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/APFloat.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/CodeGen/ISDOpcodes.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineJumpTableInfo.h"
24 #include "llvm/CodeGen/MachineMemOperand.h"
25 #include "llvm/CodeGen/MachineValueType.h"
26 #include "llvm/CodeGen/RuntimeLibcalls.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/SelectionDAGNodes.h"
29 #include "llvm/CodeGen/TargetFrameLowering.h"
30 #include "llvm/CodeGen/TargetLowering.h"
31 #include "llvm/CodeGen/TargetSubtargetInfo.h"
32 #include "llvm/CodeGen/ValueTypes.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/Metadata.h"
39 #include "llvm/IR/Type.h"
40 #include "llvm/Support/Casting.h"
41 #include "llvm/Support/Compiler.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Target/TargetMachine.h"
47 #include "llvm/Target/TargetOptions.h"
56 #define DEBUG_TYPE "legalizedag"
60 /// Keeps track of state when getting the sign of a floating-point value as an
62 struct FloatSignAsInt {
67 MachinePointerInfo IntPointerInfo;
68 MachinePointerInfo FloatPointerInfo;
74 //===----------------------------------------------------------------------===//
75 /// This takes an arbitrary SelectionDAG as input and
76 /// hacks on it until the target machine can handle it. This involves
77 /// eliminating value sizes the machine cannot handle (promoting small sizes to
78 /// large sizes or splitting up large values into small values) as well as
79 /// eliminating operations the machine cannot handle.
81 /// This code also does a small amount of optimization and recognition of idioms
82 /// as part of its processing. For example, if a target does not support a
83 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
84 /// will attempt merge setcc and brc instructions into brcc's.
85 class SelectionDAGLegalize {
86 const TargetMachine &TM;
87 const TargetLowering &TLI;
90 /// \brief The set of nodes which have already been legalized. We hold a
91 /// reference to it in order to update as necessary on node deletion.
92 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
94 /// \brief A set of all the nodes updated during legalization.
95 SmallSetVector<SDNode *, 16> *UpdatedNodes;
97 EVT getSetCCResultType(EVT VT) const {
98 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
101 // Libcall insertion helpers.
104 SelectionDAGLegalize(SelectionDAG &DAG,
105 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
106 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
107 : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
108 LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {}
110 /// \brief Legalizes the given operation.
111 void LegalizeOp(SDNode *Node);
114 SDValue OptimizeFloatStore(StoreSDNode *ST);
116 void LegalizeLoadOps(SDNode *Node);
117 void LegalizeStoreOps(SDNode *Node);
119 /// Some targets cannot handle a variable
120 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
121 /// is necessary to spill the vector being inserted into to memory, perform
122 /// the insert there, and then read the result back.
123 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
125 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx,
128 /// Return a vector shuffle operation which
129 /// performs the same shuffe in terms of order or result bytes, but on a type
130 /// whose vector element type is narrower than the original shuffle type.
131 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
132 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, const SDLoc &dl,
133 SDValue N1, SDValue N2,
134 ArrayRef<int> Mask) const;
136 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
137 bool &NeedInvert, const SDLoc &dl);
139 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
140 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
141 unsigned NumOps, bool isSigned, const SDLoc &dl);
143 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
144 SDNode *Node, bool isSigned);
145 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
146 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
147 RTLIB::Libcall Call_F128,
148 RTLIB::Libcall Call_PPCF128);
149 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
150 RTLIB::Libcall Call_I8,
151 RTLIB::Libcall Call_I16,
152 RTLIB::Libcall Call_I32,
153 RTLIB::Libcall Call_I64,
154 RTLIB::Libcall Call_I128);
155 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
156 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
158 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT,
160 SDValue ExpandBUILD_VECTOR(SDNode *Node);
161 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
162 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
163 SmallVectorImpl<SDValue> &Results);
164 void getSignAsIntValue(FloatSignAsInt &State, const SDLoc &DL,
165 SDValue Value) const;
166 SDValue modifySignAsInt(const FloatSignAsInt &State, const SDLoc &DL,
167 SDValue NewIntValue) const;
168 SDValue ExpandFCOPYSIGN(SDNode *Node) const;
169 SDValue ExpandFABS(SDNode *Node) const;
170 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
172 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
174 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
177 SDValue ExpandBITREVERSE(SDValue Op, const SDLoc &dl);
178 SDValue ExpandBSWAP(SDValue Op, const SDLoc &dl);
179 SDValue ExpandBitCount(unsigned Opc, SDValue Op, const SDLoc &dl);
181 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
182 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
183 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
185 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
186 SDValue ExpandConstant(ConstantSDNode *CP);
188 // if ExpandNode returns false, LegalizeOp falls back to ConvertNodeToLibcall
189 bool ExpandNode(SDNode *Node);
190 void ConvertNodeToLibcall(SDNode *Node);
191 void PromoteNode(SDNode *Node);
194 // Node replacement helpers
196 void ReplacedNode(SDNode *N) {
197 LegalizedNodes.erase(N);
199 UpdatedNodes->insert(N);
202 void ReplaceNode(SDNode *Old, SDNode *New) {
203 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
204 dbgs() << " with: "; New->dump(&DAG));
206 assert(Old->getNumValues() == New->getNumValues() &&
207 "Replacing one node with another that produces a different number "
209 DAG.ReplaceAllUsesWith(Old, New);
211 UpdatedNodes->insert(New);
215 void ReplaceNode(SDValue Old, SDValue New) {
216 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
217 dbgs() << " with: "; New->dump(&DAG));
219 DAG.ReplaceAllUsesWith(Old, New);
221 UpdatedNodes->insert(New.getNode());
222 ReplacedNode(Old.getNode());
225 void ReplaceNode(SDNode *Old, const SDValue *New) {
226 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG));
228 DAG.ReplaceAllUsesWith(Old, New);
229 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
230 DEBUG(dbgs() << (i == 0 ? " with: "
234 UpdatedNodes->insert(New[i].getNode());
240 } // end anonymous namespace
242 /// Return a vector shuffle operation which
243 /// performs the same shuffe in terms of order or result bytes, but on a type
244 /// whose vector element type is narrower than the original shuffle type.
245 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
246 SDValue SelectionDAGLegalize::ShuffleWithNarrowerEltType(
247 EVT NVT, EVT VT, const SDLoc &dl, SDValue N1, SDValue N2,
248 ArrayRef<int> Mask) const {
249 unsigned NumMaskElts = VT.getVectorNumElements();
250 unsigned NumDestElts = NVT.getVectorNumElements();
251 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
253 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
255 if (NumEltsGrowth == 1)
256 return DAG.getVectorShuffle(NVT, dl, N1, N2, Mask);
258 SmallVector<int, 8> NewMask;
259 for (unsigned i = 0; i != NumMaskElts; ++i) {
261 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
263 NewMask.push_back(-1);
265 NewMask.push_back(Idx * NumEltsGrowth + j);
268 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
269 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
270 return DAG.getVectorShuffle(NVT, dl, N1, N2, NewMask);
273 /// Expands the ConstantFP node to an integer constant or
274 /// a load from the constant pool.
276 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
280 // If a FP immediate is precise when represented as a float and if the
281 // target can do an extending load from float to double, we put it into
282 // the constant pool as a float, even if it's is statically typed as a
283 // double. This shrinks FP constants and canonicalizes them for targets where
284 // an FP extending load is the same cost as a normal load (such as on the x87
285 // fp stack or PPC FP unit).
286 EVT VT = CFP->getValueType(0);
287 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
289 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
290 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(), dl,
291 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
294 APFloat APF = CFP->getValueAPF();
298 // We don't want to shrink SNaNs. Converting the SNaN back to its real type
299 // can cause it to be changed into a QNaN on some platforms (e.g. on SystemZ).
300 if (!APF.isSignaling()) {
301 while (SVT != MVT::f32 && SVT != MVT::f16) {
302 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
303 if (ConstantFPSDNode::isValueValidForType(SVT, APF) &&
304 // Only do this if the target has a native EXTLOAD instruction from
306 TLI.isLoadExtLegal(ISD::EXTLOAD, OrigVT, SVT) &&
307 TLI.ShouldShrinkFPConstant(OrigVT)) {
308 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
309 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
317 DAG.getConstantPool(LLVMC, TLI.getPointerTy(DAG.getDataLayout()));
318 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
320 SDValue Result = DAG.getExtLoad(
321 ISD::EXTLOAD, dl, OrigVT, DAG.getEntryNode(), CPIdx,
322 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), VT,
326 SDValue Result = DAG.getLoad(
327 OrigVT, dl, DAG.getEntryNode(), CPIdx,
328 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Alignment);
332 /// Expands the Constant node to a load from the constant pool.
333 SDValue SelectionDAGLegalize::ExpandConstant(ConstantSDNode *CP) {
335 EVT VT = CP->getValueType(0);
336 SDValue CPIdx = DAG.getConstantPool(CP->getConstantIntValue(),
337 TLI.getPointerTy(DAG.getDataLayout()));
338 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
339 SDValue Result = DAG.getLoad(
340 VT, dl, DAG.getEntryNode(), CPIdx,
341 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Alignment);
345 /// Some target cannot handle a variable insertion index for the
346 /// INSERT_VECTOR_ELT instruction. In this case, it
347 /// is necessary to spill the vector being inserted into to memory, perform
348 /// the insert there, and then read the result back.
349 SDValue SelectionDAGLegalize::PerformInsertVectorEltInMemory(SDValue Vec,
357 // If the target doesn't support this, we have to spill the input vector
358 // to a temporary stack slot, update the element, then reload it. This is
359 // badness. We could also load the value into a vector register (either
360 // with a "move to register" or "extload into register" instruction, then
361 // permute it into place, if the idx is a constant and if the idx is
362 // supported by the target.
363 EVT VT = Tmp1.getValueType();
364 EVT EltVT = VT.getVectorElementType();
365 SDValue StackPtr = DAG.CreateStackTemporary(VT);
367 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
370 SDValue Ch = DAG.getStore(
371 DAG.getEntryNode(), dl, Tmp1, StackPtr,
372 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI));
374 SDValue StackPtr2 = TLI.getVectorElementPointer(DAG, StackPtr, VT, Tmp3);
376 // Store the scalar value.
377 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT);
378 // Load the updated vector.
379 return DAG.getLoad(VT, dl, Ch, StackPtr, MachinePointerInfo::getFixedStack(
380 DAG.getMachineFunction(), SPFI));
383 SDValue SelectionDAGLegalize::ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
386 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
387 // SCALAR_TO_VECTOR requires that the type of the value being inserted
388 // match the element type of the vector being created, except for
389 // integers in which case the inserted value can be over width.
390 EVT EltVT = Vec.getValueType().getVectorElementType();
391 if (Val.getValueType() == EltVT ||
392 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
393 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
394 Vec.getValueType(), Val);
396 unsigned NumElts = Vec.getValueType().getVectorNumElements();
397 // We generate a shuffle of InVec and ScVec, so the shuffle mask
398 // should be 0,1,2,3,4,5... with the appropriate element replaced with
400 SmallVector<int, 8> ShufOps;
401 for (unsigned i = 0; i != NumElts; ++i)
402 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
404 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec, ShufOps);
407 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
410 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
411 DEBUG(dbgs() << "Optimizing float store operations\n");
412 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
413 // FIXME: We shouldn't do this for TargetConstantFP's.
414 // FIXME: move this to the DAG Combiner! Note that we can't regress due
415 // to phase ordering between legalized code and the dag combiner. This
416 // probably means that we need to integrate dag combiner and legalizer
418 // We generally can't do this one for long doubles.
419 SDValue Chain = ST->getChain();
420 SDValue Ptr = ST->getBasePtr();
421 unsigned Alignment = ST->getAlignment();
422 MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags();
423 AAMDNodes AAInfo = ST->getAAInfo();
425 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
426 if (CFP->getValueType(0) == MVT::f32 &&
427 TLI.isTypeLegal(MVT::i32)) {
428 SDValue Con = DAG.getConstant(CFP->getValueAPF().
429 bitcastToAPInt().zextOrTrunc(32),
430 SDLoc(CFP), MVT::i32);
431 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(), Alignment,
435 if (CFP->getValueType(0) == MVT::f64) {
436 // If this target supports 64-bit registers, do a single 64-bit store.
437 if (TLI.isTypeLegal(MVT::i64)) {
438 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
439 zextOrTrunc(64), SDLoc(CFP), MVT::i64);
440 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
441 Alignment, MMOFlags, AAInfo);
444 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
445 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
446 // stores. If the target supports neither 32- nor 64-bits, this
447 // xform is certainly not worth it.
448 const APInt &IntVal = CFP->getValueAPF().bitcastToAPInt();
449 SDValue Lo = DAG.getConstant(IntVal.trunc(32), dl, MVT::i32);
450 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), dl, MVT::i32);
451 if (DAG.getDataLayout().isBigEndian())
454 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), Alignment,
456 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
457 DAG.getConstant(4, dl, Ptr.getValueType()));
458 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
459 ST->getPointerInfo().getWithOffset(4),
460 MinAlign(Alignment, 4U), MMOFlags, AAInfo);
462 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
466 return SDValue(nullptr, 0);
469 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
470 StoreSDNode *ST = cast<StoreSDNode>(Node);
471 SDValue Chain = ST->getChain();
472 SDValue Ptr = ST->getBasePtr();
475 unsigned Alignment = ST->getAlignment();
476 MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags();
477 AAMDNodes AAInfo = ST->getAAInfo();
479 if (!ST->isTruncatingStore()) {
480 DEBUG(dbgs() << "Legalizing store operation\n");
481 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
482 ReplaceNode(ST, OptStore);
486 SDValue Value = ST->getValue();
487 MVT VT = Value.getSimpleValueType();
488 switch (TLI.getOperationAction(ISD::STORE, VT)) {
489 default: llvm_unreachable("This action is not supported yet!");
490 case TargetLowering::Legal: {
491 // If this is an unaligned store and the target doesn't support it,
493 EVT MemVT = ST->getMemoryVT();
494 unsigned AS = ST->getAddressSpace();
495 unsigned Align = ST->getAlignment();
496 const DataLayout &DL = DAG.getDataLayout();
497 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
498 DEBUG(dbgs() << "Expanding unsupported unaligned store\n");
499 SDValue Result = TLI.expandUnalignedStore(ST, DAG);
500 ReplaceNode(SDValue(ST, 0), Result);
502 DEBUG(dbgs() << "Legal store\n");
505 case TargetLowering::Custom: {
506 DEBUG(dbgs() << "Trying custom lowering\n");
507 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
508 if (Res && Res != SDValue(Node, 0))
509 ReplaceNode(SDValue(Node, 0), Res);
512 case TargetLowering::Promote: {
513 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
514 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
515 "Can only promote stores to same size type");
516 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
518 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
519 Alignment, MMOFlags, AAInfo);
520 ReplaceNode(SDValue(Node, 0), Result);
527 DEBUG(dbgs() << "Legalizing truncating store operations\n");
528 SDValue Value = ST->getValue();
529 EVT StVT = ST->getMemoryVT();
530 unsigned StWidth = StVT.getSizeInBits();
531 auto &DL = DAG.getDataLayout();
533 if (StWidth != StVT.getStoreSizeInBits()) {
534 // Promote to a byte-sized store with upper bits zero if not
535 // storing an integral number of bytes. For example, promote
536 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
537 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
538 StVT.getStoreSizeInBits());
539 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
541 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(), NVT,
542 Alignment, MMOFlags, AAInfo);
543 ReplaceNode(SDValue(Node, 0), Result);
544 } else if (StWidth & (StWidth - 1)) {
545 // If not storing a power-of-2 number of bits, expand as two stores.
546 assert(!StVT.isVector() && "Unsupported truncstore!");
547 unsigned RoundWidth = 1 << Log2_32(StWidth);
548 assert(RoundWidth < StWidth);
549 unsigned ExtraWidth = StWidth - RoundWidth;
550 assert(ExtraWidth < RoundWidth);
551 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
552 "Store size not an integral number of bytes!");
553 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
554 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
556 unsigned IncrementSize;
558 if (DL.isLittleEndian()) {
559 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
560 // Store the bottom RoundWidth bits.
561 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
562 RoundVT, Alignment, MMOFlags, AAInfo);
564 // Store the remaining ExtraWidth bits.
565 IncrementSize = RoundWidth / 8;
566 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
567 DAG.getConstant(IncrementSize, dl,
568 Ptr.getValueType()));
570 ISD::SRL, dl, Value.getValueType(), Value,
571 DAG.getConstant(RoundWidth, dl,
572 TLI.getShiftAmountTy(Value.getValueType(), DL)));
573 Hi = DAG.getTruncStore(
575 ST->getPointerInfo().getWithOffset(IncrementSize), ExtraVT,
576 MinAlign(Alignment, IncrementSize), MMOFlags, AAInfo);
578 // Big endian - avoid unaligned stores.
579 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
580 // Store the top RoundWidth bits.
582 ISD::SRL, dl, Value.getValueType(), Value,
583 DAG.getConstant(ExtraWidth, dl,
584 TLI.getShiftAmountTy(Value.getValueType(), DL)));
585 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
586 RoundVT, Alignment, MMOFlags, AAInfo);
588 // Store the remaining ExtraWidth bits.
589 IncrementSize = RoundWidth / 8;
590 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
591 DAG.getConstant(IncrementSize, dl,
592 Ptr.getValueType()));
593 Lo = DAG.getTruncStore(
594 Chain, dl, Value, Ptr,
595 ST->getPointerInfo().getWithOffset(IncrementSize), ExtraVT,
596 MinAlign(Alignment, IncrementSize), MMOFlags, AAInfo);
599 // The order of the stores doesn't matter.
