1 //===-- SelectionDAG.cpp - Implement the SelectionDAG data structures -----===//
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 implements the SelectionDAG class.
12 //===----------------------------------------------------------------------===//
13 #include "llvm/CodeGen/SelectionDAG.h"
14 #include "llvm/Constants.h"
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/GlobalAlias.h"
17 #include "llvm/GlobalVariable.h"
18 #include "llvm/Intrinsics.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Assembly/Writer.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/CodeGen/MachineBasicBlock.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineModuleInfo.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/Target/TargetRegisterInfo.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Target/TargetInstrInfo.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/ManagedStatic.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/System/Mutex.h"
38 #include "llvm/ADT/SetVector.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/SmallSet.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/StringExtras.h"
47 /// makeVTList - Return an instance of the SDVTList struct initialized with the
48 /// specified members.
49 static SDVTList makeVTList(const MVT *VTs, unsigned NumVTs) {
50 SDVTList Res = {VTs, NumVTs};
54 static const fltSemantics *MVTToAPFloatSemantics(MVT VT) {
55 switch (VT.getSimpleVT()) {
56 default: assert(0 && "Unknown FP format");
57 case MVT::f32: return &APFloat::IEEEsingle;
58 case MVT::f64: return &APFloat::IEEEdouble;
59 case MVT::f80: return &APFloat::x87DoubleExtended;
60 case MVT::f128: return &APFloat::IEEEquad;
61 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
65 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
67 //===----------------------------------------------------------------------===//
68 // ConstantFPSDNode Class
69 //===----------------------------------------------------------------------===//
71 /// isExactlyValue - We don't rely on operator== working on double values, as
72 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
73 /// As such, this method can be used to do an exact bit-for-bit comparison of
74 /// two floating point values.
75 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
76 return getValueAPF().bitwiseIsEqual(V);
79 bool ConstantFPSDNode::isValueValidForType(MVT VT,
81 assert(VT.isFloatingPoint() && "Can only convert between FP types");
83 // PPC long double cannot be converted to any other type.
84 if (VT == MVT::ppcf128 ||
85 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
88 // convert modifies in place, so make a copy.
89 APFloat Val2 = APFloat(Val);
91 (void) Val2.convert(*MVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
96 //===----------------------------------------------------------------------===//
98 //===----------------------------------------------------------------------===//
100 /// isBuildVectorAllOnes - Return true if the specified node is a
101 /// BUILD_VECTOR where all of the elements are ~0 or undef.
102 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
103 // Look through a bit convert.
104 if (N->getOpcode() == ISD::BIT_CONVERT)
105 N = N->getOperand(0).getNode();
107 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
109 unsigned i = 0, e = N->getNumOperands();
111 // Skip over all of the undef values.
112 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
115 // Do not accept an all-undef vector.
116 if (i == e) return false;
118 // Do not accept build_vectors that aren't all constants or which have non-~0
120 SDValue NotZero = N->getOperand(i);
121 if (isa<ConstantSDNode>(NotZero)) {
122 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
124 } else if (isa<ConstantFPSDNode>(NotZero)) {
125 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
126 bitcastToAPInt().isAllOnesValue())
131 // Okay, we have at least one ~0 value, check to see if the rest match or are
133 for (++i; i != e; ++i)
134 if (N->getOperand(i) != NotZero &&
135 N->getOperand(i).getOpcode() != ISD::UNDEF)
141 /// isBuildVectorAllZeros - Return true if the specified node is a
142 /// BUILD_VECTOR where all of the elements are 0 or undef.
143 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
144 // Look through a bit convert.
145 if (N->getOpcode() == ISD::BIT_CONVERT)
146 N = N->getOperand(0).getNode();
148 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
150 unsigned i = 0, e = N->getNumOperands();
152 // Skip over all of the undef values.
153 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
156 // Do not accept an all-undef vector.
157 if (i == e) return false;
159 // Do not accept build_vectors that aren't all constants or which have non-0
161 SDValue Zero = N->getOperand(i);
162 if (isa<ConstantSDNode>(Zero)) {
163 if (!cast<ConstantSDNode>(Zero)->isNullValue())
165 } else if (isa<ConstantFPSDNode>(Zero)) {
166 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
171 // Okay, we have at least one 0 value, check to see if the rest match or are
173 for (++i; i != e; ++i)
174 if (N->getOperand(i) != Zero &&
175 N->getOperand(i).getOpcode() != ISD::UNDEF)
180 /// isScalarToVector - Return true if the specified node is a
181 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
182 /// element is not an undef.
183 bool ISD::isScalarToVector(const SDNode *N) {
184 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
187 if (N->getOpcode() != ISD::BUILD_VECTOR)
189 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
191 unsigned NumElems = N->getNumOperands();
192 for (unsigned i = 1; i < NumElems; ++i) {
193 SDValue V = N->getOperand(i);
194 if (V.getOpcode() != ISD::UNDEF)
201 /// isDebugLabel - Return true if the specified node represents a debug
202 /// label (i.e. ISD::DBG_LABEL or TargetInstrInfo::DBG_LABEL node).
203 bool ISD::isDebugLabel(const SDNode *N) {
205 if (N->getOpcode() == ISD::DBG_LABEL)
207 if (N->isMachineOpcode() &&
208 N->getMachineOpcode() == TargetInstrInfo::DBG_LABEL)
213 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
214 /// when given the operation for (X op Y).
215 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
216 // To perform this operation, we just need to swap the L and G bits of the
218 unsigned OldL = (Operation >> 2) & 1;
219 unsigned OldG = (Operation >> 1) & 1;
220 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
221 (OldL << 1) | // New G bit
222 (OldG << 2)); // New L bit.
225 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
226 /// 'op' is a valid SetCC operation.
227 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
228 unsigned Operation = Op;
230 Operation ^= 7; // Flip L, G, E bits, but not U.
232 Operation ^= 15; // Flip all of the condition bits.
234 if (Operation > ISD::SETTRUE2)
235 Operation &= ~8; // Don't let N and U bits get set.
237 return ISD::CondCode(Operation);
241 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
242 /// signed operation and 2 if the result is an unsigned comparison. Return zero
243 /// if the operation does not depend on the sign of the input (setne and seteq).
244 static int isSignedOp(ISD::CondCode Opcode) {
246 default: assert(0 && "Illegal integer setcc operation!");
248 case ISD::SETNE: return 0;
252 case ISD::SETGE: return 1;
256 case ISD::SETUGE: return 2;
260 /// getSetCCOrOperation - Return the result of a logical OR between different
261 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
262 /// returns SETCC_INVALID if it is not possible to represent the resultant
264 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
266 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
267 // Cannot fold a signed integer setcc with an unsigned integer setcc.
268 return ISD::SETCC_INVALID;
270 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
272 // If the N and U bits get set then the resultant comparison DOES suddenly
273 // care about orderedness, and is true when ordered.
274 if (Op > ISD::SETTRUE2)
275 Op &= ~16; // Clear the U bit if the N bit is set.
277 // Canonicalize illegal integer setcc's.
278 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
281 return ISD::CondCode(Op);
284 /// getSetCCAndOperation - Return the result of a logical AND between different
285 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
286 /// function returns zero if it is not possible to represent the resultant
288 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
290 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
291 // Cannot fold a signed setcc with an unsigned setcc.
292 return ISD::SETCC_INVALID;
294 // Combine all of the condition bits.
295 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
297 // Canonicalize illegal integer setcc's.
301 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
302 case ISD::SETOEQ: // SETEQ & SETU[LG]E
303 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
304 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
305 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
312 const TargetMachine &SelectionDAG::getTarget() const {
313 return MF->getTarget();
316 //===----------------------------------------------------------------------===//
317 // SDNode Profile Support
318 //===----------------------------------------------------------------------===//
320 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
322 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
326 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
327 /// solely with their pointer.
328 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
329 ID.AddPointer(VTList.VTs);
332 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
334 static void AddNodeIDOperands(FoldingSetNodeID &ID,
335 const SDValue *Ops, unsigned NumOps) {
336 for (; NumOps; --NumOps, ++Ops) {
337 ID.AddPointer(Ops->getNode());
338 ID.AddInteger(Ops->getResNo());
342 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
344 static void AddNodeIDOperands(FoldingSetNodeID &ID,
345 const SDUse *Ops, unsigned NumOps) {
346 for (; NumOps; --NumOps, ++Ops) {
347 ID.AddPointer(Ops->getNode());
348 ID.AddInteger(Ops->getResNo());
352 static void AddNodeIDNode(FoldingSetNodeID &ID,
353 unsigned short OpC, SDVTList VTList,
354 const SDValue *OpList, unsigned N) {
355 AddNodeIDOpcode(ID, OpC);
356 AddNodeIDValueTypes(ID, VTList);
357 AddNodeIDOperands(ID, OpList, N);
360 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
362 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
363 switch (N->getOpcode()) {
364 case ISD::TargetExternalSymbol:
365 case ISD::ExternalSymbol:
366 assert(0 && "Should only be used on nodes with operands");
367 default: break; // Normal nodes don't need extra info.
369 ID.AddInteger(cast<ARG_FLAGSSDNode>(N)->getArgFlags().getRawBits());
371 case ISD::TargetConstant:
373 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
375 case ISD::TargetConstantFP:
376 case ISD::ConstantFP: {
377 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
380 case ISD::TargetGlobalAddress:
381 case ISD::GlobalAddress:
382 case ISD::TargetGlobalTLSAddress:
383 case ISD::GlobalTLSAddress: {
384 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
385 ID.AddPointer(GA->getGlobal());
386 ID.AddInteger(GA->getOffset());
387 ID.AddInteger(GA->getTargetFlags());
390 case ISD::BasicBlock:
391 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
394 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
396 case ISD::DBG_STOPPOINT: {
397 const DbgStopPointSDNode *DSP = cast<DbgStopPointSDNode>(N);
398 ID.AddInteger(DSP->getLine());
399 ID.AddInteger(DSP->getColumn());
400 ID.AddPointer(DSP->getCompileUnit());
404 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
406 case ISD::MEMOPERAND: {
407 const MachineMemOperand &MO = cast<MemOperandSDNode>(N)->MO;
411 case ISD::FrameIndex:
412 case ISD::TargetFrameIndex:
413 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
416 case ISD::TargetJumpTable:
417 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
418 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
420 case ISD::ConstantPool:
421 case ISD::TargetConstantPool: {
422 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
423 ID.AddInteger(CP->getAlignment());
424 ID.AddInteger(CP->getOffset());
425 if (CP->isMachineConstantPoolEntry())
426 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
428 ID.AddPointer(CP->getConstVal());
429 ID.AddInteger(CP->getTargetFlags());
433 const CallSDNode *Call = cast<CallSDNode>(N);
434 ID.AddInteger(Call->getCallingConv());
435 ID.AddInteger(Call->isVarArg());
439 const LoadSDNode *LD = cast<LoadSDNode>(N);
440 ID.AddInteger(LD->getMemoryVT().getRawBits());
441 ID.AddInteger(LD->getRawSubclassData());
445 const StoreSDNode *ST = cast<StoreSDNode>(N);
446 ID.AddInteger(ST->getMemoryVT().getRawBits());
447 ID.AddInteger(ST->getRawSubclassData());
450 case ISD::ATOMIC_CMP_SWAP:
451 case ISD::ATOMIC_SWAP:
452 case ISD::ATOMIC_LOAD_ADD:
453 case ISD::ATOMIC_LOAD_SUB:
454 case ISD::ATOMIC_LOAD_AND:
455 case ISD::ATOMIC_LOAD_OR:
456 case ISD::ATOMIC_LOAD_XOR:
457 case ISD::ATOMIC_LOAD_NAND:
458 case ISD::ATOMIC_LOAD_MIN:
459 case ISD::ATOMIC_LOAD_MAX:
460 case ISD::ATOMIC_LOAD_UMIN:
461 case ISD::ATOMIC_LOAD_UMAX: {
462 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
463 ID.AddInteger(AT->getMemoryVT().getRawBits());
464 ID.AddInteger(AT->getRawSubclassData());
467 case ISD::VECTOR_SHUFFLE: {
468 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
469 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
471 ID.AddInteger(SVN->getMaskElt(i));
474 } // end switch (N->getOpcode())
477 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
479 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
480 AddNodeIDOpcode(ID, N->getOpcode());
481 // Add the return value info.
482 AddNodeIDValueTypes(ID, N->getVTList());
483 // Add the operand info.
484 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
486 // Handle SDNode leafs with special info.
487 AddNodeIDCustom(ID, N);
490 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
491 /// the CSE map that carries alignment, volatility, indexing mode, and
492 /// extension/truncation information.
494 static inline unsigned
495 encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM,
496 bool isVolatile, unsigned Alignment) {
497 assert((ConvType & 3) == ConvType &&
498 "ConvType may not require more than 2 bits!");
499 assert((AM & 7) == AM &&
500 "AM may not require more than 3 bits!");
504 ((Log2_32(Alignment) + 1) << 6);
507 //===----------------------------------------------------------------------===//
508 // SelectionDAG Class
509 //===----------------------------------------------------------------------===//
511 /// doNotCSE - Return true if CSE should not be performed for this node.
512 static bool doNotCSE(SDNode *N) {
513 if (N->getValueType(0) == MVT::Flag)
514 return true; // Never CSE anything that produces a flag.
516 switch (N->getOpcode()) {
518 case ISD::HANDLENODE:
520 case ISD::DBG_STOPPOINT:
523 return true; // Never CSE these nodes.
526 // Check that remaining values produced are not flags.
527 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
528 if (N->getValueType(i) == MVT::Flag)
529 return true; // Never CSE anything that produces a flag.
534 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
536 void SelectionDAG::RemoveDeadNodes() {
537 // Create a dummy node (which is not added to allnodes), that adds a reference
538 // to the root node, preventing it from being deleted.
539 HandleSDNode Dummy(getRoot());
541 SmallVector<SDNode*, 128> DeadNodes;
543 // Add all obviously-dead nodes to the DeadNodes worklist.
544 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
546 DeadNodes.push_back(I);
548 RemoveDeadNodes(DeadNodes);
550 // If the root changed (e.g. it was a dead load, update the root).
551 setRoot(Dummy.getValue());
554 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
555 /// given list, and any nodes that become unreachable as a result.
556 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
557 DAGUpdateListener *UpdateListener) {
559 // Process the worklist, deleting the nodes and adding their uses to the
561 while (!DeadNodes.empty()) {
562 SDNode *N = DeadNodes.pop_back_val();
565 UpdateListener->NodeDeleted(N, 0);
567 // Take the node out of the appropriate CSE map.
568 RemoveNodeFromCSEMaps(N);
570 // Next, brutally remove the operand list. This is safe to do, as there are
571 // no cycles in the graph.
572 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
574 SDNode *Operand = Use.getNode();
577 // Now that we removed this operand, see if there are no uses of it left.
578 if (Operand->use_empty())
579 DeadNodes.push_back(Operand);
586 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
587 SmallVector<SDNode*, 16> DeadNodes(1, N);
588 RemoveDeadNodes(DeadNodes, UpdateListener);
591 void SelectionDAG::DeleteNode(SDNode *N) {
592 // First take this out of the appropriate CSE map.
593 RemoveNodeFromCSEMaps(N);
595 // Finally, remove uses due to operands of this node, remove from the
596 // AllNodes list, and delete the node.
597 DeleteNodeNotInCSEMaps(N);
600 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
601 assert(N != AllNodes.begin() && "Cannot delete the entry node!");
602 assert(N->use_empty() && "Cannot delete a node that is not dead!");
604 // Drop all of the operands and decrement used node's use counts.
610 void SelectionDAG::DeallocateNode(SDNode *N) {
611 if (N->OperandsNeedDelete)
612 delete[] N->OperandList;
614 // Set the opcode to DELETED_NODE to help catch bugs when node
615 // memory is reallocated.
616 N->NodeType = ISD::DELETED_NODE;
618 NodeAllocator.Deallocate(AllNodes.remove(N));
621 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
622 /// correspond to it. This is useful when we're about to delete or repurpose
623 /// the node. We don't want future request for structurally identical nodes
624 /// to return N anymore.
625 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
627 switch (N->getOpcode()) {
628 case ISD::EntryToken:
629 assert(0 && "EntryToken should not be in CSEMaps!");
631 case ISD::HANDLENODE: return false; // noop.
633 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
634 "Cond code doesn't exist!");
635 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
636 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
638 case ISD::ExternalSymbol:
639 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
641 case ISD::TargetExternalSymbol: {
642 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
643 Erased = TargetExternalSymbols.erase(
644 std::pair<std::string,unsigned char>(ESN->getSymbol(),
645 ESN->getTargetFlags()));
648 case ISD::VALUETYPE: {
649 MVT VT = cast<VTSDNode>(N)->getVT();
650 if (VT.isExtended()) {
651 Erased = ExtendedValueTypeNodes.erase(VT);
653 Erased = ValueTypeNodes[VT.getSimpleVT()] != 0;
654 ValueTypeNodes[VT.getSimpleVT()] = 0;
659 // Remove it from the CSE Map.
660 Erased = CSEMap.RemoveNode(N);
664 // Verify that the node was actually in one of the CSE maps, unless it has a
665 // flag result (which cannot be CSE'd) or is one of the special cases that are
666 // not subject to CSE.
667 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
668 !N->isMachineOpcode() && !doNotCSE(N)) {
671 assert(0 && "Node is not in map!");
677 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
678 /// maps and modified in place. Add it back to the CSE maps, unless an identical
679 /// node already exists, in which case transfer all its users to the existing
680 /// node. This transfer can potentially trigger recursive merging.
683 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N,
684 DAGUpdateListener *UpdateListener) {
685 // For node types that aren't CSE'd, just act as if no identical node
688 SDNode *Existing = CSEMap.GetOrInsertNode(N);
690 // If there was already an existing matching node, use ReplaceAllUsesWith
691 // to replace the dead one with the existing one. This can cause
692 // recursive merging of other unrelated nodes down the line.
693 ReplaceAllUsesWith(N, Existing, UpdateListener);
695 // N is now dead. Inform the listener if it exists and delete it.
697 UpdateListener->NodeDeleted(N, Existing);
698 DeleteNodeNotInCSEMaps(N);
703 // If the node doesn't already exist, we updated it. Inform a listener if
706 UpdateListener->NodeUpdated(N);
709 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
710 /// were replaced with those specified. If this node is never memoized,
711 /// return null, otherwise return a pointer to the slot it would take. If a
712 /// node already exists with these operands, the slot will be non-null.
713 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
718 SDValue Ops[] = { Op };
720 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
721 AddNodeIDCustom(ID, N);
722 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
725 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
726 /// were replaced with those specified. If this node is never memoized,
727 /// return null, otherwise return a pointer to the slot it would take. If a
728 /// node already exists with these operands, the slot will be non-null.
729 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
730 SDValue Op1, SDValue Op2,
735 SDValue Ops[] = { Op1, Op2 };
737 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
738 AddNodeIDCustom(ID, N);
739 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
743 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
744 /// were replaced with those specified. If this node is never memoized,
745 /// return null, otherwise return a pointer to the slot it would take. If a
746 /// node already exists with these operands, the slot will be non-null.
747 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
748 const SDValue *Ops,unsigned NumOps,
754 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
755 AddNodeIDCustom(ID, N);
756 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
759 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
760 void SelectionDAG::VerifyNode(SDNode *N) {
761 switch (N->getOpcode()) {
764 case ISD::BUILD_PAIR: {
765 MVT VT = N->getValueType(0);
766 assert(N->getNumValues() == 1 && "Too many results!");
767 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
768 "Wrong return type!");
769 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
770 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
771 "Mismatched operand types!");
772 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
773 "Wrong operand type!");
774 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
775 "Wrong return type size");
778 case ISD::BUILD_VECTOR: {
779 assert(N->getNumValues() == 1 && "Too many results!");
780 assert(N->getValueType(0).isVector() && "Wrong return type!");
781 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
782 "Wrong number of operands!");
783 MVT EltVT = N->getValueType(0).getVectorElementType();
784 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
785 assert((I->getValueType() == EltVT ||
786 (EltVT.isInteger() && I->getValueType().isInteger() &&
787 EltVT.bitsLE(I->getValueType()))) &&
788 "Wrong operand type!");
794 /// getMVTAlignment - Compute the default alignment value for the
797 unsigned SelectionDAG::getMVTAlignment(MVT VT) const {
798 const Type *Ty = VT == MVT::iPTR ?
