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/TargetOptions.h"
37 #include "llvm/Target/TargetInstrInfo.h"
38 #include "llvm/Target/TargetIntrinsicInfo.h"
39 #include "llvm/Target/TargetMachine.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/ManagedStatic.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/System/Mutex.h"
47 #include "llvm/ADT/SetVector.h"
48 #include "llvm/ADT/SmallPtrSet.h"
49 #include "llvm/ADT/SmallSet.h"
50 #include "llvm/ADT/SmallVector.h"
51 #include "llvm/ADT/StringExtras.h"
56 /// makeVTList - Return an instance of the SDVTList struct initialized with the
57 /// specified members.
58 static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
59 SDVTList Res = {VTs, NumVTs};
63 static const fltSemantics *EVTToAPFloatSemantics(EVT VT) {
64 switch (VT.getSimpleVT().SimpleTy) {
65 default: llvm_unreachable("Unknown FP format");
66 case MVT::f32: return &APFloat::IEEEsingle;
67 case MVT::f64: return &APFloat::IEEEdouble;
68 case MVT::f80: return &APFloat::x87DoubleExtended;
69 case MVT::f128: return &APFloat::IEEEquad;
70 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
74 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
76 //===----------------------------------------------------------------------===//
77 // ConstantFPSDNode Class
78 //===----------------------------------------------------------------------===//
80 /// isExactlyValue - We don't rely on operator== working on double values, as
81 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
82 /// As such, this method can be used to do an exact bit-for-bit comparison of
83 /// two floating point values.
84 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
85 return getValueAPF().bitwiseIsEqual(V);
88 bool ConstantFPSDNode::isValueValidForType(EVT VT,
90 assert(VT.isFloatingPoint() && "Can only convert between FP types");
92 // PPC long double cannot be converted to any other type.
93 if (VT == MVT::ppcf128 ||
94 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
97 // convert modifies in place, so make a copy.
98 APFloat Val2 = APFloat(Val);
100 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
105 //===----------------------------------------------------------------------===//
107 //===----------------------------------------------------------------------===//
109 /// isBuildVectorAllOnes - Return true if the specified node is a
110 /// BUILD_VECTOR where all of the elements are ~0 or undef.
111 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
112 // Look through a bit convert.
113 if (N->getOpcode() == ISD::BIT_CONVERT)
114 N = N->getOperand(0).getNode();
116 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
118 unsigned i = 0, e = N->getNumOperands();
120 // Skip over all of the undef values.
121 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
124 // Do not accept an all-undef vector.
125 if (i == e) return false;
127 // Do not accept build_vectors that aren't all constants or which have non-~0
129 SDValue NotZero = N->getOperand(i);
130 if (isa<ConstantSDNode>(NotZero)) {
131 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
133 } else if (isa<ConstantFPSDNode>(NotZero)) {
134 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
135 bitcastToAPInt().isAllOnesValue())
140 // Okay, we have at least one ~0 value, check to see if the rest match or are
142 for (++i; i != e; ++i)
143 if (N->getOperand(i) != NotZero &&
144 N->getOperand(i).getOpcode() != ISD::UNDEF)
150 /// isBuildVectorAllZeros - Return true if the specified node is a
151 /// BUILD_VECTOR where all of the elements are 0 or undef.
152 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
153 // Look through a bit convert.
154 if (N->getOpcode() == ISD::BIT_CONVERT)
155 N = N->getOperand(0).getNode();
157 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
159 unsigned i = 0, e = N->getNumOperands();
161 // Skip over all of the undef values.
162 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
165 // Do not accept an all-undef vector.
166 if (i == e) return false;
168 // Do not accept build_vectors that aren't all constants or which have non-0
170 SDValue Zero = N->getOperand(i);
171 if (isa<ConstantSDNode>(Zero)) {
172 if (!cast<ConstantSDNode>(Zero)->isNullValue())
174 } else if (isa<ConstantFPSDNode>(Zero)) {
175 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
180 // Okay, we have at least one 0 value, check to see if the rest match or are
182 for (++i; i != e; ++i)
183 if (N->getOperand(i) != Zero &&
184 N->getOperand(i).getOpcode() != ISD::UNDEF)
189 /// isScalarToVector - Return true if the specified node is a
190 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
191 /// element is not an undef.
192 bool ISD::isScalarToVector(const SDNode *N) {
193 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
196 if (N->getOpcode() != ISD::BUILD_VECTOR)
198 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
200 unsigned NumElems = N->getNumOperands();
201 for (unsigned i = 1; i < NumElems; ++i) {
202 SDValue V = N->getOperand(i);
203 if (V.getOpcode() != ISD::UNDEF)
209 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
210 /// when given the operation for (X op Y).
211 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
212 // To perform this operation, we just need to swap the L and G bits of the
214 unsigned OldL = (Operation >> 2) & 1;
215 unsigned OldG = (Operation >> 1) & 1;
216 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
217 (OldL << 1) | // New G bit
218 (OldG << 2)); // New L bit.
221 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
222 /// 'op' is a valid SetCC operation.
223 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
224 unsigned Operation = Op;
226 Operation ^= 7; // Flip L, G, E bits, but not U.
228 Operation ^= 15; // Flip all of the condition bits.
230 if (Operation > ISD::SETTRUE2)
231 Operation &= ~8; // Don't let N and U bits get set.
233 return ISD::CondCode(Operation);
237 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
238 /// signed operation and 2 if the result is an unsigned comparison. Return zero
239 /// if the operation does not depend on the sign of the input (setne and seteq).
240 static int isSignedOp(ISD::CondCode Opcode) {
242 default: llvm_unreachable("Illegal integer setcc operation!");
244 case ISD::SETNE: return 0;
248 case ISD::SETGE: return 1;
252 case ISD::SETUGE: return 2;
256 /// getSetCCOrOperation - Return the result of a logical OR between different
257 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
258 /// returns SETCC_INVALID if it is not possible to represent the resultant
260 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
262 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
263 // Cannot fold a signed integer setcc with an unsigned integer setcc.
264 return ISD::SETCC_INVALID;
266 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
268 // If the N and U bits get set then the resultant comparison DOES suddenly
269 // care about orderedness, and is true when ordered.
270 if (Op > ISD::SETTRUE2)
271 Op &= ~16; // Clear the U bit if the N bit is set.
273 // Canonicalize illegal integer setcc's.
274 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
277 return ISD::CondCode(Op);
280 /// getSetCCAndOperation - Return the result of a logical AND between different
281 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
282 /// function returns zero if it is not possible to represent the resultant
284 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
286 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
287 // Cannot fold a signed setcc with an unsigned setcc.
288 return ISD::SETCC_INVALID;
290 // Combine all of the condition bits.
291 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
293 // Canonicalize illegal integer setcc's.
297 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
298 case ISD::SETOEQ: // SETEQ & SETU[LG]E
299 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
300 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
301 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
308 //===----------------------------------------------------------------------===//
309 // SDNode Profile Support
310 //===----------------------------------------------------------------------===//
312 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
314 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
318 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
319 /// solely with their pointer.
320 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
321 ID.AddPointer(VTList.VTs);
324 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
326 static void AddNodeIDOperands(FoldingSetNodeID &ID,
327 const SDValue *Ops, unsigned NumOps) {
328 for (; NumOps; --NumOps, ++Ops) {
329 ID.AddPointer(Ops->getNode());
330 ID.AddInteger(Ops->getResNo());
334 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
336 static void AddNodeIDOperands(FoldingSetNodeID &ID,
337 const SDUse *Ops, unsigned NumOps) {
338 for (; NumOps; --NumOps, ++Ops) {
339 ID.AddPointer(Ops->getNode());
340 ID.AddInteger(Ops->getResNo());
344 static void AddNodeIDNode(FoldingSetNodeID &ID,
345 unsigned short OpC, SDVTList VTList,
346 const SDValue *OpList, unsigned N) {
347 AddNodeIDOpcode(ID, OpC);
348 AddNodeIDValueTypes(ID, VTList);
349 AddNodeIDOperands(ID, OpList, N);
352 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
354 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
355 switch (N->getOpcode()) {
356 case ISD::TargetExternalSymbol:
357 case ISD::ExternalSymbol:
358 llvm_unreachable("Should only be used on nodes with operands");
359 default: break; // Normal nodes don't need extra info.
360 case ISD::TargetConstant:
362 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
364 case ISD::TargetConstantFP:
365 case ISD::ConstantFP: {
366 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
369 case ISD::TargetGlobalAddress:
370 case ISD::GlobalAddress:
371 case ISD::TargetGlobalTLSAddress:
372 case ISD::GlobalTLSAddress: {
373 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
374 ID.AddPointer(GA->getGlobal());
375 ID.AddInteger(GA->getOffset());
376 ID.AddInteger(GA->getTargetFlags());
379 case ISD::BasicBlock:
380 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
383 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
387 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
389 case ISD::FrameIndex:
390 case ISD::TargetFrameIndex:
391 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
394 case ISD::TargetJumpTable:
395 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
396 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
398 case ISD::ConstantPool:
399 case ISD::TargetConstantPool: {
400 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
401 ID.AddInteger(CP->getAlignment());
402 ID.AddInteger(CP->getOffset());
403 if (CP->isMachineConstantPoolEntry())
404 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
406 ID.AddPointer(CP->getConstVal());
407 ID.AddInteger(CP->getTargetFlags());
411 const LoadSDNode *LD = cast<LoadSDNode>(N);
412 ID.AddInteger(LD->getMemoryVT().getRawBits());
413 ID.AddInteger(LD->getRawSubclassData());
417 const StoreSDNode *ST = cast<StoreSDNode>(N);
418 ID.AddInteger(ST->getMemoryVT().getRawBits());
419 ID.AddInteger(ST->getRawSubclassData());
422 case ISD::ATOMIC_CMP_SWAP:
423 case ISD::ATOMIC_SWAP:
424 case ISD::ATOMIC_LOAD_ADD:
425 case ISD::ATOMIC_LOAD_SUB:
426 case ISD::ATOMIC_LOAD_AND:
427 case ISD::ATOMIC_LOAD_OR:
428 case ISD::ATOMIC_LOAD_XOR:
429 case ISD::ATOMIC_LOAD_NAND:
430 case ISD::ATOMIC_LOAD_MIN:
431 case ISD::ATOMIC_LOAD_MAX:
432 case ISD::ATOMIC_LOAD_UMIN:
433 case ISD::ATOMIC_LOAD_UMAX: {
434 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
435 ID.AddInteger(AT->getMemoryVT().getRawBits());
436 ID.AddInteger(AT->getRawSubclassData());
439 case ISD::VECTOR_SHUFFLE: {
440 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
441 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
443 ID.AddInteger(SVN->getMaskElt(i));
446 case ISD::TargetBlockAddress:
447 case ISD::BlockAddress: {
448 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress());
449 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags());
452 } // end switch (N->getOpcode())
455 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
457 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
458 AddNodeIDOpcode(ID, N->getOpcode());
459 // Add the return value info.
460 AddNodeIDValueTypes(ID, N->getVTList());
461 // Add the operand info.
462 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
464 // Handle SDNode leafs with special info.
465 AddNodeIDCustom(ID, N);
468 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
469 /// the CSE map that carries volatility, temporalness, indexing mode, and
470 /// extension/truncation information.
472 static inline unsigned
473 encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile,
474 bool isNonTemporal) {
475 assert((ConvType & 3) == ConvType &&
476 "ConvType may not require more than 2 bits!");
477 assert((AM & 7) == AM &&
478 "AM may not require more than 3 bits!");
482 (isNonTemporal << 6);
485 //===----------------------------------------------------------------------===//
486 // SelectionDAG Class
487 //===----------------------------------------------------------------------===//
489 /// doNotCSE - Return true if CSE should not be performed for this node.
490 static bool doNotCSE(SDNode *N) {
491 if (N->getValueType(0) == MVT::Flag)
492 return true; // Never CSE anything that produces a flag.
494 switch (N->getOpcode()) {
496 case ISD::HANDLENODE:
498 return true; // Never CSE these nodes.
501 // Check that remaining values produced are not flags.
502 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
503 if (N->getValueType(i) == MVT::Flag)
504 return true; // Never CSE anything that produces a flag.
509 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
511 void SelectionDAG::RemoveDeadNodes() {
512 // Create a dummy node (which is not added to allnodes), that adds a reference
513 // to the root node, preventing it from being deleted.
514 HandleSDNode Dummy(getRoot());
516 SmallVector<SDNode*, 128> DeadNodes;
518 // Add all obviously-dead nodes to the DeadNodes worklist.
519 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
521 DeadNodes.push_back(I);
523 RemoveDeadNodes(DeadNodes);
525 // If the root changed (e.g. it was a dead load, update the root).
526 setRoot(Dummy.getValue());
529 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
530 /// given list, and any nodes that become unreachable as a result.
531 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
532 DAGUpdateListener *UpdateListener) {
534 // Process the worklist, deleting the nodes and adding their uses to the
536 while (!DeadNodes.empty()) {
537 SDNode *N = DeadNodes.pop_back_val();
540 UpdateListener->NodeDeleted(N, 0);
542 // Take the node out of the appropriate CSE map.
543 RemoveNodeFromCSEMaps(N);
545 // Next, brutally remove the operand list. This is safe to do, as there are
546 // no cycles in the graph.
547 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
549 SDNode *Operand = Use.getNode();
552 // Now that we removed this operand, see if there are no uses of it left.
553 if (Operand->use_empty())
554 DeadNodes.push_back(Operand);
561 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
562 SmallVector<SDNode*, 16> DeadNodes(1, N);
563 RemoveDeadNodes(DeadNodes, UpdateListener);
566 void SelectionDAG::DeleteNode(SDNode *N) {
567 // First take this out of the appropriate CSE map.
568 RemoveNodeFromCSEMaps(N);
570 // Finally, remove uses due to operands of this node, remove from the
571 // AllNodes list, and delete the node.
572 DeleteNodeNotInCSEMaps(N);
575 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
576 assert(N != AllNodes.begin() && "Cannot delete the entry node!");
577 assert(N->use_empty() && "Cannot delete a node that is not dead!");
579 // Drop all of the operands and decrement used node's use counts.
585 void SelectionDAG::DeallocateNode(SDNode *N) {
586 if (N->OperandsNeedDelete)
587 delete[] N->OperandList;
589 // Set the opcode to DELETED_NODE to help catch bugs when node
590 // memory is reallocated.
591 N->NodeType = ISD::DELETED_NODE;
593 NodeAllocator.Deallocate(AllNodes.remove(N));
595 // Remove the ordering of this node.
598 // If any of the SDDbgValue nodes refer to this SDNode, invalidate them.
599 SmallVector<SDDbgValue*, 2> &DbgVals = DbgInfo->getSDDbgValues(N);
600 for (unsigned i = 0, e = DbgVals.size(); i != e; ++i)
601 DbgVals[i]->setIsInvalidated();
604 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
605 /// correspond to it. This is useful when we're about to delete or repurpose
606 /// the node. We don't want future request for structurally identical nodes
607 /// to return N anymore.
608 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
610 switch (N->getOpcode()) {
611 case ISD::EntryToken:
612 llvm_unreachable("EntryToken should not be in CSEMaps!");
614 case ISD::HANDLENODE: return false; // noop.
616 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
617 "Cond code doesn't exist!");
618 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
619 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
621 case ISD::ExternalSymbol:
622 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
624 case ISD::TargetExternalSymbol: {
625 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
626 Erased = TargetExternalSymbols.erase(
627 std::pair<std::string,unsigned char>(ESN->getSymbol(),
628 ESN->getTargetFlags()));
631 case ISD::VALUETYPE: {
632 EVT VT = cast<VTSDNode>(N)->getVT();
633 if (VT.isExtended()) {
634 Erased = ExtendedValueTypeNodes.erase(VT);
636 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0;
637 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0;
642 // Remove it from the CSE Map.
643 Erased = CSEMap.RemoveNode(N);
647 // Verify that the node was actually in one of the CSE maps, unless it has a
648 // flag result (which cannot be CSE'd) or is one of the special cases that are
649 // not subject to CSE.
650 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
651 !N->isMachineOpcode() && !doNotCSE(N)) {
654 llvm_unreachable("Node is not in map!");
660 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
661 /// maps and modified in place. Add it back to the CSE maps, unless an identical
662 /// node already exists, in which case transfer all its users to the existing
663 /// node. This transfer can potentially trigger recursive merging.
666 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N,
667 DAGUpdateListener *UpdateListener) {
668 // For node types that aren't CSE'd, just act as if no identical node
671 SDNode *Existing = CSEMap.GetOrInsertNode(N);
673 // If there was already an existing matching node, use ReplaceAllUsesWith
674 // to replace the dead one with the existing one. This can cause
675 // recursive merging of other unrelated nodes down the line.
676 ReplaceAllUsesWith(N, Existing, UpdateListener);
678 // N is now dead. Inform the listener if it exists and delete it.
680 UpdateListener->NodeDeleted(N, Existing);
681 DeleteNodeNotInCSEMaps(N);
686 // If the node doesn't already exist, we updated it. Inform a listener if
689 UpdateListener->NodeUpdated(N);
692 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
693 /// were replaced with those specified. If this node is never memoized,
694 /// return null, otherwise return a pointer to the slot it would take. If a
695 /// node already exists with these operands, the slot will be non-null.
696 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
701 SDValue Ops[] = { Op };
703 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
704 AddNodeIDCustom(ID, N);
705 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
709 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
710 /// were replaced with those specified. If this node is never memoized,
711 /// return null, otherwise return a pointer to the slot it would take. If a
712 /// node already exists with these operands, the slot will be non-null.
713 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
714 SDValue Op1, SDValue Op2,
719 SDValue Ops[] = { Op1, Op2 };
721 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
722 AddNodeIDCustom(ID, N);
723 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
728 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
729 /// were replaced with those specified. If this node is never memoized,
730 /// return null, otherwise return a pointer to the slot it would take. If a
731 /// node already exists with these operands, the slot will be non-null.
732 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
733 const SDValue *Ops,unsigned NumOps,
739 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
740 AddNodeIDCustom(ID, N);
741 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
745 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
746 void SelectionDAG::VerifyNode(SDNode *N) {
747 switch (N->getOpcode()) {
750 case ISD::BUILD_PAIR: {
751 EVT VT = N->getValueType(0);
752 assert(N->getNumValues() == 1 && "Too many results!");
753 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
754 "Wrong return type!");
755 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
756 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
757 "Mismatched operand types!");
758 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
759 "Wrong operand type!");
760 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
761 "Wrong return type size");
764 case ISD::BUILD_VECTOR: {
765 assert(N->getNumValues() == 1 && "Too many results!");
766 assert(N->getValueType(0).isVector() && "Wrong return type!");
767 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
768 "Wrong number of operands!");
769 EVT EltVT = N->getValueType(0).getVectorElementType();
770 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
771 assert((I->getValueType() == EltVT ||
772 (EltVT.isInteger() && I->getValueType().isInteger() &&
773 EltVT.bitsLE(I->getValueType()))) &&
774 "Wrong operand type!");
780 /// getEVTAlignment - Compute the default alignment value for the
783 unsigned SelectionDAG::getEVTAlignment(EVT VT) const {
784 const Type *Ty = VT == MVT::iPTR ?
785 PointerType::get(Type::getInt8Ty(*getContext()), 0) :
786 VT.getTypeForEVT(*getContext());
788 return TLI.getTargetData()->getABITypeAlignment(Ty);
791 // EntryNode could meaningfully have debug info if we can find it...
792 SelectionDAG::SelectionDAG(const TargetMachine &tm, FunctionLoweringInfo &fli)
793 : TM(tm), TLI(*tm.getTargetLowering()), FLI(fli),
794 EntryNode(ISD::EntryToken, DebugLoc(), getVTList(MVT::Other)),
795 Root(getEntryNode()), Ordering(0) {
796 AllNodes.push_back(&EntryNode);
797 Ordering = new SDNodeOrdering();
798 DbgInfo = new SDDbgInfo();
801 void SelectionDAG::init(MachineFunction &mf) {
803 Context = &mf.getFunction()->getContext();
806 SelectionDAG::~SelectionDAG() {
813 void SelectionDAG::allnodes_clear() {
814 assert(&*AllNodes.begin() == &EntryNode);
815 AllNodes.remove(AllNodes.begin());
816 while (!AllNodes.empty())
817 DeallocateNode(AllNodes.begin());
820 void SelectionDAG::clear() {
822 OperandAllocator.Reset();
825 ExtendedValueTypeNodes.clear();
826 ExternalSymbols.clear();
827 TargetExternalSymbols.clear();
828 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
829 static_cast<CondCodeSDNode*>(0));
830 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
831 static_cast<SDNode*>(0));
833 EntryNode.UseList = 0;
834 AllNodes.push_back(&EntryNode);
835 Root = getEntryNode();
837 Ordering = new SDNodeOrdering();
840 DbgInfo = new SDDbgInfo();
843 SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
844 return VT.bitsGT(Op.getValueType()) ?
845 getNode(ISD::SIGN_EXTEND, DL, VT, Op) :
846 getNode(ISD::TRUNCATE, DL, VT, Op);
849 SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
850 return VT.bitsGT(Op.getValueType()) ?
