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/SmallPtrSet.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Triple.h"
18 #include "llvm/CodeGen/Analysis.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineJumpTableInfo.h"
21 #include "llvm/DebugInfo.h"
22 #include "llvm/IR/CallingConv.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/MathExtras.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetFrameLowering.h"
33 #include "llvm/Target/TargetLowering.h"
34 #include "llvm/Target/TargetMachine.h"
37 //===----------------------------------------------------------------------===//
38 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
39 /// hacks on it until the target machine can handle it. This involves
40 /// eliminating value sizes the machine cannot handle (promoting small sizes to
41 /// large sizes or splitting up large values into small values) as well as
42 /// eliminating operations the machine cannot handle.
44 /// This code also does a small amount of optimization and recognition of idioms
45 /// as part of its processing. For example, if a target does not support a
46 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
47 /// will attempt merge setcc and brc instructions into brcc's.
50 class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
51 const TargetMachine &TM;
52 const TargetLowering &TLI;
55 /// LegalizePosition - The iterator for walking through the node list.
56 SelectionDAG::allnodes_iterator LegalizePosition;
58 /// LegalizedNodes - The set of nodes which have already been legalized.
59 SmallPtrSet<SDNode *, 16> LegalizedNodes;
61 EVT getSetCCResultType(EVT VT) const {
62 return TLI.getSetCCResultType(*DAG.getContext(), VT);
65 // Libcall insertion helpers.
68 explicit SelectionDAGLegalize(SelectionDAG &DAG);
73 /// LegalizeOp - Legalizes the given operation.
74 void LegalizeOp(SDNode *Node);
76 SDValue OptimizeFloatStore(StoreSDNode *ST);
78 void LegalizeLoadOps(SDNode *Node);
79 void LegalizeStoreOps(SDNode *Node);
81 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
82 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
83 /// is necessary to spill the vector being inserted into to memory, perform
84 /// the insert there, and then read the result back.
85 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
86 SDValue Idx, SDLoc dl);
87 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
88 SDValue Idx, SDLoc dl);
90 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
91 /// performs the same shuffe in terms of order or result bytes, but on a type
92 /// whose vector element type is narrower than the original shuffle type.
93 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
94 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
95 SDValue N1, SDValue N2,
96 ArrayRef<int> Mask) const;
98 void LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
101 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
102 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
103 unsigned NumOps, bool isSigned, SDLoc dl);
105 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
106 SDNode *Node, bool isSigned);
107 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
108 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
109 RTLIB::Libcall Call_F128,
110 RTLIB::Libcall Call_PPCF128);
111 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
112 RTLIB::Libcall Call_I8,
113 RTLIB::Libcall Call_I16,
114 RTLIB::Libcall Call_I32,
115 RTLIB::Libcall Call_I64,
116 RTLIB::Libcall Call_I128);
117 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
118 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
120 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
121 SDValue ExpandBUILD_VECTOR(SDNode *Node);
122 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
123 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
124 SmallVectorImpl<SDValue> &Results);
125 SDValue ExpandFCOPYSIGN(SDNode *Node);
126 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
128 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
130 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
133 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
134 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
136 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
137 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
138 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
140 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
142 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
144 void ExpandNode(SDNode *Node);
145 void PromoteNode(SDNode *Node);
147 void ForgetNode(SDNode *N) {
148 LegalizedNodes.erase(N);
149 if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
154 // DAGUpdateListener implementation.
155 virtual void NodeDeleted(SDNode *N, SDNode *E) {
158 virtual void NodeUpdated(SDNode *N) {}
160 // Node replacement helpers
161 void ReplacedNode(SDNode *N) {
162 if (N->use_empty()) {
163 DAG.RemoveDeadNode(N);
168 void ReplaceNode(SDNode *Old, SDNode *New) {
169 DAG.ReplaceAllUsesWith(Old, New);
172 void ReplaceNode(SDValue Old, SDValue New) {
173 DAG.ReplaceAllUsesWith(Old, New);
174 ReplacedNode(Old.getNode());
176 void ReplaceNode(SDNode *Old, const SDValue *New) {
177 DAG.ReplaceAllUsesWith(Old, New);
183 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
184 /// performs the same shuffe in terms of order or result bytes, but on a type
185 /// whose vector element type is narrower than the original shuffle type.
186 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
188 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
189 SDValue N1, SDValue N2,
190 ArrayRef<int> Mask) const {
191 unsigned NumMaskElts = VT.getVectorNumElements();
192 unsigned NumDestElts = NVT.getVectorNumElements();
193 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
195 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
197 if (NumEltsGrowth == 1)
198 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
200 SmallVector<int, 8> NewMask;
201 for (unsigned i = 0; i != NumMaskElts; ++i) {
203 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
205 NewMask.push_back(-1);
207 NewMask.push_back(Idx * NumEltsGrowth + j);
210 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
211 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
212 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
215 SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
216 : SelectionDAG::DAGUpdateListener(dag),
217 TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
221 void SelectionDAGLegalize::LegalizeDAG() {
222 DAG.AssignTopologicalOrder();
224 // Visit all the nodes. We start in topological order, so that we see
225 // nodes with their original operands intact. Legalization can produce
226 // new nodes which may themselves need to be legalized. Iterate until all
227 // nodes have been legalized.
229 bool AnyLegalized = false;
230 for (LegalizePosition = DAG.allnodes_end();
231 LegalizePosition != DAG.allnodes_begin(); ) {
234 SDNode *N = LegalizePosition;
235 if (LegalizedNodes.insert(N)) {
245 // Remove dead nodes now.
246 DAG.RemoveDeadNodes();
249 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
250 /// a load from the constant pool.
252 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
256 // If a FP immediate is precise when represented as a float and if the
257 // target can do an extending load from float to double, we put it into
258 // the constant pool as a float, even if it's is statically typed as a
259 // double. This shrinks FP constants and canonicalizes them for targets where
260 // an FP extending load is the same cost as a normal load (such as on the x87
261 // fp stack or PPC FP unit).
262 EVT VT = CFP->getValueType(0);
263 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
265 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
266 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
267 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
272 while (SVT != MVT::f32) {
273 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
274 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
275 // Only do this if the target has a native EXTLOAD instruction from
277 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
278 TLI.ShouldShrinkFPConstant(OrigVT)) {
279 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
280 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
286 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
287 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
290 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
292 CPIdx, MachinePointerInfo::getConstantPool(),
293 VT, false, false, Alignment);
297 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
298 MachinePointerInfo::getConstantPool(), false, false, false,
303 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
304 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
305 const TargetLowering &TLI,
306 SelectionDAGLegalize *DAGLegalize) {
307 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
308 "unaligned indexed stores not implemented!");
309 SDValue Chain = ST->getChain();
310 SDValue Ptr = ST->getBasePtr();
311 SDValue Val = ST->getValue();
312 EVT VT = Val.getValueType();
313 int Alignment = ST->getAlignment();
315 if (ST->getMemoryVT().isFloatingPoint() ||
316 ST->getMemoryVT().isVector()) {
317 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
318 if (TLI.isTypeLegal(intVT)) {
319 // Expand to a bitconvert of the value to the integer type of the
320 // same size, then a (misaligned) int store.
321 // FIXME: Does not handle truncating floating point stores!
322 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
323 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
324 ST->isVolatile(), ST->isNonTemporal(), Alignment);
325 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
328 // Do a (aligned) store to a stack slot, then copy from the stack slot
329 // to the final destination using (unaligned) integer loads and stores.
330 EVT StoredVT = ST->getMemoryVT();
332 TLI.getRegisterType(*DAG.getContext(),
333 EVT::getIntegerVT(*DAG.getContext(),
334 StoredVT.getSizeInBits()));
335 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
336 unsigned RegBytes = RegVT.getSizeInBits() / 8;
337 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
339 // Make sure the stack slot is also aligned for the register type.
340 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
342 // Perform the original store, only redirected to the stack slot.
343 SDValue Store = DAG.getTruncStore(Chain, dl,
344 Val, StackPtr, MachinePointerInfo(),
345 StoredVT, false, false, 0);
346 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
347 SmallVector<SDValue, 8> Stores;
350 // Do all but one copies using the full register width.
351 for (unsigned i = 1; i < NumRegs; i++) {
352 // Load one integer register's worth from the stack slot.
353 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
354 MachinePointerInfo(),
355 false, false, false, 0);
356 // Store it to the final location. Remember the store.
357 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
358 ST->getPointerInfo().getWithOffset(Offset),
359 ST->isVolatile(), ST->isNonTemporal(),
360 MinAlign(ST->getAlignment(), Offset)));
361 // Increment the pointers.
363 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
365 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
368 // The last store may be partial. Do a truncating store. On big-endian
369 // machines this requires an extending load from the stack slot to ensure
370 // that the bits are in the right place.
371 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
372 8 * (StoredBytes - Offset));
374 // Load from the stack slot.
375 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
376 MachinePointerInfo(),
377 MemVT, false, false, 0);
379 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
381 .getWithOffset(Offset),
382 MemVT, ST->isVolatile(),
384 MinAlign(ST->getAlignment(), Offset)));
385 // The order of the stores doesn't matter - say it with a TokenFactor.
387 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
389 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
392 assert(ST->getMemoryVT().isInteger() &&
393 !ST->getMemoryVT().isVector() &&
394 "Unaligned store of unknown type.");
395 // Get the half-size VT
396 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
397 int NumBits = NewStoredVT.getSizeInBits();
398 int IncrementSize = NumBits / 8;
400 // Divide the stored value in two parts.
401 SDValue ShiftAmount = DAG.getConstant(NumBits,
402 TLI.getShiftAmountTy(Val.getValueType()));
404 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
406 // Store the two parts
407 SDValue Store1, Store2;
408 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
409 ST->getPointerInfo(), NewStoredVT,
410 ST->isVolatile(), ST->isNonTemporal(), Alignment);
411 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
412 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
413 Alignment = MinAlign(Alignment, IncrementSize);
414 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
415 ST->getPointerInfo().getWithOffset(IncrementSize),
416 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
420 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
421 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
424 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
426 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
427 const TargetLowering &TLI,
428 SDValue &ValResult, SDValue &ChainResult) {
429 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
430 "unaligned indexed loads not implemented!");
431 SDValue Chain = LD->getChain();
432 SDValue Ptr = LD->getBasePtr();
433 EVT VT = LD->getValueType(0);
434 EVT LoadedVT = LD->getMemoryVT();
436 if (VT.isFloatingPoint() || VT.isVector()) {
437 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
438 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
439 // Expand to a (misaligned) integer load of the same size,
440 // then bitconvert to floating point or vector.
441 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getPointerInfo(),
444 LD->isInvariant(), LD->getAlignment());
445 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
447 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
448 ISD::ANY_EXTEND, dl, VT, Result);
455 // Copy the value to a (aligned) stack slot using (unaligned) integer
456 // loads and stores, then do a (aligned) load from the stack slot.
457 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
458 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
459 unsigned RegBytes = RegVT.getSizeInBits() / 8;
460 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
462 // Make sure the stack slot is also aligned for the register type.
463 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
465 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
466 SmallVector<SDValue, 8> Stores;
467 SDValue StackPtr = StackBase;
470 // Do all but one copies using the full register width.
471 for (unsigned i = 1; i < NumRegs; i++) {
472 // Load one integer register's worth from the original location.
473 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
474 LD->getPointerInfo().getWithOffset(Offset),
475 LD->isVolatile(), LD->isNonTemporal(),
477 MinAlign(LD->getAlignment(), Offset));
478 // Follow the load with a store to the stack slot. Remember the store.
479 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
480 MachinePointerInfo(), false, false, 0));
481 // Increment the pointers.
483 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
484 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
488 // The last copy may be partial. Do an extending load.
489 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
490 8 * (LoadedBytes - Offset));
491 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
492 LD->getPointerInfo().getWithOffset(Offset),
493 MemVT, LD->isVolatile(),
495 MinAlign(LD->getAlignment(), Offset));
496 // Follow the load with a store to the stack slot. Remember the store.
497 // On big-endian machines this requires a truncating store to ensure
498 // that the bits end up in the right place.
499 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
500 MachinePointerInfo(), MemVT,
503 // The order of the stores doesn't matter - say it with a TokenFactor.
504 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
507 // Finally, perform the original load only redirected to the stack slot.
508 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
509 MachinePointerInfo(), LoadedVT, false, false, 0);
511 // Callers expect a MERGE_VALUES node.
