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/Analysis/DebugInfo.h"
15 #include "llvm/CodeGen/Analysis.h"
16 #include "llvm/CodeGen/MachineFunction.h"
17 #include "llvm/CodeGen/MachineJumpTableInfo.h"
18 #include "llvm/CodeGen/SelectionDAG.h"
19 #include "llvm/Target/TargetFrameLowering.h"
20 #include "llvm/Target/TargetLowering.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/Target/TargetMachine.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/LLVMContext.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/ADT/DenseMap.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/SmallPtrSet.h"
36 //===----------------------------------------------------------------------===//
37 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
38 /// hacks on it until the target machine can handle it. This involves
39 /// eliminating value sizes the machine cannot handle (promoting small sizes to
40 /// large sizes or splitting up large values into small values) as well as
41 /// eliminating operations the machine cannot handle.
43 /// This code also does a small amount of optimization and recognition of idioms
44 /// as part of its processing. For example, if a target does not support a
45 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
46 /// will attempt merge setcc and brc instructions into brcc's.
49 class SelectionDAGLegalize {
50 const TargetMachine &TM;
51 const TargetLowering &TLI;
54 // Libcall insertion helpers.
56 /// LastCALLSEQ - This keeps track of the CALLSEQ_END node that has been
57 /// legalized. We use this to ensure that calls are properly serialized
58 /// against each other, including inserted libcalls.
59 SmallVector<SDValue, 8> LastCALLSEQ;
62 Legal, // The target natively supports this operation.
63 Promote, // This operation should be executed in a larger type.
64 Expand // Try to expand this to other ops, otherwise use a libcall.
67 /// ValueTypeActions - This is a bitvector that contains two bits for each
68 /// value type, where the two bits correspond to the LegalizeAction enum.
69 /// This can be queried with "getTypeAction(VT)".
70 TargetLowering::ValueTypeActionImpl ValueTypeActions;
72 /// LegalizedNodes - For nodes that are of legal width, and that have more
73 /// than one use, this map indicates what regularized operand to use. This
74 /// allows us to avoid legalizing the same thing more than once.
75 DenseMap<SDValue, SDValue> LegalizedNodes;
77 void AddLegalizedOperand(SDValue From, SDValue To) {
78 LegalizedNodes.insert(std::make_pair(From, To));
79 // If someone requests legalization of the new node, return itself.
81 LegalizedNodes.insert(std::make_pair(To, To));
83 // Transfer SDDbgValues.
84 DAG.TransferDbgValues(From, To);
88 explicit SelectionDAGLegalize(SelectionDAG &DAG);
90 /// getTypeAction - Return how we should legalize values of this type, either
91 /// it is already legal or we need to expand it into multiple registers of
92 /// smaller integer type, or we need to promote it to a larger type.
93 LegalizeAction getTypeAction(EVT VT) const {
94 return (LegalizeAction)ValueTypeActions.getTypeAction(VT);
97 /// isTypeLegal - Return true if this type is legal on this target.
99 bool isTypeLegal(EVT VT) const {
100 return getTypeAction(VT) == Legal;
106 /// LegalizeOp - We know that the specified value has a legal type.
107 /// Recursively ensure that the operands have legal types, then return the
109 SDValue LegalizeOp(SDValue O);
111 SDValue OptimizeFloatStore(StoreSDNode *ST);
113 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
114 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
115 /// is necessary to spill the vector being inserted into to memory, perform
116 /// the insert there, and then read the result back.
117 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
118 SDValue Idx, DebugLoc dl);
119 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
120 SDValue Idx, DebugLoc dl);
122 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
123 /// performs the same shuffe in terms of order or result bytes, but on a type
124 /// whose vector element type is narrower than the original shuffle type.
125 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
126 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
127 SDValue N1, SDValue N2,
128 SmallVectorImpl<int> &Mask) const;
130 bool LegalizeAllNodesNotLeadingTo(SDNode *N, SDNode *Dest,
131 SmallPtrSet<SDNode*, 32> &NodesLeadingTo);
133 void LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
136 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
137 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
138 unsigned NumOps, bool isSigned, DebugLoc dl);
140 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
141 SDNode *Node, bool isSigned);
142 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
143 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
144 RTLIB::Libcall Call_PPCF128);
145 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
146 RTLIB::Libcall Call_I8,
147 RTLIB::Libcall Call_I16,
148 RTLIB::Libcall Call_I32,
149 RTLIB::Libcall Call_I64,
150 RTLIB::Libcall Call_I128);
151 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
153 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, DebugLoc dl);
154 SDValue ExpandBUILD_VECTOR(SDNode *Node);
155 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
156 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
157 SmallVectorImpl<SDValue> &Results);
158 SDValue ExpandFCOPYSIGN(SDNode *Node);
159 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
161 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
163 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
166 SDValue ExpandBSWAP(SDValue Op, DebugLoc dl);
167 SDValue ExpandBitCount(unsigned Opc, SDValue Op, DebugLoc dl);
169 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
170 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
171 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
173 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
175 void ExpandNode(SDNode *Node, SmallVectorImpl<SDValue> &Results);
176 void PromoteNode(SDNode *Node, SmallVectorImpl<SDValue> &Results);
178 SDValue getLastCALLSEQ() { return LastCALLSEQ.back(); }
179 void setLastCALLSEQ(const SDValue s) { LastCALLSEQ.back() = s; }
180 void pushLastCALLSEQ(SDValue s) {
181 LastCALLSEQ.push_back(s);
183 void popLastCALLSEQ() {
184 LastCALLSEQ.pop_back();
189 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
190 /// performs the same shuffe in terms of order or result bytes, but on a type
191 /// whose vector element type is narrower than the original shuffle type.
192 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
194 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
195 SDValue N1, SDValue N2,
196 SmallVectorImpl<int> &Mask) const {
197 unsigned NumMaskElts = VT.getVectorNumElements();
198 unsigned NumDestElts = NVT.getVectorNumElements();
199 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
201 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
203 if (NumEltsGrowth == 1)
204 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
206 SmallVector<int, 8> NewMask;
207 for (unsigned i = 0; i != NumMaskElts; ++i) {
209 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
211 NewMask.push_back(-1);
213 NewMask.push_back(Idx * NumEltsGrowth + j);
216 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
217 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
218 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
221 SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
222 : TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
224 ValueTypeActions(TLI.getValueTypeActions()) {
225 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
226 "Too many value types for ValueTypeActions to hold!");
229 void SelectionDAGLegalize::LegalizeDAG() {
230 pushLastCALLSEQ(DAG.getEntryNode());
232 // The legalize process is inherently a bottom-up recursive process (users
233 // legalize their uses before themselves). Given infinite stack space, we
234 // could just start legalizing on the root and traverse the whole graph. In
235 // practice however, this causes us to run out of stack space on large basic
236 // blocks. To avoid this problem, compute an ordering of the nodes where each
237 // node is only legalized after all of its operands are legalized.
238 DAG.AssignTopologicalOrder();
239 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
240 E = prior(DAG.allnodes_end()); I != llvm::next(E); ++I)
241 LegalizeOp(SDValue(I, 0));
243 // Finally, it's possible the root changed. Get the new root.
244 SDValue OldRoot = DAG.getRoot();
245 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
246 DAG.setRoot(LegalizedNodes[OldRoot]);
248 LegalizedNodes.clear();
250 // Remove dead nodes now.
251 DAG.RemoveDeadNodes();
255 /// FindCallEndFromCallStart - Given a chained node that is part of a call
256 /// sequence, find the CALLSEQ_END node that terminates the call sequence.
257 static SDNode *FindCallEndFromCallStart(SDNode *Node, int depth = 0) {
258 int next_depth = depth;
259 if (Node->getOpcode() == ISD::CALLSEQ_START)
260 next_depth = depth + 1;
261 if (Node->getOpcode() == ISD::CALLSEQ_END) {
262 assert(depth > 0 && "negative depth!");
266 next_depth = depth - 1;
268 if (Node->use_empty())
269 return 0; // No CallSeqEnd
271 // The chain is usually at the end.
272 SDValue TheChain(Node, Node->getNumValues()-1);
273 if (TheChain.getValueType() != MVT::Other) {
274 // Sometimes it's at the beginning.
275 TheChain = SDValue(Node, 0);
276 if (TheChain.getValueType() != MVT::Other) {
277 // Otherwise, hunt for it.
278 for (unsigned i = 1, e = Node->getNumValues(); i != e; ++i)
279 if (Node->getValueType(i) == MVT::Other) {
280 TheChain = SDValue(Node, i);
284 // Otherwise, we walked into a node without a chain.
285 if (TheChain.getValueType() != MVT::Other)
290 for (SDNode::use_iterator UI = Node->use_begin(),
291 E = Node->use_end(); UI != E; ++UI) {
293 // Make sure to only follow users of our token chain.
295 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i)
296 if (User->getOperand(i) == TheChain)
297 if (SDNode *Result = FindCallEndFromCallStart(User, next_depth))
303 /// FindCallStartFromCallEnd - Given a chained node that is part of a call
304 /// sequence, find the CALLSEQ_START node that initiates the call sequence.
305 static SDNode *FindCallStartFromCallEnd(SDNode *Node) {
307 assert(Node && "Didn't find callseq_start for a call??");
308 while (Node->getOpcode() != ISD::CALLSEQ_START || nested) {
309 Node = Node->getOperand(0).getNode();
310 assert(Node->getOperand(0).getValueType() == MVT::Other &&
311 "Node doesn't have a token chain argument!");
312 switch (Node->getOpcode()) {
315 case ISD::CALLSEQ_START:
318 Node = Node->getOperand(0).getNode();
321 case ISD::CALLSEQ_END:
326 return (Node->getOpcode() == ISD::CALLSEQ_START) ? Node : 0;
329 /// LegalizeAllNodesNotLeadingTo - Recursively walk the uses of N, looking to
330 /// see if any uses can reach Dest. If no dest operands can get to dest,
331 /// legalize them, legalize ourself, and return false, otherwise, return true.
333 /// Keep track of the nodes we fine that actually do lead to Dest in
334 /// NodesLeadingTo. This avoids retraversing them exponential number of times.
336 bool SelectionDAGLegalize::LegalizeAllNodesNotLeadingTo(SDNode *N, SDNode *Dest,
337 SmallPtrSet<SDNode*, 32> &NodesLeadingTo) {
338 if (N == Dest) return true; // N certainly leads to Dest :)
340 // If we've already processed this node and it does lead to Dest, there is no
341 // need to reprocess it.
342 if (NodesLeadingTo.count(N)) return true;
344 // If the first result of this node has been already legalized, then it cannot
346 if (LegalizedNodes.count(SDValue(N, 0))) return false;
348 // Okay, this node has not already been legalized. Check and legalize all
349 // operands. If none lead to Dest, then we can legalize this node.
350 bool OperandsLeadToDest = false;
351 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
352 OperandsLeadToDest |= // If an operand leads to Dest, so do we.
353 LegalizeAllNodesNotLeadingTo(N->getOperand(i).getNode(), Dest,
356 if (OperandsLeadToDest) {
357 NodesLeadingTo.insert(N);
361 // Okay, this node looks safe, legalize it and return false.
362 LegalizeOp(SDValue(N, 0));
366 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
367 /// a load from the constant pool.
368 static SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP,
369 SelectionDAG &DAG, const TargetLowering &TLI) {
371 DebugLoc dl = CFP->getDebugLoc();
373 // If a FP immediate is precise when represented as a float and if the
374 // target can do an extending load from float to double, we put it into
375 // the constant pool as a float, even if it's is statically typed as a
376 // double. This shrinks FP constants and canonicalizes them for targets where
377 // an FP extending load is the same cost as a normal load (such as on the x87
378 // fp stack or PPC FP unit).
379 EVT VT = CFP->getValueType(0);
380 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
382 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
383 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
384 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
389 while (SVT != MVT::f32) {
390 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
391 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
392 // Only do this if the target has a native EXTLOAD instruction from
394 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
395 TLI.ShouldShrinkFPConstant(OrigVT)) {
396 const Type *SType = SVT.getTypeForEVT(*DAG.getContext());
397 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
403 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
404 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
406 return DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
408 CPIdx, MachinePointerInfo::getConstantPool(),
409 VT, false, false, Alignment);
410 return DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
411 MachinePointerInfo::getConstantPool(), false, false,
415 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
417 SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
418 const TargetLowering &TLI) {
419 SDValue Chain = ST->getChain();
420 SDValue Ptr = ST->getBasePtr();
421 SDValue Val = ST->getValue();
422 EVT VT = Val.getValueType();
423 int Alignment = ST->getAlignment();
424 DebugLoc dl = ST->getDebugLoc();
425 if (ST->getMemoryVT().isFloatingPoint() ||
426 ST->getMemoryVT().isVector()) {
427 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
428 if (TLI.isTypeLegal(intVT)) {
429 // Expand to a bitconvert of the value to the integer type of the
430 // same size, then a (misaligned) int store.
431 // FIXME: Does not handle truncating floating point stores!
432 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
433 return DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
434 ST->isVolatile(), ST->isNonTemporal(), Alignment);
436 // Do a (aligned) store to a stack slot, then copy from the stack slot
437 // to the final destination using (unaligned) integer loads and stores.
438 EVT StoredVT = ST->getMemoryVT();
440 TLI.getRegisterType(*DAG.getContext(),
441 EVT::getIntegerVT(*DAG.getContext(),
442 StoredVT.getSizeInBits()));
443 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
444 unsigned RegBytes = RegVT.getSizeInBits() / 8;
445 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
447 // Make sure the stack slot is also aligned for the register type.
448 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
450 // Perform the original store, only redirected to the stack slot.
451 SDValue Store = DAG.getTruncStore(Chain, dl,
452 Val, StackPtr, MachinePointerInfo(),
453 StoredVT, false, false, 0);
454 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
455 SmallVector<SDValue, 8> Stores;
458 // Do all but one copies using the full register width.
459 for (unsigned i = 1; i < NumRegs; i++) {
460 // Load one integer register's worth from the stack slot.
