1 //===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SelectionDAG::Legalize method.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineJumpTableInfo.h"
22 #include "llvm/IR/CallingConv.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DebugInfo.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Target/TargetFrameLowering.h"
34 #include "llvm/Target/TargetLowering.h"
35 #include "llvm/Target/TargetMachine.h"
38 //===----------------------------------------------------------------------===//
39 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
40 /// hacks on it until the target machine can handle it. This involves
41 /// eliminating value sizes the machine cannot handle (promoting small sizes to
42 /// large sizes or splitting up large values into small values) as well as
43 /// eliminating operations the machine cannot handle.
45 /// This code also does a small amount of optimization and recognition of idioms
46 /// as part of its processing. For example, if a target does not support a
47 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
48 /// will attempt merge setcc and brc instructions into brcc's.
51 class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
52 const TargetMachine &TM;
53 const TargetLowering &TLI;
56 /// LegalizePosition - The iterator for walking through the node list.
57 SelectionDAG::allnodes_iterator LegalizePosition;
59 /// LegalizedNodes - The set of nodes which have already been legalized.
60 SmallPtrSet<SDNode *, 16> LegalizedNodes;
62 EVT getSetCCResultType(EVT VT) const {
63 return TLI.getSetCCResultType(*DAG.getContext(), VT);
66 // Libcall insertion helpers.
69 explicit SelectionDAGLegalize(SelectionDAG &DAG);
74 /// LegalizeOp - Legalizes the given operation.
75 void LegalizeOp(SDNode *Node);
77 SDValue OptimizeFloatStore(StoreSDNode *ST);
79 void LegalizeLoadOps(SDNode *Node);
80 void LegalizeStoreOps(SDNode *Node);
82 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
83 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
84 /// is necessary to spill the vector being inserted into to memory, perform
85 /// the insert there, and then read the result back.
86 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
87 SDValue Idx, SDLoc dl);
88 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
89 SDValue Idx, SDLoc dl);
91 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
92 /// performs the same shuffe in terms of order or result bytes, but on a type
93 /// whose vector element type is narrower than the original shuffle type.
94 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
95 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
96 SDValue N1, SDValue N2,
97 ArrayRef<int> Mask) const;
99 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
100 bool &NeedInvert, SDLoc dl);
102 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
103 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
104 unsigned NumOps, bool isSigned, SDLoc dl);
106 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
107 SDNode *Node, bool isSigned);
108 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
109 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
110 RTLIB::Libcall Call_F128,
111 RTLIB::Libcall Call_PPCF128);
112 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
113 RTLIB::Libcall Call_I8,
114 RTLIB::Libcall Call_I16,
115 RTLIB::Libcall Call_I32,
116 RTLIB::Libcall Call_I64,
117 RTLIB::Libcall Call_I128);
118 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
119 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
121 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
122 SDValue ExpandBUILD_VECTOR(SDNode *Node);
123 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
124 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
125 SmallVectorImpl<SDValue> &Results);
126 SDValue ExpandFCOPYSIGN(SDNode *Node);
127 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
129 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
131 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
134 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
135 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
137 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
138 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
139 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
141 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
143 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
145 void ExpandNode(SDNode *Node);
146 void PromoteNode(SDNode *Node);
148 void ForgetNode(SDNode *N) {
149 LegalizedNodes.erase(N);
150 if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
155 // DAGUpdateListener implementation.
156 void NodeDeleted(SDNode *N, SDNode *E) override {
159 void NodeUpdated(SDNode *N) override {}
161 // Node replacement helpers
162 void ReplacedNode(SDNode *N) {
163 if (N->use_empty()) {
164 DAG.RemoveDeadNode(N);
169 void ReplaceNode(SDNode *Old, SDNode *New) {
170 DAG.ReplaceAllUsesWith(Old, New);
173 void ReplaceNode(SDValue Old, SDValue New) {
174 DAG.ReplaceAllUsesWith(Old, New);
175 ReplacedNode(Old.getNode());
177 void ReplaceNode(SDNode *Old, const SDValue *New) {
178 DAG.ReplaceAllUsesWith(Old, New);
184 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
185 /// performs the same shuffe in terms of order or result bytes, but on a type
186 /// whose vector element type is narrower than the original shuffle type.
187 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
189 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
190 SDValue N1, SDValue N2,
191 ArrayRef<int> Mask) const {
192 unsigned NumMaskElts = VT.getVectorNumElements();
193 unsigned NumDestElts = NVT.getVectorNumElements();
194 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
196 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
198 if (NumEltsGrowth == 1)
199 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
201 SmallVector<int, 8> NewMask;
202 for (unsigned i = 0; i != NumMaskElts; ++i) {
204 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
206 NewMask.push_back(-1);
208 NewMask.push_back(Idx * NumEltsGrowth + j);
211 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
212 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
213 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
216 SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
217 : SelectionDAG::DAGUpdateListener(dag),
218 TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
222 void SelectionDAGLegalize::LegalizeDAG() {
223 DAG.AssignTopologicalOrder();
225 // Visit all the nodes. We start in topological order, so that we see
226 // nodes with their original operands intact. Legalization can produce
227 // new nodes which may themselves need to be legalized. Iterate until all
228 // nodes have been legalized.
230 bool AnyLegalized = false;
231 for (LegalizePosition = DAG.allnodes_end();
232 LegalizePosition != DAG.allnodes_begin(); ) {
235 SDNode *N = LegalizePosition;
236 if (LegalizedNodes.insert(N)) {
246 // Remove dead nodes now.
247 DAG.RemoveDeadNodes();
250 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
251 /// a load from the constant pool.
253 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
257 // If a FP immediate is precise when represented as a float and if the
258 // target can do an extending load from float to double, we put it into
259 // the constant pool as a float, even if it's is statically typed as a
260 // double. This shrinks FP constants and canonicalizes them for targets where
261 // an FP extending load is the same cost as a normal load (such as on the x87
262 // fp stack or PPC FP unit).
263 EVT VT = CFP->getValueType(0);
264 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
266 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
267 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
268 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
273 while (SVT != MVT::f32) {
274 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
275 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
276 // Only do this if the target has a native EXTLOAD instruction from
278 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
279 TLI.ShouldShrinkFPConstant(OrigVT)) {
280 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
281 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
287 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
288 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
291 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
293 CPIdx, MachinePointerInfo::getConstantPool(),
294 VT, false, false, Alignment);
298 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
299 MachinePointerInfo::getConstantPool(), false, false, false,
304 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
305 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
306 const TargetLowering &TLI,
307 SelectionDAGLegalize *DAGLegalize) {
308 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
309 "unaligned indexed stores not implemented!");
310 SDValue Chain = ST->getChain();
311 SDValue Ptr = ST->getBasePtr();
312 SDValue Val = ST->getValue();
313 EVT VT = Val.getValueType();
314 int Alignment = ST->getAlignment();
315 unsigned AS = ST->getAddressSpace();
318 if (ST->getMemoryVT().isFloatingPoint() ||
319 ST->getMemoryVT().isVector()) {
320 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
321 if (TLI.isTypeLegal(intVT)) {
322 // Expand to a bitconvert of the value to the integer type of the
323 // same size, then a (misaligned) int store.
324 // FIXME: Does not handle truncating floating point stores!
325 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
326 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
327 ST->isVolatile(), ST->isNonTemporal(), Alignment);
328 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
331 // Do a (aligned) store to a stack slot, then copy from the stack slot
332 // to the final destination using (unaligned) integer loads and stores.
333 EVT StoredVT = ST->getMemoryVT();
335 TLI.getRegisterType(*DAG.getContext(),
336 EVT::getIntegerVT(*DAG.getContext(),
337 StoredVT.getSizeInBits()));
338 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
339 unsigned RegBytes = RegVT.getSizeInBits() / 8;
340 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
342 // Make sure the stack slot is also aligned for the register type.
343 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
345 // Perform the original store, only redirected to the stack slot.
346 SDValue Store = DAG.getTruncStore(Chain, dl,
347 Val, StackPtr, MachinePointerInfo(),
348 StoredVT, false, false, 0);
349 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy(AS));
350 SmallVector<SDValue, 8> Stores;
353 // Do all but one copies using the full register width.
354 for (unsigned i = 1; i < NumRegs; i++) {
355 // Load one integer register's worth from the stack slot.
356 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
357 MachinePointerInfo(),
358 false, false, false, 0);
359 // Store it to the final location. Remember the store.
360 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
361 ST->getPointerInfo().getWithOffset(Offset),
362 ST->isVolatile(), ST->isNonTemporal(),
363 MinAlign(ST->getAlignment(), Offset)));
364 // Increment the pointers.
366 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
368 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
371 // The last store may be partial. Do a truncating store. On big-endian
372 // machines this requires an extending load from the stack slot to ensure
373 // that the bits are in the right place.
374 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
375 8 * (StoredBytes - Offset));
377 // Load from the stack slot.
378 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
379 MachinePointerInfo(),
380 MemVT, false, false, 0);
382 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
384 .getWithOffset(Offset),
385 MemVT, ST->isVolatile(),
387 MinAlign(ST->getAlignment(), Offset),
389 // The order of the stores doesn't matter - say it with a TokenFactor.
391 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
393 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
396 assert(ST->getMemoryVT().isInteger() &&
397 !ST->getMemoryVT().isVector() &&
398 "Unaligned store of unknown type.");
399 // Get the half-size VT
400 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
401 int NumBits = NewStoredVT.getSizeInBits();
402 int IncrementSize = NumBits / 8;
404 // Divide the stored value in two parts.
405 SDValue ShiftAmount = DAG.getConstant(NumBits,
406 TLI.getShiftAmountTy(Val.getValueType()));
408 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
410 // Store the two parts
411 SDValue Store1, Store2;
412 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
413 ST->getPointerInfo(), NewStoredVT,
414 ST->isVolatile(), ST->isNonTemporal(), Alignment);
416 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
417 DAG.getConstant(IncrementSize, TLI.getPointerTy(AS)));
418 Alignment = MinAlign(Alignment, IncrementSize);
419 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
420 ST->getPointerInfo().getWithOffset(IncrementSize),
421 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
422 Alignment, ST->getTBAAInfo());
425 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
426 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
429 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
431 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
432 const TargetLowering &TLI,
433 SDValue &ValResult, SDValue &ChainResult) {
434 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
435 "unaligned indexed loads not implemented!");
436 SDValue Chain = LD->getChain();
437 SDValue Ptr = LD->getBasePtr();
438 EVT VT = LD->getValueType(0);
439 EVT LoadedVT = LD->getMemoryVT();
441 if (VT.isFloatingPoint() || VT.isVector()) {
442 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
443 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
444 // Expand to a (misaligned) integer load of the same size,
445 // then bitconvert to floating point or vector.
446 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
447 LD->getMemOperand());
448 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
450 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
451 ISD::ANY_EXTEND, dl, VT, Result);
458 // Copy the value to a (aligned) stack slot using (unaligned) integer
459 // loads and stores, then do a (aligned) load from the stack slot.
460 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
461 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
462 unsigned RegBytes = RegVT.getSizeInBits() / 8;
463 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
465 // Make sure the stack slot is also aligned for the register type.
466 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
468 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
469 SmallVector<SDValue, 8> Stores;
470 SDValue StackPtr = StackBase;
473 // Do all but one copies using the full register width.
474 for (unsigned i = 1; i < NumRegs; i++) {
475 // Load one integer register's worth from the original location.
476 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
477 LD->getPointerInfo().getWithOffset(Offset),
478 LD->isVolatile(), LD->isNonTemporal(),
480 MinAlign(LD->getAlignment(), Offset),
482 // Follow the load with a store to the stack slot. Remember the store.
483 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
484 MachinePointerInfo(), false, false, 0));
485 // Increment the pointers.
487 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
488 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
492 // The last copy may be partial. Do an extending load.
493 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
494 8 * (LoadedBytes - Offset));
495 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
496 LD->getPointerInfo().getWithOffset(Offset),
497 MemVT, LD->isVolatile(),
499 MinAlign(LD->getAlignment(), Offset),
501 // Follow the load with a store to the stack slot. Remember the store.
502 // On big-endian machines this requires a truncating store to ensure
503 // that the bits end up in the right place.
504 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
505 MachinePointerInfo(), MemVT,
508 // The order of the stores doesn't matter - say it with a TokenFactor.
509 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
512 // Finally, perform the original load only redirected to the stack slot.
513 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
514 MachinePointerInfo(), LoadedVT, false, false, 0);
516 // Callers expect a MERGE_VALUES node.