516 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
517 "Unaligned load of unsupported type.");
519 // Compute the new VT that is half the size of the old one. This is an
521 unsigned NumBits = LoadedVT.getSizeInBits();
523 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
526 unsigned Alignment = LD->getAlignment();
527 unsigned IncrementSize = NumBits / 8;
528 ISD::LoadExtType HiExtType = LD->getExtensionType();
530 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
531 if (HiExtType == ISD::NON_EXTLOAD)
532 HiExtType = ISD::ZEXTLOAD;
534 // Load the value in two parts
536 if (TLI.isLittleEndian()) {
537 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
538 NewLoadedVT, LD->isVolatile(),
539 LD->isNonTemporal(), Alignment);
540 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
541 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
542 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
543 LD->getPointerInfo().getWithOffset(IncrementSize),
544 NewLoadedVT, LD->isVolatile(),
545 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
547 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
548 NewLoadedVT, LD->isVolatile(),
549 LD->isNonTemporal(), Alignment);
550 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
551 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
552 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
553 LD->getPointerInfo().getWithOffset(IncrementSize),
554 NewLoadedVT, LD->isVolatile(),
555 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
558 // aggregate the two parts
559 SDValue ShiftAmount = DAG.getConstant(NumBits,
560 TLI.getShiftAmountTy(Hi.getValueType()));
561 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
562 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
564 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
571 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
572 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
573 /// is necessary to spill the vector being inserted into to memory, perform
574 /// the insert there, and then read the result back.
575 SDValue SelectionDAGLegalize::
576 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
582 // If the target doesn't support this, we have to spill the input vector
583 // to a temporary stack slot, update the element, then reload it. This is
584 // badness. We could also load the value into a vector register (either
585 // with a "move to register" or "extload into register" instruction, then
586 // permute it into place, if the idx is a constant and if the idx is
587 // supported by the target.
588 EVT VT = Tmp1.getValueType();
589 EVT EltVT = VT.getVectorElementType();
590 EVT IdxVT = Tmp3.getValueType();
591 EVT PtrVT = TLI.getPointerTy();
592 SDValue StackPtr = DAG.CreateStackTemporary(VT);
594 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
597 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
598 MachinePointerInfo::getFixedStack(SPFI),
601 // Truncate or zero extend offset to target pointer type.
602 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
603 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
604 // Add the offset to the index.
605 unsigned EltSize = EltVT.getSizeInBits()/8;
606 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
607 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
608 // Store the scalar value.
609 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
611 // Load the updated vector.
612 return DAG.getLoad(VT, dl, Ch, StackPtr,
613 MachinePointerInfo::getFixedStack(SPFI), false, false,
618 SDValue SelectionDAGLegalize::
619 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
620 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
621 // SCALAR_TO_VECTOR requires that the type of the value being inserted
622 // match the element type of the vector being created, except for
623 // integers in which case the inserted value can be over width.
624 EVT EltVT = Vec.getValueType().getVectorElementType();
625 if (Val.getValueType() == EltVT ||
626 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
627 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
628 Vec.getValueType(), Val);
630 unsigned NumElts = Vec.getValueType().getVectorNumElements();
631 // We generate a shuffle of InVec and ScVec, so the shuffle mask
632 // should be 0,1,2,3,4,5... with the appropriate element replaced with
634 SmallVector<int, 8> ShufOps;
635 for (unsigned i = 0; i != NumElts; ++i)
636 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
638 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
642 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
645 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
646 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
647 // FIXME: We shouldn't do this for TargetConstantFP's.
648 // FIXME: move this to the DAG Combiner! Note that we can't regress due
649 // to phase ordering between legalized code and the dag combiner. This
650 // probably means that we need to integrate dag combiner and legalizer
652 // We generally can't do this one for long doubles.
653 SDValue Chain = ST->getChain();
654 SDValue Ptr = ST->getBasePtr();
655 unsigned Alignment = ST->getAlignment();
656 bool isVolatile = ST->isVolatile();
657 bool isNonTemporal = ST->isNonTemporal();
659 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
660 if (CFP->getValueType(0) == MVT::f32 &&
661 TLI.isTypeLegal(MVT::i32)) {
662 SDValue Con = DAG.getConstant(CFP->getValueAPF().
663 bitcastToAPInt().zextOrTrunc(32),
665 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
666 isVolatile, isNonTemporal, Alignment);
669 if (CFP->getValueType(0) == MVT::f64) {
670 // If this target supports 64-bit registers, do a single 64-bit store.
671 if (TLI.isTypeLegal(MVT::i64)) {
672 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
673 zextOrTrunc(64), MVT::i64);
674 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
675 isVolatile, isNonTemporal, Alignment);
678 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
679 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
680 // stores. If the target supports neither 32- nor 64-bits, this
681 // xform is certainly not worth it.
682 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
683 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
684 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
685 if (TLI.isBigEndian()) std::swap(Lo, Hi);
687 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
688 isNonTemporal, Alignment);
689 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
690 DAG.getIntPtrConstant(4));
691 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
692 ST->getPointerInfo().getWithOffset(4),
693 isVolatile, isNonTemporal, MinAlign(Alignment, 4U));
695 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
699 return SDValue(0, 0);
702 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
703 StoreSDNode *ST = cast<StoreSDNode>(Node);
704 SDValue Chain = ST->getChain();
705 SDValue Ptr = ST->getBasePtr();
708 unsigned Alignment = ST->getAlignment();
709 bool isVolatile = ST->isVolatile();
710 bool isNonTemporal = ST->isNonTemporal();
712 if (!ST->isTruncatingStore()) {
713 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
714 ReplaceNode(ST, OptStore);
719 SDValue Value = ST->getValue();
720 MVT VT = Value.getSimpleValueType();
721 switch (TLI.getOperationAction(ISD::STORE, VT)) {
722 default: llvm_unreachable("This action is not supported yet!");
723 case TargetLowering::Legal:
724 // If this is an unaligned store and the target doesn't support it,
726 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
727 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
728 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
729 if (ST->getAlignment() < ABIAlignment)
730 ExpandUnalignedStore(cast<StoreSDNode>(Node),
734 case TargetLowering::Custom: {
735 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
737 ReplaceNode(SDValue(Node, 0), Res);
740 case TargetLowering::Promote: {
741 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
742 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
743 "Can only promote stores to same size type");
744 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
746 DAG.getStore(Chain, dl, Value, Ptr,
747 ST->getPointerInfo(), isVolatile,
748 isNonTemporal, Alignment);
749 ReplaceNode(SDValue(Node, 0), Result);
756 SDValue Value = ST->getValue();
758 EVT StVT = ST->getMemoryVT();
759 unsigned StWidth = StVT.getSizeInBits();
761 if (StWidth != StVT.getStoreSizeInBits()) {
762 // Promote to a byte-sized store with upper bits zero if not
763 // storing an integral number of bytes. For example, promote
764 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
765 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
766 StVT.getStoreSizeInBits());
767 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
769 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
770 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);
793 // Store the remaining ExtraWidth bits.
794 IncrementSize = RoundWidth / 8;
795 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
796 DAG.getIntPtrConstant(IncrementSize));
797 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
798 DAG.getConstant(RoundWidth,
799 TLI.getShiftAmountTy(Value.getValueType())));
800 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
801 ST->getPointerInfo().getWithOffset(IncrementSize),
802 ExtraVT, isVolatile, isNonTemporal,
803 MinAlign(Alignment, IncrementSize));
805 // Big endian - avoid unaligned stores.
806 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
807 // Store the top RoundWidth bits.
808 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
809 DAG.getConstant(ExtraWidth,
810 TLI.getShiftAmountTy(Value.getValueType())));
811 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
812 RoundVT, isVolatile, isNonTemporal, Alignment);
814 // Store the remaining ExtraWidth bits.
815 IncrementSize = RoundWidth / 8;
816 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
817 DAG.getIntPtrConstant(IncrementSize));
818 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
819 ST->getPointerInfo().getWithOffset(IncrementSize),
820 ExtraVT, isVolatile, isNonTemporal,
821 MinAlign(Alignment, IncrementSize));
824 // The order of the stores doesn't matter.
825 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
826 ReplaceNode(SDValue(Node, 0), Result);
828 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
829 StVT.getSimpleVT())) {
830 default: llvm_unreachable("This action is not supported yet!");
831 case TargetLowering::Legal:
832 // If this is an unaligned store and the target doesn't support it,
834 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
835 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
836 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
837 if (ST->getAlignment() < ABIAlignment)
838 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
841 case TargetLowering::Custom: {
842 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
844 ReplaceNode(SDValue(Node, 0), Res);
847 case TargetLowering::Expand:
848 assert(!StVT.isVector() &&
849 "Vector Stores are handled in LegalizeVectorOps");
851 // TRUNCSTORE:i16 i32 -> STORE i16
852 assert(TLI.isTypeLegal(StVT) &&
853 "Do not know how to expand this store!");
854 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
856 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
857 isVolatile, isNonTemporal, Alignment);
858 ReplaceNode(SDValue(Node, 0), Result);
865 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
866 LoadSDNode *LD = cast<LoadSDNode>(Node);
867 SDValue Chain = LD->getChain(); // The chain.
868 SDValue Ptr = LD->getBasePtr(); // The base pointer.
869 SDValue Value; // The value returned by the load op.
872 ISD::LoadExtType ExtType = LD->getExtensionType();
873 if (ExtType == ISD::NON_EXTLOAD) {
874 MVT VT = Node->getSimpleValueType(0);
875 SDValue RVal = SDValue(Node, 0);
876 SDValue RChain = SDValue(Node, 1);
878 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
879 default: llvm_unreachable("This action is not supported yet!");
880 case TargetLowering::Legal:
881 // If this is an unaligned load and the target doesn't support it,
883 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
884 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
885 unsigned ABIAlignment =
886 TLI.getDataLayout()->getABITypeAlignment(Ty);
887 if (LD->getAlignment() < ABIAlignment){
888 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
892 case TargetLowering::Custom: {
893 SDValue Res = TLI.LowerOperation(RVal, DAG);
896 RChain = Res.getValue(1);
900 case TargetLowering::Promote: {
901 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
902 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
903 "Can only promote loads to same size type");
905 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getPointerInfo(),
906 LD->isVolatile(), LD->isNonTemporal(),
907 LD->isInvariant(), LD->getAlignment());
908 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
909 RChain = Res.getValue(1);
913 if (RChain.getNode() != Node) {
914 assert(RVal.getNode() != Node && "Load must be completely replaced");
915 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
916 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
922 EVT SrcVT = LD->getMemoryVT();
923 unsigned SrcWidth = SrcVT.getSizeInBits();
924 unsigned Alignment = LD->getAlignment();
925 bool isVolatile = LD->isVolatile();
926 bool isNonTemporal = LD->isNonTemporal();
928 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
929 // Some targets pretend to have an i1 loading operation, and actually
930 // load an i8. This trick is correct for ZEXTLOAD because the top 7
931 // bits are guaranteed to be zero; it helps the optimizers understand
932 // that these bits are zero. It is also useful for EXTLOAD, since it
933 // tells the optimizers that those bits are undefined. It would be
934 // nice to have an effective generic way of getting these benefits...
935 // Until such a way is found, don't insist on promoting i1 here.
937 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
938 // Promote to a byte-sized load if not loading an integral number of
939 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
940 unsigned NewWidth = SrcVT.getStoreSizeInBits();
941 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
944 // The extra bits are guaranteed to be zero, since we stored them that
945 // way. A zext load from NVT thus automatically gives zext from SrcVT.
947 ISD::LoadExtType NewExtType =
948 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
951 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
952 Chain, Ptr, LD->getPointerInfo(),
953 NVT, isVolatile, isNonTemporal, Alignment);
955 Ch = Result.getValue(1); // The chain.
957 if (ExtType == ISD::SEXTLOAD)
958 // Having the top bits zero doesn't help when sign extending.
959 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
960 Result.getValueType(),
961 Result, DAG.getValueType(SrcVT));
962 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
963 // All the top bits are guaranteed to be zero - inform the optimizers.
964 Result = DAG.getNode(ISD::AssertZext, dl,
965 Result.getValueType(), Result,
966 DAG.getValueType(SrcVT));
970 } else if (SrcWidth & (SrcWidth - 1)) {
971 // If not loading a power-of-2 number of bits, expand as two loads.
972 assert(!SrcVT.isVector() && "Unsupported extload!");
973 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
974 assert(RoundWidth < SrcWidth);
975 unsigned ExtraWidth = SrcWidth - RoundWidth;
976 assert(ExtraWidth < RoundWidth);
977 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
978 "Load size not an integral number of bytes!");
979 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
980 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
982 unsigned IncrementSize;
984 if (TLI.isLittleEndian()) {
985 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
986 // Load the bottom RoundWidth bits.
987 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
989 LD->getPointerInfo(), RoundVT, isVolatile,
990 isNonTemporal, Alignment);
992 // Load the remaining ExtraWidth bits.
