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 "SDNodeDbgValue.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Analysis/DebugInfo.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/Function.h"
21 #include "llvm/GlobalAlias.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Assembly/Writer.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/CodeGen/MachineBasicBlock.h"
28 #include "llvm/CodeGen/MachineConstantPool.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/PseudoSourceValue.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Target/TargetFrameInfo.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetSelectionDAGInfo.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetInstrInfo.h"
39 #include "llvm/Target/TargetIntrinsicInfo.h"
40 #include "llvm/Target/TargetMachine.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/ManagedStatic.h"
45 #include "llvm/Support/MathExtras.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/System/Mutex.h"
48 #include "llvm/ADT/SetVector.h"
49 #include "llvm/ADT/SmallPtrSet.h"
50 #include "llvm/ADT/SmallSet.h"
51 #include "llvm/ADT/SmallVector.h"
52 #include "llvm/ADT/StringExtras.h"
57 /// makeVTList - Return an instance of the SDVTList struct initialized with the
58 /// specified members.
59 static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
60 SDVTList Res = {VTs, NumVTs};
64 static const fltSemantics *EVTToAPFloatSemantics(EVT VT) {
65 switch (VT.getSimpleVT().SimpleTy) {
66 default: llvm_unreachable("Unknown FP format");
67 case MVT::f32: return &APFloat::IEEEsingle;
68 case MVT::f64: return &APFloat::IEEEdouble;
69 case MVT::f80: return &APFloat::x87DoubleExtended;
70 case MVT::f128: return &APFloat::IEEEquad;
71 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
75 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
77 //===----------------------------------------------------------------------===//
78 // ConstantFPSDNode Class
79 //===----------------------------------------------------------------------===//
81 /// isExactlyValue - We don't rely on operator== working on double values, as
82 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
83 /// As such, this method can be used to do an exact bit-for-bit comparison of
84 /// two floating point values.
85 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
86 return getValueAPF().bitwiseIsEqual(V);
89 bool ConstantFPSDNode::isValueValidForType(EVT VT,
91 assert(VT.isFloatingPoint() && "Can only convert between FP types");
93 // PPC long double cannot be converted to any other type.
94 if (VT == MVT::ppcf128 ||
95 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
98 // convert modifies in place, so make a copy.
99 APFloat Val2 = APFloat(Val);
101 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
106 //===----------------------------------------------------------------------===//
108 //===----------------------------------------------------------------------===//
110 /// isBuildVectorAllOnes - Return true if the specified node is a
111 /// BUILD_VECTOR where all of the elements are ~0 or undef.
112 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
113 // Look through a bit convert.
114 if (N->getOpcode() == ISD::BIT_CONVERT)
115 N = N->getOperand(0).getNode();
117 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
119 unsigned i = 0, e = N->getNumOperands();
121 // Skip over all of the undef values.
122 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
125 // Do not accept an all-undef vector.
126 if (i == e) return false;
128 // Do not accept build_vectors that aren't all constants or which have non-~0
130 SDValue NotZero = N->getOperand(i);
131 if (isa<ConstantSDNode>(NotZero)) {
132 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
134 } else if (isa<ConstantFPSDNode>(NotZero)) {
135 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
136 bitcastToAPInt().isAllOnesValue())
141 // Okay, we have at least one ~0 value, check to see if the rest match or are
143 for (++i; i != e; ++i)
144 if (N->getOperand(i) != NotZero &&
145 N->getOperand(i).getOpcode() != ISD::UNDEF)
151 /// isBuildVectorAllZeros - Return true if the specified node is a
152 /// BUILD_VECTOR where all of the elements are 0 or undef.
153 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
154 // Look through a bit convert.
155 if (N->getOpcode() == ISD::BIT_CONVERT)
156 N = N->getOperand(0).getNode();
158 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
160 unsigned i = 0, e = N->getNumOperands();
162 // Skip over all of the undef values.
163 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
166 // Do not accept an all-undef vector.
167 if (i == e) return false;
169 // Do not accept build_vectors that aren't all constants or which have non-0
171 SDValue Zero = N->getOperand(i);
172 if (isa<ConstantSDNode>(Zero)) {
173 if (!cast<ConstantSDNode>(Zero)->isNullValue())
175 } else if (isa<ConstantFPSDNode>(Zero)) {
176 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
181 // Okay, we have at least one 0 value, check to see if the rest match or are
183 for (++i; i != e; ++i)
184 if (N->getOperand(i) != Zero &&
185 N->getOperand(i).getOpcode() != ISD::UNDEF)
190 /// isScalarToVector - Return true if the specified node is a
191 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
192 /// element is not an undef.
193 bool ISD::isScalarToVector(const SDNode *N) {
194 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
197 if (N->getOpcode() != ISD::BUILD_VECTOR)
199 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
201 unsigned NumElems = N->getNumOperands();
202 for (unsigned i = 1; i < NumElems; ++i) {
203 SDValue V = N->getOperand(i);
204 if (V.getOpcode() != ISD::UNDEF)
210 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
211 /// when given the operation for (X op Y).
212 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
213 // To perform this operation, we just need to swap the L and G bits of the
215 unsigned OldL = (Operation >> 2) & 1;
216 unsigned OldG = (Operation >> 1) & 1;
217 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
218 (OldL << 1) | // New G bit
219 (OldG << 2)); // New L bit.
222 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
223 /// 'op' is a valid SetCC operation.
224 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
225 unsigned Operation = Op;
227 Operation ^= 7; // Flip L, G, E bits, but not U.
229 Operation ^= 15; // Flip all of the condition bits.
231 if (Operation > ISD::SETTRUE2)
232 Operation &= ~8; // Don't let N and U bits get set.
234 return ISD::CondCode(Operation);
238 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
239 /// signed operation and 2 if the result is an unsigned comparison. Return zero
240 /// if the operation does not depend on the sign of the input (setne and seteq).
241 static int isSignedOp(ISD::CondCode Opcode) {
243 default: llvm_unreachable("Illegal integer setcc operation!");
245 case ISD::SETNE: return 0;
249 case ISD::SETGE: return 1;
253 case ISD::SETUGE: return 2;
257 /// getSetCCOrOperation - Return the result of a logical OR between different
258 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
259 /// returns SETCC_INVALID if it is not possible to represent the resultant
261 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
263 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
264 // Cannot fold a signed integer setcc with an unsigned integer setcc.
265 return ISD::SETCC_INVALID;
267 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
269 // If the N and U bits get set then the resultant comparison DOES suddenly
270 // care about orderedness, and is true when ordered.
271 if (Op > ISD::SETTRUE2)
272 Op &= ~16; // Clear the U bit if the N bit is set.
274 // Canonicalize illegal integer setcc's.
275 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
278 return ISD::CondCode(Op);
281 /// getSetCCAndOperation - Return the result of a logical AND between different
282 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
283 /// function returns zero if it is not possible to represent the resultant
285 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
287 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
288 // Cannot fold a signed setcc with an unsigned setcc.
289 return ISD::SETCC_INVALID;
291 // Combine all of the condition bits.
292 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
294 // Canonicalize illegal integer setcc's.
298 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
299 case ISD::SETOEQ: // SETEQ & SETU[LG]E
300 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
301 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
302 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
309 //===----------------------------------------------------------------------===//
310 // SDNode Profile Support
311 //===----------------------------------------------------------------------===//
313 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
315 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
319 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
320 /// solely with their pointer.
321 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
322 ID.AddPointer(VTList.VTs);
325 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
327 static void AddNodeIDOperands(FoldingSetNodeID &ID,
328 const SDValue *Ops, unsigned NumOps) {
329 for (; NumOps; --NumOps, ++Ops) {
330 ID.AddPointer(Ops->getNode());
331 ID.AddInteger(Ops->getResNo());
335 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
337 static void AddNodeIDOperands(FoldingSetNodeID &ID,
338 const SDUse *Ops, unsigned NumOps) {
339 for (; NumOps; --NumOps, ++Ops) {
340 ID.AddPointer(Ops->getNode());
341 ID.AddInteger(Ops->getResNo());
345 static void AddNodeIDNode(FoldingSetNodeID &ID,
346 unsigned short OpC, SDVTList VTList,
347 const SDValue *OpList, unsigned N) {
348 AddNodeIDOpcode(ID, OpC);
349 AddNodeIDValueTypes(ID, VTList);
350 AddNodeIDOperands(ID, OpList, N);
353 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
355 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
356 switch (N->getOpcode()) {
357 case ISD::TargetExternalSymbol:
358 case ISD::ExternalSymbol:
359 llvm_unreachable("Should only be used on nodes with operands");
360 default: break; // Normal nodes don't need extra info.
361 case ISD::TargetConstant:
363 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
365 case ISD::TargetConstantFP:
366 case ISD::ConstantFP: {
367 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
370 case ISD::TargetGlobalAddress:
371 case ISD::GlobalAddress:
372 case ISD::TargetGlobalTLSAddress:
373 case ISD::GlobalTLSAddress: {
374 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
375 ID.AddPointer(GA->getGlobal());
376 ID.AddInteger(GA->getOffset());
377 ID.AddInteger(GA->getTargetFlags());
380 case ISD::BasicBlock:
381 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
384 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
388 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
390 case ISD::FrameIndex:
391 case ISD::TargetFrameIndex:
392 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
395 case ISD::TargetJumpTable:
396 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
397 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
399 case ISD::ConstantPool:
400 case ISD::TargetConstantPool: {
401 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
402 ID.AddInteger(CP->getAlignment());
403 ID.AddInteger(CP->getOffset());
404 if (CP->isMachineConstantPoolEntry())
405 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
407 ID.AddPointer(CP->getConstVal());
408 ID.AddInteger(CP->getTargetFlags());
412 const LoadSDNode *LD = cast<LoadSDNode>(N);
413 ID.AddInteger(LD->getMemoryVT().getRawBits());
414 ID.AddInteger(LD->getRawSubclassData());
418 const StoreSDNode *ST = cast<StoreSDNode>(N);
419 ID.AddInteger(ST->getMemoryVT().getRawBits());
420 ID.AddInteger(ST->getRawSubclassData());
423 case ISD::ATOMIC_CMP_SWAP:
424 case ISD::ATOMIC_SWAP:
425 case ISD::ATOMIC_LOAD_ADD:
426 case ISD::ATOMIC_LOAD_SUB:
427 case ISD::ATOMIC_LOAD_AND:
428 case ISD::ATOMIC_LOAD_OR:
429 case ISD::ATOMIC_LOAD_XOR:
430 case ISD::ATOMIC_LOAD_NAND:
431 case ISD::ATOMIC_LOAD_MIN:
432 case ISD::ATOMIC_LOAD_MAX:
433 case ISD::ATOMIC_LOAD_UMIN:
434 case ISD::ATOMIC_LOAD_UMAX: {
435 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
436 ID.AddInteger(AT->getMemoryVT().getRawBits());
437 ID.AddInteger(AT->getRawSubclassData());
440 case ISD::VECTOR_SHUFFLE: {
441 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
442 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
444 ID.AddInteger(SVN->getMaskElt(i));
447 case ISD::TargetBlockAddress:
448 case ISD::BlockAddress: {
449 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress());
450 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags());
453 } // end switch (N->getOpcode())
456 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
458 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
459 AddNodeIDOpcode(ID, N->getOpcode());
460 // Add the return value info.
461 AddNodeIDValueTypes(ID, N->getVTList());
462 // Add the operand info.
463 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
465 // Handle SDNode leafs with special info.
466 AddNodeIDCustom(ID, N);
469 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
470 /// the CSE map that carries volatility, temporalness, indexing mode, and
471 /// extension/truncation information.
473 static inline unsigned
474 encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile,
475 bool isNonTemporal) {
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!");
483 (isNonTemporal << 6);
486 //===----------------------------------------------------------------------===//
487 // SelectionDAG Class
488 //===----------------------------------------------------------------------===//
490 /// doNotCSE - Return true if CSE should not be performed for this node.
491 static bool doNotCSE(SDNode *N) {
492 if (N->getValueType(0) == MVT::Flag)
493 return true; // Never CSE anything that produces a flag.
495 switch (N->getOpcode()) {
497 case ISD::HANDLENODE:
499 return true; // Never CSE these nodes.
502 // Check that remaining values produced are not flags.
503 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
504 if (N->getValueType(i) == MVT::Flag)
505 return true; // Never CSE anything that produces a flag.
510 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
512 void SelectionDAG::RemoveDeadNodes() {
513 // Create a dummy node (which is not added to allnodes), that adds a reference
514 // to the root node, preventing it from being deleted.
515 HandleSDNode Dummy(getRoot());
517 SmallVector<SDNode*, 128> DeadNodes;
519 // Add all obviously-dead nodes to the DeadNodes worklist.
520 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
522 DeadNodes.push_back(I);
524 RemoveDeadNodes(DeadNodes);
526 // If the root changed (e.g. it was a dead load, update the root).
527 setRoot(Dummy.getValue());
530 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
531 /// given list, and any nodes that become unreachable as a result.
532 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
533 DAGUpdateListener *UpdateListener) {
535 // Process the worklist, deleting the nodes and adding their uses to the
537 while (!DeadNodes.empty()) {
538 SDNode *N = DeadNodes.pop_back_val();
541 UpdateListener->NodeDeleted(N, 0);
543 // Take the node out of the appropriate CSE map.
544 RemoveNodeFromCSEMaps(N);
546 // Next, brutally remove the operand list. This is safe to do, as there are
547 // no cycles in the graph.
548 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
550 SDNode *Operand = Use.getNode();
553 // Now that we removed this operand, see if there are no uses of it left.
554 if (Operand->use_empty())
555 DeadNodes.push_back(Operand);
562 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
563 SmallVector<SDNode*, 16> DeadNodes(1, N);
564 RemoveDeadNodes(DeadNodes, UpdateListener);
567 void SelectionDAG::DeleteNode(SDNode *N) {
568 // First take this out of the appropriate CSE map.
569 RemoveNodeFromCSEMaps(N);
571 // Finally, remove uses due to operands of this node, remove from the
572 // AllNodes list, and delete the node.
573 DeleteNodeNotInCSEMaps(N);
576 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
577 assert(N != AllNodes.begin() && "Cannot delete the entry node!");
578 assert(N->use_empty() && "Cannot delete a node that is not dead!");
580 // Drop all of the operands and decrement used node's use counts.
586 void SelectionDAG::DeallocateNode(SDNode *N) {
587 if (N->OperandsNeedDelete)
588 delete[] N->OperandList;
590 // Set the opcode to DELETED_NODE to help catch bugs when node
591 // memory is reallocated.
592 N->NodeType = ISD::DELETED_NODE;
594 NodeAllocator.Deallocate(AllNodes.remove(N));
596 // Remove the ordering of this node.
599 // If any of the SDDbgValue nodes refer to this SDNode, invalidate them.
600 SmallVector<SDDbgValue*, 2> &DbgVals = DbgInfo->getSDDbgValues(N);
601 for (unsigned i = 0, e = DbgVals.size(); i != e; ++i)
602 DbgVals[i]->setIsInvalidated();
605 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
606 /// correspond to it. This is useful when we're about to delete or repurpose
607 /// the node. We don't want future request for structurally identical nodes
608 /// to return N anymore.
609 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
611 switch (N->getOpcode()) {
612 case ISD::EntryToken:
613 llvm_unreachable("EntryToken should not be in CSEMaps!");
615 case ISD::HANDLENODE: return false; // noop.
617 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
618 "Cond code doesn't exist!");
619 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
620 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
622 case ISD::ExternalSymbol:
623 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
625 case ISD::TargetExternalSymbol: {
626 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
627 Erased = TargetExternalSymbols.erase(
628 std::pair<std::string,unsigned char>(ESN->getSymbol(),
629 ESN->getTargetFlags()));
632 case ISD::VALUETYPE: {
633 EVT VT = cast<VTSDNode>(N)->getVT();
634 if (VT.isExtended()) {
635 Erased = ExtendedValueTypeNodes.erase(VT);
637 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0;
638 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0;
643 // Remove it from the CSE Map.
644 Erased = CSEMap.RemoveNode(N);
648 // Verify that the node was actually in one of the CSE maps, unless it has a
649 // flag result (which cannot be CSE'd) or is one of the special cases that are
650 // not subject to CSE.
651 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
652 !N->isMachineOpcode() && !doNotCSE(N)) {
655 llvm_unreachable("Node is not in map!");
661 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
662 /// maps and modified in place. Add it back to the CSE maps, unless an identical
663 /// node already exists, in which case transfer all its users to the existing
664 /// node. This transfer can potentially trigger recursive merging.
667 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N,
668 DAGUpdateListener *UpdateListener) {
669 // For node types that aren't CSE'd, just act as if no identical node
672 SDNode *Existing = CSEMap.GetOrInsertNode(N);
674 // If there was already an existing matching node, use ReplaceAllUsesWith
675 // to replace the dead one with the existing one. This can cause
676 // recursive merging of other unrelated nodes down the line.
677 ReplaceAllUsesWith(N, Existing, UpdateListener);
679 // N is now dead. Inform the listener if it exists and delete it.
681 UpdateListener->NodeDeleted(N, Existing);
682 DeleteNodeNotInCSEMaps(N);
687 // If the node doesn't already exist, we updated it. Inform a listener if
690 UpdateListener->NodeUpdated(N);
693 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
694 /// were replaced with those specified. If this node is never memoized,
695 /// return null, otherwise return a pointer to the slot it would take. If a
696 /// node already exists with these operands, the slot will be non-null.
697 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
702 SDValue Ops[] = { Op };
704 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
705 AddNodeIDCustom(ID, N);
706 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
710 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
711 /// were replaced with those specified. If this node is never memoized,
712 /// return null, otherwise return a pointer to the slot it would take. If a
713 /// node already exists with these operands, the slot will be non-null.
714 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
715 SDValue Op1, SDValue Op2,
720 SDValue Ops[] = { Op1, Op2 };
722 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
723 AddNodeIDCustom(ID, N);
724 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
729 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
730 /// were replaced with those specified. If this node is never memoized,
731 /// return null, otherwise return a pointer to the slot it would take. If a
732 /// node already exists with these operands, the slot will be non-null.
733 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
734 const SDValue *Ops,unsigned NumOps,
740 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
741 AddNodeIDCustom(ID, N);
742 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
746 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
747 void SelectionDAG::VerifyNode(SDNode *N) {
748 switch (N->getOpcode()) {
751 case ISD::BUILD_PAIR: {
752 EVT VT = N->getValueType(0);
753 assert(N->getNumValues() == 1 && "Too many results!");
754 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
755 "Wrong return type!");
756 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
757 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
758 "Mismatched operand types!");
759 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
760 "Wrong operand type!");
761 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
762 "Wrong return type size");
765 case ISD::BUILD_VECTOR: {
766 assert(N->getNumValues() == 1 && "Too many results!");
767 assert(N->getValueType(0).isVector() && "Wrong return type!");
768 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
769 "Wrong number of operands!");
770 EVT EltVT = N->getValueType(0).getVectorElementType();
771 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
772 assert((I->getValueType() == EltVT ||
773 (EltVT.isInteger() && I->getValueType().isInteger() &&
774 EltVT.bitsLE(I->getValueType()))) &&
775 "Wrong operand type!");
781 /// getEVTAlignment - Compute the default alignment value for the
784 unsigned SelectionDAG::getEVTAlignment(EVT VT) const {
785 const Type *Ty = VT == MVT::iPTR ?
786 PointerType::get(Type::getInt8Ty(*getContext()), 0) :
787 VT.getTypeForEVT(*getContext());
789 return TLI.getTargetData()->getABITypeAlignment(Ty);
792 // EntryNode could meaningfully have debug info if we can find it...
793 SelectionDAG::SelectionDAG(const TargetMachine &tm)
794 : TM(tm), TLI(*tm.getTargetLowering()), TSI(*tm.getSelectionDAGInfo()),
795 EntryNode(ISD::EntryToken, DebugLoc(), getVTList(MVT::Other)),
796 Root(getEntryNode()), Ordering(0) {
797 AllNodes.push_back(&EntryNode);
798 Ordering = new SDNodeOrdering();
799 DbgInfo = new SDDbgInfo();
802 void SelectionDAG::init(MachineFunction &mf) {
804 Context = &mf.getFunction()->getContext();
807 SelectionDAG::~SelectionDAG() {
814 void SelectionDAG::allnodes_clear() {
815 assert(&*AllNodes.begin() == &EntryNode);
816 AllNodes.remove(AllNodes.begin());
817 while (!AllNodes.empty())
818 DeallocateNode(AllNodes.begin());
821 void SelectionDAG::clear() {
823 OperandAllocator.Reset();
826 ExtendedValueTypeNodes.clear();
827 ExternalSymbols.clear();
828 TargetExternalSymbols.clear();
829 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
830 static_cast<CondCodeSDNode*>(0));
831 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
832 static_cast<SDNode*>(0));
834 EntryNode.UseList = 0;
835 AllNodes.push_back(&EntryNode);
836 Root = getEntryNode();
838 Ordering = new SDNodeOrdering();
841 DbgInfo = new SDDbgInfo();
844 SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
845 return VT.bitsGT(Op.getValueType()) ?
846 getNode(ISD::SIGN_EXTEND, DL, VT, Op) :
847 getNode(ISD::TRUNCATE, DL, VT, Op);
850 SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
851 return VT.bitsGT(Op.getValueType()) ?