600 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
601 ReplaceNode(SDValue(Node, 0), Result);
603 switch (TLI.getTruncStoreAction(ST->getValue().getValueType(), StVT)) {
604 default: llvm_unreachable("This action is not supported yet!");
605 case TargetLowering::Legal: {
606 EVT MemVT = ST->getMemoryVT();
607 unsigned AS = ST->getAddressSpace();
608 unsigned Align = ST->getAlignment();
609 // If this is an unaligned store and the target doesn't support it,
611 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
612 SDValue Result = TLI.expandUnalignedStore(ST, DAG);
613 ReplaceNode(SDValue(ST, 0), Result);
617 case TargetLowering::Custom: {
618 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
619 if (Res && Res != SDValue(Node, 0))
620 ReplaceNode(SDValue(Node, 0), Res);
623 case TargetLowering::Expand:
624 assert(!StVT.isVector() &&
625 "Vector Stores are handled in LegalizeVectorOps");
629 // TRUNCSTORE:i16 i32 -> STORE i16
630 if (TLI.isTypeLegal(StVT)) {
631 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
632 Result = DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
633 Alignment, MMOFlags, AAInfo);
635 // The in-memory type isn't legal. Truncate to the type it would promote
636 // to, and then do a truncstore.
637 Value = DAG.getNode(ISD::TRUNCATE, dl,
638 TLI.getTypeToTransformTo(*DAG.getContext(), StVT),
640 Result = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
641 StVT, Alignment, MMOFlags, AAInfo);
644 ReplaceNode(SDValue(Node, 0), Result);
650 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
651 LoadSDNode *LD = cast<LoadSDNode>(Node);
652 SDValue Chain = LD->getChain(); // The chain.
653 SDValue Ptr = LD->getBasePtr(); // The base pointer.
654 SDValue Value; // The value returned by the load op.
657 ISD::LoadExtType ExtType = LD->getExtensionType();
658 if (ExtType == ISD::NON_EXTLOAD) {
659 DEBUG(dbgs() << "Legalizing non-extending load operation\n");
660 MVT VT = Node->getSimpleValueType(0);
661 SDValue RVal = SDValue(Node, 0);
662 SDValue RChain = SDValue(Node, 1);
664 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
665 default: llvm_unreachable("This action is not supported yet!");
666 case TargetLowering::Legal: {
667 EVT MemVT = LD->getMemoryVT();
668 unsigned AS = LD->getAddressSpace();
669 unsigned Align = LD->getAlignment();
670 const DataLayout &DL = DAG.getDataLayout();
671 // If this is an unaligned load and the target doesn't support it,
673 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
674 std::tie(RVal, RChain) = TLI.expandUnalignedLoad(LD, DAG);
678 case TargetLowering::Custom:
679 if (SDValue Res = TLI.LowerOperation(RVal, DAG)) {
681 RChain = Res.getValue(1);
685 case TargetLowering::Promote: {
686 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
687 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
688 "Can only promote loads to same size type");
690 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
691 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
692 RChain = Res.getValue(1);
696 if (RChain.getNode() != Node) {
697 assert(RVal.getNode() != Node && "Load must be completely replaced");
698 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
699 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
701 UpdatedNodes->insert(RVal.getNode());
702 UpdatedNodes->insert(RChain.getNode());
709 DEBUG(dbgs() << "Legalizing extending load operation\n");
710 EVT SrcVT = LD->getMemoryVT();
711 unsigned SrcWidth = SrcVT.getSizeInBits();
712 unsigned Alignment = LD->getAlignment();
713 MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags();
714 AAMDNodes AAInfo = LD->getAAInfo();
716 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
717 // Some targets pretend to have an i1 loading operation, and actually
718 // load an i8. This trick is correct for ZEXTLOAD because the top 7
719 // bits are guaranteed to be zero; it helps the optimizers understand
720 // that these bits are zero. It is also useful for EXTLOAD, since it
721 // tells the optimizers that those bits are undefined. It would be
722 // nice to have an effective generic way of getting these benefits...
723 // Until such a way is found, don't insist on promoting i1 here.
725 TLI.getLoadExtAction(ExtType, Node->getValueType(0), MVT::i1) ==
726 TargetLowering::Promote)) {
727 // Promote to a byte-sized load if not loading an integral number of
728 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
729 unsigned NewWidth = SrcVT.getStoreSizeInBits();
730 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
733 // The extra bits are guaranteed to be zero, since we stored them that
734 // way. A zext load from NVT thus automatically gives zext from SrcVT.
736 ISD::LoadExtType NewExtType =
737 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
740 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0), Chain, Ptr,
741 LD->getPointerInfo(), NVT, Alignment, MMOFlags, AAInfo);
743 Ch = Result.getValue(1); // The chain.
745 if (ExtType == ISD::SEXTLOAD)
746 // Having the top bits zero doesn't help when sign extending.
747 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
748 Result.getValueType(),
749 Result, DAG.getValueType(SrcVT));
750 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
751 // All the top bits are guaranteed to be zero - inform the optimizers.
752 Result = DAG.getNode(ISD::AssertZext, dl,
753 Result.getValueType(), Result,
754 DAG.getValueType(SrcVT));
758 } else if (SrcWidth & (SrcWidth - 1)) {
759 // If not loading a power-of-2 number of bits, expand as two loads.
760 assert(!SrcVT.isVector() && "Unsupported extload!");
761 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
762 assert(RoundWidth < SrcWidth);
763 unsigned ExtraWidth = SrcWidth - RoundWidth;
764 assert(ExtraWidth < RoundWidth);
765 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
766 "Load size not an integral number of bytes!");
767 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
768 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
770 unsigned IncrementSize;
771 auto &DL = DAG.getDataLayout();
773 if (DL.isLittleEndian()) {
774 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
775 // Load the bottom RoundWidth bits.
776 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), Chain, Ptr,
777 LD->getPointerInfo(), RoundVT, Alignment, MMOFlags,
780 // Load the remaining ExtraWidth bits.
781 IncrementSize = RoundWidth / 8;
782 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
783 DAG.getConstant(IncrementSize, dl,
784 Ptr.getValueType()));
785 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
786 LD->getPointerInfo().getWithOffset(IncrementSize),
787 ExtraVT, MinAlign(Alignment, IncrementSize), MMOFlags,
790 // Build a factor node to remember that this load is independent of
792 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
795 // Move the top bits to the right place.
797 ISD::SHL, dl, Hi.getValueType(), Hi,
798 DAG.getConstant(RoundWidth, dl,
799 TLI.getShiftAmountTy(Hi.getValueType(), DL)));
801 // Join the hi and lo parts.
802 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
804 // Big endian - avoid unaligned loads.
805 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
806 // Load the top RoundWidth bits.
807 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
808 LD->getPointerInfo(), RoundVT, Alignment, MMOFlags,
811 // Load the remaining ExtraWidth bits.
812 IncrementSize = RoundWidth / 8;
813 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
814 DAG.getConstant(IncrementSize, dl,
815 Ptr.getValueType()));
816 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), Chain, Ptr,
817 LD->getPointerInfo().getWithOffset(IncrementSize),
818 ExtraVT, MinAlign(Alignment, IncrementSize), MMOFlags,
821 // Build a factor node to remember that this load is independent of
823 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
826 // Move the top bits to the right place.
828 ISD::SHL, dl, Hi.getValueType(), Hi,
829 DAG.getConstant(ExtraWidth, dl,
830 TLI.getShiftAmountTy(Hi.getValueType(), DL)));
832 // Join the hi and lo parts.
833 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
838 bool isCustom = false;
839 switch (TLI.getLoadExtAction(ExtType, Node->getValueType(0),
840 SrcVT.getSimpleVT())) {
841 default: llvm_unreachable("This action is not supported yet!");
842 case TargetLowering::Custom:
845 case TargetLowering::Legal:
846 Value = SDValue(Node, 0);
847 Chain = SDValue(Node, 1);
850 if (SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG)) {
852 Chain = Res.getValue(1);
855 // If this is an unaligned load and the target doesn't support it,
857 EVT MemVT = LD->getMemoryVT();
858 unsigned AS = LD->getAddressSpace();
859 unsigned Align = LD->getAlignment();
860 const DataLayout &DL = DAG.getDataLayout();
861 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) {
862 std::tie(Value, Chain) = TLI.expandUnalignedLoad(LD, DAG);
867 case TargetLowering::Expand: {
868 EVT DestVT = Node->getValueType(0);
869 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, DestVT, SrcVT)) {
870 // If the source type is not legal, see if there is a legal extload to
871 // an intermediate type that we can then extend further.
872 EVT LoadVT = TLI.getRegisterType(SrcVT.getSimpleVT());
873 if (TLI.isTypeLegal(SrcVT) || // Same as SrcVT == LoadVT?
874 TLI.isLoadExtLegal(ExtType, LoadVT, SrcVT)) {
875 // If we are loading a legal type, this is a non-extload followed by a
877 ISD::LoadExtType MidExtType =
878 (LoadVT == SrcVT) ? ISD::NON_EXTLOAD : ExtType;
880 SDValue Load = DAG.getExtLoad(MidExtType, dl, LoadVT, Chain, Ptr,
881 SrcVT, LD->getMemOperand());
883 ISD::getExtForLoadExtType(SrcVT.isFloatingPoint(), ExtType);
884 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
885 Chain = Load.getValue(1);
889 // Handle the special case of fp16 extloads. EXTLOAD doesn't have the
890 // normal undefined upper bits behavior to allow using an in-reg extend
891 // with the illegal FP type, so load as an integer and do the
892 // from-integer conversion.
893 if (SrcVT.getScalarType() == MVT::f16) {
894 EVT ISrcVT = SrcVT.changeTypeToInteger();
895 EVT IDestVT = DestVT.changeTypeToInteger();
896 EVT LoadVT = TLI.getRegisterType(IDestVT.getSimpleVT());
898 SDValue Result = DAG.getExtLoad(ISD::ZEXTLOAD, dl, LoadVT,
900 LD->getMemOperand());
901 Value = DAG.getNode(ISD::FP16_TO_FP, dl, DestVT, Result);
902 Chain = Result.getValue(1);
907 assert(!SrcVT.isVector() &&
908 "Vector Loads are handled in LegalizeVectorOps");
910 // FIXME: This does not work for vectors on most targets. Sign-
911 // and zero-extend operations are currently folded into extending
912 // loads, whether they are legal or not, and then we end up here
913 // without any support for legalizing them.
914 assert(ExtType != ISD::EXTLOAD &&
915 "EXTLOAD should always be supported!");
916 // Turn the unsupported load into an EXTLOAD followed by an
917 // explicit zero/sign extend inreg.
918 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
919 Node->getValueType(0),
921 LD->getMemOperand());
923 if (ExtType == ISD::SEXTLOAD)
924 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
925 Result.getValueType(),
926 Result, DAG.getValueType(SrcVT));
928 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
930 Chain = Result.getValue(1);
936 // Since loads produce two values, make sure to remember that we legalized
938 if (Chain.getNode() != Node) {
939 assert(Value.getNode() != Node && "Load must be completely replaced");
940 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
941 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
943 UpdatedNodes->insert(Value.getNode());
944 UpdatedNodes->insert(Chain.getNode());
950 static TargetLowering::LegalizeAction
951 getStrictFPOpcodeAction(const TargetLowering &TLI, unsigned Opcode, EVT VT) {
954 default: llvm_unreachable("Unexpected FP pseudo-opcode");
955 case ISD::STRICT_FSQRT: EqOpc = ISD::FSQRT; break;
956 case ISD::STRICT_FPOW: EqOpc = ISD::FPOW; break;
957 case ISD::STRICT_FPOWI: EqOpc = ISD::FPOWI; break;
958 case ISD::STRICT_FMA: EqOpc = ISD::FMA; break;
959 case ISD::STRICT_FSIN: EqOpc = ISD::FSIN; break;
960 case ISD::STRICT_FCOS: EqOpc = ISD::FCOS; break;
961 case ISD::STRICT_FEXP: EqOpc = ISD::FEXP; break;
962 case ISD::STRICT_FEXP2: EqOpc = ISD::FEXP2; break;
963 case ISD::STRICT_FLOG: EqOpc = ISD::FLOG; break;
964 case ISD::STRICT_FLOG10: EqOpc = ISD::FLOG10; break;
965 case ISD::STRICT_FLOG2: EqOpc = ISD::FLOG2; break;
966 case ISD::STRICT_FRINT: EqOpc = ISD::FRINT; break;
967 case ISD::STRICT_FNEARBYINT: EqOpc = ISD::FNEARBYINT; break;
970 auto Action = TLI.getOperationAction(EqOpc, VT);
972 // We don't currently handle Custom or Promote for strict FP pseudo-ops.
973 // For now, we just expand for those cases.
974 if (Action != TargetLowering::Legal)
975 Action = TargetLowering::Expand;
980 /// Return a legal replacement for the given operation, with all legal operands.
981 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
982 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
984 // Allow illegal target nodes and illegal registers.
985 if (Node->getOpcode() == ISD::TargetConstant ||
986 Node->getOpcode() == ISD::Register)
990 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
991 assert((TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
992 TargetLowering::TypeLegal ||
993 TLI.isTypeLegal(Node->getValueType(i))) &&
994 "Unexpected illegal type!");
996 for (const SDValue &Op : Node->op_values())
997 assert((TLI.getTypeAction(*DAG.getContext(), Op.getValueType()) ==
998 TargetLowering::TypeLegal ||
999 TLI.isTypeLegal(Op.getValueType()) ||
1000 Op.getOpcode() == ISD::TargetConstant ||
1001 Op.getOpcode() == ISD::Register) &&
1002 "Unexpected illegal type!");
1005 // Figure out the correct action; the way to query this varies by opcode
1006 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1007 bool SimpleFinishLegalizing = true;
1008 switch (Node->getOpcode()) {
1009 case ISD::INTRINSIC_W_CHAIN:
1010 case ISD::INTRINSIC_WO_CHAIN:
1011 case ISD::INTRINSIC_VOID:
1012 case ISD::STACKSAVE:
1013 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1015 case ISD::GET_DYNAMIC_AREA_OFFSET:
1016 Action = TLI.getOperationAction(Node->getOpcode(),
1017 Node->getValueType(0));
1020 Action = TLI.getOperationAction(Node->getOpcode(),
1021 Node->getValueType(0));
1022 if (Action != TargetLowering::Promote)
1023 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1025 case ISD::FP_TO_FP16:
1026 case ISD::SINT_TO_FP:
1027 case ISD::UINT_TO_FP:
1028 case ISD::EXTRACT_VECTOR_ELT:
1029 Action = TLI.getOperationAction(Node->getOpcode(),
1030 Node->getOperand(0).getValueType());
1032 case ISD::FP_ROUND_INREG:
1033 case ISD::SIGN_EXTEND_INREG: {
1034 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1035 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1038 case ISD::ATOMIC_STORE:
1039 Action = TLI.getOperationAction(Node->getOpcode(),
1040 Node->getOperand(2).getValueType());
1042 case ISD::SELECT_CC:
1045 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1046 Node->getOpcode() == ISD::SETCC ? 2 :
1047 Node->getOpcode() == ISD::SETCCE ? 3 : 1;
1048 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1049 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1050 ISD::CondCode CCCode =
1051 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1052 Action = TLI.getCondCodeAction(CCCode, OpVT);
1053 if (Action == TargetLowering::Legal) {
1054 if (Node->getOpcode() == ISD::SELECT_CC)
1055 Action = TLI.getOperationAction(Node->getOpcode(),
1056 Node->getValueType(0));
1058 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1064 // FIXME: Model these properly. LOAD and STORE are complicated, and
1065 // STORE expects the unlegalized operand in some cases.
1066 SimpleFinishLegalizing = false;
1068 case ISD::CALLSEQ_START:
1069 case ISD::CALLSEQ_END:
1070 // FIXME: This shouldn't be necessary. These nodes have special properties
1071 // dealing with the recursive nature of legalization. Removing this
1072 // special case should be done as part of making LegalizeDAG non-recursive.
1073 SimpleFinishLegalizing = false;
1075 case ISD::EXTRACT_ELEMENT:
1076 case ISD::FLT_ROUNDS_:
1077 case ISD::MERGE_VALUES:
1078 case ISD::EH_RETURN:
1079 case ISD::FRAME_TO_ARGS_OFFSET:
1080 case ISD::EH_DWARF_CFA:
1081 case ISD::EH_SJLJ_SETJMP:
1082 case ISD::EH_SJLJ_LONGJMP:
1083 case ISD::EH_SJLJ_SETUP_DISPATCH:
1084 // These operations lie about being legal: when they claim to be legal,
1085 // they should actually be expanded.
1086 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1087 if (Action == TargetLowering::Legal)
1088 Action = TargetLowering::Expand;
1090 case ISD::INIT_TRAMPOLINE:
1091 case ISD::ADJUST_TRAMPOLINE:
1092 case ISD::FRAMEADDR:
1093 case ISD::RETURNADDR:
1094 case ISD::ADDROFRETURNADDR:
1095 // These operations lie about being legal: when they claim to be legal,
1096 // they should actually be custom-lowered.
1097 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1098 if (Action == TargetLowering::Legal)
1099 Action = TargetLowering::Custom;
1101 case ISD::READCYCLECOUNTER:
1102 // READCYCLECOUNTER returns an i64, even if type legalization might have
1103 // expanded that to several smaller types.
1104 Action = TLI.getOperationAction(Node->getOpcode(), MVT::i64);
1106 case ISD::READ_REGISTER:
1107 case ISD::WRITE_REGISTER:
1108 // Named register is legal in the DAG, but blocked by register name
1109 // selection if not implemented by target (to chose the correct register)
1110 // They'll be converted to Copy(To/From)Reg.
1111 Action = TargetLowering::Legal;
1113 case ISD::DEBUGTRAP:
1114 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1115 if (Action == TargetLowering::Expand) {
1116 // replace ISD::DEBUGTRAP with ISD::TRAP
1118 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1119 Node->getOperand(0));
1120 ReplaceNode(Node, NewVal.getNode());
1121 LegalizeOp(NewVal.getNode());
1125 case ISD::STRICT_FSQRT:
1126 case ISD::STRICT_FMA:
1127 case ISD::STRICT_FPOW:
1128 case ISD::STRICT_FPOWI:
1129 case ISD::STRICT_FSIN:
1130 case ISD::STRICT_FCOS:
1131 case ISD::STRICT_FEXP:
1132 case ISD::STRICT_FEXP2:
1133 case ISD::STRICT_FLOG:
1134 case ISD::STRICT_FLOG10:
1135 case ISD::STRICT_FLOG2:
1136 case ISD::STRICT_FRINT:
1137 case ISD::STRICT_FNEARBYINT:
1138 // These pseudo-ops get legalized as if they were their non-strict
1139 // equivalent. For instance, if ISD::FSQRT is legal then ISD::STRICT_FSQRT
1140 // is also legal, but if ISD::FSQRT requires expansion then so does
1141 // ISD::STRICT_FSQRT.
