1 //===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SelectionDAG::Legalize method.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineJumpTableInfo.h"
22 #include "llvm/DebugInfo.h"
23 #include "llvm/IR/CallingConv.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.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 // Libcall insertion helpers.
65 explicit SelectionDAGLegalize(SelectionDAG &DAG);
70 /// LegalizeOp - Legalizes the given operation.
71 void LegalizeOp(SDNode *Node);
73 SDValue OptimizeFloatStore(StoreSDNode *ST);
75 void LegalizeLoadOps(SDNode *Node);
76 void LegalizeStoreOps(SDNode *Node);
78 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
79 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
80 /// is necessary to spill the vector being inserted into to memory, perform
81 /// the insert there, and then read the result back.
82 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
83 SDValue Idx, DebugLoc dl);
84 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
85 SDValue Idx, DebugLoc dl);
87 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
88 /// performs the same shuffe in terms of order or result bytes, but on a type
89 /// whose vector element type is narrower than the original shuffle type.
90 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
91 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
92 SDValue N1, SDValue N2,
93 ArrayRef<int> Mask) const;
95 void LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
98 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
99 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
100 unsigned NumOps, bool isSigned, DebugLoc dl);
102 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
103 SDNode *Node, bool isSigned);
104 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
105 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
106 RTLIB::Libcall Call_F128,
107 RTLIB::Libcall Call_PPCF128);
108 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
109 RTLIB::Libcall Call_I8,
110 RTLIB::Libcall Call_I16,
111 RTLIB::Libcall Call_I32,
112 RTLIB::Libcall Call_I64,
113 RTLIB::Libcall Call_I128);
114 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
115 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
117 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, DebugLoc dl);
118 SDValue ExpandBUILD_VECTOR(SDNode *Node);
119 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
120 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
121 SmallVectorImpl<SDValue> &Results);
122 SDValue ExpandFCOPYSIGN(SDNode *Node);
123 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
125 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
127 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
130 SDValue ExpandBSWAP(SDValue Op, DebugLoc dl);
131 SDValue ExpandBitCount(unsigned Opc, SDValue Op, DebugLoc dl);
133 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
134 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
135 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
137 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
139 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
141 void ExpandNode(SDNode *Node);
142 void PromoteNode(SDNode *Node);
144 void ForgetNode(SDNode *N) {
145 LegalizedNodes.erase(N);
146 if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
151 // DAGUpdateListener implementation.
152 virtual void NodeDeleted(SDNode *N, SDNode *E) {
155 virtual void NodeUpdated(SDNode *N) {}
157 // Node replacement helpers
158 void ReplacedNode(SDNode *N) {
159 if (N->use_empty()) {
160 DAG.RemoveDeadNode(N);
165 void ReplaceNode(SDNode *Old, SDNode *New) {
166 DAG.ReplaceAllUsesWith(Old, New);
169 void ReplaceNode(SDValue Old, SDValue New) {
170 DAG.ReplaceAllUsesWith(Old, New);
171 ReplacedNode(Old.getNode());
173 void ReplaceNode(SDNode *Old, const SDValue *New) {
174 DAG.ReplaceAllUsesWith(Old, New);
180 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
181 /// performs the same shuffe in terms of order or result bytes, but on a type
182 /// whose vector element type is narrower than the original shuffle type.
183 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
185 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
186 SDValue N1, SDValue N2,
187 ArrayRef<int> Mask) const {
188 unsigned NumMaskElts = VT.getVectorNumElements();
189 unsigned NumDestElts = NVT.getVectorNumElements();
190 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
192 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
194 if (NumEltsGrowth == 1)
195 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
197 SmallVector<int, 8> NewMask;
198 for (unsigned i = 0; i != NumMaskElts; ++i) {
200 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
202 NewMask.push_back(-1);
204 NewMask.push_back(Idx * NumEltsGrowth + j);
207 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
208 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
209 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
212 SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
213 : SelectionDAG::DAGUpdateListener(dag),
214 TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
218 void SelectionDAGLegalize::LegalizeDAG() {
219 DAG.AssignTopologicalOrder();
221 // Visit all the nodes. We start in topological order, so that we see
222 // nodes with their original operands intact. Legalization can produce
223 // new nodes which may themselves need to be legalized. Iterate until all
224 // nodes have been legalized.
226 bool AnyLegalized = false;
227 for (LegalizePosition = DAG.allnodes_end();
228 LegalizePosition != DAG.allnodes_begin(); ) {
231 SDNode *N = LegalizePosition;
232 if (LegalizedNodes.insert(N)) {
242 // Remove dead nodes now.
243 DAG.RemoveDeadNodes();
246 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
247 /// a load from the constant pool.
249 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
251 DebugLoc dl = CFP->getDebugLoc();
253 // If a FP immediate is precise when represented as a float and if the
254 // target can do an extending load from float to double, we put it into
255 // the constant pool as a float, even if it's is statically typed as a
256 // double. This shrinks FP constants and canonicalizes them for targets where
257 // an FP extending load is the same cost as a normal load (such as on the x87
258 // fp stack or PPC FP unit).
259 EVT VT = CFP->getValueType(0);
260 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
262 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
263 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
264 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
269 while (SVT != MVT::f32) {
270 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
271 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
272 // Only do this if the target has a native EXTLOAD instruction from
274 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
275 TLI.ShouldShrinkFPConstant(OrigVT)) {
276 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
277 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
283 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
284 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
287 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
289 CPIdx, MachinePointerInfo::getConstantPool(),
290 VT, false, false, Alignment);
294 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
295 MachinePointerInfo::getConstantPool(), false, false, false,
300 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
301 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
302 const TargetLowering &TLI,
303 SelectionDAGLegalize *DAGLegalize) {
304 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
305 "unaligned indexed stores not implemented!");
306 SDValue Chain = ST->getChain();
307 SDValue Ptr = ST->getBasePtr();
308 SDValue Val = ST->getValue();
309 EVT VT = Val.getValueType();
310 int Alignment = ST->getAlignment();
311 DebugLoc dl = ST->getDebugLoc();
312 if (ST->getMemoryVT().isFloatingPoint() ||
313 ST->getMemoryVT().isVector()) {
314 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
315 if (TLI.isTypeLegal(intVT)) {
316 // Expand to a bitconvert of the value to the integer type of the
317 // same size, then a (misaligned) int store.
318 // FIXME: Does not handle truncating floating point stores!
319 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
320 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
321 ST->isVolatile(), ST->isNonTemporal(), Alignment);
322 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
325 // Do a (aligned) store to a stack slot, then copy from the stack slot
326 // to the final destination using (unaligned) integer loads and stores.
327 EVT StoredVT = ST->getMemoryVT();
329 TLI.getRegisterType(*DAG.getContext(),
330 EVT::getIntegerVT(*DAG.getContext(),
331 StoredVT.getSizeInBits()));
332 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
333 unsigned RegBytes = RegVT.getSizeInBits() / 8;
334 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
336 // Make sure the stack slot is also aligned for the register type.
337 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
339 // Perform the original store, only redirected to the stack slot.
340 SDValue Store = DAG.getTruncStore(Chain, dl,
341 Val, StackPtr, MachinePointerInfo(),
342 StoredVT, false, false, 0);
343 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
344 SmallVector<SDValue, 8> Stores;
347 // Do all but one copies using the full register width.
348 for (unsigned i = 1; i < NumRegs; i++) {
349 // Load one integer register's worth from the stack slot.
350 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
351 MachinePointerInfo(),
352 false, false, false, 0);
353 // Store it to the final location. Remember the store.
354 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
355 ST->getPointerInfo().getWithOffset(Offset),
356 ST->isVolatile(), ST->isNonTemporal(),
357 MinAlign(ST->getAlignment(), Offset)));
358 // Increment the pointers.
360 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
362 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
365 // The last store may be partial. Do a truncating store. On big-endian
366 // machines this requires an extending load from the stack slot to ensure
367 // that the bits are in the right place.
368 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
369 8 * (StoredBytes - Offset));
371 // Load from the stack slot.
372 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
373 MachinePointerInfo(),
374 MemVT, false, false, 0);
376 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
378 .getWithOffset(Offset),
379 MemVT, ST->isVolatile(),
381 MinAlign(ST->getAlignment(), Offset)));
382 // The order of the stores doesn't matter - say it with a TokenFactor.
384 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
386 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
389 assert(ST->getMemoryVT().isInteger() &&
390 !ST->getMemoryVT().isVector() &&
391 "Unaligned store of unknown type.");
392 // Get the half-size VT
393 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
394 int NumBits = NewStoredVT.getSizeInBits();
395 int IncrementSize = NumBits / 8;
397 // Divide the stored value in two parts.
398 SDValue ShiftAmount = DAG.getConstant(NumBits,
399 TLI.getShiftAmountTy(Val.getValueType()));
401 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
403 // Store the two parts
404 SDValue Store1, Store2;
405 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
406 ST->getPointerInfo(), NewStoredVT,
407 ST->isVolatile(), ST->isNonTemporal(), Alignment);
408 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
409 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
410 Alignment = MinAlign(Alignment, IncrementSize);
411 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
412 ST->getPointerInfo().getWithOffset(IncrementSize),
413 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
417 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
418 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
421 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
423 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
424 const TargetLowering &TLI,
425 SDValue &ValResult, SDValue &ChainResult) {
426 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
427 "unaligned indexed loads not implemented!");
428 SDValue Chain = LD->getChain();
429 SDValue Ptr = LD->getBasePtr();
430 EVT VT = LD->getValueType(0);
431 EVT LoadedVT = LD->getMemoryVT();
432 DebugLoc dl = LD->getDebugLoc();
433 if (VT.isFloatingPoint() || VT.isVector()) {
434 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
435 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
436 // Expand to a (misaligned) integer load of the same size,
437 // then bitconvert to floating point or vector.
438 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getPointerInfo(),
441 LD->isInvariant(), LD->getAlignment());
442 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
444 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
445 ISD::ANY_EXTEND, dl, VT, Result);
452 // Copy the value to a (aligned) stack slot using (unaligned) integer
453 // loads and stores, then do a (aligned) load from the stack slot.
454 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
455 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
456 unsigned RegBytes = RegVT.getSizeInBits() / 8;
457 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
459 // Make sure the stack slot is also aligned for the register type.
460 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
462 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
463 SmallVector<SDValue, 8> Stores;
464 SDValue StackPtr = StackBase;
467 // Do all but one copies using the full register width.
468 for (unsigned i = 1; i < NumRegs; i++) {
469 // Load one integer register's worth from the original location.
470 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
471 LD->getPointerInfo().getWithOffset(Offset),
472 LD->isVolatile(), LD->isNonTemporal(),
474 MinAlign(LD->getAlignment(), Offset));
475 // Follow the load with a store to the stack slot. Remember the store.
476 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
477 MachinePointerInfo(), false, false, 0));
478 // Increment the pointers.
480 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
481 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
485 // The last copy may be partial. Do an extending load.
486 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
487 8 * (LoadedBytes - Offset));
488 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
489 LD->getPointerInfo().getWithOffset(Offset),
490 MemVT, LD->isVolatile(),
492 MinAlign(LD->getAlignment(), Offset));
493 // Follow the load with a store to the stack slot. Remember the store.
494 // On big-endian machines this requires a truncating store to ensure
495 // that the bits end up in the right place.
496 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
497 MachinePointerInfo(), MemVT,
500 // The order of the stores doesn't matter - say it with a TokenFactor.
501 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
504 // Finally, perform the original load only redirected to the stack slot.
505 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
506 MachinePointerInfo(), LoadedVT, false, false, 0);
508 // Callers expect a MERGE_VALUES node.
513 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
514 "Unaligned load of unsupported type.");
516 // Compute the new VT that is half the size of the old one. This is an
518 unsigned NumBits = LoadedVT.getSizeInBits();
520 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
523 unsigned Alignment = LD->getAlignment();
524 unsigned IncrementSize = NumBits / 8;
525 ISD::LoadExtType HiExtType = LD->getExtensionType();
527 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
528 if (HiExtType == ISD::NON_EXTLOAD)
529 HiExtType = ISD::ZEXTLOAD;
531 // Load the value in two parts
533 if (TLI.isLittleEndian()) {
534 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
535 NewLoadedVT, LD->isVolatile(),
536 LD->isNonTemporal(), Alignment);
537 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
538 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
539 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
540 LD->getPointerInfo().getWithOffset(IncrementSize),
541 NewLoadedVT, LD->isVolatile(),
542 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
544 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
545 NewLoadedVT, LD->isVolatile(),
546 LD->isNonTemporal(), Alignment);
547 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
548 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
549 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
550 LD->getPointerInfo().getWithOffset(IncrementSize),
551 NewLoadedVT, LD->isVolatile(),
552 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
555 // aggregate the two parts
556 SDValue ShiftAmount = DAG.getConstant(NumBits,
557 TLI.getShiftAmountTy(Hi.getValueType()));
558 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
559 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
561 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
568 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
569 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
570 /// is necessary to spill the vector being inserted into to memory, perform
571 /// the insert there, and then read the result back.