852 getNode(ISD::ZERO_EXTEND, DL, VT, Op) :
853 getNode(ISD::TRUNCATE, DL, VT, Op);
856 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) {
857 assert(!VT.isVector() &&
858 "getZeroExtendInReg should use the vector element type instead of "
860 if (Op.getValueType() == VT) return Op;
861 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
862 APInt Imm = APInt::getLowBitsSet(BitWidth,
864 return getNode(ISD::AND, DL, Op.getValueType(), Op,
865 getConstant(Imm, Op.getValueType()));
868 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
870 SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) {
871 EVT EltVT = VT.getScalarType();
873 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
874 return getNode(ISD::XOR, DL, VT, Val, NegOne);
877 SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) {
878 EVT EltVT = VT.getScalarType();
879 assert((EltVT.getSizeInBits() >= 64 ||
880 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
881 "getConstant with a uint64_t value that doesn't fit in the type!");
882 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
885 SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) {
886 return getConstant(*ConstantInt::get(*Context, Val), VT, isT);
889 SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
890 assert(VT.isInteger() && "Cannot create FP integer constant!");
892 EVT EltVT = VT.getScalarType();
893 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
894 "APInt size does not match type size!");
896 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
898 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
902 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
904 return SDValue(N, 0);
907 N = new (NodeAllocator) ConstantSDNode(isT, &Val, EltVT);
908 CSEMap.InsertNode(N, IP);
909 AllNodes.push_back(N);
912 SDValue Result(N, 0);
914 SmallVector<SDValue, 8> Ops;
915 Ops.assign(VT.getVectorNumElements(), Result);
916 Result = getNode(ISD::BUILD_VECTOR, DebugLoc(), VT, &Ops[0], Ops.size());
921 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
922 return getConstant(Val, TLI.getPointerTy(), isTarget);
926 SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) {
927 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget);
930 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){
931 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
933 EVT EltVT = VT.getScalarType();
935 // Do the map lookup using the actual bit pattern for the floating point
936 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
937 // we don't have issues with SNANs.
938 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
940 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
944 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
946 return SDValue(N, 0);
949 N = new (NodeAllocator) ConstantFPSDNode(isTarget, &V, EltVT);
950 CSEMap.InsertNode(N, IP);
951 AllNodes.push_back(N);
954 SDValue Result(N, 0);
956 SmallVector<SDValue, 8> Ops;
957 Ops.assign(VT.getVectorNumElements(), Result);
958 // FIXME DebugLoc info might be appropriate here
959 Result = getNode(ISD::BUILD_VECTOR, DebugLoc(), VT, &Ops[0], Ops.size());
964 SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) {
965 EVT EltVT = VT.getScalarType();
967 return getConstantFP(APFloat((float)Val), VT, isTarget);
968 else if (EltVT==MVT::f64)
969 return getConstantFP(APFloat(Val), VT, isTarget);
970 else if (EltVT==MVT::f80 || EltVT==MVT::f128) {
972 APFloat apf = APFloat(Val);
973 apf.convert(*EVTToAPFloatSemantics(EltVT), APFloat::rmNearestTiesToEven,
975 return getConstantFP(apf, VT, isTarget);
977 assert(0 && "Unsupported type in getConstantFP");
982 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
983 EVT VT, int64_t Offset,
985 unsigned char TargetFlags) {
986 assert((TargetFlags == 0 || isTargetGA) &&
987 "Cannot set target flags on target-independent globals");
989 // Truncate (with sign-extension) the offset value to the pointer size.
990 EVT PTy = TLI.getPointerTy();
991 unsigned BitWidth = PTy.getSizeInBits();
993 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
995 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
997 // If GV is an alias then use the aliasee for determining thread-localness.
998 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
999 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
1003 if (GVar && GVar->isThreadLocal())
1004 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
1006 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
1008 FoldingSetNodeID ID;
1009 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1011 ID.AddInteger(Offset);
1012 ID.AddInteger(TargetFlags);
1014 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1015 return SDValue(E, 0);
1017 SDNode *N = new (NodeAllocator) GlobalAddressSDNode(Opc, GV, VT,
1018 Offset, TargetFlags);
1019 CSEMap.InsertNode(N, IP);
1020 AllNodes.push_back(N);
1021 return SDValue(N, 0);
1024 SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1025 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1026 FoldingSetNodeID ID;
1027 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1030 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1031 return SDValue(E, 0);
1033 SDNode *N = new (NodeAllocator) FrameIndexSDNode(FI, VT, isTarget);
1034 CSEMap.InsertNode(N, IP);
1035 AllNodes.push_back(N);
1036 return SDValue(N, 0);
1039 SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1040 unsigned char TargetFlags) {
1041 assert((TargetFlags == 0 || isTarget) &&
1042 "Cannot set target flags on target-independent jump tables");
1043 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1044 FoldingSetNodeID ID;
1045 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1047 ID.AddInteger(TargetFlags);
1049 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1050 return SDValue(E, 0);
1052 SDNode *N = new (NodeAllocator) JumpTableSDNode(JTI, VT, isTarget,
1054 CSEMap.InsertNode(N, IP);
1055 AllNodes.push_back(N);
1056 return SDValue(N, 0);
1059 SDValue SelectionDAG::getConstantPool(const Constant *C, EVT VT,
1060 unsigned Alignment, int Offset,
1062 unsigned char TargetFlags) {
1063 assert((TargetFlags == 0 || isTarget) &&
1064 "Cannot set target flags on target-independent globals");
1066 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1067 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1068 FoldingSetNodeID ID;
1069 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1070 ID.AddInteger(Alignment);
1071 ID.AddInteger(Offset);
1073 ID.AddInteger(TargetFlags);
1075 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1076 return SDValue(E, 0);
1078 SDNode *N = new (NodeAllocator) ConstantPoolSDNode(isTarget, C, VT, Offset,
1079 Alignment, TargetFlags);
1080 CSEMap.InsertNode(N, IP);
1081 AllNodes.push_back(N);
1082 return SDValue(N, 0);
1086 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1087 unsigned Alignment, int Offset,
1089 unsigned char TargetFlags) {
1090 assert((TargetFlags == 0 || isTarget) &&
1091 "Cannot set target flags on target-independent globals");
1093 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1094 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1095 FoldingSetNodeID ID;
1096 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1097 ID.AddInteger(Alignment);
1098 ID.AddInteger(Offset);
1099 C->AddSelectionDAGCSEId(ID);
1100 ID.AddInteger(TargetFlags);
1102 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1103 return SDValue(E, 0);
1105 SDNode *N = new (NodeAllocator) ConstantPoolSDNode(isTarget, C, VT, Offset,
1106 Alignment, TargetFlags);
1107 CSEMap.InsertNode(N, IP);
1108 AllNodes.push_back(N);
1109 return SDValue(N, 0);
1112 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1113 FoldingSetNodeID ID;
1114 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1117 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1118 return SDValue(E, 0);
1120 SDNode *N = new (NodeAllocator) BasicBlockSDNode(MBB);
1121 CSEMap.InsertNode(N, IP);
1122 AllNodes.push_back(N);
1123 return SDValue(N, 0);
1126 SDValue SelectionDAG::getValueType(EVT VT) {
1127 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
1128 ValueTypeNodes.size())
1129 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1);
1131 SDNode *&N = VT.isExtended() ?
1132 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
1134 if (N) return SDValue(N, 0);
1135 N = new (NodeAllocator) VTSDNode(VT);
1136 AllNodes.push_back(N);
1137 return SDValue(N, 0);
1140 SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
1141 SDNode *&N = ExternalSymbols[Sym];
1142 if (N) return SDValue(N, 0);
1143 N = new (NodeAllocator) ExternalSymbolSDNode(false, Sym, 0, VT);
1144 AllNodes.push_back(N);
1145 return SDValue(N, 0);
1148 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
1149 unsigned char TargetFlags) {
1151 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym,
1153 if (N) return SDValue(N, 0);
1154 N = new (NodeAllocator) ExternalSymbolSDNode(true, Sym, TargetFlags, VT);
1155 AllNodes.push_back(N);
1156 return SDValue(N, 0);
1159 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1160 if ((unsigned)Cond >= CondCodeNodes.size())
1161 CondCodeNodes.resize(Cond+1);
1163 if (CondCodeNodes[Cond] == 0) {
1164 CondCodeSDNode *N = new (NodeAllocator) CondCodeSDNode(Cond);
1165 CondCodeNodes[Cond] = N;
1166 AllNodes.push_back(N);
1169 return SDValue(CondCodeNodes[Cond], 0);
1172 // commuteShuffle - swaps the values of N1 and N2, and swaps all indices in
1173 // the shuffle mask M that point at N1 to point at N2, and indices that point
1174 // N2 to point at N1.
1175 static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) {
1177 int NElts = M.size();
1178 for (int i = 0; i != NElts; ++i) {
1186 SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1,
1187 SDValue N2, const int *Mask) {
1188 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE");
1189 assert(VT.isVector() && N1.getValueType().isVector() &&
1190 "Vector Shuffle VTs must be a vectors");
1191 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType()
1192 && "Vector Shuffle VTs must have same element type");
1194 // Canonicalize shuffle undef, undef -> undef
1195 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF)
1196 return getUNDEF(VT);
1198 // Validate that all indices in Mask are within the range of the elements
1199 // input to the shuffle.
1200 unsigned NElts = VT.getVectorNumElements();
1201 SmallVector<int, 8> MaskVec;
1202 for (unsigned i = 0; i != NElts; ++i) {
1203 assert(Mask[i] < (int)(NElts * 2) && "Index out of range");
1204 MaskVec.push_back(Mask[i]);
1207 // Canonicalize shuffle v, v -> v, undef
1210 for (unsigned i = 0; i != NElts; ++i)
1211 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts;
1214 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1215 if (N1.getOpcode() == ISD::UNDEF)
1216 commuteShuffle(N1, N2, MaskVec);
1218 // Canonicalize all index into lhs, -> shuffle lhs, undef
1219 // Canonicalize all index into rhs, -> shuffle rhs, undef
1220 bool AllLHS = true, AllRHS = true;
1221 bool N2Undef = N2.getOpcode() == ISD::UNDEF;
1222 for (unsigned i = 0; i != NElts; ++i) {
1223 if (MaskVec[i] >= (int)NElts) {
1228 } else if (MaskVec[i] >= 0) {
1232 if (AllLHS && AllRHS)
1233 return getUNDEF(VT);
1234 if (AllLHS && !N2Undef)
1238 commuteShuffle(N1, N2, MaskVec);
1241 // If Identity shuffle, or all shuffle in to undef, return that node.
1242 bool AllUndef = true;
1243 bool Identity = true;
1244 for (unsigned i = 0; i != NElts; ++i) {
1245 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false;
1246 if (MaskVec[i] >= 0) AllUndef = false;
1248 if (Identity && NElts == N1.getValueType().getVectorNumElements())
1251 return getUNDEF(VT);
1253 FoldingSetNodeID ID;
1254 SDValue Ops[2] = { N1, N2 };
1255 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2);
1256 for (unsigned i = 0; i != NElts; ++i)
1257 ID.AddInteger(MaskVec[i]);
1260 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1261 return SDValue(E, 0);
1263 // Allocate the mask array for the node out of the BumpPtrAllocator, since
1264 // SDNode doesn't have access to it. This memory will be "leaked" when
1265 // the node is deallocated, but recovered when the NodeAllocator is released.
1266 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
1267 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int));
1269 ShuffleVectorSDNode *N =
1270 new (NodeAllocator) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc);
1271 CSEMap.InsertNode(N, IP);
1272 AllNodes.push_back(N);
1273 return SDValue(N, 0);
1276 SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl,
1277 SDValue Val, SDValue DTy,
1278 SDValue STy, SDValue Rnd, SDValue Sat,
1279 ISD::CvtCode Code) {
1280 // If the src and dest types are the same and the conversion is between
1281 // integer types of the same sign or two floats, no conversion is necessary.
1283 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF))
1286 FoldingSetNodeID ID;
1287 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat };
1288 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5);
1290 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1291 return SDValue(E, 0);
1293 CvtRndSatSDNode *N = new (NodeAllocator) CvtRndSatSDNode(VT, dl, Ops, 5,
1295 CSEMap.InsertNode(N, IP);
1296 AllNodes.push_back(N);
1297 return SDValue(N, 0);
1300 SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) {
1301 FoldingSetNodeID ID;
1302 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1303 ID.AddInteger(RegNo);
1305 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1306 return SDValue(E, 0);
1308 SDNode *N = new (NodeAllocator) RegisterSDNode(RegNo, VT);
1309 CSEMap.InsertNode(N, IP);
1310 AllNodes.push_back(N);
1311 return SDValue(N, 0);
1314 SDValue SelectionDAG::getEHLabel(DebugLoc dl, SDValue Root, MCSymbol *Label) {
1315 FoldingSetNodeID ID;
1316 SDValue Ops[] = { Root };
1317 AddNodeIDNode(ID, ISD::EH_LABEL, getVTList(MVT::Other), &Ops[0], 1);
1318 ID.AddPointer(Label);
1320 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1321 return SDValue(E, 0);
1323 SDNode *N = new (NodeAllocator) EHLabelSDNode(dl, Root, Label);
1324 CSEMap.InsertNode(N, IP);
1325 AllNodes.push_back(N);
1326 return SDValue(N, 0);
1330 SDValue SelectionDAG::getBlockAddress(const BlockAddress *BA, EVT VT,
1332 unsigned char TargetFlags) {
1333 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress;
1335 FoldingSetNodeID ID;
1336 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1338 ID.AddInteger(TargetFlags);
1340 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1341 return SDValue(E, 0);
1343 SDNode *N = new (NodeAllocator) BlockAddressSDNode(Opc, VT, BA, TargetFlags);
1344 CSEMap.InsertNode(N, IP);
1345 AllNodes.push_back(N);
1346 return SDValue(N, 0);
1349 SDValue SelectionDAG::getSrcValue(const Value *V) {
1350 assert((!V || V->getType()->isPointerTy()) &&
1351 "SrcValue is not a pointer?");
1353 FoldingSetNodeID ID;
1354 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1358 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1359 return SDValue(E, 0);
1361 SDNode *N = new (NodeAllocator) SrcValueSDNode(V);
1362 CSEMap.InsertNode(N, IP);
1363 AllNodes.push_back(N);
1364 return SDValue(N, 0);
1367 /// getMDNode - Return an MDNodeSDNode which holds an MDNode.
1368 SDValue SelectionDAG::getMDNode(const MDNode *MD) {
1369 FoldingSetNodeID ID;
1370 AddNodeIDNode(ID, ISD::MDNODE_SDNODE, getVTList(MVT::Other), 0, 0);
1374 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1375 return SDValue(E, 0);
1377 SDNode *N = new (NodeAllocator) MDNodeSDNode(MD);
1378 CSEMap.InsertNode(N, IP);
1379 AllNodes.push_back(N);
1380 return SDValue(N, 0);
1384 /// getShiftAmountOperand - Return the specified value casted to
1385 /// the target's desired shift amount type.
1386 SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) {
1387 EVT OpTy = Op.getValueType();
1388 MVT ShTy = TLI.getShiftAmountTy();
1389 if (OpTy == ShTy || OpTy.isVector()) return Op;
1391 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
1392 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op);
1395 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1396 /// specified value type.
1397 SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
1398 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1399 unsigned ByteSize = VT.getStoreSize();
1400 const Type *Ty = VT.getTypeForEVT(*getContext());
1401 unsigned StackAlign =
1402 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1404 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
1405 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1408 /// CreateStackTemporary - Create a stack temporary suitable for holding
1409 /// either of the specified value types.
1410 SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
1411 unsigned Bytes = std::max(VT1.getStoreSizeInBits(),
1412 VT2.getStoreSizeInBits())/8;
1413 const Type *Ty1 = VT1.getTypeForEVT(*getContext());
1414 const Type *Ty2 = VT2.getTypeForEVT(*getContext());
1415 const TargetData *TD = TLI.getTargetData();
1416 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1),
1417 TD->getPrefTypeAlignment(Ty2));
1419 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1420 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false);
1421 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1424 SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1,
1425 SDValue N2, ISD::CondCode Cond, DebugLoc dl) {
1426 // These setcc operations always fold.
1430 case ISD::SETFALSE2: return getConstant(0, VT);
1432 case ISD::SETTRUE2: return getConstant(1, VT);
1444 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1448 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1449 const APInt &C2 = N2C->getAPIntValue();
1450 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1451 const APInt &C1 = N1C->getAPIntValue();
1454 default: llvm_unreachable("Unknown integer setcc!");
1455 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1456 case ISD::SETNE: return getConstant(C1 != C2, VT);
1457 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1458 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1459 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1460 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1461 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1462 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1463 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1464 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1468 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1469 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1470 // No compile time operations on this type yet.
1471 if (N1C->getValueType(0) == MVT::ppcf128)
1474 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1477 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1478 return getUNDEF(VT);
1480 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1481 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1482 return getUNDEF(VT);
1484 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1485 R==APFloat::cmpLessThan, VT);
1486 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1487 return getUNDEF(VT);
1489 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1490 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1491 return getUNDEF(VT);
1493 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1494 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1495 return getUNDEF(VT);
1497 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1498 R==APFloat::cmpEqual, VT);
1499 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1500 return getUNDEF(VT);
1502 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1503 R==APFloat::cmpEqual, VT);
1504 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1505 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1506 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1507 R==APFloat::cmpEqual, VT);
1508 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1509 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1510 R==APFloat::cmpLessThan, VT);
1511 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1512 R==APFloat::cmpUnordered, VT);
1513 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1514 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1517 // Ensure that the constant occurs on the RHS.
1518 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1522 // Could not fold it.
1526 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1527 /// use this predicate to simplify operations downstream.
1528 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1529 // This predicate is not safe for vector operations.
1530 if (Op.getValueType().isVector())
1533 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
1534 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1537 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1538 /// this predicate to simplify operations downstream. Mask is known to be zero
1539 /// for bits that V cannot have.
1540 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1541 unsigned Depth) const {
1542 APInt KnownZero, KnownOne;
1543 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1544 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1545 return (KnownZero & Mask) == Mask;
1548 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1549 /// known to be either zero or one and return them in the KnownZero/KnownOne
1550 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1552 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1553 APInt &KnownZero, APInt &KnownOne,
1554 unsigned Depth) const {
1555 unsigned BitWidth = Mask.getBitWidth();
1556 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() &&
1557 "Mask size mismatches value type size!");
1559 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1560 if (Depth == 6 || Mask == 0)
1561 return; // Limit search depth.
1563 APInt KnownZero2, KnownOne2;
1565 switch (Op.getOpcode()) {
1567 // We know all of the bits for a constant!
1568 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1569 KnownZero = ~KnownOne & Mask;
1572 // If either the LHS or the RHS are Zero, the result is zero.
1573 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1574 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1575 KnownZero2, KnownOne2, Depth+1);
1576 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1577 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1579 // Output known-1 bits are only known if set in both the LHS & RHS.
1580 KnownOne &= KnownOne2;
1581 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1582 KnownZero |= KnownZero2;
1585 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1586 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1587 KnownZero2, KnownOne2, Depth+1);
1588 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1589 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1591 // Output known-0 bits are only known if clear in both the LHS & RHS.
1592 KnownZero &= KnownZero2;
1593 // Output known-1 are known to be set if set in either the LHS | RHS.
1594 KnownOne |= KnownOne2;
1597 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1598 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1599 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1600 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1602 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1603 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1604 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1605 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1606 KnownZero = KnownZeroOut;
1610 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1611 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1612 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1613 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1614 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1616 // If low bits are zero in either operand, output low known-0 bits.
1617 // Also compute a conserative estimate for high known-0 bits.
1618 // More trickiness is possible, but this is sufficient for the
1619 // interesting case of alignment computation.
1621 unsigned TrailZ = KnownZero.countTrailingOnes() +
1622 KnownZero2.countTrailingOnes();
1623 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1624 KnownZero2.countLeadingOnes(),
1625 BitWidth) - BitWidth;
1627 TrailZ = std::min(TrailZ, BitWidth);
1628 LeadZ = std::min(LeadZ, BitWidth);
1629 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1630 APInt::getHighBitsSet(BitWidth, LeadZ);
1635 // For the purposes of computing leading zeros we can conservatively
1636 // treat a udiv as a logical right shift by the power of 2 known to
1637 // be less than the denominator.
1638 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1639 ComputeMaskedBits(Op.getOperand(0),
1640 AllOnes, KnownZero2, KnownOne2, Depth+1);
1641 unsigned LeadZ = KnownZero2.countLeadingOnes();
1645 ComputeMaskedBits(Op.getOperand(1),
1646 AllOnes, KnownZero2, KnownOne2, Depth+1);
1647 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1648 if (RHSUnknownLeadingOnes != BitWidth)
1649 LeadZ = std::min(BitWidth,
1650 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1652 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1656 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1657 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1658 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1659 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1661 // Only known if known in both the LHS and RHS.
1662 KnownOne &= KnownOne2;
1663 KnownZero &= KnownZero2;
1665 case ISD::SELECT_CC:
1666 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1667 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1668 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1669 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1671 // Only known if known in both the LHS and RHS.
1672 KnownOne &= KnownOne2;
1673 KnownZero &= KnownZero2;
1681 if (Op.getResNo() != 1)
1683 // The boolean result conforms to getBooleanContents. Fall through.
1685 // If we know the result of a setcc has the top bits zero, use this info.
1686 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent &&
1688 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1691 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 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.lshr(ShAmt),
1700 KnownZero, KnownOne, Depth+1);
1701 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1702 KnownZero <<= ShAmt;
1704 // low bits known zero.
1705 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1709 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
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 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1718 KnownZero, KnownOne, Depth+1);
1719 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1720 KnownZero = KnownZero.lshr(ShAmt);
1721 KnownOne = KnownOne.lshr(ShAmt);
1723 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1724 KnownZero |= HighBits; // High bits known zero.
1728 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1729 unsigned ShAmt = SA->getZExtValue();
1731 // If the shift count is an invalid immediate, don't do anything.
1732 if (ShAmt >= BitWidth)
1735 APInt InDemandedMask = (Mask << ShAmt);
1736 // If any of the demanded bits are produced by the sign extension, we also
1737 // demand the input sign bit.
1738 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1739 if (HighBits.getBoolValue())
1740 InDemandedMask |= APInt::getSignBit(BitWidth);
1742 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1745 KnownZero = KnownZero.lshr(ShAmt);
1746 KnownOne = KnownOne.lshr(ShAmt);
1748 // Handle the sign bits.
