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/ADT/SetVector.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineJumpTableInfo.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGNodes.h"
23 #include "llvm/IR/CallingConv.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DebugInfo.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetFrameLowering.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetSubtargetInfo.h"
40 #define DEBUG_TYPE "legalizedag"
44 struct FloatSignAsInt;
46 //===----------------------------------------------------------------------===//
47 /// This takes an arbitrary SelectionDAG as input and
48 /// hacks on it until the target machine can handle it. This involves
49 /// eliminating value sizes the machine cannot handle (promoting small sizes to
50 /// large sizes or splitting up large values into small values) as well as
51 /// eliminating operations the machine cannot handle.
53 /// This code also does a small amount of optimization and recognition of idioms
54 /// as part of its processing. For example, if a target does not support a
55 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
56 /// will attempt merge setcc and brc instructions into brcc's.
58 class SelectionDAGLegalize {
59 const TargetMachine &TM;
60 const TargetLowering &TLI;
63 /// \brief The set of nodes which have already been legalized. We hold a
64 /// reference to it in order to update as necessary on node deletion.
65 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
67 /// \brief A set of all the nodes updated during legalization.
68 SmallSetVector<SDNode *, 16> *UpdatedNodes;
70 EVT getSetCCResultType(EVT VT) const {
71 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
74 // Libcall insertion helpers.
77 SelectionDAGLegalize(SelectionDAG &DAG,
78 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
79 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
80 : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
81 LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {}
83 /// \brief Legalizes the given operation.
84 void LegalizeOp(SDNode *Node);
87 SDValue OptimizeFloatStore(StoreSDNode *ST);
89 void LegalizeLoadOps(SDNode *Node);
90 void LegalizeStoreOps(SDNode *Node);
92 /// Some targets cannot handle a variable
93 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
94 /// is necessary to spill the vector being inserted into to memory, perform
95 /// the insert there, and then read the result back.
96 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
98 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx,
101 /// Return a vector shuffle operation which
102 /// performs the same shuffe in terms of order or result bytes, but on a type
103 /// whose vector element type is narrower than the original shuffle type.
104 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
105 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, const SDLoc &dl,
106 SDValue N1, SDValue N2,
107 ArrayRef<int> Mask) const;
109 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
110 bool &NeedInvert, const SDLoc &dl);
112 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
113 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
114 unsigned NumOps, bool isSigned, const SDLoc &dl);
116 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
117 SDNode *Node, bool isSigned);
118 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
119 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
120 RTLIB::Libcall Call_F128,
121 RTLIB::Libcall Call_PPCF128);
122 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
123 RTLIB::Libcall Call_I8,
124 RTLIB::Libcall Call_I16,
125 RTLIB::Libcall Call_I32,
126 RTLIB::Libcall Call_I64,
127 RTLIB::Libcall Call_I128);
128 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
129 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
131 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT,
133 SDValue ExpandBUILD_VECTOR(SDNode *Node);
134 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
135 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
136 SmallVectorImpl<SDValue> &Results);
137 void getSignAsIntValue(FloatSignAsInt &State, const SDLoc &DL,
138 SDValue Value) const;
139 SDValue modifySignAsInt(const FloatSignAsInt &State, const SDLoc &DL,
140 SDValue NewIntValue) const;
141 SDValue ExpandFCOPYSIGN(SDNode *Node) const;
142 SDValue ExpandFABS(SDNode *Node) const;
143 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
145 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
147 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
150 SDValue ExpandBITREVERSE(SDValue Op, const SDLoc &dl);
151 SDValue ExpandBSWAP(SDValue Op, const SDLoc &dl);
152 SDValue ExpandBitCount(unsigned Opc, SDValue Op, const SDLoc &dl);
154 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
155 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
156 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
158 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
159 SDValue ExpandConstant(ConstantSDNode *CP);
161 // if ExpandNode returns false, LegalizeOp falls back to ConvertNodeToLibcall
162 bool ExpandNode(SDNode *Node);
163 void ConvertNodeToLibcall(SDNode *Node);
164 void PromoteNode(SDNode *Node);
167 // Node replacement helpers
168 void ReplacedNode(SDNode *N) {
169 LegalizedNodes.erase(N);
171 UpdatedNodes->insert(N);
173 void ReplaceNode(SDNode *Old, SDNode *New) {
174 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
175 dbgs() << " with: "; New->dump(&DAG));
177 assert(Old->getNumValues() == New->getNumValues() &&
178 "Replacing one node with another that produces a different number "
180 DAG.ReplaceAllUsesWith(Old, New);
182 UpdatedNodes->insert(New);
185 void ReplaceNode(SDValue Old, SDValue New) {
186 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
187 dbgs() << " with: "; New->dump(&DAG));
189 DAG.ReplaceAllUsesWith(Old, New);
191 UpdatedNodes->insert(New.getNode());
192 ReplacedNode(Old.getNode());
194 void ReplaceNode(SDNode *Old, const SDValue *New) {
195 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG));
197 DAG.ReplaceAllUsesWith(Old, New);
198 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
199 DEBUG(dbgs() << (i == 0 ? " with: "
203 UpdatedNodes->insert(New[i].getNode());
210 /// Return a vector shuffle operation which
211 /// performs the same shuffe in terms of order or result bytes, but on a type
212 /// whose vector element type is narrower than the original shuffle type.
213 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
214 SDValue SelectionDAGLegalize::ShuffleWithNarrowerEltType(
215 EVT NVT, EVT VT, const SDLoc &dl, SDValue N1, SDValue N2,
216 ArrayRef<int> Mask) const {
217 unsigned NumMaskElts = VT.getVectorNumElements();
218 unsigned NumDestElts = NVT.getVectorNumElements();
219 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
221 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
223 if (NumEltsGrowth == 1)
224 return DAG.getVectorShuffle(NVT, dl, N1, N2, Mask);
226 SmallVector<int, 8> NewMask;
227 for (unsigned i = 0; i != NumMaskElts; ++i) {
229 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
231 NewMask.push_back(-1);
233 NewMask.push_back(Idx * NumEltsGrowth + j);
236 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
237 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
238 return DAG.getVectorShuffle(NVT, dl, N1, N2, NewMask);
241 /// Expands the ConstantFP node to an integer constant or
242 /// a load from the constant pool.
244 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
248 // If a FP immediate is precise when represented as a float and if the
249 // target can do an extending load from float to double, we put it into
250 // the constant pool as a float, even if it's is statically typed as a
251 // double. This shrinks FP constants and canonicalizes them for targets where
252 // an FP extending load is the same cost as a normal load (such as on the x87
253 // fp stack or PPC FP unit).
254 EVT VT = CFP->getValueType(0);
255 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
257 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
258 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(), dl,
259 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
262 APFloat APF = CFP->getValueAPF();
266 // We don't want to shrink SNaNs. Converting the SNaN back to its real type
267 // can cause it to be changed into a QNaN on some platforms (e.g. on SystemZ).
268 if (!APF.isSignaling()) {
269 while (SVT != MVT::f32 && SVT != MVT::f16) {
270 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
271 if (ConstantFPSDNode::isValueValidForType(SVT, APF) &&
272 // Only do this if the target has a native EXTLOAD instruction from
274 TLI.isLoadExtLegal(ISD::EXTLOAD, OrigVT, SVT) &&
275 TLI.ShouldShrinkFPConstant(OrigVT)) {
276 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
277 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
285 DAG.getConstantPool(LLVMC, TLI.getPointerTy(DAG.getDataLayout()));
286 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
288 SDValue Result = DAG.getExtLoad(
289 ISD::EXTLOAD, dl, OrigVT, DAG.getEntryNode(), CPIdx,
290 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), VT,
294 SDValue Result = DAG.getLoad(
295 OrigVT, dl, DAG.getEntryNode(), CPIdx,
296 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Alignment);
300 /// Expands the Constant node to a load from the constant pool.
301 SDValue SelectionDAGLegalize::ExpandConstant(ConstantSDNode *CP) {
303 EVT VT = CP->getValueType(0);
304 SDValue CPIdx = DAG.getConstantPool(CP->getConstantIntValue(),
305 TLI.getPointerTy(DAG.getDataLayout()));
306 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
307 SDValue Result = DAG.getLoad(
308 VT, dl, DAG.getEntryNode(), CPIdx,
309 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Alignment);
313 /// Some target cannot handle a variable insertion index for the
314 /// INSERT_VECTOR_ELT instruction. In this case, it
315 /// is necessary to spill the vector being inserted into to memory, perform
316 /// the insert there, and then read the result back.
317 SDValue SelectionDAGLegalize::PerformInsertVectorEltInMemory(SDValue Vec,
325 // If the target doesn't support this, we have to spill the input vector
326 // to a temporary stack slot, update the element, then reload it. This is
327 // badness. We could also load the value into a vector register (either
328 // with a "move to register" or "extload into register" instruction, then
329 // permute it into place, if the idx is a constant and if the idx is
330 // supported by the target.
331 EVT VT = Tmp1.getValueType();
332 EVT EltVT = VT.getVectorElementType();
333 SDValue StackPtr = DAG.CreateStackTemporary(VT);
335 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
338 SDValue Ch = DAG.getStore(
339 DAG.getEntryNode(), dl, Tmp1, StackPtr,
340 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI));
342 SDValue StackPtr2 = TLI.getVectorElementPointer(DAG, StackPtr, VT, Tmp3);
344 // Store the scalar value.
345 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT);
346 // Load the updated vector.
347 return DAG.getLoad(VT, dl, Ch, StackPtr, MachinePointerInfo::getFixedStack(
348 DAG.getMachineFunction(), SPFI));
351 SDValue SelectionDAGLegalize::ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
354 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
355 // SCALAR_TO_VECTOR requires that the type of the value being inserted
356 // match the element type of the vector being created, except for
357 // integers in which case the inserted value can be over width.
358 EVT EltVT = Vec.getValueType().getVectorElementType();
359 if (Val.getValueType() == EltVT ||
360 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
361 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
362 Vec.getValueType(), Val);
364 unsigned NumElts = Vec.getValueType().getVectorNumElements();
365 // We generate a shuffle of InVec and ScVec, so the shuffle mask
366 // should be 0,1,2,3,4,5... with the appropriate element replaced with
368 SmallVector<int, 8> ShufOps;
369 for (unsigned i = 0; i != NumElts; ++i)
370 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
372 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec, ShufOps);
375 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
378 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
379 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
380 // FIXME: We shouldn't do this for TargetConstantFP's.
381 // FIXME: move this to the DAG Combiner! Note that we can't regress due
382 // to phase ordering between legalized code and the dag combiner. This
383 // probably means that we need to integrate dag combiner and legalizer
385 // We generally can't do this one for long doubles.
386 SDValue Chain = ST->getChain();
387 SDValue Ptr = ST->getBasePtr();
388 unsigned Alignment = ST->getAlignment();
389 MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags();
390 AAMDNodes AAInfo = ST->getAAInfo();
392 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
393 if (CFP->getValueType(0) == MVT::f32 &&
394 TLI.isTypeLegal(MVT::i32)) {
395 SDValue Con = DAG.getConstant(CFP->getValueAPF().
396 bitcastToAPInt().zextOrTrunc(32),
397 SDLoc(CFP), MVT::i32);
398 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(), Alignment,
402 if (CFP->getValueType(0) == MVT::f64) {
403 // If this target supports 64-bit registers, do a single 64-bit store.
404 if (TLI.isTypeLegal(MVT::i64)) {
405 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
406 zextOrTrunc(64), SDLoc(CFP), MVT::i64);
407 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
408 Alignment, MMOFlags, AAInfo);
411 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
412 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
413 // stores. If the target supports neither 32- nor 64-bits, this
414 // xform is certainly not worth it.
415 const APInt &IntVal = CFP->getValueAPF().bitcastToAPInt();
416 SDValue Lo = DAG.getConstant(IntVal.trunc(32), dl, MVT::i32);
417 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), dl, MVT::i32);
418 if (DAG.getDataLayout().isBigEndian())
421 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), Alignment,
423 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
424 DAG.getConstant(4, dl, Ptr.getValueType()));
425 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
426 ST->getPointerInfo().getWithOffset(4),
427 MinAlign(Alignment, 4U), MMOFlags, AAInfo);
429 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
433 return SDValue(nullptr, 0);
436 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
437 StoreSDNode *ST = cast<StoreSDNode>(Node);
438 SDValue Chain = ST->getChain();
439 SDValue Ptr = ST->getBasePtr();
442 unsigned Alignment = ST->getAlignment();
443 MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags();
444 AAMDNodes AAInfo = ST->getAAInfo();
446 if (!ST->isTruncatingStore()) {
447 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
448 ReplaceNode(ST, OptStore);
453 SDValue Value = ST->getValue();
454 MVT VT = Value.getSimpleValueType();
455 switch (TLI.getOperationAction(ISD::STORE, VT)) {
456 default: llvm_unreachable("This action is not supported yet!");
457 case TargetLowering::Legal: {
458 // If this is an unaligned store and the target doesn't support it,
460 EVT MemVT = ST->getMemoryVT();
461 unsigned AS = ST->getAddressSpace();
462 unsigned Align = ST->getAlignment();
463 const DataLayout &DL = DAG.getDataLayout();
464 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
465 SDValue Result = TLI.expandUnalignedStore(ST, DAG);
466 ReplaceNode(SDValue(ST, 0), Result);
470 case TargetLowering::Custom: {
471 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
472 if (Res && Res != SDValue(Node, 0))
473 ReplaceNode(SDValue(Node, 0), Res);
476 case TargetLowering::Promote: {
477 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
478 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
479 "Can only promote stores to same size type");
480 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
482 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
483 Alignment, MMOFlags, AAInfo);
484 ReplaceNode(SDValue(Node, 0), Result);
491 SDValue Value = ST->getValue();
493 EVT StVT = ST->getMemoryVT();
494 unsigned StWidth = StVT.getSizeInBits();
495 auto &DL = DAG.getDataLayout();
497 if (StWidth != StVT.getStoreSizeInBits()) {
498 // Promote to a byte-sized store with upper bits zero if not
499 // storing an integral number of bytes. For example, promote
500 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
501 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
502 StVT.getStoreSizeInBits());
503 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
505 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(), NVT,
506 Alignment, MMOFlags, AAInfo);
507 ReplaceNode(SDValue(Node, 0), Result);
508 } else if (StWidth & (StWidth - 1)) {
509 // If not storing a power-of-2 number of bits, expand as two stores.
510 assert(!StVT.isVector() && "Unsupported truncstore!");
511 unsigned RoundWidth = 1 << Log2_32(StWidth);
512 assert(RoundWidth < StWidth);
513 unsigned ExtraWidth = StWidth - RoundWidth;
514 assert(ExtraWidth < RoundWidth);
515 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
516 "Store size not an integral number of bytes!");
517 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
518 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
520 unsigned IncrementSize;
522 if (DL.isLittleEndian()) {
523 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
524 // Store the bottom RoundWidth bits.
525 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
526 RoundVT, Alignment, MMOFlags, AAInfo);
528 // Store the remaining ExtraWidth bits.
529 IncrementSize = RoundWidth / 8;
530 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
531 DAG.getConstant(IncrementSize, dl,
532 Ptr.getValueType()));
534 ISD::SRL, dl, Value.getValueType(), Value,
535 DAG.getConstant(RoundWidth, dl,
536 TLI.getShiftAmountTy(Value.getValueType(), DL)));
537 Hi = DAG.getTruncStore(
539 ST->getPointerInfo().getWithOffset(IncrementSize), ExtraVT,
540 MinAlign(Alignment, IncrementSize), MMOFlags, AAInfo);
542 // Big endian - avoid unaligned stores.
543 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
544 // Store the top RoundWidth bits.
546 ISD::SRL, dl, Value.getValueType(), Value,
547 DAG.getConstant(ExtraWidth, dl,
548 TLI.getShiftAmountTy(Value.getValueType(), DL)));
549 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
550 RoundVT, Alignment, MMOFlags, AAInfo);
552 // Store the remaining ExtraWidth bits.
553 IncrementSize = RoundWidth / 8;
554 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
555 DAG.getConstant(IncrementSize, dl,
556 Ptr.getValueType()));
557 Lo = DAG.getTruncStore(
558 Chain, dl, Value, Ptr,
559 ST->getPointerInfo().getWithOffset(IncrementSize), ExtraVT,
560 MinAlign(Alignment, IncrementSize), MMOFlags, AAInfo);
563 // The order of the stores doesn't matter.