461 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
462 MachinePointerInfo(),
464 // Store it to the final location. Remember the store.
465 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
466 ST->getPointerInfo().getWithOffset(Offset),
467 ST->isVolatile(), ST->isNonTemporal(),
468 MinAlign(ST->getAlignment(), Offset)));
469 // Increment the pointers.
471 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
473 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
476 // The last store may be partial. Do a truncating store. On big-endian
477 // machines this requires an extending load from the stack slot to ensure
478 // that the bits are in the right place.
479 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
480 8 * (StoredBytes - Offset));
482 // Load from the stack slot.
483 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
484 MachinePointerInfo(),
485 MemVT, false, false, 0);
487 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
489 .getWithOffset(Offset),
490 MemVT, ST->isVolatile(),
492 MinAlign(ST->getAlignment(), Offset)));
493 // The order of the stores doesn't matter - say it with a TokenFactor.
494 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
497 assert(ST->getMemoryVT().isInteger() &&
498 !ST->getMemoryVT().isVector() &&
499 "Unaligned store of unknown type.");
500 // Get the half-size VT
501 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
502 int NumBits = NewStoredVT.getSizeInBits();
503 int IncrementSize = NumBits / 8;
505 // Divide the stored value in two parts.
506 SDValue ShiftAmount = DAG.getConstant(NumBits,
507 TLI.getShiftAmountTy(Val.getValueType()));
509 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
511 // Store the two parts
512 SDValue Store1, Store2;
513 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
514 ST->getPointerInfo(), NewStoredVT,
515 ST->isVolatile(), ST->isNonTemporal(), Alignment);
516 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
517 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
518 Alignment = MinAlign(Alignment, IncrementSize);
519 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
520 ST->getPointerInfo().getWithOffset(IncrementSize),
521 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
524 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
527 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
529 SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
530 const TargetLowering &TLI) {
531 SDValue Chain = LD->getChain();
532 SDValue Ptr = LD->getBasePtr();
533 EVT VT = LD->getValueType(0);
534 EVT LoadedVT = LD->getMemoryVT();
535 DebugLoc dl = LD->getDebugLoc();
536 if (VT.isFloatingPoint() || VT.isVector()) {
537 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
538 if (TLI.isTypeLegal(intVT)) {
539 // Expand to a (misaligned) integer load of the same size,
540 // then bitconvert to floating point or vector.
541 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getPointerInfo(),
543 LD->isNonTemporal(), LD->getAlignment());
544 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
545 if (VT.isFloatingPoint() && LoadedVT != VT)
546 Result = DAG.getNode(ISD::FP_EXTEND, dl, VT, Result);
548 SDValue Ops[] = { Result, Chain };
549 return DAG.getMergeValues(Ops, 2, dl);
552 // Copy the value to a (aligned) stack slot using (unaligned) integer
553 // loads and stores, then do a (aligned) load from the stack slot.
554 EVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
555 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
556 unsigned RegBytes = RegVT.getSizeInBits() / 8;
557 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
559 // Make sure the stack slot is also aligned for the register type.
560 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
562 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
563 SmallVector<SDValue, 8> Stores;
564 SDValue StackPtr = StackBase;
567 // Do all but one copies using the full register width.
568 for (unsigned i = 1; i < NumRegs; i++) {
569 // Load one integer register's worth from the original location.
570 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
571 LD->getPointerInfo().getWithOffset(Offset),
572 LD->isVolatile(), LD->isNonTemporal(),
573 MinAlign(LD->getAlignment(), Offset));
574 // Follow the load with a store to the stack slot. Remember the store.
575 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
576 MachinePointerInfo(), false, false, 0));
577 // Increment the pointers.
579 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
580 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
584 // The last copy may be partial. Do an extending load.
585 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
586 8 * (LoadedBytes - Offset));
587 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
588 LD->getPointerInfo().getWithOffset(Offset),
589 MemVT, LD->isVolatile(),
591 MinAlign(LD->getAlignment(), Offset));
592 // Follow the load with a store to the stack slot. Remember the store.
593 // On big-endian machines this requires a truncating store to ensure
594 // that the bits end up in the right place.
595 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
596 MachinePointerInfo(), MemVT,
599 // The order of the stores doesn't matter - say it with a TokenFactor.
600 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
603 // Finally, perform the original load only redirected to the stack slot.
604 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
605 MachinePointerInfo(), LoadedVT, false, false, 0);
607 // Callers expect a MERGE_VALUES node.
608 SDValue Ops[] = { Load, TF };
609 return DAG.getMergeValues(Ops, 2, dl);
611 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
612 "Unaligned load of unsupported type.");
614 // Compute the new VT that is half the size of the old one. This is an
616 unsigned NumBits = LoadedVT.getSizeInBits();
618 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
621 unsigned Alignment = LD->getAlignment();
622 unsigned IncrementSize = NumBits / 8;
623 ISD::LoadExtType HiExtType = LD->getExtensionType();
625 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
626 if (HiExtType == ISD::NON_EXTLOAD)
627 HiExtType = ISD::ZEXTLOAD;
629 // Load the value in two parts
631 if (TLI.isLittleEndian()) {
632 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
633 NewLoadedVT, LD->isVolatile(),
634 LD->isNonTemporal(), Alignment);
635 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
636 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
637 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
638 LD->getPointerInfo().getWithOffset(IncrementSize),
639 NewLoadedVT, LD->isVolatile(),
640 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
642 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
643 NewLoadedVT, LD->isVolatile(),
644 LD->isNonTemporal(), Alignment);
645 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
646 DAG.getConstant(IncrementSize, TLI.getPointerTy()));
647 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
648 LD->getPointerInfo().getWithOffset(IncrementSize),
649 NewLoadedVT, LD->isVolatile(),
650 LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
653 // aggregate the two parts
654 SDValue ShiftAmount = DAG.getConstant(NumBits,
655 TLI.getShiftAmountTy(Hi.getValueType()));
656 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
657 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
659 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
662 SDValue Ops[] = { Result, TF };
663 return DAG.getMergeValues(Ops, 2, dl);
666 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
667 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
668 /// is necessary to spill the vector being inserted into to memory, perform
669 /// the insert there, and then read the result back.
670 SDValue SelectionDAGLegalize::
671 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
677 // If the target doesn't support this, we have to spill the input vector
678 // to a temporary stack slot, update the element, then reload it. This is
679 // badness. We could also load the value into a vector register (either
680 // with a "move to register" or "extload into register" instruction, then
681 // permute it into place, if the idx is a constant and if the idx is
682 // supported by the target.
683 EVT VT = Tmp1.getValueType();
684 EVT EltVT = VT.getVectorElementType();
685 EVT IdxVT = Tmp3.getValueType();
686 EVT PtrVT = TLI.getPointerTy();
687 SDValue StackPtr = DAG.CreateStackTemporary(VT);
689 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
692 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
693 MachinePointerInfo::getFixedStack(SPFI),
696 // Truncate or zero extend offset to target pointer type.
697 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
698 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
699 // Add the offset to the index.
700 unsigned EltSize = EltVT.getSizeInBits()/8;
701 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
702 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
703 // Store the scalar value.
704 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
706 // Load the updated vector.
707 return DAG.getLoad(VT, dl, Ch, StackPtr,
708 MachinePointerInfo::getFixedStack(SPFI), false, false, 0);
712 SDValue SelectionDAGLegalize::
713 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, DebugLoc dl) {
714 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
715 // SCALAR_TO_VECTOR requires that the type of the value being inserted
716 // match the element type of the vector being created, except for
717 // integers in which case the inserted value can be over width.
718 EVT EltVT = Vec.getValueType().getVectorElementType();
719 if (Val.getValueType() == EltVT ||
720 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
721 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
722 Vec.getValueType(), Val);
724 unsigned NumElts = Vec.getValueType().getVectorNumElements();
725 // We generate a shuffle of InVec and ScVec, so the shuffle mask
726 // should be 0,1,2,3,4,5... with the appropriate element replaced with
728 SmallVector<int, 8> ShufOps;
729 for (unsigned i = 0; i != NumElts; ++i)
730 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
732 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
736 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
739 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
740 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
741 // FIXME: We shouldn't do this for TargetConstantFP's.
742 // FIXME: move this to the DAG Combiner! Note that we can't regress due
743 // to phase ordering between legalized code and the dag combiner. This
744 // probably means that we need to integrate dag combiner and legalizer
746 // We generally can't do this one for long doubles.
747 SDValue Tmp1 = ST->getChain();
748 SDValue Tmp2 = ST->getBasePtr();
750 unsigned Alignment = ST->getAlignment();
751 bool isVolatile = ST->isVolatile();
752 bool isNonTemporal = ST->isNonTemporal();
753 DebugLoc dl = ST->getDebugLoc();
754 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
755 if (CFP->getValueType(0) == MVT::f32 &&
756 getTypeAction(MVT::i32) == Legal) {
757 Tmp3 = DAG.getConstant(CFP->getValueAPF().
758 bitcastToAPInt().zextOrTrunc(32),
760 return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
761 isVolatile, isNonTemporal, Alignment);
764 if (CFP->getValueType(0) == MVT::f64) {
765 // If this target supports 64-bit registers, do a single 64-bit store.
766 if (getTypeAction(MVT::i64) == Legal) {
767 Tmp3 = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
768 zextOrTrunc(64), MVT::i64);
769 return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
770 isVolatile, isNonTemporal, Alignment);
773 if (getTypeAction(MVT::i32) == Legal && !ST->isVolatile()) {
774 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
775 // stores. If the target supports neither 32- nor 64-bits, this
776 // xform is certainly not worth it.
777 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
778 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
779 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
780 if (TLI.isBigEndian()) std::swap(Lo, Hi);
782 Lo = DAG.getStore(Tmp1, dl, Lo, Tmp2, ST->getPointerInfo(), isVolatile,
783 isNonTemporal, Alignment);
784 Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
785 DAG.getIntPtrConstant(4));
786 Hi = DAG.getStore(Tmp1, dl, Hi, Tmp2,
787 ST->getPointerInfo().getWithOffset(4),
788 isVolatile, isNonTemporal, MinAlign(Alignment, 4U));
790 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
794 return SDValue(0, 0);
797 /// LegalizeOp - We know that the specified value has a legal type, and
798 /// that its operands are legal. Now ensure that the operation itself
799 /// is legal, recursively ensuring that the operands' operations remain
801 SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) {
802 if (Op.getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
805 SDNode *Node = Op.getNode();
806 DebugLoc dl = Node->getDebugLoc();
808 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
809 assert(getTypeAction(Node->getValueType(i)) == Legal &&
810 "Unexpected illegal type!");
812 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
813 assert((isTypeLegal(Node->getOperand(i).getValueType()) ||
814 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
815 "Unexpected illegal type!");
817 // Note that LegalizeOp may be reentered even from single-use nodes, which
818 // means that we always must cache transformed nodes.
819 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
820 if (I != LegalizedNodes.end()) return I->second;
822 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
824 bool isCustom = false;
826 // Figure out the correct action; the way to query this varies by opcode
827 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
828 bool SimpleFinishLegalizing = true;
829 switch (Node->getOpcode()) {
830 case ISD::INTRINSIC_W_CHAIN:
831 case ISD::INTRINSIC_WO_CHAIN:
832 case ISD::INTRINSIC_VOID:
835 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
837 case ISD::SINT_TO_FP:
838 case ISD::UINT_TO_FP:
839 case ISD::EXTRACT_VECTOR_ELT:
840 Action = TLI.getOperationAction(Node->getOpcode(),
841 Node->getOperand(0).getValueType());
843 case ISD::FP_ROUND_INREG:
844 case ISD::SIGN_EXTEND_INREG: {
845 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
846 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
852 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
853 Node->getOpcode() == ISD::SETCC ? 2 : 1;
854 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
855 EVT OpVT = Node->getOperand(CompareOperand).getValueType();
856 ISD::CondCode CCCode =
857 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
858 Action = TLI.getCondCodeAction(CCCode, OpVT);
859 if (Action == TargetLowering::Legal) {
860 if (Node->getOpcode() == ISD::SELECT_CC)
861 Action = TLI.getOperationAction(Node->getOpcode(),
862 Node->getValueType(0));
864 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
870 // FIXME: Model these properly. LOAD and STORE are complicated, and
871 // STORE expects the unlegalized operand in some cases.
872 SimpleFinishLegalizing = false;
874 case ISD::CALLSEQ_START:
875 case ISD::CALLSEQ_END:
876 // FIXME: This shouldn't be necessary. These nodes have special properties
877 // dealing with the recursive nature of legalization. Removing this
878 // special case should be done as part of making LegalizeDAG non-recursive.
879 SimpleFinishLegalizing = false;
881 case ISD::EXTRACT_ELEMENT:
882 case ISD::FLT_ROUNDS_:
890 case ISD::MERGE_VALUES:
892 case ISD::FRAME_TO_ARGS_OFFSET:
893 case ISD::EH_SJLJ_SETJMP:
894 case ISD::EH_SJLJ_LONGJMP:
895 case ISD::EH_SJLJ_DISPATCHSETUP:
896 // These operations lie about being legal: when they claim to be legal,
897 // they should actually be expanded.
898 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
899 if (Action == TargetLowering::Legal)
900 Action = TargetLowering::Expand;
902 case ISD::TRAMPOLINE:
904 case ISD::RETURNADDR:
905 // These operations lie about being legal: when they claim to be legal,
906 // they should actually be custom-lowered.
907 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
908 if (Action == TargetLowering::Legal)
909 Action = TargetLowering::Custom;
911 case ISD::BUILD_VECTOR:
912 // A weird case: legalization for BUILD_VECTOR never legalizes the
914 // FIXME: This really sucks... changing it isn't semantically incorrect,
915 // but it massively pessimizes the code for floating-point BUILD_VECTORs
916 // because ConstantFP operands get legalized into constant pool loads
917 // before the BUILD_VECTOR code can see them. It doesn't usually bite,
918 // though, because BUILD_VECTORS usually get lowered into other nodes
919 // which get legalized properly.
