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 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "SDNodeOrdering.h"
16 #include "llvm/Constants.h"
17 #include "llvm/Analysis/ValueTracking.h"
18 #include "llvm/Function.h"
19 #include "llvm/GlobalAlias.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Assembly/Writer.h"
24 #include "llvm/CallingConv.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/PseudoSourceValue.h"
30 #include "llvm/Target/TargetRegisterInfo.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/Target/TargetFrameInfo.h"
33 #include "llvm/Target/TargetLowering.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetIntrinsicInfo.h"
37 #include "llvm/Target/TargetMachine.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/ManagedStatic.h"
42 #include "llvm/Support/MathExtras.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/System/Mutex.h"
45 #include "llvm/ADT/SetVector.h"
46 #include "llvm/ADT/SmallPtrSet.h"
47 #include "llvm/ADT/SmallSet.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/StringExtras.h"
54 /// makeVTList - Return an instance of the SDVTList struct initialized with the
55 /// specified members.
56 static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
57 SDVTList Res = {VTs, NumVTs};
61 static const fltSemantics *EVTToAPFloatSemantics(EVT VT) {
62 switch (VT.getSimpleVT().SimpleTy) {
63 default: llvm_unreachable("Unknown FP format");
64 case MVT::f32: return &APFloat::IEEEsingle;
65 case MVT::f64: return &APFloat::IEEEdouble;
66 case MVT::f80: return &APFloat::x87DoubleExtended;
67 case MVT::f128: return &APFloat::IEEEquad;
68 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
72 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
74 //===----------------------------------------------------------------------===//
75 // ConstantFPSDNode Class
76 //===----------------------------------------------------------------------===//
78 /// isExactlyValue - We don't rely on operator== working on double values, as
79 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
80 /// As such, this method can be used to do an exact bit-for-bit comparison of
81 /// two floating point values.
82 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
83 return getValueAPF().bitwiseIsEqual(V);
86 bool ConstantFPSDNode::isValueValidForType(EVT VT,
88 assert(VT.isFloatingPoint() && "Can only convert between FP types");
90 // PPC long double cannot be converted to any other type.
91 if (VT == MVT::ppcf128 ||
92 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
95 // convert modifies in place, so make a copy.
96 APFloat Val2 = APFloat(Val);
98 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
103 //===----------------------------------------------------------------------===//
105 //===----------------------------------------------------------------------===//
107 /// isBuildVectorAllOnes - Return true if the specified node is a
108 /// BUILD_VECTOR where all of the elements are ~0 or undef.
109 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
110 // Look through a bit convert.
111 if (N->getOpcode() == ISD::BIT_CONVERT)
112 N = N->getOperand(0).getNode();
114 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
116 unsigned i = 0, e = N->getNumOperands();
118 // Skip over all of the undef values.
119 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
122 // Do not accept an all-undef vector.
123 if (i == e) return false;
125 // Do not accept build_vectors that aren't all constants or which have non-~0
127 SDValue NotZero = N->getOperand(i);
128 if (isa<ConstantSDNode>(NotZero)) {
129 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
131 } else if (isa<ConstantFPSDNode>(NotZero)) {
132 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
133 bitcastToAPInt().isAllOnesValue())
138 // Okay, we have at least one ~0 value, check to see if the rest match or are
140 for (++i; i != e; ++i)
141 if (N->getOperand(i) != NotZero &&
142 N->getOperand(i).getOpcode() != ISD::UNDEF)
148 /// isBuildVectorAllZeros - Return true if the specified node is a
149 /// BUILD_VECTOR where all of the elements are 0 or undef.
150 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
151 // Look through a bit convert.
152 if (N->getOpcode() == ISD::BIT_CONVERT)
153 N = N->getOperand(0).getNode();
155 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
157 unsigned i = 0, e = N->getNumOperands();
159 // Skip over all of the undef values.
160 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
163 // Do not accept an all-undef vector.
164 if (i == e) return false;
166 // Do not accept build_vectors that aren't all constants or which have non-0
168 SDValue Zero = N->getOperand(i);
169 if (isa<ConstantSDNode>(Zero)) {
170 if (!cast<ConstantSDNode>(Zero)->isNullValue())
172 } else if (isa<ConstantFPSDNode>(Zero)) {
173 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
178 // Okay, we have at least one 0 value, check to see if the rest match or are
180 for (++i; i != e; ++i)
181 if (N->getOperand(i) != Zero &&
182 N->getOperand(i).getOpcode() != ISD::UNDEF)
187 /// isScalarToVector - Return true if the specified node is a
188 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
189 /// element is not an undef.
190 bool ISD::isScalarToVector(const SDNode *N) {
191 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
194 if (N->getOpcode() != ISD::BUILD_VECTOR)
196 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
198 unsigned NumElems = N->getNumOperands();
199 for (unsigned i = 1; i < NumElems; ++i) {
200 SDValue V = N->getOperand(i);
201 if (V.getOpcode() != ISD::UNDEF)
207 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
208 /// when given the operation for (X op Y).
209 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
210 // To perform this operation, we just need to swap the L and G bits of the
212 unsigned OldL = (Operation >> 2) & 1;
213 unsigned OldG = (Operation >> 1) & 1;
214 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
215 (OldL << 1) | // New G bit
216 (OldG << 2)); // New L bit.
219 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
220 /// 'op' is a valid SetCC operation.
221 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
222 unsigned Operation = Op;
224 Operation ^= 7; // Flip L, G, E bits, but not U.
226 Operation ^= 15; // Flip all of the condition bits.
228 if (Operation > ISD::SETTRUE2)
229 Operation &= ~8; // Don't let N and U bits get set.
231 return ISD::CondCode(Operation);
235 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
236 /// signed operation and 2 if the result is an unsigned comparison. Return zero
237 /// if the operation does not depend on the sign of the input (setne and seteq).
238 static int isSignedOp(ISD::CondCode Opcode) {
240 default: llvm_unreachable("Illegal integer setcc operation!");
242 case ISD::SETNE: return 0;
246 case ISD::SETGE: return 1;
250 case ISD::SETUGE: return 2;
254 /// getSetCCOrOperation - Return the result of a logical OR between different
255 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
256 /// returns SETCC_INVALID if it is not possible to represent the resultant
258 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
260 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
261 // Cannot fold a signed integer setcc with an unsigned integer setcc.
262 return ISD::SETCC_INVALID;
264 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
266 // If the N and U bits get set then the resultant comparison DOES suddenly
267 // care about orderedness, and is true when ordered.
268 if (Op > ISD::SETTRUE2)
269 Op &= ~16; // Clear the U bit if the N bit is set.
271 // Canonicalize illegal integer setcc's.
272 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
275 return ISD::CondCode(Op);
278 /// getSetCCAndOperation - Return the result of a logical AND between different
279 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
280 /// function returns zero if it is not possible to represent the resultant
282 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
284 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
285 // Cannot fold a signed setcc with an unsigned setcc.
286 return ISD::SETCC_INVALID;
288 // Combine all of the condition bits.
289 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
291 // Canonicalize illegal integer setcc's.
295 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
296 case ISD::SETOEQ: // SETEQ & SETU[LG]E
297 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
298 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
299 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
306 const TargetMachine &SelectionDAG::getTarget() const {
307 return MF->getTarget();
310 //===----------------------------------------------------------------------===//
311 // SDNode Profile Support
312 //===----------------------------------------------------------------------===//
314 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
316 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
320 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
321 /// solely with their pointer.
322 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
323 ID.AddPointer(VTList.VTs);
326 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
328 static void AddNodeIDOperands(FoldingSetNodeID &ID,
329 const SDValue *Ops, unsigned NumOps) {
330 for (; NumOps; --NumOps, ++Ops) {
331 ID.AddPointer(Ops->getNode());
332 ID.AddInteger(Ops->getResNo());
336 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
338 static void AddNodeIDOperands(FoldingSetNodeID &ID,
339 const SDUse *Ops, unsigned NumOps) {
340 for (; NumOps; --NumOps, ++Ops) {
341 ID.AddPointer(Ops->getNode());
342 ID.AddInteger(Ops->getResNo());
346 static void AddNodeIDNode(FoldingSetNodeID &ID,
347 unsigned short OpC, SDVTList VTList,
348 const SDValue *OpList, unsigned N) {
349 AddNodeIDOpcode(ID, OpC);
350 AddNodeIDValueTypes(ID, VTList);
351 AddNodeIDOperands(ID, OpList, N);
354 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
356 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
357 switch (N->getOpcode()) {
358 case ISD::TargetExternalSymbol:
359 case ISD::ExternalSymbol:
360 llvm_unreachable("Should only be used on nodes with operands");
361 default: break; // Normal nodes don't need extra info.
362 case ISD::TargetConstant:
364 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
366 case ISD::TargetConstantFP:
367 case ISD::ConstantFP: {
368 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
371 case ISD::TargetGlobalAddress:
372 case ISD::GlobalAddress:
373 case ISD::TargetGlobalTLSAddress:
374 case ISD::GlobalTLSAddress: {
375 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
376 ID.AddPointer(GA->getGlobal());
377 ID.AddInteger(GA->getOffset());
378 ID.AddInteger(GA->getTargetFlags());
381 case ISD::BasicBlock:
382 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
385 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
389 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
391 case ISD::FrameIndex:
392 case ISD::TargetFrameIndex:
393 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
396 case ISD::TargetJumpTable:
397 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
398 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
400 case ISD::ConstantPool:
401 case ISD::TargetConstantPool: {
402 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
403 ID.AddInteger(CP->getAlignment());
404 ID.AddInteger(CP->getOffset());
405 if (CP->isMachineConstantPoolEntry())
406 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
408 ID.AddPointer(CP->getConstVal());
409 ID.AddInteger(CP->getTargetFlags());
413 const LoadSDNode *LD = cast<LoadSDNode>(N);
414 ID.AddInteger(LD->getMemoryVT().getRawBits());
415 ID.AddInteger(LD->getRawSubclassData());
419 const StoreSDNode *ST = cast<StoreSDNode>(N);
420 ID.AddInteger(ST->getMemoryVT().getRawBits());
421 ID.AddInteger(ST->getRawSubclassData());
424 case ISD::ATOMIC_CMP_SWAP:
425 case ISD::ATOMIC_SWAP:
426 case ISD::ATOMIC_LOAD_ADD:
427 case ISD::ATOMIC_LOAD_SUB:
428 case ISD::ATOMIC_LOAD_AND:
429 case ISD::ATOMIC_LOAD_OR:
430 case ISD::ATOMIC_LOAD_XOR:
431 case ISD::ATOMIC_LOAD_NAND:
432 case ISD::ATOMIC_LOAD_MIN:
433 case ISD::ATOMIC_LOAD_MAX:
434 case ISD::ATOMIC_LOAD_UMIN:
435 case ISD::ATOMIC_LOAD_UMAX: {
436 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
437 ID.AddInteger(AT->getMemoryVT().getRawBits());
438 ID.AddInteger(AT->getRawSubclassData());
441 case ISD::VECTOR_SHUFFLE: {
442 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
443 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
445 ID.AddInteger(SVN->getMaskElt(i));
448 case ISD::TargetBlockAddress:
449 case ISD::BlockAddress: {
450 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress());
451 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags());
454 } // end switch (N->getOpcode())
457 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
459 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
460 AddNodeIDOpcode(ID, N->getOpcode());
461 // Add the return value info.
462 AddNodeIDValueTypes(ID, N->getVTList());
463 // Add the operand info.
464 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
466 // Handle SDNode leafs with special info.
467 AddNodeIDCustom(ID, N);
470 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
471 /// the CSE map that carries volatility, indexing mode, and
472 /// extension/truncation information.
474 static inline unsigned
475 encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile) {
476 assert((ConvType & 3) == ConvType &&
477 "ConvType may not require more than 2 bits!");
478 assert((AM & 7) == AM &&
479 "AM may not require more than 3 bits!");
485 //===----------------------------------------------------------------------===//
486 // SelectionDAG Class
487 //===----------------------------------------------------------------------===//
489 /// doNotCSE - Return true if CSE should not be performed for this node.
490 static bool doNotCSE(SDNode *N) {
491 if (N->getValueType(0) == MVT::Flag)
492 return true; // Never CSE anything that produces a flag.
494 switch (N->getOpcode()) {
496 case ISD::HANDLENODE:
498 return true; // Never CSE these nodes.
501 // Check that remaining values produced are not flags.
502 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
503 if (N->getValueType(i) == MVT::Flag)
504 return true; // Never CSE anything that produces a flag.
509 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
511 void SelectionDAG::RemoveDeadNodes() {
512 // Create a dummy node (which is not added to allnodes), that adds a reference
513 // to the root node, preventing it from being deleted.
514 HandleSDNode Dummy(getRoot());
516 SmallVector<SDNode*, 128> DeadNodes;
518 // Add all obviously-dead nodes to the DeadNodes worklist.
519 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
521 DeadNodes.push_back(I);
523 RemoveDeadNodes(DeadNodes);
525 // If the root changed (e.g. it was a dead load, update the root).
526 setRoot(Dummy.getValue());
529 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
530 /// given list, and any nodes that become unreachable as a result.
531 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
532 DAGUpdateListener *UpdateListener) {
534 // Process the worklist, deleting the nodes and adding their uses to the
536 while (!DeadNodes.empty()) {
537 SDNode *N = DeadNodes.pop_back_val();
540 UpdateListener->NodeDeleted(N, 0);
542 // Take the node out of the appropriate CSE map.
543 RemoveNodeFromCSEMaps(N);
545 // Next, brutally remove the operand list. This is safe to do, as there are
546 // no cycles in the graph.
547 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
549 SDNode *Operand = Use.getNode();
552 // Now that we removed this operand, see if there are no uses of it left.
553 if (Operand->use_empty())
554 DeadNodes.push_back(Operand);
561 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
562 SmallVector<SDNode*, 16> DeadNodes(1, N);
563 RemoveDeadNodes(DeadNodes, UpdateListener);
566 void SelectionDAG::DeleteNode(SDNode *N) {
567 // First take this out of the appropriate CSE map.
568 RemoveNodeFromCSEMaps(N);
570 // Finally, remove uses due to operands of this node, remove from the
571 // AllNodes list, and delete the node.
572 DeleteNodeNotInCSEMaps(N);
575 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
576 assert(N != AllNodes.begin() && "Cannot delete the entry node!");
577 assert(N->use_empty() && "Cannot delete a node that is not dead!");
579 // Drop all of the operands and decrement used node's use counts.
585 void SelectionDAG::DeallocateNode(SDNode *N) {
586 if (N->OperandsNeedDelete)
587 delete[] N->OperandList;
589 // Set the opcode to DELETED_NODE to help catch bugs when node
590 // memory is reallocated.
591 N->NodeType = ISD::DELETED_NODE;
593 NodeAllocator.Deallocate(AllNodes.remove(N));
595 // Remove the ordering of this node.
599 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
600 /// correspond to it. This is useful when we're about to delete or repurpose
601 /// the node. We don't want future request for structurally identical nodes
602 /// to return N anymore.
603 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
605 switch (N->getOpcode()) {
606 case ISD::EntryToken:
607 llvm_unreachable("EntryToken should not be in CSEMaps!");
609 case ISD::HANDLENODE: return false; // noop.
611 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
612 "Cond code doesn't exist!");
613 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
614 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
616 case ISD::ExternalSymbol:
617 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
619 case ISD::TargetExternalSymbol: {
620 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
621 Erased = TargetExternalSymbols.erase(
622 std::pair<std::string,unsigned char>(ESN->getSymbol(),
623 ESN->getTargetFlags()));
626 case ISD::VALUETYPE: {
627 EVT VT = cast<VTSDNode>(N)->getVT();
628 if (VT.isExtended()) {
629 Erased = ExtendedValueTypeNodes.erase(VT);
631 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0;
632 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0;
637 // Remove it from the CSE Map.
638 Erased = CSEMap.RemoveNode(N);
642 // Verify that the node was actually in one of the CSE maps, unless it has a
643 // flag result (which cannot be CSE'd) or is one of the special cases that are
644 // not subject to CSE.
645 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
646 !N->isMachineOpcode() && !doNotCSE(N)) {
649 llvm_unreachable("Node is not in map!");
655 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
656 /// maps and modified in place. Add it back to the CSE maps, unless an identical
657 /// node already exists, in which case transfer all its users to the existing
658 /// node. This transfer can potentially trigger recursive merging.
661 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N,
662 DAGUpdateListener *UpdateListener) {
663 // For node types that aren't CSE'd, just act as if no identical node
666 SDNode *Existing = CSEMap.GetOrInsertNode(N);
668 // If there was already an existing matching node, use ReplaceAllUsesWith
669 // to replace the dead one with the existing one. This can cause
670 // recursive merging of other unrelated nodes down the line.
671 ReplaceAllUsesWith(N, Existing, UpdateListener);
673 // N is now dead. Inform the listener if it exists and delete it.
675 UpdateListener->NodeDeleted(N, Existing);
676 DeleteNodeNotInCSEMaps(N);
681 // If the node doesn't already exist, we updated it. Inform a listener if
684 UpdateListener->NodeUpdated(N);
687 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
688 /// were replaced with those specified. If this node is never memoized,
689 /// return null, otherwise return a pointer to the slot it would take. If a
690 /// node already exists with these operands, the slot will be non-null.
691 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
696 SDValue Ops[] = { Op };
698 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
699 AddNodeIDCustom(ID, N);
700 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
704 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
705 /// were replaced with those specified. If this node is never memoized,
706 /// return null, otherwise return a pointer to the slot it would take. If a
707 /// node already exists with these operands, the slot will be non-null.
708 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
709 SDValue Op1, SDValue Op2,
714 SDValue Ops[] = { Op1, Op2 };
716 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
717 AddNodeIDCustom(ID, N);
718 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
723 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
724 /// were replaced with those specified. If this node is never memoized,
725 /// return null, otherwise return a pointer to the slot it would take. If a
726 /// node already exists with these operands, the slot will be non-null.
727 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
728 const SDValue *Ops,unsigned NumOps,
734 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
735 AddNodeIDCustom(ID, N);
736 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
740 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
741 void SelectionDAG::VerifyNode(SDNode *N) {
742 switch (N->getOpcode()) {
745 case ISD::BUILD_PAIR: {
746 EVT VT = N->getValueType(0);
747 assert(N->getNumValues() == 1 && "Too many results!");
748 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
749 "Wrong return type!");
750 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
751 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
752 "Mismatched operand types!");
753 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
754 "Wrong operand type!");
755 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
756 "Wrong return type size");
759 case ISD::BUILD_VECTOR: {
760 assert(N->getNumValues() == 1 && "Too many results!");
761 assert(N->getValueType(0).isVector() && "Wrong return type!");
762 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
763 "Wrong number of operands!");
764 EVT EltVT = N->getValueType(0).getVectorElementType();
765 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
766 assert((I->getValueType() == EltVT ||
767 (EltVT.isInteger() && I->getValueType().isInteger() &&
768 EltVT.bitsLE(I->getValueType()))) &&
769 "Wrong operand type!");
775 /// getEVTAlignment - Compute the default alignment value for the
778 unsigned SelectionDAG::getEVTAlignment(EVT VT) const {
779 const Type *Ty = VT == MVT::iPTR ?
780 PointerType::get(Type::getInt8Ty(*getContext()), 0) :
781 VT.getTypeForEVT(*getContext());
783 return TLI.getTargetData()->getABITypeAlignment(Ty);
786 // EntryNode could meaningfully have debug info if we can find it...
787 SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli)
788 : TLI(tli), FLI(fli), DW(0),
789 EntryNode(ISD::EntryToken, DebugLoc::getUnknownLoc(),
790 getVTList(MVT::Other)),
791 Root(getEntryNode()), Ordering(0) {
792 AllNodes.push_back(&EntryNode);
793 Ordering = new SDNodeOrdering();
796 void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi,
801 Context = &mf.getFunction()->getContext();
804 SelectionDAG::~SelectionDAG() {
809 void SelectionDAG::allnodes_clear() {
810 assert(&*AllNodes.begin() == &EntryNode);
811 AllNodes.remove(AllNodes.begin());
812 while (!AllNodes.empty())
813 DeallocateNode(AllNodes.begin());
816 void SelectionDAG::clear() {
818 OperandAllocator.Reset();
821 ExtendedValueTypeNodes.clear();
822 ExternalSymbols.clear();
823 TargetExternalSymbols.clear();
824 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
825 static_cast<CondCodeSDNode*>(0));
826 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
827 static_cast<SDNode*>(0));
829 EntryNode.UseList = 0;
830 AllNodes.push_back(&EntryNode);
831 Root = getEntryNode();
833 Ordering = new SDNodeOrdering();
836 SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
837 return VT.bitsGT(Op.getValueType()) ?
838 getNode(ISD::SIGN_EXTEND, DL, VT, Op) :
839 getNode(ISD::TRUNCATE, DL, VT, Op);
842 SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
843 return VT.bitsGT(Op.getValueType()) ?