799 PointerType::get(Type::Int8Ty, 0) :
800 VT.getTypeForMVT(*Context);
802 return TLI.getTargetData()->getABITypeAlignment(Ty);
805 // EntryNode could meaningfully have debug info if we can find it...
806 SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli)
807 : TLI(tli), FLI(fli), DW(0),
808 EntryNode(ISD::EntryToken, DebugLoc::getUnknownLoc(),
809 getVTList(MVT::Other)), Root(getEntryNode()) {
810 AllNodes.push_back(&EntryNode);
813 void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi,
818 Context = mf.getFunction()->getContext();
821 SelectionDAG::~SelectionDAG() {
825 void SelectionDAG::allnodes_clear() {
826 assert(&*AllNodes.begin() == &EntryNode);
827 AllNodes.remove(AllNodes.begin());
828 while (!AllNodes.empty())
829 DeallocateNode(AllNodes.begin());
832 void SelectionDAG::clear() {
834 OperandAllocator.Reset();
837 ExtendedValueTypeNodes.clear();
838 ExternalSymbols.clear();
839 TargetExternalSymbols.clear();
840 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
841 static_cast<CondCodeSDNode*>(0));
842 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
843 static_cast<SDNode*>(0));
845 EntryNode.UseList = 0;
846 AllNodes.push_back(&EntryNode);
847 Root = getEntryNode();
850 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, MVT VT) {
851 if (Op.getValueType() == VT) return Op;
852 APInt Imm = APInt::getLowBitsSet(Op.getValueSizeInBits(),
854 return getNode(ISD::AND, DL, Op.getValueType(), Op,
855 getConstant(Imm, Op.getValueType()));
858 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
860 SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, MVT VT) {
861 MVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
863 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
864 return getNode(ISD::XOR, DL, VT, Val, NegOne);
867 SDValue SelectionDAG::getConstant(uint64_t Val, MVT VT, bool isT) {
868 MVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
869 assert((EltVT.getSizeInBits() >= 64 ||
870 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
871 "getConstant with a uint64_t value that doesn't fit in the type!");
872 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
875 SDValue SelectionDAG::getConstant(const APInt &Val, MVT VT, bool isT) {
876 return getConstant(*ConstantInt::get(Val), VT, isT);
879 SDValue SelectionDAG::getConstant(const ConstantInt &Val, MVT VT, bool isT) {
880 assert(VT.isInteger() && "Cannot create FP integer constant!");
882 MVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
883 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
884 "APInt size does not match type size!");
886 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
888 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
892 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
894 return SDValue(N, 0);
896 N = NodeAllocator.Allocate<ConstantSDNode>();
897 new (N) ConstantSDNode(isT, &Val, EltVT);
898 CSEMap.InsertNode(N, IP);
899 AllNodes.push_back(N);
902 SDValue Result(N, 0);
904 SmallVector<SDValue, 8> Ops;
905 Ops.assign(VT.getVectorNumElements(), Result);
906 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
907 VT, &Ops[0], Ops.size());
912 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
913 return getConstant(Val, TLI.getPointerTy(), isTarget);
917 SDValue SelectionDAG::getConstantFP(const APFloat& V, MVT VT, bool isTarget) {
918 return getConstantFP(*ConstantFP::get(V), VT, isTarget);
921 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, MVT VT, bool isTarget){
922 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
925 VT.isVector() ? VT.getVectorElementType() : VT;
927 // Do the map lookup using the actual bit pattern for the floating point
928 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
929 // we don't have issues with SNANs.
930 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
932 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
936 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
938 return SDValue(N, 0);
940 N = NodeAllocator.Allocate<ConstantFPSDNode>();
941 new (N) ConstantFPSDNode(isTarget, &V, EltVT);
942 CSEMap.InsertNode(N, IP);
943 AllNodes.push_back(N);
946 SDValue Result(N, 0);
948 SmallVector<SDValue, 8> Ops;
949 Ops.assign(VT.getVectorNumElements(), Result);
950 // FIXME DebugLoc info might be appropriate here
951 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
952 VT, &Ops[0], Ops.size());
957 SDValue SelectionDAG::getConstantFP(double Val, MVT VT, bool isTarget) {
959 VT.isVector() ? VT.getVectorElementType() : VT;
961 return getConstantFP(APFloat((float)Val), VT, isTarget);
963 return getConstantFP(APFloat(Val), VT, isTarget);
966 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
967 MVT VT, int64_t Offset,
969 unsigned char TargetFlags) {
970 assert((TargetFlags == 0 || isTargetGA) &&
971 "Cannot set target flags on target-independent globals");
973 // Truncate (with sign-extension) the offset value to the pointer size.
974 unsigned BitWidth = TLI.getPointerTy().getSizeInBits();
976 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
978 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
980 // If GV is an alias then use the aliasee for determining thread-localness.
981 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
982 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
986 if (GVar && GVar->isThreadLocal())
987 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
989 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
992 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
994 ID.AddInteger(Offset);
995 ID.AddInteger(TargetFlags);
997 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
998 return SDValue(E, 0);
999 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>();
1000 new (N) GlobalAddressSDNode(Opc, GV, VT, Offset, TargetFlags);
1001 CSEMap.InsertNode(N, IP);
1002 AllNodes.push_back(N);
1003 return SDValue(N, 0);
1006 SDValue SelectionDAG::getFrameIndex(int FI, MVT VT, bool isTarget) {
1007 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1008 FoldingSetNodeID ID;
1009 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1012 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1013 return SDValue(E, 0);
1014 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>();
1015 new (N) FrameIndexSDNode(FI, VT, isTarget);
1016 CSEMap.InsertNode(N, IP);
1017 AllNodes.push_back(N);
1018 return SDValue(N, 0);
1021 SDValue SelectionDAG::getJumpTable(int JTI, MVT VT, bool isTarget,
1022 unsigned char TargetFlags) {
1023 assert((TargetFlags == 0 || isTarget) &&
1024 "Cannot set target flags on target-independent jump tables");
1025 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1026 FoldingSetNodeID ID;
1027 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1029 ID.AddInteger(TargetFlags);
1031 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1032 return SDValue(E, 0);
1033 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>();
1034 new (N) JumpTableSDNode(JTI, VT, isTarget, TargetFlags);
1035 CSEMap.InsertNode(N, IP);
1036 AllNodes.push_back(N);
1037 return SDValue(N, 0);
1040 SDValue SelectionDAG::getConstantPool(Constant *C, MVT VT,
1041 unsigned Alignment, int Offset,
1043 unsigned char TargetFlags) {
1044 assert((TargetFlags == 0 || isTarget) &&
1045 "Cannot set target flags on target-independent globals");
1047 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1048 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1049 FoldingSetNodeID ID;
1050 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1051 ID.AddInteger(Alignment);
1052 ID.AddInteger(Offset);
1054 ID.AddInteger(TargetFlags);
1056 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1057 return SDValue(E, 0);
1058 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1059 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1060 CSEMap.InsertNode(N, IP);
1061 AllNodes.push_back(N);
1062 return SDValue(N, 0);
1066 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, MVT VT,
1067 unsigned Alignment, int Offset,
1069 unsigned char TargetFlags) {
1070 assert((TargetFlags == 0 || isTarget) &&
1071 "Cannot set target flags on target-independent globals");
1073 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1074 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1075 FoldingSetNodeID ID;
1076 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1077 ID.AddInteger(Alignment);
1078 ID.AddInteger(Offset);
1079 C->AddSelectionDAGCSEId(ID);
1080 ID.AddInteger(TargetFlags);
1082 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1083 return SDValue(E, 0);
1084 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1085 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1086 CSEMap.InsertNode(N, IP);
1087 AllNodes.push_back(N);
1088 return SDValue(N, 0);
1091 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1092 FoldingSetNodeID ID;
1093 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1096 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1097 return SDValue(E, 0);
1098 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>();
1099 new (N) BasicBlockSDNode(MBB);
1100 CSEMap.InsertNode(N, IP);
1101 AllNodes.push_back(N);
1102 return SDValue(N, 0);
1105 SDValue SelectionDAG::getArgFlags(ISD::ArgFlagsTy Flags) {
1106 FoldingSetNodeID ID;
1107 AddNodeIDNode(ID, ISD::ARG_FLAGS, getVTList(MVT::Other), 0, 0);
1108 ID.AddInteger(Flags.getRawBits());
1110 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1111 return SDValue(E, 0);
1112 SDNode *N = NodeAllocator.Allocate<ARG_FLAGSSDNode>();
1113 new (N) ARG_FLAGSSDNode(Flags);
1114 CSEMap.InsertNode(N, IP);
1115 AllNodes.push_back(N);
1116 return SDValue(N, 0);
1119 SDValue SelectionDAG::getValueType(MVT VT) {
1120 if (VT.isSimple() && (unsigned)VT.getSimpleVT() >= ValueTypeNodes.size())
1121 ValueTypeNodes.resize(VT.getSimpleVT()+1);
1123 SDNode *&N = VT.isExtended() ?
1124 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT()];
1126 if (N) return SDValue(N, 0);
1127 N = NodeAllocator.Allocate<VTSDNode>();
1128 new (N) VTSDNode(VT);
1129 AllNodes.push_back(N);
1130 return SDValue(N, 0);
1133 SDValue SelectionDAG::getExternalSymbol(const char *Sym, MVT VT) {
1134 SDNode *&N = ExternalSymbols[Sym];
1135 if (N) return SDValue(N, 0);
1136 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1137 new (N) ExternalSymbolSDNode(false, Sym, 0, VT);
1138 AllNodes.push_back(N);
1139 return SDValue(N, 0);
1142 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, MVT VT,
1143 unsigned char TargetFlags) {
1145 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym,
1147 if (N) return SDValue(N, 0);
1148 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1149 new (N) ExternalSymbolSDNode(true, Sym, TargetFlags, VT);
1150 AllNodes.push_back(N);
1151 return SDValue(N, 0);
1154 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1155 if ((unsigned)Cond >= CondCodeNodes.size())
1156 CondCodeNodes.resize(Cond+1);
1158 if (CondCodeNodes[Cond] == 0) {
1159 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>();
1160 new (N) CondCodeSDNode(Cond);
1161 CondCodeNodes[Cond] = N;
1162 AllNodes.push_back(N);
1164 return SDValue(CondCodeNodes[Cond], 0);
1167 // commuteShuffle - swaps the values of N1 and N2, and swaps all indices in
1168 // the shuffle mask M that point at N1 to point at N2, and indices that point
1169 // N2 to point at N1.
1170 static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) {
1172 int NElts = M.size();
1173 for (int i = 0; i != NElts; ++i) {
1181 SDValue SelectionDAG::getVectorShuffle(MVT VT, DebugLoc dl, SDValue N1,
1182 SDValue N2, const int *Mask) {
1183 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE");
1184 assert(VT.isVector() && N1.getValueType().isVector() &&
1185 "Vector Shuffle VTs must be a vectors");
1186 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType()
1187 && "Vector Shuffle VTs must have same element type");
1189 // Canonicalize shuffle undef, undef -> undef
1190 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF)
1191 return getUNDEF(VT);
1193 // Validate that all indices in Mask are within the range of the elements
1194 // input to the shuffle.
1195 unsigned NElts = VT.getVectorNumElements();
1196 SmallVector<int, 8> MaskVec;
1197 for (unsigned i = 0; i != NElts; ++i) {
1198 assert(Mask[i] < (int)(NElts * 2) && "Index out of range");
1199 MaskVec.push_back(Mask[i]);
1202 // Canonicalize shuffle v, v -> v, undef
1205 for (unsigned i = 0; i != NElts; ++i)
1206 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts;
1209 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1210 if (N1.getOpcode() == ISD::UNDEF)
1211 commuteShuffle(N1, N2, MaskVec);
1213 // Canonicalize all index into lhs, -> shuffle lhs, undef
1214 // Canonicalize all index into rhs, -> shuffle rhs, undef
1215 bool AllLHS = true, AllRHS = true;
1216 bool N2Undef = N2.getOpcode() == ISD::UNDEF;
1217 for (unsigned i = 0; i != NElts; ++i) {
1218 if (MaskVec[i] >= (int)NElts) {
1223 } else if (MaskVec[i] >= 0) {
1227 if (AllLHS && AllRHS)
1228 return getUNDEF(VT);
1229 if (AllLHS && !N2Undef)
1233 commuteShuffle(N1, N2, MaskVec);
1236 // If Identity shuffle, or all shuffle in to undef, return that node.
1237 bool AllUndef = true;
1238 bool Identity = true;
1239 for (unsigned i = 0; i != NElts; ++i) {
1240 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false;
1241 if (MaskVec[i] >= 0) AllUndef = false;
1243 if (Identity && NElts == N1.getValueType().getVectorNumElements())
1246 return getUNDEF(VT);
1248 FoldingSetNodeID ID;
1249 SDValue Ops[2] = { N1, N2 };
1250 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2);
1251 for (unsigned i = 0; i != NElts; ++i)
1252 ID.AddInteger(MaskVec[i]);
1255 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1256 return SDValue(E, 0);
1258 // Allocate the mask array for the node out of the BumpPtrAllocator, since
1259 // SDNode doesn't have access to it. This memory will be "leaked" when
1260 // the node is deallocated, but recovered when the NodeAllocator is released.
1261 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
1262 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int));
1264 ShuffleVectorSDNode *N = NodeAllocator.Allocate<ShuffleVectorSDNode>();
1265 new (N) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc);
1266 CSEMap.InsertNode(N, IP);
1267 AllNodes.push_back(N);
1268 return SDValue(N, 0);
1271 SDValue SelectionDAG::getConvertRndSat(MVT VT, DebugLoc dl,
1272 SDValue Val, SDValue DTy,
1273 SDValue STy, SDValue Rnd, SDValue Sat,
1274 ISD::CvtCode Code) {
1275 // If the src and dest types are the same and the conversion is between
1276 // integer types of the same sign or two floats, no conversion is necessary.
1278 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF))
1281 FoldingSetNodeID ID;
1283 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1284 return SDValue(E, 0);
1285 CvtRndSatSDNode *N = NodeAllocator.Allocate<CvtRndSatSDNode>();
1286 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat };
1287 new (N) CvtRndSatSDNode(VT, dl, Ops, 5, Code);
1288 CSEMap.InsertNode(N, IP);
1289 AllNodes.push_back(N);
1290 return SDValue(N, 0);
1293 SDValue SelectionDAG::getRegister(unsigned RegNo, MVT VT) {
1294 FoldingSetNodeID ID;
1295 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1296 ID.AddInteger(RegNo);
1298 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1299 return SDValue(E, 0);
1300 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>();
1301 new (N) RegisterSDNode(RegNo, VT);
1302 CSEMap.InsertNode(N, IP);
1303 AllNodes.push_back(N);
1304 return SDValue(N, 0);
1307 SDValue SelectionDAG::getDbgStopPoint(DebugLoc DL, SDValue Root,
1308 unsigned Line, unsigned Col,
1310 SDNode *N = NodeAllocator.Allocate<DbgStopPointSDNode>();
1311 new (N) DbgStopPointSDNode(Root, Line, Col, CU);
1313 AllNodes.push_back(N);
1314 return SDValue(N, 0);
1317 SDValue SelectionDAG::getLabel(unsigned Opcode, DebugLoc dl,
1320 FoldingSetNodeID ID;
1321 SDValue Ops[] = { Root };
1322 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), &Ops[0], 1);
1323 ID.AddInteger(LabelID);
1325 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1326 return SDValue(E, 0);
1327 SDNode *N = NodeAllocator.Allocate<LabelSDNode>();
1328 new (N) LabelSDNode(Opcode, dl, Root, LabelID);
1329 CSEMap.InsertNode(N, IP);
1330 AllNodes.push_back(N);
1331 return SDValue(N, 0);
1334 SDValue SelectionDAG::getSrcValue(const Value *V) {
1335 assert((!V || isa<PointerType>(V->getType())) &&
1336 "SrcValue is not a pointer?");
1338 FoldingSetNodeID ID;
1339 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1343 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1344 return SDValue(E, 0);
1346 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>();
1347 new (N) SrcValueSDNode(V);
1348 CSEMap.InsertNode(N, IP);
1349 AllNodes.push_back(N);
1350 return SDValue(N, 0);
1353 SDValue SelectionDAG::getMemOperand(const MachineMemOperand &MO) {
1355 const Value *v = MO.getValue();
1356 assert((!v || isa<PointerType>(v->getType())) &&
1357 "SrcValue is not a pointer?");
1360 FoldingSetNodeID ID;
1361 AddNodeIDNode(ID, ISD::MEMOPERAND, getVTList(MVT::Other), 0, 0);
1365 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1366 return SDValue(E, 0);
1368 SDNode *N = NodeAllocator.Allocate<MemOperandSDNode>();
1369 new (N) MemOperandSDNode(MO);
1370 CSEMap.InsertNode(N, IP);
1371 AllNodes.push_back(N);
1372 return SDValue(N, 0);
1375 /// getShiftAmountOperand - Return the specified value casted to
1376 /// the target's desired shift amount type.
1377 SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) {
1378 MVT OpTy = Op.getValueType();
1379 MVT ShTy = TLI.getShiftAmountTy();
1380 if (OpTy == ShTy || OpTy.isVector()) return Op;
1382 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
1383 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op);
1386 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1387 /// specified value type.
1388 SDValue SelectionDAG::CreateStackTemporary(MVT VT, unsigned minAlign) {
1389 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1390 unsigned ByteSize = VT.getStoreSizeInBits()/8;
1391 const Type *Ty = VT.getTypeForMVT(*Context);
1392 unsigned StackAlign =
1393 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1395 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign);
1396 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1399 /// CreateStackTemporary - Create a stack temporary suitable for holding
1400 /// either of the specified value types.
1401 SDValue SelectionDAG::CreateStackTemporary(MVT VT1, MVT VT2) {
1402 unsigned Bytes = std::max(VT1.getStoreSizeInBits(),
1403 VT2.getStoreSizeInBits())/8;
1404 const Type *Ty1 = VT1.getTypeForMVT(*Context);
1405 const Type *Ty2 = VT2.getTypeForMVT(*Context);
1406 const TargetData *TD = TLI.getTargetData();
1407 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1),
1408 TD->getPrefTypeAlignment(Ty2));
1410 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1411 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align);
1412 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1415 SDValue SelectionDAG::FoldSetCC(MVT VT, SDValue N1,
1416 SDValue N2, ISD::CondCode Cond, DebugLoc dl) {
1417 // These setcc operations always fold.
1421 case ISD::SETFALSE2: return getConstant(0, VT);
1423 case ISD::SETTRUE2: return getConstant(1, VT);
1435 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1439 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1440 const APInt &C2 = N2C->getAPIntValue();
1441 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1442 const APInt &C1 = N1C->getAPIntValue();
1445 default: assert(0 && "Unknown integer setcc!");
1446 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1447 case ISD::SETNE: return getConstant(C1 != C2, VT);
1448 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1449 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1450 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1451 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1452 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1453 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1454 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1455 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1459 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1460 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1461 // No compile time operations on this type yet.
1462 if (N1C->getValueType(0) == MVT::ppcf128)
1465 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1468 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1469 return getUNDEF(VT);
1471 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1472 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1473 return getUNDEF(VT);
1475 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1476 R==APFloat::cmpLessThan, VT);
1477 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1478 return getUNDEF(VT);
1480 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1481 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1482 return getUNDEF(VT);
1484 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1485 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1486 return getUNDEF(VT);
1488 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1489 R==APFloat::cmpEqual, VT);
1490 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1491 return getUNDEF(VT);
1493 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1494 R==APFloat::cmpEqual, VT);
1495 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1496 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1497 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1498 R==APFloat::cmpEqual, VT);
1499 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1500 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1501 R==APFloat::cmpLessThan, VT);
1502 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1503 R==APFloat::cmpUnordered, VT);
1504 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1505 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1508 // Ensure that the constant occurs on the RHS.
1509 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1513 // Could not fold it.
1517 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1518 /// use this predicate to simplify operations downstream.