851 getNode(ISD::ZERO_EXTEND, DL, VT, Op) :
852 getNode(ISD::TRUNCATE, DL, VT, Op);
855 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) {
856 assert(!VT.isVector() &&
857 "getZeroExtendInReg should use the vector element type instead of "
859 if (Op.getValueType() == VT) return Op;
860 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
861 APInt Imm = APInt::getLowBitsSet(BitWidth,
863 return getNode(ISD::AND, DL, Op.getValueType(), Op,
864 getConstant(Imm, Op.getValueType()));
867 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
869 SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) {
870 EVT EltVT = VT.getScalarType();
872 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
873 return getNode(ISD::XOR, DL, VT, Val, NegOne);
876 SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) {
877 EVT EltVT = VT.getScalarType();
878 assert((EltVT.getSizeInBits() >= 64 ||
879 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
880 "getConstant with a uint64_t value that doesn't fit in the type!");
881 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
884 SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) {
885 return getConstant(*ConstantInt::get(*Context, Val), VT, isT);
888 SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
889 assert(VT.isInteger() && "Cannot create FP integer constant!");
891 EVT EltVT = VT.getScalarType();
892 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
893 "APInt size does not match type size!");
895 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
897 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
901 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
903 return SDValue(N, 0);
906 N = new (NodeAllocator) ConstantSDNode(isT, &Val, EltVT);
907 CSEMap.InsertNode(N, IP);
908 AllNodes.push_back(N);
911 SDValue Result(N, 0);
913 SmallVector<SDValue, 8> Ops;
914 Ops.assign(VT.getVectorNumElements(), Result);
915 Result = getNode(ISD::BUILD_VECTOR, DebugLoc(), VT, &Ops[0], Ops.size());
920 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
921 return getConstant(Val, TLI.getPointerTy(), isTarget);
925 SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) {
926 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget);
929 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){
930 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
932 EVT EltVT = VT.getScalarType();
934 // Do the map lookup using the actual bit pattern for the floating point
935 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
936 // we don't have issues with SNANs.
937 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
939 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
943 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
945 return SDValue(N, 0);
948 N = new (NodeAllocator) ConstantFPSDNode(isTarget, &V, EltVT);
949 CSEMap.InsertNode(N, IP);
950 AllNodes.push_back(N);
953 SDValue Result(N, 0);
955 SmallVector<SDValue, 8> Ops;
956 Ops.assign(VT.getVectorNumElements(), Result);
957 // FIXME DebugLoc info might be appropriate here
958 Result = getNode(ISD::BUILD_VECTOR, DebugLoc(), VT, &Ops[0], Ops.size());
963 SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) {
964 EVT EltVT = VT.getScalarType();
966 return getConstantFP(APFloat((float)Val), VT, isTarget);
967 else if (EltVT==MVT::f64)
968 return getConstantFP(APFloat(Val), VT, isTarget);
969 else if (EltVT==MVT::f80 || EltVT==MVT::f128) {
971 APFloat apf = APFloat(Val);
972 apf.convert(*EVTToAPFloatSemantics(EltVT), APFloat::rmNearestTiesToEven,
974 return getConstantFP(apf, VT, isTarget);
976 assert(0 && "Unsupported type in getConstantFP");
981 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
982 EVT VT, int64_t Offset,
984 unsigned char TargetFlags) {
985 assert((TargetFlags == 0 || isTargetGA) &&
986 "Cannot set target flags on target-independent globals");
988 // Truncate (with sign-extension) the offset value to the pointer size.
989 EVT PTy = TLI.getPointerTy();
990 unsigned BitWidth = PTy.getSizeInBits();
992 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
994 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
996 // If GV is an alias then use the aliasee for determining thread-localness.
997 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
998 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
1002 if (GVar && GVar->isThreadLocal())
1003 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
1005 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
1007 FoldingSetNodeID ID;
1008 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1010 ID.AddInteger(Offset);
1011 ID.AddInteger(TargetFlags);
1013 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1014 return SDValue(E, 0);
1016 SDNode *N = new (NodeAllocator) GlobalAddressSDNode(Opc, GV, VT,
1017 Offset, TargetFlags);
1018 CSEMap.InsertNode(N, IP);
1019 AllNodes.push_back(N);
1020 return SDValue(N, 0);
1023 SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1024 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1025 FoldingSetNodeID ID;
1026 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1029 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1030 return SDValue(E, 0);
1032 SDNode *N = new (NodeAllocator) FrameIndexSDNode(FI, VT, isTarget);
1033 CSEMap.InsertNode(N, IP);
1034 AllNodes.push_back(N);
1035 return SDValue(N, 0);
1038 SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1039 unsigned char TargetFlags) {
1040 assert((TargetFlags == 0 || isTarget) &&
1041 "Cannot set target flags on target-independent jump tables");
1042 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1043 FoldingSetNodeID ID;
1044 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1046 ID.AddInteger(TargetFlags);
1048 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1049 return SDValue(E, 0);
1051 SDNode *N = new (NodeAllocator) JumpTableSDNode(JTI, VT, isTarget,
1053 CSEMap.InsertNode(N, IP);
1054 AllNodes.push_back(N);
1055 return SDValue(N, 0);
1058 SDValue SelectionDAG::getConstantPool(const Constant *C, EVT VT,
1059 unsigned Alignment, int Offset,
1061 unsigned char TargetFlags) {
1062 assert((TargetFlags == 0 || isTarget) &&
1063 "Cannot set target flags on target-independent globals");
1065 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1066 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1067 FoldingSetNodeID ID;
1068 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1069 ID.AddInteger(Alignment);
1070 ID.AddInteger(Offset);
1072 ID.AddInteger(TargetFlags);
1074 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1075 return SDValue(E, 0);
1077 SDNode *N = new (NodeAllocator) ConstantPoolSDNode(isTarget, C, VT, Offset,
1078 Alignment, TargetFlags);
1079 CSEMap.InsertNode(N, IP);
1080 AllNodes.push_back(N);
1081 return SDValue(N, 0);
1085 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1086 unsigned Alignment, int Offset,
1088 unsigned char TargetFlags) {
1089 assert((TargetFlags == 0 || isTarget) &&
1090 "Cannot set target flags on target-independent globals");
1092 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1093 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1094 FoldingSetNodeID ID;
1095 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1096 ID.AddInteger(Alignment);
1097 ID.AddInteger(Offset);
1098 C->AddSelectionDAGCSEId(ID);
1099 ID.AddInteger(TargetFlags);
1101 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1102 return SDValue(E, 0);
1104 SDNode *N = new (NodeAllocator) ConstantPoolSDNode(isTarget, C, VT, Offset,
1105 Alignment, TargetFlags);
1106 CSEMap.InsertNode(N, IP);
1107 AllNodes.push_back(N);
1108 return SDValue(N, 0);
1111 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1112 FoldingSetNodeID ID;
1113 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1116 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1117 return SDValue(E, 0);
1119 SDNode *N = new (NodeAllocator) BasicBlockSDNode(MBB);
1120 CSEMap.InsertNode(N, IP);
1121 AllNodes.push_back(N);
1122 return SDValue(N, 0);
1125 SDValue SelectionDAG::getValueType(EVT VT) {
1126 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
1127 ValueTypeNodes.size())
1128 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1);
1130 SDNode *&N = VT.isExtended() ?
1131 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
1133 if (N) return SDValue(N, 0);
1134 N = new (NodeAllocator) VTSDNode(VT);
1135 AllNodes.push_back(N);
1136 return SDValue(N, 0);
1139 SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
1140 SDNode *&N = ExternalSymbols[Sym];
1141 if (N) return SDValue(N, 0);
1142 N = new (NodeAllocator) ExternalSymbolSDNode(false, Sym, 0, VT);
1143 AllNodes.push_back(N);
1144 return SDValue(N, 0);
1147 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
1148 unsigned char TargetFlags) {
1150 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym,
1152 if (N) return SDValue(N, 0);
1153 N = new (NodeAllocator) ExternalSymbolSDNode(true, Sym, TargetFlags, VT);
1154 AllNodes.push_back(N);
1155 return SDValue(N, 0);
1158 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1159 if ((unsigned)Cond >= CondCodeNodes.size())
1160 CondCodeNodes.resize(Cond+1);
1162 if (CondCodeNodes[Cond] == 0) {
1163 CondCodeSDNode *N = new (NodeAllocator) CondCodeSDNode(Cond);
1164 CondCodeNodes[Cond] = N;
1165 AllNodes.push_back(N);
1168 return SDValue(CondCodeNodes[Cond], 0);
1171 // commuteShuffle - swaps the values of N1 and N2, and swaps all indices in
1172 // the shuffle mask M that point at N1 to point at N2, and indices that point
1173 // N2 to point at N1.
1174 static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) {
1176 int NElts = M.size();
1177 for (int i = 0; i != NElts; ++i) {
1185 SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1,
1186 SDValue N2, const int *Mask) {
1187 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE");
1188 assert(VT.isVector() && N1.getValueType().isVector() &&
1189 "Vector Shuffle VTs must be a vectors");
1190 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType()
1191 && "Vector Shuffle VTs must have same element type");
1193 // Canonicalize shuffle undef, undef -> undef
1194 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF)
1195 return getUNDEF(VT);
1197 // Validate that all indices in Mask are within the range of the elements
1198 // input to the shuffle.
1199 unsigned NElts = VT.getVectorNumElements();
1200 SmallVector<int, 8> MaskVec;
1201 for (unsigned i = 0; i != NElts; ++i) {
1202 assert(Mask[i] < (int)(NElts * 2) && "Index out of range");
1203 MaskVec.push_back(Mask[i]);
1206 // Canonicalize shuffle v, v -> v, undef
1209 for (unsigned i = 0; i != NElts; ++i)
1210 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts;
1213 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1214 if (N1.getOpcode() == ISD::UNDEF)
1215 commuteShuffle(N1, N2, MaskVec);
1217 // Canonicalize all index into lhs, -> shuffle lhs, undef
1218 // Canonicalize all index into rhs, -> shuffle rhs, undef
1219 bool AllLHS = true, AllRHS = true;
1220 bool N2Undef = N2.getOpcode() == ISD::UNDEF;
1221 for (unsigned i = 0; i != NElts; ++i) {
1222 if (MaskVec[i] >= (int)NElts) {
1227 } else if (MaskVec[i] >= 0) {
1231 if (AllLHS && AllRHS)
1232 return getUNDEF(VT);
1233 if (AllLHS && !N2Undef)
1237 commuteShuffle(N1, N2, MaskVec);
1240 // If Identity shuffle, or all shuffle in to undef, return that node.
1241 bool AllUndef = true;
1242 bool Identity = true;
1243 for (unsigned i = 0; i != NElts; ++i) {
1244 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false;
1245 if (MaskVec[i] >= 0) AllUndef = false;
1247 if (Identity && NElts == N1.getValueType().getVectorNumElements())
1250 return getUNDEF(VT);
1252 FoldingSetNodeID ID;
1253 SDValue Ops[2] = { N1, N2 };
1254 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2);
1255 for (unsigned i = 0; i != NElts; ++i)
1256 ID.AddInteger(MaskVec[i]);
1259 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1260 return SDValue(E, 0);
1262 // Allocate the mask array for the node out of the BumpPtrAllocator, since
1263 // SDNode doesn't have access to it. This memory will be "leaked" when
1264 // the node is deallocated, but recovered when the NodeAllocator is released.
1265 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
1266 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int));
1268 ShuffleVectorSDNode *N =
1269 new (NodeAllocator) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc);
1270 CSEMap.InsertNode(N, IP);
1271 AllNodes.push_back(N);
1272 return SDValue(N, 0);
1275 SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl,
1276 SDValue Val, SDValue DTy,
1277 SDValue STy, SDValue Rnd, SDValue Sat,
1278 ISD::CvtCode Code) {
1279 // If the src and dest types are the same and the conversion is between
1280 // integer types of the same sign or two floats, no conversion is necessary.
1282 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF))
1285 FoldingSetNodeID ID;
1286 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat };
1287 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5);
1289 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1290 return SDValue(E, 0);
1292 CvtRndSatSDNode *N = new (NodeAllocator) CvtRndSatSDNode(VT, dl, Ops, 5,
1294 CSEMap.InsertNode(N, IP);
1295 AllNodes.push_back(N);
1296 return SDValue(N, 0);
1299 SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) {
1300 FoldingSetNodeID ID;
1301 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1302 ID.AddInteger(RegNo);
1304 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1305 return SDValue(E, 0);
1307 SDNode *N = new (NodeAllocator) RegisterSDNode(RegNo, VT);
1308 CSEMap.InsertNode(N, IP);
1309 AllNodes.push_back(N);
1310 return SDValue(N, 0);
1313 SDValue SelectionDAG::getEHLabel(DebugLoc dl, SDValue Root, MCSymbol *Label) {
1314 FoldingSetNodeID ID;
1315 SDValue Ops[] = { Root };
1316 AddNodeIDNode(ID, ISD::EH_LABEL, getVTList(MVT::Other), &Ops[0], 1);
1317 ID.AddPointer(Label);
1319 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1320 return SDValue(E, 0);
1322 SDNode *N = new (NodeAllocator) EHLabelSDNode(dl, Root, Label);
1323 CSEMap.InsertNode(N, IP);
1324 AllNodes.push_back(N);
1325 return SDValue(N, 0);
1329 SDValue SelectionDAG::getBlockAddress(const BlockAddress *BA, EVT VT,
1331 unsigned char TargetFlags) {
1332 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress;
1334 FoldingSetNodeID ID;
1335 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1337 ID.AddInteger(TargetFlags);
1339 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1340 return SDValue(E, 0);
1342 SDNode *N = new (NodeAllocator) BlockAddressSDNode(Opc, VT, BA, TargetFlags);
1343 CSEMap.InsertNode(N, IP);
1344 AllNodes.push_back(N);
1345 return SDValue(N, 0);
1348 SDValue SelectionDAG::getSrcValue(const Value *V) {
1349 assert((!V || V->getType()->isPointerTy()) &&
1350 "SrcValue is not a pointer?");
1352 FoldingSetNodeID ID;
1353 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1357 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1358 return SDValue(E, 0);
1360 SDNode *N = new (NodeAllocator) SrcValueSDNode(V);
1361 CSEMap.InsertNode(N, IP);
1362 AllNodes.push_back(N);
1363 return SDValue(N, 0);
1366 /// getMDNode - Return an MDNodeSDNode which holds an MDNode.
1367 SDValue SelectionDAG::getMDNode(const MDNode *MD) {
1368 FoldingSetNodeID ID;
1369 AddNodeIDNode(ID, ISD::MDNODE_SDNODE, getVTList(MVT::Other), 0, 0);
1373 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1374 return SDValue(E, 0);
1376 SDNode *N = new (NodeAllocator) MDNodeSDNode(MD);
1377 CSEMap.InsertNode(N, IP);
1378 AllNodes.push_back(N);
1379 return SDValue(N, 0);
1383 /// getShiftAmountOperand - Return the specified value casted to
1384 /// the target's desired shift amount type.
1385 SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) {
1386 EVT OpTy = Op.getValueType();
1387 MVT ShTy = TLI.getShiftAmountTy();
1388 if (OpTy == ShTy || OpTy.isVector()) return Op;
1390 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
1391 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op);
1394 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1395 /// specified value type.
1396 SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
1397 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1398 unsigned ByteSize = VT.getStoreSize();
1399 const Type *Ty = VT.getTypeForEVT(*getContext());
1400 unsigned StackAlign =
1401 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1403 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
1404 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1407 /// CreateStackTemporary - Create a stack temporary suitable for holding
1408 /// either of the specified value types.
1409 SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
1410 unsigned Bytes = std::max(VT1.getStoreSizeInBits(),
1411 VT2.getStoreSizeInBits())/8;
1412 const Type *Ty1 = VT1.getTypeForEVT(*getContext());
1413 const Type *Ty2 = VT2.getTypeForEVT(*getContext());
1414 const TargetData *TD = TLI.getTargetData();
1415 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1),
1416 TD->getPrefTypeAlignment(Ty2));
1418 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1419 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false);
1420 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1423 SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1,
1424 SDValue N2, ISD::CondCode Cond, DebugLoc dl) {
1425 // These setcc operations always fold.
1429 case ISD::SETFALSE2: return getConstant(0, VT);
1431 case ISD::SETTRUE2: return getConstant(1, VT);
1443 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1447 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1448 const APInt &C2 = N2C->getAPIntValue();
1449 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1450 const APInt &C1 = N1C->getAPIntValue();
1453 default: llvm_unreachable("Unknown integer setcc!");
1454 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1455 case ISD::SETNE: return getConstant(C1 != C2, VT);
1456 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1457 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1458 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1459 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1460 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1461 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1462 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1463 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1467 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1468 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1469 // No compile time operations on this type yet.
1470 if (N1C->getValueType(0) == MVT::ppcf128)
1473 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1476 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1477 return getUNDEF(VT);
1479 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1480 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1481 return getUNDEF(VT);
1483 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1484 R==APFloat::cmpLessThan, VT);
1485 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1486 return getUNDEF(VT);
1488 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1489 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1490 return getUNDEF(VT);
1492 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1493 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1494 return getUNDEF(VT);
1496 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1497 R==APFloat::cmpEqual, VT);
1498 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1499 return getUNDEF(VT);
1501 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1502 R==APFloat::cmpEqual, VT);
1503 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1504 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1505 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1506 R==APFloat::cmpEqual, VT);
1507 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1508 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1509 R==APFloat::cmpLessThan, VT);
1510 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1511 R==APFloat::cmpUnordered, VT);
1512 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1513 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1516 // Ensure that the constant occurs on the RHS.
1517 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1521 // Could not fold it.
1525 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1526 /// use this predicate to simplify operations downstream.
1527 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1528 // This predicate is not safe for vector operations.
1529 if (Op.getValueType().isVector())
1532 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
1533 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1536 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1537 /// this predicate to simplify operations downstream. Mask is known to be zero
1538 /// for bits that V cannot have.
1539 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1540 unsigned Depth) const {
1541 APInt KnownZero, KnownOne;
1542 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1543 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1544 return (KnownZero & Mask) == Mask;
1547 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1548 /// known to be either zero or one and return them in the KnownZero/KnownOne
1549 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1551 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1552 APInt &KnownZero, APInt &KnownOne,
1553 unsigned Depth) const {
1554 unsigned BitWidth = Mask.getBitWidth();
1555 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() &&
1556 "Mask size mismatches value type size!");
1558 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1559 if (Depth == 6 || Mask == 0)
1560 return; // Limit search depth.
1562 APInt KnownZero2, KnownOne2;
1564 switch (Op.getOpcode()) {
1566 // We know all of the bits for a constant!
1567 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1568 KnownZero = ~KnownOne & Mask;
1571 // If either the LHS or the RHS are Zero, the result is zero.
1572 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1573 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1574 KnownZero2, KnownOne2, Depth+1);
1575 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1576 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1578 // Output known-1 bits are only known if set in both the LHS & RHS.
1579 KnownOne &= KnownOne2;
1580 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1581 KnownZero |= KnownZero2;
1584 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1585 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1586 KnownZero2, KnownOne2, Depth+1);
1587 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1588 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1590 // Output known-0 bits are only known if clear in both the LHS & RHS.
1591 KnownZero &= KnownZero2;
1592 // Output known-1 are known to be set if set in either the LHS | RHS.
1593 KnownOne |= KnownOne2;
1596 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1597 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1598 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1599 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1601 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1602 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1603 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1604 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1605 KnownZero = KnownZeroOut;
1609 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1610 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1611 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1612 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1613 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1615 // If low bits are zero in either operand, output low known-0 bits.
1616 // Also compute a conserative estimate for high known-0 bits.
1617 // More trickiness is possible, but this is sufficient for the
1618 // interesting case of alignment computation.
1620 unsigned TrailZ = KnownZero.countTrailingOnes() +
1621 KnownZero2.countTrailingOnes();
1622 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1623 KnownZero2.countLeadingOnes(),
1624 BitWidth) - BitWidth;
1626 TrailZ = std::min(TrailZ, BitWidth);
1627 LeadZ = std::min(LeadZ, BitWidth);
1628 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1629 APInt::getHighBitsSet(BitWidth, LeadZ);
1634 // For the purposes of computing leading zeros we can conservatively
1635 // treat a udiv as a logical right shift by the power of 2 known to
1636 // be less than the denominator.
1637 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1638 ComputeMaskedBits(Op.getOperand(0),
1639 AllOnes, KnownZero2, KnownOne2, Depth+1);
1640 unsigned LeadZ = KnownZero2.countLeadingOnes();
1644 ComputeMaskedBits(Op.getOperand(1),
1645 AllOnes, KnownZero2, KnownOne2, Depth+1);
1646 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1647 if (RHSUnknownLeadingOnes != BitWidth)
1648 LeadZ = std::min(BitWidth,
1649 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1651 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1655 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1656 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1657 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1658 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1660 // Only known if known in both the LHS and RHS.
1661 KnownOne &= KnownOne2;
1662 KnownZero &= KnownZero2;
1664 case ISD::SELECT_CC:
1665 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1666 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1667 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1668 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1670 // Only known if known in both the LHS and RHS.
1671 KnownOne &= KnownOne2;
1672 KnownZero &= KnownZero2;
1680 if (Op.getResNo() != 1)
1682 // The boolean result conforms to getBooleanContents. Fall through.
1684 // If we know the result of a setcc has the top bits zero, use this info.
1685 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent &&
1687 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1690 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1691 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1692 unsigned ShAmt = SA->getZExtValue();
1694 // If the shift count is an invalid immediate, don't do anything.
1695 if (ShAmt >= BitWidth)
1698 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt),
1699 KnownZero, KnownOne, Depth+1);
1700 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1701 KnownZero <<= ShAmt;
1703 // low bits known zero.
1704 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1708 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1709 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1710 unsigned ShAmt = SA->getZExtValue();
1712 // If the shift count is an invalid immediate, don't do anything.
1713 if (ShAmt >= BitWidth)
1716 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1717 KnownZero, KnownOne, Depth+1);
1718 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1719 KnownZero = KnownZero.lshr(ShAmt);
1720 KnownOne = KnownOne.lshr(ShAmt);
1722 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1723 KnownZero |= HighBits; // High bits known zero.
1727 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1728 unsigned ShAmt = SA->getZExtValue();
1730 // If the shift count is an invalid immediate, don't do anything.
1731 if (ShAmt >= BitWidth)
1734 APInt InDemandedMask = (Mask << ShAmt);
1735 // If any of the demanded bits are produced by the sign extension, we also
1736 // demand the input sign bit.
