1 //===-- SelectionDAG.cpp - Implement the SelectionDAG data structures -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAG class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "SDNodeOrdering.h"
16 #include "llvm/Constants.h"
17 #include "llvm/Analysis/ValueTracking.h"
18 #include "llvm/Function.h"
19 #include "llvm/GlobalAlias.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Assembly/Writer.h"
24 #include "llvm/CallingConv.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/PseudoSourceValue.h"
30 #include "llvm/Target/TargetRegisterInfo.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/Target/TargetFrameInfo.h"
33 #include "llvm/Target/TargetLowering.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetIntrinsicInfo.h"
37 #include "llvm/Target/TargetMachine.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/ManagedStatic.h"
42 #include "llvm/Support/MathExtras.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/System/Mutex.h"
45 #include "llvm/ADT/SetVector.h"
46 #include "llvm/ADT/SmallPtrSet.h"
47 #include "llvm/ADT/SmallSet.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/StringExtras.h"
54 /// makeVTList - Return an instance of the SDVTList struct initialized with the
55 /// specified members.
56 static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
57 SDVTList Res = {VTs, NumVTs};
61 static const fltSemantics *EVTToAPFloatSemantics(EVT VT) {
62 switch (VT.getSimpleVT().SimpleTy) {
63 default: llvm_unreachable("Unknown FP format");
64 case MVT::f32: return &APFloat::IEEEsingle;
65 case MVT::f64: return &APFloat::IEEEdouble;
66 case MVT::f80: return &APFloat::x87DoubleExtended;
67 case MVT::f128: return &APFloat::IEEEquad;
68 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
72 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
74 //===----------------------------------------------------------------------===//
75 // ConstantFPSDNode Class
76 //===----------------------------------------------------------------------===//
78 /// isExactlyValue - We don't rely on operator== working on double values, as
79 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
80 /// As such, this method can be used to do an exact bit-for-bit comparison of
81 /// two floating point values.
82 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
83 return getValueAPF().bitwiseIsEqual(V);
86 bool ConstantFPSDNode::isValueValidForType(EVT VT,
88 assert(VT.isFloatingPoint() && "Can only convert between FP types");
90 // PPC long double cannot be converted to any other type.
91 if (VT == MVT::ppcf128 ||
92 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
95 // convert modifies in place, so make a copy.
96 APFloat Val2 = APFloat(Val);
98 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
103 //===----------------------------------------------------------------------===//
105 //===----------------------------------------------------------------------===//
107 /// isBuildVectorAllOnes - Return true if the specified node is a
108 /// BUILD_VECTOR where all of the elements are ~0 or undef.
109 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
110 // Look through a bit convert.
111 if (N->getOpcode() == ISD::BIT_CONVERT)
112 N = N->getOperand(0).getNode();
114 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
116 unsigned i = 0, e = N->getNumOperands();
118 // Skip over all of the undef values.
119 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
122 // Do not accept an all-undef vector.
123 if (i == e) return false;
125 // Do not accept build_vectors that aren't all constants or which have non-~0
127 SDValue NotZero = N->getOperand(i);
128 if (isa<ConstantSDNode>(NotZero)) {
129 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
131 } else if (isa<ConstantFPSDNode>(NotZero)) {
132 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
133 bitcastToAPInt().isAllOnesValue())
138 // Okay, we have at least one ~0 value, check to see if the rest match or are
140 for (++i; i != e; ++i)
141 if (N->getOperand(i) != NotZero &&
142 N->getOperand(i).getOpcode() != ISD::UNDEF)
148 /// isBuildVectorAllZeros - Return true if the specified node is a
149 /// BUILD_VECTOR where all of the elements are 0 or undef.
150 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
151 // Look through a bit convert.
152 if (N->getOpcode() == ISD::BIT_CONVERT)
153 N = N->getOperand(0).getNode();
155 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
157 unsigned i = 0, e = N->getNumOperands();
159 // Skip over all of the undef values.
160 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
163 // Do not accept an all-undef vector.
164 if (i == e) return false;
166 // Do not accept build_vectors that aren't all constants or which have non-0
168 SDValue Zero = N->getOperand(i);
169 if (isa<ConstantSDNode>(Zero)) {
170 if (!cast<ConstantSDNode>(Zero)->isNullValue())
172 } else if (isa<ConstantFPSDNode>(Zero)) {
173 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
178 // Okay, we have at least one 0 value, check to see if the rest match or are
180 for (++i; i != e; ++i)
181 if (N->getOperand(i) != Zero &&
182 N->getOperand(i).getOpcode() != ISD::UNDEF)
187 /// isScalarToVector - Return true if the specified node is a
188 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
189 /// element is not an undef.
190 bool ISD::isScalarToVector(const SDNode *N) {
191 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
194 if (N->getOpcode() != ISD::BUILD_VECTOR)
196 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
198 unsigned NumElems = N->getNumOperands();
199 for (unsigned i = 1; i < NumElems; ++i) {
200 SDValue V = N->getOperand(i);
201 if (V.getOpcode() != ISD::UNDEF)
207 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
208 /// when given the operation for (X op Y).
209 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
210 // To perform this operation, we just need to swap the L and G bits of the
212 unsigned OldL = (Operation >> 2) & 1;
213 unsigned OldG = (Operation >> 1) & 1;
214 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
215 (OldL << 1) | // New G bit
216 (OldG << 2)); // New L bit.
219 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
220 /// 'op' is a valid SetCC operation.
221 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
222 unsigned Operation = Op;
224 Operation ^= 7; // Flip L, G, E bits, but not U.
226 Operation ^= 15; // Flip all of the condition bits.
228 if (Operation > ISD::SETTRUE2)
229 Operation &= ~8; // Don't let N and U bits get set.
231 return ISD::CondCode(Operation);
235 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
236 /// signed operation and 2 if the result is an unsigned comparison. Return zero
237 /// if the operation does not depend on the sign of the input (setne and seteq).
238 static int isSignedOp(ISD::CondCode Opcode) {
240 default: llvm_unreachable("Illegal integer setcc operation!");
242 case ISD::SETNE: return 0;
246 case ISD::SETGE: return 1;
250 case ISD::SETUGE: return 2;
254 /// getSetCCOrOperation - Return the result of a logical OR between different
255 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
256 /// returns SETCC_INVALID if it is not possible to represent the resultant
258 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
260 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
261 // Cannot fold a signed integer setcc with an unsigned integer setcc.
262 return ISD::SETCC_INVALID;
264 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
266 // If the N and U bits get set then the resultant comparison DOES suddenly
267 // care about orderedness, and is true when ordered.
268 if (Op > ISD::SETTRUE2)
269 Op &= ~16; // Clear the U bit if the N bit is set.
271 // Canonicalize illegal integer setcc's.
272 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
275 return ISD::CondCode(Op);
278 /// getSetCCAndOperation - Return the result of a logical AND between different
279 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
280 /// function returns zero if it is not possible to represent the resultant
282 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
284 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
285 // Cannot fold a signed setcc with an unsigned setcc.
286 return ISD::SETCC_INVALID;
288 // Combine all of the condition bits.
289 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
291 // Canonicalize illegal integer setcc's.
295 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
296 case ISD::SETOEQ: // SETEQ & SETU[LG]E
297 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
298 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
299 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
306 const TargetMachine &SelectionDAG::getTarget() const {
307 return MF->getTarget();
310 //===----------------------------------------------------------------------===//
311 // SDNode Profile Support
312 //===----------------------------------------------------------------------===//
314 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
316 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
320 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
321 /// solely with their pointer.
322 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
323 ID.AddPointer(VTList.VTs);
326 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
328 static void AddNodeIDOperands(FoldingSetNodeID &ID,
329 const SDValue *Ops, unsigned NumOps) {
330 for (; NumOps; --NumOps, ++Ops) {
331 ID.AddPointer(Ops->getNode());
332 ID.AddInteger(Ops->getResNo());
336 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
338 static void AddNodeIDOperands(FoldingSetNodeID &ID,
339 const SDUse *Ops, unsigned NumOps) {
340 for (; NumOps; --NumOps, ++Ops) {
341 ID.AddPointer(Ops->getNode());
342 ID.AddInteger(Ops->getResNo());
346 static void AddNodeIDNode(FoldingSetNodeID &ID,
347 unsigned short OpC, SDVTList VTList,
348 const SDValue *OpList, unsigned N) {
349 AddNodeIDOpcode(ID, OpC);
350 AddNodeIDValueTypes(ID, VTList);
351 AddNodeIDOperands(ID, OpList, N);
354 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
356 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
357 switch (N->getOpcode()) {
358 case ISD::TargetExternalSymbol:
359 case ISD::ExternalSymbol:
360 llvm_unreachable("Should only be used on nodes with operands");
361 default: break; // Normal nodes don't need extra info.
362 case ISD::TargetConstant:
364 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
366 case ISD::TargetConstantFP:
367 case ISD::ConstantFP: {
368 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
371 case ISD::TargetGlobalAddress:
372 case ISD::GlobalAddress:
373 case ISD::TargetGlobalTLSAddress:
374 case ISD::GlobalTLSAddress: {
375 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
376 ID.AddPointer(GA->getGlobal());
377 ID.AddInteger(GA->getOffset());
378 ID.AddInteger(GA->getTargetFlags());
381 case ISD::BasicBlock:
382 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
385 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
389 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
391 case ISD::FrameIndex:
392 case ISD::TargetFrameIndex:
393 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
396 case ISD::TargetJumpTable:
397 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
398 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
400 case ISD::ConstantPool:
401 case ISD::TargetConstantPool: {
402 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
403 ID.AddInteger(CP->getAlignment());
404 ID.AddInteger(CP->getOffset());
405 if (CP->isMachineConstantPoolEntry())
406 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
408 ID.AddPointer(CP->getConstVal());
409 ID.AddInteger(CP->getTargetFlags());
413 const LoadSDNode *LD = cast<LoadSDNode>(N);
414 ID.AddInteger(LD->getMemoryVT().getRawBits());
415 ID.AddInteger(LD->getRawSubclassData());
419 const StoreSDNode *ST = cast<StoreSDNode>(N);
420 ID.AddInteger(ST->getMemoryVT().getRawBits());
421 ID.AddInteger(ST->getRawSubclassData());
424 case ISD::ATOMIC_CMP_SWAP:
425 case ISD::ATOMIC_SWAP:
426 case ISD::ATOMIC_LOAD_ADD:
427 case ISD::ATOMIC_LOAD_SUB:
428 case ISD::ATOMIC_LOAD_AND:
429 case ISD::ATOMIC_LOAD_OR:
430 case ISD::ATOMIC_LOAD_XOR:
431 case ISD::ATOMIC_LOAD_NAND:
432 case ISD::ATOMIC_LOAD_MIN:
433 case ISD::ATOMIC_LOAD_MAX:
434 case ISD::ATOMIC_LOAD_UMIN:
435 case ISD::ATOMIC_LOAD_UMAX: {
436 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
437 ID.AddInteger(AT->getMemoryVT().getRawBits());
438 ID.AddInteger(AT->getRawSubclassData());
441 case ISD::VECTOR_SHUFFLE: {
442 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
443 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
445 ID.AddInteger(SVN->getMaskElt(i));
448 case ISD::TargetBlockAddress:
449 case ISD::BlockAddress: {
450 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress());
451 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags());
454 } // end switch (N->getOpcode())
457 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
459 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
460 AddNodeIDOpcode(ID, N->getOpcode());
461 // Add the return value info.
462 AddNodeIDValueTypes(ID, N->getVTList());
463 // Add the operand info.
464 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
466 // Handle SDNode leafs with special info.
467 AddNodeIDCustom(ID, N);
470 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
471 /// the CSE map that carries volatility, temporalness, indexing mode, and
472 /// extension/truncation information.
474 static inline unsigned
475 encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile,
476 bool isNonTemporal) {
477 assert((ConvType & 3) == ConvType &&
478 "ConvType may not require more than 2 bits!");
479 assert((AM & 7) == AM &&
480 "AM may not require more than 3 bits!");
484 (isNonTemporal << 6);
487 //===----------------------------------------------------------------------===//
488 // SelectionDAG Class
489 //===----------------------------------------------------------------------===//
491 /// doNotCSE - Return true if CSE should not be performed for this node.
492 static bool doNotCSE(SDNode *N) {
493 if (N->getValueType(0) == MVT::Flag)
494 return true; // Never CSE anything that produces a flag.
496 switch (N->getOpcode()) {
498 case ISD::HANDLENODE:
500 return true; // Never CSE these nodes.
503 // Check that remaining values produced are not flags.
504 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
505 if (N->getValueType(i) == MVT::Flag)
506 return true; // Never CSE anything that produces a flag.
511 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
513 void SelectionDAG::RemoveDeadNodes() {
514 // Create a dummy node (which is not added to allnodes), that adds a reference
515 // to the root node, preventing it from being deleted.
516 HandleSDNode Dummy(getRoot());
518 SmallVector<SDNode*, 128> DeadNodes;
520 // Add all obviously-dead nodes to the DeadNodes worklist.
521 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
523 DeadNodes.push_back(I);
525 RemoveDeadNodes(DeadNodes);
527 // If the root changed (e.g. it was a dead load, update the root).
528 setRoot(Dummy.getValue());
531 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
532 /// given list, and any nodes that become unreachable as a result.