993 IncrementSize = RoundWidth / 8;
994 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
995 DAG.getIntPtrConstant(IncrementSize));
996 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
997 LD->getPointerInfo().getWithOffset(IncrementSize),
998 ExtraVT, isVolatile, isNonTemporal,
999 MinAlign(Alignment, IncrementSize));
1001 // Build a factor node to remember that this load is independent of
1003 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1006 // Move the top bits to the right place.
1007 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1008 DAG.getConstant(RoundWidth,
1009 TLI.getShiftAmountTy(Hi.getValueType())));
1011 // Join the hi and lo parts.
1012 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1014 // Big endian - avoid unaligned loads.
1015 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1016 // Load the top RoundWidth bits.
1017 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1018 LD->getPointerInfo(), RoundVT, isVolatile,
1019 isNonTemporal, Alignment);
1021 // Load the remaining ExtraWidth bits.
1022 IncrementSize = RoundWidth / 8;
1023 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1024 DAG.getIntPtrConstant(IncrementSize));
1025 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1026 dl, Node->getValueType(0), Chain, Ptr,
1027 LD->getPointerInfo().getWithOffset(IncrementSize),
1028 ExtraVT, isVolatile, isNonTemporal,
1029 MinAlign(Alignment, IncrementSize));
1031 // Build a factor node to remember that this load is independent of
1033 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1036 // Move the top bits to the right place.
1037 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1038 DAG.getConstant(ExtraWidth,
1039 TLI.getShiftAmountTy(Hi.getValueType())));
1041 // Join the hi and lo parts.
1042 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1047 bool isCustom = false;
1048 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1049 default: llvm_unreachable("This action is not supported yet!");
1050 case TargetLowering::Custom:
1053 case TargetLowering::Legal: {
1054 Value = SDValue(Node, 0);
1055 Chain = SDValue(Node, 1);
1058 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1059 if (Res.getNode()) {
1061 Chain = Res.getValue(1);
1064 // If this is an unaligned load and the target doesn't support it,
1066 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
1068 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1069 unsigned ABIAlignment =
1070 TLI.getDataLayout()->getABITypeAlignment(Ty);
1071 if (LD->getAlignment() < ABIAlignment){
1072 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1073 DAG, TLI, Value, Chain);
1079 case TargetLowering::Expand:
1080 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
1081 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1082 LD->getPointerInfo(),
1083 LD->isVolatile(), LD->isNonTemporal(),
1084 LD->isInvariant(), LD->getAlignment());
1088 ExtendOp = (SrcVT.isFloatingPoint() ?
1089 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1091 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1092 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1093 default: llvm_unreachable("Unexpected extend load type!");
1095 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1096 Chain = Load.getValue(1);
1100 assert(!SrcVT.isVector() &&
1101 "Vector Loads are handled in LegalizeVectorOps");
1103 // FIXME: This does not work for vectors on most targets. Sign- and
1104 // zero-extend operations are currently folded into extending loads,
1105 // whether they are legal or not, and then we end up here without any
1106 // support for legalizing them.
1107 assert(ExtType != ISD::EXTLOAD &&
1108 "EXTLOAD should always be supported!");
1109 // Turn the unsupported load into an EXTLOAD followed by an explicit
1110 // zero/sign extend inreg.
1111 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
1112 Chain, Ptr, LD->getPointerInfo(), SrcVT,
1113 LD->isVolatile(), LD->isNonTemporal(),
1114 LD->getAlignment());
1116 if (ExtType == ISD::SEXTLOAD)
1117 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1118 Result.getValueType(),
1119 Result, DAG.getValueType(SrcVT));
1121 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1123 Chain = Result.getValue(1);
1128 // Since loads produce two values, make sure to remember that we legalized
1130 if (Chain.getNode() != Node) {
1131 assert(Value.getNode() != Node && "Load must be completely replaced");
1132 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1133 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1138 /// LegalizeOp - Return a legal replacement for the given operation, with
1139 /// all legal operands.
1140 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1141 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1144 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1145 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1146 TargetLowering::TypeLegal &&
1147 "Unexpected illegal type!");
1149 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1150 assert((TLI.getTypeAction(*DAG.getContext(),
1151 Node->getOperand(i).getValueType()) ==
1152 TargetLowering::TypeLegal ||
1153 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1154 "Unexpected illegal type!");
1156 // Figure out the correct action; the way to query this varies by opcode
1157 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1158 bool SimpleFinishLegalizing = true;
1159 switch (Node->getOpcode()) {
1160 case ISD::INTRINSIC_W_CHAIN:
1161 case ISD::INTRINSIC_WO_CHAIN:
1162 case ISD::INTRINSIC_VOID:
1163 case ISD::STACKSAVE:
1164 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1167 Action = TLI.getOperationAction(Node->getOpcode(),
1168 Node->getValueType(0));
1169 if (Action != TargetLowering::Promote)
1170 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1172 case ISD::SINT_TO_FP:
1173 case ISD::UINT_TO_FP:
1174 case ISD::EXTRACT_VECTOR_ELT:
1175 Action = TLI.getOperationAction(Node->getOpcode(),
1176 Node->getOperand(0).getValueType());
1178 case ISD::FP_ROUND_INREG:
1179 case ISD::SIGN_EXTEND_INREG: {
1180 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1181 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1184 case ISD::ATOMIC_STORE: {
1185 Action = TLI.getOperationAction(Node->getOpcode(),
1186 Node->getOperand(2).getValueType());
1189 case ISD::SELECT_CC:
1192 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1193 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1194 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1195 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1196 ISD::CondCode CCCode =
1197 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1198 Action = TLI.getCondCodeAction(CCCode, OpVT);
1199 if (Action == TargetLowering::Legal) {
1200 if (Node->getOpcode() == ISD::SELECT_CC)
1201 Action = TLI.getOperationAction(Node->getOpcode(),
1202 Node->getValueType(0));
1204 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1210 // FIXME: Model these properly. LOAD and STORE are complicated, and
1211 // STORE expects the unlegalized operand in some cases.
1212 SimpleFinishLegalizing = false;
1214 case ISD::CALLSEQ_START:
1215 case ISD::CALLSEQ_END:
1216 // FIXME: This shouldn't be necessary. These nodes have special properties
1217 // dealing with the recursive nature of legalization. Removing this
1218 // special case should be done as part of making LegalizeDAG non-recursive.
1219 SimpleFinishLegalizing = false;
1221 case ISD::EXTRACT_ELEMENT:
1222 case ISD::FLT_ROUNDS_:
1230 case ISD::MERGE_VALUES:
1231 case ISD::EH_RETURN:
1232 case ISD::FRAME_TO_ARGS_OFFSET:
1233 case ISD::EH_SJLJ_SETJMP:
1234 case ISD::EH_SJLJ_LONGJMP:
1235 // These operations lie about being legal: when they claim to be legal,
1236 // they should actually be expanded.
1237 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1238 if (Action == TargetLowering::Legal)
1239 Action = TargetLowering::Expand;
1241 case ISD::INIT_TRAMPOLINE:
1242 case ISD::ADJUST_TRAMPOLINE:
1243 case ISD::FRAMEADDR:
1244 case ISD::RETURNADDR:
1245 // These operations lie about being legal: when they claim to be legal,
1246 // they should actually be custom-lowered.
1247 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1248 if (Action == TargetLowering::Legal)
1249 Action = TargetLowering::Custom;
1251 case ISD::DEBUGTRAP:
1252 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1253 if (Action == TargetLowering::Expand) {
1254 // replace ISD::DEBUGTRAP with ISD::TRAP
1256 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1257 Node->getOperand(0));
1258 ReplaceNode(Node, NewVal.getNode());
1259 LegalizeOp(NewVal.getNode());
1265 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1266 Action = TargetLowering::Legal;
1268 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1273 if (SimpleFinishLegalizing) {
1274 SDNode *NewNode = Node;
1275 switch (Node->getOpcode()) {
1282 // Legalizing shifts/rotates requires adjusting the shift amount
1283 // to the appropriate width.
1284 if (!Node->getOperand(1).getValueType().isVector()) {
1286 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1287 Node->getOperand(1));
1288 HandleSDNode Handle(SAO);
1289 LegalizeOp(SAO.getNode());
1290 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1294 case ISD::SRL_PARTS:
1295 case ISD::SRA_PARTS:
1296 case ISD::SHL_PARTS:
1297 // Legalizing shifts/rotates requires adjusting the shift amount
1298 // to the appropriate width.
1299 if (!Node->getOperand(2).getValueType().isVector()) {
1301 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1302 Node->getOperand(2));
1303 HandleSDNode Handle(SAO);
1304 LegalizeOp(SAO.getNode());
1305 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1306 Node->getOperand(1),
1312 if (NewNode != Node) {
1313 DAG.ReplaceAllUsesWith(Node, NewNode);
1314 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1315 DAG.TransferDbgValues(SDValue(Node, i), SDValue(NewNode, i));
1320 case TargetLowering::Legal:
1322 case TargetLowering::Custom: {
1323 // FIXME: The handling for custom lowering with multiple results is
1325 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1326 if (Res.getNode()) {
1327 SmallVector<SDValue, 8> ResultVals;
1328 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
1330 ResultVals.push_back(Res);
1332 ResultVals.push_back(Res.getValue(i));
1334 if (Res.getNode() != Node || Res.getResNo() != 0) {
1335 DAG.ReplaceAllUsesWith(Node, ResultVals.data());
1336 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1337 DAG.TransferDbgValues(SDValue(Node, i), ResultVals[i]);
1344 case TargetLowering::Expand:
1347 case TargetLowering::Promote:
1353 switch (Node->getOpcode()) {
1360 llvm_unreachable("Do not know how to legalize this operator!");
1362 case ISD::CALLSEQ_START:
1363 case ISD::CALLSEQ_END:
1366 return LegalizeLoadOps(Node);
1369 return LegalizeStoreOps(Node);
1374 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1375 SDValue Vec = Op.getOperand(0);
1376 SDValue Idx = Op.getOperand(1);
1378 // Store the value to a temporary stack slot, then LOAD the returned part.
1379 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1380 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1381 MachinePointerInfo(), false, false, 0);
1383 // Add the offset to the index.
1385 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1386 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1387 DAG.getConstant(EltSize, Idx.getValueType()));
1389 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1390 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1392 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1394 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1396 if (Op.getValueType().isVector())
1397 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1398 false, false, false, 0);
1399 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1400 MachinePointerInfo(),
1401 Vec.getValueType().getVectorElementType(),
1405 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1406 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1408 SDValue Vec = Op.getOperand(0);
1409 SDValue Part = Op.getOperand(1);
1410 SDValue Idx = Op.getOperand(2);
1413 // Store the value to a temporary stack slot, then LOAD the returned part.
1415 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1416 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1417 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1419 // First store the whole vector.
1420 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1423 // Then store the inserted part.
1425 // Add the offset to the index.
1427 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1429 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1430 DAG.getConstant(EltSize, Idx.getValueType()));
1432 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1433 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1435 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1437 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1440 // Store the subvector.
1441 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1442 MachinePointerInfo(), false, false, 0);
1444 // Finally, load the updated vector.
1445 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1446 false, false, false, 0);
1449 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1450 // We can't handle this case efficiently. Allocate a sufficiently
1451 // aligned object on the stack, store each element into it, then load
1452 // the result as a vector.
1453 // Create the stack frame object.
1454 EVT VT = Node->getValueType(0);
1455 EVT EltVT = VT.getVectorElementType();
1457 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1458 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1459 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1461 // Emit a store of each element to the stack slot.
1462 SmallVector<SDValue, 8> Stores;
1463 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1464 // Store (in the right endianness) the elements to memory.
1465 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1466 // Ignore undef elements.
1467 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1469 unsigned Offset = TypeByteSize*i;
1471 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1472 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1474 // If the destination vector element type is narrower than the source
1475 // element type, only store the bits necessary.
1476 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1477 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1478 Node->getOperand(i), Idx,
1479 PtrInfo.getWithOffset(Offset),
1480 EltVT, false, false, 0));
1482 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1483 Node->getOperand(i), Idx,
1484 PtrInfo.getWithOffset(Offset),
1489 if (!Stores.empty()) // Not all undef elements?
1490 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1491 &Stores[0], Stores.size());
1493 StoreChain = DAG.getEntryNode();
1495 // Result is a load from the stack slot.
1496 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1497 false, false, false, 0);
1500 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1502 SDValue Tmp1 = Node->getOperand(0);
1503 SDValue Tmp2 = Node->getOperand(1);
1505 // Get the sign bit of the RHS. First obtain a value that has the same
1506 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1508 EVT FloatVT = Tmp2.getValueType();
1509 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1510 if (TLI.isTypeLegal(IVT)) {
1511 // Convert to an integer with the same sign bit.