521 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
522 "Unaligned load of unsupported type.");
524 // Compute the new VT that is half the size of the old one. This is an
526 unsigned NumBits = LoadedVT.getSizeInBits();
528 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
531 unsigned Alignment = LD->getAlignment();
532 unsigned IncrementSize = NumBits / 8;
533 ISD::LoadExtType HiExtType = LD->getExtensionType();
535 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
536 if (HiExtType == ISD::NON_EXTLOAD)
537 HiExtType = ISD::ZEXTLOAD;
539 // Load the value in two parts
541 if (TLI.isLittleEndian()) {
542 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
543 NewLoadedVT, LD->isVolatile(),
544 LD->isNonTemporal(), Alignment, LD->getTBAAInfo());
545 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
546 DAG.getConstant(IncrementSize, Ptr.getValueType()));
547 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
548 LD->getPointerInfo().getWithOffset(IncrementSize),
549 NewLoadedVT, LD->isVolatile(),
550 LD->isNonTemporal(), MinAlign(Alignment, IncrementSize),
553 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
554 NewLoadedVT, LD->isVolatile(),
555 LD->isNonTemporal(), Alignment, LD->getTBAAInfo());
556 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
557 DAG.getConstant(IncrementSize, Ptr.getValueType()));
558 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
559 LD->getPointerInfo().getWithOffset(IncrementSize),
560 NewLoadedVT, LD->isVolatile(),
561 LD->isNonTemporal(), MinAlign(Alignment, IncrementSize),
565 // aggregate the two parts
566 SDValue ShiftAmount = DAG.getConstant(NumBits,
567 TLI.getShiftAmountTy(Hi.getValueType()));
568 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
569 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
571 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
578 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
579 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
580 /// is necessary to spill the vector being inserted into to memory, perform
581 /// the insert there, and then read the result back.
582 SDValue SelectionDAGLegalize::
583 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
589 // If the target doesn't support this, we have to spill the input vector
590 // to a temporary stack slot, update the element, then reload it. This is
591 // badness. We could also load the value into a vector register (either
592 // with a "move to register" or "extload into register" instruction, then
593 // permute it into place, if the idx is a constant and if the idx is
594 // supported by the target.
595 EVT VT = Tmp1.getValueType();
596 EVT EltVT = VT.getVectorElementType();
597 EVT IdxVT = Tmp3.getValueType();
598 EVT PtrVT = TLI.getPointerTy();
599 SDValue StackPtr = DAG.CreateStackTemporary(VT);
601 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
604 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
605 MachinePointerInfo::getFixedStack(SPFI),
608 // Truncate or zero extend offset to target pointer type.
609 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
610 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
611 // Add the offset to the index.
612 unsigned EltSize = EltVT.getSizeInBits()/8;
613 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
614 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
615 // Store the scalar value.
616 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
618 // Load the updated vector.
619 return DAG.getLoad(VT, dl, Ch, StackPtr,
620 MachinePointerInfo::getFixedStack(SPFI), false, false,
625 SDValue SelectionDAGLegalize::
626 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
627 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
628 // SCALAR_TO_VECTOR requires that the type of the value being inserted
629 // match the element type of the vector being created, except for
630 // integers in which case the inserted value can be over width.
631 EVT EltVT = Vec.getValueType().getVectorElementType();
632 if (Val.getValueType() == EltVT ||
633 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
634 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
635 Vec.getValueType(), Val);
637 unsigned NumElts = Vec.getValueType().getVectorNumElements();
638 // We generate a shuffle of InVec and ScVec, so the shuffle mask
639 // should be 0,1,2,3,4,5... with the appropriate element replaced with
641 SmallVector<int, 8> ShufOps;
642 for (unsigned i = 0; i != NumElts; ++i)
643 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
645 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
649 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
652 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
653 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
654 // FIXME: We shouldn't do this for TargetConstantFP's.
655 // FIXME: move this to the DAG Combiner! Note that we can't regress due
656 // to phase ordering between legalized code and the dag combiner. This
657 // probably means that we need to integrate dag combiner and legalizer
659 // We generally can't do this one for long doubles.
660 SDValue Chain = ST->getChain();
661 SDValue Ptr = ST->getBasePtr();
662 unsigned Alignment = ST->getAlignment();
663 bool isVolatile = ST->isVolatile();
664 bool isNonTemporal = ST->isNonTemporal();
665 const MDNode *TBAAInfo = ST->getTBAAInfo();
667 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
668 if (CFP->getValueType(0) == MVT::f32 &&
669 TLI.isTypeLegal(MVT::i32)) {
670 SDValue Con = DAG.getConstant(CFP->getValueAPF().
671 bitcastToAPInt().zextOrTrunc(32),
673 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
674 isVolatile, isNonTemporal, Alignment, TBAAInfo);
677 if (CFP->getValueType(0) == MVT::f64) {
678 // If this target supports 64-bit registers, do a single 64-bit store.
679 if (TLI.isTypeLegal(MVT::i64)) {
680 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
681 zextOrTrunc(64), MVT::i64);
682 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
683 isVolatile, isNonTemporal, Alignment, TBAAInfo);
686 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
687 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
688 // stores. If the target supports neither 32- nor 64-bits, this
689 // xform is certainly not worth it.
690 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
691 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
692 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
693 if (TLI.isBigEndian()) std::swap(Lo, Hi);
695 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
696 isNonTemporal, Alignment, TBAAInfo);
697 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
698 DAG.getConstant(4, Ptr.getValueType()));
699 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
700 ST->getPointerInfo().getWithOffset(4),
701 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
704 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
708 return SDValue(0, 0);
711 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
712 StoreSDNode *ST = cast<StoreSDNode>(Node);
713 SDValue Chain = ST->getChain();
714 SDValue Ptr = ST->getBasePtr();
717 unsigned Alignment = ST->getAlignment();
718 bool isVolatile = ST->isVolatile();
719 bool isNonTemporal = ST->isNonTemporal();
720 const MDNode *TBAAInfo = ST->getTBAAInfo();
722 if (!ST->isTruncatingStore()) {
723 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
724 ReplaceNode(ST, OptStore);
729 SDValue Value = ST->getValue();
730 MVT VT = Value.getSimpleValueType();
731 switch (TLI.getOperationAction(ISD::STORE, VT)) {
732 default: llvm_unreachable("This action is not supported yet!");
733 case TargetLowering::Legal: {
734 // If this is an unaligned store and the target doesn't support it,
736 unsigned AS = ST->getAddressSpace();
737 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT(), AS)) {
738 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
739 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
740 if (ST->getAlignment() < ABIAlignment)
741 ExpandUnalignedStore(cast<StoreSDNode>(Node),
746 case TargetLowering::Custom: {
747 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
749 ReplaceNode(SDValue(Node, 0), Res);
752 case TargetLowering::Promote: {
753 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
754 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
755 "Can only promote stores to same size type");
756 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
758 DAG.getStore(Chain, dl, Value, Ptr,
759 ST->getPointerInfo(), isVolatile,
760 isNonTemporal, Alignment, TBAAInfo);
761 ReplaceNode(SDValue(Node, 0), Result);
768 SDValue Value = ST->getValue();
770 EVT StVT = ST->getMemoryVT();
771 unsigned StWidth = StVT.getSizeInBits();
773 if (StWidth != StVT.getStoreSizeInBits()) {
774 // Promote to a byte-sized store with upper bits zero if not
775 // storing an integral number of bytes. For example, promote
776 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
777 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
778 StVT.getStoreSizeInBits());
779 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
781 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
782 NVT, isVolatile, isNonTemporal, Alignment,
784 ReplaceNode(SDValue(Node, 0), Result);
785 } else if (StWidth & (StWidth - 1)) {
786 // If not storing a power-of-2 number of bits, expand as two stores.
787 assert(!StVT.isVector() && "Unsupported truncstore!");
788 unsigned RoundWidth = 1 << Log2_32(StWidth);
789 assert(RoundWidth < StWidth);
790 unsigned ExtraWidth = StWidth - RoundWidth;
791 assert(ExtraWidth < RoundWidth);
792 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
793 "Store size not an integral number of bytes!");
794 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
795 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
797 unsigned IncrementSize;
799 if (TLI.isLittleEndian()) {
800 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
801 // Store the bottom RoundWidth bits.
802 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
804 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.getConstant(IncrementSize, Ptr.getValueType()));
811 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
812 DAG.getConstant(RoundWidth,
813 TLI.getShiftAmountTy(Value.getValueType())));
814 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
815 ST->getPointerInfo().getWithOffset(IncrementSize),
816 ExtraVT, isVolatile, isNonTemporal,
817 MinAlign(Alignment, IncrementSize), TBAAInfo);
819 // Big endian - avoid unaligned stores.
820 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
821 // Store the top RoundWidth bits.
822 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
823 DAG.getConstant(ExtraWidth,
824 TLI.getShiftAmountTy(Value.getValueType())));
825 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
826 RoundVT, isVolatile, isNonTemporal, Alignment,
829 // Store the remaining ExtraWidth bits.
830 IncrementSize = RoundWidth / 8;
831 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
832 DAG.getConstant(IncrementSize, Ptr.getValueType()));
833 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
834 ST->getPointerInfo().getWithOffset(IncrementSize),
835 ExtraVT, isVolatile, isNonTemporal,
836 MinAlign(Alignment, IncrementSize), TBAAInfo);
839 // The order of the stores doesn't matter.
840 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
841 ReplaceNode(SDValue(Node, 0), Result);
843 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
844 StVT.getSimpleVT())) {
845 default: llvm_unreachable("This action is not supported yet!");
846 case TargetLowering::Legal: {
847 unsigned AS = ST->getAddressSpace();
848 // If this is an unaligned store and the target doesn't support it,
850 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT(), AS)) {
851 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
852 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
853 if (ST->getAlignment() < ABIAlignment)
854 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
858 case TargetLowering::Custom: {
859 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
861 ReplaceNode(SDValue(Node, 0), Res);
864 case TargetLowering::Expand:
865 assert(!StVT.isVector() &&
866 "Vector Stores are handled in LegalizeVectorOps");
868 // TRUNCSTORE:i16 i32 -> STORE i16
869 assert(TLI.isTypeLegal(StVT) &&
870 "Do not know how to expand this store!");
871 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
873 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
874 isVolatile, isNonTemporal, Alignment, TBAAInfo);
875 ReplaceNode(SDValue(Node, 0), Result);
882 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
883 LoadSDNode *LD = cast<LoadSDNode>(Node);
884 SDValue Chain = LD->getChain(); // The chain.
885 SDValue Ptr = LD->getBasePtr(); // The base pointer.
886 SDValue Value; // The value returned by the load op.
889 ISD::LoadExtType ExtType = LD->getExtensionType();
890 if (ExtType == ISD::NON_EXTLOAD) {
891 MVT VT = Node->getSimpleValueType(0);
892 SDValue RVal = SDValue(Node, 0);
893 SDValue RChain = SDValue(Node, 1);
895 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
896 default: llvm_unreachable("This action is not supported yet!");
897 case TargetLowering::Legal: {
898 unsigned AS = LD->getAddressSpace();
899 // If this is an unaligned load and the target doesn't support it,
901 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT(), AS)) {
902 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
903 unsigned ABIAlignment =
904 TLI.getDataLayout()->getABITypeAlignment(Ty);
905 if (LD->getAlignment() < ABIAlignment){
906 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
911 case TargetLowering::Custom: {
912 SDValue Res = TLI.LowerOperation(RVal, DAG);
915 RChain = Res.getValue(1);
919 case TargetLowering::Promote: {
920 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
921 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
922 "Can only promote loads to same size type");
924 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
925 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
926 RChain = Res.getValue(1);
930 if (RChain.getNode() != Node) {
931 assert(RVal.getNode() != Node && "Load must be completely replaced");
932 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
933 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
939 EVT SrcVT = LD->getMemoryVT();
940 unsigned SrcWidth = SrcVT.getSizeInBits();
941 unsigned Alignment = LD->getAlignment();
942 bool isVolatile = LD->isVolatile();
943 bool isNonTemporal = LD->isNonTemporal();
944 const MDNode *TBAAInfo = LD->getTBAAInfo();
946 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
947 // Some targets pretend to have an i1 loading operation, and actually
948 // load an i8. This trick is correct for ZEXTLOAD because the top 7
949 // bits are guaranteed to be zero; it helps the optimizers understand
950 // that these bits are zero. It is also useful for EXTLOAD, since it
951 // tells the optimizers that those bits are undefined. It would be
952 // nice to have an effective generic way of getting these benefits...
953 // Until such a way is found, don't insist on promoting i1 here.
955 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
956 // Promote to a byte-sized load if not loading an integral number of
957 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
958 unsigned NewWidth = SrcVT.getStoreSizeInBits();
959 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
962 // The extra bits are guaranteed to be zero, since we stored them that
963 // way. A zext load from NVT thus automatically gives zext from SrcVT.
965 ISD::LoadExtType NewExtType =
966 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
969 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
970 Chain, Ptr, LD->getPointerInfo(),
971 NVT, isVolatile, isNonTemporal, Alignment, TBAAInfo);
973 Ch = Result.getValue(1); // The chain.
975 if (ExtType == ISD::SEXTLOAD)
976 // Having the top bits zero doesn't help when sign extending.
977 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
978 Result.getValueType(),
979 Result, DAG.getValueType(SrcVT));
980 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
981 // All the top bits are guaranteed to be zero - inform the optimizers.
982 Result = DAG.getNode(ISD::AssertZext, dl,
983 Result.getValueType(), Result,
984 DAG.getValueType(SrcVT));
988 } else if (SrcWidth & (SrcWidth - 1)) {
989 // If not loading a power-of-2 number of bits, expand as two loads.