533 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
534 DAGUpdateListener *UpdateListener) {
536 // Process the worklist, deleting the nodes and adding their uses to the
538 while (!DeadNodes.empty()) {
539 SDNode *N = DeadNodes.pop_back_val();
542 UpdateListener->NodeDeleted(N, 0);
544 // Take the node out of the appropriate CSE map.
545 RemoveNodeFromCSEMaps(N);
547 // Next, brutally remove the operand list. This is safe to do, as there are
548 // no cycles in the graph.
549 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
551 SDNode *Operand = Use.getNode();
554 // Now that we removed this operand, see if there are no uses of it left.
555 if (Operand->use_empty())
556 DeadNodes.push_back(Operand);
563 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
564 SmallVector<SDNode*, 16> DeadNodes(1, N);
565 RemoveDeadNodes(DeadNodes, UpdateListener);
568 void SelectionDAG::DeleteNode(SDNode *N) {
569 // First take this out of the appropriate CSE map.
570 RemoveNodeFromCSEMaps(N);
572 // Finally, remove uses due to operands of this node, remove from the
573 // AllNodes list, and delete the node.
574 DeleteNodeNotInCSEMaps(N);
577 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
578 assert(N != AllNodes.begin() && "Cannot delete the entry node!");
579 assert(N->use_empty() && "Cannot delete a node that is not dead!");
581 // Drop all of the operands and decrement used node's use counts.
587 void SelectionDAG::DeallocateNode(SDNode *N) {
588 if (N->OperandsNeedDelete)
589 delete[] N->OperandList;
591 // Set the opcode to DELETED_NODE to help catch bugs when node
592 // memory is reallocated.
593 N->NodeType = ISD::DELETED_NODE;
595 NodeAllocator.Deallocate(AllNodes.remove(N));
597 // Remove the ordering of this node.
601 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
602 /// correspond to it. This is useful when we're about to delete or repurpose
603 /// the node. We don't want future request for structurally identical nodes
604 /// to return N anymore.
605 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
607 switch (N->getOpcode()) {
608 case ISD::EntryToken:
609 llvm_unreachable("EntryToken should not be in CSEMaps!");
611 case ISD::HANDLENODE: return false; // noop.
613 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
614 "Cond code doesn't exist!");
615 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
616 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
618 case ISD::ExternalSymbol:
619 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
621 case ISD::TargetExternalSymbol: {
622 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
623 Erased = TargetExternalSymbols.erase(
624 std::pair<std::string,unsigned char>(ESN->getSymbol(),
625 ESN->getTargetFlags()));
628 case ISD::VALUETYPE: {
629 EVT VT = cast<VTSDNode>(N)->getVT();
630 if (VT.isExtended()) {
631 Erased = ExtendedValueTypeNodes.erase(VT);
633 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0;
634 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0;
639 // Remove it from the CSE Map.
640 Erased = CSEMap.RemoveNode(N);
644 // Verify that the node was actually in one of the CSE maps, unless it has a
645 // flag result (which cannot be CSE'd) or is one of the special cases that are
646 // not subject to CSE.
647 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
648 !N->isMachineOpcode() && !doNotCSE(N)) {
651 llvm_unreachable("Node is not in map!");
657 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
658 /// maps and modified in place. Add it back to the CSE maps, unless an identical
659 /// node already exists, in which case transfer all its users to the existing
660 /// node. This transfer can potentially trigger recursive merging.
663 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N,
664 DAGUpdateListener *UpdateListener) {
665 // For node types that aren't CSE'd, just act as if no identical node
668 SDNode *Existing = CSEMap.GetOrInsertNode(N);
670 // If there was already an existing matching node, use ReplaceAllUsesWith
671 // to replace the dead one with the existing one. This can cause
672 // recursive merging of other unrelated nodes down the line.
673 ReplaceAllUsesWith(N, Existing, UpdateListener);
675 // N is now dead. Inform the listener if it exists and delete it.
677 UpdateListener->NodeDeleted(N, Existing);
678 DeleteNodeNotInCSEMaps(N);
683 // If the node doesn't already exist, we updated it. Inform a listener if
686 UpdateListener->NodeUpdated(N);
689 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
690 /// were replaced with those specified. If this node is never memoized,
691 /// return null, otherwise return a pointer to the slot it would take. If a
692 /// node already exists with these operands, the slot will be non-null.
693 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
698 SDValue Ops[] = { Op };
700 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
701 AddNodeIDCustom(ID, N);
702 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
706 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
707 /// were replaced with those specified. If this node is never memoized,
708 /// return null, otherwise return a pointer to the slot it would take. If a
709 /// node already exists with these operands, the slot will be non-null.
710 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
711 SDValue Op1, SDValue Op2,
716 SDValue Ops[] = { Op1, Op2 };
718 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
719 AddNodeIDCustom(ID, N);
720 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
725 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
726 /// were replaced with those specified. If this node is never memoized,
727 /// return null, otherwise return a pointer to the slot it would take. If a
728 /// node already exists with these operands, the slot will be non-null.
729 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
730 const SDValue *Ops,unsigned NumOps,
736 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
737 AddNodeIDCustom(ID, N);
738 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
742 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
743 void SelectionDAG::VerifyNode(SDNode *N) {
744 switch (N->getOpcode()) {
747 case ISD::BUILD_PAIR: {
748 EVT VT = N->getValueType(0);
749 assert(N->getNumValues() == 1 && "Too many results!");
750 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
751 "Wrong return type!");
752 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
753 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
754 "Mismatched operand types!");
755 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
756 "Wrong operand type!");
757 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
758 "Wrong return type size");
761 case ISD::BUILD_VECTOR: {
762 assert(N->getNumValues() == 1 && "Too many results!");
763 assert(N->getValueType(0).isVector() && "Wrong return type!");
764 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
765 "Wrong number of operands!");
766 EVT EltVT = N->getValueType(0).getVectorElementType();
767 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
768 assert((I->getValueType() == EltVT ||
769 (EltVT.isInteger() && I->getValueType().isInteger() &&
770 EltVT.bitsLE(I->getValueType()))) &&
771 "Wrong operand type!");
777 /// getEVTAlignment - Compute the default alignment value for the
780 unsigned SelectionDAG::getEVTAlignment(EVT VT) const {
781 const Type *Ty = VT == MVT::iPTR ?
782 PointerType::get(Type::getInt8Ty(*getContext()), 0) :
783 VT.getTypeForEVT(*getContext());
785 return TLI.getTargetData()->getABITypeAlignment(Ty);
788 // EntryNode could meaningfully have debug info if we can find it...
789 SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli)
790 : TLI(tli), FLI(fli), DW(0),
791 EntryNode(ISD::EntryToken, DebugLoc::getUnknownLoc(),
792 getVTList(MVT::Other)),
793 Root(getEntryNode()), Ordering(0) {
794 AllNodes.push_back(&EntryNode);
795 Ordering = new SDNodeOrdering();
798 void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi,
803 Context = &mf.getFunction()->getContext();
806 SelectionDAG::~SelectionDAG() {
811 void SelectionDAG::allnodes_clear() {
812 assert(&*AllNodes.begin() == &EntryNode);
813 AllNodes.remove(AllNodes.begin());
814 while (!AllNodes.empty())
815 DeallocateNode(AllNodes.begin());
818 void SelectionDAG::clear() {
820 OperandAllocator.Reset();
823 ExtendedValueTypeNodes.clear();
824 ExternalSymbols.clear();
825 TargetExternalSymbols.clear();
826 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
827 static_cast<CondCodeSDNode*>(0));
828 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
829 static_cast<SDNode*>(0));
831 EntryNode.UseList = 0;
832 AllNodes.push_back(&EntryNode);
833 Root = getEntryNode();
835 Ordering = new SDNodeOrdering();
838 SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
839 return VT.bitsGT(Op.getValueType()) ?
840 getNode(ISD::SIGN_EXTEND, DL, VT, Op) :
841 getNode(ISD::TRUNCATE, DL, VT, Op);
844 SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
845 return VT.bitsGT(Op.getValueType()) ?
846 getNode(ISD::ZERO_EXTEND, DL, VT, Op) :
847 getNode(ISD::TRUNCATE, DL, VT, Op);
850 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) {
851 assert(!VT.isVector() &&
852 "getZeroExtendInReg should use the vector element type instead of "
854 if (Op.getValueType() == VT) return Op;
855 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
856 APInt Imm = APInt::getLowBitsSet(BitWidth,
858 return getNode(ISD::AND, DL, Op.getValueType(), Op,
859 getConstant(Imm, Op.getValueType()));
862 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
864 SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) {
865 EVT EltVT = VT.getScalarType();
867 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
868 return getNode(ISD::XOR, DL, VT, Val, NegOne);
871 SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) {
872 EVT EltVT = VT.getScalarType();
873 assert((EltVT.getSizeInBits() >= 64 ||
874 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
875 "getConstant with a uint64_t value that doesn't fit in the type!");
876 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
879 SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) {
880 return getConstant(*ConstantInt::get(*Context, Val), VT, isT);
883 SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
884 assert(VT.isInteger() && "Cannot create FP integer constant!");
886 EVT EltVT = VT.getScalarType();
887 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
888 "APInt size does not match type size!");
890 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
892 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
896 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
898 return SDValue(N, 0);
901 N = NodeAllocator.Allocate<ConstantSDNode>();
902 new (N) ConstantSDNode(isT, &Val, EltVT);
903 CSEMap.InsertNode(N, IP);
904 AllNodes.push_back(N);
907 SDValue Result(N, 0);
909 SmallVector<SDValue, 8> Ops;
910 Ops.assign(VT.getVectorNumElements(), Result);
911 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
912 VT, &Ops[0], Ops.size());
917 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
918 return getConstant(Val, TLI.getPointerTy(), isTarget);
922 SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) {
923 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget);
926 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){
927 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
929 EVT EltVT = VT.getScalarType();
931 // Do the map lookup using the actual bit pattern for the floating point
932 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
933 // we don't have issues with SNANs.
934 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
936 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
940 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
942 return SDValue(N, 0);
945 N = NodeAllocator.Allocate<ConstantFPSDNode>();
946 new (N) ConstantFPSDNode(isTarget, &V, EltVT);
947 CSEMap.InsertNode(N, IP);
948 AllNodes.push_back(N);
951 SDValue Result(N, 0);
953 SmallVector<SDValue, 8> Ops;
954 Ops.assign(VT.getVectorNumElements(), Result);
955 // FIXME DebugLoc info might be appropriate here
956 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
957 VT, &Ops[0], Ops.size());
962 SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) {
963 EVT EltVT = VT.getScalarType();
965 return getConstantFP(APFloat((float)Val), VT, isTarget);
967 return getConstantFP(APFloat(Val), VT, isTarget);
970 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
971 EVT VT, int64_t Offset,
973 unsigned char TargetFlags) {
974 assert((TargetFlags == 0 || isTargetGA) &&
975 "Cannot set target flags on target-independent globals");
977 // Truncate (with sign-extension) the offset value to the pointer size.
978 EVT PTy = TLI.getPointerTy();
979 unsigned BitWidth = PTy.getSizeInBits();
981 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
983 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
985 // If GV is an alias then use the aliasee for determining thread-localness.
986 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
987 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
991 if (GVar && GVar->isThreadLocal())
992 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
994 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
997 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
999 ID.AddInteger(Offset);
1000 ID.AddInteger(TargetFlags);
1002 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1003 return SDValue(E, 0);
1005 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>();
1006 new (N) GlobalAddressSDNode(Opc, GV, VT, Offset, TargetFlags);
1007 CSEMap.InsertNode(N, IP);
1008 AllNodes.push_back(N);
1009 return SDValue(N, 0);
1012 SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1013 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1014 FoldingSetNodeID ID;
1015 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1018 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1019 return SDValue(E, 0);
1021 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>();
1022 new (N) FrameIndexSDNode(FI, VT, isTarget);
1023 CSEMap.InsertNode(N, IP);
1024 AllNodes.push_back(N);
1025 return SDValue(N, 0);
1028 SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1029 unsigned char TargetFlags) {
1030 assert((TargetFlags == 0 || isTarget) &&
1031 "Cannot set target flags on target-independent jump tables");
1032 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1033 FoldingSetNodeID ID;
1034 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1036 ID.AddInteger(TargetFlags);
1038 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1039 return SDValue(E, 0);
1041 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>();
1042 new (N) JumpTableSDNode(JTI, VT, isTarget, TargetFlags);
1043 CSEMap.InsertNode(N, IP);
1044 AllNodes.push_back(N);
1045 return SDValue(N, 0);
1048 SDValue SelectionDAG::getConstantPool(Constant *C, EVT VT,
1049 unsigned Alignment, int Offset,
1051 unsigned char TargetFlags) {
1052 assert((TargetFlags == 0 || isTarget) &&
1053 "Cannot set target flags on target-independent globals");
1055 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1056 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1057 FoldingSetNodeID ID;
1058 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1059 ID.AddInteger(Alignment);
1060 ID.AddInteger(Offset);
1062 ID.AddInteger(TargetFlags);
1064 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1065 return SDValue(E, 0);
1067 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1068 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1069 CSEMap.InsertNode(N, IP);
1070 AllNodes.push_back(N);
1071 return SDValue(N, 0);
1075 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1076 unsigned Alignment, int Offset,
1078 unsigned char TargetFlags) {
1079 assert((TargetFlags == 0 || isTarget) &&
1080 "Cannot set target flags on target-independent globals");
1082 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1083 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1084 FoldingSetNodeID ID;
1085 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1086 ID.AddInteger(Alignment);
1087 ID.AddInteger(Offset);
1088 C->AddSelectionDAGCSEId(ID);
1089 ID.AddInteger(TargetFlags);
1091 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1092 return SDValue(E, 0);
1094 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1095 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1096 CSEMap.InsertNode(N, IP);
1097 AllNodes.push_back(N);
1098 return SDValue(N, 0);
1101 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1102 FoldingSetNodeID ID;
1103 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1106 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1107 return SDValue(E, 0);
1109 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>();
1110 new (N) BasicBlockSDNode(MBB);
1111 CSEMap.InsertNode(N, IP);
1112 AllNodes.push_back(N);
1113 return SDValue(N, 0);
1116 SDValue SelectionDAG::getValueType(EVT VT) {
1117 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
1118 ValueTypeNodes.size())
1119 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1);