920 SimpleFinishLegalizing = false;
923 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
924 Action = TargetLowering::Legal;
926 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
931 if (SimpleFinishLegalizing) {
932 SmallVector<SDValue, 8> Ops, ResultVals;
933 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
934 Ops.push_back(LegalizeOp(Node->getOperand(i)));
935 switch (Node->getOpcode()) {
942 assert(LastCALLSEQ.size() == 1 && "branch inside CALLSEQ_BEGIN/END?");
943 // Branches tweak the chain to include LastCALLSEQ
944 Ops[0] = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ops[0],
946 Ops[0] = LegalizeOp(Ops[0]);
947 setLastCALLSEQ(DAG.getEntryNode());
954 // Legalizing shifts/rotates requires adjusting the shift amount
955 // to the appropriate width.
956 if (!Ops[1].getValueType().isVector())
957 Ops[1] = LegalizeOp(DAG.getShiftAmountOperand(Ops[0].getValueType(),
963 // Legalizing shifts/rotates requires adjusting the shift amount
964 // to the appropriate width.
965 if (!Ops[2].getValueType().isVector())
966 Ops[2] = LegalizeOp(DAG.getShiftAmountOperand(Ops[0].getValueType(),
971 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(), Ops.data(),
974 case TargetLowering::Legal:
975 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
976 ResultVals.push_back(Result.getValue(i));
978 case TargetLowering::Custom:
979 // FIXME: The handling for custom lowering with multiple results is
981 Tmp1 = TLI.LowerOperation(Result, DAG);
982 if (Tmp1.getNode()) {
983 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
985 ResultVals.push_back(Tmp1);
987 ResultVals.push_back(Tmp1.getValue(i));
993 case TargetLowering::Expand:
994 ExpandNode(Result.getNode(), ResultVals);
996 case TargetLowering::Promote:
997 PromoteNode(Result.getNode(), ResultVals);
1000 if (!ResultVals.empty()) {
1001 for (unsigned i = 0, e = ResultVals.size(); i != e; ++i) {
1002 if (ResultVals[i] != SDValue(Node, i))
1003 ResultVals[i] = LegalizeOp(ResultVals[i]);
1004 AddLegalizedOperand(SDValue(Node, i), ResultVals[i]);
1006 return ResultVals[Op.getResNo()];
1010 switch (Node->getOpcode()) {
1017 assert(0 && "Do not know how to legalize this operator!");
1019 case ISD::BUILD_VECTOR:
1020 switch (TLI.getOperationAction(ISD::BUILD_VECTOR, Node->getValueType(0))) {
1021 default: assert(0 && "This action is not supported yet!");
1022 case TargetLowering::Custom:
1023 Tmp3 = TLI.LowerOperation(Result, DAG);
1024 if (Tmp3.getNode()) {
1029 case TargetLowering::Expand:
1030 Result = ExpandBUILD_VECTOR(Result.getNode());
1034 case ISD::CALLSEQ_START: {
1035 SDNode *CallEnd = FindCallEndFromCallStart(Node);
1036 assert(CallEnd && "didn't find CALLSEQ_END!");
1038 // Recursively Legalize all of the inputs of the call end that do not lead
1039 // to this call start. This ensures that any libcalls that need be inserted
1040 // are inserted *before* the CALLSEQ_START.
1041 {SmallPtrSet<SDNode*, 32> NodesLeadingTo;
1042 for (unsigned i = 0, e = CallEnd->getNumOperands(); i != e; ++i)
1043 LegalizeAllNodesNotLeadingTo(CallEnd->getOperand(i).getNode(), Node,
1047 // Now that we have legalized all of the inputs (which may have inserted
1048 // libcalls), create the new CALLSEQ_START node.
1049 Tmp1 = LegalizeOp(Node->getOperand(0)); // Legalize the chain.
1051 // Merge in the last call to ensure that this call starts after the last
1053 if (getLastCALLSEQ().getOpcode() != ISD::EntryToken) {
1054 Tmp1 = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1055 Tmp1, getLastCALLSEQ());
1056 Tmp1 = LegalizeOp(Tmp1);
1059 // Do not try to legalize the target-specific arguments (#1+).
1060 if (Tmp1 != Node->getOperand(0)) {
1061 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1063 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(), &Ops[0],
1064 Ops.size()), Result.getResNo());
1067 // Remember that the CALLSEQ_START is legalized.
1068 AddLegalizedOperand(Op.getValue(0), Result);
1069 if (Node->getNumValues() == 2) // If this has a flag result, remember it.
1070 AddLegalizedOperand(Op.getValue(1), Result.getValue(1));
1072 // Now that the callseq_start and all of the non-call nodes above this call
1073 // sequence have been legalized, legalize the call itself. During this
1074 // process, no libcalls can/will be inserted, guaranteeing that no calls
1076 // Note that we are selecting this call!
1077 setLastCALLSEQ(SDValue(CallEnd, 0));
1079 // Legalize the call, starting from the CALLSEQ_END.
1080 LegalizeOp(getLastCALLSEQ());
1083 case ISD::CALLSEQ_END:
1085 SDNode *myCALLSEQ_BEGIN = FindCallStartFromCallEnd(Node);
1087 // If the CALLSEQ_START node hasn't been legalized first, legalize it.
1088 // This will cause this node to be legalized as well as handling libcalls
1090 if (getLastCALLSEQ().getNode() != Node) {
1091 LegalizeOp(SDValue(myCALLSEQ_BEGIN, 0));
1092 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
1093 assert(I != LegalizedNodes.end() &&
1094 "Legalizing the call start should have legalized this node!");
1098 pushLastCALLSEQ(SDValue(myCALLSEQ_BEGIN, 0));
1101 // Otherwise, the call start has been legalized and everything is going
1102 // according to plan. Just legalize ourselves normally here.
1103 Tmp1 = LegalizeOp(Node->getOperand(0)); // Legalize the chain.
1104 // Do not try to legalize the target-specific arguments (#1+), except for
1105 // an optional flag input.
1106 if (Node->getOperand(Node->getNumOperands()-1).getValueType() != MVT::Glue){
1107 if (Tmp1 != Node->getOperand(0)) {
1108 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1110 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
1111 &Ops[0], Ops.size()),
1115 Tmp2 = LegalizeOp(Node->getOperand(Node->getNumOperands()-1));
1116 if (Tmp1 != Node->getOperand(0) ||
1117 Tmp2 != Node->getOperand(Node->getNumOperands()-1)) {
1118 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1121 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
1122 &Ops[0], Ops.size()),
1126 // This finishes up call legalization.
1129 // If the CALLSEQ_END node has a flag, remember that we legalized it.
1130 AddLegalizedOperand(SDValue(Node, 0), Result.getValue(0));
1131 if (Node->getNumValues() == 2)
1132 AddLegalizedOperand(SDValue(Node, 1), Result.getValue(1));
1133 return Result.getValue(Op.getResNo());
1135 LoadSDNode *LD = cast<LoadSDNode>(Node);
1136 Tmp1 = LegalizeOp(LD->getChain()); // Legalize the chain.
1137 Tmp2 = LegalizeOp(LD->getBasePtr()); // Legalize the base pointer.
1139 ISD::LoadExtType ExtType = LD->getExtensionType();
1140 if (ExtType == ISD::NON_EXTLOAD) {
1141 EVT VT = Node->getValueType(0);
1142 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
1143 Tmp1, Tmp2, LD->getOffset()),
1145 Tmp3 = Result.getValue(0);
1146 Tmp4 = Result.getValue(1);
1148 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
1149 default: assert(0 && "This action is not supported yet!");
1150 case TargetLowering::Legal:
1151 // If this is an unaligned load and the target doesn't support it,
1153 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
1154 const Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1155 unsigned ABIAlignment = TLI.getTargetData()->getABITypeAlignment(Ty);
1156 if (LD->getAlignment() < ABIAlignment){
1157 Result = ExpandUnalignedLoad(cast<LoadSDNode>(Result.getNode()),
1159 Tmp3 = Result.getOperand(0);
1160 Tmp4 = Result.getOperand(1);
1161 Tmp3 = LegalizeOp(Tmp3);
1162 Tmp4 = LegalizeOp(Tmp4);
1166 case TargetLowering::Custom:
1167 Tmp1 = TLI.LowerOperation(Tmp3, DAG);
1168 if (Tmp1.getNode()) {
1169 Tmp3 = LegalizeOp(Tmp1);
1170 Tmp4 = LegalizeOp(Tmp1.getValue(1));
1173 case TargetLowering::Promote: {
1174 // Only promote a load of vector type to another.
1175 assert(VT.isVector() && "Cannot promote this load!");
1176 // Change base type to a different vector type.
1177 EVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
1179 Tmp1 = DAG.getLoad(NVT, dl, Tmp1, Tmp2, LD->getPointerInfo(),
1180 LD->isVolatile(), LD->isNonTemporal(),
1181 LD->getAlignment());
1182 Tmp3 = LegalizeOp(DAG.getNode(ISD::BITCAST, dl, VT, Tmp1));
1183 Tmp4 = LegalizeOp(Tmp1.getValue(1));
1187 // Since loads produce two values, make sure to remember that we
1188 // legalized both of them.
1189 AddLegalizedOperand(SDValue(Node, 0), Tmp3);
1190 AddLegalizedOperand(SDValue(Node, 1), Tmp4);
1191 return Op.getResNo() ? Tmp4 : Tmp3;
1194 EVT SrcVT = LD->getMemoryVT();
1195 unsigned SrcWidth = SrcVT.getSizeInBits();
1196 unsigned Alignment = LD->getAlignment();
1197 bool isVolatile = LD->isVolatile();
1198 bool isNonTemporal = LD->isNonTemporal();
1200 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
1201 // Some targets pretend to have an i1 loading operation, and actually
1202 // load an i8. This trick is correct for ZEXTLOAD because the top 7
1203 // bits are guaranteed to be zero; it helps the optimizers understand
1204 // that these bits are zero. It is also useful for EXTLOAD, since it
1205 // tells the optimizers that those bits are undefined. It would be
1206 // nice to have an effective generic way of getting these benefits...
1207 // Until such a way is found, don't insist on promoting i1 here.
1208 (SrcVT != MVT::i1 ||
1209 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
1210 // Promote to a byte-sized load if not loading an integral number of
1211 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
1212 unsigned NewWidth = SrcVT.getStoreSizeInBits();
1213 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
1216 // The extra bits are guaranteed to be zero, since we stored them that
1217 // way. A zext load from NVT thus automatically gives zext from SrcVT.
1219 ISD::LoadExtType NewExtType =
1220 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
1222 Result = DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
1223 Tmp1, Tmp2, LD->getPointerInfo(),
1224 NVT, isVolatile, isNonTemporal, Alignment);
1226 Ch = Result.getValue(1); // The chain.
1228 if (ExtType == ISD::SEXTLOAD)
1229 // Having the top bits zero doesn't help when sign extending.
1230 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1231 Result.getValueType(),
1232 Result, DAG.getValueType(SrcVT));
1233 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
1234 // All the top bits are guaranteed to be zero - inform the optimizers.
1235 Result = DAG.getNode(ISD::AssertZext, dl,
1236 Result.getValueType(), Result,
1237 DAG.getValueType(SrcVT));
1239 Tmp1 = LegalizeOp(Result);
1240 Tmp2 = LegalizeOp(Ch);
1241 } else if (SrcWidth & (SrcWidth - 1)) {
1242 // If not loading a power-of-2 number of bits, expand as two loads.
1243 assert(!SrcVT.isVector() && "Unsupported extload!");
1244 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
1245 assert(RoundWidth < SrcWidth);
1246 unsigned ExtraWidth = SrcWidth - RoundWidth;
1247 assert(ExtraWidth < RoundWidth);
1248 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
1249 "Load size not an integral number of bytes!");
1250 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
1251 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
1253 unsigned IncrementSize;
1255 if (TLI.isLittleEndian()) {
1256 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
1257 // Load the bottom RoundWidth bits.
1258 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1260 LD->getPointerInfo(), RoundVT, isVolatile,
1261 isNonTemporal, Alignment);
1263 // Load the remaining ExtraWidth bits.
1264 IncrementSize = RoundWidth / 8;
1265 Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
1266 DAG.getIntPtrConstant(IncrementSize));
1267 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2,
1268 LD->getPointerInfo().getWithOffset(IncrementSize),
1269 ExtraVT, isVolatile, isNonTemporal,
1270 MinAlign(Alignment, IncrementSize));
1272 // Build a factor node to remember that this load is independent of
1274 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1277 // Move the top bits to the right place.
1278 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1279 DAG.getConstant(RoundWidth,
1280 TLI.getShiftAmountTy(Hi.getValueType())));
1282 // Join the hi and lo parts.
1283 Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1285 // Big endian - avoid unaligned loads.
1286 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1287 // Load the top RoundWidth bits.
1288 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2,
1289 LD->getPointerInfo(), RoundVT, isVolatile,
1290 isNonTemporal, Alignment);
1292 // Load the remaining ExtraWidth bits.
1293 IncrementSize = RoundWidth / 8;
1294 Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
1295 DAG.getIntPtrConstant(IncrementSize));
1296 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1297 dl, Node->getValueType(0), Tmp1, Tmp2,
1298 LD->getPointerInfo().getWithOffset(IncrementSize),
1299 ExtraVT, isVolatile, isNonTemporal,
1300 MinAlign(Alignment, IncrementSize));
1302 // Build a factor node to remember that this load is independent of
1304 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1307 // Move the top bits to the right place.
1308 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1309 DAG.getConstant(ExtraWidth,
1310 TLI.getShiftAmountTy(Hi.getValueType())));
1312 // Join the hi and lo parts.