1737 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1738 if (HighBits.getBoolValue())
1739 InDemandedMask |= APInt::getSignBit(BitWidth);
1741 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1743 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1744 KnownZero = KnownZero.lshr(ShAmt);
1745 KnownOne = KnownOne.lshr(ShAmt);
1747 // Handle the sign bits.
1748 APInt SignBit = APInt::getSignBit(BitWidth);
1749 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1751 if (KnownZero.intersects(SignBit)) {
1752 KnownZero |= HighBits; // New bits are known zero.
1753 } else if (KnownOne.intersects(SignBit)) {
1754 KnownOne |= HighBits; // New bits are known one.
1758 case ISD::SIGN_EXTEND_INREG: {
1759 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1760 unsigned EBits = EVT.getScalarType().getSizeInBits();
1762 // Sign extension. Compute the demanded bits in the result that are not
1763 // present in the input.
1764 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1766 APInt InSignBit = APInt::getSignBit(EBits);
1767 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1769 // If the sign extended bits are demanded, we know that the sign
1771 InSignBit.zext(BitWidth);
1772 if (NewBits.getBoolValue())
1773 InputDemandedBits |= InSignBit;
1775 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1776 KnownZero, KnownOne, Depth+1);
1777 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1779 // If the sign bit of the input is known set or clear, then we know the
1780 // top bits of the result.
1781 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1782 KnownZero |= NewBits;
1783 KnownOne &= ~NewBits;
1784 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1785 KnownOne |= NewBits;
1786 KnownZero &= ~NewBits;
1787 } else { // Input sign bit unknown
1788 KnownZero &= ~NewBits;
1789 KnownOne &= ~NewBits;
1796 unsigned LowBits = Log2_32(BitWidth)+1;
1797 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1802 if (ISD::isZEXTLoad(Op.getNode())) {
1803 LoadSDNode *LD = cast<LoadSDNode>(Op);
1804 EVT VT = LD->getMemoryVT();
1805 unsigned MemBits = VT.getScalarType().getSizeInBits();
1806 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1810 case ISD::ZERO_EXTEND: {
1811 EVT InVT = Op.getOperand(0).getValueType();
1812 unsigned InBits = InVT.getScalarType().getSizeInBits();
1813 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1814 APInt InMask = Mask;
1815 InMask.trunc(InBits);
1816 KnownZero.trunc(InBits);
1817 KnownOne.trunc(InBits);
1818 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1819 KnownZero.zext(BitWidth);
1820 KnownOne.zext(BitWidth);
1821 KnownZero |= NewBits;
1824 case ISD::SIGN_EXTEND: {
1825 EVT InVT = Op.getOperand(0).getValueType();
1826 unsigned InBits = InVT.getScalarType().getSizeInBits();
1827 APInt InSignBit = APInt::getSignBit(InBits);
1828 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1829 APInt InMask = Mask;
1830 InMask.trunc(InBits);
1832 // If any of the sign extended bits are demanded, we know that the sign
1833 // bit is demanded. Temporarily set this bit in the mask for our callee.
1834 if (NewBits.getBoolValue())
1835 InMask |= InSignBit;
1837 KnownZero.trunc(InBits);
1838 KnownOne.trunc(InBits);
1839 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1841 // Note if the sign bit is known to be zero or one.
1842 bool SignBitKnownZero = KnownZero.isNegative();
1843 bool SignBitKnownOne = KnownOne.isNegative();
1844 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1845 "Sign bit can't be known to be both zero and one!");
1847 // If the sign bit wasn't actually demanded by our caller, we don't
1848 // want it set in the KnownZero and KnownOne result values. Reset the
1849 // mask and reapply it to the result values.
1851 InMask.trunc(InBits);
1852 KnownZero &= InMask;
1855 KnownZero.zext(BitWidth);
1856 KnownOne.zext(BitWidth);
1858 // If the sign bit is known zero or one, the top bits match.
1859 if (SignBitKnownZero)
1860 KnownZero |= NewBits;
1861 else if (SignBitKnownOne)
1862 KnownOne |= NewBits;
1865 case ISD::ANY_EXTEND: {
1866 EVT InVT = Op.getOperand(0).getValueType();
1867 unsigned InBits = InVT.getScalarType().getSizeInBits();
1868 APInt InMask = Mask;
1869 InMask.trunc(InBits);
1870 KnownZero.trunc(InBits);
1871 KnownOne.trunc(InBits);
1872 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1873 KnownZero.zext(BitWidth);
1874 KnownOne.zext(BitWidth);
1877 case ISD::TRUNCATE: {
1878 EVT InVT = Op.getOperand(0).getValueType();
1879 unsigned InBits = InVT.getScalarType().getSizeInBits();
1880 APInt InMask = Mask;
1881 InMask.zext(InBits);
1882 KnownZero.zext(InBits);
1883 KnownOne.zext(InBits);
1884 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1885 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1886 KnownZero.trunc(BitWidth);
1887 KnownOne.trunc(BitWidth);
1890 case ISD::AssertZext: {
1891 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1892 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1893 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1895 KnownZero |= (~InMask) & Mask;
1899 // All bits are zero except the low bit.
1900 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1904 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1905 // We know that the top bits of C-X are clear if X contains less bits
1906 // than C (i.e. no wrap-around can happen). For example, 20-X is
1907 // positive if we can prove that X is >= 0 and < 16.
1908 if (CLHS->getAPIntValue().isNonNegative()) {
1909 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1910 // NLZ can't be BitWidth with no sign bit
1911 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1912 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1915 // If all of the MaskV bits are known to be zero, then we know the
1916 // output top bits are zero, because we now know that the output is
1918 if ((KnownZero2 & MaskV) == MaskV) {
1919 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1920 // Top bits known zero.
1921 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1928 // Output known-0 bits are known if clear or set in both the low clear bits
1929 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1930 // low 3 bits clear.
1931 APInt Mask2 = APInt::getLowBitsSet(BitWidth,
1932 BitWidth - Mask.countLeadingZeros());
1933 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1934 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1935 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1937 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1938 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1939 KnownZeroOut = std::min(KnownZeroOut,
1940 KnownZero2.countTrailingOnes());
1942 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1946 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1947 const APInt &RA = Rem->getAPIntValue().abs();
1948 if (RA.isPowerOf2()) {
1949 APInt LowBits = RA - 1;
1950 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1951 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1953 // The low bits of the first operand are unchanged by the srem.
1954 KnownZero = KnownZero2 & LowBits;
1955 KnownOne = KnownOne2 & LowBits;
1957 // If the first operand is non-negative or has all low bits zero, then
1958 // the upper bits are all zero.
1959 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1960 KnownZero |= ~LowBits;
1962 // If the first operand is negative and not all low bits are zero, then
1963 // the upper bits are all one.
1964 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1965 KnownOne |= ~LowBits;
1970 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1975 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1976 const APInt &RA = Rem->getAPIntValue();
1977 if (RA.isPowerOf2()) {
1978 APInt LowBits = (RA - 1);
1979 APInt Mask2 = LowBits & Mask;
1980 KnownZero |= ~LowBits & Mask;
1981 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1982 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1987 // Since the result is less than or equal to either operand, any leading
1988 // zero bits in either operand must also exist in the result.
1989 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1990 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne,
1992 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2,
1995 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1996 KnownZero2.countLeadingOnes());
1998 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
2002 // Allow the target to implement this method for its nodes.
2003 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
2004 case ISD::INTRINSIC_WO_CHAIN:
2005 case ISD::INTRINSIC_W_CHAIN:
2006 case ISD::INTRINSIC_VOID:
2007 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this,
2014 /// ComputeNumSignBits - Return the number of times the sign bit of the
2015 /// register is replicated into the other bits. We know that at least 1 bit
2016 /// is always equal to the sign bit (itself), but other cases can give us
2017 /// information. For example, immediately after an "SRA X, 2", we know that
2018 /// the top 3 bits are all equal to each other, so we return 3.
2019 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
2020 EVT VT = Op.getValueType();
2021 assert(VT.isInteger() && "Invalid VT!");
2022 unsigned VTBits = VT.getScalarType().getSizeInBits();
2024 unsigned FirstAnswer = 1;
2027 return 1; // Limit search depth.
2029 switch (Op.getOpcode()) {
2031 case ISD::AssertSext:
2032 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2033 return VTBits-Tmp+1;
2034 case ISD::AssertZext:
2035 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2038 case ISD::Constant: {
2039 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
2040 // If negative, return # leading ones.
2041 if (Val.isNegative())
2042 return Val.countLeadingOnes();
2044 // Return # leading zeros.
2045 return Val.countLeadingZeros();
2048 case ISD::SIGN_EXTEND:
2049 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
2050 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
2052 case ISD::SIGN_EXTEND_INREG:
2053 // Max of the input and what this extends.
2055 cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarType().getSizeInBits();
2058 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2059 return std::max(Tmp, Tmp2);
2062 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2063 // SRA X, C -> adds C sign bits.
2064 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2065 Tmp += C->getZExtValue();
2066 if (Tmp > VTBits) Tmp = VTBits;
2070 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2071 // shl destroys sign bits.
2072 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2073 if (C->getZExtValue() >= VTBits || // Bad shift.
2074 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2075 return Tmp - C->getZExtValue();
2080 case ISD::XOR: // NOT is handled here.
2081 // Logical binary ops preserve the number of sign bits at the worst.
2082 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2084 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2085 FirstAnswer = std::min(Tmp, Tmp2);
2086 // We computed what we know about the sign bits as our first
2087 // answer. Now proceed to the generic code that uses
2088 // ComputeMaskedBits, and pick whichever answer is better.
2093 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2094 if (Tmp == 1) return 1; // Early out.
2095 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
2096 return std::min(Tmp, Tmp2);
2104 if (Op.getResNo() != 1)
2106 // The boolean result conforms to getBooleanContents. Fall through.
2108 // If setcc returns 0/-1, all bits are sign bits.
2109 if (TLI.getBooleanContents() ==
2110 TargetLowering::ZeroOrNegativeOneBooleanContent)
2115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2116 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
2118 // Handle rotate right by N like a rotate left by 32-N.
2119 if (Op.getOpcode() == ISD::ROTR)
2120 RotAmt = (VTBits-RotAmt) & (VTBits-1);
2122 // If we aren't rotating out all of the known-in sign bits, return the
2123 // number that are left. This handles rotl(sext(x), 1) for example.
2124 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2125 if (Tmp > RotAmt+1) return Tmp-RotAmt;
2129 // Add can have at most one carry bit. Thus we know that the output
2130 // is, at worst, one more bit than the inputs.
2131 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2132 if (Tmp == 1) return 1; // Early out.
2134 // Special case decrementing a value (ADD X, -1):
2135 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
2136 if (CRHS->isAllOnesValue()) {
2137 APInt KnownZero, KnownOne;
2138 APInt Mask = APInt::getAllOnesValue(VTBits);
2139 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2141 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2143 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2146 // If we are subtracting one from a positive number, there is no carry
2147 // out of the result.
2148 if (KnownZero.isNegative())
2152 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2153 if (Tmp2 == 1) return 1;
2154 return std::min(Tmp, Tmp2)-1;
2158 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2159 if (Tmp2 == 1) return 1;
2162 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
2163 if (CLHS->isNullValue()) {
2164 APInt KnownZero, KnownOne;
2165 APInt Mask = APInt::getAllOnesValue(VTBits);
2166 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2167 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2169 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2172 // If the input is known to be positive (the sign bit is known clear),
2173 // the output of the NEG has the same number of sign bits as the input.
2174 if (KnownZero.isNegative())
2177 // Otherwise, we treat this like a SUB.
2180 // Sub can have at most one carry bit. Thus we know that the output
2181 // is, at worst, one more bit than the inputs.
2182 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2183 if (Tmp == 1) return 1; // Early out.
2184 return std::min(Tmp, Tmp2)-1;
2187 // FIXME: it's tricky to do anything useful for this, but it is an important
2188 // case for targets like X86.
2192 // Handle LOADX separately here. EXTLOAD case will fallthrough.
2193 if (Op.getOpcode() == ISD::LOAD) {
2194 LoadSDNode *LD = cast<LoadSDNode>(Op);
2195 unsigned ExtType = LD->getExtensionType();
2198 case ISD::SEXTLOAD: // '17' bits known
2199 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2200 return VTBits-Tmp+1;
2201 case ISD::ZEXTLOAD: // '16' bits known
2202 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2207 // Allow the target to implement this method for its nodes.
2208 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
2209 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
2210 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
2211 Op.getOpcode() == ISD::INTRINSIC_VOID) {
2212 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
2213 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
2216 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2217 // use this information.
2218 APInt KnownZero, KnownOne;
2219 APInt Mask = APInt::getAllOnesValue(VTBits);
2220 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
2222 if (KnownZero.isNegative()) { // sign bit is 0
2224 } else if (KnownOne.isNegative()) { // sign bit is 1;
2231 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2232 // the number of identical bits in the top of the input value.
2234 Mask <<= Mask.getBitWidth()-VTBits;
2235 // Return # leading zeros. We use 'min' here in case Val was zero before
2236 // shifting. We don't want to return '64' as for an i32 "0".
2237 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2240 bool SelectionDAG::isKnownNeverNaN(SDValue Op) const {
2241 // If we're told that NaNs won't happen, assume they won't.
2242 if (FiniteOnlyFPMath())
2245 // If the value is a constant, we can obviously see if it is a NaN or not.
2246 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2247 return !C->getValueAPF().isNaN();
2249 // TODO: Recognize more cases here.
2254 bool SelectionDAG::isKnownNeverZero(SDValue Op) const {
2255 // If the value is a constant, we can obviously see if it is a zero or not.
2256 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2257 return !C->isZero();
2259 // TODO: Recognize more cases here.
2264 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
2265 // Check the obvious case.
2266 if (A == B) return true;
2268 // For for negative and positive zero.
2269 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A))
2270 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B))
2271 if (CA->isZero() && CB->isZero()) return true;
2273 // Otherwise they may not be equal.
2277 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2278 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2279 if (!GA) return false;
2280 if (GA->getOffset() != 0) return false;
2281 const GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2282 if (!GV) return false;
2283 return MF->getMMI().hasDebugInfo();
2287 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2288 /// element of the result of the vector shuffle.
2289 SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N,
2291 EVT VT = N->getValueType(0);
2292 DebugLoc dl = N->getDebugLoc();
2293 if (N->getMaskElt(i) < 0)
2294 return getUNDEF(VT.getVectorElementType());
2295 unsigned Index = N->getMaskElt(i);
2296 unsigned NumElems = VT.getVectorNumElements();
2297 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2300 if (V.getOpcode() == ISD::BIT_CONVERT) {
2301 V = V.getOperand(0);
2302 EVT VVT = V.getValueType();
2303 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems)
2306 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2307 return (Index == 0) ? V.getOperand(0)
2308 : getUNDEF(VT.getVectorElementType());
2309 if (V.getOpcode() == ISD::BUILD_VECTOR)
2310 return V.getOperand(Index);
2311 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V))
2312 return getShuffleScalarElt(SVN, Index);
2317 /// getNode - Gets or creates the specified node.
2319 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) {
2320 FoldingSetNodeID ID;
2321 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2323 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2324 return SDValue(E, 0);
2326 SDNode *N = new (NodeAllocator) SDNode(Opcode, DL, getVTList(VT));
2327 CSEMap.InsertNode(N, IP);
2329 AllNodes.push_back(N);
2333 return SDValue(N, 0);
2336 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
2337 EVT VT, SDValue Operand) {
2338 // Constant fold unary operations with an integer constant operand.
2339 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2340 const APInt &Val = C->getAPIntValue();
2343 case ISD::SIGN_EXTEND:
2344 return getConstant(APInt(Val).sextOrTrunc(VT.getSizeInBits()), VT);
2345 case ISD::ANY_EXTEND:
2346 case ISD::ZERO_EXTEND:
2348 return getConstant(APInt(Val).zextOrTrunc(VT.getSizeInBits()), VT);
2349 case ISD::UINT_TO_FP:
2350 case ISD::SINT_TO_FP: {
2351 const uint64_t zero[] = {0, 0};
2352 // No compile time operations on ppcf128.
2353 if (VT == MVT::ppcf128) break;
2354 APFloat apf = APFloat(APInt(VT.getSizeInBits(), 2, zero));
2355 (void)apf.convertFromAPInt(Val,
2356 Opcode==ISD::SINT_TO_FP,
2357 APFloat::rmNearestTiesToEven);
2358 return getConstantFP(apf, VT);
2360 case ISD::BIT_CONVERT:
2361 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2362 return getConstantFP(Val.bitsToFloat(), VT);
2363 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2364 return getConstantFP(Val.bitsToDouble(), VT);
2367 return getConstant(Val.byteSwap(), VT);
2369 return getConstant(Val.countPopulation(), VT);
2371 return getConstant(Val.countLeadingZeros(), VT);
2373 return getConstant(Val.countTrailingZeros(), VT);
2377 // Constant fold unary operations with a floating point constant operand.
2378 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2379 APFloat V = C->getValueAPF(); // make copy
2380 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2384 return getConstantFP(V, VT);
2387 return getConstantFP(V, VT);
2389 case ISD::FP_EXTEND: {
2391 // This can return overflow, underflow, or inexact; we don't care.
2392 // FIXME need to be more flexible about rounding mode.
2393 (void)V.convert(*EVTToAPFloatSemantics(VT),
2394 APFloat::rmNearestTiesToEven, &ignored);
2395 return getConstantFP(V, VT);
2397 case ISD::FP_TO_SINT:
2398 case ISD::FP_TO_UINT: {
2401 assert(integerPartWidth >= 64);
2402 // FIXME need to be more flexible about rounding mode.
2403 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
2404 Opcode==ISD::FP_TO_SINT,
2405 APFloat::rmTowardZero, &ignored);
2406 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2408 APInt api(VT.getSizeInBits(), 2, x);
2409 return getConstant(api, VT);
2411 case ISD::BIT_CONVERT:
2412 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2413 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2414 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2415 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2421 unsigned OpOpcode = Operand.getNode()->getOpcode();
2423 case ISD::TokenFactor:
2424 case ISD::MERGE_VALUES:
2425 case ISD::CONCAT_VECTORS:
2426 return Operand; // Factor, merge or concat of one node? No need.
2427 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
2428 case ISD::FP_EXTEND:
2429 assert(VT.isFloatingPoint() &&
2430 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2431 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2432 assert((!VT.isVector() ||
2433 VT.getVectorNumElements() ==
2434 Operand.getValueType().getVectorNumElements()) &&
2435 "Vector element count mismatch!");
2436 if (Operand.getOpcode() == ISD::UNDEF)
2437 return getUNDEF(VT);
2439 case ISD::SIGN_EXTEND:
2440 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2441 "Invalid SIGN_EXTEND!");
2442 if (Operand.getValueType() == VT) return Operand; // noop extension
2443 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2444 "Invalid sext node, dst < src!");
2445 assert((!VT.isVector() ||
2446 VT.getVectorNumElements() ==
2447 Operand.getValueType().getVectorNumElements()) &&
2448 "Vector element count mismatch!");
2449 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2450 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2452 case ISD::ZERO_EXTEND:
2453 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2454 "Invalid ZERO_EXTEND!");
2455 if (Operand.getValueType() == VT) return Operand; // noop extension
2456 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2457 "Invalid zext node, dst < src!");
2458 assert((!VT.isVector() ||
2459 VT.getVectorNumElements() ==
2460 Operand.getValueType().getVectorNumElements()) &&
2461 "Vector element count mismatch!");
2462 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2463 return getNode(ISD::ZERO_EXTEND, DL, VT,
2464 Operand.getNode()->getOperand(0));
2466 case ISD::ANY_EXTEND:
2467 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2468 "Invalid ANY_EXTEND!");
2469 if (Operand.getValueType() == VT) return Operand; // noop extension
2470 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2471 "Invalid anyext node, dst < src!");
2472 assert((!VT.isVector() ||
2473 VT.getVectorNumElements() ==
2474 Operand.getValueType().getVectorNumElements()) &&
2475 "Vector element count mismatch!");
2476 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND)
2477 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2478 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2481 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2482 "Invalid TRUNCATE!");
2483 if (Operand.getValueType() == VT) return Operand; // noop truncate
2484 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) &&
2485 "Invalid truncate node, src < dst!");
2486 assert((!VT.isVector() ||
2487 VT.getVectorNumElements() ==
2488 Operand.getValueType().getVectorNumElements()) &&
2489 "Vector element count mismatch!");
2490 if (OpOpcode == ISD::TRUNCATE)
2491 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2492 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2493 OpOpcode == ISD::ANY_EXTEND) {
2494 // If the source is smaller than the dest, we still need an extend.
2495 if (Operand.getNode()->getOperand(0).getValueType().getScalarType()
2496 .bitsLT(VT.getScalarType()))
2497 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2498 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2499 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2501 return Operand.getNode()->getOperand(0);
2504 case ISD::BIT_CONVERT:
2505 // Basic sanity checking.