1121 SDNode *&N = VT.isExtended() ?
1122 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
1124 if (N) return SDValue(N, 0);
1125 N = NodeAllocator.Allocate<VTSDNode>();
1126 new (N) VTSDNode(VT);
1127 AllNodes.push_back(N);
1128 return SDValue(N, 0);
1131 SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
1132 SDNode *&N = ExternalSymbols[Sym];
1133 if (N) return SDValue(N, 0);
1134 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1135 new (N) ExternalSymbolSDNode(false, Sym, 0, VT);
1136 AllNodes.push_back(N);
1137 return SDValue(N, 0);
1140 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
1141 unsigned char TargetFlags) {
1143 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym,
1145 if (N) return SDValue(N, 0);
1146 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1147 new (N) ExternalSymbolSDNode(true, Sym, TargetFlags, VT);
1148 AllNodes.push_back(N);
1149 return SDValue(N, 0);
1152 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1153 if ((unsigned)Cond >= CondCodeNodes.size())
1154 CondCodeNodes.resize(Cond+1);
1156 if (CondCodeNodes[Cond] == 0) {
1157 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>();
1158 new (N) CondCodeSDNode(Cond);
1159 CondCodeNodes[Cond] = N;
1160 AllNodes.push_back(N);
1163 return SDValue(CondCodeNodes[Cond], 0);
1166 // commuteShuffle - swaps the values of N1 and N2, and swaps all indices in
1167 // the shuffle mask M that point at N1 to point at N2, and indices that point
1168 // N2 to point at N1.
1169 static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) {
1171 int NElts = M.size();
1172 for (int i = 0; i != NElts; ++i) {
1180 SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1,
1181 SDValue N2, const int *Mask) {
1182 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE");
1183 assert(VT.isVector() && N1.getValueType().isVector() &&
1184 "Vector Shuffle VTs must be a vectors");
1185 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType()
1186 && "Vector Shuffle VTs must have same element type");
1188 // Canonicalize shuffle undef, undef -> undef
1189 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF)
1190 return getUNDEF(VT);
1192 // Validate that all indices in Mask are within the range of the elements
1193 // input to the shuffle.
1194 unsigned NElts = VT.getVectorNumElements();
1195 SmallVector<int, 8> MaskVec;
1196 for (unsigned i = 0; i != NElts; ++i) {
1197 assert(Mask[i] < (int)(NElts * 2) && "Index out of range");
1198 MaskVec.push_back(Mask[i]);
1201 // Canonicalize shuffle v, v -> v, undef
1204 for (unsigned i = 0; i != NElts; ++i)
1205 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts;
1208 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1209 if (N1.getOpcode() == ISD::UNDEF)
1210 commuteShuffle(N1, N2, MaskVec);
1212 // Canonicalize all index into lhs, -> shuffle lhs, undef
1213 // Canonicalize all index into rhs, -> shuffle rhs, undef
1214 bool AllLHS = true, AllRHS = true;
1215 bool N2Undef = N2.getOpcode() == ISD::UNDEF;
1216 for (unsigned i = 0; i != NElts; ++i) {
1217 if (MaskVec[i] >= (int)NElts) {
1222 } else if (MaskVec[i] >= 0) {
1226 if (AllLHS && AllRHS)
1227 return getUNDEF(VT);
1228 if (AllLHS && !N2Undef)
1232 commuteShuffle(N1, N2, MaskVec);
1235 // If Identity shuffle, or all shuffle in to undef, return that node.
1236 bool AllUndef = true;
1237 bool Identity = true;
1238 for (unsigned i = 0; i != NElts; ++i) {
1239 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false;
1240 if (MaskVec[i] >= 0) AllUndef = false;
1242 if (Identity && NElts == N1.getValueType().getVectorNumElements())
1245 return getUNDEF(VT);
1247 FoldingSetNodeID ID;
1248 SDValue Ops[2] = { N1, N2 };
1249 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2);
1250 for (unsigned i = 0; i != NElts; ++i)
1251 ID.AddInteger(MaskVec[i]);
1254 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1255 return SDValue(E, 0);
1257 // Allocate the mask array for the node out of the BumpPtrAllocator, since
1258 // SDNode doesn't have access to it. This memory will be "leaked" when
1259 // the node is deallocated, but recovered when the NodeAllocator is released.
1260 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
1261 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int));
1263 ShuffleVectorSDNode *N = NodeAllocator.Allocate<ShuffleVectorSDNode>();
1264 new (N) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc);
1265 CSEMap.InsertNode(N, IP);
1266 AllNodes.push_back(N);
1267 return SDValue(N, 0);
1270 SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl,
1271 SDValue Val, SDValue DTy,
1272 SDValue STy, SDValue Rnd, SDValue Sat,
1273 ISD::CvtCode Code) {
1274 // If the src and dest types are the same and the conversion is between
1275 // integer types of the same sign or two floats, no conversion is necessary.
1277 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF))
1280 FoldingSetNodeID ID;
1281 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat };
1282 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5);
1284 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1285 return SDValue(E, 0);
1287 CvtRndSatSDNode *N = NodeAllocator.Allocate<CvtRndSatSDNode>();
1288 new (N) CvtRndSatSDNode(VT, dl, Ops, 5, Code);
1289 CSEMap.InsertNode(N, IP);
1290 AllNodes.push_back(N);
1291 return SDValue(N, 0);
1294 SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) {
1295 FoldingSetNodeID ID;
1296 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1297 ID.AddInteger(RegNo);
1299 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1300 return SDValue(E, 0);
1302 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>();
1303 new (N) RegisterSDNode(RegNo, VT);
1304 CSEMap.InsertNode(N, IP);
1305 AllNodes.push_back(N);
1306 return SDValue(N, 0);
1309 SDValue SelectionDAG::getLabel(unsigned Opcode, DebugLoc dl,
1312 FoldingSetNodeID ID;
1313 SDValue Ops[] = { Root };
1314 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), &Ops[0], 1);
1315 ID.AddInteger(LabelID);
1317 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1318 return SDValue(E, 0);
1320 SDNode *N = NodeAllocator.Allocate<LabelSDNode>();
1321 new (N) LabelSDNode(Opcode, dl, Root, LabelID);
1322 CSEMap.InsertNode(N, IP);
1323 AllNodes.push_back(N);
1324 return SDValue(N, 0);
1327 SDValue SelectionDAG::getBlockAddress(BlockAddress *BA, EVT VT,
1329 unsigned char TargetFlags) {
1330 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress;
1332 FoldingSetNodeID ID;
1333 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1335 ID.AddInteger(TargetFlags);
1337 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1338 return SDValue(E, 0);
1340 SDNode *N = NodeAllocator.Allocate<BlockAddressSDNode>();
1341 new (N) BlockAddressSDNode(Opc, VT, BA, TargetFlags);
1342 CSEMap.InsertNode(N, IP);
1343 AllNodes.push_back(N);
1344 return SDValue(N, 0);
1347 SDValue SelectionDAG::getSrcValue(const Value *V) {
1348 assert((!V || V->getType()->isPointerTy()) &&
1349 "SrcValue is not a pointer?");
1351 FoldingSetNodeID ID;
1352 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1356 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1357 return SDValue(E, 0);
1359 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>();
1360 new (N) SrcValueSDNode(V);
1361 CSEMap.InsertNode(N, IP);
1362 AllNodes.push_back(N);
1363 return SDValue(N, 0);
1366 /// getShiftAmountOperand - Return the specified value casted to
1367 /// the target's desired shift amount type.
1368 SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) {
1369 EVT OpTy = Op.getValueType();
1370 MVT ShTy = TLI.getShiftAmountTy();
1371 if (OpTy == ShTy || OpTy.isVector()) return Op;
1373 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
1374 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op);
1377 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1378 /// specified value type.
1379 SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
1380 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1381 unsigned ByteSize = VT.getStoreSize();
1382 const Type *Ty = VT.getTypeForEVT(*getContext());
1383 unsigned StackAlign =
1384 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1386 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
1387 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1390 /// CreateStackTemporary - Create a stack temporary suitable for holding
1391 /// either of the specified value types.
1392 SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
1393 unsigned Bytes = std::max(VT1.getStoreSizeInBits(),
1394 VT2.getStoreSizeInBits())/8;
1395 const Type *Ty1 = VT1.getTypeForEVT(*getContext());
1396 const Type *Ty2 = VT2.getTypeForEVT(*getContext());
1397 const TargetData *TD = TLI.getTargetData();
1398 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1),
1399 TD->getPrefTypeAlignment(Ty2));
1401 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1402 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false);
1403 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1406 SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1,
1407 SDValue N2, ISD::CondCode Cond, DebugLoc dl) {
1408 // These setcc operations always fold.
1412 case ISD::SETFALSE2: return getConstant(0, VT);
1414 case ISD::SETTRUE2: return getConstant(1, VT);
1426 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1430 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1431 const APInt &C2 = N2C->getAPIntValue();
1432 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1433 const APInt &C1 = N1C->getAPIntValue();
1436 default: llvm_unreachable("Unknown integer setcc!");
1437 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1438 case ISD::SETNE: return getConstant(C1 != C2, VT);
1439 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1440 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1441 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1442 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1443 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1444 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1445 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1446 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1450 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1451 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1452 // No compile time operations on this type yet.
1453 if (N1C->getValueType(0) == MVT::ppcf128)
1456 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1459 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1460 return getUNDEF(VT);
1462 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1463 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1464 return getUNDEF(VT);
1466 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1467 R==APFloat::cmpLessThan, VT);
1468 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1469 return getUNDEF(VT);
1471 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1472 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1473 return getUNDEF(VT);
1475 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1476 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1477 return getUNDEF(VT);
1479 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1480 R==APFloat::cmpEqual, VT);
1481 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1482 return getUNDEF(VT);
1484 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1485 R==APFloat::cmpEqual, VT);
1486 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1487 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1488 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1489 R==APFloat::cmpEqual, VT);
1490 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1491 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1492 R==APFloat::cmpLessThan, VT);
1493 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1494 R==APFloat::cmpUnordered, VT);
1495 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1496 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1499 // Ensure that the constant occurs on the RHS.
1500 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1504 // Could not fold it.
1508 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1509 /// use this predicate to simplify operations downstream.
1510 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1511 // This predicate is not safe for vector operations.
1512 if (Op.getValueType().isVector())
1515 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
1516 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1519 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1520 /// this predicate to simplify operations downstream. Mask is known to be zero
1521 /// for bits that V cannot have.
1522 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1523 unsigned Depth) const {
1524 APInt KnownZero, KnownOne;
1525 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1526 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1527 return (KnownZero & Mask) == Mask;
1530 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1531 /// known to be either zero or one and return them in the KnownZero/KnownOne
1532 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1534 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1535 APInt &KnownZero, APInt &KnownOne,
1536 unsigned Depth) const {
1537 unsigned BitWidth = Mask.getBitWidth();
1538 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() &&
1539 "Mask size mismatches value type size!");
1541 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1542 if (Depth == 6 || Mask == 0)
1543 return; // Limit search depth.
1545 APInt KnownZero2, KnownOne2;
1547 switch (Op.getOpcode()) {
1549 // We know all of the bits for a constant!
1550 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1551 KnownZero = ~KnownOne & Mask;
1554 // If either the LHS or the RHS are Zero, the result is zero.
1555 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1556 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1557 KnownZero2, KnownOne2, Depth+1);
1558 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1559 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1561 // Output known-1 bits are only known if set in both the LHS & RHS.
1562 KnownOne &= KnownOne2;
1563 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1564 KnownZero |= KnownZero2;
1567 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1568 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1569 KnownZero2, KnownOne2, Depth+1);
1570 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1571 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1573 // Output known-0 bits are only known if clear in both the LHS & RHS.
1574 KnownZero &= KnownZero2;
1575 // Output known-1 are known to be set if set in either the LHS | RHS.
1576 KnownOne |= KnownOne2;
1579 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1580 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1581 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1582 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1584 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1585 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1586 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1587 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1588 KnownZero = KnownZeroOut;
1592 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1593 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1594 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1595 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1596 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1598 // If low bits are zero in either operand, output low known-0 bits.
1599 // Also compute a conserative estimate for high known-0 bits.
1600 // More trickiness is possible, but this is sufficient for the
1601 // interesting case of alignment computation.
1603 unsigned TrailZ = KnownZero.countTrailingOnes() +
1604 KnownZero2.countTrailingOnes();
1605 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1606 KnownZero2.countLeadingOnes(),
1607 BitWidth) - BitWidth;
1609 TrailZ = std::min(TrailZ, BitWidth);
1610 LeadZ = std::min(LeadZ, BitWidth);
1611 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1612 APInt::getHighBitsSet(BitWidth, LeadZ);
1617 // For the purposes of computing leading zeros we can conservatively
1618 // treat a udiv as a logical right shift by the power of 2 known to
1619 // be less than the denominator.