1519 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1520 // This predicate is not safe for vector operations.
1521 if (Op.getValueType().isVector())
1524 unsigned BitWidth = Op.getValueSizeInBits();
1525 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1528 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1529 /// this predicate to simplify operations downstream. Mask is known to be zero
1530 /// for bits that V cannot have.
1531 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1532 unsigned Depth) const {
1533 APInt KnownZero, KnownOne;
1534 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1535 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1536 return (KnownZero & Mask) == Mask;
1539 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1540 /// known to be either zero or one and return them in the KnownZero/KnownOne
1541 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1543 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1544 APInt &KnownZero, APInt &KnownOne,
1545 unsigned Depth) const {
1546 unsigned BitWidth = Mask.getBitWidth();
1547 assert(BitWidth == Op.getValueType().getSizeInBits() &&
1548 "Mask size mismatches value type size!");
1550 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1551 if (Depth == 6 || Mask == 0)
1552 return; // Limit search depth.
1554 APInt KnownZero2, KnownOne2;
1556 switch (Op.getOpcode()) {
1558 // We know all of the bits for a constant!
1559 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1560 KnownZero = ~KnownOne & Mask;
1563 // If either the LHS or the RHS are Zero, the result is zero.
1564 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1565 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1566 KnownZero2, KnownOne2, Depth+1);
1567 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1568 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1570 // Output known-1 bits are only known if set in both the LHS & RHS.
1571 KnownOne &= KnownOne2;
1572 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1573 KnownZero |= KnownZero2;
1576 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1577 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1578 KnownZero2, KnownOne2, Depth+1);
1579 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1580 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1582 // Output known-0 bits are only known if clear in both the LHS & RHS.
1583 KnownZero &= KnownZero2;
1584 // Output known-1 are known to be set if set in either the LHS | RHS.
1585 KnownOne |= KnownOne2;
1588 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1589 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1590 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1591 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1593 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1594 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1595 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1596 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1597 KnownZero = KnownZeroOut;
1601 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1602 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1603 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1604 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1605 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1607 // If low bits are zero in either operand, output low known-0 bits.
1608 // Also compute a conserative estimate for high known-0 bits.
1609 // More trickiness is possible, but this is sufficient for the
1610 // interesting case of alignment computation.
1612 unsigned TrailZ = KnownZero.countTrailingOnes() +
1613 KnownZero2.countTrailingOnes();
1614 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1615 KnownZero2.countLeadingOnes(),
1616 BitWidth) - BitWidth;
1618 TrailZ = std::min(TrailZ, BitWidth);
1619 LeadZ = std::min(LeadZ, BitWidth);
1620 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1621 APInt::getHighBitsSet(BitWidth, LeadZ);
1626 // For the purposes of computing leading zeros we can conservatively
1627 // treat a udiv as a logical right shift by the power of 2 known to
1628 // be less than the denominator.
1629 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1630 ComputeMaskedBits(Op.getOperand(0),
1631 AllOnes, KnownZero2, KnownOne2, Depth+1);
1632 unsigned LeadZ = KnownZero2.countLeadingOnes();
1636 ComputeMaskedBits(Op.getOperand(1),
1637 AllOnes, KnownZero2, KnownOne2, Depth+1);
1638 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1639 if (RHSUnknownLeadingOnes != BitWidth)
1640 LeadZ = std::min(BitWidth,
1641 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1643 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1647 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1648 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1649 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1650 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1652 // Only known if known in both the LHS and RHS.
1653 KnownOne &= KnownOne2;
1654 KnownZero &= KnownZero2;
1656 case ISD::SELECT_CC:
1657 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1658 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1659 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1660 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1662 // Only known if known in both the LHS and RHS.
1663 KnownOne &= KnownOne2;
1664 KnownZero &= KnownZero2;
1672 if (Op.getResNo() != 1)
1674 // The boolean result conforms to getBooleanContents. Fall through.
1676 // If we know the result of a setcc has the top bits zero, use this info.
1677 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent &&
1679 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1682 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1683 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1684 unsigned ShAmt = SA->getZExtValue();
1686 // If the shift count is an invalid immediate, don't do anything.
1687 if (ShAmt >= BitWidth)
1690 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt),
1691 KnownZero, KnownOne, Depth+1);
1692 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1693 KnownZero <<= ShAmt;
1695 // low bits known zero.
1696 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1700 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1701 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1702 unsigned ShAmt = SA->getZExtValue();
1704 // If the shift count is an invalid immediate, don't do anything.
1705 if (ShAmt >= BitWidth)
1708 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1709 KnownZero, KnownOne, Depth+1);
1710 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1711 KnownZero = KnownZero.lshr(ShAmt);
1712 KnownOne = KnownOne.lshr(ShAmt);
1714 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1715 KnownZero |= HighBits; // High bits known zero.
1719 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1720 unsigned ShAmt = SA->getZExtValue();
1722 // If the shift count is an invalid immediate, don't do anything.
1723 if (ShAmt >= BitWidth)
1726 APInt InDemandedMask = (Mask << ShAmt);
1727 // If any of the demanded bits are produced by the sign extension, we also
1728 // demand the input sign bit.
1729 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1730 if (HighBits.getBoolValue())
1731 InDemandedMask |= APInt::getSignBit(BitWidth);
1733 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1735 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1736 KnownZero = KnownZero.lshr(ShAmt);
1737 KnownOne = KnownOne.lshr(ShAmt);
1739 // Handle the sign bits.
1740 APInt SignBit = APInt::getSignBit(BitWidth);
1741 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1743 if (KnownZero.intersects(SignBit)) {
1744 KnownZero |= HighBits; // New bits are known zero.
1745 } else if (KnownOne.intersects(SignBit)) {
1746 KnownOne |= HighBits; // New bits are known one.
1750 case ISD::SIGN_EXTEND_INREG: {
1751 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1752 unsigned EBits = EVT.getSizeInBits();
1754 // Sign extension. Compute the demanded bits in the result that are not
1755 // present in the input.
1756 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1758 APInt InSignBit = APInt::getSignBit(EBits);
1759 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1761 // If the sign extended bits are demanded, we know that the sign
1763 InSignBit.zext(BitWidth);
1764 if (NewBits.getBoolValue())
1765 InputDemandedBits |= InSignBit;
1767 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1768 KnownZero, KnownOne, Depth+1);
1769 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1771 // If the sign bit of the input is known set or clear, then we know the
1772 // top bits of the result.
1773 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1774 KnownZero |= NewBits;
1775 KnownOne &= ~NewBits;
1776 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1777 KnownOne |= NewBits;
1778 KnownZero &= ~NewBits;
1779 } else { // Input sign bit unknown
1780 KnownZero &= ~NewBits;
1781 KnownOne &= ~NewBits;
1788 unsigned LowBits = Log2_32(BitWidth)+1;
1789 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1794 if (ISD::isZEXTLoad(Op.getNode())) {
1795 LoadSDNode *LD = cast<LoadSDNode>(Op);
1796 MVT VT = LD->getMemoryVT();
1797 unsigned MemBits = VT.getSizeInBits();
1798 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1802 case ISD::ZERO_EXTEND: {
1803 MVT InVT = Op.getOperand(0).getValueType();
1804 unsigned InBits = InVT.getSizeInBits();
1805 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1806 APInt InMask = Mask;
1807 InMask.trunc(InBits);
1808 KnownZero.trunc(InBits);
1809 KnownOne.trunc(InBits);
1810 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1811 KnownZero.zext(BitWidth);
1812 KnownOne.zext(BitWidth);
1813 KnownZero |= NewBits;
1816 case ISD::SIGN_EXTEND: {
1817 MVT InVT = Op.getOperand(0).getValueType();
1818 unsigned InBits = InVT.getSizeInBits();
1819 APInt InSignBit = APInt::getSignBit(InBits);
1820 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1821 APInt InMask = Mask;
1822 InMask.trunc(InBits);
1824 // If any of the sign extended bits are demanded, we know that the sign
1825 // bit is demanded. Temporarily set this bit in the mask for our callee.
1826 if (NewBits.getBoolValue())
1827 InMask |= InSignBit;
1829 KnownZero.trunc(InBits);
1830 KnownOne.trunc(InBits);
1831 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1833 // Note if the sign bit is known to be zero or one.
1834 bool SignBitKnownZero = KnownZero.isNegative();
1835 bool SignBitKnownOne = KnownOne.isNegative();
1836 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1837 "Sign bit can't be known to be both zero and one!");
1839 // If the sign bit wasn't actually demanded by our caller, we don't
1840 // want it set in the KnownZero and KnownOne result values. Reset the
1841 // mask and reapply it to the result values.
1843 InMask.trunc(InBits);
1844 KnownZero &= InMask;
1847 KnownZero.zext(BitWidth);
1848 KnownOne.zext(BitWidth);
1850 // If the sign bit is known zero or one, the top bits match.
1851 if (SignBitKnownZero)
1852 KnownZero |= NewBits;
1853 else if (SignBitKnownOne)
1854 KnownOne |= NewBits;
1857 case ISD::ANY_EXTEND: {
1858 MVT InVT = Op.getOperand(0).getValueType();
1859 unsigned InBits = InVT.getSizeInBits();
1860 APInt InMask = Mask;
1861 InMask.trunc(InBits);
1862 KnownZero.trunc(InBits);
1863 KnownOne.trunc(InBits);
1864 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1865 KnownZero.zext(BitWidth);
1866 KnownOne.zext(BitWidth);
1869 case ISD::TRUNCATE: {
1870 MVT InVT = Op.getOperand(0).getValueType();
1871 unsigned InBits = InVT.getSizeInBits();
1872 APInt InMask = Mask;
1873 InMask.zext(InBits);
1874 KnownZero.zext(InBits);
1875 KnownOne.zext(InBits);
1876 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1877 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1878 KnownZero.trunc(BitWidth);
1879 KnownOne.trunc(BitWidth);
1882 case ISD::AssertZext: {
1883 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1884 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1885 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1887 KnownZero |= (~InMask) & Mask;
1891 // All bits are zero except the low bit.
1892 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1896 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1897 // We know that the top bits of C-X are clear if X contains less bits
1898 // than C (i.e. no wrap-around can happen). For example, 20-X is
1899 // positive if we can prove that X is >= 0 and < 16.
1900 if (CLHS->getAPIntValue().isNonNegative()) {
1901 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1902 // NLZ can't be BitWidth with no sign bit
1903 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1904 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1907 // If all of the MaskV bits are known to be zero, then we know the
1908 // output top bits are zero, because we now know that the output is
1910 if ((KnownZero2 & MaskV) == MaskV) {
1911 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1912 // Top bits known zero.
1913 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1920 // Output known-0 bits are known if clear or set in both the low clear bits
1921 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1922 // low 3 bits clear.
1923 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
1924 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1925 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1926 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1928 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1929 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1930 KnownZeroOut = std::min(KnownZeroOut,
1931 KnownZero2.countTrailingOnes());
1933 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1937 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1938 const APInt &RA = Rem->getAPIntValue();
1939 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1940 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1941 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1942 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1944 // If the sign bit of the first operand is zero, the sign bit of
1945 // the result is zero. If the first operand has no one bits below
1946 // the second operand's single 1 bit, its sign will be zero.
1947 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1948 KnownZero2 |= ~LowBits;
1950 KnownZero |= KnownZero2 & Mask;
1952 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1957 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1958 const APInt &RA = Rem->getAPIntValue();
1959 if (RA.isPowerOf2()) {
1960 APInt LowBits = (RA - 1);
1961 APInt Mask2 = LowBits & Mask;
1962 KnownZero |= ~LowBits & Mask;
1963 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1964 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1969 // Since the result is less than or equal to either operand, any leading
1970 // zero bits in either operand must also exist in the result.
1971 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1972 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne,
1974 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2,
1977 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1978 KnownZero2.countLeadingOnes());
1980 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1984 // Allow the target to implement this method for its nodes.
1985 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
1986 case ISD::INTRINSIC_WO_CHAIN:
1987 case ISD::INTRINSIC_W_CHAIN:
1988 case ISD::INTRINSIC_VOID:
1989 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this);
1995 /// ComputeNumSignBits - Return the number of times the sign bit of the
1996 /// register is replicated into the other bits. We know that at least 1 bit
1997 /// is always equal to the sign bit (itself), but other cases can give us
1998 /// information. For example, immediately after an "SRA X, 2", we know that
1999 /// the top 3 bits are all equal to each other, so we return 3.
2000 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
2001 MVT VT = Op.getValueType();
2002 assert(VT.isInteger() && "Invalid VT!");
2003 unsigned VTBits = VT.getSizeInBits();
2005 unsigned FirstAnswer = 1;
2008 return 1; // Limit search depth.
2010 switch (Op.getOpcode()) {
2012 case ISD::AssertSext:
2013 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2014 return VTBits-Tmp+1;
2015 case ISD::AssertZext:
2016 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2019 case ISD::Constant: {
2020 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
2021 // If negative, return # leading ones.
2022 if (Val.isNegative())
2023 return Val.countLeadingOnes();
2025 // Return # leading zeros.
2026 return Val.countLeadingZeros();
2029 case ISD::SIGN_EXTEND:
2030 Tmp = VTBits-Op.getOperand(0).getValueType().getSizeInBits();
2031 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
2033 case ISD::SIGN_EXTEND_INREG:
2034 // Max of the input and what this extends.
2035 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2038 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2039 return std::max(Tmp, Tmp2);
2042 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2043 // SRA X, C -> adds C sign bits.
2044 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2045 Tmp += C->getZExtValue();
2046 if (Tmp > VTBits) Tmp = VTBits;
2050 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2051 // shl destroys sign bits.
2052 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2053 if (C->getZExtValue() >= VTBits || // Bad shift.
2054 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2055 return Tmp - C->getZExtValue();
2060 case ISD::XOR: // NOT is handled here.
2061 // Logical binary ops preserve the number of sign bits at the worst.
2062 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2064 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2065 FirstAnswer = std::min(Tmp, Tmp2);
2066 // We computed what we know about the sign bits as our first
2067 // answer. Now proceed to the generic code that uses
2068 // ComputeMaskedBits, and pick whichever answer is better.
2073 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2074 if (Tmp == 1) return 1; // Early out.
2075 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
2076 return std::min(Tmp, Tmp2);
2084 if (Op.getResNo() != 1)
2086 // The boolean result conforms to getBooleanContents. Fall through.
2088 // If setcc returns 0/-1, all bits are sign bits.
2089 if (TLI.getBooleanContents() ==
2090 TargetLowering::ZeroOrNegativeOneBooleanContent)
2095 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2096 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
2098 // Handle rotate right by N like a rotate left by 32-N.
2099 if (Op.getOpcode() == ISD::ROTR)
2100 RotAmt = (VTBits-RotAmt) & (VTBits-1);
2102 // If we aren't rotating out all of the known-in sign bits, return the
2103 // number that are left. This handles rotl(sext(x), 1) for example.
2104 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2105 if (Tmp > RotAmt+1) return Tmp-RotAmt;
2109 // Add can have at most one carry bit. Thus we know that the output
2110 // is, at worst, one more bit than the inputs.
2111 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2112 if (Tmp == 1) return 1; // Early out.
2114 // Special case decrementing a value (ADD X, -1):
2115 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
2116 if (CRHS->isAllOnesValue()) {
2117 APInt KnownZero, KnownOne;
2118 APInt Mask = APInt::getAllOnesValue(VTBits);
2119 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2121 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2123 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2126 // If we are subtracting one from a positive number, there is no carry
2127 // out of the result.
2128 if (KnownZero.isNegative())
2132 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2133 if (Tmp2 == 1) return 1;
2134 return std::min(Tmp, Tmp2)-1;
2138 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2139 if (Tmp2 == 1) return 1;
2142 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
2143 if (CLHS->isNullValue()) {
2144 APInt KnownZero, KnownOne;
2145 APInt Mask = APInt::getAllOnesValue(VTBits);
2146 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2147 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2149 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2152 // If the input is known to be positive (the sign bit is known clear),
2153 // the output of the NEG has the same number of sign bits as the input.
2154 if (KnownZero.isNegative())
2157 // Otherwise, we treat this like a SUB.
2160 // Sub can have at most one carry bit. Thus we know that the output
2161 // is, at worst, one more bit than the inputs.
2162 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2163 if (Tmp == 1) return 1; // Early out.
2164 return std::min(Tmp, Tmp2)-1;
2167 // FIXME: it's tricky to do anything useful for this, but it is an important
2168 // case for targets like X86.
2172 // Handle LOADX separately here. EXTLOAD case will fallthrough.
2173 if (Op.getOpcode() == ISD::LOAD) {
2174 LoadSDNode *LD = cast<LoadSDNode>(Op);
2175 unsigned ExtType = LD->getExtensionType();
2178 case ISD::SEXTLOAD: // '17' bits known
2179 Tmp = LD->getMemoryVT().getSizeInBits();
2180 return VTBits-Tmp+1;
2181 case ISD::ZEXTLOAD: // '16' bits known
2182 Tmp = LD->getMemoryVT().getSizeInBits();
2187 // Allow the target to implement this method for its nodes.
2188 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
2189 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
2190 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
2191 Op.getOpcode() == ISD::INTRINSIC_VOID) {
2192 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
2193 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
2196 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2197 // use this information.
2198 APInt KnownZero, KnownOne;
2199 APInt Mask = APInt::getAllOnesValue(VTBits);
2200 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
2202 if (KnownZero.isNegative()) { // sign bit is 0
2204 } else if (KnownOne.isNegative()) { // sign bit is 1;
2211 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2212 // the number of identical bits in the top of the input value.
2214 Mask <<= Mask.getBitWidth()-VTBits;
2215 // Return # leading zeros. We use 'min' here in case Val was zero before
2216 // shifting. We don't want to return '64' as for an i32 "0".
2217 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2221 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2222 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2223 if (!GA) return false;
2224 if (GA->getOffset() != 0) return false;
2225 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2226 if (!GV) return false;
2227 MachineModuleInfo *MMI = getMachineModuleInfo();
2228 return MMI && MMI->hasDebugInfo();
2232 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2233 /// element of the result of the vector shuffle.
2234 SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N,
2236 MVT VT = N->getValueType(0);
2237 DebugLoc dl = N->getDebugLoc();
2238 if (N->getMaskElt(i) < 0)
2239 return getUNDEF(VT.getVectorElementType());
2240 unsigned Index = N->getMaskElt(i);
2241 unsigned NumElems = VT.getVectorNumElements();
2242 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2245 if (V.getOpcode() == ISD::BIT_CONVERT) {
2246 V = V.getOperand(0);
2247 MVT VVT = V.getValueType();
2248 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems)
2251 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2252 return (Index == 0) ? V.getOperand(0)
2253 : getUNDEF(VT.getVectorElementType());
2254 if (V.getOpcode() == ISD::BUILD_VECTOR)
2255 return V.getOperand(Index);
2256 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V))
2257 return getShuffleScalarElt(SVN, Index);
2262 /// getNode - Gets or creates the specified node.
2264 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT) {
2265 FoldingSetNodeID ID;
2266 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2268 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2269 return SDValue(E, 0);
2270 SDNode *N = NodeAllocator.Allocate<SDNode>();
2271 new (N) SDNode(Opcode, DL, getVTList(VT));
2272 CSEMap.InsertNode(N, IP);
2274 AllNodes.push_back(N);
2278 return SDValue(N, 0);
2281 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
2282 MVT VT, SDValue Operand) {
2283 // Constant fold unary operations with an integer constant operand.
2284 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2285 const APInt &Val = C->getAPIntValue();
2286 unsigned BitWidth = VT.getSizeInBits();
2289 case ISD::SIGN_EXTEND:
2290 return getConstant(APInt(Val).sextOrTrunc(BitWidth), VT);
2291 case ISD::ANY_EXTEND:
2292 case ISD::ZERO_EXTEND:
2294 return getConstant(APInt(Val).zextOrTrunc(BitWidth), VT);
2295 case ISD::UINT_TO_FP:
2296 case ISD::SINT_TO_FP: {
2297 const uint64_t zero[] = {0, 0};
2298 // No compile time operations on this type.