1142 Action = getStrictFPOpcodeAction(TLI, Node->getOpcode(),
1143 Node->getValueType(0));
1146 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1147 Action = TargetLowering::Legal;
1149 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1154 if (SimpleFinishLegalizing) {
1155 SDNode *NewNode = Node;
1156 switch (Node->getOpcode()) {
1163 // Legalizing shifts/rotates requires adjusting the shift amount
1164 // to the appropriate width.
1165 SDValue Op0 = Node->getOperand(0);
1166 SDValue Op1 = Node->getOperand(1);
1167 if (!Op1.getValueType().isVector()) {
1168 SDValue SAO = DAG.getShiftAmountOperand(Op0.getValueType(), Op1);
1169 // The getShiftAmountOperand() may create a new operand node or
1170 // return the existing one. If new operand is created we need
1171 // to update the parent node.
1172 // Do not try to legalize SAO here! It will be automatically legalized
1173 // in the next round.
1175 NewNode = DAG.UpdateNodeOperands(Node, Op0, SAO);
1179 case ISD::SRL_PARTS:
1180 case ISD::SRA_PARTS:
1181 case ISD::SHL_PARTS: {
1182 // Legalizing shifts/rotates requires adjusting the shift amount
1183 // to the appropriate width.
1184 SDValue Op0 = Node->getOperand(0);
1185 SDValue Op1 = Node->getOperand(1);
1186 SDValue Op2 = Node->getOperand(2);
1187 if (!Op2.getValueType().isVector()) {
1188 SDValue SAO = DAG.getShiftAmountOperand(Op0.getValueType(), Op2);
1189 // The getShiftAmountOperand() may create a new operand node or
1190 // return the existing one. If new operand is created we need
1191 // to update the parent node.
1193 NewNode = DAG.UpdateNodeOperands(Node, Op0, Op1, SAO);
1199 if (NewNode != Node) {
1200 ReplaceNode(Node, NewNode);
1204 case TargetLowering::Legal:
1205 DEBUG(dbgs() << "Legal node: nothing to do\n");
1207 case TargetLowering::Custom:
1208 DEBUG(dbgs() << "Trying custom legalization\n");
1209 // FIXME: The handling for custom lowering with multiple results is
1211 if (SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG)) {
1212 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1215 if (Node->getNumValues() == 1) {
1216 DEBUG(dbgs() << "Successfully custom legalized node\n");
1217 // We can just directly replace this node with the lowered value.
1218 ReplaceNode(SDValue(Node, 0), Res);
1222 SmallVector<SDValue, 8> ResultVals;
1223 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1224 ResultVals.push_back(Res.getValue(i));
1225 DEBUG(dbgs() << "Successfully custom legalized node\n");
1226 ReplaceNode(Node, ResultVals.data());
1229 DEBUG(dbgs() << "Could not custom legalize node\n");
1231 case TargetLowering::Expand:
1232 if (ExpandNode(Node))
1235 case TargetLowering::LibCall:
1236 ConvertNodeToLibcall(Node);
1238 case TargetLowering::Promote:
1244 switch (Node->getOpcode()) {
1251 llvm_unreachable("Do not know how to legalize this operator!");
1253 case ISD::CALLSEQ_START:
1254 case ISD::CALLSEQ_END:
1257 return LegalizeLoadOps(Node);
1259 return LegalizeStoreOps(Node);
1263 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1264 SDValue Vec = Op.getOperand(0);
1265 SDValue Idx = Op.getOperand(1);
1268 // Before we generate a new store to a temporary stack slot, see if there is
1269 // already one that we can use. There often is because when we scalarize
1270 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1271 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1272 // the vector. If all are expanded here, we don't want one store per vector
1275 // Caches for hasPredecessorHelper
1276 SmallPtrSet<const SDNode *, 32> Visited;
1277 SmallVector<const SDNode *, 16> Worklist;
1278 Worklist.push_back(Idx.getNode());
1279 SDValue StackPtr, Ch;
1280 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1281 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1283 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1284 if (ST->isIndexed() || ST->isTruncatingStore() ||
1285 ST->getValue() != Vec)
1288 // Make sure that nothing else could have stored into the destination of
1290 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1293 // If the index is dependent on the store we will introduce a cycle when
1294 // creating the load (the load uses the index, and by replacing the chain
1295 // we will make the index dependent on the load). Also, the store might be
1296 // dependent on the extractelement and introduce a cycle when creating
1298 if (SDNode::hasPredecessorHelper(ST, Visited, Worklist) ||
1299 ST->hasPredecessor(Op.getNode()))
1302 StackPtr = ST->getBasePtr();
1303 Ch = SDValue(ST, 0);
1308 EVT VecVT = Vec.getValueType();
1310 if (!Ch.getNode()) {
1311 // Store the value to a temporary stack slot, then LOAD the returned part.
1312 StackPtr = DAG.CreateStackTemporary(VecVT);
1313 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1314 MachinePointerInfo());
1317 StackPtr = TLI.getVectorElementPointer(DAG, StackPtr, VecVT, Idx);
1321 if (Op.getValueType().isVector())
1323 DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, MachinePointerInfo());
1325 NewLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1326 MachinePointerInfo(),
1327 VecVT.getVectorElementType());
1329 // Replace the chain going out of the store, by the one out of the load.
1330 DAG.ReplaceAllUsesOfValueWith(Ch, SDValue(NewLoad.getNode(), 1));
1332 // We introduced a cycle though, so update the loads operands, making sure
1333 // to use the original store's chain as an incoming chain.
1334 SmallVector<SDValue, 6> NewLoadOperands(NewLoad->op_begin(),
1336 NewLoadOperands[0] = Ch;
1338 SDValue(DAG.UpdateNodeOperands(NewLoad.getNode(), NewLoadOperands), 0);
1342 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1343 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1345 SDValue Vec = Op.getOperand(0);
1346 SDValue Part = Op.getOperand(1);
1347 SDValue Idx = Op.getOperand(2);
1350 // Store the value to a temporary stack slot, then LOAD the returned part.
1351 EVT VecVT = Vec.getValueType();
1352 SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
1353 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1354 MachinePointerInfo PtrInfo =
1355 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
1357 // First store the whole vector.
1358 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo);
1360 // Then store the inserted part.
1361 SDValue SubStackPtr = TLI.getVectorElementPointer(DAG, StackPtr, VecVT, Idx);
1363 // Store the subvector.
1364 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr, MachinePointerInfo());
1366 // Finally, load the updated vector.
1367 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo);
1370 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1371 // We can't handle this case efficiently. Allocate a sufficiently
1372 // aligned object on the stack, store each element into it, then load
1373 // the result as a vector.
1374 // Create the stack frame object.
1375 EVT VT = Node->getValueType(0);
1376 EVT EltVT = VT.getVectorElementType();
1378 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1379 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1380 MachinePointerInfo PtrInfo =
1381 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
1383 // Emit a store of each element to the stack slot.
1384 SmallVector<SDValue, 8> Stores;
1385 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1386 // Store (in the right endianness) the elements to memory.
1387 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1388 // Ignore undef elements.
1389 if (Node->getOperand(i).isUndef()) continue;
1391 unsigned Offset = TypeByteSize*i;
1393 SDValue Idx = DAG.getConstant(Offset, dl, FIPtr.getValueType());
1394 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1396 // If the destination vector element type is narrower than the source
1397 // element type, only store the bits necessary.
1398 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1399 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1400 Node->getOperand(i), Idx,
1401 PtrInfo.getWithOffset(Offset), EltVT));
1403 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl, Node->getOperand(i),
1404 Idx, PtrInfo.getWithOffset(Offset)));
1408 if (!Stores.empty()) // Not all undef elements?
1409 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1411 StoreChain = DAG.getEntryNode();
1413 // Result is a load from the stack slot.
1414 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo);
1417 /// Bitcast a floating-point value to an integer value. Only bitcast the part
1418 /// containing the sign bit if the target has no integer value capable of
1419 /// holding all bits of the floating-point value.
1420 void SelectionDAGLegalize::getSignAsIntValue(FloatSignAsInt &State,
1422 SDValue Value) const {
1423 EVT FloatVT = Value.getValueType();
1424 unsigned NumBits = FloatVT.getSizeInBits();
1425 State.FloatVT = FloatVT;
1426 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), NumBits);
1427 // Convert to an integer of the same size.
1428 if (TLI.isTypeLegal(IVT)) {
1429 State.IntValue = DAG.getNode(ISD::BITCAST, DL, IVT, Value);
1430 State.SignMask = APInt::getSignMask(NumBits);
1431 State.SignBit = NumBits - 1;
1435 auto &DataLayout = DAG.getDataLayout();
1436 // Store the float to memory, then load the sign part out as an integer.
1437 MVT LoadTy = TLI.getRegisterType(*DAG.getContext(), MVT::i8);
1438 // First create a temporary that is aligned for both the load and store.
1439 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1440 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1441 // Then store the float to it.
1442 State.FloatPtr = StackPtr;
1443 MachineFunction &MF = DAG.getMachineFunction();
1444 State.FloatPointerInfo = MachinePointerInfo::getFixedStack(MF, FI);
1445 State.Chain = DAG.getStore(DAG.getEntryNode(), DL, Value, State.FloatPtr,
1446 State.FloatPointerInfo);
1449 if (DataLayout.isBigEndian()) {
1450 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1451 // Load out a legal integer with the same sign bit as the float.
1453 State.IntPointerInfo = State.FloatPointerInfo;
1455 // Advance the pointer so that the loaded byte will contain the sign bit.
1456 unsigned ByteOffset = (FloatVT.getSizeInBits() / 8) - 1;
1457 IntPtr = DAG.getNode(ISD::ADD, DL, StackPtr.getValueType(), StackPtr,
1458 DAG.getConstant(ByteOffset, DL, StackPtr.getValueType()));
1459 State.IntPointerInfo = MachinePointerInfo::getFixedStack(MF, FI,
1463 State.IntPtr = IntPtr;
1464 State.IntValue = DAG.getExtLoad(ISD::EXTLOAD, DL, LoadTy, State.Chain, IntPtr,
1465 State.IntPointerInfo, MVT::i8);
1466 State.SignMask = APInt::getOneBitSet(LoadTy.getSizeInBits(), 7);
1470 /// Replace the integer value produced by getSignAsIntValue() with a new value
1471 /// and cast the result back to a floating-point type.
1472 SDValue SelectionDAGLegalize::modifySignAsInt(const FloatSignAsInt &State,
1474 SDValue NewIntValue) const {
1476 return DAG.getNode(ISD::BITCAST, DL, State.FloatVT, NewIntValue);
1478 // Override the part containing the sign bit in the value stored on the stack.
1479 SDValue Chain = DAG.getTruncStore(State.Chain, DL, NewIntValue, State.IntPtr,
1480 State.IntPointerInfo, MVT::i8);
1481 return DAG.getLoad(State.FloatVT, DL, Chain, State.FloatPtr,
1482 State.FloatPointerInfo);
1485 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode *Node) const {
1487 SDValue Mag = Node->getOperand(0);
1488 SDValue Sign = Node->getOperand(1);
1490 // Get sign bit into an integer value.
1491 FloatSignAsInt SignAsInt;
1492 getSignAsIntValue(SignAsInt, DL, Sign);
1494 EVT IntVT = SignAsInt.IntValue.getValueType();
1495 SDValue SignMask = DAG.getConstant(SignAsInt.SignMask, DL, IntVT);
1496 SDValue SignBit = DAG.getNode(ISD::AND, DL, IntVT, SignAsInt.IntValue,
1499 // If FABS is legal transform FCOPYSIGN(x, y) => sign(x) ? -FABS(x) : FABS(X)
1500 EVT FloatVT = Mag.getValueType();
1501 if (TLI.isOperationLegalOrCustom(ISD::FABS, FloatVT) &&
1502 TLI.isOperationLegalOrCustom(ISD::FNEG, FloatVT)) {
1503 SDValue AbsValue = DAG.getNode(ISD::FABS, DL, FloatVT, Mag);
1504 SDValue NegValue = DAG.getNode(ISD::FNEG, DL, FloatVT, AbsValue);
1505 SDValue Cond = DAG.getSetCC(DL, getSetCCResultType(IntVT), SignBit,
1506 DAG.getConstant(0, DL, IntVT), ISD::SETNE);
1507 return DAG.getSelect(DL, FloatVT, Cond, NegValue, AbsValue);
1510 // Transform Mag value to integer, and clear the sign bit.
1511 FloatSignAsInt MagAsInt;
1512 getSignAsIntValue(MagAsInt, DL, Mag);
1513 EVT MagVT = MagAsInt.IntValue.getValueType();
1514 SDValue ClearSignMask = DAG.getConstant(~MagAsInt.SignMask, DL, MagVT);
1515 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, MagVT, MagAsInt.IntValue,
1518 // Get the signbit at the right position for MagAsInt.
1519 int ShiftAmount = SignAsInt.SignBit - MagAsInt.SignBit;
1520 if (SignBit.getValueSizeInBits() > ClearedSign.getValueSizeInBits()) {
1521 if (ShiftAmount > 0) {
1522 SDValue ShiftCnst = DAG.getConstant(ShiftAmount, DL, IntVT);
1523 SignBit = DAG.getNode(ISD::SRL, DL, IntVT, SignBit, ShiftCnst);
1524 } else if (ShiftAmount < 0) {
1525 SDValue ShiftCnst = DAG.getConstant(-ShiftAmount, DL, IntVT);
1526 SignBit = DAG.getNode(ISD::SHL, DL, IntVT, SignBit, ShiftCnst);
1528 SignBit = DAG.getNode(ISD::TRUNCATE, DL, MagVT, SignBit);
1529 } else if (SignBit.getValueSizeInBits() < ClearedSign.getValueSizeInBits()) {
1530 SignBit = DAG.getNode(ISD::ZERO_EXTEND, DL, MagVT, SignBit);
1531 if (ShiftAmount > 0) {
1532 SDValue ShiftCnst = DAG.getConstant(ShiftAmount, DL, MagVT);
1533 SignBit = DAG.getNode(ISD::SRL, DL, MagVT, SignBit, ShiftCnst);
1534 } else if (ShiftAmount < 0) {
1535 SDValue ShiftCnst = DAG.getConstant(-ShiftAmount, DL, MagVT);
1536 SignBit = DAG.getNode(ISD::SHL, DL, MagVT, SignBit, ShiftCnst);
1540 // Store the part with the modified sign and convert back to float.
1541 SDValue CopiedSign = DAG.getNode(ISD::OR, DL, MagVT, ClearedSign, SignBit);
1542 return modifySignAsInt(MagAsInt, DL, CopiedSign);
1545 SDValue SelectionDAGLegalize::ExpandFABS(SDNode *Node) const {
1547 SDValue Value = Node->getOperand(0);
1549 // Transform FABS(x) => FCOPYSIGN(x, 0.0) if FCOPYSIGN is legal.
1550 EVT FloatVT = Value.getValueType();
1551 if (TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, FloatVT)) {
1552 SDValue Zero = DAG.getConstantFP(0.0, DL, FloatVT);
1553 return DAG.getNode(ISD::FCOPYSIGN, DL, FloatVT, Value, Zero);
1556 // Transform value to integer, clear the sign bit and transform back.
1557 FloatSignAsInt ValueAsInt;
1558 getSignAsIntValue(ValueAsInt, DL, Value);
1559 EVT IntVT = ValueAsInt.IntValue.getValueType();
1560 SDValue ClearSignMask = DAG.getConstant(~ValueAsInt.SignMask, DL, IntVT);
1561 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, ValueAsInt.IntValue,
1563 return modifySignAsInt(ValueAsInt, DL, ClearedSign);
1566 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1567 SmallVectorImpl<SDValue> &Results) {
1568 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1569 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1570 " not tell us which reg is the stack pointer!");
1572 EVT VT = Node->getValueType(0);
1573 SDValue Tmp1 = SDValue(Node, 0);
1574 SDValue Tmp2 = SDValue(Node, 1);
1575 SDValue Tmp3 = Node->getOperand(2);
1576 SDValue Chain = Tmp1.getOperand(0);
1578 // Chain the dynamic stack allocation so that it doesn't modify the stack
1579 // pointer when other instructions are using the stack.
1580 Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl);
1582 SDValue Size = Tmp2.getOperand(1);
1583 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1584 Chain = SP.getValue(1);
1585 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1586 unsigned StackAlign =
1587 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
1588 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1589 if (Align > StackAlign)
1590 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1591 DAG.getConstant(-(uint64_t)Align, dl, VT));
1592 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1594 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
1595 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
1597 Results.push_back(Tmp1);
1598 Results.push_back(Tmp2);
1601 /// Legalize a SETCC with given LHS and RHS and condition code CC on the current
1604 /// If the SETCC has been legalized using AND / OR, then the legalized node
1605 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1606 /// will be set to false.
1608 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1609 /// then the values of LHS and RHS will be swapped, CC will be set to the
1610 /// new condition, and NeedInvert will be set to false.
1612 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1613 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1614 /// will be set to true. The caller must invert the result of the SETCC with
1615 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1616 /// of a true/false result.
1618 /// \returns true if the SetCC has been legalized, false if it hasn't.
1619 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT, SDValue &LHS,
1620 SDValue &RHS, SDValue &CC,
1623 MVT OpVT = LHS.getSimpleValueType();
1624 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1626 bool NeedSwap = false;
1627 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1628 default: llvm_unreachable("Unknown condition code action!");
1629 case TargetLowering::Legal:
1632 case TargetLowering::Expand: {
1633 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1634 if (TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) {
1635 std::swap(LHS, RHS);
1636 CC = DAG.getCondCode(InvCC);
1639 // Swapping operands didn't work. Try inverting the condition.
