 OVT.getSizeInBits(), NVT));
4126 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4129 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4130 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4131 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
4132 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
4133 DAG.getConstant(DiffBits, 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));
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)));
4256 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4257 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4258 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4259 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4260 Tmp3, DAG.getIntPtrConstant(0)));
4267 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4268 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4269 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4270 Tmp2, DAG.getIntPtrConstant(0)));
4275 // Replace the original node with the legalized result.
4276 if (!Results.empty())
4277 ReplaceNode(Node, Results.data());
4280 // SelectionDAG::Legalize - This is the entry point for the file.
4282 void SelectionDAG::Legalize() {
4283 AssignTopologicalOrder();
4285 allnodes_iterator LegalizePosition;
4286 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4287 SelectionDAGLegalize Legalizer(*this, LegalizePosition, LegalizedNodes);
4289 // Visit all the nodes. We start in topological order, so that we see
4290 // nodes with their original operands intact. Legalization can produce
4291 // new nodes which may themselves need to be legalized. Iterate until all
4292 // nodes have been legalized.
4294 bool AnyLegalized = false;
4295 for (LegalizePosition = allnodes_end();
4296 LegalizePosition != allnodes_begin(); ) {
4299 SDNode *N = LegalizePosition;
4300 if (LegalizedNodes.insert(N)) {
4301 AnyLegalized = true;
4302 Legalizer.LegalizeOp(N);
4310 // Remove dead nodes now.
4314 bool SelectionDAG::LegalizeOp(SDNode *N,
4315 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4316 allnodes_iterator LegalizePosition(N);
4317 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4318 SelectionDAGLegalize Legalizer(*this, LegalizePosition, LegalizedNodes,
4321 // Directly insert the node in question, and legalize it. This will recurse
4322 // as needed through operands.
4323 LegalizedNodes.insert(N);
4324 Legalizer.LegalizeOp(N);
4326 return LegalizedNodes.count(N);