1512 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1514 // Store the float to memory, then load the sign part out as an integer.
1515 MVT LoadTy = TLI.getPointerTy();
1516 // First create a temporary that is aligned for both the load and store.
1517 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1518 // Then store the float to it.
1520 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1522 if (TLI.isBigEndian()) {
1523 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1524 // Load out a legal integer with the same sign bit as the float.
1525 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1526 false, false, false, 0);
1527 } else { // Little endian
1528 SDValue LoadPtr = StackPtr;
1529 // The float may be wider than the integer we are going to load. Advance
1530 // the pointer so that the loaded integer will contain the sign bit.
1531 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1532 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1533 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(),
1534 LoadPtr, DAG.getIntPtrConstant(ByteOffset));
1535 // Load a legal integer containing the sign bit.
1536 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1537 false, false, false, 0);
1538 // Move the sign bit to the top bit of the loaded integer.
1539 unsigned BitShift = LoadTy.getSizeInBits() -
1540 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1541 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1543 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1544 DAG.getConstant(BitShift,
1545 TLI.getShiftAmountTy(SignBit.getValueType())));
1548 // Now get the sign bit proper, by seeing whether the value is negative.
1549 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1550 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1552 // Get the absolute value of the result.
1553 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1554 // Select between the nabs and abs value based on the sign bit of
1556 return DAG.getNode(ISD::SELECT, dl, AbsVal.getValueType(), SignBit,
1557 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1561 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1562 SmallVectorImpl<SDValue> &Results) {
1563 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1564 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1565 " not tell us which reg is the stack pointer!");
1567 EVT VT = Node->getValueType(0);
1568 SDValue Tmp1 = SDValue(Node, 0);
1569 SDValue Tmp2 = SDValue(Node, 1);
1570 SDValue Tmp3 = Node->getOperand(2);
1571 SDValue Chain = Tmp1.getOperand(0);
1573 // Chain the dynamic stack allocation so that it doesn't modify the stack
1574 // pointer when other instructions are using the stack.
1575 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1577 SDValue Size = Tmp2.getOperand(1);
1578 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1579 Chain = SP.getValue(1);
1580 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1581 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1582 if (Align > StackAlign)
1583 SP = DAG.getNode(ISD::AND, dl, VT, SP,
1584 DAG.getConstant(-(uint64_t)Align, VT));
1585 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1586 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1588 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1589 DAG.getIntPtrConstant(0, true), SDValue());
1591 Results.push_back(Tmp1);
1592 Results.push_back(Tmp2);
1595 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1596 /// condition code CC on the current target. This routine expands SETCC with
1597 /// illegal condition code into AND / OR of multiple SETCC values.
1598 void SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1599 SDValue &LHS, SDValue &RHS,
1602 MVT OpVT = LHS.getSimpleValueType();
1603 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1604 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1605 default: llvm_unreachable("Unknown condition code action!");
1606 case TargetLowering::Legal:
1609 case TargetLowering::Expand: {
1610 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1611 ISD::CondCode InvCC = ISD::SETCC_INVALID;
1614 default: llvm_unreachable("Don't know how to expand this condition!");
1616 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1617 == TargetLowering::Legal
1618 && "If SETO is expanded, SETOEQ must be legal!");
1619 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1621 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1622 == TargetLowering::Legal
1623 && "If SETUO is expanded, SETUNE must be legal!");
1624 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1637 // If we are floating point, assign and break, otherwise fall through.
1638 if (!OpVT.isInteger()) {
1639 // We can use the 4th bit to tell if we are the unordered
1640 // or ordered version of the opcode.
1641 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1642 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1643 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1646 // Fallthrough if we are unsigned integer.
1653 InvCC = ISD::getSetCCSwappedOperands(CCCode);
1654 if (TLI.getCondCodeAction(InvCC, OpVT) == TargetLowering::Expand) {
1655 // We only support using the inverted operation and not a
1656 // different manner of supporting expanding these cases.
1657 llvm_unreachable("Don't know how to expand this condition!");
1659 LHS = DAG.getSetCC(dl, VT, RHS, LHS, InvCC);
1665 SDValue SetCC1, SetCC2;
1666 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1667 // If we aren't the ordered or unorder operation,
1668 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1669 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1670 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1672 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1673 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1674 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1676 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1684 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1685 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1686 /// a load from the stack slot to DestVT, extending it if needed.
1687 /// The resultant code need not be legal.
1688 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1692 // Create the stack frame object.
1694 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1695 getTypeForEVT(*DAG.getContext()));
1696 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1698 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1699 int SPFI = StackPtrFI->getIndex();
1700 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1702 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1703 unsigned SlotSize = SlotVT.getSizeInBits();
1704 unsigned DestSize = DestVT.getSizeInBits();
1705 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1706 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1708 // Emit a store to the stack slot. Use a truncstore if the input value is
1709 // later than DestVT.
1712 if (SrcSize > SlotSize)
1713 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1714 PtrInfo, SlotVT, false, false, SrcAlign);
1716 assert(SrcSize == SlotSize && "Invalid store");
1717 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1718 PtrInfo, false, false, SrcAlign);
1721 // Result is a load from the stack slot.
1722 if (SlotSize == DestSize)
1723 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1724 false, false, false, DestAlign);
1726 assert(SlotSize < DestSize && "Unknown extension!");
1727 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1728 PtrInfo, SlotVT, false, false, DestAlign);
1731 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1733 // Create a vector sized/aligned stack slot, store the value to element #0,
1734 // then load the whole vector back out.
1735 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1737 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1738 int SPFI = StackPtrFI->getIndex();
1740 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1742 MachinePointerInfo::getFixedStack(SPFI),
1743 Node->getValueType(0).getVectorElementType(),
1745 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1746 MachinePointerInfo::getFixedStack(SPFI),
1747 false, false, false, 0);
1751 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1752 /// support the operation, but do support the resultant vector type.
1753 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1754 unsigned NumElems = Node->getNumOperands();
1755 SDValue Value1, Value2;
1757 EVT VT = Node->getValueType(0);
1758 EVT OpVT = Node->getOperand(0).getValueType();
1759 EVT EltVT = VT.getVectorElementType();
1761 // If the only non-undef value is the low element, turn this into a
1762 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1763 bool isOnlyLowElement = true;
1764 bool MoreThanTwoValues = false;
1765 bool isConstant = true;
1766 for (unsigned i = 0; i < NumElems; ++i) {
1767 SDValue V = Node->getOperand(i);
1768 if (V.getOpcode() == ISD::UNDEF)
1771 isOnlyLowElement = false;
1772 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1775 if (!Value1.getNode()) {
1777 } else if (!Value2.getNode()) {
1780 } else if (V != Value1 && V != Value2) {
1781 MoreThanTwoValues = true;
1785 if (!Value1.getNode())
1786 return DAG.getUNDEF(VT);
1788 if (isOnlyLowElement)
1789 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1791 // If all elements are constants, create a load from the constant pool.
1793 SmallVector<Constant*, 16> CV;
1794 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1795 if (ConstantFPSDNode *V =
1796 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1797 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1798 } else if (ConstantSDNode *V =
1799 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1801 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1803 // If OpVT and EltVT don't match, EltVT is not legal and the
1804 // element values have been promoted/truncated earlier. Undo this;
1805 // we don't want a v16i8 to become a v16i32 for example.
1806 const ConstantInt *CI = V->getConstantIntValue();
1807 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1808 CI->getZExtValue()));
1811 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1812 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1813 CV.push_back(UndefValue::get(OpNTy));
1816 Constant *CP = ConstantVector::get(CV);
1817 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1818 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1819 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1820 MachinePointerInfo::getConstantPool(),
1821 false, false, false, Alignment);
1824 if (!MoreThanTwoValues) {
1825 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1826 for (unsigned i = 0; i < NumElems; ++i) {
1827 SDValue V = Node->getOperand(i);
1828 if (V.getOpcode() == ISD::UNDEF)
1830 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1832 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1833 // Get the splatted value into the low element of a vector register.
1834 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1836 if (Value2.getNode())
1837 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1839 Vec2 = DAG.getUNDEF(VT);
1841 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1842 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1846 // Otherwise, we can't handle this case efficiently.
1847 return ExpandVectorBuildThroughStack(Node);
1850 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
1851 // does not fit into a register, return the lo part and set the hi part to the
1852 // by-reg argument. If it does fit into a single register, return the result
1853 // and leave the Hi part unset.
1854 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
1856 TargetLowering::ArgListTy Args;
1857 TargetLowering::ArgListEntry Entry;
1858 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1859 EVT ArgVT = Node->getOperand(i).getValueType();
1860 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1861 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
1862 Entry.isSExt = isSigned;
1863 Entry.isZExt = !isSigned;
1864 Args.push_back(Entry);
1866 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1867 TLI.getPointerTy());
1869 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1871 // By default, the input chain to this libcall is the entry node of the
1872 // function. If the libcall is going to be emitted as a tail call then
1873 // TLI.isUsedByReturnOnly will change it to the right chain if the return
1874 // node which is being folded has a non-entry input chain.
1875 SDValue InChain = DAG.getEntryNode();
1877 // isTailCall may be true since the callee does not reference caller stack
1878 // frame. Check if it's in the right position.
1879 SDValue TCChain = InChain;
1880 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
1885 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1886 0, TLI.getLibcallCallingConv(LC), isTailCall,
1887 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1888 Callee, Args, DAG, SDLoc(Node));
1889 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1892 if (!CallInfo.second.getNode())
1893 // It's a tailcall, return the chain (which is the DAG root).
1894 return DAG.getRoot();
1896 return CallInfo.first;
1899 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
1900 /// and returning a result of type RetVT.
1901 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
1902 const SDValue *Ops, unsigned NumOps,
1903 bool isSigned, SDLoc dl) {
1904 TargetLowering::ArgListTy Args;
1905 Args.reserve(NumOps);
1907 TargetLowering::ArgListEntry Entry;
1908 for (unsigned i = 0; i != NumOps; ++i) {
1909 Entry.Node = Ops[i];
1910 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
1911 Entry.isSExt = isSigned;
1912 Entry.isZExt = !isSigned;
1913 Args.push_back(Entry);
1915 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1916 TLI.getPointerTy());
1918 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
1920 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
1921 false, 0, TLI.getLibcallCallingConv(LC),
1922 /*isTailCall=*/false,
1923 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1924 Callee, Args, DAG, dl);
1925 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
1927 return CallInfo.first;
1930 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
1931 // ExpandLibCall except that the first operand is the in-chain.
1932 std::pair<SDValue, SDValue>
1933 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
1936 SDValue InChain = Node->getOperand(0);
1938 TargetLowering::ArgListTy Args;
1939 TargetLowering::ArgListEntry Entry;
1940 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
1941 EVT ArgVT = Node->getOperand(i).getValueType();
1942 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1943 Entry.Node = Node->getOperand(i);
1945 Entry.isSExt = isSigned;
1946 Entry.isZExt = !isSigned;
1947 Args.push_back(Entry);
1949 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1950 TLI.getPointerTy());
1952 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1954 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1955 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
1956 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1957 Callee, Args, DAG, SDLoc(Node));
1958 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1963 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
1964 RTLIB::Libcall Call_F32,
1965 RTLIB::Libcall Call_F64,
1966 RTLIB::Libcall Call_F80,
1967 RTLIB::Libcall Call_F128,
1968 RTLIB::Libcall Call_PPCF128) {
1970 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1971 default: llvm_unreachable("Unexpected request for libcall!");
1972 case MVT::f32: LC = Call_F32; break;
1973 case MVT::f64: LC = Call_F64; break;
1974 case MVT::f80: LC = Call_F80; break;
1975 case MVT::f128: LC = Call_F128; break;
1976 case MVT::ppcf128: LC = Call_PPCF128; break;
1978 return ExpandLibCall(LC, Node, false);
1981 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
1982 RTLIB::Libcall Call_I8,
1983 RTLIB::Libcall Call_I16,
1984 RTLIB::Libcall Call_I32,
1985 RTLIB::Libcall Call_I64,
1986 RTLIB::Libcall Call_I128) {
1988 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1989 default: llvm_unreachable("Unexpected request for libcall!");
1990 case MVT::i8: LC = Call_I8; break;
1991 case MVT::i16: LC = Call_I16; break;
1992 case MVT::i32: LC = Call_I32; break;
1993 case MVT::i64: LC = Call_I64; break;
1994 case MVT::i128: LC = Call_I128; break;
1996 return ExpandLibCall(LC, Node, isSigned);
1999 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
2000 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2001 const TargetLowering &TLI) {
2003 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2004 default: llvm_unreachable("Unexpected request for libcall!");
2005 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2006 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2007 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2008 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2009 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2012 return TLI.getLibcallName(LC) != 0;
2015 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2017 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2018 // The other use might have been replaced with a divrem already.