1749 APInt SignBit = APInt::getSignBit(BitWidth);
1750 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1752 if (KnownZero.intersects(SignBit)) {
1753 KnownZero |= HighBits; // New bits are known zero.
1754 } else if (KnownOne.intersects(SignBit)) {
1755 KnownOne |= HighBits; // New bits are known one.
1759 case ISD::SIGN_EXTEND_INREG: {
1760 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1761 unsigned EBits = EVT.getScalarType().getSizeInBits();
1763 // Sign extension. Compute the demanded bits in the result that are not
1764 // present in the input.
1765 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1767 APInt InSignBit = APInt::getSignBit(EBits);
1768 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1770 // If the sign extended bits are demanded, we know that the sign
1772 InSignBit.zext(BitWidth);
1773 if (NewBits.getBoolValue())
1774 InputDemandedBits |= InSignBit;
1776 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1777 KnownZero, KnownOne, Depth+1);
1778 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1780 // If the sign bit of the input is known set or clear, then we know the
1781 // top bits of the result.
1782 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1783 KnownZero |= NewBits;
1784 KnownOne &= ~NewBits;
1785 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1786 KnownOne |= NewBits;
1787 KnownZero &= ~NewBits;
1788 } else { // Input sign bit unknown
1789 KnownZero &= ~NewBits;
1790 KnownOne &= ~NewBits;
1797 unsigned LowBits = Log2_32(BitWidth)+1;
1798 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1803 if (ISD::isZEXTLoad(Op.getNode())) {
1804 LoadSDNode *LD = cast<LoadSDNode>(Op);
1805 EVT VT = LD->getMemoryVT();
1806 unsigned MemBits = VT.getScalarType().getSizeInBits();
1807 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1811 case ISD::ZERO_EXTEND: {
1812 EVT InVT = Op.getOperand(0).getValueType();
1813 unsigned InBits = InVT.getScalarType().getSizeInBits();
1814 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1815 APInt InMask = Mask;
1816 InMask.trunc(InBits);
1817 KnownZero.trunc(InBits);
1818 KnownOne.trunc(InBits);
1819 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1820 KnownZero.zext(BitWidth);
1821 KnownOne.zext(BitWidth);
1822 KnownZero |= NewBits;
1825 case ISD::SIGN_EXTEND: {
1826 EVT InVT = Op.getOperand(0).getValueType();
1827 unsigned InBits = InVT.getScalarType().getSizeInBits();
1828 APInt InSignBit = APInt::getSignBit(InBits);
1829 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1830 APInt InMask = Mask;
1831 InMask.trunc(InBits);
1833 // If any of the sign extended bits are demanded, we know that the sign
1834 // bit is demanded. Temporarily set this bit in the mask for our callee.
1835 if (NewBits.getBoolValue())
1836 InMask |= InSignBit;
1838 KnownZero.trunc(InBits);
1839 KnownOne.trunc(InBits);
1840 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1842 // Note if the sign bit is known to be zero or one.
1843 bool SignBitKnownZero = KnownZero.isNegative();
1844 bool SignBitKnownOne = KnownOne.isNegative();
1845 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1846 "Sign bit can't be known to be both zero and one!");
1848 // If the sign bit wasn't actually demanded by our caller, we don't
1849 // want it set in the KnownZero and KnownOne result values. Reset the
1850 // mask and reapply it to the result values.
1852 InMask.trunc(InBits);
1853 KnownZero &= InMask;
1856 KnownZero.zext(BitWidth);
1857 KnownOne.zext(BitWidth);
1859 // If the sign bit is known zero or one, the top bits match.
1860 if (SignBitKnownZero)
1861 KnownZero |= NewBits;
1862 else if (SignBitKnownOne)
1863 KnownOne |= NewBits;
1866 case ISD::ANY_EXTEND: {
1867 EVT InVT = Op.getOperand(0).getValueType();
1868 unsigned InBits = InVT.getScalarType().getSizeInBits();
1869 APInt InMask = Mask;
1870 InMask.trunc(InBits);
1871 KnownZero.trunc(InBits);
1872 KnownOne.trunc(InBits);
1873 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1874 KnownZero.zext(BitWidth);
1875 KnownOne.zext(BitWidth);
1878 case ISD::TRUNCATE: {
1879 EVT InVT = Op.getOperand(0).getValueType();
1880 unsigned InBits = InVT.getScalarType().getSizeInBits();
1881 APInt InMask = Mask;
1882 InMask.zext(InBits);
1883 KnownZero.zext(InBits);
1884 KnownOne.zext(InBits);
1885 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1886 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1887 KnownZero.trunc(BitWidth);
1888 KnownOne.trunc(BitWidth);
1891 case ISD::AssertZext: {
1892 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1893 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1894 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1896 KnownZero |= (~InMask) & Mask;
1900 // All bits are zero except the low bit.
1901 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1905 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1906 // We know that the top bits of C-X are clear if X contains less bits
1907 // than C (i.e. no wrap-around can happen). For example, 20-X is
1908 // positive if we can prove that X is >= 0 and < 16.
1909 if (CLHS->getAPIntValue().isNonNegative()) {
1910 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1911 // NLZ can't be BitWidth with no sign bit
1912 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1913 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1916 // If all of the MaskV bits are known to be zero, then we know the
1917 // output top bits are zero, because we now know that the output is
1919 if ((KnownZero2 & MaskV) == MaskV) {
1920 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1921 // Top bits known zero.
1922 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1929 // Output known-0 bits are known if clear or set in both the low clear bits
1930 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1931 // low 3 bits clear.
1932 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
1933 BitWidth - Mask.countLeadingZeros());
1934 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1935 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1936 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1938 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1939 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1940 KnownZeroOut = std::min(KnownZeroOut,
1941 KnownZero2.countTrailingOnes());
1943 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1947 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1948 const APInt &RA = Rem->getAPIntValue().abs();
1949 if (RA.isPowerOf2()) {
1950 APInt LowBits = RA - 1;
1951 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1952 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1954 // The low bits of the first operand are unchanged by the srem.
1955 KnownZero = KnownZero2 & LowBits;
1956 KnownOne = KnownOne2 & LowBits;
1958 // If the first operand is non-negative or has all low bits zero, then
1959 // the upper bits are all zero.
1960 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1961 KnownZero |= ~LowBits;
1963 // If the first operand is negative and not all low bits are zero, then
1964 // the upper bits are all one.
1965 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1966 KnownOne |= ~LowBits;
1971 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1976 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1977 const APInt &RA = Rem->getAPIntValue();
1978 if (RA.isPowerOf2()) {
1979 APInt LowBits = (RA - 1);
1980 APInt Mask2 = LowBits & Mask;
1981 KnownZero |= ~LowBits & Mask;
1982 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1983 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1988 // Since the result is less than or equal to either operand, any leading
1989 // zero bits in either operand must also exist in the result.
1990 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1991 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne,
1993 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2,
1996 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1997 KnownZero2.countLeadingOnes());
1999 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
2003 // Allow the target to implement this method for its nodes.
2004 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
2005 case ISD::INTRINSIC_WO_CHAIN:
2006 case ISD::INTRINSIC_W_CHAIN:
2007 case ISD::INTRINSIC_VOID:
2008 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this,
2015 /// ComputeNumSignBits - Return the number of times the sign bit of the
2016 /// register is replicated into the other bits. We know that at least 1 bit
2017 /// is always equal to the sign bit (itself), but other cases can give us
2018 /// information. For example, immediately after an "SRA X, 2", we know that
2019 /// the top 3 bits are all equal to each other, so we return 3.
2020 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
2021 EVT VT = Op.getValueType();
2022 assert(VT.isInteger() && "Invalid VT!");
2023 unsigned VTBits = VT.getScalarType().getSizeInBits();
2025 unsigned FirstAnswer = 1;
2028 return 1; // Limit search depth.
2030 switch (Op.getOpcode()) {
2032 case ISD::AssertSext:
2033 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2034 return VTBits-Tmp+1;
2035 case ISD::AssertZext:
2036 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2039 case ISD::Constant: {
2040 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
2041 // If negative, return # leading ones.
2042 if (Val.isNegative())
2043 return Val.countLeadingOnes();
2045 // Return # leading zeros.
2046 return Val.countLeadingZeros();
2049 case ISD::SIGN_EXTEND:
2050 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
2051 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
2053 case ISD::SIGN_EXTEND_INREG:
2054 // Max of the input and what this extends.
2056 cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarType().getSizeInBits();
2059 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2060 return std::max(Tmp, Tmp2);
2063 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2064 // SRA X, C -> adds C sign bits.
2065 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2066 Tmp += C->getZExtValue();
2067 if (Tmp > VTBits) Tmp = VTBits;
2071 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2072 // shl destroys sign bits.
2073 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2074 if (C->getZExtValue() >= VTBits || // Bad shift.
2075 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2076 return Tmp - C->getZExtValue();
2081 case ISD::XOR: // NOT is handled here.
2082 // Logical binary ops preserve the number of sign bits at the worst.
2083 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2085 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2086 FirstAnswer = std::min(Tmp, Tmp2);
2087 // We computed what we know about the sign bits as our first
2088 // answer. Now proceed to the generic code that uses
2089 // ComputeMaskedBits, and pick whichever answer is better.
2094 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2095 if (Tmp == 1) return 1; // Early out.
2096 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
2097 return std::min(Tmp, Tmp2);
2105 if (Op.getResNo() != 1)
2107 // The boolean result conforms to getBooleanContents. Fall through.
2109 // If setcc returns 0/-1, all bits are sign bits.
2110 if (TLI.getBooleanContents() ==
2111 TargetLowering::ZeroOrNegativeOneBooleanContent)
2116 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2117 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
2119 // Handle rotate right by N like a rotate left by 32-N.
2120 if (Op.getOpcode() == ISD::ROTR)
2121 RotAmt = (VTBits-RotAmt) & (VTBits-1);
2123 // If we aren't rotating out all of the known-in sign bits, return the
2124 // number that are left. This handles rotl(sext(x), 1) for example.
2125 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2126 if (Tmp > RotAmt+1) return Tmp-RotAmt;
2130 // Add can have at most one carry bit. Thus we know that the output
2131 // is, at worst, one more bit than the inputs.
2132 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2133 if (Tmp == 1) return 1; // Early out.
2135 // Special case decrementing a value (ADD X, -1):
2136 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
2137 if (CRHS->isAllOnesValue()) {
2138 APInt KnownZero, KnownOne;
2139 APInt Mask = APInt::getAllOnesValue(VTBits);
2140 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2142 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2144 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2147 // If we are subtracting one from a positive number, there is no carry
2148 // out of the result.
2149 if (KnownZero.isNegative())
2153 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2154 if (Tmp2 == 1) return 1;
2155 return std::min(Tmp, Tmp2)-1;
2159 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2160 if (Tmp2 == 1) return 1;
2163 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
2164 if (CLHS->isNullValue()) {
2165 APInt KnownZero, KnownOne;
2166 APInt Mask = APInt::getAllOnesValue(VTBits);
2167 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2168 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2170 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2173 // If the input is known to be positive (the sign bit is known clear),
2174 // the output of the NEG has the same number of sign bits as the input.
2175 if (KnownZero.isNegative())
2178 // Otherwise, we treat this like a SUB.
2181 // Sub can have at most one carry bit. Thus we know that the output
2182 // is, at worst, one more bit than the inputs.
2183 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2184 if (Tmp == 1) return 1; // Early out.
2185 return std::min(Tmp, Tmp2)-1;
2188 // FIXME: it's tricky to do anything useful for this, but it is an important
2189 // case for targets like X86.
2193 // Handle LOADX separately here. EXTLOAD case will fallthrough.
2194 if (Op.getOpcode() == ISD::LOAD) {
2195 LoadSDNode *LD = cast<LoadSDNode>(Op);
2196 unsigned ExtType = LD->getExtensionType();
2199 case ISD::SEXTLOAD: // '17' bits known
2200 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2201 return VTBits-Tmp+1;
2202 case ISD::ZEXTLOAD: // '16' bits known
2203 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2208 // Allow the target to implement this method for its nodes.
2209 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
2210 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
2211 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
2212 Op.getOpcode() == ISD::INTRINSIC_VOID) {
2213 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
2214 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
2217 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2218 // use this information.
2219 APInt KnownZero, KnownOne;
2220 APInt Mask = APInt::getAllOnesValue(VTBits);
2221 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
2223 if (KnownZero.isNegative()) { // sign bit is 0
2225 } else if (KnownOne.isNegative()) { // sign bit is 1;
2232 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2233 // the number of identical bits in the top of the input value.
2235 Mask <<= Mask.getBitWidth()-VTBits;
2236 // Return # leading zeros. We use 'min' here in case Val was zero before
2237 // shifting. We don't want to return '64' as for an i32 "0".
2238 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2241 bool SelectionDAG::isKnownNeverNaN(SDValue Op) const {
2242 // If we're told that NaNs won't happen, assume they won't.
2243 if (FiniteOnlyFPMath())
2246 // If the value is a constant, we can obviously see if it is a NaN or not.
2247 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2248 return !C->getValueAPF().isNaN();
2250 // TODO: Recognize more cases here.
2255 bool SelectionDAG::isKnownNeverZero(SDValue Op) const {
2256 // If the value is a constant, we can obviously see if it is a zero or not.
2257 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2258 return !C->isZero();
2260 // TODO: Recognize more cases here.
2265 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
2266 // Check the obvious case.
2267 if (A == B) return true;
2269 // For for negative and positive zero.
2270 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A))
2271 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B))
2272 if (CA->isZero() && CB->isZero()) return true;
2274 // Otherwise they may not be equal.
2278 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2279 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2280 if (!GA) return false;
2281 if (GA->getOffset() != 0) return false;
2282 const GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2283 if (!GV) return false;
2284 return MF->getMMI().hasDebugInfo();
2288 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2289 /// element of the result of the vector shuffle.
2290 SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N,
2292 EVT VT = N->getValueType(0);
2293 DebugLoc dl = N->getDebugLoc();
2294 if (N->getMaskElt(i) < 0)
2295 return getUNDEF(VT.getVectorElementType());
2296 unsigned Index = N->getMaskElt(i);
2297 unsigned NumElems = VT.getVectorNumElements();
2298 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2301 if (V.getOpcode() == ISD::BIT_CONVERT) {
2302 V = V.getOperand(0);
2303 EVT VVT = V.getValueType();
2304 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems)
2307 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2308 return (Index == 0) ? V.getOperand(0)
2309 : getUNDEF(VT.getVectorElementType());
2310 if (V.getOpcode() == ISD::BUILD_VECTOR)
2311 return V.getOperand(Index);
2312 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V))
2313 return getShuffleScalarElt(SVN, Index);
2318 /// getNode - Gets or creates the specified node.
2320 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) {
2321 FoldingSetNodeID ID;
2322 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2324 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2325 return SDValue(E, 0);
2327 SDNode *N = new (NodeAllocator) SDNode(Opcode, DL, getVTList(VT));
2328 CSEMap.InsertNode(N, IP);
2330 AllNodes.push_back(N);
2334 return SDValue(N, 0);
2337 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
2338 EVT VT, SDValue Operand) {
2339 // Constant fold unary operations with an integer constant operand.
2340 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2341 const APInt &Val = C->getAPIntValue();
2344 case ISD::SIGN_EXTEND:
2345 return getConstant(APInt(Val).sextOrTrunc(VT.getSizeInBits()), VT);
2346 case ISD::ANY_EXTEND:
2347 case ISD::ZERO_EXTEND:
2349 return getConstant(APInt(Val).zextOrTrunc(VT.getSizeInBits()), VT);
2350 case ISD::UINT_TO_FP:
2351 case ISD::SINT_TO_FP: {
2352 const uint64_t zero[] = {0, 0};
2353 // No compile time operations on ppcf128.
2354 if (VT == MVT::ppcf128) break;
2355 APFloat apf = APFloat(APInt(VT.getSizeInBits(), 2, zero));
2356 (void)apf.convertFromAPInt(Val,
2357 Opcode==ISD::SINT_TO_FP,
2358 APFloat::rmNearestTiesToEven);
2359 return getConstantFP(apf, VT);
2361 case ISD::BIT_CONVERT:
2362 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2363 return getConstantFP(Val.bitsToFloat(), VT);
2364 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2365 return getConstantFP(Val.bitsToDouble(), VT);
2368 return getConstant(Val.byteSwap(), VT);
2370 return getConstant(Val.countPopulation(), VT);
2372 return getConstant(Val.countLeadingZeros(), VT);
2374 return getConstant(Val.countTrailingZeros(), VT);
2378 // Constant fold unary operations with a floating point constant operand.
2379 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2380 APFloat V = C->getValueAPF(); // make copy
2381 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2385 return getConstantFP(V, VT);
2388 return getConstantFP(V, VT);
2390 case ISD::FP_EXTEND: {
2392 // This can return overflow, underflow, or inexact; we don't care.
2393 // FIXME need to be more flexible about rounding mode.
2394 (void)V.convert(*EVTToAPFloatSemantics(VT),
2395 APFloat::rmNearestTiesToEven, &ignored);
2396 return getConstantFP(V, VT);
2398 case ISD::FP_TO_SINT:
2399 case ISD::FP_TO_UINT: {
2402 assert(integerPartWidth >= 64);
2403 // FIXME need to be more flexible about rounding mode.
2404 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
2405 Opcode==ISD::FP_TO_SINT,
2406 APFloat::rmTowardZero, &ignored);
2407 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2409 APInt api(VT.getSizeInBits(), 2, x);
2410 return getConstant(api, VT);
2412 case ISD::BIT_CONVERT:
2413 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2414 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2415 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2416 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2422 unsigned OpOpcode = Operand.getNode()->getOpcode();
2424 case ISD::TokenFactor:
2425 case ISD::MERGE_VALUES:
2426 case ISD::CONCAT_VECTORS:
2427 return Operand; // Factor, merge or concat of one node? No need.
2428 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
2429 case ISD::FP_EXTEND:
2430 assert(VT.isFloatingPoint() &&
2431 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2432 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2433 assert((!VT.isVector() ||
2434 VT.getVectorNumElements() ==
2435 Operand.getValueType().getVectorNumElements()) &&
2436 "Vector element count mismatch!");
2437 if (Operand.getOpcode() == ISD::UNDEF)
2438 return getUNDEF(VT);
2440 case ISD::SIGN_EXTEND:
2441 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2442 "Invalid SIGN_EXTEND!");
2443 if (Operand.getValueType() == VT) return Operand; // noop extension
2444 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2445 "Invalid sext node, dst < src!");
2446 assert((!VT.isVector() ||
2447 VT.getVectorNumElements() ==
2448 Operand.getValueType().getVectorNumElements()) &&
2449 "Vector element count mismatch!");
2450 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2451 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2453 case ISD::ZERO_EXTEND:
2454 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2455 "Invalid ZERO_EXTEND!");
2456 if (Operand.getValueType() == VT) return Operand; // noop extension
2457 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2458 "Invalid zext node, dst < src!");
2459 assert((!VT.isVector() ||
2460 VT.getVectorNumElements() ==
2461 Operand.getValueType().getVectorNumElements()) &&
2462 "Vector element count mismatch!");
2463 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2464 return getNode(ISD::ZERO_EXTEND, DL, VT,
2465 Operand.getNode()->getOperand(0));
2467 case ISD::ANY_EXTEND:
2468 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2469 "Invalid ANY_EXTEND!");
2470 if (Operand.getValueType() == VT) return Operand; // noop extension
2471 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2472 "Invalid anyext node, dst < src!");
2473 assert((!VT.isVector() ||
2474 VT.getVectorNumElements() ==
2475 Operand.getValueType().getVectorNumElements()) &&
2476 "Vector element count mismatch!");
2477 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2478 OpOpcode == ISD::ANY_EXTEND)
2479 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2480 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2483 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2484 "Invalid TRUNCATE!");
2485 if (Operand.getValueType() == VT) return Operand; // noop truncate
2486 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) &&
2487 "Invalid truncate node, src < dst!");
2488 assert((!VT.isVector() ||
2489 VT.getVectorNumElements() ==
2490 Operand.getValueType().getVectorNumElements()) &&
2491 "Vector element count mismatch!");
2492 if (OpOpcode == ISD::TRUNCATE)
2493 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2494 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2495 OpOpcode == ISD::ANY_EXTEND) {
2496 // If the source is smaller than the dest, we still need an extend.
2497 if (Operand.getNode()->getOperand(0).getValueType().getScalarType()
2498 .bitsLT(VT.getScalarType()))
2499 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2500 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2501 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2503 return Operand.getNode()->getOperand(0);
2506 case ISD::BIT_CONVERT:
2507 // Basic sanity checking.
2508 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2509 && "Cannot BIT_CONVERT between types of different sizes!");
2510 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2511 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2512 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0));
2513 if (OpOpcode == ISD::UNDEF)
2514 return getUNDEF(VT);
2516 case ISD::SCALAR_TO_VECTOR:
2517 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2518 (VT.getVectorElementType() == Operand.getValueType() ||
2519 (VT.getVectorElementType().isInteger() &&
2520 Operand.getValueType().isInteger() &&
2521 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
2522 "Illegal SCALAR_TO_VECTOR node!");
2523 if (OpOpcode == ISD::UNDEF)
2524 return getUNDEF(VT);
2525 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2526 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2527 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2528 Operand.getConstantOperandVal(1) == 0 &&
2529 Operand.getOperand(0).getValueType() == VT)
2530 return Operand.getOperand(0);
2533 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0
2534 if (UnsafeFPMath && OpOpcode == ISD::FSUB)
2535 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1),
2536 Operand.getNode()->getOperand(0));
2537 if (OpOpcode == ISD::FNEG) // --X -> X
2538 return Operand.getNode()->getOperand(0);
2541 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2542 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0));
2547 SDVTList VTs = getVTList(VT);
2548 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2549 FoldingSetNodeID ID;
2550 SDValue Ops[1] = { Operand };
2551 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2553 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2554 return SDValue(E, 0);
2556 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTs, Operand);
2557 CSEMap.InsertNode(N, IP);
2559 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTs, Operand);
2562 AllNodes.push_back(N);
2566 return SDValue(N, 0);
2569 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2571 ConstantSDNode *Cst1,
2572 ConstantSDNode *Cst2) {
2573 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2576 case ISD::ADD: return getConstant(C1 + C2, VT);
2577 case ISD::SUB: return getConstant(C1 - C2, VT);
2578 case ISD::MUL: return getConstant(C1 * C2, VT);
2580 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2583 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2586 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2589 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2591 case ISD::AND: return getConstant(C1 & C2, VT);
2592 case ISD::OR: return getConstant(C1 | C2, VT);
2593 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2594 case ISD::SHL: return getConstant(C1 << C2, VT);
2595 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2596 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2597 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2598 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2605 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2606 SDValue N1, SDValue N2) {
2607 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2608 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2611 case ISD::TokenFactor:
2612 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2613 N2.getValueType() == MVT::Other && "Invalid token factor!");
2614 // Fold trivial token factors.