2506 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2507 && "Cannot BIT_CONVERT between types of different sizes!");
2508 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2509 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2510 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0));
2511 if (OpOpcode == ISD::UNDEF)
2512 return getUNDEF(VT);
2514 case ISD::SCALAR_TO_VECTOR:
2515 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2516 (VT.getVectorElementType() == Operand.getValueType() ||
2517 (VT.getVectorElementType().isInteger() &&
2518 Operand.getValueType().isInteger() &&
2519 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
2520 "Illegal SCALAR_TO_VECTOR node!");
2521 if (OpOpcode == ISD::UNDEF)
2522 return getUNDEF(VT);
2523 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2524 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2525 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2526 Operand.getConstantOperandVal(1) == 0 &&
2527 Operand.getOperand(0).getValueType() == VT)
2528 return Operand.getOperand(0);
2531 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0
2532 if (UnsafeFPMath && OpOpcode == ISD::FSUB)
2533 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1),
2534 Operand.getNode()->getOperand(0));
2535 if (OpOpcode == ISD::FNEG) // --X -> X
2536 return Operand.getNode()->getOperand(0);
2539 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2540 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0));
2545 SDVTList VTs = getVTList(VT);
2546 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2547 FoldingSetNodeID ID;
2548 SDValue Ops[1] = { Operand };
2549 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2551 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2552 return SDValue(E, 0);
2554 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTs, Operand);
2555 CSEMap.InsertNode(N, IP);
2557 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTs, Operand);
2560 AllNodes.push_back(N);
2564 return SDValue(N, 0);
2567 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2569 ConstantSDNode *Cst1,
2570 ConstantSDNode *Cst2) {
2571 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2574 case ISD::ADD: return getConstant(C1 + C2, VT);
2575 case ISD::SUB: return getConstant(C1 - C2, VT);
2576 case ISD::MUL: return getConstant(C1 * C2, VT);
2578 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2581 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2584 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2587 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2589 case ISD::AND: return getConstant(C1 & C2, VT);
2590 case ISD::OR: return getConstant(C1 | C2, VT);
2591 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2592 case ISD::SHL: return getConstant(C1 << C2, VT);
2593 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2594 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2595 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2596 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2603 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2604 SDValue N1, SDValue N2) {
2605 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2606 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2609 case ISD::TokenFactor:
2610 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2611 N2.getValueType() == MVT::Other && "Invalid token factor!");
2612 // Fold trivial token factors.
2613 if (N1.getOpcode() == ISD::EntryToken) return N2;
2614 if (N2.getOpcode() == ISD::EntryToken) return N1;
2615 if (N1 == N2) return N1;
2617 case ISD::CONCAT_VECTORS:
2618 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2619 // one big BUILD_VECTOR.
2620 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2621 N2.getOpcode() == ISD::BUILD_VECTOR) {
2622 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2623 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2624 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2628 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2629 N1.getValueType() == VT && "Binary operator types must match!");
2630 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2631 // worth handling here.
2632 if (N2C && N2C->isNullValue())
2634 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2641 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2642 N1.getValueType() == VT && "Binary operator types must match!");
2643 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2644 // it's worth handling here.
2645 if (N2C && N2C->isNullValue())
2655 assert(VT.isInteger() && "This operator does not apply to FP types!");
2663 if (Opcode == ISD::FADD) {
2665 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1))
2666 if (CFP->getValueAPF().isZero())
2669 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2670 if (CFP->getValueAPF().isZero())
2672 } else if (Opcode == ISD::FSUB) {
2674 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2675 if (CFP->getValueAPF().isZero())
2679 assert(N1.getValueType() == N2.getValueType() &&
2680 N1.getValueType() == VT && "Binary operator types must match!");
2682 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2683 assert(N1.getValueType() == VT &&
2684 N1.getValueType().isFloatingPoint() &&
2685 N2.getValueType().isFloatingPoint() &&
2686 "Invalid FCOPYSIGN!");
2693 assert(VT == N1.getValueType() &&
2694 "Shift operators return type must be the same as their first arg");
2695 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2696 "Shifts only work on integers");
2698 // Always fold shifts of i1 values so the code generator doesn't need to
2699 // handle them. Since we know the size of the shift has to be less than the
2700 // size of the value, the shift/rotate count is guaranteed to be zero.
2703 if (N2C && N2C->isNullValue())
2706 case ISD::FP_ROUND_INREG: {
2707 EVT EVT = cast<VTSDNode>(N2)->getVT();
2708 assert(VT == N1.getValueType() && "Not an inreg round!");
2709 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2710 "Cannot FP_ROUND_INREG integer types");
2711 assert(EVT.isVector() == VT.isVector() &&
2712 "FP_ROUND_INREG type should be vector iff the operand "
2714 assert((!EVT.isVector() ||
2715 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2716 "Vector element counts must match in FP_ROUND_INREG");
2717 assert(EVT.bitsLE(VT) && "Not rounding down!");
2718 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2722 assert(VT.isFloatingPoint() &&
2723 N1.getValueType().isFloatingPoint() &&
2724 VT.bitsLE(N1.getValueType()) &&
2725 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2726 if (N1.getValueType() == VT) return N1; // noop conversion.
2728 case ISD::AssertSext:
2729 case ISD::AssertZext: {
2730 EVT EVT = cast<VTSDNode>(N2)->getVT();
2731 assert(VT == N1.getValueType() && "Not an inreg extend!");
2732 assert(VT.isInteger() && EVT.isInteger() &&
2733 "Cannot *_EXTEND_INREG FP types");
2734 assert(!EVT.isVector() &&
2735 "AssertSExt/AssertZExt type should be the vector element type "
2736 "rather than the vector type!");
2737 assert(EVT.bitsLE(VT) && "Not extending!");
2738 if (VT == EVT) return N1; // noop assertion.
2741 case ISD::SIGN_EXTEND_INREG: {
2742 EVT EVT = cast<VTSDNode>(N2)->getVT();
2743 assert(VT == N1.getValueType() && "Not an inreg extend!");
2744 assert(VT.isInteger() && EVT.isInteger() &&
2745 "Cannot *_EXTEND_INREG FP types");
2746 assert(EVT.isVector() == VT.isVector() &&
2747 "SIGN_EXTEND_INREG type should be vector iff the operand "
2749 assert((!EVT.isVector() ||
2750 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2751 "Vector element counts must match in SIGN_EXTEND_INREG");
2752 assert(EVT.bitsLE(VT) && "Not extending!");
2753 if (EVT == VT) return N1; // Not actually extending
2756 APInt Val = N1C->getAPIntValue();
2757 unsigned FromBits = EVT.getScalarType().getSizeInBits();
2758 Val <<= Val.getBitWidth()-FromBits;
2759 Val = Val.ashr(Val.getBitWidth()-FromBits);
2760 return getConstant(Val, VT);
2764 case ISD::EXTRACT_VECTOR_ELT:
2765 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2766 if (N1.getOpcode() == ISD::UNDEF)
2767 return getUNDEF(VT);
2769 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2770 // expanding copies of large vectors from registers.
2772 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2773 N1.getNumOperands() > 0) {
2775 N1.getOperand(0).getValueType().getVectorNumElements();
2776 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
2777 N1.getOperand(N2C->getZExtValue() / Factor),
2778 getConstant(N2C->getZExtValue() % Factor,
2779 N2.getValueType()));
2782 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2783 // expanding large vector constants.
2784 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) {
2785 SDValue Elt = N1.getOperand(N2C->getZExtValue());
2786 EVT VEltTy = N1.getValueType().getVectorElementType();
2787 if (Elt.getValueType() != VEltTy) {
2788 // If the vector element type is not legal, the BUILD_VECTOR operands
2789 // are promoted and implicitly truncated. Make that explicit here.
2790 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt);
2793 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT
2794 // result is implicitly extended.
2795 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt);
2800 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2801 // operations are lowered to scalars.
2802 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2803 // If the indices are the same, return the inserted element else
2804 // if the indices are known different, extract the element from
2805 // the original vector.
2806 SDValue N1Op2 = N1.getOperand(2);
2807 ConstantSDNode *N1Op2C = dyn_cast<ConstantSDNode>(N1Op2.getNode());
2809 if (N1Op2C && N2C) {
2810 if (N1Op2C->getZExtValue() == N2C->getZExtValue()) {
2811 if (VT == N1.getOperand(1).getValueType())
2812 return N1.getOperand(1);
2814 return getSExtOrTrunc(N1.getOperand(1), DL, VT);
2817 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
2821 case ISD::EXTRACT_ELEMENT:
2822 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2823 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2824 (N1.getValueType().isInteger() == VT.isInteger()) &&
2825 "Wrong types for EXTRACT_ELEMENT!");
2827 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2828 // 64-bit integers into 32-bit parts. Instead of building the extract of
2829 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2830 if (N1.getOpcode() == ISD::BUILD_PAIR)
2831 return N1.getOperand(N2C->getZExtValue());
2833 // EXTRACT_ELEMENT of a constant int is also very common.
2834 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2835 unsigned ElementSize = VT.getSizeInBits();
2836 unsigned Shift = ElementSize * N2C->getZExtValue();
2837 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2838 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2841 case ISD::EXTRACT_SUBVECTOR:
2842 if (N1.getValueType() == VT) // Trivial extraction.
2849 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2850 if (SV.getNode()) return SV;
2851 } else { // Cannonicalize constant to RHS if commutative
2852 if (isCommutativeBinOp(Opcode)) {
2853 std::swap(N1C, N2C);
2859 // Constant fold FP operations.
2860 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2861 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2863 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2864 // Cannonicalize constant to RHS if commutative
2865 std::swap(N1CFP, N2CFP);
2867 } else if (N2CFP && VT != MVT::ppcf128) {
2868 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2869 APFloat::opStatus s;
2872 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2873 if (s != APFloat::opInvalidOp)
2874 return getConstantFP(V1, VT);
2877 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2878 if (s!=APFloat::opInvalidOp)
2879 return getConstantFP(V1, VT);
2882 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2883 if (s!=APFloat::opInvalidOp)
2884 return getConstantFP(V1, VT);
2887 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2888 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2889 return getConstantFP(V1, VT);
2892 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2893 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2894 return getConstantFP(V1, VT);
2896 case ISD::FCOPYSIGN:
2898 return getConstantFP(V1, VT);
2904 // Canonicalize an UNDEF to the RHS, even over a constant.
2905 if (N1.getOpcode() == ISD::UNDEF) {
2906 if (isCommutativeBinOp(Opcode)) {
2910 case ISD::FP_ROUND_INREG:
2911 case ISD::SIGN_EXTEND_INREG:
2917 return N1; // fold op(undef, arg2) -> undef
2925 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2926 // For vectors, we can't easily build an all zero vector, just return
2933 // Fold a bunch of operators when the RHS is undef.
2934 if (N2.getOpcode() == ISD::UNDEF) {
2937 if (N1.getOpcode() == ISD::UNDEF)
2938 // Handle undef ^ undef -> 0 special case. This is a common
2940 return getConstant(0, VT);
2950 return N2; // fold op(arg1, undef) -> undef
2964 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2965 // For vectors, we can't easily build an all zero vector, just return
2970 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
2971 // For vectors, we can't easily build an all one vector, just return
2979 // Memoize this node if possible.
2981 SDVTList VTs = getVTList(VT);
2982 if (VT != MVT::Flag) {
2983 SDValue Ops[] = { N1, N2 };
2984 FoldingSetNodeID ID;
2985 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2987 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2988 return SDValue(E, 0);
2990 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTs, N1, N2);
2991 CSEMap.InsertNode(N, IP);
2993 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTs, N1, N2);
2996 AllNodes.push_back(N);
3000 return SDValue(N, 0);
3003 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3004 SDValue N1, SDValue N2, SDValue N3) {
3005 // Perform various simplifications.
3006 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
3007 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
3009 case ISD::CONCAT_VECTORS:
3010 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
3011 // one big BUILD_VECTOR.
3012 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
3013 N2.getOpcode() == ISD::BUILD_VECTOR &&
3014 N3.getOpcode() == ISD::BUILD_VECTOR) {
3015 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
3016 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
3017 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
3018 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
3022 // Use FoldSetCC to simplify SETCC's.
3023 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL);
3024 if (Simp.getNode()) return Simp;
3029 if (N1C->getZExtValue())
3030 return N2; // select true, X, Y -> X
3032 return N3; // select false, X, Y -> Y
3035 if (N2 == N3) return N2; // select C, X, X -> X
3039 if (N2C->getZExtValue()) // Unconditional branch
3040 return getNode(ISD::BR, DL, MVT::Other, N1, N3);
3042 return N1; // Never-taken branch
3045 case ISD::VECTOR_SHUFFLE:
3046 llvm_unreachable("should use getVectorShuffle constructor!");
3048 case ISD::BIT_CONVERT:
3049 // Fold bit_convert nodes from a type to themselves.
3050 if (N1.getValueType() == VT)
3055 // Memoize node if it doesn't produce a flag.
3057 SDVTList VTs = getVTList(VT);
3058 if (VT != MVT::Flag) {
3059 SDValue Ops[] = { N1, N2, N3 };
3060 FoldingSetNodeID ID;
3061 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3063 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3064 return SDValue(E, 0);
3066 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3067 CSEMap.InsertNode(N, IP);
3069 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3072 AllNodes.push_back(N);
3076 return SDValue(N, 0);
3079 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3080 SDValue N1, SDValue N2, SDValue N3,
3082 SDValue Ops[] = { N1, N2, N3, N4 };
3083 return getNode(Opcode, DL, VT, Ops, 4);
3086 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3087 SDValue N1, SDValue N2, SDValue N3,
3088 SDValue N4, SDValue N5) {
3089 SDValue Ops[] = { N1, N2, N3, N4, N5 };
3090 return getNode(Opcode, DL, VT, Ops, 5);
3093 /// getStackArgumentTokenFactor - Compute a TokenFactor to force all
3094 /// the incoming stack arguments to be loaded from the stack.
3095 SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
3096 SmallVector<SDValue, 8> ArgChains;
3098 // Include the original chain at the beginning of the list. When this is
3099 // used by target LowerCall hooks, this helps legalize find the
3100 // CALLSEQ_BEGIN node.
3101 ArgChains.push_back(Chain);
3103 // Add a chain value for each stack argument.
3104 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(),
3105 UE = getEntryNode().getNode()->use_end(); U != UE; ++U)
3106 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
3107 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
3108 if (FI->getIndex() < 0)
3109 ArgChains.push_back(SDValue(L, 1));
3111 // Build a tokenfactor for all the chains.
3112 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
3113 &ArgChains[0], ArgChains.size());
3116 /// getMemsetValue - Vectorized representation of the memset value
3118 static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
3120 assert(Value.getOpcode() != ISD::UNDEF);
3122 unsigned NumBits = VT.getScalarType().getSizeInBits();
3123 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3124 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
3126 for (unsigned i = NumBits; i > 8; i >>= 1) {
3127 Val = (Val << Shift) | Val;
3131 return DAG.getConstant(Val, VT);
3132 return DAG.getConstantFP(APFloat(Val), VT);
3135 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3136 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
3138 for (unsigned i = NumBits; i > 8; i >>= 1) {
3139 Value = DAG.getNode(ISD::OR, dl, VT,
3140 DAG.getNode(ISD::SHL, dl, VT, Value,
3141 DAG.getConstant(Shift,
3142 TLI.getShiftAmountTy())),
3150 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3151 /// used when a memcpy is turned into a memset when the source is a constant
3153 static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG,
3154 const TargetLowering &TLI,
3155 std::string &Str, unsigned Offset) {
3156 // Handle vector with all elements zero.
3159 return DAG.getConstant(0, VT);
3160 else if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
3161 VT.getSimpleVT().SimpleTy == MVT::f64)
3162 return DAG.getConstantFP(0.0, VT);
3163 else if (VT.isVector()) {
3164 unsigned NumElts = VT.getVectorNumElements();
3165 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3166 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3167 DAG.getConstant(0, EVT::getVectorVT(*DAG.getContext(),
3170 llvm_unreachable("Expected type!");
3173 assert(!VT.isVector() && "Can't handle vector type here!");
3174 unsigned NumBits = VT.getSizeInBits();
3175 unsigned MSB = NumBits / 8;
3177 if (TLI.isLittleEndian())
3178 Offset = Offset + MSB - 1;
3179 for (unsigned i = 0; i != MSB; ++i) {
3180 Val = (Val << 8) | (unsigned char)Str[Offset];
3181 Offset += TLI.isLittleEndian() ? -1 : 1;
3183 return DAG.getConstant(Val, VT);
3186 /// getMemBasePlusOffset - Returns base and offset node for the
3188 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
3189 SelectionDAG &DAG) {
3190 EVT VT = Base.getValueType();
3191 return DAG.getNode(ISD::ADD, Base.getDebugLoc(),
3192 VT, Base, DAG.getConstant(Offset, VT));
3195 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
3197 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
3198 unsigned SrcDelta = 0;
3199 GlobalAddressSDNode *G = NULL;
3200 if (Src.getOpcode() == ISD::GlobalAddress)
3201 G = cast<GlobalAddressSDNode>(Src);
3202 else if (Src.getOpcode() == ISD::ADD &&
3203 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3204 Src.getOperand(1).getOpcode() == ISD::Constant) {
3205 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
3206 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
3211 const GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3212 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
3218 /// FindOptimalMemOpLowering - Determines the optimial series memory ops
3219 /// to replace the memset / memcpy. Return true if the number of memory ops
3220 /// is below the threshold. It returns the types of the sequence of
3221 /// memory ops to perform memset / memcpy by reference.
3222 static bool FindOptimalMemOpLowering(std::vector<EVT> &MemOps,
3223 unsigned Limit, uint64_t Size,
3224 unsigned DstAlign, unsigned SrcAlign,
3225 bool NonScalarIntSafe,
3228 const TargetLowering &TLI) {
3229 assert((SrcAlign == 0 || SrcAlign >= DstAlign) &&
3230 "Expecting memcpy / memset source to meet alignment requirement!");
3231 // If 'SrcAlign' is zero, that means the memory operation does not need load
3232 // the value, i.e. memset or memcpy from constant string. Otherwise, it's
3233 // the inferred alignment of the source. 'DstAlign', on the other hand, is the
3234 // specified alignment of the memory operation. If it is zero, that means
3235 // it's possible to change the alignment of the destination. 'MemcpyStrSrc'
3236 // indicates whether the memcpy source is constant so it does not need to be
3238 EVT VT = TLI.getOptimalMemOpType(Size, DstAlign, SrcAlign,
3239 NonScalarIntSafe, MemcpyStrSrc,
3240 DAG.getMachineFunction());
3242 if (VT == MVT::Other) {
3243 if (DstAlign >= TLI.getTargetData()->getPointerPrefAlignment() ||
3244 TLI.allowsUnalignedMemoryAccesses(VT)) {
3245 VT = TLI.getPointerTy();
3247 switch (DstAlign & 7) {
3248 case 0: VT = MVT::i64; break;
3249 case 4: VT = MVT::i32; break;
3250 case 2: VT = MVT::i16; break;
3251 default: VT = MVT::i8; break;
3256 while (!TLI.isTypeLegal(LVT))
3257 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1);
3258 assert(LVT.isInteger());
3264 unsigned NumMemOps = 0;
3266 unsigned VTSize = VT.getSizeInBits() / 8;
3267 while (VTSize > Size) {
3268 // For now, only use non-vector load / store's for the left-over pieces.
3269 if (VT.isVector() || VT.isFloatingPoint()) {
3271 while (!TLI.isTypeLegal(VT))
3272 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3273 VTSize = VT.getSizeInBits() / 8;
3275 // This can result in a type that is not legal on the target, e.g.
3276 // 1 or 2 bytes on PPC.
3277 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3282 if (++NumMemOps > Limit)
3284 MemOps.push_back(VT);
3291 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3292 SDValue Chain, SDValue Dst,
3293 SDValue Src, uint64_t Size,
3294 unsigned Align, bool isVol,
3296 const Value *DstSV, uint64_t DstSVOff,
3297 const Value *SrcSV, uint64_t SrcSVOff) {
3298 // Turn a memcpy of undef to nop.
3299 if (Src.getOpcode() == ISD::UNDEF)
3302 // Expand memcpy to a series of load and store ops if the size operand falls
3303 // below a certain threshold.
3304 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3305 std::vector<EVT> MemOps;
3306 bool DstAlignCanChange = false;
3307 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3308 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
3309 if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
3310 DstAlignCanChange = true;
3311 unsigned SrcAlign = DAG.InferPtrAlignment(Src);
3312 if (Align > SrcAlign)
3315 bool CopyFromStr = isMemSrcFromString(Src, Str);
3316 bool isZeroStr = CopyFromStr && Str.empty();
3317 uint64_t Limit = -1ULL;
3319 Limit = TLI.getMaxStoresPerMemcpy();
3320 if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
3321 (DstAlignCanChange ? 0 : Align),
3322 (isZeroStr ? 0 : SrcAlign),
3323 true, CopyFromStr, DAG, TLI))
3326 if (DstAlignCanChange) {
3327 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
3328 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3329 if (NewAlign > Align) {
3330 // Give the stack frame object a larger alignment if needed.
3331 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign)
3332 MFI->setObjectAlignment(FI->getIndex(), NewAlign);
3337 SmallVector<SDValue, 8> OutChains;
3338 unsigned NumMemOps = MemOps.size();
3339 uint64_t SrcOff = 0, DstOff = 0;
3340 for (unsigned i = 0; i != NumMemOps; ++i) {
3342 unsigned VTSize = VT.getSizeInBits() / 8;
3343 SDValue Value, Store;
3346 (isZeroStr || (VT.isInteger() && !VT.isVector()))) {
3347 // It's unlikely a store of a vector immediate can be done in a single
3348 // instruction. It would require a load from a constantpool first.
3349 // We only handle zero vectors here.
3350 // FIXME: Handle other cases where store of vector immediate is done in
3351 // a single instruction.
3352 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff);
3353 Store = DAG.getStore(Chain, dl, Value,
3354 getMemBasePlusOffset(Dst, DstOff, DAG),
3355 DstSV, DstSVOff + DstOff, isVol, false, Align);
3357 // The type might not be legal for the target. This should only happen
3358 // if the type is smaller than a legal type, as on PPC, so the right
3359 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
3360 // to Load/Store if NVT==VT.
3361 // FIXME does the case above also need this?
3362 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
3363 assert(NVT.bitsGE(VT));
3364 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain,
3365 getMemBasePlusOffset(Src, SrcOff, DAG),
3366 SrcSV, SrcSVOff + SrcOff, VT, isVol, false,
3367 MinAlign(SrcAlign, SrcOff));
3368 Store = DAG.getTruncStore(Chain, dl, Value,
3369 getMemBasePlusOffset(Dst, DstOff, DAG),
3370 DstSV, DstSVOff + DstOff, VT, isVol, false,
3373 OutChains.push_back(Store);
3378 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3379 &OutChains[0], OutChains.size());
3382 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3383 SDValue Chain, SDValue Dst,
3384 SDValue Src, uint64_t Size,
3385 unsigned Align, bool isVol,
3387 const Value *DstSV, uint64_t DstSVOff,
3388 const Value *SrcSV, uint64_t SrcSVOff) {
3389 // Turn a memmove of undef to nop.
