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/DenseMap.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineJumpTableInfo.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DataLayout.h"
24 #include "llvm/DebugInfo.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/LLVMContext.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetFrameLowering.h"
32 #include "llvm/Target/TargetLowering.h"
33 #include "llvm/Target/TargetMachine.h"
36 //===----------------------------------------------------------------------===//
37 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
38 /// hacks on it until the target machine can handle it. This involves
39 /// eliminating value sizes the machine cannot handle (promoting small sizes to
40 /// large sizes or splitting up large values into small values) as well as
41 /// eliminating operations the machine cannot handle.
43 /// This code also does a small amount of optimization and recognition of idioms
44 /// as part of its processing. For example, if a target does not support a
45 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
46 /// will attempt merge setcc and brc instructions into brcc's.
49 class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
50 const TargetMachine &TM;
51 const TargetLowering &TLI;
54 /// LegalizePosition - The iterator for walking through the node list.
55 SelectionDAG::allnodes_iterator LegalizePosition;
57 /// LegalizedNodes - The set of nodes which have already been legalized.
58 SmallPtrSet<SDNode *, 16> LegalizedNodes;
60 // Libcall insertion helpers.
63 explicit SelectionDAGLegalize(SelectionDAG &DAG);
68 /// LegalizeOp - Legalizes the given operation.
69 void LegalizeOp(SDNode *Node);
71 SDValue OptimizeFloatStore(StoreSDNode *ST);
73 void LegalizeLoadOps(SDNode *Node);
74 void LegalizeStoreOps(SDNode *Node);
76 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
77 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
78 /// is necessary to spill the vector being inserted into to memory, perform
79 /// the insert there, and then read the result back.
80 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
81 SDValue Idx, DebugLoc dl);
82 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
83 SDValue Idx, DebugLoc dl);
85 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
86 /// performs the same shuffe in terms of order or result bytes, but on a type
87 /// whose vector element type is narrower than the original shuffle type.
88 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
89 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
90 SDValue N1, SDValue N2,
91 ArrayRef<int> Mask) const;
93 void LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
96 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
97 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
98 unsigned NumOps, bool isSigned, DebugLoc dl);
100 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
101 SDNode *Node, bool isSigned);
102 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
103 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
104 RTLIB::Libcall Call_PPCF128);
105 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
106 RTLIB::Libcall Call_I8,
107 RTLIB::Libcall Call_I16,
108 RTLIB::Libcall Call_I32,
109 RTLIB::Libcall Call_I64,
110 RTLIB::Libcall Call_I128);
111 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
113 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, DebugLoc dl);
114 SDValue ExpandBUILD_VECTOR(SDNode *Node);
115 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
116 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
117 SmallVectorImpl<SDValue> &Results);
118 SDValue ExpandFCOPYSIGN(SDNode *Node);
119 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
121 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
123 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
126 SDValue ExpandBSWAP(SDValue Op, DebugLoc dl);
127 SDValue ExpandBitCount(unsigned Opc, SDValue Op, DebugLoc dl);
129 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
130 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
131 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
133 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
135 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
137 void ExpandNode(SDNode *Node);
138 void PromoteNode(SDNode *Node);
140 void ForgetNode(SDNode *N) {
141 LegalizedNodes.erase(N);
142 if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
147 // DAGUpdateListener implementation.
148 virtual void NodeDeleted(SDNode *N, SDNode *E) {
151 virtual void NodeUpdated(SDNode *N) {}
153 // Node replacement helpers
154 void ReplacedNode(SDNode *N) {
155 if (N->use_empty()) {
156 DAG.RemoveDeadNode(N);
161 void ReplaceNode(SDNode *Old, SDNode *New) {
162 DAG.ReplaceAllUsesWith(Old, New);
165 void ReplaceNode(SDValue Old, SDValue New) {
166 DAG.ReplaceAllUsesWith(Old, New);
167 ReplacedNode(Old.getNode());
169 void ReplaceNode(SDNode *Old, const SDValue *New) {
170 DAG.ReplaceAllUsesWith(Old, New);
176 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
177 /// performs the same shuffe in terms of order or result bytes, but on a type
178 /// whose vector element type is narrower than the original shuffle type.
179 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
181 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
182 SDValue N1, SDValue N2,
183 ArrayRef<int> Mask) const {
184 unsigned NumMaskElts = VT.getVectorNumElements();
185 unsigned NumDestElts = NVT.getVectorNumElements();
186 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
188 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
190 if (NumEltsGrowth == 1)
191 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
193 SmallVector<int, 8> NewMask;
194 for (unsigned i = 0; i != NumMaskElts; ++i) {
196 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
198 NewMask.push_back(-1);
200 NewMask.push_back(Idx * NumEltsGrowth + j);
203 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
204 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
205 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
208 SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
209 : SelectionDAG::DAGUpdateListener(dag),
210 TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
214 void SelectionDAGLegalize::LegalizeDAG() {
215 DAG.AssignTopologicalOrder();
217 // Visit all the nodes. We start in topological order, so that we see
218 // nodes with their original operands intact. Legalization can produce
219 // new nodes which may themselves need to be legalized. Iterate until all
220 // nodes have been legalized.
222 bool AnyLegalized = false;
223 for (LegalizePosition = DAG.allnodes_end();
224 LegalizePosition != DAG.allnodes_begin(); ) {
227 SDNode *N = LegalizePosition;
228 if (LegalizedNodes.insert(N)) {
238 // Remove dead nodes now.
239 DAG.RemoveDeadNodes();
242 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
243 /// a load from the constant pool.
245 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
247 DebugLoc dl = CFP->getDebugLoc();
249 // If a FP immediate is precise when represented as a float and if the
250 // target can do an extending load from float to double, we put it into
251 // the constant pool as a float, even if it's is statically typed as a
252 // double. This shrinks FP constants and canonicalizes them for targets where
253 // an FP extending load is the same cost as a normal load (such as on the x87
254 // fp stack or PPC FP unit).
255 EVT VT = CFP->getValueType(0);
256 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
258 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
259 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
260 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
265 while (SVT != MVT::f32) {
266 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
267 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
268 // Only do this if the target has a native EXTLOAD instruction from
270 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
271 TLI.ShouldShrinkFPConstant(OrigVT)) {
272 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
273 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
279 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
280 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
283 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
285 CPIdx, MachinePointerInfo::getConstantPool(),
286 VT, false, false, Alignment);
290 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
291 MachinePointerInfo::getConstantPool(), false, false, false,
296 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
297 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
298 const TargetLowering &TLI,
299 SelectionDAGLegalize *DAGLegalize) {
300 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
301 "unaligned indexed stores not implemented!");
302 SDValue Chain = ST->getChain();
303 SDValue Ptr = ST->getBasePtr();
304 SDValue Val = ST->getValue();
305 EVT VT = Val.getValueType();
306 int Alignment = ST->getAlignment();
307 DebugLoc dl = ST->getDebugLoc();
308 if (ST->getMemoryVT().isFloatingPoint() ||
309 ST->getMemoryVT().isVector()) {
310 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
311 if (TLI.isTypeLegal(intVT)) {
312 // Expand to a bitconvert of the value to the integer type of the
313 // same size, then a (misaligned) int store.
314 // FIXME: Does not handle truncating floating point stores!
315 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
316 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
317 ST->isVolatile(), ST->isNonTemporal(), Alignment);
318 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
321 // Do a (aligned) store to a stack slot, then copy from the stack slot
322 // to the final destination using (unaligned) integer loads and stores.
323 EVT StoredVT = ST->getMemoryVT();
325 TLI.getRegisterType(*DAG.getContext(),
326 EVT::getIntegerVT(*DAG.getContext(),
327 StoredVT.getSizeInBits()));
328 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
329 unsigned RegBytes = RegVT.getSizeInBits() / 8;
330 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
332 // Make sure the stack slot is also aligned for the register type.
333 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
335 // Perform the original store, only redirected to the stack slot.
336 SDValue Store = DAG.getTruncStore(Chain, dl,
337 Val, StackPtr, MachinePointerInfo(),
338 StoredVT, false, false, 0);
339 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
340 SmallVector<SDValue, 8> Stores;
343 // Do all but one copies using the full register width.
344 for (unsigned i = 1; i < NumRegs; i++) {
345 // Load one integer register's worth from the stack slot.
346 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
347 MachinePointerInfo(),
348 false, false, false, 0);
349 // Store it to the final location. Remember the store.
350 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
351 ST->getPointerInfo().getWithOffset(Offset),
352 ST->isVolatile(), ST->isNonTemporal(),
353 MinAlign(ST->getAlignment(), Offset)));
354 // Increment the pointers.
356 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
358 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
361 // The last store may be partial. Do a truncating store. On big-endian
362 // machines this requires an extending load from the stack slot to ensure
363 // that the bits are in the right place.
364 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
365 8 * (StoredBytes - Offset));
367 // Load from the stack slot.
368 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
369 MachinePointerInfo(),
370 MemVT, false, false, 0);
372 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
374 .getWithOffset(Offset),
375 MemVT, ST->isVolatile(),
377 MinAlign(ST->getAlignment(), Offset)));
378 // The order of the stores doesn't matter - say it with a TokenFactor.
380 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
382 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
385 assert(ST->getMemoryVT().isInteger() &&
386 !ST->getMemoryVT().isVector() &&
387 "Unaligned store of unknown type.");
388 // Get the half-size VT
389 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
390 int NumBits = NewStoredVT.getSizeInBits();
391 int IncrementSize = NumBits / 8;
393 // Divide the stored value in two parts.
394 SDValue ShiftAmount = DAG.getConstant(NumBits,
395 TLI.getShiftAmountTy(Val.getValueType()));
397 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
399 // Store the two parts
400 SDValue Store1, Store2;
401 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
402 ST->getPointerInfo(), NewStoredVT,
403 ST->isVolatile(), ST->isNonTemporal(), Alignment);
404 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
405 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
406 Alignment = MinAlign(Alignment, IncrementSize);
407 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
408 ST->getPointerInfo().getWithOffset(IncrementSize),
409 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
413 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
414 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
417 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
419 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
420 const TargetLowering &TLI,
421 SDValue &ValResult, SDValue &ChainResult) {
422 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
423 "unaligned indexed loads not implemented!");
424 SDValue Chain = LD->getChain();
425 SDValue Ptr = LD->getBasePtr();
426 EVT VT = LD->getValueType(0);
427 EVT LoadedVT = LD->getMemoryVT();
428 DebugLoc dl = LD->getDebugLoc();
429 if (VT.isFloatingPoint() || VT.isVector()) {
430 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
431 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
432 // Expand to a (misaligned) integer load of the same size,
433 // then bitconvert to floating point or vector.
434 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getPointerInfo(),
437 LD->isInvariant(), LD->getAlignment());
438 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
440 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
441 ISD::ANY_EXTEND, dl, VT, Result);
448 // Copy the value to a (aligned) stack slot using (unaligned) integer
449 // loads and stores, then do a (aligned) load from the stack slot.
450 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
451 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
452 unsigned RegBytes = RegVT.getSizeInBits() / 8;
453 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
455 // Make sure the stack slot is also aligned for the register type.
456 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
458 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
459 SmallVector<SDValue, 8> Stores;
460 SDValue StackPtr = StackBase;
463 // Do all but one copies using the full register width.
464 for (unsigned i = 1; i < NumRegs; i++) {
465 // Load one integer register's worth from the original location.
466 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
467 LD->getPointerInfo().getWithOffset(Offset),
468 LD->isVolatile(), LD->isNonTemporal(),
470 MinAlign(LD->getAlignment(), Offset));
471 // Follow the load with a store to the stack slot. Remember the store.
472 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
473 MachinePointerInfo(), false, false, 0));
474 // Increment the pointers.
476 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
477 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
481 // The last copy may be partial. Do an extending load.
482 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
483 8 * (LoadedBytes - Offset));
484 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
485 LD->getPointerInfo().getWithOffset(Offset),
486 MemVT, LD->isVolatile(),
488 MinAlign(LD->getAlignment(), Offset));
489 // Follow the load with a store to the stack slot. Remember the store.
490 // On big-endian machines this requires a truncating store to ensure
491 // that the bits end up in the right place.
492 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
493 MachinePointerInfo(), MemVT,
496 // The order of the stores doesn't matter - say it with a TokenFactor.
497 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
500 // Finally, perform the original load only redirected to the stack slot.
501 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
502 MachinePointerInfo(), LoadedVT, false, false, 0);
504 // Callers expect a MERGE_VALUES node.