1620 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1621 ComputeMaskedBits(Op.getOperand(0),
1622 AllOnes, KnownZero2, KnownOne2, Depth+1);
1623 unsigned LeadZ = KnownZero2.countLeadingOnes();
1627 ComputeMaskedBits(Op.getOperand(1),
1628 AllOnes, KnownZero2, KnownOne2, Depth+1);
1629 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1630 if (RHSUnknownLeadingOnes != BitWidth)
1631 LeadZ = std::min(BitWidth,
1632 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1634 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1638 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1639 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1640 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1641 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1643 // Only known if known in both the LHS and RHS.
1644 KnownOne &= KnownOne2;
1645 KnownZero &= KnownZero2;
1647 case ISD::SELECT_CC:
1648 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1649 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1650 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1651 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1653 // Only known if known in both the LHS and RHS.
1654 KnownOne &= KnownOne2;
1655 KnownZero &= KnownZero2;
1663 if (Op.getResNo() != 1)
1665 // The boolean result conforms to getBooleanContents. Fall through.
1667 // If we know the result of a setcc has the top bits zero, use this info.
1668 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent &&
1670 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1673 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1674 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1675 unsigned ShAmt = SA->getZExtValue();
1677 // If the shift count is an invalid immediate, don't do anything.
1678 if (ShAmt >= BitWidth)
1681 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt),
1682 KnownZero, KnownOne, Depth+1);
1683 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1684 KnownZero <<= ShAmt;
1686 // low bits known zero.
1687 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1691 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1692 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1693 unsigned ShAmt = SA->getZExtValue();
1695 // If the shift count is an invalid immediate, don't do anything.
1696 if (ShAmt >= BitWidth)
1699 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1700 KnownZero, KnownOne, Depth+1);
1701 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1702 KnownZero = KnownZero.lshr(ShAmt);
1703 KnownOne = KnownOne.lshr(ShAmt);
1705 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1706 KnownZero |= HighBits; // High bits known zero.
1710 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1711 unsigned ShAmt = SA->getZExtValue();
1713 // If the shift count is an invalid immediate, don't do anything.
1714 if (ShAmt >= BitWidth)
1717 APInt InDemandedMask = (Mask << ShAmt);
1718 // If any of the demanded bits are produced by the sign extension, we also
1719 // demand the input sign bit.
1720 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1721 if (HighBits.getBoolValue())
1722 InDemandedMask |= APInt::getSignBit(BitWidth);
1724 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1726 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1727 KnownZero = KnownZero.lshr(ShAmt);
1728 KnownOne = KnownOne.lshr(ShAmt);
1730 // Handle the sign bits.
1731 APInt SignBit = APInt::getSignBit(BitWidth);
1732 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1734 if (KnownZero.intersects(SignBit)) {
1735 KnownZero |= HighBits; // New bits are known zero.
1736 } else if (KnownOne.intersects(SignBit)) {
1737 KnownOne |= HighBits; // New bits are known one.
1741 case ISD::SIGN_EXTEND_INREG: {
1742 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1743 unsigned EBits = EVT.getScalarType().getSizeInBits();
1745 // Sign extension. Compute the demanded bits in the result that are not
1746 // present in the input.
1747 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1749 APInt InSignBit = APInt::getSignBit(EBits);
1750 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1752 // If the sign extended bits are demanded, we know that the sign
1754 InSignBit.zext(BitWidth);
1755 if (NewBits.getBoolValue())
1756 InputDemandedBits |= InSignBit;
1758 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1759 KnownZero, KnownOne, Depth+1);
1760 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1762 // If the sign bit of the input is known set or clear, then we know the
1763 // top bits of the result.
1764 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1765 KnownZero |= NewBits;
1766 KnownOne &= ~NewBits;
1767 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1768 KnownOne |= NewBits;
1769 KnownZero &= ~NewBits;
1770 } else { // Input sign bit unknown
1771 KnownZero &= ~NewBits;
1772 KnownOne &= ~NewBits;
1779 unsigned LowBits = Log2_32(BitWidth)+1;
1780 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1785 if (ISD::isZEXTLoad(Op.getNode())) {
1786 LoadSDNode *LD = cast<LoadSDNode>(Op);
1787 EVT VT = LD->getMemoryVT();
1788 unsigned MemBits = VT.getScalarType().getSizeInBits();
1789 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1793 case ISD::ZERO_EXTEND: {
1794 EVT InVT = Op.getOperand(0).getValueType();
1795 unsigned InBits = InVT.getScalarType().getSizeInBits();
1796 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1797 APInt InMask = Mask;
1798 InMask.trunc(InBits);
1799 KnownZero.trunc(InBits);
1800 KnownOne.trunc(InBits);
1801 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1802 KnownZero.zext(BitWidth);
1803 KnownOne.zext(BitWidth);
1804 KnownZero |= NewBits;
1807 case ISD::SIGN_EXTEND: {
1808 EVT InVT = Op.getOperand(0).getValueType();
1809 unsigned InBits = InVT.getScalarType().getSizeInBits();
1810 APInt InSignBit = APInt::getSignBit(InBits);
1811 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1812 APInt InMask = Mask;
1813 InMask.trunc(InBits);
1815 // If any of the sign extended bits are demanded, we know that the sign
1816 // bit is demanded. Temporarily set this bit in the mask for our callee.
1817 if (NewBits.getBoolValue())
1818 InMask |= InSignBit;
1820 KnownZero.trunc(InBits);
1821 KnownOne.trunc(InBits);
1822 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1824 // Note if the sign bit is known to be zero or one.
1825 bool SignBitKnownZero = KnownZero.isNegative();
1826 bool SignBitKnownOne = KnownOne.isNegative();
1827 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1828 "Sign bit can't be known to be both zero and one!");
1830 // If the sign bit wasn't actually demanded by our caller, we don't
1831 // want it set in the KnownZero and KnownOne result values. Reset the
1832 // mask and reapply it to the result values.
1834 InMask.trunc(InBits);
1835 KnownZero &= InMask;
1838 KnownZero.zext(BitWidth);
1839 KnownOne.zext(BitWidth);
1841 // If the sign bit is known zero or one, the top bits match.
1842 if (SignBitKnownZero)
1843 KnownZero |= NewBits;
1844 else if (SignBitKnownOne)
1845 KnownOne |= NewBits;
1848 case ISD::ANY_EXTEND: {
1849 EVT InVT = Op.getOperand(0).getValueType();
1850 unsigned InBits = InVT.getScalarType().getSizeInBits();
1851 APInt InMask = Mask;
1852 InMask.trunc(InBits);
1853 KnownZero.trunc(InBits);
1854 KnownOne.trunc(InBits);
1855 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1856 KnownZero.zext(BitWidth);
1857 KnownOne.zext(BitWidth);
1860 case ISD::TRUNCATE: {
1861 EVT InVT = Op.getOperand(0).getValueType();
1862 unsigned InBits = InVT.getScalarType().getSizeInBits();
1863 APInt InMask = Mask;
1864 InMask.zext(InBits);
1865 KnownZero.zext(InBits);
1866 KnownOne.zext(InBits);
1867 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1868 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1869 KnownZero.trunc(BitWidth);
1870 KnownOne.trunc(BitWidth);
1873 case ISD::AssertZext: {
1874 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1875 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1876 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1878 KnownZero |= (~InMask) & Mask;
1882 // All bits are zero except the low bit.
1883 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1887 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1888 // We know that the top bits of C-X are clear if X contains less bits
1889 // than C (i.e. no wrap-around can happen). For example, 20-X is
1890 // positive if we can prove that X is >= 0 and < 16.
1891 if (CLHS->getAPIntValue().isNonNegative()) {
1892 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1893 // NLZ can't be BitWidth with no sign bit
1894 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1895 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1898 // If all of the MaskV bits are known to be zero, then we know the
1899 // output top bits are zero, because we now know that the output is
1901 if ((KnownZero2 & MaskV) == MaskV) {
1902 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1903 // Top bits known zero.
1904 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1911 // Output known-0 bits are known if clear or set in both the low clear bits
1912 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1913 // low 3 bits clear.
1914 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
1915 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1916 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1917 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1919 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1920 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1921 KnownZeroOut = std::min(KnownZeroOut,
1922 KnownZero2.countTrailingOnes());
1924 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1928 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1929 const APInt &RA = Rem->getAPIntValue().abs();
1930 if (RA.isPowerOf2()) {
1931 APInt LowBits = RA - 1;
1932 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1933 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1935 // The low bits of the first operand are unchanged by the srem.
1936 KnownZero = KnownZero2 & LowBits;
1937 KnownOne = KnownOne2 & LowBits;
1939 // If the first operand is non-negative or has all low bits zero, then
1940 // the upper bits are all zero.
1941 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1942 KnownZero |= ~LowBits;
1944 // If the first operand is negative and not all low bits are zero, then
1945 // the upper bits are all one.
1946 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1947 KnownOne |= ~LowBits;
1952 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1957 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1958 const APInt &RA = Rem->getAPIntValue();
1959 if (RA.isPowerOf2()) {
1960 APInt LowBits = (RA - 1);
1961 APInt Mask2 = LowBits & Mask;
1962 KnownZero |= ~LowBits & Mask;
1963 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1964 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1969 // Since the result is less than or equal to either operand, any leading
1970 // zero bits in either operand must also exist in the result.
1971 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1972 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne,
1974 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2,
1977 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1978 KnownZero2.countLeadingOnes());
1980 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1984 // Allow the target to implement this method for its nodes.
1985 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
1986 case ISD::INTRINSIC_WO_CHAIN:
1987 case ISD::INTRINSIC_W_CHAIN:
1988 case ISD::INTRINSIC_VOID:
1989 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this,
1996 /// ComputeNumSignBits - Return the number of times the sign bit of the
1997 /// register is replicated into the other bits. We know that at least 1 bit
1998 /// is always equal to the sign bit (itself), but other cases can give us
1999 /// information. For example, immediately after an "SRA X, 2", we know that
2000 /// the top 3 bits are all equal to each other, so we return 3.
2001 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
2002 EVT VT = Op.getValueType();
2003 assert(VT.isInteger() && "Invalid VT!");
2004 unsigned VTBits = VT.getScalarType().getSizeInBits();
2006 unsigned FirstAnswer = 1;
2009 return 1; // Limit search depth.
2011 switch (Op.getOpcode()) {
2013 case ISD::AssertSext:
2014 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2015 return VTBits-Tmp+1;
2016 case ISD::AssertZext:
2017 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2020 case ISD::Constant: {
2021 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
2022 // If negative, return # leading ones.
2023 if (Val.isNegative())
2024 return Val.countLeadingOnes();
2026 // Return # leading zeros.
2027 return Val.countLeadingZeros();
2030 case ISD::SIGN_EXTEND:
2031 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
2032 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
2034 case ISD::SIGN_EXTEND_INREG:
2035 // Max of the input and what this extends.
2037 cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarType().getSizeInBits();
2040 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2041 return std::max(Tmp, Tmp2);
2044 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2045 // SRA X, C -> adds C sign bits.
2046 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2047 Tmp += C->getZExtValue();
2048 if (Tmp > VTBits) Tmp = VTBits;
2052 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2053 // shl destroys sign bits.
2054 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2055 if (C->getZExtValue() >= VTBits || // Bad shift.
2056 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2057 return Tmp - C->getZExtValue();
2062 case ISD::XOR: // NOT is handled here.
2063 // Logical binary ops preserve the number of sign bits at the worst.
2064 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2066 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2067 FirstAnswer = std::min(Tmp, Tmp2);
2068 // We computed what we know about the sign bits as our first
2069 // answer. Now proceed to the generic code that uses
2070 // ComputeMaskedBits, and pick whichever answer is better.
2075 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2076 if (Tmp == 1) return 1; // Early out.
2077 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
2078 return std::min(Tmp, Tmp2);
2086 if (Op.getResNo() != 1)
2088 // The boolean result conforms to getBooleanContents. Fall through.
2090 // If setcc returns 0/-1, all bits are sign bits.
2091 if (TLI.getBooleanContents() ==
2092 TargetLowering::ZeroOrNegativeOneBooleanContent)
2097 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2098 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
2100 // Handle rotate right by N like a rotate left by 32-N.
2101 if (Op.getOpcode() == ISD::ROTR)
2102 RotAmt = (VTBits-RotAmt) & (VTBits-1);
2104 // If we aren't rotating out all of the known-in sign bits, return the
2105 // number that are left. This handles rotl(sext(x), 1) for example.
2106 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2107 if (Tmp > RotAmt+1) return Tmp-RotAmt;
2111 // Add can have at most one carry bit. Thus we know that the output
2112 // is, at worst, one more bit than the inputs.
2113 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2114 if (Tmp == 1) return 1; // Early out.
2116 // Special case decrementing a value (ADD X, -1):
2117 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
2118 if (CRHS->isAllOnesValue()) {
2119 APInt KnownZero, KnownOne;
2120 APInt Mask = APInt::getAllOnesValue(VTBits);
2121 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2123 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2125 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2128 // If we are subtracting one from a positive number, there is no carry
2129 // out of the result.