1313 Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1316 Tmp1 = LegalizeOp(Result);
1317 Tmp2 = LegalizeOp(Ch);
1319 switch (TLI.getLoadExtAction(ExtType, SrcVT)) {
1320 default: assert(0 && "This action is not supported yet!");
1321 case TargetLowering::Custom:
1324 case TargetLowering::Legal:
1325 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
1326 Tmp1, Tmp2, LD->getOffset()),
1328 Tmp1 = Result.getValue(0);
1329 Tmp2 = Result.getValue(1);
1332 Tmp3 = TLI.LowerOperation(Result, DAG);
1333 if (Tmp3.getNode()) {
1334 Tmp1 = LegalizeOp(Tmp3);
1335 Tmp2 = LegalizeOp(Tmp3.getValue(1));
1338 // If this is an unaligned load and the target doesn't support it,
1340 if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
1342 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1343 unsigned ABIAlignment =
1344 TLI.getTargetData()->getABITypeAlignment(Ty);
1345 if (LD->getAlignment() < ABIAlignment){
1346 Result = ExpandUnalignedLoad(cast<LoadSDNode>(Result.getNode()),
1348 Tmp1 = Result.getOperand(0);
1349 Tmp2 = Result.getOperand(1);
1350 Tmp1 = LegalizeOp(Tmp1);
1351 Tmp2 = LegalizeOp(Tmp2);
1356 case TargetLowering::Expand:
1357 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && isTypeLegal(SrcVT)) {
1358 SDValue Load = DAG.getLoad(SrcVT, dl, Tmp1, Tmp2,
1359 LD->getPointerInfo(),
1360 LD->isVolatile(), LD->isNonTemporal(),
1361 LD->getAlignment());
1365 ExtendOp = (SrcVT.isFloatingPoint() ?
1366 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1368 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1369 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1370 default: llvm_unreachable("Unexpected extend load type!");
1372 Result = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1373 Tmp1 = LegalizeOp(Result); // Relegalize new nodes.
1374 Tmp2 = LegalizeOp(Load.getValue(1));
1377 // FIXME: This does not work for vectors on most targets. Sign- and
1378 // zero-extend operations are currently folded into extending loads,
1379 // whether they are legal or not, and then we end up here without any
1380 // support for legalizing them.
1381 assert(ExtType != ISD::EXTLOAD &&
1382 "EXTLOAD should always be supported!");
1383 // Turn the unsupported load into an EXTLOAD followed by an explicit
1384 // zero/sign extend inreg.
1385 Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
1386 Tmp1, Tmp2, LD->getPointerInfo(), SrcVT,
1387 LD->isVolatile(), LD->isNonTemporal(),
1388 LD->getAlignment());
1390 if (ExtType == ISD::SEXTLOAD)
1391 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1392 Result.getValueType(),
1393 Result, DAG.getValueType(SrcVT));
1395 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1396 Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
1397 Tmp2 = LegalizeOp(Result.getValue(1)); // Relegalize new nodes.
1402 // Since loads produce two values, make sure to remember that we legalized
1404 AddLegalizedOperand(SDValue(Node, 0), Tmp1);
1405 AddLegalizedOperand(SDValue(Node, 1), Tmp2);
1406 return Op.getResNo() ? Tmp2 : Tmp1;
1409 StoreSDNode *ST = cast<StoreSDNode>(Node);
1410 Tmp1 = LegalizeOp(ST->getChain()); // Legalize the chain.
1411 Tmp2 = LegalizeOp(ST->getBasePtr()); // Legalize the pointer.
1412 unsigned Alignment = ST->getAlignment();
1413 bool isVolatile = ST->isVolatile();
1414 bool isNonTemporal = ST->isNonTemporal();
1416 if (!ST->isTruncatingStore()) {
1417 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
1418 Result = SDValue(OptStore, 0);
1423 Tmp3 = LegalizeOp(ST->getValue());
1424 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
1429 EVT VT = Tmp3.getValueType();
1430 switch (TLI.getOperationAction(ISD::STORE, VT)) {
1431 default: assert(0 && "This action is not supported yet!");
1432 case TargetLowering::Legal:
1433 // If this is an unaligned store and the target doesn't support it,
1435 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
1436 const Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
1437 unsigned ABIAlignment= TLI.getTargetData()->getABITypeAlignment(Ty);
1438 if (ST->getAlignment() < ABIAlignment)
1439 Result = ExpandUnalignedStore(cast<StoreSDNode>(Result.getNode()),
1443 case TargetLowering::Custom:
1444 Tmp1 = TLI.LowerOperation(Result, DAG);
1445 if (Tmp1.getNode()) Result = Tmp1;
1447 case TargetLowering::Promote:
1448 assert(VT.isVector() && "Unknown legal promote case!");
1449 Tmp3 = DAG.getNode(ISD::BITCAST, dl,
1450 TLI.getTypeToPromoteTo(ISD::STORE, VT), Tmp3);
1451 Result = DAG.getStore(Tmp1, dl, Tmp3, Tmp2,
1452 ST->getPointerInfo(), isVolatile,
1453 isNonTemporal, Alignment);
1459 Tmp3 = LegalizeOp(ST->getValue());
1461 EVT StVT = ST->getMemoryVT();
1462 unsigned StWidth = StVT.getSizeInBits();
1464 if (StWidth != StVT.getStoreSizeInBits()) {
1465 // Promote to a byte-sized store with upper bits zero if not
1466 // storing an integral number of bytes. For example, promote
1467 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
1468 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
1469 StVT.getStoreSizeInBits());
1470 Tmp3 = DAG.getZeroExtendInReg(Tmp3, dl, StVT);
1471 Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
1472 NVT, isVolatile, isNonTemporal, Alignment);
1473 } else if (StWidth & (StWidth - 1)) {
1474 // If not storing a power-of-2 number of bits, expand as two stores.
1475 assert(!StVT.isVector() && "Unsupported truncstore!");
1476 unsigned RoundWidth = 1 << Log2_32(StWidth);
1477 assert(RoundWidth < StWidth);
1478 unsigned ExtraWidth = StWidth - RoundWidth;
1479 assert(ExtraWidth < RoundWidth);
1480 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
1481 "Store size not an integral number of bytes!");
1482 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
1483 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
1485 unsigned IncrementSize;
1487 if (TLI.isLittleEndian()) {
1488 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
1489 // Store the bottom RoundWidth bits.
1490 Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
1492 isVolatile, isNonTemporal, Alignment);
1494 // Store the remaining ExtraWidth bits.
1495 IncrementSize = RoundWidth / 8;
1496 Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
1497 DAG.getIntPtrConstant(IncrementSize));
1498 Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3,
1499 DAG.getConstant(RoundWidth,
1500 TLI.getShiftAmountTy(Tmp3.getValueType())));
1501 Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2,
1502 ST->getPointerInfo().getWithOffset(IncrementSize),
1503 ExtraVT, isVolatile, isNonTemporal,
1504 MinAlign(Alignment, IncrementSize));
1506 // Big endian - avoid unaligned stores.
1507 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
1508 // Store the top RoundWidth bits.
1509 Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3,
1510 DAG.getConstant(ExtraWidth,
1511 TLI.getShiftAmountTy(Tmp3.getValueType())));
1512 Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2, ST->getPointerInfo(),
1513 RoundVT, isVolatile, isNonTemporal, Alignment);
1515 // Store the remaining ExtraWidth bits.
1516 IncrementSize = RoundWidth / 8;
1517 Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
1518 DAG.getIntPtrConstant(IncrementSize));
1519 Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
1520 ST->getPointerInfo().getWithOffset(IncrementSize),
1521 ExtraVT, isVolatile, isNonTemporal,
1522 MinAlign(Alignment, IncrementSize));
1525 // The order of the stores doesn't matter.
1526 Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
1528 if (Tmp1 != ST->getChain() || Tmp3 != ST->getValue() ||
1529 Tmp2 != ST->getBasePtr())
1530 Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
1535 switch (TLI.getTruncStoreAction(ST->getValue().getValueType(), StVT)) {
1536 default: assert(0 && "This action is not supported yet!");
1537 case TargetLowering::Legal:
1538 // If this is an unaligned store and the target doesn't support it,
1540 if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
1541 const Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
1542 unsigned ABIAlignment= TLI.getTargetData()->getABITypeAlignment(Ty);
1543 if (ST->getAlignment() < ABIAlignment)
1544 Result = ExpandUnalignedStore(cast<StoreSDNode>(Result.getNode()),
1548 case TargetLowering::Custom:
1549 Result = TLI.LowerOperation(Result, DAG);
1552 // TRUNCSTORE:i16 i32 -> STORE i16
1553 assert(isTypeLegal(StVT) && "Do not know how to expand this store!");
1554 Tmp3 = DAG.getNode(ISD::TRUNCATE, dl, StVT, Tmp3);
1555 Result = DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
1556 isVolatile, isNonTemporal, Alignment);
1564 assert(Result.getValueType() == Op.getValueType() &&
1565 "Bad legalization!");
1567 // Make sure that the generated code is itself legal.
1569 Result = LegalizeOp(Result);
1571 // Note that LegalizeOp may be reentered even from single-use nodes, which
1572 // means that we always must cache transformed nodes.
1573 AddLegalizedOperand(Op, Result);
1577 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1578 SDValue Vec = Op.getOperand(0);
1579 SDValue Idx = Op.getOperand(1);
1580 DebugLoc dl = Op.getDebugLoc();
1581 // Store the value to a temporary stack slot, then LOAD the returned part.
1582 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1583 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1584 MachinePointerInfo(), false, false, 0);
1586 // Add the offset to the index.
1588 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1589 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1590 DAG.getConstant(EltSize, Idx.getValueType()));
1592 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1593 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1595 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1597 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1599 if (Op.getValueType().isVector())
1600 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1602 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1603 MachinePointerInfo(),
1604 Vec.getValueType().getVectorElementType(),
1608 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1609 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1611 SDValue Vec = Op.getOperand(0);
1612 SDValue Part = Op.getOperand(1);
1613 SDValue Idx = Op.getOperand(2);
1614 DebugLoc dl = Op.getDebugLoc();
1616 // Store the value to a temporary stack slot, then LOAD the returned part.
1618 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1619 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1620 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1622 // First store the whole vector.
1623 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1626 // Then store the inserted part.
1628 // Add the offset to the index.
1630 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1632 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1633 DAG.getConstant(EltSize, Idx.getValueType()));
1635 if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
1636 Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
1638 Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
1640 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1643 // Store the subvector.
1644 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1645 MachinePointerInfo(), false, false, 0);
1647 // Finally, load the updated vector.
1648 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1652 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1653 // We can't handle this case efficiently. Allocate a sufficiently
1654 // aligned object on the stack, store each element into it, then load
1655 // the result as a vector.
1656 // Create the stack frame object.
1657 EVT VT = Node->getValueType(0);
1658 EVT EltVT = VT.getVectorElementType();
1659 DebugLoc dl = Node->getDebugLoc();
1660 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1661 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1662 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1664 // Emit a store of each element to the stack slot.
1665 SmallVector<SDValue, 8> Stores;
1666 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1667 // Store (in the right endianness) the elements to memory.
1668 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1669 // Ignore undef elements.
1670 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1672 unsigned Offset = TypeByteSize*i;
1674 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1675 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1677 // If the destination vector element type is narrower than the source
1678 // element type, only store the bits necessary.
1679 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1680 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1681 Node->getOperand(i), Idx,
1682 PtrInfo.getWithOffset(Offset),
1683 EltVT, false, false, 0));
1685 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1686 Node->getOperand(i), Idx,
1687 PtrInfo.getWithOffset(Offset),
1692 if (!Stores.empty()) // Not all undef elements?
1693 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1694 &Stores[0], Stores.size());
1696 StoreChain = DAG.getEntryNode();
1698 // Result is a load from the stack slot.
1699 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo, false, false, 0);
1702 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1703 DebugLoc dl = Node->getDebugLoc();
1704 SDValue Tmp1 = Node->getOperand(0);
1705 SDValue Tmp2 = Node->getOperand(1);
1707 // Get the sign bit of the RHS. First obtain a value that has the same
1708 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1710 EVT FloatVT = Tmp2.getValueType();
1711 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1712 if (isTypeLegal(IVT)) {
1713 // Convert to an integer with the same sign bit.
1714 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1716 // Store the float to memory, then load the sign part out as an integer.
1717 MVT LoadTy = TLI.getPointerTy();
1718 // First create a temporary that is aligned for both the load and store.
1719 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1720 // Then store the float to it.
1722 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1724 if (TLI.isBigEndian()) {
1725 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1726 // Load out a legal integer with the same sign bit as the float.
1727 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1729 } else { // Little endian
1730 SDValue LoadPtr = StackPtr;
1731 // The float may be wider than the integer we are going to load. Advance
1732 // the pointer so that the loaded integer will contain the sign bit.
1733 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1734 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1735 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(),
1736 LoadPtr, DAG.getIntPtrConstant(ByteOffset));
1737 // Load a legal integer containing the sign bit.
1738 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1740 // Move the sign bit to the top bit of the loaded integer.
1741 unsigned BitShift = LoadTy.getSizeInBits() -
1742 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1743 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1745 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1746 DAG.getConstant(BitShift,
1747 TLI.getShiftAmountTy(SignBit.getValueType())));
1750 // Now get the sign bit proper, by seeing whether the value is negative.
1751 SignBit = DAG.getSetCC(dl, TLI.getSetCCResultType(SignBit.getValueType()),
1752 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1754 // Get the absolute value of the result.
1755 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1756 // Select between the nabs and abs value based on the sign bit of
1758 return DAG.getNode(ISD::SELECT, dl, AbsVal.getValueType(), SignBit,
1759 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1763 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1764 SmallVectorImpl<SDValue> &Results) {
1765 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1766 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1767 " not tell us which reg is the stack pointer!");
1768 DebugLoc dl = Node->getDebugLoc();
1769 EVT VT = Node->getValueType(0);
1770 SDValue Tmp1 = SDValue(Node, 0);
1771 SDValue Tmp2 = SDValue(Node, 1);
1772 SDValue Tmp3 = Node->getOperand(2);
1773 SDValue Chain = Tmp1.getOperand(0);
1775 // Chain the dynamic stack allocation so that it doesn't modify the stack
1776 // pointer when other instructions are using the stack.