2019 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2020 unsigned OtherOpcode = 0;
2022 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2024 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2026 SDValue Op0 = Node->getOperand(0);
2027 SDValue Op1 = Node->getOperand(1);
2028 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2029 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2033 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2034 User->getOperand(0) == Op0 &&
2035 User->getOperand(1) == Op1)
2041 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2044 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2045 SmallVectorImpl<SDValue> &Results) {
2046 unsigned Opcode = Node->getOpcode();
2047 bool isSigned = Opcode == ISD::SDIVREM;
2050 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2051 default: llvm_unreachable("Unexpected request for libcall!");
2052 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2053 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2054 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2055 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2056 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2059 // The input chain to this libcall is the entry node of the function.
2060 // Legalizing the call will automatically add the previous call to the
2062 SDValue InChain = DAG.getEntryNode();
2064 EVT RetVT = Node->getValueType(0);
2065 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2067 TargetLowering::ArgListTy Args;
2068 TargetLowering::ArgListEntry Entry;
2069 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2070 EVT ArgVT = Node->getOperand(i).getValueType();
2071 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2072 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2073 Entry.isSExt = isSigned;
2074 Entry.isZExt = !isSigned;
2075 Args.push_back(Entry);
2078 // Also pass the return address of the remainder.
2079 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2081 Entry.Ty = RetTy->getPointerTo();
2082 Entry.isSExt = isSigned;
2083 Entry.isZExt = !isSigned;
2084 Args.push_back(Entry);
2086 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2087 TLI.getPointerTy());
2091 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
2092 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2093 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2094 Callee, Args, DAG, dl);
2095 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2097 // Remainder is loaded back from the stack frame.
2098 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2099 MachinePointerInfo(), false, false, false, 0);
2100 Results.push_back(CallInfo.first);
2101 Results.push_back(Rem);
2104 /// isSinCosLibcallAvailable - Return true if sincos libcall is available.
2105 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2107 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2108 default: llvm_unreachable("Unexpected request for libcall!");
2109 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2110 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2111 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2112 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2113 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2115 return TLI.getLibcallName(LC) != 0;
2118 /// canCombineSinCosLibcall - Return true if sincos libcall is available and
2119 /// can be used to combine sin and cos.
2120 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2121 const TargetMachine &TM) {
2122 if (!isSinCosLibcallAvailable(Node, TLI))
2124 // GNU sin/cos functions set errno while sincos does not. Therefore
2125 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2126 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2127 if (isGNU && !TM.Options.UnsafeFPMath)
2132 /// useSinCos - Only issue sincos libcall if both sin and cos are
2134 static bool useSinCos(SDNode *Node) {
2135 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2136 ? ISD::FCOS : ISD::FSIN;
2138 SDValue Op0 = Node->getOperand(0);
2139 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2140 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2144 // The other user might have been turned into sincos already.
2145 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2151 /// ExpandSinCosLibCall - Issue libcalls to sincos to compute sin / cos
2154 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2155 SmallVectorImpl<SDValue> &Results) {
2157 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2158 default: llvm_unreachable("Unexpected request for libcall!");
2159 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2160 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2161 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2162 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2163 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2166 // The input chain to this libcall is the entry node of the function.
2167 // Legalizing the call will automatically add the previous call to the
2169 SDValue InChain = DAG.getEntryNode();
2171 EVT RetVT = Node->getValueType(0);
2172 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2174 TargetLowering::ArgListTy Args;
2175 TargetLowering::ArgListEntry Entry;
2177 // Pass the argument.
2178 Entry.Node = Node->getOperand(0);
2180 Entry.isSExt = false;
2181 Entry.isZExt = false;
2182 Args.push_back(Entry);
2184 // Pass the return address of sin.
2185 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2186 Entry.Node = SinPtr;
2187 Entry.Ty = RetTy->getPointerTo();
2188 Entry.isSExt = false;
2189 Entry.isZExt = false;
2190 Args.push_back(Entry);
2192 // Also pass the return address of the cos.
2193 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2194 Entry.Node = CosPtr;
2195 Entry.Ty = RetTy->getPointerTo();
2196 Entry.isSExt = false;
2197 Entry.isZExt = false;
2198 Args.push_back(Entry);
2200 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2201 TLI.getPointerTy());
2205 CallLoweringInfo CLI(InChain, Type::getVoidTy(*DAG.getContext()),
2206 false, false, false, false,
2207 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2208 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2209 Callee, Args, DAG, dl);
2210 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2212 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2213 MachinePointerInfo(), false, false, false, 0));
2214 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2215 MachinePointerInfo(), false, false, false, 0));
2218 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2219 /// INT_TO_FP operation of the specified operand when the target requests that
2220 /// we expand it. At this point, we know that the result and operand types are
2221 /// legal for the target.
2222 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2226 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2227 // simple 32-bit [signed|unsigned] integer to float/double expansion
2229 // Get the stack frame index of a 8 byte buffer.
2230 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2232 // word offset constant for Hi/Lo address computation
2233 SDValue WordOff = DAG.getConstant(sizeof(int), TLI.getPointerTy());
2234 // set up Hi and Lo (into buffer) address based on endian
2235 SDValue Hi = StackSlot;
2236 SDValue Lo = DAG.getNode(ISD::ADD, dl,
2237 TLI.getPointerTy(), StackSlot, WordOff);
2238 if (TLI.isLittleEndian())
2241 // if signed map to unsigned space
2244 // constant used to invert sign bit (signed to unsigned mapping)
2245 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2246 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2250 // store the lo of the constructed double - based on integer input
2251 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2252 Op0Mapped, Lo, MachinePointerInfo(),
2254 // initial hi portion of constructed double
2255 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2256 // store the hi of the constructed double - biased exponent
2257 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2258 MachinePointerInfo(),
2260 // load the constructed double
2261 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2262 MachinePointerInfo(), false, false, false, 0);
2263 // FP constant to bias correct the final result
2264 SDValue Bias = DAG.getConstantFP(isSigned ?
2265 BitsToDouble(0x4330000080000000ULL) :
2266 BitsToDouble(0x4330000000000000ULL),
2268 // subtract the bias
2269 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2272 // handle final rounding
2273 if (DestVT == MVT::f64) {
2276 } else if (DestVT.bitsLT(MVT::f64)) {
2277 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2278 DAG.getIntPtrConstant(0));
2279 } else if (DestVT.bitsGT(MVT::f64)) {
2280 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2284 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2285 // Code below here assumes !isSigned without checking again.
2287 // Implementation of unsigned i64 to f64 following the algorithm in
2288 // __floatundidf in compiler_rt. This implementation has the advantage
2289 // of performing rounding correctly, both in the default rounding mode
2290 // and in all alternate rounding modes.
2291 // TODO: Generalize this for use with other types.
2292 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2294 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2295 SDValue TwoP84PlusTwoP52 =
2296 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2298 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2300 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2301 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2302 DAG.getConstant(32, MVT::i64));
2303 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2304 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2305 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2306 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2307 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2309 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2312 // Implementation of unsigned i64 to f32.
2313 // TODO: Generalize this for use with other types.
2314 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2315 // For unsigned conversions, convert them to signed conversions using the
2316 // algorithm from the x86_64 __floatundidf in compiler_rt.
2318 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2320 SDValue ShiftConst =
2321 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2322 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2323 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2324 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2325 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2327 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2328 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2330 // TODO: This really should be implemented using a branch rather than a
2331 // select. We happen to get lucky and machinesink does the right
2332 // thing most of the time. This would be a good candidate for a
2333 //pseudo-op, or, even better, for whole-function isel.
2334 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2335 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2336 return DAG.getNode(ISD::SELECT, dl, MVT::f32, SignBitTest, Slow, Fast);
2339 // Otherwise, implement the fully general conversion.
2341 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2342 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2343 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2344 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2345 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2346 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2347 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2348 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2349 SDValue Sel = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ne, Or, Op0);
2350 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2351 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2353 SDValue Sel2 = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ge, Sel, Op0);
2354 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2356 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2357 DAG.getConstant(32, SHVT));
2358 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2359 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2361 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2362 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2363 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2364 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2365 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2366 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2367 DAG.getIntPtrConstant(0));
2370 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2372 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2373 Op0, DAG.getConstant(0, Op0.getValueType()),
2375 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2376 SDValue CstOffset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(),
2377 SignSet, Four, Zero);
2379 // If the sign bit of the integer is set, the large number will be treated
2380 // as a negative number. To counteract this, the dynamic code adds an
2381 // offset depending on the data type.
2383 switch (Op0.getValueType().getSimpleVT().SimpleTy) {
2384 default: llvm_unreachable("Unsupported integer type!");
2385 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2386 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2387 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2388 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2390 if (TLI.isLittleEndian()) FF <<= 32;
2391 Constant *FudgeFactor = ConstantInt::get(
2392 Type::getInt64Ty(*DAG.getContext()), FF);
2394 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2395 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2396 CPIdx = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), CPIdx, CstOffset);
2397 Alignment = std::min(Alignment, 4u);
2399 if (DestVT == MVT::f32)
2400 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2401 MachinePointerInfo::getConstantPool(),
2402 false, false, false, Alignment);
2404 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2405 DAG.getEntryNode(), CPIdx,
2406 MachinePointerInfo::getConstantPool(),
2407 MVT::f32, false, false, Alignment);
2408 HandleSDNode Handle(Load);
2409 LegalizeOp(Load.getNode());
2410 FudgeInReg = Handle.getValue();
2413 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2416 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2417 /// *INT_TO_FP operation of the specified operand when the target requests that
2418 /// we promote it. At this point, we know that the result and operand types are
2419 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2420 /// operation that takes a larger input.
2421 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2425 // First step, figure out the appropriate *INT_TO_FP operation to use.
2426 EVT NewInTy = LegalOp.getValueType();
2428 unsigned OpToUse = 0;
2430 // Scan for the appropriate larger type to use.
2432 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2433 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2435 // If the target supports SINT_TO_FP of this type, use it.
2436 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2437 OpToUse = ISD::SINT_TO_FP;
2440 if (isSigned) continue;
2442 // If the target supports UINT_TO_FP of this type, use it.
2443 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2444 OpToUse = ISD::UINT_TO_FP;
2448 // Otherwise, try a larger type.
2451 // Okay, we found the operation and type to use. Zero extend our input to the
2452 // desired type then run the operation on it.
2453 return DAG.getNode(OpToUse, dl, DestVT,
2454 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2455 dl, NewInTy, LegalOp));
2458 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2459 /// FP_TO_*INT operation of the specified operand when the target requests that
2460 /// we promote it. At this point, we know that the result and operand types are
2461 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2462 /// operation that returns a larger result.
2463 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2467 // First step, figure out the appropriate FP_TO*INT operation to use.
2468 EVT NewOutTy = DestVT;
2470 unsigned OpToUse = 0;
2472 // Scan for the appropriate larger type to use.
2474 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2475 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2477 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2478 OpToUse = ISD::FP_TO_SINT;
2482 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2483 OpToUse = ISD::FP_TO_UINT;
2487 // Otherwise, try a larger type.
2491 // Okay, we found the operation and type to use.
2492 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2494 // Truncate the result of the extended FP_TO_*INT operation to the desired
2496 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2499 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2501 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2502 EVT VT = Op.getValueType();
2503 EVT SHVT = TLI.getShiftAmountTy(VT);
2504 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2505 switch (VT.getSimpleVT().SimpleTy) {
2506 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2508 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2509 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2510 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2512 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2513 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2514 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2515 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2516 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2517 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2518 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2519 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2520 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2522 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2523 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2524 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2525 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2526 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2527 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2528 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2529 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2530 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2531 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2532 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2533 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2534 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2535 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2536 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2537 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2538 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2539 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2540 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2541 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2542 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2546 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2548 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2551 default: llvm_unreachable("Cannot expand this yet!");
2553 EVT VT = Op.getValueType();
2554 EVT ShVT = TLI.getShiftAmountTy(VT);
2555 unsigned Len = VT.getSizeInBits();
2557 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2558 "CTPOP not implemented for this type.");
2560 // This is the "best" algorithm from
2561 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2563 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2564 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2565 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2566 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2568 // v = v - ((v >> 1) & 0x55555555...)