572 SDValue SelectionDAGLegalize::
573 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
579 // If the target doesn't support this, we have to spill the input vector
580 // to a temporary stack slot, update the element, then reload it. This is
581 // badness. We could also load the value into a vector register (either
582 // with a "move to register" or "extload into register" instruction, then
583 // permute it into place, if the idx is a constant and if the idx is
584 // supported by the target.
585 EVT VT = Tmp1.getValueType();
586 EVT EltVT = VT.getVectorElementType();
587 EVT IdxVT = Tmp3.getValueType();
588 EVT PtrVT = TLI.getPointerTy();
589 SDValue StackPtr = DAG.CreateStackTemporary(VT);
591 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
594 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
595 MachinePointerInfo::getFixedStack(SPFI),
598 // Truncate or zero extend offset to target pointer type.
599 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
600 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
601 // Add the offset to the index.
602 unsigned EltSize = EltVT.getSizeInBits()/8;
603 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
604 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
605 // Store the scalar value.
606 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
608 // Load the updated vector.
609 return DAG.getLoad(VT, dl, Ch, StackPtr,
610 MachinePointerInfo::getFixedStack(SPFI), false, false,
615 SDValue SelectionDAGLegalize::
616 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, DebugLoc dl) {
617 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
618 // SCALAR_TO_VECTOR requires that the type of the value being inserted
619 // match the element type of the vector being created, except for
620 // integers in which case the inserted value can be over width.
621 EVT EltVT = Vec.getValueType().getVectorElementType();
622 if (Val.getValueType() == EltVT ||
623 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
624 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
625 Vec.getValueType(), Val);
627 unsigned NumElts = Vec.getValueType().getVectorNumElements();
628 // We generate a shuffle of InVec and ScVec, so the shuffle mask
629 // should be 0,1,2,3,4,5... with the appropriate element replaced with
631 SmallVector<int, 8> ShufOps;
632 for (unsigned i = 0; i != NumElts; ++i)
633 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
635 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
639 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
642 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
643 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
644 // FIXME: We shouldn't do this for TargetConstantFP's.
645 // FIXME: move this to the DAG Combiner! Note that we can't regress due
646 // to phase ordering between legalized code and the dag combiner. This
647 // probably means that we need to integrate dag combiner and legalizer
649 // We generally can't do this one for long doubles.
650 SDValue Chain = ST->getChain();
651 SDValue Ptr = ST->getBasePtr();
652 unsigned Alignment = ST->getAlignment();
653 bool isVolatile = ST->isVolatile();
654 bool isNonTemporal = ST->isNonTemporal();
655 DebugLoc dl = ST->getDebugLoc();
656 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
657 if (CFP->getValueType(0) == MVT::f32 &&
658 TLI.isTypeLegal(MVT::i32)) {
659 SDValue Con = DAG.getConstant(CFP->getValueAPF().
660 bitcastToAPInt().zextOrTrunc(32),
662 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
663 isVolatile, isNonTemporal, Alignment);
666 if (CFP->getValueType(0) == MVT::f64) {
667 // If this target supports 64-bit registers, do a single 64-bit store.
668 if (TLI.isTypeLegal(MVT::i64)) {
669 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
670 zextOrTrunc(64), MVT::i64);
671 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
672 isVolatile, isNonTemporal, Alignment);
675 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
676 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
677 // stores. If the target supports neither 32- nor 64-bits, this
678 // xform is certainly not worth it.
679 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
680 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
681 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
682 if (TLI.isBigEndian()) std::swap(Lo, Hi);
684 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
685 isNonTemporal, Alignment);
686 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
687 DAG.getIntPtrConstant(4));
688 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
689 ST->getPointerInfo().getWithOffset(4),
690 isVolatile, isNonTemporal, MinAlign(Alignment, 4U));
692 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
696 return SDValue(0, 0);
699 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
700 StoreSDNode *ST = cast<StoreSDNode>(Node);
701 SDValue Chain = ST->getChain();
702 SDValue Ptr = ST->getBasePtr();
703 DebugLoc dl = Node->getDebugLoc();
705 unsigned Alignment = ST->getAlignment();
706 bool isVolatile = ST->isVolatile();
707 bool isNonTemporal = ST->isNonTemporal();
709 if (!ST->isTruncatingStore()) {
710 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
711 ReplaceNode(ST, OptStore);
716 SDValue Value = ST->getValue();
717 MVT VT = Value.getSimpleValueType();
718 switch (TLI.getOperationAction(ISD::STORE, VT)) {
719 default: llvm_unreachable("This action is not supported yet!");
720 case TargetLowering::Legal:
721 // If this is an unaligned store and the target doesn't support it,
723 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
724 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
725 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
726 if (ST->getAlignment() < ABIAlignment)
727 ExpandUnalignedStore(cast<StoreSDNode>(Node),
731 case TargetLowering::Custom: {
732 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
734 ReplaceNode(SDValue(Node, 0), Res);
737 case TargetLowering::Promote: {
738 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
739 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
740 "Can only promote stores to same size type");
741 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
743 DAG.getStore(Chain, dl, Value, Ptr,
744 ST->getPointerInfo(), isVolatile,
745 isNonTemporal, Alignment);
746 ReplaceNode(SDValue(Node, 0), Result);
753 SDValue Value = ST->getValue();
755 EVT StVT = ST->getMemoryVT();
756 unsigned StWidth = StVT.getSizeInBits();
758 if (StWidth != StVT.getStoreSizeInBits()) {
759 // Promote to a byte-sized store with upper bits zero if not
760 // storing an integral number of bytes. For example, promote
761 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
762 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
763 StVT.getStoreSizeInBits());
764 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
766 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
767 NVT, isVolatile, isNonTemporal, Alignment);
768 ReplaceNode(SDValue(Node, 0), Result);
769 } else if (StWidth & (StWidth - 1)) {
770 // If not storing a power-of-2 number of bits, expand as two stores.
771 assert(!StVT.isVector() && "Unsupported truncstore!");
772 unsigned RoundWidth = 1 << Log2_32(StWidth);
773 assert(RoundWidth < StWidth);
774 unsigned ExtraWidth = StWidth - RoundWidth;
775 assert(ExtraWidth < RoundWidth);
776 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
777 "Store size not an integral number of bytes!");
778 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
779 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
781 unsigned IncrementSize;
783 if (TLI.isLittleEndian()) {
784 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
785 // Store the bottom RoundWidth bits.
786 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
788 isVolatile, isNonTemporal, Alignment);
790 // Store the remaining ExtraWidth bits.
791 IncrementSize = RoundWidth / 8;
792 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
793 DAG.getIntPtrConstant(IncrementSize));
794 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
795 DAG.getConstant(RoundWidth,
796 TLI.getShiftAmountTy(Value.getValueType())));
797 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
798 ST->getPointerInfo().getWithOffset(IncrementSize),
799 ExtraVT, isVolatile, isNonTemporal,
800 MinAlign(Alignment, IncrementSize));
802 // Big endian - avoid unaligned stores.
803 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
804 // Store the top RoundWidth bits.
805 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
806 DAG.getConstant(ExtraWidth,
807 TLI.getShiftAmountTy(Value.getValueType())));
808 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
809 RoundVT, isVolatile, isNonTemporal, Alignment);
811 // Store the remaining ExtraWidth bits.
812 IncrementSize = RoundWidth / 8;
813 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
814 DAG.getIntPtrConstant(IncrementSize));
815 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
816 ST->getPointerInfo().getWithOffset(IncrementSize),
817 ExtraVT, isVolatile, isNonTemporal,
818 MinAlign(Alignment, IncrementSize));
821 // The order of the stores doesn't matter.
822 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
823 ReplaceNode(SDValue(Node, 0), Result);
825 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
826 StVT.getSimpleVT())) {
827 default: llvm_unreachable("This action is not supported yet!");
828 case TargetLowering::Legal:
829 // If this is an unaligned store and the target doesn't support it,
831 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
832 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
833 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
834 if (ST->getAlignment() < ABIAlignment)
835 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
838 case TargetLowering::Custom: {
839 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
841 ReplaceNode(SDValue(Node, 0), Res);
844 case TargetLowering::Expand:
845 assert(!StVT.isVector() &&
846 "Vector Stores are handled in LegalizeVectorOps");
848 // TRUNCSTORE:i16 i32 -> STORE i16
849 assert(TLI.isTypeLegal(StVT) &&
850 "Do not know how to expand this store!");
851 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
853 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
854 isVolatile, isNonTemporal, Alignment);
855 ReplaceNode(SDValue(Node, 0), Result);
862 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
863 LoadSDNode *LD = cast<LoadSDNode>(Node);
864 SDValue Chain = LD->getChain(); // The chain.
865 SDValue Ptr = LD->getBasePtr(); // The base pointer.
866 SDValue Value; // The value returned by the load op.
867 DebugLoc dl = Node->getDebugLoc();
869 ISD::LoadExtType ExtType = LD->getExtensionType();
870 if (ExtType == ISD::NON_EXTLOAD) {
871 MVT VT = Node->getSimpleValueType(0);
872 SDValue RVal = SDValue(Node, 0);
873 SDValue RChain = SDValue(Node, 1);
875 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
876 default: llvm_unreachable("This action is not supported yet!");
877 case TargetLowering::Legal:
878 // If this is an unaligned load and the target doesn't support it,
880 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
881 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
882 unsigned ABIAlignment =
883 TLI.getDataLayout()->getABITypeAlignment(Ty);
884 if (LD->getAlignment() < ABIAlignment){
885 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
889 case TargetLowering::Custom: {
890 SDValue Res = TLI.LowerOperation(RVal, DAG);
893 RChain = Res.getValue(1);
897 case TargetLowering::Promote: {
898 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
899 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
900 "Can only promote loads to same size type");
902 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getPointerInfo(),
903 LD->isVolatile(), LD->isNonTemporal(),
904 LD->isInvariant(), LD->getAlignment());
905 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
906 RChain = Res.getValue(1);
910 if (RChain.getNode() != Node) {
911 assert(RVal.getNode() != Node && "Load must be completely replaced");
912 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
913 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
919 EVT SrcVT = LD->getMemoryVT();
920 unsigned SrcWidth = SrcVT.getSizeInBits();
921 unsigned Alignment = LD->getAlignment();
922 bool isVolatile = LD->isVolatile();
923 bool isNonTemporal = LD->isNonTemporal();
925 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
926 // Some targets pretend to have an i1 loading operation, and actually
927 // load an i8. This trick is correct for ZEXTLOAD because the top 7
928 // bits are guaranteed to be zero; it helps the optimizers understand
929 // that these bits are zero. It is also useful for EXTLOAD, since it
930 // tells the optimizers that those bits are undefined. It would be
931 // nice to have an effective generic way of getting these benefits...
932 // Until such a way is found, don't insist on promoting i1 here.
934 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
935 // Promote to a byte-sized load if not loading an integral number of
936 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
937 unsigned NewWidth = SrcVT.getStoreSizeInBits();
938 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
941 // The extra bits are guaranteed to be zero, since we stored them that
942 // way. A zext load from NVT thus automatically gives zext from SrcVT.
944 ISD::LoadExtType NewExtType =
945 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
948 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
949 Chain, Ptr, LD->getPointerInfo(),
950 NVT, isVolatile, isNonTemporal, Alignment);
952 Ch = Result.getValue(1); // The chain.
954 if (ExtType == ISD::SEXTLOAD)
955 // Having the top bits zero doesn't help when sign extending.
956 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
957 Result.getValueType(),
958 Result, DAG.getValueType(SrcVT));
959 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
960 // All the top bits are guaranteed to be zero - inform the optimizers.
961 Result = DAG.getNode(ISD::AssertZext, dl,
962 Result.getValueType(), Result,
963 DAG.getValueType(SrcVT));
967 } else if (SrcWidth & (SrcWidth - 1)) {
968 // If not loading a power-of-2 number of bits, expand as two loads.
969 assert(!SrcVT.isVector() && "Unsupported extload!");
970 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
971 assert(RoundWidth < SrcWidth);
972 unsigned ExtraWidth = SrcWidth - RoundWidth;
973 assert(ExtraWidth < RoundWidth);
974 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
975 "Load size not an integral number of bytes!");
976 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
977 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
979 unsigned IncrementSize;
981 if (TLI.isLittleEndian()) {
982 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
983 // Load the bottom RoundWidth bits.
984 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
986 LD->getPointerInfo(), RoundVT, isVolatile,
987 isNonTemporal, Alignment);
989 // Load the remaining ExtraWidth bits.
990 IncrementSize = RoundWidth / 8;
991 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
992 DAG.getIntPtrConstant(IncrementSize));
993 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
994 LD->getPointerInfo().getWithOffset(IncrementSize),
995 ExtraVT, isVolatile, isNonTemporal,
996 MinAlign(Alignment, IncrementSize));
998 // Build a factor node to remember that this load is independent of
1000 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1003 // Move the top bits to the right place.