1777 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1779 SDValue Size = Tmp2.getOperand(1);
1780 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1781 Chain = SP.getValue(1);
1782 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1783 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1784 if (Align > StackAlign)
1785 SP = DAG.getNode(ISD::AND, dl, VT, SP,
1786 DAG.getConstant(-(uint64_t)Align, VT));
1787 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1788 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1790 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1791 DAG.getIntPtrConstant(0, true), SDValue());
1793 Results.push_back(Tmp1);
1794 Results.push_back(Tmp2);
1797 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1798 /// condition code CC on the current target. This routine expands SETCC with
1799 /// illegal condition code into AND / OR of multiple SETCC values.
1800 void SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1801 SDValue &LHS, SDValue &RHS,
1804 EVT OpVT = LHS.getValueType();
1805 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1806 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1807 default: assert(0 && "Unknown condition code action!");
1808 case TargetLowering::Legal:
1811 case TargetLowering::Expand: {
1812 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1815 default: assert(0 && "Don't know how to expand this condition!");
1816 case ISD::SETOEQ: CC1 = ISD::SETEQ; CC2 = ISD::SETO; Opc = ISD::AND; break;
1817 case ISD::SETOGT: CC1 = ISD::SETGT; CC2 = ISD::SETO; Opc = ISD::AND; break;
1818 case ISD::SETOGE: CC1 = ISD::SETGE; CC2 = ISD::SETO; Opc = ISD::AND; break;
1819 case ISD::SETOLT: CC1 = ISD::SETLT; CC2 = ISD::SETO; Opc = ISD::AND; break;
1820 case ISD::SETOLE: CC1 = ISD::SETLE; CC2 = ISD::SETO; Opc = ISD::AND; break;
1821 case ISD::SETONE: CC1 = ISD::SETNE; CC2 = ISD::SETO; Opc = ISD::AND; break;
1822 case ISD::SETUEQ: CC1 = ISD::SETEQ; CC2 = ISD::SETUO; Opc = ISD::OR; break;
1823 case ISD::SETUGT: CC1 = ISD::SETGT; CC2 = ISD::SETUO; Opc = ISD::OR; break;
1824 case ISD::SETUGE: CC1 = ISD::SETGE; CC2 = ISD::SETUO; Opc = ISD::OR; break;
1825 case ISD::SETULT: CC1 = ISD::SETLT; CC2 = ISD::SETUO; Opc = ISD::OR; break;
1826 case ISD::SETULE: CC1 = ISD::SETLE; CC2 = ISD::SETUO; Opc = ISD::OR; break;
1827 case ISD::SETUNE: CC1 = ISD::SETNE; CC2 = ISD::SETUO; Opc = ISD::OR; break;
1828 // FIXME: Implement more expansions.
1831 SDValue SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1832 SDValue SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1833 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1841 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1842 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1843 /// a load from the stack slot to DestVT, extending it if needed.
1844 /// The resultant code need not be legal.
1845 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1849 // Create the stack frame object.
1851 TLI.getTargetData()->getPrefTypeAlignment(SrcOp.getValueType().
1852 getTypeForEVT(*DAG.getContext()));
1853 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1855 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1856 int SPFI = StackPtrFI->getIndex();
1857 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1859 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1860 unsigned SlotSize = SlotVT.getSizeInBits();
1861 unsigned DestSize = DestVT.getSizeInBits();
1862 const Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1863 unsigned DestAlign = TLI.getTargetData()->getPrefTypeAlignment(DestType);
1865 // Emit a store to the stack slot. Use a truncstore if the input value is
1866 // later than DestVT.
1869 if (SrcSize > SlotSize)
1870 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1871 PtrInfo, SlotVT, false, false, SrcAlign);
1873 assert(SrcSize == SlotSize && "Invalid store");
1874 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1875 PtrInfo, false, false, SrcAlign);
1878 // Result is a load from the stack slot.
1879 if (SlotSize == DestSize)
1880 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1881 false, false, DestAlign);
1883 assert(SlotSize < DestSize && "Unknown extension!");
1884 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1885 PtrInfo, SlotVT, false, false, DestAlign);
1888 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1889 DebugLoc dl = Node->getDebugLoc();
1890 // Create a vector sized/aligned stack slot, store the value to element #0,
1891 // then load the whole vector back out.
1892 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1894 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1895 int SPFI = StackPtrFI->getIndex();
1897 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1899 MachinePointerInfo::getFixedStack(SPFI),
1900 Node->getValueType(0).getVectorElementType(),
1902 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1903 MachinePointerInfo::getFixedStack(SPFI),
1908 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1909 /// support the operation, but do support the resultant vector type.
1910 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1911 unsigned NumElems = Node->getNumOperands();
1912 SDValue Value1, Value2;
1913 DebugLoc dl = Node->getDebugLoc();
1914 EVT VT = Node->getValueType(0);
1915 EVT OpVT = Node->getOperand(0).getValueType();
1916 EVT EltVT = VT.getVectorElementType();
1918 // If the only non-undef value is the low element, turn this into a
1919 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1920 bool isOnlyLowElement = true;
1921 bool MoreThanTwoValues = false;
1922 bool isConstant = true;
1923 for (unsigned i = 0; i < NumElems; ++i) {
1924 SDValue V = Node->getOperand(i);
1925 if (V.getOpcode() == ISD::UNDEF)
1928 isOnlyLowElement = false;
1929 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1932 if (!Value1.getNode()) {
1934 } else if (!Value2.getNode()) {
1937 } else if (V != Value1 && V != Value2) {
1938 MoreThanTwoValues = true;
1942 if (!Value1.getNode())
1943 return DAG.getUNDEF(VT);
1945 if (isOnlyLowElement)
1946 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1948 // If all elements are constants, create a load from the constant pool.
1950 std::vector<Constant*> CV;
1951 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1952 if (ConstantFPSDNode *V =
1953 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1954 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1955 } else if (ConstantSDNode *V =
1956 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1958 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1960 // If OpVT and EltVT don't match, EltVT is not legal and the
1961 // element values have been promoted/truncated earlier. Undo this;
1962 // we don't want a v16i8 to become a v16i32 for example.
1963 const ConstantInt *CI = V->getConstantIntValue();
1964 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1965 CI->getZExtValue()));
1968 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1969 const Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1970 CV.push_back(UndefValue::get(OpNTy));
1973 Constant *CP = ConstantVector::get(CV);
1974 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1975 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1976 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1977 MachinePointerInfo::getConstantPool(),
1978 false, false, Alignment);
1981 if (!MoreThanTwoValues) {
1982 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1983 for (unsigned i = 0; i < NumElems; ++i) {
1984 SDValue V = Node->getOperand(i);
1985 if (V.getOpcode() == ISD::UNDEF)
1987 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
1989 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
1990 // Get the splatted value into the low element of a vector register.
1991 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
1993 if (Value2.getNode())
1994 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
1996 Vec2 = DAG.getUNDEF(VT);
1998 // Return shuffle(LowValVec, undef, <0,0,0,0>)
1999 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2003 // Otherwise, we can't handle this case efficiently.
2004 return ExpandVectorBuildThroughStack(Node);
2007 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
2008 // does not fit into a register, return the lo part and set the hi part to the
2009 // by-reg argument. If it does fit into a single register, return the result
2010 // and leave the Hi part unset.
2011 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2013 // The input chain to this libcall is the entry node of the function.
2014 // Legalizing the call will automatically add the previous call to the
2016 SDValue InChain = DAG.getEntryNode();
2018 TargetLowering::ArgListTy Args;
2019 TargetLowering::ArgListEntry Entry;
2020 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2021 EVT ArgVT = Node->getOperand(i).getValueType();
2022 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2023 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2024 Entry.isSExt = isSigned;
2025 Entry.isZExt = !isSigned;
2026 Args.push_back(Entry);
2028 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2029 TLI.getPointerTy());
2031 // Splice the libcall in wherever FindInputOutputChains tells us to.
2032 const Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2034 // isTailCall may be true since the callee does not reference caller stack
2035 // frame. Check if it's in the right position.
2036 bool isTailCall = isInTailCallPosition(DAG, Node, TLI);
2037 std::pair<SDValue, SDValue> CallInfo =
2038 TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
2039 0, TLI.getLibcallCallingConv(LC), isTailCall,
2040 /*isReturnValueUsed=*/true,
2041 Callee, Args, DAG, Node->getDebugLoc());
2043 if (!CallInfo.second.getNode())
2044 // It's a tailcall, return the chain (which is the DAG root).
2045 return DAG.getRoot();
2047 // Legalize the call sequence, starting with the chain. This will advance
2048 // the LastCALLSEQ to the legalized version of the CALLSEQ_END node that
2049 // was added by LowerCallTo (guaranteeing proper serialization of calls).
2050 LegalizeOp(CallInfo.second);
2051 return CallInfo.first;
2054 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
2055 /// and returning a result of type RetVT.
2056 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2057 const SDValue *Ops, unsigned NumOps,
2058 bool isSigned, DebugLoc dl) {
2059 TargetLowering::ArgListTy Args;
2060 Args.reserve(NumOps);
2062 TargetLowering::ArgListEntry Entry;
2063 for (unsigned i = 0; i != NumOps; ++i) {
2064 Entry.Node = Ops[i];
2065 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2066 Entry.isSExt = isSigned;
2067 Entry.isZExt = !isSigned;
2068 Args.push_back(Entry);
2070 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2071 TLI.getPointerTy());
2073 const Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2074 std::pair<SDValue,SDValue> CallInfo =
2075 TLI.LowerCallTo(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
2076 false, 0, TLI.getLibcallCallingConv(LC), false,
2077 /*isReturnValueUsed=*/true,
2078 Callee, Args, DAG, dl);
2080 // Legalize the call sequence, starting with the chain. This will advance
2081 // the LastCALLSEQ_END to the legalized version of the CALLSEQ_END node that
2082 // was added by LowerCallTo (guaranteeing proper serialization of calls).
2083 LegalizeOp(CallInfo.second);
2085 return CallInfo.first;
2088 // ExpandChainLibCall - 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 const 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());
2110 // Splice the libcall in wherever FindInputOutputChains tells us to.
2111 const Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2112 std::pair<SDValue, SDValue> CallInfo =
2113 TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
2114 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2115 /*isReturnValueUsed=*/true,
2116 Callee, Args, DAG, Node->getDebugLoc());
2118 // Legalize the call sequence, starting with the chain. This will advance
2119 // the LastCALLSEQ to the legalized version of the CALLSEQ_END node that
2120 // was added by LowerCallTo (guaranteeing proper serialization of calls).
2121 LegalizeOp(CallInfo.second);
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_PPCF128) {
2131 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2132 default: assert(0 && "Unexpected request for libcall!");
2133 case MVT::f32: LC = Call_F32; break;
2134 case MVT::f64: LC = Call_F64; break;
2135 case MVT::f80: LC = Call_F80; break;
2136 case MVT::ppcf128: LC = Call_PPCF128; break;
2138 return ExpandLibCall(LC, Node, false);
2141 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2142 RTLIB::Libcall Call_I8,
2143 RTLIB::Libcall Call_I16,
2144 RTLIB::Libcall Call_I32,
2145 RTLIB::Libcall Call_I64,
2146 RTLIB::Libcall Call_I128) {
2148 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2149 default: assert(0 && "Unexpected request for libcall!");
2150 case MVT::i8: LC = Call_I8; break;
2151 case MVT::i16: LC = Call_I16; break;
2152 case MVT::i32: LC = Call_I32; break;
2153 case MVT::i64: LC = Call_I64; break;
2154 case MVT::i128: LC = Call_I128; break;
2156 return ExpandLibCall(LC, Node, isSigned);
2159 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
2160 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2161 const TargetLowering &TLI) {
2163 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2164 default: assert(0 && "Unexpected request for libcall!");
2165 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2166 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2167 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2168 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2169 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2172 return TLI.getLibcallName(LC) != 0;
2175 /// UseDivRem - Only issue divrem libcall if both quotient and remainder are
2177 static bool UseDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2178 unsigned OtherOpcode = 0;
2180 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2182 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2184 SDValue Op0 = Node->getOperand(0);
2185 SDValue Op1 = Node->getOperand(1);
2186 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2187 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2191 if (User->getOpcode() == OtherOpcode &&
2192 User->getOperand(0) == Op0 &&
2193 User->getOperand(1) == Op1)
2199 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2202 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2203 SmallVectorImpl<SDValue> &Results) {
2204 unsigned Opcode = Node->getOpcode();
2205 bool isSigned = Opcode == ISD::SDIVREM;
2208 switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
2209 default: assert(0 && "Unexpected request for libcall!");
2210 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2211 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2212 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2213 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2214 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2217 // The input chain to this libcall is the entry node of the function.
2218 // Legalizing the call will automatically add the previous call to the
2220 SDValue InChain = DAG.getEntryNode();
2222 EVT RetVT = Node->getValueType(0);
2223 const Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2225 TargetLowering::ArgListTy Args;
2226 TargetLowering::ArgListEntry Entry;
2227 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2228 EVT ArgVT = Node->getOperand(i).getValueType();
2229 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2230 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2231 Entry.isSExt = isSigned;
2232 Entry.isZExt = !isSigned;
2233 Args.push_back(Entry);
2236 // Also pass the return address of the remainder.
2237 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2239 Entry.Ty = RetTy->getPointerTo();
2240 Entry.isSExt = isSigned;
2241 Entry.isZExt = !isSigned;
2242 Args.push_back(Entry);
2244 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2245 TLI.getPointerTy());
2247 // Splice the libcall in wherever FindInputOutputChains tells us to.
2248 DebugLoc dl = Node->getDebugLoc();
2249 std::pair<SDValue, SDValue> CallInfo =
2250 TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
2251 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
2252 /*isReturnValueUsed=*/true, Callee, Args, DAG, dl);
2254 // Legalize the call sequence, starting with the chain. This will advance
2255 // the LastCALLSEQ to the legalized version of the CALLSEQ_END node that
2256 // was added by LowerCallTo (guaranteeing proper serialization of calls).