844 getNode(ISD::ZERO_EXTEND, DL, VT, Op) :
845 getNode(ISD::TRUNCATE, DL, VT, Op);
848 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) {
849 assert(!VT.isVector() &&
850 "getZeroExtendInReg should use the vector element type instead of "
852 if (Op.getValueType() == VT) return Op;
853 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
854 APInt Imm = APInt::getLowBitsSet(BitWidth,
856 return getNode(ISD::AND, DL, Op.getValueType(), Op,
857 getConstant(Imm, Op.getValueType()));
860 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
862 SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) {
863 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
865 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
866 return getNode(ISD::XOR, DL, VT, Val, NegOne);
869 SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) {
870 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
871 assert((EltVT.getSizeInBits() >= 64 ||
872 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
873 "getConstant with a uint64_t value that doesn't fit in the type!");
874 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
877 SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) {
878 return getConstant(*ConstantInt::get(*Context, Val), VT, isT);
881 SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
882 assert(VT.isInteger() && "Cannot create FP integer constant!");
884 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
885 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
886 "APInt size does not match type size!");
888 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
890 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
894 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
896 return SDValue(N, 0);
899 N = NodeAllocator.Allocate<ConstantSDNode>();
900 new (N) ConstantSDNode(isT, &Val, EltVT);
901 CSEMap.InsertNode(N, IP);
902 AllNodes.push_back(N);
905 SDValue Result(N, 0);
907 SmallVector<SDValue, 8> Ops;
908 Ops.assign(VT.getVectorNumElements(), Result);
909 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
910 VT, &Ops[0], Ops.size());
915 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
916 return getConstant(Val, TLI.getPointerTy(), isTarget);
920 SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) {
921 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget);
924 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){
925 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
928 VT.isVector() ? VT.getVectorElementType() : VT;
930 // Do the map lookup using the actual bit pattern for the floating point
931 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
932 // we don't have issues with SNANs.
933 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
935 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
939 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
941 return SDValue(N, 0);
944 N = NodeAllocator.Allocate<ConstantFPSDNode>();
945 new (N) ConstantFPSDNode(isTarget, &V, EltVT);
946 CSEMap.InsertNode(N, IP);
947 AllNodes.push_back(N);
950 SDValue Result(N, 0);
952 SmallVector<SDValue, 8> Ops;
953 Ops.assign(VT.getVectorNumElements(), Result);
954 // FIXME DebugLoc info might be appropriate here
955 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
956 VT, &Ops[0], Ops.size());
961 SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) {
963 VT.isVector() ? VT.getVectorElementType() : VT;
965 return getConstantFP(APFloat((float)Val), VT, isTarget);
967 return getConstantFP(APFloat(Val), VT, isTarget);
970 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
971 EVT VT, int64_t Offset,
973 unsigned char TargetFlags) {
974 assert((TargetFlags == 0 || isTargetGA) &&
975 "Cannot set target flags on target-independent globals");
977 // Truncate (with sign-extension) the offset value to the pointer size.
978 EVT PTy = TLI.getPointerTy();
979 unsigned BitWidth = PTy.getSizeInBits();
981 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
983 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
985 // If GV is an alias then use the aliasee for determining thread-localness.
986 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
987 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
991 if (GVar && GVar->isThreadLocal())
992 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
994 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
997 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
999 ID.AddInteger(Offset);
1000 ID.AddInteger(TargetFlags);
1002 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1003 return SDValue(E, 0);
1005 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>();
1006 new (N) GlobalAddressSDNode(Opc, GV, VT, Offset, TargetFlags);
1007 CSEMap.InsertNode(N, IP);
1008 AllNodes.push_back(N);
1009 return SDValue(N, 0);
1012 SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1013 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1014 FoldingSetNodeID ID;
1015 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1018 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1019 return SDValue(E, 0);
1021 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>();
1022 new (N) FrameIndexSDNode(FI, VT, isTarget);
1023 CSEMap.InsertNode(N, IP);
1024 AllNodes.push_back(N);
1025 return SDValue(N, 0);
1028 SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1029 unsigned char TargetFlags) {
1030 assert((TargetFlags == 0 || isTarget) &&
1031 "Cannot set target flags on target-independent jump tables");
1032 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1033 FoldingSetNodeID ID;
1034 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1036 ID.AddInteger(TargetFlags);
1038 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1039 return SDValue(E, 0);
1041 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>();
1042 new (N) JumpTableSDNode(JTI, VT, isTarget, TargetFlags);
1043 CSEMap.InsertNode(N, IP);
1044 AllNodes.push_back(N);
1045 return SDValue(N, 0);
1048 SDValue SelectionDAG::getConstantPool(Constant *C, EVT VT,
1049 unsigned Alignment, int Offset,
1051 unsigned char TargetFlags) {
1052 assert((TargetFlags == 0 || isTarget) &&
1053 "Cannot set target flags on target-independent globals");
1055 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1056 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1057 FoldingSetNodeID ID;
1058 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1059 ID.AddInteger(Alignment);
1060 ID.AddInteger(Offset);
1062 ID.AddInteger(TargetFlags);
1064 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1065 return SDValue(E, 0);
1067 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1068 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1069 CSEMap.InsertNode(N, IP);
1070 AllNodes.push_back(N);
1071 return SDValue(N, 0);
1075 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1076 unsigned Alignment, int Offset,
1078 unsigned char TargetFlags) {
1079 assert((TargetFlags == 0 || isTarget) &&
1080 "Cannot set target flags on target-independent globals");
1082 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1083 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1084 FoldingSetNodeID ID;
1085 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1086 ID.AddInteger(Alignment);
1087 ID.AddInteger(Offset);
1088 C->AddSelectionDAGCSEId(ID);
1089 ID.AddInteger(TargetFlags);
1091 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1092 return SDValue(E, 0);
1094 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1095 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1096 CSEMap.InsertNode(N, IP);
1097 AllNodes.push_back(N);
1098 return SDValue(N, 0);
1101 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1102 FoldingSetNodeID ID;
1103 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1106 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1107 return SDValue(E, 0);
1109 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>();
1110 new (N) BasicBlockSDNode(MBB);
1111 CSEMap.InsertNode(N, IP);
1112 AllNodes.push_back(N);
1113 return SDValue(N, 0);
1116 SDValue SelectionDAG::getValueType(EVT VT) {
1117 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
1118 ValueTypeNodes.size())
1119 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1);
1121 SDNode *&N = VT.isExtended() ?
1122 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
1124 if (N) return SDValue(N, 0);
1125 N = NodeAllocator.Allocate<VTSDNode>();
1126 new (N) VTSDNode(VT);
1127 AllNodes.push_back(N);
1128 return SDValue(N, 0);
1131 SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
1132 SDNode *&N = ExternalSymbols[Sym];
1133 if (N) return SDValue(N, 0);
1134 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1135 new (N) ExternalSymbolSDNode(false, Sym, 0, VT);
1136 AllNodes.push_back(N);
1137 return SDValue(N, 0);
1140 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
1141 unsigned char TargetFlags) {
1143 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym,
1145 if (N) return SDValue(N, 0);
1146 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1147 new (N) ExternalSymbolSDNode(true, Sym, TargetFlags, VT);
1148 AllNodes.push_back(N);
1149 return SDValue(N, 0);
1152 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1153 if ((unsigned)Cond >= CondCodeNodes.size())
1154 CondCodeNodes.resize(Cond+1);
1156 if (CondCodeNodes[Cond] == 0) {
1157 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>();
1158 new (N) CondCodeSDNode(Cond);
1159 CondCodeNodes[Cond] = N;
1160 AllNodes.push_back(N);
1163 return SDValue(CondCodeNodes[Cond], 0);
1166 // commuteShuffle - swaps the values of N1 and N2, and swaps all indices in
1167 // the shuffle mask M that point at N1 to point at N2, and indices that point
1168 // N2 to point at N1.
1169 static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) {
1171 int NElts = M.size();
1172 for (int i = 0; i != NElts; ++i) {
1180 SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1,
1181 SDValue N2, const int *Mask) {
1182 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE");
1183 assert(VT.isVector() && N1.getValueType().isVector() &&
1184 "Vector Shuffle VTs must be a vectors");
1185 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType()
1186 && "Vector Shuffle VTs must have same element type");
1188 // Canonicalize shuffle undef, undef -> undef
1189 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF)
1190 return getUNDEF(VT);
1192 // Validate that all indices in Mask are within the range of the elements
1193 // input to the shuffle.
1194 unsigned NElts = VT.getVectorNumElements();
1195 SmallVector<int, 8> MaskVec;
1196 for (unsigned i = 0; i != NElts; ++i) {
1197 assert(Mask[i] < (int)(NElts * 2) && "Index out of range");
1198 MaskVec.push_back(Mask[i]);
1201 // Canonicalize shuffle v, v -> v, undef
1204 for (unsigned i = 0; i != NElts; ++i)
1205 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts;
1208 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1209 if (N1.getOpcode() == ISD::UNDEF)
1210 commuteShuffle(N1, N2, MaskVec);
1212 // Canonicalize all index into lhs, -> shuffle lhs, undef
1213 // Canonicalize all index into rhs, -> shuffle rhs, undef
1214 bool AllLHS = true, AllRHS = true;
1215 bool N2Undef = N2.getOpcode() == ISD::UNDEF;
1216 for (unsigned i = 0; i != NElts; ++i) {
1217 if (MaskVec[i] >= (int)NElts) {
1222 } else if (MaskVec[i] >= 0) {
1226 if (AllLHS && AllRHS)
1227 return getUNDEF(VT);
1228 if (AllLHS && !N2Undef)
1232 commuteShuffle(N1, N2, MaskVec);
1235 // If Identity shuffle, or all shuffle in to undef, return that node.
1236 bool AllUndef = true;
1237 bool Identity = true;
1238 for (unsigned i = 0; i != NElts; ++i) {
1239 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false;
1240 if (MaskVec[i] >= 0) AllUndef = false;
1242 if (Identity && NElts == N1.getValueType().getVectorNumElements())
1245 return getUNDEF(VT);
1247 FoldingSetNodeID ID;
1248 SDValue Ops[2] = { N1, N2 };
1249 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2);
1250 for (unsigned i = 0; i != NElts; ++i)
1251 ID.AddInteger(MaskVec[i]);
1254 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1255 return SDValue(E, 0);
1257 // Allocate the mask array for the node out of the BumpPtrAllocator, since
1258 // SDNode doesn't have access to it. This memory will be "leaked" when
1259 // the node is deallocated, but recovered when the NodeAllocator is released.
1260 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
1261 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int));
1263 ShuffleVectorSDNode *N = NodeAllocator.Allocate<ShuffleVectorSDNode>();
1264 new (N) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc);
1265 CSEMap.InsertNode(N, IP);
1266 AllNodes.push_back(N);
1267 return SDValue(N, 0);
1270 SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl,
1271 SDValue Val, SDValue DTy,
1272 SDValue STy, SDValue Rnd, SDValue Sat,
1273 ISD::CvtCode Code) {
1274 // If the src and dest types are the same and the conversion is between
1275 // integer types of the same sign or two floats, no conversion is necessary.
1277 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF))
1280 FoldingSetNodeID ID;
1281 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat };
1282 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5);
1284 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1285 return SDValue(E, 0);
1287 CvtRndSatSDNode *N = NodeAllocator.Allocate<CvtRndSatSDNode>();
1288 new (N) CvtRndSatSDNode(VT, dl, Ops, 5, Code);
1289 CSEMap.InsertNode(N, IP);
1290 AllNodes.push_back(N);
1291 return SDValue(N, 0);
1294 SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) {
1295 FoldingSetNodeID ID;
1296 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1297 ID.AddInteger(RegNo);
1299 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1300 return SDValue(E, 0);
1302 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>();
1303 new (N) RegisterSDNode(RegNo, VT);
1304 CSEMap.InsertNode(N, IP);
1305 AllNodes.push_back(N);
1306 return SDValue(N, 0);
1309 SDValue SelectionDAG::getLabel(unsigned Opcode, DebugLoc dl,
1312 FoldingSetNodeID ID;
1313 SDValue Ops[] = { Root };
1314 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), &Ops[0], 1);
1315 ID.AddInteger(LabelID);
1317 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1318 return SDValue(E, 0);
1320 SDNode *N = NodeAllocator.Allocate<LabelSDNode>();
1321 new (N) LabelSDNode(Opcode, dl, Root, LabelID);
1322 CSEMap.InsertNode(N, IP);
1323 AllNodes.push_back(N);
1324 return SDValue(N, 0);
1327 SDValue SelectionDAG::getBlockAddress(BlockAddress *BA, EVT VT,
1329 unsigned char TargetFlags) {
1330 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress;
1332 FoldingSetNodeID ID;
1333 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1335 ID.AddInteger(TargetFlags);
1337 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1338 return SDValue(E, 0);
1340 SDNode *N = NodeAllocator.Allocate<BlockAddressSDNode>();
1341 new (N) BlockAddressSDNode(Opc, VT, BA, TargetFlags);
1342 CSEMap.InsertNode(N, IP);
1343 AllNodes.push_back(N);
1344 return SDValue(N, 0);
1347 SDValue SelectionDAG::getSrcValue(const Value *V) {
1348 assert((!V || V->getType()->isPointerTy()) &&
1349 "SrcValue is not a pointer?");
1351 FoldingSetNodeID ID;
1352 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1356 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1357 return SDValue(E, 0);
1359 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>();
1360 new (N) SrcValueSDNode(V);
1361 CSEMap.InsertNode(N, IP);
1362 AllNodes.push_back(N);
1363 return SDValue(N, 0);
1366 /// getShiftAmountOperand - Return the specified value casted to
1367 /// the target's desired shift amount type.
1368 SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) {
1369 EVT OpTy = Op.getValueType();
1370 MVT ShTy = TLI.getShiftAmountTy();
1371 if (OpTy == ShTy || OpTy.isVector()) return Op;
1373 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
1374 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op);
1377 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1378 /// specified value type.
1379 SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
1380 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1381 unsigned ByteSize = VT.getStoreSize();
1382 const Type *Ty = VT.getTypeForEVT(*getContext());
1383 unsigned StackAlign =
1384 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1386 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
1387 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1390 /// CreateStackTemporary - Create a stack temporary suitable for holding
1391 /// either of the specified value types.
1392 SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
1393 unsigned Bytes = std::max(VT1.getStoreSizeInBits(),
1394 VT2.getStoreSizeInBits())/8;
1395 const Type *Ty1 = VT1.getTypeForEVT(*getContext());
1396 const Type *Ty2 = VT2.getTypeForEVT(*getContext());
1397 const TargetData *TD = TLI.getTargetData();
1398 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1),
1399 TD->getPrefTypeAlignment(Ty2));
1401 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1402 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false);
1403 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1406 SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1,
1407 SDValue N2, ISD::CondCode Cond, DebugLoc dl) {
1408 // These setcc operations always fold.
1412 case ISD::SETFALSE2: return getConstant(0, VT);
1414 case ISD::SETTRUE2: return getConstant(1, VT);
1426 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1430 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1431 const APInt &C2 = N2C->getAPIntValue();
1432 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1433 const APInt &C1 = N1C->getAPIntValue();
1436 default: llvm_unreachable("Unknown integer setcc!");
1437 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1438 case ISD::SETNE: return getConstant(C1 != C2, VT);
1439 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1440 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1441 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1442 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1443 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1444 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1445 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1446 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1450 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1451 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1452 // No compile time operations on this type yet.
1453 if (N1C->getValueType(0) == MVT::ppcf128)
1456 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1459 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1460 return getUNDEF(VT);
1462 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1463 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1464 return getUNDEF(VT);
1466 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1467 R==APFloat::cmpLessThan, VT);
1468 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1469 return getUNDEF(VT);
1471 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1472 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1473 return getUNDEF(VT);
1475 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1476 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1477 return getUNDEF(VT);
1479 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1480 R==APFloat::cmpEqual, VT);
1481 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1482 return getUNDEF(VT);
1484 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1485 R==APFloat::cmpEqual, VT);
1486 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1487 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1488 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1489 R==APFloat::cmpEqual, VT);
1490 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1491 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1492 R==APFloat::cmpLessThan, VT);
1493 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1494 R==APFloat::cmpUnordered, VT);
1495 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1496 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1499 // Ensure that the constant occurs on the RHS.
1500 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1504 // Could not fold it.
1508 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1509 /// use this predicate to simplify operations downstream.
1510 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1511 // This predicate is not safe for vector operations.
1512 if (Op.getValueType().isVector())
1515 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
1516 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1519 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1520 /// this predicate to simplify operations downstream. Mask is known to be zero
1521 /// for bits that V cannot have.
1522 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1523 unsigned Depth) const {
1524 APInt KnownZero, KnownOne;
1525 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1526 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1527 return (KnownZero & Mask) == Mask;
1530 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1531 /// known to be either zero or one and return them in the KnownZero/KnownOne
1532 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1534 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1535 APInt &KnownZero, APInt &KnownOne,
1536 unsigned Depth) const {
1537 unsigned BitWidth = Mask.getBitWidth();
1538 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() &&
1539 "Mask size mismatches value type size!");
1541 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1542 if (Depth == 6 || Mask == 0)
1543 return; // Limit search depth.
1545 APInt KnownZero2, KnownOne2;
1547 switch (Op.getOpcode()) {
1549 // We know all of the bits for a constant!
1550 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1551 KnownZero = ~KnownOne & Mask;
1554 // If either the LHS or the RHS are Zero, the result is zero.
1555 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1556 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1557 KnownZero2, KnownOne2, Depth+1);
1558 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1559 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1561 // Output known-1 bits are only known if set in both the LHS & RHS.
1562 KnownOne &= KnownOne2;
1563 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1564 KnownZero |= KnownZero2;
1567 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1568 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1569 KnownZero2, KnownOne2, Depth+1);
1570 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1571 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1573 // Output known-0 bits are only known if clear in both the LHS & RHS.
1574 KnownZero &= KnownZero2;
1575 // Output known-1 are known to be set if set in either the LHS | RHS.
1576 KnownOne |= KnownOne2;
1579 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1580 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1581 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1582 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1584 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1585 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1586 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1587 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1588 KnownZero = KnownZeroOut;
1592 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1593 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1594 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1595 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1596 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1598 // If low bits are zero in either operand, output low known-0 bits.
1599 // Also compute a conserative estimate for high known-0 bits.
1600 // More trickiness is possible, but this is sufficient for the
1601 // interesting case of alignment computation.
1603 unsigned TrailZ = KnownZero.countTrailingOnes() +
1604 KnownZero2.countTrailingOnes();
1605 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1606 KnownZero2.countLeadingOnes(),
1607 BitWidth) - BitWidth;
1609 TrailZ = std::min(TrailZ, BitWidth);
1610 LeadZ = std::min(LeadZ, BitWidth);
1611 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1612 APInt::getHighBitsSet(BitWidth, LeadZ);
1617 // For the purposes of computing leading zeros we can conservatively
1618 // treat a udiv as a logical right shift by the power of 2 known to
1619 // be less than the denominator.
1620 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1621 ComputeMaskedBits(Op.getOperand(0),
1622 AllOnes, KnownZero2, KnownOne2, Depth+1);
1623 unsigned LeadZ = KnownZero2.countLeadingOnes();
1627 ComputeMaskedBits(Op.getOperand(1),
1628 AllOnes, KnownZero2, KnownOne2, Depth+1);
1629 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1630 if (RHSUnknownLeadingOnes != BitWidth)
1631 LeadZ = std::min(BitWidth,
1632 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1634 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1638 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1639 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1640 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1641 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1643 // Only known if known in both the LHS and RHS.
1644 KnownOne &= KnownOne2;
1645 KnownZero &= KnownZero2;
1647 case ISD::SELECT_CC:
1648 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1649 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1650 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1651 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1653 // Only known if known in both the LHS and RHS.
1654 KnownOne &= KnownOne2;
1655 KnownZero &= KnownZero2;
1663 if (Op.getResNo() != 1)
1665 // The boolean result conforms to getBooleanContents. Fall through.
1667 // If we know the result of a setcc has the top bits zero, use this info.
1668 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent &&
1670 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1673 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1674 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1675 unsigned ShAmt = SA->getZExtValue();
1677 // If the shift count is an invalid immediate, don't do anything.
1678 if (ShAmt >= BitWidth)
1681 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt),
1682 KnownZero, KnownOne, Depth+1);
1683 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1684 KnownZero <<= ShAmt;
1686 // low bits known zero.
1687 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1691 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1692 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1693 unsigned ShAmt = SA->getZExtValue();
1695 // If the shift count is an invalid immediate, don't do anything.
1696 if (ShAmt >= BitWidth)
1699 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1700 KnownZero, KnownOne, Depth+1);
1701 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1702 KnownZero = KnownZero.lshr(ShAmt);
1703 KnownOne = KnownOne.lshr(ShAmt);
1705 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1706 KnownZero |= HighBits; // High bits known zero.
1710 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1711 unsigned ShAmt = SA->getZExtValue();
1713 // If the shift count is an invalid immediate, don't do anything.