564 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
565 ReplaceNode(SDValue(Node, 0), Result);
567 switch (TLI.getTruncStoreAction(ST->getValue().getValueType(), StVT)) {
568 default: llvm_unreachable("This action is not supported yet!");
569 case TargetLowering::Legal: {
570 EVT MemVT = ST->getMemoryVT();
571 unsigned AS = ST->getAddressSpace();
572 unsigned Align = ST->getAlignment();
573 // If this is an unaligned store and the target doesn't support it,
575 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
576 SDValue Result = TLI.expandUnalignedStore(ST, DAG);
577 ReplaceNode(SDValue(ST, 0), Result);
581 case TargetLowering::Custom: {
582 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
583 if (Res && Res != SDValue(Node, 0))
584 ReplaceNode(SDValue(Node, 0), Res);
587 case TargetLowering::Expand:
588 assert(!StVT.isVector() &&
589 "Vector Stores are handled in LegalizeVectorOps");
591 // TRUNCSTORE:i16 i32 -> STORE i16
592 assert(TLI.isTypeLegal(StVT) &&
593 "Do not know how to expand this store!");
594 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
596 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
597 Alignment, MMOFlags, AAInfo);
598 ReplaceNode(SDValue(Node, 0), Result);
605 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
606 LoadSDNode *LD = cast<LoadSDNode>(Node);
607 SDValue Chain = LD->getChain(); // The chain.
608 SDValue Ptr = LD->getBasePtr(); // The base pointer.
609 SDValue Value; // The value returned by the load op.
612 ISD::LoadExtType ExtType = LD->getExtensionType();
613 if (ExtType == ISD::NON_EXTLOAD) {
614 MVT VT = Node->getSimpleValueType(0);
615 SDValue RVal = SDValue(Node, 0);
616 SDValue RChain = SDValue(Node, 1);
618 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
619 default: llvm_unreachable("This action is not supported yet!");
620 case TargetLowering::Legal: {
621 EVT MemVT = LD->getMemoryVT();
622 unsigned AS = LD->getAddressSpace();
623 unsigned Align = LD->getAlignment();
624 const DataLayout &DL = DAG.getDataLayout();
625 // If this is an unaligned load and the target doesn't support it,
627 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
628 std::tie(RVal, RChain) = TLI.expandUnalignedLoad(LD, DAG);
632 case TargetLowering::Custom: {
633 if (SDValue Res = TLI.LowerOperation(RVal, DAG)) {
635 RChain = Res.getValue(1);
639 case TargetLowering::Promote: {
640 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
641 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
642 "Can only promote loads to same size type");
644 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
645 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
646 RChain = Res.getValue(1);
650 if (RChain.getNode() != Node) {
651 assert(RVal.getNode() != Node && "Load must be completely replaced");
652 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
653 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
655 UpdatedNodes->insert(RVal.getNode());
656 UpdatedNodes->insert(RChain.getNode());
663 EVT SrcVT = LD->getMemoryVT();
664 unsigned SrcWidth = SrcVT.getSizeInBits();
665 unsigned Alignment = LD->getAlignment();
666 MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags();
667 AAMDNodes AAInfo = LD->getAAInfo();
669 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
670 // Some targets pretend to have an i1 loading operation, and actually
671 // load an i8. This trick is correct for ZEXTLOAD because the top 7
672 // bits are guaranteed to be zero; it helps the optimizers understand
673 // that these bits are zero. It is also useful for EXTLOAD, since it
674 // tells the optimizers that those bits are undefined. It would be
675 // nice to have an effective generic way of getting these benefits...
676 // Until such a way is found, don't insist on promoting i1 here.
678 TLI.getLoadExtAction(ExtType, Node->getValueType(0), MVT::i1) ==
679 TargetLowering::Promote)) {
680 // Promote to a byte-sized load if not loading an integral number of
681 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
682 unsigned NewWidth = SrcVT.getStoreSizeInBits();
683 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
686 // The extra bits are guaranteed to be zero, since we stored them that
687 // way. A zext load from NVT thus automatically gives zext from SrcVT.
689 ISD::LoadExtType NewExtType =
690 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
693 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0), Chain, Ptr,
694 LD->getPointerInfo(), NVT, Alignment, MMOFlags, AAInfo);
696 Ch = Result.getValue(1); // The chain.
698 if (ExtType == ISD::SEXTLOAD)
699 // Having the top bits zero doesn't help when sign extending.
700 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
701 Result.getValueType(),
702 Result, DAG.getValueType(SrcVT));
703 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
704 // All the top bits are guaranteed to be zero - inform the optimizers.
705 Result = DAG.getNode(ISD::AssertZext, dl,
706 Result.getValueType(), Result,
707 DAG.getValueType(SrcVT));
711 } else if (SrcWidth & (SrcWidth - 1)) {
712 // If not loading a power-of-2 number of bits, expand as two loads.
713 assert(!SrcVT.isVector() && "Unsupported extload!");
714 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
715 assert(RoundWidth < SrcWidth);
716 unsigned ExtraWidth = SrcWidth - RoundWidth;
717 assert(ExtraWidth < RoundWidth);
718 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
719 "Load size not an integral number of bytes!");
720 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
721 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
723 unsigned IncrementSize;
724 auto &DL = DAG.getDataLayout();
726 if (DL.isLittleEndian()) {
727 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
728 // Load the bottom RoundWidth bits.
729 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), Chain, Ptr,
730 LD->getPointerInfo(), RoundVT, Alignment, MMOFlags,
733 // Load the remaining ExtraWidth bits.
734 IncrementSize = RoundWidth / 8;
735 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
736 DAG.getConstant(IncrementSize, dl,
737 Ptr.getValueType()));
738 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
739 LD->getPointerInfo().getWithOffset(IncrementSize),
740 ExtraVT, MinAlign(Alignment, IncrementSize), MMOFlags,
743 // Build a factor node to remember that this load is independent of
745 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
748 // Move the top bits to the right place.
750 ISD::SHL, dl, Hi.getValueType(), Hi,
751 DAG.getConstant(RoundWidth, dl,
752 TLI.getShiftAmountTy(Hi.getValueType(), DL)));
754 // Join the hi and lo parts.
755 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
757 // Big endian - avoid unaligned loads.
758 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
759 // Load the top RoundWidth bits.
760 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
761 LD->getPointerInfo(), RoundVT, Alignment, MMOFlags,
764 // Load the remaining ExtraWidth bits.
765 IncrementSize = RoundWidth / 8;
766 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
767 DAG.getConstant(IncrementSize, dl,
768 Ptr.getValueType()));
769 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), Chain, Ptr,
770 LD->getPointerInfo().getWithOffset(IncrementSize),
771 ExtraVT, MinAlign(Alignment, IncrementSize), MMOFlags,
774 // Build a factor node to remember that this load is independent of
776 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
779 // Move the top bits to the right place.
781 ISD::SHL, dl, Hi.getValueType(), Hi,
782 DAG.getConstant(ExtraWidth, dl,
783 TLI.getShiftAmountTy(Hi.getValueType(), DL)));
785 // Join the hi and lo parts.
786 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
791 bool isCustom = false;
792 switch (TLI.getLoadExtAction(ExtType, Node->getValueType(0),
793 SrcVT.getSimpleVT())) {
794 default: llvm_unreachable("This action is not supported yet!");
795 case TargetLowering::Custom:
798 case TargetLowering::Legal: {
799 Value = SDValue(Node, 0);
800 Chain = SDValue(Node, 1);
803 if (SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG)) {
805 Chain = Res.getValue(1);
808 // If this is an unaligned load and the target doesn't support it,
810 EVT MemVT = LD->getMemoryVT();
811 unsigned AS = LD->getAddressSpace();
812 unsigned Align = LD->getAlignment();
813 const DataLayout &DL = DAG.getDataLayout();
814 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
815 std::tie(Value, Chain) = TLI.expandUnalignedLoad(LD, DAG);
820 case TargetLowering::Expand:
821 EVT DestVT = Node->getValueType(0);
822 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, DestVT, SrcVT)) {
823 // If the source type is not legal, see if there is a legal extload to
824 // an intermediate type that we can then extend further.
825 EVT LoadVT = TLI.getRegisterType(SrcVT.getSimpleVT());
826 if (TLI.isTypeLegal(SrcVT) || // Same as SrcVT == LoadVT?
827 TLI.isLoadExtLegal(ExtType, LoadVT, SrcVT)) {
828 // If we are loading a legal type, this is a non-extload followed by a
830 ISD::LoadExtType MidExtType =
831 (LoadVT == SrcVT) ? ISD::NON_EXTLOAD : ExtType;
833 SDValue Load = DAG.getExtLoad(MidExtType, dl, LoadVT, Chain, Ptr,
834 SrcVT, LD->getMemOperand());
836 ISD::getExtForLoadExtType(SrcVT.isFloatingPoint(), ExtType);
837 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
838 Chain = Load.getValue(1);
842 // Handle the special case of fp16 extloads. EXTLOAD doesn't have the
843 // normal undefined upper bits behavior to allow using an in-reg extend
844 // with the illegal FP type, so load as an integer and do the
845 // from-integer conversion.
846 if (SrcVT.getScalarType() == MVT::f16) {
847 EVT ISrcVT = SrcVT.changeTypeToInteger();
848 EVT IDestVT = DestVT.changeTypeToInteger();
849 EVT LoadVT = TLI.getRegisterType(IDestVT.getSimpleVT());
851 SDValue Result = DAG.getExtLoad(ISD::ZEXTLOAD, dl, LoadVT,
853 LD->getMemOperand());
854 Value = DAG.getNode(ISD::FP16_TO_FP, dl, DestVT, Result);
855 Chain = Result.getValue(1);
860 assert(!SrcVT.isVector() &&
861 "Vector Loads are handled in LegalizeVectorOps");
863 // FIXME: This does not work for vectors on most targets. Sign-
864 // and zero-extend operations are currently folded into extending
865 // loads, whether they are legal or not, and then we end up here
866 // without any support for legalizing them.
867 assert(ExtType != ISD::EXTLOAD &&
868 "EXTLOAD should always be supported!");
869 // Turn the unsupported load into an EXTLOAD followed by an
870 // explicit zero/sign extend inreg.
871 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
872 Node->getValueType(0),
874 LD->getMemOperand());
876 if (ExtType == ISD::SEXTLOAD)
877 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
878 Result.getValueType(),
879 Result, DAG.getValueType(SrcVT));
881 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
883 Chain = Result.getValue(1);
888 // Since loads produce two values, make sure to remember that we legalized
890 if (Chain.getNode() != Node) {
891 assert(Value.getNode() != Node && "Load must be completely replaced");
892 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
893 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
895 UpdatedNodes->insert(Value.getNode());
896 UpdatedNodes->insert(Chain.getNode());
902 static TargetLowering::LegalizeAction
903 getStrictFPOpcodeAction(const TargetLowering &TLI, unsigned Opcode, EVT VT) {
906 default: llvm_unreachable("Unexpected FP pseudo-opcode");
907 case ISD::STRICT_FSQRT: EqOpc = ISD::FSQRT; break;
908 case ISD::STRICT_FPOW: EqOpc = ISD::FPOW; break;
909 case ISD::STRICT_FPOWI: EqOpc = ISD::FPOWI; break;
910 case ISD::STRICT_FSIN: EqOpc = ISD::FSIN; break;
911 case ISD::STRICT_FCOS: EqOpc = ISD::FCOS; break;
912 case ISD::STRICT_FEXP: EqOpc = ISD::FEXP; break;
913 case ISD::STRICT_FEXP2: EqOpc = ISD::FEXP2; break;
914 case ISD::STRICT_FLOG: EqOpc = ISD::FLOG; break;
915 case ISD::STRICT_FLOG10: EqOpc = ISD::FLOG10; break;
916 case ISD::STRICT_FLOG2: EqOpc = ISD::FLOG2; break;
917 case ISD::STRICT_FRINT: EqOpc = ISD::FRINT; break;
918 case ISD::STRICT_FNEARBYINT: EqOpc = ISD::FNEARBYINT; break;
921 auto Action = TLI.getOperationAction(EqOpc, VT);
923 // We don't currently handle Custom or Promote for strict FP pseudo-ops.
924 // For now, we just expand for those cases.
925 if (Action != TargetLowering::Legal)
926 Action = TargetLowering::Expand;
928 // ISD::FPOWI returns 'Legal' even though it should be expanded.
929 if (Opcode == ISD::STRICT_FPOWI && Action == TargetLowering::Legal)
930 Action = TargetLowering::Expand;
935 /// Return a legal replacement for the given operation, with all legal operands.
936 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
937 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
939 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
943 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
944 assert((TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
945 TargetLowering::TypeLegal ||
946 TLI.isTypeLegal(Node->getValueType(i))) &&
947 "Unexpected illegal type!");
949 for (const SDValue &Op : Node->op_values())
950 assert((TLI.getTypeAction(*DAG.getContext(), Op.getValueType()) ==
951 TargetLowering::TypeLegal ||
952 TLI.isTypeLegal(Op.getValueType()) ||
953 Op.getOpcode() == ISD::TargetConstant) &&
954 "Unexpected illegal type!");
957 // Figure out the correct action; the way to query this varies by opcode
958 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
959 bool SimpleFinishLegalizing = true;
960 switch (Node->getOpcode()) {
961 case ISD::INTRINSIC_W_CHAIN:
962 case ISD::INTRINSIC_WO_CHAIN:
963 case ISD::INTRINSIC_VOID:
965 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
967 case ISD::GET_DYNAMIC_AREA_OFFSET:
968 Action = TLI.getOperationAction(Node->getOpcode(),
969 Node->getValueType(0));
972 Action = TLI.getOperationAction(Node->getOpcode(),
973 Node->getValueType(0));
974 if (Action != TargetLowering::Promote)
975 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
977 case ISD::FP_TO_FP16:
978 case ISD::SINT_TO_FP:
979 case ISD::UINT_TO_FP:
980 case ISD::EXTRACT_VECTOR_ELT:
981 Action = TLI.getOperationAction(Node->getOpcode(),
982 Node->getOperand(0).getValueType());
984 case ISD::FP_ROUND_INREG:
985 case ISD::SIGN_EXTEND_INREG: {
986 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
987 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
990 case ISD::ATOMIC_STORE: {
991 Action = TLI.getOperationAction(Node->getOpcode(),
992 Node->getOperand(2).getValueType());
998 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
999 Node->getOpcode() == ISD::SETCC ? 2 :
1000 Node->getOpcode() == ISD::SETCCE ? 3 : 1;
1001 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1002 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1003 ISD::CondCode CCCode =
1004 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1005 Action = TLI.getCondCodeAction(CCCode, OpVT);
1006 if (Action == TargetLowering::Legal) {
1007 if (Node->getOpcode() == ISD::SELECT_CC)
1008 Action = TLI.getOperationAction(Node->getOpcode(),
1009 Node->getValueType(0));
1011 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1017 // FIXME: Model these properly. LOAD and STORE are complicated, and
1018 // STORE expects the unlegalized operand in some cases.
1019 SimpleFinishLegalizing = false;
1021 case ISD::CALLSEQ_START:
1022 case ISD::CALLSEQ_END:
1023 // FIXME: This shouldn't be necessary. These nodes have special properties
1024 // dealing with the recursive nature of legalization. Removing this
1025 // special case should be done as part of making LegalizeDAG non-recursive.
1026 SimpleFinishLegalizing = false;
1028 case ISD::EXTRACT_ELEMENT:
1029 case ISD::FLT_ROUNDS_:
1031 case ISD::MERGE_VALUES:
1032 case ISD::EH_RETURN:
1033 case ISD::FRAME_TO_ARGS_OFFSET:
1034 case ISD::EH_DWARF_CFA:
1035 case ISD::EH_SJLJ_SETJMP:
1036 case ISD::EH_SJLJ_LONGJMP:
1037 case ISD::EH_SJLJ_SETUP_DISPATCH:
1038 // These operations lie about being legal: when they claim to be legal,
1039 // they should actually be expanded.
1040 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1041 if (Action == TargetLowering::Legal)
1042 Action = TargetLowering::Expand;
1044 case ISD::INIT_TRAMPOLINE:
1045 case ISD::ADJUST_TRAMPOLINE:
1046 case ISD::FRAMEADDR:
1047 case ISD::RETURNADDR:
1048 case ISD::ADDROFRETURNADDR:
1049 // These operations lie about being legal: when they claim to be legal,
1050 // they should actually be custom-lowered.
1051 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1052 if (Action == TargetLowering::Legal)
1053 Action = TargetLowering::Custom;
1055 case ISD::READCYCLECOUNTER:
1056 // READCYCLECOUNTER returns an i64, even if type legalization might have
1057 // expanded that to several smaller types.
1058 Action = TLI.getOperationAction(Node->getOpcode(), MVT::i64);
1060 case ISD::READ_REGISTER:
1061 case ISD::WRITE_REGISTER:
1062 // Named register is legal in the DAG, but blocked by register name
1063 // selection if not implemented by target (to chose the correct register)
1064 // They'll be converted to Copy(To/From)Reg.
1065 Action = TargetLowering::Legal;
1067 case ISD::DEBUGTRAP:
1068 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1069 if (Action == TargetLowering::Expand) {
1070 // replace ISD::DEBUGTRAP with ISD::TRAP
1072 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1073 Node->getOperand(0));
1074 ReplaceNode(Node, NewVal.getNode());
1075 LegalizeOp(NewVal.getNode());
1079 case ISD::STRICT_FSQRT:
1080 case ISD::STRICT_FPOW:
1081 case ISD::STRICT_FPOWI:
1082 case ISD::STRICT_FSIN:
1083 case ISD::STRICT_FCOS:
1084 case ISD::STRICT_FEXP:
1085 case ISD::STRICT_FEXP2:
1086 case ISD::STRICT_FLOG:
1087 case ISD::STRICT_FLOG10:
1088 case ISD::STRICT_FLOG2:
1089 case ISD::STRICT_FRINT:
1090 case ISD::STRICT_FNEARBYINT:
1091 // These pseudo-ops get legalized as if they were their non-strict
1092 // equivalent. For instance, if ISD::FSQRT is legal then ISD::STRICT_FSQRT
1093 // is also legal, but if ISD::FSQRT requires expansion then so does
1094 // ISD::STRICT_FSQRT.