2299 if (VT==MVT::ppcf128)
2301 APFloat apf = APFloat(APInt(BitWidth, 2, zero));
2302 (void)apf.convertFromAPInt(Val,
2303 Opcode==ISD::SINT_TO_FP,
2304 APFloat::rmNearestTiesToEven);
2305 return getConstantFP(apf, VT);
2307 case ISD::BIT_CONVERT:
2308 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2309 return getConstantFP(Val.bitsToFloat(), VT);
2310 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2311 return getConstantFP(Val.bitsToDouble(), VT);
2314 return getConstant(Val.byteSwap(), VT);
2316 return getConstant(Val.countPopulation(), VT);
2318 return getConstant(Val.countLeadingZeros(), VT);
2320 return getConstant(Val.countTrailingZeros(), VT);
2324 // Constant fold unary operations with a floating point constant operand.
2325 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2326 APFloat V = C->getValueAPF(); // make copy
2327 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2331 return getConstantFP(V, VT);
2334 return getConstantFP(V, VT);
2336 case ISD::FP_EXTEND: {
2338 // This can return overflow, underflow, or inexact; we don't care.
2339 // FIXME need to be more flexible about rounding mode.
2340 (void)V.convert(*MVTToAPFloatSemantics(VT),
2341 APFloat::rmNearestTiesToEven, &ignored);
2342 return getConstantFP(V, VT);
2344 case ISD::FP_TO_SINT:
2345 case ISD::FP_TO_UINT: {
2348 assert(integerPartWidth >= 64);
2349 // FIXME need to be more flexible about rounding mode.
2350 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
2351 Opcode==ISD::FP_TO_SINT,
2352 APFloat::rmTowardZero, &ignored);
2353 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2355 APInt api(VT.getSizeInBits(), 2, x);
2356 return getConstant(api, VT);
2358 case ISD::BIT_CONVERT:
2359 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2360 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2361 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2362 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2368 unsigned OpOpcode = Operand.getNode()->getOpcode();
2370 case ISD::TokenFactor:
2371 case ISD::MERGE_VALUES:
2372 case ISD::CONCAT_VECTORS:
2373 return Operand; // Factor, merge or concat of one node? No need.
2374 case ISD::FP_ROUND: assert(0 && "Invalid method to make FP_ROUND node");
2375 case ISD::FP_EXTEND:
2376 assert(VT.isFloatingPoint() &&
2377 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2378 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2379 if (Operand.getOpcode() == ISD::UNDEF)
2380 return getUNDEF(VT);
2382 case ISD::SIGN_EXTEND:
2383 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2384 "Invalid SIGN_EXTEND!");
2385 if (Operand.getValueType() == VT) return Operand; // noop extension
2386 assert(Operand.getValueType().bitsLT(VT)
2387 && "Invalid sext node, dst < src!");
2388 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2389 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2391 case ISD::ZERO_EXTEND:
2392 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2393 "Invalid ZERO_EXTEND!");
2394 if (Operand.getValueType() == VT) return Operand; // noop extension
2395 assert(Operand.getValueType().bitsLT(VT)
2396 && "Invalid zext node, dst < src!");
2397 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2398 return getNode(ISD::ZERO_EXTEND, DL, VT,
2399 Operand.getNode()->getOperand(0));
2401 case ISD::ANY_EXTEND:
2402 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2403 "Invalid ANY_EXTEND!");
2404 if (Operand.getValueType() == VT) return Operand; // noop extension
2405 assert(Operand.getValueType().bitsLT(VT)
2406 && "Invalid anyext node, dst < src!");
2407 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND)
2408 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2409 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2412 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2413 "Invalid TRUNCATE!");
2414 if (Operand.getValueType() == VT) return Operand; // noop truncate
2415 assert(Operand.getValueType().bitsGT(VT)
2416 && "Invalid truncate node, src < dst!");
2417 if (OpOpcode == ISD::TRUNCATE)
2418 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2419 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2420 OpOpcode == ISD::ANY_EXTEND) {
2421 // If the source is smaller than the dest, we still need an extend.
2422 if (Operand.getNode()->getOperand(0).getValueType().bitsLT(VT))
2423 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2424 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2425 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2427 return Operand.getNode()->getOperand(0);
2430 case ISD::BIT_CONVERT:
2431 // Basic sanity checking.
2432 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2433 && "Cannot BIT_CONVERT between types of different sizes!");
2434 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2435 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2436 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0));
2437 if (OpOpcode == ISD::UNDEF)
2438 return getUNDEF(VT);
2440 case ISD::SCALAR_TO_VECTOR:
2441 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2442 (VT.getVectorElementType() == Operand.getValueType() ||
2443 (VT.getVectorElementType().isInteger() &&
2444 Operand.getValueType().isInteger() &&
2445 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
2446 "Illegal SCALAR_TO_VECTOR node!");
2447 if (OpOpcode == ISD::UNDEF)
2448 return getUNDEF(VT);
2449 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2450 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2451 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2452 Operand.getConstantOperandVal(1) == 0 &&
2453 Operand.getOperand(0).getValueType() == VT)
2454 return Operand.getOperand(0);
2457 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0
2458 if (UnsafeFPMath && OpOpcode == ISD::FSUB)
2459 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1),
2460 Operand.getNode()->getOperand(0));
2461 if (OpOpcode == ISD::FNEG) // --X -> X
2462 return Operand.getNode()->getOperand(0);
2465 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2466 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0));
2471 SDVTList VTs = getVTList(VT);
2472 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2473 FoldingSetNodeID ID;
2474 SDValue Ops[1] = { Operand };
2475 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2477 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2478 return SDValue(E, 0);
2479 N = NodeAllocator.Allocate<UnarySDNode>();
2480 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2481 CSEMap.InsertNode(N, IP);
2483 N = NodeAllocator.Allocate<UnarySDNode>();
2484 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2487 AllNodes.push_back(N);
2491 return SDValue(N, 0);
2494 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2496 ConstantSDNode *Cst1,
2497 ConstantSDNode *Cst2) {
2498 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2501 case ISD::ADD: return getConstant(C1 + C2, VT);
2502 case ISD::SUB: return getConstant(C1 - C2, VT);
2503 case ISD::MUL: return getConstant(C1 * C2, VT);
2505 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2508 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2511 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2514 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2516 case ISD::AND: return getConstant(C1 & C2, VT);
2517 case ISD::OR: return getConstant(C1 | C2, VT);
2518 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2519 case ISD::SHL: return getConstant(C1 << C2, VT);
2520 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2521 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2522 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2523 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2530 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT,
2531 SDValue N1, SDValue N2) {
2532 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2533 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2536 case ISD::TokenFactor:
2537 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2538 N2.getValueType() == MVT::Other && "Invalid token factor!");
2539 // Fold trivial token factors.
2540 if (N1.getOpcode() == ISD::EntryToken) return N2;
2541 if (N2.getOpcode() == ISD::EntryToken) return N1;
2542 if (N1 == N2) return N1;
2544 case ISD::CONCAT_VECTORS:
2545 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2546 // one big BUILD_VECTOR.
2547 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2548 N2.getOpcode() == ISD::BUILD_VECTOR) {
2549 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2550 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2551 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2555 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2556 N1.getValueType() == VT && "Binary operator types must match!");
2557 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2558 // worth handling here.
2559 if (N2C && N2C->isNullValue())
2561 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2568 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2569 N1.getValueType() == VT && "Binary operator types must match!");
2570 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2571 // it's worth handling here.
2572 if (N2C && N2C->isNullValue())
2582 assert(VT.isInteger() && "This operator does not apply to FP types!");
2590 if (Opcode == ISD::FADD) {
2592 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1))
2593 if (CFP->getValueAPF().isZero())
2596 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2597 if (CFP->getValueAPF().isZero())
2599 } else if (Opcode == ISD::FSUB) {
2601 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2602 if (CFP->getValueAPF().isZero())
2606 assert(N1.getValueType() == N2.getValueType() &&
2607 N1.getValueType() == VT && "Binary operator types must match!");
2609 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2610 assert(N1.getValueType() == VT &&
2611 N1.getValueType().isFloatingPoint() &&
2612 N2.getValueType().isFloatingPoint() &&
2613 "Invalid FCOPYSIGN!");
2620 assert(VT == N1.getValueType() &&
2621 "Shift operators return type must be the same as their first arg");
2622 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2623 "Shifts only work on integers");
2625 // Always fold shifts of i1 values so the code generator doesn't need to
2626 // handle them. Since we know the size of the shift has to be less than the
2627 // size of the value, the shift/rotate count is guaranteed to be zero.
2631 case ISD::FP_ROUND_INREG: {
2632 MVT EVT = cast<VTSDNode>(N2)->getVT();
2633 assert(VT == N1.getValueType() && "Not an inreg round!");
2634 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2635 "Cannot FP_ROUND_INREG integer types");
2636 assert(EVT.bitsLE(VT) && "Not rounding down!");
2637 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2641 assert(VT.isFloatingPoint() &&
2642 N1.getValueType().isFloatingPoint() &&
2643 VT.bitsLE(N1.getValueType()) &&
2644 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2645 if (N1.getValueType() == VT) return N1; // noop conversion.
2647 case ISD::AssertSext:
2648 case ISD::AssertZext: {
2649 MVT EVT = cast<VTSDNode>(N2)->getVT();
2650 assert(VT == N1.getValueType() && "Not an inreg extend!");
2651 assert(VT.isInteger() && EVT.isInteger() &&
2652 "Cannot *_EXTEND_INREG FP types");
2653 assert(EVT.bitsLE(VT) && "Not extending!");
2654 if (VT == EVT) return N1; // noop assertion.
2657 case ISD::SIGN_EXTEND_INREG: {
2658 MVT EVT = cast<VTSDNode>(N2)->getVT();
2659 assert(VT == N1.getValueType() && "Not an inreg extend!");
2660 assert(VT.isInteger() && EVT.isInteger() &&
2661 "Cannot *_EXTEND_INREG FP types");
2662 assert(EVT.bitsLE(VT) && "Not extending!");
2663 if (EVT == VT) return N1; // Not actually extending
2666 APInt Val = N1C->getAPIntValue();
2667 unsigned FromBits = cast<VTSDNode>(N2)->getVT().getSizeInBits();
2668 Val <<= Val.getBitWidth()-FromBits;
2669 Val = Val.ashr(Val.getBitWidth()-FromBits);
2670 return getConstant(Val, VT);
2674 case ISD::EXTRACT_VECTOR_ELT:
2675 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2676 if (N1.getOpcode() == ISD::UNDEF)
2677 return getUNDEF(VT);
2679 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2680 // expanding copies of large vectors from registers.
2682 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2683 N1.getNumOperands() > 0) {
2685 N1.getOperand(0).getValueType().getVectorNumElements();
2686 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
2687 N1.getOperand(N2C->getZExtValue() / Factor),
2688 getConstant(N2C->getZExtValue() % Factor,
2689 N2.getValueType()));
2692 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2693 // expanding large vector constants.
2694 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) {
2695 SDValue Elt = N1.getOperand(N2C->getZExtValue());
2696 MVT VEltTy = N1.getValueType().getVectorElementType();
2697 if (Elt.getValueType() != VEltTy) {
2698 // If the vector element type is not legal, the BUILD_VECTOR operands
2699 // are promoted and implicitly truncated. Make that explicit here.
2700 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt);
2703 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT
2704 // result is implicitly extended.
2705 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt);
2710 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2711 // operations are lowered to scalars.
2712 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2713 // If the indices are the same, return the inserted element.
2714 if (N1.getOperand(2) == N2)
2715 return N1.getOperand(1);
2716 // If the indices are known different, extract the element from
2717 // the original vector.
2718 else if (isa<ConstantSDNode>(N1.getOperand(2)) &&
2719 isa<ConstantSDNode>(N2))
2720 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
2723 case ISD::EXTRACT_ELEMENT:
2724 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2725 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2726 (N1.getValueType().isInteger() == VT.isInteger()) &&
2727 "Wrong types for EXTRACT_ELEMENT!");
2729 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2730 // 64-bit integers into 32-bit parts. Instead of building the extract of
2731 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2732 if (N1.getOpcode() == ISD::BUILD_PAIR)
2733 return N1.getOperand(N2C->getZExtValue());
2735 // EXTRACT_ELEMENT of a constant int is also very common.
2736 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2737 unsigned ElementSize = VT.getSizeInBits();
2738 unsigned Shift = ElementSize * N2C->getZExtValue();
2739 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2740 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2743 case ISD::EXTRACT_SUBVECTOR:
2744 if (N1.getValueType() == VT) // Trivial extraction.
2751 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2752 if (SV.getNode()) return SV;
2753 } else { // Cannonicalize constant to RHS if commutative
2754 if (isCommutativeBinOp(Opcode)) {
2755 std::swap(N1C, N2C);
2761 // Constant fold FP operations.
2762 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2763 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2765 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2766 // Cannonicalize constant to RHS if commutative
2767 std::swap(N1CFP, N2CFP);
2769 } else if (N2CFP && VT != MVT::ppcf128) {
2770 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2771 APFloat::opStatus s;
2774 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2775 if (s != APFloat::opInvalidOp)
2776 return getConstantFP(V1, VT);
2779 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2780 if (s!=APFloat::opInvalidOp)
2781 return getConstantFP(V1, VT);
2784 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2785 if (s!=APFloat::opInvalidOp)
2786 return getConstantFP(V1, VT);
2789 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2790 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2791 return getConstantFP(V1, VT);
2794 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2795 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2796 return getConstantFP(V1, VT);
2798 case ISD::FCOPYSIGN:
2800 return getConstantFP(V1, VT);
2806 // Canonicalize an UNDEF to the RHS, even over a constant.
2807 if (N1.getOpcode() == ISD::UNDEF) {
2808 if (isCommutativeBinOp(Opcode)) {
2812 case ISD::FP_ROUND_INREG:
2813 case ISD::SIGN_EXTEND_INREG:
2819 return N1; // fold op(undef, arg2) -> undef
2827 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2828 // For vectors, we can't easily build an all zero vector, just return
2835 // Fold a bunch of operators when the RHS is undef.
2836 if (N2.getOpcode() == ISD::UNDEF) {
2839 if (N1.getOpcode() == ISD::UNDEF)
2840 // Handle undef ^ undef -> 0 special case. This is a common
2842 return getConstant(0, VT);
2852 return N2; // fold op(arg1, undef) -> undef
2866 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2867 // For vectors, we can't easily build an all zero vector, just return
2872 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
2873 // For vectors, we can't easily build an all one vector, just return
2881 // Memoize this node if possible.
2883 SDVTList VTs = getVTList(VT);
2884 if (VT != MVT::Flag) {
2885 SDValue Ops[] = { N1, N2 };
2886 FoldingSetNodeID ID;
2887 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2889 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2890 return SDValue(E, 0);
2891 N = NodeAllocator.Allocate<BinarySDNode>();
2892 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2893 CSEMap.InsertNode(N, IP);
2895 N = NodeAllocator.Allocate<BinarySDNode>();
2896 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2899 AllNodes.push_back(N);
2903 return SDValue(N, 0);
2906 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT,
2907 SDValue N1, SDValue N2, SDValue N3) {
2908 // Perform various simplifications.
2909 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2910 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2912 case ISD::CONCAT_VECTORS:
2913 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2914 // one big BUILD_VECTOR.
2915 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2916 N2.getOpcode() == ISD::BUILD_VECTOR &&
2917 N3.getOpcode() == ISD::BUILD_VECTOR) {
2918 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2919 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2920 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
2921 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2925 // Use FoldSetCC to simplify SETCC's.
2926 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL);
2927 if (Simp.getNode()) return Simp;
2932 if (N1C->getZExtValue())
2933 return N2; // select true, X, Y -> X
2935 return N3; // select false, X, Y -> Y
2938 if (N2 == N3) return N2; // select C, X, X -> X
2942 if (N2C->getZExtValue()) // Unconditional branch
2943 return getNode(ISD::BR, DL, MVT::Other, N1, N3);
2945 return N1; // Never-taken branch
2948 case ISD::VECTOR_SHUFFLE:
2949 assert(0 && "should use getVectorShuffle constructor!");
2951 case ISD::BIT_CONVERT:
2952 // Fold bit_convert nodes from a type to themselves.
2953 if (N1.getValueType() == VT)
2958 // Memoize node if it doesn't produce a flag.
2960 SDVTList VTs = getVTList(VT);
2961 if (VT != MVT::Flag) {
2962 SDValue Ops[] = { N1, N2, N3 };
2963 FoldingSetNodeID ID;
2964 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
2966 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2967 return SDValue(E, 0);
2968 N = NodeAllocator.Allocate<TernarySDNode>();
2969 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
2970 CSEMap.InsertNode(N, IP);
2972 N = NodeAllocator.Allocate<TernarySDNode>();
2973 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
2975 AllNodes.push_back(N);
2979 return SDValue(N, 0);
2982 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT,
2983 SDValue N1, SDValue N2, SDValue N3,
2985 SDValue Ops[] = { N1, N2, N3, N4 };
2986 return getNode(Opcode, DL, VT, Ops, 4);
2989 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT,
2990 SDValue N1, SDValue N2, SDValue N3,
2991 SDValue N4, SDValue N5) {
2992 SDValue Ops[] = { N1, N2, N3, N4, N5 };
2993 return getNode(Opcode, DL, VT, Ops, 5);
2996 /// getMemsetValue - Vectorized representation of the memset value
2998 static SDValue getMemsetValue(SDValue Value, MVT VT, SelectionDAG &DAG,
3000 unsigned NumBits = VT.isVector() ?
3001 VT.getVectorElementType().getSizeInBits() : VT.getSizeInBits();
3002 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3003 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
3005 for (unsigned i = NumBits; i > 8; i >>= 1) {
3006 Val = (Val << Shift) | Val;
3010 return DAG.getConstant(Val, VT);
3011 return DAG.getConstantFP(APFloat(Val), VT);
3014 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3015 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
3017 for (unsigned i = NumBits; i > 8; i >>= 1) {
3018 Value = DAG.getNode(ISD::OR, dl, VT,
3019 DAG.getNode(ISD::SHL, dl, VT, Value,
3020 DAG.getConstant(Shift,
3021 TLI.getShiftAmountTy())),
3029 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3030 /// used when a memcpy is turned into a memset when the source is a constant
3032 static SDValue getMemsetStringVal(MVT VT, DebugLoc dl, SelectionDAG &DAG,
3033 const TargetLowering &TLI,
3034 std::string &Str, unsigned Offset) {
3035 // Handle vector with all elements zero.
3038 return DAG.getConstant(0, VT);
3039 unsigned NumElts = VT.getVectorNumElements();
3040 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3041 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3042 DAG.getConstant(0, MVT::getVectorVT(EltVT, NumElts)));
3045 assert(!VT.isVector() && "Can't handle vector type here!");
3046 unsigned NumBits = VT.getSizeInBits();
3047 unsigned MSB = NumBits / 8;
3049 if (TLI.isLittleEndian())
3050 Offset = Offset + MSB - 1;
3051 for (unsigned i = 0; i != MSB; ++i) {
3052 Val = (Val << 8) | (unsigned char)Str[Offset];
3053 Offset += TLI.isLittleEndian() ? -1 : 1;
3055 return DAG.getConstant(Val, VT);
3058 /// getMemBasePlusOffset - Returns base and offset node for the
3060 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
3061 SelectionDAG &DAG) {
3062 MVT VT = Base.getValueType();
3063 return DAG.getNode(ISD::ADD, Base.getDebugLoc(),
3064 VT, Base, DAG.getConstant(Offset, VT));
3067 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
3069 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
3070 unsigned SrcDelta = 0;
3071 GlobalAddressSDNode *G = NULL;
3072 if (Src.getOpcode() == ISD::GlobalAddress)
3073 G = cast<GlobalAddressSDNode>(Src);
3074 else if (Src.getOpcode() == ISD::ADD &&
3075 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3076 Src.getOperand(1).getOpcode() == ISD::Constant) {
3077 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
3078 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
3083 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3084 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
3090 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
3091 /// to replace the memset / memcpy is below the threshold. It also returns the
3092 /// types of the sequence of memory ops to perform memset / memcpy.