1714 if (ShAmt >= BitWidth)
1717 APInt InDemandedMask = (Mask << ShAmt);
1718 // If any of the demanded bits are produced by the sign extension, we also
1719 // demand the input sign bit.
1720 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1721 if (HighBits.getBoolValue())
1722 InDemandedMask |= APInt::getSignBit(BitWidth);
1724 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1726 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1727 KnownZero = KnownZero.lshr(ShAmt);
1728 KnownOne = KnownOne.lshr(ShAmt);
1730 // Handle the sign bits.
1731 APInt SignBit = APInt::getSignBit(BitWidth);
1732 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1734 if (KnownZero.intersects(SignBit)) {
1735 KnownZero |= HighBits; // New bits are known zero.
1736 } else if (KnownOne.intersects(SignBit)) {
1737 KnownOne |= HighBits; // New bits are known one.
1741 case ISD::SIGN_EXTEND_INREG: {
1742 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1743 unsigned EBits = EVT.getScalarType().getSizeInBits();
1745 // Sign extension. Compute the demanded bits in the result that are not
1746 // present in the input.
1747 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1749 APInt InSignBit = APInt::getSignBit(EBits);
1750 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1752 // If the sign extended bits are demanded, we know that the sign
1754 InSignBit.zext(BitWidth);
1755 if (NewBits.getBoolValue())
1756 InputDemandedBits |= InSignBit;
1758 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1759 KnownZero, KnownOne, Depth+1);
1760 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1762 // If the sign bit of the input is known set or clear, then we know the
1763 // top bits of the result.
1764 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1765 KnownZero |= NewBits;
1766 KnownOne &= ~NewBits;
1767 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1768 KnownOne |= NewBits;
1769 KnownZero &= ~NewBits;
1770 } else { // Input sign bit unknown
1771 KnownZero &= ~NewBits;
1772 KnownOne &= ~NewBits;
1779 unsigned LowBits = Log2_32(BitWidth)+1;
1780 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1785 if (ISD::isZEXTLoad(Op.getNode())) {
1786 LoadSDNode *LD = cast<LoadSDNode>(Op);
1787 EVT VT = LD->getMemoryVT();
1788 unsigned MemBits = VT.getScalarType().getSizeInBits();
1789 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1793 case ISD::ZERO_EXTEND: {
1794 EVT InVT = Op.getOperand(0).getValueType();
1795 unsigned InBits = InVT.getScalarType().getSizeInBits();
1796 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1797 APInt InMask = Mask;
1798 InMask.trunc(InBits);
1799 KnownZero.trunc(InBits);
1800 KnownOne.trunc(InBits);
1801 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1802 KnownZero.zext(BitWidth);
1803 KnownOne.zext(BitWidth);
1804 KnownZero |= NewBits;
1807 case ISD::SIGN_EXTEND: {
1808 EVT InVT = Op.getOperand(0).getValueType();
1809 unsigned InBits = InVT.getScalarType().getSizeInBits();
1810 APInt InSignBit = APInt::getSignBit(InBits);
1811 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1812 APInt InMask = Mask;
1813 InMask.trunc(InBits);
1815 // If any of the sign extended bits are demanded, we know that the sign
1816 // bit is demanded. Temporarily set this bit in the mask for our callee.
1817 if (NewBits.getBoolValue())
1818 InMask |= InSignBit;
1820 KnownZero.trunc(InBits);
1821 KnownOne.trunc(InBits);
1822 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1824 // Note if the sign bit is known to be zero or one.
1825 bool SignBitKnownZero = KnownZero.isNegative();
1826 bool SignBitKnownOne = KnownOne.isNegative();
1827 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1828 "Sign bit can't be known to be both zero and one!");
1830 // If the sign bit wasn't actually demanded by our caller, we don't
1831 // want it set in the KnownZero and KnownOne result values. Reset the
1832 // mask and reapply it to the result values.
1834 InMask.trunc(InBits);
1835 KnownZero &= InMask;
1838 KnownZero.zext(BitWidth);
1839 KnownOne.zext(BitWidth);
1841 // If the sign bit is known zero or one, the top bits match.
1842 if (SignBitKnownZero)
1843 KnownZero |= NewBits;
1844 else if (SignBitKnownOne)
1845 KnownOne |= NewBits;
1848 case ISD::ANY_EXTEND: {
1849 EVT InVT = Op.getOperand(0).getValueType();
1850 unsigned InBits = InVT.getScalarType().getSizeInBits();
1851 APInt InMask = Mask;
1852 InMask.trunc(InBits);
1853 KnownZero.trunc(InBits);
1854 KnownOne.trunc(InBits);
1855 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1856 KnownZero.zext(BitWidth);
1857 KnownOne.zext(BitWidth);
1860 case ISD::TRUNCATE: {
1861 EVT InVT = Op.getOperand(0).getValueType();
1862 unsigned InBits = InVT.getScalarType().getSizeInBits();
1863 APInt InMask = Mask;
1864 InMask.zext(InBits);
1865 KnownZero.zext(InBits);
1866 KnownOne.zext(InBits);
1867 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1868 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1869 KnownZero.trunc(BitWidth);
1870 KnownOne.trunc(BitWidth);
1873 case ISD::AssertZext: {
1874 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1875 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1876 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1878 KnownZero |= (~InMask) & Mask;
1882 // All bits are zero except the low bit.
1883 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1887 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1888 // We know that the top bits of C-X are clear if X contains less bits
1889 // than C (i.e. no wrap-around can happen). For example, 20-X is
1890 // positive if we can prove that X is >= 0 and < 16.
1891 if (CLHS->getAPIntValue().isNonNegative()) {
1892 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1893 // NLZ can't be BitWidth with no sign bit
1894 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1895 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1898 // If all of the MaskV bits are known to be zero, then we know the
1899 // output top bits are zero, because we now know that the output is
1901 if ((KnownZero2 & MaskV) == MaskV) {
1902 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1903 // Top bits known zero.
1904 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1911 // Output known-0 bits are known if clear or set in both the low clear bits
1912 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1913 // low 3 bits clear.
1914 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
1915 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1916 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1917 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1919 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1920 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1921 KnownZeroOut = std::min(KnownZeroOut,
1922 KnownZero2.countTrailingOnes());
1924 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1928 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1929 const APInt &RA = Rem->getAPIntValue().abs();
1930 if (RA.isPowerOf2()) {
1931 APInt LowBits = RA - 1;
1932 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1933 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1935 // The low bits of the first operand are unchanged by the srem.
1936 KnownZero = KnownZero2 & LowBits;
1937 KnownOne = KnownOne2 & LowBits;
1939 // If the first operand is non-negative or has all low bits zero, then
1940 // the upper bits are all zero.
1941 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1942 KnownZero |= ~LowBits;
1944 // If the first operand is negative and not all low bits are zero, then
1945 // the upper bits are all one.
1946 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1947 KnownOne |= ~LowBits;
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,
1996 /// ComputeNumSignBits - Return the number of times the sign bit of the
1997 /// register is replicated into the other bits. We know that at least 1 bit
1998 /// is always equal to the sign bit (itself), but other cases can give us
1999 /// information. For example, immediately after an "SRA X, 2", we know that
2000 /// the top 3 bits are all equal to each other, so we return 3.
2001 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
2002 EVT VT = Op.getValueType();
2003 assert(VT.isInteger() && "Invalid VT!");
2004 unsigned VTBits = VT.getScalarType().getSizeInBits();
2006 unsigned FirstAnswer = 1;
2009 return 1; // Limit search depth.
2011 switch (Op.getOpcode()) {
2013 case ISD::AssertSext:
2014 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2015 return VTBits-Tmp+1;
2016 case ISD::AssertZext:
2017 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2020 case ISD::Constant: {
2021 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
2022 // If negative, return # leading ones.
2023 if (Val.isNegative())
2024 return Val.countLeadingOnes();
2026 // Return # leading zeros.
2027 return Val.countLeadingZeros();
2030 case ISD::SIGN_EXTEND:
2031 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
2032 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
2034 case ISD::SIGN_EXTEND_INREG:
2035 // Max of the input and what this extends.
2037 cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarType().getSizeInBits();
2040 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2041 return std::max(Tmp, Tmp2);
2044 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2045 // SRA X, C -> adds C sign bits.
2046 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2047 Tmp += C->getZExtValue();
2048 if (Tmp > VTBits) Tmp = VTBits;
2052 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2053 // shl destroys sign bits.
2054 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2055 if (C->getZExtValue() >= VTBits || // Bad shift.
2056 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2057 return Tmp - C->getZExtValue();
2062 case ISD::XOR: // NOT is handled here.
2063 // Logical binary ops preserve the number of sign bits at the worst.
2064 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2066 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2067 FirstAnswer = std::min(Tmp, Tmp2);
2068 // We computed what we know about the sign bits as our first
2069 // answer. Now proceed to the generic code that uses
2070 // ComputeMaskedBits, and pick whichever answer is better.
2075 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2076 if (Tmp == 1) return 1; // Early out.
2077 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
2078 return std::min(Tmp, Tmp2);
2086 if (Op.getResNo() != 1)
2088 // The boolean result conforms to getBooleanContents. Fall through.
2090 // If setcc returns 0/-1, all bits are sign bits.
2091 if (TLI.getBooleanContents() ==
2092 TargetLowering::ZeroOrNegativeOneBooleanContent)
2097 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2098 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
2100 // Handle rotate right by N like a rotate left by 32-N.
2101 if (Op.getOpcode() == ISD::ROTR)
2102 RotAmt = (VTBits-RotAmt) & (VTBits-1);
2104 // If we aren't rotating out all of the known-in sign bits, return the
2105 // number that are left. This handles rotl(sext(x), 1) for example.
2106 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2107 if (Tmp > RotAmt+1) return Tmp-RotAmt;
2111 // Add can have at most one carry bit. Thus we know that the output
2112 // is, at worst, one more bit than the inputs.
2113 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2114 if (Tmp == 1) return 1; // Early out.
2116 // Special case decrementing a value (ADD X, -1):
2117 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
2118 if (CRHS->isAllOnesValue()) {
2119 APInt KnownZero, KnownOne;
2120 APInt Mask = APInt::getAllOnesValue(VTBits);
2121 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2123 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2125 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2128 // If we are subtracting one from a positive number, there is no carry
2129 // out of the result.
2130 if (KnownZero.isNegative())
2134 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2135 if (Tmp2 == 1) return 1;
2136 return std::min(Tmp, Tmp2)-1;
2140 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2141 if (Tmp2 == 1) return 1;
2144 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
2145 if (CLHS->isNullValue()) {
2146 APInt KnownZero, KnownOne;
2147 APInt Mask = APInt::getAllOnesValue(VTBits);
2148 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2149 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2151 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2154 // If the input is known to be positive (the sign bit is known clear),
2155 // the output of the NEG has the same number of sign bits as the input.
2156 if (KnownZero.isNegative())
2159 // Otherwise, we treat this like a SUB.
2162 // Sub can have at most one carry bit. Thus we know that the output
2163 // is, at worst, one more bit than the inputs.
2164 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2165 if (Tmp == 1) return 1; // Early out.
2166 return std::min(Tmp, Tmp2)-1;
2169 // FIXME: it's tricky to do anything useful for this, but it is an important
2170 // case for targets like X86.
2174 // Handle LOADX separately here. EXTLOAD case will fallthrough.
2175 if (Op.getOpcode() == ISD::LOAD) {
2176 LoadSDNode *LD = cast<LoadSDNode>(Op);
2177 unsigned ExtType = LD->getExtensionType();
2180 case ISD::SEXTLOAD: // '17' bits known
2181 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2182 return VTBits-Tmp+1;
2183 case ISD::ZEXTLOAD: // '16' bits known
2184 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2189 // Allow the target to implement this method for its nodes.
2190 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
2191 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
2192 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
2193 Op.getOpcode() == ISD::INTRINSIC_VOID) {
2194 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
2195 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
2198 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2199 // use this information.
2200 APInt KnownZero, KnownOne;
2201 APInt Mask = APInt::getAllOnesValue(VTBits);
2202 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
2204 if (KnownZero.isNegative()) { // sign bit is 0
2206 } else if (KnownOne.isNegative()) { // sign bit is 1;
2213 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2214 // the number of identical bits in the top of the input value.
2216 Mask <<= Mask.getBitWidth()-VTBits;
2217 // Return # leading zeros. We use 'min' here in case Val was zero before
2218 // shifting. We don't want to return '64' as for an i32 "0".
2219 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2222 bool SelectionDAG::isKnownNeverNaN(SDValue Op) const {
2223 // If we're told that NaNs won't happen, assume they won't.
2224 if (FiniteOnlyFPMath())
2227 // If the value is a constant, we can obviously see if it is a NaN or not.
2228 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2229 return !C->getValueAPF().isNaN();
2231 // TODO: Recognize more cases here.
2236 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2237 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2238 if (!GA) return false;
2239 if (GA->getOffset() != 0) return false;
2240 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2241 if (!GV) return false;
2242 MachineModuleInfo *MMI = getMachineModuleInfo();
2243 return MMI && MMI->hasDebugInfo();
2247 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2248 /// element of the result of the vector shuffle.
2249 SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N,
2251 EVT VT = N->getValueType(0);
2252 DebugLoc dl = N->getDebugLoc();
2253 if (N->getMaskElt(i) < 0)
2254 return getUNDEF(VT.getVectorElementType());
2255 unsigned Index = N->getMaskElt(i);
2256 unsigned NumElems = VT.getVectorNumElements();
2257 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2260 if (V.getOpcode() == ISD::BIT_CONVERT) {
2261 V = V.getOperand(0);
2262 EVT VVT = V.getValueType();
2263 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems)
2266 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2267 return (Index == 0) ? V.getOperand(0)
2268 : getUNDEF(VT.getVectorElementType());
2269 if (V.getOpcode() == ISD::BUILD_VECTOR)
2270 return V.getOperand(Index);
2271 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V))
2272 return getShuffleScalarElt(SVN, Index);
2277 /// getNode - Gets or creates the specified node.
2279 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) {
2280 FoldingSetNodeID ID;
2281 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2283 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2284 return SDValue(E, 0);
2286 SDNode *N = NodeAllocator.Allocate<SDNode>();
2287 new (N) SDNode(Opcode, DL, getVTList(VT));
2288 CSEMap.InsertNode(N, IP);
2290 AllNodes.push_back(N);
2294 return SDValue(N, 0);
2297 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
2298 EVT VT, SDValue Operand) {
2299 // Constant fold unary operations with an integer constant operand.
2300 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2301 const APInt &Val = C->getAPIntValue();
2302 unsigned BitWidth = VT.getSizeInBits();
2305 case ISD::SIGN_EXTEND:
2306 return getConstant(APInt(Val).sextOrTrunc(BitWidth), VT);
2307 case ISD::ANY_EXTEND:
2308 case ISD::ZERO_EXTEND:
2310 return getConstant(APInt(Val).zextOrTrunc(BitWidth), VT);
2311 case ISD::UINT_TO_FP:
2312 case ISD::SINT_TO_FP: {
2313 const uint64_t zero[] = {0, 0};
2314 // No compile time operations on this type.
2315 if (VT==MVT::ppcf128)
2317 APFloat apf = APFloat(APInt(BitWidth, 2, zero));
2318 (void)apf.convertFromAPInt(Val,
2319 Opcode==ISD::SINT_TO_FP,
2320 APFloat::rmNearestTiesToEven);
2321 return getConstantFP(apf, VT);
2323 case ISD::BIT_CONVERT:
2324 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2325 return getConstantFP(Val.bitsToFloat(), VT);
2326 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2327 return getConstantFP(Val.bitsToDouble(), VT);
2330 return getConstant(Val.byteSwap(), VT);
2332 return getConstant(Val.countPopulation(), VT);
2334 return getConstant(Val.countLeadingZeros(), VT);
2336 return getConstant(Val.countTrailingZeros(), VT);
2340 // Constant fold unary operations with a floating point constant operand.
2341 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2342 APFloat V = C->getValueAPF(); // make copy
2343 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2347 return getConstantFP(V, VT);
2350 return getConstantFP(V, VT);
2352 case ISD::FP_EXTEND: {
2354 // This can return overflow, underflow, or inexact; we don't care.
2355 // FIXME need to be more flexible about rounding mode.
2356 (void)V.convert(*EVTToAPFloatSemantics(VT),
2357 APFloat::rmNearestTiesToEven, &ignored);
2358 return getConstantFP(V, VT);
2360 case ISD::FP_TO_SINT:
2361 case ISD::FP_TO_UINT: {
2364 assert(integerPartWidth >= 64);
2365 // FIXME need to be more flexible about rounding mode.
2366 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
2367 Opcode==ISD::FP_TO_SINT,
2368 APFloat::rmTowardZero, &ignored);
2369 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2371 APInt api(VT.getSizeInBits(), 2, x);
2372 return getConstant(api, VT);
2374 case ISD::BIT_CONVERT:
2375 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2376 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2377 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2378 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2384 unsigned OpOpcode = Operand.getNode()->getOpcode();
2386 case ISD::TokenFactor:
2387 case ISD::MERGE_VALUES:
2388 case ISD::CONCAT_VECTORS:
2389 return Operand; // Factor, merge or concat of one node? No need.
2390 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
2391 case ISD::FP_EXTEND:
2392 assert(VT.isFloatingPoint() &&
2393 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2394 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2395 assert((!VT.isVector() ||
2396 VT.getVectorNumElements() ==
2397 Operand.getValueType().getVectorNumElements()) &&
2398 "Vector element count mismatch!");
2399 if (Operand.getOpcode() == ISD::UNDEF)
2400 return getUNDEF(VT);
2402 case ISD::SIGN_EXTEND:
2403 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2404 "Invalid SIGN_EXTEND!");
2405 if (Operand.getValueType() == VT) return Operand; // noop extension
2406 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2407 "Invalid sext node, dst < src!");
2408 assert((!VT.isVector() ||
2409 VT.getVectorNumElements() ==
2410 Operand.getValueType().getVectorNumElements()) &&
2411 "Vector element count mismatch!");
2412 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2413 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2415 case ISD::ZERO_EXTEND:
2416 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2417 "Invalid ZERO_EXTEND!");
2418 if (Operand.getValueType() == VT) return Operand; // noop extension
2419 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2420 "Invalid zext node, dst < src!");
2421 assert((!VT.isVector() ||
2422 VT.getVectorNumElements() ==
2423 Operand.getValueType().getVectorNumElements()) &&
2424 "Vector element count mismatch!");
2425 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2426 return getNode(ISD::ZERO_EXTEND, DL, VT,
2427 Operand.getNode()->getOperand(0));
2429 case ISD::ANY_EXTEND:
2430 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2431 "Invalid ANY_EXTEND!");
2432 if (Operand.getValueType() == VT) return Operand; // noop extension
2433 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2434 "Invalid anyext node, dst < src!");
2435 assert((!VT.isVector() ||
2436 VT.getVectorNumElements() ==
2437 Operand.getValueType().getVectorNumElements()) &&
2438 "Vector element count mismatch!");
2439 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND)
2440 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2441 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2444 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2445 "Invalid TRUNCATE!");
2446 if (Operand.getValueType() == VT) return Operand; // noop truncate
2447 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) &&
2448 "Invalid truncate node, src < dst!");
2449 assert((!VT.isVector() ||
2450 VT.getVectorNumElements() ==
2451 Operand.getValueType().getVectorNumElements()) &&
2452 "Vector element count mismatch!");
2453 if (OpOpcode == ISD::TRUNCATE)
2454 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2455 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2456 OpOpcode == ISD::ANY_EXTEND) {
2457 // If the source is smaller than the dest, we still need an extend.
2458 if (Operand.getNode()->getOperand(0).getValueType().getScalarType()
2459 .bitsLT(VT.getScalarType()))
2460 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2461 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2462 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2464 return Operand.getNode()->getOperand(0);
2467 case ISD::BIT_CONVERT:
2468 // Basic sanity checking.