1095 Action = getStrictFPOpcodeAction(TLI, Node->getOpcode(),
1096 Node->getValueType(0));
1100 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1101 Action = TargetLowering::Legal;
1103 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1108 if (SimpleFinishLegalizing) {
1109 SDNode *NewNode = Node;
1110 switch (Node->getOpcode()) {
1117 // Legalizing shifts/rotates requires adjusting the shift amount
1118 // to the appropriate width.
1119 SDValue Op0 = Node->getOperand(0);
1120 SDValue Op1 = Node->getOperand(1);
1121 if (!Op1.getValueType().isVector()) {
1122 SDValue SAO = DAG.getShiftAmountOperand(Op0.getValueType(), Op1);
1123 // The getShiftAmountOperand() may create a new operand node or
1124 // return the existing one. If new operand is created we need
1125 // to update the parent node.
1126 // Do not try to legalize SAO here! It will be automatically legalized
1127 // in the next round.
1129 NewNode = DAG.UpdateNodeOperands(Node, Op0, SAO);
1133 case ISD::SRL_PARTS:
1134 case ISD::SRA_PARTS:
1135 case ISD::SHL_PARTS: {
1136 // Legalizing shifts/rotates requires adjusting the shift amount
1137 // to the appropriate width.
1138 SDValue Op0 = Node->getOperand(0);
1139 SDValue Op1 = Node->getOperand(1);
1140 SDValue Op2 = Node->getOperand(2);
1141 if (!Op2.getValueType().isVector()) {
1142 SDValue SAO = DAG.getShiftAmountOperand(Op0.getValueType(), Op2);
1143 // The getShiftAmountOperand() may create a new operand node or
1144 // return the existing one. If new operand is created we need
1145 // to update the parent node.
1147 NewNode = DAG.UpdateNodeOperands(Node, Op0, Op1, SAO);
1153 if (NewNode != Node) {
1154 ReplaceNode(Node, NewNode);
1158 case TargetLowering::Legal:
1160 case TargetLowering::Custom: {
1161 // FIXME: The handling for custom lowering with multiple results is
1163 if (SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG)) {
1164 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1167 if (Node->getNumValues() == 1) {
1168 // We can just directly replace this node with the lowered value.
1169 ReplaceNode(SDValue(Node, 0), Res);
1173 SmallVector<SDValue, 8> ResultVals;
1174 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1175 ResultVals.push_back(Res.getValue(i));
1176 ReplaceNode(Node, ResultVals.data());
1181 case TargetLowering::Expand:
1182 if (ExpandNode(Node))
1185 case TargetLowering::LibCall:
1186 ConvertNodeToLibcall(Node);
1188 case TargetLowering::Promote:
1194 switch (Node->getOpcode()) {
1201 llvm_unreachable("Do not know how to legalize this operator!");
1203 case ISD::CALLSEQ_START:
1204 case ISD::CALLSEQ_END:
1207 return LegalizeLoadOps(Node);
1210 return LegalizeStoreOps(Node);
1215 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1216 SDValue Vec = Op.getOperand(0);
1217 SDValue Idx = Op.getOperand(1);
1220 // Before we generate a new store to a temporary stack slot, see if there is
1221 // already one that we can use. There often is because when we scalarize
1222 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1223 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1224 // the vector. If all are expanded here, we don't want one store per vector
1227 // Caches for hasPredecessorHelper
1228 SmallPtrSet<const SDNode *, 32> Visited;
1229 SmallVector<const SDNode *, 16> Worklist;
1230 Worklist.push_back(Idx.getNode());
1231 SDValue StackPtr, Ch;
1232 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1233 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1235 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1236 if (ST->isIndexed() || ST->isTruncatingStore() ||
1237 ST->getValue() != Vec)
1240 // Make sure that nothing else could have stored into the destination of
1242 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1245 // If the index is dependent on the store we will introduce a cycle when
1246 // creating the load (the load uses the index, and by replacing the chain
1247 // we will make the index dependent on the load). Also, the store might be
1248 // dependent on the extractelement and introduce a cycle when creating
1250 if (SDNode::hasPredecessorHelper(ST, Visited, Worklist) ||
1251 ST->hasPredecessor(Op.getNode()))
1254 StackPtr = ST->getBasePtr();
1255 Ch = SDValue(ST, 0);
1260 EVT VecVT = Vec.getValueType();
1262 if (!Ch.getNode()) {
1263 // Store the value to a temporary stack slot, then LOAD the returned part.
1264 StackPtr = DAG.CreateStackTemporary(VecVT);
1265 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1266 MachinePointerInfo());
1269 StackPtr = TLI.getVectorElementPointer(DAG, StackPtr, VecVT, Idx);
1273 if (Op.getValueType().isVector())
1275 DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, MachinePointerInfo());
1277 NewLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1278 MachinePointerInfo(),
1279 VecVT.getVectorElementType());
1281 // Replace the chain going out of the store, by the one out of the load.
1282 DAG.ReplaceAllUsesOfValueWith(Ch, SDValue(NewLoad.getNode(), 1));
1284 // We introduced a cycle though, so update the loads operands, making sure
1285 // to use the original store's chain as an incoming chain.
1286 SmallVector<SDValue, 6> NewLoadOperands(NewLoad->op_begin(),
1288 NewLoadOperands[0] = Ch;
1290 SDValue(DAG.UpdateNodeOperands(NewLoad.getNode(), NewLoadOperands), 0);
1294 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1295 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1297 SDValue Vec = Op.getOperand(0);
1298 SDValue Part = Op.getOperand(1);
1299 SDValue Idx = Op.getOperand(2);
1302 // Store the value to a temporary stack slot, then LOAD the returned part.
1303 EVT VecVT = Vec.getValueType();
1304 SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
1305 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1306 MachinePointerInfo PtrInfo =
1307 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
1309 // First store the whole vector.
1310 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo);
1312 // Then store the inserted part.
1313 SDValue SubStackPtr = TLI.getVectorElementPointer(DAG, StackPtr, VecVT, Idx);
1315 // Store the subvector.
1316 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr, MachinePointerInfo());
1318 // Finally, load the updated vector.
1319 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo);
1322 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1323 // We can't handle this case efficiently. Allocate a sufficiently
1324 // aligned object on the stack, store each element into it, then load
1325 // the result as a vector.
1326 // Create the stack frame object.
1327 EVT VT = Node->getValueType(0);
1328 EVT EltVT = VT.getVectorElementType();
1330 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1331 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1332 MachinePointerInfo PtrInfo =
1333 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
1335 // Emit a store of each element to the stack slot.
1336 SmallVector<SDValue, 8> Stores;
1337 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1338 // Store (in the right endianness) the elements to memory.
1339 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1340 // Ignore undef elements.
1341 if (Node->getOperand(i).isUndef()) continue;
1343 unsigned Offset = TypeByteSize*i;
1345 SDValue Idx = DAG.getConstant(Offset, dl, FIPtr.getValueType());
1346 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1348 // If the destination vector element type is narrower than the source
1349 // element type, only store the bits necessary.
1350 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1351 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1352 Node->getOperand(i), Idx,
1353 PtrInfo.getWithOffset(Offset), EltVT));
1355 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl, Node->getOperand(i),
1356 Idx, PtrInfo.getWithOffset(Offset)));
1360 if (!Stores.empty()) // Not all undef elements?
1361 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1363 StoreChain = DAG.getEntryNode();
1365 // Result is a load from the stack slot.
1366 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo);
1370 /// Keeps track of state when getting the sign of a floating-point value as an
1372 struct FloatSignAsInt {
1377 MachinePointerInfo IntPointerInfo;
1378 MachinePointerInfo FloatPointerInfo;
1385 /// Bitcast a floating-point value to an integer value. Only bitcast the part
1386 /// containing the sign bit if the target has no integer value capable of
1387 /// holding all bits of the floating-point value.
1388 void SelectionDAGLegalize::getSignAsIntValue(FloatSignAsInt &State,
1390 SDValue Value) const {
1391 EVT FloatVT = Value.getValueType();
1392 unsigned NumBits = FloatVT.getSizeInBits();
1393 State.FloatVT = FloatVT;
1394 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), NumBits);
1395 // Convert to an integer of the same size.
1396 if (TLI.isTypeLegal(IVT)) {
1397 State.IntValue = DAG.getNode(ISD::BITCAST, DL, IVT, Value);
1398 State.SignMask = APInt::getSignMask(NumBits);
1399 State.SignBit = NumBits - 1;
1403 auto &DataLayout = DAG.getDataLayout();
1404 // Store the float to memory, then load the sign part out as an integer.
1405 MVT LoadTy = TLI.getRegisterType(*DAG.getContext(), MVT::i8);
1406 // First create a temporary that is aligned for both the load and store.
1407 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1408 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1409 // Then store the float to it.
1410 State.FloatPtr = StackPtr;
1411 MachineFunction &MF = DAG.getMachineFunction();
1412 State.FloatPointerInfo = MachinePointerInfo::getFixedStack(MF, FI);
1413 State.Chain = DAG.getStore(DAG.getEntryNode(), DL, Value, State.FloatPtr,
1414 State.FloatPointerInfo);
1417 if (DataLayout.isBigEndian()) {
1418 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1419 // Load out a legal integer with the same sign bit as the float.
1421 State.IntPointerInfo = State.FloatPointerInfo;
1423 // Advance the pointer so that the loaded byte will contain the sign bit.
1424 unsigned ByteOffset = (FloatVT.getSizeInBits() / 8) - 1;
1425 IntPtr = DAG.getNode(ISD::ADD, DL, StackPtr.getValueType(), StackPtr,
1426 DAG.getConstant(ByteOffset, DL, StackPtr.getValueType()));
1427 State.IntPointerInfo = MachinePointerInfo::getFixedStack(MF, FI,
1431 State.IntPtr = IntPtr;
1432 State.IntValue = DAG.getExtLoad(ISD::EXTLOAD, DL, LoadTy, State.Chain, IntPtr,
1433 State.IntPointerInfo, MVT::i8);
1434 State.SignMask = APInt::getOneBitSet(LoadTy.getSizeInBits(), 7);
1438 /// Replace the integer value produced by getSignAsIntValue() with a new value
1439 /// and cast the result back to a floating-point type.
1440 SDValue SelectionDAGLegalize::modifySignAsInt(const FloatSignAsInt &State,
1442 SDValue NewIntValue) const {
1444 return DAG.getNode(ISD::BITCAST, DL, State.FloatVT, NewIntValue);
1446 // Override the part containing the sign bit in the value stored on the stack.
1447 SDValue Chain = DAG.getTruncStore(State.Chain, DL, NewIntValue, State.IntPtr,
1448 State.IntPointerInfo, MVT::i8);
1449 return DAG.getLoad(State.FloatVT, DL, Chain, State.FloatPtr,
1450 State.FloatPointerInfo);
1453 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode *Node) const {
1455 SDValue Mag = Node->getOperand(0);
1456 SDValue Sign = Node->getOperand(1);
1458 // Get sign bit into an integer value.
1459 FloatSignAsInt SignAsInt;
1460 getSignAsIntValue(SignAsInt, DL, Sign);
1462 EVT IntVT = SignAsInt.IntValue.getValueType();
1463 SDValue SignMask = DAG.getConstant(SignAsInt.SignMask, DL, IntVT);
1464 SDValue SignBit = DAG.getNode(ISD::AND, DL, IntVT, SignAsInt.IntValue,
1467 // If FABS is legal transform FCOPYSIGN(x, y) => sign(x) ? -FABS(x) : FABS(X)
1468 EVT FloatVT = Mag.getValueType();
1469 if (TLI.isOperationLegalOrCustom(ISD::FABS, FloatVT) &&
1470 TLI.isOperationLegalOrCustom(ISD::FNEG, FloatVT)) {
1471 SDValue AbsValue = DAG.getNode(ISD::FABS, DL, FloatVT, Mag);
1472 SDValue NegValue = DAG.getNode(ISD::FNEG, DL, FloatVT, AbsValue);
1473 SDValue Cond = DAG.getSetCC(DL, getSetCCResultType(IntVT), SignBit,
1474 DAG.getConstant(0, DL, IntVT), ISD::SETNE);
1475 return DAG.getSelect(DL, FloatVT, Cond, NegValue, AbsValue);
1478 // Transform Mag value to integer, and clear the sign bit.
1479 FloatSignAsInt MagAsInt;
1480 getSignAsIntValue(MagAsInt, DL, Mag);
1481 EVT MagVT = MagAsInt.IntValue.getValueType();
1482 SDValue ClearSignMask = DAG.getConstant(~MagAsInt.SignMask, DL, MagVT);
1483 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, MagVT, MagAsInt.IntValue,
1486 // Get the signbit at the right position for MagAsInt.
1487 int ShiftAmount = SignAsInt.SignBit - MagAsInt.SignBit;
1488 if (SignBit.getValueSizeInBits() > ClearedSign.getValueSizeInBits()) {
1489 if (ShiftAmount > 0) {
1490 SDValue ShiftCnst = DAG.getConstant(ShiftAmount, DL, IntVT);
1491 SignBit = DAG.getNode(ISD::SRL, DL, IntVT, SignBit, ShiftCnst);
1492 } else if (ShiftAmount < 0) {
1493 SDValue ShiftCnst = DAG.getConstant(-ShiftAmount, DL, IntVT);
1494 SignBit = DAG.getNode(ISD::SHL, DL, IntVT, SignBit, ShiftCnst);
1496 SignBit = DAG.getNode(ISD::TRUNCATE, DL, MagVT, SignBit);
1497 } else if (SignBit.getValueSizeInBits() < ClearedSign.getValueSizeInBits()) {
1498 SignBit = DAG.getNode(ISD::ZERO_EXTEND, DL, MagVT, SignBit);
1499 if (ShiftAmount > 0) {
1500 SDValue ShiftCnst = DAG.getConstant(ShiftAmount, DL, MagVT);
1501 SignBit = DAG.getNode(ISD::SRL, DL, MagVT, SignBit, ShiftCnst);
1502 } else if (ShiftAmount < 0) {
1503 SDValue ShiftCnst = DAG.getConstant(-ShiftAmount, DL, MagVT);
1504 SignBit = DAG.getNode(ISD::SHL, DL, MagVT, SignBit, ShiftCnst);
1508 // Store the part with the modified sign and convert back to float.
1509 SDValue CopiedSign = DAG.getNode(ISD::OR, DL, MagVT, ClearedSign, SignBit);
1510 return modifySignAsInt(MagAsInt, DL, CopiedSign);
1513 SDValue SelectionDAGLegalize::ExpandFABS(SDNode *Node) const {
1515 SDValue Value = Node->getOperand(0);
1517 // Transform FABS(x) => FCOPYSIGN(x, 0.0) if FCOPYSIGN is legal.
1518 EVT FloatVT = Value.getValueType();
1519 if (TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, FloatVT)) {
1520 SDValue Zero = DAG.getConstantFP(0.0, DL, FloatVT);
1521 return DAG.getNode(ISD::FCOPYSIGN, DL, FloatVT, Value, Zero);
1524 // Transform value to integer, clear the sign bit and transform back.
1525 FloatSignAsInt ValueAsInt;
1526 getSignAsIntValue(ValueAsInt, DL, Value);
1527 EVT IntVT = ValueAsInt.IntValue.getValueType();
1528 SDValue ClearSignMask = DAG.getConstant(~ValueAsInt.SignMask, DL, IntVT);
1529 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, ValueAsInt.IntValue,
1531 return modifySignAsInt(ValueAsInt, DL, ClearedSign);
1534 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1535 SmallVectorImpl<SDValue> &Results) {
1536 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1537 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1538 " not tell us which reg is the stack pointer!");
1540 EVT VT = Node->getValueType(0);
1541 SDValue Tmp1 = SDValue(Node, 0);
1542 SDValue Tmp2 = SDValue(Node, 1);
1543 SDValue Tmp3 = Node->getOperand(2);
1544 SDValue Chain = Tmp1.getOperand(0);
1546 // Chain the dynamic stack allocation so that it doesn't modify the stack
1547 // pointer when other instructions are using the stack.
1548 Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl);
1550 SDValue Size = Tmp2.getOperand(1);
1551 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1552 Chain = SP.getValue(1);
1553 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1554 unsigned StackAlign =
1555 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
1556 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1557 if (Align > StackAlign)
1558 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1559 DAG.getConstant(-(uint64_t)Align, dl, VT));
1560 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1562 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
1563 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
1565 Results.push_back(Tmp1);
1566 Results.push_back(Tmp2);
1569 /// Legalize a SETCC with given LHS and RHS and condition code CC on the current
1572 /// If the SETCC has been legalized using AND / OR, then the legalized node
1573 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1574 /// will be set to false.
1576 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1577 /// then the values of LHS and RHS will be swapped, CC will be set to the
1578 /// new condition, and NeedInvert will be set to false.
1580 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1581 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1582 /// will be set to true. The caller must invert the result of the SETCC with
1583 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1584 /// of a true/false result.
1586 /// \returns true if the SetCC has been legalized, false if it hasn't.