1640 InvCC = getSetCCInverse(CCCode, OpVT.isInteger());
1641 if (!TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) {
1642 // If inverting the condition is not enough, try swapping operands
1644 InvCC = ISD::getSetCCSwappedOperands(InvCC);
1647 if (TLI.isCondCodeLegalOrCustom(InvCC, OpVT)) {
1648 CC = DAG.getCondCode(InvCC);
1651 std::swap(LHS, RHS);
1655 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1658 default: llvm_unreachable("Don't know how to expand this condition!");
1660 assert(TLI.isCondCodeLegal(ISD::SETOEQ, OpVT)
1661 && "If SETO is expanded, SETOEQ must be legal!");
1662 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1664 assert(TLI.isCondCodeLegal(ISD::SETUNE, OpVT)
1665 && "If SETUO is expanded, SETUNE must be legal!");
1666 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1679 // If we are floating point, assign and break, otherwise fall through.
1680 if (!OpVT.isInteger()) {
1681 // We can use the 4th bit to tell if we are the unordered
1682 // or ordered version of the opcode.
1683 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1684 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1685 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1688 // Fallthrough if we are unsigned integer.
1696 // If all combinations of inverting the condition and swapping operands
1697 // didn't work then we have no means to expand the condition.
1698 llvm_unreachable("Don't know how to expand this condition!");
1701 SDValue SetCC1, SetCC2;
1702 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1703 // If we aren't the ordered or unorder operation,
1704 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1705 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1706 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1708 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1709 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1710 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1712 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1721 /// Emit a store/load combination to the stack. This stores
1722 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1723 /// a load from the stack slot to DestVT, extending it if needed.
1724 /// The resultant code need not be legal.
1725 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, EVT SlotVT,
1726 EVT DestVT, const SDLoc &dl) {
1727 // Create the stack frame object.
1728 unsigned SrcAlign = DAG.getDataLayout().getPrefTypeAlignment(
1729 SrcOp.getValueType().getTypeForEVT(*DAG.getContext()));
1730 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1732 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1733 int SPFI = StackPtrFI->getIndex();
1734 MachinePointerInfo PtrInfo =
1735 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI);
1737 unsigned SrcSize = SrcOp.getValueSizeInBits();
1738 unsigned SlotSize = SlotVT.getSizeInBits();
1739 unsigned DestSize = DestVT.getSizeInBits();
1740 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1741 unsigned DestAlign = DAG.getDataLayout().getPrefTypeAlignment(DestType);
1743 // Emit a store to the stack slot. Use a truncstore if the input value is
1744 // later than DestVT.
1747 if (SrcSize > SlotSize)
1748 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, PtrInfo,
1751 assert(SrcSize == SlotSize && "Invalid store");
1753 DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, PtrInfo, SrcAlign);
1756 // Result is a load from the stack slot.
1757 if (SlotSize == DestSize)
1758 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo, DestAlign);
1760 assert(SlotSize < DestSize && "Unknown extension!");
1761 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr, PtrInfo, SlotVT,
1765 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1767 // Create a vector sized/aligned stack slot, store the value to element #0,
1768 // then load the whole vector back out.
1769 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1771 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1772 int SPFI = StackPtrFI->getIndex();
1774 SDValue Ch = DAG.getTruncStore(
1775 DAG.getEntryNode(), dl, Node->getOperand(0), StackPtr,
1776 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI),
1777 Node->getValueType(0).getVectorElementType());
1779 Node->getValueType(0), dl, Ch, StackPtr,
1780 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI));
1784 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1785 const TargetLowering &TLI, SDValue &Res) {
1786 unsigned NumElems = Node->getNumOperands();
1788 EVT VT = Node->getValueType(0);
1790 // Try to group the scalars into pairs, shuffle the pairs together, then
1791 // shuffle the pairs of pairs together, etc. until the vector has
1792 // been built. This will work only if all of the necessary shuffle masks
1795 // We do this in two phases; first to check the legality of the shuffles,
1796 // and next, assuming that all shuffles are legal, to create the new nodes.
1797 for (int Phase = 0; Phase < 2; ++Phase) {
1798 SmallVector<std::pair<SDValue, SmallVector<int, 16>>, 16> IntermedVals,
1800 for (unsigned i = 0; i < NumElems; ++i) {
1801 SDValue V = Node->getOperand(i);
1807 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1808 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1811 while (IntermedVals.size() > 2) {
1812 NewIntermedVals.clear();
1813 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1814 // This vector and the next vector are shuffled together (simply to
1815 // append the one to the other).
1816 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1818 SmallVector<int, 16> FinalIndices;
1819 FinalIndices.reserve(IntermedVals[i].second.size() +
1820 IntermedVals[i+1].second.size());
1823 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1826 FinalIndices.push_back(IntermedVals[i].second[j]);
1828 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1830 ShuffleVec[k] = NumElems + j;
1831 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1836 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1837 IntermedVals[i+1].first,
1839 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1841 NewIntermedVals.push_back(
1842 std::make_pair(Shuffle, std::move(FinalIndices)));
1845 // If we had an odd number of defined values, then append the last
1846 // element to the array of new vectors.
1847 if ((IntermedVals.size() & 1) != 0)
1848 NewIntermedVals.push_back(IntermedVals.back());
1850 IntermedVals.swap(NewIntermedVals);
1853 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1854 "Invalid number of intermediate vectors");
1855 SDValue Vec1 = IntermedVals[0].first;
1857 if (IntermedVals.size() > 1)
1858 Vec2 = IntermedVals[1].first;
1860 Vec2 = DAG.getUNDEF(VT);
1862 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1863 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1864 ShuffleVec[IntermedVals[0].second[i]] = i;
1865 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1866 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1869 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec);
1870 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1877 /// Expand a BUILD_VECTOR node on targets that don't
1878 /// support the operation, but do support the resultant vector type.
1879 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1880 unsigned NumElems = Node->getNumOperands();
1881 SDValue Value1, Value2;
1883 EVT VT = Node->getValueType(0);
1884 EVT OpVT = Node->getOperand(0).getValueType();
1885 EVT EltVT = VT.getVectorElementType();
1887 // If the only non-undef value is the low element, turn this into a
1888 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1889 bool isOnlyLowElement = true;
1890 bool MoreThanTwoValues = false;
1891 bool isConstant = true;
1892 for (unsigned i = 0; i < NumElems; ++i) {
1893 SDValue V = Node->getOperand(i);
1897 isOnlyLowElement = false;
1898 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1901 if (!Value1.getNode()) {
1903 } else if (!Value2.getNode()) {
1906 } else if (V != Value1 && V != Value2) {
1907 MoreThanTwoValues = true;
1911 if (!Value1.getNode())
1912 return DAG.getUNDEF(VT);
1914 if (isOnlyLowElement)
1915 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1917 // If all elements are constants, create a load from the constant pool.
1919 SmallVector<Constant*, 16> CV;
1920 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1921 if (ConstantFPSDNode *V =
1922 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1923 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1924 } else if (ConstantSDNode *V =
1925 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1927 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1929 // If OpVT and EltVT don't match, EltVT is not legal and the
1930 // element values have been promoted/truncated earlier. Undo this;
1931 // we don't want a v16i8 to become a v16i32 for example.
1932 const ConstantInt *CI = V->getConstantIntValue();
1933 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1934 CI->getZExtValue()));
1937 assert(Node->getOperand(i).isUndef());
1938 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1939 CV.push_back(UndefValue::get(OpNTy));
1942 Constant *CP = ConstantVector::get(CV);
1944 DAG.getConstantPool(CP, TLI.getPointerTy(DAG.getDataLayout()));
1945 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1947 VT, dl, DAG.getEntryNode(), CPIdx,
1948 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
1952 SmallSet<SDValue, 16> DefinedValues;
1953 for (unsigned i = 0; i < NumElems; ++i) {
1954 if (Node->getOperand(i).isUndef())
1956 DefinedValues.insert(Node->getOperand(i));
1959 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
1960 if (!MoreThanTwoValues) {
1961 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1962 for (unsigned i = 0; i < NumElems; ++i) {
1963 SDValue V = Node->getOperand(i);
1966 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1968 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1969 // Get the splatted value into the low element of a vector register.
1970 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1972 if (Value2.getNode())
1973 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1975 Vec2 = DAG.getUNDEF(VT);
1977 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1978 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec);
1982 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
1987 // Otherwise, we can't handle this case efficiently.
1988 return ExpandVectorBuildThroughStack(Node);
1991 // Expand a node into a call to a libcall. If the result value
1992 // does not fit into a register, return the lo part and set the hi part to the
1993 // by-reg argument. If it does fit into a single register, return the result
1994 // and leave the Hi part unset.
1995 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
1997 TargetLowering::ArgListTy Args;
1998 TargetLowering::ArgListEntry Entry;
1999 for (const SDValue &Op : Node->op_values()) {
2000 EVT ArgVT = Op.getValueType();
2001 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2004 Entry.IsSExt = TLI.shouldSignExtendTypeInLibCall(ArgVT, isSigned);
2005 Entry.IsZExt = !TLI.shouldSignExtendTypeInLibCall(ArgVT, isSigned);
2006 Args.push_back(Entry);
2008 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2009 TLI.getPointerTy(DAG.getDataLayout()));
2011 EVT RetVT = Node->getValueType(0);
2012 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2014 // By default, the input chain to this libcall is the entry node of the
2015 // function. If the libcall is going to be emitted as a tail call then
2016 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2017 // node which is being folded has a non-entry input chain.
2018 SDValue InChain = DAG.getEntryNode();
2020 // isTailCall may be true since the callee does not reference caller stack
2021 // frame. Check if it's in the right position and that the return types match.
2022 SDValue TCChain = InChain;
2023 const Function &F = DAG.getMachineFunction().getFunction();
2025 TLI.isInTailCallPosition(DAG, Node, TCChain) &&
2026 (RetTy == F.getReturnType() || F.getReturnType()->isVoidTy());
2030 TargetLowering::CallLoweringInfo CLI(DAG);
2031 bool signExtend = TLI.shouldSignExtendTypeInLibCall(RetVT, isSigned);
2032 CLI.setDebugLoc(SDLoc(Node))
2034 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2036 .setTailCall(isTailCall)
2037 .setSExtResult(signExtend)
2038 .setZExtResult(!signExtend)
2039 .setIsPostTypeLegalization(true);
2041 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2043 if (!CallInfo.second.getNode()) {
2044 DEBUG(dbgs() << "Created tailcall: "; DAG.getRoot().dump());
2045 // It's a tailcall, return the chain (which is the DAG root).
2046 return DAG.getRoot();
2049 DEBUG(dbgs() << "Created libcall: "; CallInfo.first.dump());
2050 return CallInfo.first;
2053 /// Generate a libcall taking the given operands as arguments
2054 /// and returning a result of type RetVT.
2055 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2056 const SDValue *Ops, unsigned NumOps,
2057 bool isSigned, const SDLoc &dl) {
2058 TargetLowering::ArgListTy Args;
2059 Args.reserve(NumOps);
2061 TargetLowering::ArgListEntry Entry;
2062 for (unsigned i = 0; i != NumOps; ++i) {
2063 Entry.Node = Ops[i];
2064 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2065 Entry.IsSExt = isSigned;
2066 Entry.IsZExt = !isSigned;
2067 Args.push_back(Entry);
2069 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2070 TLI.getPointerTy(DAG.getDataLayout()));
2072 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2074 TargetLowering::CallLoweringInfo CLI(DAG);
2076 .setChain(DAG.getEntryNode())
2077 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2079 .setSExtResult(isSigned)
2080 .setZExtResult(!isSigned)
2081 .setIsPostTypeLegalization(true);
2083 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2085 return CallInfo.first;
2088 // Expand a node into a call to a libcall. Similar to
2089 // ExpandLibCall except that the first operand is the in-chain.
2090 std::pair<SDValue, SDValue>
2091 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2094 SDValue InChain = Node->getOperand(0);
2096 TargetLowering::ArgListTy Args;
2097 TargetLowering::ArgListEntry Entry;
2098 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2099 EVT ArgVT = Node->getOperand(i).getValueType();
2100 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2101 Entry.Node = Node->getOperand(i);
2103 Entry.IsSExt = isSigned;
2104 Entry.IsZExt = !isSigned;
2105 Args.push_back(Entry);
2107 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2108 TLI.getPointerTy(DAG.getDataLayout()));
2110 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2112 TargetLowering::CallLoweringInfo CLI(DAG);
2113 CLI.setDebugLoc(SDLoc(Node))
2115 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2117 .setSExtResult(isSigned)
2118 .setZExtResult(!isSigned);
2120 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2125 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2126 RTLIB::Libcall Call_F32,
2127 RTLIB::Libcall Call_F64,
2128 RTLIB::Libcall Call_F80,
2129 RTLIB::Libcall Call_F128,
2130 RTLIB::Libcall Call_PPCF128) {
2131 if (Node->isStrictFPOpcode())
2132 Node = DAG.mutateStrictFPToFP(Node);
2135 switch (Node->getSimpleValueType(0).SimpleTy) {
2136 default: llvm_unreachable("Unexpected request for libcall!");
2137 case MVT::f32: LC = Call_F32; break;
2138 case MVT::f64: LC = Call_F64; break;
2139 case MVT::f80: LC = Call_F80; break;
2140 case MVT::f128: LC = Call_F128; break;
2141 case MVT::ppcf128: LC = Call_PPCF128; break;
2143 return ExpandLibCall(LC, Node, false);
2146 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2147 RTLIB::Libcall Call_I8,
2148 RTLIB::Libcall Call_I16,
2149 RTLIB::Libcall Call_I32,
2150 RTLIB::Libcall Call_I64,
2151 RTLIB::Libcall Call_I128) {
2153 switch (Node->getSimpleValueType(0).SimpleTy) {
2154 default: llvm_unreachable("Unexpected request for libcall!");
2155 case MVT::i8: LC = Call_I8; break;
2156 case MVT::i16: LC = Call_I16; break;
2157 case MVT::i32: LC = Call_I32; break;
2158 case MVT::i64: LC = Call_I64; break;
2159 case MVT::i128: LC = Call_I128; break;
2161 return ExpandLibCall(LC, Node, isSigned);
2164 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
2166 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2167 SmallVectorImpl<SDValue> &Results) {
2168 unsigned Opcode = Node->getOpcode();
2169 bool isSigned = Opcode == ISD::SDIVREM;
2172 switch (Node->getSimpleValueType(0).SimpleTy) {
2173 default: llvm_unreachable("Unexpected request for libcall!");
2174 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2175 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2176 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2177 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2178 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2181 // The input chain to this libcall is the entry node of the function.
2182 // Legalizing the call will automatically add the previous call to the
2184 SDValue InChain = DAG.getEntryNode();
2186 EVT RetVT = Node->getValueType(0);
2187 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2189 TargetLowering::ArgListTy Args;
2190 TargetLowering::ArgListEntry Entry;
2191 for (const SDValue &Op : Node->op_values()) {
2192 EVT ArgVT = Op.getValueType();
2193 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2196 Entry.IsSExt = isSigned;
2197 Entry.IsZExt = !isSigned;
2198 Args.push_back(Entry);
2201 // Also pass the return address of the remainder.
2202 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2204 Entry.Ty = RetTy->getPointerTo();
2205 Entry.IsSExt = isSigned;
2206 Entry.IsZExt = !isSigned;
2207 Args.push_back(Entry);
2209 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2210 TLI.getPointerTy(DAG.getDataLayout()));
2213 TargetLowering::CallLoweringInfo CLI(DAG);
2216 .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
2218 .setSExtResult(isSigned)
2219 .setZExtResult(!isSigned);
2221 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2223 // Remainder is loaded back from the stack frame.
2225 DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr, MachinePointerInfo());
2226 Results.push_back(CallInfo.first);
2227 Results.push_back(Rem);
2230 /// Return true if sincos libcall is available.
2231 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2233 switch (Node->getSimpleValueType(0).SimpleTy) {
2234 default: llvm_unreachable("Unexpected request for libcall!");
2235 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2236 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2237 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2238 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2239 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2241 return TLI.getLibcallName(LC) != nullptr;
2244 /// Only issue sincos libcall if both sin and cos are needed.
2245 static bool useSinCos(SDNode *Node) {
2246 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2247 ? ISD::FCOS : ISD::FSIN;
2249 SDValue Op0 = Node->getOperand(0);
2250 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2251 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2255 // The other user might have been turned into sincos already.
2256 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2262 /// Issue libcalls to sincos to compute sin / cos pairs.
2264 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2265 SmallVectorImpl<SDValue> &Results) {
2267 switch (Node->getSimpleValueType(0).SimpleTy) {
2268 default: llvm_unreachable("Unexpected request for libcall!");
2269 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2270 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2271 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2272 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2273 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2276 // The input chain to this libcall is the entry node of the function.
2277 // Legalizing the call will automatically add the previous call to the
2279 SDValue InChain = DAG.getEntryNode();
2281 EVT RetVT = Node->getValueType(0);
2282 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2284 TargetLowering::ArgListTy Args;
2285 TargetLowering::ArgListEntry Entry;
2287 // Pass the argument.
2288 Entry.Node = Node->getOperand(0);
2290 Entry.IsSExt = false;
2291 Entry.IsZExt = false;
2292 Args.push_back(Entry);
2294 // Pass the return address of sin.
2295 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2296 Entry.Node = SinPtr;
2297 Entry.Ty = RetTy->getPointerTo();
2298 Entry.IsSExt = false;
2299 Entry.IsZExt = false;
2300 Args.push_back(Entry);
2302 // Also pass the return address of the cos.
2303 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2304 Entry.Node = CosPtr;
2305 Entry.Ty = RetTy->getPointerTo();
2306 Entry.IsSExt = false;
2307 Entry.IsZExt = false;
2308 Args.push_back(Entry);
2310 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2311 TLI.getPointerTy(DAG.getDataLayout()));
2314 TargetLowering::CallLoweringInfo CLI(DAG);
2315 CLI.setDebugLoc(dl).setChain(InChain).setLibCallee(
2316 TLI.getLibcallCallingConv(LC), Type::getVoidTy(*DAG.getContext()), Callee,
2319 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2322 DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr, MachinePointerInfo()));
2324 DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr, MachinePointerInfo()));
2327 /// This function is responsible for legalizing a
2328 /// INT_TO_FP operation of the specified operand when the target requests that
2329 /// we expand it. At this point, we know that the result and operand types are
2330 /// legal for the target.
2331 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned, SDValue Op0,
2334 // TODO: Should any fast-math-flags be set for the created nodes?