2469 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2470 && "Cannot BIT_CONVERT between types of different sizes!");
2471 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2472 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2473 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0));
2474 if (OpOpcode == ISD::UNDEF)
2475 return getUNDEF(VT);
2477 case ISD::SCALAR_TO_VECTOR:
2478 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2479 (VT.getVectorElementType() == Operand.getValueType() ||
2480 (VT.getVectorElementType().isInteger() &&
2481 Operand.getValueType().isInteger() &&
2482 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
2483 "Illegal SCALAR_TO_VECTOR node!");
2484 if (OpOpcode == ISD::UNDEF)
2485 return getUNDEF(VT);
2486 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2487 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2488 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2489 Operand.getConstantOperandVal(1) == 0 &&
2490 Operand.getOperand(0).getValueType() == VT)
2491 return Operand.getOperand(0);
2494 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0
2495 if (UnsafeFPMath && OpOpcode == ISD::FSUB)
2496 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1),
2497 Operand.getNode()->getOperand(0));
2498 if (OpOpcode == ISD::FNEG) // --X -> X
2499 return Operand.getNode()->getOperand(0);
2502 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2503 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0));
2508 SDVTList VTs = getVTList(VT);
2509 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2510 FoldingSetNodeID ID;
2511 SDValue Ops[1] = { Operand };
2512 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2514 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2515 return SDValue(E, 0);
2517 N = NodeAllocator.Allocate<UnarySDNode>();
2518 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2519 CSEMap.InsertNode(N, IP);
2521 N = NodeAllocator.Allocate<UnarySDNode>();
2522 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2525 AllNodes.push_back(N);
2529 return SDValue(N, 0);
2532 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2534 ConstantSDNode *Cst1,
2535 ConstantSDNode *Cst2) {
2536 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2539 case ISD::ADD: return getConstant(C1 + C2, VT);
2540 case ISD::SUB: return getConstant(C1 - C2, VT);
2541 case ISD::MUL: return getConstant(C1 * C2, VT);
2543 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2546 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2549 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2552 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2554 case ISD::AND: return getConstant(C1 & C2, VT);
2555 case ISD::OR: return getConstant(C1 | C2, VT);
2556 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2557 case ISD::SHL: return getConstant(C1 << C2, VT);
2558 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2559 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2560 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2561 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2568 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2569 SDValue N1, SDValue N2) {
2570 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2571 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2574 case ISD::TokenFactor:
2575 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2576 N2.getValueType() == MVT::Other && "Invalid token factor!");
2577 // Fold trivial token factors.
2578 if (N1.getOpcode() == ISD::EntryToken) return N2;
2579 if (N2.getOpcode() == ISD::EntryToken) return N1;
2580 if (N1 == N2) return N1;
2582 case ISD::CONCAT_VECTORS:
2583 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2584 // one big BUILD_VECTOR.
2585 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2586 N2.getOpcode() == ISD::BUILD_VECTOR) {
2587 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2588 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2589 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2593 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2594 N1.getValueType() == VT && "Binary operator types must match!");
2595 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2596 // worth handling here.
2597 if (N2C && N2C->isNullValue())
2599 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2606 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2607 N1.getValueType() == VT && "Binary operator types must match!");
2608 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2609 // it's worth handling here.
2610 if (N2C && N2C->isNullValue())
2620 assert(VT.isInteger() && "This operator does not apply to FP types!");
2628 if (Opcode == ISD::FADD) {
2630 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1))
2631 if (CFP->getValueAPF().isZero())
2634 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2635 if (CFP->getValueAPF().isZero())
2637 } else if (Opcode == ISD::FSUB) {
2639 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2640 if (CFP->getValueAPF().isZero())
2644 assert(N1.getValueType() == N2.getValueType() &&
2645 N1.getValueType() == VT && "Binary operator types must match!");
2647 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2648 assert(N1.getValueType() == VT &&
2649 N1.getValueType().isFloatingPoint() &&
2650 N2.getValueType().isFloatingPoint() &&
2651 "Invalid FCOPYSIGN!");
2658 assert(VT == N1.getValueType() &&
2659 "Shift operators return type must be the same as their first arg");
2660 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2661 "Shifts only work on integers");
2663 // Always fold shifts of i1 values so the code generator doesn't need to
2664 // handle them. Since we know the size of the shift has to be less than the
2665 // size of the value, the shift/rotate count is guaranteed to be zero.
2668 if (N2C && N2C->isNullValue())
2671 case ISD::FP_ROUND_INREG: {
2672 EVT EVT = cast<VTSDNode>(N2)->getVT();
2673 assert(VT == N1.getValueType() && "Not an inreg round!");
2674 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2675 "Cannot FP_ROUND_INREG integer types");
2676 assert(EVT.isVector() == VT.isVector() &&
2677 "FP_ROUND_INREG type should be vector iff the operand "
2679 assert((!EVT.isVector() ||
2680 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2681 "Vector element counts must match in FP_ROUND_INREG");
2682 assert(EVT.bitsLE(VT) && "Not rounding down!");
2683 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2687 assert(VT.isFloatingPoint() &&
2688 N1.getValueType().isFloatingPoint() &&
2689 VT.bitsLE(N1.getValueType()) &&
2690 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2691 if (N1.getValueType() == VT) return N1; // noop conversion.
2693 case ISD::AssertSext:
2694 case ISD::AssertZext: {
2695 EVT EVT = cast<VTSDNode>(N2)->getVT();
2696 assert(VT == N1.getValueType() && "Not an inreg extend!");
2697 assert(VT.isInteger() && EVT.isInteger() &&
2698 "Cannot *_EXTEND_INREG FP types");
2699 assert(!EVT.isVector() &&
2700 "AssertSExt/AssertZExt type should be the vector element type "
2701 "rather than the vector type!");
2702 assert(EVT.bitsLE(VT) && "Not extending!");
2703 if (VT == EVT) return N1; // noop assertion.
2706 case ISD::SIGN_EXTEND_INREG: {
2707 EVT EVT = cast<VTSDNode>(N2)->getVT();
2708 assert(VT == N1.getValueType() && "Not an inreg extend!");
2709 assert(VT.isInteger() && EVT.isInteger() &&
2710 "Cannot *_EXTEND_INREG FP types");
2711 assert(EVT.isVector() == VT.isVector() &&
2712 "SIGN_EXTEND_INREG type should be vector iff the operand "
2714 assert((!EVT.isVector() ||
2715 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2716 "Vector element counts must match in SIGN_EXTEND_INREG");
2717 assert(EVT.bitsLE(VT) && "Not extending!");
2718 if (EVT == VT) return N1; // Not actually extending
2721 APInt Val = N1C->getAPIntValue();
2722 unsigned FromBits = EVT.getScalarType().getSizeInBits();
2723 Val <<= Val.getBitWidth()-FromBits;
2724 Val = Val.ashr(Val.getBitWidth()-FromBits);
2725 return getConstant(Val, VT);
2729 case ISD::EXTRACT_VECTOR_ELT:
2730 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2731 if (N1.getOpcode() == ISD::UNDEF)
2732 return getUNDEF(VT);
2734 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2735 // expanding copies of large vectors from registers.
2737 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2738 N1.getNumOperands() > 0) {
2740 N1.getOperand(0).getValueType().getVectorNumElements();
2741 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
2742 N1.getOperand(N2C->getZExtValue() / Factor),
2743 getConstant(N2C->getZExtValue() % Factor,
2744 N2.getValueType()));
2747 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2748 // expanding large vector constants.
2749 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) {
2750 SDValue Elt = N1.getOperand(N2C->getZExtValue());
2751 EVT VEltTy = N1.getValueType().getVectorElementType();
2752 if (Elt.getValueType() != VEltTy) {
2753 // If the vector element type is not legal, the BUILD_VECTOR operands
2754 // are promoted and implicitly truncated. Make that explicit here.
2755 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt);
2758 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT
2759 // result is implicitly extended.
2760 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt);
2765 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2766 // operations are lowered to scalars.
2767 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2768 // If the indices are the same, return the inserted element else
2769 // if the indices are known different, extract the element from
2770 // the original vector.
2771 if (N1.getOperand(2) == N2) {
2772 if (VT == N1.getOperand(1).getValueType())
2773 return N1.getOperand(1);
2775 return getSExtOrTrunc(N1.getOperand(1), DL, VT);
2776 } else if (isa<ConstantSDNode>(N1.getOperand(2)) &&
2777 isa<ConstantSDNode>(N2))
2778 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
2781 case ISD::EXTRACT_ELEMENT:
2782 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2783 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2784 (N1.getValueType().isInteger() == VT.isInteger()) &&
2785 "Wrong types for EXTRACT_ELEMENT!");
2787 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2788 // 64-bit integers into 32-bit parts. Instead of building the extract of
2789 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2790 if (N1.getOpcode() == ISD::BUILD_PAIR)
2791 return N1.getOperand(N2C->getZExtValue());
2793 // EXTRACT_ELEMENT of a constant int is also very common.
2794 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2795 unsigned ElementSize = VT.getSizeInBits();
2796 unsigned Shift = ElementSize * N2C->getZExtValue();
2797 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2798 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2801 case ISD::EXTRACT_SUBVECTOR:
2802 if (N1.getValueType() == VT) // Trivial extraction.
2809 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2810 if (SV.getNode()) return SV;
2811 } else { // Cannonicalize constant to RHS if commutative
2812 if (isCommutativeBinOp(Opcode)) {
2813 std::swap(N1C, N2C);
2819 // Constant fold FP operations.
2820 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2821 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2823 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2824 // Cannonicalize constant to RHS if commutative
2825 std::swap(N1CFP, N2CFP);
2827 } else if (N2CFP && VT != MVT::ppcf128) {
2828 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2829 APFloat::opStatus s;
2832 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2833 if (s != APFloat::opInvalidOp)
2834 return getConstantFP(V1, VT);
2837 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2838 if (s!=APFloat::opInvalidOp)
2839 return getConstantFP(V1, VT);
2842 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2843 if (s!=APFloat::opInvalidOp)
2844 return getConstantFP(V1, VT);
2847 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2848 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2849 return getConstantFP(V1, VT);
2852 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2853 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2854 return getConstantFP(V1, VT);
2856 case ISD::FCOPYSIGN:
2858 return getConstantFP(V1, VT);
2864 // Canonicalize an UNDEF to the RHS, even over a constant.
2865 if (N1.getOpcode() == ISD::UNDEF) {
2866 if (isCommutativeBinOp(Opcode)) {
2870 case ISD::FP_ROUND_INREG:
2871 case ISD::SIGN_EXTEND_INREG:
2877 return N1; // fold op(undef, arg2) -> undef
2885 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2886 // For vectors, we can't easily build an all zero vector, just return
2893 // Fold a bunch of operators when the RHS is undef.
2894 if (N2.getOpcode() == ISD::UNDEF) {
2897 if (N1.getOpcode() == ISD::UNDEF)
2898 // Handle undef ^ undef -> 0 special case. This is a common
2900 return getConstant(0, VT);
2910 return N2; // fold op(arg1, undef) -> undef
2924 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2925 // For vectors, we can't easily build an all zero vector, just return
2930 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
2931 // For vectors, we can't easily build an all one vector, just return
2939 // Memoize this node if possible.
2941 SDVTList VTs = getVTList(VT);
2942 if (VT != MVT::Flag) {
2943 SDValue Ops[] = { N1, N2 };
2944 FoldingSetNodeID ID;
2945 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2947 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2948 return SDValue(E, 0);
2950 N = NodeAllocator.Allocate<BinarySDNode>();
2951 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2952 CSEMap.InsertNode(N, IP);
2954 N = NodeAllocator.Allocate<BinarySDNode>();
2955 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2958 AllNodes.push_back(N);
2962 return SDValue(N, 0);
2965 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2966 SDValue N1, SDValue N2, SDValue N3) {
2967 // Perform various simplifications.
2968 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2969 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2971 case ISD::CONCAT_VECTORS:
2972 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2973 // one big BUILD_VECTOR.
2974 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2975 N2.getOpcode() == ISD::BUILD_VECTOR &&
2976 N3.getOpcode() == ISD::BUILD_VECTOR) {
2977 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2978 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2979 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
2980 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2984 // Use FoldSetCC to simplify SETCC's.
2985 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL);
2986 if (Simp.getNode()) return Simp;
2991 if (N1C->getZExtValue())
2992 return N2; // select true, X, Y -> X
2994 return N3; // select false, X, Y -> Y
2997 if (N2 == N3) return N2; // select C, X, X -> X
3001 if (N2C->getZExtValue()) // Unconditional branch
3002 return getNode(ISD::BR, DL, MVT::Other, N1, N3);
3004 return N1; // Never-taken branch
3007 case ISD::VECTOR_SHUFFLE:
3008 llvm_unreachable("should use getVectorShuffle constructor!");
3010 case ISD::BIT_CONVERT:
3011 // Fold bit_convert nodes from a type to themselves.
3012 if (N1.getValueType() == VT)
3017 // Memoize node if it doesn't produce a flag.
3019 SDVTList VTs = getVTList(VT);
3020 if (VT != MVT::Flag) {
3021 SDValue Ops[] = { N1, N2, N3 };
3022 FoldingSetNodeID ID;
3023 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3025 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3026 return SDValue(E, 0);
3028 N = NodeAllocator.Allocate<TernarySDNode>();
3029 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3030 CSEMap.InsertNode(N, IP);
3032 N = NodeAllocator.Allocate<TernarySDNode>();
3033 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3036 AllNodes.push_back(N);
3040 return SDValue(N, 0);
3043 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3044 SDValue N1, SDValue N2, SDValue N3,
3046 SDValue Ops[] = { N1, N2, N3, N4 };
3047 return getNode(Opcode, DL, VT, Ops, 4);
3050 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3051 SDValue N1, SDValue N2, SDValue N3,
3052 SDValue N4, SDValue N5) {
3053 SDValue Ops[] = { N1, N2, N3, N4, N5 };
3054 return getNode(Opcode, DL, VT, Ops, 5);
3057 /// getStackArgumentTokenFactor - Compute a TokenFactor to force all
3058 /// the incoming stack arguments to be loaded from the stack.
3059 SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
3060 SmallVector<SDValue, 8> ArgChains;
3062 // Include the original chain at the beginning of the list. When this is
3063 // used by target LowerCall hooks, this helps legalize find the
3064 // CALLSEQ_BEGIN node.
3065 ArgChains.push_back(Chain);
3067 // Add a chain value for each stack argument.
3068 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(),
3069 UE = getEntryNode().getNode()->use_end(); U != UE; ++U)
3070 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
3071 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
3072 if (FI->getIndex() < 0)
3073 ArgChains.push_back(SDValue(L, 1));
3075 // Build a tokenfactor for all the chains.
3076 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
3077 &ArgChains[0], ArgChains.size());
3080 /// getMemsetValue - Vectorized representation of the memset value
3082 static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
3084 unsigned NumBits = VT.isVector() ?
3085 VT.getVectorElementType().getSizeInBits() : VT.getSizeInBits();
3086 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3087 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
3089 for (unsigned i = NumBits; i > 8; i >>= 1) {
3090 Val = (Val << Shift) | Val;
3094 return DAG.getConstant(Val, VT);
3095 return DAG.getConstantFP(APFloat(Val), VT);
3098 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3099 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
3101 for (unsigned i = NumBits; i > 8; i >>= 1) {
3102 Value = DAG.getNode(ISD::OR, dl, VT,
3103 DAG.getNode(ISD::SHL, dl, VT, Value,
3104 DAG.getConstant(Shift,
3105 TLI.getShiftAmountTy())),
3113 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3114 /// used when a memcpy is turned into a memset when the source is a constant
3116 static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG,
3117 const TargetLowering &TLI,
3118 std::string &Str, unsigned Offset) {
3119 // Handle vector with all elements zero.
3122 return DAG.getConstant(0, VT);
3123 unsigned NumElts = VT.getVectorNumElements();
3124 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3125 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3127 EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts)));
3130 assert(!VT.isVector() && "Can't handle vector type here!");
3131 unsigned NumBits = VT.getSizeInBits();
3132 unsigned MSB = NumBits / 8;
3134 if (TLI.isLittleEndian())
3135 Offset = Offset + MSB - 1;
3136 for (unsigned i = 0; i != MSB; ++i) {
3137 Val = (Val << 8) | (unsigned char)Str[Offset];
3138 Offset += TLI.isLittleEndian() ? -1 : 1;
3140 return DAG.getConstant(Val, VT);
3143 /// getMemBasePlusOffset - Returns base and offset node for the
3145 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
3146 SelectionDAG &DAG) {
3147 EVT VT = Base.getValueType();
3148 return DAG.getNode(ISD::ADD, Base.getDebugLoc(),
3149 VT, Base, DAG.getConstant(Offset, VT));
3152 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
3154 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
3155 unsigned SrcDelta = 0;
3156 GlobalAddressSDNode *G = NULL;
3157 if (Src.getOpcode() == ISD::GlobalAddress)
3158 G = cast<GlobalAddressSDNode>(Src);
3159 else if (Src.getOpcode() == ISD::ADD &&
3160 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3161 Src.getOperand(1).getOpcode() == ISD::Constant) {
3162 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
3163 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
3168 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3169 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
3175 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
3176 /// to replace the memset / memcpy is below the threshold. It also returns the
3177 /// types of the sequence of memory ops to perform memset / memcpy.
3179 bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps,
3180 SDValue Dst, SDValue Src,
3181 unsigned Limit, uint64_t Size, unsigned &Align,
3182 std::string &Str, bool &isSrcStr,
3184 const TargetLowering &TLI) {
3185 isSrcStr = isMemSrcFromString(Src, Str);
3186 bool isSrcConst = isa<ConstantSDNode>(Src);
3187 EVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG);
3188 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses(VT);
3189 if (VT != MVT::iAny) {
3190 const Type *Ty = VT.getTypeForEVT(*DAG.getContext());
3191 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3192 // If source is a string constant, this will require an unaligned load.
3193 if (NewAlign > Align && (isSrcConst || AllowUnalign)) {
3194 if (Dst.getOpcode() != ISD::FrameIndex) {
3195 // Can't change destination alignment. It requires a unaligned store.
3199 int FI = cast<FrameIndexSDNode>(Dst)->getIndex();
3200 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3201 if (MFI->isFixedObjectIndex(FI)) {
3202 // Can't change destination alignment. It requires a unaligned store.
3206 // Give the stack frame object a larger alignment if needed.
3207 if (MFI->getObjectAlignment(FI) < NewAlign)
3208 MFI->setObjectAlignment(FI, NewAlign);
3215 if (VT == MVT::iAny) {
3216 if (TLI.allowsUnalignedMemoryAccesses(MVT::i64)) {
3219 switch (Align & 7) {
3220 case 0: VT = MVT::i64; break;
3221 case 4: VT = MVT::i32; break;
3222 case 2: VT = MVT::i16; break;
3223 default: VT = MVT::i8; break;
3228 while (!TLI.isTypeLegal(LVT))
3229 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1);
3230 assert(LVT.isInteger());
3236 unsigned NumMemOps = 0;
3238 unsigned VTSize = VT.getSizeInBits() / 8;
3239 while (VTSize > Size) {
3240 // For now, only use non-vector load / store's for the left-over pieces.
3241 if (VT.isVector()) {
3243 while (!TLI.isTypeLegal(VT))
3244 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3245 VTSize = VT.getSizeInBits() / 8;
3247 // This can result in a type that is not legal on the target, e.g.
3248 // 1 or 2 bytes on PPC.
3249 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3254 if (++NumMemOps > Limit)
3256 MemOps.push_back(VT);
3263 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3264 SDValue Chain, SDValue Dst,
3265 SDValue Src, uint64_t Size,
3266 unsigned Align, bool AlwaysInline,
3267 const Value *DstSV, uint64_t DstSVOff,
3268 const Value *SrcSV, uint64_t SrcSVOff){
3269 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3271 // Expand memcpy to a series of load and store ops if the size operand falls
3272 // below a certain threshold.
3273 std::vector<EVT> MemOps;
3274 uint64_t Limit = -1ULL;
3276 Limit = TLI.getMaxStoresPerMemcpy();
3277 unsigned DstAlign = Align; // Destination alignment can change.
3280 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3281 Str, CopyFromStr, DAG, TLI))
3285 bool isZeroStr = CopyFromStr && Str.empty();
3286 SmallVector<SDValue, 8> OutChains;
3287 unsigned NumMemOps = MemOps.size();
3288 uint64_t SrcOff = 0, DstOff = 0;
3289 for (unsigned i = 0; i != NumMemOps; ++i) {
3291 unsigned VTSize = VT.getSizeInBits() / 8;
3292 SDValue Value, Store;
3294 if (CopyFromStr && (isZeroStr || !VT.isVector())) {
3295 // It's unlikely a store of a vector immediate can be done in a single
3296 // instruction. It would require a load from a constantpool first.
3297 // We also handle store a vector with all zero's.
3298 // FIXME: Handle other cases where store of vector immediate is done in
3299 // a single instruction.
3300 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff);
3301 Store = DAG.getStore(Chain, dl, Value,
3302 getMemBasePlusOffset(Dst, DstOff, DAG),
3303 DstSV, DstSVOff + DstOff, false, false, DstAlign);
3305 // The type might not be legal for the target. This should only happen
3306 // if the type is smaller than a legal type, as on PPC, so the right
3307 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
3308 // to Load/Store if NVT==VT.
3309 // FIXME does the case above also need this?
3310 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
3311 assert(NVT.bitsGE(VT));
3312 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain,
3313 getMemBasePlusOffset(Src, SrcOff, DAG),
3314 SrcSV, SrcSVOff + SrcOff, VT, false, false, Align);
3315 Store = DAG.getTruncStore(Chain, dl, Value,
3316 getMemBasePlusOffset(Dst, DstOff, DAG),
3317 DstSV, DstSVOff + DstOff, VT, false, false,
3320 OutChains.push_back(Store);
3325 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3326 &OutChains[0], OutChains.size());
3329 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3330 SDValue Chain, SDValue Dst,
3331 SDValue Src, uint64_t Size,
3332 unsigned Align, bool AlwaysInline,
3333 const Value *DstSV, uint64_t DstSVOff,
3334 const Value *SrcSV, uint64_t SrcSVOff){
3335 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3337 // Expand memmove to a series of load and store ops if the size operand falls
3338 // below a certain threshold.