1587 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT, SDValue &LHS,
1588 SDValue &RHS, SDValue &CC,
1591 MVT OpVT = LHS.getSimpleValueType();
1592 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1594 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1595 default: llvm_unreachable("Unknown condition code action!");
1596 case TargetLowering::Legal:
1599 case TargetLowering::Expand: {
1600 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1601 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1602 std::swap(LHS, RHS);
1603 CC = DAG.getCondCode(InvCC);
1606 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1609 default: llvm_unreachable("Don't know how to expand this condition!");
1611 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1612 == TargetLowering::Legal
1613 && "If SETO is expanded, SETOEQ must be legal!");
1614 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1616 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1617 == TargetLowering::Legal
1618 && "If SETUO is expanded, SETUNE must be legal!");
1619 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1632 // If we are floating point, assign and break, otherwise fall through.
1633 if (!OpVT.isInteger()) {
1634 // We can use the 4th bit to tell if we are the unordered
1635 // or ordered version of the opcode.
1636 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1637 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1638 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1641 // Fallthrough if we are unsigned integer.
1647 // We only support using the inverted operation, which is computed above
1648 // and not a different manner of supporting expanding these cases.
1649 llvm_unreachable("Don't know how to expand this condition!");
1652 // Try inverting the result of the inverse condition.
1653 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1654 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1655 CC = DAG.getCondCode(InvCC);
1659 // If inverting the condition didn't work then we have no means to expand
1661 llvm_unreachable("Don't know how to expand this condition!");
1664 SDValue SetCC1, SetCC2;
1665 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1666 // If we aren't the ordered or unorder operation,
1667 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1668 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1669 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1671 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1672 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1673 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1675 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1684 /// Emit a store/load combination to the stack. This stores
1685 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1686 /// a load from the stack slot to DestVT, extending it if needed.
1687 /// The resultant code need not be legal.
1688 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, EVT SlotVT,
1689 EVT DestVT, const SDLoc &dl) {
1690 // Create the stack frame object.
1691 unsigned SrcAlign = DAG.getDataLayout().getPrefTypeAlignment(
1692 SrcOp.getValueType().getTypeForEVT(*DAG.getContext()));
1693 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1695 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1696 int SPFI = StackPtrFI->getIndex();
1697 MachinePointerInfo PtrInfo =
1698 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI);
1700 unsigned SrcSize = SrcOp.getValueSizeInBits();
1701 unsigned SlotSize = SlotVT.getSizeInBits();
1702 unsigned DestSize = DestVT.getSizeInBits();
1703 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1704 unsigned DestAlign = DAG.getDataLayout().getPrefTypeAlignment(DestType);
1706 // Emit a store to the stack slot. Use a truncstore if the input value is
1707 // later than DestVT.
1710 if (SrcSize > SlotSize)
1711 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, PtrInfo,
1714 assert(SrcSize == SlotSize && "Invalid store");
1716 DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, PtrInfo, SrcAlign);
1719 // Result is a load from the stack slot.
1720 if (SlotSize == DestSize)
1721 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo, DestAlign);
1723 assert(SlotSize < DestSize && "Unknown extension!");
1724 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr, PtrInfo, SlotVT,
1728 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
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(
1738 DAG.getEntryNode(), dl, Node->getOperand(0), StackPtr,
1739 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI),
1740 Node->getValueType(0).getVectorElementType());
1742 Node->getValueType(0), dl, Ch, StackPtr,
1743 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI));
1747 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1748 const TargetLowering &TLI, SDValue &Res) {
1749 unsigned NumElems = Node->getNumOperands();
1751 EVT VT = Node->getValueType(0);
1753 // Try to group the scalars into pairs, shuffle the pairs together, then
1754 // shuffle the pairs of pairs together, etc. until the vector has
1755 // been built. This will work only if all of the necessary shuffle masks
1758 // We do this in two phases; first to check the legality of the shuffles,
1759 // and next, assuming that all shuffles are legal, to create the new nodes.
1760 for (int Phase = 0; Phase < 2; ++Phase) {
1761 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1763 for (unsigned i = 0; i < NumElems; ++i) {
1764 SDValue V = Node->getOperand(i);
1770 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1771 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1774 while (IntermedVals.size() > 2) {
1775 NewIntermedVals.clear();
1776 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1777 // This vector and the next vector are shuffled together (simply to
1778 // append the one to the other).
1779 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1781 SmallVector<int, 16> FinalIndices;
1782 FinalIndices.reserve(IntermedVals[i].second.size() +
1783 IntermedVals[i+1].second.size());
1786 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1789 FinalIndices.push_back(IntermedVals[i].second[j]);
1791 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1793 ShuffleVec[k] = NumElems + j;
1794 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1799 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1800 IntermedVals[i+1].first,
1802 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1804 NewIntermedVals.push_back(
1805 std::make_pair(Shuffle, std::move(FinalIndices)));
1808 // If we had an odd number of defined values, then append the last
1809 // element to the array of new vectors.
1810 if ((IntermedVals.size() & 1) != 0)
1811 NewIntermedVals.push_back(IntermedVals.back());
1813 IntermedVals.swap(NewIntermedVals);
1816 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1817 "Invalid number of intermediate vectors");
1818 SDValue Vec1 = IntermedVals[0].first;
1820 if (IntermedVals.size() > 1)
1821 Vec2 = IntermedVals[1].first;
1823 Vec2 = DAG.getUNDEF(VT);
1825 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1826 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1827 ShuffleVec[IntermedVals[0].second[i]] = i;
1828 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1829 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1832 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec);
1833 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1840 /// Expand a BUILD_VECTOR node on targets that don't
1841 /// support the operation, but do support the resultant vector type.
1842 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1843 unsigned NumElems = Node->getNumOperands();
1844 SDValue Value1, Value2;
1846 EVT VT = Node->getValueType(0);
1847 EVT OpVT = Node->getOperand(0).getValueType();
1848 EVT EltVT = VT.getVectorElementType();
1850 // If the only non-undef value is the low element, turn this into a
1851 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1852 bool isOnlyLowElement = true;
1853 bool MoreThanTwoValues = false;
1854 bool isConstant = true;
1855 for (unsigned i = 0; i < NumElems; ++i) {
1856 SDValue V = Node->getOperand(i);
1860 isOnlyLowElement = false;
1861 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1864 if (!Value1.getNode()) {
1866 } else if (!Value2.getNode()) {
1869 } else if (V != Value1 && V != Value2) {
1870 MoreThanTwoValues = true;
1874 if (!Value1.getNode())
1875 return DAG.getUNDEF(VT);
1877 if (isOnlyLowElement)
1878 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1880 // If all elements are constants, create a load from the constant pool.
1882 SmallVector<Constant*, 16> CV;
1883 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1884 if (ConstantFPSDNode *V =
1885 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1886 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1887 } else if (ConstantSDNode *V =
1888 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1890 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1892 // If OpVT and EltVT don't match, EltVT is not legal and the
1893 // element values have been promoted/truncated earlier. Undo this;
1894 // we don't want a v16i8 to become a v16i32 for example.
1895 const ConstantInt *CI = V->getConstantIntValue();
1896 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1897 CI->getZExtValue()));
1900 assert(Node->getOperand(i).isUndef());
1901 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1902 CV.push_back(UndefValue::get(OpNTy));
1905 Constant *CP = ConstantVector::get(CV);
1907 DAG.getConstantPool(CP, TLI.getPointerTy(DAG.getDataLayout()));
1908 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1910 VT, dl, DAG.getEntryNode(), CPIdx,
1911 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
1915 SmallSet<SDValue, 16> DefinedValues;
1916 for (unsigned i = 0; i < NumElems; ++i) {
1917 if (Node->getOperand(i).isUndef())
1919 DefinedValues.insert(Node->getOperand(i));
1922 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
1923 if (!MoreThanTwoValues) {
1924 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1925 for (unsigned i = 0; i < NumElems; ++i) {
1926 SDValue V = Node->getOperand(i);
1929 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1931 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1932 // Get the splatted value into the low element of a vector register.
1933 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1935 if (Value2.getNode())
1936 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1938 Vec2 = DAG.getUNDEF(VT);
1940 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1941 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec);
1945 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
1950 // Otherwise, we can't handle this case efficiently.
1951 return ExpandVectorBuildThroughStack(Node);
1954 // Expand a node into a call to a libcall. If the result value
1955 // does not fit into a register, return the lo part and set the hi part to the
1956 // by-reg argument. If it does fit into a single register, return the result
1957 // and leave the Hi part unset.
1958 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
1960 TargetLowering::ArgListTy Args;
1961 TargetLowering::ArgListEntry Entry;
1962 for (const SDValue &Op : Node->op_values()) {
1963 EVT ArgVT = Op.getValueType();
1964 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1967 Entry.IsSExt = isSigned;
1968 Entry.IsZExt = !isSigned;
1969 Args.push_back(Entry);
1971 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
1972 TLI.getPointerTy(DAG.getDataLayout()));
1974 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
1976 // By default, the input chain to this libcall is the entry node of the
1977 // function. If the libcall is going to be emitted as a tail call then
1978 // TLI.isUsedByReturnOnly will change it to the right chain if the return
1979 // node which is being folded has a non-entry input chain.
1980 SDValue InChain = DAG.getEntryNode();
1982 // isTailCall may be true since the callee does not reference caller stack
1983 // frame. Check if it's in the right position and that the return types match.
1984 SDValue TCChain = InChain;
1985 const Function *F = DAG.getMachineFunction().getFunction();
1987 TLI.isInTailCallPosition(DAG, Node, TCChain) &&
1988 (RetTy == F->getReturnType() || F->getReturnType()->isVoidTy());
1992 TargetLowering::CallLoweringInfo CLI(DAG);
1993 CLI.setDebugLoc(SDLoc(Node))
1995 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
1997 .setTailCall(isTailCall)
1998 .setSExtResult(isSigned)
1999 .setZExtResult(!isSigned);
2001 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2003 if (!CallInfo.second.getNode())
2004 // It's a tailcall, return the chain (which is the DAG root).
2005 return DAG.getRoot();
2007 return CallInfo.first;
2010 /// Generate a libcall taking the given operands as arguments
2011 /// and returning a result of type RetVT.
2012 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2013 const SDValue *Ops, unsigned NumOps,
2014 bool isSigned, const SDLoc &dl) {
2015 TargetLowering::ArgListTy Args;
2016 Args.reserve(NumOps);
2018 TargetLowering::ArgListEntry Entry;
2019 for (unsigned i = 0; i != NumOps; ++i) {
2020 Entry.Node = Ops[i];
2021 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2022 Entry.IsSExt = isSigned;
2023 Entry.IsZExt = !isSigned;
2024 Args.push_back(Entry);
2026 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2027 TLI.getPointerTy(DAG.getDataLayout()));
2029 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2031 TargetLowering::CallLoweringInfo CLI(DAG);
2033 .setChain(DAG.getEntryNode())
2034 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2036 .setSExtResult(isSigned)
2037 .setZExtResult(!isSigned);
2039 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2041 return CallInfo.first;
2044 // Expand a node into a call to a libcall. Similar to
2045 // ExpandLibCall except that the first operand is the in-chain.
2046 std::pair<SDValue, SDValue>
2047 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2050 SDValue InChain = Node->getOperand(0);
2052 TargetLowering::ArgListTy Args;
2053 TargetLowering::ArgListEntry Entry;
2054 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2055 EVT ArgVT = Node->getOperand(i).getValueType();
2056 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2057 Entry.Node = Node->getOperand(i);
2059 Entry.IsSExt = isSigned;
2060 Entry.IsZExt = !isSigned;
2061 Args.push_back(Entry);
2063 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2064 TLI.getPointerTy(DAG.getDataLayout()));
2066 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2068 TargetLowering::CallLoweringInfo CLI(DAG);
2069 CLI.setDebugLoc(SDLoc(Node))
2071 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2073 .setSExtResult(isSigned)
2074 .setZExtResult(!isSigned);
2076 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2081 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2082 RTLIB::Libcall Call_F32,
2083 RTLIB::Libcall Call_F64,
2084 RTLIB::Libcall Call_F80,
2085 RTLIB::Libcall Call_F128,
2086 RTLIB::Libcall Call_PPCF128) {
2087 if (Node->isStrictFPOpcode())
2088 Node = DAG.mutateStrictFPToFP(Node);
2091 switch (Node->getSimpleValueType(0).SimpleTy) {
2092 default: llvm_unreachable("Unexpected request for libcall!");
2093 case MVT::f32: LC = Call_F32; break;
2094 case MVT::f64: LC = Call_F64; break;
2095 case MVT::f80: LC = Call_F80; break;
2096 case MVT::f128: LC = Call_F128; break;
2097 case MVT::ppcf128: LC = Call_PPCF128; break;
2099 return ExpandLibCall(LC, Node, false);
2102 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2103 RTLIB::Libcall Call_I8,
2104 RTLIB::Libcall Call_I16,
2105 RTLIB::Libcall Call_I32,
2106 RTLIB::Libcall Call_I64,
2107 RTLIB::Libcall Call_I128) {
2109 switch (Node->getSimpleValueType(0).SimpleTy) {
2110 default: llvm_unreachable("Unexpected request for libcall!");
2111 case MVT::i8: LC = Call_I8; break;
2112 case MVT::i16: LC = Call_I16; break;
2113 case MVT::i32: LC = Call_I32; break;
2114 case MVT::i64: LC = Call_I64; break;
2115 case MVT::i128: LC = Call_I128; break;
2117 return ExpandLibCall(LC, Node, isSigned);
2120 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
2122 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2123 SmallVectorImpl<SDValue> &Results) {
2124 unsigned Opcode = Node->getOpcode();
2125 bool isSigned = Opcode == ISD::SDIVREM;
2128 switch (Node->getSimpleValueType(0).SimpleTy) {
2129 default: llvm_unreachable("Unexpected request for libcall!");
2130 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2131 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2132 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2133 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2134 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2137 // The input chain to this libcall is the entry node of the function.
2138 // Legalizing the call will automatically add the previous call to the
2140 SDValue InChain = DAG.getEntryNode();
2142 EVT RetVT = Node->getValueType(0);
2143 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2145 TargetLowering::ArgListTy Args;
2146 TargetLowering::ArgListEntry Entry;
2147 for (const SDValue &Op : Node->op_values()) {
2148 EVT ArgVT = Op.getValueType();
2149 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2152 Entry.IsSExt = isSigned;
2153 Entry.IsZExt = !isSigned;
2154 Args.push_back(Entry);
2157 // Also pass the return address of the remainder.
2158 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2160 Entry.Ty = RetTy->getPointerTo();
2161 Entry.IsSExt = isSigned;
2162 Entry.IsZExt = !isSigned;
2163 Args.push_back(Entry);
2165 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2166 TLI.getPointerTy(DAG.getDataLayout()));
2169 TargetLowering::CallLoweringInfo CLI(DAG);
2172 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2174 .setSExtResult(isSigned)
2175 .setZExtResult(!isSigned);
2177 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2179 // Remainder is loaded back from the stack frame.
2181 DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr, MachinePointerInfo());
2182 Results.push_back(CallInfo.first);
2183 Results.push_back(Rem);
2186 /// Return true if sincos libcall is available.
2187 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2189 switch (Node->getSimpleValueType(0).SimpleTy) {
2190 default: llvm_unreachable("Unexpected request for libcall!");
2191 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2192 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2193 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2194 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2195 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2197 return TLI.getLibcallName(LC) != nullptr;
2200 /// Return true if sincos libcall is available and can be used to combine sin
2202 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2203 const TargetMachine &TM) {
2204 if (!isSinCosLibcallAvailable(Node, TLI))
2206 // GNU sin/cos functions set errno while sincos does not. Therefore
2207 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2208 if (TM.getTargetTriple().isGNUEnvironment() && !TM.Options.UnsafeFPMath)
2213 /// Only issue sincos libcall if both sin and cos are needed.
2214 static bool useSinCos(SDNode *Node) {
2215 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2216 ? ISD::FCOS : ISD::FSIN;
2218 SDValue Op0 = Node->getOperand(0);
2219 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2220 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2224 // The other user might have been turned into sincos already.
2225 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2231 /// Issue libcalls to sincos to compute sin / cos pairs.
2233 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2234 SmallVectorImpl<SDValue> &Results) {
2236 switch (Node->getSimpleValueType(0).SimpleTy) {
2237 default: llvm_unreachable("Unexpected request for libcall!");
2238 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2239 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2240 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2241 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2242 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2245 // The input chain to this libcall is the entry node of the function.
2246 // Legalizing the call will automatically add the previous call to the
2248 SDValue InChain = DAG.getEntryNode();
2250 EVT RetVT = Node->getValueType(0);
2251 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2253 TargetLowering::ArgListTy Args;
2254 TargetLowering::ArgListEntry Entry;
2256 // Pass the argument.
2257 Entry.Node = Node->getOperand(0);
2259 Entry.IsSExt = false;
2260 Entry.IsZExt = false;
2261 Args.push_back(Entry);
2263 // Pass the return address of sin.
2264 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2265 Entry.Node = SinPtr;
2266 Entry.Ty = RetTy->getPointerTo();
2267 Entry.IsSExt = false;
2268 Entry.IsZExt = false;
2269 Args.push_back(Entry);
2271 // Also pass the return address of the cos.