2335 DEBUG(dbgs() << "Legalizing INT_TO_FP\n");
2336 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2337 DEBUG(dbgs() << "32-bit [signed|unsigned] integer to float/double "
2340 // Get the stack frame index of a 8 byte buffer.
2341 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2343 // word offset constant for Hi/Lo address computation
2344 SDValue WordOff = DAG.getConstant(sizeof(int), dl,
2345 StackSlot.getValueType());
2346 // set up Hi and Lo (into buffer) address based on endian
2347 SDValue Hi = StackSlot;
2348 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2349 StackSlot, WordOff);
2350 if (DAG.getDataLayout().isLittleEndian())
2353 // if signed map to unsigned space
2356 // constant used to invert sign bit (signed to unsigned mapping)
2357 SDValue SignBit = DAG.getConstant(0x80000000u, dl, MVT::i32);
2358 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2362 // store the lo of the constructed double - based on integer input
2363 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op0Mapped, Lo,
2364 MachinePointerInfo());
2365 // initial hi portion of constructed double
2366 SDValue InitialHi = DAG.getConstant(0x43300000u, dl, MVT::i32);
2367 // store the hi of the constructed double - biased exponent
2369 DAG.getStore(Store1, dl, InitialHi, Hi, MachinePointerInfo());
2370 // load the constructed double
2372 DAG.getLoad(MVT::f64, dl, Store2, StackSlot, MachinePointerInfo());
2373 // FP constant to bias correct the final result
2374 SDValue Bias = DAG.getConstantFP(isSigned ?
2375 BitsToDouble(0x4330000080000000ULL) :
2376 BitsToDouble(0x4330000000000000ULL),
2378 // subtract the bias
2379 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2382 // handle final rounding
2383 if (DestVT == MVT::f64) {
2386 } else if (DestVT.bitsLT(MVT::f64)) {
2387 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2388 DAG.getIntPtrConstant(0, dl));
2389 } else if (DestVT.bitsGT(MVT::f64)) {
2390 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2394 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2395 // Code below here assumes !isSigned without checking again.
2397 // Implementation of unsigned i64 to f64 following the algorithm in
2398 // __floatundidf in compiler_rt. This implementation has the advantage
2399 // of performing rounding correctly, both in the default rounding mode
2400 // and in all alternate rounding modes.
2401 // TODO: Generalize this for use with other types.
2402 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2403 DEBUG(dbgs() << "Converting unsigned i64 to f64\n");
2405 DAG.getConstant(UINT64_C(0x4330000000000000), dl, MVT::i64);
2406 SDValue TwoP84PlusTwoP52 =
2407 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), dl,
2410 DAG.getConstant(UINT64_C(0x4530000000000000), dl, MVT::i64);
2412 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2413 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2414 DAG.getConstant(32, dl, MVT::i64));
2415 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2416 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2417 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2418 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2419 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2421 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2424 // TODO: Generalize this for use with other types.
2425 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2426 DEBUG(dbgs() << "Converting unsigned i64 to f32\n");
2427 // For unsigned conversions, convert them to signed conversions using the
2428 // algorithm from the x86_64 __floatundidf in compiler_rt.
2430 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2432 SDValue ShiftConst = DAG.getConstant(
2433 1, dl, TLI.getShiftAmountTy(Op0.getValueType(), DAG.getDataLayout()));
2434 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2435 SDValue AndConst = DAG.getConstant(1, dl, MVT::i64);
2436 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2437 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2439 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2440 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2442 // TODO: This really should be implemented using a branch rather than a
2443 // select. We happen to get lucky and machinesink does the right
2444 // thing most of the time. This would be a good candidate for a
2445 //pseudo-op, or, even better, for whole-function isel.
2446 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2447 Op0, DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
2448 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2451 // Otherwise, implement the fully general conversion.
2453 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2454 DAG.getConstant(UINT64_C(0xfffffffffffff800), dl, MVT::i64));
2455 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2456 DAG.getConstant(UINT64_C(0x800), dl, MVT::i64));
2457 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2458 DAG.getConstant(UINT64_C(0x7ff), dl, MVT::i64));
2459 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), And2,
2460 DAG.getConstant(UINT64_C(0), dl, MVT::i64),
2462 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2463 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), Op0,
2464 DAG.getConstant(UINT64_C(0x0020000000000000), dl,
2467 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2468 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType(), DAG.getDataLayout());
2470 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2471 DAG.getConstant(32, dl, SHVT));
2472 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2473 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2475 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), dl,
2477 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2478 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2479 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2480 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2481 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2482 DAG.getIntPtrConstant(0, dl));
2485 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2487 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2489 DAG.getConstant(0, dl, Op0.getValueType()),
2491 SDValue Zero = DAG.getIntPtrConstant(0, dl),
2492 Four = DAG.getIntPtrConstant(4, dl);
2493 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2494 SignSet, Four, Zero);
2496 // If the sign bit of the integer is set, the large number will be treated
2497 // as a negative number. To counteract this, the dynamic code adds an
2498 // offset depending on the data type.
2500 switch (Op0.getSimpleValueType().SimpleTy) {
2501 default: llvm_unreachable("Unsupported integer type!");
2502 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2503 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2504 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2505 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2507 if (DAG.getDataLayout().isLittleEndian())
2509 Constant *FudgeFactor = ConstantInt::get(
2510 Type::getInt64Ty(*DAG.getContext()), FF);
2513 DAG.getConstantPool(FudgeFactor, TLI.getPointerTy(DAG.getDataLayout()));
2514 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2515 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2516 Alignment = std::min(Alignment, 4u);
2518 if (DestVT == MVT::f32)
2519 FudgeInReg = DAG.getLoad(
2520 MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2521 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2524 SDValue Load = DAG.getExtLoad(
2525 ISD::EXTLOAD, dl, DestVT, DAG.getEntryNode(), CPIdx,
2526 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
2528 HandleSDNode Handle(Load);
2529 LegalizeOp(Load.getNode());
2530 FudgeInReg = Handle.getValue();
2533 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2536 /// This function is responsible for legalizing a
2537 /// *INT_TO_FP operation of the specified operand when the target requests that
2538 /// we promote it. At this point, we know that the result and operand types are
2539 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2540 /// operation that takes a larger input.
2541 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT,
2544 // First step, figure out the appropriate *INT_TO_FP operation to use.
2545 EVT NewInTy = LegalOp.getValueType();
2547 unsigned OpToUse = 0;
2549 // Scan for the appropriate larger type to use.
2551 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2552 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2554 // If the target supports SINT_TO_FP of this type, use it.
2555 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2556 OpToUse = ISD::SINT_TO_FP;
2559 if (isSigned) continue;
2561 // If the target supports UINT_TO_FP of this type, use it.
2562 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2563 OpToUse = ISD::UINT_TO_FP;
2567 // Otherwise, try a larger type.
2570 // Okay, we found the operation and type to use. Zero extend our input to the
2571 // desired type then run the operation on it.
2572 return DAG.getNode(OpToUse, dl, DestVT,
2573 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2574 dl, NewInTy, LegalOp));
2577 /// This function is responsible for legalizing a
2578 /// FP_TO_*INT operation of the specified operand when the target requests that
2579 /// we promote it. At this point, we know that the result and operand types are
2580 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2581 /// operation that returns a larger result.
2582 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT,
2585 // First step, figure out the appropriate FP_TO*INT operation to use.
2586 EVT NewOutTy = DestVT;
2588 unsigned OpToUse = 0;
2590 // Scan for the appropriate larger type to use.
2592 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2593 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2595 // A larger signed type can hold all unsigned values of the requested type,
2596 // so using FP_TO_SINT is valid
2597 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2598 OpToUse = ISD::FP_TO_SINT;
2602 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2603 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2604 OpToUse = ISD::FP_TO_UINT;
2608 // Otherwise, try a larger type.
2611 // Okay, we found the operation and type to use.
2612 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2614 // Truncate the result of the extended FP_TO_*INT operation to the desired
2616 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2619 /// Legalize a BITREVERSE scalar/vector operation as a series of mask + shifts.
2620 SDValue SelectionDAGLegalize::ExpandBITREVERSE(SDValue Op, const SDLoc &dl) {
2621 EVT VT = Op.getValueType();
2622 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2623 unsigned Sz = VT.getScalarSizeInBits();
2625 SDValue Tmp, Tmp2, Tmp3;
2627 // If we can, perform BSWAP first and then the mask+swap the i4, then i2
2628 // and finally the i1 pairs.
2629 // TODO: We can easily support i4/i2 legal types if any target ever does.
2630 if (Sz >= 8 && isPowerOf2_32(Sz)) {
2631 // Create the masks - repeating the pattern every byte.
2632 APInt MaskHi4(Sz, 0), MaskHi2(Sz, 0), MaskHi1(Sz, 0);
2633 APInt MaskLo4(Sz, 0), MaskLo2(Sz, 0), MaskLo1(Sz, 0);
2634 for (unsigned J = 0; J != Sz; J += 8) {
2635 MaskHi4 = MaskHi4 | (0xF0ull << J);
2636 MaskLo4 = MaskLo4 | (0x0Full << J);
2637 MaskHi2 = MaskHi2 | (0xCCull << J);
2638 MaskLo2 = MaskLo2 | (0x33ull << J);
2639 MaskHi1 = MaskHi1 | (0xAAull << J);
2640 MaskLo1 = MaskLo1 | (0x55ull << J);
2643 // BSWAP if the type is wider than a single byte.
2644 Tmp = (Sz > 8 ? DAG.getNode(ISD::BSWAP, dl, VT, Op) : Op);
2646 // swap i4: ((V & 0xF0) >> 4) | ((V & 0x0F) << 4)
2647 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskHi4, dl, VT));
2648 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskLo4, dl, VT));
2649 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp2, DAG.getConstant(4, dl, VT));
2650 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(4, dl, VT));
2651 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
2653 // swap i2: ((V & 0xCC) >> 2) | ((V & 0x33) << 2)
2654 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskHi2, dl, VT));
2655 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskLo2, dl, VT));
2656 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp2, DAG.getConstant(2, dl, VT));
2657 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(2, dl, VT));
2658 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
2660 // swap i1: ((V & 0xAA) >> 1) | ((V & 0x55) << 1)
2661 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskHi1, dl, VT));
2662 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp, DAG.getConstant(MaskLo1, dl, VT));
2663 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Tmp2, DAG.getConstant(1, dl, VT));
2664 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Tmp3, DAG.getConstant(1, dl, VT));
2665 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
2669 Tmp = DAG.getConstant(0, dl, VT);
2670 for (unsigned I = 0, J = Sz-1; I < Sz; ++I, --J) {
2673 DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(J - I, dl, SHVT));
2676 DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(I - J, dl, SHVT));
2680 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Shift, dl, VT));
2681 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp, Tmp2);
2687 /// Open code the operations for BSWAP of the specified operation.
2688 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, const SDLoc &dl) {
2689 EVT VT = Op.getValueType();
2690 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2691 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2692 switch (VT.getSimpleVT().getScalarType().SimpleTy) {
2693 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2695 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2696 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2697 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2699 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2700 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2701 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2702 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2703 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2704 DAG.getConstant(0xFF0000, dl, VT));
2705 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT));
2706 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2707 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2708 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2710 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2711 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2712 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2713 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2714 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2715 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2716 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2717 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2718 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7,
2719 DAG.getConstant(255ULL<<48, dl, VT));
2720 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6,
2721 DAG.getConstant(255ULL<<40, dl, VT));
2722 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5,
2723 DAG.getConstant(255ULL<<32, dl, VT));
2724 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4,
2725 DAG.getConstant(255ULL<<24, dl, VT));
2726 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2727 DAG.getConstant(255ULL<<16, dl, VT));
2728 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2,
2729 DAG.getConstant(255ULL<<8 , dl, VT));
2730 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2731 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2732 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2733 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2734 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2735 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2736 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2740 /// Expand the specified bitcount instruction into operations.
2741 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2744 default: llvm_unreachable("Cannot expand this yet!");
2746 EVT VT = Op.getValueType();
2747 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2748 unsigned Len = VT.getSizeInBits();
2750 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2751 "CTPOP not implemented for this type.");
2753 // This is the "best" algorithm from
2754 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2756 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)),
2758 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)),
2760 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)),
2762 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)),
2765 // v = v - ((v >> 1) & 0x55555555...)
2766 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2767 DAG.getNode(ISD::AND, dl, VT,
2768 DAG.getNode(ISD::SRL, dl, VT, Op,
2769 DAG.getConstant(1, dl, ShVT)),
2771 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2772 Op = DAG.getNode(ISD::ADD, dl, VT,
2773 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2774 DAG.getNode(ISD::AND, dl, VT,
2775 DAG.getNode(ISD::SRL, dl, VT, Op,
2776 DAG.getConstant(2, dl, ShVT)),
2778 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2779 Op = DAG.getNode(ISD::AND, dl, VT,
2780 DAG.getNode(ISD::ADD, dl, VT, Op,
2781 DAG.getNode(ISD::SRL, dl, VT, Op,
2782 DAG.getConstant(4, dl, ShVT))),
2784 // v = (v * 0x01010101...) >> (Len - 8)
2785 Op = DAG.getNode(ISD::SRL, dl, VT,
2786 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2787 DAG.getConstant(Len - 8, dl, ShVT));
2791 case ISD::CTLZ_ZERO_UNDEF:
2792 // This trivially expands to CTLZ.
2793 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2795 EVT VT = Op.getValueType();
2796 unsigned Len = VT.getSizeInBits();
2798 if (TLI.isOperationLegalOrCustom(ISD::CTLZ_ZERO_UNDEF, VT)) {
2799 EVT SetCCVT = getSetCCResultType(VT);
2800 SDValue CTLZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, dl, VT, Op);
2801 SDValue Zero = DAG.getConstant(0, dl, VT);
2802 SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ);
2803 return DAG.getNode(ISD::SELECT, dl, VT, SrcIsZero,
2804 DAG.getConstant(Len, dl, VT), CTLZ);
2807 // for now, we do this:
2808 // x = x | (x >> 1);
2809 // x = x | (x >> 2);
2811 // x = x | (x >>16);
2812 // x = x | (x >>32); // for 64-bit input
2813 // return popcount(~x);
2815 // Ref: "Hacker's Delight" by Henry Warren
2816 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2817 for (unsigned i = 0; (1U << i) <= (Len / 2); ++i) {
2818 SDValue Tmp3 = DAG.getConstant(1ULL << i, dl, ShVT);
2819 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2820 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2822 Op = DAG.getNOT(dl, Op, VT);
2823 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2825 case ISD::CTTZ_ZERO_UNDEF:
2826 // This trivially expands to CTTZ.
2827 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2829 EVT VT = Op.getValueType();
2830 unsigned Len = VT.getSizeInBits();
2832 if (TLI.isOperationLegalOrCustom(ISD::CTTZ_ZERO_UNDEF, VT)) {
2833 EVT SetCCVT = getSetCCResultType(VT);
2834 SDValue CTTZ = DAG.getNode(ISD::CTTZ_ZERO_UNDEF, dl, VT, Op);
2835 SDValue Zero = DAG.getConstant(0, dl, VT);
2836 SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ);
2837 return DAG.getNode(ISD::SELECT, dl, VT, SrcIsZero,
2838 DAG.getConstant(Len, dl, VT), CTTZ);
2841 // for now, we use: { return popcount(~x & (x - 1)); }
2842 // unless the target has ctlz but not ctpop, in which case we use:
2843 // { return 32 - nlz(~x & (x-1)); }
2844 // Ref: "Hacker's Delight" by Henry Warren
2845 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2846 DAG.getNOT(dl, Op, VT),
2847 DAG.getNode(ISD::SUB, dl, VT, Op,
2848 DAG.getConstant(1, dl, VT)));
2849 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2850 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2851 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2852 return DAG.getNode(ISD::SUB, dl, VT,
2853 DAG.getConstant(VT.getSizeInBits(), dl, VT),
2854 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2855 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2860 bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2861 DEBUG(dbgs() << "Trying to expand node\n");
2862 SmallVector<SDValue, 8> Results;
2864 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2866 switch (Node->getOpcode()) {
2869 case ISD::CTLZ_ZERO_UNDEF:
2871 case ISD::CTTZ_ZERO_UNDEF:
2872 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2873 Results.push_back(Tmp1);
2875 case ISD::BITREVERSE:
2876 Results.push_back(ExpandBITREVERSE(Node->getOperand(0), dl));
2879 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2881 case ISD::FRAMEADDR:
2882 case ISD::RETURNADDR:
2883 case ISD::FRAME_TO_ARGS_OFFSET:
2884 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0)));
2886 case ISD::EH_DWARF_CFA: {
2887 SDValue CfaArg = DAG.getSExtOrTrunc(Node->getOperand(0), dl,
2888 TLI.getPointerTy(DAG.getDataLayout()));
2889 SDValue Offset = DAG.getNode(ISD::ADD, dl,
2890 CfaArg.getValueType(),
2891 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
2892 CfaArg.getValueType()),
2894 SDValue FA = DAG.getNode(
2895 ISD::FRAMEADDR, dl, TLI.getPointerTy(DAG.getDataLayout()),
2896 DAG.getConstant(0, dl, TLI.getPointerTy(DAG.getDataLayout())));
2897 Results.push_back(DAG.getNode(ISD::ADD, dl, FA.getValueType(),
2901 case ISD::FLT_ROUNDS_:
2902 Results.push_back(DAG.getConstant(1, dl, Node->getValueType(0)));
2904 case ISD::EH_RETURN:
2908 case ISD::EH_SJLJ_LONGJMP:
2909 // If the target didn't expand these, there's nothing to do, so just
2910 // preserve the chain and be done.