1004 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1005 DAG.getConstant(RoundWidth,
1006 TLI.getShiftAmountTy(Hi.getValueType())));
1008 // Join the hi and lo parts.
1009 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1011 // Big endian - avoid unaligned loads.
1012 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1013 // Load the top RoundWidth bits.
1014 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1015 LD->getPointerInfo(), RoundVT, isVolatile,
1016 isNonTemporal, Alignment);
1018 // Load the remaining ExtraWidth bits.
1019 IncrementSize = RoundWidth / 8;
1020 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1021 DAG.getIntPtrConstant(IncrementSize));
1022 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1023 dl, Node->getValueType(0), Chain, Ptr,
1024 LD->getPointerInfo().getWithOffset(IncrementSize),
1025 ExtraVT, isVolatile, isNonTemporal,
1026 MinAlign(Alignment, IncrementSize));
1028 // Build a factor node to remember that this load is independent of
1030 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1033 // Move the top bits to the right place.
1034 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1035 DAG.getConstant(ExtraWidth,
1036 TLI.getShiftAmountTy(Hi.getValueType())));
1038 // Join the hi and lo parts.
1039 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1044 bool isCustom = false;
1045 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1046 default: llvm_unreachable("This action is not supported yet!");
1047 case TargetLowering::Custom:
1050 case TargetLowering::Legal: {
1051 Value = SDValue(Node, 0);
1052 Chain = SDValue(Node, 1);
1055 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1056 if (Res.getNode()) {
1058 Chain = Res.getValue(1);
1061 // If this is an unaligned load and the target doesn't support it,
1063 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
1065 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1066 unsigned ABIAlignment =
1067 TLI.getDataLayout()->getABITypeAlignment(Ty);
1068 if (LD->getAlignment() < ABIAlignment){
1069 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1070 DAG, TLI, Value, Chain);
1076 case TargetLowering::Expand:
1077 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
1078 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1079 LD->getPointerInfo(),
1080 LD->isVolatile(), LD->isNonTemporal(),
1081 LD->isInvariant(), LD->getAlignment());
1085 ExtendOp = (SrcVT.isFloatingPoint() ?
1086 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1088 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1089 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1090 default: llvm_unreachable("Unexpected extend load type!");
1092 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1093 Chain = Load.getValue(1);
1097 assert(!SrcVT.isVector() &&
1098 "Vector Loads are handled in LegalizeVectorOps");
1100 // FIXME: This does not work for vectors on most targets. Sign- and
1101 // zero-extend operations are currently folded into extending loads,
1102 // whether they are legal or not, and then we end up here without any
1103 // support for legalizing them.
1104 assert(ExtType != ISD::EXTLOAD &&
1105 "EXTLOAD should always be supported!");
1106 // Turn the unsupported load into an EXTLOAD followed by an explicit
1107 // zero/sign extend inreg.
1108 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
1109 Chain, Ptr, LD->getPointerInfo(), SrcVT,
1110 LD->isVolatile(), LD->isNonTemporal(),
1111 LD->getAlignment());
1113 if (ExtType == ISD::SEXTLOAD)
1114 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1115 Result.getValueType(),
1116 Result, DAG.getValueType(SrcVT));
1118 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1120 Chain = Result.getValue(1);
1125 // Since loads produce two values, make sure to remember that we legalized
1127 if (Chain.getNode() != Node) {
1128 assert(Value.getNode() != Node && "Load must be completely replaced");
1129 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1130 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1135 /// LegalizeOp - Return a legal replacement for the given operation, with
1136 /// all legal operands.
1137 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1138 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1141 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1142 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1143 TargetLowering::TypeLegal &&
1144 "Unexpected illegal type!");
1146 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1147 assert((TLI.getTypeAction(*DAG.getContext(),
1148 Node->getOperand(i).getValueType()) ==
1149 TargetLowering::TypeLegal ||
1150 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1151 "Unexpected illegal type!");
1153 // Figure out the correct action; the way to query this varies by opcode
1154 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1155 bool SimpleFinishLegalizing = true;
1156 switch (Node->getOpcode()) {
1157 case ISD::INTRINSIC_W_CHAIN:
1158 case ISD::INTRINSIC_WO_CHAIN:
1159 case ISD::INTRINSIC_VOID:
1160 case ISD::STACKSAVE:
1161 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1164 Action = TLI.getOperationAction(Node->getOpcode(),
1165 Node->getValueType(0));
1166 if (Action != TargetLowering::Promote)
1167 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1169 case ISD::SINT_TO_FP:
1170 case ISD::UINT_TO_FP:
1171 case ISD::EXTRACT_VECTOR_ELT:
1172 Action = TLI.getOperationAction(Node->getOpcode(),
1173 Node->getOperand(0).getValueType());
1175 case ISD::FP_ROUND_INREG:
1176 case ISD::SIGN_EXTEND_INREG: {
1177 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1178 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1181 case ISD::ATOMIC_STORE: {
1182 Action = TLI.getOperationAction(Node->getOpcode(),
1183 Node->getOperand(2).getValueType());
1186 case ISD::SELECT_CC:
1189 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1190 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1191 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1192 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1193 ISD::CondCode CCCode =
1194 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1195 Action = TLI.getCondCodeAction(CCCode, OpVT);
1196 if (Action == TargetLowering::Legal) {
1197 if (Node->getOpcode() == ISD::SELECT_CC)
1198 Action = TLI.getOperationAction(Node->getOpcode(),
1199 Node->getValueType(0));
1201 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1207 // FIXME: Model these properly. LOAD and STORE are complicated, and
1208 // STORE expects the unlegalized operand in some cases.
1209 SimpleFinishLegalizing = false;
1211 case ISD::CALLSEQ_START:
1212 case ISD::CALLSEQ_END:
1213 // FIXME: This shouldn't be necessary. These nodes have special properties
1214 // dealing with the recursive nature of legalization. Removing this
1215 // special case should be done as part of making LegalizeDAG non-recursive.
1216 SimpleFinishLegalizing = false;
1218 case ISD::EXTRACT_ELEMENT:
1219 case ISD::FLT_ROUNDS_:
1227 case ISD::MERGE_VALUES:
1228 case ISD::EH_RETURN:
1229 case ISD::FRAME_TO_ARGS_OFFSET:
1230 case ISD::EH_SJLJ_SETJMP:
1231 case ISD::EH_SJLJ_LONGJMP:
1232 // These operations lie about being legal: when they claim to be legal,
1233 // they should actually be expanded.
1234 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1235 if (Action == TargetLowering::Legal)
1236 Action = TargetLowering::Expand;
1238 case ISD::INIT_TRAMPOLINE:
1239 case ISD::ADJUST_TRAMPOLINE:
1240 case ISD::FRAMEADDR:
1241 case ISD::RETURNADDR:
1242 // These operations lie about being legal: when they claim to be legal,
1243 // they should actually be custom-lowered.
1244 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1245 if (Action == TargetLowering::Legal)
1246 Action = TargetLowering::Custom;
1248 case ISD::DEBUGTRAP:
1249 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1250 if (Action == TargetLowering::Expand) {
1251 // replace ISD::DEBUGTRAP with ISD::TRAP
1253 NewVal = DAG.getNode(ISD::TRAP, Node->getDebugLoc(), Node->getVTList(),
1254 Node->getOperand(0));
1255 ReplaceNode(Node, NewVal.getNode());
1256 LegalizeOp(NewVal.getNode());
1262 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1263 Action = TargetLowering::Legal;
1265 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1270 if (SimpleFinishLegalizing) {
1271 SDNode *NewNode = Node;
1272 switch (Node->getOpcode()) {
1279 // Legalizing shifts/rotates requires adjusting the shift amount
1280 // to the appropriate width.
1281 if (!Node->getOperand(1).getValueType().isVector()) {
1283 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1284 Node->getOperand(1));
1285 HandleSDNode Handle(SAO);
1286 LegalizeOp(SAO.getNode());
1287 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1291 case ISD::SRL_PARTS:
1292 case ISD::SRA_PARTS:
1293 case ISD::SHL_PARTS:
1294 // Legalizing shifts/rotates requires adjusting the shift amount
1295 // to the appropriate width.
1296 if (!Node->getOperand(2).getValueType().isVector()) {
1298 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1299 Node->getOperand(2));
1300 HandleSDNode Handle(SAO);
1301 LegalizeOp(SAO.getNode());
1302 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1303 Node->getOperand(1),
1309 if (NewNode != Node) {
1310 DAG.ReplaceAllUsesWith(Node, NewNode);
1311 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1312 DAG.TransferDbgValues(SDValue(Node, i), SDValue(NewNode, i));
1317 case TargetLowering::Legal:
1319 case TargetLowering::Custom: {
1320 // FIXME: The handling for custom lowering with multiple results is
1322 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1323 if (Res.getNode()) {
1324 SmallVector<SDValue, 8> ResultVals;
1325 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
1327 ResultVals.push_back(Res);
1329 ResultVals.push_back(Res.getValue(i));
1331 if (Res.getNode() != Node || Res.getResNo() != 0) {
1332 DAG.ReplaceAllUsesWith(Node, ResultVals.data());
1333 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1334 DAG.TransferDbgValues(SDValue(Node, i), ResultVals[i]);
1341 case TargetLowering::Expand:
1344 case TargetLowering::Promote:
1350 switch (Node->getOpcode()) {
1357 llvm_unreachable("Do not know how to legalize this operator!");
1359 case ISD::CALLSEQ_START:
1360 case ISD::CALLSEQ_END:
1363 return LegalizeLoadOps(Node);
1366 return LegalizeStoreOps(Node);
1371 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1372 SDValue Vec = Op.getOperand(0);
1373 SDValue Idx = Op.getOperand(1);
1374 DebugLoc dl = Op.getDebugLoc();
1375 // Store the value to a temporary stack slot, then LOAD the returned part.
1376 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1377 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1378 MachinePointerInfo(), false, false, 0);
1380 // Add the offset to the index.
1382 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1383 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1384 DAG.getConstant(EltSize, Idx.getValueType()));
1386 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1387 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1389 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1391 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1393 if (Op.getValueType().isVector())
1394 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1395 false, false, false, 0);
1396 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1397 MachinePointerInfo(),
1398 Vec.getValueType().getVectorElementType(),
1402 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1403 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1405 SDValue Vec = Op.getOperand(0);
1406 SDValue Part = Op.getOperand(1);
1407 SDValue Idx = Op.getOperand(2);
1408 DebugLoc dl = Op.getDebugLoc();
1410 // Store the value to a temporary stack slot, then LOAD the returned part.
1412 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1413 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1414 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1416 // First store the whole vector.
1417 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1420 // Then store the inserted part.
1422 // Add the offset to the index.
1424 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1426 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1427 DAG.getConstant(EltSize, Idx.getValueType()));
1429 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1430 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1432 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1434 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1437 // Store the subvector.
1438 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1439 MachinePointerInfo(), false, false, 0);
1441 // Finally, load the updated vector.
1442 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1443 false, false, false, 0);
1446 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1447 // We can't handle this case efficiently. Allocate a sufficiently
1448 // aligned object on the stack, store each element into it, then load
1449 // the result as a vector.
1450 // Create the stack frame object.
1451 EVT VT = Node->getValueType(0);
1452 EVT EltVT = VT.getVectorElementType();
1453 DebugLoc dl = Node->getDebugLoc();
1454 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1455 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1456 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1458 // Emit a store of each element to the stack slot.
1459 SmallVector<SDValue, 8> Stores;
1460 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1461 // Store (in the right endianness) the elements to memory.
1462 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1463 // Ignore undef elements.
1464 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1466 unsigned Offset = TypeByteSize*i;
1468 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1469 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1471 // If the destination vector element type is narrower than the source
1472 // element type, only store the bits necessary.
1473 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1474 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1475 Node->getOperand(i), Idx,
1476 PtrInfo.getWithOffset(Offset),
1477 EltVT, false, false, 0));
1479 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1480 Node->getOperand(i), Idx,
1481 PtrInfo.getWithOffset(Offset),
1486 if (!Stores.empty()) // Not all undef elements?
1487 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1488 &Stores[0], Stores.size());
1490 StoreChain = DAG.getEntryNode();
1492 // Result is a load from the stack slot.
1493 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1494 false, false, false, 0);
1497 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1498 DebugLoc dl = Node->getDebugLoc();
1499 SDValue Tmp1 = Node->getOperand(0);
1500 SDValue Tmp2 = Node->getOperand(1);
1502 // Get the sign bit of the RHS. First obtain a value that has the same
1503 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1505 EVT FloatVT = Tmp2.getValueType();
1506 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1507 if (TLI.isTypeLegal(IVT)) {
1508 // Convert to an integer with the same sign bit.
1509 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1511 // Store the float to memory, then load the sign part out as an integer.