509 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
510 "Unaligned load of unsupported type.");
512 // Compute the new VT that is half the size of the old one. This is an
514 unsigned NumBits = LoadedVT.getSizeInBits();
516 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
519 unsigned Alignment = LD->getAlignment();
520 unsigned IncrementSize = NumBits / 8;
521 ISD::LoadExtType HiExtType = LD->getExtensionType();
523 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
524 if (HiExtType == ISD::NON_EXTLOAD)
525 HiExtType = ISD::ZEXTLOAD;
527 // Load the value in two parts
529 if (TLI.isLittleEndian()) {
530 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
531 NewLoadedVT, LD->isVolatile(),
532 LD->isNonTemporal(), Alignment);
533 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
534 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
535 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
536 LD->getPointerInfo().getWithOffset(IncrementSize),
537 NewLoadedVT, LD->isVolatile(),
538 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
540 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
541 NewLoadedVT, LD->isVolatile(),
542 LD->isNonTemporal(), Alignment);
543 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
544 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
545 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
546 LD->getPointerInfo().getWithOffset(IncrementSize),
547 NewLoadedVT, LD->isVolatile(),
548 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
551 // aggregate the two parts
552 SDValue ShiftAmount = DAG.getConstant(NumBits,
553 TLI.getShiftAmountTy(Hi.getValueType()));
554 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
555 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
557 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
564 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
565 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
566 /// is necessary to spill the vector being inserted into to memory, perform
567 /// the insert there, and then read the result back.
568 SDValue SelectionDAGLegalize::
569 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
575 // If the target doesn't support this, we have to spill the input vector
576 // to a temporary stack slot, update the element, then reload it. This is
577 // badness. We could also load the value into a vector register (either
578 // with a "move to register" or "extload into register" instruction, then
579 // permute it into place, if the idx is a constant and if the idx is
580 // supported by the target.
581 EVT VT = Tmp1.getValueType();
582 EVT EltVT = VT.getVectorElementType();
583 EVT IdxVT = Tmp3.getValueType();
584 EVT PtrVT = TLI.getPointerTy();
585 SDValue StackPtr = DAG.CreateStackTemporary(VT);
587 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
590 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
591 MachinePointerInfo::getFixedStack(SPFI),
594 // Truncate or zero extend offset to target pointer type.
595 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
596 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
597 // Add the offset to the index.
598 unsigned EltSize = EltVT.getSizeInBits()/8;
599 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
600 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
601 // Store the scalar value.
602 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
604 // Load the updated vector.
605 return DAG.getLoad(VT, dl, Ch, StackPtr,
606 MachinePointerInfo::getFixedStack(SPFI), false, false,
611 SDValue SelectionDAGLegalize::
612 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, DebugLoc dl) {
613 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
614 // SCALAR_TO_VECTOR requires that the type of the value being inserted
615 // match the element type of the vector being created, except for
616 // integers in which case the inserted value can be over width.
617 EVT EltVT = Vec.getValueType().getVectorElementType();
618 if (Val.getValueType() == EltVT ||
619 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
620 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
621 Vec.getValueType(), Val);
623 unsigned NumElts = Vec.getValueType().getVectorNumElements();
624 // We generate a shuffle of InVec and ScVec, so the shuffle mask
625 // should be 0,1,2,3,4,5... with the appropriate element replaced with
627 SmallVector<int, 8> ShufOps;
628 for (unsigned i = 0; i != NumElts; ++i)
629 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
631 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
635 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
638 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
639 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
640 // FIXME: We shouldn't do this for TargetConstantFP's.
641 // FIXME: move this to the DAG Combiner! Note that we can't regress due
642 // to phase ordering between legalized code and the dag combiner. This
643 // probably means that we need to integrate dag combiner and legalizer
645 // We generally can't do this one for long doubles.
646 SDValue Chain = ST->getChain();
647 SDValue Ptr = ST->getBasePtr();
648 unsigned Alignment = ST->getAlignment();
649 bool isVolatile = ST->isVolatile();
650 bool isNonTemporal = ST->isNonTemporal();
651 DebugLoc dl = ST->getDebugLoc();
652 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
653 if (CFP->getValueType(0) == MVT::f32 &&
654 TLI.isTypeLegal(MVT::i32)) {
655 SDValue Con = DAG.getConstant(CFP->getValueAPF().
656 bitcastToAPInt().zextOrTrunc(32),
658 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
659 isVolatile, isNonTemporal, Alignment);
662 if (CFP->getValueType(0) == MVT::f64) {
663 // If this target supports 64-bit registers, do a single 64-bit store.
664 if (TLI.isTypeLegal(MVT::i64)) {
665 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
666 zextOrTrunc(64), MVT::i64);
667 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
668 isVolatile, isNonTemporal, Alignment);
671 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
672 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
673 // stores. If the target supports neither 32- nor 64-bits, this
674 // xform is certainly not worth it.
675 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
676 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
677 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
678 if (TLI.isBigEndian()) std::swap(Lo, Hi);
680 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
681 isNonTemporal, Alignment);
682 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
683 DAG.getIntPtrConstant(4));
684 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
685 ST->getPointerInfo().getWithOffset(4),
686 isVolatile, isNonTemporal, MinAlign(Alignment, 4U));
688 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
692 return SDValue(0, 0);
695 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
696 StoreSDNode *ST = cast<StoreSDNode>(Node);
697 SDValue Chain = ST->getChain();
698 SDValue Ptr = ST->getBasePtr();
699 DebugLoc dl = Node->getDebugLoc();
701 unsigned Alignment = ST->getAlignment();
702 bool isVolatile = ST->isVolatile();
703 bool isNonTemporal = ST->isNonTemporal();
705 if (!ST->isTruncatingStore()) {
706 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
707 ReplaceNode(ST, OptStore);
712 SDValue Value = ST->getValue();
713 MVT VT = Value.getSimpleValueType();
714 switch (TLI.getOperationAction(ISD::STORE, VT)) {
715 default: llvm_unreachable("This action is not supported yet!");
716 case TargetLowering::Legal:
717 // If this is an unaligned store and the target doesn't support it,
719 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
720 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
721 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
722 if (ST->getAlignment() < ABIAlignment)
723 ExpandUnalignedStore(cast<StoreSDNode>(Node),
727 case TargetLowering::Custom: {
728 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
730 ReplaceNode(SDValue(Node, 0), Res);
733 case TargetLowering::Promote: {
734 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
735 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
736 "Can only promote stores to same size type");
737 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
739 DAG.getStore(Chain, dl, Value, Ptr,
740 ST->getPointerInfo(), isVolatile,
741 isNonTemporal, Alignment);
742 ReplaceNode(SDValue(Node, 0), Result);
749 SDValue Value = ST->getValue();
751 EVT StVT = ST->getMemoryVT();
752 unsigned StWidth = StVT.getSizeInBits();
754 if (StWidth != StVT.getStoreSizeInBits()) {
755 // Promote to a byte-sized store with upper bits zero if not
756 // storing an integral number of bytes. For example, promote
757 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
758 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
759 StVT.getStoreSizeInBits());
760 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
762 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
763 NVT, isVolatile, isNonTemporal, Alignment);
764 ReplaceNode(SDValue(Node, 0), Result);
765 } else if (StWidth & (StWidth - 1)) {
766 // If not storing a power-of-2 number of bits, expand as two stores.
767 assert(!StVT.isVector() && "Unsupported truncstore!");
768 unsigned RoundWidth = 1 << Log2_32(StWidth);
769 assert(RoundWidth < StWidth);
770 unsigned ExtraWidth = StWidth - RoundWidth;
771 assert(ExtraWidth < RoundWidth);
772 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
773 "Store size not an integral number of bytes!");
774 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
775 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
777 unsigned IncrementSize;
779 if (TLI.isLittleEndian()) {
780 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
781 // Store the bottom RoundWidth bits.
782 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
784 isVolatile, isNonTemporal, Alignment);
786 // Store the remaining ExtraWidth bits.
787 IncrementSize = RoundWidth / 8;
788 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
789 DAG.getIntPtrConstant(IncrementSize));
790 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
791 DAG.getConstant(RoundWidth,
792 TLI.getShiftAmountTy(Value.getValueType())));
793 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
794 ST->getPointerInfo().getWithOffset(IncrementSize),
795 ExtraVT, isVolatile, isNonTemporal,
796 MinAlign(Alignment, IncrementSize));
798 // Big endian - avoid unaligned stores.
799 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
800 // Store the top RoundWidth bits.
801 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
802 DAG.getConstant(ExtraWidth,
803 TLI.getShiftAmountTy(Value.getValueType())));
804 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
805 RoundVT, isVolatile, isNonTemporal, Alignment);
807 // Store the remaining ExtraWidth bits.
808 IncrementSize = RoundWidth / 8;
809 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
810 DAG.getIntPtrConstant(IncrementSize));
811 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
812 ST->getPointerInfo().getWithOffset(IncrementSize),
813 ExtraVT, isVolatile, isNonTemporal,
814 MinAlign(Alignment, IncrementSize));
817 // The order of the stores doesn't matter.
818 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
819 ReplaceNode(SDValue(Node, 0), Result);
821 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
822 StVT.getSimpleVT())) {
823 default: llvm_unreachable("This action is not supported yet!");
824 case TargetLowering::Legal:
825 // If this is an unaligned store and the target doesn't support it,
827 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
828 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
829 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
830 if (ST->getAlignment() < ABIAlignment)
831 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
834 case TargetLowering::Custom: {
835 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
837 ReplaceNode(SDValue(Node, 0), Res);
840 case TargetLowering::Expand:
841 assert(!StVT.isVector() &&
842 "Vector Stores are handled in LegalizeVectorOps");
844 // TRUNCSTORE:i16 i32 -> STORE i16
845 assert(TLI.isTypeLegal(StVT) &&
846 "Do not know how to expand this store!");
847 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
849 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
850 isVolatile, isNonTemporal, Alignment);
851 ReplaceNode(SDValue(Node, 0), Result);
858 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
859 LoadSDNode *LD = cast<LoadSDNode>(Node);
860 SDValue Chain = LD->getChain(); // The chain.
861 SDValue Ptr = LD->getBasePtr(); // The base pointer.
862 SDValue Value; // The value returned by the load op.
863 DebugLoc dl = Node->getDebugLoc();
865 ISD::LoadExtType ExtType = LD->getExtensionType();
866 if (ExtType == ISD::NON_EXTLOAD) {
867 MVT VT = Node->getSimpleValueType(0);
868 SDValue RVal = SDValue(Node, 0);
869 SDValue RChain = SDValue(Node, 1);
871 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
872 default: llvm_unreachable("This action is not supported yet!");
873 case TargetLowering::Legal:
874 // If this is an unaligned load and the target doesn't support it,
876 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
877 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
878 unsigned ABIAlignment =
879 TLI.getDataLayout()->getABITypeAlignment(Ty);
880 if (LD->getAlignment() < ABIAlignment){
881 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
885 case TargetLowering::Custom: {
886 SDValue Res = TLI.LowerOperation(RVal, DAG);
889 RChain = Res.getValue(1);
893 case TargetLowering::Promote: {
894 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
895 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
896 "Can only promote loads to same size type");
898 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getPointerInfo(),
899 LD->isVolatile(), LD->isNonTemporal(),
900 LD->isInvariant(), LD->getAlignment());
901 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
902 RChain = Res.getValue(1);
906 if (RChain.getNode() != Node) {
907 assert(RVal.getNode() != Node && "Load must be completely replaced");
908 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
909 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
915 EVT SrcVT = LD->getMemoryVT();
916 unsigned SrcWidth = SrcVT.getSizeInBits();
917 unsigned Alignment = LD->getAlignment();
918 bool isVolatile = LD->isVolatile();
919 bool isNonTemporal = LD->isNonTemporal();
921 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
922 // Some targets pretend to have an i1 loading operation, and actually
923 // load an i8. This trick is correct for ZEXTLOAD because the top 7
924 // bits are guaranteed to be zero; it helps the optimizers understand
925 // that these bits are zero. It is also useful for EXTLOAD, since it
926 // tells the optimizers that those bits are undefined. It would be
927 // nice to have an effective generic way of getting these benefits...
928 // Until such a way is found, don't insist on promoting i1 here.
930 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
931 // Promote to a byte-sized load if not loading an integral number of
932 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
933 unsigned NewWidth = SrcVT.getStoreSizeInBits();
934 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
937 // The extra bits are guaranteed to be zero, since we stored them that
938 // way. A zext load from NVT thus automatically gives zext from SrcVT.