2257 LegalizeOp(CallInfo.second);
2259 // Remainder is loaded back from the stack frame.
2260 SDValue Rem = DAG.getLoad(RetVT, dl, getLastCALLSEQ(), FIPtr,
2261 MachinePointerInfo(), false, false, 0);
2262 Results.push_back(CallInfo.first);
2263 Results.push_back(Rem);
2266 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2267 /// INT_TO_FP operation of the specified operand when the target requests that
2268 /// we expand it. At this point, we know that the result and operand types are
2269 /// legal for the target.
2270 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2274 if (Op0.getValueType() == MVT::i32) {
2275 // simple 32-bit [signed|unsigned] integer to float/double expansion
2277 // Get the stack frame index of a 8 byte buffer.
2278 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2280 // word offset constant for Hi/Lo address computation
2281 SDValue WordOff = DAG.getConstant(sizeof(int), TLI.getPointerTy());
2282 // set up Hi and Lo (into buffer) address based on endian
2283 SDValue Hi = StackSlot;
2284 SDValue Lo = DAG.getNode(ISD::ADD, dl,
2285 TLI.getPointerTy(), StackSlot, WordOff);
2286 if (TLI.isLittleEndian())
2289 // if signed map to unsigned space
2292 // constant used to invert sign bit (signed to unsigned mapping)
2293 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2294 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2298 // store the lo of the constructed double - based on integer input
2299 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2300 Op0Mapped, Lo, MachinePointerInfo(),
2302 // initial hi portion of constructed double
2303 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2304 // store the hi of the constructed double - biased exponent
2305 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2306 MachinePointerInfo(),
2308 // load the constructed double
2309 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2310 MachinePointerInfo(), false, false, 0);
2311 // FP constant to bias correct the final result
2312 SDValue Bias = DAG.getConstantFP(isSigned ?
2313 BitsToDouble(0x4330000080000000ULL) :
2314 BitsToDouble(0x4330000000000000ULL),
2316 // subtract the bias
2317 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2320 // handle final rounding
2321 if (DestVT == MVT::f64) {
2324 } else if (DestVT.bitsLT(MVT::f64)) {
2325 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2326 DAG.getIntPtrConstant(0));
2327 } else if (DestVT.bitsGT(MVT::f64)) {
2328 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2332 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2333 // Code below here assumes !isSigned without checking again.
2335 // Implementation of unsigned i64 to f64 following the algorithm in
2336 // __floatundidf in compiler_rt. This implementation has the advantage
2337 // of performing rounding correctly, both in the default rounding mode
2338 // and in all alternate rounding modes.
2339 // TODO: Generalize this for use with other types.
2340 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2342 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2343 SDValue TwoP84PlusTwoP52 =
2344 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2346 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2348 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2349 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2350 DAG.getConstant(32, MVT::i64));
2351 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2352 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2353 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2354 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2355 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2357 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2360 // Implementation of unsigned i64 to f32.
2361 // TODO: Generalize this for use with other types.
2362 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2363 // For unsigned conversions, convert them to signed conversions using the
2364 // algorithm from the x86_64 __floatundidf in compiler_rt.
2366 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2368 SDValue ShiftConst =
2369 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2370 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2371 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2372 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2373 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2375 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2376 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2378 // TODO: This really should be implemented using a branch rather than a
2379 // select. We happen to get lucky and machinesink does the right
2380 // thing most of the time. This would be a good candidate for a
2381 //pseudo-op, or, even better, for whole-function isel.
2382 SDValue SignBitTest = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2383 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2384 return DAG.getNode(ISD::SELECT, dl, MVT::f32, SignBitTest, Slow, Fast);
2387 // Otherwise, implement the fully general conversion.
2389 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2390 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2391 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2392 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2393 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2394 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2395 SDValue Ne = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2396 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2397 SDValue Sel = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ne, Or, Op0);
2398 SDValue Ge = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
2399 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2401 SDValue Sel2 = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ge, Sel, Op0);
2402 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2404 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2405 DAG.getConstant(32, SHVT));
2406 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2407 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2409 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2410 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2411 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2412 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2413 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2414 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2415 DAG.getIntPtrConstant(0));
2418 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2420 SDValue SignSet = DAG.getSetCC(dl, TLI.getSetCCResultType(Op0.getValueType()),
2421 Op0, DAG.getConstant(0, Op0.getValueType()),
2423 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2424 SDValue CstOffset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(),
2425 SignSet, Four, Zero);
2427 // If the sign bit of the integer is set, the large number will be treated
2428 // as a negative number. To counteract this, the dynamic code adds an
2429 // offset depending on the data type.
2431 switch (Op0.getValueType().getSimpleVT().SimpleTy) {
2432 default: assert(0 && "Unsupported integer type!");
2433 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2434 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2435 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2436 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2438 if (TLI.isLittleEndian()) FF <<= 32;
2439 Constant *FudgeFactor = ConstantInt::get(
2440 Type::getInt64Ty(*DAG.getContext()), FF);
2442 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2443 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2444 CPIdx = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), CPIdx, CstOffset);
2445 Alignment = std::min(Alignment, 4u);
2447 if (DestVT == MVT::f32)
2448 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2449 MachinePointerInfo::getConstantPool(),
2450 false, false, Alignment);
2453 LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2454 DAG.getEntryNode(), CPIdx,
2455 MachinePointerInfo::getConstantPool(),
2456 MVT::f32, false, false, Alignment));
2459 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2462 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2463 /// *INT_TO_FP operation of the specified operand when the target requests that
2464 /// we promote it. At this point, we know that the result and operand types are
2465 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2466 /// operation that takes a larger input.
2467 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2471 // First step, figure out the appropriate *INT_TO_FP operation to use.
2472 EVT NewInTy = LegalOp.getValueType();
2474 unsigned OpToUse = 0;
2476 // Scan for the appropriate larger type to use.
2478 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2479 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2481 // If the target supports SINT_TO_FP of this type, use it.
2482 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2483 OpToUse = ISD::SINT_TO_FP;
2486 if (isSigned) continue;
2488 // If the target supports UINT_TO_FP of this type, use it.
2489 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2490 OpToUse = ISD::UINT_TO_FP;
2494 // Otherwise, try a larger type.
2497 // Okay, we found the operation and type to use. Zero extend our input to the
2498 // desired type then run the operation on it.
2499 return DAG.getNode(OpToUse, dl, DestVT,
2500 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2501 dl, NewInTy, LegalOp));
2504 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2505 /// FP_TO_*INT operation of the specified operand when the target requests that
2506 /// we promote it. At this point, we know that the result and operand types are
2507 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2508 /// operation that returns a larger result.
2509 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2513 // First step, figure out the appropriate FP_TO*INT operation to use.
2514 EVT NewOutTy = DestVT;
2516 unsigned OpToUse = 0;
2518 // Scan for the appropriate larger type to use.
2520 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2521 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2523 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2524 OpToUse = ISD::FP_TO_SINT;
2528 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2529 OpToUse = ISD::FP_TO_UINT;
2533 // Otherwise, try a larger type.
2537 // Okay, we found the operation and type to use.
2538 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2540 // Truncate the result of the extended FP_TO_*INT operation to the desired
2542 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2545 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2547 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, DebugLoc dl) {
2548 EVT VT = Op.getValueType();
2549 EVT SHVT = TLI.getShiftAmountTy(VT);
2550 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2551 switch (VT.getSimpleVT().SimpleTy) {
2552 default: assert(0 && "Unhandled Expand type in BSWAP!");
2554 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2555 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2556 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2558 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2559 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2560 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2561 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2562 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2563 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2564 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2565 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2566 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2568 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2569 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2570 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2571 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2572 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2573 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2574 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2575 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2576 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2577 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2578 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2579 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2580 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2581 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2582 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2583 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2584 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2585 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2586 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2587 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2588 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2592 /// SplatByte - Distribute ByteVal over NumBits bits.
2593 // FIXME: Move this helper to a common place.
2594 static APInt SplatByte(unsigned NumBits, uint8_t ByteVal) {
2595 APInt Val = APInt(NumBits, ByteVal);
2597 for (unsigned i = NumBits; i > 8; i >>= 1) {
2598 Val = (Val << Shift) | Val;
2604 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2606 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2609 default: assert(0 && "Cannot expand this yet!");
2611 EVT VT = Op.getValueType();
2612 EVT ShVT = TLI.getShiftAmountTy(VT);
2613 unsigned Len = VT.getSizeInBits();
2615 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2616 "CTPOP not implemented for this type.");
2618 // This is the "best" algorithm from
2619 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2621 SDValue Mask55 = DAG.getConstant(SplatByte(Len, 0x55), VT);
2622 SDValue Mask33 = DAG.getConstant(SplatByte(Len, 0x33), VT);
2623 SDValue Mask0F = DAG.getConstant(SplatByte(Len, 0x0F), VT);
2624 SDValue Mask01 = DAG.getConstant(SplatByte(Len, 0x01), VT);
2626 // v = v - ((v >> 1) & 0x55555555...)
2627 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2628 DAG.getNode(ISD::AND, dl, VT,
2629 DAG.getNode(ISD::SRL, dl, VT, Op,
2630 DAG.getConstant(1, ShVT)),
2632 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2633 Op = DAG.getNode(ISD::ADD, dl, VT,
2634 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2635 DAG.getNode(ISD::AND, dl, VT,
2636 DAG.getNode(ISD::SRL, dl, VT, Op,
2637 DAG.getConstant(2, ShVT)),
2639 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2640 Op = DAG.getNode(ISD::AND, dl, VT,
2641 DAG.getNode(ISD::ADD, dl, VT, Op,
2642 DAG.getNode(ISD::SRL, dl, VT, Op,
2643 DAG.getConstant(4, ShVT))),
2645 // v = (v * 0x01010101...) >> (Len - 8)
2646 Op = DAG.getNode(ISD::SRL, dl, VT,
2647 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2648 DAG.getConstant(Len - 8, ShVT));
2653 // for now, we do this:
2654 // x = x | (x >> 1);
2655 // x = x | (x >> 2);
2657 // x = x | (x >>16);
2658 // x = x | (x >>32); // for 64-bit input
2659 // return popcount(~x);
2661 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2662 EVT VT = Op.getValueType();
2663 EVT ShVT = TLI.getShiftAmountTy(VT);
2664 unsigned len = VT.getSizeInBits();
2665 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2666 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2667 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2668 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2670 Op = DAG.getNOT(dl, Op, VT);
2671 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2674 // for now, we use: { return popcount(~x & (x - 1)); }
2675 // unless the target has ctlz but not ctpop, in which case we use:
2676 // { return 32 - nlz(~x & (x-1)); }
2677 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2678 EVT VT = Op.getValueType();
2679 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2680 DAG.getNOT(dl, Op, VT),
2681 DAG.getNode(ISD::SUB, dl, VT, Op,
2682 DAG.getConstant(1, VT)));
2683 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2684 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2685 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2686 return DAG.getNode(ISD::SUB, dl, VT,
2687 DAG.getConstant(VT.getSizeInBits(), VT),
2688 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2689 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2694 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2695 unsigned Opc = Node->getOpcode();
2696 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2701 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2703 case ISD::ATOMIC_SWAP:
2704 switch (VT.SimpleTy) {
2705 default: llvm_unreachable("Unexpected value type for atomic!");
2706 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2707 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2708 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2709 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2712 case ISD::ATOMIC_CMP_SWAP:
2713 switch (VT.SimpleTy) {
2714 default: llvm_unreachable("Unexpected value type for atomic!");
2715 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2716 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2717 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2718 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2721 case ISD::ATOMIC_LOAD_ADD:
2722 switch (VT.SimpleTy) {
2723 default: llvm_unreachable("Unexpected value type for atomic!");
2724 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2725 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2726 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2727 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2730 case ISD::ATOMIC_LOAD_SUB:
2731 switch (VT.SimpleTy) {
2732 default: llvm_unreachable("Unexpected value type for atomic!");
2733 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2734 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2735 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2736 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2739 case ISD::ATOMIC_LOAD_AND:
2740 switch (VT.SimpleTy) {
2741 default: llvm_unreachable("Unexpected value type for atomic!");
2742 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2743 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2744 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2745 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2748 case ISD::ATOMIC_LOAD_OR:
2749 switch (VT.SimpleTy) {
2750 default: llvm_unreachable("Unexpected value type for atomic!");
2751 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2752 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2753 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2754 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2757 case ISD::ATOMIC_LOAD_XOR:
2758 switch (VT.SimpleTy) {
2759 default: llvm_unreachable("Unexpected value type for atomic!");
2760 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2761 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2762 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2763 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2766 case ISD::ATOMIC_LOAD_NAND:
2767 switch (VT.SimpleTy) {
2768 default: llvm_unreachable("Unexpected value type for atomic!");
2769 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2770 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2771 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2772 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2777 return ExpandChainLibCall(LC, Node, false);
2780 void SelectionDAGLegalize::ExpandNode(SDNode *Node,
2781 SmallVectorImpl<SDValue> &Results) {
2782 DebugLoc dl = Node->getDebugLoc();
2783 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2784 switch (Node->getOpcode()) {
2788 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2789 Results.push_back(Tmp1);
2792 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2794 case ISD::FRAMEADDR:
2795 case ISD::RETURNADDR:
2796 case ISD::FRAME_TO_ARGS_OFFSET:
2797 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2799 case ISD::FLT_ROUNDS_:
2800 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2802 case ISD::EH_RETURN:
2806 case ISD::EH_SJLJ_LONGJMP:
2807 case ISD::EH_SJLJ_DISPATCHSETUP:
2808 // If the target didn't expand these, there's nothing to do, so just
2809 // preserve the chain and be done.