1512 MVT LoadTy = TLI.getPointerTy();
1513 // First create a temporary that is aligned for both the load and store.
1514 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1515 // Then store the float to it.
1517 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1519 if (TLI.isBigEndian()) {
1520 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1521 // Load out a legal integer with the same sign bit as the float.
1522 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1523 false, false, false, 0);
1524 } else { // Little endian
1525 SDValue LoadPtr = StackPtr;
1526 // The float may be wider than the integer we are going to load. Advance
1527 // the pointer so that the loaded integer will contain the sign bit.
1528 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1529 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1530 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(),
1531 LoadPtr, DAG.getIntPtrConstant(ByteOffset));
1532 // Load a legal integer containing the sign bit.
1533 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1534 false, false, false, 0);
1535 // Move the sign bit to the top bit of the loaded integer.
1536 unsigned BitShift = LoadTy.getSizeInBits() -
1537 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1538 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1540 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1541 DAG.getConstant(BitShift,
1542 TLI.getShiftAmountTy(SignBit.getValueType())));
1545 // Now get the sign bit proper, by seeing whether the value is negative.
1546 SignBit = DAG.getSetCC(dl, TLI.getSetCCResultType(SignBit.getValueType()),
1547 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1549 // Get the absolute value of the result.
1550 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1551 // Select between the nabs and abs value based on the sign bit of
1553 return DAG.getNode(ISD::SELECT, dl, AbsVal.getValueType(), SignBit,
1554 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1558 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1559 SmallVectorImpl<SDValue> &Results) {
1560 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1561 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1562 " not tell us which reg is the stack pointer!");
1563 DebugLoc dl = Node->getDebugLoc();
1564 EVT VT = Node->getValueType(0);
1565 SDValue Tmp1 = SDValue(Node, 0);
1566 SDValue Tmp2 = SDValue(Node, 1);
1567 SDValue Tmp3 = Node->getOperand(2);
1568 SDValue Chain = Tmp1.getOperand(0);
1570 // Chain the dynamic stack allocation so that it doesn't modify the stack
1571 // pointer when other instructions are using the stack.
1572 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1574 SDValue Size = Tmp2.getOperand(1);
1575 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1576 Chain = SP.getValue(1);
1577 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1578 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1579 if (Align > StackAlign)
1580 SP = DAG.getNode(ISD::AND, dl, VT, SP,
1581 DAG.getConstant(-(uint64_t)Align, VT));
1582 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1583 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1585 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1586 DAG.getIntPtrConstant(0, true), SDValue());
1588 Results.push_back(Tmp1);
1589 Results.push_back(Tmp2);
1592 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1593 /// condition code CC on the current target. This routine expands SETCC with
1594 /// illegal condition code into AND / OR of multiple SETCC values.
1595 void SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1596 SDValue &LHS, SDValue &RHS,
1599 MVT OpVT = LHS.getSimpleValueType();
1600 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1601 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1602 default: llvm_unreachable("Unknown condition code action!");
1603 case TargetLowering::Legal:
1606 case TargetLowering::Expand: {
1607 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1608 ISD::CondCode InvCC = ISD::SETCC_INVALID;
1611 default: llvm_unreachable("Don't know how to expand this condition!");
1613 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1614 == TargetLowering::Legal
1615 && "If SETO is expanded, SETOEQ must be legal!");
1616 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1618 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1619 == TargetLowering::Legal
1620 && "If SETUO is expanded, SETUNE must be legal!");
1621 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1634 // If we are floating point, assign and break, otherwise fall through.
1635 if (!OpVT.isInteger()) {
1636 // We can use the 4th bit to tell if we are the unordered
1637 // or ordered version of the opcode.
1638 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1639 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1640 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1643 // Fallthrough if we are unsigned integer.
1650 InvCC = ISD::getSetCCSwappedOperands(CCCode);
1651 if (TLI.getCondCodeAction(InvCC, OpVT) == TargetLowering::Expand) {
1652 // We only support using the inverted operation and not a
1653 // different manner of supporting expanding these cases.
1654 llvm_unreachable("Don't know how to expand this condition!");
1656 LHS = DAG.getSetCC(dl, VT, RHS, LHS, InvCC);
1662 SDValue SetCC1, SetCC2;
1663 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1664 // If we aren't the ordered or unorder operation,
1665 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1666 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1667 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1669 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1670 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1671 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1673 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1681 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1682 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1683 /// a load from the stack slot to DestVT, extending it if needed.
1684 /// The resultant code need not be legal.
1685 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1689 // Create the stack frame object.
1691 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1692 getTypeForEVT(*DAG.getContext()));
1693 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1695 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1696 int SPFI = StackPtrFI->getIndex();
1697 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1699 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1700 unsigned SlotSize = SlotVT.getSizeInBits();
1701 unsigned DestSize = DestVT.getSizeInBits();
1702 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1703 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1705 // Emit a store to the stack slot. Use a truncstore if the input value is
1706 // later than DestVT.
1709 if (SrcSize > SlotSize)
1710 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1711 PtrInfo, SlotVT, false, false, SrcAlign);
1713 assert(SrcSize == SlotSize && "Invalid store");
1714 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1715 PtrInfo, false, false, SrcAlign);
1718 // Result is a load from the stack slot.
1719 if (SlotSize == DestSize)
1720 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1721 false, false, false, DestAlign);
1723 assert(SlotSize < DestSize && "Unknown extension!");
1724 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1725 PtrInfo, SlotVT, false, false, DestAlign);
1728 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1729 DebugLoc dl = Node->getDebugLoc();
1730 // Create a vector sized/aligned stack slot, store the value to element #0,
1731 // then load the whole vector back out.
1732 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1734 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1735 int SPFI = StackPtrFI->getIndex();
1737 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1739 MachinePointerInfo::getFixedStack(SPFI),
1740 Node->getValueType(0).getVectorElementType(),
1742 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1743 MachinePointerInfo::getFixedStack(SPFI),
1744 false, false, false, 0);
1748 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1749 /// support the operation, but do support the resultant vector type.
1750 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1751 unsigned NumElems = Node->getNumOperands();
1752 SDValue Value1, Value2;
1753 DebugLoc dl = Node->getDebugLoc();
1754 EVT VT = Node->getValueType(0);
1755 EVT OpVT = Node->getOperand(0).getValueType();
1756 EVT EltVT = VT.getVectorElementType();
1758 // If the only non-undef value is the low element, turn this into a
1759 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1760 bool isOnlyLowElement = true;
1761 bool MoreThanTwoValues = false;
1762 bool isConstant = true;
1763 for (unsigned i = 0; i < NumElems; ++i) {
1764 SDValue V = Node->getOperand(i);
1765 if (V.getOpcode() == ISD::UNDEF)
1768 isOnlyLowElement = false;
1769 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1772 if (!Value1.getNode()) {
1774 } else if (!Value2.getNode()) {
1777 } else if (V != Value1 && V != Value2) {
1778 MoreThanTwoValues = true;
1782 if (!Value1.getNode())
1783 return DAG.getUNDEF(VT);
1785 if (isOnlyLowElement)
1786 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1788 // If all elements are constants, create a load from the constant pool.
1790 SmallVector<Constant*, 16> CV;
1791 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1792 if (ConstantFPSDNode *V =
1793 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1794 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1795 } else if (ConstantSDNode *V =
1796 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1798 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1800 // If OpVT and EltVT don't match, EltVT is not legal and the
1801 // element values have been promoted/truncated earlier. Undo this;
1802 // we don't want a v16i8 to become a v16i32 for example.
1803 const ConstantInt *CI = V->getConstantIntValue();
1804 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1805 CI->getZExtValue()));
1808 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1809 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1810 CV.push_back(UndefValue::get(OpNTy));
1813 Constant *CP = ConstantVector::get(CV);
1814 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1815 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1816 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1817 MachinePointerInfo::getConstantPool(),
1818 false, false, false, Alignment);
1821 if (!MoreThanTwoValues) {
1822 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1823 for (unsigned i = 0; i < NumElems; ++i) {
1824 SDValue V = Node->getOperand(i);
1825 if (V.getOpcode() == ISD::UNDEF)
1827 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1829 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1830 // Get the splatted value into the low element of a vector register.
1831 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1833 if (Value2.getNode())
1834 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1836 Vec2 = DAG.getUNDEF(VT);
1838 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1839 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1843 // Otherwise, we can't handle this case efficiently.
1844 return ExpandVectorBuildThroughStack(Node);
1847 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
1848 // does not fit into a register, return the lo part and set the hi part to the
1849 // by-reg argument. If it does fit into a single register, return the result
1850 // and leave the Hi part unset.
1851 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
1853 TargetLowering::ArgListTy Args;
1854 TargetLowering::ArgListEntry Entry;
1855 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1856 EVT ArgVT = Node->getOperand(i).getValueType();
1857 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1858 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
1859 Entry.isSExt = isSigned;
1860 Entry.isZExt = !isSigned;
1861 Args.push_back(Entry);
1863 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1864 TLI.getPointerTy());
1866 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1868 // By default, the input chain to this libcall is the entry node of the
1869 // function. If the libcall is going to be emitted as a tail call then
1870 // TLI.isUsedByReturnOnly will change it to the right chain if the return
1871 // node which is being folded has a non-entry input chain.
1872 SDValue InChain = DAG.getEntryNode();
1874 // isTailCall may be true since the callee does not reference caller stack
1875 // frame. Check if it's in the right position.
1876 SDValue TCChain = InChain;
1877 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
1882 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1883 0, TLI.getLibcallCallingConv(LC), isTailCall,
1884 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1885 Callee, Args, DAG, Node->getDebugLoc());
1886 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1889 if (!CallInfo.second.getNode())
1890 // It's a tailcall, return the chain (which is the DAG root).
1891 return DAG.getRoot();
1893 return CallInfo.first;
1896 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
1897 /// and returning a result of type RetVT.
1898 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
1899 const SDValue *Ops, unsigned NumOps,
1900 bool isSigned, DebugLoc dl) {
1901 TargetLowering::ArgListTy Args;
1902 Args.reserve(NumOps);
1904 TargetLowering::ArgListEntry Entry;
1905 for (unsigned i = 0; i != NumOps; ++i) {
1906 Entry.Node = Ops[i];
1907 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
1908 Entry.isSExt = isSigned;
1909 Entry.isZExt = !isSigned;
1910 Args.push_back(Entry);
1912 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1913 TLI.getPointerTy());
1915 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
1917 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
1918 false, 0, TLI.getLibcallCallingConv(LC),
1919 /*isTailCall=*/false,
1920 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1921 Callee, Args, DAG, dl);
1922 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
1924 return CallInfo.first;
1927 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
1928 // ExpandLibCall except that the first operand is the in-chain.
1929 std::pair<SDValue, SDValue>
1930 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
1933 SDValue InChain = Node->getOperand(0);
1935 TargetLowering::ArgListTy Args;
1936 TargetLowering::ArgListEntry Entry;
1937 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
1938 EVT ArgVT = Node->getOperand(i).getValueType();
1939 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1940 Entry.Node = Node->getOperand(i);
1942 Entry.isSExt = isSigned;
1943 Entry.isZExt = !isSigned;
1944 Args.push_back(Entry);
1946 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1947 TLI.getPointerTy());
1949 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1951 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
1952 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
1953 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1954 Callee, Args, DAG, Node->getDebugLoc());
1955 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
1960 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
1961 RTLIB::Libcall Call_F32,
1962 RTLIB::Libcall Call_F64,
1963 RTLIB::Libcall Call_F80,
1964 RTLIB::Libcall Call_F128,
1965 RTLIB::Libcall Call_PPCF128) {
1967 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1968 default: llvm_unreachable("Unexpected request for libcall!");
1969 case MVT::f32: LC = Call_F32; break;
1970 case MVT::f64: LC = Call_F64; break;
1971 case MVT::f80: LC = Call_F80; break;
1972 case MVT::f128: LC = Call_F128; break;
1973 case MVT::ppcf128: LC = Call_PPCF128; break;
1975 return ExpandLibCall(LC, Node, false);
1978 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
1979 RTLIB::Libcall Call_I8,
1980 RTLIB::Libcall Call_I16,
1981 RTLIB::Libcall Call_I32,
1982 RTLIB::Libcall Call_I64,
1983 RTLIB::Libcall Call_I128) {
1985 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
1986 default: llvm_unreachable("Unexpected request for libcall!");
1987 case MVT::i8: LC = Call_I8; break;
1988 case MVT::i16: LC = Call_I16; break;
1989 case MVT::i32: LC = Call_I32; break;
1990 case MVT::i64: LC = Call_I64; break;
1991 case MVT::i128: LC = Call_I128; break;
1993 return ExpandLibCall(LC, Node, isSigned);
1996 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
1997 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
1998 const TargetLowering &TLI) {
2000 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2001 default: llvm_unreachable("Unexpected request for libcall!");
2002 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2003 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2004 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2005 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2006 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2009 return TLI.getLibcallName(LC) != 0;
2012 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2014 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2015 // The other use might have been replaced with a divrem already.