2569 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2570 DAG.getNode(ISD::AND, dl, VT,
2571 DAG.getNode(ISD::SRL, dl, VT, Op,
2572 DAG.getConstant(1, ShVT)),
2574 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2575 Op = DAG.getNode(ISD::ADD, dl, VT,
2576 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2577 DAG.getNode(ISD::AND, dl, VT,
2578 DAG.getNode(ISD::SRL, dl, VT, Op,
2579 DAG.getConstant(2, ShVT)),
2581 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2582 Op = DAG.getNode(ISD::AND, dl, VT,
2583 DAG.getNode(ISD::ADD, dl, VT, Op,
2584 DAG.getNode(ISD::SRL, dl, VT, Op,
2585 DAG.getConstant(4, ShVT))),
2587 // v = (v * 0x01010101...) >> (Len - 8)
2588 Op = DAG.getNode(ISD::SRL, dl, VT,
2589 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2590 DAG.getConstant(Len - 8, ShVT));
2594 case ISD::CTLZ_ZERO_UNDEF:
2595 // This trivially expands to CTLZ.
2596 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2598 // for now, we do this:
2599 // x = x | (x >> 1);
2600 // x = x | (x >> 2);
2602 // x = x | (x >>16);
2603 // x = x | (x >>32); // for 64-bit input
2604 // return popcount(~x);
2606 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2607 EVT VT = Op.getValueType();
2608 EVT ShVT = TLI.getShiftAmountTy(VT);
2609 unsigned len = VT.getSizeInBits();
2610 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2611 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2612 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2613 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2615 Op = DAG.getNOT(dl, Op, VT);
2616 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2618 case ISD::CTTZ_ZERO_UNDEF:
2619 // This trivially expands to CTTZ.
2620 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2622 // for now, we use: { return popcount(~x & (x - 1)); }
2623 // unless the target has ctlz but not ctpop, in which case we use:
2624 // { return 32 - nlz(~x & (x-1)); }
2625 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2626 EVT VT = Op.getValueType();
2627 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2628 DAG.getNOT(dl, Op, VT),
2629 DAG.getNode(ISD::SUB, dl, VT, Op,
2630 DAG.getConstant(1, VT)));
2631 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2632 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2633 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2634 return DAG.getNode(ISD::SUB, dl, VT,
2635 DAG.getConstant(VT.getSizeInBits(), VT),
2636 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2637 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2642 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2643 unsigned Opc = Node->getOpcode();
2644 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2649 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2650 case ISD::ATOMIC_SWAP:
2651 switch (VT.SimpleTy) {
2652 default: llvm_unreachable("Unexpected value type for atomic!");
2653 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2654 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2655 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2656 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2659 case ISD::ATOMIC_CMP_SWAP:
2660 switch (VT.SimpleTy) {
2661 default: llvm_unreachable("Unexpected value type for atomic!");
2662 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2663 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2664 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2665 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2668 case ISD::ATOMIC_LOAD_ADD:
2669 switch (VT.SimpleTy) {
2670 default: llvm_unreachable("Unexpected value type for atomic!");
2671 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2672 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2673 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2674 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2677 case ISD::ATOMIC_LOAD_SUB:
2678 switch (VT.SimpleTy) {
2679 default: llvm_unreachable("Unexpected value type for atomic!");
2680 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2681 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2682 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2683 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2686 case ISD::ATOMIC_LOAD_AND:
2687 switch (VT.SimpleTy) {
2688 default: llvm_unreachable("Unexpected value type for atomic!");
2689 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2690 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2691 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2692 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2695 case ISD::ATOMIC_LOAD_OR:
2696 switch (VT.SimpleTy) {
2697 default: llvm_unreachable("Unexpected value type for atomic!");
2698 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2699 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2700 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2701 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2704 case ISD::ATOMIC_LOAD_XOR:
2705 switch (VT.SimpleTy) {
2706 default: llvm_unreachable("Unexpected value type for atomic!");
2707 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2708 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2709 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2710 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2713 case ISD::ATOMIC_LOAD_NAND:
2714 switch (VT.SimpleTy) {
2715 default: llvm_unreachable("Unexpected value type for atomic!");
2716 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2717 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2718 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2719 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2724 return ExpandChainLibCall(LC, Node, false);
2727 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2728 SmallVector<SDValue, 8> Results;
2730 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2731 switch (Node->getOpcode()) {
2734 case ISD::CTLZ_ZERO_UNDEF:
2736 case ISD::CTTZ_ZERO_UNDEF:
2737 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2738 Results.push_back(Tmp1);
2741 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2743 case ISD::FRAMEADDR:
2744 case ISD::RETURNADDR:
2745 case ISD::FRAME_TO_ARGS_OFFSET:
2746 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2748 case ISD::FLT_ROUNDS_:
2749 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2751 case ISD::EH_RETURN:
2755 case ISD::EH_SJLJ_LONGJMP:
2756 // If the target didn't expand these, there's nothing to do, so just
2757 // preserve the chain and be done.
2758 Results.push_back(Node->getOperand(0));
2760 case ISD::EH_SJLJ_SETJMP:
2761 // If the target didn't expand this, just return 'zero' and preserve the
2763 Results.push_back(DAG.getConstant(0, MVT::i32));
2764 Results.push_back(Node->getOperand(0));
2766 case ISD::ATOMIC_FENCE: {
2767 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2768 // FIXME: handle "fence singlethread" more efficiently.
2769 TargetLowering::ArgListTy Args;
2771 CallLoweringInfo CLI(Node->getOperand(0),
2772 Type::getVoidTy(*DAG.getContext()),
2773 false, false, false, false, 0, CallingConv::C,
2774 /*isTailCall=*/false,
2775 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2776 DAG.getExternalSymbol("__sync_synchronize",
2777 TLI.getPointerTy()),
2779 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2781 Results.push_back(CallResult.second);
2784 case ISD::ATOMIC_LOAD: {
2785 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2786 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
2787 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
2788 cast<AtomicSDNode>(Node)->getMemoryVT(),
2789 Node->getOperand(0),
2790 Node->getOperand(1), Zero, Zero,
2791 cast<AtomicSDNode>(Node)->getMemOperand(),
2792 cast<AtomicSDNode>(Node)->getOrdering(),
2793 cast<AtomicSDNode>(Node)->getSynchScope());
2794 Results.push_back(Swap.getValue(0));
2795 Results.push_back(Swap.getValue(1));
2798 case ISD::ATOMIC_STORE: {
2799 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2800 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2801 cast<AtomicSDNode>(Node)->getMemoryVT(),
2802 Node->getOperand(0),
2803 Node->getOperand(1), Node->getOperand(2),
2804 cast<AtomicSDNode>(Node)->getMemOperand(),
2805 cast<AtomicSDNode>(Node)->getOrdering(),
2806 cast<AtomicSDNode>(Node)->getSynchScope());
2807 Results.push_back(Swap.getValue(1));
2810 // By default, atomic intrinsics are marked Legal and lowered. Targets
2811 // which don't support them directly, however, may want libcalls, in which
2812 // case they mark them Expand, and we get here.
2813 case ISD::ATOMIC_SWAP:
2814 case ISD::ATOMIC_LOAD_ADD:
2815 case ISD::ATOMIC_LOAD_SUB:
2816 case ISD::ATOMIC_LOAD_AND:
2817 case ISD::ATOMIC_LOAD_OR:
2818 case ISD::ATOMIC_LOAD_XOR:
2819 case ISD::ATOMIC_LOAD_NAND:
2820 case ISD::ATOMIC_LOAD_MIN:
2821 case ISD::ATOMIC_LOAD_MAX:
2822 case ISD::ATOMIC_LOAD_UMIN:
2823 case ISD::ATOMIC_LOAD_UMAX:
2824 case ISD::ATOMIC_CMP_SWAP: {
2825 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
2826 Results.push_back(Tmp.first);
2827 Results.push_back(Tmp.second);
2830 case ISD::DYNAMIC_STACKALLOC:
2831 ExpandDYNAMIC_STACKALLOC(Node, Results);
2833 case ISD::MERGE_VALUES:
2834 for (unsigned i = 0; i < Node->getNumValues(); i++)
2835 Results.push_back(Node->getOperand(i));
2838 EVT VT = Node->getValueType(0);
2840 Results.push_back(DAG.getConstant(0, VT));
2842 assert(VT.isFloatingPoint() && "Unknown value type!");
2843 Results.push_back(DAG.getConstantFP(0, VT));
2848 // If this operation is not supported, lower it to 'abort()' call
2849 TargetLowering::ArgListTy Args;
2851 CallLoweringInfo CLI(Node->getOperand(0),
2852 Type::getVoidTy(*DAG.getContext()),
2853 false, false, false, false, 0, CallingConv::C,
2854 /*isTailCall=*/false,
2855 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2856 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
2858 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2860 Results.push_back(CallResult.second);
2865 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
2866 Node->getValueType(0), dl);
2867 Results.push_back(Tmp1);
2869 case ISD::FP_EXTEND:
2870 Tmp1 = EmitStackConvert(Node->getOperand(0),
2871 Node->getOperand(0).getValueType(),
2872 Node->getValueType(0), dl);
2873 Results.push_back(Tmp1);
2875 case ISD::SIGN_EXTEND_INREG: {
2876 // NOTE: we could fall back on load/store here too for targets without
2877 // SAR. However, it is doubtful that any exist.
2878 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2879 EVT VT = Node->getValueType(0);
2880 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
2883 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
2884 ExtraVT.getScalarType().getSizeInBits();
2885 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
2886 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
2887 Node->getOperand(0), ShiftCst);
2888 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
2889 Results.push_back(Tmp1);
2892 case ISD::FP_ROUND_INREG: {
2893 // The only way we can lower this is to turn it into a TRUNCSTORE,
2894 // EXTLOAD pair, targeting a temporary location (a stack slot).
2896 // NOTE: there is a choice here between constantly creating new stack
2897 // slots and always reusing the same one. We currently always create
2898 // new ones, as reuse may inhibit scheduling.
2899 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2900 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
2901 Node->getValueType(0), dl);
2902 Results.push_back(Tmp1);
2905 case ISD::SINT_TO_FP:
2906 case ISD::UINT_TO_FP:
2907 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
2908 Node->getOperand(0), Node->getValueType(0), dl);
2909 Results.push_back(Tmp1);
2911 case ISD::FP_TO_UINT: {
2912 SDValue True, False;
2913 EVT VT = Node->getOperand(0).getValueType();
2914 EVT NVT = Node->getValueType(0);
2915 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
2916 APInt::getNullValue(VT.getSizeInBits()));
2917 APInt x = APInt::getSignBit(NVT.getSizeInBits());
2918 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
2919 Tmp1 = DAG.getConstantFP(apf, VT);
2920 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
2921 Node->getOperand(0),
2923 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
2924 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
2925 DAG.getNode(ISD::FSUB, dl, VT,
2926 Node->getOperand(0), Tmp1));
2927 False = DAG.getNode(ISD::XOR, dl, NVT, False,
2928 DAG.getConstant(x, NVT));
2929 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2, True, False);
2930 Results.push_back(Tmp1);
2934 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2935 EVT VT = Node->getValueType(0);
2936 Tmp1 = Node->getOperand(0);
2937 Tmp2 = Node->getOperand(1);
2938 unsigned Align = Node->getConstantOperandVal(3);
2940 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
2941 MachinePointerInfo(V),
2942 false, false, false, 0);
2943 SDValue VAList = VAListLoad;
2945 if (Align > TLI.getMinStackArgumentAlignment()) {
2946 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
2948 VAList = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2949 DAG.getConstant(Align - 1,
2950 TLI.getPointerTy()));
2952 VAList = DAG.getNode(ISD::AND, dl, TLI.getPointerTy(), VAList,
2953 DAG.getConstant(-(int64_t)Align,
2954 TLI.getPointerTy()));
2957 // Increment the pointer, VAList, to the next vaarg
2958 Tmp3 = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2959 DAG.getConstant(TLI.getDataLayout()->
2960 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
2961 TLI.getPointerTy()));
2962 // Store the incremented VAList to the legalized pointer
2963 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
2964 MachinePointerInfo(V), false, false, 0);
2965 // Load the actual argument out of the pointer VAList
2966 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
2967 false, false, false, 0));
2968 Results.push_back(Results[0].getValue(1));
2972 // This defaults to loading a pointer from the input and storing it to the
2973 // output, returning the chain.
2974 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
2975 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
2976 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
2977 Node->getOperand(2), MachinePointerInfo(VS),
2978 false, false, false, 0);
2979 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
2980 MachinePointerInfo(VD), false, false, 0);
2981 Results.push_back(Tmp1);
2984 case ISD::EXTRACT_VECTOR_ELT:
2985 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
2986 // This must be an access of the only element. Return it.