990 assert(!SrcVT.isVector() && "Unsupported extload!");
991 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
992 assert(RoundWidth < SrcWidth);
993 unsigned ExtraWidth = SrcWidth - RoundWidth;
994 assert(ExtraWidth < RoundWidth);
995 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
996 "Load size not an integral number of bytes!");
997 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
998 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
1000 unsigned IncrementSize;
1002 if (TLI.isLittleEndian()) {
1003 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
1004 // Load the bottom RoundWidth bits.
1005 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1007 LD->getPointerInfo(), RoundVT, isVolatile,
1008 isNonTemporal, Alignment, TBAAInfo);
1010 // Load the remaining ExtraWidth bits.
1011 IncrementSize = RoundWidth / 8;
1012 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1013 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1014 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1015 LD->getPointerInfo().getWithOffset(IncrementSize),
1016 ExtraVT, isVolatile, isNonTemporal,
1017 MinAlign(Alignment, IncrementSize), TBAAInfo);
1019 // Build a factor node to remember that this load is independent of
1021 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1024 // Move the top bits to the right place.
1025 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1026 DAG.getConstant(RoundWidth,
1027 TLI.getShiftAmountTy(Hi.getValueType())));
1029 // Join the hi and lo parts.
1030 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1032 // Big endian - avoid unaligned loads.
1033 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1034 // Load the top RoundWidth bits.
1035 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1036 LD->getPointerInfo(), RoundVT, isVolatile,
1037 isNonTemporal, Alignment, TBAAInfo);
1039 // Load the remaining ExtraWidth bits.
1040 IncrementSize = RoundWidth / 8;
1041 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1042 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1043 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1044 dl, Node->getValueType(0), Chain, Ptr,
1045 LD->getPointerInfo().getWithOffset(IncrementSize),
1046 ExtraVT, isVolatile, isNonTemporal,
1047 MinAlign(Alignment, IncrementSize), TBAAInfo);
1049 // Build a factor node to remember that this load is independent of
1051 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1054 // Move the top bits to the right place.
1055 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1056 DAG.getConstant(ExtraWidth,
1057 TLI.getShiftAmountTy(Hi.getValueType())));
1059 // Join the hi and lo parts.
1060 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1065 bool isCustom = false;
1066 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1067 default: llvm_unreachable("This action is not supported yet!");
1068 case TargetLowering::Custom:
1071 case TargetLowering::Legal: {
1072 Value = SDValue(Node, 0);
1073 Chain = SDValue(Node, 1);
1076 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1077 if (Res.getNode()) {
1079 Chain = Res.getValue(1);
1082 // If this is an unaligned load and the target doesn't support
1084 EVT MemVT = LD->getMemoryVT();
1085 unsigned AS = LD->getAddressSpace();
1086 if (!TLI.allowsUnalignedMemoryAccesses(MemVT, AS)) {
1088 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1089 unsigned ABIAlignment =
1090 TLI.getDataLayout()->getABITypeAlignment(Ty);
1091 if (LD->getAlignment() < ABIAlignment){
1092 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1093 DAG, TLI, Value, Chain);
1099 case TargetLowering::Expand:
1100 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) &&
1101 TLI.isTypeLegal(SrcVT)) {
1102 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1103 LD->getMemOperand());
1107 ExtendOp = (SrcVT.isFloatingPoint() ?
1108 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1110 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1111 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1112 default: llvm_unreachable("Unexpected extend load type!");
1114 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1115 Chain = Load.getValue(1);
1119 assert(!SrcVT.isVector() &&
1120 "Vector Loads are handled in LegalizeVectorOps");
1122 // FIXME: This does not work for vectors on most targets. Sign-
1123 // and zero-extend operations are currently folded into extending
1124 // loads, whether they are legal or not, and then we end up here
1125 // without any support for legalizing them.
1126 assert(ExtType != ISD::EXTLOAD &&
1127 "EXTLOAD should always be supported!");
1128 // Turn the unsupported load into an EXTLOAD followed by an
1129 // explicit zero/sign extend inreg.
1130 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1131 Node->getValueType(0),
1133 LD->getMemOperand());
1135 if (ExtType == ISD::SEXTLOAD)
1136 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1137 Result.getValueType(),
1138 Result, DAG.getValueType(SrcVT));
1140 ValRes = DAG.getZeroExtendInReg(Result, dl,
1141 SrcVT.getScalarType());
1143 Chain = Result.getValue(1);
1148 // Since loads produce two values, make sure to remember that we legalized
1150 if (Chain.getNode() != Node) {
1151 assert(Value.getNode() != Node && "Load must be completely replaced");
1152 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1153 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1158 /// LegalizeOp - Return a legal replacement for the given operation, with
1159 /// all legal operands.
1160 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1161 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1164 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1165 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1166 TargetLowering::TypeLegal &&
1167 "Unexpected illegal type!");
1169 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1170 assert((TLI.getTypeAction(*DAG.getContext(),
1171 Node->getOperand(i).getValueType()) ==
1172 TargetLowering::TypeLegal ||
1173 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1174 "Unexpected illegal type!");
1176 // Figure out the correct action; the way to query this varies by opcode
1177 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1178 bool SimpleFinishLegalizing = true;
1179 switch (Node->getOpcode()) {
1180 case ISD::INTRINSIC_W_CHAIN:
1181 case ISD::INTRINSIC_WO_CHAIN:
1182 case ISD::INTRINSIC_VOID:
1183 case ISD::STACKSAVE:
1184 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1187 Action = TLI.getOperationAction(Node->getOpcode(),
1188 Node->getValueType(0));
1189 if (Action != TargetLowering::Promote)
1190 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1192 case ISD::SINT_TO_FP:
1193 case ISD::UINT_TO_FP:
1194 case ISD::EXTRACT_VECTOR_ELT:
1195 Action = TLI.getOperationAction(Node->getOpcode(),
1196 Node->getOperand(0).getValueType());
1198 case ISD::FP_ROUND_INREG:
1199 case ISD::SIGN_EXTEND_INREG: {
1200 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1201 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1204 case ISD::ATOMIC_STORE: {
1205 Action = TLI.getOperationAction(Node->getOpcode(),
1206 Node->getOperand(2).getValueType());
1209 case ISD::SELECT_CC:
1212 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1213 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1214 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1215 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1216 ISD::CondCode CCCode =
1217 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1218 Action = TLI.getCondCodeAction(CCCode, OpVT);
1219 if (Action == TargetLowering::Legal) {
1220 if (Node->getOpcode() == ISD::SELECT_CC)
1221 Action = TLI.getOperationAction(Node->getOpcode(),
1222 Node->getValueType(0));
1224 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1230 // FIXME: Model these properly. LOAD and STORE are complicated, and
1231 // STORE expects the unlegalized operand in some cases.
1232 SimpleFinishLegalizing = false;
1234 case ISD::CALLSEQ_START:
1235 case ISD::CALLSEQ_END:
1236 // FIXME: This shouldn't be necessary. These nodes have special properties
1237 // dealing with the recursive nature of legalization. Removing this
1238 // special case should be done as part of making LegalizeDAG non-recursive.
1239 SimpleFinishLegalizing = false;
1241 case ISD::EXTRACT_ELEMENT:
1242 case ISD::FLT_ROUNDS_:
1250 case ISD::MERGE_VALUES:
1251 case ISD::EH_RETURN:
1252 case ISD::FRAME_TO_ARGS_OFFSET:
1253 case ISD::EH_SJLJ_SETJMP:
1254 case ISD::EH_SJLJ_LONGJMP:
1255 // These operations lie about being legal: when they claim to be legal,
1256 // they should actually be expanded.
1257 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1258 if (Action == TargetLowering::Legal)
1259 Action = TargetLowering::Expand;
1261 case ISD::INIT_TRAMPOLINE:
1262 case ISD::ADJUST_TRAMPOLINE:
1263 case ISD::FRAMEADDR:
1264 case ISD::RETURNADDR:
1265 // These operations lie about being legal: when they claim to be legal,
1266 // they should actually be custom-lowered.
1267 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1268 if (Action == TargetLowering::Legal)
1269 Action = TargetLowering::Custom;
1271 case ISD::DEBUGTRAP:
1272 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1273 if (Action == TargetLowering::Expand) {
1274 // replace ISD::DEBUGTRAP with ISD::TRAP
1276 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1277 Node->getOperand(0));
1278 ReplaceNode(Node, NewVal.getNode());
1279 LegalizeOp(NewVal.getNode());
1285 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1286 Action = TargetLowering::Legal;
1288 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1293 if (SimpleFinishLegalizing) {
1294 SDNode *NewNode = Node;
1295 switch (Node->getOpcode()) {
1302 // Legalizing shifts/rotates requires adjusting the shift amount
1303 // to the appropriate width.
1304 if (!Node->getOperand(1).getValueType().isVector()) {
1306 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1307 Node->getOperand(1));
1308 HandleSDNode Handle(SAO);
1309 LegalizeOp(SAO.getNode());
1310 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1314 case ISD::SRL_PARTS:
1315 case ISD::SRA_PARTS:
1316 case ISD::SHL_PARTS:
1317 // Legalizing shifts/rotates requires adjusting the shift amount
1318 // to the appropriate width.
1319 if (!Node->getOperand(2).getValueType().isVector()) {
1321 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1322 Node->getOperand(2));
1323 HandleSDNode Handle(SAO);
1324 LegalizeOp(SAO.getNode());
1325 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1326 Node->getOperand(1),
1332 if (NewNode != Node) {
1333 DAG.ReplaceAllUsesWith(Node, NewNode);
1334 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1335 DAG.TransferDbgValues(SDValue(Node, i), SDValue(NewNode, i));
1340 case TargetLowering::Legal:
1342 case TargetLowering::Custom: {
1343 // FIXME: The handling for custom lowering with multiple results is
1345 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1346 if (Res.getNode()) {
1347 SmallVector<SDValue, 8> ResultVals;
1348 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
1350 ResultVals.push_back(Res);
1352 ResultVals.push_back(Res.getValue(i));
1354 if (Res.getNode() != Node || Res.getResNo() != 0) {
1355 DAG.ReplaceAllUsesWith(Node, ResultVals.data());
1356 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1357 DAG.TransferDbgValues(SDValue(Node, i), ResultVals[i]);
1364 case TargetLowering::Expand:
1367 case TargetLowering::Promote:
1373 switch (Node->getOpcode()) {
1380 llvm_unreachable("Do not know how to legalize this operator!");
1382 case ISD::CALLSEQ_START:
1383 case ISD::CALLSEQ_END:
1386 return LegalizeLoadOps(Node);
1389 return LegalizeStoreOps(Node);
1394 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1395 SDValue Vec = Op.getOperand(0);
1396 SDValue Idx = Op.getOperand(1);
1399 // Before we generate a new store to a temporary stack slot, see if there is
1400 // already one that we can use. There often is because when we scalarize
1401 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1402 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1403 // the vector. If all are expanded here, we don't want one store per vector
1405 SDValue StackPtr, Ch;
1406 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1407 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1409 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1410 if (ST->isIndexed() || ST->isTruncatingStore() ||
1411 ST->getValue() != Vec)
1414 // Make sure that nothing else could have stored into the destination of
1416 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1419 StackPtr = ST->getBasePtr();
1420 Ch = SDValue(ST, 0);
1425 if (!Ch.getNode()) {
1426 // Store the value to a temporary stack slot, then LOAD the returned part.
1427 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1428 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1429 MachinePointerInfo(), false, false, 0);
1432 // Add the offset to the index.
1434 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1435 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1436 DAG.getConstant(EltSize, Idx.getValueType()));
1438 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1439 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1441 if (Op.getValueType().isVector())
1442 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1443 false, false, false, 0);
1444 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1445 MachinePointerInfo(),
1446 Vec.getValueType().getVectorElementType(),
1450 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1451 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1453 SDValue Vec = Op.getOperand(0);
1454 SDValue Part = Op.getOperand(1);
1455 SDValue Idx = Op.getOperand(2);
1458 // Store the value to a temporary stack slot, then LOAD the returned part.
1460 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1461 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1462 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1464 // First store the whole vector.
1465 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1468 // Then store the inserted part.
1470 // Add the offset to the index.
1472 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1474 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1475 DAG.getConstant(EltSize, Idx.getValueType()));
1476 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1478 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1481 // Store the subvector.
1482 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1483 MachinePointerInfo(), false, false, 0);
1485 // Finally, load the updated vector.
1486 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1487 false, false, false, 0);
1490 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1491 // We can't handle this case efficiently. Allocate a sufficiently
1492 // aligned object on the stack, store each element into it, then load
1493 // the result as a vector.
1494 // Create the stack frame object.
1495 EVT VT = Node->getValueType(0);
1496 EVT EltVT = VT.getVectorElementType();
1498 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1499 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1500 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1502 // Emit a store of each element to the stack slot.
1503 SmallVector<SDValue, 8> Stores;
1504 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1505 // Store (in the right endianness) the elements to memory.
1506 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1507 // Ignore undef elements.
1508 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1510 unsigned Offset = TypeByteSize*i;
1512 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1513 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1515 // If the destination vector element type is narrower than the source
1516 // element type, only store the bits necessary.
1517 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1518 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1519 Node->getOperand(i), Idx,
1520 PtrInfo.getWithOffset(Offset),
1521 EltVT, false, false, 0));
1523 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1524 Node->getOperand(i), Idx,
1525 PtrInfo.getWithOffset(Offset),
1530 if (!Stores.empty()) // Not all undef elements?