2272 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2273 Entry.Node = CosPtr;
2274 Entry.Ty = RetTy->getPointerTo();
2275 Entry.IsSExt = false;
2276 Entry.IsZExt = false;
2277 Args.push_back(Entry);
2279 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2280 TLI.getPointerTy(DAG.getDataLayout()));
2283 TargetLowering::CallLoweringInfo CLI(DAG);
2284 CLI.setDebugLoc(dl).setChain(InChain).setLibCallee(
2285 TLI.getLibcallCallingConv(LC), Type::getVoidTy(*DAG.getContext()), Callee,
2288 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2291 DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr, MachinePointerInfo()));
2293 DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr, MachinePointerInfo()));
2296 /// This function is responsible for legalizing a
2297 /// INT_TO_FP operation of the specified operand when the target requests that
2298 /// we expand it. At this point, we know that the result and operand types are
2299 /// legal for the target.
2300 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned, SDValue Op0,
2303 // TODO: Should any fast-math-flags be set for the created nodes?
2305 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2306 // simple 32-bit [signed|unsigned] integer to float/double expansion
2308 // Get the stack frame index of a 8 byte buffer.
2309 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2311 // word offset constant for Hi/Lo address computation
2312 SDValue WordOff = DAG.getConstant(sizeof(int), dl,
2313 StackSlot.getValueType());
2314 // set up Hi and Lo (into buffer) address based on endian
2315 SDValue Hi = StackSlot;
2316 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2317 StackSlot, WordOff);
2318 if (DAG.getDataLayout().isLittleEndian())
2321 // if signed map to unsigned space
2324 // constant used to invert sign bit (signed to unsigned mapping)
2325 SDValue SignBit = DAG.getConstant(0x80000000u, dl, MVT::i32);
2326 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2330 // store the lo of the constructed double - based on integer input
2331 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op0Mapped, Lo,
2332 MachinePointerInfo());
2333 // initial hi portion of constructed double
2334 SDValue InitialHi = DAG.getConstant(0x43300000u, dl, MVT::i32);
2335 // store the hi of the constructed double - biased exponent
2337 DAG.getStore(Store1, dl, InitialHi, Hi, MachinePointerInfo());
2338 // load the constructed double
2340 DAG.getLoad(MVT::f64, dl, Store2, StackSlot, MachinePointerInfo());
2341 // FP constant to bias correct the final result
2342 SDValue Bias = DAG.getConstantFP(isSigned ?
2343 BitsToDouble(0x4330000080000000ULL) :
2344 BitsToDouble(0x4330000000000000ULL),
2346 // subtract the bias
2347 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2350 // handle final rounding
2351 if (DestVT == MVT::f64) {
2354 } else if (DestVT.bitsLT(MVT::f64)) {
2355 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2356 DAG.getIntPtrConstant(0, dl));
2357 } else if (DestVT.bitsGT(MVT::f64)) {
2358 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2362 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2363 // Code below here assumes !isSigned without checking again.
2365 // Implementation of unsigned i64 to f64 following the algorithm in
2366 // __floatundidf in compiler_rt. This implementation has the advantage
2367 // of performing rounding correctly, both in the default rounding mode
2368 // and in all alternate rounding modes.
2369 // TODO: Generalize this for use with other types.
2370 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2372 DAG.getConstant(UINT64_C(0x4330000000000000), dl, MVT::i64);
2373 SDValue TwoP84PlusTwoP52 =
2374 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), dl,
2377 DAG.getConstant(UINT64_C(0x4530000000000000), dl, MVT::i64);
2379 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2380 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2381 DAG.getConstant(32, dl, MVT::i64));
2382 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2383 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2384 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2385 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2386 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2388 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2391 // Implementation of unsigned i64 to f32.
2392 // TODO: Generalize this for use with other types.
2393 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2394 // For unsigned conversions, convert them to signed conversions using the
2395 // algorithm from the x86_64 __floatundidf in compiler_rt.
2397 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2399 SDValue ShiftConst = DAG.getConstant(
2400 1, dl, TLI.getShiftAmountTy(Op0.getValueType(), DAG.getDataLayout()));
2401 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2402 SDValue AndConst = DAG.getConstant(1, dl, MVT::i64);
2403 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2404 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2406 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2407 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2409 // TODO: This really should be implemented using a branch rather than a
2410 // select. We happen to get lucky and machinesink does the right
2411 // thing most of the time. This would be a good candidate for a
2412 //pseudo-op, or, even better, for whole-function isel.
2413 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2414 Op0, DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
2415 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2418 // Otherwise, implement the fully general conversion.
2420 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2421 DAG.getConstant(UINT64_C(0xfffffffffffff800), dl, MVT::i64));
2422 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2423 DAG.getConstant(UINT64_C(0x800), dl, MVT::i64));
2424 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2425 DAG.getConstant(UINT64_C(0x7ff), dl, MVT::i64));
2426 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), And2,
2427 DAG.getConstant(UINT64_C(0), dl, MVT::i64),
2429 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2430 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), Op0,
2431 DAG.getConstant(UINT64_C(0x0020000000000000), dl,
2434 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2435 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType(), DAG.getDataLayout());
2437 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2438 DAG.getConstant(32, dl, SHVT));
2439 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2440 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2442 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), dl,
2444 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2445 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2446 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2447 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2448 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2449 DAG.getIntPtrConstant(0, dl));
2452 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2454 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2456 DAG.getConstant(0, dl, Op0.getValueType()),
2458 SDValue Zero = DAG.getIntPtrConstant(0, dl),
2459 Four = DAG.getIntPtrConstant(4, dl);
2460 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2461 SignSet, Four, Zero);
2463 // If the sign bit of the integer is set, the large number will be treated
2464 // as a negative number. To counteract this, the dynamic code adds an
2465 // offset depending on the data type.
2467 switch (Op0.getSimpleValueType().SimpleTy) {
2468 default: llvm_unreachable("Unsupported integer type!");
2469 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2470 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2471 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2472 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2474 if (DAG.getDataLayout().isLittleEndian())
2476 Constant *FudgeFactor = ConstantInt::get(
2477 Type::getInt64Ty(*DAG.getContext()), FF);
2480 DAG.getConstantPool(FudgeFactor, TLI.getPointerTy(DAG.getDataLayout()));
2481 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2482 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2483 Alignment = std::min(Alignment, 4u);
2485 if (DestVT == MVT::f32)
2486 FudgeInReg = DAG.getLoad(
2487 MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2488 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2491 SDValue Load = DAG.getExtLoad(
2492 ISD::EXTLOAD, dl, DestVT, DAG.getEntryNode(), CPIdx,
2493 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
2495 HandleSDNode Handle(Load);
2496 LegalizeOp(Load.getNode());
2497 FudgeInReg = Handle.getValue();
2500 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2503 /// This function is responsible for legalizing a
2504 /// *INT_TO_FP operation of the specified operand when the target requests that
2505 /// we promote it. At this point, we know that the result and operand types are
2506 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2507 /// operation that takes a larger input.
2508 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT,
2511 // First step, figure out the appropriate *INT_TO_FP operation to use.
2512 EVT NewInTy = LegalOp.getValueType();
2514 unsigned OpToUse = 0;
2516 // Scan for the appropriate larger type to use.
2518 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2519 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2521 // If the target supports SINT_TO_FP of this type, use it.
2522 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2523 OpToUse = ISD::SINT_TO_FP;
2526 if (isSigned) continue;
2528 // If the target supports UINT_TO_FP of this type, use it.
2529 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2530 OpToUse = ISD::UINT_TO_FP;
2534 // Otherwise, try a larger type.
2537 // Okay, we found the operation and type to use. Zero extend our input to the
2538 // desired type then run the operation on it.
2539 return DAG.getNode(OpToUse, dl, DestVT,
2540 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2541 dl, NewInTy, LegalOp));
2544 /// This function is responsible for legalizing a
2545 /// FP_TO_*INT operation of the specified operand when the target requests that
2546 /// we promote it. At this point, we know that the result and operand types are
2547 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2548 /// operation that returns a larger result.
2549 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT,
2552 // First step, figure out the appropriate FP_TO*INT operation to use.
2553 EVT NewOutTy = DestVT;
2555 unsigned OpToUse = 0;
2557 // Scan for the appropriate larger type to use.
2559 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2560 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2562 // A larger signed type can hold all unsigned values of the requested type,
2563 // so using FP_TO_SINT is valid
2564 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2565 OpToUse = ISD::FP_TO_SINT;
2569 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2570 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2571 OpToUse = ISD::FP_TO_UINT;
2575 // Otherwise, try a larger type.
2579 // Okay, we found the operation and type to use.
2580 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2582 // Truncate the result of the extended FP_TO_*INT operation to the desired
2584 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2587 /// Legalize a BITREVERSE scalar/vector operation as a series of mask + shifts.
2588 SDValue SelectionDAGLegalize::ExpandBITREVERSE(SDValue Op, const SDLoc &dl) {
2589 EVT VT = Op.getValueType();
2590 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2591 unsigned Sz = VT.getScalarSizeInBits();
2593 SDValue Tmp, Tmp2, Tmp3;
2595 // If we can, perform BSWAP first and then the mask+swap the i4, then i2
2596 // and finally the i1 pairs.
2597 // TODO: We can easily support i4/i2 legal types if any target ever does.
2598 if (Sz >= 8 && isPowerOf2_32(Sz)) {
2599 // Create the masks - repeating the pattern every byte.
2600 APInt MaskHi4(Sz, 0), MaskHi2(Sz, 0), MaskHi1(Sz, 0);
2601 APInt MaskLo4(Sz, 0), MaskLo2(Sz, 0), MaskLo1(Sz, 0);
2602 for (unsigned J = 0; J != Sz; J += 8) {
2603 MaskHi4 = MaskHi4 | (0xF0ull << J);
2604 MaskLo4 = MaskLo4 | (0x0Full << J);
2605 MaskHi2 = MaskHi2 | (0xCCull << J);
2606 MaskLo2 = MaskLo2 | (0x33ull << J);
2607 MaskHi1 = MaskHi1 | (0xAAull << J);
2608 MaskLo1 = MaskLo1 | (0x55ull << J);
2611 // BSWAP if the type is wider than a single byte.
2612 Tmp = (Sz > 8 ? DAG.getNode(ISD::BSWAP, dl, VT, Op) : Op);
2614 // swap i4: ((V & 0xF0) >> 4) | ((V & 0x0F) << 4)
2615 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskHi4, dl, VT));
2616 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskLo4, dl, VT));
2617 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp2, DAG.getConstant(4, dl, VT));
2618 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(4, dl, VT));
2619 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
2621 // swap i2: ((V & 0xCC) >> 2) | ((V & 0x33) << 2)
2622 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskHi2, dl, VT));
2623 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskLo2, dl, VT));
2624 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp2, DAG.getConstant(2, dl, VT));
2625 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(2, dl, VT));
2626 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
2628 // swap i1: ((V & 0xAA) >> 1) | ((V & 0x55) << 1)
2629 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskHi1, dl, VT));
2630 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskLo1, dl, VT));
2631 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp2, DAG.getConstant(1, dl, VT));
2632 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(1, dl, VT));
2633 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
2637 Tmp = DAG.getConstant(0, dl, VT);
2638 for (unsigned I = 0, J = Sz-1; I < Sz; ++I, --J) {
2641 DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(J - I, dl, SHVT));
2644 DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(I - J, dl, SHVT));
2648 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Shift, dl, VT));
2649 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp, Tmp2);
2655 /// Open code the operations for BSWAP of the specified operation.
2656 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, const SDLoc &dl) {
2657 EVT VT = Op.getValueType();
2658 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2659 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2660 switch (VT.getSimpleVT().getScalarType().SimpleTy) {
2661 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2663 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2664 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2665 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2667 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2668 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2669 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2670 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2671 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2672 DAG.getConstant(0xFF0000, dl, VT));
2673 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT));
2674 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2675 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2676 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2678 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2679 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2680 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2681 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2682 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2683 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2684 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2685 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2686 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7,
2687 DAG.getConstant(255ULL<<48, dl, VT));
2688 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6,
2689 DAG.getConstant(255ULL<<40, dl, VT));
2690 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5,
2691 DAG.getConstant(255ULL<<32, dl, VT));
2692 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4,
2693 DAG.getConstant(255ULL<<24, dl, VT));
2694 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2695 DAG.getConstant(255ULL<<16, dl, VT));
2696 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2,
2697 DAG.getConstant(255ULL<<8 , dl, VT));
2698 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2699 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2700 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2701 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2702 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2703 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2704 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2708 /// Expand the specified bitcount instruction into operations.
2709 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2712 default: llvm_unreachable("Cannot expand this yet!");
2714 EVT VT = Op.getValueType();
2715 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2716 unsigned Len = VT.getSizeInBits();
2718 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2719 "CTPOP not implemented for this type.");
2721 // This is the "best" algorithm from
2722 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2724 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)),
2726 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)),
2728 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)),
2730 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)),
2733 // v = v - ((v >> 1) & 0x55555555...)
2734 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2735 DAG.getNode(ISD::AND, dl, VT,
2736 DAG.getNode(ISD::SRL, dl, VT, Op,
2737 DAG.getConstant(1, dl, ShVT)),
2739 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2740 Op = DAG.getNode(ISD::ADD, dl, VT,
2741 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2742 DAG.getNode(ISD::AND, dl, VT,
2743 DAG.getNode(ISD::SRL, dl, VT, Op,
2744 DAG.getConstant(2, dl, ShVT)),
2746 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2747 Op = DAG.getNode(ISD::AND, dl, VT,
2748 DAG.getNode(ISD::ADD, dl, VT, Op,
2749 DAG.getNode(ISD::SRL, dl, VT, Op,
2750 DAG.getConstant(4, dl, ShVT))),
2752 // v = (v * 0x01010101...) >> (Len - 8)
2753 Op = DAG.getNode(ISD::SRL, dl, VT,
2754 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2755 DAG.getConstant(Len - 8, dl, ShVT));
2759 case ISD::CTLZ_ZERO_UNDEF:
2760 // This trivially expands to CTLZ.
2761 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2763 EVT VT = Op.getValueType();
2764 unsigned len = VT.getSizeInBits();
2766 if (TLI.isOperationLegalOrCustom(ISD::CTLZ_ZERO_UNDEF, VT)) {
2767 EVT SetCCVT = getSetCCResultType(VT);
2768 SDValue CTLZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, dl, VT, Op);
2769 SDValue Zero = DAG.getConstant(0, dl, VT);
2770 SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ);
2771 return DAG.getNode(ISD::SELECT, dl, VT, SrcIsZero,
2772 DAG.getConstant(len, dl, VT), CTLZ);
2775 // for now, we do this:
2776 // x = x | (x >> 1);
2777 // x = x | (x >> 2);
2779 // x = x | (x >>16);
2780 // x = x | (x >>32); // for 64-bit input
2781 // return popcount(~x);
2783 // Ref: "Hacker's Delight" by Henry Warren
2784 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2785 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2786 SDValue Tmp3 = DAG.getConstant(1ULL << i, dl, ShVT);
2787 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2788 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2790 Op = DAG.getNOT(dl, Op, VT);
2791 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2793 case ISD::CTTZ_ZERO_UNDEF:
2794 // This trivially expands to CTTZ.
2795 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2797 // for now, we use: { return popcount(~x & (x - 1)); }
2798 // unless the target has ctlz but not ctpop, in which case we use:
2799 // { return 32 - nlz(~x & (x-1)); }
2800 // Ref: "Hacker's Delight" by Henry Warren
2801 EVT VT = Op.getValueType();
2802 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2803 DAG.getNOT(dl, Op, VT),
2804 DAG.getNode(ISD::SUB, dl, VT, Op,
2805 DAG.getConstant(1, dl, VT)));
2806 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2807 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2808 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2809 return DAG.getNode(ISD::SUB, dl, VT,
2810 DAG.getConstant(VT.getSizeInBits(), dl, VT),
2811 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2812 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2817 bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2818 SmallVector<SDValue, 8> Results;
2820 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2822 switch (Node->getOpcode()) {
2825 case ISD::CTLZ_ZERO_UNDEF:
2827 case ISD::CTTZ_ZERO_UNDEF:
2828 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2829 Results.push_back(Tmp1);
2831 case ISD::BITREVERSE:
2832 Results.push_back(ExpandBITREVERSE(Node->getOperand(0), dl));
2835 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2837 case ISD::FRAMEADDR:
2838 case ISD::RETURNADDR:
2839 case ISD::FRAME_TO_ARGS_OFFSET:
2840 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0)));
2842 case ISD::EH_DWARF_CFA: {
2843 SDValue CfaArg = DAG.getSExtOrTrunc(Node->getOperand(0), dl,
2844 TLI.getPointerTy(DAG.getDataLayout()));
2845 SDValue Offset = DAG.getNode(ISD::ADD, dl,
2846 CfaArg.getValueType(),
2847 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
2848 CfaArg.getValueType()),
2850 SDValue FA = DAG.getNode(
2851 ISD::FRAMEADDR, dl, TLI.getPointerTy(DAG.getDataLayout()),
2852 DAG.getConstant(0, dl, TLI.getPointerTy(DAG.getDataLayout())));
2853 Results.push_back(DAG.getNode(ISD::ADD, dl, FA.getValueType(),
2857 case ISD::FLT_ROUNDS_:
2858 Results.push_back(DAG.getConstant(1, dl, Node->getValueType(0)));
2860 case ISD::EH_RETURN:
2864 case ISD::EH_SJLJ_LONGJMP:
2865 // If the target didn't expand these, there's nothing to do, so just
2866 // preserve the chain and be done.