940 ISD::LoadExtType NewExtType =
941 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
944 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
945 Chain, Ptr, LD->getPointerInfo(),
946 NVT, isVolatile, isNonTemporal, Alignment);
948 Ch = Result.getValue(1); // The chain.
950 if (ExtType == ISD::SEXTLOAD)
951 // Having the top bits zero doesn't help when sign extending.
952 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
953 Result.getValueType(),
954 Result, DAG.getValueType(SrcVT));
955 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
956 // All the top bits are guaranteed to be zero - inform the optimizers.
957 Result = DAG.getNode(ISD::AssertZext, dl,
958 Result.getValueType(), Result,
959 DAG.getValueType(SrcVT));
963 } else if (SrcWidth & (SrcWidth - 1)) {
964 // If not loading a power-of-2 number of bits, expand as two loads.
965 assert(!SrcVT.isVector() && "Unsupported extload!");
966 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
967 assert(RoundWidth < SrcWidth);
968 unsigned ExtraWidth = SrcWidth - RoundWidth;
969 assert(ExtraWidth < RoundWidth);
970 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
971 "Load size not an integral number of bytes!");
972 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
973 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
975 unsigned IncrementSize;
977 if (TLI.isLittleEndian()) {
978 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
979 // Load the bottom RoundWidth bits.
980 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
982 LD->getPointerInfo(), RoundVT, isVolatile,
983 isNonTemporal, Alignment);
985 // Load the remaining ExtraWidth bits.
986 IncrementSize = RoundWidth / 8;
987 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
988 DAG.getIntPtrConstant(IncrementSize));
989 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
990 LD->getPointerInfo().getWithOffset(IncrementSize),
991 ExtraVT, isVolatile, isNonTemporal,
992 MinAlign(Alignment, IncrementSize));
994 // Build a factor node to remember that this load is independent of
996 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
999 // Move the top bits to the right place.
1000 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1001 DAG.getConstant(RoundWidth,
1002 TLI.getShiftAmountTy(Hi.getValueType())));
1004 // Join the hi and lo parts.
1005 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1007 // Big endian - avoid unaligned loads.
1008 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1009 // Load the top RoundWidth bits.
1010 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1011 LD->getPointerInfo(), RoundVT, isVolatile,
1012 isNonTemporal, Alignment);
1014 // Load the remaining ExtraWidth bits.
1015 IncrementSize = RoundWidth / 8;
1016 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1017 DAG.getIntPtrConstant(IncrementSize));
1018 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1019 dl, Node->getValueType(0), Chain, Ptr,
1020 LD->getPointerInfo().getWithOffset(IncrementSize),
1021 ExtraVT, isVolatile, isNonTemporal,
1022 MinAlign(Alignment, IncrementSize));
1024 // Build a factor node to remember that this load is independent of
1026 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1029 // Move the top bits to the right place.
1030 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1031 DAG.getConstant(ExtraWidth,
1032 TLI.getShiftAmountTy(Hi.getValueType())));
1034 // Join the hi and lo parts.
1035 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1040 bool isCustom = false;
1041 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1042 default: llvm_unreachable("This action is not supported yet!");
1043 case TargetLowering::Custom:
1046 case TargetLowering::Legal: {
1047 Value = SDValue(Node, 0);
1048 Chain = SDValue(Node, 1);
1051 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1052 if (Res.getNode()) {
1054 Chain = Res.getValue(1);
1057 // If this is an unaligned load and the target doesn't support it,
1059 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
1061 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1062 unsigned ABIAlignment =
1063 TLI.getDataLayout()->getABITypeAlignment(Ty);
1064 if (LD->getAlignment() < ABIAlignment){
1065 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1066 DAG, TLI, Value, Chain);
1072 case TargetLowering::Expand:
1073 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
1074 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1075 LD->getPointerInfo(),
1076 LD->isVolatile(), LD->isNonTemporal(),
1077 LD->isInvariant(), LD->getAlignment());
1081 ExtendOp = (SrcVT.isFloatingPoint() ?
1082 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1084 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1085 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1086 default: llvm_unreachable("Unexpected extend load type!");
1088 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1089 Chain = Load.getValue(1);
1093 assert(!SrcVT.isVector() &&
1094 "Vector Loads are handled in LegalizeVectorOps");
1096 // FIXME: This does not work for vectors on most targets. Sign- and
1097 // zero-extend operations are currently folded into extending loads,
1098 // whether they are legal or not, and then we end up here without any
1099 // support for legalizing them.
1100 assert(ExtType != ISD::EXTLOAD &&
1101 "EXTLOAD should always be supported!");
1102 // Turn the unsupported load into an EXTLOAD followed by an explicit
1103 // zero/sign extend inreg.
1104 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
1105 Chain, Ptr, LD->getPointerInfo(), SrcVT,
1106 LD->isVolatile(), LD->isNonTemporal(),
1107 LD->getAlignment());
1109 if (ExtType == ISD::SEXTLOAD)
1110 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1111 Result.getValueType(),
1112 Result, DAG.getValueType(SrcVT));
1114 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1116 Chain = Result.getValue(1);
1121 // Since loads produce two values, make sure to remember that we legalized
1123 if (Chain.getNode() != Node) {
1124 assert(Value.getNode() != Node && "Load must be completely replaced");
1125 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1126 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1131 /// LegalizeOp - Return a legal replacement for the given operation, with
1132 /// all legal operands.
1133 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1134 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1137 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1138 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1139 TargetLowering::TypeLegal &&
1140 "Unexpected illegal type!");
1142 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1143 assert((TLI.getTypeAction(*DAG.getContext(),
1144 Node->getOperand(i).getValueType()) ==
1145 TargetLowering::TypeLegal ||
1146 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1147 "Unexpected illegal type!");
1149 // Figure out the correct action; the way to query this varies by opcode
1150 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1151 bool SimpleFinishLegalizing = true;
1152 switch (Node->getOpcode()) {
1153 case ISD::INTRINSIC_W_CHAIN:
1154 case ISD::INTRINSIC_WO_CHAIN:
1155 case ISD::INTRINSIC_VOID:
1156 case ISD::STACKSAVE:
1157 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1160 Action = TLI.getOperationAction(Node->getOpcode(),
1161 Node->getValueType(0));
1162 if (Action != TargetLowering::Promote)
1163 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1165 case ISD::SINT_TO_FP:
1166 case ISD::UINT_TO_FP:
1167 case ISD::EXTRACT_VECTOR_ELT:
1168 Action = TLI.getOperationAction(Node->getOpcode(),
1169 Node->getOperand(0).getValueType());
1171 case ISD::FP_ROUND_INREG:
1172 case ISD::SIGN_EXTEND_INREG: {
1173 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1174 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1177 case ISD::ATOMIC_STORE: {
1178 Action = TLI.getOperationAction(Node->getOpcode(),
1179 Node->getOperand(2).getValueType());
1182 case ISD::SELECT_CC:
1185 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1186 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1187 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1188 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1189 ISD::CondCode CCCode =
1190 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1191 Action = TLI.getCondCodeAction(CCCode, OpVT);
1192 if (Action == TargetLowering::Legal) {
1193 if (Node->getOpcode() == ISD::SELECT_CC)
1194 Action = TLI.getOperationAction(Node->getOpcode(),
1195 Node->getValueType(0));
1197 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1203 // FIXME: Model these properly. LOAD and STORE are complicated, and
1204 // STORE expects the unlegalized operand in some cases.
1205 SimpleFinishLegalizing = false;
1207 case ISD::CALLSEQ_START:
1208 case ISD::CALLSEQ_END:
1209 // FIXME: This shouldn't be necessary. These nodes have special properties
1210 // dealing with the recursive nature of legalization. Removing this
1211 // special case should be done as part of making LegalizeDAG non-recursive.
1212 SimpleFinishLegalizing = false;
1214 case ISD::EXTRACT_ELEMENT:
1215 case ISD::FLT_ROUNDS_:
1223 case ISD::MERGE_VALUES:
1224 case ISD::EH_RETURN:
1225 case ISD::FRAME_TO_ARGS_OFFSET:
1226 case ISD::EH_SJLJ_SETJMP:
1227 case ISD::EH_SJLJ_LONGJMP:
1228 // These operations lie about being legal: when they claim to be legal,
1229 // they should actually be expanded.
1230 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1231 if (Action == TargetLowering::Legal)
1232 Action = TargetLowering::Expand;
1234 case ISD::INIT_TRAMPOLINE:
1235 case ISD::ADJUST_TRAMPOLINE:
1236 case ISD::FRAMEADDR:
1237 case ISD::RETURNADDR:
1238 // These operations lie about being legal: when they claim to be legal,
1239 // they should actually be custom-lowered.
1240 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1241 if (Action == TargetLowering::Legal)
1242 Action = TargetLowering::Custom;
1244 case ISD::DEBUGTRAP:
1245 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1246 if (Action == TargetLowering::Expand) {
1247 // replace ISD::DEBUGTRAP with ISD::TRAP
1249 NewVal = DAG.getNode(ISD::TRAP, Node->getDebugLoc(), Node->getVTList(),
1250 Node->getOperand(0));
1251 ReplaceNode(Node, NewVal.getNode());
1252 LegalizeOp(NewVal.getNode());
1258 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1259 Action = TargetLowering::Legal;
1261 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1266 if (SimpleFinishLegalizing) {
1267 SDNode *NewNode = Node;
1268 switch (Node->getOpcode()) {
1275 // Legalizing shifts/rotates requires adjusting the shift amount
1276 // to the appropriate width.
1277 if (!Node->getOperand(1).getValueType().isVector()) {
1279 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1280 Node->getOperand(1));
1281 HandleSDNode Handle(SAO);
1282 LegalizeOp(SAO.getNode());
1283 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1287 case ISD::SRL_PARTS:
1288 case ISD::SRA_PARTS:
1289 case ISD::SHL_PARTS:
1290 // Legalizing shifts/rotates requires adjusting the shift amount
1291 // to the appropriate width.
1292 if (!Node->getOperand(2).getValueType().isVector()) {
1294 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1295 Node->getOperand(2));
1296 HandleSDNode Handle(SAO);
1297 LegalizeOp(SAO.getNode());
1298 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1299 Node->getOperand(1),
1305 if (NewNode != Node) {
1306 DAG.ReplaceAllUsesWith(Node, NewNode);
1307 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1308 DAG.TransferDbgValues(SDValue(Node, i), SDValue(NewNode, i));
1313 case TargetLowering::Legal:
1315 case TargetLowering::Custom: {
1316 // FIXME: The handling for custom lowering with multiple results is
1318 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1319 if (Res.getNode()) {
1320 SmallVector<SDValue, 8> ResultVals;
1321 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
1323 ResultVals.push_back(Res);
1325 ResultVals.push_back(Res.getValue(i));
1327 if (Res.getNode() != Node || Res.getResNo() != 0) {
1328 DAG.ReplaceAllUsesWith(Node, ResultVals.data());
1329 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1330 DAG.TransferDbgValues(SDValue(Node, i), ResultVals[i]);
1337 case TargetLowering::Expand:
1340 case TargetLowering::Promote:
1346 switch (Node->getOpcode()) {
1353 llvm_unreachable("Do not know how to legalize this operator!");
1355 case ISD::CALLSEQ_START:
1356 case ISD::CALLSEQ_END:
1359 return LegalizeLoadOps(Node);
1362 return LegalizeStoreOps(Node);
1367 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1368 SDValue Vec = Op.getOperand(0);
1369 SDValue Idx = Op.getOperand(1);
1370 DebugLoc dl = Op.getDebugLoc();
1371 // Store the value to a temporary stack slot, then LOAD the returned part.
1372 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1373 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1374 MachinePointerInfo(), false, false, 0);
1376 // Add the offset to the index.
1378 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1379 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1380 DAG.getConstant(EltSize, Idx.getValueType()));
1382 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1383 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1385 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1387 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1389 if (Op.getValueType().isVector())
1390 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1391 false, false, false, 0);
1392 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1393 MachinePointerInfo(),
1394 Vec.getValueType().getVectorElementType(),
1398 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1399 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1401 SDValue Vec = Op.getOperand(0);
1402 SDValue Part = Op.getOperand(1);
1403 SDValue Idx = Op.getOperand(2);
1404 DebugLoc dl = Op.getDebugLoc();
1406 // Store the value to a temporary stack slot, then LOAD the returned part.
1408 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1409 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1410 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1412 // First store the whole vector.
1413 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1416 // Then store the inserted part.
1418 // Add the offset to the index.
1420 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1422 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1423 DAG.getConstant(EltSize, Idx.getValueType()));
1425 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1426 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1428 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1430 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1433 // Store the subvector.