3339 std::vector<EVT> MemOps;
3340 uint64_t Limit = -1ULL;
3342 Limit = TLI.getMaxStoresPerMemmove();
3343 unsigned DstAlign = Align; // Destination alignment can change.
3346 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3347 Str, CopyFromStr, DAG, TLI))
3350 uint64_t SrcOff = 0, DstOff = 0;
3352 SmallVector<SDValue, 8> LoadValues;
3353 SmallVector<SDValue, 8> LoadChains;
3354 SmallVector<SDValue, 8> OutChains;
3355 unsigned NumMemOps = MemOps.size();
3356 for (unsigned i = 0; i < NumMemOps; i++) {
3358 unsigned VTSize = VT.getSizeInBits() / 8;
3359 SDValue Value, Store;
3361 Value = DAG.getLoad(VT, dl, Chain,
3362 getMemBasePlusOffset(Src, SrcOff, DAG),
3363 SrcSV, SrcSVOff + SrcOff, false, false, Align);
3364 LoadValues.push_back(Value);
3365 LoadChains.push_back(Value.getValue(1));
3368 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3369 &LoadChains[0], LoadChains.size());
3371 for (unsigned i = 0; i < NumMemOps; i++) {
3373 unsigned VTSize = VT.getSizeInBits() / 8;
3374 SDValue Value, Store;
3376 Store = DAG.getStore(Chain, dl, LoadValues[i],
3377 getMemBasePlusOffset(Dst, DstOff, DAG),
3378 DstSV, DstSVOff + DstOff, false, false, DstAlign);
3379 OutChains.push_back(Store);
3383 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3384 &OutChains[0], OutChains.size());
3387 static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl,
3388 SDValue Chain, SDValue Dst,
3389 SDValue Src, uint64_t Size,
3391 const Value *DstSV, uint64_t DstSVOff) {
3392 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3394 // Expand memset to a series of load/store ops if the size operand
3395 // falls below a certain threshold.
3396 std::vector<EVT> MemOps;
3399 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(),
3400 Size, Align, Str, CopyFromStr, DAG, TLI))
3403 SmallVector<SDValue, 8> OutChains;
3404 uint64_t DstOff = 0;
3406 unsigned NumMemOps = MemOps.size();
3407 for (unsigned i = 0; i < NumMemOps; i++) {
3409 unsigned VTSize = VT.getSizeInBits() / 8;
3410 SDValue Value = getMemsetValue(Src, VT, DAG, dl);
3411 SDValue Store = DAG.getStore(Chain, dl, Value,
3412 getMemBasePlusOffset(Dst, DstOff, DAG),
3413 DstSV, DstSVOff + DstOff, false, false, 0);
3414 OutChains.push_back(Store);
3418 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3419 &OutChains[0], OutChains.size());
3422 SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst,
3423 SDValue Src, SDValue Size,
3424 unsigned Align, bool AlwaysInline,
3425 const Value *DstSV, uint64_t DstSVOff,
3426 const Value *SrcSV, uint64_t SrcSVOff) {
3428 // Check to see if we should lower the memcpy to loads and stores first.
3429 // For cases within the target-specified limits, this is the best choice.
3430 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3432 // Memcpy with size zero? Just return the original chain.
3433 if (ConstantSize->isNullValue())
3437 getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3438 ConstantSize->getZExtValue(),
3439 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3440 if (Result.getNode())
3444 // Then check to see if we should lower the memcpy with target-specific
3445 // code. If the target chooses to do this, this is the next best.
3447 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align,
3449 DstSV, DstSVOff, SrcSV, SrcSVOff);
3450 if (Result.getNode())
3453 // If we really need inline code and the target declined to provide it,
3454 // use a (potentially long) sequence of loads and stores.
3456 assert(ConstantSize && "AlwaysInline requires a constant size!");
3457 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3458 ConstantSize->getZExtValue(), Align, true,
3459 DstSV, DstSVOff, SrcSV, SrcSVOff);
3462 // Emit a library call.
3463 TargetLowering::ArgListTy Args;
3464 TargetLowering::ArgListEntry Entry;
3465 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3466 Entry.Node = Dst; Args.push_back(Entry);
3467 Entry.Node = Src; Args.push_back(Entry);
3468 Entry.Node = Size; Args.push_back(Entry);
3469 // FIXME: pass in DebugLoc
3470 std::pair<SDValue,SDValue> CallResult =
3471 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3472 false, false, false, false, 0,
3473 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false,
3474 /*isReturnValueUsed=*/false,
3475 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY),
3476 TLI.getPointerTy()),
3477 Args, *this, dl, GetOrdering(Chain.getNode()));
3478 return CallResult.second;
3481 SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst,
3482 SDValue Src, SDValue Size,
3484 const Value *DstSV, uint64_t DstSVOff,
3485 const Value *SrcSV, uint64_t SrcSVOff) {
3487 // Check to see if we should lower the memmove to loads and stores first.
3488 // For cases within the target-specified limits, this is the best choice.
3489 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3491 // Memmove with size zero? Just return the original chain.
3492 if (ConstantSize->isNullValue())
3496 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src,
3497 ConstantSize->getZExtValue(),
3498 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3499 if (Result.getNode())
3503 // Then check to see if we should lower the memmove with target-specific
3504 // code. If the target chooses to do this, this is the next best.
3506 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align,
3507 DstSV, DstSVOff, SrcSV, SrcSVOff);
3508 if (Result.getNode())
3511 // Emit a library call.
3512 TargetLowering::ArgListTy Args;
3513 TargetLowering::ArgListEntry Entry;
3514 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3515 Entry.Node = Dst; Args.push_back(Entry);
3516 Entry.Node = Src; Args.push_back(Entry);
3517 Entry.Node = Size; Args.push_back(Entry);
3518 // FIXME: pass in DebugLoc
3519 std::pair<SDValue,SDValue> CallResult =
3520 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3521 false, false, false, false, 0,
3522 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false,
3523 /*isReturnValueUsed=*/false,
3524 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE),
3525 TLI.getPointerTy()),
3526 Args, *this, dl, GetOrdering(Chain.getNode()));
3527 return CallResult.second;
3530 SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst,
3531 SDValue Src, SDValue Size,
3533 const Value *DstSV, uint64_t DstSVOff) {
3535 // Check to see if we should lower the memset to stores first.
3536 // For cases within the target-specified limits, this is the best choice.
3537 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3539 // Memset with size zero? Just return the original chain.
3540 if (ConstantSize->isNullValue())
3544 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
3545 Align, DstSV, DstSVOff);
3546 if (Result.getNode())
3550 // Then check to see if we should lower the memset with target-specific
3551 // code. If the target chooses to do this, this is the next best.
3553 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align,
3555 if (Result.getNode())
3558 // Emit a library call.
3559 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext());
3560 TargetLowering::ArgListTy Args;
3561 TargetLowering::ArgListEntry Entry;
3562 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3563 Args.push_back(Entry);
3564 // Extend or truncate the argument to be an i32 value for the call.
3565 if (Src.getValueType().bitsGT(MVT::i32))
3566 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
3568 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
3570 Entry.Ty = Type::getInt32Ty(*getContext());
3571 Entry.isSExt = true;
3572 Args.push_back(Entry);
3574 Entry.Ty = IntPtrTy;
3575 Entry.isSExt = false;
3576 Args.push_back(Entry);
3577 // FIXME: pass in DebugLoc
3578 std::pair<SDValue,SDValue> CallResult =
3579 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3580 false, false, false, false, 0,
3581 TLI.getLibcallCallingConv(RTLIB::MEMSET), false,
3582 /*isReturnValueUsed=*/false,
3583 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET),
3584 TLI.getPointerTy()),
3585 Args, *this, dl, GetOrdering(Chain.getNode()));
3586 return CallResult.second;
3589 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3591 SDValue Ptr, SDValue Cmp,
3592 SDValue Swp, const Value* PtrVal,
3593 unsigned Alignment) {
3594 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3595 Alignment = getEVTAlignment(MemVT);
3597 // Check if the memory reference references a frame index
3599 if (const FrameIndexSDNode *FI =
3600 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3601 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3603 MachineFunction &MF = getMachineFunction();
3604 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3606 // For now, atomics are considered to be volatile always.
3607 Flags |= MachineMemOperand::MOVolatile;
3609 MachineMemOperand *MMO =
3610 MF.getMachineMemOperand(PtrVal, Flags, 0,
3611 MemVT.getStoreSize(), Alignment);
3613 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3616 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3618 SDValue Ptr, SDValue Cmp,
3619 SDValue Swp, MachineMemOperand *MMO) {
3620 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op");
3621 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3623 EVT VT = Cmp.getValueType();
3625 SDVTList VTs = getVTList(VT, MVT::Other);
3626 FoldingSetNodeID ID;
3627 ID.AddInteger(MemVT.getRawBits());
3628 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3629 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3631 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3632 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3633 return SDValue(E, 0);
3635 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3636 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3637 CSEMap.InsertNode(N, IP);
3638 AllNodes.push_back(N);
3639 return SDValue(N, 0);
3642 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3644 SDValue Ptr, SDValue Val,
3645 const Value* PtrVal,
3646 unsigned Alignment) {
3647 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3648 Alignment = getEVTAlignment(MemVT);
3650 // Check if the memory reference references a frame index
3652 if (const FrameIndexSDNode *FI =
3653 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3654 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3656 MachineFunction &MF = getMachineFunction();
3657 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3659 // For now, atomics are considered to be volatile always.
3660 Flags |= MachineMemOperand::MOVolatile;
3662 MachineMemOperand *MMO =
3663 MF.getMachineMemOperand(PtrVal, Flags, 0,
3664 MemVT.getStoreSize(), Alignment);
3666 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO);
3669 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3671 SDValue Ptr, SDValue Val,
3672 MachineMemOperand *MMO) {
3673 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
3674 Opcode == ISD::ATOMIC_LOAD_SUB ||
3675 Opcode == ISD::ATOMIC_LOAD_AND ||
3676 Opcode == ISD::ATOMIC_LOAD_OR ||
3677 Opcode == ISD::ATOMIC_LOAD_XOR ||
3678 Opcode == ISD::ATOMIC_LOAD_NAND ||
3679 Opcode == ISD::ATOMIC_LOAD_MIN ||
3680 Opcode == ISD::ATOMIC_LOAD_MAX ||
3681 Opcode == ISD::ATOMIC_LOAD_UMIN ||
3682 Opcode == ISD::ATOMIC_LOAD_UMAX ||
3683 Opcode == ISD::ATOMIC_SWAP) &&
3684 "Invalid Atomic Op");
3686 EVT VT = Val.getValueType();
3688 SDVTList VTs = getVTList(VT, MVT::Other);
3689 FoldingSetNodeID ID;
3690 ID.AddInteger(MemVT.getRawBits());
3691 SDValue Ops[] = {Chain, Ptr, Val};
3692 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3694 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3695 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3696 return SDValue(E, 0);
3698 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3699 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Val, MMO);
3700 CSEMap.InsertNode(N, IP);
3701 AllNodes.push_back(N);
3702 return SDValue(N, 0);
3705 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3706 /// Allowed to return something different (and simpler) if Simplify is true.
3707 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3712 SmallVector<EVT, 4> VTs;
3713 VTs.reserve(NumOps);
3714 for (unsigned i = 0; i < NumOps; ++i)
3715 VTs.push_back(Ops[i].getValueType());
3716 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps),
3721 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl,
3722 const EVT *VTs, unsigned NumVTs,
3723 const SDValue *Ops, unsigned NumOps,
3724 EVT MemVT, const Value *srcValue, int SVOff,
3725 unsigned Align, bool Vol,
3726 bool ReadMem, bool WriteMem) {
3727 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps,
3728 MemVT, srcValue, SVOff, Align, Vol,
3733 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3734 const SDValue *Ops, unsigned NumOps,
3735 EVT MemVT, const Value *srcValue, int SVOff,
3736 unsigned Align, bool Vol,
3737 bool ReadMem, bool WriteMem) {
3738 if (Align == 0) // Ensure that codegen never sees alignment 0
3739 Align = getEVTAlignment(MemVT);
3741 MachineFunction &MF = getMachineFunction();
3744 Flags |= MachineMemOperand::MOStore;
3746 Flags |= MachineMemOperand::MOLoad;
3748 Flags |= MachineMemOperand::MOVolatile;
3749 MachineMemOperand *MMO =
3750 MF.getMachineMemOperand(srcValue, Flags, SVOff,
3751 MemVT.getStoreSize(), Align);
3753 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3757 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3758 const SDValue *Ops, unsigned NumOps,
3759 EVT MemVT, MachineMemOperand *MMO) {
3760 assert((Opcode == ISD::INTRINSIC_VOID ||
3761 Opcode == ISD::INTRINSIC_W_CHAIN ||
3762 (Opcode <= INT_MAX &&
3763 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) &&
3764 "Opcode is not a memory-accessing opcode!");
3766 // Memoize the node unless it returns a flag.
3767 MemIntrinsicSDNode *N;
3768 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3769 FoldingSetNodeID ID;
3770 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3772 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3773 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO);
3774 return SDValue(E, 0);
3777 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3778 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3779 CSEMap.InsertNode(N, IP);
3781 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3782 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3784 AllNodes.push_back(N);
3785 return SDValue(N, 0);
3789 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3790 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3791 SDValue Ptr, SDValue Offset,
3792 const Value *SV, int SVOffset, EVT MemVT,
3793 bool isVolatile, bool isNonTemporal,
3794 unsigned Alignment) {
3795 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3796 Alignment = getEVTAlignment(VT);
3798 // Check if the memory reference references a frame index
3800 if (const FrameIndexSDNode *FI =
3801 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3802 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3804 MachineFunction &MF = getMachineFunction();
3805 unsigned Flags = MachineMemOperand::MOLoad;
3807 Flags |= MachineMemOperand::MOVolatile;
3809 Flags |= MachineMemOperand::MONonTemporal;
3810 MachineMemOperand *MMO =
3811 MF.getMachineMemOperand(SV, Flags, SVOffset,
3812 MemVT.getStoreSize(), Alignment);
3813 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO);
3817 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3818 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3819 SDValue Ptr, SDValue Offset, EVT MemVT,
3820 MachineMemOperand *MMO) {
3822 ExtType = ISD::NON_EXTLOAD;
3823 } else if (ExtType == ISD::NON_EXTLOAD) {
3824 assert(VT == MemVT && "Non-extending load from different memory type!");
3827 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) &&
3828 "Should only be an extending load, not truncating!");
3829 assert(VT.isInteger() == MemVT.isInteger() &&
3830 "Cannot convert from FP to Int or Int -> FP!");
3831 assert(VT.isVector() == MemVT.isVector() &&
3832 "Cannot use trunc store to convert to or from a vector!");
3833 assert((!VT.isVector() ||
3834 VT.getVectorNumElements() == MemVT.getVectorNumElements()) &&
3835 "Cannot use trunc store to change the number of vector elements!");
3838 bool Indexed = AM != ISD::UNINDEXED;
3839 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3840 "Unindexed load with an offset!");
3842 SDVTList VTs = Indexed ?
3843 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3844 SDValue Ops[] = { Chain, Ptr, Offset };
3845 FoldingSetNodeID ID;
3846 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3847 ID.AddInteger(MemVT.getRawBits());
3848 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile()));
3850 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3851 cast<LoadSDNode>(E)->refineAlignment(MMO);
3852 return SDValue(E, 0);
3854 SDNode *N = NodeAllocator.Allocate<LoadSDNode>();
3855 new (N) LoadSDNode(Ops, dl, VTs, AM, ExtType, MemVT, MMO);
3856 CSEMap.InsertNode(N, IP);
3857 AllNodes.push_back(N);
3858 return SDValue(N, 0);
3861 SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl,
3862 SDValue Chain, SDValue Ptr,
3863 const Value *SV, int SVOffset,
3864 bool isVolatile, bool isNonTemporal,
3865 unsigned Alignment) {
3866 SDValue Undef = getUNDEF(Ptr.getValueType());
3867 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3868 SV, SVOffset, VT, isVolatile, isNonTemporal, Alignment);
3871 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT,
3872 SDValue Chain, SDValue Ptr,
3874 int SVOffset, EVT MemVT,
3875 bool isVolatile, bool isNonTemporal,
3876 unsigned Alignment) {
3877 SDValue Undef = getUNDEF(Ptr.getValueType());
3878 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef,
3879 SV, SVOffset, MemVT, isVolatile, isNonTemporal, Alignment);
3883 SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base,
3884 SDValue Offset, ISD::MemIndexedMode AM) {
3885 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
3886 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
3887 "Load is already a indexed load!");
3888 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(),
3889 LD->getChain(), Base, Offset, LD->getSrcValue(),
3890 LD->getSrcValueOffset(), LD->getMemoryVT(),
3891 LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment());
3894 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3895 SDValue Ptr, const Value *SV, int SVOffset,
3896 bool isVolatile, bool isNonTemporal,
3897 unsigned Alignment) {
3898 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3899 Alignment = getEVTAlignment(Val.getValueType());
3901 // Check if the memory reference references a frame index
3903 if (const FrameIndexSDNode *FI =
3904 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3905 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3907 MachineFunction &MF = getMachineFunction();
3908 unsigned Flags = MachineMemOperand::MOStore;
3910 Flags |= MachineMemOperand::MOVolatile;
3912 Flags |= MachineMemOperand::MONonTemporal;
3913 MachineMemOperand *MMO =
3914 MF.getMachineMemOperand(SV, Flags, SVOffset,
3915 Val.getValueType().getStoreSize(), Alignment);
3917 return getStore(Chain, dl, Val, Ptr, MMO);
3920 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3921 SDValue Ptr, MachineMemOperand *MMO) {
3922 EVT VT = Val.getValueType();
3923 SDVTList VTs = getVTList(MVT::Other);
3924 SDValue Undef = getUNDEF(Ptr.getValueType());
3925 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3926 FoldingSetNodeID ID;
3927 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3928 ID.AddInteger(VT.getRawBits());
3929 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile()));
3931 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3932 cast<StoreSDNode>(E)->refineAlignment(MMO);
3933 return SDValue(E, 0);
3935 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3936 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, false, VT, MMO);
3937 CSEMap.InsertNode(N, IP);
3938 AllNodes.push_back(N);
3939 return SDValue(N, 0);
3942 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3943 SDValue Ptr, const Value *SV,
3944 int SVOffset, EVT SVT,
3945 bool isVolatile, bool isNonTemporal,
3946 unsigned Alignment) {
3947 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3948 Alignment = getEVTAlignment(SVT);
3950 // Check if the memory reference references a frame index
3952 if (const FrameIndexSDNode *FI =
3953 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3954 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3956 MachineFunction &MF = getMachineFunction();
3957 unsigned Flags = MachineMemOperand::MOStore;
3959 Flags |= MachineMemOperand::MOVolatile;
3961 Flags |= MachineMemOperand::MONonTemporal;
3962 MachineMemOperand *MMO =
3963 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment);
3965 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
3968 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3969 SDValue Ptr, EVT SVT,
3970 MachineMemOperand *MMO) {
3971 EVT VT = Val.getValueType();
3974 return getStore(Chain, dl, Val, Ptr, MMO);
3976 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
3977 "Should only be a truncating store, not extending!");
3978 assert(VT.isInteger() == SVT.isInteger() &&
3979 "Can't do FP-INT conversion!");
3980 assert(VT.isVector() == SVT.isVector() &&
3981 "Cannot use trunc store to convert to or from a vector!");
3982 assert((!VT.isVector() ||
3983 VT.getVectorNumElements() == SVT.getVectorNumElements()) &&
3984 "Cannot use trunc store to change the number of vector elements!");
3986 SDVTList VTs = getVTList(MVT::Other);
3987 SDValue Undef = getUNDEF(Ptr.getValueType());
3988 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3989 FoldingSetNodeID ID;
3990 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3991 ID.AddInteger(SVT.getRawBits());
3992 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile()));
3994 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3995 cast<StoreSDNode>(E)->refineAlignment(MMO);
3996 return SDValue(E, 0);
3998 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3999 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, true, SVT, MMO);
4000 CSEMap.InsertNode(N, IP);
4001 AllNodes.push_back(N);
4002 return SDValue(N, 0);
4006 SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base,
4007 SDValue Offset, ISD::MemIndexedMode AM) {
4008 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
4009 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
4010 "Store is already a indexed store!");
4011 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
4012 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
4013 FoldingSetNodeID ID;
4014 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4015 ID.AddInteger(ST->getMemoryVT().getRawBits());
4016 ID.AddInteger(ST->getRawSubclassData());
4018 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4019 return SDValue(E, 0);
4021 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
4022 new (N) StoreSDNode(Ops, dl, VTs, AM,
4023 ST->isTruncatingStore(), ST->getMemoryVT(),
4024 ST->getMemOperand());
4025 CSEMap.InsertNode(N, IP);
4026 AllNodes.push_back(N);
4027 return SDValue(N, 0);
4030 SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl,
4031 SDValue Chain, SDValue Ptr,
4033 SDValue Ops[] = { Chain, Ptr, SV };
4034 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3);
4037 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4038 const SDUse *Ops, unsigned NumOps) {
4040 case 0: return getNode(Opcode, DL, VT);
4041 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4042 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4043 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4047 // Copy from an SDUse array into an SDValue array for use with
4048 // the regular getNode logic.