2810 Results.push_back(Node->getOperand(0));
2812 case ISD::EH_SJLJ_SETJMP:
2813 // If the target didn't expand this, just return 'zero' and preserve the
2815 Results.push_back(DAG.getConstant(0, MVT::i32));
2816 Results.push_back(Node->getOperand(0));
2818 case ISD::MEMBARRIER: {
2819 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2820 TargetLowering::ArgListTy Args;
2821 std::pair<SDValue, SDValue> CallResult =
2822 TLI.LowerCallTo(Node->getOperand(0), Type::getVoidTy(*DAG.getContext()),
2823 false, false, false, false, 0, CallingConv::C,
2824 /*isTailCall=*/false,
2825 /*isReturnValueUsed=*/true,
2826 DAG.getExternalSymbol("__sync_synchronize",
2827 TLI.getPointerTy()),
2829 Results.push_back(CallResult.second);
2832 // By default, atomic intrinsics are marked Legal and lowered. Targets
2833 // which don't support them directly, however, may want libcalls, in which
2834 // case they mark them Expand, and we get here.
2835 case ISD::ATOMIC_SWAP:
2836 case ISD::ATOMIC_LOAD_ADD:
2837 case ISD::ATOMIC_LOAD_SUB:
2838 case ISD::ATOMIC_LOAD_AND:
2839 case ISD::ATOMIC_LOAD_OR:
2840 case ISD::ATOMIC_LOAD_XOR:
2841 case ISD::ATOMIC_LOAD_NAND:
2842 case ISD::ATOMIC_LOAD_MIN:
2843 case ISD::ATOMIC_LOAD_MAX:
2844 case ISD::ATOMIC_LOAD_UMIN:
2845 case ISD::ATOMIC_LOAD_UMAX:
2846 case ISD::ATOMIC_CMP_SWAP: {
2847 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
2848 Results.push_back(Tmp.first);
2849 Results.push_back(Tmp.second);
2852 case ISD::DYNAMIC_STACKALLOC:
2853 ExpandDYNAMIC_STACKALLOC(Node, Results);
2855 case ISD::MERGE_VALUES:
2856 for (unsigned i = 0; i < Node->getNumValues(); i++)
2857 Results.push_back(Node->getOperand(i));
2860 EVT VT = Node->getValueType(0);
2862 Results.push_back(DAG.getConstant(0, VT));
2864 assert(VT.isFloatingPoint() && "Unknown value type!");
2865 Results.push_back(DAG.getConstantFP(0, VT));
2870 // If this operation is not supported, lower it to 'abort()' call
2871 TargetLowering::ArgListTy Args;
2872 std::pair<SDValue, SDValue> CallResult =
2873 TLI.LowerCallTo(Node->getOperand(0), Type::getVoidTy(*DAG.getContext()),
2874 false, false, false, false, 0, CallingConv::C,
2875 /*isTailCall=*/false,
2876 /*isReturnValueUsed=*/true,
2877 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
2879 Results.push_back(CallResult.second);
2884 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
2885 Node->getValueType(0), dl);
2886 Results.push_back(Tmp1);
2888 case ISD::FP_EXTEND:
2889 Tmp1 = EmitStackConvert(Node->getOperand(0),
2890 Node->getOperand(0).getValueType(),
2891 Node->getValueType(0), dl);
2892 Results.push_back(Tmp1);
2894 case ISD::SIGN_EXTEND_INREG: {
2895 // NOTE: we could fall back on load/store here too for targets without
2896 // SAR. However, it is doubtful that any exist.
2897 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2898 EVT VT = Node->getValueType(0);
2899 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
2902 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
2903 ExtraVT.getScalarType().getSizeInBits();
2904 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
2905 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
2906 Node->getOperand(0), ShiftCst);
2907 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
2908 Results.push_back(Tmp1);
2911 case ISD::FP_ROUND_INREG: {
2912 // The only way we can lower this is to turn it into a TRUNCSTORE,
2913 // EXTLOAD pair, targeting a temporary location (a stack slot).
2915 // NOTE: there is a choice here between constantly creating new stack
2916 // slots and always reusing the same one. We currently always create
2917 // new ones, as reuse may inhibit scheduling.
2918 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
2919 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
2920 Node->getValueType(0), dl);
2921 Results.push_back(Tmp1);
2924 case ISD::SINT_TO_FP:
2925 case ISD::UINT_TO_FP:
2926 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
2927 Node->getOperand(0), Node->getValueType(0), dl);
2928 Results.push_back(Tmp1);
2930 case ISD::FP_TO_UINT: {
2931 SDValue True, False;
2932 EVT VT = Node->getOperand(0).getValueType();
2933 EVT NVT = Node->getValueType(0);
2934 APFloat apf(APInt::getNullValue(VT.getSizeInBits()));
2935 APInt x = APInt::getSignBit(NVT.getSizeInBits());
2936 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
2937 Tmp1 = DAG.getConstantFP(apf, VT);
2938 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(VT),
2939 Node->getOperand(0),
2941 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
2942 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
2943 DAG.getNode(ISD::FSUB, dl, VT,
2944 Node->getOperand(0), Tmp1));
2945 False = DAG.getNode(ISD::XOR, dl, NVT, False,
2946 DAG.getConstant(x, NVT));
2947 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2, True, False);
2948 Results.push_back(Tmp1);
2952 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2953 EVT VT = Node->getValueType(0);
2954 Tmp1 = Node->getOperand(0);
2955 Tmp2 = Node->getOperand(1);
2956 unsigned Align = Node->getConstantOperandVal(3);
2958 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
2959 MachinePointerInfo(V), false, false, 0);
2960 SDValue VAList = VAListLoad;
2962 if (Align > TLI.getMinStackArgumentAlignment()) {
2963 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
2965 VAList = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2966 DAG.getConstant(Align - 1,
2967 TLI.getPointerTy()));
2969 VAList = DAG.getNode(ISD::AND, dl, TLI.getPointerTy(), VAList,
2970 DAG.getConstant(-(int64_t)Align,
2971 TLI.getPointerTy()));
2974 // Increment the pointer, VAList, to the next vaarg
2975 Tmp3 = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
2976 DAG.getConstant(TLI.getTargetData()->
2977 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
2978 TLI.getPointerTy()));
2979 // Store the incremented VAList to the legalized pointer
2980 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
2981 MachinePointerInfo(V), false, false, 0);
2982 // Load the actual argument out of the pointer VAList
2983 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
2985 Results.push_back(Results[0].getValue(1));
2989 // This defaults to loading a pointer from the input and storing it to the
2990 // output, returning the chain.
2991 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
2992 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
2993 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
2994 Node->getOperand(2), MachinePointerInfo(VS),
2996 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
2997 MachinePointerInfo(VD), false, false, 0);
2998 Results.push_back(Tmp1);
3001 case ISD::EXTRACT_VECTOR_ELT:
3002 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3003 // This must be an access of the only element. Return it.
3004 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3005 Node->getOperand(0));
3007 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3008 Results.push_back(Tmp1);
3010 case ISD::EXTRACT_SUBVECTOR:
3011 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3013 case ISD::INSERT_SUBVECTOR:
3014 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3016 case ISD::CONCAT_VECTORS: {
3017 Results.push_back(ExpandVectorBuildThroughStack(Node));
3020 case ISD::SCALAR_TO_VECTOR:
3021 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3023 case ISD::INSERT_VECTOR_ELT:
3024 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3025 Node->getOperand(1),
3026 Node->getOperand(2), dl));
3028 case ISD::VECTOR_SHUFFLE: {
3029 SmallVector<int, 8> Mask;
3030 cast<ShuffleVectorSDNode>(Node)->getMask(Mask);
3032 EVT VT = Node->getValueType(0);
3033 EVT EltVT = VT.getVectorElementType();
3034 if (getTypeAction(EltVT) == Promote)
3035 EltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3036 unsigned NumElems = VT.getVectorNumElements();
3037 SmallVector<SDValue, 8> Ops;
3038 for (unsigned i = 0; i != NumElems; ++i) {
3040 Ops.push_back(DAG.getUNDEF(EltVT));
3043 unsigned Idx = Mask[i];
3045 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3046 Node->getOperand(0),
3047 DAG.getIntPtrConstant(Idx)));
3049 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3050 Node->getOperand(1),
3051 DAG.getIntPtrConstant(Idx - NumElems)));
3053 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
3054 Results.push_back(Tmp1);
3057 case ISD::EXTRACT_ELEMENT: {
3058 EVT OpTy = Node->getOperand(0).getValueType();
3059 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3061 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3062 DAG.getConstant(OpTy.getSizeInBits()/2,
3063 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3064 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3067 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3068 Node->getOperand(0));
3070 Results.push_back(Tmp1);
3073 case ISD::STACKSAVE:
3074 // Expand to CopyFromReg if the target set
3075 // StackPointerRegisterToSaveRestore.
3076 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3077 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3078 Node->getValueType(0)));
3079 Results.push_back(Results[0].getValue(1));
3081 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3082 Results.push_back(Node->getOperand(0));
3085 case ISD::STACKRESTORE:
3086 // Expand to CopyToReg if the target set
3087 // StackPointerRegisterToSaveRestore.
3088 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3089 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3090 Node->getOperand(1)));
3092 Results.push_back(Node->getOperand(0));
3095 case ISD::FCOPYSIGN:
3096 Results.push_back(ExpandFCOPYSIGN(Node));
3099 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3100 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3101 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3102 Node->getOperand(0));
3103 Results.push_back(Tmp1);
3106 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3107 EVT VT = Node->getValueType(0);
3108 Tmp1 = Node->getOperand(0);
3109 Tmp2 = DAG.getConstantFP(0.0, VT);
3110 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(Tmp1.getValueType()),
3111 Tmp1, Tmp2, ISD::SETUGT);
3112 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3113 Tmp1 = DAG.getNode(ISD::SELECT, dl, VT, Tmp2, Tmp1, Tmp3);
3114 Results.push_back(Tmp1);
3118 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3119 RTLIB::SQRT_F80, RTLIB::SQRT_PPCF128));
3122 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3123 RTLIB::SIN_F80, RTLIB::SIN_PPCF128));
3126 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3127 RTLIB::COS_F80, RTLIB::COS_PPCF128));
3130 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3131 RTLIB::LOG_F80, RTLIB::LOG_PPCF128));
3134 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3135 RTLIB::LOG2_F80, RTLIB::LOG2_PPCF128));
3138 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3139 RTLIB::LOG10_F80, RTLIB::LOG10_PPCF128));
3142 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3143 RTLIB::EXP_F80, RTLIB::EXP_PPCF128));
3146 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3147 RTLIB::EXP2_F80, RTLIB::EXP2_PPCF128));
3150 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3151 RTLIB::TRUNC_F80, RTLIB::TRUNC_PPCF128));
3154 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3155 RTLIB::FLOOR_F80, RTLIB::FLOOR_PPCF128));
3158 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3159 RTLIB::CEIL_F80, RTLIB::CEIL_PPCF128));
3162 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3163 RTLIB::RINT_F80, RTLIB::RINT_PPCF128));
3165 case ISD::FNEARBYINT:
3166 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3167 RTLIB::NEARBYINT_F64,
3168 RTLIB::NEARBYINT_F80,
3169 RTLIB::NEARBYINT_PPCF128));
3172 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3173 RTLIB::POWI_F80, RTLIB::POWI_PPCF128));
3176 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3177 RTLIB::POW_F80, RTLIB::POW_PPCF128));
3180 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3181 RTLIB::DIV_F80, RTLIB::DIV_PPCF128));
3184 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3185 RTLIB::REM_F80, RTLIB::REM_PPCF128));
3187 case ISD::FP16_TO_FP32:
3188 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3190 case ISD::FP32_TO_FP16:
3191 Results.push_back(ExpandLibCall(RTLIB::FPROUND_F32_F16, Node, false));
3193 case ISD::ConstantFP: {
3194 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3195 // Check to see if this FP immediate is already legal.
3196 // If this is a legal constant, turn it into a TargetConstantFP node.
3197 if (TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3198 Results.push_back(SDValue(Node, 0));
3200 Results.push_back(ExpandConstantFP(CFP, true, DAG, TLI));
3203 case ISD::EHSELECTION: {
3204 unsigned Reg = TLI.getExceptionSelectorRegister();
3205 assert(Reg && "Can't expand to unknown register!");
3206 Results.push_back(DAG.getCopyFromReg(Node->getOperand(1), dl, Reg,
3207 Node->getValueType(0)));
3208 Results.push_back(Results[0].getValue(1));
3211 case ISD::EXCEPTIONADDR: {
3212 unsigned Reg = TLI.getExceptionAddressRegister();
3213 assert(Reg && "Can't expand to unknown register!");
3214 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, Reg,
3215 Node->getValueType(0)));
3216 Results.push_back(Results[0].getValue(1));
3220 EVT VT = Node->getValueType(0);
3221 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3222 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3223 "Don't know how to expand this subtraction!");
3224 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3225 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3226 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp2, DAG.getConstant(1, VT));
3227 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3232 EVT VT = Node->getValueType(0);
3233 SDVTList VTs = DAG.getVTList(VT, VT);
3234 bool isSigned = Node->getOpcode() == ISD::SREM;
3235 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3236 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3237 Tmp2 = Node->getOperand(0);
3238 Tmp3 = Node->getOperand(1);
3239 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3240 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3241 UseDivRem(Node, isSigned, false))) {
3242 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3243 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3245 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3246 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3247 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3248 } else if (isSigned)
3249 Tmp1 = ExpandIntLibCall(Node, true,
3251 RTLIB::SREM_I16, RTLIB::SREM_I32,
3252 RTLIB::SREM_I64, RTLIB::SREM_I128);
3254 Tmp1 = ExpandIntLibCall(Node, false,
3256 RTLIB::UREM_I16, RTLIB::UREM_I32,
3257 RTLIB::UREM_I64, RTLIB::UREM_I128);
3258 Results.push_back(Tmp1);
3263 bool isSigned = Node->getOpcode() == ISD::SDIV;
3264 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3265 EVT VT = Node->getValueType(0);
3266 SDVTList VTs = DAG.getVTList(VT, VT);
3267 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3268 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3269 UseDivRem(Node, isSigned, true)))
3270 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3271 Node->getOperand(1));
3273 Tmp1 = ExpandIntLibCall(Node, true,
3275 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3276 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3278 Tmp1 = ExpandIntLibCall(Node, false,
3280 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3281 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3282 Results.push_back(Tmp1);
3287 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3289 EVT VT = Node->getValueType(0);
3290 SDVTList VTs = DAG.getVTList(VT, VT);
3291 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3292 "If this wasn't legal, it shouldn't have been created!");
3293 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3294 Node->getOperand(1));
3295 Results.push_back(Tmp1.getValue(1));
3300 // Expand into divrem libcall
3301 ExpandDivRemLibCall(Node, Results);
3304 EVT VT = Node->getValueType(0);
3305 SDVTList VTs = DAG.getVTList(VT, VT);
3306 // See if multiply or divide can be lowered using two-result operations.