1531 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1532 &Stores[0], Stores.size());
1534 StoreChain = DAG.getEntryNode();
1536 // Result is a load from the stack slot.
1537 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1538 false, false, false, 0);
1541 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1543 SDValue Tmp1 = Node->getOperand(0);
1544 SDValue Tmp2 = Node->getOperand(1);
1546 // Get the sign bit of the RHS. First obtain a value that has the same
1547 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1549 EVT FloatVT = Tmp2.getValueType();
1550 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1551 if (TLI.isTypeLegal(IVT)) {
1552 // Convert to an integer with the same sign bit.
1553 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1555 // Store the float to memory, then load the sign part out as an integer.
1556 MVT LoadTy = TLI.getPointerTy();
1557 // First create a temporary that is aligned for both the load and store.
1558 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1559 // Then store the float to it.
1561 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1563 if (TLI.isBigEndian()) {
1564 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1565 // Load out a legal integer with the same sign bit as the float.
1566 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1567 false, false, false, 0);
1568 } else { // Little endian
1569 SDValue LoadPtr = StackPtr;
1570 // The float may be wider than the integer we are going to load. Advance
1571 // the pointer so that the loaded integer will contain the sign bit.
1572 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1573 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1574 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1575 DAG.getConstant(ByteOffset, LoadPtr.getValueType()));
1576 // Load a legal integer containing the sign bit.
1577 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1578 false, false, false, 0);
1579 // Move the sign bit to the top bit of the loaded integer.
1580 unsigned BitShift = LoadTy.getSizeInBits() -
1581 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1582 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1584 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1585 DAG.getConstant(BitShift,
1586 TLI.getShiftAmountTy(SignBit.getValueType())));
1589 // Now get the sign bit proper, by seeing whether the value is negative.
1590 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1591 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1593 // Get the absolute value of the result.
1594 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1595 // Select between the nabs and abs value based on the sign bit of
1597 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1598 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1602 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1603 SmallVectorImpl<SDValue> &Results) {
1604 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1605 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1606 " not tell us which reg is the stack pointer!");
1608 EVT VT = Node->getValueType(0);
1609 SDValue Tmp1 = SDValue(Node, 0);
1610 SDValue Tmp2 = SDValue(Node, 1);
1611 SDValue Tmp3 = Node->getOperand(2);
1612 SDValue Chain = Tmp1.getOperand(0);
1614 // Chain the dynamic stack allocation so that it doesn't modify the stack
1615 // pointer when other instructions are using the stack.
1616 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
1619 SDValue Size = Tmp2.getOperand(1);
1620 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1621 Chain = SP.getValue(1);
1622 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1623 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1624 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1625 if (Align > StackAlign)
1626 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1627 DAG.getConstant(-(uint64_t)Align, VT));
1628 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1630 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1631 DAG.getIntPtrConstant(0, true), SDValue(),
1634 Results.push_back(Tmp1);
1635 Results.push_back(Tmp2);
1638 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1639 /// condition code CC on the current target.
1641 /// If the SETCC has been legalized using AND / OR, then the legalized node
1642 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1643 /// will be set to false.
1645 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1646 /// then the values of LHS and RHS will be swapped, CC will be set to the
1647 /// new condition, and NeedInvert will be set to false.
1649 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1650 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1651 /// will be set to true. The caller must invert the result of the SETCC with
1652 /// SelectionDAG::getNOT() or take equivalent action to swap the effect of a
1653 /// true/false result.
1655 /// \returns true if the SetCC has been legalized, false if it hasn't.
1656 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1657 SDValue &LHS, SDValue &RHS,
1661 MVT OpVT = LHS.getSimpleValueType();
1662 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1664 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1665 default: llvm_unreachable("Unknown condition code action!");
1666 case TargetLowering::Legal:
1669 case TargetLowering::Expand: {
1670 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1671 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1672 std::swap(LHS, RHS);
1673 CC = DAG.getCondCode(InvCC);
1676 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1679 default: llvm_unreachable("Don't know how to expand this condition!");
1681 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1682 == TargetLowering::Legal
1683 && "If SETO is expanded, SETOEQ must be legal!");
1684 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1686 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1687 == TargetLowering::Legal
1688 && "If SETUO is expanded, SETUNE must be legal!");
1689 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1702 // If we are floating point, assign and break, otherwise fall through.
1703 if (!OpVT.isInteger()) {
1704 // We can use the 4th bit to tell if we are the unordered
1705 // or ordered version of the opcode.
1706 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1707 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1708 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1711 // Fallthrough if we are unsigned integer.
1716 // We only support using the inverted operation, which is computed above
1717 // and not a different manner of supporting expanding these cases.
1718 llvm_unreachable("Don't know how to expand this condition!");
1721 // Try inverting the result of the inverse condition.
1722 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1723 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1724 CC = DAG.getCondCode(InvCC);
1728 // If inverting the condition didn't work then we have no means to expand
1730 llvm_unreachable("Don't know how to expand this condition!");
1733 SDValue SetCC1, SetCC2;
1734 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1735 // If we aren't the ordered or unorder operation,
1736 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1737 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1738 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1740 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1741 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1742 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1744 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1753 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1754 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1755 /// a load from the stack slot to DestVT, extending it if needed.
1756 /// The resultant code need not be legal.
1757 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1761 // Create the stack frame object.
1763 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1764 getTypeForEVT(*DAG.getContext()));
1765 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1767 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1768 int SPFI = StackPtrFI->getIndex();
1769 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1771 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1772 unsigned SlotSize = SlotVT.getSizeInBits();
1773 unsigned DestSize = DestVT.getSizeInBits();
1774 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1775 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1777 // Emit a store to the stack slot. Use a truncstore if the input value is
1778 // later than DestVT.
1781 if (SrcSize > SlotSize)
1782 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1783 PtrInfo, SlotVT, false, false, SrcAlign);
1785 assert(SrcSize == SlotSize && "Invalid store");
1786 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1787 PtrInfo, false, false, SrcAlign);
1790 // Result is a load from the stack slot.
1791 if (SlotSize == DestSize)
1792 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1793 false, false, false, DestAlign);
1795 assert(SlotSize < DestSize && "Unknown extension!");
1796 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1797 PtrInfo, SlotVT, false, false, DestAlign);
1800 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1802 // Create a vector sized/aligned stack slot, store the value to element #0,
1803 // then load the whole vector back out.
1804 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1806 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1807 int SPFI = StackPtrFI->getIndex();
1809 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1811 MachinePointerInfo::getFixedStack(SPFI),
1812 Node->getValueType(0).getVectorElementType(),
1814 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1815 MachinePointerInfo::getFixedStack(SPFI),
1816 false, false, false, 0);
1820 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1821 /// support the operation, but do support the resultant vector type.
1822 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1823 unsigned NumElems = Node->getNumOperands();
1824 SDValue Value1, Value2;
1826 EVT VT = Node->getValueType(0);
1827 EVT OpVT = Node->getOperand(0).getValueType();
1828 EVT EltVT = VT.getVectorElementType();
1830 // If the only non-undef value is the low element, turn this into a
1831 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1832 bool isOnlyLowElement = true;
1833 bool MoreThanTwoValues = false;
1834 bool isConstant = true;
1835 for (unsigned i = 0; i < NumElems; ++i) {
1836 SDValue V = Node->getOperand(i);
1837 if (V.getOpcode() == ISD::UNDEF)
1840 isOnlyLowElement = false;
1841 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1844 if (!Value1.getNode()) {
1846 } else if (!Value2.getNode()) {
1849 } else if (V != Value1 && V != Value2) {
1850 MoreThanTwoValues = true;
1854 if (!Value1.getNode())
1855 return DAG.getUNDEF(VT);
1857 if (isOnlyLowElement)
1858 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1860 // If all elements are constants, create a load from the constant pool.
1862 SmallVector<Constant*, 16> CV;
1863 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1864 if (ConstantFPSDNode *V =
1865 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1866 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1867 } else if (ConstantSDNode *V =
1868 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1870 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1872 // If OpVT and EltVT don't match, EltVT is not legal and the
1873 // element values have been promoted/truncated earlier. Undo this;
1874 // we don't want a v16i8 to become a v16i32 for example.
1875 const ConstantInt *CI = V->getConstantIntValue();
1876 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1877 CI->getZExtValue()));
1880 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1881 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1882 CV.push_back(UndefValue::get(OpNTy));
1885 Constant *CP = ConstantVector::get(CV);
1886 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1887 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1888 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1889 MachinePointerInfo::getConstantPool(),
1890 false, false, false, Alignment);
1893 SmallSet<SDValue, 16> DefinedValues;
1894 for (unsigned i = 0; i < NumElems; ++i) {
1895 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
1897 DefinedValues.insert(Node->getOperand(i));
1900 if (!MoreThanTwoValues &&
1901 TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
1902 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1903 for (unsigned i = 0; i < NumElems; ++i) {
1904 SDValue V = Node->getOperand(i);
1905 if (V.getOpcode() == ISD::UNDEF)
1907 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1909 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1910 // Get the splatted value into the low element of a vector register.
1911 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1913 if (Value2.getNode())
1914 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1916 Vec2 = DAG.getUNDEF(VT);
1918 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1919 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1923 // Otherwise, we can't handle this case efficiently.
1924 return ExpandVectorBuildThroughStack(Node);
1927 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
1928 // does not fit into a register, return the lo part and set the hi part to the
1929 // by-reg argument. If it does fit into a single register, return the result
1930 // and leave the Hi part unset.
1931 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
1933 TargetLowering::ArgListTy Args;
1934 TargetLowering::ArgListEntry Entry;
1935 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1936 EVT ArgVT = Node->getOperand(i).getValueType();
1937 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1938 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
1939 Entry.isSExt = isSigned;
1940 Entry.isZExt = !isSigned;
1941 Args.push_back(Entry);
1943 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1944 TLI.getPointerTy());
1946 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1948 // By default, the input chain to this libcall is the entry node of the
1949 // function. If the libcall is going to be emitted as a tail call then
1950 // TLI.isUsedByReturnOnly will change it to the right chain if the return
1951 // node which is being folded has a non-entry input chain.
1952 SDValue InChain = DAG.getEntryNode();
1954 // isTailCall may be true since the callee does not reference caller stack
1955 // frame. Check if it's in the right position.
1956 SDValue TCChain = InChain;
1957 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
1962 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1963 0, TLI.getLibcallCallingConv(LC), isTailCall,
1964 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1965 Callee, Args, DAG, SDLoc(Node));
1966 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1969 if (!CallInfo.second.getNode())
1970 // It's a tailcall, return the chain (which is the DAG root).
1971 return DAG.getRoot();
1973 return CallInfo.first;
1976 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
1977 /// and returning a result of type RetVT.
1978 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
1979 const SDValue *Ops, unsigned NumOps,
1980 bool isSigned, SDLoc dl) {
1981 TargetLowering::ArgListTy Args;
1982 Args.reserve(NumOps);
1984 TargetLowering::ArgListEntry Entry;
1985 for (unsigned i = 0; i != NumOps; ++i) {
1986 Entry.Node = Ops[i];
1987 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
1988 Entry.isSExt = isSigned;
1989 Entry.isZExt = !isSigned;
1990 Args.push_back(Entry);
1992 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1993 TLI.getPointerTy());
1995 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
1997 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
1998 false, 0, TLI.getLibcallCallingConv(LC),
1999 /*isTailCall=*/false,
2000 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2001 Callee, Args, DAG, dl);
2002 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2004 return CallInfo.first;
2007 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
2008 // ExpandLibCall except that the first operand is the in-chain.
2009 std::pair<SDValue, SDValue>
2010 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2013 SDValue InChain = Node->getOperand(0);
2015 TargetLowering::ArgListTy Args;
2016 TargetLowering::ArgListEntry Entry;
2017 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2018 EVT ArgVT = Node->getOperand(i).getValueType();
2019 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2020 Entry.Node = Node->getOperand(i);
2022 Entry.isSExt = isSigned;
2023 Entry.isZExt = !isSigned;
2024 Args.push_back(Entry);
2026 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2027 TLI.getPointerTy());
2029 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2031 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
2032 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2033 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2034 Callee, Args, DAG, SDLoc(Node));
2035 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2040 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2041 RTLIB::Libcall Call_F32,
2042 RTLIB::Libcall Call_F64,
2043 RTLIB::Libcall Call_F80,
2044 RTLIB::Libcall Call_F128,
2045 RTLIB::Libcall Call_PPCF128) {
2047 switch (Node->getSimpleValueType(0).SimpleTy) {
2048 default: llvm_unreachable("Unexpected request for libcall!");
2049 case MVT::f32: LC = Call_F32; break;
2050 case MVT::f64: LC = Call_F64; break;
2051 case MVT::f80: LC = Call_F80; break;
2052 case MVT::f128: LC = Call_F128; break;
2053 case MVT::ppcf128: LC = Call_PPCF128; break;
2055 return ExpandLibCall(LC, Node, false);
2058 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2059 RTLIB::Libcall Call_I8,
2060 RTLIB::Libcall Call_I16,
2061 RTLIB::Libcall Call_I32,
2062 RTLIB::Libcall Call_I64,
2063 RTLIB::Libcall Call_I128) {
2065 switch (Node->getSimpleValueType(0).SimpleTy) {
2066 default: llvm_unreachable("Unexpected request for libcall!");
2067 case MVT::i8: LC = Call_I8; break;
2068 case MVT::i16: LC = Call_I16; break;
2069 case MVT::i32: LC = Call_I32; break;
2070 case MVT::i64: LC = Call_I64; break;
2071 case MVT::i128: LC = Call_I128; break;
2073 return ExpandLibCall(LC, Node, isSigned);
2076 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
2077 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2078 const TargetLowering &TLI) {
2080 switch (Node->getSimpleValueType(0).SimpleTy) {
2081 default: llvm_unreachable("Unexpected request for libcall!");
2082 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2083 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2084 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2085 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2086 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2089 return TLI.getLibcallName(LC) != 0;
2092 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2094 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2095 // The other use might have been replaced with a divrem already.