4049 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
4050 return getNode(Opcode, DL, VT, &NewOps[0], NumOps);
4053 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4054 const SDValue *Ops, unsigned NumOps) {
4056 case 0: return getNode(Opcode, DL, VT);
4057 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4058 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4059 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4065 case ISD::SELECT_CC: {
4066 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
4067 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
4068 "LHS and RHS of condition must have same type!");
4069 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4070 "True and False arms of SelectCC must have same type!");
4071 assert(Ops[2].getValueType() == VT &&
4072 "select_cc node must be of same type as true and false value!");
4076 assert(NumOps == 5 && "BR_CC takes 5 operands!");
4077 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4078 "LHS/RHS of comparison should match types!");
4085 SDVTList VTs = getVTList(VT);
4087 if (VT != MVT::Flag) {
4088 FoldingSetNodeID ID;
4089 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
4092 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4093 return SDValue(E, 0);
4095 N = NodeAllocator.Allocate<SDNode>();
4096 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
4097 CSEMap.InsertNode(N, IP);
4099 N = NodeAllocator.Allocate<SDNode>();
4100 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
4103 AllNodes.push_back(N);
4107 return SDValue(N, 0);
4110 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4111 const std::vector<EVT> &ResultTys,
4112 const SDValue *Ops, unsigned NumOps) {
4113 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()),
4117 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4118 const EVT *VTs, unsigned NumVTs,
4119 const SDValue *Ops, unsigned NumOps) {
4121 return getNode(Opcode, DL, VTs[0], Ops, NumOps);
4122 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps);
4125 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4126 const SDValue *Ops, unsigned NumOps) {
4127 if (VTList.NumVTs == 1)
4128 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps);
4132 // FIXME: figure out how to safely handle things like
4133 // int foo(int x) { return 1 << (x & 255); }
4134 // int bar() { return foo(256); }
4135 case ISD::SRA_PARTS:
4136 case ISD::SRL_PARTS:
4137 case ISD::SHL_PARTS:
4138 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4139 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
4140 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4141 else if (N3.getOpcode() == ISD::AND)
4142 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
4143 // If the and is only masking out bits that cannot effect the shift,
4144 // eliminate the and.
4145 unsigned NumBits = VT.getScalarType().getSizeInBits()*2;
4146 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
4147 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4153 // Memoize the node unless it returns a flag.
4155 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4156 FoldingSetNodeID ID;
4157 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4159 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4160 return SDValue(E, 0);
4163 N = NodeAllocator.Allocate<UnarySDNode>();
4164 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4165 } else if (NumOps == 2) {
4166 N = NodeAllocator.Allocate<BinarySDNode>();
4167 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4168 } else if (NumOps == 3) {
4169 N = NodeAllocator.Allocate<TernarySDNode>();
4170 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
4172 N = NodeAllocator.Allocate<SDNode>();
4173 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
4175 CSEMap.InsertNode(N, IP);
4178 N = NodeAllocator.Allocate<UnarySDNode>();
4179 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4180 } else if (NumOps == 2) {
4181 N = NodeAllocator.Allocate<BinarySDNode>();
4182 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4183 } else if (NumOps == 3) {
4184 N = NodeAllocator.Allocate<TernarySDNode>();
4185 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
4187 N = NodeAllocator.Allocate<SDNode>();
4188 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
4191 AllNodes.push_back(N);
4195 return SDValue(N, 0);
4198 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) {
4199 return getNode(Opcode, DL, VTList, 0, 0);
4202 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4204 SDValue Ops[] = { N1 };
4205 return getNode(Opcode, DL, VTList, Ops, 1);
4208 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4209 SDValue N1, SDValue N2) {
4210 SDValue Ops[] = { N1, N2 };
4211 return getNode(Opcode, DL, VTList, Ops, 2);
4214 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4215 SDValue N1, SDValue N2, SDValue N3) {
4216 SDValue Ops[] = { N1, N2, N3 };
4217 return getNode(Opcode, DL, VTList, Ops, 3);
4220 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4221 SDValue N1, SDValue N2, SDValue N3,
4223 SDValue Ops[] = { N1, N2, N3, N4 };
4224 return getNode(Opcode, DL, VTList, Ops, 4);
4227 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4228 SDValue N1, SDValue N2, SDValue N3,
4229 SDValue N4, SDValue N5) {
4230 SDValue Ops[] = { N1, N2, N3, N4, N5 };
4231 return getNode(Opcode, DL, VTList, Ops, 5);
4234 SDVTList SelectionDAG::getVTList(EVT VT) {
4235 return makeVTList(SDNode::getValueTypeList(VT), 1);
4238 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
4239 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4240 E = VTList.rend(); I != E; ++I)
4241 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
4244 EVT *Array = Allocator.Allocate<EVT>(2);
4247 SDVTList Result = makeVTList(Array, 2);
4248 VTList.push_back(Result);
4252 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
4253 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4254 E = VTList.rend(); I != E; ++I)
4255 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4259 EVT *Array = Allocator.Allocate<EVT>(3);
4263 SDVTList Result = makeVTList(Array, 3);
4264 VTList.push_back(Result);
4268 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
4269 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4270 E = VTList.rend(); I != E; ++I)
4271 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4272 I->VTs[2] == VT3 && I->VTs[3] == VT4)
4275 EVT *Array = Allocator.Allocate<EVT>(4);
4280 SDVTList Result = makeVTList(Array, 4);
4281 VTList.push_back(Result);
4285 SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) {
4287 case 0: llvm_unreachable("Cannot have nodes without results!");
4288 case 1: return getVTList(VTs[0]);
4289 case 2: return getVTList(VTs[0], VTs[1]);
4290 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
4291 case 4: return getVTList(VTs[0], VTs[1], VTs[2], VTs[3]);
4295 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4296 E = VTList.rend(); I != E; ++I) {
4297 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
4300 bool NoMatch = false;
4301 for (unsigned i = 2; i != NumVTs; ++i)
4302 if (VTs[i] != I->VTs[i]) {
4310 EVT *Array = Allocator.Allocate<EVT>(NumVTs);
4311 std::copy(VTs, VTs+NumVTs, Array);
4312 SDVTList Result = makeVTList(Array, NumVTs);
4313 VTList.push_back(Result);
4318 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
4319 /// specified operands. If the resultant node already exists in the DAG,
4320 /// this does not modify the specified node, instead it returns the node that
4321 /// already exists. If the resultant node does not exist in the DAG, the
4322 /// input node is returned. As a degenerate case, if you specify the same
4323 /// input operands as the node already has, the input node is returned.
4324 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
4325 SDNode *N = InN.getNode();
4326 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
4328 // Check to see if there is no change.
4329 if (Op == N->getOperand(0)) return InN;
4331 // See if the modified node already exists.
4332 void *InsertPos = 0;
4333 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
4334 return SDValue(Existing, InN.getResNo());
4336 // Nope it doesn't. Remove the node from its current place in the maps.
4338 if (!RemoveNodeFromCSEMaps(N))
4341 // Now we update the operands.
4342 N->OperandList[0].set(Op);
4344 // If this gets put into a CSE map, add it.
4345 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4349 SDValue SelectionDAG::
4350 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
4351 SDNode *N = InN.getNode();
4352 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
4354 // Check to see if there is no change.
4355 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
4356 return InN; // No operands changed, just return the input node.
4358 // See if the modified node already exists.
4359 void *InsertPos = 0;
4360 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
4361 return SDValue(Existing, InN.getResNo());
4363 // Nope it doesn't. Remove the node from its current place in the maps.
4365 if (!RemoveNodeFromCSEMaps(N))
4368 // Now we update the operands.
4369 if (N->OperandList[0] != Op1)
4370 N->OperandList[0].set(Op1);
4371 if (N->OperandList[1] != Op2)
4372 N->OperandList[1].set(Op2);
4374 // If this gets put into a CSE map, add it.
4375 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4379 SDValue SelectionDAG::
4380 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
4381 SDValue Ops[] = { Op1, Op2, Op3 };
4382 return UpdateNodeOperands(N, Ops, 3);
4385 SDValue SelectionDAG::
4386 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4387 SDValue Op3, SDValue Op4) {
4388 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
4389 return UpdateNodeOperands(N, Ops, 4);
4392 SDValue SelectionDAG::
4393 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4394 SDValue Op3, SDValue Op4, SDValue Op5) {
4395 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
4396 return UpdateNodeOperands(N, Ops, 5);
4399 SDValue SelectionDAG::
4400 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
4401 SDNode *N = InN.getNode();
4402 assert(N->getNumOperands() == NumOps &&
4403 "Update with wrong number of operands");
4405 // Check to see if there is no change.
4406 bool AnyChange = false;
4407 for (unsigned i = 0; i != NumOps; ++i) {
4408 if (Ops[i] != N->getOperand(i)) {
4414 // No operands changed, just return the input node.
4415 if (!AnyChange) return InN;
4417 // See if the modified node already exists.
4418 void *InsertPos = 0;
4419 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
4420 return SDValue(Existing, InN.getResNo());
4422 // Nope it doesn't. Remove the node from its current place in the maps.
4424 if (!RemoveNodeFromCSEMaps(N))
4427 // Now we update the operands.
4428 for (unsigned i = 0; i != NumOps; ++i)
4429 if (N->OperandList[i] != Ops[i])
4430 N->OperandList[i].set(Ops[i]);
4432 // If this gets put into a CSE map, add it.
4433 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4437 /// DropOperands - Release the operands and set this node to have
4439 void SDNode::DropOperands() {
4440 // Unlike the code in MorphNodeTo that does this, we don't need to
4441 // watch for dead nodes here.
4442 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
4448 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
4451 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4453 SDVTList VTs = getVTList(VT);
4454 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4457 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4458 EVT VT, SDValue Op1) {
4459 SDVTList VTs = getVTList(VT);
4460 SDValue Ops[] = { Op1 };
4461 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4464 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4465 EVT VT, SDValue Op1,
4467 SDVTList VTs = getVTList(VT);
4468 SDValue Ops[] = { Op1, Op2 };
4469 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4472 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4473 EVT VT, SDValue Op1,
4474 SDValue Op2, SDValue Op3) {
4475 SDVTList VTs = getVTList(VT);
4476 SDValue Ops[] = { Op1, Op2, Op3 };
4477 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4480 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4481 EVT VT, const SDValue *Ops,
4483 SDVTList VTs = getVTList(VT);
4484 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4487 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4488 EVT VT1, EVT VT2, const SDValue *Ops,
4490 SDVTList VTs = getVTList(VT1, VT2);
4491 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4494 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4496 SDVTList VTs = getVTList(VT1, VT2);
4497 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4500 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4501 EVT VT1, EVT VT2, EVT VT3,
4502 const SDValue *Ops, unsigned NumOps) {
4503 SDVTList VTs = getVTList(VT1, VT2, VT3);
4504 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4507 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4508 EVT VT1, EVT VT2, EVT VT3, EVT VT4,
4509 const SDValue *Ops, unsigned NumOps) {
4510 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4511 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4514 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4517 SDVTList VTs = getVTList(VT1, VT2);
4518 SDValue Ops[] = { Op1 };
4519 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4522 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4524 SDValue Op1, SDValue Op2) {
4525 SDVTList VTs = getVTList(VT1, VT2);
4526 SDValue Ops[] = { Op1, Op2 };
4527 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4530 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4532 SDValue Op1, SDValue Op2,
4534 SDVTList VTs = getVTList(VT1, VT2);
4535 SDValue Ops[] = { Op1, Op2, Op3 };
4536 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4539 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4540 EVT VT1, EVT VT2, EVT VT3,
4541 SDValue Op1, SDValue Op2,
4543 SDVTList VTs = getVTList(VT1, VT2, VT3);
4544 SDValue Ops[] = { Op1, Op2, Op3 };
4545 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4548 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4549 SDVTList VTs, const SDValue *Ops,
4551 return MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4554 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4556 SDVTList VTs = getVTList(VT);
4557 return MorphNodeTo(N, Opc, VTs, 0, 0);
4560 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4561 EVT VT, SDValue Op1) {
4562 SDVTList VTs = getVTList(VT);
4563 SDValue Ops[] = { Op1 };
4564 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4567 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4568 EVT VT, SDValue Op1,
4570 SDVTList VTs = getVTList(VT);
4571 SDValue Ops[] = { Op1, Op2 };
4572 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4575 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4576 EVT VT, SDValue Op1,
4577 SDValue Op2, SDValue Op3) {
4578 SDVTList VTs = getVTList(VT);
4579 SDValue Ops[] = { Op1, Op2, Op3 };
4580 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4583 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4584 EVT VT, const SDValue *Ops,
4586 SDVTList VTs = getVTList(VT);
4587 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4590 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4591 EVT VT1, EVT VT2, const SDValue *Ops,
4593 SDVTList VTs = getVTList(VT1, VT2);
4594 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4597 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4599 SDVTList VTs = getVTList(VT1, VT2);
4600 return MorphNodeTo(N, Opc, VTs, (SDValue *)0, 0);
4603 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4604 EVT VT1, EVT VT2, EVT VT3,
4605 const SDValue *Ops, unsigned NumOps) {
4606 SDVTList VTs = getVTList(VT1, VT2, VT3);
4607 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4610 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4613 SDVTList VTs = getVTList(VT1, VT2);
4614 SDValue Ops[] = { Op1 };
4615 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4618 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4620 SDValue Op1, SDValue Op2) {
4621 SDVTList VTs = getVTList(VT1, VT2);
4622 SDValue Ops[] = { Op1, Op2 };
4623 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4626 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4628 SDValue Op1, SDValue Op2,
4630 SDVTList VTs = getVTList(VT1, VT2);
4631 SDValue Ops[] = { Op1, Op2, Op3 };
4632 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4635 /// MorphNodeTo - These *mutate* the specified node to have the specified
4636 /// return type, opcode, and operands.
4638 /// Note that MorphNodeTo returns the resultant node. If there is already a
4639 /// node of the specified opcode and operands, it returns that node instead of
4640 /// the current one. Note that the DebugLoc need not be the same.
4642 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4643 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4644 /// node, and because it doesn't require CSE recalculation for any of
4645 /// the node's users.
4647 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4648 SDVTList VTs, const SDValue *Ops,
4650 // If an identical node already exists, use it.
4652 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4653 FoldingSetNodeID ID;
4654 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4655 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4659 if (!RemoveNodeFromCSEMaps(N))
4662 // Start the morphing.
4664 N->ValueList = VTs.VTs;
4665 N->NumValues = VTs.NumVTs;
4667 // Clear the operands list, updating used nodes to remove this from their
4668 // use list. Keep track of any operands that become dead as a result.
4669 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4670 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
4672 SDNode *Used = Use.getNode();
4674 if (Used->use_empty())
4675 DeadNodeSet.insert(Used);
4678 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) {
4679 // Initialize the memory references information.
4680 MN->setMemRefs(0, 0);
4681 // If NumOps is larger than the # of operands we can have in a
4682 // MachineSDNode, reallocate the operand list.
4683 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) {
4684 if (MN->OperandsNeedDelete)
4685 delete[] MN->OperandList;
4686 if (NumOps > array_lengthof(MN->LocalOperands))
4687 // We're creating a final node that will live unmorphed for the
4688 // remainder of the current SelectionDAG iteration, so we can allocate
4689 // the operands directly out of a pool with no recycling metadata.
4690 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4693 MN->InitOperands(MN->LocalOperands, Ops, NumOps);
4694 MN->OperandsNeedDelete = false;
4696 MN->InitOperands(MN->OperandList, Ops, NumOps);
4698 // If NumOps is larger than the # of operands we currently have, reallocate
4699 // the operand list.
4700 if (NumOps > N->NumOperands) {
4701 if (N->OperandsNeedDelete)
4702 delete[] N->OperandList;
4703 N->InitOperands(new SDUse[NumOps], Ops, NumOps);
4704 N->OperandsNeedDelete = true;
4706 N->InitOperands(N->OperandList, Ops, NumOps);
4709 // Delete any nodes that are still dead after adding the uses for the
4711 SmallVector<SDNode *, 16> DeadNodes;
4712 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4713 E = DeadNodeSet.end(); I != E; ++I)
4714 if ((*I)->use_empty())
4715 DeadNodes.push_back(*I);
4716 RemoveDeadNodes(DeadNodes);
4719 CSEMap.InsertNode(N, IP); // Memoize the new node.
4724 /// getMachineNode - These are used for target selectors to create a new node
4725 /// with specified return type(s), MachineInstr opcode, and operands.
4727 /// Note that getMachineNode returns the resultant node. If there is already a
4728 /// node of the specified opcode and operands, it returns that node instead of
4729 /// the current one.
4731 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) {
4732 SDVTList VTs = getVTList(VT);
4733 return getMachineNode(Opcode, dl, VTs, 0, 0);
4737 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) {
4738 SDVTList VTs = getVTList(VT);
4739 SDValue Ops[] = { Op1 };
4740 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4744 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4745 SDValue Op1, SDValue Op2) {
4746 SDVTList VTs = getVTList(VT);
4747 SDValue Ops[] = { Op1, Op2 };
4748 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4752 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4753 SDValue Op1, SDValue Op2, SDValue Op3) {
4754 SDVTList VTs = getVTList(VT);
4755 SDValue Ops[] = { Op1, Op2, Op3 };
4756 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4760 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4761 const SDValue *Ops, unsigned NumOps) {
4762 SDVTList VTs = getVTList(VT);
4763 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4767 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) {
4768 SDVTList VTs = getVTList(VT1, VT2);
4769 return getMachineNode(Opcode, dl, VTs, 0, 0);
4773 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4774 EVT VT1, EVT VT2, SDValue Op1) {
4775 SDVTList VTs = getVTList(VT1, VT2);
4776 SDValue Ops[] = { Op1 };
4777 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4781 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4782 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) {
4783 SDVTList VTs = getVTList(VT1, VT2);
4784 SDValue Ops[] = { Op1, Op2 };
4785 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4789 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4790 EVT VT1, EVT VT2, SDValue Op1,
4791 SDValue Op2, SDValue Op3) {
4792 SDVTList VTs = getVTList(VT1, VT2);
4793 SDValue Ops[] = { Op1, Op2, Op3 };
4794 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4798 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4800 const SDValue *Ops, unsigned NumOps) {
4801 SDVTList VTs = getVTList(VT1, VT2);
4802 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4806 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4807 EVT VT1, EVT VT2, EVT VT3,
4808 SDValue Op1, SDValue Op2) {
4809 SDVTList VTs = getVTList(VT1, VT2, VT3);
4810 SDValue Ops[] = { Op1, Op2 };
4811 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4815 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4816 EVT VT1, EVT VT2, EVT VT3,
4817 SDValue Op1, SDValue Op2, SDValue Op3) {
4818 SDVTList VTs = getVTList(VT1, VT2, VT3);
4819 SDValue Ops[] = { Op1, Op2, Op3 };
4820 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4824 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4825 EVT VT1, EVT VT2, EVT VT3,
4826 const SDValue *Ops, unsigned NumOps) {
4827 SDVTList VTs = getVTList(VT1, VT2, VT3);
4828 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4832 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1,
4833 EVT VT2, EVT VT3, EVT VT4,
4834 const SDValue *Ops, unsigned NumOps) {
4835 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4836 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4840 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4841 const std::vector<EVT> &ResultTys,
4842 const SDValue *Ops, unsigned NumOps) {
4843 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size());
4844 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4848 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs,
4849 const SDValue *Ops, unsigned NumOps) {
4850 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag;
4855 FoldingSetNodeID ID;
4856 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps);
4858 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4859 return cast<MachineSDNode>(E);
4862 // Allocate a new MachineSDNode.
4863 N = NodeAllocator.Allocate<MachineSDNode>();
4864 new (N) MachineSDNode(~Opcode, DL, VTs);
4866 // Initialize the operands list.
4867 if (NumOps > array_lengthof(N->LocalOperands))
4868 // We're creating a final node that will live unmorphed for the
4869 // remainder of the current SelectionDAG iteration, so we can allocate
4870 // the operands directly out of a pool with no recycling metadata.