2016 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2017 unsigned OtherOpcode = 0;
2019 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2021 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2023 SDValue Op0 = Node->getOperand(0);
2024 SDValue Op1 = Node->getOperand(1);
2025 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2026 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2030 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2031 User->getOperand(0) == Op0 &&
2032 User->getOperand(1) == Op1)
2038 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2041 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2042 SmallVectorImpl<SDValue> &Results) {
2043 unsigned Opcode = Node->getOpcode();
2044 bool isSigned = Opcode == ISD::SDIVREM;
2047 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2048 default: llvm_unreachable("Unexpected request for libcall!");
2049 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2050 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2051 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2052 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2053 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2056 // The input chain to this libcall is the entry node of the function.
2057 // Legalizing the call will automatically add the previous call to the
2059 SDValue InChain = DAG.getEntryNode();
2061 EVT RetVT = Node->getValueType(0);
2062 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2064 TargetLowering::ArgListTy Args;
2065 TargetLowering::ArgListEntry Entry;
2066 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2067 EVT ArgVT = Node->getOperand(i).getValueType();
2068 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2069 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2070 Entry.isSExt = isSigned;
2071 Entry.isZExt = !isSigned;
2072 Args.push_back(Entry);
2075 // Also pass the return address of the remainder.
2076 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2078 Entry.Ty = RetTy->getPointerTo();
2079 Entry.isSExt = isSigned;
2080 Entry.isZExt = !isSigned;
2081 Args.push_back(Entry);
2083 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2084 TLI.getPointerTy());
2086 DebugLoc dl = Node->getDebugLoc();
2088 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
2089 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2090 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2091 Callee, Args, DAG, dl);
2092 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2094 // Remainder is loaded back from the stack frame.
2095 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2096 MachinePointerInfo(), false, false, false, 0);
2097 Results.push_back(CallInfo.first);
2098 Results.push_back(Rem);
2101 /// isSinCosLibcallAvailable - Return true if sincos libcall is available.
2102 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2104 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2105 default: llvm_unreachable("Unexpected request for libcall!");
2106 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2107 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2108 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2109 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2110 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2112 return TLI.getLibcallName(LC) != 0;
2115 /// canCombineSinCosLibcall - Return true if sincos libcall is available and
2116 /// can be used to combine sin and cos.
2117 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2118 const TargetMachine &TM) {
2119 if (!isSinCosLibcallAvailable(Node, TLI))
2121 // GNU sin/cos functions set errno while sincos does not. Therefore
2122 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2123 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2124 if (isGNU && !TM.Options.UnsafeFPMath)
2129 /// useSinCos - Only issue sincos libcall if both sin and cos are
2131 static bool useSinCos(SDNode *Node) {
2132 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2133 ? ISD::FCOS : ISD::FSIN;
2135 SDValue Op0 = Node->getOperand(0);
2136 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2137 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2141 // The other user might have been turned into sincos already.
2142 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2148 /// ExpandSinCosLibCall - Issue libcalls to sincos to compute sin / cos
2151 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2152 SmallVectorImpl<SDValue> &Results) {
2154 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2155 default: llvm_unreachable("Unexpected request for libcall!");
2156 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2157 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2158 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2159 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2160 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2163 // The input chain to this libcall is the entry node of the function.
2164 // Legalizing the call will automatically add the previous call to the
2166 SDValue InChain = DAG.getEntryNode();
2168 EVT RetVT = Node->getValueType(0);
2169 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2171 TargetLowering::ArgListTy Args;
2172 TargetLowering::ArgListEntry Entry;
2174 // Pass the argument.
2175 Entry.Node = Node->getOperand(0);
2177 Entry.isSExt = false;
2178 Entry.isZExt = false;
2179 Args.push_back(Entry);
2181 // Pass the return address of sin.
2182 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2183 Entry.Node = SinPtr;
2184 Entry.Ty = RetTy->getPointerTo();
2185 Entry.isSExt = false;
2186 Entry.isZExt = false;
2187 Args.push_back(Entry);
2189 // Also pass the return address of the cos.
2190 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2191 Entry.Node = CosPtr;
2192 Entry.Ty = RetTy->getPointerTo();
2193 Entry.isSExt = false;
2194 Entry.isZExt = false;
2195 Args.push_back(Entry);
2197 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2198 TLI.getPointerTy());
2200 DebugLoc dl = Node->getDebugLoc();
2202 CallLoweringInfo CLI(InChain, Type::getVoidTy(*DAG.getContext()),
2203 false, false, false, false,
2204 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2205 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2206 Callee, Args, DAG, dl);
2207 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2209 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2210 MachinePointerInfo(), false, false, false, 0));
2211 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2212 MachinePointerInfo(), false, false, false, 0));
2215 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2216 /// INT_TO_FP operation of the specified operand when the target requests that
2217 /// we expand it. At this point, we know that the result and operand types are
2218 /// legal for the target.
2219 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2223 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2224 // simple 32-bit [signed|unsigned] integer to float/double expansion
2226 // Get the stack frame index of a 8 byte buffer.
2227 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2229 // word offset constant for Hi/Lo address computation
2230 SDValue WordOff = DAG.getConstant(sizeof(int), TLI.getPointerTy());
2231 // set up Hi and Lo (into buffer) address based on endian
2232 SDValue Hi = StackSlot;
2233 SDValue Lo = DAG.getNode(ISD::ADD, dl,
2234 TLI.getPointerTy(), StackSlot, WordOff);
2235 if (TLI.isLittleEndian())
2238 // if signed map to unsigned space
2241 // constant used to invert sign bit (signed to unsigned mapping)
2242 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2243 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2247 // store the lo of the constructed double - based on integer input
2248 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2249 Op0Mapped, Lo, MachinePointerInfo(),
2251 // initial hi portion of constructed double
2252 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2253 // store the hi of the constructed double - biased exponent
2254 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2255 MachinePointerInfo(),
2257 // load the constructed double
2258 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2259 MachinePointerInfo(), false, false, false, 0);
2260 // FP constant to bias correct the final result
2261 SDValue Bias = DAG.getConstantFP(isSigned ?
2262 BitsToDouble(0x4330000080000000ULL) :
2263 BitsToDouble(0x4330000000000000ULL),
2265 // subtract the bias
2266 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2269 // handle final rounding
2270 if (DestVT == MVT::f64) {
2273 } else if (DestVT.bitsLT(MVT::f64)) {
2274 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2275 DAG.getIntPtrConstant(0));
2276 } else if (DestVT.bitsGT(MVT::f64)) {
2277 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2281 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2282 // Code below here assumes !isSigned without checking again.
2284 // Implementation of unsigned i64 to f64 following the algorithm in
2285 // __floatundidf in compiler_rt. This implementation has the advantage
2286 // of performing rounding correctly, both in the default rounding mode
2287 // and in all alternate rounding modes.
2288 // TODO: Generalize this for use with other types.
2289 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2291 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2292 SDValue TwoP84PlusTwoP52 =
2293 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2295 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2297 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2298 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2299 DAG.getConstant(32, MVT::i64));
2300 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2301 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2302 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2303 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2304 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2306 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2309 // Implementation of unsigned i64 to f32.
2310 // TODO: Generalize this for use with other types.
2311 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2312 // For unsigned conversions, convert them to signed conversions using the
2313 // algorithm from the x86_64 __floatundidf in compiler_rt.
2315 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2317 SDValue ShiftConst =
2318 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2319 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2320 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2321 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2322 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2324 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2325 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2327 // TODO: This really should be implemented using a branch rather than a
2328 // select. We happen to get lucky and machinesink does the right
2329 // thing most of the time. This would be a good candidate for a
2330 //pseudo-op, or, even better, for whole-function isel.
2331 SDValue SignBitTest = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2332 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2333 return DAG.getNode(ISD::SELECT, dl, MVT::f32, SignBitTest, Slow, Fast);
2336 // Otherwise, implement the fully general conversion.
2338 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2339 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2340 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2341 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2342 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2343 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2344 SDValue Ne = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2345 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2346 SDValue Sel = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ne, Or, Op0);
2347 SDValue Ge = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2348 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2350 SDValue Sel2 = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ge, Sel, Op0);
2351 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2353 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2354 DAG.getConstant(32, SHVT));
2355 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2356 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2358 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2359 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2360 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2361 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2362 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2363 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2364 DAG.getIntPtrConstant(0));
2367 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2369 SDValue SignSet = DAG.getSetCC(dl, TLI.getSetCCResultType(Op0.getValueType()),
2370 Op0, DAG.getConstant(0, Op0.getValueType()),
2372 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2373 SDValue CstOffset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(),
2374 SignSet, Four, Zero);
2376 // If the sign bit of the integer is set, the large number will be treated
2377 // as a negative number. To counteract this, the dynamic code adds an
2378 // offset depending on the data type.
2380 switch (Op0.getValueType().getSimpleVT().SimpleTy) {
2381 default: llvm_unreachable("Unsupported integer type!");
2382 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2383 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2384 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2385 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2387 if (TLI.isLittleEndian()) FF <<= 32;
2388 Constant *FudgeFactor = ConstantInt::get(
2389 Type::getInt64Ty(*DAG.getContext()), FF);
2391 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2392 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2393 CPIdx = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), CPIdx, CstOffset);
2394 Alignment = std::min(Alignment, 4u);
2396 if (DestVT == MVT::f32)
2397 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2398 MachinePointerInfo::getConstantPool(),
2399 false, false, false, Alignment);
2401 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2402 DAG.getEntryNode(), CPIdx,
2403 MachinePointerInfo::getConstantPool(),
2404 MVT::f32, false, false, Alignment);
2405 HandleSDNode Handle(Load);
2406 LegalizeOp(Load.getNode());
2407 FudgeInReg = Handle.getValue();
2410 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2413 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2414 /// *INT_TO_FP operation of the specified operand when the target requests that
2415 /// we promote it. At this point, we know that the result and operand types are
2416 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2417 /// operation that takes a larger input.
2418 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2422 // First step, figure out the appropriate *INT_TO_FP operation to use.
2423 EVT NewInTy = LegalOp.getValueType();
2425 unsigned OpToUse = 0;
2427 // Scan for the appropriate larger type to use.
2429 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2430 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2432 // If the target supports SINT_TO_FP of this type, use it.
2433 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2434 OpToUse = ISD::SINT_TO_FP;
2437 if (isSigned) continue;
2439 // If the target supports UINT_TO_FP of this type, use it.
2440 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2441 OpToUse = ISD::UINT_TO_FP;
2445 // Otherwise, try a larger type.
2448 // Okay, we found the operation and type to use. Zero extend our input to the
2449 // desired type then run the operation on it.
2450 return DAG.getNode(OpToUse, dl, DestVT,
2451 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2452 dl, NewInTy, LegalOp));
2455 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2456 /// FP_TO_*INT operation of the specified operand when the target requests that
2457 /// we promote it. At this point, we know that the result and operand types are
2458 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2459 /// operation that returns a larger result.
2460 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2464 // First step, figure out the appropriate FP_TO*INT operation to use.
2465 EVT NewOutTy = DestVT;
2467 unsigned OpToUse = 0;
2469 // Scan for the appropriate larger type to use.
2471 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2472 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2474 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2475 OpToUse = ISD::FP_TO_SINT;
2479 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2480 OpToUse = ISD::FP_TO_UINT;
2484 // Otherwise, try a larger type.
2488 // Okay, we found the operation and type to use.
2489 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2491 // Truncate the result of the extended FP_TO_*INT operation to the desired
2493 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2496 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2498 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, DebugLoc dl) {
2499 EVT VT = Op.getValueType();
2500 EVT SHVT = TLI.getShiftAmountTy(VT);
2501 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2502 switch (VT.getSimpleVT().SimpleTy) {
2503 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2505 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2506 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2507 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2509 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2510 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2511 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2512 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2513 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2514 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2515 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2516 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2517 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2519 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2520 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2521 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2522 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2523 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2524 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2525 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2526 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2527 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2528 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2529 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2530 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2531 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2532 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2533 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2534 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2535 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2536 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2537 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2538 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2539 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2543 /// SplatByte - Distribute ByteVal over NumBits bits.
2544 // FIXME: Move this helper to a common place.
2545 static APInt SplatByte(unsigned NumBits, uint8_t ByteVal) {
2546 APInt Val = APInt(NumBits, ByteVal);
2548 for (unsigned i = NumBits; i > 8; i >>= 1) {
2549 Val = (Val << Shift) | Val;
2555 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2557 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2560 default: llvm_unreachable("Cannot expand this yet!");
2562 EVT VT = Op.getValueType();
2563 EVT ShVT = TLI.getShiftAmountTy(VT);
2564 unsigned Len = VT.getSizeInBits();
2566 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2567 "CTPOP not implemented for this type.");
2569 // This is the "best" algorithm from
2570 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2572 SDValue Mask55 = DAG.getConstant(SplatByte(Len, 0x55), VT);
2573 SDValue Mask33 = DAG.getConstant(SplatByte(Len, 0x33), VT);
2574 SDValue Mask0F = DAG.getConstant(SplatByte(Len, 0x0F), VT);
2575 SDValue Mask01 = DAG.getConstant(SplatByte(Len, 0x01), VT);
2577 // v = v - ((v >> 1) & 0x55555555...)