3094 bool MeetsMaxMemopRequirement(std::vector<MVT> &MemOps,
3095 SDValue Dst, SDValue Src,
3096 unsigned Limit, uint64_t Size, unsigned &Align,
3097 std::string &Str, bool &isSrcStr,
3099 const TargetLowering &TLI) {
3100 isSrcStr = isMemSrcFromString(Src, Str);
3101 bool isSrcConst = isa<ConstantSDNode>(Src);
3102 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses();
3103 MVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG);
3104 if (VT != MVT::iAny) {
3105 unsigned NewAlign = (unsigned)
3106 TLI.getTargetData()->getABITypeAlignment(VT.getTypeForMVT(
3107 *DAG.getContext()));
3108 // If source is a string constant, this will require an unaligned load.
3109 if (NewAlign > Align && (isSrcConst || AllowUnalign)) {
3110 if (Dst.getOpcode() != ISD::FrameIndex) {
3111 // Can't change destination alignment. It requires a unaligned store.
3115 int FI = cast<FrameIndexSDNode>(Dst)->getIndex();
3116 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3117 if (MFI->isFixedObjectIndex(FI)) {
3118 // Can't change destination alignment. It requires a unaligned store.
3122 // Give the stack frame object a larger alignment if needed.
3123 if (MFI->getObjectAlignment(FI) < NewAlign)
3124 MFI->setObjectAlignment(FI, NewAlign);
3131 if (VT == MVT::iAny) {
3135 switch (Align & 7) {
3136 case 0: VT = MVT::i64; break;
3137 case 4: VT = MVT::i32; break;
3138 case 2: VT = MVT::i16; break;
3139 default: VT = MVT::i8; break;
3144 while (!TLI.isTypeLegal(LVT))
3145 LVT = (MVT::SimpleValueType)(LVT.getSimpleVT() - 1);
3146 assert(LVT.isInteger());
3152 unsigned NumMemOps = 0;
3154 unsigned VTSize = VT.getSizeInBits() / 8;
3155 while (VTSize > Size) {
3156 // For now, only use non-vector load / store's for the left-over pieces.
3157 if (VT.isVector()) {
3159 while (!TLI.isTypeLegal(VT))
3160 VT = (MVT::SimpleValueType)(VT.getSimpleVT() - 1);
3161 VTSize = VT.getSizeInBits() / 8;
3163 // This can result in a type that is not legal on the target, e.g.
3164 // 1 or 2 bytes on PPC.
3165 VT = (MVT::SimpleValueType)(VT.getSimpleVT() - 1);
3170 if (++NumMemOps > Limit)
3172 MemOps.push_back(VT);
3179 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3180 SDValue Chain, SDValue Dst,
3181 SDValue Src, uint64_t Size,
3182 unsigned Align, bool AlwaysInline,
3183 const Value *DstSV, uint64_t DstSVOff,
3184 const Value *SrcSV, uint64_t SrcSVOff){
3185 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3187 // Expand memcpy to a series of load and store ops if the size operand falls
3188 // below a certain threshold.
3189 std::vector<MVT> MemOps;
3190 uint64_t Limit = -1ULL;
3192 Limit = TLI.getMaxStoresPerMemcpy();
3193 unsigned DstAlign = Align; // Destination alignment can change.
3196 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3197 Str, CopyFromStr, DAG, TLI))
3201 bool isZeroStr = CopyFromStr && Str.empty();
3202 SmallVector<SDValue, 8> OutChains;
3203 unsigned NumMemOps = MemOps.size();
3204 uint64_t SrcOff = 0, DstOff = 0;
3205 for (unsigned i = 0; i < NumMemOps; i++) {
3207 unsigned VTSize = VT.getSizeInBits() / 8;
3208 SDValue Value, Store;
3210 if (CopyFromStr && (isZeroStr || !VT.isVector())) {
3211 // It's unlikely a store of a vector immediate can be done in a single
3212 // instruction. It would require a load from a constantpool first.
3213 // We also handle store a vector with all zero's.
3214 // FIXME: Handle other cases where store of vector immediate is done in
3215 // a single instruction.
3216 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff);
3217 Store = DAG.getStore(Chain, dl, Value,
3218 getMemBasePlusOffset(Dst, DstOff, DAG),
3219 DstSV, DstSVOff + DstOff, false, DstAlign);
3221 // The type might not be legal for the target. This should only happen
3222 // if the type is smaller than a legal type, as on PPC, so the right
3223 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
3224 // to Load/Store if NVT==VT.
3225 // FIXME does the case above also need this?
3226 MVT NVT = TLI.getTypeToTransformTo(VT);
3227 assert(NVT.bitsGE(VT));
3228 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain,
3229 getMemBasePlusOffset(Src, SrcOff, DAG),
3230 SrcSV, SrcSVOff + SrcOff, VT, false, Align);
3231 Store = DAG.getTruncStore(Chain, dl, Value,
3232 getMemBasePlusOffset(Dst, DstOff, DAG),
3233 DstSV, DstSVOff + DstOff, VT, false, DstAlign);
3235 OutChains.push_back(Store);
3240 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3241 &OutChains[0], OutChains.size());
3244 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3245 SDValue Chain, SDValue Dst,
3246 SDValue Src, uint64_t Size,
3247 unsigned Align, bool AlwaysInline,
3248 const Value *DstSV, uint64_t DstSVOff,
3249 const Value *SrcSV, uint64_t SrcSVOff){
3250 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3252 // Expand memmove to a series of load and store ops if the size operand falls
3253 // below a certain threshold.
3254 std::vector<MVT> MemOps;
3255 uint64_t Limit = -1ULL;
3257 Limit = TLI.getMaxStoresPerMemmove();
3258 unsigned DstAlign = Align; // Destination alignment can change.
3261 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3262 Str, CopyFromStr, DAG, TLI))
3265 uint64_t SrcOff = 0, DstOff = 0;
3267 SmallVector<SDValue, 8> LoadValues;
3268 SmallVector<SDValue, 8> LoadChains;
3269 SmallVector<SDValue, 8> OutChains;
3270 unsigned NumMemOps = MemOps.size();
3271 for (unsigned i = 0; i < NumMemOps; i++) {
3273 unsigned VTSize = VT.getSizeInBits() / 8;
3274 SDValue Value, Store;
3276 Value = DAG.getLoad(VT, dl, Chain,
3277 getMemBasePlusOffset(Src, SrcOff, DAG),
3278 SrcSV, SrcSVOff + SrcOff, false, Align);
3279 LoadValues.push_back(Value);
3280 LoadChains.push_back(Value.getValue(1));
3283 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3284 &LoadChains[0], LoadChains.size());
3286 for (unsigned i = 0; i < NumMemOps; i++) {
3288 unsigned VTSize = VT.getSizeInBits() / 8;
3289 SDValue Value, Store;
3291 Store = DAG.getStore(Chain, dl, LoadValues[i],
3292 getMemBasePlusOffset(Dst, DstOff, DAG),
3293 DstSV, DstSVOff + DstOff, false, DstAlign);
3294 OutChains.push_back(Store);
3298 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3299 &OutChains[0], OutChains.size());
3302 static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl,
3303 SDValue Chain, SDValue Dst,
3304 SDValue Src, uint64_t Size,
3306 const Value *DstSV, uint64_t DstSVOff) {
3307 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3309 // Expand memset to a series of load/store ops if the size operand
3310 // falls below a certain threshold.
3311 std::vector<MVT> MemOps;
3314 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(),
3315 Size, Align, Str, CopyFromStr, DAG, TLI))
3318 SmallVector<SDValue, 8> OutChains;
3319 uint64_t DstOff = 0;
3321 unsigned NumMemOps = MemOps.size();
3322 for (unsigned i = 0; i < NumMemOps; i++) {
3324 unsigned VTSize = VT.getSizeInBits() / 8;
3325 SDValue Value = getMemsetValue(Src, VT, DAG, dl);
3326 SDValue Store = DAG.getStore(Chain, dl, Value,
3327 getMemBasePlusOffset(Dst, DstOff, DAG),
3328 DstSV, DstSVOff + DstOff);
3329 OutChains.push_back(Store);
3333 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3334 &OutChains[0], OutChains.size());
3337 SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst,
3338 SDValue Src, SDValue Size,
3339 unsigned Align, bool AlwaysInline,
3340 const Value *DstSV, uint64_t DstSVOff,
3341 const Value *SrcSV, uint64_t SrcSVOff) {
3343 // Check to see if we should lower the memcpy to loads and stores first.
3344 // For cases within the target-specified limits, this is the best choice.
3345 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3347 // Memcpy with size zero? Just return the original chain.
3348 if (ConstantSize->isNullValue())
3352 getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3353 ConstantSize->getZExtValue(),
3354 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3355 if (Result.getNode())
3359 // Then check to see if we should lower the memcpy with target-specific
3360 // code. If the target chooses to do this, this is the next best.
3362 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align,
3364 DstSV, DstSVOff, SrcSV, SrcSVOff);
3365 if (Result.getNode())
3368 // If we really need inline code and the target declined to provide it,
3369 // use a (potentially long) sequence of loads and stores.
3371 assert(ConstantSize && "AlwaysInline requires a constant size!");
3372 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3373 ConstantSize->getZExtValue(), Align, true,
3374 DstSV, DstSVOff, SrcSV, SrcSVOff);
3377 // Emit a library call.
3378 TargetLowering::ArgListTy Args;
3379 TargetLowering::ArgListEntry Entry;
3380 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3381 Entry.Node = Dst; Args.push_back(Entry);
3382 Entry.Node = Src; Args.push_back(Entry);
3383 Entry.Node = Size; Args.push_back(Entry);
3384 // FIXME: pass in DebugLoc
3385 std::pair<SDValue,SDValue> CallResult =
3386 TLI.LowerCallTo(Chain, Type::VoidTy,
3387 false, false, false, false, 0, CallingConv::C, false,
3388 getExternalSymbol("memcpy", TLI.getPointerTy()),
3390 return CallResult.second;
3393 SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst,
3394 SDValue Src, SDValue Size,
3396 const Value *DstSV, uint64_t DstSVOff,
3397 const Value *SrcSV, uint64_t SrcSVOff) {
3399 // Check to see if we should lower the memmove to loads and stores first.
3400 // For cases within the target-specified limits, this is the best choice.
3401 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3403 // Memmove with size zero? Just return the original chain.
3404 if (ConstantSize->isNullValue())
3408 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src,
3409 ConstantSize->getZExtValue(),
3410 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3411 if (Result.getNode())
3415 // Then check to see if we should lower the memmove with target-specific
3416 // code. If the target chooses to do this, this is the next best.
3418 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align,
3419 DstSV, DstSVOff, SrcSV, SrcSVOff);
3420 if (Result.getNode())
3423 // Emit a library call.
3424 TargetLowering::ArgListTy Args;
3425 TargetLowering::ArgListEntry Entry;
3426 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3427 Entry.Node = Dst; Args.push_back(Entry);
3428 Entry.Node = Src; Args.push_back(Entry);
3429 Entry.Node = Size; Args.push_back(Entry);
3430 // FIXME: pass in DebugLoc
3431 std::pair<SDValue,SDValue> CallResult =
3432 TLI.LowerCallTo(Chain, Type::VoidTy,
3433 false, false, false, false, 0, CallingConv::C, false,
3434 getExternalSymbol("memmove", TLI.getPointerTy()),
3436 return CallResult.second;
3439 SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst,
3440 SDValue Src, SDValue Size,
3442 const Value *DstSV, uint64_t DstSVOff) {
3444 // Check to see if we should lower the memset to stores first.
3445 // For cases within the target-specified limits, this is the best choice.
3446 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3448 // Memset with size zero? Just return the original chain.
3449 if (ConstantSize->isNullValue())
3453 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
3454 Align, DstSV, DstSVOff);
3455 if (Result.getNode())
3459 // Then check to see if we should lower the memset with target-specific
3460 // code. If the target chooses to do this, this is the next best.
3462 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align,
3464 if (Result.getNode())
3467 // Emit a library call.
3468 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType();
3469 TargetLowering::ArgListTy Args;
3470 TargetLowering::ArgListEntry Entry;
3471 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3472 Args.push_back(Entry);
3473 // Extend or truncate the argument to be an i32 value for the call.
3474 if (Src.getValueType().bitsGT(MVT::i32))
3475 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
3477 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
3478 Entry.Node = Src; Entry.Ty = Type::Int32Ty; Entry.isSExt = true;
3479 Args.push_back(Entry);
3480 Entry.Node = Size; Entry.Ty = IntPtrTy; Entry.isSExt = false;
3481 Args.push_back(Entry);
3482 // FIXME: pass in DebugLoc
3483 std::pair<SDValue,SDValue> CallResult =
3484 TLI.LowerCallTo(Chain, Type::VoidTy,
3485 false, false, false, false, 0, CallingConv::C, false,
3486 getExternalSymbol("memset", TLI.getPointerTy()),
3488 return CallResult.second;
3491 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, MVT MemVT,
3493 SDValue Ptr, SDValue Cmp,
3494 SDValue Swp, const Value* PtrVal,
3495 unsigned Alignment) {
3496 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op");
3497 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3499 MVT VT = Cmp.getValueType();
3501 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3502 Alignment = getMVTAlignment(MemVT);
3504 SDVTList VTs = getVTList(VT, MVT::Other);
3505 FoldingSetNodeID ID;
3506 ID.AddInteger(MemVT.getRawBits());
3507 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3508 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3510 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3511 return SDValue(E, 0);
3512 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3513 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT,
3514 Chain, Ptr, Cmp, Swp, PtrVal, Alignment);
3515 CSEMap.InsertNode(N, IP);
3516 AllNodes.push_back(N);
3517 return SDValue(N, 0);
3520 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, MVT MemVT,
3522 SDValue Ptr, SDValue Val,
3523 const Value* PtrVal,
3524 unsigned Alignment) {
3525 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
3526 Opcode == ISD::ATOMIC_LOAD_SUB ||
3527 Opcode == ISD::ATOMIC_LOAD_AND ||
3528 Opcode == ISD::ATOMIC_LOAD_OR ||
3529 Opcode == ISD::ATOMIC_LOAD_XOR ||
3530 Opcode == ISD::ATOMIC_LOAD_NAND ||
3531 Opcode == ISD::ATOMIC_LOAD_MIN ||
3532 Opcode == ISD::ATOMIC_LOAD_MAX ||
3533 Opcode == ISD::ATOMIC_LOAD_UMIN ||
3534 Opcode == ISD::ATOMIC_LOAD_UMAX ||
3535 Opcode == ISD::ATOMIC_SWAP) &&
3536 "Invalid Atomic Op");
3538 MVT VT = Val.getValueType();
3540 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3541 Alignment = getMVTAlignment(MemVT);
3543 SDVTList VTs = getVTList(VT, MVT::Other);
3544 FoldingSetNodeID ID;
3545 ID.AddInteger(MemVT.getRawBits());
3546 SDValue Ops[] = {Chain, Ptr, Val};
3547 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3549 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3550 return SDValue(E, 0);
3551 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3552 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT,
3553 Chain, Ptr, Val, PtrVal, Alignment);
3554 CSEMap.InsertNode(N, IP);
3555 AllNodes.push_back(N);
3556 return SDValue(N, 0);
3559 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3560 /// Allowed to return something different (and simpler) if Simplify is true.
3561 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3566 SmallVector<MVT, 4> VTs;
3567 VTs.reserve(NumOps);
3568 for (unsigned i = 0; i < NumOps; ++i)
3569 VTs.push_back(Ops[i].getValueType());
3570 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps),
3575 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl,
3576 const MVT *VTs, unsigned NumVTs,
3577 const SDValue *Ops, unsigned NumOps,
3578 MVT MemVT, const Value *srcValue, int SVOff,
3579 unsigned Align, bool Vol,
3580 bool ReadMem, bool WriteMem) {
3581 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps,
3582 MemVT, srcValue, SVOff, Align, Vol,
3587 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3588 const SDValue *Ops, unsigned NumOps,
3589 MVT MemVT, const Value *srcValue, int SVOff,
3590 unsigned Align, bool Vol,
3591 bool ReadMem, bool WriteMem) {
3592 // Memoize the node unless it returns a flag.
3593 MemIntrinsicSDNode *N;
3594 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3595 FoldingSetNodeID ID;
3596 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3598 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3599 return SDValue(E, 0);
3601 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3602 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT,
3603 srcValue, SVOff, Align, Vol, ReadMem, WriteMem);
3604 CSEMap.InsertNode(N, IP);
3606 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3607 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT,
3608 srcValue, SVOff, Align, Vol, ReadMem, WriteMem);
3610 AllNodes.push_back(N);
3611 return SDValue(N, 0);
3615 SelectionDAG::getCall(unsigned CallingConv, DebugLoc dl, bool IsVarArgs,
3616 bool IsTailCall, bool IsInreg, SDVTList VTs,
3617 const SDValue *Operands, unsigned NumOperands,
3618 unsigned NumFixedArgs) {
3619 // Do not include isTailCall in the folding set profile.
3620 FoldingSetNodeID ID;
3621 AddNodeIDNode(ID, ISD::CALL, VTs, Operands, NumOperands);
3622 ID.AddInteger(CallingConv);
3623 ID.AddInteger(IsVarArgs);
3625 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3626 // Instead of including isTailCall in the folding set, we just
3627 // set the flag of the existing node.
3629 cast<CallSDNode>(E)->setNotTailCall();
3630 return SDValue(E, 0);
3632 SDNode *N = NodeAllocator.Allocate<CallSDNode>();
3633 new (N) CallSDNode(CallingConv, dl, IsVarArgs, IsTailCall, IsInreg,
3634 VTs, Operands, NumOperands, NumFixedArgs);
3635 CSEMap.InsertNode(N, IP);
3636 AllNodes.push_back(N);
3637 return SDValue(N, 0);
3641 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3642 ISD::LoadExtType ExtType, MVT VT, SDValue Chain,
3643 SDValue Ptr, SDValue Offset,
3644 const Value *SV, int SVOffset, MVT EVT,
3645 bool isVolatile, unsigned Alignment) {
3646 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3647 Alignment = getMVTAlignment(VT);
3650 ExtType = ISD::NON_EXTLOAD;
3651 } else if (ExtType == ISD::NON_EXTLOAD) {
3652 assert(VT == EVT && "Non-extending load from different memory type!");
3656 assert(EVT.getVectorNumElements() == VT.getVectorNumElements() &&
3657 "Invalid vector extload!");
3659 assert(EVT.bitsLT(VT) &&
3660 "Should only be an extending load, not truncating!");
3661 assert((ExtType == ISD::EXTLOAD || VT.isInteger()) &&
3662 "Cannot sign/zero extend a FP/Vector load!");
3663 assert(VT.isInteger() == EVT.isInteger() &&
3664 "Cannot convert from FP to Int or Int -> FP!");
3667 bool Indexed = AM != ISD::UNINDEXED;
3668 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3669 "Unindexed load with an offset!");