2615 if (N1.getOpcode() == ISD::EntryToken) return N2;
2616 if (N2.getOpcode() == ISD::EntryToken) return N1;
2617 if (N1 == N2) return N1;
2619 case ISD::CONCAT_VECTORS:
2620 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2621 // one big BUILD_VECTOR.
2622 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2623 N2.getOpcode() == ISD::BUILD_VECTOR) {
2624 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2625 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2626 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2630 assert(VT.isInteger() && "This operator does not apply to FP types!");
2631 assert(N1.getValueType() == N2.getValueType() &&
2632 N1.getValueType() == VT && "Binary operator types must match!");
2633 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2634 // worth handling here.
2635 if (N2C && N2C->isNullValue())
2637 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2644 assert(VT.isInteger() && "This operator does not apply to FP types!");
2645 assert(N1.getValueType() == N2.getValueType() &&
2646 N1.getValueType() == VT && "Binary operator types must match!");
2647 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2648 // it's worth handling here.
2649 if (N2C && N2C->isNullValue())
2659 assert(VT.isInteger() && "This operator does not apply to FP types!");
2660 assert(N1.getValueType() == N2.getValueType() &&
2661 N1.getValueType() == VT && "Binary operator types must match!");
2669 if (Opcode == ISD::FADD) {
2671 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1))
2672 if (CFP->getValueAPF().isZero())
2675 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2676 if (CFP->getValueAPF().isZero())
2678 } else if (Opcode == ISD::FSUB) {
2680 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2681 if (CFP->getValueAPF().isZero())
2685 assert(VT.isFloatingPoint() && "This operator only applies to FP types!");
2686 assert(N1.getValueType() == N2.getValueType() &&
2687 N1.getValueType() == VT && "Binary operator types must match!");
2689 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2690 assert(N1.getValueType() == VT &&
2691 N1.getValueType().isFloatingPoint() &&
2692 N2.getValueType().isFloatingPoint() &&
2693 "Invalid FCOPYSIGN!");
2700 assert(VT == N1.getValueType() &&
2701 "Shift operators return type must be the same as their first arg");
2702 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2703 "Shifts only work on integers");
2705 // Always fold shifts of i1 values so the code generator doesn't need to
2706 // handle them. Since we know the size of the shift has to be less than the
2707 // size of the value, the shift/rotate count is guaranteed to be zero.
2710 if (N2C && N2C->isNullValue())
2713 case ISD::FP_ROUND_INREG: {
2714 EVT EVT = cast<VTSDNode>(N2)->getVT();
2715 assert(VT == N1.getValueType() && "Not an inreg round!");
2716 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2717 "Cannot FP_ROUND_INREG integer types");
2718 assert(EVT.isVector() == VT.isVector() &&
2719 "FP_ROUND_INREG type should be vector iff the operand "
2721 assert((!EVT.isVector() ||
2722 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2723 "Vector element counts must match in FP_ROUND_INREG");
2724 assert(EVT.bitsLE(VT) && "Not rounding down!");
2725 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2729 assert(VT.isFloatingPoint() &&
2730 N1.getValueType().isFloatingPoint() &&
2731 VT.bitsLE(N1.getValueType()) &&
2732 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2733 if (N1.getValueType() == VT) return N1; // noop conversion.
2735 case ISD::AssertSext:
2736 case ISD::AssertZext: {
2737 EVT EVT = cast<VTSDNode>(N2)->getVT();
2738 assert(VT == N1.getValueType() && "Not an inreg extend!");
2739 assert(VT.isInteger() && EVT.isInteger() &&
2740 "Cannot *_EXTEND_INREG FP types");
2741 assert(!EVT.isVector() &&
2742 "AssertSExt/AssertZExt type should be the vector element type "
2743 "rather than the vector type!");
2744 assert(EVT.bitsLE(VT) && "Not extending!");
2745 if (VT == EVT) return N1; // noop assertion.
2748 case ISD::SIGN_EXTEND_INREG: {
2749 EVT EVT = cast<VTSDNode>(N2)->getVT();
2750 assert(VT == N1.getValueType() && "Not an inreg extend!");
2751 assert(VT.isInteger() && EVT.isInteger() &&
2752 "Cannot *_EXTEND_INREG FP types");
2753 assert(EVT.isVector() == VT.isVector() &&
2754 "SIGN_EXTEND_INREG type should be vector iff the operand "
2756 assert((!EVT.isVector() ||
2757 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2758 "Vector element counts must match in SIGN_EXTEND_INREG");
2759 assert(EVT.bitsLE(VT) && "Not extending!");
2760 if (EVT == VT) return N1; // Not actually extending
2763 APInt Val = N1C->getAPIntValue();
2764 unsigned FromBits = EVT.getScalarType().getSizeInBits();
2765 Val <<= Val.getBitWidth()-FromBits;
2766 Val = Val.ashr(Val.getBitWidth()-FromBits);
2767 return getConstant(Val, VT);
2771 case ISD::EXTRACT_VECTOR_ELT:
2772 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2773 if (N1.getOpcode() == ISD::UNDEF)
2774 return getUNDEF(VT);
2776 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2777 // expanding copies of large vectors from registers.
2779 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2780 N1.getNumOperands() > 0) {
2782 N1.getOperand(0).getValueType().getVectorNumElements();
2783 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
2784 N1.getOperand(N2C->getZExtValue() / Factor),
2785 getConstant(N2C->getZExtValue() % Factor,
2786 N2.getValueType()));
2789 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2790 // expanding large vector constants.
2791 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) {
2792 SDValue Elt = N1.getOperand(N2C->getZExtValue());
2793 EVT VEltTy = N1.getValueType().getVectorElementType();
2794 if (Elt.getValueType() != VEltTy) {
2795 // If the vector element type is not legal, the BUILD_VECTOR operands
2796 // are promoted and implicitly truncated. Make that explicit here.
2797 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt);
2800 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT
2801 // result is implicitly extended.
2802 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt);
2807 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2808 // operations are lowered to scalars.
2809 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2810 // If the indices are the same, return the inserted element else
2811 // if the indices are known different, extract the element from
2812 // the original vector.
2813 SDValue N1Op2 = N1.getOperand(2);
2814 ConstantSDNode *N1Op2C = dyn_cast<ConstantSDNode>(N1Op2.getNode());
2816 if (N1Op2C && N2C) {
2817 if (N1Op2C->getZExtValue() == N2C->getZExtValue()) {
2818 if (VT == N1.getOperand(1).getValueType())
2819 return N1.getOperand(1);
2821 return getSExtOrTrunc(N1.getOperand(1), DL, VT);
2824 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
2828 case ISD::EXTRACT_ELEMENT:
2829 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2830 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2831 (N1.getValueType().isInteger() == VT.isInteger()) &&
2832 "Wrong types for EXTRACT_ELEMENT!");
2834 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2835 // 64-bit integers into 32-bit parts. Instead of building the extract of
2836 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2837 if (N1.getOpcode() == ISD::BUILD_PAIR)
2838 return N1.getOperand(N2C->getZExtValue());
2840 // EXTRACT_ELEMENT of a constant int is also very common.
2841 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2842 unsigned ElementSize = VT.getSizeInBits();
2843 unsigned Shift = ElementSize * N2C->getZExtValue();
2844 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2845 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2848 case ISD::EXTRACT_SUBVECTOR:
2849 if (N1.getValueType() == VT) // Trivial extraction.
2856 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2857 if (SV.getNode()) return SV;
2858 } else { // Cannonicalize constant to RHS if commutative
2859 if (isCommutativeBinOp(Opcode)) {
2860 std::swap(N1C, N2C);
2866 // Constant fold FP operations.
2867 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2868 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2870 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2871 // Cannonicalize constant to RHS if commutative
2872 std::swap(N1CFP, N2CFP);
2874 } else if (N2CFP && VT != MVT::ppcf128) {
2875 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2876 APFloat::opStatus s;
2879 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2880 if (s != APFloat::opInvalidOp)
2881 return getConstantFP(V1, VT);
2884 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2885 if (s!=APFloat::opInvalidOp)
2886 return getConstantFP(V1, VT);
2889 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2890 if (s!=APFloat::opInvalidOp)
2891 return getConstantFP(V1, VT);
2894 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2895 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2896 return getConstantFP(V1, VT);
2899 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2900 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2901 return getConstantFP(V1, VT);
2903 case ISD::FCOPYSIGN:
2905 return getConstantFP(V1, VT);
2911 // Canonicalize an UNDEF to the RHS, even over a constant.
2912 if (N1.getOpcode() == ISD::UNDEF) {
2913 if (isCommutativeBinOp(Opcode)) {
2917 case ISD::FP_ROUND_INREG:
2918 case ISD::SIGN_EXTEND_INREG:
2924 return N1; // fold op(undef, arg2) -> undef
2932 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2933 // For vectors, we can't easily build an all zero vector, just return
2940 // Fold a bunch of operators when the RHS is undef.
2941 if (N2.getOpcode() == ISD::UNDEF) {
2944 if (N1.getOpcode() == ISD::UNDEF)
2945 // Handle undef ^ undef -> 0 special case. This is a common
2947 return getConstant(0, VT);
2957 return N2; // fold op(arg1, undef) -> undef
2971 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2972 // For vectors, we can't easily build an all zero vector, just return
2977 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
2978 // For vectors, we can't easily build an all one vector, just return
2986 // Memoize this node if possible.
2988 SDVTList VTs = getVTList(VT);
2989 if (VT != MVT::Flag) {
2990 SDValue Ops[] = { N1, N2 };
2991 FoldingSetNodeID ID;
2992 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2994 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2995 return SDValue(E, 0);
2997 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTs, N1, N2);
2998 CSEMap.InsertNode(N, IP);
3000 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTs, N1, N2);
3003 AllNodes.push_back(N);
3007 return SDValue(N, 0);
3010 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3011 SDValue N1, SDValue N2, SDValue N3) {
3012 // Perform various simplifications.
3013 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
3014 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
3016 case ISD::CONCAT_VECTORS:
3017 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
3018 // one big BUILD_VECTOR.
3019 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
3020 N2.getOpcode() == ISD::BUILD_VECTOR &&
3021 N3.getOpcode() == ISD::BUILD_VECTOR) {
3022 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
3023 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
3024 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
3025 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
3029 // Use FoldSetCC to simplify SETCC's.
3030 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL);
3031 if (Simp.getNode()) return Simp;
3036 if (N1C->getZExtValue())
3037 return N2; // select true, X, Y -> X
3039 return N3; // select false, X, Y -> Y
3042 if (N2 == N3) return N2; // select C, X, X -> X
3046 if (N2C->getZExtValue()) // Unconditional branch
3047 return getNode(ISD::BR, DL, MVT::Other, N1, N3);
3049 return N1; // Never-taken branch
3052 case ISD::VECTOR_SHUFFLE:
3053 llvm_unreachable("should use getVectorShuffle constructor!");
3055 case ISD::BIT_CONVERT:
3056 // Fold bit_convert nodes from a type to themselves.
3057 if (N1.getValueType() == VT)
3062 // Memoize node if it doesn't produce a flag.
3064 SDVTList VTs = getVTList(VT);
3065 if (VT != MVT::Flag) {
3066 SDValue Ops[] = { N1, N2, N3 };
3067 FoldingSetNodeID ID;
3068 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3070 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3071 return SDValue(E, 0);
3073 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3074 CSEMap.InsertNode(N, IP);
3076 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3079 AllNodes.push_back(N);
3083 return SDValue(N, 0);
3086 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3087 SDValue N1, SDValue N2, SDValue N3,
3089 SDValue Ops[] = { N1, N2, N3, N4 };
3090 return getNode(Opcode, DL, VT, Ops, 4);
3093 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3094 SDValue N1, SDValue N2, SDValue N3,
3095 SDValue N4, SDValue N5) {
3096 SDValue Ops[] = { N1, N2, N3, N4, N5 };
3097 return getNode(Opcode, DL, VT, Ops, 5);
3100 /// getStackArgumentTokenFactor - Compute a TokenFactor to force all
3101 /// the incoming stack arguments to be loaded from the stack.
3102 SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
3103 SmallVector<SDValue, 8> ArgChains;
3105 // Include the original chain at the beginning of the list. When this is
3106 // used by target LowerCall hooks, this helps legalize find the
3107 // CALLSEQ_BEGIN node.
3108 ArgChains.push_back(Chain);
3110 // Add a chain value for each stack argument.
3111 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(),
3112 UE = getEntryNode().getNode()->use_end(); U != UE; ++U)
3113 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
3114 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
3115 if (FI->getIndex() < 0)
3116 ArgChains.push_back(SDValue(L, 1));
3118 // Build a tokenfactor for all the chains.
3119 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
3120 &ArgChains[0], ArgChains.size());
3123 /// getMemsetValue - Vectorized representation of the memset value
3125 static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
3127 assert(Value.getOpcode() != ISD::UNDEF);
3129 unsigned NumBits = VT.getScalarType().getSizeInBits();
3130 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3131 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
3133 for (unsigned i = NumBits; i > 8; i >>= 1) {
3134 Val = (Val << Shift) | Val;
3138 return DAG.getConstant(Val, VT);
3139 return DAG.getConstantFP(APFloat(Val), VT);
3142 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3143 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
3145 for (unsigned i = NumBits; i > 8; i >>= 1) {
3146 Value = DAG.getNode(ISD::OR, dl, VT,
3147 DAG.getNode(ISD::SHL, dl, VT, Value,
3148 DAG.getConstant(Shift,
3149 TLI.getShiftAmountTy())),
3157 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3158 /// used when a memcpy is turned into a memset when the source is a constant
3160 static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG,
3161 const TargetLowering &TLI,
3162 std::string &Str, unsigned Offset) {
3163 // Handle vector with all elements zero.
3166 return DAG.getConstant(0, VT);
3167 else if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
3168 VT.getSimpleVT().SimpleTy == MVT::f64)
3169 return DAG.getConstantFP(0.0, VT);
3170 else if (VT.isVector()) {
3171 unsigned NumElts = VT.getVectorNumElements();
3172 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3173 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3174 DAG.getConstant(0, EVT::getVectorVT(*DAG.getContext(),
3177 llvm_unreachable("Expected type!");
3180 assert(!VT.isVector() && "Can't handle vector type here!");
3181 unsigned NumBits = VT.getSizeInBits();
3182 unsigned MSB = NumBits / 8;
3184 if (TLI.isLittleEndian())
3185 Offset = Offset + MSB - 1;
3186 for (unsigned i = 0; i != MSB; ++i) {
3187 Val = (Val << 8) | (unsigned char)Str[Offset];
3188 Offset += TLI.isLittleEndian() ? -1 : 1;
3190 return DAG.getConstant(Val, VT);
3193 /// getMemBasePlusOffset - Returns base and offset node for the
3195 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
3196 SelectionDAG &DAG) {
3197 EVT VT = Base.getValueType();
3198 return DAG.getNode(ISD::ADD, Base.getDebugLoc(),
3199 VT, Base, DAG.getConstant(Offset, VT));
3202 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
3204 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
3205 unsigned SrcDelta = 0;
3206 GlobalAddressSDNode *G = NULL;
3207 if (Src.getOpcode() == ISD::GlobalAddress)
3208 G = cast<GlobalAddressSDNode>(Src);
3209 else if (Src.getOpcode() == ISD::ADD &&
3210 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3211 Src.getOperand(1).getOpcode() == ISD::Constant) {
3212 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
3213 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
3218 const GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3219 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
3225 /// FindOptimalMemOpLowering - Determines the optimial series memory ops
3226 /// to replace the memset / memcpy. Return true if the number of memory ops
3227 /// is below the threshold. It returns the types of the sequence of
3228 /// memory ops to perform memset / memcpy by reference.
3229 static bool FindOptimalMemOpLowering(std::vector<EVT> &MemOps,
3230 unsigned Limit, uint64_t Size,
3231 unsigned DstAlign, unsigned SrcAlign,
3232 bool NonScalarIntSafe,
3235 const TargetLowering &TLI) {
3236 assert((SrcAlign == 0 || SrcAlign >= DstAlign) &&
3237 "Expecting memcpy / memset source to meet alignment requirement!");
3238 // If 'SrcAlign' is zero, that means the memory operation does not need load
3239 // the value, i.e. memset or memcpy from constant string. Otherwise, it's
3240 // the inferred alignment of the source. 'DstAlign', on the other hand, is the
3241 // specified alignment of the memory operation. If it is zero, that means
3242 // it's possible to change the alignment of the destination. 'MemcpyStrSrc'
3243 // indicates whether the memcpy source is constant so it does not need to be
3245 EVT VT = TLI.getOptimalMemOpType(Size, DstAlign, SrcAlign,
3246 NonScalarIntSafe, MemcpyStrSrc,
3247 DAG.getMachineFunction());
3249 if (VT == MVT::Other) {
3250 if (DstAlign >= TLI.getTargetData()->getPointerPrefAlignment() ||
3251 TLI.allowsUnalignedMemoryAccesses(VT)) {
3252 VT = TLI.getPointerTy();
3254 switch (DstAlign & 7) {
3255 case 0: VT = MVT::i64; break;
3256 case 4: VT = MVT::i32; break;
3257 case 2: VT = MVT::i16; break;
3258 default: VT = MVT::i8; break;
3263 while (!TLI.isTypeLegal(LVT))
3264 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1);
3265 assert(LVT.isInteger());
3271 // If we're optimizing for size, and there is a limit, bump the maximum number
3272 // of operations inserted down to 4. This is a wild guess that approximates
3273 // the size of a call to memcpy or memset (3 arguments + call).
3275 const Function *F = DAG.getMachineFunction().getFunction();
3276 if (F->hasFnAttr(Attribute::OptimizeForSize))
3280 unsigned NumMemOps = 0;
3282 unsigned VTSize = VT.getSizeInBits() / 8;
3283 while (VTSize > Size) {
3284 // For now, only use non-vector load / store's for the left-over pieces.
3285 if (VT.isVector() || VT.isFloatingPoint()) {
3287 while (!TLI.isTypeLegal(VT))
3288 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3289 VTSize = VT.getSizeInBits() / 8;
3291 // This can result in a type that is not legal on the target, e.g.
3292 // 1 or 2 bytes on PPC.
3293 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3298 if (++NumMemOps > Limit)
3300 MemOps.push_back(VT);
3307 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3308 SDValue Chain, SDValue Dst,
3309 SDValue Src, uint64_t Size,
3310 unsigned Align, bool isVol,
3312 const Value *DstSV, uint64_t DstSVOff,
3313 const Value *SrcSV, uint64_t SrcSVOff) {
3314 // Turn a memcpy of undef to nop.
3315 if (Src.getOpcode() == ISD::UNDEF)
3318 // Expand memcpy to a series of load and store ops if the size operand falls
3319 // below a certain threshold.
3320 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3321 std::vector<EVT> MemOps;
3322 bool DstAlignCanChange = false;
3323 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3324 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
3325 if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
3326 DstAlignCanChange = true;
3327 unsigned SrcAlign = DAG.InferPtrAlignment(Src);
3328 if (Align > SrcAlign)
3331 bool CopyFromStr = isMemSrcFromString(Src, Str);
3332 bool isZeroStr = CopyFromStr && Str.empty();
3333 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemcpy();
3335 if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
3336 (DstAlignCanChange ? 0 : Align),
3337 (isZeroStr ? 0 : SrcAlign),
3338 true, CopyFromStr, DAG, TLI))
3341 if (DstAlignCanChange) {
3342 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
3343 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3344 if (NewAlign > Align) {
3345 // Give the stack frame object a larger alignment if needed.
3346 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign)
3347 MFI->setObjectAlignment(FI->getIndex(), NewAlign);
3352 SmallVector<SDValue, 8> OutChains;
3353 unsigned NumMemOps = MemOps.size();
3354 uint64_t SrcOff = 0, DstOff = 0;
3355 for (unsigned i = 0; i != NumMemOps; ++i) {
3357 unsigned VTSize = VT.getSizeInBits() / 8;
3358 SDValue Value, Store;
3361 (isZeroStr || (VT.isInteger() && !VT.isVector()))) {
3362 // It's unlikely a store of a vector immediate can be done in a single
3363 // instruction. It would require a load from a constantpool first.
3364 // We only handle zero vectors here.
3365 // FIXME: Handle other cases where store of vector immediate is done in
3366 // a single instruction.
3367 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff);
3368 Store = DAG.getStore(Chain, dl, Value,
3369 getMemBasePlusOffset(Dst, DstOff, DAG),
3370 DstSV, DstSVOff + DstOff, isVol, false, Align);
3372 // The type might not be legal for the target. This should only happen
3373 // if the type is smaller than a legal type, as on PPC, so the right
3374 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
3375 // to Load/Store if NVT==VT.
3376 // FIXME does the case above also need this?
3377 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
3378 assert(NVT.bitsGE(VT));
3379 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain,
3380 getMemBasePlusOffset(Src, SrcOff, DAG),
3381 SrcSV, SrcSVOff + SrcOff, VT, isVol, false,
3382 MinAlign(SrcAlign, SrcOff));
3383 Store = DAG.getTruncStore(Chain, dl, Value,
3384 getMemBasePlusOffset(Dst, DstOff, DAG),
3385 DstSV, DstSVOff + DstOff, VT, isVol, false,
3388 OutChains.push_back(Store);
3393 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3394 &OutChains[0], OutChains.size());
3397 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3398 SDValue Chain, SDValue Dst,
3399 SDValue Src, uint64_t Size,
3400 unsigned Align, bool isVol,
3402 const Value *DstSV, uint64_t DstSVOff,
3403 const Value *SrcSV, uint64_t SrcSVOff) {
3404 // Turn a memmove of undef to nop.