2578 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2579 DAG.getNode(ISD::AND, dl, VT,
2580 DAG.getNode(ISD::SRL, dl, VT, Op,
2581 DAG.getConstant(1, ShVT)),
2583 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2584 Op = DAG.getNode(ISD::ADD, dl, VT,
2585 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2586 DAG.getNode(ISD::AND, dl, VT,
2587 DAG.getNode(ISD::SRL, dl, VT, Op,
2588 DAG.getConstant(2, ShVT)),
2590 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2591 Op = DAG.getNode(ISD::AND, dl, VT,
2592 DAG.getNode(ISD::ADD, dl, VT, Op,
2593 DAG.getNode(ISD::SRL, dl, VT, Op,
2594 DAG.getConstant(4, ShVT))),
2596 // v = (v * 0x01010101...) >> (Len - 8)
2597 Op = DAG.getNode(ISD::SRL, dl, VT,
2598 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2599 DAG.getConstant(Len - 8, ShVT));
2603 case ISD::CTLZ_ZERO_UNDEF:
2604 // This trivially expands to CTLZ.
2605 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2607 // for now, we do this:
2608 // x = x | (x >> 1);
2609 // x = x | (x >> 2);
2611 // x = x | (x >>16);
2612 // x = x | (x >>32); // for 64-bit input
2613 // return popcount(~x);
2615 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2616 EVT VT = Op.getValueType();
2617 EVT ShVT = TLI.getShiftAmountTy(VT);
2618 unsigned len = VT.getSizeInBits();
2619 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2620 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2621 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2622 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2624 Op = DAG.getNOT(dl, Op, VT);
2625 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2627 case ISD::CTTZ_ZERO_UNDEF:
2628 // This trivially expands to CTTZ.
2629 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2631 // for now, we use: { return popcount(~x & (x - 1)); }
2632 // unless the target has ctlz but not ctpop, in which case we use:
2633 // { return 32 - nlz(~x & (x-1)); }
2634 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2635 EVT VT = Op.getValueType();
2636 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2637 DAG.getNOT(dl, Op, VT),
2638 DAG.getNode(ISD::SUB, dl, VT, Op,
2639 DAG.getConstant(1, VT)));
2640 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2641 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2642 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2643 return DAG.getNode(ISD::SUB, dl, VT,
2644 DAG.getConstant(VT.getSizeInBits(), VT),
2645 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2646 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2651 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2652 unsigned Opc = Node->getOpcode();
2653 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2658 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2659 case ISD::ATOMIC_SWAP:
2660 switch (VT.SimpleTy) {
2661 default: llvm_unreachable("Unexpected value type for atomic!");
2662 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2663 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2664 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2665 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2668 case ISD::ATOMIC_CMP_SWAP:
2669 switch (VT.SimpleTy) {
2670 default: llvm_unreachable("Unexpected value type for atomic!");
2671 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2672 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2673 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2674 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2677 case ISD::ATOMIC_LOAD_ADD:
2678 switch (VT.SimpleTy) {
2679 default: llvm_unreachable("Unexpected value type for atomic!");
2680 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2681 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2682 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2683 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2686 case ISD::ATOMIC_LOAD_SUB:
2687 switch (VT.SimpleTy) {
2688 default: llvm_unreachable("Unexpected value type for atomic!");
2689 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2690 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2691 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2692 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2695 case ISD::ATOMIC_LOAD_AND:
2696 switch (VT.SimpleTy) {
2697 default: llvm_unreachable("Unexpected value type for atomic!");
2698 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2699 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2700 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2701 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2704 case ISD::ATOMIC_LOAD_OR:
2705 switch (VT.SimpleTy) {
2706 default: llvm_unreachable("Unexpected value type for atomic!");
2707 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2708 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2709 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2710 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2713 case ISD::ATOMIC_LOAD_XOR:
2714 switch (VT.SimpleTy) {
2715 default: llvm_unreachable("Unexpected value type for atomic!");
2716 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2717 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2718 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2719 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2722 case ISD::ATOMIC_LOAD_NAND:
2723 switch (VT.SimpleTy) {
2724 default: llvm_unreachable("Unexpected value type for atomic!");
2725 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2726 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2727 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2728 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2733 return ExpandChainLibCall(LC, Node, false);
2736 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2737 SmallVector<SDValue, 8> Results;
2738 DebugLoc dl = Node->getDebugLoc();
2739 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2740 switch (Node->getOpcode()) {
2743 case ISD::CTLZ_ZERO_UNDEF:
2745 case ISD::CTTZ_ZERO_UNDEF:
2746 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2747 Results.push_back(Tmp1);
2750 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2752 case ISD::FRAMEADDR:
2753 case ISD::RETURNADDR:
2754 case ISD::FRAME_TO_ARGS_OFFSET:
2755 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2757 case ISD::FLT_ROUNDS_:
2758 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2760 case ISD::EH_RETURN:
2764 case ISD::EH_SJLJ_LONGJMP:
2765 // If the target didn't expand these, there's nothing to do, so just
2766 // preserve the chain and be done.
2767 Results.push_back(Node->getOperand(0));
2769 case ISD::EH_SJLJ_SETJMP:
2770 // If the target didn't expand this, just return 'zero' and preserve the
2772 Results.push_back(DAG.getConstant(0, MVT::i32));
2773 Results.push_back(Node->getOperand(0));
2775 case ISD::ATOMIC_FENCE:
2776 case ISD::MEMBARRIER: {
2777 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2778 // FIXME: handle "fence singlethread" more efficiently.
2779 TargetLowering::ArgListTy Args;
2781 CallLoweringInfo CLI(Node->getOperand(0),
2782 Type::getVoidTy(*DAG.getContext()),
2783 false, false, false, false, 0, CallingConv::C,
2784 /*isTailCall=*/false,
2785 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2786 DAG.getExternalSymbol("__sync_synchronize",
2787 TLI.getPointerTy()),
2789 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2791 Results.push_back(CallResult.second);
2794 case ISD::ATOMIC_LOAD: {
2795 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2796 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
2797 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
2798 cast<AtomicSDNode>(Node)->getMemoryVT(),
2799 Node->getOperand(0),
2800 Node->getOperand(1), Zero, Zero,
2801 cast<AtomicSDNode>(Node)->getMemOperand(),
2802 cast<AtomicSDNode>(Node)->getOrdering(),
2803 cast<AtomicSDNode>(Node)->getSynchScope());
2804 Results.push_back(Swap.getValue(0));
2805 Results.push_back(Swap.getValue(1));
2808 case ISD::ATOMIC_STORE: {
2809 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2810 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2811 cast<AtomicSDNode>(Node)->getMemoryVT(),
2812 Node->getOperand(0),
2813 Node->getOperand(1), Node->getOperand(2),
2814 cast<AtomicSDNode>(Node)->getMemOperand(),
2815 cast<AtomicSDNode>(Node)->getOrdering(),
2816 cast<AtomicSDNode>(Node)->getSynchScope());
2817 Results.push_back(Swap.getValue(1));
2820 // By default, atomic intrinsics are marked Legal and lowered. Targets
2821 // which don't support them directly, however, may want libcalls, in which
2822 // case they mark them Expand, and we get here.
2823 case ISD::ATOMIC_SWAP:
2824 case ISD::ATOMIC_LOAD_ADD:
2825 case ISD::ATOMIC_LOAD_SUB:
2826 case ISD::ATOMIC_LOAD_AND:
2827 case ISD::ATOMIC_LOAD_OR:
2828 case ISD::ATOMIC_LOAD_XOR:
2829 case ISD::ATOMIC_LOAD_NAND:
2830 case ISD::ATOMIC_LOAD_MIN:
2831 case ISD::ATOMIC_LOAD_MAX:
2832 case ISD::ATOMIC_LOAD_UMIN:
2833 case ISD::ATOMIC_LOAD_UMAX:
2834 case ISD::ATOMIC_CMP_SWAP: {
2835 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
2836 Results.push_back(Tmp.first);
2837 Results.push_back(Tmp.second);
2840 case ISD::DYNAMIC_STACKALLOC:
2841 ExpandDYNAMIC_STACKALLOC(Node, Results);
2843 case ISD::MERGE_VALUES:
2844 for (unsigned i = 0; i < Node->getNumValues(); i++)
2845 Results.push_back(Node->getOperand(i));
2848 EVT VT = Node->getValueType(0);
2850 Results.push_back(DAG.getConstant(0, VT));
2852 assert(VT.isFloatingPoint() && "Unknown value type!");
2853 Results.push_back(DAG.getConstantFP(0, VT));
2858 // If this operation is not supported, lower it to 'abort()' call
2859 TargetLowering::ArgListTy Args;
2861 CallLoweringInfo CLI(Node->getOperand(0),
2862 Type::getVoidTy(*DAG.getContext()),
2863 false, false, false, false, 0, CallingConv::C,
2864 /*isTailCall=*/false,
2865 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2866 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
2868 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2870 Results.push_back(CallResult.second);
2875 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
2876 Node->getValueType(0), dl);
2877 Results.push_back(Tmp1);
2879 case ISD::FP_EXTEND:
2880 Tmp1 = EmitStackConvert(Node->getOperand(0),
2881 Node->getOperand(0).getValueType(),
2882 Node->getValueType(0), dl);
2883 Results.push_back(Tmp1);
2885 case ISD::SIGN_EXTEND_INREG: {
2886 // NOTE: we could fall back on load/store here too for targets without
2887 // SAR. However, it is doubtful that any exist.
2888 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2889 EVT VT = Node->getValueType(0);
2890 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
2893 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
2894 ExtraVT.getScalarType().getSizeInBits();
2895 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
2896 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
2897 Node->getOperand(0), ShiftCst);
2898 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
2899 Results.push_back(Tmp1);
2902 case ISD::FP_ROUND_INREG: {
2903 // The only way we can lower this is to turn it into a TRUNCSTORE,
2904 // EXTLOAD pair, targeting a temporary location (a stack slot).
2906 // NOTE: there is a choice here between constantly creating new stack
2907 // slots and always reusing the same one. We currently always create
2908 // new ones, as reuse may inhibit scheduling.
2909 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2910 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
2911 Node->getValueType(0), dl);
2912 Results.push_back(Tmp1);
2915 case ISD::SINT_TO_FP:
2916 case ISD::UINT_TO_FP:
2917 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
2918 Node->getOperand(0), Node->getValueType(0), dl);
2919 Results.push_back(Tmp1);
2921 case ISD::FP_TO_UINT: {
2922 SDValue True, False;
2923 EVT VT = Node->getOperand(0).getValueType();
2924 EVT NVT = Node->getValueType(0);
2925 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
2926 APInt::getNullValue(VT.getSizeInBits()));
2927 APInt x = APInt::getSignBit(NVT.getSizeInBits());
2928 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
2929 Tmp1 = DAG.getConstantFP(apf, VT);
2930 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(VT),
2931 Node->getOperand(0),
2933 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
2934 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
2935 DAG.getNode(ISD::FSUB, dl, VT,
2936 Node->getOperand(0), Tmp1));
2937 False = DAG.getNode(ISD::XOR, dl, NVT, False,
2938 DAG.getConstant(x, NVT));
2939 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2, True, False);
2940 Results.push_back(Tmp1);
2944 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2945 EVT VT = Node->getValueType(0);
2946 Tmp1 = Node->getOperand(0);
2947 Tmp2 = Node->getOperand(1);
2948 unsigned Align = Node->getConstantOperandVal(3);
2950 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
2951 MachinePointerInfo(V),
2952 false, false, false, 0);
2953 SDValue VAList = VAListLoad;
2955 if (Align > TLI.getMinStackArgumentAlignment()) {
2956 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
2958 VAList = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2959 DAG.getConstant(Align - 1,
2960 TLI.getPointerTy()));
2962 VAList = DAG.getNode(ISD::AND, dl, TLI.getPointerTy(), VAList,
2963 DAG.getConstant(-(int64_t)Align,
2964 TLI.getPointerTy()));
2967 // Increment the pointer, VAList, to the next vaarg
2968 Tmp3 = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2969 DAG.getConstant(TLI.getDataLayout()->
2970 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
2971 TLI.getPointerTy()));
2972 // Store the incremented VAList to the legalized pointer
2973 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
2974 MachinePointerInfo(V), false, false, 0);
2975 // Load the actual argument out of the pointer VAList
2976 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
2977 false, false, false, 0));
2978 Results.push_back(Results[0].getValue(1));
2982 // This defaults to loading a pointer from the input and storing it to the
2983 // output, returning the chain.