3671 SDVTList VTs = Indexed ?
3672 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3673 SDValue Ops[] = { Chain, Ptr, Offset };
3674 FoldingSetNodeID ID;
3675 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3676 ID.AddInteger(EVT.getRawBits());
3677 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, isVolatile, Alignment));
3679 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3680 return SDValue(E, 0);
3681 SDNode *N = NodeAllocator.Allocate<LoadSDNode>();
3682 new (N) LoadSDNode(Ops, dl, VTs, AM, ExtType, EVT, SV, SVOffset,
3683 Alignment, isVolatile);
3684 CSEMap.InsertNode(N, IP);
3685 AllNodes.push_back(N);
3686 return SDValue(N, 0);
3689 SDValue SelectionDAG::getLoad(MVT VT, DebugLoc dl,
3690 SDValue Chain, SDValue Ptr,
3691 const Value *SV, int SVOffset,
3692 bool isVolatile, unsigned Alignment) {
3693 SDValue Undef = getUNDEF(Ptr.getValueType());
3694 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3695 SV, SVOffset, VT, isVolatile, Alignment);
3698 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, MVT VT,
3699 SDValue Chain, SDValue Ptr,
3701 int SVOffset, MVT EVT,
3702 bool isVolatile, unsigned Alignment) {
3703 SDValue Undef = getUNDEF(Ptr.getValueType());
3704 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef,
3705 SV, SVOffset, EVT, isVolatile, Alignment);
3709 SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base,
3710 SDValue Offset, ISD::MemIndexedMode AM) {
3711 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
3712 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
3713 "Load is already a indexed load!");
3714 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(),
3715 LD->getChain(), Base, Offset, LD->getSrcValue(),
3716 LD->getSrcValueOffset(), LD->getMemoryVT(),
3717 LD->isVolatile(), LD->getAlignment());
3720 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3721 SDValue Ptr, const Value *SV, int SVOffset,
3722 bool isVolatile, unsigned Alignment) {
3723 MVT VT = Val.getValueType();
3725 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3726 Alignment = getMVTAlignment(VT);
3728 SDVTList VTs = getVTList(MVT::Other);
3729 SDValue Undef = getUNDEF(Ptr.getValueType());
3730 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3731 FoldingSetNodeID ID;
3732 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3733 ID.AddInteger(VT.getRawBits());
3734 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED,
3735 isVolatile, Alignment));
3737 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3738 return SDValue(E, 0);
3739 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3740 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, false,
3741 VT, SV, SVOffset, Alignment, isVolatile);
3742 CSEMap.InsertNode(N, IP);
3743 AllNodes.push_back(N);
3744 return SDValue(N, 0);
3747 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3748 SDValue Ptr, const Value *SV,
3749 int SVOffset, MVT SVT,
3750 bool isVolatile, unsigned Alignment) {
3751 MVT VT = Val.getValueType();
3754 return getStore(Chain, dl, Val, Ptr, SV, SVOffset, isVolatile, Alignment);
3756 assert(VT.bitsGT(SVT) && "Not a truncation?");
3757 assert(VT.isInteger() == SVT.isInteger() &&
3758 "Can't do FP-INT conversion!");
3760 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3761 Alignment = getMVTAlignment(VT);
3763 SDVTList VTs = getVTList(MVT::Other);
3764 SDValue Undef = getUNDEF(Ptr.getValueType());
3765 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3766 FoldingSetNodeID ID;
3767 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3768 ID.AddInteger(SVT.getRawBits());
3769 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED,
3770 isVolatile, Alignment));
3772 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3773 return SDValue(E, 0);
3774 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3775 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, true,
3776 SVT, SV, SVOffset, Alignment, isVolatile);
3777 CSEMap.InsertNode(N, IP);
3778 AllNodes.push_back(N);
3779 return SDValue(N, 0);
3783 SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base,
3784 SDValue Offset, ISD::MemIndexedMode AM) {
3785 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
3786 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
3787 "Store is already a indexed store!");
3788 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
3789 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
3790 FoldingSetNodeID ID;
3791 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3792 ID.AddInteger(ST->getMemoryVT().getRawBits());
3793 ID.AddInteger(ST->getRawSubclassData());
3795 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3796 return SDValue(E, 0);
3797 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3798 new (N) StoreSDNode(Ops, dl, VTs, AM,
3799 ST->isTruncatingStore(), ST->getMemoryVT(),
3800 ST->getSrcValue(), ST->getSrcValueOffset(),
3801 ST->getAlignment(), ST->isVolatile());
3802 CSEMap.InsertNode(N, IP);
3803 AllNodes.push_back(N);
3804 return SDValue(N, 0);
3807 SDValue SelectionDAG::getVAArg(MVT VT, DebugLoc dl,
3808 SDValue Chain, SDValue Ptr,
3810 SDValue Ops[] = { Chain, Ptr, SV };
3811 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3);
3814 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT,
3815 const SDUse *Ops, unsigned NumOps) {
3817 case 0: return getNode(Opcode, DL, VT);
3818 case 1: return getNode(Opcode, DL, VT, Ops[0]);
3819 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
3820 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
3824 // Copy from an SDUse array into an SDValue array for use with
3825 // the regular getNode logic.
3826 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
3827 return getNode(Opcode, DL, VT, &NewOps[0], NumOps);
3830 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, MVT VT,
3831 const SDValue *Ops, unsigned NumOps) {
3833 case 0: return getNode(Opcode, DL, VT);
3834 case 1: return getNode(Opcode, DL, VT, Ops[0]);
3835 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
3836 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
3842 case ISD::SELECT_CC: {
3843 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
3844 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
3845 "LHS and RHS of condition must have same type!");
3846 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
3847 "True and False arms of SelectCC must have same type!");
3848 assert(Ops[2].getValueType() == VT &&
3849 "select_cc node must be of same type as true and false value!");
3853 assert(NumOps == 5 && "BR_CC takes 5 operands!");
3854 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
3855 "LHS/RHS of comparison should match types!");
3862 SDVTList VTs = getVTList(VT);
3864 if (VT != MVT::Flag) {
3865 FoldingSetNodeID ID;
3866 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
3869 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3870 return SDValue(E, 0);
3872 N = NodeAllocator.Allocate<SDNode>();
3873 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
3874 CSEMap.InsertNode(N, IP);
3876 N = NodeAllocator.Allocate<SDNode>();
3877 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
3880 AllNodes.push_back(N);
3884 return SDValue(N, 0);
3887 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
3888 const std::vector<MVT> &ResultTys,
3889 const SDValue *Ops, unsigned NumOps) {
3890 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()),
3894 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
3895 const MVT *VTs, unsigned NumVTs,
3896 const SDValue *Ops, unsigned NumOps) {
3898 return getNode(Opcode, DL, VTs[0], Ops, NumOps);
3899 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps);
3902 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
3903 const SDValue *Ops, unsigned NumOps) {
3904 if (VTList.NumVTs == 1)
3905 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps);
3908 // FIXME: figure out how to safely handle things like
3909 // int foo(int x) { return 1 << (x & 255); }
3910 // int bar() { return foo(256); }
3912 case ISD::SRA_PARTS:
3913 case ISD::SRL_PARTS:
3914 case ISD::SHL_PARTS:
3915 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3916 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
3917 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
3918 else if (N3.getOpcode() == ISD::AND)
3919 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
3920 // If the and is only masking out bits that cannot effect the shift,
3921 // eliminate the and.
3922 unsigned NumBits = VT.getSizeInBits()*2;
3923 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
3924 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
3930 // Memoize the node unless it returns a flag.
3932 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3933 FoldingSetNodeID ID;
3934 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3936 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3937 return SDValue(E, 0);
3939 N = NodeAllocator.Allocate<UnarySDNode>();
3940 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
3941 } else if (NumOps == 2) {
3942 N = NodeAllocator.Allocate<BinarySDNode>();
3943 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
3944 } else if (NumOps == 3) {
3945 N = NodeAllocator.Allocate<TernarySDNode>();
3946 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
3948 N = NodeAllocator.Allocate<SDNode>();
3949 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
3951 CSEMap.InsertNode(N, IP);
3954 N = NodeAllocator.Allocate<UnarySDNode>();
3955 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
3956 } else if (NumOps == 2) {
3957 N = NodeAllocator.Allocate<BinarySDNode>();
3958 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
3959 } else if (NumOps == 3) {
3960 N = NodeAllocator.Allocate<TernarySDNode>();
3961 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
3963 N = NodeAllocator.Allocate<SDNode>();
3964 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
3967 AllNodes.push_back(N);
3971 return SDValue(N, 0);
3974 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) {
3975 return getNode(Opcode, DL, VTList, 0, 0);
3978 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
3980 SDValue Ops[] = { N1 };
3981 return getNode(Opcode, DL, VTList, Ops, 1);
3984 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
3985 SDValue N1, SDValue N2) {
3986 SDValue Ops[] = { N1, N2 };
3987 return getNode(Opcode, DL, VTList, Ops, 2);
3990 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
3991 SDValue N1, SDValue N2, SDValue N3) {
3992 SDValue Ops[] = { N1, N2, N3 };
3993 return getNode(Opcode, DL, VTList, Ops, 3);
3996 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
3997 SDValue N1, SDValue N2, SDValue N3,
3999 SDValue Ops[] = { N1, N2, N3, N4 };
4000 return getNode(Opcode, DL, VTList, Ops, 4);
4003 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4004 SDValue N1, SDValue N2, SDValue N3,
4005 SDValue N4, SDValue N5) {
4006 SDValue Ops[] = { N1, N2, N3, N4, N5 };
4007 return getNode(Opcode, DL, VTList, Ops, 5);
4010 SDVTList SelectionDAG::getVTList(MVT VT) {
4011 return makeVTList(SDNode::getValueTypeList(VT), 1);
4014 SDVTList SelectionDAG::getVTList(MVT VT1, MVT VT2) {
4015 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4016 E = VTList.rend(); I != E; ++I)
4017 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
4020 MVT *Array = Allocator.Allocate<MVT>(2);
4023 SDVTList Result = makeVTList(Array, 2);
4024 VTList.push_back(Result);
4028 SDVTList SelectionDAG::getVTList(MVT VT1, MVT VT2, MVT VT3) {
4029 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4030 E = VTList.rend(); I != E; ++I)
4031 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4035 MVT *Array = Allocator.Allocate<MVT>(3);
4039 SDVTList Result = makeVTList(Array, 3);
4040 VTList.push_back(Result);
4044 SDVTList SelectionDAG::getVTList(MVT VT1, MVT VT2, MVT VT3, MVT VT4) {
4045 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4046 E = VTList.rend(); I != E; ++I)
4047 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4048 I->VTs[2] == VT3 && I->VTs[3] == VT4)
4051 MVT *Array = Allocator.Allocate<MVT>(3);
4056 SDVTList Result = makeVTList(Array, 4);
4057 VTList.push_back(Result);
4061 SDVTList SelectionDAG::getVTList(const MVT *VTs, unsigned NumVTs) {
4063 case 0: assert(0 && "Cannot have nodes without results!");
4064 case 1: return getVTList(VTs[0]);
4065 case 2: return getVTList(VTs[0], VTs[1]);
4066 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
4070 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4071 E = VTList.rend(); I != E; ++I) {
4072 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
4075 bool NoMatch = false;
4076 for (unsigned i = 2; i != NumVTs; ++i)
4077 if (VTs[i] != I->VTs[i]) {
4085 MVT *Array = Allocator.Allocate<MVT>(NumVTs);
4086 std::copy(VTs, VTs+NumVTs, Array);
4087 SDVTList Result = makeVTList(Array, NumVTs);
4088 VTList.push_back(Result);
4093 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
4094 /// specified operands. If the resultant node already exists in the DAG,
4095 /// this does not modify the specified node, instead it returns the node that
4096 /// already exists. If the resultant node does not exist in the DAG, the
4097 /// input node is returned. As a degenerate case, if you specify the same
4098 /// input operands as the node already has, the input node is returned.
4099 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
4100 SDNode *N = InN.getNode();
4101 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
4103 // Check to see if there is no change.
4104 if (Op == N->getOperand(0)) return InN;
4106 // See if the modified node already exists.
4107 void *InsertPos = 0;
4108 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
4109 return SDValue(Existing, InN.getResNo());
4111 // Nope it doesn't. Remove the node from its current place in the maps.
4113 if (!RemoveNodeFromCSEMaps(N))
4116 // Now we update the operands.
4117 N->OperandList[0].set(Op);
4119 // If this gets put into a CSE map, add it.
4120 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4124 SDValue SelectionDAG::
4125 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
4126 SDNode *N = InN.getNode();
4127 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
4129 // Check to see if there is no change.
4130 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
4131 return InN; // No operands changed, just return the input node.
4133 // See if the modified node already exists.
4134 void *InsertPos = 0;
4135 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
4136 return SDValue(Existing, InN.getResNo());
4138 // Nope it doesn't. Remove the node from its current place in the maps.
4140 if (!RemoveNodeFromCSEMaps(N))
4143 // Now we update the operands.
4144 if (N->OperandList[0] != Op1)
4145 N->OperandList[0].set(Op1);
4146 if (N->OperandList[1] != Op2)
4147 N->OperandList[1].set(Op2);
4149 // If this gets put into a CSE map, add it.
4150 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4154 SDValue SelectionDAG::
4155 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
4156 SDValue Ops[] = { Op1, Op2, Op3 };
4157 return UpdateNodeOperands(N, Ops, 3);
4160 SDValue SelectionDAG::
4161 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4162 SDValue Op3, SDValue Op4) {
4163 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
4164 return UpdateNodeOperands(N, Ops, 4);
4167 SDValue SelectionDAG::
4168 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4169 SDValue Op3, SDValue Op4, SDValue Op5) {
4170 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
4171 return UpdateNodeOperands(N, Ops, 5);
4174 SDValue SelectionDAG::
4175 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
4176 SDNode *N = InN.getNode();
4177 assert(N->getNumOperands() == NumOps &&
4178 "Update with wrong number of operands");
4180 // Check to see if there is no change.
4181 bool AnyChange = false;
4182 for (unsigned i = 0; i != NumOps; ++i) {
4183 if (Ops[i] != N->getOperand(i)) {
4189 // No operands changed, just return the input node.
4190 if (!AnyChange) return InN;
4192 // See if the modified node already exists.
4193 void *InsertPos = 0;
4194 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
4195 return SDValue(Existing, InN.getResNo());
4197 // Nope it doesn't. Remove the node from its current place in the maps.
4199 if (!RemoveNodeFromCSEMaps(N))
4202 // Now we update the operands.
4203 for (unsigned i = 0; i != NumOps; ++i)
4204 if (N->OperandList[i] != Ops[i])
4205 N->OperandList[i].set(Ops[i]);
4207 // If this gets put into a CSE map, add it.
4208 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4212 /// DropOperands - Release the operands and set this node to have
4214 void SDNode::DropOperands() {
4215 // Unlike the code in MorphNodeTo that does this, we don't need to
4216 // watch for dead nodes here.
4217 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
4223 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
4226 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4228 SDVTList VTs = getVTList(VT);
4229 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4232 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4233 MVT VT, SDValue Op1) {
4234 SDVTList VTs = getVTList(VT);
4235 SDValue Ops[] = { Op1 };
4236 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4239 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4240 MVT VT, SDValue Op1,
4242 SDVTList VTs = getVTList(VT);
4243 SDValue Ops[] = { Op1, Op2 };
4244 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4247 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4248 MVT VT, SDValue Op1,
4249 SDValue Op2, SDValue Op3) {
4250 SDVTList VTs = getVTList(VT);
4251 SDValue Ops[] = { Op1, Op2, Op3 };
4252 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4255 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4256 MVT VT, const SDValue *Ops,
4258 SDVTList VTs = getVTList(VT);
4259 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4262 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4263 MVT VT1, MVT VT2, const SDValue *Ops,
4265 SDVTList VTs = getVTList(VT1, VT2);
4266 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4269 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4271 SDVTList VTs = getVTList(VT1, VT2);
4272 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4275 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4276 MVT VT1, MVT VT2, MVT VT3,
4277 const SDValue *Ops, unsigned NumOps) {
4278 SDVTList VTs = getVTList(VT1, VT2, VT3);
4279 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4282 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4283 MVT VT1, MVT VT2, MVT VT3, MVT VT4,
4284 const SDValue *Ops, unsigned NumOps) {
4285 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4286 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4289 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4292 SDVTList VTs = getVTList(VT1, VT2);
4293 SDValue Ops[] = { Op1 };
4294 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4297 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4299 SDValue Op1, SDValue Op2) {
4300 SDVTList VTs = getVTList(VT1, VT2);
4301 SDValue Ops[] = { Op1, Op2 };
4302 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4305 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4307 SDValue Op1, SDValue Op2,
4309 SDVTList VTs = getVTList(VT1, VT2);
4310 SDValue Ops[] = { Op1, Op2, Op3 };
4311 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4314 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4315 MVT VT1, MVT VT2, MVT VT3,
4316 SDValue Op1, SDValue Op2,
4318 SDVTList VTs = getVTList(VT1, VT2, VT3);
4319 SDValue Ops[] = { Op1, Op2, Op3 };
4320 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4323 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4324 SDVTList VTs, const SDValue *Ops,
4326 return MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4329 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4331 SDVTList VTs = getVTList(VT);
4332 return MorphNodeTo(N, Opc, VTs, 0, 0);
4335 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4336 MVT VT, SDValue Op1) {
4337 SDVTList VTs = getVTList(VT);
4338 SDValue Ops[] = { Op1 };
4339 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4342 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4343 MVT VT, SDValue Op1,
4345 SDVTList VTs = getVTList(VT);
4346 SDValue Ops[] = { Op1, Op2 };
4347 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4350 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4351 MVT VT, SDValue Op1,
4352 SDValue Op2, SDValue Op3) {
4353 SDVTList VTs = getVTList(VT);
4354 SDValue Ops[] = { Op1, Op2, Op3 };
4355 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4358 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4359 MVT VT, const SDValue *Ops,
4361 SDVTList VTs = getVTList(VT);
4362 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4365 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4366 MVT VT1, MVT VT2, const SDValue *Ops,
4368 SDVTList VTs = getVTList(VT1, VT2);
4369 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4372 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4374 SDVTList VTs = getVTList(VT1, VT2);
4375 return MorphNodeTo(N, Opc, VTs, (SDValue *)0, 0);
4378 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4379 MVT VT1, MVT VT2, MVT VT3,
4380 const SDValue *Ops, unsigned NumOps) {
4381 SDVTList VTs = getVTList(VT1, VT2, VT3);
4382 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4385 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4388 SDVTList VTs = getVTList(VT1, VT2);
4389 SDValue Ops[] = { Op1 };
4390 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4393 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4395 SDValue Op1, SDValue Op2) {
4396 SDVTList VTs = getVTList(VT1, VT2);
4397 SDValue Ops[] = { Op1, Op2 };
4398 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4401 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4403 SDValue Op1, SDValue Op2,
4405 SDVTList VTs = getVTList(VT1, VT2);
4406 SDValue Ops[] = { Op1, Op2, Op3 };
4407 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4410 /// MorphNodeTo - These *mutate* the specified node to have the specified
4411 /// return type, opcode, and operands.
4413 /// Note that MorphNodeTo returns the resultant node. If there is already a
4414 /// node of the specified opcode and operands, it returns that node instead of
4415 /// the current one. Note that the DebugLoc need not be the same.
4417 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4418 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4419 /// node, and because it doesn't require CSE recalculation for any of
4420 /// the node's users.
4422 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4423 SDVTList VTs, const SDValue *Ops,
4425 // If an identical node already exists, use it.
4427 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4428 FoldingSetNodeID ID;
4429 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4430 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4434 if (!RemoveNodeFromCSEMaps(N))
4437 // Start the morphing.
4439 N->ValueList = VTs.VTs;
4440 N->NumValues = VTs.NumVTs;
4442 // Clear the operands list, updating used nodes to remove this from their
4443 // use list. Keep track of any operands that become dead as a result.
4444 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4445 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
4447 SDNode *Used = Use.getNode();
4449 if (Used->use_empty())
4450 DeadNodeSet.insert(Used);
4453 // If NumOps is larger than the # of operands we currently have, reallocate
4454 // the operand list.
4455 if (NumOps > N->NumOperands) {
4456 if (N->OperandsNeedDelete)
4457 delete[] N->OperandList;
4459 if (N->isMachineOpcode()) {
4460 // We're creating a final node that will live unmorphed for the
4461 // remainder of the current SelectionDAG iteration, so we can allocate
4462 // the operands directly out of a pool with no recycling metadata.
4463 N->OperandList = OperandAllocator.Allocate<SDUse>(NumOps);
4464 N->OperandsNeedDelete = false;
4466 N->OperandList = new SDUse[NumOps];
4467 N->OperandsNeedDelete = true;
4471 // Assign the new operands.
4472 N->NumOperands = NumOps;
4473 for (unsigned i = 0, e = NumOps; i != e; ++i) {
4474 N->OperandList[i].setUser(N);
4475 N->OperandList[i].setInitial(Ops[i]);
4478 // Delete any nodes that are still dead after adding the uses for the
4480 SmallVector<SDNode *, 16> DeadNodes;
4481 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4482 E = DeadNodeSet.end(); I != E; ++I)
4483 if ((*I)->use_empty())
4484 DeadNodes.push_back(*I);
4485 RemoveDeadNodes(DeadNodes);
4488 CSEMap.InsertNode(N, IP); // Memoize the new node.
4493 /// getTargetNode - These are used for target selectors to create a new node
4494 /// with specified return type(s), target opcode, and operands.
4496 /// Note that getTargetNode returns the resultant node. If there is already a
4497 /// node of the specified opcode and operands, it returns that node instead of
4498 /// the current one.