2867 Results.push_back(Node->getOperand(0));
2869 case ISD::READCYCLECOUNTER:
2870 // If the target didn't expand this, just return 'zero' and preserve the
2872 Results.append(Node->getNumValues() - 1,
2873 DAG.getConstant(0, dl, Node->getValueType(0)));
2874 Results.push_back(Node->getOperand(0));
2876 case ISD::EH_SJLJ_SETJMP:
2877 // If the target didn't expand this, just return 'zero' and preserve the
2879 Results.push_back(DAG.getConstant(0, dl, MVT::i32));
2880 Results.push_back(Node->getOperand(0));
2882 case ISD::ATOMIC_LOAD: {
2883 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2884 SDValue Zero = DAG.getConstant(0, dl, Node->getValueType(0));
2885 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
2886 SDValue Swap = DAG.getAtomicCmpSwap(
2887 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
2888 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
2889 cast<AtomicSDNode>(Node)->getMemOperand());
2890 Results.push_back(Swap.getValue(0));
2891 Results.push_back(Swap.getValue(1));
2894 case ISD::ATOMIC_STORE: {
2895 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2896 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2897 cast<AtomicSDNode>(Node)->getMemoryVT(),
2898 Node->getOperand(0),
2899 Node->getOperand(1), Node->getOperand(2),
2900 cast<AtomicSDNode>(Node)->getMemOperand());
2901 Results.push_back(Swap.getValue(1));
2904 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
2905 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
2906 // splits out the success value as a comparison. Expanding the resulting
2907 // ATOMIC_CMP_SWAP will produce a libcall.
2908 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
2909 SDValue Res = DAG.getAtomicCmpSwap(
2910 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
2911 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
2912 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand());
2914 SDValue ExtRes = Res;
2916 SDValue RHS = Node->getOperand(1);
2918 EVT AtomicType = cast<AtomicSDNode>(Node)->getMemoryVT();
2919 EVT OuterType = Node->getValueType(0);
2920 switch (TLI.getExtendForAtomicOps()) {
2921 case ISD::SIGN_EXTEND:
2922 LHS = DAG.getNode(ISD::AssertSext, dl, OuterType, Res,
2923 DAG.getValueType(AtomicType));
2924 RHS = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, OuterType,
2925 Node->getOperand(2), DAG.getValueType(AtomicType));
2928 case ISD::ZERO_EXTEND:
2929 LHS = DAG.getNode(ISD::AssertZext, dl, OuterType, Res,
2930 DAG.getValueType(AtomicType));
2931 RHS = DAG.getNode(ISD::ZERO_EXTEND, dl, OuterType, Node->getOperand(2));
2934 case ISD::ANY_EXTEND:
2935 LHS = DAG.getZeroExtendInReg(Res, dl, AtomicType);
2936 RHS = DAG.getNode(ISD::ZERO_EXTEND, dl, OuterType, Node->getOperand(2));
2939 llvm_unreachable("Invalid atomic op extension");
2943 DAG.getSetCC(dl, Node->getValueType(1), LHS, RHS, ISD::SETEQ);
2945 Results.push_back(ExtRes.getValue(0));
2946 Results.push_back(Success);
2947 Results.push_back(Res.getValue(1));
2950 case ISD::DYNAMIC_STACKALLOC:
2951 ExpandDYNAMIC_STACKALLOC(Node, Results);
2953 case ISD::MERGE_VALUES:
2954 for (unsigned i = 0; i < Node->getNumValues(); i++)
2955 Results.push_back(Node->getOperand(i));
2958 EVT VT = Node->getValueType(0);
2960 Results.push_back(DAG.getConstant(0, dl, VT));
2962 assert(VT.isFloatingPoint() && "Unknown value type!");
2963 Results.push_back(DAG.getConstantFP(0, dl, VT));
2969 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
2970 Node->getValueType(0), dl);
2971 Results.push_back(Tmp1);
2973 case ISD::FP_EXTEND:
2974 Tmp1 = EmitStackConvert(Node->getOperand(0),
2975 Node->getOperand(0).getValueType(),
2976 Node->getValueType(0), dl);
2977 Results.push_back(Tmp1);
2979 case ISD::SIGN_EXTEND_INREG: {
2980 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2981 EVT VT = Node->getValueType(0);
2983 // An in-register sign-extend of a boolean is a negation:
2984 // 'true' (1) sign-extended is -1.
2985 // 'false' (0) sign-extended is 0.
2986 // However, we must mask the high bits of the source operand because the
2987 // SIGN_EXTEND_INREG does not guarantee that the high bits are already zero.
2989 // TODO: Do this for vectors too?
2990 if (ExtraVT.getSizeInBits() == 1) {
2991 SDValue One = DAG.getConstant(1, dl, VT);
2992 SDValue And = DAG.getNode(ISD::AND, dl, VT, Node->getOperand(0), One);
2993 SDValue Zero = DAG.getConstant(0, dl, VT);
2994 SDValue Neg = DAG.getNode(ISD::SUB, dl, VT, Zero, And);
2995 Results.push_back(Neg);
2999 // NOTE: we could fall back on load/store here too for targets without
3000 // SRA. However, it is doubtful that any exist.
3001 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
3004 unsigned BitsDiff = VT.getScalarSizeInBits() -
3005 ExtraVT.getScalarSizeInBits();
3006 SDValue ShiftCst = DAG.getConstant(BitsDiff, dl, ShiftAmountTy);
3007 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3008 Node->getOperand(0), ShiftCst);
3009 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3010 Results.push_back(Tmp1);
3013 case ISD::FP_ROUND_INREG: {
3014 // The only way we can lower this is to turn it into a TRUNCSTORE,
3015 // EXTLOAD pair, targeting a temporary location (a stack slot).
3017 // NOTE: there is a choice here between constantly creating new stack
3018 // slots and always reusing the same one. We currently always create
3019 // new ones, as reuse may inhibit scheduling.
3020 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3021 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3022 Node->getValueType(0), dl);
3023 Results.push_back(Tmp1);
3026 case ISD::SINT_TO_FP:
3027 case ISD::UINT_TO_FP:
3028 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3029 Node->getOperand(0), Node->getValueType(0), dl);
3030 Results.push_back(Tmp1);
3032 case ISD::FP_TO_SINT:
3033 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3034 Results.push_back(Tmp1);
3036 case ISD::FP_TO_UINT: {
3037 SDValue True, False;
3038 EVT VT = Node->getOperand(0).getValueType();
3039 EVT NVT = Node->getValueType(0);
3040 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3041 APInt::getNullValue(VT.getSizeInBits()));
3042 APInt x = APInt::getSignMask(NVT.getSizeInBits());
3043 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3044 Tmp1 = DAG.getConstantFP(apf, dl, VT);
3045 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3046 Node->getOperand(0),
3048 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3049 // TODO: Should any fast-math-flags be set for the FSUB?
3050 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3051 DAG.getNode(ISD::FSUB, dl, VT,
3052 Node->getOperand(0), Tmp1));
3053 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3054 DAG.getConstant(x, dl, NVT));
3055 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3056 Results.push_back(Tmp1);
3060 Results.push_back(DAG.expandVAArg(Node));
3061 Results.push_back(Results[0].getValue(1));
3064 Results.push_back(DAG.expandVACopy(Node));
3066 case ISD::EXTRACT_VECTOR_ELT:
3067 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3068 // This must be an access of the only element. Return it.
3069 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3070 Node->getOperand(0));
3072 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3073 Results.push_back(Tmp1);
3075 case ISD::EXTRACT_SUBVECTOR:
3076 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3078 case ISD::INSERT_SUBVECTOR:
3079 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3081 case ISD::CONCAT_VECTORS: {
3082 Results.push_back(ExpandVectorBuildThroughStack(Node));
3085 case ISD::SCALAR_TO_VECTOR:
3086 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3088 case ISD::INSERT_VECTOR_ELT:
3089 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3090 Node->getOperand(1),
3091 Node->getOperand(2), dl));
3093 case ISD::VECTOR_SHUFFLE: {
3094 SmallVector<int, 32> NewMask;
3095 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3097 EVT VT = Node->getValueType(0);
3098 EVT EltVT = VT.getVectorElementType();
3099 SDValue Op0 = Node->getOperand(0);
3100 SDValue Op1 = Node->getOperand(1);
3101 if (!TLI.isTypeLegal(EltVT)) {
3103 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3105 // BUILD_VECTOR operands are allowed to be wider than the element type.
3106 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3108 if (NewEltVT.bitsLT(EltVT)) {
3110 // Convert shuffle node.
3111 // If original node was v4i64 and the new EltVT is i32,
3112 // cast operands to v8i32 and re-build the mask.
3114 // Calculate new VT, the size of the new VT should be equal to original.
3116 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3117 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3118 assert(NewVT.bitsEq(VT));
3120 // cast operands to new VT
3121 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3122 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3124 // Convert the shuffle mask
3125 unsigned int factor =
3126 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3128 // EltVT gets smaller
3131 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3133 for (unsigned fi = 0; fi < factor; ++fi)
3134 NewMask.push_back(Mask[i]);
3137 for (unsigned fi = 0; fi < factor; ++fi)
3138 NewMask.push_back(Mask[i]*factor+fi);
3146 unsigned NumElems = VT.getVectorNumElements();
3147 SmallVector<SDValue, 16> Ops;
3148 for (unsigned i = 0; i != NumElems; ++i) {
3150 Ops.push_back(DAG.getUNDEF(EltVT));
3153 unsigned Idx = Mask[i];
3155 Ops.push_back(DAG.getNode(
3156 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op0,
3157 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))));
3159 Ops.push_back(DAG.getNode(
3160 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op1,
3161 DAG.getConstant(Idx - NumElems, dl,
3162 TLI.getVectorIdxTy(DAG.getDataLayout()))));
3165 Tmp1 = DAG.getBuildVector(VT, dl, Ops);
3166 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3167 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3168 Results.push_back(Tmp1);
3171 case ISD::EXTRACT_ELEMENT: {
3172 EVT OpTy = Node->getOperand(0).getValueType();
3173 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3175 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3176 DAG.getConstant(OpTy.getSizeInBits() / 2, dl,
3177 TLI.getShiftAmountTy(
3178 Node->getOperand(0).getValueType(),
3179 DAG.getDataLayout())));
3180 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3183 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3184 Node->getOperand(0));
3186 Results.push_back(Tmp1);
3189 case ISD::STACKSAVE:
3190 // Expand to CopyFromReg if the target set
3191 // StackPointerRegisterToSaveRestore.
3192 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3193 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3194 Node->getValueType(0)));
3195 Results.push_back(Results[0].getValue(1));
3197 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3198 Results.push_back(Node->getOperand(0));
3201 case ISD::STACKRESTORE:
3202 // Expand to CopyToReg if the target set
3203 // StackPointerRegisterToSaveRestore.
3204 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3205 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3206 Node->getOperand(1)));
3208 Results.push_back(Node->getOperand(0));
3211 case ISD::GET_DYNAMIC_AREA_OFFSET:
3212 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0)));
3213 Results.push_back(Results[0].getValue(0));
3215 case ISD::FCOPYSIGN:
3216 Results.push_back(ExpandFCOPYSIGN(Node));
3219 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3220 Tmp1 = DAG.getConstantFP(-0.0, dl, Node->getValueType(0));
3221 // TODO: If FNEG has fast-math-flags, propagate them to the FSUB.
3222 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3223 Node->getOperand(0));
3224 Results.push_back(Tmp1);
3227 Results.push_back(ExpandFABS(Node));
3233 // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B
3235 switch (Node->getOpcode()) {
3236 default: llvm_unreachable("How did we get here?");
3237 case ISD::SMAX: Pred = ISD::SETGT; break;
3238 case ISD::SMIN: Pred = ISD::SETLT; break;
3239 case ISD::UMAX: Pred = ISD::SETUGT; break;
3240 case ISD::UMIN: Pred = ISD::SETULT; break;
3242 Tmp1 = Node->getOperand(0);
3243 Tmp2 = Node->getOperand(1);
3244 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp1, Tmp2, Pred);
3245 Results.push_back(Tmp1);
3251 EVT VT = Node->getValueType(0);
3252 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3253 // fcos which share the same operand and both are used.
3254 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3255 canCombineSinCosLibcall(Node, TLI, TM))
3256 && useSinCos(Node)) {
3257 SDVTList VTs = DAG.getVTList(VT, VT);
3258 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3259 if (Node->getOpcode() == ISD::FCOS)
3260 Tmp1 = Tmp1.getValue(1);
3261 Results.push_back(Tmp1);
3266 llvm_unreachable("Illegal fmad should never be formed");
3268 case ISD::FP16_TO_FP:
3269 if (Node->getValueType(0) != MVT::f32) {
3270 // We can extend to types bigger than f32 in two steps without changing
3271 // the result. Since "f16 -> f32" is much more commonly available, give
3272 // CodeGen the option of emitting that before resorting to a libcall.
3274 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3276 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3279 case ISD::FP_TO_FP16:
3280 if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) {
3281 SDValue Op = Node->getOperand(0);
3282 MVT SVT = Op.getSimpleValueType();
3283 if ((SVT == MVT::f64 || SVT == MVT::f80) &&
3284 TLI.isOperationLegalOrCustom(ISD::FP_TO_FP16, MVT::f32)) {
3285 // Under fastmath, we can expand this node into a fround followed by
3286 // a float-half conversion.
3287 SDValue FloatVal = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Op,
3288 DAG.getIntPtrConstant(0, dl));
3290 DAG.getNode(ISD::FP_TO_FP16, dl, Node->getValueType(0), FloatVal));
3294 case ISD::ConstantFP: {
3295 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3296 // Check to see if this FP immediate is already legal.
3297 // If this is a legal constant, turn it into a TargetConstantFP node.
3298 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3299 Results.push_back(ExpandConstantFP(CFP, true));
3302 case ISD::Constant: {
3303 ConstantSDNode *CP = cast<ConstantSDNode>(Node);
3304 Results.push_back(ExpandConstant(CP));
3308 EVT VT = Node->getValueType(0);
3309 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3310 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3311 const SDNodeFlags Flags = Node->getFlags();
3312 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3313 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1, Flags);
3314 Results.push_back(Tmp1);
3319 EVT VT = Node->getValueType(0);
3320 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3321 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3322 "Don't know how to expand this subtraction!");
3323 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3324 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
3326 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, dl, VT));
3327 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3332 EVT VT = Node->getValueType(0);
3333 bool isSigned = Node->getOpcode() == ISD::SREM;
3334 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3335 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3336 Tmp2 = Node->getOperand(0);
3337 Tmp3 = Node->getOperand(1);
3338 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
3339 SDVTList VTs = DAG.getVTList(VT, VT);
3340 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3341 Results.push_back(Tmp1);
3342 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3344 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3345 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3346 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3347 Results.push_back(Tmp1);
3353 bool isSigned = Node->getOpcode() == ISD::SDIV;
3354 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3355 EVT VT = Node->getValueType(0);
3356 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
3357 SDVTList VTs = DAG.getVTList(VT, VT);
3358 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3359 Node->getOperand(1));
3360 Results.push_back(Tmp1);
3366 unsigned ExpandOpcode =
3367 Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI : ISD::SMUL_LOHI;
3368 EVT VT = Node->getValueType(0);
3369 SDVTList VTs = DAG.getVTList(VT, VT);
3371 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3372 Node->getOperand(1));
3373 Results.push_back(Tmp1.getValue(1));
3376 case ISD::UMUL_LOHI:
3377 case ISD::SMUL_LOHI: {
3378 SDValue LHS = Node->getOperand(0);
3379 SDValue RHS = Node->getOperand(1);
3380 MVT VT = LHS.getSimpleValueType();
3381 unsigned MULHOpcode =
3382 Node->getOpcode() == ISD::UMUL_LOHI ? ISD::MULHU : ISD::MULHS;
3384 if (TLI.isOperationLegalOrCustom(MULHOpcode, VT)) {
3385 Results.push_back(DAG.getNode(ISD::MUL, dl, VT, LHS, RHS));
3386 Results.push_back(DAG.getNode(MULHOpcode, dl, VT, LHS, RHS));
3390 SmallVector<SDValue, 4> Halves;
3391 EVT HalfType = EVT(VT).getHalfSizedIntegerVT(*DAG.getContext());
3392 assert(TLI.isTypeLegal(HalfType));
3393 if (TLI.expandMUL_LOHI(Node->getOpcode(), VT, Node, LHS, RHS, Halves,
3395 TargetLowering::MulExpansionKind::Always)) {
3396 for (unsigned i = 0; i < 2; ++i) {
3397 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Halves[2 * i]);
3398 SDValue Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Halves[2 * i + 1]);
3399 SDValue Shift = DAG.getConstant(
3400 HalfType.getScalarSizeInBits(), dl,
3401 TLI.getShiftAmountTy(HalfType, DAG.getDataLayout()));
3402 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3403 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3410 EVT VT = Node->getValueType(0);
3411 SDVTList VTs = DAG.getVTList(VT, VT);
3412 // See if multiply or divide can be lowered using two-result operations.