2984 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
2985 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
2986 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
2987 Node->getOperand(2), MachinePointerInfo(VS),
2988 false, false, false, 0);
2989 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
2990 MachinePointerInfo(VD), false, false, 0);
2991 Results.push_back(Tmp1);
2994 case ISD::EXTRACT_VECTOR_ELT:
2995 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
2996 // This must be an access of the only element. Return it.
2997 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
2998 Node->getOperand(0));
3000 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3001 Results.push_back(Tmp1);
3003 case ISD::EXTRACT_SUBVECTOR:
3004 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3006 case ISD::INSERT_SUBVECTOR:
3007 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3009 case ISD::CONCAT_VECTORS: {
3010 Results.push_back(ExpandVectorBuildThroughStack(Node));
3013 case ISD::SCALAR_TO_VECTOR:
3014 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3016 case ISD::INSERT_VECTOR_ELT:
3017 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3018 Node->getOperand(1),
3019 Node->getOperand(2), dl));
3021 case ISD::VECTOR_SHUFFLE: {
3022 SmallVector<int, 32> NewMask;
3023 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3025 EVT VT = Node->getValueType(0);
3026 EVT EltVT = VT.getVectorElementType();
3027 SDValue Op0 = Node->getOperand(0);
3028 SDValue Op1 = Node->getOperand(1);
3029 if (!TLI.isTypeLegal(EltVT)) {
3031 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3033 // BUILD_VECTOR operands are allowed to be wider than the element type.
3034 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept it
3035 if (NewEltVT.bitsLT(EltVT)) {
3037 // Convert shuffle node.
3038 // If original node was v4i64 and the new EltVT is i32,
3039 // cast operands to v8i32 and re-build the mask.
3041 // Calculate new VT, the size of the new VT should be equal to original.
3042 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3043 VT.getSizeInBits()/NewEltVT.getSizeInBits());
3044 assert(NewVT.bitsEq(VT));
3046 // cast operands to new VT
3047 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3048 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3050 // Convert the shuffle mask
3051 unsigned int factor = NewVT.getVectorNumElements()/VT.getVectorNumElements();
3053 // EltVT gets smaller
3056 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3058 for (unsigned fi = 0; fi < factor; ++fi)
3059 NewMask.push_back(Mask[i]);
3062 for (unsigned fi = 0; fi < factor; ++fi)
3063 NewMask.push_back(Mask[i]*factor+fi);
3071 unsigned NumElems = VT.getVectorNumElements();
3072 SmallVector<SDValue, 16> Ops;
3073 for (unsigned i = 0; i != NumElems; ++i) {
3075 Ops.push_back(DAG.getUNDEF(EltVT));
3078 unsigned Idx = Mask[i];
3080 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3082 DAG.getIntPtrConstant(Idx)));
3084 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3086 DAG.getIntPtrConstant(Idx - NumElems)));
3089 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
3090 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3091 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3092 Results.push_back(Tmp1);
3095 case ISD::EXTRACT_ELEMENT: {
3096 EVT OpTy = Node->getOperand(0).getValueType();
3097 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3099 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3100 DAG.getConstant(OpTy.getSizeInBits()/2,
3101 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3102 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3105 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3106 Node->getOperand(0));
3108 Results.push_back(Tmp1);
3111 case ISD::STACKSAVE:
3112 // Expand to CopyFromReg if the target set
3113 // StackPointerRegisterToSaveRestore.
3114 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3115 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3116 Node->getValueType(0)));
3117 Results.push_back(Results[0].getValue(1));
3119 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3120 Results.push_back(Node->getOperand(0));
3123 case ISD::STACKRESTORE:
3124 // Expand to CopyToReg if the target set
3125 // StackPointerRegisterToSaveRestore.
3126 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3127 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3128 Node->getOperand(1)));
3130 Results.push_back(Node->getOperand(0));
3133 case ISD::FCOPYSIGN:
3134 Results.push_back(ExpandFCOPYSIGN(Node));
3137 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3138 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3139 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3140 Node->getOperand(0));
3141 Results.push_back(Tmp1);
3144 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3145 EVT VT = Node->getValueType(0);
3146 Tmp1 = Node->getOperand(0);
3147 Tmp2 = DAG.getConstantFP(0.0, VT);
3148 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(Tmp1.getValueType()),
3149 Tmp1, Tmp2, ISD::SETUGT);
3150 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3151 Tmp1 = DAG.getNode(ISD::SELECT, dl, VT, Tmp2, Tmp1, Tmp3);
3152 Results.push_back(Tmp1);
3156 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3157 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3158 RTLIB::SQRT_PPCF128));
3162 EVT VT = Node->getValueType(0);
3163 bool isSIN = Node->getOpcode() == ISD::FSIN;
3164 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3165 // fcos which share the same operand and both are used.
3166 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3167 canCombineSinCosLibcall(Node, TLI, TM))
3168 && useSinCos(Node)) {
3169 SDVTList VTs = DAG.getVTList(VT, VT);
3170 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3172 Tmp1 = Tmp1.getValue(1);
3173 Results.push_back(Tmp1);
3175 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3176 RTLIB::SIN_F80, RTLIB::SIN_F128,
3177 RTLIB::SIN_PPCF128));
3179 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3180 RTLIB::COS_F80, RTLIB::COS_F128,
3181 RTLIB::COS_PPCF128));
3186 // Expand into sincos libcall.
3187 ExpandSinCosLibCall(Node, Results);
3190 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3191 RTLIB::LOG_F80, RTLIB::LOG_F128,
3192 RTLIB::LOG_PPCF128));
3195 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3196 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3197 RTLIB::LOG2_PPCF128));
3200 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3201 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3202 RTLIB::LOG10_PPCF128));
3205 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3206 RTLIB::EXP_F80, RTLIB::EXP_F128,
3207 RTLIB::EXP_PPCF128));
3210 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3211 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3212 RTLIB::EXP2_PPCF128));
3215 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3216 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3217 RTLIB::TRUNC_PPCF128));
3220 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3221 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3222 RTLIB::FLOOR_PPCF128));
3225 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3226 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3227 RTLIB::CEIL_PPCF128));
3230 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3231 RTLIB::RINT_F80, RTLIB::RINT_F128,
3232 RTLIB::RINT_PPCF128));
3234 case ISD::FNEARBYINT:
3235 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3236 RTLIB::NEARBYINT_F64,
3237 RTLIB::NEARBYINT_F80,
3238 RTLIB::NEARBYINT_F128,
3239 RTLIB::NEARBYINT_PPCF128));
3242 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3243 RTLIB::POWI_F80, RTLIB::POWI_F128,
3244 RTLIB::POWI_PPCF128));
3247 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3248 RTLIB::POW_F80, RTLIB::POW_F128,
3249 RTLIB::POW_PPCF128));
3252 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3253 RTLIB::DIV_F80, RTLIB::DIV_F128,
3254 RTLIB::DIV_PPCF128));
3257 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3258 RTLIB::REM_F80, RTLIB::REM_F128,
3259 RTLIB::REM_PPCF128));
3262 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3263 RTLIB::FMA_F80, RTLIB::FMA_F128,
3264 RTLIB::FMA_PPCF128));
3266 case ISD::FP16_TO_FP32:
3267 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3269 case ISD::FP32_TO_FP16:
3270 Results.push_back(ExpandLibCall(RTLIB::FPROUND_F32_F16, Node, false));
3272 case ISD::ConstantFP: {
3273 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3274 // Check to see if this FP immediate is already legal.
3275 // If this is a legal constant, turn it into a TargetConstantFP node.
3276 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3277 Results.push_back(ExpandConstantFP(CFP, true));
3280 case ISD::EHSELECTION: {
3281 unsigned Reg = TLI.getExceptionSelectorRegister();
3282 assert(Reg && "Can't expand to unknown register!");
3283 Results.push_back(DAG.getCopyFromReg(Node->getOperand(1), dl, Reg,
3284 Node->getValueType(0)));
3285 Results.push_back(Results[0].getValue(1));
3288 case ISD::EXCEPTIONADDR: {
3289 unsigned Reg = TLI.getExceptionPointerRegister();
3290 assert(Reg && "Can't expand to unknown register!");
3291 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, Reg,
3292 Node->getValueType(0)));
3293 Results.push_back(Results[0].getValue(1));
3297 EVT VT = Node->getValueType(0);
3298 assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3299 TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
3300 "Don't know how to expand this FP subtraction!");
3301 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3302 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3303 Results.push_back(Tmp1);
3307 EVT VT = Node->getValueType(0);
3308 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3309 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3310 "Don't know how to expand this subtraction!");
3311 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3312 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3313 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3314 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3319 EVT VT = Node->getValueType(0);
3320 bool isSigned = Node->getOpcode() == ISD::SREM;
3321 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3322 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3323 Tmp2 = Node->getOperand(0);
3324 Tmp3 = Node->getOperand(1);
3325 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3326 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3327 // If div is legal, it's better to do the normal expansion
3328 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3329 useDivRem(Node, isSigned, false))) {
3330 SDVTList VTs = DAG.getVTList(VT, VT);
3331 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3332 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3334 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3335 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3336 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3337 } else if (isSigned)
3338 Tmp1 = ExpandIntLibCall(Node, true,
3340 RTLIB::SREM_I16, RTLIB::SREM_I32,
3341 RTLIB::SREM_I64, RTLIB::SREM_I128);
3343 Tmp1 = ExpandIntLibCall(Node, false,
3345 RTLIB::UREM_I16, RTLIB::UREM_I32,
3346 RTLIB::UREM_I64, RTLIB::UREM_I128);
3347 Results.push_back(Tmp1);
3352 bool isSigned = Node->getOpcode() == ISD::SDIV;
3353 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3354 EVT VT = Node->getValueType(0);
3355 SDVTList VTs = DAG.getVTList(VT, VT);
3356 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3357 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3358 useDivRem(Node, isSigned, true)))
3359 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3360 Node->getOperand(1));
3362 Tmp1 = ExpandIntLibCall(Node, true,
3364 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3365 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3367 Tmp1 = ExpandIntLibCall(Node, false,
3369 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3370 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3371 Results.push_back(Tmp1);
3376 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3378 EVT VT = Node->getValueType(0);
3379 SDVTList VTs = DAG.getVTList(VT, VT);
3380 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3381 "If this wasn't legal, it shouldn't have been created!");
3382 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3383 Node->getOperand(1));
3384 Results.push_back(Tmp1.getValue(1));
3389 // Expand into divrem libcall
3390 ExpandDivRemLibCall(Node, Results);
3393 EVT VT = Node->getValueType(0);
3394 SDVTList VTs = DAG.getVTList(VT, VT);
3395 // See if multiply or divide can be lowered using two-result operations.
3396 // We just need the low half of the multiply; try both the signed
3397 // and unsigned forms. If the target supports both SMUL_LOHI and
3398 // UMUL_LOHI, form a preference by checking which forms of plain
3399 // MULH it supports.
3400 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3401 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3402 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3403 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3404 unsigned OpToUse = 0;
3405 if (HasSMUL_LOHI && !HasMULHS) {
3406 OpToUse = ISD::SMUL_LOHI;
3407 } else if (HasUMUL_LOHI && !HasMULHU) {
3408 OpToUse = ISD::UMUL_LOHI;
3409 } else if (HasSMUL_LOHI) {
3410 OpToUse = ISD::SMUL_LOHI;
3411 } else if (HasUMUL_LOHI) {
3412 OpToUse = ISD::UMUL_LOHI;
3415 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3416 Node->getOperand(1)));
3419 Tmp1 = ExpandIntLibCall(Node, false,
3421 RTLIB::MUL_I16, RTLIB::MUL_I32,
3422 RTLIB::MUL_I64, RTLIB::MUL_I128);
3423 Results.push_back(Tmp1);
3428 SDValue LHS = Node->getOperand(0);
3429 SDValue RHS = Node->getOperand(1);
3430 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3431 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3433 Results.push_back(Sum);
3434 EVT OType = Node->getValueType(1);
3436 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3438 // LHSSign -> LHS >= 0
3439 // RHSSign -> RHS >= 0
3440 // SumSign -> Sum >= 0
3443 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3445 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3447 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3448 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3449 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3450 Node->getOpcode() == ISD::SADDO ?