2130 if (KnownZero.isNegative())
2134 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2135 if (Tmp2 == 1) return 1;
2136 return std::min(Tmp, Tmp2)-1;
2140 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2141 if (Tmp2 == 1) return 1;
2144 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
2145 if (CLHS->isNullValue()) {
2146 APInt KnownZero, KnownOne;
2147 APInt Mask = APInt::getAllOnesValue(VTBits);
2148 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2149 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2151 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2154 // If the input is known to be positive (the sign bit is known clear),
2155 // the output of the NEG has the same number of sign bits as the input.
2156 if (KnownZero.isNegative())
2159 // Otherwise, we treat this like a SUB.
2162 // Sub can have at most one carry bit. Thus we know that the output
2163 // is, at worst, one more bit than the inputs.
2164 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2165 if (Tmp == 1) return 1; // Early out.
2166 return std::min(Tmp, Tmp2)-1;
2169 // FIXME: it's tricky to do anything useful for this, but it is an important
2170 // case for targets like X86.
2174 // Handle LOADX separately here. EXTLOAD case will fallthrough.
2175 if (Op.getOpcode() == ISD::LOAD) {
2176 LoadSDNode *LD = cast<LoadSDNode>(Op);
2177 unsigned ExtType = LD->getExtensionType();
2180 case ISD::SEXTLOAD: // '17' bits known
2181 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2182 return VTBits-Tmp+1;
2183 case ISD::ZEXTLOAD: // '16' bits known
2184 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits();
2189 // Allow the target to implement this method for its nodes.
2190 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
2191 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
2192 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
2193 Op.getOpcode() == ISD::INTRINSIC_VOID) {
2194 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
2195 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
2198 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2199 // use this information.
2200 APInt KnownZero, KnownOne;
2201 APInt Mask = APInt::getAllOnesValue(VTBits);
2202 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
2204 if (KnownZero.isNegative()) { // sign bit is 0
2206 } else if (KnownOne.isNegative()) { // sign bit is 1;
2213 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2214 // the number of identical bits in the top of the input value.
2216 Mask <<= Mask.getBitWidth()-VTBits;
2217 // Return # leading zeros. We use 'min' here in case Val was zero before
2218 // shifting. We don't want to return '64' as for an i32 "0".
2219 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2222 bool SelectionDAG::isKnownNeverNaN(SDValue Op) const {
2223 // If we're told that NaNs won't happen, assume they won't.
2224 if (FiniteOnlyFPMath())
2227 // If the value is a constant, we can obviously see if it is a NaN or not.
2228 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2229 return !C->getValueAPF().isNaN();
2231 // TODO: Recognize more cases here.
2236 bool SelectionDAG::isKnownNeverZero(SDValue Op) const {
2237 // If the value is a constant, we can obviously see if it is a zero or not.
2238 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2239 return !C->isZero();
2241 // TODO: Recognize more cases here.
2246 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
2247 // Check the obvious case.
2248 if (A == B) return true;
2250 // For for negative and positive zero.
2251 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A))
2252 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B))
2253 if (CA->isZero() && CB->isZero()) return true;
2255 // Otherwise they may not be equal.
2259 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2260 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2261 if (!GA) return false;
2262 if (GA->getOffset() != 0) return false;
2263 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2264 if (!GV) return false;
2265 MachineModuleInfo *MMI = getMachineModuleInfo();
2266 return MMI && MMI->hasDebugInfo();
2270 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2271 /// element of the result of the vector shuffle.
2272 SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N,
2274 EVT VT = N->getValueType(0);
2275 DebugLoc dl = N->getDebugLoc();
2276 if (N->getMaskElt(i) < 0)
2277 return getUNDEF(VT.getVectorElementType());
2278 unsigned Index = N->getMaskElt(i);
2279 unsigned NumElems = VT.getVectorNumElements();
2280 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2283 if (V.getOpcode() == ISD::BIT_CONVERT) {
2284 V = V.getOperand(0);
2285 EVT VVT = V.getValueType();
2286 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems)
2289 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2290 return (Index == 0) ? V.getOperand(0)
2291 : getUNDEF(VT.getVectorElementType());
2292 if (V.getOpcode() == ISD::BUILD_VECTOR)
2293 return V.getOperand(Index);
2294 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V))
2295 return getShuffleScalarElt(SVN, Index);
2300 /// getNode - Gets or creates the specified node.
2302 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) {
2303 FoldingSetNodeID ID;
2304 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2306 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2307 return SDValue(E, 0);
2309 SDNode *N = NodeAllocator.Allocate<SDNode>();
2310 new (N) SDNode(Opcode, DL, getVTList(VT));
2311 CSEMap.InsertNode(N, IP);
2313 AllNodes.push_back(N);
2317 return SDValue(N, 0);
2320 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
2321 EVT VT, SDValue Operand) {
2322 // Constant fold unary operations with an integer constant operand.
2323 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2324 const APInt &Val = C->getAPIntValue();
2325 unsigned BitWidth = VT.getSizeInBits();
2328 case ISD::SIGN_EXTEND:
2329 return getConstant(APInt(Val).sextOrTrunc(BitWidth), VT);
2330 case ISD::ANY_EXTEND:
2331 case ISD::ZERO_EXTEND:
2333 return getConstant(APInt(Val).zextOrTrunc(BitWidth), VT);
2334 case ISD::UINT_TO_FP:
2335 case ISD::SINT_TO_FP: {
2336 const uint64_t zero[] = {0, 0};
2337 // No compile time operations on this type.
2338 if (VT==MVT::ppcf128)
2340 APFloat apf = APFloat(APInt(BitWidth, 2, zero));
2341 (void)apf.convertFromAPInt(Val,
2342 Opcode==ISD::SINT_TO_FP,
2343 APFloat::rmNearestTiesToEven);
2344 return getConstantFP(apf, VT);
2346 case ISD::BIT_CONVERT:
2347 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2348 return getConstantFP(Val.bitsToFloat(), VT);
2349 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2350 return getConstantFP(Val.bitsToDouble(), VT);
2353 return getConstant(Val.byteSwap(), VT);
2355 return getConstant(Val.countPopulation(), VT);
2357 return getConstant(Val.countLeadingZeros(), VT);
2359 return getConstant(Val.countTrailingZeros(), VT);
2363 // Constant fold unary operations with a floating point constant operand.
2364 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2365 APFloat V = C->getValueAPF(); // make copy
2366 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2370 return getConstantFP(V, VT);
2373 return getConstantFP(V, VT);
2375 case ISD::FP_EXTEND: {
2377 // This can return overflow, underflow, or inexact; we don't care.
2378 // FIXME need to be more flexible about rounding mode.
2379 (void)V.convert(*EVTToAPFloatSemantics(VT),
2380 APFloat::rmNearestTiesToEven, &ignored);
2381 return getConstantFP(V, VT);
2383 case ISD::FP_TO_SINT:
2384 case ISD::FP_TO_UINT: {
2387 assert(integerPartWidth >= 64);
2388 // FIXME need to be more flexible about rounding mode.
2389 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
2390 Opcode==ISD::FP_TO_SINT,
2391 APFloat::rmTowardZero, &ignored);
2392 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2394 APInt api(VT.getSizeInBits(), 2, x);
2395 return getConstant(api, VT);
2397 case ISD::BIT_CONVERT:
2398 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2399 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2400 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2401 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2407 unsigned OpOpcode = Operand.getNode()->getOpcode();
2409 case ISD::TokenFactor:
2410 case ISD::MERGE_VALUES:
2411 case ISD::CONCAT_VECTORS:
2412 return Operand; // Factor, merge or concat of one node? No need.
2413 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
2414 case ISD::FP_EXTEND:
2415 assert(VT.isFloatingPoint() &&
2416 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2417 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2418 assert((!VT.isVector() ||
2419 VT.getVectorNumElements() ==
2420 Operand.getValueType().getVectorNumElements()) &&
2421 "Vector element count mismatch!");
2422 if (Operand.getOpcode() == ISD::UNDEF)
2423 return getUNDEF(VT);
2425 case ISD::SIGN_EXTEND:
2426 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2427 "Invalid SIGN_EXTEND!");
2428 if (Operand.getValueType() == VT) return Operand; // noop extension
2429 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2430 "Invalid sext node, dst < src!");
2431 assert((!VT.isVector() ||
2432 VT.getVectorNumElements() ==
2433 Operand.getValueType().getVectorNumElements()) &&
2434 "Vector element count mismatch!");
2435 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2436 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2438 case ISD::ZERO_EXTEND:
2439 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2440 "Invalid ZERO_EXTEND!");
2441 if (Operand.getValueType() == VT) return Operand; // noop extension
2442 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2443 "Invalid zext node, dst < src!");
2444 assert((!VT.isVector() ||
2445 VT.getVectorNumElements() ==
2446 Operand.getValueType().getVectorNumElements()) &&
2447 "Vector element count mismatch!");
2448 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2449 return getNode(ISD::ZERO_EXTEND, DL, VT,
2450 Operand.getNode()->getOperand(0));
2452 case ISD::ANY_EXTEND:
2453 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2454 "Invalid ANY_EXTEND!");
2455 if (Operand.getValueType() == VT) return Operand; // noop extension
2456 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) &&
2457 "Invalid anyext 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 || OpOpcode == ISD::SIGN_EXTEND)
2463 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2464 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2467 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2468 "Invalid TRUNCATE!");
2469 if (Operand.getValueType() == VT) return Operand; // noop truncate
2470 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) &&
2471 "Invalid truncate node, src < dst!");
2472 assert((!VT.isVector() ||
2473 VT.getVectorNumElements() ==
2474 Operand.getValueType().getVectorNumElements()) &&
2475 "Vector element count mismatch!");
2476 if (OpOpcode == ISD::TRUNCATE)
2477 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2478 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2479 OpOpcode == ISD::ANY_EXTEND) {
2480 // If the source is smaller than the dest, we still need an extend.
2481 if (Operand.getNode()->getOperand(0).getValueType().getScalarType()
2482 .bitsLT(VT.getScalarType()))
2483 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2484 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2485 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2487 return Operand.getNode()->getOperand(0);
2490 case ISD::BIT_CONVERT:
2491 // Basic sanity checking.
2492 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2493 && "Cannot BIT_CONVERT between types of different sizes!");
2494 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2495 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2496 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0));
2497 if (OpOpcode == ISD::UNDEF)
2498 return getUNDEF(VT);
2500 case ISD::SCALAR_TO_VECTOR:
2501 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2502 (VT.getVectorElementType() == Operand.getValueType() ||
2503 (VT.getVectorElementType().isInteger() &&
2504 Operand.getValueType().isInteger() &&
2505 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
2506 "Illegal SCALAR_TO_VECTOR node!");
2507 if (OpOpcode == ISD::UNDEF)
2508 return getUNDEF(VT);
2509 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2510 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2511 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2512 Operand.getConstantOperandVal(1) == 0 &&
2513 Operand.getOperand(0).getValueType() == VT)
2514 return Operand.getOperand(0);
2517 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0
2518 if (UnsafeFPMath && OpOpcode == ISD::FSUB)
2519 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1),
2520 Operand.getNode()->getOperand(0));
2521 if (OpOpcode == ISD::FNEG) // --X -> X
2522 return Operand.getNode()->getOperand(0);
2525 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2526 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0));
2531 SDVTList VTs = getVTList(VT);
2532 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2533 FoldingSetNodeID ID;
2534 SDValue Ops[1] = { Operand };
2535 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2537 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2538 return SDValue(E, 0);
2540 N = NodeAllocator.Allocate<UnarySDNode>();
2541 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2542 CSEMap.InsertNode(N, IP);
2544 N = NodeAllocator.Allocate<UnarySDNode>();
2545 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2548 AllNodes.push_back(N);
2552 return SDValue(N, 0);
2555 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2557 ConstantSDNode *Cst1,
2558 ConstantSDNode *Cst2) {
2559 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2562 case ISD::ADD: return getConstant(C1 + C2, VT);
2563 case ISD::SUB: return getConstant(C1 - C2, VT);
2564 case ISD::MUL: return getConstant(C1 * C2, VT);
2566 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2569 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2572 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2575 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2577 case ISD::AND: return getConstant(C1 & C2, VT);
2578 case ISD::OR: return getConstant(C1 | C2, VT);
2579 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2580 case ISD::SHL: return getConstant(C1 << C2, VT);
2581 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2582 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2583 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2584 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2591 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2592 SDValue N1, SDValue N2) {
2593 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2594 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2597 case ISD::TokenFactor:
2598 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2599 N2.getValueType() == MVT::Other && "Invalid token factor!");
2600 // Fold trivial token factors.
2601 if (N1.getOpcode() == ISD::EntryToken) return N2;
2602 if (N2.getOpcode() == ISD::EntryToken) return N1;
2603 if (N1 == N2) return N1;
2605 case ISD::CONCAT_VECTORS:
2606 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2607 // one big BUILD_VECTOR.
2608 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2609 N2.getOpcode() == ISD::BUILD_VECTOR) {
2610 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2611 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2612 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2616 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2617 N1.getValueType() == VT && "Binary operator types must match!");
2618 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2619 // worth handling here.
2620 if (N2C && N2C->isNullValue())
2622 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2629 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2630 N1.getValueType() == VT && "Binary operator types must match!");
2631 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2632 // it's worth handling here.