3413 // We just need the low half of the multiply; try both the signed
3414 // and unsigned forms. If the target supports both SMUL_LOHI and
3415 // UMUL_LOHI, form a preference by checking which forms of plain
3416 // MULH it supports.
3417 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3418 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3419 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3420 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3421 unsigned OpToUse = 0;
3422 if (HasSMUL_LOHI && !HasMULHS) {
3423 OpToUse = ISD::SMUL_LOHI;
3424 } else if (HasUMUL_LOHI && !HasMULHU) {
3425 OpToUse = ISD::UMUL_LOHI;
3426 } else if (HasSMUL_LOHI) {
3427 OpToUse = ISD::SMUL_LOHI;
3428 } else if (HasUMUL_LOHI) {
3429 OpToUse = ISD::UMUL_LOHI;
3432 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3433 Node->getOperand(1)));
3438 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3439 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3440 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3441 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3442 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3443 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG,
3444 TargetLowering::MulExpansionKind::OnlyLegalOrCustom)) {
3445 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3446 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3448 DAG.getConstant(HalfType.getSizeInBits(), dl,
3449 TLI.getShiftAmountTy(HalfType, DAG.getDataLayout()));
3450 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3451 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3457 SDValue LHS = Node->getOperand(0);
3458 SDValue RHS = Node->getOperand(1);
3459 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3460 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3462 Results.push_back(Sum);
3463 EVT ResultType = Node->getValueType(1);
3464 EVT OType = getSetCCResultType(Node->getValueType(0));
3466 SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType());
3468 // LHSSign -> LHS >= 0
3469 // RHSSign -> RHS >= 0
3470 // SumSign -> Sum >= 0
3473 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3475 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3477 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3478 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3479 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3480 Node->getOpcode() == ISD::SADDO ?
3481 ISD::SETEQ : ISD::SETNE);
3483 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3484 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3486 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3487 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3492 SDValue LHS = Node->getOperand(0);
3493 SDValue RHS = Node->getOperand(1);
3494 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3495 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3497 Results.push_back(Sum);
3499 EVT ResultType = Node->getValueType(1);
3500 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3502 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3503 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3505 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3510 EVT VT = Node->getValueType(0);
3511 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3512 SDValue LHS = Node->getOperand(0);
3513 SDValue RHS = Node->getOperand(1);
3516 static const unsigned Ops[2][3] =
3517 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3518 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3519 bool isSigned = Node->getOpcode() == ISD::SMULO;
3520 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3521 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3522 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3523 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3524 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3526 TopHalf = BottomHalf.getValue(1);
3527 } else if (TLI.isTypeLegal(WideVT)) {
3528 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3529 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3530 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3531 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3532 DAG.getIntPtrConstant(0, dl));
3533 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3534 DAG.getIntPtrConstant(1, dl));
3536 // We can fall back to a libcall with an illegal type for the MUL if we
3537 // have a libcall big enough.
3538 // Also, we can fall back to a division in some cases, but that's a big
3539 // performance hit in the general case.
3540 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3541 if (WideVT == MVT::i16)
3542 LC = RTLIB::MUL_I16;
3543 else if (WideVT == MVT::i32)
3544 LC = RTLIB::MUL_I32;
3545 else if (WideVT == MVT::i64)
3546 LC = RTLIB::MUL_I64;
3547 else if (WideVT == MVT::i128)
3548 LC = RTLIB::MUL_I128;
3549 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3551 // The high part is obtained by SRA'ing all but one of the bits of low
3553 unsigned LoSize = VT.getSizeInBits();
3555 DAG.getNode(ISD::SRA, dl, VT, LHS,
3556 DAG.getConstant(LoSize - 1, dl,
3557 TLI.getPointerTy(DAG.getDataLayout())));
3559 DAG.getNode(ISD::SRA, dl, VT, RHS,
3560 DAG.getConstant(LoSize - 1, dl,
3561 TLI.getPointerTy(DAG.getDataLayout())));
3563 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3564 // pre-lowered to the correct types. This all depends upon WideVT not
3565 // being a legal type for the architecture and thus has to be split to
3568 if(DAG.getDataLayout().isLittleEndian()) {
3569 // Halves of WideVT are packed into registers in different order
3570 // depending on platform endianness. This is usually handled by
3571 // the C calling convention, but we can't defer to it in
3573 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3574 Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3576 SDValue Args[] = { HiLHS, LHS, HiRHS, RHS };
3577 Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3579 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3580 DAG.getIntPtrConstant(0, dl));
3581 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3582 DAG.getIntPtrConstant(1, dl));
3583 // Ret is a node with an illegal type. Because such things are not
3584 // generally permitted during this phase of legalization, make sure the
3585 // node has no more uses. The above EXTRACT_ELEMENT nodes should have been
3587 assert(Ret->use_empty() &&
3588 "Unexpected uses of illegally type from expanded lib call.");
3592 Tmp1 = DAG.getConstant(
3593 VT.getSizeInBits() - 1, dl,
3594 TLI.getShiftAmountTy(BottomHalf.getValueType(), DAG.getDataLayout()));
3595 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3596 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3599 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3600 DAG.getConstant(0, dl, VT), ISD::SETNE);
3603 // Truncate the result if SetCC returns a larger type than needed.
3604 EVT RType = Node->getValueType(1);
3605 if (RType.getSizeInBits() < TopHalf.getValueSizeInBits())
3606 TopHalf = DAG.getNode(ISD::TRUNCATE, dl, RType, TopHalf);
3608 assert(RType.getSizeInBits() == TopHalf.getValueSizeInBits() &&
3609 "Unexpected result type for S/UMULO legalization");
3611 Results.push_back(BottomHalf);
3612 Results.push_back(TopHalf);
3615 case ISD::BUILD_PAIR: {
3616 EVT PairTy = Node->getValueType(0);
3617 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3618 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3620 ISD::SHL, dl, PairTy, Tmp2,
3621 DAG.getConstant(PairTy.getSizeInBits() / 2, dl,
3622 TLI.getShiftAmountTy(PairTy, DAG.getDataLayout())));
3623 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3627 Tmp1 = Node->getOperand(0);
3628 Tmp2 = Node->getOperand(1);
3629 Tmp3 = Node->getOperand(2);
3630 if (Tmp1.getOpcode() == ISD::SETCC) {
3631 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3633 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3635 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3636 DAG.getConstant(0, dl, Tmp1.getValueType()),
3637 Tmp2, Tmp3, ISD::SETNE);
3639 Results.push_back(Tmp1);
3642 SDValue Chain = Node->getOperand(0);
3643 SDValue Table = Node->getOperand(1);
3644 SDValue Index = Node->getOperand(2);
3646 const DataLayout &TD = DAG.getDataLayout();
3647 EVT PTy = TLI.getPointerTy(TD);
3649 unsigned EntrySize =
3650 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3652 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(), Index,
3653 DAG.getConstant(EntrySize, dl, Index.getValueType()));
3654 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3657 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3658 SDValue LD = DAG.getExtLoad(
3659 ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3660 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()), MemVT);
3662 if (TLI.isJumpTableRelative()) {
3663 // For PIC, the sequence is:
3664 // BRIND(load(Jumptable + index) + RelocBase)
3665 // RelocBase can be JumpTable, GOT or some sort of global base.
3666 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3667 TLI.getPICJumpTableRelocBase(Table, DAG));
3669 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3670 Results.push_back(Tmp1);
3674 // Expand brcond's setcc into its constituent parts and create a BR_CC
3676 Tmp1 = Node->getOperand(0);
3677 Tmp2 = Node->getOperand(1);
3678 if (Tmp2.getOpcode() == ISD::SETCC) {
3679 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3680 Tmp1, Tmp2.getOperand(2),
3681 Tmp2.getOperand(0), Tmp2.getOperand(1),
3682 Node->getOperand(2));
3684 // We test only the i1 bit. Skip the AND if UNDEF.
3685 Tmp3 = (Tmp2.isUndef()) ? Tmp2 :
3686 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3687 DAG.getConstant(1, dl, Tmp2.getValueType()));
3688 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3689 DAG.getCondCode(ISD::SETNE), Tmp3,
3690 DAG.getConstant(0, dl, Tmp3.getValueType()),
3691 Node->getOperand(2));
3693 Results.push_back(Tmp1);
3696 Tmp1 = Node->getOperand(0);
3697 Tmp2 = Node->getOperand(1);
3698 Tmp3 = Node->getOperand(2);
3699 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3700 Tmp3, NeedInvert, dl);
3703 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3704 // condition code, create a new SETCC node.
3706 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3709 // If we expanded the SETCC by inverting the condition code, then wrap
3710 // the existing SETCC in a NOT to restore the intended condition.
3712 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
3714 Results.push_back(Tmp1);
3718 // Otherwise, SETCC for the given comparison type must be completely
3719 // illegal; expand it into a SELECT_CC.
3720 EVT VT = Node->getValueType(0);
3722 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
3723 case TargetLowering::ZeroOrOneBooleanContent:
3724 case TargetLowering::UndefinedBooleanContent:
3727 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3731 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3732 DAG.getConstant(TrueValue, dl, VT),
3733 DAG.getConstant(0, dl, VT),
3735 Results.push_back(Tmp1);
3738 case ISD::SELECT_CC: {
3739 Tmp1 = Node->getOperand(0); // LHS
3740 Tmp2 = Node->getOperand(1); // RHS
3741 Tmp3 = Node->getOperand(2); // True
3742 Tmp4 = Node->getOperand(3); // False
3743 EVT VT = Node->getValueType(0);
3744 SDValue CC = Node->getOperand(4);
3745 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
3747 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) {
3748 // If the condition code is legal, then we need to expand this
3749 // node using SETCC and SELECT.
3750 EVT CmpVT = Tmp1.getValueType();
3751 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
3752 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
3755 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT);
3756 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
3757 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
3761 // SELECT_CC is legal, so the condition code must not be.
3762 bool Legalized = false;
3763 // Try to legalize by inverting the condition. This is for targets that
3764 // might support an ordered version of a condition, but not the unordered
3765 // version (or vice versa).
3766 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
3767 Tmp1.getValueType().isInteger());
3768 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
3769 // Use the new condition code and swap true and false
3771 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
3773 // If The inverse is not legal, then try to swap the arguments using
3774 // the inverse condition code.
3775 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
3776 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
3777 // The swapped inverse condition is legal, so swap true and false,
3780 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
3785 Legalized = LegalizeSetCCCondCode(
3786 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
3789 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
3791 // If we expanded the SETCC by inverting the condition code, then swap
3792 // the True/False operands to match.
3794 std::swap(Tmp3, Tmp4);
3796 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3797 // condition code, create a new SELECT_CC node.
3799 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
3800 Tmp1, Tmp2, Tmp3, Tmp4, CC);
3802 Tmp2 = DAG.getConstant(0, dl, Tmp1.getValueType());
3803 CC = DAG.getCondCode(ISD::SETNE);
3804 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
3805 Tmp2, Tmp3, Tmp4, CC);
3808 Results.push_back(Tmp1);
3812 Tmp1 = Node->getOperand(0); // Chain
3813 Tmp2 = Node->getOperand(2); // LHS
3814 Tmp3 = Node->getOperand(3); // RHS
3815 Tmp4 = Node->getOperand(1); // CC
3817 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
3818 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
3820 assert(Legalized && "Can't legalize BR_CC with legal condition!");
3822 // If we expanded the SETCC by inverting the condition code, then wrap
3823 // the existing SETCC in a NOT to restore the intended condition.
3825 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
3827 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
3829 if (Tmp4.getNode()) {
3830 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
3831 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
3833 Tmp3 = DAG.getConstant(0, dl, Tmp2.getValueType());
3834 Tmp4 = DAG.getCondCode(ISD::SETNE);
3835 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
3836 Tmp2, Tmp3, Node->getOperand(4));
3838 Results.push_back(Tmp1);
3841 case ISD::BUILD_VECTOR:
3842 Results.push_back(ExpandBUILD_VECTOR(Node));
3847 // Scalarize vector SRA/SRL/SHL.
3848 EVT VT = Node->getValueType(0);
3849 assert(VT.isVector() && "Unable to legalize non-vector shift");
3850 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
3851 unsigned NumElem = VT.getVectorNumElements();
3853 SmallVector<SDValue, 8> Scalars;
3854 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
3855 SDValue Ex = DAG.getNode(
3856 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(0),
3857 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
3858 SDValue Sh = DAG.getNode(
3859 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(1),
3860 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
3861 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
3862 VT.getScalarType(), Ex, Sh));
3865 SDValue Result = DAG.getBuildVector(Node->getValueType(0), dl, Scalars);
3866 ReplaceNode(SDValue(Node, 0), Result);
3869 case ISD::GLOBAL_OFFSET_TABLE:
3870 case ISD::GlobalAddress:
3871 case ISD::GlobalTLSAddress:
3872 case ISD::ExternalSymbol:
3873 case ISD::ConstantPool:
3874 case ISD::JumpTable:
3875 case ISD::INTRINSIC_W_CHAIN:
3876 case ISD::INTRINSIC_WO_CHAIN:
3877 case ISD::INTRINSIC_VOID:
3878 // FIXME: Custom lowering for these operations shouldn't return null!
3882 // Replace the original node with the legalized result.
3883 if (Results.empty())
3886 ReplaceNode(Node, Results.data());
3890 void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) {
3891 SmallVector<SDValue, 8> Results;
3893 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
3894 unsigned Opc = Node->getOpcode();
3896 case ISD::ATOMIC_FENCE: {
3897 // If the target didn't lower this, lower it to '__sync_synchronize()' call
3898 // FIXME: handle "fence singlethread" more efficiently.
3899 TargetLowering::ArgListTy Args;
3901 TargetLowering::CallLoweringInfo CLI(DAG);
3903 .setChain(Node->getOperand(0))
3905 CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3906 DAG.getExternalSymbol("__sync_synchronize",
3907 TLI.getPointerTy(DAG.getDataLayout())),
3910 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3912 Results.push_back(CallResult.second);
3915 // By default, atomic intrinsics are marked Legal and lowered. Targets
3916 // which don't support them directly, however, may want libcalls, in which
3917 // case they mark them Expand, and we get here.
3918 case ISD::ATOMIC_SWAP:
3919 case ISD::ATOMIC_LOAD_ADD:
3920 case ISD::ATOMIC_LOAD_SUB:
3921 case ISD::ATOMIC_LOAD_AND:
3922 case ISD::ATOMIC_LOAD_OR:
3923 case ISD::ATOMIC_LOAD_XOR:
3924 case ISD::ATOMIC_LOAD_NAND:
3925 case ISD::ATOMIC_LOAD_MIN:
3926 case ISD::ATOMIC_LOAD_MAX:
3927 case ISD::ATOMIC_LOAD_UMIN:
3928 case ISD::ATOMIC_LOAD_UMAX:
3929 case ISD::ATOMIC_CMP_SWAP: {
3930 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
3931 RTLIB::Libcall LC = RTLIB::getSYNC(Opc, VT);
3932 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!");
3934 std::pair<SDValue, SDValue> Tmp = ExpandChainLibCall(LC, Node, false);
3935 Results.push_back(Tmp.first);
3936 Results.push_back(Tmp.second);
3940 // If this operation is not supported, lower it to 'abort()' call
3941 TargetLowering::ArgListTy Args;
3942 TargetLowering::CallLoweringInfo CLI(DAG);
3944 .setChain(Node->getOperand(0))
3945 .setLibCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3946 DAG.getExternalSymbol(
3947 "abort", TLI.getPointerTy(DAG.getDataLayout())),
3949 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3951 Results.push_back(CallResult.second);
3955 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
3956 RTLIB::FMIN_F80, RTLIB::FMIN_F128,
3957 RTLIB::FMIN_PPCF128));
3960 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
3961 RTLIB::FMAX_F80, RTLIB::FMAX_F128,
3962 RTLIB::FMAX_PPCF128));
3965 case ISD::STRICT_FSQRT:
3966 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3967 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3968 RTLIB::SQRT_PPCF128));
3971 case ISD::STRICT_FSIN:
3972 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3973 RTLIB::SIN_F80, RTLIB::SIN_F128,
3974 RTLIB::SIN_PPCF128));
3977 case ISD::STRICT_FCOS:
3978 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3979 RTLIB::COS_F80, RTLIB::COS_F128,
3980 RTLIB::COS_PPCF128));
3983 // Expand into sincos libcall.