3451 ISD::SETEQ : ISD::SETNE);
3453 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3454 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3456 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3457 Results.push_back(Cmp);
3462 SDValue LHS = Node->getOperand(0);
3463 SDValue RHS = Node->getOperand(1);
3464 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3465 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3467 Results.push_back(Sum);
3468 Results.push_back(DAG.getSetCC(dl, Node->getValueType(1), Sum, LHS,
3469 Node->getOpcode () == ISD::UADDO ?
3470 ISD::SETULT : ISD::SETUGT));
3475 EVT VT = Node->getValueType(0);
3476 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3477 SDValue LHS = Node->getOperand(0);
3478 SDValue RHS = Node->getOperand(1);
3481 static const unsigned Ops[2][3] =
3482 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3483 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3484 bool isSigned = Node->getOpcode() == ISD::SMULO;
3485 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3486 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3487 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3488 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3489 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3491 TopHalf = BottomHalf.getValue(1);
3492 } else if (TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
3493 VT.getSizeInBits() * 2))) {
3494 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3495 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3496 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3497 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3498 DAG.getIntPtrConstant(0));
3499 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3500 DAG.getIntPtrConstant(1));
3502 // We can fall back to a libcall with an illegal type for the MUL if we
3503 // have a libcall big enough.
3504 // Also, we can fall back to a division in some cases, but that's a big
3505 // performance hit in the general case.
3506 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3507 if (WideVT == MVT::i16)
3508 LC = RTLIB::MUL_I16;
3509 else if (WideVT == MVT::i32)
3510 LC = RTLIB::MUL_I32;
3511 else if (WideVT == MVT::i64)
3512 LC = RTLIB::MUL_I64;
3513 else if (WideVT == MVT::i128)
3514 LC = RTLIB::MUL_I128;
3515 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3517 // The high part is obtained by SRA'ing all but one of the bits of low
3519 unsigned LoSize = VT.getSizeInBits();
3520 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3521 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3522 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3523 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3525 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3526 // pre-lowered to the correct types. This all depends upon WideVT not
3527 // being a legal type for the architecture and thus has to be split to
3529 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3530 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3531 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3532 DAG.getIntPtrConstant(0));
3533 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3534 DAG.getIntPtrConstant(1));
3535 // Ret is a node with an illegal type. Because such things are not
3536 // generally permitted during this phase of legalization, delete the
3537 // node. The above EXTRACT_ELEMENT nodes should have been folded.
3538 DAG.DeleteNode(Ret.getNode());
3542 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3543 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3544 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3545 TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf, Tmp1,
3548 TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf,
3549 DAG.getConstant(0, VT), ISD::SETNE);
3551 Results.push_back(BottomHalf);
3552 Results.push_back(TopHalf);
3555 case ISD::BUILD_PAIR: {
3556 EVT PairTy = Node->getValueType(0);
3557 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3558 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3559 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3560 DAG.getConstant(PairTy.getSizeInBits()/2,
3561 TLI.getShiftAmountTy(PairTy)));
3562 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3566 Tmp1 = Node->getOperand(0);
3567 Tmp2 = Node->getOperand(1);
3568 Tmp3 = Node->getOperand(2);
3569 if (Tmp1.getOpcode() == ISD::SETCC) {
3570 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3572 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3574 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3575 DAG.getConstant(0, Tmp1.getValueType()),
3576 Tmp2, Tmp3, ISD::SETNE);
3578 Results.push_back(Tmp1);
3581 SDValue Chain = Node->getOperand(0);
3582 SDValue Table = Node->getOperand(1);
3583 SDValue Index = Node->getOperand(2);
3585 EVT PTy = TLI.getPointerTy();
3587 const DataLayout &TD = *TLI.getDataLayout();
3588 unsigned EntrySize =
3589 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3591 Index = DAG.getNode(ISD::MUL, dl, PTy,
3592 Index, DAG.getConstant(EntrySize, PTy));
3593 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3595 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3596 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3597 MachinePointerInfo::getJumpTable(), MemVT,
3600 if (TM.getRelocationModel() == Reloc::PIC_) {
3601 // For PIC, the sequence is:
3602 // BRIND(load(Jumptable + index) + RelocBase)
3603 // RelocBase can be JumpTable, GOT or some sort of global base.
3604 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3605 TLI.getPICJumpTableRelocBase(Table, DAG));
3607 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3608 Results.push_back(Tmp1);
3612 // Expand brcond's setcc into its constituent parts and create a BR_CC
3614 Tmp1 = Node->getOperand(0);
3615 Tmp2 = Node->getOperand(1);
3616 if (Tmp2.getOpcode() == ISD::SETCC) {
3617 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3618 Tmp1, Tmp2.getOperand(2),
3619 Tmp2.getOperand(0), Tmp2.getOperand(1),
3620 Node->getOperand(2));
3622 // We test only the i1 bit. Skip the AND if UNDEF.
3623 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3624 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3625 DAG.getConstant(1, Tmp2.getValueType()));
3626 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3627 DAG.getCondCode(ISD::SETNE), Tmp3,
3628 DAG.getConstant(0, Tmp3.getValueType()),
3629 Node->getOperand(2));
3631 Results.push_back(Tmp1);
3634 Tmp1 = Node->getOperand(0);
3635 Tmp2 = Node->getOperand(1);
3636 Tmp3 = Node->getOperand(2);
3637 LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2, Tmp3, dl);
3639 // If we expanded the SETCC into an AND/OR, return the new node
3640 if (Tmp2.getNode() == 0) {
3641 Results.push_back(Tmp1);
3645 // Otherwise, SETCC for the given comparison type must be completely
3646 // illegal; expand it into a SELECT_CC.
3647 EVT VT = Node->getValueType(0);
3648 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3649 DAG.getConstant(1, VT), DAG.getConstant(0, VT), Tmp3);
3650 Results.push_back(Tmp1);
3653 case ISD::SELECT_CC: {
3654 Tmp1 = Node->getOperand(0); // LHS
3655 Tmp2 = Node->getOperand(1); // RHS
3656 Tmp3 = Node->getOperand(2); // True
3657 Tmp4 = Node->getOperand(3); // False
3658 SDValue CC = Node->getOperand(4);
3660 LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp1.getValueType()),
3661 Tmp1, Tmp2, CC, dl);
3663 assert(!Tmp2.getNode() && "Can't legalize SELECT_CC with legal condition!");
3664 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
3665 CC = DAG.getCondCode(ISD::SETNE);
3666 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1, Tmp2,
3668 Results.push_back(Tmp1);
3672 Tmp1 = Node->getOperand(0); // Chain
3673 Tmp2 = Node->getOperand(2); // LHS
3674 Tmp3 = Node->getOperand(3); // RHS
3675 Tmp4 = Node->getOperand(1); // CC
3677 LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp2.getValueType()),
3678 Tmp2, Tmp3, Tmp4, dl);
3680 assert(!Tmp3.getNode() && "Can't legalize BR_CC with legal condition!");
3681 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
3682 Tmp4 = DAG.getCondCode(ISD::SETNE);
3683 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4, Tmp2,
3684 Tmp3, Node->getOperand(4));
3685 Results.push_back(Tmp1);
3688 case ISD::BUILD_VECTOR:
3689 Results.push_back(ExpandBUILD_VECTOR(Node));
3694 // Scalarize vector SRA/SRL/SHL.
3695 EVT VT = Node->getValueType(0);
3696 assert(VT.isVector() && "Unable to legalize non-vector shift");
3697 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
3698 unsigned NumElem = VT.getVectorNumElements();
3700 SmallVector<SDValue, 8> Scalars;
3701 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
3702 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3704 Node->getOperand(0), DAG.getIntPtrConstant(Idx));
3705 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
3707 Node->getOperand(1), DAG.getIntPtrConstant(Idx));
3708 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
3709 VT.getScalarType(), Ex, Sh));
3712 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
3713 &Scalars[0], Scalars.size());
3714 ReplaceNode(SDValue(Node, 0), Result);
3717 case ISD::GLOBAL_OFFSET_TABLE:
3718 case ISD::GlobalAddress:
3719 case ISD::GlobalTLSAddress:
3720 case ISD::ExternalSymbol:
3721 case ISD::ConstantPool:
3722 case ISD::JumpTable:
3723 case ISD::INTRINSIC_W_CHAIN:
3724 case ISD::INTRINSIC_WO_CHAIN:
3725 case ISD::INTRINSIC_VOID:
3726 // FIXME: Custom lowering for these operations shouldn't return null!
3730 // Replace the original node with the legalized result.
3731 if (!Results.empty())
3732 ReplaceNode(Node, Results.data());
3735 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
3736 SmallVector<SDValue, 8> Results;
3737 MVT OVT = Node->getSimpleValueType(0);
3738 if (Node->getOpcode() == ISD::UINT_TO_FP ||
3739 Node->getOpcode() == ISD::SINT_TO_FP ||
3740 Node->getOpcode() == ISD::SETCC) {
3741 OVT = Node->getOperand(0).getSimpleValueType();
3743 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
3744 DebugLoc dl = Node->getDebugLoc();
3745 SDValue Tmp1, Tmp2, Tmp3;
3746 switch (Node->getOpcode()) {
3748 case ISD::CTTZ_ZERO_UNDEF:
3750 case ISD::CTLZ_ZERO_UNDEF:
3752 // Zero extend the argument.
3753 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3754 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
3755 // already the correct result.
3756 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3757 if (Node->getOpcode() == ISD::CTTZ) {
3758 // FIXME: This should set a bit in the zero extended value instead.
3759 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(NVT),
3760 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
3762 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2,
3763 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
3764 } else if (Node->getOpcode() == ISD::CTLZ ||
3765 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
3766 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
3767 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
3768 DAG.getConstant(NVT.getSizeInBits() -
3769 OVT.getSizeInBits(), NVT));
3771 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
3774 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
3775 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3776 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
3777 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
3778 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
3779 Results.push_back(Tmp1);
3782 case ISD::FP_TO_UINT:
3783 case ISD::FP_TO_SINT:
3784 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
3785 Node->getOpcode() == ISD::FP_TO_SINT, dl);
3786 Results.push_back(Tmp1);
3788 case ISD::UINT_TO_FP:
3789 case ISD::SINT_TO_FP:
3790 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
3791 Node->getOpcode() == ISD::SINT_TO_FP, dl);
3792 Results.push_back(Tmp1);
3795 SDValue Chain = Node->getOperand(0); // Get the chain.
3796 SDValue Ptr = Node->getOperand(1); // Get the pointer.
3799 if (OVT.isVector()) {
3800 TruncOp = ISD::BITCAST;
3802 assert(OVT.isInteger()
3803 && "VAARG promotion is supported only for vectors or integer types");
3804 TruncOp = ISD::TRUNCATE;
3807 // Perform the larger operation, then convert back
3808 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
3809 Node->getConstantOperandVal(3));
3810 Chain = Tmp1.getValue(1);
3812 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
3814 // Modified the chain result - switch anything that used the old chain to
3816 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
3817 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
3824 unsigned ExtOp, TruncOp;
3825 if (OVT.isVector()) {
3826 ExtOp = ISD::BITCAST;
3827 TruncOp = ISD::BITCAST;
3829 assert(OVT.isInteger() && "Cannot promote logic operation");
3830 ExtOp = ISD::ANY_EXTEND;
3831 TruncOp = ISD::TRUNCATE;
3833 // Promote each of the values to the new type.
3834 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3835 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3836 // Perform the larger operation, then convert back
3837 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3838 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
3842 unsigned ExtOp, TruncOp;
3843 if (Node->getValueType(0).isVector()) {
3844 ExtOp = ISD::BITCAST;
3845 TruncOp = ISD::BITCAST;
3846 } else if (Node->getValueType(0).isInteger()) {
3847 ExtOp = ISD::ANY_EXTEND;
3848 TruncOp = ISD::TRUNCATE;
3850 ExtOp = ISD::FP_EXTEND;
3851 TruncOp = ISD::FP_ROUND;
3853 Tmp1 = Node->getOperand(0);
3854 // Promote each of the values to the new type.
3855 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3856 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
3857 // Perform the larger operation, then round down.
3858 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp1, Tmp2, Tmp3);
3859 if (TruncOp != ISD::FP_ROUND)
3860 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
3862 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
3863 DAG.getIntPtrConstant(0));
3864 Results.push_back(Tmp1);
3867 case ISD::VECTOR_SHUFFLE: {
3868 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3870 // Cast the two input vectors.
3871 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
3872 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
3874 // Convert the shuffle mask to the right # elements.
3875 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
3876 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
3877 Results.push_back(Tmp1);
3881 unsigned ExtOp = ISD::FP_EXTEND;
3882 if (NVT.isInteger()) {
3883 ISD::CondCode CCCode =
3884 cast<CondCodeSDNode>(Node->getOperand(2))->get();
3885 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3887 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3888 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3889 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3890 Tmp1, Tmp2, Node->getOperand(2)));
3896 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
3897 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
3898 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3899 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
3900 Tmp3, DAG.getIntPtrConstant(0)));
3907 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
3908 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3909 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
3910 Tmp2, DAG.getIntPtrConstant(0)));
3915 // Replace the original node with the legalized result.
3916 if (!Results.empty())
3917 ReplaceNode(Node, Results.data());
3920 // SelectionDAG::Legalize - This is the entry point for the file.
3922 void SelectionDAG::Legalize() {
3923 /// run - This is the main entry point to this class.
3925 SelectionDAGLegalize(*this).LegalizeDAG();