2096 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2097 unsigned OtherOpcode = 0;
2099 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2101 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2103 SDValue Op0 = Node->getOperand(0);
2104 SDValue Op1 = Node->getOperand(1);
2105 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2106 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2110 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2111 User->getOperand(0) == Op0 &&
2112 User->getOperand(1) == Op1)
2118 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2121 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2122 SmallVectorImpl<SDValue> &Results) {
2123 unsigned Opcode = Node->getOpcode();
2124 bool isSigned = Opcode == ISD::SDIVREM;
2127 switch (Node->getSimpleValueType(0).SimpleTy) {
2128 default: llvm_unreachable("Unexpected request for libcall!");
2129 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2130 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2131 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2132 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2133 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2136 // The input chain to this libcall is the entry node of the function.
2137 // Legalizing the call will automatically add the previous call to the
2139 SDValue InChain = DAG.getEntryNode();
2141 EVT RetVT = Node->getValueType(0);
2142 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2144 TargetLowering::ArgListTy Args;
2145 TargetLowering::ArgListEntry Entry;
2146 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2147 EVT ArgVT = Node->getOperand(i).getValueType();
2148 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2149 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2150 Entry.isSExt = isSigned;
2151 Entry.isZExt = !isSigned;
2152 Args.push_back(Entry);
2155 // Also pass the return address of the remainder.
2156 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2158 Entry.Ty = RetTy->getPointerTo();
2159 Entry.isSExt = isSigned;
2160 Entry.isZExt = !isSigned;
2161 Args.push_back(Entry);
2163 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2164 TLI.getPointerTy());
2168 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
2169 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2170 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2171 Callee, Args, DAG, dl);
2172 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2174 // Remainder is loaded back from the stack frame.
2175 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2176 MachinePointerInfo(), false, false, false, 0);
2177 Results.push_back(CallInfo.first);
2178 Results.push_back(Rem);
2181 /// isSinCosLibcallAvailable - Return true if sincos libcall is available.
2182 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2184 switch (Node->getSimpleValueType(0).SimpleTy) {
2185 default: llvm_unreachable("Unexpected request for libcall!");
2186 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2187 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2188 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2189 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2190 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2192 return TLI.getLibcallName(LC) != 0;
2195 /// canCombineSinCosLibcall - Return true if sincos libcall is available and
2196 /// can be used to combine sin and cos.
2197 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2198 const TargetMachine &TM) {
2199 if (!isSinCosLibcallAvailable(Node, TLI))
2201 // GNU sin/cos functions set errno while sincos does not. Therefore
2202 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2203 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2204 if (isGNU && !TM.Options.UnsafeFPMath)
2209 /// useSinCos - Only issue sincos libcall if both sin and cos are
2211 static bool useSinCos(SDNode *Node) {
2212 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2213 ? ISD::FCOS : ISD::FSIN;
2215 SDValue Op0 = Node->getOperand(0);
2216 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2217 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2221 // The other user might have been turned into sincos already.
2222 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2228 /// ExpandSinCosLibCall - Issue libcalls to sincos to compute sin / cos
2231 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2232 SmallVectorImpl<SDValue> &Results) {
2234 switch (Node->getSimpleValueType(0).SimpleTy) {
2235 default: llvm_unreachable("Unexpected request for libcall!");
2236 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2237 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2238 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2239 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2240 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2243 // The input chain to this libcall is the entry node of the function.
2244 // Legalizing the call will automatically add the previous call to the
2246 SDValue InChain = DAG.getEntryNode();
2248 EVT RetVT = Node->getValueType(0);
2249 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2251 TargetLowering::ArgListTy Args;
2252 TargetLowering::ArgListEntry Entry;
2254 // Pass the argument.
2255 Entry.Node = Node->getOperand(0);
2257 Entry.isSExt = false;
2258 Entry.isZExt = false;
2259 Args.push_back(Entry);
2261 // Pass the return address of sin.
2262 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2263 Entry.Node = SinPtr;
2264 Entry.Ty = RetTy->getPointerTo();
2265 Entry.isSExt = false;
2266 Entry.isZExt = false;
2267 Args.push_back(Entry);
2269 // Also pass the return address of the cos.
2270 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2271 Entry.Node = CosPtr;
2272 Entry.Ty = RetTy->getPointerTo();
2273 Entry.isSExt = false;
2274 Entry.isZExt = false;
2275 Args.push_back(Entry);
2277 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2278 TLI.getPointerTy());
2282 CallLoweringInfo CLI(InChain, Type::getVoidTy(*DAG.getContext()),
2283 false, false, false, false,
2284 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2285 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2286 Callee, Args, DAG, dl);
2287 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2289 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2290 MachinePointerInfo(), false, false, false, 0));
2291 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2292 MachinePointerInfo(), false, false, false, 0));
2295 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2296 /// INT_TO_FP operation of the specified operand when the target requests that
2297 /// we expand it. At this point, we know that the result and operand types are
2298 /// legal for the target.
2299 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2303 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2304 // simple 32-bit [signed|unsigned] integer to float/double expansion
2306 // Get the stack frame index of a 8 byte buffer.
2307 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2309 // word offset constant for Hi/Lo address computation
2310 SDValue WordOff = DAG.getConstant(sizeof(int), StackSlot.getValueType());
2311 // set up Hi and Lo (into buffer) address based on endian
2312 SDValue Hi = StackSlot;
2313 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2314 StackSlot, WordOff);
2315 if (TLI.isLittleEndian())
2318 // if signed map to unsigned space
2321 // constant used to invert sign bit (signed to unsigned mapping)
2322 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2323 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2327 // store the lo of the constructed double - based on integer input
2328 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2329 Op0Mapped, Lo, MachinePointerInfo(),
2331 // initial hi portion of constructed double
2332 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2333 // store the hi of the constructed double - biased exponent
2334 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2335 MachinePointerInfo(),
2337 // load the constructed double
2338 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2339 MachinePointerInfo(), false, false, false, 0);
2340 // FP constant to bias correct the final result
2341 SDValue Bias = DAG.getConstantFP(isSigned ?
2342 BitsToDouble(0x4330000080000000ULL) :
2343 BitsToDouble(0x4330000000000000ULL),
2345 // subtract the bias
2346 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2349 // handle final rounding
2350 if (DestVT == MVT::f64) {
2353 } else if (DestVT.bitsLT(MVT::f64)) {
2354 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2355 DAG.getIntPtrConstant(0));
2356 } else if (DestVT.bitsGT(MVT::f64)) {
2357 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2361 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2362 // Code below here assumes !isSigned without checking again.
2364 // Implementation of unsigned i64 to f64 following the algorithm in
2365 // __floatundidf in compiler_rt. This implementation has the advantage
2366 // of performing rounding correctly, both in the default rounding mode
2367 // and in all alternate rounding modes.
2368 // TODO: Generalize this for use with other types.
2369 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2371 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2372 SDValue TwoP84PlusTwoP52 =
2373 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2375 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2377 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2378 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2379 DAG.getConstant(32, MVT::i64));
2380 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2381 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2382 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2383 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2384 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2386 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2389 // Implementation of unsigned i64 to f32.
2390 // TODO: Generalize this for use with other types.
2391 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2392 // For unsigned conversions, convert them to signed conversions using the
2393 // algorithm from the x86_64 __floatundidf in compiler_rt.
2395 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2397 SDValue ShiftConst =
2398 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2399 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2400 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2401 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2402 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2404 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2405 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2407 // TODO: This really should be implemented using a branch rather than a
2408 // select. We happen to get lucky and machinesink does the right
2409 // thing most of the time. This would be a good candidate for a
2410 //pseudo-op, or, even better, for whole-function isel.
2411 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2412 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2413 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2416 // Otherwise, implement the fully general conversion.
2418 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2419 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2420 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2421 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2422 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2423 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2424 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2425 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2426 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2427 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2428 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2430 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2431 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2433 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2434 DAG.getConstant(32, SHVT));
2435 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2436 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2438 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2439 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2440 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2441 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2442 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2443 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2444 DAG.getIntPtrConstant(0));
2447 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2449 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2450 Op0, DAG.getConstant(0, Op0.getValueType()),
2452 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2453 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2454 SignSet, Four, Zero);
2456 // If the sign bit of the integer is set, the large number will be treated
2457 // as a negative number. To counteract this, the dynamic code adds an
2458 // offset depending on the data type.
2460 switch (Op0.getSimpleValueType().SimpleTy) {
2461 default: llvm_unreachable("Unsupported integer type!");
2462 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2463 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2464 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2465 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2467 if (TLI.isLittleEndian()) FF <<= 32;
2468 Constant *FudgeFactor = ConstantInt::get(
2469 Type::getInt64Ty(*DAG.getContext()), FF);
2471 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2472 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2473 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2474 Alignment = std::min(Alignment, 4u);
2476 if (DestVT == MVT::f32)
2477 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2478 MachinePointerInfo::getConstantPool(),
2479 false, false, false, Alignment);
2481 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2482 DAG.getEntryNode(), CPIdx,
2483 MachinePointerInfo::getConstantPool(),
2484 MVT::f32, false, false, Alignment);
2485 HandleSDNode Handle(Load);
2486 LegalizeOp(Load.getNode());
2487 FudgeInReg = Handle.getValue();
2490 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2493 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2494 /// *INT_TO_FP operation of the specified operand when the target requests that
2495 /// we promote it. At this point, we know that the result and operand types are
2496 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2497 /// operation that takes a larger input.
2498 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2502 // First step, figure out the appropriate *INT_TO_FP operation to use.
2503 EVT NewInTy = LegalOp.getValueType();
2505 unsigned OpToUse = 0;
2507 // Scan for the appropriate larger type to use.
2509 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2510 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2512 // If the target supports SINT_TO_FP of this type, use it.
2513 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2514 OpToUse = ISD::SINT_TO_FP;
2517 if (isSigned) continue;
2519 // If the target supports UINT_TO_FP of this type, use it.
2520 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2521 OpToUse = ISD::UINT_TO_FP;
2525 // Otherwise, try a larger type.
2528 // Okay, we found the operation and type to use. Zero extend our input to the
2529 // desired type then run the operation on it.
2530 return DAG.getNode(OpToUse, dl, DestVT,
2531 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2532 dl, NewInTy, LegalOp));
2535 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2536 /// FP_TO_*INT operation of the specified operand when the target requests that
2537 /// we promote it. At this point, we know that the result and operand types are
2538 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2539 /// operation that returns a larger result.
2540 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2544 // First step, figure out the appropriate FP_TO*INT operation to use.
2545 EVT NewOutTy = DestVT;
2547 unsigned OpToUse = 0;
2549 // Scan for the appropriate larger type to use.
2551 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2552 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2554 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2555 OpToUse = ISD::FP_TO_SINT;
2559 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2560 OpToUse = ISD::FP_TO_UINT;
2564 // Otherwise, try a larger type.
2568 // Okay, we found the operation and type to use.
2569 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2571 // Truncate the result of the extended FP_TO_*INT operation to the desired
2573 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2576 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2578 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2579 EVT VT = Op.getValueType();
2580 EVT SHVT = TLI.getShiftAmountTy(VT);
2581 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2582 switch (VT.getSimpleVT().SimpleTy) {
2583 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2585 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2586 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2587 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2589 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2590 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2591 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2592 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2593 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2594 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2595 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2596 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2597 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2599 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2600 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2601 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2602 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2603 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2604 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2605 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2606 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2607 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2608 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2609 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2610 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2611 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2612 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2613 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2614 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2615 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2616 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2617 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2618 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2619 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2623 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2625 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2628 default: llvm_unreachable("Cannot expand this yet!");
2630 EVT VT = Op.getValueType();
2631 EVT ShVT = TLI.getShiftAmountTy(VT);
2632 unsigned Len = VT.getSizeInBits();
2634 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2635 "CTPOP not implemented for this type.");
2637 // This is the "best" algorithm from
2638 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2640 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2641 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2642 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2643 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2645 // v = v - ((v >> 1) & 0x55555555...)