4871 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4874 N->InitOperands(N->LocalOperands, Ops, NumOps);
4875 N->OperandsNeedDelete = false;
4878 CSEMap.InsertNode(N, IP);
4880 AllNodes.push_back(N);
4887 /// getTargetExtractSubreg - A convenience function for creating
4888 /// TargetOpcode::EXTRACT_SUBREG nodes.
4890 SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT,
4892 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4893 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
4894 VT, Operand, SRIdxVal);
4895 return SDValue(Subreg, 0);
4898 /// getTargetInsertSubreg - A convenience function for creating
4899 /// TargetOpcode::INSERT_SUBREG nodes.
4901 SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT,
4902 SDValue Operand, SDValue Subreg) {
4903 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4904 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
4905 VT, Operand, Subreg, SRIdxVal);
4906 return SDValue(Result, 0);
4909 /// getNodeIfExists - Get the specified node if it's already available, or
4910 /// else return NULL.
4911 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4912 const SDValue *Ops, unsigned NumOps) {
4913 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4914 FoldingSetNodeID ID;
4915 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4917 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4923 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4924 /// This can cause recursive merging of nodes in the DAG.
4926 /// This version assumes From has a single result value.
4928 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
4929 DAGUpdateListener *UpdateListener) {
4930 SDNode *From = FromN.getNode();
4931 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
4932 "Cannot replace with this method!");
4933 assert(From != To.getNode() && "Cannot replace uses of with self");
4935 // Iterate over all the existing uses of From. New uses will be added
4936 // to the beginning of the use list, which we avoid visiting.
4937 // This specifically avoids visiting uses of From that arise while the
4938 // replacement is happening, because any such uses would be the result
4939 // of CSE: If an existing node looks like From after one of its operands
4940 // is replaced by To, we don't want to replace of all its users with To
4941 // too. See PR3018 for more info.
4942 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4946 // This node is about to morph, remove its old self from the CSE maps.
4947 RemoveNodeFromCSEMaps(User);
4949 // A user can appear in a use list multiple times, and when this
4950 // happens the uses are usually next to each other in the list.
4951 // To help reduce the number of CSE recomputations, process all
4952 // the uses of this user that we can find this way.
4954 SDUse &Use = UI.getUse();
4957 } while (UI != UE && *UI == User);
4959 // Now that we have modified User, add it back to the CSE maps. If it
4960 // already exists there, recursively merge the results together.
4961 AddModifiedNodeToCSEMaps(User, UpdateListener);
4965 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4966 /// This can cause recursive merging of nodes in the DAG.
4968 /// This version assumes that for each value of From, there is a
4969 /// corresponding value in To in the same position with the same type.
4971 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
4972 DAGUpdateListener *UpdateListener) {
4974 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
4975 assert((!From->hasAnyUseOfValue(i) ||
4976 From->getValueType(i) == To->getValueType(i)) &&
4977 "Cannot use this version of ReplaceAllUsesWith!");
4980 // Handle the trivial case.
4984 // Iterate over just the existing users of From. See the comments in
4985 // the ReplaceAllUsesWith above.
4986 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4990 // This node is about to morph, remove its old self from the CSE maps.
4991 RemoveNodeFromCSEMaps(User);
4993 // A user can appear in a use list multiple times, and when this
4994 // happens the uses are usually next to each other in the list.
4995 // To help reduce the number of CSE recomputations, process all
4996 // the uses of this user that we can find this way.
4998 SDUse &Use = UI.getUse();
5001 } while (UI != UE && *UI == User);
5003 // Now that we have modified User, add it back to the CSE maps. If it
5004 // already exists there, recursively merge the results together.
5005 AddModifiedNodeToCSEMaps(User, UpdateListener);
5009 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5010 /// This can cause recursive merging of nodes in the DAG.
5012 /// This version can replace From with any result values. To must match the
5013 /// number and types of values returned by From.
5014 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
5016 DAGUpdateListener *UpdateListener) {
5017 if (From->getNumValues() == 1) // Handle the simple case efficiently.
5018 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
5020 // Iterate over just the existing users of From. See the comments in
5021 // the ReplaceAllUsesWith above.
5022 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5026 // This node is about to morph, remove its old self from the CSE maps.
5027 RemoveNodeFromCSEMaps(User);
5029 // A user can appear in a use list multiple times, and when this
5030 // happens the uses are usually next to each other in the list.
5031 // To help reduce the number of CSE recomputations, process all
5032 // the uses of this user that we can find this way.
5034 SDUse &Use = UI.getUse();
5035 const SDValue &ToOp = To[Use.getResNo()];
5038 } while (UI != UE && *UI == User);
5040 // Now that we have modified User, add it back to the CSE maps. If it
5041 // already exists there, recursively merge the results together.
5042 AddModifiedNodeToCSEMaps(User, UpdateListener);
5046 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
5047 /// uses of other values produced by From.getNode() alone. The Deleted
5048 /// vector is handled the same way as for ReplaceAllUsesWith.
5049 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
5050 DAGUpdateListener *UpdateListener){
5051 // Handle the really simple, really trivial case efficiently.
5052 if (From == To) return;
5054 // Handle the simple, trivial, case efficiently.
5055 if (From.getNode()->getNumValues() == 1) {
5056 ReplaceAllUsesWith(From, To, UpdateListener);
5060 // Iterate over just the existing users of From. See the comments in
5061 // the ReplaceAllUsesWith above.
5062 SDNode::use_iterator UI = From.getNode()->use_begin(),
5063 UE = From.getNode()->use_end();
5066 bool UserRemovedFromCSEMaps = false;
5068 // A user can appear in a use list multiple times, and when this
5069 // happens the uses are usually next to each other in the list.
5070 // To help reduce the number of CSE recomputations, process all
5071 // the uses of this user that we can find this way.
5073 SDUse &Use = UI.getUse();
5075 // Skip uses of different values from the same node.
5076 if (Use.getResNo() != From.getResNo()) {
5081 // If this node hasn't been modified yet, it's still in the CSE maps,
5082 // so remove its old self from the CSE maps.
5083 if (!UserRemovedFromCSEMaps) {
5084 RemoveNodeFromCSEMaps(User);
5085 UserRemovedFromCSEMaps = true;
5090 } while (UI != UE && *UI == User);
5092 // We are iterating over all uses of the From node, so if a use
5093 // doesn't use the specific value, no changes are made.
5094 if (!UserRemovedFromCSEMaps)
5097 // Now that we have modified User, add it back to the CSE maps. If it
5098 // already exists there, recursively merge the results together.
5099 AddModifiedNodeToCSEMaps(User, UpdateListener);
5104 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
5105 /// to record information about a use.
5112 /// operator< - Sort Memos by User.
5113 bool operator<(const UseMemo &L, const UseMemo &R) {
5114 return (intptr_t)L.User < (intptr_t)R.User;
5118 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
5119 /// uses of other values produced by From.getNode() alone. The same value
5120 /// may appear in both the From and To list. The Deleted vector is
5121 /// handled the same way as for ReplaceAllUsesWith.
5122 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
5125 DAGUpdateListener *UpdateListener){
5126 // Handle the simple, trivial case efficiently.
5128 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
5130 // Read up all the uses and make records of them. This helps
5131 // processing new uses that are introduced during the
5132 // replacement process.
5133 SmallVector<UseMemo, 4> Uses;
5134 for (unsigned i = 0; i != Num; ++i) {
5135 unsigned FromResNo = From[i].getResNo();
5136 SDNode *FromNode = From[i].getNode();
5137 for (SDNode::use_iterator UI = FromNode->use_begin(),
5138 E = FromNode->use_end(); UI != E; ++UI) {
5139 SDUse &Use = UI.getUse();
5140 if (Use.getResNo() == FromResNo) {
5141 UseMemo Memo = { *UI, i, &Use };
5142 Uses.push_back(Memo);
5147 // Sort the uses, so that all the uses from a given User are together.
5148 std::sort(Uses.begin(), Uses.end());
5150 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
5151 UseIndex != UseIndexEnd; ) {
5152 // We know that this user uses some value of From. If it is the right
5153 // value, update it.
5154 SDNode *User = Uses[UseIndex].User;
5156 // This node is about to morph, remove its old self from the CSE maps.
5157 RemoveNodeFromCSEMaps(User);
5159 // The Uses array is sorted, so all the uses for a given User
5160 // are next to each other in the list.
5161 // To help reduce the number of CSE recomputations, process all
5162 // the uses of this user that we can find this way.
5164 unsigned i = Uses[UseIndex].Index;
5165 SDUse &Use = *Uses[UseIndex].Use;
5169 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
5171 // Now that we have modified User, add it back to the CSE maps. If it
5172 // already exists there, recursively merge the results together.
5173 AddModifiedNodeToCSEMaps(User, UpdateListener);
5177 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
5178 /// based on their topological order. It returns the maximum id and a vector
5179 /// of the SDNodes* in assigned order by reference.
5180 unsigned SelectionDAG::AssignTopologicalOrder() {
5182 unsigned DAGSize = 0;
5184 // SortedPos tracks the progress of the algorithm. Nodes before it are
5185 // sorted, nodes after it are unsorted. When the algorithm completes
5186 // it is at the end of the list.
5187 allnodes_iterator SortedPos = allnodes_begin();
5189 // Visit all the nodes. Move nodes with no operands to the front of
5190 // the list immediately. Annotate nodes that do have operands with their
5191 // operand count. Before we do this, the Node Id fields of the nodes
5192 // may contain arbitrary values. After, the Node Id fields for nodes
5193 // before SortedPos will contain the topological sort index, and the
5194 // Node Id fields for nodes At SortedPos and after will contain the
5195 // count of outstanding operands.
5196 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
5199 unsigned Degree = N->getNumOperands();
5201 // A node with no uses, add it to the result array immediately.
5202 N->setNodeId(DAGSize++);
5203 allnodes_iterator Q = N;
5205 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
5206 assert(SortedPos != AllNodes.end() && "Overran node list");
5209 // Temporarily use the Node Id as scratch space for the degree count.
5210 N->setNodeId(Degree);
5214 // Visit all the nodes. As we iterate, moves nodes into sorted order,
5215 // such that by the time the end is reached all nodes will be sorted.
5216 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
5219 // N is in sorted position, so all its uses have one less operand
5220 // that needs to be sorted.
5221 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5224 unsigned Degree = P->getNodeId();
5225 assert(Degree != 0 && "Invalid node degree");
5228 // All of P's operands are sorted, so P may sorted now.
5229 P->setNodeId(DAGSize++);
5231 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
5232 assert(SortedPos != AllNodes.end() && "Overran node list");
5235 // Update P's outstanding operand count.
5236 P->setNodeId(Degree);
5239 if (I == SortedPos) {
5242 dbgs() << "Overran sorted position:\n";
5245 llvm_unreachable(0);
5249 assert(SortedPos == AllNodes.end() &&
5250 "Topological sort incomplete!");
5251 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
5252 "First node in topological sort is not the entry token!");
5253 assert(AllNodes.front().getNodeId() == 0 &&
5254 "First node in topological sort has non-zero id!");
5255 assert(AllNodes.front().getNumOperands() == 0 &&
5256 "First node in topological sort has operands!");
5257 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
5258 "Last node in topologic sort has unexpected id!");
5259 assert(AllNodes.back().use_empty() &&
5260 "Last node in topologic sort has users!");
5261 assert(DAGSize == allnodes_size() && "Node count mismatch!");
5265 /// AssignOrdering - Assign an order to the SDNode.
5266 void SelectionDAG::AssignOrdering(const SDNode *SD, unsigned Order) {
5267 assert(SD && "Trying to assign an order to a null node!");
5268 Ordering->add(SD, Order);
5271 /// GetOrdering - Get the order for the SDNode.
5272 unsigned SelectionDAG::GetOrdering(const SDNode *SD) const {
5273 assert(SD && "Trying to get the order of a null node!");
5274 return Ordering->getOrder(SD);
5278 //===----------------------------------------------------------------------===//
5280 //===----------------------------------------------------------------------===//
5282 HandleSDNode::~HandleSDNode() {
5286 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA,
5287 EVT VT, int64_t o, unsigned char TF)
5288 : SDNode(Opc, DebugLoc::getUnknownLoc(), getSDVTList(VT)),
5289 Offset(o), TargetFlags(TF) {
5290 TheGlobal = const_cast<GlobalValue*>(GA);
5293 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt,
5294 MachineMemOperand *mmo)
5295 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) {
5296 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile());
5297 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5298 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5301 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs,
5302 const SDValue *Ops, unsigned NumOps, EVT memvt,
5303 MachineMemOperand *mmo)
5304 : SDNode(Opc, dl, VTs, Ops, NumOps),
5305 MemoryVT(memvt), MMO(mmo) {
5306 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile());
5307 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5308 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5311 /// Profile - Gather unique data for the node.
5313 void SDNode::Profile(FoldingSetNodeID &ID) const {
5314 AddNodeIDNode(ID, this);
5319 std::vector<EVT> VTs;
5322 VTs.reserve(MVT::LAST_VALUETYPE);
5323 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i)
5324 VTs.push_back(MVT((MVT::SimpleValueType)i));
5329 static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs;
5330 static ManagedStatic<EVTArray> SimpleVTArray;
5331 static ManagedStatic<sys::SmartMutex<true> > VTMutex;
5333 /// getValueTypeList - Return a pointer to the specified value type.
5335 const EVT *SDNode::getValueTypeList(EVT VT) {
5336 if (VT.isExtended()) {
5337 sys::SmartScopedLock<true> Lock(*VTMutex);
5338 return &(*EVTs->insert(VT).first);
5340 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy];
5344 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
5345 /// indicated value. This method ignores uses of other values defined by this
5347 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
5348 assert(Value < getNumValues() && "Bad value!");
5350 // TODO: Only iterate over uses of a given value of the node
5351 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
5352 if (UI.getUse().getResNo() == Value) {
5359 // Found exactly the right number of uses?
5364 /// hasAnyUseOfValue - Return true if there are any use of the indicated
5365 /// value. This method ignores uses of other values defined by this operation.
5366 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
5367 assert(Value < getNumValues() && "Bad value!");
5369 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
5370 if (UI.getUse().getResNo() == Value)
5377 /// isOnlyUserOf - Return true if this node is the only use of N.
5379 bool SDNode::isOnlyUserOf(SDNode *N) const {
5381 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
5392 /// isOperand - Return true if this node is an operand of N.
5394 bool SDValue::isOperandOf(SDNode *N) const {
5395 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5396 if (*this == N->getOperand(i))
5401 bool SDNode::isOperandOf(SDNode *N) const {
5402 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
5403 if (this == N->OperandList[i].getNode())
5408 /// reachesChainWithoutSideEffects - Return true if this operand (which must
5409 /// be a chain) reaches the specified operand without crossing any
5410 /// side-effecting instructions. In practice, this looks through token
5411 /// factors and non-volatile loads. In order to remain efficient, this only
5412 /// looks a couple of nodes in, it does not do an exhaustive search.
5413 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
5414 unsigned Depth) const {
5415 if (*this == Dest) return true;
5417 // Don't search too deeply, we just want to be able to see through
5418 // TokenFactor's etc.
5419 if (Depth == 0) return false;
5421 // If this is a token factor, all inputs to the TF happen in parallel. If any
5422 // of the operands of the TF reach dest, then we can do the xform.
5423 if (getOpcode() == ISD::TokenFactor) {
5424 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5425 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
5430 // Loads don't have side effects, look through them.
5431 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
5432 if (!Ld->isVolatile())
5433 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
5438 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
5439 /// is either an operand of N or it can be reached by traversing up the operands.
5440 /// NOTE: this is an expensive method. Use it carefully.
5441 bool SDNode::isPredecessorOf(SDNode *N) const {
5442 SmallPtrSet<SDNode *, 32> Visited;
5443 SmallVector<SDNode *, 16> Worklist;
5444 Worklist.push_back(N);
5447 N = Worklist.pop_back_val();
5448 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5449 SDNode *Op = N->getOperand(i).getNode();
5452 if (Visited.insert(Op))
5453 Worklist.push_back(Op);
5455 } while (!Worklist.empty());
5460 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
5461 assert(Num < NumOperands && "Invalid child # of SDNode!");
5462 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
5465 std::string SDNode::getOperationName(const SelectionDAG *G) const {
5466 switch (getOpcode()) {
5468 if (getOpcode() < ISD::BUILTIN_OP_END)
5469 return "<<Unknown DAG Node>>";
5470 if (isMachineOpcode()) {
5472 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
5473 if (getMachineOpcode() < TII->getNumOpcodes())
5474 return TII->get(getMachineOpcode()).getName();
5475 return "<<Unknown Machine Node>>";
5478 const TargetLowering &TLI = G->getTargetLoweringInfo();
5479 const char *Name = TLI.getTargetNodeName(getOpcode());
5480 if (Name) return Name;
5481 return "<<Unknown Target Node>>";
5483 return "<<Unknown Node>>";
5486 case ISD::DELETED_NODE:
5487 return "<<Deleted Node!>>";
5489 case ISD::PREFETCH: return "Prefetch";
5490 case ISD::MEMBARRIER: return "MemBarrier";
5491 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap";
5492 case ISD::ATOMIC_SWAP: return "AtomicSwap";
5493 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd";
5494 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub";
5495 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd";
5496 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr";
5497 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor";
5498 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand";
5499 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin";
5500 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax";
5501 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin";
5502 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax";
5503 case ISD::PCMARKER: return "PCMarker";
5504 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5505 case ISD::SRCVALUE: return "SrcValue";
5506 case ISD::EntryToken: return "EntryToken";
5507 case ISD::TokenFactor: return "TokenFactor";
5508 case ISD::AssertSext: return "AssertSext";
5509 case ISD::AssertZext: return "AssertZext";
5511 case ISD::BasicBlock: return "BasicBlock";
5512 case ISD::VALUETYPE: return "ValueType";
5513 case ISD::Register: return "Register";
5515 case ISD::Constant: return "Constant";
5516 case ISD::ConstantFP: return "ConstantFP";
5517 case ISD::GlobalAddress: return "GlobalAddress";
5518 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5519 case ISD::FrameIndex: return "FrameIndex";
5520 case ISD::JumpTable: return "JumpTable";
5521 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5522 case ISD::RETURNADDR: return "RETURNADDR";
5523 case ISD::FRAMEADDR: return "FRAMEADDR";
5524 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5525 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5526 case ISD::LSDAADDR: return "LSDAADDR";
5527 case ISD::EHSELECTION: return "EHSELECTION";
5528 case ISD::EH_RETURN: return "EH_RETURN";
5529 case ISD::ConstantPool: return "ConstantPool";
5530 case ISD::ExternalSymbol: return "ExternalSymbol";
5531 case ISD::BlockAddress: return "BlockAddress";
5532 case ISD::INTRINSIC_WO_CHAIN:
5533 case ISD::INTRINSIC_VOID:
5534 case ISD::INTRINSIC_W_CHAIN: {
5535 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1;
5536 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue();
5537 if (IID < Intrinsic::num_intrinsics)
5538 return Intrinsic::getName((Intrinsic::ID)IID);
5539 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo())
5540 return TII->getName(IID);
5541 llvm_unreachable("Invalid intrinsic ID");
5544 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5545 case ISD::TargetConstant: return "TargetConstant";
5546 case ISD::TargetConstantFP:return "TargetConstantFP";
5547 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5548 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5549 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5550 case ISD::TargetJumpTable: return "TargetJumpTable";
5551 case ISD::TargetConstantPool: return "TargetConstantPool";
5552 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5553 case ISD::TargetBlockAddress: return "TargetBlockAddress";
5555 case ISD::CopyToReg: return "CopyToReg";
5556 case ISD::CopyFromReg: return "CopyFromReg";
5557 case ISD::UNDEF: return "undef";
5558 case ISD::MERGE_VALUES: return "merge_values";
5559 case ISD::INLINEASM: return "inlineasm";
5560 case ISD::EH_LABEL: return "eh_label";
5561 case ISD::HANDLENODE: return "handlenode";
5564 case ISD::FABS: return "fabs";
5565 case ISD::FNEG: return "fneg";
5566 case ISD::FSQRT: return "fsqrt";
5567 case ISD::FSIN: return "fsin";
5568 case ISD::FCOS: return "fcos";
5569 case ISD::FPOWI: return "fpowi";
5570 case ISD::FPOW: return "fpow";
5571 case ISD::FTRUNC: return "ftrunc";
5572 case ISD::FFLOOR: return "ffloor";
5573 case ISD::FCEIL: return "fceil";
5574 case ISD::FRINT: return "frint";
5575 case ISD::FNEARBYINT: return "fnearbyint";
5578 case ISD::ADD: return "add";
5579 case ISD::SUB: return "sub";
5580 case ISD::MUL: return "mul";
5581 case ISD::MULHU: return "mulhu";
5582 case ISD::MULHS: return "mulhs";
5583 case ISD::SDIV: return "sdiv";
5584 case ISD::UDIV: return "udiv";
5585 case ISD::SREM: return "srem";
5586 case ISD::UREM: return "urem";
5587 case ISD::SMUL_LOHI: return "smul_lohi";
5588 case ISD::UMUL_LOHI: return "umul_lohi";
5589 case ISD::SDIVREM: return "sdivrem";
5590 case ISD::UDIVREM: return "udivrem";
5591 case ISD::AND: return "and";
5592 case ISD::OR: return "or";
5593 case ISD::XOR: return "xor";
5594 case ISD::SHL: return "shl";
5595 case ISD::SRA: return "sra";
5596 case ISD::SRL: return "srl";
5597 case ISD::ROTL: return "rotl";
5598 case ISD::ROTR: return "rotr";
5599 case ISD::FADD: return "fadd";
5600 case ISD::FSUB: return "fsub";
5601 case ISD::FMUL: return "fmul";
5602 case ISD::FDIV: return "fdiv";
5603 case ISD::FREM: return "frem";
5604 case ISD::FCOPYSIGN: return "fcopysign";
5605 case ISD::FGETSIGN: return "fgetsign";
5607 case ISD::SETCC: return "setcc";
5608 case ISD::VSETCC: return "vsetcc";
5609 case ISD::SELECT: return "select";
5610 case ISD::SELECT_CC: return "select_cc";
5611 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5612 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5613 case ISD::CONCAT_VECTORS: return "concat_vectors";
5614 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5615 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5616 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5617 case ISD::CARRY_FALSE: return "carry_false";
5618 case ISD::ADDC: return "addc";
5619 case ISD::ADDE: return "adde";
5620 case ISD::SADDO: return "saddo";
5621 case ISD::UADDO: return "uaddo";
5622 case ISD::SSUBO: return "ssubo";
5623 case ISD::USUBO: return "usubo";
5624 case ISD::SMULO: return "smulo";
5625 case ISD::UMULO: return "umulo";
5626 case ISD::SUBC: return "subc";
5627 case ISD::SUBE: return "sube";
5628 case ISD::SHL_PARTS: return "shl_parts";
5629 case ISD::SRA_PARTS: return "sra_parts";
5630 case ISD::SRL_PARTS: return "srl_parts";
5632 // Conversion operators.