3405 if (Src.getOpcode() == ISD::UNDEF)
3408 // Expand memmove to a series of load and store ops if the size operand falls
3409 // below a certain threshold.
3410 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3411 std::vector<EVT> MemOps;
3412 bool DstAlignCanChange = false;
3413 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3414 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
3415 if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
3416 DstAlignCanChange = true;
3417 unsigned SrcAlign = DAG.InferPtrAlignment(Src);
3418 if (Align > SrcAlign)
3420 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemmove();
3422 if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
3423 (DstAlignCanChange ? 0 : Align),
3424 SrcAlign, true, false, DAG, TLI))
3427 if (DstAlignCanChange) {
3428 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
3429 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3430 if (NewAlign > Align) {
3431 // Give the stack frame object a larger alignment if needed.
3432 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign)
3433 MFI->setObjectAlignment(FI->getIndex(), NewAlign);
3438 uint64_t SrcOff = 0, DstOff = 0;
3439 SmallVector<SDValue, 8> LoadValues;
3440 SmallVector<SDValue, 8> LoadChains;
3441 SmallVector<SDValue, 8> OutChains;
3442 unsigned NumMemOps = MemOps.size();
3443 for (unsigned i = 0; i < NumMemOps; i++) {
3445 unsigned VTSize = VT.getSizeInBits() / 8;
3446 SDValue Value, Store;
3448 Value = DAG.getLoad(VT, dl, Chain,
3449 getMemBasePlusOffset(Src, SrcOff, DAG),
3450 SrcSV, SrcSVOff + SrcOff, isVol, false, SrcAlign);
3451 LoadValues.push_back(Value);
3452 LoadChains.push_back(Value.getValue(1));
3455 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3456 &LoadChains[0], LoadChains.size());
3458 for (unsigned i = 0; i < NumMemOps; i++) {
3460 unsigned VTSize = VT.getSizeInBits() / 8;
3461 SDValue Value, Store;
3463 Store = DAG.getStore(Chain, dl, LoadValues[i],
3464 getMemBasePlusOffset(Dst, DstOff, DAG),
3465 DstSV, DstSVOff + DstOff, isVol, false, Align);
3466 OutChains.push_back(Store);
3470 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3471 &OutChains[0], OutChains.size());
3474 static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl,
3475 SDValue Chain, SDValue Dst,
3476 SDValue Src, uint64_t Size,
3477 unsigned Align, bool isVol,
3478 const Value *DstSV, uint64_t DstSVOff) {
3479 // Turn a memset of undef to nop.
3480 if (Src.getOpcode() == ISD::UNDEF)
3483 // Expand memset to a series of load/store ops if the size operand
3484 // falls below a certain threshold.
3485 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3486 std::vector<EVT> MemOps;
3487 bool DstAlignCanChange = false;
3488 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3489 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
3490 if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
3491 DstAlignCanChange = true;
3492 bool NonScalarIntSafe =
3493 isa<ConstantSDNode>(Src) && cast<ConstantSDNode>(Src)->isNullValue();
3494 if (!FindOptimalMemOpLowering(MemOps, TLI.getMaxStoresPerMemset(),
3495 Size, (DstAlignCanChange ? 0 : Align), 0,
3496 NonScalarIntSafe, false, DAG, TLI))
3499 if (DstAlignCanChange) {
3500 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
3501 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3502 if (NewAlign > Align) {
3503 // Give the stack frame object a larger alignment if needed.
3504 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign)
3505 MFI->setObjectAlignment(FI->getIndex(), NewAlign);
3510 SmallVector<SDValue, 8> OutChains;
3511 uint64_t DstOff = 0;
3512 unsigned NumMemOps = MemOps.size();
3513 for (unsigned i = 0; i < NumMemOps; i++) {
3515 unsigned VTSize = VT.getSizeInBits() / 8;
3516 SDValue Value = getMemsetValue(Src, VT, DAG, dl);
3517 SDValue Store = DAG.getStore(Chain, dl, Value,
3518 getMemBasePlusOffset(Dst, DstOff, DAG),
3519 DstSV, DstSVOff + DstOff, isVol, false, 0);
3520 OutChains.push_back(Store);
3524 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3525 &OutChains[0], OutChains.size());
3528 SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst,
3529 SDValue Src, SDValue Size,
3530 unsigned Align, bool isVol, bool AlwaysInline,
3531 const Value *DstSV, uint64_t DstSVOff,
3532 const Value *SrcSV, uint64_t SrcSVOff) {
3534 // Check to see if we should lower the memcpy to loads and stores first.
3535 // For cases within the target-specified limits, this is the best choice.
3536 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3538 // Memcpy with size zero? Just return the original chain.
3539 if (ConstantSize->isNullValue())
3542 SDValue Result = getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3543 ConstantSize->getZExtValue(),Align,
3544 isVol, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3545 if (Result.getNode())
3549 // Then check to see if we should lower the memcpy with target-specific
3550 // code. If the target chooses to do this, this is the next best.
3552 TSI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align,
3553 isVol, AlwaysInline,
3554 DstSV, DstSVOff, SrcSV, SrcSVOff);
3555 if (Result.getNode())
3558 // If we really need inline code and the target declined to provide it,
3559 // use a (potentially long) sequence of loads and stores.
3561 assert(ConstantSize && "AlwaysInline requires a constant size!");
3562 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3563 ConstantSize->getZExtValue(), Align, isVol,
3564 true, DstSV, DstSVOff, SrcSV, SrcSVOff);
3567 // FIXME: If the memcpy is volatile (isVol), lowering it to a plain libc
3568 // memcpy is not guaranteed to be safe. libc memcpys aren't required to
3569 // respect volatile, so they may do things like read or write memory
3570 // beyond the given memory regions. But fixing this isn't easy, and most
3571 // people don't care.
3573 // Emit a library call.
3574 TargetLowering::ArgListTy Args;
3575 TargetLowering::ArgListEntry Entry;
3576 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3577 Entry.Node = Dst; Args.push_back(Entry);
3578 Entry.Node = Src; Args.push_back(Entry);
3579 Entry.Node = Size; Args.push_back(Entry);
3580 // FIXME: pass in DebugLoc
3581 std::pair<SDValue,SDValue> CallResult =
3582 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3583 false, false, false, false, 0,
3584 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false,
3585 /*isReturnValueUsed=*/false,
3586 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY),
3587 TLI.getPointerTy()),
3589 return CallResult.second;
3592 SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst,
3593 SDValue Src, SDValue Size,
3594 unsigned Align, bool isVol,
3595 const Value *DstSV, uint64_t DstSVOff,
3596 const Value *SrcSV, uint64_t SrcSVOff) {
3598 // Check to see if we should lower the memmove to loads and stores first.
3599 // For cases within the target-specified limits, this is the best choice.
3600 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3602 // Memmove with size zero? Just return the original chain.
3603 if (ConstantSize->isNullValue())
3607 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src,
3608 ConstantSize->getZExtValue(), Align, isVol,
3609 false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3610 if (Result.getNode())
3614 // Then check to see if we should lower the memmove with target-specific
3615 // code. If the target chooses to do this, this is the next best.
3617 TSI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align, isVol,
3618 DstSV, DstSVOff, SrcSV, SrcSVOff);
3619 if (Result.getNode())
3622 // FIXME: If the memmove is volatile, lowering it to plain libc memmove may
3623 // not be safe. See memcpy above for more details.
3625 // Emit a library call.
3626 TargetLowering::ArgListTy Args;
3627 TargetLowering::ArgListEntry Entry;
3628 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3629 Entry.Node = Dst; Args.push_back(Entry);
3630 Entry.Node = Src; Args.push_back(Entry);
3631 Entry.Node = Size; Args.push_back(Entry);
3632 // FIXME: pass in DebugLoc
3633 std::pair<SDValue,SDValue> CallResult =
3634 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3635 false, false, false, false, 0,
3636 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false,
3637 /*isReturnValueUsed=*/false,
3638 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE),
3639 TLI.getPointerTy()),
3641 return CallResult.second;
3644 SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst,
3645 SDValue Src, SDValue Size,
3646 unsigned Align, bool isVol,
3647 const Value *DstSV, uint64_t DstSVOff) {
3649 // Check to see if we should lower the memset to stores first.
3650 // For cases within the target-specified limits, this is the best choice.
3651 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3653 // Memset with size zero? Just return the original chain.
3654 if (ConstantSize->isNullValue())
3658 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
3659 Align, isVol, DstSV, DstSVOff);
3661 if (Result.getNode())
3665 // Then check to see if we should lower the memset with target-specific
3666 // code. If the target chooses to do this, this is the next best.
3668 TSI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align, isVol,
3670 if (Result.getNode())
3673 // Emit a library call.
3674 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext());
3675 TargetLowering::ArgListTy Args;
3676 TargetLowering::ArgListEntry Entry;
3677 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3678 Args.push_back(Entry);
3679 // Extend or truncate the argument to be an i32 value for the call.
3680 if (Src.getValueType().bitsGT(MVT::i32))
3681 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
3683 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
3685 Entry.Ty = Type::getInt32Ty(*getContext());
3686 Entry.isSExt = true;
3687 Args.push_back(Entry);
3689 Entry.Ty = IntPtrTy;
3690 Entry.isSExt = false;
3691 Args.push_back(Entry);
3692 // FIXME: pass in DebugLoc
3693 std::pair<SDValue,SDValue> CallResult =
3694 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3695 false, false, false, false, 0,
3696 TLI.getLibcallCallingConv(RTLIB::MEMSET), false,
3697 /*isReturnValueUsed=*/false,
3698 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET),
3699 TLI.getPointerTy()),
3701 return CallResult.second;
3704 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3706 SDValue Ptr, SDValue Cmp,
3707 SDValue Swp, const Value* PtrVal,
3708 unsigned Alignment) {
3709 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3710 Alignment = getEVTAlignment(MemVT);
3712 // Check if the memory reference references a frame index
3714 if (const FrameIndexSDNode *FI =
3715 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3716 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3718 MachineFunction &MF = getMachineFunction();
3719 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3721 // For now, atomics are considered to be volatile always.
3722 Flags |= MachineMemOperand::MOVolatile;
3724 MachineMemOperand *MMO =
3725 MF.getMachineMemOperand(PtrVal, Flags, 0,
3726 MemVT.getStoreSize(), Alignment);
3728 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3731 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3733 SDValue Ptr, SDValue Cmp,
3734 SDValue Swp, MachineMemOperand *MMO) {
3735 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op");
3736 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3738 EVT VT = Cmp.getValueType();
3740 SDVTList VTs = getVTList(VT, MVT::Other);
3741 FoldingSetNodeID ID;
3742 ID.AddInteger(MemVT.getRawBits());
3743 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3744 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3746 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3747 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3748 return SDValue(E, 0);
3750 SDNode *N = new (NodeAllocator) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain,
3751 Ptr, Cmp, Swp, MMO);
3752 CSEMap.InsertNode(N, IP);
3753 AllNodes.push_back(N);
3754 return SDValue(N, 0);
3757 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3759 SDValue Ptr, SDValue Val,
3760 const Value* PtrVal,
3761 unsigned Alignment) {
3762 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3763 Alignment = getEVTAlignment(MemVT);
3765 // Check if the memory reference references a frame index
3767 if (const FrameIndexSDNode *FI =
3768 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3769 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3771 MachineFunction &MF = getMachineFunction();
3772 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3774 // For now, atomics are considered to be volatile always.
3775 Flags |= MachineMemOperand::MOVolatile;
3777 MachineMemOperand *MMO =
3778 MF.getMachineMemOperand(PtrVal, Flags, 0,
3779 MemVT.getStoreSize(), Alignment);
3781 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO);
3784 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3786 SDValue Ptr, SDValue Val,
3787 MachineMemOperand *MMO) {
3788 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
3789 Opcode == ISD::ATOMIC_LOAD_SUB ||
3790 Opcode == ISD::ATOMIC_LOAD_AND ||
3791 Opcode == ISD::ATOMIC_LOAD_OR ||
3792 Opcode == ISD::ATOMIC_LOAD_XOR ||
3793 Opcode == ISD::ATOMIC_LOAD_NAND ||
3794 Opcode == ISD::ATOMIC_LOAD_MIN ||
3795 Opcode == ISD::ATOMIC_LOAD_MAX ||
3796 Opcode == ISD::ATOMIC_LOAD_UMIN ||
3797 Opcode == ISD::ATOMIC_LOAD_UMAX ||
3798 Opcode == ISD::ATOMIC_SWAP) &&
3799 "Invalid Atomic Op");
3801 EVT VT = Val.getValueType();
3803 SDVTList VTs = getVTList(VT, MVT::Other);
3804 FoldingSetNodeID ID;
3805 ID.AddInteger(MemVT.getRawBits());
3806 SDValue Ops[] = {Chain, Ptr, Val};
3807 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3809 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3810 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3811 return SDValue(E, 0);
3813 SDNode *N = new (NodeAllocator) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain,
3815 CSEMap.InsertNode(N, IP);
3816 AllNodes.push_back(N);
3817 return SDValue(N, 0);
3820 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3821 /// Allowed to return something different (and simpler) if Simplify is true.
3822 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3827 SmallVector<EVT, 4> VTs;
3828 VTs.reserve(NumOps);
3829 for (unsigned i = 0; i < NumOps; ++i)
3830 VTs.push_back(Ops[i].getValueType());
3831 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps),
3836 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl,
3837 const EVT *VTs, unsigned NumVTs,
3838 const SDValue *Ops, unsigned NumOps,
3839 EVT MemVT, const Value *srcValue, int SVOff,
3840 unsigned Align, bool Vol,
3841 bool ReadMem, bool WriteMem) {
3842 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps,
3843 MemVT, srcValue, SVOff, Align, Vol,
3848 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3849 const SDValue *Ops, unsigned NumOps,
3850 EVT MemVT, const Value *srcValue, int SVOff,
3851 unsigned Align, bool Vol,
3852 bool ReadMem, bool WriteMem) {
3853 if (Align == 0) // Ensure that codegen never sees alignment 0
3854 Align = getEVTAlignment(MemVT);
3856 MachineFunction &MF = getMachineFunction();
3859 Flags |= MachineMemOperand::MOStore;
3861 Flags |= MachineMemOperand::MOLoad;
3863 Flags |= MachineMemOperand::MOVolatile;
3864 MachineMemOperand *MMO =
3865 MF.getMachineMemOperand(srcValue, Flags, SVOff,
3866 MemVT.getStoreSize(), Align);
3868 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3872 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3873 const SDValue *Ops, unsigned NumOps,
3874 EVT MemVT, MachineMemOperand *MMO) {
3875 assert((Opcode == ISD::INTRINSIC_VOID ||
3876 Opcode == ISD::INTRINSIC_W_CHAIN ||
3877 (Opcode <= INT_MAX &&
3878 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) &&
3879 "Opcode is not a memory-accessing opcode!");
3881 // Memoize the node unless it returns a flag.
3882 MemIntrinsicSDNode *N;
3883 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3884 FoldingSetNodeID ID;
3885 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3887 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3888 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO);
3889 return SDValue(E, 0);
3892 N = new (NodeAllocator) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps,
3894 CSEMap.InsertNode(N, IP);
3896 N = new (NodeAllocator) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps,
3899 AllNodes.push_back(N);
3900 return SDValue(N, 0);
3904 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3905 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3906 SDValue Ptr, SDValue Offset,
3907 const Value *SV, int SVOffset, EVT MemVT,
3908 bool isVolatile, bool isNonTemporal,
3909 unsigned Alignment) {
3910 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3911 Alignment = getEVTAlignment(VT);
3913 // Check if the memory reference references a frame index
3915 if (const FrameIndexSDNode *FI =
3916 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3917 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3919 MachineFunction &MF = getMachineFunction();
3920 unsigned Flags = MachineMemOperand::MOLoad;
3922 Flags |= MachineMemOperand::MOVolatile;
3924 Flags |= MachineMemOperand::MONonTemporal;
3925 MachineMemOperand *MMO =
3926 MF.getMachineMemOperand(SV, Flags, SVOffset,
3927 MemVT.getStoreSize(), Alignment);
3928 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO);
3932 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3933 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3934 SDValue Ptr, SDValue Offset, EVT MemVT,
3935 MachineMemOperand *MMO) {
3937 ExtType = ISD::NON_EXTLOAD;
3938 } else if (ExtType == ISD::NON_EXTLOAD) {
3939 assert(VT == MemVT && "Non-extending load from different memory type!");
3942 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) &&
3943 "Should only be an extending load, not truncating!");
3944 assert(VT.isInteger() == MemVT.isInteger() &&
3945 "Cannot convert from FP to Int or Int -> FP!");
3946 assert(VT.isVector() == MemVT.isVector() &&
3947 "Cannot use trunc store to convert to or from a vector!");
3948 assert((!VT.isVector() ||
3949 VT.getVectorNumElements() == MemVT.getVectorNumElements()) &&
3950 "Cannot use trunc store to change the number of vector elements!");
3953 bool Indexed = AM != ISD::UNINDEXED;
3954 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3955 "Unindexed load with an offset!");
3957 SDVTList VTs = Indexed ?
3958 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3959 SDValue Ops[] = { Chain, Ptr, Offset };
3960 FoldingSetNodeID ID;
3961 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3962 ID.AddInteger(MemVT.getRawBits());
3963 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile(),
3964 MMO->isNonTemporal()));
3966 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3967 cast<LoadSDNode>(E)->refineAlignment(MMO);
3968 return SDValue(E, 0);
3970 SDNode *N = new (NodeAllocator) LoadSDNode(Ops, dl, VTs, AM, ExtType,
3972 CSEMap.InsertNode(N, IP);
3973 AllNodes.push_back(N);
3974 return SDValue(N, 0);
3977 SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl,
3978 SDValue Chain, SDValue Ptr,
3979 const Value *SV, int SVOffset,
3980 bool isVolatile, bool isNonTemporal,
3981 unsigned Alignment) {
3982 SDValue Undef = getUNDEF(Ptr.getValueType());
3983 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3984 SV, SVOffset, VT, isVolatile, isNonTemporal, Alignment);
3987 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT,
3988 SDValue Chain, SDValue Ptr,
3990 int SVOffset, EVT MemVT,
3991 bool isVolatile, bool isNonTemporal,
3992 unsigned Alignment) {
3993 SDValue Undef = getUNDEF(Ptr.getValueType());
3994 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef,
3995 SV, SVOffset, MemVT, isVolatile, isNonTemporal, Alignment);
3999 SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base,
4000 SDValue Offset, ISD::MemIndexedMode AM) {
4001 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
4002 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
4003 "Load is already a indexed load!");
4004 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(),
4005 LD->getChain(), Base, Offset, LD->getSrcValue(),
4006 LD->getSrcValueOffset(), LD->getMemoryVT(),
4007 LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment());
4010 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
4011 SDValue Ptr, const Value *SV, int SVOffset,
4012 bool isVolatile, bool isNonTemporal,
4013 unsigned Alignment) {
4014 if (Alignment == 0) // Ensure that codegen never sees alignment 0
4015 Alignment = getEVTAlignment(Val.getValueType());
4017 // Check if the memory reference references a frame index
4019 if (const FrameIndexSDNode *FI =
4020 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
4021 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
4023 MachineFunction &MF = getMachineFunction();
4024 unsigned Flags = MachineMemOperand::MOStore;
4026 Flags |= MachineMemOperand::MOVolatile;
4028 Flags |= MachineMemOperand::MONonTemporal;
4029 MachineMemOperand *MMO =
4030 MF.getMachineMemOperand(SV, Flags, SVOffset,
4031 Val.getValueType().getStoreSize(), Alignment);
4033 return getStore(Chain, dl, Val, Ptr, MMO);
4036 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
4037 SDValue Ptr, MachineMemOperand *MMO) {
4038 EVT VT = Val.getValueType();
4039 SDVTList VTs = getVTList(MVT::Other);
4040 SDValue Undef = getUNDEF(Ptr.getValueType());
4041 SDValue Ops[] = { Chain, Val, Ptr, Undef };
4042 FoldingSetNodeID ID;
4043 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4044 ID.AddInteger(VT.getRawBits());
4045 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile(),
4046 MMO->isNonTemporal()));
4048 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
4049 cast<StoreSDNode>(E)->refineAlignment(MMO);
4050 return SDValue(E, 0);
4052 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED,
4054 CSEMap.InsertNode(N, IP);
4055 AllNodes.push_back(N);
4056 return SDValue(N, 0);
4059 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
4060 SDValue Ptr, const Value *SV,
4061 int SVOffset, EVT SVT,
4062 bool isVolatile, bool isNonTemporal,
4063 unsigned Alignment) {
4064 if (Alignment == 0) // Ensure that codegen never sees alignment 0
4065 Alignment = getEVTAlignment(SVT);
4067 // Check if the memory reference references a frame index
4069 if (const FrameIndexSDNode *FI =
4070 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
4071 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
4073 MachineFunction &MF = getMachineFunction();
4074 unsigned Flags = MachineMemOperand::MOStore;
4076 Flags |= MachineMemOperand::MOVolatile;
4078 Flags |= MachineMemOperand::MONonTemporal;
4079 MachineMemOperand *MMO =
4080 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment);
4082 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
4085 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
4086 SDValue Ptr, EVT SVT,
4087 MachineMemOperand *MMO) {
4088 EVT VT = Val.getValueType();
4091 return getStore(Chain, dl, Val, Ptr, MMO);
4093 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
4094 "Should only be a truncating store, not extending!");
4095 assert(VT.isInteger() == SVT.isInteger() &&
4096 "Can't do FP-INT conversion!");
4097 assert(VT.isVector() == SVT.isVector() &&
4098 "Cannot use trunc store to convert to or from a vector!");
4099 assert((!VT.isVector() ||
4100 VT.getVectorNumElements() == SVT.getVectorNumElements()) &&
4101 "Cannot use trunc store to change the number of vector elements!");
4103 SDVTList VTs = getVTList(MVT::Other);
4104 SDValue Undef = getUNDEF(Ptr.getValueType());
4105 SDValue Ops[] = { Chain, Val, Ptr, Undef };
4106 FoldingSetNodeID ID;
4107 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4108 ID.AddInteger(SVT.getRawBits());
4109 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile(),
4110 MMO->isNonTemporal()));
4112 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
4113 cast<StoreSDNode>(E)->refineAlignment(MMO);
4114 return SDValue(E, 0);
4116 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED,
4118 CSEMap.InsertNode(N, IP);
4119 AllNodes.push_back(N);
4120 return SDValue(N, 0);
4124 SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base,
4125 SDValue Offset, ISD::MemIndexedMode AM) {
4126 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
4127 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
4128 "Store is already a indexed store!");
4129 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
4130 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
4131 FoldingSetNodeID ID;
4132 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4133 ID.AddInteger(ST->getMemoryVT().getRawBits());
4134 ID.AddInteger(ST->getRawSubclassData());
4136 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4137 return SDValue(E, 0);
4139 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, AM,
4140 ST->isTruncatingStore(),
4142 ST->getMemOperand());
4143 CSEMap.InsertNode(N, IP);
4144 AllNodes.push_back(N);
4145 return SDValue(N, 0);
4148 SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl,
4149 SDValue Chain, SDValue Ptr,
4151 SDValue Ops[] = { Chain, Ptr, SV };
4152 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3);
4155 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4156 const SDUse *Ops, unsigned NumOps) {
4158 case 0: return getNode(Opcode, DL, VT);
4159 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4160 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4161 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4165 // Copy from an SDUse array into an SDValue array for use with
4166 // the regular getNode logic.