2987 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
2988 Node->getOperand(0));
2990 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
2991 Results.push_back(Tmp1);
2993 case ISD::EXTRACT_SUBVECTOR:
2994 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
2996 case ISD::INSERT_SUBVECTOR:
2997 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
2999 case ISD::CONCAT_VECTORS: {
3000 Results.push_back(ExpandVectorBuildThroughStack(Node));
3003 case ISD::SCALAR_TO_VECTOR:
3004 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3006 case ISD::INSERT_VECTOR_ELT:
3007 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3008 Node->getOperand(1),
3009 Node->getOperand(2), dl));
3011 case ISD::VECTOR_SHUFFLE: {
3012 SmallVector<int, 32> NewMask;
3013 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3015 EVT VT = Node->getValueType(0);
3016 EVT EltVT = VT.getVectorElementType();
3017 SDValue Op0 = Node->getOperand(0);
3018 SDValue Op1 = Node->getOperand(1);
3019 if (!TLI.isTypeLegal(EltVT)) {
3021 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3023 // BUILD_VECTOR operands are allowed to be wider than the element type.
3024 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept it
3025 if (NewEltVT.bitsLT(EltVT)) {
3027 // Convert shuffle node.
3028 // If original node was v4i64 and the new EltVT is i32,
3029 // cast operands to v8i32 and re-build the mask.
3031 // Calculate new VT, the size of the new VT should be equal to original.
3032 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3033 VT.getSizeInBits()/NewEltVT.getSizeInBits());
3034 assert(NewVT.bitsEq(VT));
3036 // cast operands to new VT
3037 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3038 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3040 // Convert the shuffle mask
3041 unsigned int factor = NewVT.getVectorNumElements()/VT.getVectorNumElements();
3043 // EltVT gets smaller
3046 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3048 for (unsigned fi = 0; fi < factor; ++fi)
3049 NewMask.push_back(Mask[i]);
3052 for (unsigned fi = 0; fi < factor; ++fi)
3053 NewMask.push_back(Mask[i]*factor+fi);
3061 unsigned NumElems = VT.getVectorNumElements();
3062 SmallVector<SDValue, 16> Ops;
3063 for (unsigned i = 0; i != NumElems; ++i) {
3065 Ops.push_back(DAG.getUNDEF(EltVT));
3068 unsigned Idx = Mask[i];
3070 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3072 DAG.getIntPtrConstant(Idx)));
3074 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3076 DAG.getIntPtrConstant(Idx - NumElems)));
3079 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
3080 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3081 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3082 Results.push_back(Tmp1);
3085 case ISD::EXTRACT_ELEMENT: {
3086 EVT OpTy = Node->getOperand(0).getValueType();
3087 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3089 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3090 DAG.getConstant(OpTy.getSizeInBits()/2,
3091 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3092 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3095 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3096 Node->getOperand(0));
3098 Results.push_back(Tmp1);
3101 case ISD::STACKSAVE:
3102 // Expand to CopyFromReg if the target set
3103 // StackPointerRegisterToSaveRestore.
3104 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3105 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3106 Node->getValueType(0)));
3107 Results.push_back(Results[0].getValue(1));
3109 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3110 Results.push_back(Node->getOperand(0));
3113 case ISD::STACKRESTORE:
3114 // Expand to CopyToReg if the target set
3115 // StackPointerRegisterToSaveRestore.
3116 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3117 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3118 Node->getOperand(1)));
3120 Results.push_back(Node->getOperand(0));
3123 case ISD::FCOPYSIGN:
3124 Results.push_back(ExpandFCOPYSIGN(Node));
3127 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3128 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3129 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3130 Node->getOperand(0));
3131 Results.push_back(Tmp1);
3134 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3135 EVT VT = Node->getValueType(0);
3136 Tmp1 = Node->getOperand(0);
3137 Tmp2 = DAG.getConstantFP(0.0, VT);
3138 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3139 Tmp1, Tmp2, ISD::SETUGT);
3140 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3141 Tmp1 = DAG.getNode(ISD::SELECT, dl, VT, Tmp2, Tmp1, Tmp3);
3142 Results.push_back(Tmp1);
3146 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3147 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3148 RTLIB::SQRT_PPCF128));
3152 EVT VT = Node->getValueType(0);
3153 bool isSIN = Node->getOpcode() == ISD::FSIN;
3154 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3155 // fcos which share the same operand and both are used.
3156 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3157 canCombineSinCosLibcall(Node, TLI, TM))
3158 && useSinCos(Node)) {
3159 SDVTList VTs = DAG.getVTList(VT, VT);
3160 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3162 Tmp1 = Tmp1.getValue(1);
3163 Results.push_back(Tmp1);
3165 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3166 RTLIB::SIN_F80, RTLIB::SIN_F128,
3167 RTLIB::SIN_PPCF128));
3169 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3170 RTLIB::COS_F80, RTLIB::COS_F128,
3171 RTLIB::COS_PPCF128));
3176 // Expand into sincos libcall.
3177 ExpandSinCosLibCall(Node, Results);
3180 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3181 RTLIB::LOG_F80, RTLIB::LOG_F128,
3182 RTLIB::LOG_PPCF128));
3185 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3186 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3187 RTLIB::LOG2_PPCF128));
3190 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3191 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3192 RTLIB::LOG10_PPCF128));
3195 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3196 RTLIB::EXP_F80, RTLIB::EXP_F128,
3197 RTLIB::EXP_PPCF128));
3200 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3201 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3202 RTLIB::EXP2_PPCF128));
3205 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3206 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3207 RTLIB::TRUNC_PPCF128));
3210 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3211 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3212 RTLIB::FLOOR_PPCF128));
3215 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3216 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3217 RTLIB::CEIL_PPCF128));
3220 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3221 RTLIB::RINT_F80, RTLIB::RINT_F128,
3222 RTLIB::RINT_PPCF128));
3224 case ISD::FNEARBYINT:
3225 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3226 RTLIB::NEARBYINT_F64,
3227 RTLIB::NEARBYINT_F80,
3228 RTLIB::NEARBYINT_F128,
3229 RTLIB::NEARBYINT_PPCF128));
3232 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3233 RTLIB::POWI_F80, RTLIB::POWI_F128,
3234 RTLIB::POWI_PPCF128));
3237 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3238 RTLIB::POW_F80, RTLIB::POW_F128,
3239 RTLIB::POW_PPCF128));
3242 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3243 RTLIB::DIV_F80, RTLIB::DIV_F128,
3244 RTLIB::DIV_PPCF128));
3247 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3248 RTLIB::REM_F80, RTLIB::REM_F128,
3249 RTLIB::REM_PPCF128));
3252 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3253 RTLIB::FMA_F80, RTLIB::FMA_F128,
3254 RTLIB::FMA_PPCF128));
3256 case ISD::FP16_TO_FP32:
3257 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3259 case ISD::FP32_TO_FP16:
3260 Results.push_back(ExpandLibCall(RTLIB::FPROUND_F32_F16, Node, false));
3262 case ISD::ConstantFP: {
3263 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3264 // Check to see if this FP immediate is already legal.
3265 // If this is a legal constant, turn it into a TargetConstantFP node.
3266 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3267 Results.push_back(ExpandConstantFP(CFP, true));
3270 case ISD::EHSELECTION: {
3271 unsigned Reg = TLI.getExceptionSelectorRegister();
3272 assert(Reg && "Can't expand to unknown register!");
3273 Results.push_back(DAG.getCopyFromReg(Node->getOperand(1), dl, Reg,
3274 Node->getValueType(0)));
3275 Results.push_back(Results[0].getValue(1));
3278 case ISD::EXCEPTIONADDR: {
3279 unsigned Reg = TLI.getExceptionPointerRegister();
3280 assert(Reg && "Can't expand to unknown register!");
3281 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, Reg,
3282 Node->getValueType(0)));
3283 Results.push_back(Results[0].getValue(1));
3287 EVT VT = Node->getValueType(0);
3288 assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3289 TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
3290 "Don't know how to expand this FP subtraction!");
3291 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3292 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3293 Results.push_back(Tmp1);
3297 EVT VT = Node->getValueType(0);
3298 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3299 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3300 "Don't know how to expand this subtraction!");
3301 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3302 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3303 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3304 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3309 EVT VT = Node->getValueType(0);
3310 bool isSigned = Node->getOpcode() == ISD::SREM;
3311 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3312 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3313 Tmp2 = Node->getOperand(0);
3314 Tmp3 = Node->getOperand(1);
3315 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3316 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3317 // If div is legal, it's better to do the normal expansion
3318 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3319 useDivRem(Node, isSigned, false))) {
3320 SDVTList VTs = DAG.getVTList(VT, VT);
3321 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3322 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3324 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3325 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3326 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3327 } else if (isSigned)
3328 Tmp1 = ExpandIntLibCall(Node, true,
3330 RTLIB::SREM_I16, RTLIB::SREM_I32,
3331 RTLIB::SREM_I64, RTLIB::SREM_I128);
3333 Tmp1 = ExpandIntLibCall(Node, false,
3335 RTLIB::UREM_I16, RTLIB::UREM_I32,
3336 RTLIB::UREM_I64, RTLIB::UREM_I128);
3337 Results.push_back(Tmp1);
3342 bool isSigned = Node->getOpcode() == ISD::SDIV;
3343 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3344 EVT VT = Node->getValueType(0);
3345 SDVTList VTs = DAG.getVTList(VT, VT);
3346 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3347 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3348 useDivRem(Node, isSigned, true)))
3349 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3350 Node->getOperand(1));
3352 Tmp1 = ExpandIntLibCall(Node, true,
3354 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3355 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3357 Tmp1 = ExpandIntLibCall(Node, false,
3359 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3360 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3361 Results.push_back(Tmp1);
3366 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3368 EVT VT = Node->getValueType(0);
3369 SDVTList VTs = DAG.getVTList(VT, VT);
3370 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3371 "If this wasn't legal, it shouldn't have been created!");
3372 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3373 Node->getOperand(1));
3374 Results.push_back(Tmp1.getValue(1));
3379 // Expand into divrem libcall
3380 ExpandDivRemLibCall(Node, Results);
3383 EVT VT = Node->getValueType(0);
3384 SDVTList VTs = DAG.getVTList(VT, VT);
3385 // See if multiply or divide can be lowered using two-result operations.
3386 // We just need the low half of the multiply; try both the signed
3387 // and unsigned forms. If the target supports both SMUL_LOHI and
3388 // UMUL_LOHI, form a preference by checking which forms of plain
3389 // MULH it supports.
3390 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3391 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3392 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3393 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3394 unsigned OpToUse = 0;
3395 if (HasSMUL_LOHI && !HasMULHS) {
3396 OpToUse = ISD::SMUL_LOHI;
3397 } else if (HasUMUL_LOHI && !HasMULHU) {
3398 OpToUse = ISD::UMUL_LOHI;
3399 } else if (HasSMUL_LOHI) {
3400 OpToUse = ISD::SMUL_LOHI;
3401 } else if (HasUMUL_LOHI) {
3402 OpToUse = ISD::UMUL_LOHI;
3405 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3406 Node->getOperand(1)));
3409 Tmp1 = ExpandIntLibCall(Node, false,
3411 RTLIB::MUL_I16, RTLIB::MUL_I32,
3412 RTLIB::MUL_I64, RTLIB::MUL_I128);
3413 Results.push_back(Tmp1);
3418 SDValue LHS = Node->getOperand(0);
3419 SDValue RHS = Node->getOperand(1);
3420 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3421 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3423 Results.push_back(Sum);
3424 EVT OType = Node->getValueType(1);
3426 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3428 // LHSSign -> LHS >= 0
3429 // RHSSign -> RHS >= 0
3430 // SumSign -> Sum >= 0
3433 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3435 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3437 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3438 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3439 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3440 Node->getOpcode() == ISD::SADDO ?
3441 ISD::SETEQ : ISD::SETNE);
3443 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3444 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3446 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3447 Results.push_back(Cmp);
3452 SDValue LHS = Node->getOperand(0);
3453 SDValue RHS = Node->getOperand(1);
3454 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3455 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3457 Results.push_back(Sum);
3458 Results.push_back(DAG.getSetCC(dl, Node->getValueType(1), Sum, LHS,
3459 Node->getOpcode () == ISD::UADDO ?
3460 ISD::SETULT : ISD::SETUGT));
3465 EVT VT = Node->getValueType(0);
3466 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3467 SDValue LHS = Node->getOperand(0);
3468 SDValue RHS = Node->getOperand(1);
3471 static const unsigned Ops[2][3] =
3472 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3473 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3474 bool isSigned = Node->getOpcode() == ISD::SMULO;
3475 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3476 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3477 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3478 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3479 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3481 TopHalf = BottomHalf.getValue(1);
3482 } else if (TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
3483 VT.getSizeInBits() * 2))) {
3484 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3485 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3486 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3487 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3488 DAG.getIntPtrConstant(0));
3489 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3490 DAG.getIntPtrConstant(1));
3492 // We can fall back to a libcall with an illegal type for the MUL if we
3493 // have a libcall big enough.