5633 case ISD::SIGN_EXTEND: return "sign_extend";
5634 case ISD::ZERO_EXTEND: return "zero_extend";
5635 case ISD::ANY_EXTEND: return "any_extend";
5636 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5637 case ISD::TRUNCATE: return "truncate";
5638 case ISD::FP_ROUND: return "fp_round";
5639 case ISD::FLT_ROUNDS_: return "flt_rounds";
5640 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5641 case ISD::FP_EXTEND: return "fp_extend";
5643 case ISD::SINT_TO_FP: return "sint_to_fp";
5644 case ISD::UINT_TO_FP: return "uint_to_fp";
5645 case ISD::FP_TO_SINT: return "fp_to_sint";
5646 case ISD::FP_TO_UINT: return "fp_to_uint";
5647 case ISD::BIT_CONVERT: return "bit_convert";
5649 case ISD::CONVERT_RNDSAT: {
5650 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) {
5651 default: llvm_unreachable("Unknown cvt code!");
5652 case ISD::CVT_FF: return "cvt_ff";
5653 case ISD::CVT_FS: return "cvt_fs";
5654 case ISD::CVT_FU: return "cvt_fu";
5655 case ISD::CVT_SF: return "cvt_sf";
5656 case ISD::CVT_UF: return "cvt_uf";
5657 case ISD::CVT_SS: return "cvt_ss";
5658 case ISD::CVT_SU: return "cvt_su";
5659 case ISD::CVT_US: return "cvt_us";
5660 case ISD::CVT_UU: return "cvt_uu";
5664 // Control flow instructions
5665 case ISD::BR: return "br";
5666 case ISD::BRIND: return "brind";
5667 case ISD::BR_JT: return "br_jt";
5668 case ISD::BRCOND: return "brcond";
5669 case ISD::BR_CC: return "br_cc";
5670 case ISD::CALLSEQ_START: return "callseq_start";
5671 case ISD::CALLSEQ_END: return "callseq_end";
5674 case ISD::LOAD: return "load";
5675 case ISD::STORE: return "store";
5676 case ISD::VAARG: return "vaarg";
5677 case ISD::VACOPY: return "vacopy";
5678 case ISD::VAEND: return "vaend";
5679 case ISD::VASTART: return "vastart";
5680 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5681 case ISD::EXTRACT_ELEMENT: return "extract_element";
5682 case ISD::BUILD_PAIR: return "build_pair";
5683 case ISD::STACKSAVE: return "stacksave";
5684 case ISD::STACKRESTORE: return "stackrestore";
5685 case ISD::TRAP: return "trap";
5688 case ISD::BSWAP: return "bswap";
5689 case ISD::CTPOP: return "ctpop";
5690 case ISD::CTTZ: return "cttz";
5691 case ISD::CTLZ: return "ctlz";
5694 case ISD::TRAMPOLINE: return "trampoline";
5697 switch (cast<CondCodeSDNode>(this)->get()) {
5698 default: llvm_unreachable("Unknown setcc condition!");
5699 case ISD::SETOEQ: return "setoeq";
5700 case ISD::SETOGT: return "setogt";
5701 case ISD::SETOGE: return "setoge";
5702 case ISD::SETOLT: return "setolt";
5703 case ISD::SETOLE: return "setole";
5704 case ISD::SETONE: return "setone";
5706 case ISD::SETO: return "seto";
5707 case ISD::SETUO: return "setuo";
5708 case ISD::SETUEQ: return "setue";
5709 case ISD::SETUGT: return "setugt";
5710 case ISD::SETUGE: return "setuge";
5711 case ISD::SETULT: return "setult";
5712 case ISD::SETULE: return "setule";
5713 case ISD::SETUNE: return "setune";
5715 case ISD::SETEQ: return "seteq";
5716 case ISD::SETGT: return "setgt";
5717 case ISD::SETGE: return "setge";
5718 case ISD::SETLT: return "setlt";
5719 case ISD::SETLE: return "setle";
5720 case ISD::SETNE: return "setne";
5725 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5734 return "<post-inc>";
5736 return "<post-dec>";
5740 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5741 std::string S = "< ";
5755 if (getByValAlign())
5756 S += "byval-align:" + utostr(getByValAlign()) + " ";
5758 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5760 S += "byval-size:" + utostr(getByValSize()) + " ";
5764 void SDNode::dump() const { dump(0); }
5765 void SDNode::dump(const SelectionDAG *G) const {
5769 void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const {
5770 OS << (void*)this << ": ";
5772 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5774 if (getValueType(i) == MVT::Other)
5777 OS << getValueType(i).getEVTString();
5779 OS << " = " << getOperationName(G);
5782 void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const {
5783 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) {
5784 if (!MN->memoperands_empty()) {
5787 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(),
5788 e = MN->memoperands_end(); i != e; ++i) {
5795 } else if (const ShuffleVectorSDNode *SVN =
5796 dyn_cast<ShuffleVectorSDNode>(this)) {
5798 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) {
5799 int Idx = SVN->getMaskElt(i);
5807 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5808 OS << '<' << CSDN->getAPIntValue() << '>';
5809 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5810 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5811 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5812 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5813 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5816 CSDN->getValueAPF().bitcastToAPInt().dump();
5819 } else if (const GlobalAddressSDNode *GADN =
5820 dyn_cast<GlobalAddressSDNode>(this)) {
5821 int64_t offset = GADN->getOffset();
5823 WriteAsOperand(OS, GADN->getGlobal());
5826 OS << " + " << offset;
5828 OS << " " << offset;
5829 if (unsigned int TF = GADN->getTargetFlags())
5830 OS << " [TF=" << TF << ']';
5831 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5832 OS << "<" << FIDN->getIndex() << ">";
5833 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5834 OS << "<" << JTDN->getIndex() << ">";
5835 if (unsigned int TF = JTDN->getTargetFlags())
5836 OS << " [TF=" << TF << ']';
5837 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5838 int offset = CP->getOffset();
5839 if (CP->isMachineConstantPoolEntry())
5840 OS << "<" << *CP->getMachineCPVal() << ">";
5842 OS << "<" << *CP->getConstVal() << ">";
5844 OS << " + " << offset;
5846 OS << " " << offset;
5847 if (unsigned int TF = CP->getTargetFlags())
5848 OS << " [TF=" << TF << ']';
5849 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5851 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5853 OS << LBB->getName() << " ";
5854 OS << (const void*)BBDN->getBasicBlock() << ">";
5855 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5856 if (G && R->getReg() &&
5857 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5858 OS << " %" << G->getTarget().getRegisterInfo()->getName(R->getReg());
5860 OS << " %reg" << R->getReg();
5862 } else if (const ExternalSymbolSDNode *ES =
5863 dyn_cast<ExternalSymbolSDNode>(this)) {
5864 OS << "'" << ES->getSymbol() << "'";
5865 if (unsigned int TF = ES->getTargetFlags())
5866 OS << " [TF=" << TF << ']';
5867 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5869 OS << "<" << M->getValue() << ">";
5872 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5873 OS << ":" << N->getVT().getEVTString();
5875 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5876 OS << "<" << *LD->getMemOperand();
5879 switch (LD->getExtensionType()) {
5880 default: doExt = false; break;
5881 case ISD::EXTLOAD: OS << ", anyext"; break;
5882 case ISD::SEXTLOAD: OS << ", sext"; break;
5883 case ISD::ZEXTLOAD: OS << ", zext"; break;
5886 OS << " from " << LD->getMemoryVT().getEVTString();
5888 const char *AM = getIndexedModeName(LD->getAddressingMode());
5893 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
5894 OS << "<" << *ST->getMemOperand();
5896 if (ST->isTruncatingStore())
5897 OS << ", trunc to " << ST->getMemoryVT().getEVTString();
5899 const char *AM = getIndexedModeName(ST->getAddressingMode());
5904 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) {
5905 OS << "<" << *M->getMemOperand() << ">";
5906 } else if (const BlockAddressSDNode *BA =
5907 dyn_cast<BlockAddressSDNode>(this)) {
5909 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false);
5911 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false);
5913 if (unsigned int TF = BA->getTargetFlags())
5914 OS << " [TF=" << TF << ']';
5918 if (unsigned Order = G->GetOrdering(this))
5919 OS << " [ORD=" << Order << ']';
5922 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
5924 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
5925 if (i) OS << ", "; else OS << " ";
5926 OS << (void*)getOperand(i).getNode();
5927 if (unsigned RN = getOperand(i).getResNo())
5930 print_details(OS, G);
5933 static void printrWithDepthHelper(raw_ostream &OS, const SDNode *N,
5934 const SelectionDAG *G, unsigned depth,
5947 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5949 printrWithDepthHelper(OS, N->getOperand(i).getNode(), G, depth-1, indent+2);
5953 void SDNode::printrWithDepth(raw_ostream &OS, const SelectionDAG *G,
5954 unsigned depth) const {
5955 printrWithDepthHelper(OS, this, G, depth, 0);
5958 void SDNode::printrFull(raw_ostream &OS, const SelectionDAG *G) const {
5959 // Don't print impossibly deep things.
5960 printrWithDepth(OS, G, 100);
5963 void SDNode::dumprWithDepth(const SelectionDAG *G, unsigned depth) const {
5964 printrWithDepth(dbgs(), G, depth);
5967 void SDNode::dumprFull(const SelectionDAG *G) const {
5968 // Don't print impossibly deep things.
5969 dumprWithDepth(G, 100);
5972 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
5973 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5974 if (N->getOperand(i).getNode()->hasOneUse())
5975 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
5977 dbgs() << "\n" << std::string(indent+2, ' ')
5978 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
5982 dbgs().indent(indent);
5986 SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) {
5987 assert(N->getNumValues() == 1 &&
5988 "Can't unroll a vector with multiple results!");
5990 EVT VT = N->getValueType(0);
5991 unsigned NE = VT.getVectorNumElements();
5992 EVT EltVT = VT.getVectorElementType();
5993 DebugLoc dl = N->getDebugLoc();
5995 SmallVector<SDValue, 8> Scalars;
5996 SmallVector<SDValue, 4> Operands(N->getNumOperands());
5998 // If ResNE is 0, fully unroll the vector op.
6001 else if (NE > ResNE)
6005 for (i= 0; i != NE; ++i) {
6006 for (unsigned j = 0; j != N->getNumOperands(); ++j) {
6007 SDValue Operand = N->getOperand(j);
6008 EVT OperandVT = Operand.getValueType();
6009 if (OperandVT.isVector()) {
6010 // A vector operand; extract a single element.
6011 EVT OperandEltVT = OperandVT.getVectorElementType();
6012 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl,
6015 getConstant(i, MVT::i32));
6017 // A scalar operand; just use it as is.
6018 Operands[j] = Operand;
6022 switch (N->getOpcode()) {
6024 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
6025 &Operands[0], Operands.size()));
6032 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0],
6033 getShiftAmountOperand(Operands[1])));
6035 case ISD::SIGN_EXTEND_INREG:
6036 case ISD::FP_ROUND_INREG: {
6037 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType();
6038 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
6040 getValueType(ExtVT)));
6045 for (; i < ResNE; ++i)
6046 Scalars.push_back(getUNDEF(EltVT));
6048 return getNode(ISD::BUILD_VECTOR, dl,
6049 EVT::getVectorVT(*getContext(), EltVT, ResNE),
6050 &Scalars[0], Scalars.size());
6054 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
6055 /// location that is 'Dist' units away from the location that the 'Base' load
6056 /// is loading from.
6057 bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
6058 unsigned Bytes, int Dist) const {
6059 if (LD->getChain() != Base->getChain())
6061 EVT VT = LD->getValueType(0);
6062 if (VT.getSizeInBits() / 8 != Bytes)
6065 SDValue Loc = LD->getOperand(1);
6066 SDValue BaseLoc = Base->getOperand(1);
6067 if (Loc.getOpcode() == ISD::FrameIndex) {
6068 if (BaseLoc.getOpcode() != ISD::FrameIndex)
6070 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo();
6071 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
6072 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
6073 int FS = MFI->getObjectSize(FI);
6074 int BFS = MFI->getObjectSize(BFI);
6075 if (FS != BFS || FS != (int)Bytes) return false;
6076 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
6078 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
6079 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
6080 if (V && (V->getSExtValue() == Dist*Bytes))
6084 GlobalValue *GV1 = NULL;
6085 GlobalValue *GV2 = NULL;
6086 int64_t Offset1 = 0;
6087 int64_t Offset2 = 0;
6088 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
6089 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
6090 if (isGA1 && isGA2 && GV1 == GV2)
6091 return Offset1 == (Offset2 + Dist*Bytes);
6096 /// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if
6097 /// it cannot be inferred.
6098 unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const {
6099 // If this is a GlobalAddress + cst, return the alignment.
6101 int64_t GVOffset = 0;
6102 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset))
6103 return MinAlign(GV->getAlignment(), GVOffset);
6105 // If this is a direct reference to a stack slot, use information about the
6106 // stack slot's alignment.
6107 int FrameIdx = 1 << 31;
6108 int64_t FrameOffset = 0;
6109 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
6110 FrameIdx = FI->getIndex();
6111 } else if (Ptr.getOpcode() == ISD::ADD &&
6112 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
6113 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
6114 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
6115 FrameOffset = Ptr.getConstantOperandVal(1);
6118 if (FrameIdx != (1 << 31)) {
6119 // FIXME: Handle FI+CST.
6120 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo();
6121 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
6123 if (MFI.isFixedObjectIndex(FrameIdx)) {
6124 int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx) + FrameOffset;
6126 // The alignment of the frame index can be determined from its offset from
6127 // the incoming frame position. If the frame object is at offset 32 and
6128 // the stack is guaranteed to be 16-byte aligned, then we know that the
6129 // object is 16-byte aligned.
6130 unsigned StackAlign = getTarget().getFrameInfo()->getStackAlignment();
6131 unsigned Align = MinAlign(ObjectOffset, StackAlign);
6133 // Finally, the frame object itself may have a known alignment. Factor
6134 // the alignment + offset into a new alignment. For example, if we know
6135 // the FI is 8 byte aligned, but the pointer is 4 off, we really have a
6136 // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte
6137 // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc.
6138 return std::max(Align, FIInfoAlign);
6146 void SelectionDAG::dump() const {
6147 dbgs() << "SelectionDAG has " << AllNodes.size() << " nodes:";
6149 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
6151 const SDNode *N = I;
6152 if (!N->hasOneUse() && N != getRoot().getNode())
6153 DumpNodes(N, 2, this);
6156 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
6161 void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
6163 print_details(OS, G);
6166 typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet;
6167 static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent,
6168 const SelectionDAG *G, VisitedSDNodeSet &once) {
6169 if (!once.insert(N)) // If we've been here before, return now.
6172 // Dump the current SDNode, but don't end the line yet.
6173 OS << std::string(indent, ' ');
6176 // Having printed this SDNode, walk the children:
6177 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6178 const SDNode *child = N->getOperand(i).getNode();
6183 if (child->getNumOperands() == 0) {
6184 // This child has no grandchildren; print it inline right here.
6185 child->printr(OS, G);
6187 } else { // Just the address. FIXME: also print the child's opcode.
6189 if (unsigned RN = N->getOperand(i).getResNo())
6196 // Dump children that have grandchildren on their own line(s).
6197 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6198 const SDNode *child = N->getOperand(i).getNode();
6199 DumpNodesr(OS, child, indent+2, G, once);
6203 void SDNode::dumpr() const {
6204 VisitedSDNodeSet once;
6205 DumpNodesr(dbgs(), this, 0, 0, once);
6208 void SDNode::dumpr(const SelectionDAG *G) const {
6209 VisitedSDNodeSet once;
6210 DumpNodesr(dbgs(), this, 0, G, once);
6214 // getAddressSpace - Return the address space this GlobalAddress belongs to.
6215 unsigned GlobalAddressSDNode::getAddressSpace() const {
6216 return getGlobal()->getType()->getAddressSpace();
6220 const Type *ConstantPoolSDNode::getType() const {
6221 if (isMachineConstantPoolEntry())
6222 return Val.MachineCPVal->getType();
6223 return Val.ConstVal->getType();
6226 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue,
6228 unsigned &SplatBitSize,
6230 unsigned MinSplatBits,
6232 EVT VT = getValueType(0);
6233 assert(VT.isVector() && "Expected a vector type");
6234 unsigned sz = VT.getSizeInBits();
6235 if (MinSplatBits > sz)
6238 SplatValue = APInt(sz, 0);
6239 SplatUndef = APInt(sz, 0);
6241 // Get the bits. Bits with undefined values (when the corresponding element
6242 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
6243 // in SplatValue. If any of the values are not constant, give up and return
6245 unsigned int nOps = getNumOperands();
6246 assert(nOps > 0 && "isConstantSplat has 0-size build vector");
6247 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits();
6249 for (unsigned j = 0; j < nOps; ++j) {
6250 unsigned i = isBigEndian ? nOps-1-j : j;
6251 SDValue OpVal = getOperand(i);
6252 unsigned BitPos = j * EltBitSize;
6254 if (OpVal.getOpcode() == ISD::UNDEF)
6255 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize);
6256 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal))
6257 SplatValue |= (APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize).
6258 zextOrTrunc(sz) << BitPos);
6259 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal))
6260 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos;
6265 // The build_vector is all constants or undefs. Find the smallest element
6266 // size that splats the vector.
6268 HasAnyUndefs = (SplatUndef != 0);
6271 unsigned HalfSize = sz / 2;
6272 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize);
6273 APInt LowValue = APInt(SplatValue).trunc(HalfSize);
6274 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize);
6275 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize);
6277 // If the two halves do not match (ignoring undef bits), stop here.
6278 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
6279 MinSplatBits > HalfSize)
6282 SplatValue = HighValue | LowValue;
6283 SplatUndef = HighUndef & LowUndef;
6292 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) {
6293 // Find the first non-undef value in the shuffle mask.
6295 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
6298 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!");
6300 // Make sure all remaining elements are either undef or the same as the first
6302 for (int Idx = Mask[i]; i != e; ++i)
6303 if (Mask[i] >= 0 && Mask[i] != Idx)
6308 static void checkForCyclesHelper(const SDNode *N,
6309 std::set<const SDNode *> &visited) {
6310 if (visited.find(N) != visited.end()) {
6311 dbgs() << "Offending node:\n";
6313 assert(0 && "Detected cycle in SelectionDAG");
6316 std::set<const SDNode*>::iterator i;
6319 tie(i, inserted) = visited.insert(N);
6320 assert(inserted && "Missed cycle");
6322 for(unsigned i = 0; i < N->getNumOperands(); ++i) {
6323 checkForCyclesHelper(N->getOperand(i).getNode(), visited);
6328 void llvm::checkForCycles(const llvm::SDNode *N) {
6330 assert(N && "Checking nonexistant SDNode");
6331 std::set<const SDNode *> visited;
6332 checkForCyclesHelper(N, visited);
6336 void llvm::checkForCycles(const llvm::SelectionDAG *DAG) {
6337 checkForCycles(DAG->getRoot().getNode());