1434 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1435 MachinePointerInfo(), false, false, 0);
1437 // Finally, load the updated vector.
1438 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1439 false, false, false, 0);
1442 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1443 // We can't handle this case efficiently. Allocate a sufficiently
1444 // aligned object on the stack, store each element into it, then load
1445 // the result as a vector.
1446 // Create the stack frame object.
1447 EVT VT = Node->getValueType(0);
1448 EVT EltVT = VT.getVectorElementType();
1449 DebugLoc dl = Node->getDebugLoc();
1450 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1451 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1452 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1454 // Emit a store of each element to the stack slot.
1455 SmallVector<SDValue, 8> Stores;
1456 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1457 // Store (in the right endianness) the elements to memory.
1458 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1459 // Ignore undef elements.
1460 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1462 unsigned Offset = TypeByteSize*i;
1464 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1465 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1467 // If the destination vector element type is narrower than the source
1468 // element type, only store the bits necessary.
1469 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1470 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1471 Node->getOperand(i), Idx,
1472 PtrInfo.getWithOffset(Offset),
1473 EltVT, false, false, 0));
1475 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1476 Node->getOperand(i), Idx,
1477 PtrInfo.getWithOffset(Offset),
1482 if (!Stores.empty()) // Not all undef elements?
1483 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1484 &Stores[0], Stores.size());
1486 StoreChain = DAG.getEntryNode();
1488 // Result is a load from the stack slot.
1489 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1490 false, false, false, 0);
1493 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1494 DebugLoc dl = Node->getDebugLoc();
1495 SDValue Tmp1 = Node->getOperand(0);
1496 SDValue Tmp2 = Node->getOperand(1);
1498 // Get the sign bit of the RHS. First obtain a value that has the same
1499 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1501 EVT FloatVT = Tmp2.getValueType();
1502 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1503 if (TLI.isTypeLegal(IVT)) {
1504 // Convert to an integer with the same sign bit.
1505 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1507 // Store the float to memory, then load the sign part out as an integer.
1508 MVT LoadTy = TLI.getPointerTy();
1509 // First create a temporary that is aligned for both the load and store.
1510 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1511 // Then store the float to it.
1513 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1515 if (TLI.isBigEndian()) {
1516 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1517 // Load out a legal integer with the same sign bit as the float.
1518 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1519 false, false, false, 0);
1520 } else { // Little endian
1521 SDValue LoadPtr = StackPtr;
1522 // The float may be wider than the integer we are going to load. Advance
1523 // the pointer so that the loaded integer will contain the sign bit.
1524 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1525 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1526 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(),
1527 LoadPtr, DAG.getIntPtrConstant(ByteOffset));
1528 // Load a legal integer containing the sign bit.
1529 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1530 false, false, false, 0);
1531 // Move the sign bit to the top bit of the loaded integer.
1532 unsigned BitShift = LoadTy.getSizeInBits() -
1533 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1534 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1536 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1537 DAG.getConstant(BitShift,
1538 TLI.getShiftAmountTy(SignBit.getValueType())));
1541 // Now get the sign bit proper, by seeing whether the value is negative.
1542 SignBit = DAG.getSetCC(dl, TLI.getSetCCResultType(SignBit.getValueType()),
1543 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1545 // Get the absolute value of the result.
1546 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1547 // Select between the nabs and abs value based on the sign bit of
1549 return DAG.getNode(ISD::SELECT, dl, AbsVal.getValueType(), SignBit,
1550 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1554 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1555 SmallVectorImpl<SDValue> &Results) {
1556 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1557 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1558 " not tell us which reg is the stack pointer!");
1559 DebugLoc dl = Node->getDebugLoc();
1560 EVT VT = Node->getValueType(0);
1561 SDValue Tmp1 = SDValue(Node, 0);
1562 SDValue Tmp2 = SDValue(Node, 1);
1563 SDValue Tmp3 = Node->getOperand(2);
1564 SDValue Chain = Tmp1.getOperand(0);
1566 // Chain the dynamic stack allocation so that it doesn't modify the stack
1567 // pointer when other instructions are using the stack.
1568 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1570 SDValue Size = Tmp2.getOperand(1);
1571 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1572 Chain = SP.getValue(1);
1573 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1574 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1575 if (Align > StackAlign)
1576 SP = DAG.getNode(ISD::AND, dl, VT, SP,
1577 DAG.getConstant(-(uint64_t)Align, VT));
1578 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1579 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1581 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1582 DAG.getIntPtrConstant(0, true), SDValue());
1584 Results.push_back(Tmp1);
1585 Results.push_back(Tmp2);
1588 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1589 /// condition code CC on the current target. This routine expands SETCC with
1590 /// illegal condition code into AND / OR of multiple SETCC values.
1591 void SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1592 SDValue &LHS, SDValue &RHS,
1595 MVT OpVT = LHS.getSimpleValueType();
1596 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1597 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1598 default: llvm_unreachable("Unknown condition code action!");
1599 case TargetLowering::Legal:
1602 case TargetLowering::Expand: {
1603 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1604 ISD::CondCode InvCC = ISD::SETCC_INVALID;
1607 default: llvm_unreachable("Don't know how to expand this condition!");
1609 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1610 == TargetLowering::Legal
1611 && "If SETO is expanded, SETOEQ must be legal!");
1612 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1614 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1615 == TargetLowering::Legal
1616 && "If SETUO is expanded, SETUNE must be legal!");
1617 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1630 // If we are floating point, assign and break, otherwise fall through.
1631 if (!OpVT.isInteger()) {
1632 // We can use the 4th bit to tell if we are the unordered
1633 // or ordered version of the opcode.
1634 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1635 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1636 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1639 // Fallthrough if we are unsigned integer.
1646 InvCC = ISD::getSetCCSwappedOperands(CCCode);
1647 if (TLI.getCondCodeAction(InvCC, OpVT) == TargetLowering::Expand) {
1648 // We only support using the inverted operation and not a
1649 // different manner of supporting expanding these cases.
1650 llvm_unreachable("Don't know how to expand this condition!");
1652 LHS = DAG.getSetCC(dl, VT, RHS, LHS, InvCC);
1658 SDValue SetCC1, SetCC2;
1659 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1660 // If we aren't the ordered or unorder operation,
1661 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1662 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1663 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1665 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1666 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1667 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1669 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1677 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1678 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1679 /// a load from the stack slot to DestVT, extending it if needed.
1680 /// The resultant code need not be legal.
1681 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1685 // Create the stack frame object.
1687 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1688 getTypeForEVT(*DAG.getContext()));
1689 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1691 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1692 int SPFI = StackPtrFI->getIndex();
1693 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1695 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1696 unsigned SlotSize = SlotVT.getSizeInBits();
1697 unsigned DestSize = DestVT.getSizeInBits();
1698 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1699 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1701 // Emit a store to the stack slot. Use a truncstore if the input value is
1702 // later than DestVT.
1705 if (SrcSize > SlotSize)
1706 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1707 PtrInfo, SlotVT, false, false, SrcAlign);
1709 assert(SrcSize == SlotSize && "Invalid store");
1710 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1711 PtrInfo, false, false, SrcAlign);
1714 // Result is a load from the stack slot.
1715 if (SlotSize == DestSize)
1716 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1717 false, false, false, DestAlign);
1719 assert(SlotSize < DestSize && "Unknown extension!");
1720 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1721 PtrInfo, SlotVT, false, false, DestAlign);
1724 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1725 DebugLoc dl = Node->getDebugLoc();
1726 // Create a vector sized/aligned stack slot, store the value to element #0,
1727 // then load the whole vector back out.
1728 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1730 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1731 int SPFI = StackPtrFI->getIndex();
1733 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1735 MachinePointerInfo::getFixedStack(SPFI),
1736 Node->getValueType(0).getVectorElementType(),
1738 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1739 MachinePointerInfo::getFixedStack(SPFI),
1740 false, false, false, 0);
1744 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1745 /// support the operation, but do support the resultant vector type.
1746 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1747 unsigned NumElems = Node->getNumOperands();
1748 SDValue Value1, Value2;
1749 DebugLoc dl = Node->getDebugLoc();
1750 EVT VT = Node->getValueType(0);
1751 EVT OpVT = Node->getOperand(0).getValueType();
1752 EVT EltVT = VT.getVectorElementType();
1754 // If the only non-undef value is the low element, turn this into a
1755 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1756 bool isOnlyLowElement = true;
1757 bool MoreThanTwoValues = false;
1758 bool isConstant = true;
1759 for (unsigned i = 0; i < NumElems; ++i) {
1760 SDValue V = Node->getOperand(i);
1761 if (V.getOpcode() == ISD::UNDEF)
1764 isOnlyLowElement = false;
1765 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1768 if (!Value1.getNode()) {
1770 } else if (!Value2.getNode()) {
1773 } else if (V != Value1 && V != Value2) {
1774 MoreThanTwoValues = true;
1778 if (!Value1.getNode())
1779 return DAG.getUNDEF(VT);
1781 if (isOnlyLowElement)
1782 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1784 // If all elements are constants, create a load from the constant pool.
1786 SmallVector<Constant*, 16> CV;
1787 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1788 if (ConstantFPSDNode *V =
1789 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1790 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1791 } else if (ConstantSDNode *V =
1792 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1794 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1796 // If OpVT and EltVT don't match, EltVT is not legal and the
1797 // element values have been promoted/truncated earlier. Undo this;
1798 // we don't want a v16i8 to become a v16i32 for example.
1799 const ConstantInt *CI = V->getConstantIntValue();
1800 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1801 CI->getZExtValue()));
1804 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1805 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1806 CV.push_back(UndefValue::get(OpNTy));
1809 Constant *CP = ConstantVector::get(CV);
1810 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1811 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1812 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1813 MachinePointerInfo::getConstantPool(),
1814 false, false, false, Alignment);
1817 if (!MoreThanTwoValues) {
1818 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1819 for (unsigned i = 0; i < NumElems; ++i) {
1820 SDValue V = Node->getOperand(i);
1821 if (V.getOpcode() == ISD::UNDEF)
1823 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1825 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1826 // Get the splatted value into the low element of a vector register.
1827 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1829 if (Value2.getNode())
1830 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1832 Vec2 = DAG.getUNDEF(VT);
1834 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1835 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1839 // Otherwise, we can't handle this case efficiently.
1840 return ExpandVectorBuildThroughStack(Node);
1843 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
1844 // does not fit into a register, return the lo part and set the hi part to the
1845 // by-reg argument. If it does fit into a single register, return the result
1846 // and leave the Hi part unset.
1847 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
1849 TargetLowering::ArgListTy Args;
1850 TargetLowering::ArgListEntry Entry;
1851 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1852 EVT ArgVT = Node->getOperand(i).getValueType();
1853 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1854 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
1855 Entry.isSExt = isSigned;
1856 Entry.isZExt = !isSigned;
1857 Args.push_back(Entry);
1859 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1860 TLI.getPointerTy());
1862 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1864 // By default, the input chain to this libcall is the entry node of the
1865 // function. If the libcall is going to be emitted as a tail call then
1866 // TLI.isUsedByReturnOnly will change it to the right chain if the return
1867 // node which is being folded has a non-entry input chain.
1868 SDValue InChain = DAG.getEntryNode();
1870 // isTailCall may be true since the callee does not reference caller stack
1871 // frame. Check if it's in the right position.
1872 SDValue TCChain = InChain;
1873 bool isTailCall = isInTailCallPosition(DAG, Node, TCChain, TLI);
1878 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1879 0, TLI.getLibcallCallingConv(LC), isTailCall,
1880 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1881 Callee, Args, DAG, Node->getDebugLoc());
1882 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1885 if (!CallInfo.second.getNode())
1886 // It's a tailcall, return the chain (which is the DAG root).
1887 return DAG.getRoot();
1889 return CallInfo.first;
1892 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
1893 /// and returning a result of type RetVT.
1894 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
1895 const SDValue *Ops, unsigned NumOps,
1896 bool isSigned, DebugLoc dl) {
1897 TargetLowering::ArgListTy Args;
1898 Args.reserve(NumOps);
1900 TargetLowering::ArgListEntry Entry;
1901 for (unsigned i = 0; i != NumOps; ++i) {
1902 Entry.Node = Ops[i];
1903 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
1904 Entry.isSExt = isSigned;
1905 Entry.isZExt = !isSigned;
1906 Args.push_back(Entry);
1908 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1909 TLI.getPointerTy());
1911 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
1913 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
1914 false, 0, TLI.getLibcallCallingConv(LC),
1915 /*isTailCall=*/false,
1916 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1917 Callee, Args, DAG, dl);
1918 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
1920 return CallInfo.first;
1923 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
1924 // ExpandLibCall except that the first operand is the in-chain.