2646 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2647 DAG.getNode(ISD::AND, dl, VT,
2648 DAG.getNode(ISD::SRL, dl, VT, Op,
2649 DAG.getConstant(1, ShVT)),
2651 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2652 Op = DAG.getNode(ISD::ADD, dl, VT,
2653 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2654 DAG.getNode(ISD::AND, dl, VT,
2655 DAG.getNode(ISD::SRL, dl, VT, Op,
2656 DAG.getConstant(2, ShVT)),
2658 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2659 Op = DAG.getNode(ISD::AND, dl, VT,
2660 DAG.getNode(ISD::ADD, dl, VT, Op,
2661 DAG.getNode(ISD::SRL, dl, VT, Op,
2662 DAG.getConstant(4, ShVT))),
2664 // v = (v * 0x01010101...) >> (Len - 8)
2665 Op = DAG.getNode(ISD::SRL, dl, VT,
2666 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2667 DAG.getConstant(Len - 8, ShVT));
2671 case ISD::CTLZ_ZERO_UNDEF:
2672 // This trivially expands to CTLZ.
2673 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2675 // for now, we do this:
2676 // x = x | (x >> 1);
2677 // x = x | (x >> 2);
2679 // x = x | (x >>16);
2680 // x = x | (x >>32); // for 64-bit input
2681 // return popcount(~x);
2683 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2684 EVT VT = Op.getValueType();
2685 EVT ShVT = TLI.getShiftAmountTy(VT);
2686 unsigned len = VT.getSizeInBits();
2687 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2688 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2689 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2690 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2692 Op = DAG.getNOT(dl, Op, VT);
2693 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2695 case ISD::CTTZ_ZERO_UNDEF:
2696 // This trivially expands to CTTZ.
2697 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2699 // for now, we use: { return popcount(~x & (x - 1)); }
2700 // unless the target has ctlz but not ctpop, in which case we use:
2701 // { return 32 - nlz(~x & (x-1)); }
2702 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2703 EVT VT = Op.getValueType();
2704 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2705 DAG.getNOT(dl, Op, VT),
2706 DAG.getNode(ISD::SUB, dl, VT, Op,
2707 DAG.getConstant(1, VT)));
2708 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2709 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2710 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2711 return DAG.getNode(ISD::SUB, dl, VT,
2712 DAG.getConstant(VT.getSizeInBits(), VT),
2713 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2714 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2719 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2720 unsigned Opc = Node->getOpcode();
2721 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2726 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2727 case ISD::ATOMIC_SWAP:
2728 switch (VT.SimpleTy) {
2729 default: llvm_unreachable("Unexpected value type for atomic!");
2730 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2731 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2732 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2733 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2734 case MVT::i128:LC = RTLIB::SYNC_LOCK_TEST_AND_SET_16;break;
2737 case ISD::ATOMIC_CMP_SWAP:
2738 switch (VT.SimpleTy) {
2739 default: llvm_unreachable("Unexpected value type for atomic!");
2740 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2741 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2742 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2743 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2744 case MVT::i128:LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16;break;
2747 case ISD::ATOMIC_LOAD_ADD:
2748 switch (VT.SimpleTy) {
2749 default: llvm_unreachable("Unexpected value type for atomic!");
2750 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2751 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2752 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2753 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2754 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_ADD_16;break;
2757 case ISD::ATOMIC_LOAD_SUB:
2758 switch (VT.SimpleTy) {
2759 default: llvm_unreachable("Unexpected value type for atomic!");
2760 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2761 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2762 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2763 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2764 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_SUB_16;break;
2767 case ISD::ATOMIC_LOAD_AND:
2768 switch (VT.SimpleTy) {
2769 default: llvm_unreachable("Unexpected value type for atomic!");
2770 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2771 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2772 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2773 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2774 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_AND_16;break;
2777 case ISD::ATOMIC_LOAD_OR:
2778 switch (VT.SimpleTy) {
2779 default: llvm_unreachable("Unexpected value type for atomic!");
2780 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2781 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2782 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2783 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2784 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_OR_16;break;
2787 case ISD::ATOMIC_LOAD_XOR:
2788 switch (VT.SimpleTy) {
2789 default: llvm_unreachable("Unexpected value type for atomic!");
2790 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2791 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2792 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2793 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2794 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_XOR_16;break;
2797 case ISD::ATOMIC_LOAD_NAND:
2798 switch (VT.SimpleTy) {
2799 default: llvm_unreachable("Unexpected value type for atomic!");
2800 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2801 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2802 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2803 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2804 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_NAND_16;break;
2807 case ISD::ATOMIC_LOAD_MAX:
2808 switch (VT.SimpleTy) {
2809 default: llvm_unreachable("Unexpected value type for atomic!");
2810 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MAX_1; break;
2811 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MAX_2; break;
2812 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MAX_4; break;
2813 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MAX_8; break;
2814 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MAX_16;break;
2817 case ISD::ATOMIC_LOAD_UMAX:
2818 switch (VT.SimpleTy) {
2819 default: llvm_unreachable("Unexpected value type for atomic!");
2820 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMAX_1; break;
2821 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMAX_2; break;
2822 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMAX_4; break;
2823 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMAX_8; break;
2824 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMAX_16;break;
2827 case ISD::ATOMIC_LOAD_MIN:
2828 switch (VT.SimpleTy) {
2829 default: llvm_unreachable("Unexpected value type for atomic!");
2830 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MIN_1; break;
2831 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MIN_2; break;
2832 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MIN_4; break;
2833 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MIN_8; break;
2834 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MIN_16;break;
2837 case ISD::ATOMIC_LOAD_UMIN:
2838 switch (VT.SimpleTy) {
2839 default: llvm_unreachable("Unexpected value type for atomic!");
2840 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMIN_1; break;
2841 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMIN_2; break;
2842 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMIN_4; break;
2843 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMIN_8; break;
2844 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMIN_16;break;
2849 return ExpandChainLibCall(LC, Node, false);
2852 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2853 SmallVector<SDValue, 8> Results;
2855 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2857 switch (Node->getOpcode()) {
2860 case ISD::CTLZ_ZERO_UNDEF:
2862 case ISD::CTTZ_ZERO_UNDEF:
2863 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2864 Results.push_back(Tmp1);
2867 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2869 case ISD::FRAMEADDR:
2870 case ISD::RETURNADDR:
2871 case ISD::FRAME_TO_ARGS_OFFSET:
2872 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2874 case ISD::FLT_ROUNDS_:
2875 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2877 case ISD::EH_RETURN:
2881 case ISD::EH_SJLJ_LONGJMP:
2882 // If the target didn't expand these, there's nothing to do, so just
2883 // preserve the chain and be done.
2884 Results.push_back(Node->getOperand(0));
2886 case ISD::EH_SJLJ_SETJMP:
2887 // If the target didn't expand this, just return 'zero' and preserve the
2889 Results.push_back(DAG.getConstant(0, MVT::i32));
2890 Results.push_back(Node->getOperand(0));
2892 case ISD::ATOMIC_FENCE: {
2893 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2894 // FIXME: handle "fence singlethread" more efficiently.
2895 TargetLowering::ArgListTy Args;
2897 CallLoweringInfo CLI(Node->getOperand(0),
2898 Type::getVoidTy(*DAG.getContext()),
2899 false, false, false, false, 0, CallingConv::C,
2900 /*isTailCall=*/false,
2901 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2902 DAG.getExternalSymbol("__sync_synchronize",
2903 TLI.getPointerTy()),
2905 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2907 Results.push_back(CallResult.second);
2910 case ISD::ATOMIC_LOAD: {
2911 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2912 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
2913 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
2914 cast<AtomicSDNode>(Node)->getMemoryVT(),
2915 Node->getOperand(0),
2916 Node->getOperand(1), Zero, Zero,
2917 cast<AtomicSDNode>(Node)->getMemOperand(),
2918 cast<AtomicSDNode>(Node)->getOrdering(),
2919 cast<AtomicSDNode>(Node)->getOrdering(),
2920 cast<AtomicSDNode>(Node)->getSynchScope());
2921 Results.push_back(Swap.getValue(0));
2922 Results.push_back(Swap.getValue(1));
2925 case ISD::ATOMIC_STORE: {
2926 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2927 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2928 cast<AtomicSDNode>(Node)->getMemoryVT(),
2929 Node->getOperand(0),
2930 Node->getOperand(1), Node->getOperand(2),
2931 cast<AtomicSDNode>(Node)->getMemOperand(),
2932 cast<AtomicSDNode>(Node)->getOrdering(),
2933 cast<AtomicSDNode>(Node)->getSynchScope());
2934 Results.push_back(Swap.getValue(1));
2937 // By default, atomic intrinsics are marked Legal and lowered. Targets
2938 // which don't support them directly, however, may want libcalls, in which
2939 // case they mark them Expand, and we get here.
2940 case ISD::ATOMIC_SWAP:
2941 case ISD::ATOMIC_LOAD_ADD:
2942 case ISD::ATOMIC_LOAD_SUB:
2943 case ISD::ATOMIC_LOAD_AND:
2944 case ISD::ATOMIC_LOAD_OR:
2945 case ISD::ATOMIC_LOAD_XOR:
2946 case ISD::ATOMIC_LOAD_NAND:
2947 case ISD::ATOMIC_LOAD_MIN:
2948 case ISD::ATOMIC_LOAD_MAX:
2949 case ISD::ATOMIC_LOAD_UMIN:
2950 case ISD::ATOMIC_LOAD_UMAX:
2951 case ISD::ATOMIC_CMP_SWAP: {
2952 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
2953 Results.push_back(Tmp.first);
2954 Results.push_back(Tmp.second);
2957 case ISD::DYNAMIC_STACKALLOC:
2958 ExpandDYNAMIC_STACKALLOC(Node, Results);
2960 case ISD::MERGE_VALUES:
2961 for (unsigned i = 0; i < Node->getNumValues(); i++)
2962 Results.push_back(Node->getOperand(i));
2965 EVT VT = Node->getValueType(0);
2967 Results.push_back(DAG.getConstant(0, VT));
2969 assert(VT.isFloatingPoint() && "Unknown value type!");
2970 Results.push_back(DAG.getConstantFP(0, VT));
2975 // If this operation is not supported, lower it to 'abort()' call
2976 TargetLowering::ArgListTy Args;
2978 CallLoweringInfo CLI(Node->getOperand(0),
2979 Type::getVoidTy(*DAG.getContext()),
2980 false, false, false, false, 0, CallingConv::C,
2981 /*isTailCall=*/false,
2982 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2983 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
2985 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2987 Results.push_back(CallResult.second);
2992 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
2993 Node->getValueType(0), dl);
2994 Results.push_back(Tmp1);
2996 case ISD::FP_EXTEND:
2997 Tmp1 = EmitStackConvert(Node->getOperand(0),
2998 Node->getOperand(0).getValueType(),
2999 Node->getValueType(0), dl);
3000 Results.push_back(Tmp1);
3002 case ISD::SIGN_EXTEND_INREG: {
3003 // NOTE: we could fall back on load/store here too for targets without
3004 // SAR. However, it is doubtful that any exist.
3005 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3006 EVT VT = Node->getValueType(0);
3007 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3010 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3011 ExtraVT.getScalarType().getSizeInBits();
3012 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
3013 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3014 Node->getOperand(0), ShiftCst);
3015 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3016 Results.push_back(Tmp1);
3019 case ISD::FP_ROUND_INREG: {
3020 // The only way we can lower this is to turn it into a TRUNCSTORE,
3021 // EXTLOAD pair, targeting a temporary location (a stack slot).
3023 // NOTE: there is a choice here between constantly creating new stack
3024 // slots and always reusing the same one. We currently always create
3025 // new ones, as reuse may inhibit scheduling.
3026 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3027 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3028 Node->getValueType(0), dl);
3029 Results.push_back(Tmp1);
3032 case ISD::SINT_TO_FP:
3033 case ISD::UINT_TO_FP:
3034 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3035 Node->getOperand(0), Node->getValueType(0), dl);
3036 Results.push_back(Tmp1);
3038 case ISD::FP_TO_UINT: {
3039 SDValue True, False;
3040 EVT VT = Node->getOperand(0).getValueType();
3041 EVT NVT = Node->getValueType(0);
3042 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3043 APInt::getNullValue(VT.getSizeInBits()));
3044 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3045 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3046 Tmp1 = DAG.getConstantFP(apf, VT);
3047 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3048 Node->getOperand(0),
3050 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3051 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3052 DAG.getNode(ISD::FSUB, dl, VT,
3053 Node->getOperand(0), Tmp1));
3054 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3055 DAG.getConstant(x, NVT));
3056 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3057 Results.push_back(Tmp1);
3061 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3062 EVT VT = Node->getValueType(0);
3063 Tmp1 = Node->getOperand(0);
3064 Tmp2 = Node->getOperand(1);
3065 unsigned Align = Node->getConstantOperandVal(3);
3067 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
3068 MachinePointerInfo(V),
3069 false, false, false, 0);
3070 SDValue VAList = VAListLoad;
3072 if (Align > TLI.getMinStackArgumentAlignment()) {
3073 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3075 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3076 DAG.getConstant(Align - 1,
3077 VAList.getValueType()));
3079 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3080 DAG.getConstant(-(int64_t)Align,
3081 VAList.getValueType()));
3084 // Increment the pointer, VAList, to the next vaarg
3085 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3086 DAG.getConstant(TLI.getDataLayout()->
3087 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
3088 VAList.getValueType()));
3089 // Store the incremented VAList to the legalized pointer
3090 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3091 MachinePointerInfo(V), false, false, 0);
3092 // Load the actual argument out of the pointer VAList
3093 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3094 false, false, false, 0));
3095 Results.push_back(Results[0].getValue(1));
3099 // This defaults to loading a pointer from the input and storing it to the
3100 // output, returning the chain.