4167 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
4168 return getNode(Opcode, DL, VT, &NewOps[0], NumOps);
4171 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4172 const SDValue *Ops, unsigned NumOps) {
4174 case 0: return getNode(Opcode, DL, VT);
4175 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4176 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4177 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4183 case ISD::SELECT_CC: {
4184 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
4185 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
4186 "LHS and RHS of condition must have same type!");
4187 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4188 "True and False arms of SelectCC must have same type!");
4189 assert(Ops[2].getValueType() == VT &&
4190 "select_cc node must be of same type as true and false value!");
4194 assert(NumOps == 5 && "BR_CC takes 5 operands!");
4195 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4196 "LHS/RHS of comparison should match types!");
4203 SDVTList VTs = getVTList(VT);
4205 if (VT != MVT::Flag) {
4206 FoldingSetNodeID ID;
4207 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
4210 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4211 return SDValue(E, 0);
4213 N = new (NodeAllocator) SDNode(Opcode, DL, VTs, Ops, NumOps);
4214 CSEMap.InsertNode(N, IP);
4216 N = new (NodeAllocator) SDNode(Opcode, DL, VTs, Ops, NumOps);
4219 AllNodes.push_back(N);
4223 return SDValue(N, 0);
4226 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4227 const std::vector<EVT> &ResultTys,
4228 const SDValue *Ops, unsigned NumOps) {
4229 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()),
4233 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4234 const EVT *VTs, unsigned NumVTs,
4235 const SDValue *Ops, unsigned NumOps) {
4237 return getNode(Opcode, DL, VTs[0], Ops, NumOps);
4238 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps);
4241 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4242 const SDValue *Ops, unsigned NumOps) {
4243 if (VTList.NumVTs == 1)
4244 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps);
4248 // FIXME: figure out how to safely handle things like
4249 // int foo(int x) { return 1 << (x & 255); }
4250 // int bar() { return foo(256); }
4251 case ISD::SRA_PARTS:
4252 case ISD::SRL_PARTS:
4253 case ISD::SHL_PARTS:
4254 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4255 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
4256 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4257 else if (N3.getOpcode() == ISD::AND)
4258 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
4259 // If the and is only masking out bits that cannot effect the shift,
4260 // eliminate the and.
4261 unsigned NumBits = VT.getScalarType().getSizeInBits()*2;
4262 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
4263 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4269 // Memoize the node unless it returns a flag.
4271 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4272 FoldingSetNodeID ID;
4273 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4275 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4276 return SDValue(E, 0);
4279 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4280 } else if (NumOps == 2) {
4281 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4282 } else if (NumOps == 3) {
4283 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1],
4286 N = new (NodeAllocator) SDNode(Opcode, DL, VTList, Ops, NumOps);
4288 CSEMap.InsertNode(N, IP);
4291 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4292 } else if (NumOps == 2) {
4293 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4294 } else if (NumOps == 3) {
4295 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1],
4298 N = new (NodeAllocator) SDNode(Opcode, DL, VTList, Ops, NumOps);
4301 AllNodes.push_back(N);
4305 return SDValue(N, 0);
4308 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) {
4309 return getNode(Opcode, DL, VTList, 0, 0);
4312 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4314 SDValue Ops[] = { N1 };
4315 return getNode(Opcode, DL, VTList, Ops, 1);
4318 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4319 SDValue N1, SDValue N2) {
4320 SDValue Ops[] = { N1, N2 };
4321 return getNode(Opcode, DL, VTList, Ops, 2);
4324 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4325 SDValue N1, SDValue N2, SDValue N3) {
4326 SDValue Ops[] = { N1, N2, N3 };
4327 return getNode(Opcode, DL, VTList, Ops, 3);
4330 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4331 SDValue N1, SDValue N2, SDValue N3,
4333 SDValue Ops[] = { N1, N2, N3, N4 };
4334 return getNode(Opcode, DL, VTList, Ops, 4);
4337 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4338 SDValue N1, SDValue N2, SDValue N3,
4339 SDValue N4, SDValue N5) {
4340 SDValue Ops[] = { N1, N2, N3, N4, N5 };
4341 return getNode(Opcode, DL, VTList, Ops, 5);
4344 SDVTList SelectionDAG::getVTList(EVT VT) {
4345 return makeVTList(SDNode::getValueTypeList(VT), 1);
4348 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
4349 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4350 E = VTList.rend(); I != E; ++I)
4351 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
4354 EVT *Array = Allocator.Allocate<EVT>(2);
4357 SDVTList Result = makeVTList(Array, 2);
4358 VTList.push_back(Result);
4362 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
4363 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4364 E = VTList.rend(); I != E; ++I)
4365 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4369 EVT *Array = Allocator.Allocate<EVT>(3);
4373 SDVTList Result = makeVTList(Array, 3);
4374 VTList.push_back(Result);
4378 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
4379 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4380 E = VTList.rend(); I != E; ++I)
4381 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4382 I->VTs[2] == VT3 && I->VTs[3] == VT4)
4385 EVT *Array = Allocator.Allocate<EVT>(4);
4390 SDVTList Result = makeVTList(Array, 4);
4391 VTList.push_back(Result);
4395 SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) {
4397 case 0: llvm_unreachable("Cannot have nodes without results!");
4398 case 1: return getVTList(VTs[0]);
4399 case 2: return getVTList(VTs[0], VTs[1]);
4400 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
4401 case 4: return getVTList(VTs[0], VTs[1], VTs[2], VTs[3]);
4405 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4406 E = VTList.rend(); I != E; ++I) {
4407 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
4410 bool NoMatch = false;
4411 for (unsigned i = 2; i != NumVTs; ++i)
4412 if (VTs[i] != I->VTs[i]) {
4420 EVT *Array = Allocator.Allocate<EVT>(NumVTs);
4421 std::copy(VTs, VTs+NumVTs, Array);
4422 SDVTList Result = makeVTList(Array, NumVTs);
4423 VTList.push_back(Result);
4428 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
4429 /// specified operands. If the resultant node already exists in the DAG,
4430 /// this does not modify the specified node, instead it returns the node that
4431 /// already exists. If the resultant node does not exist in the DAG, the
4432 /// input node is returned. As a degenerate case, if you specify the same
4433 /// input operands as the node already has, the input node is returned.
4434 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
4435 SDNode *N = InN.getNode();
4436 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
4438 // Check to see if there is no change.
4439 if (Op == N->getOperand(0)) return InN;
4441 // See if the modified node already exists.
4442 void *InsertPos = 0;
4443 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
4444 return SDValue(Existing, InN.getResNo());
4446 // Nope it doesn't. Remove the node from its current place in the maps.
4448 if (!RemoveNodeFromCSEMaps(N))
4451 // Now we update the operands.
4452 N->OperandList[0].set(Op);
4454 // If this gets put into a CSE map, add it.
4455 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4459 SDValue SelectionDAG::
4460 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
4461 SDNode *N = InN.getNode();
4462 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
4464 // Check to see if there is no change.
4465 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
4466 return InN; // No operands changed, just return the input node.
4468 // See if the modified node already exists.
4469 void *InsertPos = 0;
4470 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
4471 return SDValue(Existing, InN.getResNo());
4473 // Nope it doesn't. Remove the node from its current place in the maps.
4475 if (!RemoveNodeFromCSEMaps(N))
4478 // Now we update the operands.
4479 if (N->OperandList[0] != Op1)
4480 N->OperandList[0].set(Op1);
4481 if (N->OperandList[1] != Op2)
4482 N->OperandList[1].set(Op2);
4484 // If this gets put into a CSE map, add it.
4485 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4489 SDValue SelectionDAG::
4490 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
4491 SDValue Ops[] = { Op1, Op2, Op3 };
4492 return UpdateNodeOperands(N, Ops, 3);
4495 SDValue SelectionDAG::
4496 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4497 SDValue Op3, SDValue Op4) {
4498 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
4499 return UpdateNodeOperands(N, Ops, 4);
4502 SDValue SelectionDAG::
4503 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4504 SDValue Op3, SDValue Op4, SDValue Op5) {
4505 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
4506 return UpdateNodeOperands(N, Ops, 5);
4509 SDValue SelectionDAG::
4510 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
4511 SDNode *N = InN.getNode();
4512 assert(N->getNumOperands() == NumOps &&
4513 "Update with wrong number of operands");
4515 // Check to see if there is no change.
4516 bool AnyChange = false;
4517 for (unsigned i = 0; i != NumOps; ++i) {
4518 if (Ops[i] != N->getOperand(i)) {
4524 // No operands changed, just return the input node.
4525 if (!AnyChange) return InN;
4527 // See if the modified node already exists.
4528 void *InsertPos = 0;
4529 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
4530 return SDValue(Existing, InN.getResNo());
4532 // Nope it doesn't. Remove the node from its current place in the maps.
4534 if (!RemoveNodeFromCSEMaps(N))
4537 // Now we update the operands.
4538 for (unsigned i = 0; i != NumOps; ++i)
4539 if (N->OperandList[i] != Ops[i])
4540 N->OperandList[i].set(Ops[i]);
4542 // If this gets put into a CSE map, add it.
4543 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4547 /// DropOperands - Release the operands and set this node to have
4549 void SDNode::DropOperands() {
4550 // Unlike the code in MorphNodeTo that does this, we don't need to
4551 // watch for dead nodes here.
4552 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
4558 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
4561 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4563 SDVTList VTs = getVTList(VT);
4564 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4567 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4568 EVT VT, SDValue Op1) {
4569 SDVTList VTs = getVTList(VT);
4570 SDValue Ops[] = { Op1 };
4571 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4574 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4575 EVT VT, SDValue Op1,
4577 SDVTList VTs = getVTList(VT);
4578 SDValue Ops[] = { Op1, Op2 };
4579 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4582 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4583 EVT VT, SDValue Op1,
4584 SDValue Op2, SDValue Op3) {
4585 SDVTList VTs = getVTList(VT);
4586 SDValue Ops[] = { Op1, Op2, Op3 };
4587 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4590 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4591 EVT VT, const SDValue *Ops,
4593 SDVTList VTs = getVTList(VT);
4594 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4597 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4598 EVT VT1, EVT VT2, const SDValue *Ops,
4600 SDVTList VTs = getVTList(VT1, VT2);
4601 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4604 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4606 SDVTList VTs = getVTList(VT1, VT2);
4607 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4610 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4611 EVT VT1, EVT VT2, EVT VT3,
4612 const SDValue *Ops, unsigned NumOps) {
4613 SDVTList VTs = getVTList(VT1, VT2, VT3);
4614 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4617 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4618 EVT VT1, EVT VT2, EVT VT3, EVT VT4,
4619 const SDValue *Ops, unsigned NumOps) {
4620 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4621 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4624 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4627 SDVTList VTs = getVTList(VT1, VT2);
4628 SDValue Ops[] = { Op1 };
4629 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4632 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4634 SDValue Op1, SDValue Op2) {
4635 SDVTList VTs = getVTList(VT1, VT2);
4636 SDValue Ops[] = { Op1, Op2 };
4637 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4640 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4642 SDValue Op1, SDValue Op2,
4644 SDVTList VTs = getVTList(VT1, VT2);
4645 SDValue Ops[] = { Op1, Op2, Op3 };
4646 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4649 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4650 EVT VT1, EVT VT2, EVT VT3,
4651 SDValue Op1, SDValue Op2,
4653 SDVTList VTs = getVTList(VT1, VT2, VT3);
4654 SDValue Ops[] = { Op1, Op2, Op3 };
4655 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4658 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4659 SDVTList VTs, const SDValue *Ops,
4661 N = MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4662 // Reset the NodeID to -1.
4667 /// MorphNodeTo - This *mutates* the specified node to have the specified
4668 /// return type, opcode, and operands.
4670 /// Note that MorphNodeTo returns the resultant node. If there is already a
4671 /// node of the specified opcode and operands, it returns that node instead of
4672 /// the current one. Note that the DebugLoc need not be the same.
4674 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4675 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4676 /// node, and because it doesn't require CSE recalculation for any of
4677 /// the node's users.
4679 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4680 SDVTList VTs, const SDValue *Ops,
4682 // If an identical node already exists, use it.
4684 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4685 FoldingSetNodeID ID;
4686 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4687 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4691 if (!RemoveNodeFromCSEMaps(N))
4694 // Start the morphing.
4696 N->ValueList = VTs.VTs;
4697 N->NumValues = VTs.NumVTs;
4699 // Clear the operands list, updating used nodes to remove this from their
4700 // use list. Keep track of any operands that become dead as a result.
4701 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4702 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
4704 SDNode *Used = Use.getNode();
4706 if (Used->use_empty())
4707 DeadNodeSet.insert(Used);
4710 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) {
4711 // Initialize the memory references information.
4712 MN->setMemRefs(0, 0);
4713 // If NumOps is larger than the # of operands we can have in a
4714 // MachineSDNode, reallocate the operand list.
4715 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) {
4716 if (MN->OperandsNeedDelete)
4717 delete[] MN->OperandList;
4718 if (NumOps > array_lengthof(MN->LocalOperands))
4719 // We're creating a final node that will live unmorphed for the
4720 // remainder of the current SelectionDAG iteration, so we can allocate
4721 // the operands directly out of a pool with no recycling metadata.
4722 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4725 MN->InitOperands(MN->LocalOperands, Ops, NumOps);
4726 MN->OperandsNeedDelete = false;
4728 MN->InitOperands(MN->OperandList, Ops, NumOps);
4730 // If NumOps is larger than the # of operands we currently have, reallocate
4731 // the operand list.
4732 if (NumOps > N->NumOperands) {
4733 if (N->OperandsNeedDelete)
4734 delete[] N->OperandList;
4735 N->InitOperands(new SDUse[NumOps], Ops, NumOps);
4736 N->OperandsNeedDelete = true;
4738 N->InitOperands(N->OperandList, Ops, NumOps);
4741 // Delete any nodes that are still dead after adding the uses for the
4743 if (!DeadNodeSet.empty()) {
4744 SmallVector<SDNode *, 16> DeadNodes;
4745 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4746 E = DeadNodeSet.end(); I != E; ++I)
4747 if ((*I)->use_empty())
4748 DeadNodes.push_back(*I);
4749 RemoveDeadNodes(DeadNodes);
4753 CSEMap.InsertNode(N, IP); // Memoize the new node.
4758 /// getMachineNode - These are used for target selectors to create a new node
4759 /// with specified return type(s), MachineInstr opcode, and operands.
4761 /// Note that getMachineNode returns the resultant node. If there is already a
4762 /// node of the specified opcode and operands, it returns that node instead of
4763 /// the current one.
4765 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) {
4766 SDVTList VTs = getVTList(VT);
4767 return getMachineNode(Opcode, dl, VTs, 0, 0);
4771 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) {
4772 SDVTList VTs = getVTList(VT);
4773 SDValue Ops[] = { Op1 };
4774 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4778 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4779 SDValue Op1, SDValue Op2) {
4780 SDVTList VTs = getVTList(VT);
4781 SDValue Ops[] = { Op1, Op2 };
4782 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4786 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4787 SDValue Op1, SDValue Op2, SDValue Op3) {
4788 SDVTList VTs = getVTList(VT);
4789 SDValue Ops[] = { Op1, Op2, Op3 };
4790 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4794 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4795 const SDValue *Ops, unsigned NumOps) {
4796 SDVTList VTs = getVTList(VT);
4797 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4801 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) {
4802 SDVTList VTs = getVTList(VT1, VT2);
4803 return getMachineNode(Opcode, dl, VTs, 0, 0);
4807 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4808 EVT VT1, EVT VT2, SDValue Op1) {
4809 SDVTList VTs = getVTList(VT1, VT2);
4810 SDValue Ops[] = { Op1 };
4811 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4815 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4816 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) {
4817 SDVTList VTs = getVTList(VT1, VT2);
4818 SDValue Ops[] = { Op1, Op2 };
4819 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4823 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4824 EVT VT1, EVT VT2, SDValue Op1,
4825 SDValue Op2, SDValue Op3) {
4826 SDVTList VTs = getVTList(VT1, VT2);
4827 SDValue Ops[] = { Op1, Op2, Op3 };
4828 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4832 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4834 const SDValue *Ops, unsigned NumOps) {
4835 SDVTList VTs = getVTList(VT1, VT2);
4836 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4840 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4841 EVT VT1, EVT VT2, EVT VT3,
4842 SDValue Op1, SDValue Op2) {
4843 SDVTList VTs = getVTList(VT1, VT2, VT3);
4844 SDValue Ops[] = { Op1, Op2 };
4845 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4849 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4850 EVT VT1, EVT VT2, EVT VT3,
4851 SDValue Op1, SDValue Op2, SDValue Op3) {
4852 SDVTList VTs = getVTList(VT1, VT2, VT3);
4853 SDValue Ops[] = { Op1, Op2, Op3 };
4854 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4858 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4859 EVT VT1, EVT VT2, EVT VT3,
4860 const SDValue *Ops, unsigned NumOps) {
4861 SDVTList VTs = getVTList(VT1, VT2, VT3);
4862 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4866 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1,
4867 EVT VT2, EVT VT3, EVT VT4,
4868 const SDValue *Ops, unsigned NumOps) {
4869 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4870 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4874 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4875 const std::vector<EVT> &ResultTys,
4876 const SDValue *Ops, unsigned NumOps) {
4877 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size());
4878 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4882 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs,
4883 const SDValue *Ops, unsigned NumOps) {
4884 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag;
4889 FoldingSetNodeID ID;
4890 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps);
4892 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4893 return cast<MachineSDNode>(E);
4896 // Allocate a new MachineSDNode.
4897 N = new (NodeAllocator) MachineSDNode(~Opcode, DL, VTs);
4899 // Initialize the operands list.
4900 if (NumOps > array_lengthof(N->LocalOperands))
4901 // We're creating a final node that will live unmorphed for the
4902 // remainder of the current SelectionDAG iteration, so we can allocate
4903 // the operands directly out of a pool with no recycling metadata.
4904 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4907 N->InitOperands(N->LocalOperands, Ops, NumOps);
4908 N->OperandsNeedDelete = false;
4911 CSEMap.InsertNode(N, IP);
4913 AllNodes.push_back(N);
4920 /// getTargetExtractSubreg - A convenience function for creating
4921 /// TargetOpcode::EXTRACT_SUBREG nodes.
4923 SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT,
4925 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4926 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
4927 VT, Operand, SRIdxVal);
4928 return SDValue(Subreg, 0);
4931 /// getTargetInsertSubreg - A convenience function for creating
4932 /// TargetOpcode::INSERT_SUBREG nodes.
4934 SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT,
4935 SDValue Operand, SDValue Subreg) {
4936 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4937 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
4938 VT, Operand, Subreg, SRIdxVal);
4939 return SDValue(Result, 0);
4942 /// getNodeIfExists - Get the specified node if it's already available, or
4943 /// else return NULL.
4944 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4945 const SDValue *Ops, unsigned NumOps) {
4946 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4947 FoldingSetNodeID ID;
4948 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4950 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4956 /// getDbgValue - Creates a SDDbgValue node.
4959 SelectionDAG::getDbgValue(MDNode *MDPtr, SDNode *N, unsigned R, uint64_t Off,
4960 DebugLoc DL, unsigned O) {
4961 return new (Allocator) SDDbgValue(MDPtr, N, R, Off, DL, O);
4965 SelectionDAG::getDbgValue(MDNode *MDPtr, const Value *C, uint64_t Off,
4966 DebugLoc DL, unsigned O) {
4967 return new (Allocator) SDDbgValue(MDPtr, C, Off, DL, O);
4971 SelectionDAG::getDbgValue(MDNode *MDPtr, unsigned FI, uint64_t Off,
4972 DebugLoc DL, unsigned O) {
4973 return new (Allocator) SDDbgValue(MDPtr, FI, Off, DL, O);
4978 /// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node
4979 /// pointed to by a use iterator is deleted, increment the use iterator
4980 /// so that it doesn't dangle.
4982 /// This class also manages a "downlink" DAGUpdateListener, to forward
4983 /// messages to ReplaceAllUsesWith's callers.