2911 Results.push_back(Node->getOperand(0));
2913 case ISD::READCYCLECOUNTER:
2914 // If the target didn't expand this, just return 'zero' and preserve the
2916 Results.append(Node->getNumValues() - 1,
2917 DAG.getConstant(0, dl, Node->getValueType(0)));
2918 Results.push_back(Node->getOperand(0));
2920 case ISD::EH_SJLJ_SETJMP:
2921 // If the target didn't expand this, just return 'zero' and preserve the
2923 Results.push_back(DAG.getConstant(0, dl, MVT::i32));
2924 Results.push_back(Node->getOperand(0));
2926 case ISD::ATOMIC_LOAD: {
2927 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
2928 SDValue Zero = DAG.getConstant(0, dl, Node->getValueType(0));
2929 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
2930 SDValue Swap = DAG.getAtomicCmpSwap(
2931 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
2932 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
2933 cast<AtomicSDNode>(Node)->getMemOperand());
2934 Results.push_back(Swap.getValue(0));
2935 Results.push_back(Swap.getValue(1));
2938 case ISD::ATOMIC_STORE: {
2939 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
2940 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
2941 cast<AtomicSDNode>(Node)->getMemoryVT(),
2942 Node->getOperand(0),
2943 Node->getOperand(1), Node->getOperand(2),
2944 cast<AtomicSDNode>(Node)->getMemOperand());
2945 Results.push_back(Swap.getValue(1));
2948 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
2949 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
2950 // splits out the success value as a comparison. Expanding the resulting
2951 // ATOMIC_CMP_SWAP will produce a libcall.
2952 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
2953 SDValue Res = DAG.getAtomicCmpSwap(
2954 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
2955 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
2956 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand());
2958 SDValue ExtRes = Res;
2960 SDValue RHS = Node->getOperand(1);
2962 EVT AtomicType = cast<AtomicSDNode>(Node)->getMemoryVT();
2963 EVT OuterType = Node->getValueType(0);
2964 switch (TLI.getExtendForAtomicOps()) {
2965 case ISD::SIGN_EXTEND:
2966 LHS = DAG.getNode(ISD::AssertSext, dl, OuterType, Res,
2967 DAG.getValueType(AtomicType));
2968 RHS = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, OuterType,
2969 Node->getOperand(2), DAG.getValueType(AtomicType));
2972 case ISD::ZERO_EXTEND:
2973 LHS = DAG.getNode(ISD::AssertZext, dl, OuterType, Res,
2974 DAG.getValueType(AtomicType));
2975 RHS = DAG.getZeroExtendInReg(Node->getOperand(2), dl, AtomicType);
2978 case ISD::ANY_EXTEND:
2979 LHS = DAG.getZeroExtendInReg(Res, dl, AtomicType);
2980 RHS = DAG.getZeroExtendInReg(Node->getOperand(2), dl, AtomicType);
2983 llvm_unreachable("Invalid atomic op extension");
2987 DAG.getSetCC(dl, Node->getValueType(1), LHS, RHS, ISD::SETEQ);
2989 Results.push_back(ExtRes.getValue(0));
2990 Results.push_back(Success);
2991 Results.push_back(Res.getValue(1));
2994 case ISD::DYNAMIC_STACKALLOC:
2995 ExpandDYNAMIC_STACKALLOC(Node, Results);
2997 case ISD::MERGE_VALUES:
2998 for (unsigned i = 0; i < Node->getNumValues(); i++)
2999 Results.push_back(Node->getOperand(i));
3002 EVT VT = Node->getValueType(0);
3004 Results.push_back(DAG.getConstant(0, dl, VT));
3006 assert(VT.isFloatingPoint() && "Unknown value type!");
3007 Results.push_back(DAG.getConstantFP(0, dl, VT));
3013 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3014 Node->getValueType(0), dl);
3015 Results.push_back(Tmp1);
3017 case ISD::FP_EXTEND:
3018 Tmp1 = EmitStackConvert(Node->getOperand(0),
3019 Node->getOperand(0).getValueType(),
3020 Node->getValueType(0), dl);
3021 Results.push_back(Tmp1);
3023 case ISD::SIGN_EXTEND_INREG: {
3024 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3025 EVT VT = Node->getValueType(0);
3027 // An in-register sign-extend of a boolean is a negation:
3028 // 'true' (1) sign-extended is -1.
3029 // 'false' (0) sign-extended is 0.
3030 // However, we must mask the high bits of the source operand because the
3031 // SIGN_EXTEND_INREG does not guarantee that the high bits are already zero.
3033 // TODO: Do this for vectors too?
3034 if (ExtraVT.getSizeInBits() == 1) {
3035 SDValue One = DAG.getConstant(1, dl, VT);
3036 SDValue And = DAG.getNode(ISD::AND, dl, VT, Node->getOperand(0), One);
3037 SDValue Zero = DAG.getConstant(0, dl, VT);
3038 SDValue Neg = DAG.getNode(ISD::SUB, dl, VT, Zero, And);
3039 Results.push_back(Neg);
3043 // NOTE: we could fall back on load/store here too for targets without
3044 // SRA. However, it is doubtful that any exist.
3045 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
3046 unsigned BitsDiff = VT.getScalarSizeInBits() -
3047 ExtraVT.getScalarSizeInBits();
3048 SDValue ShiftCst = DAG.getConstant(BitsDiff, dl, ShiftAmountTy);
3049 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3050 Node->getOperand(0), ShiftCst);
3051 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3052 Results.push_back(Tmp1);
3055 case ISD::FP_ROUND_INREG: {
3056 // The only way we can lower this is to turn it into a TRUNCSTORE,
3057 // EXTLOAD pair, targeting a temporary location (a stack slot).
3059 // NOTE: there is a choice here between constantly creating new stack
3060 // slots and always reusing the same one. We currently always create
3061 // new ones, as reuse may inhibit scheduling.
3062 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3063 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3064 Node->getValueType(0), dl);
3065 Results.push_back(Tmp1);
3068 case ISD::SINT_TO_FP:
3069 case ISD::UINT_TO_FP:
3070 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3071 Node->getOperand(0), Node->getValueType(0), dl);
3072 Results.push_back(Tmp1);
3074 case ISD::FP_TO_SINT:
3075 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3076 Results.push_back(Tmp1);
3078 case ISD::FP_TO_UINT: {
3079 SDValue True, False;
3080 EVT VT = Node->getOperand(0).getValueType();
3081 EVT NVT = Node->getValueType(0);
3082 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3083 APInt::getNullValue(VT.getSizeInBits()));
3084 APInt x = APInt::getSignMask(NVT.getSizeInBits());
3085 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3086 Tmp1 = DAG.getConstantFP(apf, dl, VT);
3087 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3088 Node->getOperand(0),
3090 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3091 // TODO: Should any fast-math-flags be set for the FSUB?
3092 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3093 DAG.getNode(ISD::FSUB, dl, VT,
3094 Node->getOperand(0), Tmp1));
3095 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3096 DAG.getConstant(x, dl, NVT));
3097 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3098 Results.push_back(Tmp1);
3102 Results.push_back(DAG.expandVAArg(Node));
3103 Results.push_back(Results[0].getValue(1));
3106 Results.push_back(DAG.expandVACopy(Node));
3108 case ISD::EXTRACT_VECTOR_ELT:
3109 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3110 // This must be an access of the only element. Return it.
3111 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3112 Node->getOperand(0));
3114 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3115 Results.push_back(Tmp1);
3117 case ISD::EXTRACT_SUBVECTOR:
3118 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3120 case ISD::INSERT_SUBVECTOR:
3121 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3123 case ISD::CONCAT_VECTORS:
3124 Results.push_back(ExpandVectorBuildThroughStack(Node));
3126 case ISD::SCALAR_TO_VECTOR:
3127 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3129 case ISD::INSERT_VECTOR_ELT:
3130 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3131 Node->getOperand(1),
3132 Node->getOperand(2), dl));
3134 case ISD::VECTOR_SHUFFLE: {
3135 SmallVector<int, 32> NewMask;
3136 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3138 EVT VT = Node->getValueType(0);
3139 EVT EltVT = VT.getVectorElementType();
3140 SDValue Op0 = Node->getOperand(0);
3141 SDValue Op1 = Node->getOperand(1);
3142 if (!TLI.isTypeLegal(EltVT)) {
3143 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3145 // BUILD_VECTOR operands are allowed to be wider than the element type.
3146 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3148 if (NewEltVT.bitsLT(EltVT)) {
3149 // Convert shuffle node.
3150 // If original node was v4i64 and the new EltVT is i32,
3151 // cast operands to v8i32 and re-build the mask.
3153 // Calculate new VT, the size of the new VT should be equal to original.
3155 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3156 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3157 assert(NewVT.bitsEq(VT));
3159 // cast operands to new VT
3160 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3161 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3163 // Convert the shuffle mask
3164 unsigned int factor =
3165 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3167 // EltVT gets smaller
3170 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3172 for (unsigned fi = 0; fi < factor; ++fi)
3173 NewMask.push_back(Mask[i]);
3176 for (unsigned fi = 0; fi < factor; ++fi)
3177 NewMask.push_back(Mask[i]*factor+fi);
3185 unsigned NumElems = VT.getVectorNumElements();
3186 SmallVector<SDValue, 16> Ops;
3187 for (unsigned i = 0; i != NumElems; ++i) {
3189 Ops.push_back(DAG.getUNDEF(EltVT));
3192 unsigned Idx = Mask[i];
3194 Ops.push_back(DAG.getNode(
3195 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op0,
3196 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))));
3198 Ops.push_back(DAG.getNode(
3199 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op1,
3200 DAG.getConstant(Idx - NumElems, dl,
3201 TLI.getVectorIdxTy(DAG.getDataLayout()))));
3204 Tmp1 = DAG.getBuildVector(VT, dl, Ops);
3205 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3206 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3207 Results.push_back(Tmp1);
3210 case ISD::EXTRACT_ELEMENT: {
3211 EVT OpTy = Node->getOperand(0).getValueType();
3212 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3214 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3215 DAG.getConstant(OpTy.getSizeInBits() / 2, dl,
3216 TLI.getShiftAmountTy(
3217 Node->getOperand(0).getValueType(),
3218 DAG.getDataLayout())));
3219 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3222 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3223 Node->getOperand(0));
3225 Results.push_back(Tmp1);
3228 case ISD::STACKSAVE:
3229 // Expand to CopyFromReg if the target set
3230 // StackPointerRegisterToSaveRestore.
3231 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3232 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3233 Node->getValueType(0)));
3234 Results.push_back(Results[0].getValue(1));
3236 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3237 Results.push_back(Node->getOperand(0));
3240 case ISD::STACKRESTORE:
3241 // Expand to CopyToReg if the target set
3242 // StackPointerRegisterToSaveRestore.
3243 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3244 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3245 Node->getOperand(1)));
3247 Results.push_back(Node->getOperand(0));
3250 case ISD::GET_DYNAMIC_AREA_OFFSET:
3251 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0)));
3252 Results.push_back(Results[0].getValue(0));
3254 case ISD::FCOPYSIGN:
3255 Results.push_back(ExpandFCOPYSIGN(Node));
3258 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3259 Tmp1 = DAG.getConstantFP(-0.0, dl, Node->getValueType(0));
3260 // TODO: If FNEG has fast-math-flags, propagate them to the FSUB.
3261 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3262 Node->getOperand(0));
3263 Results.push_back(Tmp1);
3266 Results.push_back(ExpandFABS(Node));
3272 // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B
3274 switch (Node->getOpcode()) {
3275 default: llvm_unreachable("How did we get here?");
3276 case ISD::SMAX: Pred = ISD::SETGT; break;
3277 case ISD::SMIN: Pred = ISD::SETLT; break;
3278 case ISD::UMAX: Pred = ISD::SETUGT; break;
3279 case ISD::UMIN: Pred = ISD::SETULT; break;
3281 Tmp1 = Node->getOperand(0);
3282 Tmp2 = Node->getOperand(1);
3283 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp1, Tmp2, Pred);
3284 Results.push_back(Tmp1);
3290 EVT VT = Node->getValueType(0);
3291 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3292 // fcos which share the same operand and both are used.
3293 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3294 isSinCosLibcallAvailable(Node, TLI))
3295 && useSinCos(Node)) {
3296 SDVTList VTs = DAG.getVTList(VT, VT);
3297 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3298 if (Node->getOpcode() == ISD::FCOS)
3299 Tmp1 = Tmp1.getValue(1);
3300 Results.push_back(Tmp1);
3305 llvm_unreachable("Illegal fmad should never be formed");
3307 case ISD::FP16_TO_FP:
3308 if (Node->getValueType(0) != MVT::f32) {
3309 // We can extend to types bigger than f32 in two steps without changing
3310 // the result. Since "f16 -> f32" is much more commonly available, give
3311 // CodeGen the option of emitting that before resorting to a libcall.
3313 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3315 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3318 case ISD::FP_TO_FP16:
3319 DEBUG(dbgs() << "Legalizing FP_TO_FP16\n");
3320 if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) {
3321 SDValue Op = Node->getOperand(0);
3322 MVT SVT = Op.getSimpleValueType();
3323 if ((SVT == MVT::f64 || SVT == MVT::f80) &&
3324 TLI.isOperationLegalOrCustom(ISD::FP_TO_FP16, MVT::f32)) {
3325 // Under fastmath, we can expand this node into a fround followed by
3326 // a float-half conversion.
3327 SDValue FloatVal = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Op,
3328 DAG.getIntPtrConstant(0, dl));
3330 DAG.getNode(ISD::FP_TO_FP16, dl, Node->getValueType(0), FloatVal));
3334 case ISD::ConstantFP: {
3335 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3336 // Check to see if this FP immediate is already legal.
3337 // If this is a legal constant, turn it into a TargetConstantFP node.
3338 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3339 Results.push_back(ExpandConstantFP(CFP, true));
3342 case ISD::Constant: {
3343 ConstantSDNode *CP = cast<ConstantSDNode>(Node);
3344 Results.push_back(ExpandConstant(CP));
3348 EVT VT = Node->getValueType(0);
3349 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3350 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3351 const SDNodeFlags Flags = Node->getFlags();
3352 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3353 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1, Flags);
3354 Results.push_back(Tmp1);
3359 EVT VT = Node->getValueType(0);
3360 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3361 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3362 "Don't know how to expand this subtraction!");
3363 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3364 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
3366 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, dl, VT));
3367 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3372 EVT VT = Node->getValueType(0);
3373 bool isSigned = Node->getOpcode() == ISD::SREM;
3374 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3375 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3376 Tmp2 = Node->getOperand(0);
3377 Tmp3 = Node->getOperand(1);
3378 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
3379 SDVTList VTs = DAG.getVTList(VT, VT);
3380 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3381 Results.push_back(Tmp1);
3382 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3384 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3385 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3386 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3387 Results.push_back(Tmp1);
3393 bool isSigned = Node->getOpcode() == ISD::SDIV;
3394 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3395 EVT VT = Node->getValueType(0);
3396 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
3397 SDVTList VTs = DAG.getVTList(VT, VT);
3398 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3399 Node->getOperand(1));
3400 Results.push_back(Tmp1);
3406 unsigned ExpandOpcode =
3407 Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI : ISD::SMUL_LOHI;
3408 EVT VT = Node->getValueType(0);
3409 SDVTList VTs = DAG.getVTList(VT, VT);
3411 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3412 Node->getOperand(1));
3413 Results.push_back(Tmp1.getValue(1));
3416 case ISD::UMUL_LOHI:
3417 case ISD::SMUL_LOHI: {
3418 SDValue LHS = Node->getOperand(0);
3419 SDValue RHS = Node->getOperand(1);
3420 MVT VT = LHS.getSimpleValueType();
3421 unsigned MULHOpcode =
3422 Node->getOpcode() == ISD::UMUL_LOHI ? ISD::MULHU : ISD::MULHS;
3424 if (TLI.isOperationLegalOrCustom(MULHOpcode, VT)) {
3425 Results.push_back(DAG.getNode(ISD::MUL, dl, VT, LHS, RHS));
3426 Results.push_back(DAG.getNode(MULHOpcode, dl, VT, LHS, RHS));
3430 SmallVector<SDValue, 4> Halves;
3431 EVT HalfType = EVT(VT).getHalfSizedIntegerVT(*DAG.getContext());
3432 assert(TLI.isTypeLegal(HalfType));
3433 if (TLI.expandMUL_LOHI(Node->getOpcode(), VT, Node, LHS, RHS, Halves,
3435 TargetLowering::MulExpansionKind::Always)) {
3436 for (unsigned i = 0; i < 2; ++i) {
3437 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Halves[2 * i]);
3438 SDValue Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Halves[2 * i + 1]);
3439 SDValue Shift = DAG.getConstant(
3440 HalfType.getScalarSizeInBits(), dl,
3441 TLI.getShiftAmountTy(HalfType, DAG.getDataLayout()));
3442 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3443 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3450 EVT VT = Node->getValueType(0);
3451 SDVTList VTs = DAG.getVTList(VT, VT);
3452 // See if multiply or divide can be lowered using two-result operations.
3453 // We just need the low half of the multiply; try both the signed
3454 // and unsigned forms. If the target supports both SMUL_LOHI and
3455 // UMUL_LOHI, form a preference by checking which forms of plain
3456 // MULH it supports.
3457 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3458 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3459 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3460 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3461 unsigned OpToUse = 0;
3462 if (HasSMUL_LOHI && !HasMULHS) {
3463 OpToUse = ISD::SMUL_LOHI;
3464 } else if (HasUMUL_LOHI && !HasMULHU) {
3465 OpToUse = ISD::UMUL_LOHI;
3466 } else if (HasSMUL_LOHI) {
3467 OpToUse = ISD::SMUL_LOHI;
3468 } else if (HasUMUL_LOHI) {
3469 OpToUse = ISD::UMUL_LOHI;
3472 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3473 Node->getOperand(1)));
3478 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3479 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3480 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3481 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3482 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3483 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG,
3484 TargetLowering::MulExpansionKind::OnlyLegalOrCustom)) {
3485 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3486 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3488 DAG.getConstant(HalfType.getSizeInBits(), dl,
3489 TLI.getShiftAmountTy(HalfType, DAG.getDataLayout()));
3490 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3491 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3497 SDValue LHS = Node->getOperand(0);
3498 SDValue RHS = Node->getOperand(1);
3499 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3500 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3502 Results.push_back(Sum);
3503 EVT ResultType = Node->getValueType(1);
3504 EVT OType = getSetCCResultType(Node->getValueType(0));
3506 SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType());
3508 // LHSSign -> LHS >= 0
3509 // RHSSign -> RHS >= 0
3510 // SumSign -> Sum >= 0
3513 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3515 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3516 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3517 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3518 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3519 Node->getOpcode() == ISD::SADDO ?