4499 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT) {
4500 return getNode(~Opcode, dl, VT).getNode();
4503 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT,
4505 return getNode(~Opcode, dl, VT, Op1).getNode();
4508 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT,
4509 SDValue Op1, SDValue Op2) {
4510 return getNode(~Opcode, dl, VT, Op1, Op2).getNode();
4513 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT,
4514 SDValue Op1, SDValue Op2,
4516 return getNode(~Opcode, dl, VT, Op1, Op2, Op3).getNode();
4519 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT,
4520 const SDValue *Ops, unsigned NumOps) {
4521 return getNode(~Opcode, dl, VT, Ops, NumOps).getNode();
4524 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl,
4526 SDVTList VTs = getVTList(VT1, VT2);
4528 return getNode(~Opcode, dl, VTs, &Op, 0).getNode();
4531 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT1,
4532 MVT VT2, SDValue Op1) {
4533 SDVTList VTs = getVTList(VT1, VT2);
4534 return getNode(~Opcode, dl, VTs, &Op1, 1).getNode();
4537 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT1,
4538 MVT VT2, SDValue Op1,
4540 SDVTList VTs = getVTList(VT1, VT2);
4541 SDValue Ops[] = { Op1, Op2 };
4542 return getNode(~Opcode, dl, VTs, Ops, 2).getNode();
4545 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT1,
4546 MVT VT2, SDValue Op1,
4547 SDValue Op2, SDValue Op3) {
4548 SDVTList VTs = getVTList(VT1, VT2);
4549 SDValue Ops[] = { Op1, Op2, Op3 };
4550 return getNode(~Opcode, dl, VTs, Ops, 3).getNode();
4553 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl,
4555 const SDValue *Ops, unsigned NumOps) {
4556 SDVTList VTs = getVTList(VT1, VT2);
4557 return getNode(~Opcode, dl, VTs, Ops, NumOps).getNode();
4560 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl,
4561 MVT VT1, MVT VT2, MVT VT3,
4562 SDValue Op1, SDValue Op2) {
4563 SDVTList VTs = getVTList(VT1, VT2, VT3);
4564 SDValue Ops[] = { Op1, Op2 };
4565 return getNode(~Opcode, dl, VTs, Ops, 2).getNode();
4568 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl,
4569 MVT VT1, MVT VT2, MVT VT3,
4570 SDValue Op1, SDValue Op2,
4572 SDVTList VTs = getVTList(VT1, VT2, VT3);
4573 SDValue Ops[] = { Op1, Op2, Op3 };
4574 return getNode(~Opcode, dl, VTs, Ops, 3).getNode();
4577 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl,
4578 MVT VT1, MVT VT2, MVT VT3,
4579 const SDValue *Ops, unsigned NumOps) {
4580 SDVTList VTs = getVTList(VT1, VT2, VT3);
4581 return getNode(~Opcode, dl, VTs, Ops, NumOps).getNode();
4584 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl, MVT VT1,
4585 MVT VT2, MVT VT3, MVT VT4,
4586 const SDValue *Ops, unsigned NumOps) {
4587 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4588 return getNode(~Opcode, dl, VTs, Ops, NumOps).getNode();
4591 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, DebugLoc dl,
4592 const std::vector<MVT> &ResultTys,
4593 const SDValue *Ops, unsigned NumOps) {
4594 return getNode(~Opcode, dl, ResultTys, Ops, NumOps).getNode();
4597 /// getNodeIfExists - Get the specified node if it's already available, or
4598 /// else return NULL.
4599 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4600 const SDValue *Ops, unsigned NumOps) {
4601 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4602 FoldingSetNodeID ID;
4603 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4605 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4611 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4612 /// This can cause recursive merging of nodes in the DAG.
4614 /// This version assumes From has a single result value.
4616 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
4617 DAGUpdateListener *UpdateListener) {
4618 SDNode *From = FromN.getNode();
4619 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
4620 "Cannot replace with this method!");
4621 assert(From != To.getNode() && "Cannot replace uses of with self");
4623 // Iterate over all the existing uses of From. New uses will be added
4624 // to the beginning of the use list, which we avoid visiting.
4625 // This specifically avoids visiting uses of From that arise while the
4626 // replacement is happening, because any such uses would be the result
4627 // of CSE: If an existing node looks like From after one of its operands
4628 // is replaced by To, we don't want to replace of all its users with To
4629 // too. See PR3018 for more info.
4630 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4634 // This node is about to morph, remove its old self from the CSE maps.
4635 RemoveNodeFromCSEMaps(User);
4637 // A user can appear in a use list multiple times, and when this
4638 // happens the uses are usually next to each other in the list.
4639 // To help reduce the number of CSE recomputations, process all
4640 // the uses of this user that we can find this way.
4642 SDUse &Use = UI.getUse();
4645 } while (UI != UE && *UI == User);
4647 // Now that we have modified User, add it back to the CSE maps. If it
4648 // already exists there, recursively merge the results together.
4649 AddModifiedNodeToCSEMaps(User, UpdateListener);
4653 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4654 /// This can cause recursive merging of nodes in the DAG.
4656 /// This version assumes that for each value of From, there is a
4657 /// corresponding value in To in the same position with the same type.
4659 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
4660 DAGUpdateListener *UpdateListener) {
4662 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
4663 assert((!From->hasAnyUseOfValue(i) ||
4664 From->getValueType(i) == To->getValueType(i)) &&
4665 "Cannot use this version of ReplaceAllUsesWith!");
4668 // Handle the trivial case.
4672 // Iterate over just the existing users of From. See the comments in
4673 // the ReplaceAllUsesWith above.
4674 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4678 // This node is about to morph, remove its old self from the CSE maps.
4679 RemoveNodeFromCSEMaps(User);
4681 // A user can appear in a use list multiple times, and when this
4682 // happens the uses are usually next to each other in the list.
4683 // To help reduce the number of CSE recomputations, process all
4684 // the uses of this user that we can find this way.
4686 SDUse &Use = UI.getUse();
4689 } while (UI != UE && *UI == User);
4691 // Now that we have modified User, add it back to the CSE maps. If it
4692 // already exists there, recursively merge the results together.
4693 AddModifiedNodeToCSEMaps(User, UpdateListener);
4697 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4698 /// This can cause recursive merging of nodes in the DAG.
4700 /// This version can replace From with any result values. To must match the
4701 /// number and types of values returned by From.
4702 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
4704 DAGUpdateListener *UpdateListener) {
4705 if (From->getNumValues() == 1) // Handle the simple case efficiently.
4706 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
4708 // Iterate over just the existing users of From. See the comments in
4709 // the ReplaceAllUsesWith above.
4710 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4714 // This node is about to morph, remove its old self from the CSE maps.
4715 RemoveNodeFromCSEMaps(User);
4717 // A user can appear in a use list multiple times, and when this
4718 // happens the uses are usually next to each other in the list.
4719 // To help reduce the number of CSE recomputations, process all
4720 // the uses of this user that we can find this way.
4722 SDUse &Use = UI.getUse();
4723 const SDValue &ToOp = To[Use.getResNo()];
4726 } while (UI != UE && *UI == User);
4728 // Now that we have modified User, add it back to the CSE maps. If it
4729 // already exists there, recursively merge the results together.
4730 AddModifiedNodeToCSEMaps(User, UpdateListener);
4734 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
4735 /// uses of other values produced by From.getNode() alone. The Deleted
4736 /// vector is handled the same way as for ReplaceAllUsesWith.
4737 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
4738 DAGUpdateListener *UpdateListener){
4739 // Handle the really simple, really trivial case efficiently.
4740 if (From == To) return;
4742 // Handle the simple, trivial, case efficiently.
4743 if (From.getNode()->getNumValues() == 1) {
4744 ReplaceAllUsesWith(From, To, UpdateListener);
4748 // Iterate over just the existing users of From. See the comments in
4749 // the ReplaceAllUsesWith above.
4750 SDNode::use_iterator UI = From.getNode()->use_begin(),
4751 UE = From.getNode()->use_end();
4754 bool UserRemovedFromCSEMaps = false;
4756 // A user can appear in a use list multiple times, and when this
4757 // happens the uses are usually next to each other in the list.
4758 // To help reduce the number of CSE recomputations, process all
4759 // the uses of this user that we can find this way.
4761 SDUse &Use = UI.getUse();
4763 // Skip uses of different values from the same node.
4764 if (Use.getResNo() != From.getResNo()) {
4769 // If this node hasn't been modified yet, it's still in the CSE maps,
4770 // so remove its old self from the CSE maps.
4771 if (!UserRemovedFromCSEMaps) {
4772 RemoveNodeFromCSEMaps(User);
4773 UserRemovedFromCSEMaps = true;
4778 } while (UI != UE && *UI == User);
4780 // We are iterating over all uses of the From node, so if a use
4781 // doesn't use the specific value, no changes are made.
4782 if (!UserRemovedFromCSEMaps)
4785 // Now that we have modified User, add it back to the CSE maps. If it
4786 // already exists there, recursively merge the results together.
4787 AddModifiedNodeToCSEMaps(User, UpdateListener);
4792 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
4793 /// to record information about a use.
4800 /// operator< - Sort Memos by User.
4801 bool operator<(const UseMemo &L, const UseMemo &R) {
4802 return (intptr_t)L.User < (intptr_t)R.User;
4806 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
4807 /// uses of other values produced by From.getNode() alone. The same value
4808 /// may appear in both the From and To list. The Deleted vector is
4809 /// handled the same way as for ReplaceAllUsesWith.
4810 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
4813 DAGUpdateListener *UpdateListener){
4814 // Handle the simple, trivial case efficiently.
4816 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
4818 // Read up all the uses and make records of them. This helps
4819 // processing new uses that are introduced during the
4820 // replacement process.
4821 SmallVector<UseMemo, 4> Uses;
4822 for (unsigned i = 0; i != Num; ++i) {
4823 unsigned FromResNo = From[i].getResNo();
4824 SDNode *FromNode = From[i].getNode();
4825 for (SDNode::use_iterator UI = FromNode->use_begin(),
4826 E = FromNode->use_end(); UI != E; ++UI) {
4827 SDUse &Use = UI.getUse();
4828 if (Use.getResNo() == FromResNo) {
4829 UseMemo Memo = { *UI, i, &Use };
4830 Uses.push_back(Memo);
4835 // Sort the uses, so that all the uses from a given User are together.
4836 std::sort(Uses.begin(), Uses.end());
4838 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
4839 UseIndex != UseIndexEnd; ) {
4840 // We know that this user uses some value of From. If it is the right
4841 // value, update it.
4842 SDNode *User = Uses[UseIndex].User;
4844 // This node is about to morph, remove its old self from the CSE maps.
4845 RemoveNodeFromCSEMaps(User);
4847 // The Uses array is sorted, so all the uses for a given User
4848 // are next to each other in the list.
4849 // To help reduce the number of CSE recomputations, process all
4850 // the uses of this user that we can find this way.
4852 unsigned i = Uses[UseIndex].Index;
4853 SDUse &Use = *Uses[UseIndex].Use;
4857 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
4859 // Now that we have modified User, add it back to the CSE maps. If it
4860 // already exists there, recursively merge the results together.
4861 AddModifiedNodeToCSEMaps(User, UpdateListener);
4865 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
4866 /// based on their topological order. It returns the maximum id and a vector
4867 /// of the SDNodes* in assigned order by reference.
4868 unsigned SelectionDAG::AssignTopologicalOrder() {
4870 unsigned DAGSize = 0;
4872 // SortedPos tracks the progress of the algorithm. Nodes before it are
4873 // sorted, nodes after it are unsorted. When the algorithm completes
4874 // it is at the end of the list.
4875 allnodes_iterator SortedPos = allnodes_begin();
4877 // Visit all the nodes. Move nodes with no operands to the front of
4878 // the list immediately. Annotate nodes that do have operands with their
4879 // operand count. Before we do this, the Node Id fields of the nodes
4880 // may contain arbitrary values. After, the Node Id fields for nodes
4881 // before SortedPos will contain the topological sort index, and the
4882 // Node Id fields for nodes At SortedPos and after will contain the
4883 // count of outstanding operands.
4884 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
4886 unsigned Degree = N->getNumOperands();
4888 // A node with no uses, add it to the result array immediately.
4889 N->setNodeId(DAGSize++);
4890 allnodes_iterator Q = N;
4892 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
4895 // Temporarily use the Node Id as scratch space for the degree count.
4896 N->setNodeId(Degree);
4900 // Visit all the nodes. As we iterate, moves nodes into sorted order,
4901 // such that by the time the end is reached all nodes will be sorted.
4902 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
4904 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4907 unsigned Degree = P->getNodeId();
4910 // All of P's operands are sorted, so P may sorted now.
4911 P->setNodeId(DAGSize++);
4913 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
4916 // Update P's outstanding operand count.
4917 P->setNodeId(Degree);
4922 assert(SortedPos == AllNodes.end() &&
4923 "Topological sort incomplete!");
4924 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
4925 "First node in topological sort is not the entry token!");
4926 assert(AllNodes.front().getNodeId() == 0 &&
4927 "First node in topological sort has non-zero id!");
4928 assert(AllNodes.front().getNumOperands() == 0 &&
4929 "First node in topological sort has operands!");
4930 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
4931 "Last node in topologic sort has unexpected id!");
4932 assert(AllNodes.back().use_empty() &&
4933 "Last node in topologic sort has users!");
4934 assert(DAGSize == allnodes_size() && "Node count mismatch!");
4940 //===----------------------------------------------------------------------===//
4942 //===----------------------------------------------------------------------===//
4944 HandleSDNode::~HandleSDNode() {
4948 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA,
4949 MVT VT, int64_t o, unsigned char TF)
4950 : SDNode(Opc, DebugLoc::getUnknownLoc(), getSDVTList(VT)),
4951 Offset(o), TargetFlags(TF) {
4952 TheGlobal = const_cast<GlobalValue*>(GA);
4955 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, MVT memvt,
4956 const Value *srcValue, int SVO,
4957 unsigned alignment, bool vol)
4958 : SDNode(Opc, dl, VTs), MemoryVT(memvt), SrcValue(srcValue), SVOffset(SVO) {
4959 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, vol, alignment);
4960 assert(isPowerOf2_32(alignment) && "Alignment is not a power of 2!");
4961 assert(getAlignment() == alignment && "Alignment representation error!");
4962 assert(isVolatile() == vol && "Volatile representation error!");
4965 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs,
4967 unsigned NumOps, MVT memvt, const Value *srcValue,
4968 int SVO, unsigned alignment, bool vol)
4969 : SDNode(Opc, dl, VTs, Ops, NumOps),
4970 MemoryVT(memvt), SrcValue(srcValue), SVOffset(SVO) {
4971 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, vol, alignment);
4972 assert(isPowerOf2_32(alignment) && "Alignment is not a power of 2!");
4973 assert(getAlignment() == alignment && "Alignment representation error!");
4974 assert(isVolatile() == vol && "Volatile representation error!");
4977 /// getMemOperand - Return a MachineMemOperand object describing the memory
4978 /// reference performed by this memory reference.
4979 MachineMemOperand MemSDNode::getMemOperand() const {
4981 if (isa<LoadSDNode>(this))
4982 Flags = MachineMemOperand::MOLoad;
4983 else if (isa<StoreSDNode>(this))
4984 Flags = MachineMemOperand::MOStore;
4985 else if (isa<AtomicSDNode>(this)) {
4986 Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
4989 const MemIntrinsicSDNode* MemIntrinNode = dyn_cast<MemIntrinsicSDNode>(this);
4990 assert(MemIntrinNode && "Unknown MemSDNode opcode!");
4991 if (MemIntrinNode->readMem()) Flags |= MachineMemOperand::MOLoad;
4992 if (MemIntrinNode->writeMem()) Flags |= MachineMemOperand::MOStore;
4995 int Size = (getMemoryVT().getSizeInBits() + 7) >> 3;
4996 if (isVolatile()) Flags |= MachineMemOperand::MOVolatile;
4998 // Check if the memory reference references a frame index
4999 const FrameIndexSDNode *FI =
5000 dyn_cast<const FrameIndexSDNode>(getBasePtr().getNode());
5001 if (!getSrcValue() && FI)
5002 return MachineMemOperand(PseudoSourceValue::getFixedStack(FI->getIndex()),
5003 Flags, 0, Size, getAlignment());
5005 return MachineMemOperand(getSrcValue(), Flags, getSrcValueOffset(),
5006 Size, getAlignment());
5009 /// Profile - Gather unique data for the node.
5011 void SDNode::Profile(FoldingSetNodeID &ID) const {
5012 AddNodeIDNode(ID, this);
5015 static ManagedStatic<std::set<MVT, MVT::compareRawBits> > EVTs;
5016 static MVT VTs[MVT::LAST_VALUETYPE];
5017 static ManagedStatic<sys::SmartMutex<true> > VTMutex;
5019 /// getValueTypeList - Return a pointer to the specified value type.
5021 const MVT *SDNode::getValueTypeList(MVT VT) {
5022 sys::SmartScopedLock<true> Lock(*VTMutex);
5023 if (VT.isExtended()) {
5024 return &(*EVTs->insert(VT).first);
5026 VTs[VT.getSimpleVT()] = VT;
5027 return &VTs[VT.getSimpleVT()];
5031 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
5032 /// indicated value. This method ignores uses of other values defined by this
5034 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
5035 assert(Value < getNumValues() && "Bad value!");
5037 // TODO: Only iterate over uses of a given value of the node
5038 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
5039 if (UI.getUse().getResNo() == Value) {
5046 // Found exactly the right number of uses?
5051 /// hasAnyUseOfValue - Return true if there are any use of the indicated
5052 /// value. This method ignores uses of other values defined by this operation.
5053 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
5054 assert(Value < getNumValues() && "Bad value!");
5056 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
5057 if (UI.getUse().getResNo() == Value)
5064 /// isOnlyUserOf - Return true if this node is the only use of N.
5066 bool SDNode::isOnlyUserOf(SDNode *N) const {
5068 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
5079 /// isOperand - Return true if this node is an operand of N.
5081 bool SDValue::isOperandOf(SDNode *N) const {
5082 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5083 if (*this == N->getOperand(i))
5088 bool SDNode::isOperandOf(SDNode *N) const {
5089 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
5090 if (this == N->OperandList[i].getNode())
5095 /// reachesChainWithoutSideEffects - Return true if this operand (which must
5096 /// be a chain) reaches the specified operand without crossing any
5097 /// side-effecting instructions. In practice, this looks through token
5098 /// factors and non-volatile loads. In order to remain efficient, this only
5099 /// looks a couple of nodes in, it does not do an exhaustive search.
5100 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
5101 unsigned Depth) const {
5102 if (*this == Dest) return true;
5104 // Don't search too deeply, we just want to be able to see through
5105 // TokenFactor's etc.
5106 if (Depth == 0) return false;
5108 // If this is a token factor, all inputs to the TF happen in parallel. If any
5109 // of the operands of the TF reach dest, then we can do the xform.
5110 if (getOpcode() == ISD::TokenFactor) {
5111 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5112 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
5117 // Loads don't have side effects, look through them.
5118 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
5119 if (!Ld->isVolatile())
5120 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
5126 static void findPredecessor(SDNode *N, const SDNode *P, bool &found,
5127 SmallPtrSet<SDNode *, 32> &Visited) {
5128 if (found || !Visited.insert(N))
5131 for (unsigned i = 0, e = N->getNumOperands(); !found && i != e; ++i) {
5132 SDNode *Op = N->getOperand(i).getNode();
5137 findPredecessor(Op, P, found, Visited);
5141 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
5142 /// is either an operand of N or it can be reached by recursively traversing
5143 /// up the operands.
5144 /// NOTE: this is an expensive method. Use it carefully.