2633 if (N2C && N2C->isNullValue())
2643 assert(VT.isInteger() && "This operator does not apply to FP types!");
2651 if (Opcode == ISD::FADD) {
2653 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1))
2654 if (CFP->getValueAPF().isZero())
2657 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2658 if (CFP->getValueAPF().isZero())
2660 } else if (Opcode == ISD::FSUB) {
2662 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2663 if (CFP->getValueAPF().isZero())
2667 assert(N1.getValueType() == N2.getValueType() &&
2668 N1.getValueType() == VT && "Binary operator types must match!");
2670 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2671 assert(N1.getValueType() == VT &&
2672 N1.getValueType().isFloatingPoint() &&
2673 N2.getValueType().isFloatingPoint() &&
2674 "Invalid FCOPYSIGN!");
2681 assert(VT == N1.getValueType() &&
2682 "Shift operators return type must be the same as their first arg");
2683 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2684 "Shifts only work on integers");
2686 // Always fold shifts of i1 values so the code generator doesn't need to
2687 // handle them. Since we know the size of the shift has to be less than the
2688 // size of the value, the shift/rotate count is guaranteed to be zero.
2691 if (N2C && N2C->isNullValue())
2694 case ISD::FP_ROUND_INREG: {
2695 EVT EVT = cast<VTSDNode>(N2)->getVT();
2696 assert(VT == N1.getValueType() && "Not an inreg round!");
2697 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2698 "Cannot FP_ROUND_INREG integer types");
2699 assert(EVT.isVector() == VT.isVector() &&
2700 "FP_ROUND_INREG type should be vector iff the operand "
2702 assert((!EVT.isVector() ||
2703 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2704 "Vector element counts must match in FP_ROUND_INREG");
2705 assert(EVT.bitsLE(VT) && "Not rounding down!");
2706 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2710 assert(VT.isFloatingPoint() &&
2711 N1.getValueType().isFloatingPoint() &&
2712 VT.bitsLE(N1.getValueType()) &&
2713 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2714 if (N1.getValueType() == VT) return N1; // noop conversion.
2716 case ISD::AssertSext:
2717 case ISD::AssertZext: {
2718 EVT EVT = cast<VTSDNode>(N2)->getVT();
2719 assert(VT == N1.getValueType() && "Not an inreg extend!");
2720 assert(VT.isInteger() && EVT.isInteger() &&
2721 "Cannot *_EXTEND_INREG FP types");
2722 assert(!EVT.isVector() &&
2723 "AssertSExt/AssertZExt type should be the vector element type "
2724 "rather than the vector type!");
2725 assert(EVT.bitsLE(VT) && "Not extending!");
2726 if (VT == EVT) return N1; // noop assertion.
2729 case ISD::SIGN_EXTEND_INREG: {
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() == VT.isVector() &&
2735 "SIGN_EXTEND_INREG type should be vector iff the operand "
2737 assert((!EVT.isVector() ||
2738 EVT.getVectorNumElements() == VT.getVectorNumElements()) &&
2739 "Vector element counts must match in SIGN_EXTEND_INREG");
2740 assert(EVT.bitsLE(VT) && "Not extending!");
2741 if (EVT == VT) return N1; // Not actually extending
2744 APInt Val = N1C->getAPIntValue();
2745 unsigned FromBits = EVT.getScalarType().getSizeInBits();
2746 Val <<= Val.getBitWidth()-FromBits;
2747 Val = Val.ashr(Val.getBitWidth()-FromBits);
2748 return getConstant(Val, VT);
2752 case ISD::EXTRACT_VECTOR_ELT:
2753 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2754 if (N1.getOpcode() == ISD::UNDEF)
2755 return getUNDEF(VT);
2757 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2758 // expanding copies of large vectors from registers.
2760 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2761 N1.getNumOperands() > 0) {
2763 N1.getOperand(0).getValueType().getVectorNumElements();
2764 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
2765 N1.getOperand(N2C->getZExtValue() / Factor),
2766 getConstant(N2C->getZExtValue() % Factor,
2767 N2.getValueType()));
2770 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2771 // expanding large vector constants.
2772 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) {
2773 SDValue Elt = N1.getOperand(N2C->getZExtValue());
2774 EVT VEltTy = N1.getValueType().getVectorElementType();
2775 if (Elt.getValueType() != VEltTy) {
2776 // If the vector element type is not legal, the BUILD_VECTOR operands
2777 // are promoted and implicitly truncated. Make that explicit here.
2778 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt);
2781 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT
2782 // result is implicitly extended.
2783 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt);
2788 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2789 // operations are lowered to scalars.
2790 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2791 // If the indices are the same, return the inserted element else
2792 // if the indices are known different, extract the element from
2793 // the original vector.
2794 if (N1.getOperand(2) == N2) {
2795 if (VT == N1.getOperand(1).getValueType())
2796 return N1.getOperand(1);
2798 return getSExtOrTrunc(N1.getOperand(1), DL, VT);
2799 } else if (isa<ConstantSDNode>(N1.getOperand(2)) &&
2800 isa<ConstantSDNode>(N2))
2801 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
2804 case ISD::EXTRACT_ELEMENT:
2805 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2806 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2807 (N1.getValueType().isInteger() == VT.isInteger()) &&
2808 "Wrong types for EXTRACT_ELEMENT!");
2810 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2811 // 64-bit integers into 32-bit parts. Instead of building the extract of
2812 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2813 if (N1.getOpcode() == ISD::BUILD_PAIR)
2814 return N1.getOperand(N2C->getZExtValue());
2816 // EXTRACT_ELEMENT of a constant int is also very common.
2817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2818 unsigned ElementSize = VT.getSizeInBits();
2819 unsigned Shift = ElementSize * N2C->getZExtValue();
2820 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2821 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2824 case ISD::EXTRACT_SUBVECTOR:
2825 if (N1.getValueType() == VT) // Trivial extraction.
2832 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2833 if (SV.getNode()) return SV;
2834 } else { // Cannonicalize constant to RHS if commutative
2835 if (isCommutativeBinOp(Opcode)) {
2836 std::swap(N1C, N2C);
2842 // Constant fold FP operations.
2843 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2844 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2846 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2847 // Cannonicalize constant to RHS if commutative
2848 std::swap(N1CFP, N2CFP);
2850 } else if (N2CFP && VT != MVT::ppcf128) {
2851 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2852 APFloat::opStatus s;
2855 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2856 if (s != APFloat::opInvalidOp)
2857 return getConstantFP(V1, VT);
2860 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2861 if (s!=APFloat::opInvalidOp)
2862 return getConstantFP(V1, VT);
2865 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2866 if (s!=APFloat::opInvalidOp)
2867 return getConstantFP(V1, VT);
2870 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2871 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2872 return getConstantFP(V1, VT);
2875 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2876 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2877 return getConstantFP(V1, VT);
2879 case ISD::FCOPYSIGN:
2881 return getConstantFP(V1, VT);
2887 // Canonicalize an UNDEF to the RHS, even over a constant.
2888 if (N1.getOpcode() == ISD::UNDEF) {
2889 if (isCommutativeBinOp(Opcode)) {
2893 case ISD::FP_ROUND_INREG:
2894 case ISD::SIGN_EXTEND_INREG:
2900 return N1; // fold op(undef, arg2) -> undef
2908 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2909 // For vectors, we can't easily build an all zero vector, just return
2916 // Fold a bunch of operators when the RHS is undef.
2917 if (N2.getOpcode() == ISD::UNDEF) {
2920 if (N1.getOpcode() == ISD::UNDEF)
2921 // Handle undef ^ undef -> 0 special case. This is a common
2923 return getConstant(0, VT);
2933 return N2; // fold op(arg1, undef) -> undef
2947 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2948 // For vectors, we can't easily build an all zero vector, just return
2953 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
2954 // For vectors, we can't easily build an all one vector, just return
2962 // Memoize this node if possible.
2964 SDVTList VTs = getVTList(VT);
2965 if (VT != MVT::Flag) {
2966 SDValue Ops[] = { N1, N2 };
2967 FoldingSetNodeID ID;
2968 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2970 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2971 return SDValue(E, 0);
2973 N = NodeAllocator.Allocate<BinarySDNode>();
2974 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2975 CSEMap.InsertNode(N, IP);
2977 N = NodeAllocator.Allocate<BinarySDNode>();
2978 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2981 AllNodes.push_back(N);
2985 return SDValue(N, 0);
2988 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2989 SDValue N1, SDValue N2, SDValue N3) {
2990 // Perform various simplifications.
2991 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2992 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2994 case ISD::CONCAT_VECTORS:
2995 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2996 // one big BUILD_VECTOR.
2997 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2998 N2.getOpcode() == ISD::BUILD_VECTOR &&
2999 N3.getOpcode() == ISD::BUILD_VECTOR) {
3000 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
3001 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
3002 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
3003 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
3007 // Use FoldSetCC to simplify SETCC's.
3008 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL);
3009 if (Simp.getNode()) return Simp;
3014 if (N1C->getZExtValue())
3015 return N2; // select true, X, Y -> X
3017 return N3; // select false, X, Y -> Y
3020 if (N2 == N3) return N2; // select C, X, X -> X
3024 if (N2C->getZExtValue()) // Unconditional branch
3025 return getNode(ISD::BR, DL, MVT::Other, N1, N3);
3027 return N1; // Never-taken branch
3030 case ISD::VECTOR_SHUFFLE:
3031 llvm_unreachable("should use getVectorShuffle constructor!");
3033 case ISD::BIT_CONVERT:
3034 // Fold bit_convert nodes from a type to themselves.
3035 if (N1.getValueType() == VT)
3040 // Memoize node if it doesn't produce a flag.
3042 SDVTList VTs = getVTList(VT);
3043 if (VT != MVT::Flag) {
3044 SDValue Ops[] = { N1, N2, N3 };
3045 FoldingSetNodeID ID;
3046 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3048 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3049 return SDValue(E, 0);
3051 N = NodeAllocator.Allocate<TernarySDNode>();
3052 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3053 CSEMap.InsertNode(N, IP);
3055 N = NodeAllocator.Allocate<TernarySDNode>();
3056 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
3059 AllNodes.push_back(N);
3063 return SDValue(N, 0);
3066 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3067 SDValue N1, SDValue N2, SDValue N3,
3069 SDValue Ops[] = { N1, N2, N3, N4 };
3070 return getNode(Opcode, DL, VT, Ops, 4);
3073 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3074 SDValue N1, SDValue N2, SDValue N3,
3075 SDValue N4, SDValue N5) {
3076 SDValue Ops[] = { N1, N2, N3, N4, N5 };
3077 return getNode(Opcode, DL, VT, Ops, 5);
3080 /// getStackArgumentTokenFactor - Compute a TokenFactor to force all
3081 /// the incoming stack arguments to be loaded from the stack.
3082 SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
3083 SmallVector<SDValue, 8> ArgChains;
3085 // Include the original chain at the beginning of the list. When this is
3086 // used by target LowerCall hooks, this helps legalize find the
3087 // CALLSEQ_BEGIN node.
3088 ArgChains.push_back(Chain);
3090 // Add a chain value for each stack argument.
3091 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(),
3092 UE = getEntryNode().getNode()->use_end(); U != UE; ++U)
3093 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
3094 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
3095 if (FI->getIndex() < 0)
3096 ArgChains.push_back(SDValue(L, 1));
3098 // Build a tokenfactor for all the chains.
3099 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
3100 &ArgChains[0], ArgChains.size());
3103 /// getMemsetValue - Vectorized representation of the memset value
3105 static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
3107 unsigned NumBits = VT.getScalarType().getSizeInBits();
3108 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3109 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
3111 for (unsigned i = NumBits; i > 8; i >>= 1) {
3112 Val = (Val << Shift) | Val;
3116 return DAG.getConstant(Val, VT);
3117 return DAG.getConstantFP(APFloat(Val), VT);
3120 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3121 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
3123 for (unsigned i = NumBits; i > 8; i >>= 1) {
3124 Value = DAG.getNode(ISD::OR, dl, VT,
3125 DAG.getNode(ISD::SHL, dl, VT, Value,
3126 DAG.getConstant(Shift,
3127 TLI.getShiftAmountTy())),
3135 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3136 /// used when a memcpy is turned into a memset when the source is a constant
3138 static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG,
3139 const TargetLowering &TLI,
3140 std::string &Str, unsigned Offset) {
3141 // Handle vector with all elements zero.
3144 return DAG.getConstant(0, VT);
3145 unsigned NumElts = VT.getVectorNumElements();
3146 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3147 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3149 EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts)));
3152 assert(!VT.isVector() && "Can't handle vector type here!");
3153 unsigned NumBits = VT.getSizeInBits();
3154 unsigned MSB = NumBits / 8;
3156 if (TLI.isLittleEndian())
3157 Offset = Offset + MSB - 1;
3158 for (unsigned i = 0; i != MSB; ++i) {
3159 Val = (Val << 8) | (unsigned char)Str[Offset];
3160 Offset += TLI.isLittleEndian() ? -1 : 1;
3162 return DAG.getConstant(Val, VT);
3165 /// getMemBasePlusOffset - Returns base and offset node for the
3167 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
3168 SelectionDAG &DAG) {
3169 EVT VT = Base.getValueType();
3170 return DAG.getNode(ISD::ADD, Base.getDebugLoc(),
3171 VT, Base, DAG.getConstant(Offset, VT));
3174 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
3176 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
3177 unsigned SrcDelta = 0;
3178 GlobalAddressSDNode *G = NULL;
3179 if (Src.getOpcode() == ISD::GlobalAddress)
3180 G = cast<GlobalAddressSDNode>(Src);
3181 else if (Src.getOpcode() == ISD::ADD &&
3182 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3183 Src.getOperand(1).getOpcode() == ISD::Constant) {
3184 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
3185 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
3190 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3191 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
3197 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
3198 /// to replace the memset / memcpy is below the threshold. It also returns the
3199 /// types of the sequence of memory ops to perform memset / memcpy.