3390 if (Src.getOpcode() == ISD::UNDEF)
3393 // Expand memmove to a series of load and store ops if the size operand falls
3394 // below a certain threshold.
3395 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3396 std::vector<EVT> MemOps;
3397 uint64_t Limit = -1ULL;
3399 Limit = TLI.getMaxStoresPerMemmove();
3400 bool DstAlignCanChange = false;
3401 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3402 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
3403 if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
3404 DstAlignCanChange = true;
3405 unsigned SrcAlign = DAG.InferPtrAlignment(Src);
3406 if (Align > SrcAlign)
3409 if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
3410 (DstAlignCanChange ? 0 : Align),
3411 SrcAlign, true, false, DAG, TLI))
3414 if (DstAlignCanChange) {
3415 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
3416 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3417 if (NewAlign > Align) {
3418 // Give the stack frame object a larger alignment if needed.
3419 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign)
3420 MFI->setObjectAlignment(FI->getIndex(), NewAlign);
3425 uint64_t SrcOff = 0, DstOff = 0;
3426 SmallVector<SDValue, 8> LoadValues;
3427 SmallVector<SDValue, 8> LoadChains;
3428 SmallVector<SDValue, 8> OutChains;
3429 unsigned NumMemOps = MemOps.size();
3430 for (unsigned i = 0; i < NumMemOps; i++) {
3432 unsigned VTSize = VT.getSizeInBits() / 8;
3433 SDValue Value, Store;
3435 Value = DAG.getLoad(VT, dl, Chain,
3436 getMemBasePlusOffset(Src, SrcOff, DAG),
3437 SrcSV, SrcSVOff + SrcOff, isVol, false, SrcAlign);
3438 LoadValues.push_back(Value);
3439 LoadChains.push_back(Value.getValue(1));
3442 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3443 &LoadChains[0], LoadChains.size());
3445 for (unsigned i = 0; i < NumMemOps; i++) {
3447 unsigned VTSize = VT.getSizeInBits() / 8;
3448 SDValue Value, Store;
3450 Store = DAG.getStore(Chain, dl, LoadValues[i],
3451 getMemBasePlusOffset(Dst, DstOff, DAG),
3452 DstSV, DstSVOff + DstOff, isVol, false, Align);
3453 OutChains.push_back(Store);
3457 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3458 &OutChains[0], OutChains.size());
3461 static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl,
3462 SDValue Chain, SDValue Dst,
3463 SDValue Src, uint64_t Size,
3464 unsigned Align, bool isVol,
3465 const Value *DstSV, uint64_t DstSVOff) {
3466 // Turn a memset of undef to nop.
3467 if (Src.getOpcode() == ISD::UNDEF)
3470 // Expand memset to a series of load/store ops if the size operand
3471 // falls below a certain threshold.
3472 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3473 std::vector<EVT> MemOps;
3474 bool DstAlignCanChange = false;
3475 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3476 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
3477 if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
3478 DstAlignCanChange = true;
3479 bool NonScalarIntSafe =
3480 isa<ConstantSDNode>(Src) && cast<ConstantSDNode>(Src)->isNullValue();
3481 if (!FindOptimalMemOpLowering(MemOps, TLI.getMaxStoresPerMemset(),
3482 Size, (DstAlignCanChange ? 0 : Align), 0,
3483 NonScalarIntSafe, false, DAG, TLI))
3486 if (DstAlignCanChange) {
3487 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
3488 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3489 if (NewAlign > Align) {
3490 // Give the stack frame object a larger alignment if needed.
3491 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign)
3492 MFI->setObjectAlignment(FI->getIndex(), NewAlign);
3497 SmallVector<SDValue, 8> OutChains;
3498 uint64_t DstOff = 0;
3499 unsigned NumMemOps = MemOps.size();
3500 for (unsigned i = 0; i < NumMemOps; i++) {
3502 unsigned VTSize = VT.getSizeInBits() / 8;
3503 SDValue Value = getMemsetValue(Src, VT, DAG, dl);
3504 SDValue Store = DAG.getStore(Chain, dl, Value,
3505 getMemBasePlusOffset(Dst, DstOff, DAG),
3506 DstSV, DstSVOff + DstOff, isVol, false, 0);
3507 OutChains.push_back(Store);
3511 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3512 &OutChains[0], OutChains.size());
3515 SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst,
3516 SDValue Src, SDValue Size,
3517 unsigned Align, bool isVol, bool AlwaysInline,
3518 const Value *DstSV, uint64_t DstSVOff,
3519 const Value *SrcSV, uint64_t SrcSVOff) {
3521 // Check to see if we should lower the memcpy to loads and stores first.
3522 // For cases within the target-specified limits, this is the best choice.
3523 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3525 // Memcpy with size zero? Just return the original chain.
3526 if (ConstantSize->isNullValue())
3529 SDValue Result = getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3530 ConstantSize->getZExtValue(),Align,
3531 isVol, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3532 if (Result.getNode())
3536 // Then check to see if we should lower the memcpy with target-specific
3537 // code. If the target chooses to do this, this is the next best.
3539 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align,
3540 isVol, AlwaysInline,
3541 DstSV, DstSVOff, SrcSV, SrcSVOff);
3542 if (Result.getNode())
3545 // If we really need inline code and the target declined to provide it,
3546 // use a (potentially long) sequence of loads and stores.
3548 assert(ConstantSize && "AlwaysInline requires a constant size!");
3549 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3550 ConstantSize->getZExtValue(), Align, isVol,
3551 true, DstSV, DstSVOff, SrcSV, SrcSVOff);
3554 // FIXME: If the memcpy is volatile (isVol), lowering it to a plain libc
3555 // memcpy is not guaranteed to be safe. libc memcpys aren't required to
3556 // respect volatile, so they may do things like read or write memory
3557 // beyond the given memory regions. But fixing this isn't easy, and most
3558 // people don't care.
3560 // Emit a library call.
3561 TargetLowering::ArgListTy Args;
3562 TargetLowering::ArgListEntry Entry;
3563 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3564 Entry.Node = Dst; Args.push_back(Entry);
3565 Entry.Node = Src; Args.push_back(Entry);
3566 Entry.Node = Size; Args.push_back(Entry);
3567 // FIXME: pass in DebugLoc
3568 std::pair<SDValue,SDValue> CallResult =
3569 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3570 false, false, false, false, 0,
3571 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false,
3572 /*isReturnValueUsed=*/false,
3573 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY),
3574 TLI.getPointerTy()),
3576 return CallResult.second;
3579 SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst,
3580 SDValue Src, SDValue Size,
3581 unsigned Align, bool isVol,
3582 const Value *DstSV, uint64_t DstSVOff,
3583 const Value *SrcSV, uint64_t SrcSVOff) {
3585 // Check to see if we should lower the memmove to loads and stores first.
3586 // For cases within the target-specified limits, this is the best choice.
3587 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3589 // Memmove with size zero? Just return the original chain.
3590 if (ConstantSize->isNullValue())
3594 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src,
3595 ConstantSize->getZExtValue(), Align, isVol,
3596 false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3597 if (Result.getNode())
3601 // Then check to see if we should lower the memmove with target-specific
3602 // code. If the target chooses to do this, this is the next best.
3604 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align, isVol,
3605 DstSV, DstSVOff, SrcSV, SrcSVOff);
3606 if (Result.getNode())
3609 // FIXME: If the memmove is volatile, lowering it to plain libc memmove may
3610 // not be safe. See memcpy above for more details.
3612 // Emit a library call.
3613 TargetLowering::ArgListTy Args;
3614 TargetLowering::ArgListEntry Entry;
3615 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3616 Entry.Node = Dst; Args.push_back(Entry);
3617 Entry.Node = Src; Args.push_back(Entry);
3618 Entry.Node = Size; Args.push_back(Entry);
3619 // FIXME: pass in DebugLoc
3620 std::pair<SDValue,SDValue> CallResult =
3621 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3622 false, false, false, false, 0,
3623 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false,
3624 /*isReturnValueUsed=*/false,
3625 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE),
3626 TLI.getPointerTy()),
3628 return CallResult.second;
3631 SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst,
3632 SDValue Src, SDValue Size,
3633 unsigned Align, bool isVol,
3634 const Value *DstSV, uint64_t DstSVOff) {
3636 // Check to see if we should lower the memset to stores first.
3637 // For cases within the target-specified limits, this is the best choice.
3638 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3640 // Memset with size zero? Just return the original chain.
3641 if (ConstantSize->isNullValue())
3645 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
3646 Align, isVol, DstSV, DstSVOff);
3648 if (Result.getNode())
3652 // Then check to see if we should lower the memset with target-specific
3653 // code. If the target chooses to do this, this is the next best.
3655 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align, isVol,
3657 if (Result.getNode())
3660 // Emit a library call.
3661 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext());
3662 TargetLowering::ArgListTy Args;
3663 TargetLowering::ArgListEntry Entry;
3664 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3665 Args.push_back(Entry);
3666 // Extend or truncate the argument to be an i32 value for the call.
3667 if (Src.getValueType().bitsGT(MVT::i32))
3668 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
3670 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
3672 Entry.Ty = Type::getInt32Ty(*getContext());
3673 Entry.isSExt = true;
3674 Args.push_back(Entry);
3676 Entry.Ty = IntPtrTy;
3677 Entry.isSExt = false;
3678 Args.push_back(Entry);
3679 // FIXME: pass in DebugLoc
3680 std::pair<SDValue,SDValue> CallResult =
3681 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3682 false, false, false, false, 0,
3683 TLI.getLibcallCallingConv(RTLIB::MEMSET), false,
3684 /*isReturnValueUsed=*/false,
3685 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET),
3686 TLI.getPointerTy()),
3688 return CallResult.second;
3691 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3693 SDValue Ptr, SDValue Cmp,
3694 SDValue Swp, const Value* PtrVal,
3695 unsigned Alignment) {
3696 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3697 Alignment = getEVTAlignment(MemVT);
3699 // Check if the memory reference references a frame index
3701 if (const FrameIndexSDNode *FI =
3702 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3703 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3705 MachineFunction &MF = getMachineFunction();
3706 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3708 // For now, atomics are considered to be volatile always.
3709 Flags |= MachineMemOperand::MOVolatile;
3711 MachineMemOperand *MMO =
3712 MF.getMachineMemOperand(PtrVal, Flags, 0,
3713 MemVT.getStoreSize(), Alignment);
3715 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3718 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3720 SDValue Ptr, SDValue Cmp,
3721 SDValue Swp, MachineMemOperand *MMO) {
3722 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op");
3723 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3725 EVT VT = Cmp.getValueType();
3727 SDVTList VTs = getVTList(VT, MVT::Other);
3728 FoldingSetNodeID ID;
3729 ID.AddInteger(MemVT.getRawBits());
3730 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3731 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3733 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3734 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3735 return SDValue(E, 0);
3737 SDNode *N = new (NodeAllocator) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain,
3738 Ptr, Cmp, Swp, MMO);
3739 CSEMap.InsertNode(N, IP);
3740 AllNodes.push_back(N);
3741 return SDValue(N, 0);
3744 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3746 SDValue Ptr, SDValue Val,
3747 const Value* PtrVal,
3748 unsigned Alignment) {
3749 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3750 Alignment = getEVTAlignment(MemVT);
3752 // Check if the memory reference references a frame index
3754 if (const FrameIndexSDNode *FI =
3755 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3756 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3758 MachineFunction &MF = getMachineFunction();
3759 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3761 // For now, atomics are considered to be volatile always.
3762 Flags |= MachineMemOperand::MOVolatile;
3764 MachineMemOperand *MMO =
3765 MF.getMachineMemOperand(PtrVal, Flags, 0,
3766 MemVT.getStoreSize(), Alignment);
3768 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO);
3771 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3773 SDValue Ptr, SDValue Val,
3774 MachineMemOperand *MMO) {
3775 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
3776 Opcode == ISD::ATOMIC_LOAD_SUB ||
3777 Opcode == ISD::ATOMIC_LOAD_AND ||
3778 Opcode == ISD::ATOMIC_LOAD_OR ||
3779 Opcode == ISD::ATOMIC_LOAD_XOR ||
3780 Opcode == ISD::ATOMIC_LOAD_NAND ||
3781 Opcode == ISD::ATOMIC_LOAD_MIN ||
3782 Opcode == ISD::ATOMIC_LOAD_MAX ||
3783 Opcode == ISD::ATOMIC_LOAD_UMIN ||
3784 Opcode == ISD::ATOMIC_LOAD_UMAX ||
3785 Opcode == ISD::ATOMIC_SWAP) &&
3786 "Invalid Atomic Op");
3788 EVT VT = Val.getValueType();
3790 SDVTList VTs = getVTList(VT, MVT::Other);
3791 FoldingSetNodeID ID;
3792 ID.AddInteger(MemVT.getRawBits());
3793 SDValue Ops[] = {Chain, Ptr, Val};
3794 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3796 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3797 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3798 return SDValue(E, 0);
3800 SDNode *N = new (NodeAllocator) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain,
3802 CSEMap.InsertNode(N, IP);
3803 AllNodes.push_back(N);
3804 return SDValue(N, 0);
3807 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3808 /// Allowed to return something different (and simpler) if Simplify is true.
3809 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3814 SmallVector<EVT, 4> VTs;
3815 VTs.reserve(NumOps);
3816 for (unsigned i = 0; i < NumOps; ++i)
3817 VTs.push_back(Ops[i].getValueType());
3818 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps),
3823 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl,
3824 const EVT *VTs, unsigned NumVTs,
3825 const SDValue *Ops, unsigned NumOps,
3826 EVT MemVT, const Value *srcValue, int SVOff,
3827 unsigned Align, bool Vol,
3828 bool ReadMem, bool WriteMem) {
3829 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps,
3830 MemVT, srcValue, SVOff, Align, Vol,
3835 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3836 const SDValue *Ops, unsigned NumOps,
3837 EVT MemVT, const Value *srcValue, int SVOff,
3838 unsigned Align, bool Vol,
3839 bool ReadMem, bool WriteMem) {
3840 if (Align == 0) // Ensure that codegen never sees alignment 0
3841 Align = getEVTAlignment(MemVT);
3843 MachineFunction &MF = getMachineFunction();
3846 Flags |= MachineMemOperand::MOStore;
3848 Flags |= MachineMemOperand::MOLoad;
3850 Flags |= MachineMemOperand::MOVolatile;
3851 MachineMemOperand *MMO =
3852 MF.getMachineMemOperand(srcValue, Flags, SVOff,
3853 MemVT.getStoreSize(), Align);
3855 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3859 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3860 const SDValue *Ops, unsigned NumOps,
3861 EVT MemVT, MachineMemOperand *MMO) {
3862 assert((Opcode == ISD::INTRINSIC_VOID ||
3863 Opcode == ISD::INTRINSIC_W_CHAIN ||
3864 (Opcode <= INT_MAX &&
3865 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) &&
3866 "Opcode is not a memory-accessing opcode!");
3868 // Memoize the node unless it returns a flag.
3869 MemIntrinsicSDNode *N;
3870 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3871 FoldingSetNodeID ID;
3872 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3874 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3875 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO);
3876 return SDValue(E, 0);
3879 N = new (NodeAllocator) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps,
3881 CSEMap.InsertNode(N, IP);
3883 N = new (NodeAllocator) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps,
3886 AllNodes.push_back(N);
3887 return SDValue(N, 0);
3891 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3892 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3893 SDValue Ptr, SDValue Offset,
3894 const Value *SV, int SVOffset, EVT MemVT,
3895 bool isVolatile, bool isNonTemporal,
3896 unsigned Alignment) {
3897 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3898 Alignment = getEVTAlignment(VT);
3900 // Check if the memory reference references a frame index
3902 if (const FrameIndexSDNode *FI =
3903 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3904 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3906 MachineFunction &MF = getMachineFunction();
3907 unsigned Flags = MachineMemOperand::MOLoad;
3909 Flags |= MachineMemOperand::MOVolatile;
3911 Flags |= MachineMemOperand::MONonTemporal;
3912 MachineMemOperand *MMO =
3913 MF.getMachineMemOperand(SV, Flags, SVOffset,
3914 MemVT.getStoreSize(), Alignment);
3915 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO);
3919 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3920 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3921 SDValue Ptr, SDValue Offset, EVT MemVT,
3922 MachineMemOperand *MMO) {
3924 ExtType = ISD::NON_EXTLOAD;
3925 } else if (ExtType == ISD::NON_EXTLOAD) {
3926 assert(VT == MemVT && "Non-extending load from different memory type!");
3929 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) &&
3930 "Should only be an extending load, not truncating!");
3931 assert(VT.isInteger() == MemVT.isInteger() &&
3932 "Cannot convert from FP to Int or Int -> FP!");
3933 assert(VT.isVector() == MemVT.isVector() &&
3934 "Cannot use trunc store to convert to or from a vector!");
3935 assert((!VT.isVector() ||
3936 VT.getVectorNumElements() == MemVT.getVectorNumElements()) &&
3937 "Cannot use trunc store to change the number of vector elements!");
3940 bool Indexed = AM != ISD::UNINDEXED;
3941 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3942 "Unindexed load with an offset!");
3944 SDVTList VTs = Indexed ?
3945 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3946 SDValue Ops[] = { Chain, Ptr, Offset };
3947 FoldingSetNodeID ID;
3948 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3949 ID.AddInteger(MemVT.getRawBits());
3950 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile(),
3951 MMO->isNonTemporal()));
3953 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3954 cast<LoadSDNode>(E)->refineAlignment(MMO);
3955 return SDValue(E, 0);
3957 SDNode *N = new (NodeAllocator) LoadSDNode(Ops, dl, VTs, AM, ExtType,
3959 CSEMap.InsertNode(N, IP);
3960 AllNodes.push_back(N);
3961 return SDValue(N, 0);
3964 SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl,
3965 SDValue Chain, SDValue Ptr,
3966 const Value *SV, int SVOffset,
3967 bool isVolatile, bool isNonTemporal,
3968 unsigned Alignment) {
3969 SDValue Undef = getUNDEF(Ptr.getValueType());
3970 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3971 SV, SVOffset, VT, isVolatile, isNonTemporal, Alignment);
3974 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT,
3975 SDValue Chain, SDValue Ptr,
3977 int SVOffset, EVT MemVT,
3978 bool isVolatile, bool isNonTemporal,
3979 unsigned Alignment) {
3980 SDValue Undef = getUNDEF(Ptr.getValueType());
3981 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef,
3982 SV, SVOffset, MemVT, isVolatile, isNonTemporal, Alignment);
3986 SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base,
3987 SDValue Offset, ISD::MemIndexedMode AM) {
3988 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
3989 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
3990 "Load is already a indexed load!");
3991 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(),
3992 LD->getChain(), Base, Offset, LD->getSrcValue(),
3993 LD->getSrcValueOffset(), LD->getMemoryVT(),
3994 LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment());
3997 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3998 SDValue Ptr, const Value *SV, int SVOffset,
3999 bool isVolatile, bool isNonTemporal,
4000 unsigned Alignment) {
4001 if (Alignment == 0) // Ensure that codegen never sees alignment 0
4002 Alignment = getEVTAlignment(Val.getValueType());
4004 // Check if the memory reference references a frame index
4006 if (const FrameIndexSDNode *FI =
4007 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
4008 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
4010 MachineFunction &MF = getMachineFunction();
4011 unsigned Flags = MachineMemOperand::MOStore;
4013 Flags |= MachineMemOperand::MOVolatile;
4015 Flags |= MachineMemOperand::MONonTemporal;
4016 MachineMemOperand *MMO =
4017 MF.getMachineMemOperand(SV, Flags, SVOffset,
4018 Val.getValueType().getStoreSize(), Alignment);
4020 return getStore(Chain, dl, Val, Ptr, MMO);
4023 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
4024 SDValue Ptr, MachineMemOperand *MMO) {
4025 EVT VT = Val.getValueType();
4026 SDVTList VTs = getVTList(MVT::Other);
4027 SDValue Undef = getUNDEF(Ptr.getValueType());
4028 SDValue Ops[] = { Chain, Val, Ptr, Undef };
4029 FoldingSetNodeID ID;
4030 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4031 ID.AddInteger(VT.getRawBits());
4032 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile(),
4033 MMO->isNonTemporal()));
4035 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
4036 cast<StoreSDNode>(E)->refineAlignment(MMO);
4037 return SDValue(E, 0);
4039 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED,
4041 CSEMap.InsertNode(N, IP);
4042 AllNodes.push_back(N);
4043 return SDValue(N, 0);
4046 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
4047 SDValue Ptr, const Value *SV,
4048 int SVOffset, EVT SVT,
4049 bool isVolatile, bool isNonTemporal,
4050 unsigned Alignment) {
4051 if (Alignment == 0) // Ensure that codegen never sees alignment 0
4052 Alignment = getEVTAlignment(SVT);
4054 // Check if the memory reference references a frame index
4056 if (const FrameIndexSDNode *FI =
4057 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
4058 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
4060 MachineFunction &MF = getMachineFunction();
4061 unsigned Flags = MachineMemOperand::MOStore;
4063 Flags |= MachineMemOperand::MOVolatile;
4065 Flags |= MachineMemOperand::MONonTemporal;
4066 MachineMemOperand *MMO =
4067 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment);
4069 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
4072 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
4073 SDValue Ptr, EVT SVT,
4074 MachineMemOperand *MMO) {
4075 EVT VT = Val.getValueType();
4078 return getStore(Chain, dl, Val, Ptr, MMO);
4080 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
4081 "Should only be a truncating store, not extending!");
4082 assert(VT.isInteger() == SVT.isInteger() &&
4083 "Can't do FP-INT conversion!");
4084 assert(VT.isVector() == SVT.isVector() &&
4085 "Cannot use trunc store to convert to or from a vector!");
4086 assert((!VT.isVector() ||
4087 VT.getVectorNumElements() == SVT.getVectorNumElements()) &&
4088 "Cannot use trunc store to change the number of vector elements!");
4090 SDVTList VTs = getVTList(MVT::Other);
4091 SDValue Undef = getUNDEF(Ptr.getValueType());
4092 SDValue Ops[] = { Chain, Val, Ptr, Undef };
4093 FoldingSetNodeID ID;
4094 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4095 ID.AddInteger(SVT.getRawBits());
4096 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile(),
4097 MMO->isNonTemporal()));
4099 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
4100 cast<StoreSDNode>(E)->refineAlignment(MMO);
4101 return SDValue(E, 0);
4103 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED,
4105 CSEMap.InsertNode(N, IP);
4106 AllNodes.push_back(N);
4107 return SDValue(N, 0);
4111 SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base,
4112 SDValue Offset, ISD::MemIndexedMode AM) {
4113 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
4114 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
4115 "Store is already a indexed store!");
4116 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
4117 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
4118 FoldingSetNodeID ID;
4119 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4120 ID.AddInteger(ST->getMemoryVT().getRawBits());
4121 ID.AddInteger(ST->getRawSubclassData());
4123 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4124 return SDValue(E, 0);
4126 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, AM,
4127 ST->isTruncatingStore(),
4129 ST->getMemOperand());
4130 CSEMap.InsertNode(N, IP);
4131 AllNodes.push_back(N);
4132 return SDValue(N, 0);
4135 SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl,
4136 SDValue Chain, SDValue Ptr,
4138 SDValue Ops[] = { Chain, Ptr, SV };
4139 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3);
4142 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4143 const SDUse *Ops, unsigned NumOps) {
4145 case 0: return getNode(Opcode, DL, VT);
4146 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4147 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4148 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4152 // Copy from an SDUse array into an SDValue array for use with
4153 // the regular getNode logic.