3101 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3102 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3103 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
3104 Node->getOperand(2), MachinePointerInfo(VS),
3105 false, false, false, 0);
3106 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3107 MachinePointerInfo(VD), false, false, 0);
3108 Results.push_back(Tmp1);
3111 case ISD::EXTRACT_VECTOR_ELT:
3112 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3113 // This must be an access of the only element. Return it.
3114 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3115 Node->getOperand(0));
3117 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3118 Results.push_back(Tmp1);
3120 case ISD::EXTRACT_SUBVECTOR:
3121 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3123 case ISD::INSERT_SUBVECTOR:
3124 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3126 case ISD::CONCAT_VECTORS: {
3127 Results.push_back(ExpandVectorBuildThroughStack(Node));
3130 case ISD::SCALAR_TO_VECTOR:
3131 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3133 case ISD::INSERT_VECTOR_ELT:
3134 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3135 Node->getOperand(1),
3136 Node->getOperand(2), dl));
3138 case ISD::VECTOR_SHUFFLE: {
3139 SmallVector<int, 32> NewMask;
3140 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3142 EVT VT = Node->getValueType(0);
3143 EVT EltVT = VT.getVectorElementType();
3144 SDValue Op0 = Node->getOperand(0);
3145 SDValue Op1 = Node->getOperand(1);
3146 if (!TLI.isTypeLegal(EltVT)) {
3148 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3150 // BUILD_VECTOR operands are allowed to be wider than the element type.
3151 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3153 if (NewEltVT.bitsLT(EltVT)) {
3155 // Convert shuffle node.
3156 // If original node was v4i64 and the new EltVT is i32,
3157 // cast operands to v8i32 and re-build the mask.
3159 // Calculate new VT, the size of the new VT should be equal to original.
3161 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3162 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3163 assert(NewVT.bitsEq(VT));
3165 // cast operands to new VT
3166 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3167 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3169 // Convert the shuffle mask
3170 unsigned int factor =
3171 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3173 // EltVT gets smaller
3176 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3178 for (unsigned fi = 0; fi < factor; ++fi)
3179 NewMask.push_back(Mask[i]);
3182 for (unsigned fi = 0; fi < factor; ++fi)
3183 NewMask.push_back(Mask[i]*factor+fi);
3191 unsigned NumElems = VT.getVectorNumElements();
3192 SmallVector<SDValue, 16> Ops;
3193 for (unsigned i = 0; i != NumElems; ++i) {
3195 Ops.push_back(DAG.getUNDEF(EltVT));
3198 unsigned Idx = Mask[i];
3200 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3202 DAG.getConstant(Idx, TLI.getVectorIdxTy())));
3204 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3206 DAG.getConstant(Idx - NumElems,
3207 TLI.getVectorIdxTy())));
3210 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
3211 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3212 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3213 Results.push_back(Tmp1);
3216 case ISD::EXTRACT_ELEMENT: {
3217 EVT OpTy = Node->getOperand(0).getValueType();
3218 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3220 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3221 DAG.getConstant(OpTy.getSizeInBits()/2,
3222 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3223 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3226 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3227 Node->getOperand(0));
3229 Results.push_back(Tmp1);
3232 case ISD::STACKSAVE:
3233 // Expand to CopyFromReg if the target set
3234 // StackPointerRegisterToSaveRestore.
3235 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3236 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3237 Node->getValueType(0)));
3238 Results.push_back(Results[0].getValue(1));
3240 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3241 Results.push_back(Node->getOperand(0));
3244 case ISD::STACKRESTORE:
3245 // Expand to CopyToReg if the target set
3246 // StackPointerRegisterToSaveRestore.
3247 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3248 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3249 Node->getOperand(1)));
3251 Results.push_back(Node->getOperand(0));
3254 case ISD::FCOPYSIGN:
3255 Results.push_back(ExpandFCOPYSIGN(Node));
3258 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3259 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3260 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3261 Node->getOperand(0));
3262 Results.push_back(Tmp1);
3265 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3266 EVT VT = Node->getValueType(0);
3267 Tmp1 = Node->getOperand(0);
3268 Tmp2 = DAG.getConstantFP(0.0, VT);
3269 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3270 Tmp1, Tmp2, ISD::SETUGT);
3271 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3272 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3273 Results.push_back(Tmp1);
3277 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3278 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3279 RTLIB::SQRT_PPCF128));
3283 EVT VT = Node->getValueType(0);
3284 bool isSIN = Node->getOpcode() == ISD::FSIN;
3285 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3286 // fcos which share the same operand and both are used.
3287 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3288 canCombineSinCosLibcall(Node, TLI, TM))
3289 && useSinCos(Node)) {
3290 SDVTList VTs = DAG.getVTList(VT, VT);
3291 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3293 Tmp1 = Tmp1.getValue(1);
3294 Results.push_back(Tmp1);
3296 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3297 RTLIB::SIN_F80, RTLIB::SIN_F128,
3298 RTLIB::SIN_PPCF128));
3300 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3301 RTLIB::COS_F80, RTLIB::COS_F128,
3302 RTLIB::COS_PPCF128));
3307 // Expand into sincos libcall.
3308 ExpandSinCosLibCall(Node, Results);
3311 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3312 RTLIB::LOG_F80, RTLIB::LOG_F128,
3313 RTLIB::LOG_PPCF128));
3316 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3317 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3318 RTLIB::LOG2_PPCF128));
3321 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3322 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3323 RTLIB::LOG10_PPCF128));
3326 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3327 RTLIB::EXP_F80, RTLIB::EXP_F128,
3328 RTLIB::EXP_PPCF128));
3331 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3332 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3333 RTLIB::EXP2_PPCF128));
3336 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3337 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3338 RTLIB::TRUNC_PPCF128));
3341 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3342 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3343 RTLIB::FLOOR_PPCF128));
3346 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3347 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3348 RTLIB::CEIL_PPCF128));
3351 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3352 RTLIB::RINT_F80, RTLIB::RINT_F128,
3353 RTLIB::RINT_PPCF128));
3355 case ISD::FNEARBYINT:
3356 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3357 RTLIB::NEARBYINT_F64,
3358 RTLIB::NEARBYINT_F80,
3359 RTLIB::NEARBYINT_F128,
3360 RTLIB::NEARBYINT_PPCF128));
3363 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3367 RTLIB::ROUND_PPCF128));
3370 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3371 RTLIB::POWI_F80, RTLIB::POWI_F128,
3372 RTLIB::POWI_PPCF128));
3375 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3376 RTLIB::POW_F80, RTLIB::POW_F128,
3377 RTLIB::POW_PPCF128));
3380 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3381 RTLIB::DIV_F80, RTLIB::DIV_F128,
3382 RTLIB::DIV_PPCF128));
3385 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3386 RTLIB::REM_F80, RTLIB::REM_F128,
3387 RTLIB::REM_PPCF128));
3390 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3391 RTLIB::FMA_F80, RTLIB::FMA_F128,
3392 RTLIB::FMA_PPCF128));
3394 case ISD::FP16_TO_FP32:
3395 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3397 case ISD::FP32_TO_FP16:
3398 Results.push_back(ExpandLibCall(RTLIB::FPROUND_F32_F16, Node, false));
3400 case ISD::ConstantFP: {
3401 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3402 // Check to see if this FP immediate is already legal.
3403 // If this is a legal constant, turn it into a TargetConstantFP node.
3404 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3405 Results.push_back(ExpandConstantFP(CFP, true));
3409 EVT VT = Node->getValueType(0);
3410 assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3411 TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
3412 "Don't know how to expand this FP subtraction!");
3413 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3414 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3415 Results.push_back(Tmp1);
3419 EVT VT = Node->getValueType(0);
3420 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3421 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3422 "Don't know how to expand this subtraction!");
3423 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3424 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3425 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3426 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3431 EVT VT = Node->getValueType(0);
3432 bool isSigned = Node->getOpcode() == ISD::SREM;
3433 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3434 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3435 Tmp2 = Node->getOperand(0);
3436 Tmp3 = Node->getOperand(1);
3437 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3438 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3439 // If div is legal, it's better to do the normal expansion
3440 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3441 useDivRem(Node, isSigned, false))) {
3442 SDVTList VTs = DAG.getVTList(VT, VT);
3443 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3444 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3446 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3447 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3448 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3449 } else if (isSigned)
3450 Tmp1 = ExpandIntLibCall(Node, true,
3452 RTLIB::SREM_I16, RTLIB::SREM_I32,
3453 RTLIB::SREM_I64, RTLIB::SREM_I128);
3455 Tmp1 = ExpandIntLibCall(Node, false,
3457 RTLIB::UREM_I16, RTLIB::UREM_I32,
3458 RTLIB::UREM_I64, RTLIB::UREM_I128);
3459 Results.push_back(Tmp1);
3464 bool isSigned = Node->getOpcode() == ISD::SDIV;
3465 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3466 EVT VT = Node->getValueType(0);
3467 SDVTList VTs = DAG.getVTList(VT, VT);
3468 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3469 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3470 useDivRem(Node, isSigned, true)))
3471 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3472 Node->getOperand(1));
3474 Tmp1 = ExpandIntLibCall(Node, true,
3476 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3477 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3479 Tmp1 = ExpandIntLibCall(Node, false,
3481 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3482 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3483 Results.push_back(Tmp1);
3488 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3490 EVT VT = Node->getValueType(0);
3491 SDVTList VTs = DAG.getVTList(VT, VT);
3492 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3493 "If this wasn't legal, it shouldn't have been created!");
3494 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3495 Node->getOperand(1));
3496 Results.push_back(Tmp1.getValue(1));
3501 // Expand into divrem libcall
3502 ExpandDivRemLibCall(Node, Results);
3505 EVT VT = Node->getValueType(0);
3506 SDVTList VTs = DAG.getVTList(VT, VT);
3507 // See if multiply or divide can be lowered using two-result operations.
3508 // We just need the low half of the multiply; try both the signed
3509 // and unsigned forms. If the target supports both SMUL_LOHI and
3510 // UMUL_LOHI, form a preference by checking which forms of plain
3511 // MULH it supports.
3512 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3513 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3514 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3515 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3516 unsigned OpToUse = 0;
3517 if (HasSMUL_LOHI && !HasMULHS) {
3518 OpToUse = ISD::SMUL_LOHI;
3519 } else if (HasUMUL_LOHI && !HasMULHU) {
3520 OpToUse = ISD::UMUL_LOHI;
3521 } else if (HasSMUL_LOHI) {
3522 OpToUse = ISD::SMUL_LOHI;
3523 } else if (HasUMUL_LOHI) {
3524 OpToUse = ISD::UMUL_LOHI;
3527 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3528 Node->getOperand(1)));
3531 Tmp1 = ExpandIntLibCall(Node, false,
3533 RTLIB::MUL_I16, RTLIB::MUL_I32,
3534 RTLIB::MUL_I64, RTLIB::MUL_I128);
3535 Results.push_back(Tmp1);
3540 SDValue LHS = Node->getOperand(0);
3541 SDValue RHS = Node->getOperand(1);
3542 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3543 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3545 Results.push_back(Sum);
3546 EVT OType = Node->getValueType(1);
3548 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3550 // LHSSign -> LHS >= 0
3551 // RHSSign -> RHS >= 0
3552 // SumSign -> Sum >= 0
3555 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3557 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3559 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3560 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3561 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3562 Node->getOpcode() == ISD::SADDO ?
3563 ISD::SETEQ : ISD::SETNE);
3565 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3566 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3568 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3569 Results.push_back(Cmp);
3574 SDValue LHS = Node->getOperand(0);
3575 SDValue RHS = Node->getOperand(1);
3576 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3577 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3579 Results.push_back(Sum);
3580 Results.push_back(DAG.getSetCC(dl, Node->getValueType(1), Sum, LHS,
3581 Node->getOpcode () == ISD::UADDO ?
3582 ISD::SETULT : ISD::SETUGT));
3587 EVT VT = Node->getValueType(0);
3588 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3589 SDValue LHS = Node->getOperand(0);
3590 SDValue RHS = Node->getOperand(1);
3593 static const unsigned Ops[2][3] =
3594 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3595 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3596 bool isSigned = Node->getOpcode() == ISD::SMULO;
3597 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3598 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3599 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3600 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3601 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3603 TopHalf = BottomHalf.getValue(1);
3604 } else if (TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
3605 VT.getSizeInBits() * 2))) {
3606 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3607 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3608 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3609 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3610 DAG.getIntPtrConstant(0));
3611 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3612 DAG.getIntPtrConstant(1));
3614 // We can fall back to a libcall with an illegal type for the MUL if we
3615 // have a libcall big enough.
3616 // Also, we can fall back to a division in some cases, but that's a big
3617 // performance hit in the general case.