1925 std::pair<SDValue, SDValue>
1926 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
1929 SDValue InChain = Node->getOperand(0);
1931 TargetLowering::ArgListTy Args;
1932 TargetLowering::ArgListEntry Entry;
1933 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
1934 EVT ArgVT = Node->getOperand(i).getValueType();
1935 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1936 Entry.Node = Node->getOperand(i);
1938 Entry.isSExt = isSigned;
1939 Entry.isZExt = !isSigned;
1940 Args.push_back(Entry);
1942 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1943 TLI.getPointerTy());
1945 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1947 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1948 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
1949 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1950 Callee, Args, DAG, Node->getDebugLoc());
1951 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1956 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
1957 RTLIB::Libcall Call_F32,
1958 RTLIB::Libcall Call_F64,
1959 RTLIB::Libcall Call_F80,
1960 RTLIB::Libcall Call_PPCF128) {
1962 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1963 default: llvm_unreachable("Unexpected request for libcall!");
1964 case MVT::f32: LC = Call_F32; break;
1965 case MVT::f64: LC = Call_F64; break;
1966 case MVT::f80: LC = Call_F80; break;
1967 case MVT::ppcf128: LC = Call_PPCF128; break;
1969 return ExpandLibCall(LC, Node, false);
1972 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
1973 RTLIB::Libcall Call_I8,
1974 RTLIB::Libcall Call_I16,
1975 RTLIB::Libcall Call_I32,
1976 RTLIB::Libcall Call_I64,
1977 RTLIB::Libcall Call_I128) {
1979 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1980 default: llvm_unreachable("Unexpected request for libcall!");
1981 case MVT::i8: LC = Call_I8; break;
1982 case MVT::i16: LC = Call_I16; break;
1983 case MVT::i32: LC = Call_I32; break;
1984 case MVT::i64: LC = Call_I64; break;
1985 case MVT::i128: LC = Call_I128; break;
1987 return ExpandLibCall(LC, Node, isSigned);
1990 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
1991 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
1992 const TargetLowering &TLI) {
1994 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1995 default: llvm_unreachable("Unexpected request for libcall!");
1996 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
1997 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
1998 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
1999 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2000 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2003 return TLI.getLibcallName(LC) != 0;
2006 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2008 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2009 // The other use might have been replaced with a divrem already.
2010 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2011 unsigned OtherOpcode = 0;
2013 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2015 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2017 SDValue Op0 = Node->getOperand(0);
2018 SDValue Op1 = Node->getOperand(1);
2019 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2020 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2024 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2025 User->getOperand(0) == Op0 &&
2026 User->getOperand(1) == Op1)
2032 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2035 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2036 SmallVectorImpl<SDValue> &Results) {
2037 unsigned Opcode = Node->getOpcode();
2038 bool isSigned = Opcode == ISD::SDIVREM;
2041 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2042 default: llvm_unreachable("Unexpected request for libcall!");
2043 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2044 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2045 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2046 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2047 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2050 // The input chain to this libcall is the entry node of the function.
2051 // Legalizing the call will automatically add the previous call to the
2053 SDValue InChain = DAG.getEntryNode();
2055 EVT RetVT = Node->getValueType(0);
2056 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2058 TargetLowering::ArgListTy Args;
2059 TargetLowering::ArgListEntry Entry;
2060 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2061 EVT ArgVT = Node->getOperand(i).getValueType();
2062 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2063 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2064 Entry.isSExt = isSigned;
2065 Entry.isZExt = !isSigned;
2066 Args.push_back(Entry);
2069 // Also pass the return address of the remainder.
2070 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2072 Entry.Ty = RetTy->getPointerTo();
2073 Entry.isSExt = isSigned;
2074 Entry.isZExt = !isSigned;
2075 Args.push_back(Entry);
2077 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2078 TLI.getPointerTy());
2080 DebugLoc dl = Node->getDebugLoc();
2082 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
2083 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2084 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2085 Callee, Args, DAG, dl);
2086 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2088 // Remainder is loaded back from the stack frame.
2089 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2090 MachinePointerInfo(), false, false, false, 0);
2091 Results.push_back(CallInfo.first);
2092 Results.push_back(Rem);
2095 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2096 /// INT_TO_FP operation of the specified operand when the target requests that
2097 /// we expand it. At this point, we know that the result and operand types are
2098 /// legal for the target.
2099 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2103 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2104 // simple 32-bit [signed|unsigned] integer to float/double expansion
2106 // Get the stack frame index of a 8 byte buffer.
2107 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2109 // word offset constant for Hi/Lo address computation
2110 SDValue WordOff = DAG.getConstant(sizeof(int), TLI.getPointerTy());
2111 // set up Hi and Lo (into buffer) address based on endian
2112 SDValue Hi = StackSlot;
2113 SDValue Lo = DAG.getNode(ISD::ADD, dl,
2114 TLI.getPointerTy(), StackSlot, WordOff);
2115 if (TLI.isLittleEndian())
2118 // if signed map to unsigned space
2121 // constant used to invert sign bit (signed to unsigned mapping)
2122 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2123 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2127 // store the lo of the constructed double - based on integer input
2128 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2129 Op0Mapped, Lo, MachinePointerInfo(),
2131 // initial hi portion of constructed double
2132 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2133 // store the hi of the constructed double - biased exponent
2134 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2135 MachinePointerInfo(),
2137 // load the constructed double
2138 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2139 MachinePointerInfo(), false, false, false, 0);
2140 // FP constant to bias correct the final result
2141 SDValue Bias = DAG.getConstantFP(isSigned ?
2142 BitsToDouble(0x4330000080000000ULL) :
2143 BitsToDouble(0x4330000000000000ULL),
2145 // subtract the bias
2146 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2149 // handle final rounding
2150 if (DestVT == MVT::f64) {
2153 } else if (DestVT.bitsLT(MVT::f64)) {
2154 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2155 DAG.getIntPtrConstant(0));
2156 } else if (DestVT.bitsGT(MVT::f64)) {
2157 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2161 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2162 // Code below here assumes !isSigned without checking again.
2164 // Implementation of unsigned i64 to f64 following the algorithm in
2165 // __floatundidf in compiler_rt. This implementation has the advantage
2166 // of performing rounding correctly, both in the default rounding mode
2167 // and in all alternate rounding modes.
2168 // TODO: Generalize this for use with other types.
2169 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2171 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2172 SDValue TwoP84PlusTwoP52 =
2173 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2175 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2177 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2178 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2179 DAG.getConstant(32, MVT::i64));
2180 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2181 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2182 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2183 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2184 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2186 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2189 // Implementation of unsigned i64 to f32.
2190 // TODO: Generalize this for use with other types.
2191 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2192 // For unsigned conversions, convert them to signed conversions using the
2193 // algorithm from the x86_64 __floatundidf in compiler_rt.
2195 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2197 SDValue ShiftConst =
2198 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2199 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2200 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2201 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2202 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2204 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2205 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2207 // TODO: This really should be implemented using a branch rather than a
2208 // select. We happen to get lucky and machinesink does the right
2209 // thing most of the time. This would be a good candidate for a
2210 //pseudo-op, or, even better, for whole-function isel.
2211 SDValue SignBitTest = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2212 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2213 return DAG.getNode(ISD::SELECT, dl, MVT::f32, SignBitTest, Slow, Fast);
2216 // Otherwise, implement the fully general conversion.
2218 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2219 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2220 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2221 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2222 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2223 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2224 SDValue Ne = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2225 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2226 SDValue Sel = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ne, Or, Op0);
2227 SDValue Ge = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2228 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2230 SDValue Sel2 = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ge, Sel, Op0);
2231 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2233 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2234 DAG.getConstant(32, SHVT));
2235 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2236 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2238 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2239 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2240 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2241 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2242 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2243 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2244 DAG.getIntPtrConstant(0));
2247 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2249 SDValue SignSet = DAG.getSetCC(dl, TLI.getSetCCResultType(Op0.getValueType()),
2250 Op0, DAG.getConstant(0, Op0.getValueType()),
2252 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2253 SDValue CstOffset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(),
2254 SignSet, Four, Zero);
2256 // If the sign bit of the integer is set, the large number will be treated
2257 // as a negative number. To counteract this, the dynamic code adds an
2258 // offset depending on the data type.
2260 switch (Op0.getValueType().getSimpleVT().SimpleTy) {
2261 default: llvm_unreachable("Unsupported integer type!");
2262 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2263 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2264 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2265 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2267 if (TLI.isLittleEndian()) FF <<= 32;
2268 Constant *FudgeFactor = ConstantInt::get(
2269 Type::getInt64Ty(*DAG.getContext()), FF);
2271 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2272 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2273 CPIdx = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), CPIdx, CstOffset);
2274 Alignment = std::min(Alignment, 4u);
2276 if (DestVT == MVT::f32)
2277 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2278 MachinePointerInfo::getConstantPool(),
2279 false, false, false, Alignment);
2281 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2282 DAG.getEntryNode(), CPIdx,
2283 MachinePointerInfo::getConstantPool(),
2284 MVT::f32, false, false, Alignment);
2285 HandleSDNode Handle(Load);
2286 LegalizeOp(Load.getNode());
2287 FudgeInReg = Handle.getValue();
2290 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2293 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2294 /// *INT_TO_FP operation of the specified operand when the target requests that
2295 /// we promote it. At this point, we know that the result and operand types are
2296 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2297 /// operation that takes a larger input.
2298 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2302 // First step, figure out the appropriate *INT_TO_FP operation to use.
2303 EVT NewInTy = LegalOp.getValueType();
2305 unsigned OpToUse = 0;
2307 // Scan for the appropriate larger type to use.
2309 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2310 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2312 // If the target supports SINT_TO_FP of this type, use it.
2313 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2314 OpToUse = ISD::SINT_TO_FP;
2317 if (isSigned) continue;
2319 // If the target supports UINT_TO_FP of this type, use it.
2320 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2321 OpToUse = ISD::UINT_TO_FP;
2325 // Otherwise, try a larger type.
2328 // Okay, we found the operation and type to use. Zero extend our input to the
2329 // desired type then run the operation on it.
2330 return DAG.getNode(OpToUse, dl, DestVT,
2331 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2332 dl, NewInTy, LegalOp));
2335 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2336 /// FP_TO_*INT operation of the specified operand when the target requests that
2337 /// we promote it. At this point, we know that the result and operand types are
2338 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2339 /// operation that returns a larger result.
2340 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2344 // First step, figure out the appropriate FP_TO*INT operation to use.
2345 EVT NewOutTy = DestVT;
2347 unsigned OpToUse = 0;
2349 // Scan for the appropriate larger type to use.
2351 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2352 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2354 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2355 OpToUse = ISD::FP_TO_SINT;
2359 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2360 OpToUse = ISD::FP_TO_UINT;
2364 // Otherwise, try a larger type.
2368 // Okay, we found the operation and type to use.
2369 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2371 // Truncate the result of the extended FP_TO_*INT operation to the desired
2373 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2376 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2378 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, DebugLoc dl) {
2379 EVT VT = Op.getValueType();
2380 EVT SHVT = TLI.getShiftAmountTy(VT);
2381 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2382 switch (VT.getSimpleVT().SimpleTy) {
2383 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2385 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2386 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2387 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2389 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2390 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2391 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2392 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2393 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2394 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2395 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2396 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2397 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2399 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2400 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2401 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2402 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2403 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2404 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2405 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2406 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2407 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2408 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2409 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2410 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2411 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2412 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2413 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2414 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2415 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2416 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2417 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2418 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2419 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2423 /// SplatByte - Distribute ByteVal over NumBits bits.
2424 // FIXME: Move this helper to a common place.
2425 static APInt SplatByte(unsigned NumBits, uint8_t ByteVal) {
2426 APInt Val = APInt(NumBits, ByteVal);
2428 for (unsigned i = NumBits; i > 8; i >>= 1) {
2429 Val = (Val << Shift) | Val;
2435 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2437 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2440 default: llvm_unreachable("Cannot expand this yet!");
2442 EVT VT = Op.getValueType();
2443 EVT ShVT = TLI.getShiftAmountTy(VT);
2444 unsigned Len = VT.getSizeInBits();
2446 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2447 "CTPOP not implemented for this type.");
2449 // This is the "best" algorithm from
2450 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2452 SDValue Mask55 = DAG.getConstant(SplatByte(Len, 0x55), VT);
2453 SDValue Mask33 = DAG.getConstant(SplatByte(Len, 0x33), VT);
2454 SDValue Mask0F = DAG.getConstant(SplatByte(Len, 0x0F), VT);
2455 SDValue Mask01 = DAG.getConstant(SplatByte(Len, 0x01), VT);
2457 // v = v - ((v >> 1) & 0x55555555...)