4154 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
4155 return getNode(Opcode, DL, VT, &NewOps[0], NumOps);
4158 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4159 const SDValue *Ops, unsigned NumOps) {
4161 case 0: return getNode(Opcode, DL, VT);
4162 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4163 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4164 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4170 case ISD::SELECT_CC: {
4171 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
4172 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
4173 "LHS and RHS of condition must have same type!");
4174 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4175 "True and False arms of SelectCC must have same type!");
4176 assert(Ops[2].getValueType() == VT &&
4177 "select_cc node must be of same type as true and false value!");
4181 assert(NumOps == 5 && "BR_CC takes 5 operands!");
4182 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4183 "LHS/RHS of comparison should match types!");
4190 SDVTList VTs = getVTList(VT);
4192 if (VT != MVT::Flag) {
4193 FoldingSetNodeID ID;
4194 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
4197 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4198 return SDValue(E, 0);
4200 N = new (NodeAllocator) SDNode(Opcode, DL, VTs, Ops, NumOps);
4201 CSEMap.InsertNode(N, IP);
4203 N = new (NodeAllocator) SDNode(Opcode, DL, VTs, Ops, NumOps);
4206 AllNodes.push_back(N);
4210 return SDValue(N, 0);
4213 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4214 const std::vector<EVT> &ResultTys,
4215 const SDValue *Ops, unsigned NumOps) {
4216 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()),
4220 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4221 const EVT *VTs, unsigned NumVTs,
4222 const SDValue *Ops, unsigned NumOps) {
4224 return getNode(Opcode, DL, VTs[0], Ops, NumOps);
4225 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps);
4228 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4229 const SDValue *Ops, unsigned NumOps) {
4230 if (VTList.NumVTs == 1)
4231 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps);
4235 // FIXME: figure out how to safely handle things like
4236 // int foo(int x) { return 1 << (x & 255); }
4237 // int bar() { return foo(256); }
4238 case ISD::SRA_PARTS:
4239 case ISD::SRL_PARTS:
4240 case ISD::SHL_PARTS:
4241 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4242 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
4243 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4244 else if (N3.getOpcode() == ISD::AND)
4245 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
4246 // If the and is only masking out bits that cannot effect the shift,
4247 // eliminate the and.
4248 unsigned NumBits = VT.getScalarType().getSizeInBits()*2;
4249 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
4250 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4256 // Memoize the node unless it returns a flag.
4258 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4259 FoldingSetNodeID ID;
4260 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4262 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4263 return SDValue(E, 0);
4266 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4267 } else if (NumOps == 2) {
4268 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4269 } else if (NumOps == 3) {
4270 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1],
4273 N = new (NodeAllocator) SDNode(Opcode, DL, VTList, Ops, NumOps);
4275 CSEMap.InsertNode(N, IP);
4278 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4279 } else if (NumOps == 2) {
4280 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4281 } else if (NumOps == 3) {
4282 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1],
4285 N = new (NodeAllocator) SDNode(Opcode, DL, VTList, Ops, NumOps);
4288 AllNodes.push_back(N);
4292 return SDValue(N, 0);
4295 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) {
4296 return getNode(Opcode, DL, VTList, 0, 0);
4299 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4301 SDValue Ops[] = { N1 };
4302 return getNode(Opcode, DL, VTList, Ops, 1);
4305 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4306 SDValue N1, SDValue N2) {
4307 SDValue Ops[] = { N1, N2 };
4308 return getNode(Opcode, DL, VTList, Ops, 2);
4311 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4312 SDValue N1, SDValue N2, SDValue N3) {
4313 SDValue Ops[] = { N1, N2, N3 };
4314 return getNode(Opcode, DL, VTList, Ops, 3);
4317 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4318 SDValue N1, SDValue N2, SDValue N3,
4320 SDValue Ops[] = { N1, N2, N3, N4 };
4321 return getNode(Opcode, DL, VTList, Ops, 4);
4324 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4325 SDValue N1, SDValue N2, SDValue N3,
4326 SDValue N4, SDValue N5) {
4327 SDValue Ops[] = { N1, N2, N3, N4, N5 };
4328 return getNode(Opcode, DL, VTList, Ops, 5);
4331 SDVTList SelectionDAG::getVTList(EVT VT) {
4332 return makeVTList(SDNode::getValueTypeList(VT), 1);
4335 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
4336 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4337 E = VTList.rend(); I != E; ++I)
4338 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
4341 EVT *Array = Allocator.Allocate<EVT>(2);
4344 SDVTList Result = makeVTList(Array, 2);
4345 VTList.push_back(Result);
4349 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
4350 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4351 E = VTList.rend(); I != E; ++I)
4352 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4356 EVT *Array = Allocator.Allocate<EVT>(3);
4360 SDVTList Result = makeVTList(Array, 3);
4361 VTList.push_back(Result);
4365 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
4366 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4367 E = VTList.rend(); I != E; ++I)
4368 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4369 I->VTs[2] == VT3 && I->VTs[3] == VT4)
4372 EVT *Array = Allocator.Allocate<EVT>(4);
4377 SDVTList Result = makeVTList(Array, 4);
4378 VTList.push_back(Result);
4382 SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) {
4384 case 0: llvm_unreachable("Cannot have nodes without results!");
4385 case 1: return getVTList(VTs[0]);
4386 case 2: return getVTList(VTs[0], VTs[1]);
4387 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
4388 case 4: return getVTList(VTs[0], VTs[1], VTs[2], VTs[3]);
4392 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4393 E = VTList.rend(); I != E; ++I) {
4394 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
4397 bool NoMatch = false;
4398 for (unsigned i = 2; i != NumVTs; ++i)
4399 if (VTs[i] != I->VTs[i]) {
4407 EVT *Array = Allocator.Allocate<EVT>(NumVTs);
4408 std::copy(VTs, VTs+NumVTs, Array);
4409 SDVTList Result = makeVTList(Array, NumVTs);
4410 VTList.push_back(Result);
4415 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
4416 /// specified operands. If the resultant node already exists in the DAG,
4417 /// this does not modify the specified node, instead it returns the node that
4418 /// already exists. If the resultant node does not exist in the DAG, the
4419 /// input node is returned. As a degenerate case, if you specify the same
4420 /// input operands as the node already has, the input node is returned.
4421 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
4422 SDNode *N = InN.getNode();
4423 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
4425 // Check to see if there is no change.
4426 if (Op == N->getOperand(0)) return InN;
4428 // See if the modified node already exists.
4429 void *InsertPos = 0;
4430 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
4431 return SDValue(Existing, InN.getResNo());
4433 // Nope it doesn't. Remove the node from its current place in the maps.
4435 if (!RemoveNodeFromCSEMaps(N))
4438 // Now we update the operands.
4439 N->OperandList[0].set(Op);
4441 // If this gets put into a CSE map, add it.
4442 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4446 SDValue SelectionDAG::
4447 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
4448 SDNode *N = InN.getNode();
4449 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
4451 // Check to see if there is no change.
4452 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
4453 return InN; // No operands changed, just return the input node.
4455 // See if the modified node already exists.
4456 void *InsertPos = 0;
4457 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
4458 return SDValue(Existing, InN.getResNo());
4460 // Nope it doesn't. Remove the node from its current place in the maps.
4462 if (!RemoveNodeFromCSEMaps(N))
4465 // Now we update the operands.
4466 if (N->OperandList[0] != Op1)
4467 N->OperandList[0].set(Op1);
4468 if (N->OperandList[1] != Op2)
4469 N->OperandList[1].set(Op2);
4471 // If this gets put into a CSE map, add it.
4472 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4476 SDValue SelectionDAG::
4477 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
4478 SDValue Ops[] = { Op1, Op2, Op3 };
4479 return UpdateNodeOperands(N, Ops, 3);
4482 SDValue SelectionDAG::
4483 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4484 SDValue Op3, SDValue Op4) {
4485 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
4486 return UpdateNodeOperands(N, Ops, 4);
4489 SDValue SelectionDAG::
4490 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4491 SDValue Op3, SDValue Op4, SDValue Op5) {
4492 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
4493 return UpdateNodeOperands(N, Ops, 5);
4496 SDValue SelectionDAG::
4497 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
4498 SDNode *N = InN.getNode();
4499 assert(N->getNumOperands() == NumOps &&
4500 "Update with wrong number of operands");
4502 // Check to see if there is no change.
4503 bool AnyChange = false;
4504 for (unsigned i = 0; i != NumOps; ++i) {
4505 if (Ops[i] != N->getOperand(i)) {
4511 // No operands changed, just return the input node.
4512 if (!AnyChange) return InN;
4514 // See if the modified node already exists.
4515 void *InsertPos = 0;
4516 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
4517 return SDValue(Existing, InN.getResNo());
4519 // Nope it doesn't. Remove the node from its current place in the maps.
4521 if (!RemoveNodeFromCSEMaps(N))
4524 // Now we update the operands.
4525 for (unsigned i = 0; i != NumOps; ++i)
4526 if (N->OperandList[i] != Ops[i])
4527 N->OperandList[i].set(Ops[i]);
4529 // If this gets put into a CSE map, add it.
4530 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4534 /// DropOperands - Release the operands and set this node to have
4536 void SDNode::DropOperands() {
4537 // Unlike the code in MorphNodeTo that does this, we don't need to
4538 // watch for dead nodes here.
4539 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
4545 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
4548 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4550 SDVTList VTs = getVTList(VT);
4551 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4554 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4555 EVT VT, SDValue Op1) {
4556 SDVTList VTs = getVTList(VT);
4557 SDValue Ops[] = { Op1 };
4558 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4561 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4562 EVT VT, SDValue Op1,
4564 SDVTList VTs = getVTList(VT);
4565 SDValue Ops[] = { Op1, Op2 };
4566 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4569 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4570 EVT VT, SDValue Op1,
4571 SDValue Op2, SDValue Op3) {
4572 SDVTList VTs = getVTList(VT);
4573 SDValue Ops[] = { Op1, Op2, Op3 };
4574 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4577 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4578 EVT VT, const SDValue *Ops,
4580 SDVTList VTs = getVTList(VT);
4581 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4584 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4585 EVT VT1, EVT VT2, const SDValue *Ops,
4587 SDVTList VTs = getVTList(VT1, VT2);
4588 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4591 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4593 SDVTList VTs = getVTList(VT1, VT2);
4594 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4597 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4598 EVT VT1, EVT VT2, EVT VT3,
4599 const SDValue *Ops, unsigned NumOps) {
4600 SDVTList VTs = getVTList(VT1, VT2, VT3);
4601 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4604 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4605 EVT VT1, EVT VT2, EVT VT3, EVT VT4,
4606 const SDValue *Ops, unsigned NumOps) {
4607 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4608 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4611 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4614 SDVTList VTs = getVTList(VT1, VT2);
4615 SDValue Ops[] = { Op1 };
4616 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4619 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4621 SDValue Op1, SDValue Op2) {
4622 SDVTList VTs = getVTList(VT1, VT2);
4623 SDValue Ops[] = { Op1, Op2 };
4624 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4627 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4629 SDValue Op1, SDValue Op2,
4631 SDVTList VTs = getVTList(VT1, VT2);
4632 SDValue Ops[] = { Op1, Op2, Op3 };
4633 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4636 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4637 EVT VT1, EVT VT2, EVT VT3,
4638 SDValue Op1, SDValue Op2,
4640 SDVTList VTs = getVTList(VT1, VT2, VT3);
4641 SDValue Ops[] = { Op1, Op2, Op3 };
4642 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4645 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4646 SDVTList VTs, const SDValue *Ops,
4648 N = MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4649 // Reset the NodeID to -1.
4654 /// MorphNodeTo - This *mutates* the specified node to have the specified
4655 /// return type, opcode, and operands.
4657 /// Note that MorphNodeTo returns the resultant node. If there is already a
4658 /// node of the specified opcode and operands, it returns that node instead of
4659 /// the current one. Note that the DebugLoc need not be the same.
4661 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4662 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4663 /// node, and because it doesn't require CSE recalculation for any of
4664 /// the node's users.
4666 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4667 SDVTList VTs, const SDValue *Ops,
4669 // If an identical node already exists, use it.
4671 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4672 FoldingSetNodeID ID;
4673 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4674 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4678 if (!RemoveNodeFromCSEMaps(N))
4681 // Start the morphing.
4683 N->ValueList = VTs.VTs;
4684 N->NumValues = VTs.NumVTs;
4686 // Clear the operands list, updating used nodes to remove this from their
4687 // use list. Keep track of any operands that become dead as a result.
4688 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4689 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
4691 SDNode *Used = Use.getNode();
4693 if (Used->use_empty())
4694 DeadNodeSet.insert(Used);
4697 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) {
4698 // Initialize the memory references information.
4699 MN->setMemRefs(0, 0);
4700 // If NumOps is larger than the # of operands we can have in a
4701 // MachineSDNode, reallocate the operand list.
4702 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) {
4703 if (MN->OperandsNeedDelete)
4704 delete[] MN->OperandList;
4705 if (NumOps > array_lengthof(MN->LocalOperands))
4706 // We're creating a final node that will live unmorphed for the
4707 // remainder of the current SelectionDAG iteration, so we can allocate
4708 // the operands directly out of a pool with no recycling metadata.
4709 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4712 MN->InitOperands(MN->LocalOperands, Ops, NumOps);
4713 MN->OperandsNeedDelete = false;
4715 MN->InitOperands(MN->OperandList, Ops, NumOps);
4717 // If NumOps is larger than the # of operands we currently have, reallocate
4718 // the operand list.
4719 if (NumOps > N->NumOperands) {
4720 if (N->OperandsNeedDelete)
4721 delete[] N->OperandList;
4722 N->InitOperands(new SDUse[NumOps], Ops, NumOps);
4723 N->OperandsNeedDelete = true;
4725 N->InitOperands(N->OperandList, Ops, NumOps);
4728 // Delete any nodes that are still dead after adding the uses for the
4730 if (!DeadNodeSet.empty()) {
4731 SmallVector<SDNode *, 16> DeadNodes;
4732 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4733 E = DeadNodeSet.end(); I != E; ++I)
4734 if ((*I)->use_empty())
4735 DeadNodes.push_back(*I);
4736 RemoveDeadNodes(DeadNodes);
4740 CSEMap.InsertNode(N, IP); // Memoize the new node.
4745 /// getMachineNode - These are used for target selectors to create a new node
4746 /// with specified return type(s), MachineInstr opcode, and operands.
4748 /// Note that getMachineNode returns the resultant node. If there is already a
4749 /// node of the specified opcode and operands, it returns that node instead of
4750 /// the current one.
4752 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) {
4753 SDVTList VTs = getVTList(VT);
4754 return getMachineNode(Opcode, dl, VTs, 0, 0);
4758 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) {
4759 SDVTList VTs = getVTList(VT);
4760 SDValue Ops[] = { Op1 };
4761 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4765 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4766 SDValue Op1, SDValue Op2) {
4767 SDVTList VTs = getVTList(VT);
4768 SDValue Ops[] = { Op1, Op2 };
4769 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4773 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4774 SDValue Op1, SDValue Op2, SDValue Op3) {
4775 SDVTList VTs = getVTList(VT);
4776 SDValue Ops[] = { Op1, Op2, Op3 };
4777 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4781 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4782 const SDValue *Ops, unsigned NumOps) {
4783 SDVTList VTs = getVTList(VT);
4784 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4788 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) {
4789 SDVTList VTs = getVTList(VT1, VT2);
4790 return getMachineNode(Opcode, dl, VTs, 0, 0);
4794 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4795 EVT VT1, EVT VT2, SDValue Op1) {
4796 SDVTList VTs = getVTList(VT1, VT2);
4797 SDValue Ops[] = { Op1 };
4798 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4802 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4803 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) {
4804 SDVTList VTs = getVTList(VT1, VT2);
4805 SDValue Ops[] = { Op1, Op2 };
4806 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4810 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4811 EVT VT1, EVT VT2, SDValue Op1,
4812 SDValue Op2, SDValue Op3) {
4813 SDVTList VTs = getVTList(VT1, VT2);
4814 SDValue Ops[] = { Op1, Op2, Op3 };
4815 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4819 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4821 const SDValue *Ops, unsigned NumOps) {
4822 SDVTList VTs = getVTList(VT1, VT2);
4823 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4827 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4828 EVT VT1, EVT VT2, EVT VT3,
4829 SDValue Op1, SDValue Op2) {
4830 SDVTList VTs = getVTList(VT1, VT2, VT3);
4831 SDValue Ops[] = { Op1, Op2 };
4832 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4836 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4837 EVT VT1, EVT VT2, EVT VT3,
4838 SDValue Op1, SDValue Op2, SDValue Op3) {
4839 SDVTList VTs = getVTList(VT1, VT2, VT3);
4840 SDValue Ops[] = { Op1, Op2, Op3 };
4841 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4845 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4846 EVT VT1, EVT VT2, EVT VT3,
4847 const SDValue *Ops, unsigned NumOps) {
4848 SDVTList VTs = getVTList(VT1, VT2, VT3);
4849 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4853 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1,
4854 EVT VT2, EVT VT3, EVT VT4,
4855 const SDValue *Ops, unsigned NumOps) {
4856 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4857 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4861 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4862 const std::vector<EVT> &ResultTys,
4863 const SDValue *Ops, unsigned NumOps) {
4864 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size());
4865 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4869 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs,
4870 const SDValue *Ops, unsigned NumOps) {
4871 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag;
4876 FoldingSetNodeID ID;
4877 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps);
4879 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4880 return cast<MachineSDNode>(E);
4883 // Allocate a new MachineSDNode.
4884 N = new (NodeAllocator) MachineSDNode(~Opcode, DL, VTs);
4886 // Initialize the operands list.
4887 if (NumOps > array_lengthof(N->LocalOperands))
4888 // We're creating a final node that will live unmorphed for the
4889 // remainder of the current SelectionDAG iteration, so we can allocate
4890 // the operands directly out of a pool with no recycling metadata.
4891 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4894 N->InitOperands(N->LocalOperands, Ops, NumOps);
4895 N->OperandsNeedDelete = false;
4898 CSEMap.InsertNode(N, IP);
4900 AllNodes.push_back(N);
4907 /// getTargetExtractSubreg - A convenience function for creating
4908 /// TargetOpcode::EXTRACT_SUBREG nodes.
4910 SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT,
4912 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4913 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
4914 VT, Operand, SRIdxVal);
4915 return SDValue(Subreg, 0);
4918 /// getTargetInsertSubreg - A convenience function for creating
4919 /// TargetOpcode::INSERT_SUBREG nodes.
4921 SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT,
4922 SDValue Operand, SDValue Subreg) {
4923 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4924 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
4925 VT, Operand, Subreg, SRIdxVal);
4926 return SDValue(Result, 0);
4929 /// getNodeIfExists - Get the specified node if it's already available, or
4930 /// else return NULL.
4931 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4932 const SDValue *Ops, unsigned NumOps) {
4933 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4934 FoldingSetNodeID ID;
4935 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4937 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4943 /// getDbgValue - Creates a SDDbgValue node.
4946 SelectionDAG::getDbgValue(MDNode *MDPtr, SDNode *N, unsigned R, uint64_t Off,
4947 DebugLoc DL, unsigned O) {
4948 return new (Allocator) SDDbgValue(MDPtr, N, R, Off, DL, O);
4952 SelectionDAG::getDbgValue(MDNode *MDPtr, const Value *C, uint64_t Off,
4953 DebugLoc DL, unsigned O) {
4954 return new (Allocator) SDDbgValue(MDPtr, C, Off, DL, O);
4958 SelectionDAG::getDbgValue(MDNode *MDPtr, unsigned FI, uint64_t Off,
4959 DebugLoc DL, unsigned O) {
4960 return new (Allocator) SDDbgValue(MDPtr, FI, Off, DL, O);
4965 /// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node
4966 /// pointed to by a use iterator is deleted, increment the use iterator
4967 /// so that it doesn't dangle.