3618 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3619 if (WideVT == MVT::i16)
3620 LC = RTLIB::MUL_I16;
3621 else if (WideVT == MVT::i32)
3622 LC = RTLIB::MUL_I32;
3623 else if (WideVT == MVT::i64)
3624 LC = RTLIB::MUL_I64;
3625 else if (WideVT == MVT::i128)
3626 LC = RTLIB::MUL_I128;
3627 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3629 // The high part is obtained by SRA'ing all but one of the bits of low
3631 unsigned LoSize = VT.getSizeInBits();
3632 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3633 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3634 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3635 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3637 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3638 // pre-lowered to the correct types. This all depends upon WideVT not
3639 // being a legal type for the architecture and thus has to be split to
3641 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3642 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3643 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3644 DAG.getIntPtrConstant(0));
3645 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3646 DAG.getIntPtrConstant(1));
3647 // Ret is a node with an illegal type. Because such things are not
3648 // generally permitted during this phase of legalization, delete the
3649 // node. The above EXTRACT_ELEMENT nodes should have been folded.
3650 DAG.DeleteNode(Ret.getNode());
3654 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3655 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3656 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3657 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3660 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3661 DAG.getConstant(0, VT), ISD::SETNE);
3663 Results.push_back(BottomHalf);
3664 Results.push_back(TopHalf);
3667 case ISD::BUILD_PAIR: {
3668 EVT PairTy = Node->getValueType(0);
3669 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3670 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3671 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3672 DAG.getConstant(PairTy.getSizeInBits()/2,
3673 TLI.getShiftAmountTy(PairTy)));
3674 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3678 Tmp1 = Node->getOperand(0);
3679 Tmp2 = Node->getOperand(1);
3680 Tmp3 = Node->getOperand(2);
3681 if (Tmp1.getOpcode() == ISD::SETCC) {
3682 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3684 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3686 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3687 DAG.getConstant(0, Tmp1.getValueType()),
3688 Tmp2, Tmp3, ISD::SETNE);
3690 Results.push_back(Tmp1);
3693 SDValue Chain = Node->getOperand(0);
3694 SDValue Table = Node->getOperand(1);
3695 SDValue Index = Node->getOperand(2);
3697 EVT PTy = TLI.getPointerTy();
3699 const DataLayout &TD = *TLI.getDataLayout();
3700 unsigned EntrySize =
3701 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3703 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(),
3704 Index, DAG.getConstant(EntrySize, Index.getValueType()));
3705 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3708 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3709 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3710 MachinePointerInfo::getJumpTable(), MemVT,
3713 if (TM.getRelocationModel() == Reloc::PIC_) {
3714 // For PIC, the sequence is:
3715 // BRIND(load(Jumptable + index) + RelocBase)
3716 // RelocBase can be JumpTable, GOT or some sort of global base.
3717 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3718 TLI.getPICJumpTableRelocBase(Table, DAG));
3720 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3721 Results.push_back(Tmp1);
3725 // Expand brcond's setcc into its constituent parts and create a BR_CC
3727 Tmp1 = Node->getOperand(0);
3728 Tmp2 = Node->getOperand(1);
3729 if (Tmp2.getOpcode() == ISD::SETCC) {
3730 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3731 Tmp1, Tmp2.getOperand(2),
3732 Tmp2.getOperand(0), Tmp2.getOperand(1),
3733 Node->getOperand(2));
3735 // We test only the i1 bit. Skip the AND if UNDEF.
3736 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3737 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3738 DAG.getConstant(1, Tmp2.getValueType()));
3739 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3740 DAG.getCondCode(ISD::SETNE), Tmp3,
3741 DAG.getConstant(0, Tmp3.getValueType()),
3742 Node->getOperand(2));
3744 Results.push_back(Tmp1);
3747 Tmp1 = Node->getOperand(0);
3748 Tmp2 = Node->getOperand(1);
3749 Tmp3 = Node->getOperand(2);
3750 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3751 Tmp3, NeedInvert, dl);
3754 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3755 // condition code, create a new SETCC node.
3757 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3760 // If we expanded the SETCC by inverting the condition code, then wrap
3761 // the existing SETCC in a NOT to restore the intended condition.
3763 Tmp1 = DAG.getNOT(dl, Tmp1, Tmp1->getValueType(0));
3765 Results.push_back(Tmp1);
3769 // Otherwise, SETCC for the given comparison type must be completely
3770 // illegal; expand it into a SELECT_CC.
3771 EVT VT = Node->getValueType(0);
3773 switch (TLI.getBooleanContents(VT.isVector())) {
3774 case TargetLowering::ZeroOrOneBooleanContent:
3775 case TargetLowering::UndefinedBooleanContent:
3778 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3782 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3783 DAG.getConstant(TrueValue, VT), DAG.getConstant(0, VT),
3785 Results.push_back(Tmp1);
3788 case ISD::SELECT_CC: {
3789 Tmp1 = Node->getOperand(0); // LHS
3790 Tmp2 = Node->getOperand(1); // RHS
3791 Tmp3 = Node->getOperand(2); // True
3792 Tmp4 = Node->getOperand(3); // False
3793 SDValue CC = Node->getOperand(4);
3795 bool Legalized = false;
3796 // Try to legalize by inverting the condition. This is for targets that
3797 // might support an ordered version of a condition, but not the unordered
3798 // version (or vice versa).
3799 ISD::CondCode InvCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3800 Tmp1.getValueType().isInteger());
3801 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
3802 // Use the new condition code and swap true and false
3804 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
3806 // If The inverse is not legal, then try to swap the arguments using
3807 // the inverse condition code.
3808 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
3809 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
3810 // The swapped inverse condition is legal, so swap true and false,
3813 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
3818 Legalized = LegalizeSetCCCondCode(
3819 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
3822 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
3824 // If we expanded the SETCC by inverting the condition code, then swap
3825 // the True/False operands to match.
3827 std::swap(Tmp3, Tmp4);
3829 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3830 // condition code, create a new SELECT_CC node.
3832 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
3833 Tmp1, Tmp2, Tmp3, Tmp4, CC);
3835 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
3836 CC = DAG.getCondCode(ISD::SETNE);
3837 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
3838 Tmp2, Tmp3, Tmp4, CC);
3841 Results.push_back(Tmp1);
3845 Tmp1 = Node->getOperand(0); // Chain
3846 Tmp2 = Node->getOperand(2); // LHS
3847 Tmp3 = Node->getOperand(3); // RHS
3848 Tmp4 = Node->getOperand(1); // CC
3850 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
3851 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
3853 assert(Legalized && "Can't legalize BR_CC with legal condition!");
3855 // If we expanded the SETCC by inverting the condition code, then wrap
3856 // the existing SETCC in a NOT to restore the intended condition.
3858 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
3860 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
3862 if (Tmp4.getNode()) {
3863 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
3864 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
3866 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
3867 Tmp4 = DAG.getCondCode(ISD::SETNE);
3868 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
3869 Tmp2, Tmp3, Node->getOperand(4));
3871 Results.push_back(Tmp1);
3874 case ISD::BUILD_VECTOR:
3875 Results.push_back(ExpandBUILD_VECTOR(Node));
3880 // Scalarize vector SRA/SRL/SHL.
3881 EVT VT = Node->getValueType(0);
3882 assert(VT.isVector() && "Unable to legalize non-vector shift");
3883 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
3884 unsigned NumElem = VT.getVectorNumElements();
3886 SmallVector<SDValue, 8> Scalars;
3887 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
3888 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3890 Node->getOperand(0), DAG.getConstant(Idx,
3891 TLI.getVectorIdxTy()));
3892 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3894 Node->getOperand(1), DAG.getConstant(Idx,
3895 TLI.getVectorIdxTy()));
3896 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
3897 VT.getScalarType(), Ex, Sh));
3900 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
3901 &Scalars[0], Scalars.size());
3902 ReplaceNode(SDValue(Node, 0), Result);
3905 case ISD::GLOBAL_OFFSET_TABLE:
3906 case ISD::GlobalAddress:
3907 case ISD::GlobalTLSAddress:
3908 case ISD::ExternalSymbol:
3909 case ISD::ConstantPool:
3910 case ISD::JumpTable:
3911 case ISD::INTRINSIC_W_CHAIN:
3912 case ISD::INTRINSIC_WO_CHAIN:
3913 case ISD::INTRINSIC_VOID:
3914 // FIXME: Custom lowering for these operations shouldn't return null!
3918 // Replace the original node with the legalized result.
3919 if (!Results.empty())
3920 ReplaceNode(Node, Results.data());
3923 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
3924 SmallVector<SDValue, 8> Results;
3925 MVT OVT = Node->getSimpleValueType(0);
3926 if (Node->getOpcode() == ISD::UINT_TO_FP ||
3927 Node->getOpcode() == ISD::SINT_TO_FP ||
3928 Node->getOpcode() == ISD::SETCC) {
3929 OVT = Node->getOperand(0).getSimpleValueType();
3931 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
3933 SDValue Tmp1, Tmp2, Tmp3;
3934 switch (Node->getOpcode()) {
3936 case ISD::CTTZ_ZERO_UNDEF:
3938 case ISD::CTLZ_ZERO_UNDEF:
3940 // Zero extend the argument.
3941 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3942 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
3943 // already the correct result.
3944 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3945 if (Node->getOpcode() == ISD::CTTZ) {
3946 // FIXME: This should set a bit in the zero extended value instead.
3947 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
3948 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
3950 Tmp1 = DAG.getSelect(dl, NVT, Tmp2,
3951 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
3952 } else if (Node->getOpcode() == ISD::CTLZ ||
3953 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
3954 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
3955 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
3956 DAG.getConstant(NVT.getSizeInBits() -
3957 OVT.getSizeInBits(), NVT));
3959 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
3962 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
3963 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3964 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
3965 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
3966 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
3967 Results.push_back(Tmp1);
3970 case ISD::FP_TO_UINT:
3971 case ISD::FP_TO_SINT:
3972 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
3973 Node->getOpcode() == ISD::FP_TO_SINT, dl);
3974 Results.push_back(Tmp1);
3976 case ISD::UINT_TO_FP:
3977 case ISD::SINT_TO_FP:
3978 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
3979 Node->getOpcode() == ISD::SINT_TO_FP, dl);
3980 Results.push_back(Tmp1);
3983 SDValue Chain = Node->getOperand(0); // Get the chain.
3984 SDValue Ptr = Node->getOperand(1); // Get the pointer.
3987 if (OVT.isVector()) {
3988 TruncOp = ISD::BITCAST;
3990 assert(OVT.isInteger()
3991 && "VAARG promotion is supported only for vectors or integer types");
3992 TruncOp = ISD::TRUNCATE;
3995 // Perform the larger operation, then convert back
3996 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
3997 Node->getConstantOperandVal(3));
3998 Chain = Tmp1.getValue(1);
4000 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4002 // Modified the chain result - switch anything that used the old chain to
4004 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4005 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4012 unsigned ExtOp, TruncOp;
4013 if (OVT.isVector()) {
4014 ExtOp = ISD::BITCAST;
4015 TruncOp = ISD::BITCAST;
4017 assert(OVT.isInteger() && "Cannot promote logic operation");
4018 ExtOp = ISD::ANY_EXTEND;
4019 TruncOp = ISD::TRUNCATE;
4021 // Promote each of the values to the new type.
4022 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4023 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4024 // Perform the larger operation, then convert back
4025 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4026 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4030 unsigned ExtOp, TruncOp;
4031 if (Node->getValueType(0).isVector() ||
4032 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4033 ExtOp = ISD::BITCAST;
4034 TruncOp = ISD::BITCAST;
4035 } else if (Node->getValueType(0).isInteger()) {
4036 ExtOp = ISD::ANY_EXTEND;
4037 TruncOp = ISD::TRUNCATE;
4039 ExtOp = ISD::FP_EXTEND;
4040 TruncOp = ISD::FP_ROUND;
4042 Tmp1 = Node->getOperand(0);
4043 // Promote each of the values to the new type.
4044 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4045 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4046 // Perform the larger operation, then round down.
4047 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4048 if (TruncOp != ISD::FP_ROUND)
4049 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4051 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4052 DAG.getIntPtrConstant(0));
4053 Results.push_back(Tmp1);
4056 case ISD::VECTOR_SHUFFLE: {
4057 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4059 // Cast the two input vectors.
4060 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4061 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4063 // Convert the shuffle mask to the right # elements.
4064 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4065 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4066 Results.push_back(Tmp1);
4070 unsigned ExtOp = ISD::FP_EXTEND;
4071 if (NVT.isInteger()) {
4072 ISD::CondCode CCCode =
4073 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4074 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4076 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4077 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4078 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4079 Tmp1, Tmp2, Node->getOperand(2)));
4085 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4086 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4087 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4088 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4089 Tmp3, DAG.getIntPtrConstant(0)));
4096 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4097 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4098 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4099 Tmp2, DAG.getIntPtrConstant(0)));
4104 // Replace the original node with the legalized result.
4105 if (!Results.empty())
4106 ReplaceNode(Node, Results.data());
4109 // SelectionDAG::Legalize - This is the entry point for the file.
4111 void SelectionDAG::Legalize() {
4112 /// run - This is the main entry point to this class.
4114 SelectionDAGLegalize(*this).LegalizeDAG();