3494 // Also, we can fall back to a division in some cases, but that's a big
3495 // performance hit in the general case.
3496 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3497 if (WideVT == MVT::i16)
3498 LC = RTLIB::MUL_I16;
3499 else if (WideVT == MVT::i32)
3500 LC = RTLIB::MUL_I32;
3501 else if (WideVT == MVT::i64)
3502 LC = RTLIB::MUL_I64;
3503 else if (WideVT == MVT::i128)
3504 LC = RTLIB::MUL_I128;
3505 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3507 // The high part is obtained by SRA'ing all but one of the bits of low
3509 unsigned LoSize = VT.getSizeInBits();
3510 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3511 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3512 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3513 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3515 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3516 // pre-lowered to the correct types. This all depends upon WideVT not
3517 // being a legal type for the architecture and thus has to be split to
3519 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3520 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3521 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3522 DAG.getIntPtrConstant(0));
3523 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3524 DAG.getIntPtrConstant(1));
3525 // Ret is a node with an illegal type. Because such things are not
3526 // generally permitted during this phase of legalization, delete the
3527 // node. The above EXTRACT_ELEMENT nodes should have been folded.
3528 DAG.DeleteNode(Ret.getNode());
3532 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3533 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3534 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3535 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3538 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3539 DAG.getConstant(0, VT), ISD::SETNE);
3541 Results.push_back(BottomHalf);
3542 Results.push_back(TopHalf);
3545 case ISD::BUILD_PAIR: {
3546 EVT PairTy = Node->getValueType(0);
3547 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3548 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3549 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3550 DAG.getConstant(PairTy.getSizeInBits()/2,
3551 TLI.getShiftAmountTy(PairTy)));
3552 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3556 Tmp1 = Node->getOperand(0);
3557 Tmp2 = Node->getOperand(1);
3558 Tmp3 = Node->getOperand(2);
3559 if (Tmp1.getOpcode() == ISD::SETCC) {
3560 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3562 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3564 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3565 DAG.getConstant(0, Tmp1.getValueType()),
3566 Tmp2, Tmp3, ISD::SETNE);
3568 Results.push_back(Tmp1);
3571 SDValue Chain = Node->getOperand(0);
3572 SDValue Table = Node->getOperand(1);
3573 SDValue Index = Node->getOperand(2);
3575 EVT PTy = TLI.getPointerTy();
3577 const DataLayout &TD = *TLI.getDataLayout();
3578 unsigned EntrySize =
3579 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3581 Index = DAG.getNode(ISD::MUL, dl, PTy,
3582 Index, DAG.getConstant(EntrySize, PTy));
3583 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3585 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3586 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3587 MachinePointerInfo::getJumpTable(), MemVT,
3590 if (TM.getRelocationModel() == Reloc::PIC_) {
3591 // For PIC, the sequence is:
3592 // BRIND(load(Jumptable + index) + RelocBase)
3593 // RelocBase can be JumpTable, GOT or some sort of global base.
3594 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3595 TLI.getPICJumpTableRelocBase(Table, DAG));
3597 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3598 Results.push_back(Tmp1);
3602 // Expand brcond's setcc into its constituent parts and create a BR_CC
3604 Tmp1 = Node->getOperand(0);
3605 Tmp2 = Node->getOperand(1);
3606 if (Tmp2.getOpcode() == ISD::SETCC) {
3607 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3608 Tmp1, Tmp2.getOperand(2),
3609 Tmp2.getOperand(0), Tmp2.getOperand(1),
3610 Node->getOperand(2));
3612 // We test only the i1 bit. Skip the AND if UNDEF.
3613 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3614 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3615 DAG.getConstant(1, Tmp2.getValueType()));
3616 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3617 DAG.getCondCode(ISD::SETNE), Tmp3,
3618 DAG.getConstant(0, Tmp3.getValueType()),
3619 Node->getOperand(2));
3621 Results.push_back(Tmp1);
3624 Tmp1 = Node->getOperand(0);
3625 Tmp2 = Node->getOperand(1);
3626 Tmp3 = Node->getOperand(2);
3627 LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2, Tmp3, dl);
3629 // If we expanded the SETCC into an AND/OR, return the new node
3630 if (Tmp2.getNode() == 0) {
3631 Results.push_back(Tmp1);
3635 // Otherwise, SETCC for the given comparison type must be completely
3636 // illegal; expand it into a SELECT_CC.
3637 EVT VT = Node->getValueType(0);
3639 switch (TLI.getBooleanContents(VT.isVector())) {
3640 case TargetLowering::ZeroOrOneBooleanContent:
3641 case TargetLowering::UndefinedBooleanContent:
3644 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3648 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3649 DAG.getConstant(TrueValue, VT), DAG.getConstant(0, VT),
3651 Results.push_back(Tmp1);
3654 case ISD::SELECT_CC: {
3655 Tmp1 = Node->getOperand(0); // LHS
3656 Tmp2 = Node->getOperand(1); // RHS
3657 Tmp3 = Node->getOperand(2); // True
3658 Tmp4 = Node->getOperand(3); // False
3659 SDValue CC = Node->getOperand(4);
3661 LegalizeSetCCCondCode(getSetCCResultType(Tmp1.getValueType()),
3662 Tmp1, Tmp2, CC, dl);
3664 assert(!Tmp2.getNode() && "Can't legalize SELECT_CC with legal condition!");
3665 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
3666 CC = DAG.getCondCode(ISD::SETNE);
3667 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1, Tmp2,
3669 Results.push_back(Tmp1);
3673 Tmp1 = Node->getOperand(0); // Chain
3674 Tmp2 = Node->getOperand(2); // LHS
3675 Tmp3 = Node->getOperand(3); // RHS
3676 Tmp4 = Node->getOperand(1); // CC
3678 LegalizeSetCCCondCode(getSetCCResultType(Tmp2.getValueType()),
3679 Tmp2, Tmp3, Tmp4, dl);
3681 assert(!Tmp3.getNode() && "Can't legalize BR_CC with legal condition!");
3682 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
3683 Tmp4 = DAG.getCondCode(ISD::SETNE);
3684 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4, Tmp2,
3685 Tmp3, Node->getOperand(4));
3686 Results.push_back(Tmp1);
3689 case ISD::BUILD_VECTOR:
3690 Results.push_back(ExpandBUILD_VECTOR(Node));
3695 // Scalarize vector SRA/SRL/SHL.
3696 EVT VT = Node->getValueType(0);
3697 assert(VT.isVector() && "Unable to legalize non-vector shift");
3698 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
3699 unsigned NumElem = VT.getVectorNumElements();
3701 SmallVector<SDValue, 8> Scalars;
3702 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
3703 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3705 Node->getOperand(0), DAG.getIntPtrConstant(Idx));
3706 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3708 Node->getOperand(1), DAG.getIntPtrConstant(Idx));
3709 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
3710 VT.getScalarType(), Ex, Sh));
3713 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
3714 &Scalars[0], Scalars.size());
3715 ReplaceNode(SDValue(Node, 0), Result);
3718 case ISD::GLOBAL_OFFSET_TABLE:
3719 case ISD::GlobalAddress:
3720 case ISD::GlobalTLSAddress:
3721 case ISD::ExternalSymbol:
3722 case ISD::ConstantPool:
3723 case ISD::JumpTable:
3724 case ISD::INTRINSIC_W_CHAIN:
3725 case ISD::INTRINSIC_WO_CHAIN:
3726 case ISD::INTRINSIC_VOID:
3727 // FIXME: Custom lowering for these operations shouldn't return null!
3731 // Replace the original node with the legalized result.
3732 if (!Results.empty())
3733 ReplaceNode(Node, Results.data());
3736 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
3737 SmallVector<SDValue, 8> Results;
3738 MVT OVT = Node->getSimpleValueType(0);
3739 if (Node->getOpcode() == ISD::UINT_TO_FP ||
3740 Node->getOpcode() == ISD::SINT_TO_FP ||
3741 Node->getOpcode() == ISD::SETCC) {
3742 OVT = Node->getOperand(0).getSimpleValueType();
3744 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
3746 SDValue Tmp1, Tmp2, Tmp3;
3747 switch (Node->getOpcode()) {
3749 case ISD::CTTZ_ZERO_UNDEF:
3751 case ISD::CTLZ_ZERO_UNDEF:
3753 // Zero extend the argument.
3754 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3755 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
3756 // already the correct result.
3757 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3758 if (Node->getOpcode() == ISD::CTTZ) {
3759 // FIXME: This should set a bit in the zero extended value instead.
3760 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
3761 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
3763 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2,
3764 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
3765 } else if (Node->getOpcode() == ISD::CTLZ ||
3766 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
3767 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
3768 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
3769 DAG.getConstant(NVT.getSizeInBits() -
3770 OVT.getSizeInBits(), NVT));
3772 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
3775 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
3776 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3777 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
3778 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
3779 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
3780 Results.push_back(Tmp1);
3783 case ISD::FP_TO_UINT:
3784 case ISD::FP_TO_SINT:
3785 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
3786 Node->getOpcode() == ISD::FP_TO_SINT, dl);
3787 Results.push_back(Tmp1);
3789 case ISD::UINT_TO_FP:
3790 case ISD::SINT_TO_FP:
3791 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
3792 Node->getOpcode() == ISD::SINT_TO_FP, dl);
3793 Results.push_back(Tmp1);
3796 SDValue Chain = Node->getOperand(0); // Get the chain.
3797 SDValue Ptr = Node->getOperand(1); // Get the pointer.
3800 if (OVT.isVector()) {
3801 TruncOp = ISD::BITCAST;
3803 assert(OVT.isInteger()
3804 && "VAARG promotion is supported only for vectors or integer types");
3805 TruncOp = ISD::TRUNCATE;
3808 // Perform the larger operation, then convert back
3809 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
3810 Node->getConstantOperandVal(3));
3811 Chain = Tmp1.getValue(1);
3813 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
3815 // Modified the chain result - switch anything that used the old chain to
3817 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
3818 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
3825 unsigned ExtOp, TruncOp;
3826 if (OVT.isVector()) {
3827 ExtOp = ISD::BITCAST;
3828 TruncOp = ISD::BITCAST;
3830 assert(OVT.isInteger() && "Cannot promote logic operation");
3831 ExtOp = ISD::ANY_EXTEND;
3832 TruncOp = ISD::TRUNCATE;
3834 // Promote each of the values to the new type.
3835 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3836 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3837 // Perform the larger operation, then convert back
3838 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3839 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
3843 unsigned ExtOp, TruncOp;
3844 if (Node->getValueType(0).isVector()) {
3845 ExtOp = ISD::BITCAST;
3846 TruncOp = ISD::BITCAST;
3847 } else if (Node->getValueType(0).isInteger()) {
3848 ExtOp = ISD::ANY_EXTEND;
3849 TruncOp = ISD::TRUNCATE;
3851 ExtOp = ISD::FP_EXTEND;
3852 TruncOp = ISD::FP_ROUND;
3854 Tmp1 = Node->getOperand(0);
3855 // Promote each of the values to the new type.
3856 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3857 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
3858 // Perform the larger operation, then round down.
3859 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp1, Tmp2, Tmp3);
3860 if (TruncOp != ISD::FP_ROUND)
3861 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
3863 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
3864 DAG.getIntPtrConstant(0));
3865 Results.push_back(Tmp1);
3868 case ISD::VECTOR_SHUFFLE: {
3869 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3871 // Cast the two input vectors.
3872 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
3873 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
3875 // Convert the shuffle mask to the right # elements.
3876 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
3877 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
3878 Results.push_back(Tmp1);
3882 unsigned ExtOp = ISD::FP_EXTEND;
3883 if (NVT.isInteger()) {
3884 ISD::CondCode CCCode =
3885 cast<CondCodeSDNode>(Node->getOperand(2))->get();
3886 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3888 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3889 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3890 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3891 Tmp1, Tmp2, Node->getOperand(2)));
3897 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
3898 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
3899 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3900 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
3901 Tmp3, DAG.getIntPtrConstant(0)));
3908 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
3909 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3910 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
3911 Tmp2, DAG.getIntPtrConstant(0)));
3916 // Replace the original node with the legalized result.
3917 if (!Results.empty())
3918 ReplaceNode(Node, Results.data());
3921 // SelectionDAG::Legalize - This is the entry point for the file.
3923 void SelectionDAG::Legalize() {
3924 /// run - This is the main entry point to this class.
3926 SelectionDAGLegalize(*this).LegalizeDAG();