4985 class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener {
4986 SelectionDAG::DAGUpdateListener *DownLink;
4987 SDNode::use_iterator &UI;
4988 SDNode::use_iterator &UE;
4990 virtual void NodeDeleted(SDNode *N, SDNode *E) {
4991 // Increment the iterator as needed.
4992 while (UI != UE && N == *UI)
4995 // Then forward the message.
4996 if (DownLink) DownLink->NodeDeleted(N, E);
4999 virtual void NodeUpdated(SDNode *N) {
5000 // Just forward the message.
5001 if (DownLink) DownLink->NodeUpdated(N);
5005 RAUWUpdateListener(SelectionDAG::DAGUpdateListener *dl,
5006 SDNode::use_iterator &ui,
5007 SDNode::use_iterator &ue)
5008 : DownLink(dl), UI(ui), UE(ue) {}
5013 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5014 /// This can cause recursive merging of nodes in the DAG.
5016 /// This version assumes From has a single result value.
5018 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
5019 DAGUpdateListener *UpdateListener) {
5020 SDNode *From = FromN.getNode();
5021 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
5022 "Cannot replace with this method!");
5023 assert(From != To.getNode() && "Cannot replace uses of with self");
5025 // Iterate over all the existing uses of From. New uses will be added
5026 // to the beginning of the use list, which we avoid visiting.
5027 // This specifically avoids visiting uses of From that arise while the
5028 // replacement is happening, because any such uses would be the result
5029 // of CSE: If an existing node looks like From after one of its operands
5030 // is replaced by To, we don't want to replace of all its users with To
5031 // too. See PR3018 for more info.
5032 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5033 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5037 // This node is about to morph, remove its old self from the CSE maps.
5038 RemoveNodeFromCSEMaps(User);
5040 // A user can appear in a use list multiple times, and when this
5041 // happens the uses are usually next to each other in the list.
5042 // To help reduce the number of CSE recomputations, process all
5043 // the uses of this user that we can find this way.
5045 SDUse &Use = UI.getUse();
5048 } while (UI != UE && *UI == User);
5050 // Now that we have modified User, add it back to the CSE maps. If it
5051 // already exists there, recursively merge the results together.
5052 AddModifiedNodeToCSEMaps(User, &Listener);
5056 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5057 /// This can cause recursive merging of nodes in the DAG.
5059 /// This version assumes that for each value of From, there is a
5060 /// corresponding value in To in the same position with the same type.
5062 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
5063 DAGUpdateListener *UpdateListener) {
5065 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
5066 assert((!From->hasAnyUseOfValue(i) ||
5067 From->getValueType(i) == To->getValueType(i)) &&
5068 "Cannot use this version of ReplaceAllUsesWith!");
5071 // Handle the trivial case.
5075 // Iterate over just the existing users of From. See the comments in
5076 // the ReplaceAllUsesWith above.
5077 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5078 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5082 // This node is about to morph, remove its old self from the CSE maps.
5083 RemoveNodeFromCSEMaps(User);
5085 // A user can appear in a use list multiple times, and when this
5086 // happens the uses are usually next to each other in the list.
5087 // To help reduce the number of CSE recomputations, process all
5088 // the uses of this user that we can find this way.
5090 SDUse &Use = UI.getUse();
5093 } while (UI != UE && *UI == User);
5095 // Now that we have modified User, add it back to the CSE maps. If it
5096 // already exists there, recursively merge the results together.
5097 AddModifiedNodeToCSEMaps(User, &Listener);
5101 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5102 /// This can cause recursive merging of nodes in the DAG.
5104 /// This version can replace From with any result values. To must match the
5105 /// number and types of values returned by From.
5106 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
5108 DAGUpdateListener *UpdateListener) {
5109 if (From->getNumValues() == 1) // Handle the simple case efficiently.
5110 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
5112 // Iterate over just the existing users of From. See the comments in
5113 // the ReplaceAllUsesWith above.
5114 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5115 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5119 // This node is about to morph, remove its old self from the CSE maps.
5120 RemoveNodeFromCSEMaps(User);
5122 // A user can appear in a use list multiple times, and when this
5123 // happens the uses are usually next to each other in the list.
5124 // To help reduce the number of CSE recomputations, process all
5125 // the uses of this user that we can find this way.
5127 SDUse &Use = UI.getUse();
5128 const SDValue &ToOp = To[Use.getResNo()];
5131 } while (UI != UE && *UI == User);
5133 // Now that we have modified User, add it back to the CSE maps. If it
5134 // already exists there, recursively merge the results together.
5135 AddModifiedNodeToCSEMaps(User, &Listener);
5139 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
5140 /// uses of other values produced by From.getNode() alone. The Deleted
5141 /// vector is handled the same way as for ReplaceAllUsesWith.
5142 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
5143 DAGUpdateListener *UpdateListener){
5144 // Handle the really simple, really trivial case efficiently.
5145 if (From == To) return;
5147 // Handle the simple, trivial, case efficiently.
5148 if (From.getNode()->getNumValues() == 1) {
5149 ReplaceAllUsesWith(From, To, UpdateListener);
5153 // Iterate over just the existing users of From. See the comments in
5154 // the ReplaceAllUsesWith above.
5155 SDNode::use_iterator UI = From.getNode()->use_begin(),
5156 UE = From.getNode()->use_end();
5157 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5160 bool UserRemovedFromCSEMaps = false;
5162 // A user can appear in a use list multiple times, and when this
5163 // happens the uses are usually next to each other in the list.
5164 // To help reduce the number of CSE recomputations, process all
5165 // the uses of this user that we can find this way.
5167 SDUse &Use = UI.getUse();
5169 // Skip uses of different values from the same node.
5170 if (Use.getResNo() != From.getResNo()) {
5175 // If this node hasn't been modified yet, it's still in the CSE maps,
5176 // so remove its old self from the CSE maps.
5177 if (!UserRemovedFromCSEMaps) {
5178 RemoveNodeFromCSEMaps(User);
5179 UserRemovedFromCSEMaps = true;
5184 } while (UI != UE && *UI == User);
5186 // We are iterating over all uses of the From node, so if a use
5187 // doesn't use the specific value, no changes are made.
5188 if (!UserRemovedFromCSEMaps)
5191 // Now that we have modified User, add it back to the CSE maps. If it
5192 // already exists there, recursively merge the results together.
5193 AddModifiedNodeToCSEMaps(User, &Listener);
5198 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
5199 /// to record information about a use.
5206 /// operator< - Sort Memos by User.
5207 bool operator<(const UseMemo &L, const UseMemo &R) {
5208 return (intptr_t)L.User < (intptr_t)R.User;
5212 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
5213 /// uses of other values produced by From.getNode() alone. The same value
5214 /// may appear in both the From and To list. The Deleted vector is
5215 /// handled the same way as for ReplaceAllUsesWith.
5216 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
5219 DAGUpdateListener *UpdateListener){
5220 // Handle the simple, trivial case efficiently.
5222 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
5224 // Read up all the uses and make records of them. This helps
5225 // processing new uses that are introduced during the
5226 // replacement process.
5227 SmallVector<UseMemo, 4> Uses;
5228 for (unsigned i = 0; i != Num; ++i) {
5229 unsigned FromResNo = From[i].getResNo();
5230 SDNode *FromNode = From[i].getNode();
5231 for (SDNode::use_iterator UI = FromNode->use_begin(),
5232 E = FromNode->use_end(); UI != E; ++UI) {
5233 SDUse &Use = UI.getUse();
5234 if (Use.getResNo() == FromResNo) {
5235 UseMemo Memo = { *UI, i, &Use };
5236 Uses.push_back(Memo);
5241 // Sort the uses, so that all the uses from a given User are together.
5242 std::sort(Uses.begin(), Uses.end());
5244 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
5245 UseIndex != UseIndexEnd; ) {
5246 // We know that this user uses some value of From. If it is the right
5247 // value, update it.
5248 SDNode *User = Uses[UseIndex].User;
5250 // This node is about to morph, remove its old self from the CSE maps.
5251 RemoveNodeFromCSEMaps(User);
5253 // The Uses array is sorted, so all the uses for a given User
5254 // are next to each other in the list.
5255 // To help reduce the number of CSE recomputations, process all
5256 // the uses of this user that we can find this way.
5258 unsigned i = Uses[UseIndex].Index;
5259 SDUse &Use = *Uses[UseIndex].Use;
5263 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
5265 // Now that we have modified User, add it back to the CSE maps. If it
5266 // already exists there, recursively merge the results together.
5267 AddModifiedNodeToCSEMaps(User, UpdateListener);
5271 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
5272 /// based on their topological order. It returns the maximum id and a vector
5273 /// of the SDNodes* in assigned order by reference.
5274 unsigned SelectionDAG::AssignTopologicalOrder() {
5276 unsigned DAGSize = 0;
5278 // SortedPos tracks the progress of the algorithm. Nodes before it are
5279 // sorted, nodes after it are unsorted. When the algorithm completes
5280 // it is at the end of the list.
5281 allnodes_iterator SortedPos = allnodes_begin();
5283 // Visit all the nodes. Move nodes with no operands to the front of
5284 // the list immediately. Annotate nodes that do have operands with their
5285 // operand count. Before we do this, the Node Id fields of the nodes
5286 // may contain arbitrary values. After, the Node Id fields for nodes
5287 // before SortedPos will contain the topological sort index, and the
5288 // Node Id fields for nodes At SortedPos and after will contain the
5289 // count of outstanding operands.
5290 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
5293 unsigned Degree = N->getNumOperands();
5295 // A node with no uses, add it to the result array immediately.
5296 N->setNodeId(DAGSize++);
5297 allnodes_iterator Q = N;
5299 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
5300 assert(SortedPos != AllNodes.end() && "Overran node list");
5303 // Temporarily use the Node Id as scratch space for the degree count.
5304 N->setNodeId(Degree);
5308 // Visit all the nodes. As we iterate, moves nodes into sorted order,
5309 // such that by the time the end is reached all nodes will be sorted.
5310 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
5313 // N is in sorted position, so all its uses have one less operand
5314 // that needs to be sorted.
5315 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5318 unsigned Degree = P->getNodeId();
5319 assert(Degree != 0 && "Invalid node degree");
5322 // All of P's operands are sorted, so P may sorted now.
5323 P->setNodeId(DAGSize++);
5325 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
5326 assert(SortedPos != AllNodes.end() && "Overran node list");
5329 // Update P's outstanding operand count.
5330 P->setNodeId(Degree);
5333 if (I == SortedPos) {
5336 dbgs() << "Overran sorted position:\n";
5339 llvm_unreachable(0);
5343 assert(SortedPos == AllNodes.end() &&
5344 "Topological sort incomplete!");
5345 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
5346 "First node in topological sort is not the entry token!");
5347 assert(AllNodes.front().getNodeId() == 0 &&
5348 "First node in topological sort has non-zero id!");
5349 assert(AllNodes.front().getNumOperands() == 0 &&
5350 "First node in topological sort has operands!");
5351 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
5352 "Last node in topologic sort has unexpected id!");
5353 assert(AllNodes.back().use_empty() &&
5354 "Last node in topologic sort has users!");
5355 assert(DAGSize == allnodes_size() && "Node count mismatch!");
5359 /// AssignOrdering - Assign an order to the SDNode.
5360 void SelectionDAG::AssignOrdering(const SDNode *SD, unsigned Order) {
5361 assert(SD && "Trying to assign an order to a null node!");
5362 Ordering->add(SD, Order);
5365 /// GetOrdering - Get the order for the SDNode.
5366 unsigned SelectionDAG::GetOrdering(const SDNode *SD) const {
5367 assert(SD && "Trying to get the order of a null node!");
5368 return Ordering->getOrder(SD);
5371 /// AddDbgValue - Add a dbg_value SDNode. If SD is non-null that means the
5372 /// value is produced by SD.
5373 void SelectionDAG::AddDbgValue(SDDbgValue *DB, SDNode *SD, bool isParameter) {
5374 DbgInfo->add(DB, SD, isParameter);
5376 SD->setHasDebugValue(true);
5379 //===----------------------------------------------------------------------===//
5381 //===----------------------------------------------------------------------===//
5383 HandleSDNode::~HandleSDNode() {
5387 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA,
5388 EVT VT, int64_t o, unsigned char TF)
5389 : SDNode(Opc, DebugLoc(), getSDVTList(VT)), Offset(o), TargetFlags(TF) {
5393 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt,
5394 MachineMemOperand *mmo)
5395 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) {
5396 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(),
5397 MMO->isNonTemporal());
5398 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5399 assert(isNonTemporal() == MMO->isNonTemporal() &&
5400 "Non-temporal encoding error!");
5401 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5404 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs,
5405 const SDValue *Ops, unsigned NumOps, EVT memvt,
5406 MachineMemOperand *mmo)
5407 : SDNode(Opc, dl, VTs, Ops, NumOps),
5408 MemoryVT(memvt), MMO(mmo) {
5409 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(),
5410 MMO->isNonTemporal());
5411 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5412 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5415 /// Profile - Gather unique data for the node.
5417 void SDNode::Profile(FoldingSetNodeID &ID) const {
5418 AddNodeIDNode(ID, this);
5423 std::vector<EVT> VTs;
5426 VTs.reserve(MVT::LAST_VALUETYPE);
5427 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i)
5428 VTs.push_back(MVT((MVT::SimpleValueType)i));
5433 static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs;
5434 static ManagedStatic<EVTArray> SimpleVTArray;
5435 static ManagedStatic<sys::SmartMutex<true> > VTMutex;
5437 /// getValueTypeList - Return a pointer to the specified value type.
5439 const EVT *SDNode::getValueTypeList(EVT VT) {
5440 if (VT.isExtended()) {
5441 sys::SmartScopedLock<true> Lock(*VTMutex);
5442 return &(*EVTs->insert(VT).first);
5444 assert(VT.getSimpleVT().SimpleTy < MVT::LAST_VALUETYPE &&
5445 "Value type out of range!");
5446 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy];
5450 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
5451 /// indicated value. This method ignores uses of other values defined by this
5453 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
5454 assert(Value < getNumValues() && "Bad value!");
5456 // TODO: Only iterate over uses of a given value of the node
5457 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
5458 if (UI.getUse().getResNo() == Value) {
5465 // Found exactly the right number of uses?
5470 /// hasAnyUseOfValue - Return true if there are any use of the indicated
5471 /// value. This method ignores uses of other values defined by this operation.
5472 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
5473 assert(Value < getNumValues() && "Bad value!");
5475 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
5476 if (UI.getUse().getResNo() == Value)
5483 /// isOnlyUserOf - Return true if this node is the only use of N.
5485 bool SDNode::isOnlyUserOf(SDNode *N) const {
5487 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
5498 /// isOperand - Return true if this node is an operand of N.
5500 bool SDValue::isOperandOf(SDNode *N) const {
5501 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5502 if (*this == N->getOperand(i))
5507 bool SDNode::isOperandOf(SDNode *N) const {
5508 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
5509 if (this == N->OperandList[i].getNode())
5514 /// reachesChainWithoutSideEffects - Return true if this operand (which must
5515 /// be a chain) reaches the specified operand without crossing any
5516 /// side-effecting instructions. In practice, this looks through token
5517 /// factors and non-volatile loads. In order to remain efficient, this only
5518 /// looks a couple of nodes in, it does not do an exhaustive search.
5519 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
5520 unsigned Depth) const {
5521 if (*this == Dest) return true;
5523 // Don't search too deeply, we just want to be able to see through
5524 // TokenFactor's etc.
5525 if (Depth == 0) return false;
5527 // If this is a token factor, all inputs to the TF happen in parallel. If any
5528 // of the operands of the TF reach dest, then we can do the xform.
5529 if (getOpcode() == ISD::TokenFactor) {
5530 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5531 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
5536 // Loads don't have side effects, look through them.
5537 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
5538 if (!Ld->isVolatile())
5539 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
5544 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
5545 /// is either an operand of N or it can be reached by traversing up the operands.
5546 /// NOTE: this is an expensive method. Use it carefully.
5547 bool SDNode::isPredecessorOf(SDNode *N) const {
5548 SmallPtrSet<SDNode *, 32> Visited;
5549 SmallVector<SDNode *, 16> Worklist;
5550 Worklist.push_back(N);
5553 N = Worklist.pop_back_val();
5554 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5555 SDNode *Op = N->getOperand(i).getNode();
5558 if (Visited.insert(Op))
5559 Worklist.push_back(Op);
5561 } while (!Worklist.empty());
5566 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
5567 assert(Num < NumOperands && "Invalid child # of SDNode!");
5568 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
5571 std::string SDNode::getOperationName(const SelectionDAG *G) const {
5572 switch (getOpcode()) {
5574 if (getOpcode() < ISD::BUILTIN_OP_END)
5575 return "<<Unknown DAG Node>>";
5576 if (isMachineOpcode()) {
5578 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
5579 if (getMachineOpcode() < TII->getNumOpcodes())
5580 return TII->get(getMachineOpcode()).getName();
5581 return "<<Unknown Machine Node #" + utostr(getOpcode()) + ">>";
5584 const TargetLowering &TLI = G->getTargetLoweringInfo();
5585 const char *Name = TLI.getTargetNodeName(getOpcode());
5586 if (Name) return Name;
5587 return "<<Unknown Target Node #" + utostr(getOpcode()) + ">>";
5589 return "<<Unknown Node #" + utostr(getOpcode()) + ">>";
5592 case ISD::DELETED_NODE:
5593 return "<<Deleted Node!>>";
5595 case ISD::PREFETCH: return "Prefetch";
5596 case ISD::MEMBARRIER: return "MemBarrier";
5597 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap";
5598 case ISD::ATOMIC_SWAP: return "AtomicSwap";
5599 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd";
5600 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub";
5601 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd";
5602 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr";
5603 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor";
5604 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand";
5605 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin";
5606 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax";
5607 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin";
5608 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax";
5609 case ISD::PCMARKER: return "PCMarker";
5610 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5611 case ISD::SRCVALUE: return "SrcValue";
5612 case ISD::MDNODE_SDNODE: return "MDNode";
5613 case ISD::EntryToken: return "EntryToken";
5614 case ISD::TokenFactor: return "TokenFactor";
5615 case ISD::AssertSext: return "AssertSext";
5616 case ISD::AssertZext: return "AssertZext";
5618 case ISD::BasicBlock: return "BasicBlock";
5619 case ISD::VALUETYPE: return "ValueType";
5620 case ISD::Register: return "Register";
5622 case ISD::Constant: return "Constant";
5623 case ISD::ConstantFP: return "ConstantFP";
5624 case ISD::GlobalAddress: return "GlobalAddress";
5625 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5626 case ISD::FrameIndex: return "FrameIndex";
5627 case ISD::JumpTable: return "JumpTable";
5628 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5629 case ISD::RETURNADDR: return "RETURNADDR";
5630 case ISD::FRAMEADDR: return "FRAMEADDR";
5631 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5632 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5633 case ISD::LSDAADDR: return "LSDAADDR";
5634 case ISD::EHSELECTION: return "EHSELECTION";
5635 case ISD::EH_RETURN: return "EH_RETURN";
5636 case ISD::EH_SJLJ_SETJMP: return "EH_SJLJ_SETJMP";
5637 case ISD::EH_SJLJ_LONGJMP: return "EH_SJLJ_LONGJMP";
5638 case ISD::ConstantPool: return "ConstantPool";
5639 case ISD::ExternalSymbol: return "ExternalSymbol";
5640 case ISD::BlockAddress: return "BlockAddress";
5641 case ISD::INTRINSIC_WO_CHAIN:
5642 case ISD::INTRINSIC_VOID:
5643 case ISD::INTRINSIC_W_CHAIN: {
5644 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1;
5645 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue();
5646 if (IID < Intrinsic::num_intrinsics)
5647 return Intrinsic::getName((Intrinsic::ID)IID);
5648 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo())
5649 return TII->getName(IID);
5650 llvm_unreachable("Invalid intrinsic ID");
5653 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5654 case ISD::TargetConstant: return "TargetConstant";
5655 case ISD::TargetConstantFP:return "TargetConstantFP";
5656 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5657 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5658 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5659 case ISD::TargetJumpTable: return "TargetJumpTable";
5660 case ISD::TargetConstantPool: return "TargetConstantPool";
5661 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5662 case ISD::TargetBlockAddress: return "TargetBlockAddress";
5664 case ISD::CopyToReg: return "CopyToReg";
5665 case ISD::CopyFromReg: return "CopyFromReg";
5666 case ISD::UNDEF: return "undef";
5667 case ISD::MERGE_VALUES: return "merge_values";
5668 case ISD::INLINEASM: return "inlineasm";
5669 case ISD::EH_LABEL: return "eh_label";
5670 case ISD::HANDLENODE: return "handlenode";
5673 case ISD::FABS: return "fabs";
5674 case ISD::FNEG: return "fneg";
5675 case ISD::FSQRT: return "fsqrt";
5676 case ISD::FSIN: return "fsin";
5677 case ISD::FCOS: return "fcos";
5678 case ISD::FTRUNC: return "ftrunc";
5679 case ISD::FFLOOR: return "ffloor";
5680 case ISD::FCEIL: return "fceil";
5681 case ISD::FRINT: return "frint";
5682 case ISD::FNEARBYINT: return "fnearbyint";
5683 case ISD::FEXP: return "fexp";
5684 case ISD::FEXP2: return "fexp2";
5685 case ISD::FLOG: return "flog";
5686 case ISD::FLOG2: return "flog2";
5687 case ISD::FLOG10: return "flog10";
5690 case ISD::ADD: return "add";
5691 case ISD::SUB: return "sub";
5692 case ISD::MUL: return "mul";
5693 case ISD::MULHU: return "mulhu";
5694 case ISD::MULHS: return "mulhs";
5695 case ISD::SDIV: return "sdiv";
5696 case ISD::UDIV: return "udiv";
5697 case ISD::SREM: return "srem";
5698 case ISD::UREM: return "urem";
5699 case ISD::SMUL_LOHI: return "smul_lohi";
5700 case ISD::UMUL_LOHI: return "umul_lohi";
5701 case ISD::SDIVREM: return "sdivrem";
5702 case ISD::UDIVREM: return "udivrem";
5703 case ISD::AND: return "and";
5704 case ISD::OR: return "or";
5705 case ISD::XOR: return "xor";
5706 case ISD::SHL: return "shl";
5707 case ISD::SRA: return "sra";
5708 case ISD::SRL: return "srl";
5709 case ISD::ROTL: return "rotl";
5710 case ISD::ROTR: return "rotr";
5711 case ISD::FADD: return "fadd";
5712 case ISD::FSUB: return "fsub";
5713 case ISD::FMUL: return "fmul";
5714 case ISD::FDIV: return "fdiv";
5715 case ISD::FREM: return "frem";
5716 case ISD::FCOPYSIGN: return "fcopysign";
5717 case ISD::FGETSIGN: return "fgetsign";
5718 case ISD::FPOW: return "fpow";
5720 case ISD::FPOWI: return "fpowi";
5721 case ISD::SETCC: return "setcc";
5722 case ISD::VSETCC: return "vsetcc";
5723 case ISD::SELECT: return "select";
5724 case ISD::SELECT_CC: return "select_cc";
5725 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5726 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5727 case ISD::CONCAT_VECTORS: return "concat_vectors";
5728 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5729 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5730 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5731 case ISD::CARRY_FALSE: return "carry_false";
5732 case ISD::ADDC: return "addc";
5733 case ISD::ADDE: return "adde";
5734 case ISD::SADDO: return "saddo";
5735 case ISD::UADDO: return "uaddo";
5736 case ISD::SSUBO: return "ssubo";
5737 case ISD::USUBO: return "usubo";
5738 case ISD::SMULO: return "smulo";
5739 case ISD::UMULO: return "umulo";
5740 case ISD::SUBC: return "subc";
5741 case ISD::SUBE: return "sube";
5742 case ISD::SHL_PARTS: return "shl_parts";
5743 case ISD::SRA_PARTS: return "sra_parts";
5744 case ISD::SRL_PARTS: return "srl_parts";
5746 // Conversion operators.