2458 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2459 DAG.getNode(ISD::AND, dl, VT,
2460 DAG.getNode(ISD::SRL, dl, VT, Op,
2461 DAG.getConstant(1, ShVT)),
2463 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2464 Op = DAG.getNode(ISD::ADD, dl, VT,
2465 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2466 DAG.getNode(ISD::AND, dl, VT,
2467 DAG.getNode(ISD::SRL, dl, VT, Op,
2468 DAG.getConstant(2, ShVT)),
2470 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2471 Op = DAG.getNode(ISD::AND, dl, VT,
2472 DAG.getNode(ISD::ADD, dl, VT, Op,
2473 DAG.getNode(ISD::SRL, dl, VT, Op,
2474 DAG.getConstant(4, ShVT))),
2476 // v = (v * 0x01010101...) >> (Len - 8)
2477 Op = DAG.getNode(ISD::SRL, dl, VT,
2478 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2479 DAG.getConstant(Len - 8, ShVT));
2483 case ISD::CTLZ_ZERO_UNDEF:
2484 // This trivially expands to CTLZ.
2485 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2487 // for now, we do this:
2488 // x = x | (x >> 1);
2489 // x = x | (x >> 2);
2491 // x = x | (x >>16);
2492 // x = x | (x >>32); // for 64-bit input
2493 // return popcount(~x);
2495 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2496 EVT VT = Op.getValueType();
2497 EVT ShVT = TLI.getShiftAmountTy(VT);
2498 unsigned len = VT.getSizeInBits();
2499 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2500 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2501 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2502 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2504 Op = DAG.getNOT(dl, Op, VT);
2505 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2507 case ISD::CTTZ_ZERO_UNDEF:
2508 // This trivially expands to CTTZ.
2509 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2511 // for now, we use: { return popcount(~x & (x - 1)); }
2512 // unless the target has ctlz but not ctpop, in which case we use:
2513 // { return 32 - nlz(~x & (x-1)); }
2514 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2515 EVT VT = Op.getValueType();
2516 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2517 DAG.getNOT(dl, Op, VT),
2518 DAG.getNode(ISD::SUB, dl, VT, Op,
2519 DAG.getConstant(1, VT)));
2520 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2521 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2522 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2523 return DAG.getNode(ISD::SUB, dl, VT,
2524 DAG.getConstant(VT.getSizeInBits(), VT),
2525 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2526 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2531 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2532 unsigned Opc = Node->getOpcode();
2533 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2538 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2539 case ISD::ATOMIC_SWAP:
2540 switch (VT.SimpleTy) {
2541 default: llvm_unreachable("Unexpected value type for atomic!");
2542 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2543 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2544 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2545 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2548 case ISD::ATOMIC_CMP_SWAP:
2549 switch (VT.SimpleTy) {
2550 default: llvm_unreachable("Unexpected value type for atomic!");
2551 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2552 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2553 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2554 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2557 case ISD::ATOMIC_LOAD_ADD:
2558 switch (VT.SimpleTy) {
2559 default: llvm_unreachable("Unexpected value type for atomic!");
2560 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2561 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2562 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2563 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2566 case ISD::ATOMIC_LOAD_SUB:
2567 switch (VT.SimpleTy) {
2568 default: llvm_unreachable("Unexpected value type for atomic!");
2569 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2570 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2571 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2572 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2575 case ISD::ATOMIC_LOAD_AND:
2576 switch (VT.SimpleTy) {
2577 default: llvm_unreachable("Unexpected value type for atomic!");
2578 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2579 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2580 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2581 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2584 case ISD::ATOMIC_LOAD_OR:
2585 switch (VT.SimpleTy) {
2586 default: llvm_unreachable("Unexpected value type for atomic!");
2587 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2588 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2589 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2590 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2593 case ISD::ATOMIC_LOAD_XOR:
2594 switch (VT.SimpleTy) {
2595 default: llvm_unreachable("Unexpected value type for atomic!");
2596 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2597 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2598 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2599 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2602 case ISD::ATOMIC_LOAD_NAND:
2603 switch (VT.SimpleTy) {
2604 default: llvm_unreachable("Unexpected value type for atomic!");
2605 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2606 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2607 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2608 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2613 return ExpandChainLibCall(LC, Node, false);
2616 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2617 SmallVector<SDValue, 8> Results;
2618 DebugLoc dl = Node->getDebugLoc();
2619 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2620 switch (Node->getOpcode()) {
2623 case ISD::CTLZ_ZERO_UNDEF:
2625 case ISD::CTTZ_ZERO_UNDEF:
2626 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2627 Results.push_back(Tmp1);
2630 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2632 case ISD::FRAMEADDR:
2633 case ISD::RETURNADDR:
2634 case ISD::FRAME_TO_ARGS_OFFSET:
2635 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2637 case ISD::FLT_ROUNDS_:
2638 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2640 case ISD::EH_RETURN:
2644 case ISD::EH_SJLJ_LONGJMP:
2645 // If the target didn't expand these, there's nothing to do, so just
2646 // preserve the chain and be done.
2647 Results.push_back(Node->getOperand(0));
2649 case ISD::EH_SJLJ_SETJMP:
2650 // If the target didn't expand this, just return 'zero' and preserve the
2652 Results.push_back(DAG.getConstant(0, MVT::i32));
2653 Results.push_back(Node->getOperand(0));
2655 case ISD::ATOMIC_FENCE:
2656 case ISD::MEMBARRIER: {
2657 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2658 // FIXME: handle "fence singlethread" more efficiently.
2659 TargetLowering::ArgListTy Args;
2661 CallLoweringInfo CLI(Node->getOperand(0),
2662 Type::getVoidTy(*DAG.getContext()),
2663 false, false, false, false, 0, CallingConv::C,
2664 /*isTailCall=*/false,
2665 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2666 DAG.getExternalSymbol("__sync_synchronize",
2667 TLI.getPointerTy()),
2669 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2671 Results.push_back(CallResult.second);
2674 case ISD::ATOMIC_LOAD: {
2675 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2676 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
2677 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
2678 cast<AtomicSDNode>(Node)->getMemoryVT(),
2679 Node->getOperand(0),
2680 Node->getOperand(1), Zero, Zero,
2681 cast<AtomicSDNode>(Node)->getMemOperand(),
2682 cast<AtomicSDNode>(Node)->getOrdering(),
2683 cast<AtomicSDNode>(Node)->getSynchScope());
2684 Results.push_back(Swap.getValue(0));
2685 Results.push_back(Swap.getValue(1));
2688 case ISD::ATOMIC_STORE: {
2689 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2690 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2691 cast<AtomicSDNode>(Node)->getMemoryVT(),
2692 Node->getOperand(0),
2693 Node->getOperand(1), Node->getOperand(2),
2694 cast<AtomicSDNode>(Node)->getMemOperand(),
2695 cast<AtomicSDNode>(Node)->getOrdering(),
2696 cast<AtomicSDNode>(Node)->getSynchScope());
2697 Results.push_back(Swap.getValue(1));
2700 // By default, atomic intrinsics are marked Legal and lowered. Targets
2701 // which don't support them directly, however, may want libcalls, in which
2702 // case they mark them Expand, and we get here.
2703 case ISD::ATOMIC_SWAP:
2704 case ISD::ATOMIC_LOAD_ADD:
2705 case ISD::ATOMIC_LOAD_SUB:
2706 case ISD::ATOMIC_LOAD_AND:
2707 case ISD::ATOMIC_LOAD_OR:
2708 case ISD::ATOMIC_LOAD_XOR:
2709 case ISD::ATOMIC_LOAD_NAND:
2710 case ISD::ATOMIC_LOAD_MIN:
2711 case ISD::ATOMIC_LOAD_MAX:
2712 case ISD::ATOMIC_LOAD_UMIN:
2713 case ISD::ATOMIC_LOAD_UMAX:
2714 case ISD::ATOMIC_CMP_SWAP: {
2715 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
2716 Results.push_back(Tmp.first);
2717 Results.push_back(Tmp.second);
2720 case ISD::DYNAMIC_STACKALLOC:
2721 ExpandDYNAMIC_STACKALLOC(Node, Results);
2723 case ISD::MERGE_VALUES:
2724 for (unsigned i = 0; i < Node->getNumValues(); i++)
2725 Results.push_back(Node->getOperand(i));
2728 EVT VT = Node->getValueType(0);
2730 Results.push_back(DAG.getConstant(0, VT));
2732 assert(VT.isFloatingPoint() && "Unknown value type!");
2733 Results.push_back(DAG.getConstantFP(0, VT));
2738 // If this operation is not supported, lower it to 'abort()' call
2739 TargetLowering::ArgListTy Args;
2741 CallLoweringInfo CLI(Node->getOperand(0),
2742 Type::getVoidTy(*DAG.getContext()),
2743 false, false, false, false, 0, CallingConv::C,
2744 /*isTailCall=*/false,
2745 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2746 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
2748 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2750 Results.push_back(CallResult.second);
2755 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
2756 Node->getValueType(0), dl);
2757 Results.push_back(Tmp1);
2759 case ISD::FP_EXTEND:
2760 Tmp1 = EmitStackConvert(Node->getOperand(0),
2761 Node->getOperand(0).getValueType(),
2762 Node->getValueType(0), dl);
2763 Results.push_back(Tmp1);
2765 case ISD::SIGN_EXTEND_INREG: {
2766 // NOTE: we could fall back on load/store here too for targets without
2767 // SAR. However, it is doubtful that any exist.
2768 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2769 EVT VT = Node->getValueType(0);
2770 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
2773 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
2774 ExtraVT.getScalarType().getSizeInBits();
2775 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
2776 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
2777 Node->getOperand(0), ShiftCst);
2778 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
2779 Results.push_back(Tmp1);
2782 case ISD::FP_ROUND_INREG: {
2783 // The only way we can lower this is to turn it into a TRUNCSTORE,
2784 // EXTLOAD pair, targeting a temporary location (a stack slot).
2786 // NOTE: there is a choice here between constantly creating new stack
2787 // slots and always reusing the same one. We currently always create
2788 // new ones, as reuse may inhibit scheduling.
2789 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2790 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
2791 Node->getValueType(0), dl);
2792 Results.push_back(Tmp1);
2795 case ISD::SINT_TO_FP:
2796 case ISD::UINT_TO_FP:
2797 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
2798 Node->getOperand(0), Node->getValueType(0), dl);
2799 Results.push_back(Tmp1);
2801 case ISD::FP_TO_UINT: {
2802 SDValue True, False;
2803 EVT VT = Node->getOperand(0).getValueType();
2804 EVT NVT = Node->getValueType(0);
2805 APFloat apf(APInt::getNullValue(VT.getSizeInBits()));
2806 APInt x = APInt::getSignBit(NVT.getSizeInBits());
2807 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
2808 Tmp1 = DAG.getConstantFP(apf, VT);
2809 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(VT),
2810 Node->getOperand(0),
2812 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
2813 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
2814 DAG.getNode(ISD::FSUB, dl, VT,
2815 Node->getOperand(0), Tmp1));
2816 False = DAG.getNode(ISD::XOR, dl, NVT, False,
2817 DAG.getConstant(x, NVT));
2818 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2, True, False);
2819 Results.push_back(Tmp1);
2823 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2824 EVT VT = Node->getValueType(0);
2825 Tmp1 = Node->getOperand(0);
2826 Tmp2 = Node->getOperand(1);
2827 unsigned Align = Node->getConstantOperandVal(3);
2829 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
2830 MachinePointerInfo(V),
2831 false, false, false, 0);
2832 SDValue VAList = VAListLoad;
2834 if (Align > TLI.getMinStackArgumentAlignment()) {
2835 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
2837 VAList = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2838 DAG.getConstant(Align - 1,
2839 TLI.getPointerTy()));
2841 VAList = DAG.getNode(ISD::AND, dl, TLI.getPointerTy(), VAList,
2842 DAG.getConstant(-(int64_t)Align,
2843 TLI.getPointerTy()));
2846 // Increment the pointer, VAList, to the next vaarg
2847 Tmp3 = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2848 DAG.getConstant(TLI.getDataLayout()->
2849 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
2850 TLI.getPointerTy()));
2851 // Store the incremented VAList to the legalized pointer
2852 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
2853 MachinePointerInfo(V), false, false, 0);
2854 // Load the actual argument out of the pointer VAList
2855 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
2856 false, false, false, 0));
2857 Results.push_back(Results[0].getValue(1));
2861 // This defaults to loading a pointer from the input and storing it to the
2862 // output, returning the chain.