3307 // We just need the low half of the multiply; try both the signed
3308 // and unsigned forms. If the target supports both SMUL_LOHI and
3309 // UMUL_LOHI, form a preference by checking which forms of plain
3310 // MULH it supports.
3311 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3312 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3313 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3314 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3315 unsigned OpToUse = 0;
3316 if (HasSMUL_LOHI && !HasMULHS) {
3317 OpToUse = ISD::SMUL_LOHI;
3318 } else if (HasUMUL_LOHI && !HasMULHU) {
3319 OpToUse = ISD::UMUL_LOHI;
3320 } else if (HasSMUL_LOHI) {
3321 OpToUse = ISD::SMUL_LOHI;
3322 } else if (HasUMUL_LOHI) {
3323 OpToUse = ISD::UMUL_LOHI;
3326 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3327 Node->getOperand(1)));
3330 Tmp1 = ExpandIntLibCall(Node, false,
3332 RTLIB::MUL_I16, RTLIB::MUL_I32,
3333 RTLIB::MUL_I64, RTLIB::MUL_I128);
3334 Results.push_back(Tmp1);
3339 SDValue LHS = Node->getOperand(0);
3340 SDValue RHS = Node->getOperand(1);
3341 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3342 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3344 Results.push_back(Sum);
3345 EVT OType = Node->getValueType(1);
3347 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3349 // LHSSign -> LHS >= 0
3350 // RHSSign -> RHS >= 0
3351 // SumSign -> Sum >= 0
3354 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3356 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3358 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3359 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3360 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3361 Node->getOpcode() == ISD::SADDO ?
3362 ISD::SETEQ : ISD::SETNE);
3364 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3365 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3367 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3368 Results.push_back(Cmp);
3373 SDValue LHS = Node->getOperand(0);
3374 SDValue RHS = Node->getOperand(1);
3375 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3376 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3378 Results.push_back(Sum);
3379 Results.push_back(DAG.getSetCC(dl, Node->getValueType(1), Sum, LHS,
3380 Node->getOpcode () == ISD::UADDO ?
3381 ISD::SETULT : ISD::SETUGT));
3386 EVT VT = Node->getValueType(0);
3387 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3388 SDValue LHS = Node->getOperand(0);
3389 SDValue RHS = Node->getOperand(1);
3392 static const unsigned Ops[2][3] =
3393 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3394 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3395 bool isSigned = Node->getOpcode() == ISD::SMULO;
3396 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3397 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3398 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3399 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3400 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3402 TopHalf = BottomHalf.getValue(1);
3403 } else if (TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
3404 VT.getSizeInBits() * 2))) {
3405 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3406 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3407 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3408 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3409 DAG.getIntPtrConstant(0));
3410 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3411 DAG.getIntPtrConstant(1));
3413 // We can fall back to a libcall with an illegal type for the MUL if we
3414 // have a libcall big enough.
3415 // Also, we can fall back to a division in some cases, but that's a big
3416 // performance hit in the general case.
3417 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3418 if (WideVT == MVT::i16)
3419 LC = RTLIB::MUL_I16;
3420 else if (WideVT == MVT::i32)
3421 LC = RTLIB::MUL_I32;
3422 else if (WideVT == MVT::i64)
3423 LC = RTLIB::MUL_I64;
3424 else if (WideVT == MVT::i128)
3425 LC = RTLIB::MUL_I128;
3426 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3428 // The high part is obtained by SRA'ing all but one of the bits of low
3430 unsigned LoSize = VT.getSizeInBits();
3431 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3432 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3433 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3434 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3436 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3437 // pre-lowered to the correct types. This all depends upon WideVT not
3438 // being a legal type for the architecture and thus has to be split to
3440 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3441 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3442 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3443 DAG.getIntPtrConstant(0));
3444 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3445 DAG.getIntPtrConstant(1));
3449 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3450 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3451 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3452 TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf, Tmp1,
3455 TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf,
3456 DAG.getConstant(0, VT), ISD::SETNE);
3458 Results.push_back(BottomHalf);
3459 Results.push_back(TopHalf);
3462 case ISD::BUILD_PAIR: {
3463 EVT PairTy = Node->getValueType(0);
3464 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3465 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3466 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3467 DAG.getConstant(PairTy.getSizeInBits()/2,
3468 TLI.getShiftAmountTy(PairTy)));
3469 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3473 Tmp1 = Node->getOperand(0);
3474 Tmp2 = Node->getOperand(1);
3475 Tmp3 = Node->getOperand(2);
3476 if (Tmp1.getOpcode() == ISD::SETCC) {
3477 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3479 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3481 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3482 DAG.getConstant(0, Tmp1.getValueType()),
3483 Tmp2, Tmp3, ISD::SETNE);
3485 Results.push_back(Tmp1);
3488 SDValue Chain = Node->getOperand(0);
3489 SDValue Table = Node->getOperand(1);
3490 SDValue Index = Node->getOperand(2);
3492 EVT PTy = TLI.getPointerTy();
3494 const TargetData &TD = *TLI.getTargetData();
3495 unsigned EntrySize =
3496 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3498 Index = DAG.getNode(ISD::MUL, dl, PTy,
3499 Index, DAG.getConstant(EntrySize, PTy));
3500 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3502 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3503 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3504 MachinePointerInfo::getJumpTable(), MemVT,
3507 if (TM.getRelocationModel() == Reloc::PIC_) {
3508 // For PIC, the sequence is:
3509 // BRIND(load(Jumptable + index) + RelocBase)
3510 // RelocBase can be JumpTable, GOT or some sort of global base.
3511 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3512 TLI.getPICJumpTableRelocBase(Table, DAG));
3514 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3515 Results.push_back(Tmp1);
3519 // Expand brcond's setcc into its constituent parts and create a BR_CC
3521 Tmp1 = Node->getOperand(0);
3522 Tmp2 = Node->getOperand(1);
3523 if (Tmp2.getOpcode() == ISD::SETCC) {
3524 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3525 Tmp1, Tmp2.getOperand(2),
3526 Tmp2.getOperand(0), Tmp2.getOperand(1),
3527 Node->getOperand(2));
3529 // We test only the i1 bit. Skip the AND if UNDEF.
3530 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3531 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3532 DAG.getConstant(1, Tmp2.getValueType()));
3533 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3534 DAG.getCondCode(ISD::SETNE), Tmp3,
3535 DAG.getConstant(0, Tmp3.getValueType()),
3536 Node->getOperand(2));
3538 Results.push_back(Tmp1);
3541 Tmp1 = Node->getOperand(0);
3542 Tmp2 = Node->getOperand(1);
3543 Tmp3 = Node->getOperand(2);
3544 LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2, Tmp3, dl);
3546 // If we expanded the SETCC into an AND/OR, return the new node
3547 if (Tmp2.getNode() == 0) {
3548 Results.push_back(Tmp1);
3552 // Otherwise, SETCC for the given comparison type must be completely
3553 // illegal; expand it into a SELECT_CC.
3554 EVT VT = Node->getValueType(0);
3555 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3556 DAG.getConstant(1, VT), DAG.getConstant(0, VT), Tmp3);
3557 Results.push_back(Tmp1);
3560 case ISD::SELECT_CC: {
3561 Tmp1 = Node->getOperand(0); // LHS
3562 Tmp2 = Node->getOperand(1); // RHS
3563 Tmp3 = Node->getOperand(2); // True
3564 Tmp4 = Node->getOperand(3); // False
3565 SDValue CC = Node->getOperand(4);
3567 LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp1.getValueType()),
3568 Tmp1, Tmp2, CC, dl);
3570 assert(!Tmp2.getNode() && "Can't legalize SELECT_CC with legal condition!");
3571 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
3572 CC = DAG.getCondCode(ISD::SETNE);
3573 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1, Tmp2,
3575 Results.push_back(Tmp1);
3579 Tmp1 = Node->getOperand(0); // Chain
3580 Tmp2 = Node->getOperand(2); // LHS
3581 Tmp3 = Node->getOperand(3); // RHS
3582 Tmp4 = Node->getOperand(1); // CC
3584 LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp2.getValueType()),
3585 Tmp2, Tmp3, Tmp4, dl);
3586 assert(LastCALLSEQ.size() == 1 && "branch inside CALLSEQ_BEGIN/END?");
3587 setLastCALLSEQ(DAG.getEntryNode());
3589 assert(!Tmp3.getNode() && "Can't legalize BR_CC with legal condition!");
3590 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
3591 Tmp4 = DAG.getCondCode(ISD::SETNE);
3592 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4, Tmp2,
3593 Tmp3, Node->getOperand(4));
3594 Results.push_back(Tmp1);
3597 case ISD::GLOBAL_OFFSET_TABLE:
3598 case ISD::GlobalAddress:
3599 case ISD::GlobalTLSAddress:
3600 case ISD::ExternalSymbol:
3601 case ISD::ConstantPool:
3602 case ISD::JumpTable:
3603 case ISD::INTRINSIC_W_CHAIN:
3604 case ISD::INTRINSIC_WO_CHAIN:
3605 case ISD::INTRINSIC_VOID:
3606 // FIXME: Custom lowering for these operations shouldn't return null!
3607 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
3608 Results.push_back(SDValue(Node, i));
3612 void SelectionDAGLegalize::PromoteNode(SDNode *Node,
3613 SmallVectorImpl<SDValue> &Results) {
3614 EVT OVT = Node->getValueType(0);
3615 if (Node->getOpcode() == ISD::UINT_TO_FP ||
3616 Node->getOpcode() == ISD::SINT_TO_FP ||
3617 Node->getOpcode() == ISD::SETCC) {
3618 OVT = Node->getOperand(0).getValueType();
3620 EVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
3621 DebugLoc dl = Node->getDebugLoc();
3622 SDValue Tmp1, Tmp2, Tmp3;
3623 switch (Node->getOpcode()) {
3627 // Zero extend the argument.
3628 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3629 // Perform the larger operation.
3630 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
3631 if (Node->getOpcode() == ISD::CTTZ) {
3632 //if Tmp1 == sizeinbits(NVT) then Tmp1 = sizeinbits(Old VT)
3633 Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(NVT),
3634 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
3636 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2,
3637 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
3638 } else if (Node->getOpcode() == ISD::CTLZ) {
3639 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
3640 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
3641 DAG.getConstant(NVT.getSizeInBits() -
3642 OVT.getSizeInBits(), NVT));
3644 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
3647 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
3648 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
3649 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
3650 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
3651 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
3652 Results.push_back(Tmp1);
3655 case ISD::FP_TO_UINT:
3656 case ISD::FP_TO_SINT:
3657 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
3658 Node->getOpcode() == ISD::FP_TO_SINT, dl);
3659 Results.push_back(Tmp1);
3661 case ISD::UINT_TO_FP:
3662 case ISD::SINT_TO_FP:
3663 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
3664 Node->getOpcode() == ISD::SINT_TO_FP, dl);
3665 Results.push_back(Tmp1);
3670 unsigned ExtOp, TruncOp;
3671 if (OVT.isVector()) {
3672 ExtOp = ISD::BITCAST;
3673 TruncOp = ISD::BITCAST;
3675 assert(OVT.isInteger() && "Cannot promote logic operation");
3676 ExtOp = ISD::ANY_EXTEND;
3677 TruncOp = ISD::TRUNCATE;
3679 // Promote each of the values to the new type.
3680 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3681 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3682 // Perform the larger operation, then convert back
3683 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
3684 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
3688 unsigned ExtOp, TruncOp;
3689 if (Node->getValueType(0).isVector()) {
3690 ExtOp = ISD::BITCAST;
3691 TruncOp = ISD::BITCAST;
3692 } else if (Node->getValueType(0).isInteger()) {
3693 ExtOp = ISD::ANY_EXTEND;
3694 TruncOp = ISD::TRUNCATE;
3696 ExtOp = ISD::FP_EXTEND;
3697 TruncOp = ISD::FP_ROUND;
3699 Tmp1 = Node->getOperand(0);
3700 // Promote each of the values to the new type.
3701 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3702 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
3703 // Perform the larger operation, then round down.
3704 Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp1, Tmp2, Tmp3);
3705 if (TruncOp != ISD::FP_ROUND)
3706 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
3708 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
3709 DAG.getIntPtrConstant(0));
3710 Results.push_back(Tmp1);
3713 case ISD::VECTOR_SHUFFLE: {
3714 SmallVector<int, 8> Mask;
3715 cast<ShuffleVectorSDNode>(Node)->getMask(Mask);
3717 // Cast the two input vectors.
3718 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
3719 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
3721 // Convert the shuffle mask to the right # elements.
3722 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
3723 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
3724 Results.push_back(Tmp1);
3728 unsigned ExtOp = ISD::FP_EXTEND;
3729 if (NVT.isInteger()) {
3730 ISD::CondCode CCCode =
3731 cast<CondCodeSDNode>(Node->getOperand(2))->get();
3732 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3734 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
3735 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
3736 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3737 Tmp1, Tmp2, Node->getOperand(2)));
3743 // SelectionDAG::Legalize - This is the entry point for the file.
3745 void SelectionDAG::Legalize() {
3746 /// run - This is the main entry point to this class.
3748 SelectionDAGLegalize(*this).LegalizeDAG();