5747 case ISD::SIGN_EXTEND: return "sign_extend";
5748 case ISD::ZERO_EXTEND: return "zero_extend";
5749 case ISD::ANY_EXTEND: return "any_extend";
5750 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5751 case ISD::TRUNCATE: return "truncate";
5752 case ISD::FP_ROUND: return "fp_round";
5753 case ISD::FLT_ROUNDS_: return "flt_rounds";
5754 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5755 case ISD::FP_EXTEND: return "fp_extend";
5757 case ISD::SINT_TO_FP: return "sint_to_fp";
5758 case ISD::UINT_TO_FP: return "uint_to_fp";
5759 case ISD::FP_TO_SINT: return "fp_to_sint";
5760 case ISD::FP_TO_UINT: return "fp_to_uint";
5761 case ISD::BIT_CONVERT: return "bit_convert";
5762 case ISD::FP16_TO_FP32: return "fp16_to_fp32";
5763 case ISD::FP32_TO_FP16: return "fp32_to_fp16";
5765 case ISD::CONVERT_RNDSAT: {
5766 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) {
5767 default: llvm_unreachable("Unknown cvt code!");
5768 case ISD::CVT_FF: return "cvt_ff";
5769 case ISD::CVT_FS: return "cvt_fs";
5770 case ISD::CVT_FU: return "cvt_fu";
5771 case ISD::CVT_SF: return "cvt_sf";
5772 case ISD::CVT_UF: return "cvt_uf";
5773 case ISD::CVT_SS: return "cvt_ss";
5774 case ISD::CVT_SU: return "cvt_su";
5775 case ISD::CVT_US: return "cvt_us";
5776 case ISD::CVT_UU: return "cvt_uu";
5780 // Control flow instructions
5781 case ISD::BR: return "br";
5782 case ISD::BRIND: return "brind";
5783 case ISD::BR_JT: return "br_jt";
5784 case ISD::BRCOND: return "brcond";
5785 case ISD::BR_CC: return "br_cc";
5786 case ISD::CALLSEQ_START: return "callseq_start";
5787 case ISD::CALLSEQ_END: return "callseq_end";
5790 case ISD::LOAD: return "load";
5791 case ISD::STORE: return "store";
5792 case ISD::VAARG: return "vaarg";
5793 case ISD::VACOPY: return "vacopy";
5794 case ISD::VAEND: return "vaend";
5795 case ISD::VASTART: return "vastart";
5796 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5797 case ISD::EXTRACT_ELEMENT: return "extract_element";
5798 case ISD::BUILD_PAIR: return "build_pair";
5799 case ISD::STACKSAVE: return "stacksave";
5800 case ISD::STACKRESTORE: return "stackrestore";
5801 case ISD::TRAP: return "trap";
5804 case ISD::BSWAP: return "bswap";
5805 case ISD::CTPOP: return "ctpop";
5806 case ISD::CTTZ: return "cttz";
5807 case ISD::CTLZ: return "ctlz";
5810 case ISD::TRAMPOLINE: return "trampoline";
5813 switch (cast<CondCodeSDNode>(this)->get()) {
5814 default: llvm_unreachable("Unknown setcc condition!");
5815 case ISD::SETOEQ: return "setoeq";
5816 case ISD::SETOGT: return "setogt";
5817 case ISD::SETOGE: return "setoge";
5818 case ISD::SETOLT: return "setolt";
5819 case ISD::SETOLE: return "setole";
5820 case ISD::SETONE: return "setone";
5822 case ISD::SETO: return "seto";
5823 case ISD::SETUO: return "setuo";
5824 case ISD::SETUEQ: return "setue";
5825 case ISD::SETUGT: return "setugt";
5826 case ISD::SETUGE: return "setuge";
5827 case ISD::SETULT: return "setult";
5828 case ISD::SETULE: return "setule";
5829 case ISD::SETUNE: return "setune";
5831 case ISD::SETEQ: return "seteq";
5832 case ISD::SETGT: return "setgt";
5833 case ISD::SETGE: return "setge";
5834 case ISD::SETLT: return "setlt";
5835 case ISD::SETLE: return "setle";
5836 case ISD::SETNE: return "setne";
5841 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5850 return "<post-inc>";
5852 return "<post-dec>";
5856 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5857 std::string S = "< ";
5871 if (getByValAlign())
5872 S += "byval-align:" + utostr(getByValAlign()) + " ";
5874 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5876 S += "byval-size:" + utostr(getByValSize()) + " ";
5880 void SDNode::dump() const { dump(0); }
5881 void SDNode::dump(const SelectionDAG *G) const {
5885 void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const {
5886 OS << (void*)this << ": ";
5888 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5890 if (getValueType(i) == MVT::Other)
5893 OS << getValueType(i).getEVTString();
5895 OS << " = " << getOperationName(G);
5898 void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const {
5899 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) {
5900 if (!MN->memoperands_empty()) {
5903 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(),
5904 e = MN->memoperands_end(); i != e; ++i) {
5911 } else if (const ShuffleVectorSDNode *SVN =
5912 dyn_cast<ShuffleVectorSDNode>(this)) {
5914 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) {
5915 int Idx = SVN->getMaskElt(i);
5923 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5924 OS << '<' << CSDN->getAPIntValue() << '>';
5925 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5926 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5927 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5928 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5929 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5932 CSDN->getValueAPF().bitcastToAPInt().dump();
5935 } else if (const GlobalAddressSDNode *GADN =
5936 dyn_cast<GlobalAddressSDNode>(this)) {
5937 int64_t offset = GADN->getOffset();
5939 WriteAsOperand(OS, GADN->getGlobal());
5942 OS << " + " << offset;
5944 OS << " " << offset;
5945 if (unsigned int TF = GADN->getTargetFlags())
5946 OS << " [TF=" << TF << ']';
5947 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5948 OS << "<" << FIDN->getIndex() << ">";
5949 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5950 OS << "<" << JTDN->getIndex() << ">";
5951 if (unsigned int TF = JTDN->getTargetFlags())
5952 OS << " [TF=" << TF << ']';
5953 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5954 int offset = CP->getOffset();
5955 if (CP->isMachineConstantPoolEntry())
5956 OS << "<" << *CP->getMachineCPVal() << ">";
5958 OS << "<" << *CP->getConstVal() << ">";
5960 OS << " + " << offset;
5962 OS << " " << offset;
5963 if (unsigned int TF = CP->getTargetFlags())
5964 OS << " [TF=" << TF << ']';
5965 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5967 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5969 OS << LBB->getName() << " ";
5970 OS << (const void*)BBDN->getBasicBlock() << ">";
5971 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5972 if (G && R->getReg() &&
5973 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5974 OS << " %" << G->getTarget().getRegisterInfo()->getName(R->getReg());
5976 OS << " %reg" << R->getReg();
5978 } else if (const ExternalSymbolSDNode *ES =
5979 dyn_cast<ExternalSymbolSDNode>(this)) {
5980 OS << "'" << ES->getSymbol() << "'";
5981 if (unsigned int TF = ES->getTargetFlags())
5982 OS << " [TF=" << TF << ']';
5983 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5985 OS << "<" << M->getValue() << ">";
5988 } else if (const MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(this)) {
5990 OS << "<" << MD->getMD() << ">";
5993 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5994 OS << ":" << N->getVT().getEVTString();
5996 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5997 OS << "<" << *LD->getMemOperand();
6000 switch (LD->getExtensionType()) {
6001 default: doExt = false; break;
6002 case ISD::EXTLOAD: OS << ", anyext"; break;
6003 case ISD::SEXTLOAD: OS << ", sext"; break;
6004 case ISD::ZEXTLOAD: OS << ", zext"; break;
6007 OS << " from " << LD->getMemoryVT().getEVTString();
6009 const char *AM = getIndexedModeName(LD->getAddressingMode());
6014 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
6015 OS << "<" << *ST->getMemOperand();
6017 if (ST->isTruncatingStore())
6018 OS << ", trunc to " << ST->getMemoryVT().getEVTString();
6020 const char *AM = getIndexedModeName(ST->getAddressingMode());
6025 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) {
6026 OS << "<" << *M->getMemOperand() << ">";
6027 } else if (const BlockAddressSDNode *BA =
6028 dyn_cast<BlockAddressSDNode>(this)) {
6030 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false);
6032 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false);
6034 if (unsigned int TF = BA->getTargetFlags())
6035 OS << " [TF=" << TF << ']';
6039 if (unsigned Order = G->GetOrdering(this))
6040 OS << " [ORD=" << Order << ']';
6042 if (getNodeId() != -1)
6043 OS << " [ID=" << getNodeId() << ']';
6045 DebugLoc dl = getDebugLoc();
6046 if (G && !dl.isUnknown()) {
6048 Scope(dl.getScope(G->getMachineFunction().getFunction()->getContext()));
6050 // Omit the directory, since it's usually long and uninteresting.
6052 OS << Scope.getFilename();
6055 OS << ':' << dl.getLine();
6056 if (dl.getCol() != 0)
6057 OS << ':' << dl.getCol();
6061 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
6063 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
6064 if (i) OS << ", "; else OS << " ";
6065 OS << (void*)getOperand(i).getNode();
6066 if (unsigned RN = getOperand(i).getResNo())
6069 print_details(OS, G);
6072 static void printrWithDepthHelper(raw_ostream &OS, const SDNode *N,
6073 const SelectionDAG *G, unsigned depth,
6086 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6088 printrWithDepthHelper(OS, N->getOperand(i).getNode(), G, depth-1, indent+2);
6092 void SDNode::printrWithDepth(raw_ostream &OS, const SelectionDAG *G,
6093 unsigned depth) const {
6094 printrWithDepthHelper(OS, this, G, depth, 0);
6097 void SDNode::printrFull(raw_ostream &OS, const SelectionDAG *G) const {
6098 // Don't print impossibly deep things.
6099 printrWithDepth(OS, G, 100);
6102 void SDNode::dumprWithDepth(const SelectionDAG *G, unsigned depth) const {
6103 printrWithDepth(dbgs(), G, depth);
6106 void SDNode::dumprFull(const SelectionDAG *G) const {
6107 // Don't print impossibly deep things.
6108 dumprWithDepth(G, 100);
6111 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
6112 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
6113 if (N->getOperand(i).getNode()->hasOneUse())
6114 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
6116 dbgs() << "\n" << std::string(indent+2, ' ')
6117 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
6121 dbgs().indent(indent);
6125 SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) {
6126 assert(N->getNumValues() == 1 &&
6127 "Can't unroll a vector with multiple results!");
6129 EVT VT = N->getValueType(0);
6130 unsigned NE = VT.getVectorNumElements();
6131 EVT EltVT = VT.getVectorElementType();
6132 DebugLoc dl = N->getDebugLoc();
6134 SmallVector<SDValue, 8> Scalars;
6135 SmallVector<SDValue, 4> Operands(N->getNumOperands());
6137 // If ResNE is 0, fully unroll the vector op.
6140 else if (NE > ResNE)
6144 for (i= 0; i != NE; ++i) {
6145 for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) {
6146 SDValue Operand = N->getOperand(j);
6147 EVT OperandVT = Operand.getValueType();
6148 if (OperandVT.isVector()) {
6149 // A vector operand; extract a single element.
6150 EVT OperandEltVT = OperandVT.getVectorElementType();
6151 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl,
6154 getConstant(i, MVT::i32));
6156 // A scalar operand; just use it as is.
6157 Operands[j] = Operand;
6161 switch (N->getOpcode()) {
6163 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
6164 &Operands[0], Operands.size()));
6171 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0],
6172 getShiftAmountOperand(Operands[1])));
6174 case ISD::SIGN_EXTEND_INREG:
6175 case ISD::FP_ROUND_INREG: {
6176 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType();
6177 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
6179 getValueType(ExtVT)));
6184 for (; i < ResNE; ++i)
6185 Scalars.push_back(getUNDEF(EltVT));
6187 return getNode(ISD::BUILD_VECTOR, dl,
6188 EVT::getVectorVT(*getContext(), EltVT, ResNE),
6189 &Scalars[0], Scalars.size());
6193 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
6194 /// location that is 'Dist' units away from the location that the 'Base' load
6195 /// is loading from.
6196 bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
6197 unsigned Bytes, int Dist) const {
6198 if (LD->getChain() != Base->getChain())
6200 EVT VT = LD->getValueType(0);
6201 if (VT.getSizeInBits() / 8 != Bytes)
6204 SDValue Loc = LD->getOperand(1);
6205 SDValue BaseLoc = Base->getOperand(1);
6206 if (Loc.getOpcode() == ISD::FrameIndex) {
6207 if (BaseLoc.getOpcode() != ISD::FrameIndex)
6209 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo();
6210 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
6211 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
6212 int FS = MFI->getObjectSize(FI);
6213 int BFS = MFI->getObjectSize(BFI);
6214 if (FS != BFS || FS != (int)Bytes) return false;
6215 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
6217 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
6218 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
6219 if (V && (V->getSExtValue() == Dist*Bytes))
6223 const GlobalValue *GV1 = NULL;
6224 const GlobalValue *GV2 = NULL;
6225 int64_t Offset1 = 0;
6226 int64_t Offset2 = 0;
6227 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
6228 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
6229 if (isGA1 && isGA2 && GV1 == GV2)
6230 return Offset1 == (Offset2 + Dist*Bytes);
6235 /// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if
6236 /// it cannot be inferred.
6237 unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const {
6238 // If this is a GlobalAddress + cst, return the alignment.
6239 const GlobalValue *GV;
6240 int64_t GVOffset = 0;
6241 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) {
6242 // If GV has specified alignment, then use it. Otherwise, use the preferred
6244 unsigned Align = GV->getAlignment();
6246 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
6247 if (GVar->hasInitializer()) {
6248 const TargetData *TD = TLI.getTargetData();
6249 Align = TD->getPreferredAlignment(GVar);
6253 return MinAlign(Align, GVOffset);
6256 // If this is a direct reference to a stack slot, use information about the
6257 // stack slot's alignment.
6258 int FrameIdx = 1 << 31;
6259 int64_t FrameOffset = 0;
6260 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
6261 FrameIdx = FI->getIndex();
6262 } else if (Ptr.getOpcode() == ISD::ADD &&
6263 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
6264 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
6265 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
6266 FrameOffset = Ptr.getConstantOperandVal(1);
6269 if (FrameIdx != (1 << 31)) {
6270 // FIXME: Handle FI+CST.
6271 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo();
6272 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
6274 if (MFI.isFixedObjectIndex(FrameIdx)) {
6275 int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx) + FrameOffset;
6277 // The alignment of the frame index can be determined from its offset from
6278 // the incoming frame position. If the frame object is at offset 32 and
6279 // the stack is guaranteed to be 16-byte aligned, then we know that the
6280 // object is 16-byte aligned.
6281 unsigned StackAlign = getTarget().getFrameInfo()->getStackAlignment();
6282 unsigned Align = MinAlign(ObjectOffset, StackAlign);
6284 // Finally, the frame object itself may have a known alignment. Factor
6285 // the alignment + offset into a new alignment. For example, if we know
6286 // the FI is 8 byte aligned, but the pointer is 4 off, we really have a
6287 // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte
6288 // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc.
6289 return std::max(Align, FIInfoAlign);
6297 void SelectionDAG::dump() const {
6298 dbgs() << "SelectionDAG has " << AllNodes.size() << " nodes:";
6300 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
6302 const SDNode *N = I;
6303 if (!N->hasOneUse() && N != getRoot().getNode())
6304 DumpNodes(N, 2, this);
6307 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
6312 void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
6314 print_details(OS, G);
6317 typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet;
6318 static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent,
6319 const SelectionDAG *G, VisitedSDNodeSet &once) {
6320 if (!once.insert(N)) // If we've been here before, return now.
6323 // Dump the current SDNode, but don't end the line yet.
6324 OS << std::string(indent, ' ');
6327 // Having printed this SDNode, walk the children:
6328 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6329 const SDNode *child = N->getOperand(i).getNode();
6334 if (child->getNumOperands() == 0) {
6335 // This child has no grandchildren; print it inline right here.
6336 child->printr(OS, G);
6338 } else { // Just the address. FIXME: also print the child's opcode.
6340 if (unsigned RN = N->getOperand(i).getResNo())
6347 // Dump children that have grandchildren on their own line(s).
6348 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6349 const SDNode *child = N->getOperand(i).getNode();
6350 DumpNodesr(OS, child, indent+2, G, once);
6354 void SDNode::dumpr() const {
6355 VisitedSDNodeSet once;
6356 DumpNodesr(dbgs(), this, 0, 0, once);
6359 void SDNode::dumpr(const SelectionDAG *G) const {
6360 VisitedSDNodeSet once;
6361 DumpNodesr(dbgs(), this, 0, G, once);
6365 // getAddressSpace - Return the address space this GlobalAddress belongs to.
6366 unsigned GlobalAddressSDNode::getAddressSpace() const {
6367 return getGlobal()->getType()->getAddressSpace();
6371 const Type *ConstantPoolSDNode::getType() const {
6372 if (isMachineConstantPoolEntry())
6373 return Val.MachineCPVal->getType();
6374 return Val.ConstVal->getType();
6377 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue,
6379 unsigned &SplatBitSize,
6381 unsigned MinSplatBits,
6383 EVT VT = getValueType(0);
6384 assert(VT.isVector() && "Expected a vector type");
6385 unsigned sz = VT.getSizeInBits();
6386 if (MinSplatBits > sz)
6389 SplatValue = APInt(sz, 0);
6390 SplatUndef = APInt(sz, 0);
6392 // Get the bits. Bits with undefined values (when the corresponding element
6393 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
6394 // in SplatValue. If any of the values are not constant, give up and return
6396 unsigned int nOps = getNumOperands();
6397 assert(nOps > 0 && "isConstantSplat has 0-size build vector");
6398 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits();
6400 for (unsigned j = 0; j < nOps; ++j) {
6401 unsigned i = isBigEndian ? nOps-1-j : j;
6402 SDValue OpVal = getOperand(i);
6403 unsigned BitPos = j * EltBitSize;
6405 if (OpVal.getOpcode() == ISD::UNDEF)
6406 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize);
6407 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal))
6408 SplatValue |= APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize).
6409 zextOrTrunc(sz) << BitPos;
6410 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal))
6411 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos;
6416 // The build_vector is all constants or undefs. Find the smallest element
6417 // size that splats the vector.
6419 HasAnyUndefs = (SplatUndef != 0);
6422 unsigned HalfSize = sz / 2;
6423 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize);
6424 APInt LowValue = APInt(SplatValue).trunc(HalfSize);
6425 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize);
6426 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize);
6428 // If the two halves do not match (ignoring undef bits), stop here.
6429 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
6430 MinSplatBits > HalfSize)
6433 SplatValue = HighValue | LowValue;
6434 SplatUndef = HighUndef & LowUndef;
6443 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) {
6444 // Find the first non-undef value in the shuffle mask.
6446 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
6449 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!");
6451 // Make sure all remaining elements are either undef or the same as the first
6453 for (int Idx = Mask[i]; i != e; ++i)
6454 if (Mask[i] >= 0 && Mask[i] != Idx)
6460 static void checkForCyclesHelper(const SDNode *N,
6461 SmallPtrSet<const SDNode*, 32> &Visited,
6462 SmallPtrSet<const SDNode*, 32> &Checked) {
6463 // If this node has already been checked, don't check it again.
6464 if (Checked.count(N))
6467 // If a node has already been visited on this depth-first walk, reject it as
6469 if (!Visited.insert(N)) {
6470 dbgs() << "Offending node:\n";
6472 errs() << "Detected cycle in SelectionDAG\n";
6476 for(unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
6477 checkForCyclesHelper(N->getOperand(i).getNode(), Visited, Checked);
6484 void llvm::checkForCycles(const llvm::SDNode *N) {
6486 assert(N && "Checking nonexistant SDNode");
6487 SmallPtrSet<const SDNode*, 32> visited;
6488 SmallPtrSet<const SDNode*, 32> checked;
6489 checkForCyclesHelper(N, visited, checked);
6493 void llvm::checkForCycles(const llvm::SelectionDAG *DAG) {
6494 checkForCycles(DAG->getRoot().getNode());