4969 /// This class also manages a "downlink" DAGUpdateListener, to forward
4970 /// messages to ReplaceAllUsesWith's callers.
4972 class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener {
4973 SelectionDAG::DAGUpdateListener *DownLink;
4974 SDNode::use_iterator &UI;
4975 SDNode::use_iterator &UE;
4977 virtual void NodeDeleted(SDNode *N, SDNode *E) {
4978 // Increment the iterator as needed.
4979 while (UI != UE && N == *UI)
4982 // Then forward the message.
4983 if (DownLink) DownLink->NodeDeleted(N, E);
4986 virtual void NodeUpdated(SDNode *N) {
4987 // Just forward the message.
4988 if (DownLink) DownLink->NodeUpdated(N);
4992 RAUWUpdateListener(SelectionDAG::DAGUpdateListener *dl,
4993 SDNode::use_iterator &ui,
4994 SDNode::use_iterator &ue)
4995 : DownLink(dl), UI(ui), UE(ue) {}
5000 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5001 /// This can cause recursive merging of nodes in the DAG.
5003 /// This version assumes From has a single result value.
5005 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
5006 DAGUpdateListener *UpdateListener) {
5007 SDNode *From = FromN.getNode();
5008 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
5009 "Cannot replace with this method!");
5010 assert(From != To.getNode() && "Cannot replace uses of with self");
5012 // Iterate over all the existing uses of From. New uses will be added
5013 // to the beginning of the use list, which we avoid visiting.
5014 // This specifically avoids visiting uses of From that arise while the
5015 // replacement is happening, because any such uses would be the result
5016 // of CSE: If an existing node looks like From after one of its operands
5017 // is replaced by To, we don't want to replace of all its users with To
5018 // too. See PR3018 for more info.
5019 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5020 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5024 // This node is about to morph, remove its old self from the CSE maps.
5025 RemoveNodeFromCSEMaps(User);
5027 // A user can appear in a use list multiple times, and when this
5028 // happens the uses are usually next to each other in the list.
5029 // To help reduce the number of CSE recomputations, process all
5030 // the uses of this user that we can find this way.
5032 SDUse &Use = UI.getUse();
5035 } while (UI != UE && *UI == User);
5037 // Now that we have modified User, add it back to the CSE maps. If it
5038 // already exists there, recursively merge the results together.
5039 AddModifiedNodeToCSEMaps(User, &Listener);
5043 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5044 /// This can cause recursive merging of nodes in the DAG.
5046 /// This version assumes that for each value of From, there is a
5047 /// corresponding value in To in the same position with the same type.
5049 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
5050 DAGUpdateListener *UpdateListener) {
5052 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
5053 assert((!From->hasAnyUseOfValue(i) ||
5054 From->getValueType(i) == To->getValueType(i)) &&
5055 "Cannot use this version of ReplaceAllUsesWith!");
5058 // Handle the trivial case.
5062 // Iterate over just the existing users of From. See the comments in
5063 // the ReplaceAllUsesWith above.
5064 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5065 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5069 // This node is about to morph, remove its old self from the CSE maps.
5070 RemoveNodeFromCSEMaps(User);
5072 // A user can appear in a use list multiple times, and when this
5073 // happens the uses are usually next to each other in the list.
5074 // To help reduce the number of CSE recomputations, process all
5075 // the uses of this user that we can find this way.
5077 SDUse &Use = UI.getUse();
5080 } while (UI != UE && *UI == User);
5082 // Now that we have modified User, add it back to the CSE maps. If it
5083 // already exists there, recursively merge the results together.
5084 AddModifiedNodeToCSEMaps(User, &Listener);
5088 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
5089 /// This can cause recursive merging of nodes in the DAG.
5091 /// This version can replace From with any result values. To must match the
5092 /// number and types of values returned by From.
5093 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
5095 DAGUpdateListener *UpdateListener) {
5096 if (From->getNumValues() == 1) // Handle the simple case efficiently.
5097 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
5099 // Iterate over just the existing users of From. See the comments in
5100 // the ReplaceAllUsesWith above.
5101 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
5102 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5106 // This node is about to morph, remove its old self from the CSE maps.
5107 RemoveNodeFromCSEMaps(User);
5109 // A user can appear in a use list multiple times, and when this
5110 // happens the uses are usually next to each other in the list.
5111 // To help reduce the number of CSE recomputations, process all
5112 // the uses of this user that we can find this way.
5114 SDUse &Use = UI.getUse();
5115 const SDValue &ToOp = To[Use.getResNo()];
5118 } while (UI != UE && *UI == User);
5120 // Now that we have modified User, add it back to the CSE maps. If it
5121 // already exists there, recursively merge the results together.
5122 AddModifiedNodeToCSEMaps(User, &Listener);
5126 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
5127 /// uses of other values produced by From.getNode() alone. The Deleted
5128 /// vector is handled the same way as for ReplaceAllUsesWith.
5129 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
5130 DAGUpdateListener *UpdateListener){
5131 // Handle the really simple, really trivial case efficiently.
5132 if (From == To) return;
5134 // Handle the simple, trivial, case efficiently.
5135 if (From.getNode()->getNumValues() == 1) {
5136 ReplaceAllUsesWith(From, To, UpdateListener);
5140 // Iterate over just the existing users of From. See the comments in
5141 // the ReplaceAllUsesWith above.
5142 SDNode::use_iterator UI = From.getNode()->use_begin(),
5143 UE = From.getNode()->use_end();
5144 RAUWUpdateListener Listener(UpdateListener, UI, UE);
5147 bool UserRemovedFromCSEMaps = false;
5149 // A user can appear in a use list multiple times, and when this
5150 // happens the uses are usually next to each other in the list.
5151 // To help reduce the number of CSE recomputations, process all
5152 // the uses of this user that we can find this way.
5154 SDUse &Use = UI.getUse();
5156 // Skip uses of different values from the same node.
5157 if (Use.getResNo() != From.getResNo()) {
5162 // If this node hasn't been modified yet, it's still in the CSE maps,
5163 // so remove its old self from the CSE maps.
5164 if (!UserRemovedFromCSEMaps) {
5165 RemoveNodeFromCSEMaps(User);
5166 UserRemovedFromCSEMaps = true;
5171 } while (UI != UE && *UI == User);
5173 // We are iterating over all uses of the From node, so if a use
5174 // doesn't use the specific value, no changes are made.
5175 if (!UserRemovedFromCSEMaps)
5178 // Now that we have modified User, add it back to the CSE maps. If it
5179 // already exists there, recursively merge the results together.
5180 AddModifiedNodeToCSEMaps(User, &Listener);
5185 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
5186 /// to record information about a use.
5193 /// operator< - Sort Memos by User.
5194 bool operator<(const UseMemo &L, const UseMemo &R) {
5195 return (intptr_t)L.User < (intptr_t)R.User;
5199 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
5200 /// uses of other values produced by From.getNode() alone. The same value
5201 /// may appear in both the From and To list. The Deleted vector is
5202 /// handled the same way as for ReplaceAllUsesWith.
5203 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
5206 DAGUpdateListener *UpdateListener){
5207 // Handle the simple, trivial case efficiently.
5209 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
5211 // Read up all the uses and make records of them. This helps
5212 // processing new uses that are introduced during the
5213 // replacement process.
5214 SmallVector<UseMemo, 4> Uses;
5215 for (unsigned i = 0; i != Num; ++i) {
5216 unsigned FromResNo = From[i].getResNo();
5217 SDNode *FromNode = From[i].getNode();
5218 for (SDNode::use_iterator UI = FromNode->use_begin(),
5219 E = FromNode->use_end(); UI != E; ++UI) {
5220 SDUse &Use = UI.getUse();
5221 if (Use.getResNo() == FromResNo) {
5222 UseMemo Memo = { *UI, i, &Use };
5223 Uses.push_back(Memo);
5228 // Sort the uses, so that all the uses from a given User are together.
5229 std::sort(Uses.begin(), Uses.end());
5231 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
5232 UseIndex != UseIndexEnd; ) {
5233 // We know that this user uses some value of From. If it is the right
5234 // value, update it.
5235 SDNode *User = Uses[UseIndex].User;
5237 // This node is about to morph, remove its old self from the CSE maps.
5238 RemoveNodeFromCSEMaps(User);
5240 // The Uses array is sorted, so all the uses for a given User
5241 // are next to each other in the list.
5242 // To help reduce the number of CSE recomputations, process all
5243 // the uses of this user that we can find this way.
5245 unsigned i = Uses[UseIndex].Index;
5246 SDUse &Use = *Uses[UseIndex].Use;
5250 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
5252 // Now that we have modified User, add it back to the CSE maps. If it
5253 // already exists there, recursively merge the results together.
5254 AddModifiedNodeToCSEMaps(User, UpdateListener);
5258 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
5259 /// based on their topological order. It returns the maximum id and a vector
5260 /// of the SDNodes* in assigned order by reference.
5261 unsigned SelectionDAG::AssignTopologicalOrder() {
5263 unsigned DAGSize = 0;
5265 // SortedPos tracks the progress of the algorithm. Nodes before it are
5266 // sorted, nodes after it are unsorted. When the algorithm completes
5267 // it is at the end of the list.
5268 allnodes_iterator SortedPos = allnodes_begin();
5270 // Visit all the nodes. Move nodes with no operands to the front of
5271 // the list immediately. Annotate nodes that do have operands with their
5272 // operand count. Before we do this, the Node Id fields of the nodes
5273 // may contain arbitrary values. After, the Node Id fields for nodes
5274 // before SortedPos will contain the topological sort index, and the
5275 // Node Id fields for nodes At SortedPos and after will contain the
5276 // count of outstanding operands.
5277 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
5280 unsigned Degree = N->getNumOperands();
5282 // A node with no uses, add it to the result array immediately.
5283 N->setNodeId(DAGSize++);
5284 allnodes_iterator Q = N;
5286 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
5287 assert(SortedPos != AllNodes.end() && "Overran node list");
5290 // Temporarily use the Node Id as scratch space for the degree count.
5291 N->setNodeId(Degree);
5295 // Visit all the nodes. As we iterate, moves nodes into sorted order,
5296 // such that by the time the end is reached all nodes will be sorted.
5297 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
5300 // N is in sorted position, so all its uses have one less operand
5301 // that needs to be sorted.
5302 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5305 unsigned Degree = P->getNodeId();
5306 assert(Degree != 0 && "Invalid node degree");
5309 // All of P's operands are sorted, so P may sorted now.
5310 P->setNodeId(DAGSize++);
5312 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
5313 assert(SortedPos != AllNodes.end() && "Overran node list");
5316 // Update P's outstanding operand count.
5317 P->setNodeId(Degree);
5320 if (I == SortedPos) {
5323 dbgs() << "Overran sorted position:\n";
5326 llvm_unreachable(0);
5330 assert(SortedPos == AllNodes.end() &&
5331 "Topological sort incomplete!");
5332 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
5333 "First node in topological sort is not the entry token!");
5334 assert(AllNodes.front().getNodeId() == 0 &&
5335 "First node in topological sort has non-zero id!");
5336 assert(AllNodes.front().getNumOperands() == 0 &&
5337 "First node in topological sort has operands!");
5338 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
5339 "Last node in topologic sort has unexpected id!");
5340 assert(AllNodes.back().use_empty() &&
5341 "Last node in topologic sort has users!");
5342 assert(DAGSize == allnodes_size() && "Node count mismatch!");
5346 /// AssignOrdering - Assign an order to the SDNode.
5347 void SelectionDAG::AssignOrdering(const SDNode *SD, unsigned Order) {
5348 assert(SD && "Trying to assign an order to a null node!");
5349 Ordering->add(SD, Order);
5352 /// GetOrdering - Get the order for the SDNode.
5353 unsigned SelectionDAG::GetOrdering(const SDNode *SD) const {
5354 assert(SD && "Trying to get the order of a null node!");
5355 return Ordering->getOrder(SD);
5358 /// AddDbgValue - Add a dbg_value SDNode. If SD is non-null that means the
5359 /// value is produced by SD.
5360 void SelectionDAG::AddDbgValue(SDDbgValue *DB, SDNode *SD, bool isParameter) {
5361 DbgInfo->add(DB, SD, isParameter);
5363 SD->setHasDebugValue(true);
5366 //===----------------------------------------------------------------------===//
5368 //===----------------------------------------------------------------------===//
5370 HandleSDNode::~HandleSDNode() {
5374 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA,
5375 EVT VT, int64_t o, unsigned char TF)
5376 : SDNode(Opc, DebugLoc(), getSDVTList(VT)), Offset(o), TargetFlags(TF) {
5380 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt,
5381 MachineMemOperand *mmo)
5382 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) {
5383 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(),
5384 MMO->isNonTemporal());
5385 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5386 assert(isNonTemporal() == MMO->isNonTemporal() &&
5387 "Non-temporal encoding error!");
5388 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5391 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs,
5392 const SDValue *Ops, unsigned NumOps, EVT memvt,
5393 MachineMemOperand *mmo)
5394 : SDNode(Opc, dl, VTs, Ops, NumOps),
5395 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(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5402 /// Profile - Gather unique data for the node.
5404 void SDNode::Profile(FoldingSetNodeID &ID) const {
5405 AddNodeIDNode(ID, this);
5410 std::vector<EVT> VTs;
5413 VTs.reserve(MVT::LAST_VALUETYPE);
5414 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i)
5415 VTs.push_back(MVT((MVT::SimpleValueType)i));
5420 static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs;
5421 static ManagedStatic<EVTArray> SimpleVTArray;
5422 static ManagedStatic<sys::SmartMutex<true> > VTMutex;
5424 /// getValueTypeList - Return a pointer to the specified value type.
5426 const EVT *SDNode::getValueTypeList(EVT VT) {
5427 if (VT.isExtended()) {
5428 sys::SmartScopedLock<true> Lock(*VTMutex);
5429 return &(*EVTs->insert(VT).first);
5431 assert(VT.getSimpleVT().SimpleTy < MVT::LAST_VALUETYPE &&
5432 "Value type out of range!");
5433 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy];
5437 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
5438 /// indicated value. This method ignores uses of other values defined by this
5440 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
5441 assert(Value < getNumValues() && "Bad value!");
5443 // TODO: Only iterate over uses of a given value of the node
5444 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
5445 if (UI.getUse().getResNo() == Value) {
5452 // Found exactly the right number of uses?
5457 /// hasAnyUseOfValue - Return true if there are any use of the indicated
5458 /// value. This method ignores uses of other values defined by this operation.
5459 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
5460 assert(Value < getNumValues() && "Bad value!");
5462 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
5463 if (UI.getUse().getResNo() == Value)
5470 /// isOnlyUserOf - Return true if this node is the only use of N.
5472 bool SDNode::isOnlyUserOf(SDNode *N) const {
5474 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
5485 /// isOperand - Return true if this node is an operand of N.
5487 bool SDValue::isOperandOf(SDNode *N) const {
5488 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5489 if (*this == N->getOperand(i))
5494 bool SDNode::isOperandOf(SDNode *N) const {
5495 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
5496 if (this == N->OperandList[i].getNode())
5501 /// reachesChainWithoutSideEffects - Return true if this operand (which must
5502 /// be a chain) reaches the specified operand without crossing any
5503 /// side-effecting instructions. In practice, this looks through token
5504 /// factors and non-volatile loads. In order to remain efficient, this only
5505 /// looks a couple of nodes in, it does not do an exhaustive search.
5506 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
5507 unsigned Depth) const {
5508 if (*this == Dest) return true;
5510 // Don't search too deeply, we just want to be able to see through
5511 // TokenFactor's etc.
5512 if (Depth == 0) return false;
5514 // If this is a token factor, all inputs to the TF happen in parallel. If any
5515 // of the operands of the TF reach dest, then we can do the xform.
5516 if (getOpcode() == ISD::TokenFactor) {
5517 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5518 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
5523 // Loads don't have side effects, look through them.
5524 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
5525 if (!Ld->isVolatile())
5526 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
5531 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
5532 /// is either an operand of N or it can be reached by traversing up the operands.
5533 /// NOTE: this is an expensive method. Use it carefully.
5534 bool SDNode::isPredecessorOf(SDNode *N) const {
5535 SmallPtrSet<SDNode *, 32> Visited;
5536 SmallVector<SDNode *, 16> Worklist;
5537 Worklist.push_back(N);
5540 N = Worklist.pop_back_val();
5541 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5542 SDNode *Op = N->getOperand(i).getNode();
5545 if (Visited.insert(Op))
5546 Worklist.push_back(Op);
5548 } while (!Worklist.empty());
5553 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
5554 assert(Num < NumOperands && "Invalid child # of SDNode!");
5555 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
5558 std::string SDNode::getOperationName(const SelectionDAG *G) const {
5559 switch (getOpcode()) {
5561 if (getOpcode() < ISD::BUILTIN_OP_END)
5562 return "<<Unknown DAG Node>>";
5563 if (isMachineOpcode()) {
5565 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
5566 if (getMachineOpcode() < TII->getNumOpcodes())
5567 return TII->get(getMachineOpcode()).getName();
5568 return "<<Unknown Machine Node #" + utostr(getOpcode()) + ">>";
5571 const TargetLowering &TLI = G->getTargetLoweringInfo();
5572 const char *Name = TLI.getTargetNodeName(getOpcode());
5573 if (Name) return Name;
5574 return "<<Unknown Target Node #" + utostr(getOpcode()) + ">>";
5576 return "<<Unknown Node #" + utostr(getOpcode()) + ">>";
5579 case ISD::DELETED_NODE:
5580 return "<<Deleted Node!>>";
5582 case ISD::PREFETCH: return "Prefetch";
5583 case ISD::MEMBARRIER: return "MemBarrier";
5584 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap";
5585 case ISD::ATOMIC_SWAP: return "AtomicSwap";
5586 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd";
5587 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub";
5588 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd";
5589 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr";
5590 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor";
5591 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand";
5592 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin";
5593 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax";
5594 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin";
5595 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax";
5596 case ISD::PCMARKER: return "PCMarker";
5597 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5598 case ISD::SRCVALUE: return "SrcValue";
5599 case ISD::MDNODE_SDNODE: return "MDNode";
5600 case ISD::EntryToken: return "EntryToken";
5601 case ISD::TokenFactor: return "TokenFactor";
5602 case ISD::AssertSext: return "AssertSext";
5603 case ISD::AssertZext: return "AssertZext";
5605 case ISD::BasicBlock: return "BasicBlock";
5606 case ISD::VALUETYPE: return "ValueType";
5607 case ISD::Register: return "Register";
5609 case ISD::Constant: return "Constant";
5610 case ISD::ConstantFP: return "ConstantFP";
5611 case ISD::GlobalAddress: return "GlobalAddress";
5612 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5613 case ISD::FrameIndex: return "FrameIndex";
5614 case ISD::JumpTable: return "JumpTable";
5615 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5616 case ISD::RETURNADDR: return "RETURNADDR";
5617 case ISD::FRAMEADDR: return "FRAMEADDR";
5618 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5619 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5620 case ISD::LSDAADDR: return "LSDAADDR";
5621 case ISD::EHSELECTION: return "EHSELECTION";
5622 case ISD::EH_RETURN: return "EH_RETURN";
5623 case ISD::ConstantPool: return "ConstantPool";
5624 case ISD::ExternalSymbol: return "ExternalSymbol";
5625 case ISD::BlockAddress: return "BlockAddress";
5626 case ISD::INTRINSIC_WO_CHAIN:
5627 case ISD::INTRINSIC_VOID:
5628 case ISD::INTRINSIC_W_CHAIN: {
5629 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1;
5630 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue();
5631 if (IID < Intrinsic::num_intrinsics)
5632 return Intrinsic::getName((Intrinsic::ID)IID);
5633 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo())
5634 return TII->getName(IID);
5635 llvm_unreachable("Invalid intrinsic ID");
5638 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5639 case ISD::TargetConstant: return "TargetConstant";
5640 case ISD::TargetConstantFP:return "TargetConstantFP";
5641 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5642 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5643 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5644 case ISD::TargetJumpTable: return "TargetJumpTable";
5645 case ISD::TargetConstantPool: return "TargetConstantPool";
5646 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5647 case ISD::TargetBlockAddress: return "TargetBlockAddress";
5649 case ISD::CopyToReg: return "CopyToReg";
5650 case ISD::CopyFromReg: return "CopyFromReg";
5651 case ISD::UNDEF: return "undef";
5652 case ISD::MERGE_VALUES: return "merge_values";
5653 case ISD::INLINEASM: return "inlineasm";
5654 case ISD::EH_LABEL: return "eh_label";
5655 case ISD::HANDLENODE: return "handlenode";
5658 case ISD::FABS: return "fabs";
5659 case ISD::FNEG: return "fneg";
5660 case ISD::FSQRT: return "fsqrt";
5661 case ISD::FSIN: return "fsin";
5662 case ISD::FCOS: return "fcos";
5663 case ISD::FPOWI: return "fpowi";
5664 case ISD::FPOW: return "fpow";
5665 case ISD::FTRUNC: return "ftrunc";
5666 case ISD::FFLOOR: return "ffloor";
5667 case ISD::FCEIL: return "fceil";
5668 case ISD::FRINT: return "frint";
5669 case ISD::FNEARBYINT: return "fnearbyint";
5672 case ISD::ADD: return "add";
5673 case ISD::SUB: return "sub";
5674 case ISD::MUL: return "mul";
5675 case ISD::MULHU: return "mulhu";
5676 case ISD::MULHS: return "mulhs";
5677 case ISD::SDIV: return "sdiv";
5678 case ISD::UDIV: return "udiv";
5679 case ISD::SREM: return "srem";
5680 case ISD::UREM: return "urem";
5681 case ISD::SMUL_LOHI: return "smul_lohi";
5682 case ISD::UMUL_LOHI: return "umul_lohi";
5683 case ISD::SDIVREM: return "sdivrem";
5684 case ISD::UDIVREM: return "udivrem";
5685 case ISD::AND: return "and";
5686 case ISD::OR: return "or";
5687 case ISD::XOR: return "xor";
5688 case ISD::SHL: return "shl";
5689 case ISD::SRA: return "sra";
5690 case ISD::SRL: return "srl";
5691 case ISD::ROTL: return "rotl";
5692 case ISD::ROTR: return "rotr";
5693 case ISD::FADD: return "fadd";
5694 case ISD::FSUB: return "fsub";
5695 case ISD::FMUL: return "fmul";
5696 case ISD::FDIV: return "fdiv";
5697 case ISD::FREM: return "frem";
5698 case ISD::FCOPYSIGN: return "fcopysign";
5699 case ISD::FGETSIGN: return "fgetsign";
5701 case ISD::SETCC: return "setcc";
5702 case ISD::VSETCC: return "vsetcc";
5703 case ISD::SELECT: return "select";
5704 case ISD::SELECT_CC: return "select_cc";
5705 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5706 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5707 case ISD::CONCAT_VECTORS: return "concat_vectors";
5708 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5709 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5710 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5711 case ISD::CARRY_FALSE: return "carry_false";
5712 case ISD::ADDC: return "addc";
5713 case ISD::ADDE: return "adde";
5714 case ISD::SADDO: return "saddo";
5715 case ISD::UADDO: return "uaddo";
5716 case ISD::SSUBO: return "ssubo";
5717 case ISD::USUBO: return "usubo";
5718 case ISD::SMULO: return "smulo";
5719 case ISD::UMULO: return "umulo";
5720 case ISD::SUBC: return "subc";
5721 case ISD::SUBE: return "sube";
5722 case ISD::SHL_PARTS: return "shl_parts";
5723 case ISD::SRA_PARTS: return "sra_parts";
5724 case ISD::SRL_PARTS: return "srl_parts";
5726 // Conversion operators.