3201 bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps,
3202 SDValue Dst, SDValue Src,
3203 unsigned Limit, uint64_t Size, unsigned &Align,
3204 std::string &Str, bool &isSrcStr,
3206 const TargetLowering &TLI) {
3207 isSrcStr = isMemSrcFromString(Src, Str);
3208 bool isSrcConst = isa<ConstantSDNode>(Src);
3209 EVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG);
3210 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses(VT);
3211 if (VT != MVT::iAny) {
3212 const Type *Ty = VT.getTypeForEVT(*DAG.getContext());
3213 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3214 // If source is a string constant, this will require an unaligned load.
3215 if (NewAlign > Align && (isSrcConst || AllowUnalign)) {
3216 if (Dst.getOpcode() != ISD::FrameIndex) {
3217 // Can't change destination alignment. It requires a unaligned store.
3221 int FI = cast<FrameIndexSDNode>(Dst)->getIndex();
3222 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3223 if (MFI->isFixedObjectIndex(FI)) {
3224 // Can't change destination alignment. It requires a unaligned store.
3228 // Give the stack frame object a larger alignment if needed.
3229 if (MFI->getObjectAlignment(FI) < NewAlign)
3230 MFI->setObjectAlignment(FI, NewAlign);
3237 if (VT == MVT::iAny) {
3238 if (TLI.allowsUnalignedMemoryAccesses(MVT::i64)) {
3241 switch (Align & 7) {
3242 case 0: VT = MVT::i64; break;
3243 case 4: VT = MVT::i32; break;
3244 case 2: VT = MVT::i16; break;
3245 default: VT = MVT::i8; break;
3250 while (!TLI.isTypeLegal(LVT))
3251 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1);
3252 assert(LVT.isInteger());
3258 unsigned NumMemOps = 0;
3260 unsigned VTSize = VT.getSizeInBits() / 8;
3261 while (VTSize > Size) {
3262 // For now, only use non-vector load / store's for the left-over pieces.
3263 if (VT.isVector()) {
3265 while (!TLI.isTypeLegal(VT))
3266 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3267 VTSize = VT.getSizeInBits() / 8;
3269 // This can result in a type that is not legal on the target, e.g.
3270 // 1 or 2 bytes on PPC.
3271 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3276 if (++NumMemOps > Limit)
3278 MemOps.push_back(VT);
3285 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3286 SDValue Chain, SDValue Dst,
3287 SDValue Src, uint64_t Size,
3288 unsigned Align, bool AlwaysInline,
3289 const Value *DstSV, uint64_t DstSVOff,
3290 const Value *SrcSV, uint64_t SrcSVOff){
3291 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3293 // Expand memcpy to a series of load and store ops if the size operand falls
3294 // below a certain threshold.
3295 std::vector<EVT> MemOps;
3296 uint64_t Limit = -1ULL;
3298 Limit = TLI.getMaxStoresPerMemcpy();
3299 unsigned DstAlign = Align; // Destination alignment can change.
3302 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3303 Str, CopyFromStr, DAG, TLI))
3307 bool isZeroStr = CopyFromStr && Str.empty();
3308 SmallVector<SDValue, 8> OutChains;
3309 unsigned NumMemOps = MemOps.size();
3310 uint64_t SrcOff = 0, DstOff = 0;
3311 for (unsigned i = 0; i != NumMemOps; ++i) {
3313 unsigned VTSize = VT.getSizeInBits() / 8;
3314 SDValue Value, Store;
3316 if (CopyFromStr && (isZeroStr || !VT.isVector())) {
3317 // It's unlikely a store of a vector immediate can be done in a single
3318 // instruction. It would require a load from a constantpool first.
3319 // We also handle store a vector with all zero's.
3320 // FIXME: Handle other cases where store of vector immediate is done in
3321 // a single instruction.
3322 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff);
3323 Store = DAG.getStore(Chain, dl, Value,
3324 getMemBasePlusOffset(Dst, DstOff, DAG),
3325 DstSV, DstSVOff + DstOff, false, false, DstAlign);
3327 // The type might not be legal for the target. This should only happen
3328 // if the type is smaller than a legal type, as on PPC, so the right
3329 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
3330 // to Load/Store if NVT==VT.
3331 // FIXME does the case above also need this?
3332 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
3333 assert(NVT.bitsGE(VT));
3334 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain,
3335 getMemBasePlusOffset(Src, SrcOff, DAG),
3336 SrcSV, SrcSVOff + SrcOff, VT, false, false, Align);
3337 Store = DAG.getTruncStore(Chain, dl, Value,
3338 getMemBasePlusOffset(Dst, DstOff, DAG),
3339 DstSV, DstSVOff + DstOff, VT, false, false,
3342 OutChains.push_back(Store);
3347 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3348 &OutChains[0], OutChains.size());
3351 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3352 SDValue Chain, SDValue Dst,
3353 SDValue Src, uint64_t Size,
3354 unsigned Align, bool AlwaysInline,
3355 const Value *DstSV, uint64_t DstSVOff,
3356 const Value *SrcSV, uint64_t SrcSVOff){
3357 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3359 // Expand memmove to a series of load and store ops if the size operand falls
3360 // below a certain threshold.
3361 std::vector<EVT> MemOps;
3362 uint64_t Limit = -1ULL;
3364 Limit = TLI.getMaxStoresPerMemmove();
3365 unsigned DstAlign = Align; // Destination alignment can change.
3368 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3369 Str, CopyFromStr, DAG, TLI))
3372 uint64_t SrcOff = 0, DstOff = 0;
3374 SmallVector<SDValue, 8> LoadValues;
3375 SmallVector<SDValue, 8> LoadChains;
3376 SmallVector<SDValue, 8> OutChains;
3377 unsigned NumMemOps = MemOps.size();
3378 for (unsigned i = 0; i < NumMemOps; i++) {
3380 unsigned VTSize = VT.getSizeInBits() / 8;
3381 SDValue Value, Store;
3383 Value = DAG.getLoad(VT, dl, Chain,
3384 getMemBasePlusOffset(Src, SrcOff, DAG),
3385 SrcSV, SrcSVOff + SrcOff, false, false, Align);
3386 LoadValues.push_back(Value);
3387 LoadChains.push_back(Value.getValue(1));
3390 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3391 &LoadChains[0], LoadChains.size());
3393 for (unsigned i = 0; i < NumMemOps; i++) {
3395 unsigned VTSize = VT.getSizeInBits() / 8;
3396 SDValue Value, Store;
3398 Store = DAG.getStore(Chain, dl, LoadValues[i],
3399 getMemBasePlusOffset(Dst, DstOff, DAG),
3400 DstSV, DstSVOff + DstOff, false, false, DstAlign);
3401 OutChains.push_back(Store);
3405 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3406 &OutChains[0], OutChains.size());
3409 static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl,
3410 SDValue Chain, SDValue Dst,
3411 SDValue Src, uint64_t Size,
3413 const Value *DstSV, uint64_t DstSVOff) {
3414 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3416 // Expand memset to a series of load/store ops if the size operand
3417 // falls below a certain threshold.
3418 std::vector<EVT> MemOps;
3421 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(),
3422 Size, Align, Str, CopyFromStr, DAG, TLI))
3425 SmallVector<SDValue, 8> OutChains;
3426 uint64_t DstOff = 0;
3428 unsigned NumMemOps = MemOps.size();
3429 for (unsigned i = 0; i < NumMemOps; i++) {
3431 unsigned VTSize = VT.getSizeInBits() / 8;
3432 SDValue Value = getMemsetValue(Src, VT, DAG, dl);
3433 SDValue Store = DAG.getStore(Chain, dl, Value,
3434 getMemBasePlusOffset(Dst, DstOff, DAG),
3435 DstSV, DstSVOff + DstOff, false, false, 0);
3436 OutChains.push_back(Store);
3440 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3441 &OutChains[0], OutChains.size());
3444 SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst,
3445 SDValue Src, SDValue Size,
3446 unsigned Align, bool AlwaysInline,
3447 const Value *DstSV, uint64_t DstSVOff,
3448 const Value *SrcSV, uint64_t SrcSVOff) {
3450 // Check to see if we should lower the memcpy to loads and stores first.
3451 // For cases within the target-specified limits, this is the best choice.
3452 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3454 // Memcpy with size zero? Just return the original chain.
3455 if (ConstantSize->isNullValue())
3459 getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3460 ConstantSize->getZExtValue(),
3461 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3462 if (Result.getNode())
3466 // Then check to see if we should lower the memcpy with target-specific
3467 // code. If the target chooses to do this, this is the next best.
3469 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align,
3471 DstSV, DstSVOff, SrcSV, SrcSVOff);
3472 if (Result.getNode())
3475 // If we really need inline code and the target declined to provide it,
3476 // use a (potentially long) sequence of loads and stores.
3478 assert(ConstantSize && "AlwaysInline requires a constant size!");
3479 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3480 ConstantSize->getZExtValue(), Align, true,
3481 DstSV, DstSVOff, SrcSV, SrcSVOff);
3484 // Emit a library call.
3485 TargetLowering::ArgListTy Args;
3486 TargetLowering::ArgListEntry Entry;
3487 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3488 Entry.Node = Dst; Args.push_back(Entry);
3489 Entry.Node = Src; Args.push_back(Entry);
3490 Entry.Node = Size; Args.push_back(Entry);
3491 // FIXME: pass in DebugLoc
3492 std::pair<SDValue,SDValue> CallResult =
3493 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3494 false, false, false, false, 0,
3495 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false,
3496 /*isReturnValueUsed=*/false,
3497 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY),
3498 TLI.getPointerTy()),
3499 Args, *this, dl, GetOrdering(Chain.getNode()));
3500 return CallResult.second;
3503 SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst,
3504 SDValue Src, SDValue Size,
3506 const Value *DstSV, uint64_t DstSVOff,
3507 const Value *SrcSV, uint64_t SrcSVOff) {
3509 // Check to see if we should lower the memmove to loads and stores first.
3510 // For cases within the target-specified limits, this is the best choice.
3511 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3513 // Memmove with size zero? Just return the original chain.
3514 if (ConstantSize->isNullValue())
3518 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src,
3519 ConstantSize->getZExtValue(),
3520 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3521 if (Result.getNode())
3525 // Then check to see if we should lower the memmove with target-specific
3526 // code. If the target chooses to do this, this is the next best.
3528 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align,
3529 DstSV, DstSVOff, SrcSV, SrcSVOff);
3530 if (Result.getNode())
3533 // Emit a library call.
3534 TargetLowering::ArgListTy Args;
3535 TargetLowering::ArgListEntry Entry;
3536 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3537 Entry.Node = Dst; Args.push_back(Entry);
3538 Entry.Node = Src; Args.push_back(Entry);
3539 Entry.Node = Size; Args.push_back(Entry);
3540 // FIXME: pass in DebugLoc
3541 std::pair<SDValue,SDValue> CallResult =
3542 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3543 false, false, false, false, 0,
3544 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false,
3545 /*isReturnValueUsed=*/false,
3546 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE),
3547 TLI.getPointerTy()),
3548 Args, *this, dl, GetOrdering(Chain.getNode()));
3549 return CallResult.second;
3552 SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst,
3553 SDValue Src, SDValue Size,
3555 const Value *DstSV, uint64_t DstSVOff) {
3557 // Check to see if we should lower the memset to stores first.
3558 // For cases within the target-specified limits, this is the best choice.
3559 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3561 // Memset with size zero? Just return the original chain.
3562 if (ConstantSize->isNullValue())
3566 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
3567 Align, DstSV, DstSVOff);
3568 if (Result.getNode())
3572 // Then check to see if we should lower the memset with target-specific
3573 // code. If the target chooses to do this, this is the next best.
3575 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align,
3577 if (Result.getNode())
3580 // Emit a library call.
3581 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext());
3582 TargetLowering::ArgListTy Args;
3583 TargetLowering::ArgListEntry Entry;
3584 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3585 Args.push_back(Entry);
3586 // Extend or truncate the argument to be an i32 value for the call.
3587 if (Src.getValueType().bitsGT(MVT::i32))
3588 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
3590 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
3592 Entry.Ty = Type::getInt32Ty(*getContext());
3593 Entry.isSExt = true;
3594 Args.push_back(Entry);
3596 Entry.Ty = IntPtrTy;
3597 Entry.isSExt = false;
3598 Args.push_back(Entry);
3599 // FIXME: pass in DebugLoc
3600 std::pair<SDValue,SDValue> CallResult =
3601 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3602 false, false, false, false, 0,
3603 TLI.getLibcallCallingConv(RTLIB::MEMSET), false,
3604 /*isReturnValueUsed=*/false,
3605 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET),
3606 TLI.getPointerTy()),
3607 Args, *this, dl, GetOrdering(Chain.getNode()));
3608 return CallResult.second;
3611 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3613 SDValue Ptr, SDValue Cmp,
3614 SDValue Swp, const Value* PtrVal,
3615 unsigned Alignment) {
3616 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3617 Alignment = getEVTAlignment(MemVT);
3619 // Check if the memory reference references a frame index
3621 if (const FrameIndexSDNode *FI =
3622 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3623 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3625 MachineFunction &MF = getMachineFunction();
3626 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3628 // For now, atomics are considered to be volatile always.