2863 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
2864 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
2865 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
2866 Node->getOperand(2), MachinePointerInfo(VS),
2867 false, false, false, 0);
2868 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
2869 MachinePointerInfo(VD), false, false, 0);
2870 Results.push_back(Tmp1);
2873 case ISD::EXTRACT_VECTOR_ELT:
2874 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
2875 // This must be an access of the only element. Return it.
2876 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
2877 Node->getOperand(0));
2879 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
2880 Results.push_back(Tmp1);
2882 case ISD::EXTRACT_SUBVECTOR:
2883 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
2885 case ISD::INSERT_SUBVECTOR:
2886 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
2888 case ISD::CONCAT_VECTORS: {
2889 Results.push_back(ExpandVectorBuildThroughStack(Node));
2892 case ISD::SCALAR_TO_VECTOR:
2893 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
2895 case ISD::INSERT_VECTOR_ELT:
2896 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
2897 Node->getOperand(1),
2898 Node->getOperand(2), dl));
2900 case ISD::VECTOR_SHUFFLE: {
2901 SmallVector<int, 32> NewMask;
2902 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
2904 EVT VT = Node->getValueType(0);
2905 EVT EltVT = VT.getVectorElementType();
2906 SDValue Op0 = Node->getOperand(0);
2907 SDValue Op1 = Node->getOperand(1);
2908 if (!TLI.isTypeLegal(EltVT)) {
2910 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
2912 // BUILD_VECTOR operands are allowed to be wider than the element type.
2913 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept it
2914 if (NewEltVT.bitsLT(EltVT)) {
2916 // Convert shuffle node.
2917 // If original node was v4i64 and the new EltVT is i32,
2918 // cast operands to v8i32 and re-build the mask.
2920 // Calculate new VT, the size of the new VT should be equal to original.
2921 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), NewEltVT,
2922 VT.getSizeInBits()/NewEltVT.getSizeInBits());
2923 assert(NewVT.bitsEq(VT));
2925 // cast operands to new VT
2926 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
2927 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
2929 // Convert the shuffle mask
2930 unsigned int factor = NewVT.getVectorNumElements()/VT.getVectorNumElements();
2932 // EltVT gets smaller
2935 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
2937 for (unsigned fi = 0; fi < factor; ++fi)
2938 NewMask.push_back(Mask[i]);
2941 for (unsigned fi = 0; fi < factor; ++fi)
2942 NewMask.push_back(Mask[i]*factor+fi);
2950 unsigned NumElems = VT.getVectorNumElements();
2951 SmallVector<SDValue, 16> Ops;
2952 for (unsigned i = 0; i != NumElems; ++i) {
2954 Ops.push_back(DAG.getUNDEF(EltVT));
2957 unsigned Idx = Mask[i];
2959 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
2961 DAG.getIntPtrConstant(Idx)));
2963 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
2965 DAG.getIntPtrConstant(Idx - NumElems)));
2968 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
2969 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
2970 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
2971 Results.push_back(Tmp1);
2974 case ISD::EXTRACT_ELEMENT: {
2975 EVT OpTy = Node->getOperand(0).getValueType();
2976 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
2978 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
2979 DAG.getConstant(OpTy.getSizeInBits()/2,
2980 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
2981 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
2984 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
2985 Node->getOperand(0));
2987 Results.push_back(Tmp1);
2990 case ISD::STACKSAVE:
2991 // Expand to CopyFromReg if the target set
2992 // StackPointerRegisterToSaveRestore.
2993 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
2994 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
2995 Node->getValueType(0)));
2996 Results.push_back(Results[0].getValue(1));
2998 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
2999 Results.push_back(Node->getOperand(0));
3002 case ISD::STACKRESTORE:
3003 // Expand to CopyToReg if the target set
3004 // StackPointerRegisterToSaveRestore.
3005 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3006 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3007 Node->getOperand(1)));
3009 Results.push_back(Node->getOperand(0));
3012 case ISD::FCOPYSIGN:
3013 Results.push_back(ExpandFCOPYSIGN(Node));
3016 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3017 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3018 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3019 Node->getOperand(0));
3020 Results.push_back(Tmp1);
3023 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3024 EVT VT = Node->getValueType(0);
3025 Tmp1 = Node->getOperand(0);
3026 Tmp2 = DAG.getConstantFP(0.0, VT);
3027 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(Tmp1.getValueType()),
3028 Tmp1, Tmp2, ISD::SETUGT);
3029 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3030 Tmp1 = DAG.getNode(ISD::SELECT, dl, VT, Tmp2, Tmp1, Tmp3);
3031 Results.push_back(Tmp1);
3035 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3036 RTLIB::SQRT_F80, RTLIB::SQRT_PPCF128));
3039 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3040 RTLIB::SIN_F80, RTLIB::SIN_PPCF128));
3043 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3044 RTLIB::COS_F80, RTLIB::COS_PPCF128));
3047 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3048 RTLIB::LOG_F80, RTLIB::LOG_PPCF128));
3051 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3052 RTLIB::LOG2_F80, RTLIB::LOG2_PPCF128));
3055 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3056 RTLIB::LOG10_F80, RTLIB::LOG10_PPCF128));
3059 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3060 RTLIB::EXP_F80, RTLIB::EXP_PPCF128));
3063 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3064 RTLIB::EXP2_F80, RTLIB::EXP2_PPCF128));
3067 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3068 RTLIB::TRUNC_F80, RTLIB::TRUNC_PPCF128));
3071 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3072 RTLIB::FLOOR_F80, RTLIB::FLOOR_PPCF128));
3075 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3076 RTLIB::CEIL_F80, RTLIB::CEIL_PPCF128));
3079 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3080 RTLIB::RINT_F80, RTLIB::RINT_PPCF128));
3082 case ISD::FNEARBYINT:
3083 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3084 RTLIB::NEARBYINT_F64,
3085 RTLIB::NEARBYINT_F80,
3086 RTLIB::NEARBYINT_PPCF128));
3089 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3090 RTLIB::POWI_F80, RTLIB::POWI_PPCF128));
3093 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3094 RTLIB::POW_F80, RTLIB::POW_PPCF128));
3097 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3098 RTLIB::DIV_F80, RTLIB::DIV_PPCF128));
3101 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3102 RTLIB::REM_F80, RTLIB::REM_PPCF128));
3105 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3106 RTLIB::FMA_F80, RTLIB::FMA_PPCF128));
3108 case ISD::FP16_TO_FP32:
3109 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3111 case ISD::FP32_TO_FP16:
3112 Results.push_back(ExpandLibCall(RTLIB::FPROUND_F32_F16, Node, false));
3114 case ISD::ConstantFP: {
3115 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3116 // Check to see if this FP immediate is already legal.
3117 // If this is a legal constant, turn it into a TargetConstantFP node.
3118 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3119 Results.push_back(ExpandConstantFP(CFP, true));
3122 case ISD::EHSELECTION: {
3123 unsigned Reg = TLI.getExceptionSelectorRegister();
3124 assert(Reg && "Can't expand to unknown register!");
3125 Results.push_back(DAG.getCopyFromReg(Node->getOperand(1), dl, Reg,
3126 Node->getValueType(0)));
3127 Results.push_back(Results[0].getValue(1));
3130 case ISD::EXCEPTIONADDR: {
3131 unsigned Reg = TLI.getExceptionPointerRegister();
3132 assert(Reg && "Can't expand to unknown register!");
3133 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, Reg,
3134 Node->getValueType(0)));
3135 Results.push_back(Results[0].getValue(1));
3139 EVT VT = Node->getValueType(0);
3140 assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3141 TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
3142 "Don't know how to expand this FP subtraction!");
3143 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3144 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3145 Results.push_back(Tmp1);
3149 EVT VT = Node->getValueType(0);
3150 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3151 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3152 "Don't know how to expand this subtraction!");
3153 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3154 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3155 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3156 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3161 EVT VT = Node->getValueType(0);
3162 SDVTList VTs = DAG.getVTList(VT, VT);
3163 bool isSigned = Node->getOpcode() == ISD::SREM;
3164 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3165 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3166 Tmp2 = Node->getOperand(0);
3167 Tmp3 = Node->getOperand(1);
3168 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3169 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3170 // If div is legal, it's better to do the normal expansion
3171 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3172 useDivRem(Node, isSigned, false))) {
3173 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3174 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3176 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3177 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3178 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3179 } else if (isSigned)
3180 Tmp1 = ExpandIntLibCall(Node, true,
3182 RTLIB::SREM_I16, RTLIB::SREM_I32,
3183 RTLIB::SREM_I64, RTLIB::SREM_I128);
3185 Tmp1 = ExpandIntLibCall(Node, false,
3187 RTLIB::UREM_I16, RTLIB::UREM_I32,
3188 RTLIB::UREM_I64, RTLIB::UREM_I128);
3189 Results.push_back(Tmp1);
3194 bool isSigned = Node->getOpcode() == ISD::SDIV;
3195 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3196 EVT VT = Node->getValueType(0);
3197 SDVTList VTs = DAG.getVTList(VT, VT);
3198 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3199 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3200 useDivRem(Node, isSigned, true)))
3201 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3202 Node->getOperand(1));
3204 Tmp1 = ExpandIntLibCall(Node, true,
3206 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3207 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3209 Tmp1 = ExpandIntLibCall(Node, false,
3211 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3212 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3213 Results.push_back(Tmp1);
3218 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3220 EVT VT = Node->getValueType(0);
3221 SDVTList VTs = DAG.getVTList(VT, VT);
3222 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3223 "If this wasn't legal, it shouldn't have been created!");
3224 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3225 Node->getOperand(1));
3226 Results.push_back(Tmp1.getValue(1));
3231 // Expand into divrem libcall
3232 ExpandDivRemLibCall(Node, Results);
3235 EVT VT = Node->getValueType(0);
3236 SDVTList VTs = DAG.getVTList(VT, VT);
3237 // See if multiply or divide can be lowered using two-result operations.
3238 // We just need the low half of the multiply; try both the signed
3239 // and unsigned forms. If the target supports both SMUL_LOHI and
3240 // UMUL_LOHI, form a preference by checking which forms of plain
3241 // MULH it supports.
3242 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3243 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3244 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3245 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3246 unsigned OpToUse = 0;
3247 if (HasSMUL_LOHI && !HasMULHS) {
3248 OpToUse = ISD::SMUL_LOHI;
3249 } else if (HasUMUL_LOHI && !HasMULHU) {
3250 OpToUse = ISD::UMUL_LOHI;
3251 } else if (HasSMUL_LOHI) {
3252 OpToUse = ISD::SMUL_LOHI;
3253 } else if (HasUMUL_LOHI) {
3254 OpToUse = ISD::UMUL_LOHI;
3257 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3258 Node->getOperand(1)));
3261 Tmp1 = ExpandIntLibCall(Node, false,
3263 RTLIB::MUL_I16, RTLIB::MUL_I32,
3264 RTLIB::MUL_I64, RTLIB::MUL_I128);
3265 Results.push_back(Tmp1);
3270 SDValue LHS = Node->getOperand(0);
3271 SDValue RHS = Node->getOperand(1);
3272 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3273 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3275 Results.push_back(Sum);
3276 EVT OType = Node->getValueType(1);
3278 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3280 // LHSSign -> LHS >= 0
3281 // RHSSign -> RHS >= 0
3282 // SumSign -> Sum >= 0
3285 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3287 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3289 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3290 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3291 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3292 Node->getOpcode() == ISD::SADDO ?
3293 ISD::SETEQ : ISD::SETNE);
3295 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3296 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3298 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3299 Results.push_back(Cmp);
3304 SDValue LHS = Node->getOperand(0);
3305 SDValue RHS = Node->getOperand(1);
3306 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3307 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3309 Results.push_back(Sum);
3310 Results.push_back(DAG.getSetCC(dl, Node->getValueType(1), Sum, LHS,
3311 Node->getOpcode () == ISD::UADDO ?
3312 ISD::SETULT : ISD::SETUGT));
3317 EVT VT = Node->getValueType(0);
3318 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3319 SDValue LHS = Node->getOperand(0);
3320 SDValue RHS = Node->getOperand(1);
3323 static const unsigned Ops[2][3] =
3324 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3325 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3326 bool isSigned = Node->getOpcode() == ISD::SMULO;
3327 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3328 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3329 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3330 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3331 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3333 TopHalf = BottomHalf.getValue(1);
3334 } else if (TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
3335 VT.getSizeInBits() * 2))) {
3336 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3337 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3338 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3339 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3340 DAG.getIntPtrConstant(0));
3341 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3342 DAG.getIntPtrConstant(1));
3344 // We can fall back to a libcall with an illegal type for the MUL if we
3345 // have a libcall big enough.