3984 ExpandSinCosLibCall(Node, Results);
3987 case ISD::STRICT_FLOG:
3988 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3989 RTLIB::LOG_F80, RTLIB::LOG_F128,
3990 RTLIB::LOG_PPCF128));
3993 case ISD::STRICT_FLOG2:
3994 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3995 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3996 RTLIB::LOG2_PPCF128));
3999 case ISD::STRICT_FLOG10:
4000 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
4001 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
4002 RTLIB::LOG10_PPCF128));
4005 case ISD::STRICT_FEXP:
4006 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
4007 RTLIB::EXP_F80, RTLIB::EXP_F128,
4008 RTLIB::EXP_PPCF128));
4011 case ISD::STRICT_FEXP2:
4012 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
4013 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
4014 RTLIB::EXP2_PPCF128));
4017 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
4018 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
4019 RTLIB::TRUNC_PPCF128));
4022 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
4023 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
4024 RTLIB::FLOOR_PPCF128));
4027 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
4028 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
4029 RTLIB::CEIL_PPCF128));
4032 case ISD::STRICT_FRINT:
4033 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
4034 RTLIB::RINT_F80, RTLIB::RINT_F128,
4035 RTLIB::RINT_PPCF128));
4037 case ISD::FNEARBYINT:
4038 case ISD::STRICT_FNEARBYINT:
4039 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
4040 RTLIB::NEARBYINT_F64,
4041 RTLIB::NEARBYINT_F80,
4042 RTLIB::NEARBYINT_F128,
4043 RTLIB::NEARBYINT_PPCF128));
4046 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
4050 RTLIB::ROUND_PPCF128));
4053 case ISD::STRICT_FPOWI:
4054 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
4055 RTLIB::POWI_F80, RTLIB::POWI_F128,
4056 RTLIB::POWI_PPCF128));
4059 case ISD::STRICT_FPOW:
4060 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
4061 RTLIB::POW_F80, RTLIB::POW_F128,
4062 RTLIB::POW_PPCF128));
4065 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
4066 RTLIB::DIV_F80, RTLIB::DIV_F128,
4067 RTLIB::DIV_PPCF128));
4070 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
4071 RTLIB::REM_F80, RTLIB::REM_F128,
4072 RTLIB::REM_PPCF128));
4075 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
4076 RTLIB::FMA_F80, RTLIB::FMA_F128,
4077 RTLIB::FMA_PPCF128));
4080 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
4081 RTLIB::ADD_F80, RTLIB::ADD_F128,
4082 RTLIB::ADD_PPCF128));
4085 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
4086 RTLIB::MUL_F80, RTLIB::MUL_F128,
4087 RTLIB::MUL_PPCF128));
4089 case ISD::FP16_TO_FP:
4090 if (Node->getValueType(0) == MVT::f32) {
4091 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
4094 case ISD::FP_TO_FP16: {
4096 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
4097 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
4098 Results.push_back(ExpandLibCall(LC, Node, false));
4102 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
4103 RTLIB::SUB_F80, RTLIB::SUB_F128,
4104 RTLIB::SUB_PPCF128));
4107 Results.push_back(ExpandIntLibCall(Node, true,
4109 RTLIB::SREM_I16, RTLIB::SREM_I32,
4110 RTLIB::SREM_I64, RTLIB::SREM_I128));
4113 Results.push_back(ExpandIntLibCall(Node, false,
4115 RTLIB::UREM_I16, RTLIB::UREM_I32,
4116 RTLIB::UREM_I64, RTLIB::UREM_I128));
4119 Results.push_back(ExpandIntLibCall(Node, true,
4121 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
4122 RTLIB::SDIV_I64, RTLIB::SDIV_I128));
4125 Results.push_back(ExpandIntLibCall(Node, false,
4127 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
4128 RTLIB::UDIV_I64, RTLIB::UDIV_I128));
4132 // Expand into divrem libcall
4133 ExpandDivRemLibCall(Node, Results);
4136 Results.push_back(ExpandIntLibCall(Node, false,
4138 RTLIB::MUL_I16, RTLIB::MUL_I32,
4139 RTLIB::MUL_I64, RTLIB::MUL_I128));
4143 // Replace the original node with the legalized result.
4144 if (!Results.empty())
4145 ReplaceNode(Node, Results.data());
4148 // Determine the vector type to use in place of an original scalar element when
4149 // promoting equally sized vectors.
4150 static MVT getPromotedVectorElementType(const TargetLowering &TLI,
4151 MVT EltVT, MVT NewEltVT) {
4152 unsigned OldEltsPerNewElt = EltVT.getSizeInBits() / NewEltVT.getSizeInBits();
4153 MVT MidVT = MVT::getVectorVT(NewEltVT, OldEltsPerNewElt);
4154 assert(TLI.isTypeLegal(MidVT) && "unexpected");
4158 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4159 SmallVector<SDValue, 8> Results;
4160 MVT OVT = Node->getSimpleValueType(0);
4161 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4162 Node->getOpcode() == ISD::SINT_TO_FP ||
4163 Node->getOpcode() == ISD::SETCC ||
4164 Node->getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
4165 Node->getOpcode() == ISD::INSERT_VECTOR_ELT) {
4166 OVT = Node->getOperand(0).getSimpleValueType();
4168 if (Node->getOpcode() == ISD::BR_CC)
4169 OVT = Node->getOperand(2).getSimpleValueType();
4170 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4172 SDValue Tmp1, Tmp2, Tmp3;
4173 switch (Node->getOpcode()) {
4175 case ISD::CTTZ_ZERO_UNDEF:
4177 case ISD::CTLZ_ZERO_UNDEF:
4179 // Zero extend the argument.
4180 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4181 if (Node->getOpcode() == ISD::CTTZ) {
4182 // The count is the same in the promoted type except if the original
4183 // value was zero. This can be handled by setting the bit just off
4184 // the top of the original type.
4185 auto TopBit = APInt::getOneBitSet(NVT.getSizeInBits(),
4186 OVT.getSizeInBits());
4187 Tmp1 = DAG.getNode(ISD::OR, dl, NVT, Tmp1,
4188 DAG.getConstant(TopBit, dl, NVT));
4190 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4191 // already the correct result.
4192 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4193 if (Node->getOpcode() == ISD::CTLZ ||
4194 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4195 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4196 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4197 DAG.getConstant(NVT.getSizeInBits() -
4198 OVT.getSizeInBits(), dl, NVT));
4200 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4202 case ISD::BITREVERSE:
4204 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4205 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4206 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4208 ISD::SRL, dl, NVT, Tmp1,
4209 DAG.getConstant(DiffBits, dl,
4210 TLI.getShiftAmountTy(NVT, DAG.getDataLayout())));
4211 Results.push_back(Tmp1);
4214 case ISD::FP_TO_UINT:
4215 case ISD::FP_TO_SINT:
4216 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
4217 Node->getOpcode() == ISD::FP_TO_SINT, dl);
4218 Results.push_back(Tmp1);
4220 case ISD::UINT_TO_FP:
4221 case ISD::SINT_TO_FP:
4222 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
4223 Node->getOpcode() == ISD::SINT_TO_FP, dl);
4224 Results.push_back(Tmp1);
4227 SDValue Chain = Node->getOperand(0); // Get the chain.
4228 SDValue Ptr = Node->getOperand(1); // Get the pointer.
4231 if (OVT.isVector()) {
4232 TruncOp = ISD::BITCAST;
4234 assert(OVT.isInteger()
4235 && "VAARG promotion is supported only for vectors or integer types");
4236 TruncOp = ISD::TRUNCATE;
4239 // Perform the larger operation, then convert back
4240 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
4241 Node->getConstantOperandVal(3));
4242 Chain = Tmp1.getValue(1);
4244 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4246 // Modified the chain result - switch anything that used the old chain to
4248 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4249 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4251 UpdatedNodes->insert(Tmp2.getNode());
4252 UpdatedNodes->insert(Chain.getNode());
4265 unsigned ExtOp, TruncOp;
4266 if (OVT.isVector()) {
4267 ExtOp = ISD::BITCAST;
4268 TruncOp = ISD::BITCAST;
4270 assert(OVT.isInteger() && "Cannot promote logic operation");
4272 switch (Node->getOpcode()) {
4274 ExtOp = ISD::ANY_EXTEND;
4278 ExtOp = ISD::SIGN_EXTEND;
4282 ExtOp = ISD::ZERO_EXTEND;
4285 TruncOp = ISD::TRUNCATE;
4287 // Promote each of the values to the new type.
4288 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4289 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4290 // Perform the larger operation, then convert back
4291 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4292 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4295 case ISD::UMUL_LOHI:
4296 case ISD::SMUL_LOHI: {
4297 // Promote to a multiply in a wider integer type.
4298 unsigned ExtOp = Node->getOpcode() == ISD::UMUL_LOHI ? ISD::ZERO_EXTEND
4300 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4301 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4302 Tmp1 = DAG.getNode(ISD::MUL, dl, NVT, Tmp1, Tmp2);
4304 auto &DL = DAG.getDataLayout();
4305 unsigned OriginalSize = OVT.getScalarSizeInBits();
4307 ISD::SRL, dl, NVT, Tmp1,
4308 DAG.getConstant(OriginalSize, dl, TLI.getScalarShiftAmountTy(DL, NVT)));
4309 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4310 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp2));
4314 unsigned ExtOp, TruncOp;
4315 if (Node->getValueType(0).isVector() ||
4316 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4317 ExtOp = ISD::BITCAST;
4318 TruncOp = ISD::BITCAST;
4319 } else if (Node->getValueType(0).isInteger()) {
4320 ExtOp = ISD::ANY_EXTEND;
4321 TruncOp = ISD::TRUNCATE;
4323 ExtOp = ISD::FP_EXTEND;
4324 TruncOp = ISD::FP_ROUND;
4326 Tmp1 = Node->getOperand(0);
4327 // Promote each of the values to the new type.
4328 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4329 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4330 // Perform the larger operation, then round down.
4331 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4332 if (TruncOp != ISD::FP_ROUND)
4333 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4335 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4336 DAG.getIntPtrConstant(0, dl));
4337 Results.push_back(Tmp1);
4340 case ISD::VECTOR_SHUFFLE: {
4341 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4343 // Cast the two input vectors.
4344 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4345 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4347 // Convert the shuffle mask to the right # elements.
4348 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4349 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4350 Results.push_back(Tmp1);
4354 unsigned ExtOp = ISD::FP_EXTEND;
4355 if (NVT.isInteger()) {
4356 ISD::CondCode CCCode =
4357 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4358 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4360 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4361 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4362 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4363 Tmp1, Tmp2, Node->getOperand(2)));
4367 unsigned ExtOp = ISD::FP_EXTEND;
4368 if (NVT.isInteger()) {
4369 ISD::CondCode CCCode =
4370 cast<CondCodeSDNode>(Node->getOperand(1))->get();
4371 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4373 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4374 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(3));
4375 Results.push_back(DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0),
4376 Node->getOperand(0), Node->getOperand(1),
4377 Tmp1, Tmp2, Node->getOperand(4)));
4388 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4389 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4390 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2,
4392 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4393 Tmp3, DAG.getIntPtrConstant(0, dl)));
4397 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4398 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4399 Tmp3 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(2));
4401 DAG.getNode(ISD::FP_ROUND, dl, OVT,
4402 DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2, Tmp3),
4403 DAG.getIntPtrConstant(0, dl)));
4406 case ISD::FCOPYSIGN:
4408 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4409 Tmp2 = Node->getOperand(1);
4410 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4412 // fcopysign doesn't change anything but the sign bit, so
4413 // (fp_round (fcopysign (fpext a), b))
4415 // (fp_round (fpext a))
4416 // which is a no-op. Mark it as a TRUNCating FP_ROUND.
4417 const bool isTrunc = (Node->getOpcode() == ISD::FCOPYSIGN);
4418 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4419 Tmp3, DAG.getIntPtrConstant(isTrunc, dl)));
4425 case ISD::FNEARBYINT:
4438 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4439 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4440 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4441 Tmp2, DAG.getIntPtrConstant(0, dl)));
4444 case ISD::BUILD_VECTOR: {
4445 MVT EltVT = OVT.getVectorElementType();
4446 MVT NewEltVT = NVT.getVectorElementType();
4448 // Handle bitcasts to a different vector type with the same total bit size
4450 // e.g. v2i64 = build_vector i64:x, i64:y => v4i32
4452 // v4i32 = concat_vectors (v2i32 (bitcast i64:x)), (v2i32 (bitcast i64:y))
4454 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
4455 "Invalid promote type for build_vector");
4456 assert(NewEltVT.bitsLT(EltVT) && "not handled");
4458 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4460 SmallVector<SDValue, 8> NewOps;
4461 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I) {
4462 SDValue Op = Node->getOperand(I);
4463 NewOps.push_back(DAG.getNode(ISD::BITCAST, SDLoc(Op), MidVT, Op));
4467 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewOps);
4468 SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat);
4469 Results.push_back(CvtVec);
4472 case ISD::EXTRACT_VECTOR_ELT: {
4473 MVT EltVT = OVT.getVectorElementType();
4474 MVT NewEltVT = NVT.getVectorElementType();
4476 // Handle bitcasts to a different vector type with the same total bit size.
4478 // e.g. v2i64 = extract_vector_elt x:v2i64, y:i32
4480 // v4i32:castx = bitcast x:v2i64
4483 // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))),
4484 // (i32 (extract_vector_elt castx, (2 * y + 1)))
4487 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
4488 "Invalid promote type for extract_vector_elt");
4489 assert(NewEltVT.bitsLT(EltVT) && "not handled");
4491 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4492 unsigned NewEltsPerOldElt = MidVT.getVectorNumElements();
4494 SDValue Idx = Node->getOperand(1);
4495 EVT IdxVT = Idx.getValueType();
4497 SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SL, IdxVT);
4498 SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor);
4500 SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0));
4502 SmallVector<SDValue, 8> NewOps;
4503 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
4504 SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT);
4505 SDValue TmpIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset);
4507 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT,
4509 NewOps.push_back(Elt);
4512 SDValue NewVec = DAG.getBuildVector(MidVT, SL, NewOps);
4513 Results.push_back(DAG.getNode(ISD::BITCAST, SL, EltVT, NewVec));
4516 case ISD::INSERT_VECTOR_ELT: {
4517 MVT EltVT = OVT.getVectorElementType();
4518 MVT NewEltVT = NVT.getVectorElementType();
4520 // Handle bitcasts to a different vector type with the same total bit size
4522 // e.g. v2i64 = insert_vector_elt x:v2i64, y:i64, z:i32
4524 // v4i32:castx = bitcast x:v2i64
4525 // v2i32:casty = bitcast y:i64
4528 // (v4i32 insert_vector_elt
4529 // (v4i32 insert_vector_elt v4i32:castx,
4530 // (extract_vector_elt casty, 0), 2 * z),
4531 // (extract_vector_elt casty, 1), (2 * z + 1))
4533 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
4534 "Invalid promote type for insert_vector_elt");
4535 assert(NewEltVT.bitsLT(EltVT) && "not handled");
4537 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4538 unsigned NewEltsPerOldElt = MidVT.getVectorNumElements();
4540 SDValue Val = Node->getOperand(1);
4541 SDValue Idx = Node->getOperand(2);
4542 EVT IdxVT = Idx.getValueType();
4545 SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SDLoc(), IdxVT);
4546 SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor);
4548 SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0));
4549 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val);
4551 SDValue NewVec = CastVec;
4552 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
4553 SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT);
4554 SDValue InEltIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset);
4556 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT,
4557 CastVal, IdxOffset);
4559 NewVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, SL, NVT,
4560 NewVec, Elt, InEltIdx);
4563 Results.push_back(DAG.getNode(ISD::BITCAST, SL, OVT, NewVec));
4566 case ISD::SCALAR_TO_VECTOR: {
4567 MVT EltVT = OVT.getVectorElementType();
4568 MVT NewEltVT = NVT.getVectorElementType();
4570 // Handle bitcasts to different vector type with the same total bit size.
4572 // e.g. v2i64 = scalar_to_vector x:i64
4574 // concat_vectors (v2i32 bitcast x:i64), (v2i32 undef)
4577 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4578 SDValue Val = Node->getOperand(0);
4581 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val);
4582 SDValue Undef = DAG.getUNDEF(MidVT);
4584 SmallVector<SDValue, 8> NewElts;
4585 NewElts.push_back(CastVal);
4586 for (unsigned I = 1, NElts = OVT.getVectorNumElements(); I != NElts; ++I)
4587 NewElts.push_back(Undef);
4589 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewElts);
4590 SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat);
4591 Results.push_back(CvtVec);
4596 // Replace the original node with the legalized result.
4597 if (!Results.empty())
4598 ReplaceNode(Node, Results.data());
4601 /// This is the entry point for the file.
4602 void SelectionDAG::Legalize() {
4603 AssignTopologicalOrder();
4605 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4606 SelectionDAGLegalize Legalizer(*this, LegalizedNodes);
4608 // Visit all the nodes. We start in topological order, so that we see
4609 // nodes with their original operands intact. Legalization can produce
4610 // new nodes which may themselves need to be legalized. Iterate until all
4611 // nodes have been legalized.
4613 bool AnyLegalized = false;
4614 for (auto NI = allnodes_end(); NI != allnodes_begin();) {
4618 if (N->use_empty() && N != getRoot().getNode()) {
4624 if (LegalizedNodes.insert(N).second) {
4625 AnyLegalized = true;
4626 Legalizer.LegalizeOp(N);
4628 if (N->use_empty() && N != getRoot().getNode()) {
4639 // Remove dead nodes now.
4643 bool SelectionDAG::LegalizeOp(SDNode *N,
4644 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4645 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4646 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes);
4648 // Directly insert the node in question, and legalize it. This will recurse
4649 // as needed through operands.
4650 LegalizedNodes.insert(N);
4651 Legalizer.LegalizeOp(N);
4653 return LegalizedNodes.count(N);