3520 ISD::SETEQ : ISD::SETNE);
3522 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3523 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3525 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3526 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3531 SDValue LHS = Node->getOperand(0);
3532 SDValue RHS = Node->getOperand(1);
3533 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3534 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3536 Results.push_back(Sum);
3538 EVT ResultType = Node->getValueType(1);
3539 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3541 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3542 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3544 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3549 EVT VT = Node->getValueType(0);
3550 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3551 SDValue LHS = Node->getOperand(0);
3552 SDValue RHS = Node->getOperand(1);
3555 static const unsigned Ops[2][3] =
3556 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3557 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3558 bool isSigned = Node->getOpcode() == ISD::SMULO;
3559 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3560 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3561 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3562 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3563 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3565 TopHalf = BottomHalf.getValue(1);
3566 } else if (TLI.isTypeLegal(WideVT)) {
3567 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3568 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3569 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3570 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3571 DAG.getIntPtrConstant(0, dl));
3572 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3573 DAG.getIntPtrConstant(1, dl));
3575 // We can fall back to a libcall with an illegal type for the MUL if we
3576 // have a libcall big enough.
3577 // Also, we can fall back to a division in some cases, but that's a big
3578 // performance hit in the general case.
3579 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3580 if (WideVT == MVT::i16)
3581 LC = RTLIB::MUL_I16;
3582 else if (WideVT == MVT::i32)
3583 LC = RTLIB::MUL_I32;
3584 else if (WideVT == MVT::i64)
3585 LC = RTLIB::MUL_I64;
3586 else if (WideVT == MVT::i128)
3587 LC = RTLIB::MUL_I128;
3588 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3593 // The high part is obtained by SRA'ing all but one of the bits of low
3595 unsigned LoSize = VT.getSizeInBits();
3597 DAG.getNode(ISD::SRA, dl, VT, LHS,
3598 DAG.getConstant(LoSize - 1, dl,
3599 TLI.getPointerTy(DAG.getDataLayout())));
3601 DAG.getNode(ISD::SRA, dl, VT, RHS,
3602 DAG.getConstant(LoSize - 1, dl,
3603 TLI.getPointerTy(DAG.getDataLayout())));
3605 HiLHS = DAG.getConstant(0, dl, VT);
3606 HiRHS = DAG.getConstant(0, dl, VT);
3609 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3610 // pre-lowered to the correct types. This all depends upon WideVT not
3611 // being a legal type for the architecture and thus has to be split to
3614 if(DAG.getDataLayout().isLittleEndian()) {
3615 // Halves of WideVT are packed into registers in different order
3616 // depending on platform endianness. This is usually handled by
3617 // the C calling convention, but we can't defer to it in
3619 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3620 Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3622 SDValue Args[] = { HiLHS, LHS, HiRHS, RHS };
3623 Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3625 assert(Ret.getOpcode() == ISD::MERGE_VALUES &&
3626 "Ret value is a collection of constituent nodes holding result.");
3627 BottomHalf = Ret.getOperand(0);
3628 TopHalf = Ret.getOperand(1);
3632 Tmp1 = DAG.getConstant(
3633 VT.getSizeInBits() - 1, dl,
3634 TLI.getShiftAmountTy(BottomHalf.getValueType(), DAG.getDataLayout()));
3635 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3636 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3639 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3640 DAG.getConstant(0, dl, VT), ISD::SETNE);
3643 // Truncate the result if SetCC returns a larger type than needed.
3644 EVT RType = Node->getValueType(1);
3645 if (RType.getSizeInBits() < TopHalf.getValueSizeInBits())
3646 TopHalf = DAG.getNode(ISD::TRUNCATE, dl, RType, TopHalf);
3648 assert(RType.getSizeInBits() == TopHalf.getValueSizeInBits() &&
3649 "Unexpected result type for S/UMULO legalization");
3651 Results.push_back(BottomHalf);
3652 Results.push_back(TopHalf);
3655 case ISD::BUILD_PAIR: {
3656 EVT PairTy = Node->getValueType(0);
3657 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3658 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3660 ISD::SHL, dl, PairTy, Tmp2,
3661 DAG.getConstant(PairTy.getSizeInBits() / 2, dl,
3662 TLI.getShiftAmountTy(PairTy, DAG.getDataLayout())));
3663 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3667 Tmp1 = Node->getOperand(0);
3668 Tmp2 = Node->getOperand(1);
3669 Tmp3 = Node->getOperand(2);
3670 if (Tmp1.getOpcode() == ISD::SETCC) {
3671 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3673 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3675 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3676 DAG.getConstant(0, dl, Tmp1.getValueType()),
3677 Tmp2, Tmp3, ISD::SETNE);
3679 Results.push_back(Tmp1);
3682 SDValue Chain = Node->getOperand(0);
3683 SDValue Table = Node->getOperand(1);
3684 SDValue Index = Node->getOperand(2);
3686 const DataLayout &TD = DAG.getDataLayout();
3687 EVT PTy = TLI.getPointerTy(TD);
3689 unsigned EntrySize =
3690 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3692 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(), Index,
3693 DAG.getConstant(EntrySize, dl, Index.getValueType()));
3694 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3697 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3698 SDValue LD = DAG.getExtLoad(
3699 ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3700 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()), MemVT);
3702 if (TLI.isJumpTableRelative()) {
3703 // For PIC, the sequence is:
3704 // BRIND(load(Jumptable + index) + RelocBase)
3705 // RelocBase can be JumpTable, GOT or some sort of global base.
3706 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3707 TLI.getPICJumpTableRelocBase(Table, DAG));
3709 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3710 Results.push_back(Tmp1);
3714 // Expand brcond's setcc into its constituent parts and create a BR_CC
3716 Tmp1 = Node->getOperand(0);
3717 Tmp2 = Node->getOperand(1);
3718 if (Tmp2.getOpcode() == ISD::SETCC) {
3719 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3720 Tmp1, Tmp2.getOperand(2),
3721 Tmp2.getOperand(0), Tmp2.getOperand(1),
3722 Node->getOperand(2));
3724 // We test only the i1 bit. Skip the AND if UNDEF or another AND.
3725 if (Tmp2.isUndef() ||
3726 (Tmp2.getOpcode() == ISD::AND &&
3727 isa<ConstantSDNode>(Tmp2.getOperand(1)) &&
3728 dyn_cast<ConstantSDNode>(Tmp2.getOperand(1))->getZExtValue() == 1))
3731 Tmp3 = DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3732 DAG.getConstant(1, dl, Tmp2.getValueType()));
3733 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3734 DAG.getCondCode(ISD::SETNE), Tmp3,
3735 DAG.getConstant(0, dl, Tmp3.getValueType()),
3736 Node->getOperand(2));
3738 Results.push_back(Tmp1);
3741 Tmp1 = Node->getOperand(0);
3742 Tmp2 = Node->getOperand(1);
3743 Tmp3 = Node->getOperand(2);
3744 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3745 Tmp3, NeedInvert, dl);
3748 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3749 // condition code, create a new SETCC node.
3751 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3754 // If we expanded the SETCC by inverting the condition code, then wrap
3755 // the existing SETCC in a NOT to restore the intended condition.
3757 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
3759 Results.push_back(Tmp1);
3763 // Otherwise, SETCC for the given comparison type must be completely
3764 // illegal; expand it into a SELECT_CC.
3765 EVT VT = Node->getValueType(0);
3767 switch (TLI.getBooleanContents(Tmp1.getValueType())) {
3768 case TargetLowering::ZeroOrOneBooleanContent:
3769 case TargetLowering::UndefinedBooleanContent:
3772 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3776 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3777 DAG.getConstant(TrueValue, dl, VT),
3778 DAG.getConstant(0, dl, VT),
3780 Results.push_back(Tmp1);
3783 case ISD::SELECT_CC: {
3784 Tmp1 = Node->getOperand(0); // LHS
3785 Tmp2 = Node->getOperand(1); // RHS
3786 Tmp3 = Node->getOperand(2); // True
3787 Tmp4 = Node->getOperand(3); // False
3788 EVT VT = Node->getValueType(0);
3789 SDValue CC = Node->getOperand(4);
3790 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
3792 if (TLI.isCondCodeLegalOrCustom(CCOp, Tmp1.getSimpleValueType())) {
3793 // If the condition code is legal, then we need to expand this
3794 // node using SETCC and SELECT.
3795 EVT CmpVT = Tmp1.getValueType();
3796 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
3797 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
3800 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT);
3801 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
3802 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
3806 // SELECT_CC is legal, so the condition code must not be.
3807 bool Legalized = false;
3808 // Try to legalize by inverting the condition. This is for targets that
3809 // might support an ordered version of a condition, but not the unordered
3810 // version (or vice versa).
3811 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
3812 Tmp1.getValueType().isInteger());
3813 if (TLI.isCondCodeLegalOrCustom(InvCC, Tmp1.getSimpleValueType())) {
3814 // Use the new condition code and swap true and false
3816 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
3818 // If The inverse is not legal, then try to swap the arguments using
3819 // the inverse condition code.
3820 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
3821 if (TLI.isCondCodeLegalOrCustom(SwapInvCC, Tmp1.getSimpleValueType())) {
3822 // The swapped inverse condition is legal, so swap true and false,
3825 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
3830 Legalized = LegalizeSetCCCondCode(
3831 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
3834 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
3836 // If we expanded the SETCC by inverting the condition code, then swap
3837 // the True/False operands to match.
3839 std::swap(Tmp3, Tmp4);
3841 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3842 // condition code, create a new SELECT_CC node.
3844 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
3845 Tmp1, Tmp2, Tmp3, Tmp4, CC);
3847 Tmp2 = DAG.getConstant(0, dl, Tmp1.getValueType());
3848 CC = DAG.getCondCode(ISD::SETNE);
3849 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
3850 Tmp2, Tmp3, Tmp4, CC);
3853 Results.push_back(Tmp1);
3857 Tmp1 = Node->getOperand(0); // Chain
3858 Tmp2 = Node->getOperand(2); // LHS
3859 Tmp3 = Node->getOperand(3); // RHS
3860 Tmp4 = Node->getOperand(1); // CC
3862 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
3863 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
3865 assert(Legalized && "Can't legalize BR_CC with legal condition!");
3867 // If we expanded the SETCC by inverting the condition code, then wrap
3868 // the existing SETCC in a NOT to restore the intended condition.
3870 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
3872 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
3874 if (Tmp4.getNode()) {
3875 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
3876 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
3878 Tmp3 = DAG.getConstant(0, dl, Tmp2.getValueType());
3879 Tmp4 = DAG.getCondCode(ISD::SETNE);
3880 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
3881 Tmp2, Tmp3, Node->getOperand(4));
3883 Results.push_back(Tmp1);
3886 case ISD::BUILD_VECTOR:
3887 Results.push_back(ExpandBUILD_VECTOR(Node));
3892 // Scalarize vector SRA/SRL/SHL.
3893 EVT VT = Node->getValueType(0);
3894 assert(VT.isVector() && "Unable to legalize non-vector shift");
3895 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
3896 unsigned NumElem = VT.getVectorNumElements();
3898 SmallVector<SDValue, 8> Scalars;
3899 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
3900 SDValue Ex = DAG.getNode(
3901 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(0),
3902 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
3903 SDValue Sh = DAG.getNode(
3904 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(1),
3905 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
3906 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
3907 VT.getScalarType(), Ex, Sh));
3910 SDValue Result = DAG.getBuildVector(Node->getValueType(0), dl, Scalars);
3911 ReplaceNode(SDValue(Node, 0), Result);
3914 case ISD::GLOBAL_OFFSET_TABLE:
3915 case ISD::GlobalAddress:
3916 case ISD::GlobalTLSAddress:
3917 case ISD::ExternalSymbol:
3918 case ISD::ConstantPool:
3919 case ISD::JumpTable:
3920 case ISD::INTRINSIC_W_CHAIN:
3921 case ISD::INTRINSIC_WO_CHAIN:
3922 case ISD::INTRINSIC_VOID:
3923 // FIXME: Custom lowering for these operations shouldn't return null!
3927 // Replace the original node with the legalized result.
3928 if (Results.empty()) {
3929 DEBUG(dbgs() << "Cannot expand node\n");
3933 DEBUG(dbgs() << "Succesfully expanded node\n");
3934 ReplaceNode(Node, Results.data());
3938 void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) {
3939 DEBUG(dbgs() << "Trying to convert node to libcall\n");
3940 SmallVector<SDValue, 8> Results;
3942 // FIXME: Check flags on the node to see if we can use a finite call.
3943 bool CanUseFiniteLibCall = TM.Options.NoInfsFPMath && TM.Options.NoNaNsFPMath;
3944 unsigned Opc = Node->getOpcode();
3946 case ISD::ATOMIC_FENCE: {
3947 // If the target didn't lower this, lower it to '__sync_synchronize()' call
3948 // FIXME: handle "fence singlethread" more efficiently.
3949 TargetLowering::ArgListTy Args;
3951 TargetLowering::CallLoweringInfo CLI(DAG);
3953 .setChain(Node->getOperand(0))
3955 CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3956 DAG.getExternalSymbol("__sync_synchronize",
3957 TLI.getPointerTy(DAG.getDataLayout())),
3960 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3962 Results.push_back(CallResult.second);
3965 // By default, atomic intrinsics are marked Legal and lowered. Targets
3966 // which don't support them directly, however, may want libcalls, in which
3967 // case they mark them Expand, and we get here.
3968 case ISD::ATOMIC_SWAP:
3969 case ISD::ATOMIC_LOAD_ADD:
3970 case ISD::ATOMIC_LOAD_SUB:
3971 case ISD::ATOMIC_LOAD_AND:
3972 case ISD::ATOMIC_LOAD_CLR:
3973 case ISD::ATOMIC_LOAD_OR:
3974 case ISD::ATOMIC_LOAD_XOR:
3975 case ISD::ATOMIC_LOAD_NAND:
3976 case ISD::ATOMIC_LOAD_MIN:
3977 case ISD::ATOMIC_LOAD_MAX:
3978 case ISD::ATOMIC_LOAD_UMIN:
3979 case ISD::ATOMIC_LOAD_UMAX:
3980 case ISD::ATOMIC_CMP_SWAP: {
3981 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
3982 RTLIB::Libcall LC = RTLIB::getSYNC(Opc, VT);
3983 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!");
3985 std::pair<SDValue, SDValue> Tmp = ExpandChainLibCall(LC, Node, false);
3986 Results.push_back(Tmp.first);
3987 Results.push_back(Tmp.second);
3991 // If this operation is not supported, lower it to 'abort()' call
3992 TargetLowering::ArgListTy Args;
3993 TargetLowering::CallLoweringInfo CLI(DAG);
3995 .setChain(Node->getOperand(0))
3996 .setLibCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3997 DAG.getExternalSymbol(
3998 "abort", TLI.getPointerTy(DAG.getDataLayout())),
4000 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
4002 Results.push_back(CallResult.second);
4006 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
4007 RTLIB::FMIN_F80, RTLIB::FMIN_F128,
4008 RTLIB::FMIN_PPCF128));
4011 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
4012 RTLIB::FMAX_F80, RTLIB::FMAX_F128,
4013 RTLIB::FMAX_PPCF128));
4016 case ISD::STRICT_FSQRT:
4017 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
4018 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
4019 RTLIB::SQRT_PPCF128));
4022 case ISD::STRICT_FSIN:
4023 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
4024 RTLIB::SIN_F80, RTLIB::SIN_F128,
4025 RTLIB::SIN_PPCF128));
4028 case ISD::STRICT_FCOS:
4029 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
4030 RTLIB::COS_F80, RTLIB::COS_F128,
4031 RTLIB::COS_PPCF128));
4034 // Expand into sincos libcall.