5727 case ISD::SIGN_EXTEND: return "sign_extend";
5728 case ISD::ZERO_EXTEND: return "zero_extend";
5729 case ISD::ANY_EXTEND: return "any_extend";
5730 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5731 case ISD::TRUNCATE: return "truncate";
5732 case ISD::FP_ROUND: return "fp_round";
5733 case ISD::FLT_ROUNDS_: return "flt_rounds";
5734 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5735 case ISD::FP_EXTEND: return "fp_extend";
5737 case ISD::SINT_TO_FP: return "sint_to_fp";
5738 case ISD::UINT_TO_FP: return "uint_to_fp";
5739 case ISD::FP_TO_SINT: return "fp_to_sint";
5740 case ISD::FP_TO_UINT: return "fp_to_uint";
5741 case ISD::BIT_CONVERT: return "bit_convert";
5742 case ISD::FP16_TO_FP32: return "fp16_to_fp32";
5743 case ISD::FP32_TO_FP16: return "fp32_to_fp16";
5745 case ISD::CONVERT_RNDSAT: {
5746 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) {
5747 default: llvm_unreachable("Unknown cvt code!");
5748 case ISD::CVT_FF: return "cvt_ff";
5749 case ISD::CVT_FS: return "cvt_fs";
5750 case ISD::CVT_FU: return "cvt_fu";
5751 case ISD::CVT_SF: return "cvt_sf";
5752 case ISD::CVT_UF: return "cvt_uf";
5753 case ISD::CVT_SS: return "cvt_ss";
5754 case ISD::CVT_SU: return "cvt_su";
5755 case ISD::CVT_US: return "cvt_us";
5756 case ISD::CVT_UU: return "cvt_uu";
5760 // Control flow instructions
5761 case ISD::BR: return "br";
5762 case ISD::BRIND: return "brind";
5763 case ISD::BR_JT: return "br_jt";
5764 case ISD::BRCOND: return "brcond";
5765 case ISD::BR_CC: return "br_cc";
5766 case ISD::CALLSEQ_START: return "callseq_start";
5767 case ISD::CALLSEQ_END: return "callseq_end";
5770 case ISD::LOAD: return "load";
5771 case ISD::STORE: return "store";
5772 case ISD::VAARG: return "vaarg";
5773 case ISD::VACOPY: return "vacopy";
5774 case ISD::VAEND: return "vaend";
5775 case ISD::VASTART: return "vastart";
5776 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5777 case ISD::EXTRACT_ELEMENT: return "extract_element";
5778 case ISD::BUILD_PAIR: return "build_pair";
5779 case ISD::STACKSAVE: return "stacksave";
5780 case ISD::STACKRESTORE: return "stackrestore";
5781 case ISD::TRAP: return "trap";
5784 case ISD::BSWAP: return "bswap";
5785 case ISD::CTPOP: return "ctpop";
5786 case ISD::CTTZ: return "cttz";
5787 case ISD::CTLZ: return "ctlz";
5790 case ISD::TRAMPOLINE: return "trampoline";
5793 switch (cast<CondCodeSDNode>(this)->get()) {
5794 default: llvm_unreachable("Unknown setcc condition!");
5795 case ISD::SETOEQ: return "setoeq";
5796 case ISD::SETOGT: return "setogt";
5797 case ISD::SETOGE: return "setoge";
5798 case ISD::SETOLT: return "setolt";
5799 case ISD::SETOLE: return "setole";
5800 case ISD::SETONE: return "setone";
5802 case ISD::SETO: return "seto";
5803 case ISD::SETUO: return "setuo";
5804 case ISD::SETUEQ: return "setue";
5805 case ISD::SETUGT: return "setugt";
5806 case ISD::SETUGE: return "setuge";
5807 case ISD::SETULT: return "setult";
5808 case ISD::SETULE: return "setule";
5809 case ISD::SETUNE: return "setune";
5811 case ISD::SETEQ: return "seteq";
5812 case ISD::SETGT: return "setgt";
5813 case ISD::SETGE: return "setge";
5814 case ISD::SETLT: return "setlt";
5815 case ISD::SETLE: return "setle";
5816 case ISD::SETNE: return "setne";
5821 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5830 return "<post-inc>";
5832 return "<post-dec>";
5836 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5837 std::string S = "< ";
5851 if (getByValAlign())
5852 S += "byval-align:" + utostr(getByValAlign()) + " ";
5854 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5856 S += "byval-size:" + utostr(getByValSize()) + " ";
5860 void SDNode::dump() const { dump(0); }
5861 void SDNode::dump(const SelectionDAG *G) const {
5865 void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const {
5866 OS << (void*)this << ": ";
5868 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5870 if (getValueType(i) == MVT::Other)
5873 OS << getValueType(i).getEVTString();
5875 OS << " = " << getOperationName(G);
5878 void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const {
5879 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) {
5880 if (!MN->memoperands_empty()) {
5883 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(),
5884 e = MN->memoperands_end(); i != e; ++i) {
5891 } else if (const ShuffleVectorSDNode *SVN =
5892 dyn_cast<ShuffleVectorSDNode>(this)) {
5894 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) {
5895 int Idx = SVN->getMaskElt(i);
5903 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5904 OS << '<' << CSDN->getAPIntValue() << '>';
5905 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5906 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5907 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5908 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5909 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5912 CSDN->getValueAPF().bitcastToAPInt().dump();
5915 } else if (const GlobalAddressSDNode *GADN =
5916 dyn_cast<GlobalAddressSDNode>(this)) {
5917 int64_t offset = GADN->getOffset();
5919 WriteAsOperand(OS, GADN->getGlobal());
5922 OS << " + " << offset;
5924 OS << " " << offset;
5925 if (unsigned int TF = GADN->getTargetFlags())
5926 OS << " [TF=" << TF << ']';
5927 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5928 OS << "<" << FIDN->getIndex() << ">";
5929 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5930 OS << "<" << JTDN->getIndex() << ">";
5931 if (unsigned int TF = JTDN->getTargetFlags())
5932 OS << " [TF=" << TF << ']';
5933 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5934 int offset = CP->getOffset();
5935 if (CP->isMachineConstantPoolEntry())
5936 OS << "<" << *CP->getMachineCPVal() << ">";
5938 OS << "<" << *CP->getConstVal() << ">";
5940 OS << " + " << offset;
5942 OS << " " << offset;
5943 if (unsigned int TF = CP->getTargetFlags())
5944 OS << " [TF=" << TF << ']';
5945 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5947 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5949 OS << LBB->getName() << " ";
5950 OS << (const void*)BBDN->getBasicBlock() << ">";
5951 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5952 if (G && R->getReg() &&
5953 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5954 OS << " %" << G->getTarget().getRegisterInfo()->getName(R->getReg());
5956 OS << " %reg" << R->getReg();
5958 } else if (const ExternalSymbolSDNode *ES =
5959 dyn_cast<ExternalSymbolSDNode>(this)) {
5960 OS << "'" << ES->getSymbol() << "'";
5961 if (unsigned int TF = ES->getTargetFlags())
5962 OS << " [TF=" << TF << ']';
5963 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5965 OS << "<" << M->getValue() << ">";
5968 } else if (const MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(this)) {
5970 OS << "<" << MD->getMD() << ">";
5973 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5974 OS << ":" << N->getVT().getEVTString();
5976 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5977 OS << "<" << *LD->getMemOperand();
5980 switch (LD->getExtensionType()) {
5981 default: doExt = false; break;
5982 case ISD::EXTLOAD: OS << ", anyext"; break;
5983 case ISD::SEXTLOAD: OS << ", sext"; break;
5984 case ISD::ZEXTLOAD: OS << ", zext"; break;
5987 OS << " from " << LD->getMemoryVT().getEVTString();
5989 const char *AM = getIndexedModeName(LD->getAddressingMode());
5994 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
5995 OS << "<" << *ST->getMemOperand();
5997 if (ST->isTruncatingStore())
5998 OS << ", trunc to " << ST->getMemoryVT().getEVTString();
6000 const char *AM = getIndexedModeName(ST->getAddressingMode());
6005 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) {
6006 OS << "<" << *M->getMemOperand() << ">";
6007 } else if (const BlockAddressSDNode *BA =
6008 dyn_cast<BlockAddressSDNode>(this)) {
6010 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false);
6012 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false);
6014 if (unsigned int TF = BA->getTargetFlags())
6015 OS << " [TF=" << TF << ']';
6019 if (unsigned Order = G->GetOrdering(this))
6020 OS << " [ORD=" << Order << ']';
6022 if (getNodeId() != -1)
6023 OS << " [ID=" << getNodeId() << ']';
6025 DebugLoc dl = getDebugLoc();
6026 if (G && !dl.isUnknown()) {
6028 Scope(dl.getScope(G->getMachineFunction().getFunction()->getContext()));
6030 // Omit the directory, since it's usually long and uninteresting.
6032 OS << Scope.getFilename();
6035 OS << ':' << dl.getLine();
6036 if (dl.getCol() != 0)
6037 OS << ':' << dl.getCol();
6041 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
6043 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
6044 if (i) OS << ", "; else OS << " ";
6045 OS << (void*)getOperand(i).getNode();
6046 if (unsigned RN = getOperand(i).getResNo())
6049 print_details(OS, G);
6052 static void printrWithDepthHelper(raw_ostream &OS, const SDNode *N,
6053 const SelectionDAG *G, unsigned depth,
6066 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6068 printrWithDepthHelper(OS, N->getOperand(i).getNode(), G, depth-1, indent+2);
6072 void SDNode::printrWithDepth(raw_ostream &OS, const SelectionDAG *G,
6073 unsigned depth) const {
6074 printrWithDepthHelper(OS, this, G, depth, 0);
6077 void SDNode::printrFull(raw_ostream &OS, const SelectionDAG *G) const {
6078 // Don't print impossibly deep things.
6079 printrWithDepth(OS, G, 100);
6082 void SDNode::dumprWithDepth(const SelectionDAG *G, unsigned depth) const {
6083 printrWithDepth(dbgs(), G, depth);
6086 void SDNode::dumprFull(const SelectionDAG *G) const {
6087 // Don't print impossibly deep things.
6088 dumprWithDepth(G, 100);
6091 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
6092 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
6093 if (N->getOperand(i).getNode()->hasOneUse())
6094 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
6096 dbgs() << "\n" << std::string(indent+2, ' ')
6097 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
6101 dbgs().indent(indent);
6105 SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) {
6106 assert(N->getNumValues() == 1 &&
6107 "Can't unroll a vector with multiple results!");
6109 EVT VT = N->getValueType(0);
6110 unsigned NE = VT.getVectorNumElements();
6111 EVT EltVT = VT.getVectorElementType();
6112 DebugLoc dl = N->getDebugLoc();
6114 SmallVector<SDValue, 8> Scalars;
6115 SmallVector<SDValue, 4> Operands(N->getNumOperands());
6117 // If ResNE is 0, fully unroll the vector op.
6120 else if (NE > ResNE)
6124 for (i= 0; i != NE; ++i) {
6125 for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) {
6126 SDValue Operand = N->getOperand(j);
6127 EVT OperandVT = Operand.getValueType();
6128 if (OperandVT.isVector()) {
6129 // A vector operand; extract a single element.
6130 EVT OperandEltVT = OperandVT.getVectorElementType();
6131 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl,
6134 getConstant(i, MVT::i32));
6136 // A scalar operand; just use it as is.
6137 Operands[j] = Operand;
6141 switch (N->getOpcode()) {
6143 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
6144 &Operands[0], Operands.size()));
6151 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0],
6152 getShiftAmountOperand(Operands[1])));
6154 case ISD::SIGN_EXTEND_INREG:
6155 case ISD::FP_ROUND_INREG: {
6156 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType();
6157 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
6159 getValueType(ExtVT)));
6164 for (; i < ResNE; ++i)
6165 Scalars.push_back(getUNDEF(EltVT));
6167 return getNode(ISD::BUILD_VECTOR, dl,
6168 EVT::getVectorVT(*getContext(), EltVT, ResNE),
6169 &Scalars[0], Scalars.size());
6173 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
6174 /// location that is 'Dist' units away from the location that the 'Base' load
6175 /// is loading from.
6176 bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
6177 unsigned Bytes, int Dist) const {
6178 if (LD->getChain() != Base->getChain())
6180 EVT VT = LD->getValueType(0);
6181 if (VT.getSizeInBits() / 8 != Bytes)
6184 SDValue Loc = LD->getOperand(1);
6185 SDValue BaseLoc = Base->getOperand(1);
6186 if (Loc.getOpcode() == ISD::FrameIndex) {
6187 if (BaseLoc.getOpcode() != ISD::FrameIndex)
6189 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo();
6190 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
6191 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
6192 int FS = MFI->getObjectSize(FI);
6193 int BFS = MFI->getObjectSize(BFI);
6194 if (FS != BFS || FS != (int)Bytes) return false;
6195 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
6197 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
6198 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
6199 if (V && (V->getSExtValue() == Dist*Bytes))
6203 const GlobalValue *GV1 = NULL;
6204 const GlobalValue *GV2 = NULL;
6205 int64_t Offset1 = 0;
6206 int64_t Offset2 = 0;
6207 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
6208 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
6209 if (isGA1 && isGA2 && GV1 == GV2)
6210 return Offset1 == (Offset2 + Dist*Bytes);
6215 /// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if
6216 /// it cannot be inferred.
6217 unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const {
6218 // If this is a GlobalAddress + cst, return the alignment.
6219 const GlobalValue *GV;
6220 int64_t GVOffset = 0;
6221 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) {
6222 // If GV has specified alignment, then use it. Otherwise, use the preferred
6224 unsigned Align = GV->getAlignment();
6226 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
6227 if (GVar->hasInitializer()) {
6228 const TargetData *TD = TLI.getTargetData();
6229 Align = TD->getPreferredAlignment(GVar);
6233 return MinAlign(Align, GVOffset);
6236 // If this is a direct reference to a stack slot, use information about the
6237 // stack slot's alignment.
6238 int FrameIdx = 1 << 31;
6239 int64_t FrameOffset = 0;
6240 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
6241 FrameIdx = FI->getIndex();
6242 } else if (Ptr.getOpcode() == ISD::ADD &&
6243 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
6244 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
6245 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
6246 FrameOffset = Ptr.getConstantOperandVal(1);
6249 if (FrameIdx != (1 << 31)) {
6250 // FIXME: Handle FI+CST.
6251 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo();
6252 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
6254 if (MFI.isFixedObjectIndex(FrameIdx)) {
6255 int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx) + FrameOffset;
6257 // The alignment of the frame index can be determined from its offset from
6258 // the incoming frame position. If the frame object is at offset 32 and
6259 // the stack is guaranteed to be 16-byte aligned, then we know that the
6260 // object is 16-byte aligned.
6261 unsigned StackAlign = getTarget().getFrameInfo()->getStackAlignment();
6262 unsigned Align = MinAlign(ObjectOffset, StackAlign);
6264 // Finally, the frame object itself may have a known alignment. Factor
6265 // the alignment + offset into a new alignment. For example, if we know
6266 // the FI is 8 byte aligned, but the pointer is 4 off, we really have a
6267 // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte
6268 // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc.
6269 return std::max(Align, FIInfoAlign);
6277 void SelectionDAG::dump() const {
6278 dbgs() << "SelectionDAG has " << AllNodes.size() << " nodes:";
6280 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
6282 const SDNode *N = I;
6283 if (!N->hasOneUse() && N != getRoot().getNode())
6284 DumpNodes(N, 2, this);
6287 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
6292 void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
6294 print_details(OS, G);
6297 typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet;
6298 static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent,
6299 const SelectionDAG *G, VisitedSDNodeSet &once) {
6300 if (!once.insert(N)) // If we've been here before, return now.
6303 // Dump the current SDNode, but don't end the line yet.
6304 OS << std::string(indent, ' ');
6307 // Having printed this SDNode, walk the children:
6308 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6309 const SDNode *child = N->getOperand(i).getNode();
6314 if (child->getNumOperands() == 0) {
6315 // This child has no grandchildren; print it inline right here.
6316 child->printr(OS, G);
6318 } else { // Just the address. FIXME: also print the child's opcode.
6320 if (unsigned RN = N->getOperand(i).getResNo())
6327 // Dump children that have grandchildren on their own line(s).
6328 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6329 const SDNode *child = N->getOperand(i).getNode();
6330 DumpNodesr(OS, child, indent+2, G, once);
6334 void SDNode::dumpr() const {
6335 VisitedSDNodeSet once;
6336 DumpNodesr(dbgs(), this, 0, 0, once);
6339 void SDNode::dumpr(const SelectionDAG *G) const {
6340 VisitedSDNodeSet once;
6341 DumpNodesr(dbgs(), this, 0, G, once);
6345 // getAddressSpace - Return the address space this GlobalAddress belongs to.
6346 unsigned GlobalAddressSDNode::getAddressSpace() const {
6347 return getGlobal()->getType()->getAddressSpace();
6351 const Type *ConstantPoolSDNode::getType() const {
6352 if (isMachineConstantPoolEntry())
6353 return Val.MachineCPVal->getType();
6354 return Val.ConstVal->getType();
6357 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue,
6359 unsigned &SplatBitSize,
6361 unsigned MinSplatBits,
6363 EVT VT = getValueType(0);
6364 assert(VT.isVector() && "Expected a vector type");
6365 unsigned sz = VT.getSizeInBits();
6366 if (MinSplatBits > sz)
6369 SplatValue = APInt(sz, 0);
6370 SplatUndef = APInt(sz, 0);
6372 // Get the bits. Bits with undefined values (when the corresponding element
6373 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
6374 // in SplatValue. If any of the values are not constant, give up and return
6376 unsigned int nOps = getNumOperands();
6377 assert(nOps > 0 && "isConstantSplat has 0-size build vector");
6378 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits();
6380 for (unsigned j = 0; j < nOps; ++j) {
6381 unsigned i = isBigEndian ? nOps-1-j : j;
6382 SDValue OpVal = getOperand(i);
6383 unsigned BitPos = j * EltBitSize;
6385 if (OpVal.getOpcode() == ISD::UNDEF)
6386 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize);
6387 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal))
6388 SplatValue |= APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize).
6389 zextOrTrunc(sz) << BitPos;
6390 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal))
6391 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos;
6396 // The build_vector is all constants or undefs. Find the smallest element
6397 // size that splats the vector.
6399 HasAnyUndefs = (SplatUndef != 0);
6402 unsigned HalfSize = sz / 2;
6403 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize);
6404 APInt LowValue = APInt(SplatValue).trunc(HalfSize);
6405 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize);
6406 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize);
6408 // If the two halves do not match (ignoring undef bits), stop here.
6409 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
6410 MinSplatBits > HalfSize)
6413 SplatValue = HighValue | LowValue;
6414 SplatUndef = HighUndef & LowUndef;
6423 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) {
6424 // Find the first non-undef value in the shuffle mask.
6426 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
6429 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!");
6431 // Make sure all remaining elements are either undef or the same as the first
6433 for (int Idx = Mask[i]; i != e; ++i)
6434 if (Mask[i] >= 0 && Mask[i] != Idx)
6440 static void checkForCyclesHelper(const SDNode *N,
6441 SmallPtrSet<const SDNode*, 32> &Visited,
6442 SmallPtrSet<const SDNode*, 32> &Checked) {
6443 // If this node has already been checked, don't check it again.
6444 if (Checked.count(N))
6447 // If a node has already been visited on this depth-first walk, reject it as
6449 if (!Visited.insert(N)) {
6450 dbgs() << "Offending node:\n";
6452 errs() << "Detected cycle in SelectionDAG\n";
6456 for(unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
6457 checkForCyclesHelper(N->getOperand(i).getNode(), Visited, Checked);
6464 void llvm::checkForCycles(const llvm::SDNode *N) {
6466 assert(N && "Checking nonexistant SDNode");
6467 SmallPtrSet<const SDNode*, 32> visited;
6468 SmallPtrSet<const SDNode*, 32> checked;
6469 checkForCyclesHelper(N, visited, checked);
6473 void llvm::checkForCycles(const llvm::SelectionDAG *DAG) {
6474 checkForCycles(DAG->getRoot().getNode());