5145 bool SDNode::isPredecessorOf(SDNode *N) const {
5146 SmallPtrSet<SDNode *, 32> Visited;
5148 findPredecessor(N, this, found, Visited);
5152 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
5153 assert(Num < NumOperands && "Invalid child # of SDNode!");
5154 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
5157 std::string SDNode::getOperationName(const SelectionDAG *G) const {
5158 switch (getOpcode()) {
5160 if (getOpcode() < ISD::BUILTIN_OP_END)
5161 return "<<Unknown DAG Node>>";
5162 if (isMachineOpcode()) {
5164 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
5165 if (getMachineOpcode() < TII->getNumOpcodes())
5166 return TII->get(getMachineOpcode()).getName();
5167 return "<<Unknown Machine Node>>";
5170 const TargetLowering &TLI = G->getTargetLoweringInfo();
5171 const char *Name = TLI.getTargetNodeName(getOpcode());
5172 if (Name) return Name;
5173 return "<<Unknown Target Node>>";
5175 return "<<Unknown Node>>";
5178 case ISD::DELETED_NODE:
5179 return "<<Deleted Node!>>";
5181 case ISD::PREFETCH: return "Prefetch";
5182 case ISD::MEMBARRIER: return "MemBarrier";
5183 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap";
5184 case ISD::ATOMIC_SWAP: return "AtomicSwap";
5185 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd";
5186 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub";
5187 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd";
5188 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr";
5189 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor";
5190 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand";
5191 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin";
5192 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax";
5193 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin";
5194 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax";
5195 case ISD::PCMARKER: return "PCMarker";
5196 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5197 case ISD::SRCVALUE: return "SrcValue";
5198 case ISD::MEMOPERAND: return "MemOperand";
5199 case ISD::EntryToken: return "EntryToken";
5200 case ISD::TokenFactor: return "TokenFactor";
5201 case ISD::AssertSext: return "AssertSext";
5202 case ISD::AssertZext: return "AssertZext";
5204 case ISD::BasicBlock: return "BasicBlock";
5205 case ISD::ARG_FLAGS: return "ArgFlags";
5206 case ISD::VALUETYPE: return "ValueType";
5207 case ISD::Register: return "Register";
5209 case ISD::Constant: return "Constant";
5210 case ISD::ConstantFP: return "ConstantFP";
5211 case ISD::GlobalAddress: return "GlobalAddress";
5212 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5213 case ISD::FrameIndex: return "FrameIndex";
5214 case ISD::JumpTable: return "JumpTable";
5215 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5216 case ISD::RETURNADDR: return "RETURNADDR";
5217 case ISD::FRAMEADDR: return "FRAMEADDR";
5218 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5219 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5220 case ISD::EHSELECTION: return "EHSELECTION";
5221 case ISD::EH_RETURN: return "EH_RETURN";
5222 case ISD::ConstantPool: return "ConstantPool";
5223 case ISD::ExternalSymbol: return "ExternalSymbol";
5224 case ISD::INTRINSIC_WO_CHAIN: {
5225 unsigned IID = cast<ConstantSDNode>(getOperand(0))->getZExtValue();
5226 return Intrinsic::getName((Intrinsic::ID)IID);
5228 case ISD::INTRINSIC_VOID:
5229 case ISD::INTRINSIC_W_CHAIN: {
5230 unsigned IID = cast<ConstantSDNode>(getOperand(1))->getZExtValue();
5231 return Intrinsic::getName((Intrinsic::ID)IID);
5234 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5235 case ISD::TargetConstant: return "TargetConstant";
5236 case ISD::TargetConstantFP:return "TargetConstantFP";
5237 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5238 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5239 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5240 case ISD::TargetJumpTable: return "TargetJumpTable";
5241 case ISD::TargetConstantPool: return "TargetConstantPool";
5242 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5244 case ISD::CopyToReg: return "CopyToReg";
5245 case ISD::CopyFromReg: return "CopyFromReg";
5246 case ISD::UNDEF: return "undef";
5247 case ISD::MERGE_VALUES: return "merge_values";
5248 case ISD::INLINEASM: return "inlineasm";
5249 case ISD::DBG_LABEL: return "dbg_label";
5250 case ISD::EH_LABEL: return "eh_label";
5251 case ISD::DECLARE: return "declare";
5252 case ISD::HANDLENODE: return "handlenode";
5253 case ISD::FORMAL_ARGUMENTS: return "formal_arguments";
5254 case ISD::CALL: return "call";
5257 case ISD::FABS: return "fabs";
5258 case ISD::FNEG: return "fneg";
5259 case ISD::FSQRT: return "fsqrt";
5260 case ISD::FSIN: return "fsin";
5261 case ISD::FCOS: return "fcos";
5262 case ISD::FPOWI: return "fpowi";
5263 case ISD::FPOW: return "fpow";
5264 case ISD::FTRUNC: return "ftrunc";
5265 case ISD::FFLOOR: return "ffloor";
5266 case ISD::FCEIL: return "fceil";
5267 case ISD::FRINT: return "frint";
5268 case ISD::FNEARBYINT: return "fnearbyint";
5271 case ISD::ADD: return "add";
5272 case ISD::SUB: return "sub";
5273 case ISD::MUL: return "mul";
5274 case ISD::MULHU: return "mulhu";
5275 case ISD::MULHS: return "mulhs";
5276 case ISD::SDIV: return "sdiv";
5277 case ISD::UDIV: return "udiv";
5278 case ISD::SREM: return "srem";
5279 case ISD::UREM: return "urem";
5280 case ISD::SMUL_LOHI: return "smul_lohi";
5281 case ISD::UMUL_LOHI: return "umul_lohi";
5282 case ISD::SDIVREM: return "sdivrem";
5283 case ISD::UDIVREM: return "udivrem";
5284 case ISD::AND: return "and";
5285 case ISD::OR: return "or";
5286 case ISD::XOR: return "xor";
5287 case ISD::SHL: return "shl";
5288 case ISD::SRA: return "sra";
5289 case ISD::SRL: return "srl";
5290 case ISD::ROTL: return "rotl";
5291 case ISD::ROTR: return "rotr";
5292 case ISD::FADD: return "fadd";
5293 case ISD::FSUB: return "fsub";
5294 case ISD::FMUL: return "fmul";
5295 case ISD::FDIV: return "fdiv";
5296 case ISD::FREM: return "frem";
5297 case ISD::FCOPYSIGN: return "fcopysign";
5298 case ISD::FGETSIGN: return "fgetsign";
5300 case ISD::SETCC: return "setcc";
5301 case ISD::VSETCC: return "vsetcc";
5302 case ISD::SELECT: return "select";
5303 case ISD::SELECT_CC: return "select_cc";
5304 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5305 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5306 case ISD::CONCAT_VECTORS: return "concat_vectors";
5307 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5308 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5309 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5310 case ISD::CARRY_FALSE: return "carry_false";
5311 case ISD::ADDC: return "addc";
5312 case ISD::ADDE: return "adde";
5313 case ISD::SADDO: return "saddo";
5314 case ISD::UADDO: return "uaddo";
5315 case ISD::SSUBO: return "ssubo";
5316 case ISD::USUBO: return "usubo";
5317 case ISD::SMULO: return "smulo";
5318 case ISD::UMULO: return "umulo";
5319 case ISD::SUBC: return "subc";
5320 case ISD::SUBE: return "sube";
5321 case ISD::SHL_PARTS: return "shl_parts";
5322 case ISD::SRA_PARTS: return "sra_parts";
5323 case ISD::SRL_PARTS: return "srl_parts";
5325 // Conversion operators.
5326 case ISD::SIGN_EXTEND: return "sign_extend";
5327 case ISD::ZERO_EXTEND: return "zero_extend";
5328 case ISD::ANY_EXTEND: return "any_extend";
5329 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5330 case ISD::TRUNCATE: return "truncate";
5331 case ISD::FP_ROUND: return "fp_round";
5332 case ISD::FLT_ROUNDS_: return "flt_rounds";
5333 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5334 case ISD::FP_EXTEND: return "fp_extend";
5336 case ISD::SINT_TO_FP: return "sint_to_fp";
5337 case ISD::UINT_TO_FP: return "uint_to_fp";
5338 case ISD::FP_TO_SINT: return "fp_to_sint";
5339 case ISD::FP_TO_UINT: return "fp_to_uint";
5340 case ISD::BIT_CONVERT: return "bit_convert";
5342 case ISD::CONVERT_RNDSAT: {
5343 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) {
5344 default: assert(0 && "Unknown cvt code!");
5345 case ISD::CVT_FF: return "cvt_ff";
5346 case ISD::CVT_FS: return "cvt_fs";
5347 case ISD::CVT_FU: return "cvt_fu";
5348 case ISD::CVT_SF: return "cvt_sf";
5349 case ISD::CVT_UF: return "cvt_uf";
5350 case ISD::CVT_SS: return "cvt_ss";
5351 case ISD::CVT_SU: return "cvt_su";
5352 case ISD::CVT_US: return "cvt_us";
5353 case ISD::CVT_UU: return "cvt_uu";
5357 // Control flow instructions
5358 case ISD::BR: return "br";
5359 case ISD::BRIND: return "brind";
5360 case ISD::BR_JT: return "br_jt";
5361 case ISD::BRCOND: return "brcond";
5362 case ISD::BR_CC: return "br_cc";
5363 case ISD::RET: return "ret";
5364 case ISD::CALLSEQ_START: return "callseq_start";
5365 case ISD::CALLSEQ_END: return "callseq_end";
5368 case ISD::LOAD: return "load";
5369 case ISD::STORE: return "store";
5370 case ISD::VAARG: return "vaarg";
5371 case ISD::VACOPY: return "vacopy";
5372 case ISD::VAEND: return "vaend";
5373 case ISD::VASTART: return "vastart";
5374 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5375 case ISD::EXTRACT_ELEMENT: return "extract_element";
5376 case ISD::BUILD_PAIR: return "build_pair";
5377 case ISD::STACKSAVE: return "stacksave";
5378 case ISD::STACKRESTORE: return "stackrestore";
5379 case ISD::TRAP: return "trap";
5382 case ISD::BSWAP: return "bswap";
5383 case ISD::CTPOP: return "ctpop";
5384 case ISD::CTTZ: return "cttz";
5385 case ISD::CTLZ: return "ctlz";
5388 case ISD::DBG_STOPPOINT: return "dbg_stoppoint";
5389 case ISD::DEBUG_LOC: return "debug_loc";
5392 case ISD::TRAMPOLINE: return "trampoline";
5395 switch (cast<CondCodeSDNode>(this)->get()) {
5396 default: assert(0 && "Unknown setcc condition!");
5397 case ISD::SETOEQ: return "setoeq";
5398 case ISD::SETOGT: return "setogt";
5399 case ISD::SETOGE: return "setoge";
5400 case ISD::SETOLT: return "setolt";
5401 case ISD::SETOLE: return "setole";
5402 case ISD::SETONE: return "setone";
5404 case ISD::SETO: return "seto";
5405 case ISD::SETUO: return "setuo";
5406 case ISD::SETUEQ: return "setue";
5407 case ISD::SETUGT: return "setugt";
5408 case ISD::SETUGE: return "setuge";
5409 case ISD::SETULT: return "setult";
5410 case ISD::SETULE: return "setule";
5411 case ISD::SETUNE: return "setune";
5413 case ISD::SETEQ: return "seteq";
5414 case ISD::SETGT: return "setgt";
5415 case ISD::SETGE: return "setge";
5416 case ISD::SETLT: return "setlt";
5417 case ISD::SETLE: return "setle";
5418 case ISD::SETNE: return "setne";
5423 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5432 return "<post-inc>";
5434 return "<post-dec>";
5438 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5439 std::string S = "< ";
5453 if (getByValAlign())
5454 S += "byval-align:" + utostr(getByValAlign()) + " ";
5456 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5458 S += "byval-size:" + utostr(getByValSize()) + " ";
5462 void SDNode::dump() const { dump(0); }
5463 void SDNode::dump(const SelectionDAG *G) const {
5467 void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const {
5468 OS << (void*)this << ": ";
5470 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5472 if (getValueType(i) == MVT::Other)
5475 OS << getValueType(i).getMVTString();
5477 OS << " = " << getOperationName(G);
5480 void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const {
5481 if (!isTargetOpcode() && getOpcode() == ISD::VECTOR_SHUFFLE) {
5482 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(this);
5484 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) {
5485 int Idx = SVN->getMaskElt(i);
5495 if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5496 OS << '<' << CSDN->getAPIntValue() << '>';
5497 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5498 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5499 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5500 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5501 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5504 CSDN->getValueAPF().bitcastToAPInt().dump();
5507 } else if (const GlobalAddressSDNode *GADN =
5508 dyn_cast<GlobalAddressSDNode>(this)) {
5509 int64_t offset = GADN->getOffset();
5511 WriteAsOperand(OS, GADN->getGlobal());
5514 OS << " + " << offset;
5516 OS << " " << offset;
5517 if (unsigned char TF = GADN->getTargetFlags())
5518 OS << " [TF=" << TF << ']';
5519 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5520 OS << "<" << FIDN->getIndex() << ">";
5521 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5522 OS << "<" << JTDN->getIndex() << ">";
5523 if (unsigned char TF = JTDN->getTargetFlags())
5524 OS << " [TF=" << TF << ']';
5525 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5526 int offset = CP->getOffset();
5527 if (CP->isMachineConstantPoolEntry())
5528 OS << "<" << *CP->getMachineCPVal() << ">";
5530 OS << "<" << *CP->getConstVal() << ">";
5532 OS << " + " << offset;
5534 OS << " " << offset;
5535 if (unsigned char TF = CP->getTargetFlags())
5536 OS << " [TF=" << TF << ']';
5537 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5539 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5541 OS << LBB->getName() << " ";
5542 OS << (const void*)BBDN->getBasicBlock() << ">";
5543 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5544 if (G && R->getReg() &&
5545 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5546 OS << " " << G->getTarget().getRegisterInfo()->getName(R->getReg());
5548 OS << " #" << R->getReg();
5550 } else if (const ExternalSymbolSDNode *ES =
5551 dyn_cast<ExternalSymbolSDNode>(this)) {
5552 OS << "'" << ES->getSymbol() << "'";
5553 if (unsigned char TF = ES->getTargetFlags())
5554 OS << " [TF=" << TF << ']';
5555 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5557 OS << "<" << M->getValue() << ">";
5560 } else if (const MemOperandSDNode *M = dyn_cast<MemOperandSDNode>(this)) {
5561 if (M->MO.getValue())
5562 OS << "<" << M->MO.getValue() << ":" << M->MO.getOffset() << ">";
5564 OS << "<null:" << M->MO.getOffset() << ">";
5565 } else if (const ARG_FLAGSSDNode *N = dyn_cast<ARG_FLAGSSDNode>(this)) {
5566 OS << N->getArgFlags().getArgFlagsString();
5567 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5568 OS << ":" << N->getVT().getMVTString();
5570 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5571 const Value *SrcValue = LD->getSrcValue();
5572 int SrcOffset = LD->getSrcValueOffset();
5578 OS << ":" << SrcOffset << ">";
5581 switch (LD->getExtensionType()) {
5582 default: doExt = false; break;
5583 case ISD::EXTLOAD: OS << " <anyext "; break;
5584 case ISD::SEXTLOAD: OS << " <sext "; break;
5585 case ISD::ZEXTLOAD: OS << " <zext "; break;
5588 OS << LD->getMemoryVT().getMVTString() << ">";
5590 const char *AM = getIndexedModeName(LD->getAddressingMode());
5593 if (LD->isVolatile())
5594 OS << " <volatile>";
5595 OS << " alignment=" << LD->getAlignment();
5596 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
5597 const Value *SrcValue = ST->getSrcValue();
5598 int SrcOffset = ST->getSrcValueOffset();
5604 OS << ":" << SrcOffset << ">";
5606 if (ST->isTruncatingStore())
5607 OS << " <trunc " << ST->getMemoryVT().getMVTString() << ">";
5609 const char *AM = getIndexedModeName(ST->getAddressingMode());
5612 if (ST->isVolatile())
5613 OS << " <volatile>";
5614 OS << " alignment=" << ST->getAlignment();
5615 } else if (const AtomicSDNode* AT = dyn_cast<AtomicSDNode>(this)) {
5616 const Value *SrcValue = AT->getSrcValue();
5617 int SrcOffset = AT->getSrcValueOffset();
5623 OS << ":" << SrcOffset << ">";
5624 if (AT->isVolatile())
5625 OS << " <volatile>";
5626 OS << " alignment=" << AT->getAlignment();
5630 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
5633 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
5635 OS << (void*)getOperand(i).getNode();
5636 if (unsigned RN = getOperand(i).getResNo())
5639 print_details(OS, G);
5642 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
5643 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5644 if (N->getOperand(i).getNode()->hasOneUse())
5645 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
5647 cerr << "\n" << std::string(indent+2, ' ')
5648 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
5651 cerr << "\n" << std::string(indent, ' ');
5655 void SelectionDAG::dump() const {
5656 cerr << "SelectionDAG has " << AllNodes.size() << " nodes:";
5658 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
5660 const SDNode *N = I;
5661 if (!N->hasOneUse() && N != getRoot().getNode())
5662 DumpNodes(N, 2, this);
5665 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
5670 void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
5672 print_details(OS, G);
5675 typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet;
5676 static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent,
5677 const SelectionDAG *G, VisitedSDNodeSet &once) {
5678 if (!once.insert(N)) // If we've been here before, return now.
5680 // Dump the current SDNode, but don't end the line yet.
5681 OS << std::string(indent, ' ');
5683 // Having printed this SDNode, walk the children:
5684 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5685 const SDNode *child = N->getOperand(i).getNode();
5688 if (child->getNumOperands() == 0) {
5689 // This child has no grandchildren; print it inline right here.
5690 child->printr(OS, G);
5692 } else { // Just the address. FIXME: also print the child's opcode
5694 if (unsigned RN = N->getOperand(i).getResNo())
5699 // Dump children that have grandchildren on their own line(s).
5700 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5701 const SDNode *child = N->getOperand(i).getNode();
5702 DumpNodesr(OS, child, indent+2, G, once);
5706 void SDNode::dumpr() const {
5707 VisitedSDNodeSet once;
5708 DumpNodesr(errs(), this, 0, 0, once);
5712 // getAddressSpace - Return the address space this GlobalAddress belongs to.
5713 unsigned GlobalAddressSDNode::getAddressSpace() const {
5714 return getGlobal()->getType()->getAddressSpace();
5718 const Type *ConstantPoolSDNode::getType() const {
5719 if (isMachineConstantPoolEntry())
5720 return Val.MachineCPVal->getType();
5721 return Val.ConstVal->getType();
5724 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue,
5726 unsigned &SplatBitSize,
5728 unsigned MinSplatBits) {
5729 MVT VT = getValueType(0);
5730 assert(VT.isVector() && "Expected a vector type");
5731 unsigned sz = VT.getSizeInBits();
5732 if (MinSplatBits > sz)
5735 SplatValue = APInt(sz, 0);
5736 SplatUndef = APInt(sz, 0);
5738 // Get the bits. Bits with undefined values (when the corresponding element
5739 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
5740 // in SplatValue. If any of the values are not constant, give up and return
5742 unsigned int nOps = getNumOperands();
5743 assert(nOps > 0 && "isConstantSplat has 0-size build vector");
5744 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits();
5745 for (unsigned i = 0; i < nOps; ++i) {
5746 SDValue OpVal = getOperand(i);
5747 unsigned BitPos = i * EltBitSize;
5749 if (OpVal.getOpcode() == ISD::UNDEF)
5750 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos +EltBitSize);
5751 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal))
5752 SplatValue |= (APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize).
5753 zextOrTrunc(sz) << BitPos);
5754 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal))
5755 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos;
5760 // The build_vector is all constants or undefs. Find the smallest element
5761 // size that splats the vector.
5763 HasAnyUndefs = (SplatUndef != 0);
5766 unsigned HalfSize = sz / 2;
5767 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize);
5768 APInt LowValue = APInt(SplatValue).trunc(HalfSize);
5769 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize);
5770 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize);
5772 // If the two halves do not match (ignoring undef bits), stop here.
5773 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
5774 MinSplatBits > HalfSize)
5777 SplatValue = HighValue | LowValue;
5778 SplatUndef = HighUndef & LowUndef;
5787 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, MVT VT) {
5788 // Find the first non-undef value in the shuffle mask.
5790 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
5793 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!");
5795 // Make sure all remaining elements are either undef or the same as the first
5797 for (int Idx = Mask[i]; i != e; ++i)
5798 if (Mask[i] >= 0 && Mask[i] != Idx)