4035 ExpandSinCosLibCall(Node, Results);
4038 case ISD::STRICT_FLOG:
4039 if (CanUseFiniteLibCall && DAG.getLibInfo().has(LibFunc_log_finite))
4040 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_FINITE_F32,
4041 RTLIB::LOG_FINITE_F64,
4042 RTLIB::LOG_FINITE_F80,
4043 RTLIB::LOG_FINITE_F128,
4044 RTLIB::LOG_FINITE_PPCF128));
4046 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
4047 RTLIB::LOG_F80, RTLIB::LOG_F128,
4048 RTLIB::LOG_PPCF128));
4051 case ISD::STRICT_FLOG2:
4052 if (CanUseFiniteLibCall && DAG.getLibInfo().has(LibFunc_log2_finite))
4053 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_FINITE_F32,
4054 RTLIB::LOG2_FINITE_F64,
4055 RTLIB::LOG2_FINITE_F80,
4056 RTLIB::LOG2_FINITE_F128,
4057 RTLIB::LOG2_FINITE_PPCF128));
4059 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
4060 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
4061 RTLIB::LOG2_PPCF128));
4064 case ISD::STRICT_FLOG10:
4065 if (CanUseFiniteLibCall && DAG.getLibInfo().has(LibFunc_log10_finite))
4066 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_FINITE_F32,
4067 RTLIB::LOG10_FINITE_F64,
4068 RTLIB::LOG10_FINITE_F80,
4069 RTLIB::LOG10_FINITE_F128,
4070 RTLIB::LOG10_FINITE_PPCF128));
4072 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
4073 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
4074 RTLIB::LOG10_PPCF128));
4077 case ISD::STRICT_FEXP:
4078 if (CanUseFiniteLibCall && DAG.getLibInfo().has(LibFunc_exp_finite))
4079 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_FINITE_F32,
4080 RTLIB::EXP_FINITE_F64,
4081 RTLIB::EXP_FINITE_F80,
4082 RTLIB::EXP_FINITE_F128,
4083 RTLIB::EXP_FINITE_PPCF128));
4085 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
4086 RTLIB::EXP_F80, RTLIB::EXP_F128,
4087 RTLIB::EXP_PPCF128));
4090 case ISD::STRICT_FEXP2:
4091 if (CanUseFiniteLibCall && DAG.getLibInfo().has(LibFunc_exp2_finite))
4092 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_FINITE_F32,
4093 RTLIB::EXP2_FINITE_F64,
4094 RTLIB::EXP2_FINITE_F80,
4095 RTLIB::EXP2_FINITE_F128,
4096 RTLIB::EXP2_FINITE_PPCF128));
4098 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
4099 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
4100 RTLIB::EXP2_PPCF128));
4103 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
4104 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
4105 RTLIB::TRUNC_PPCF128));
4108 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
4109 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
4110 RTLIB::FLOOR_PPCF128));
4113 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
4114 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
4115 RTLIB::CEIL_PPCF128));
4118 case ISD::STRICT_FRINT:
4119 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
4120 RTLIB::RINT_F80, RTLIB::RINT_F128,
4121 RTLIB::RINT_PPCF128));
4123 case ISD::FNEARBYINT:
4124 case ISD::STRICT_FNEARBYINT:
4125 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
4126 RTLIB::NEARBYINT_F64,
4127 RTLIB::NEARBYINT_F80,
4128 RTLIB::NEARBYINT_F128,
4129 RTLIB::NEARBYINT_PPCF128));
4132 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
4136 RTLIB::ROUND_PPCF128));
4139 case ISD::STRICT_FPOWI:
4140 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
4141 RTLIB::POWI_F80, RTLIB::POWI_F128,
4142 RTLIB::POWI_PPCF128));
4145 case ISD::STRICT_FPOW:
4146 if (CanUseFiniteLibCall && DAG.getLibInfo().has(LibFunc_pow_finite))
4147 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_FINITE_F32,
4148 RTLIB::POW_FINITE_F64,
4149 RTLIB::POW_FINITE_F80,
4150 RTLIB::POW_FINITE_F128,
4151 RTLIB::POW_FINITE_PPCF128));
4153 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
4154 RTLIB::POW_F80, RTLIB::POW_F128,
4155 RTLIB::POW_PPCF128));
4158 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
4159 RTLIB::DIV_F80, RTLIB::DIV_F128,
4160 RTLIB::DIV_PPCF128));
4163 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
4164 RTLIB::REM_F80, RTLIB::REM_F128,
4165 RTLIB::REM_PPCF128));
4168 case ISD::STRICT_FMA:
4169 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
4170 RTLIB::FMA_F80, RTLIB::FMA_F128,
4171 RTLIB::FMA_PPCF128));
4174 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
4175 RTLIB::ADD_F80, RTLIB::ADD_F128,
4176 RTLIB::ADD_PPCF128));
4179 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
4180 RTLIB::MUL_F80, RTLIB::MUL_F128,
4181 RTLIB::MUL_PPCF128));
4183 case ISD::FP16_TO_FP:
4184 if (Node->getValueType(0) == MVT::f32) {
4185 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
4188 case ISD::FP_TO_FP16: {
4190 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
4191 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
4192 Results.push_back(ExpandLibCall(LC, Node, false));
4196 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
4197 RTLIB::SUB_F80, RTLIB::SUB_F128,
4198 RTLIB::SUB_PPCF128));
4201 Results.push_back(ExpandIntLibCall(Node, true,
4203 RTLIB::SREM_I16, RTLIB::SREM_I32,
4204 RTLIB::SREM_I64, RTLIB::SREM_I128));
4207 Results.push_back(ExpandIntLibCall(Node, false,
4209 RTLIB::UREM_I16, RTLIB::UREM_I32,
4210 RTLIB::UREM_I64, RTLIB::UREM_I128));
4213 Results.push_back(ExpandIntLibCall(Node, true,
4215 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
4216 RTLIB::SDIV_I64, RTLIB::SDIV_I128));
4219 Results.push_back(ExpandIntLibCall(Node, false,
4221 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
4222 RTLIB::UDIV_I64, RTLIB::UDIV_I128));
4226 // Expand into divrem libcall
4227 ExpandDivRemLibCall(Node, Results);
4230 Results.push_back(ExpandIntLibCall(Node, false,
4232 RTLIB::MUL_I16, RTLIB::MUL_I32,
4233 RTLIB::MUL_I64, RTLIB::MUL_I128));
4237 // Replace the original node with the legalized result.
4238 if (!Results.empty()) {
4239 DEBUG(dbgs() << "Successfully converted node to libcall\n");
4240 ReplaceNode(Node, Results.data());
4242 DEBUG(dbgs() << "Could not convert node to libcall\n");
4245 // Determine the vector type to use in place of an original scalar element when
4246 // promoting equally sized vectors.
4247 static MVT getPromotedVectorElementType(const TargetLowering &TLI,
4248 MVT EltVT, MVT NewEltVT) {
4249 unsigned OldEltsPerNewElt = EltVT.getSizeInBits() / NewEltVT.getSizeInBits();
4250 MVT MidVT = MVT::getVectorVT(NewEltVT, OldEltsPerNewElt);
4251 assert(TLI.isTypeLegal(MidVT) && "unexpected");
4255 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4256 DEBUG(dbgs() << "Trying to promote node\n");
4257 SmallVector<SDValue, 8> Results;
4258 MVT OVT = Node->getSimpleValueType(0);
4259 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4260 Node->getOpcode() == ISD::SINT_TO_FP ||
4261 Node->getOpcode() == ISD::SETCC ||
4262 Node->getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
4263 Node->getOpcode() == ISD::INSERT_VECTOR_ELT) {
4264 OVT = Node->getOperand(0).getSimpleValueType();
4266 if (Node->getOpcode() == ISD::BR_CC)
4267 OVT = Node->getOperand(2).getSimpleValueType();
4268 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4270 SDValue Tmp1, Tmp2, Tmp3;
4271 switch (Node->getOpcode()) {
4273 case ISD::CTTZ_ZERO_UNDEF:
4275 case ISD::CTLZ_ZERO_UNDEF:
4277 // Zero extend the argument.
4278 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4279 if (Node->getOpcode() == ISD::CTTZ) {
4280 // The count is the same in the promoted type except if the original
4281 // value was zero. This can be handled by setting the bit just off
4282 // the top of the original type.
4283 auto TopBit = APInt::getOneBitSet(NVT.getSizeInBits(),
4284 OVT.getSizeInBits());
4285 Tmp1 = DAG.getNode(ISD::OR, dl, NVT, Tmp1,
4286 DAG.getConstant(TopBit, dl, NVT));
4288 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4289 // already the correct result.
4290 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4291 if (Node->getOpcode() == ISD::CTLZ ||
4292 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4293 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4294 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4295 DAG.getConstant(NVT.getSizeInBits() -
4296 OVT.getSizeInBits(), dl, NVT));
4298 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4300 case ISD::BITREVERSE:
4302 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4303 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4304 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4306 ISD::SRL, dl, NVT, Tmp1,
4307 DAG.getConstant(DiffBits, dl,
4308 TLI.getShiftAmountTy(NVT, DAG.getDataLayout())));
4310 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4313 case ISD::FP_TO_UINT:
4314 case ISD::FP_TO_SINT:
4315 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
4316 Node->getOpcode() == ISD::FP_TO_SINT, dl);
4317 Results.push_back(Tmp1);
4319 case ISD::UINT_TO_FP:
4320 case ISD::SINT_TO_FP:
4321 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
4322 Node->getOpcode() == ISD::SINT_TO_FP, dl);
4323 Results.push_back(Tmp1);
4326 SDValue Chain = Node->getOperand(0); // Get the chain.
4327 SDValue Ptr = Node->getOperand(1); // Get the pointer.
4330 if (OVT.isVector()) {
4331 TruncOp = ISD::BITCAST;
4333 assert(OVT.isInteger()
4334 && "VAARG promotion is supported only for vectors or integer types");
4335 TruncOp = ISD::TRUNCATE;
4338 // Perform the larger operation, then convert back
4339 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
4340 Node->getConstantOperandVal(3));
4341 Chain = Tmp1.getValue(1);
4343 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4345 // Modified the chain result - switch anything that used the old chain to
4347 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4348 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4350 UpdatedNodes->insert(Tmp2.getNode());
4351 UpdatedNodes->insert(Chain.getNode());
4364 unsigned ExtOp, TruncOp;
4365 if (OVT.isVector()) {
4366 ExtOp = ISD::BITCAST;
4367 TruncOp = ISD::BITCAST;
4369 assert(OVT.isInteger() && "Cannot promote logic operation");
4371 switch (Node->getOpcode()) {
4373 ExtOp = ISD::ANY_EXTEND;
4377 ExtOp = ISD::SIGN_EXTEND;
4381 ExtOp = ISD::ZERO_EXTEND;
4384 TruncOp = ISD::TRUNCATE;
4386 // Promote each of the values to the new type.
4387 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4388 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4389 // Perform the larger operation, then convert back
4390 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4391 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4394 case ISD::UMUL_LOHI:
4395 case ISD::SMUL_LOHI: {
4396 // Promote to a multiply in a wider integer type.
4397 unsigned ExtOp = Node->getOpcode() == ISD::UMUL_LOHI ? ISD::ZERO_EXTEND
4399 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4400 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4401 Tmp1 = DAG.getNode(ISD::MUL, dl, NVT, Tmp1, Tmp2);
4403 auto &DL = DAG.getDataLayout();
4404 unsigned OriginalSize = OVT.getScalarSizeInBits();
4406 ISD::SRL, dl, NVT, Tmp1,
4407 DAG.getConstant(OriginalSize, dl, TLI.getScalarShiftAmountTy(DL, NVT)));
4408 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4409 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp2));
4413 unsigned ExtOp, TruncOp;
4414 if (Node->getValueType(0).isVector() ||
4415 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4416 ExtOp = ISD::BITCAST;
4417 TruncOp = ISD::BITCAST;
4418 } else if (Node->getValueType(0).isInteger()) {
4419 ExtOp = ISD::ANY_EXTEND;
4420 TruncOp = ISD::TRUNCATE;
4422 ExtOp = ISD::FP_EXTEND;
4423 TruncOp = ISD::FP_ROUND;
4425 Tmp1 = Node->getOperand(0);
4426 // Promote each of the values to the new type.
4427 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4428 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4429 // Perform the larger operation, then round down.
4430 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4431 if (TruncOp != ISD::FP_ROUND)
4432 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4434 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4435 DAG.getIntPtrConstant(0, dl));
4436 Results.push_back(Tmp1);
4439 case ISD::VECTOR_SHUFFLE: {
4440 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4442 // Cast the two input vectors.
4443 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4444 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4446 // Convert the shuffle mask to the right # elements.
4447 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4448 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4449 Results.push_back(Tmp1);
4453 unsigned ExtOp = ISD::FP_EXTEND;
4454 if (NVT.isInteger()) {
4455 ISD::CondCode CCCode =
4456 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4457 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4459 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4460 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4461 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4462 Tmp1, Tmp2, Node->getOperand(2)));
4466 unsigned ExtOp = ISD::FP_EXTEND;
4467 if (NVT.isInteger()) {
4468 ISD::CondCode CCCode =
4469 cast<CondCodeSDNode>(Node->getOperand(1))->get();
4470 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4472 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4473 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(3));
4474 Results.push_back(DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0),
4475 Node->getOperand(0), Node->getOperand(1),
4476 Tmp1, Tmp2, Node->getOperand(4)));
4487 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4488 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4489 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2,
4491 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4492 Tmp3, DAG.getIntPtrConstant(0, dl)));
4495 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4496 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4497 Tmp3 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(2));
4499 DAG.getNode(ISD::FP_ROUND, dl, OVT,
4500 DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2, Tmp3),
4501 DAG.getIntPtrConstant(0, dl)));
4503 case ISD::FCOPYSIGN:
4505 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4506 Tmp2 = Node->getOperand(1);
4507 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4509 // fcopysign doesn't change anything but the sign bit, so
4510 // (fp_round (fcopysign (fpext a), b))
4512 // (fp_round (fpext a))
4513 // which is a no-op. Mark it as a TRUNCating FP_ROUND.
4514 const bool isTrunc = (Node->getOpcode() == ISD::FCOPYSIGN);
4515 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4516 Tmp3, DAG.getIntPtrConstant(isTrunc, dl)));
4522 case ISD::FNEARBYINT:
4535 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4536 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4537 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4538 Tmp2, DAG.getIntPtrConstant(0, dl)));
4540 case ISD::BUILD_VECTOR: {
4541 MVT EltVT = OVT.getVectorElementType();
4542 MVT NewEltVT = NVT.getVectorElementType();
4544 // Handle bitcasts to a different vector type with the same total bit size
4546 // e.g. v2i64 = build_vector i64:x, i64:y => v4i32
4548 // v4i32 = concat_vectors (v2i32 (bitcast i64:x)), (v2i32 (bitcast i64:y))
4550 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
4551 "Invalid promote type for build_vector");
4552 assert(NewEltVT.bitsLT(EltVT) && "not handled");
4554 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4556 SmallVector<SDValue, 8> NewOps;
4557 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I) {
4558 SDValue Op = Node->getOperand(I);
4559 NewOps.push_back(DAG.getNode(ISD::BITCAST, SDLoc(Op), MidVT, Op));
4563 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewOps);
4564 SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat);
4565 Results.push_back(CvtVec);
4568 case ISD::EXTRACT_VECTOR_ELT: {
4569 MVT EltVT = OVT.getVectorElementType();
4570 MVT NewEltVT = NVT.getVectorElementType();
4572 // Handle bitcasts to a different vector type with the same total bit size.
4574 // e.g. v2i64 = extract_vector_elt x:v2i64, y:i32
4576 // v4i32:castx = bitcast x:v2i64
4579 // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))),
4580 // (i32 (extract_vector_elt castx, (2 * y + 1)))
4583 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
4584 "Invalid promote type for extract_vector_elt");
4585 assert(NewEltVT.bitsLT(EltVT) && "not handled");
4587 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4588 unsigned NewEltsPerOldElt = MidVT.getVectorNumElements();
4590 SDValue Idx = Node->getOperand(1);
4591 EVT IdxVT = Idx.getValueType();
4593 SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SL, IdxVT);
4594 SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor);
4596 SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0));
4598 SmallVector<SDValue, 8> NewOps;
4599 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
4600 SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT);
4601 SDValue TmpIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset);
4603 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT,
4605 NewOps.push_back(Elt);
4608 SDValue NewVec = DAG.getBuildVector(MidVT, SL, NewOps);
4609 Results.push_back(DAG.getNode(ISD::BITCAST, SL, EltVT, NewVec));
4612 case ISD::INSERT_VECTOR_ELT: {
4613 MVT EltVT = OVT.getVectorElementType();
4614 MVT NewEltVT = NVT.getVectorElementType();
4616 // Handle bitcasts to a different vector type with the same total bit size
4618 // e.g. v2i64 = insert_vector_elt x:v2i64, y:i64, z:i32
4620 // v4i32:castx = bitcast x:v2i64
4621 // v2i32:casty = bitcast y:i64
4624 // (v4i32 insert_vector_elt
4625 // (v4i32 insert_vector_elt v4i32:castx,
4626 // (extract_vector_elt casty, 0), 2 * z),
4627 // (extract_vector_elt casty, 1), (2 * z + 1))
4629 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
4630 "Invalid promote type for insert_vector_elt");
4631 assert(NewEltVT.bitsLT(EltVT) && "not handled");
4633 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4634 unsigned NewEltsPerOldElt = MidVT.getVectorNumElements();
4636 SDValue Val = Node->getOperand(1);
4637 SDValue Idx = Node->getOperand(2);
4638 EVT IdxVT = Idx.getValueType();
4641 SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SDLoc(), IdxVT);
4642 SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor);
4644 SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0));
4645 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val);
4647 SDValue NewVec = CastVec;
4648 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
4649 SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT);
4650 SDValue InEltIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset);
4652 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT,
4653 CastVal, IdxOffset);
4655 NewVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, SL, NVT,
4656 NewVec, Elt, InEltIdx);
4659 Results.push_back(DAG.getNode(ISD::BITCAST, SL, OVT, NewVec));
4662 case ISD::SCALAR_TO_VECTOR: {
4663 MVT EltVT = OVT.getVectorElementType();
4664 MVT NewEltVT = NVT.getVectorElementType();
4666 // Handle bitcasts to different vector type with the same total bit size.
4668 // e.g. v2i64 = scalar_to_vector x:i64
4670 // concat_vectors (v2i32 bitcast x:i64), (v2i32 undef)
4673 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
4674 SDValue Val = Node->getOperand(0);
4677 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val);
4678 SDValue Undef = DAG.getUNDEF(MidVT);
4680 SmallVector<SDValue, 8> NewElts;
4681 NewElts.push_back(CastVal);
4682 for (unsigned I = 1, NElts = OVT.getVectorNumElements(); I != NElts; ++I)
4683 NewElts.push_back(Undef);
4685 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewElts);
4686 SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat);
4687 Results.push_back(CvtVec);
4692 // Replace the original node with the legalized result.
4693 if (!Results.empty()) {
4694 DEBUG(dbgs() << "Successfully promoted node\n");
4695 ReplaceNode(Node, Results.data());
4697 DEBUG(dbgs() << "Could not promote node\n");
4700 /// This is the entry point for the file.
4701 void SelectionDAG::Legalize() {
4702 AssignTopologicalOrder();
4704 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4705 // Use a delete listener to remove nodes which were deleted during
4706 // legalization from LegalizeNodes. This is needed to handle the situation
4707 // where a new node is allocated by the object pool to the same address of a
4708 // previously deleted node.
4709 DAGNodeDeletedListener DeleteListener(
4711 [&LegalizedNodes](SDNode *N, SDNode *E) { LegalizedNodes.erase(N); });
4713 SelectionDAGLegalize Legalizer(*this, LegalizedNodes);
4715 // Visit all the nodes. We start in topological order, so that we see
4716 // nodes with their original operands intact. Legalization can produce
4717 // new nodes which may themselves need to be legalized. Iterate until all
4718 // nodes have been legalized.
4720 bool AnyLegalized = false;
4721 for (auto NI = allnodes_end(); NI != allnodes_begin();) {
4725 if (N->use_empty() && N != getRoot().getNode()) {
4731 if (LegalizedNodes.insert(N).second) {
4732 AnyLegalized = true;
4733 Legalizer.LegalizeOp(N);
4735 if (N->use_empty() && N != getRoot().getNode()) {
4746 // Remove dead nodes now.
4750 bool SelectionDAG::LegalizeOp(SDNode *N,
4751 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4752 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4753 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes);
4755 // Directly insert the node in question, and legalize it. This will recurse
4756 // as needed through operands.
4757 LegalizedNodes.insert(N);
4758 Legalizer.LegalizeOp(N);
4760 return LegalizedNodes.count(N);