3346 // Also, we can fall back to a division in some cases, but that's a big
3347 // performance hit in the general case.
3348 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3349 if (WideVT == MVT::i16)
3350 LC = RTLIB::MUL_I16;
3351 else if (WideVT == MVT::i32)
3352 LC = RTLIB::MUL_I32;
3353 else if (WideVT == MVT::i64)
3354 LC = RTLIB::MUL_I64;
3355 else if (WideVT == MVT::i128)
3356 LC = RTLIB::MUL_I128;
3357 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3359 // The high part is obtained by SRA'ing all but one of the bits of low
3361 unsigned LoSize = VT.getSizeInBits();
3362 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3363 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3364 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3365 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3367 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3368 // pre-lowered to the correct types. This all depends upon WideVT not
3369 // being a legal type for the architecture and thus has to be split to
3371 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3372 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3373 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3374 DAG.getIntPtrConstant(0));
3375 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3376 DAG.getIntPtrConstant(1));
3377 // Ret is a node with an illegal type. Because such things are not
3378 // generally permitted during this phase of legalization, delete the
3379 // node. The above EXTRACT_ELEMENT nodes should have been folded.
3380 DAG.DeleteNode(Ret.getNode());
3384 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3385 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3386 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3387 TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf, Tmp1,
3390 TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf,
3391 DAG.getConstant(0, VT), ISD::SETNE);
3393 Results.push_back(BottomHalf);
3394 Results.push_back(TopHalf);
3397 case ISD::BUILD_PAIR: {
3398 EVT PairTy = Node->getValueType(0);
3399 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3400 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3401 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3402 DAG.getConstant(PairTy.getSizeInBits()/2,
3403 TLI.getShiftAmountTy(PairTy)));
3404 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3408 Tmp1 = Node->getOperand(0);
3409 Tmp2 = Node->getOperand(1);
3410 Tmp3 = Node->getOperand(2);
3411 if (Tmp1.getOpcode() == ISD::SETCC) {
3412 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3414 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3416 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3417 DAG.getConstant(0, Tmp1.getValueType()),
3418 Tmp2, Tmp3, ISD::SETNE);
3420 Results.push_back(Tmp1);
3423 SDValue Chain = Node->getOperand(0);
3424 SDValue Table = Node->getOperand(1);
3425 SDValue Index = Node->getOperand(2);
3427 EVT PTy = TLI.getPointerTy();
3429 const DataLayout &TD = *TLI.getDataLayout();
3430 unsigned EntrySize =
3431 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3433 Index = DAG.getNode(ISD::MUL, dl, PTy,
3434 Index, DAG.getConstant(EntrySize, PTy));
3435 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3437 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3438 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3439 MachinePointerInfo::getJumpTable(), MemVT,
3442 if (TM.getRelocationModel() == Reloc::PIC_) {
3443 // For PIC, the sequence is:
3444 // BRIND(load(Jumptable + index) + RelocBase)
3445 // RelocBase can be JumpTable, GOT or some sort of global base.
3446 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3447 TLI.getPICJumpTableRelocBase(Table, DAG));
3449 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3450 Results.push_back(Tmp1);
3454 // Expand brcond's setcc into its constituent parts and create a BR_CC
3456 Tmp1 = Node->getOperand(0);
3457 Tmp2 = Node->getOperand(1);
3458 if (Tmp2.getOpcode() == ISD::SETCC) {
3459 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3460 Tmp1, Tmp2.getOperand(2),
3461 Tmp2.getOperand(0), Tmp2.getOperand(1),
3462 Node->getOperand(2));
3464 // We test only the i1 bit. Skip the AND if UNDEF.
3465 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3466 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3467 DAG.getConstant(1, Tmp2.getValueType()));
3468 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3469 DAG.getCondCode(ISD::SETNE), Tmp3,
3470 DAG.getConstant(0, Tmp3.getValueType()),
3471 Node->getOperand(2));
3473 Results.push_back(Tmp1);
3476 Tmp1 = Node->getOperand(0);
3477 Tmp2 = Node->getOperand(1);
3478 Tmp3 = Node->getOperand(2);
3479 LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2, Tmp3, dl);
3481 // If we expanded the SETCC into an AND/OR, return the new node
3482 if (Tmp2.getNode() == 0) {
3483 Results.push_back(Tmp1);
3487 // Otherwise, SETCC for the given comparison type must be completely
3488 // illegal; expand it into a SELECT_CC.
3489 EVT VT = Node->getValueType(0);
3490 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3491 DAG.getConstant(1, VT), DAG.getConstant(0, VT), Tmp3);
3492 Results.push_back(Tmp1);
3495 case ISD::SELECT_CC: {
3496 Tmp1 = Node->getOperand(0); // LHS
3497 Tmp2 = Node->getOperand(1); // RHS
3498 Tmp3 = Node->getOperand(2); // True
3499 Tmp4 = Node->getOperand(3); // False
3500 SDValue CC = Node->getOperand(4);
3502 LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp1.getValueType()),
3503 Tmp1, Tmp2, CC, dl);
3505 assert(!Tmp2.getNode() && "Can't legalize SELECT_CC with legal condition!");
3506 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
3507 CC = DAG.getCondCode(ISD::SETNE);
3508 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1, Tmp2,
3510 Results.push_back(Tmp1);
3514 Tmp1 = Node->getOperand(0); // Chain
3515 Tmp2 = Node->getOperand(2); // LHS
3516 Tmp3 = Node->getOperand(3); // RHS
3517 Tmp4 = Node->getOperand(1); // CC
3519 LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp2.getValueType()),
3520 Tmp2, Tmp3, Tmp4, dl);
3522 assert(!Tmp3.getNode() && "Can't legalize BR_CC with legal condition!");
3523 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
3524 Tmp4 = DAG.getCondCode(ISD::SETNE);
3525 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4, Tmp2,
3526 Tmp3, Node->getOperand(4));
3527 Results.push_back(Tmp1);
3530 case ISD::BUILD_VECTOR:
3531 Results.push_back(ExpandBUILD_VECTOR(Node));
3536 // Scalarize vector SRA/SRL/SHL.
3537 EVT VT = Node->getValueType(0);
3538 assert(VT.isVector() && "Unable to legalize non-vector shift");
3539 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
3540 unsigned NumElem = VT.getVectorNumElements();
3542 SmallVector<SDValue, 8> Scalars;
3543 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
3544 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3546 Node->getOperand(0), DAG.getIntPtrConstant(Idx));
3547 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3549 Node->getOperand(1), DAG.getIntPtrConstant(Idx));
3550 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
3551 VT.getScalarType(), Ex, Sh));
3554 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
3555 &Scalars[0], Scalars.size());
3556 ReplaceNode(SDValue(Node, 0), Result);
3559 case ISD::GLOBAL_OFFSET_TABLE:
3560 case ISD::GlobalAddress:
3561 case ISD::GlobalTLSAddress:
3562 case ISD::ExternalSymbol:
3563 case ISD::ConstantPool:
3564 case ISD::JumpTable:
3565 case ISD::INTRINSIC_W_CHAIN:
3566 case ISD::INTRINSIC_WO_CHAIN:
3567 case ISD::INTRINSIC_VOID:
3568 // FIXME: Custom lowering for these operations shouldn't return null!
3572 // Replace the original node with the legalized result.
3573 if (!Results.empty())
3574 ReplaceNode(Node, Results.data());
3577 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
3578 SmallVector<SDValue, 8> Results;
3579 MVT OVT = Node->getSimpleValueType(0);
3580 if (Node->getOpcode() == ISD::UINT_TO_FP ||
3581 Node->getOpcode() == ISD::SINT_TO_FP ||
3582 Node->getOpcode() == ISD::SETCC) {
3583 OVT = Node->getOperand(0).getSimpleValueType();
3585 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
3586 DebugLoc dl = Node->getDebugLoc();
3587 SDValue Tmp1, Tmp2, Tmp3;
3588 switch (Node->getOpcode()) {
3590 case ISD::CTTZ_ZERO_UNDEF:
3592 case ISD::CTLZ_ZERO_UNDEF:
3594 // Zero extend the argument.
3595 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3596 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
3597 // already the correct result.
3598 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3599 if (Node->getOpcode() == ISD::CTTZ) {
3600 // FIXME: This should set a bit in the zero extended value instead.
3601 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(NVT),
3602 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
3604 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2,
3605 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
3606 } else if (Node->getOpcode() == ISD::CTLZ ||
3607 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
3608 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
3609 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
3610 DAG.getConstant(NVT.getSizeInBits() -
3611 OVT.getSizeInBits(), NVT));
3613 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
3616 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
3617 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3618 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
3619 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
3620 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
3621 Results.push_back(Tmp1);
3624 case ISD::FP_TO_UINT:
3625 case ISD::FP_TO_SINT:
3626 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
3627 Node->getOpcode() == ISD::FP_TO_SINT, dl);
3628 Results.push_back(Tmp1);
3630 case ISD::UINT_TO_FP:
3631 case ISD::SINT_TO_FP:
3632 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
3633 Node->getOpcode() == ISD::SINT_TO_FP, dl);
3634 Results.push_back(Tmp1);
3637 SDValue Chain = Node->getOperand(0); // Get the chain.
3638 SDValue Ptr = Node->getOperand(1); // Get the pointer.
3641 if (OVT.isVector()) {
3642 TruncOp = ISD::BITCAST;
3644 assert(OVT.isInteger()
3645 && "VAARG promotion is supported only for vectors or integer types");
3646 TruncOp = ISD::TRUNCATE;
3649 // Perform the larger operation, then convert back
3650 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
3651 Node->getConstantOperandVal(3));
3652 Chain = Tmp1.getValue(1);
3654 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
3656 // Modified the chain result - switch anything that used the old chain to
3658 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
3659 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
3666 unsigned ExtOp, TruncOp;
3667 if (OVT.isVector()) {
3668 ExtOp = ISD::BITCAST;
3669 TruncOp = ISD::BITCAST;
3671 assert(OVT.isInteger() && "Cannot promote logic operation");
3672 ExtOp = ISD::ANY_EXTEND;
3673 TruncOp = ISD::TRUNCATE;
3675 // Promote each of the values to the new type.
3676 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3677 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3678 // Perform the larger operation, then convert back
3679 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3680 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
3684 unsigned ExtOp, TruncOp;
3685 if (Node->getValueType(0).isVector()) {
3686 ExtOp = ISD::BITCAST;
3687 TruncOp = ISD::BITCAST;
3688 } else if (Node->getValueType(0).isInteger()) {
3689 ExtOp = ISD::ANY_EXTEND;
3690 TruncOp = ISD::TRUNCATE;
3692 ExtOp = ISD::FP_EXTEND;
3693 TruncOp = ISD::FP_ROUND;
3695 Tmp1 = Node->getOperand(0);
3696 // Promote each of the values to the new type.
3697 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3698 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
3699 // Perform the larger operation, then round down.
3700 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp1, Tmp2, Tmp3);
3701 if (TruncOp != ISD::FP_ROUND)
3702 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
3704 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
3705 DAG.getIntPtrConstant(0));
3706 Results.push_back(Tmp1);
3709 case ISD::VECTOR_SHUFFLE: {
3710 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3712 // Cast the two input vectors.
3713 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
3714 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
3716 // Convert the shuffle mask to the right # elements.
3717 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
3718 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
3719 Results.push_back(Tmp1);
3723 unsigned ExtOp = ISD::FP_EXTEND;
3724 if (NVT.isInteger()) {
3725 ISD::CondCode CCCode =
3726 cast<CondCodeSDNode>(Node->getOperand(2))->get();
3727 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3729 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3730 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3731 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3732 Tmp1, Tmp2, Node->getOperand(2)));
3738 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
3739 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
3740 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3741 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
3742 Tmp3, DAG.getIntPtrConstant(0)));
3749 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
3750 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3751 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
3752 Tmp2, DAG.getIntPtrConstant(0)));
3757 // Replace the original node with the legalized result.
3758 if (!Results.empty())
3759 ReplaceNode(Node, Results.data());
3762 // SelectionDAG::Legalize - This is the entry point for the file.
3764 void SelectionDAG::Legalize() {
3765 /// run - This is the main entry point to this class.
3767 SelectionDAGLegalize(*this).LegalizeDAG();