3629 Flags |= MachineMemOperand::MOVolatile;
3631 MachineMemOperand *MMO =
3632 MF.getMachineMemOperand(PtrVal, Flags, 0,
3633 MemVT.getStoreSize(), Alignment);
3635 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3638 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3640 SDValue Ptr, SDValue Cmp,
3641 SDValue Swp, MachineMemOperand *MMO) {
3642 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op");
3643 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3645 EVT VT = Cmp.getValueType();
3647 SDVTList VTs = getVTList(VT, MVT::Other);
3648 FoldingSetNodeID ID;
3649 ID.AddInteger(MemVT.getRawBits());
3650 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3651 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3653 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3654 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3655 return SDValue(E, 0);
3657 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3658 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3659 CSEMap.InsertNode(N, IP);
3660 AllNodes.push_back(N);
3661 return SDValue(N, 0);
3664 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3666 SDValue Ptr, SDValue Val,
3667 const Value* PtrVal,
3668 unsigned Alignment) {
3669 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3670 Alignment = getEVTAlignment(MemVT);
3672 // Check if the memory reference references a frame index
3674 if (const FrameIndexSDNode *FI =
3675 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3676 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3678 MachineFunction &MF = getMachineFunction();
3679 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3681 // For now, atomics are considered to be volatile always.
3682 Flags |= MachineMemOperand::MOVolatile;
3684 MachineMemOperand *MMO =
3685 MF.getMachineMemOperand(PtrVal, Flags, 0,
3686 MemVT.getStoreSize(), Alignment);
3688 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO);
3691 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3693 SDValue Ptr, SDValue Val,
3694 MachineMemOperand *MMO) {
3695 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
3696 Opcode == ISD::ATOMIC_LOAD_SUB ||
3697 Opcode == ISD::ATOMIC_LOAD_AND ||
3698 Opcode == ISD::ATOMIC_LOAD_OR ||
3699 Opcode == ISD::ATOMIC_LOAD_XOR ||
3700 Opcode == ISD::ATOMIC_LOAD_NAND ||
3701 Opcode == ISD::ATOMIC_LOAD_MIN ||
3702 Opcode == ISD::ATOMIC_LOAD_MAX ||
3703 Opcode == ISD::ATOMIC_LOAD_UMIN ||
3704 Opcode == ISD::ATOMIC_LOAD_UMAX ||
3705 Opcode == ISD::ATOMIC_SWAP) &&
3706 "Invalid Atomic Op");
3708 EVT VT = Val.getValueType();
3710 SDVTList VTs = getVTList(VT, MVT::Other);
3711 FoldingSetNodeID ID;
3712 ID.AddInteger(MemVT.getRawBits());
3713 SDValue Ops[] = {Chain, Ptr, Val};
3714 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3716 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3717 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3718 return SDValue(E, 0);
3720 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3721 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Val, MMO);
3722 CSEMap.InsertNode(N, IP);
3723 AllNodes.push_back(N);
3724 return SDValue(N, 0);
3727 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3728 /// Allowed to return something different (and simpler) if Simplify is true.
3729 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3734 SmallVector<EVT, 4> VTs;
3735 VTs.reserve(NumOps);
3736 for (unsigned i = 0; i < NumOps; ++i)
3737 VTs.push_back(Ops[i].getValueType());
3738 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps),
3743 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl,
3744 const EVT *VTs, unsigned NumVTs,
3745 const SDValue *Ops, unsigned NumOps,
3746 EVT MemVT, const Value *srcValue, int SVOff,
3747 unsigned Align, bool Vol,
3748 bool ReadMem, bool WriteMem) {
3749 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps,
3750 MemVT, srcValue, SVOff, Align, Vol,
3755 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3756 const SDValue *Ops, unsigned NumOps,
3757 EVT MemVT, const Value *srcValue, int SVOff,
3758 unsigned Align, bool Vol,
3759 bool ReadMem, bool WriteMem) {
3760 if (Align == 0) // Ensure that codegen never sees alignment 0
3761 Align = getEVTAlignment(MemVT);
3763 MachineFunction &MF = getMachineFunction();
3766 Flags |= MachineMemOperand::MOStore;
3768 Flags |= MachineMemOperand::MOLoad;
3770 Flags |= MachineMemOperand::MOVolatile;
3771 MachineMemOperand *MMO =
3772 MF.getMachineMemOperand(srcValue, Flags, SVOff,
3773 MemVT.getStoreSize(), Align);
3775 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3779 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3780 const SDValue *Ops, unsigned NumOps,
3781 EVT MemVT, MachineMemOperand *MMO) {
3782 assert((Opcode == ISD::INTRINSIC_VOID ||
3783 Opcode == ISD::INTRINSIC_W_CHAIN ||
3784 (Opcode <= INT_MAX &&
3785 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) &&
3786 "Opcode is not a memory-accessing opcode!");
3788 // Memoize the node unless it returns a flag.
3789 MemIntrinsicSDNode *N;
3790 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3791 FoldingSetNodeID ID;
3792 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3794 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3795 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO);
3796 return SDValue(E, 0);
3799 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3800 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3801 CSEMap.InsertNode(N, IP);
3803 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3804 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3806 AllNodes.push_back(N);
3807 return SDValue(N, 0);
3811 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3812 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3813 SDValue Ptr, SDValue Offset,
3814 const Value *SV, int SVOffset, EVT MemVT,
3815 bool isVolatile, bool isNonTemporal,
3816 unsigned Alignment) {
3817 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3818 Alignment = getEVTAlignment(VT);
3820 // Check if the memory reference references a frame index
3822 if (const FrameIndexSDNode *FI =
3823 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3824 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3826 MachineFunction &MF = getMachineFunction();
3827 unsigned Flags = MachineMemOperand::MOLoad;
3829 Flags |= MachineMemOperand::MOVolatile;
3831 Flags |= MachineMemOperand::MONonTemporal;
3832 MachineMemOperand *MMO =
3833 MF.getMachineMemOperand(SV, Flags, SVOffset,
3834 MemVT.getStoreSize(), Alignment);
3835 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO);
3839 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3840 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3841 SDValue Ptr, SDValue Offset, EVT MemVT,
3842 MachineMemOperand *MMO) {
3844 ExtType = ISD::NON_EXTLOAD;
3845 } else if (ExtType == ISD::NON_EXTLOAD) {
3846 assert(VT == MemVT && "Non-extending load from different memory type!");
3849 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) &&
3850 "Should only be an extending load, not truncating!");
3851 assert(VT.isInteger() == MemVT.isInteger() &&
3852 "Cannot convert from FP to Int or Int -> FP!");
3853 assert(VT.isVector() == MemVT.isVector() &&
3854 "Cannot use trunc store to convert to or from a vector!");
3855 assert((!VT.isVector() ||
3856 VT.getVectorNumElements() == MemVT.getVectorNumElements()) &&
3857 "Cannot use trunc store to change the number of vector elements!");
3860 bool Indexed = AM != ISD::UNINDEXED;
3861 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3862 "Unindexed load with an offset!");
3864 SDVTList VTs = Indexed ?
3865 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3866 SDValue Ops[] = { Chain, Ptr, Offset };
3867 FoldingSetNodeID ID;
3868 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3869 ID.AddInteger(MemVT.getRawBits());
3870 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile(),
3871 MMO->isNonTemporal()));
3873 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3874 cast<LoadSDNode>(E)->refineAlignment(MMO);
3875 return SDValue(E, 0);
3877 SDNode *N = NodeAllocator.Allocate<LoadSDNode>();
3878 new (N) LoadSDNode(Ops, dl, VTs, AM, ExtType, MemVT, MMO);
3879 CSEMap.InsertNode(N, IP);
3880 AllNodes.push_back(N);
3881 return SDValue(N, 0);
3884 SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl,
3885 SDValue Chain, SDValue Ptr,
3886 const Value *SV, int SVOffset,
3887 bool isVolatile, bool isNonTemporal,
3888 unsigned Alignment) {
3889 SDValue Undef = getUNDEF(Ptr.getValueType());
3890 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3891 SV, SVOffset, VT, isVolatile, isNonTemporal, Alignment);
3894 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT,
3895 SDValue Chain, SDValue Ptr,
3897 int SVOffset, EVT MemVT,
3898 bool isVolatile, bool isNonTemporal,
3899 unsigned Alignment) {
3900 SDValue Undef = getUNDEF(Ptr.getValueType());
3901 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef,
3902 SV, SVOffset, MemVT, isVolatile, isNonTemporal, Alignment);
3906 SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base,
3907 SDValue Offset, ISD::MemIndexedMode AM) {
3908 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
3909 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
3910 "Load is already a indexed load!");
3911 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(),
3912 LD->getChain(), Base, Offset, LD->getSrcValue(),
3913 LD->getSrcValueOffset(), LD->getMemoryVT(),
3914 LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment());
3917 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3918 SDValue Ptr, const Value *SV, int SVOffset,
3919 bool isVolatile, bool isNonTemporal,
3920 unsigned Alignment) {
3921 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3922 Alignment = getEVTAlignment(Val.getValueType());
3924 // Check if the memory reference references a frame index
3926 if (const FrameIndexSDNode *FI =
3927 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3928 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3930 MachineFunction &MF = getMachineFunction();
3931 unsigned Flags = MachineMemOperand::MOStore;
3933 Flags |= MachineMemOperand::MOVolatile;
3935 Flags |= MachineMemOperand::MONonTemporal;
3936 MachineMemOperand *MMO =
3937 MF.getMachineMemOperand(SV, Flags, SVOffset,
3938 Val.getValueType().getStoreSize(), Alignment);
3940 return getStore(Chain, dl, Val, Ptr, MMO);
3943 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3944 SDValue Ptr, MachineMemOperand *MMO) {
3945 EVT VT = Val.getValueType();
3946 SDVTList VTs = getVTList(MVT::Other);
3947 SDValue Undef = getUNDEF(Ptr.getValueType());
3948 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3949 FoldingSetNodeID ID;
3950 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3951 ID.AddInteger(VT.getRawBits());
3952 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile(),
3953 MMO->isNonTemporal()));
3955 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3956 cast<StoreSDNode>(E)->refineAlignment(MMO);
3957 return SDValue(E, 0);
3959 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3960 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, false, VT, MMO);
3961 CSEMap.InsertNode(N, IP);
3962 AllNodes.push_back(N);
3963 return SDValue(N, 0);
3966 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3967 SDValue Ptr, const Value *SV,
3968 int SVOffset, EVT SVT,
3969 bool isVolatile, bool isNonTemporal,
3970 unsigned Alignment) {
3971 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3972 Alignment = getEVTAlignment(SVT);
3974 // Check if the memory reference references a frame index
3976 if (const FrameIndexSDNode *FI =
3977 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3978 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3980 MachineFunction &MF = getMachineFunction();
3981 unsigned Flags = MachineMemOperand::MOStore;
3983 Flags |= MachineMemOperand::MOVolatile;
3985 Flags |= MachineMemOperand::MONonTemporal;
3986 MachineMemOperand *MMO =
3987 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment);
3989 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
3992 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3993 SDValue Ptr, EVT SVT,
3994 MachineMemOperand *MMO) {
3995 EVT VT = Val.getValueType();
3998 return getStore(Chain, dl, Val, Ptr, MMO);
4000 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
4001 "Should only be a truncating store, not extending!");
4002 assert(VT.isInteger() == SVT.isInteger() &&
4003 "Can't do FP-INT conversion!");
4004 assert(VT.isVector() == SVT.isVector() &&
4005 "Cannot use trunc store to convert to or from a vector!");
4006 assert((!VT.isVector() ||
4007 VT.getVectorNumElements() == SVT.getVectorNumElements()) &&
4008 "Cannot use trunc store to change the number of vector elements!");
4010 SDVTList VTs = getVTList(MVT::Other);
4011 SDValue Undef = getUNDEF(Ptr.getValueType());
4012 SDValue Ops[] = { Chain, Val, Ptr, Undef };
4013 FoldingSetNodeID ID;
4014 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4015 ID.AddInteger(SVT.getRawBits());
4016 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile(),
4017 MMO->isNonTemporal()));
4019 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
4020 cast<StoreSDNode>(E)->refineAlignment(MMO);
4021 return SDValue(E, 0);
4023 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
4024 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, true, SVT, MMO);
4025 CSEMap.InsertNode(N, IP);
4026 AllNodes.push_back(N);
4027 return SDValue(N, 0);
4031 SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base,
4032 SDValue Offset, ISD::MemIndexedMode AM) {
4033 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
4034 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
4035 "Store is already a indexed store!");
4036 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
4037 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
4038 FoldingSetNodeID ID;
4039 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
4040 ID.AddInteger(ST->getMemoryVT().getRawBits());
4041 ID.AddInteger(ST->getRawSubclassData());
4043 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4044 return SDValue(E, 0);
4046 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
4047 new (N) StoreSDNode(Ops, dl, VTs, AM,
4048 ST->isTruncatingStore(), ST->getMemoryVT(),
4049 ST->getMemOperand());
4050 CSEMap.InsertNode(N, IP);
4051 AllNodes.push_back(N);
4052 return SDValue(N, 0);
4055 SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl,
4056 SDValue Chain, SDValue Ptr,
4058 SDValue Ops[] = { Chain, Ptr, SV };
4059 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3);
4062 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4063 const SDUse *Ops, unsigned NumOps) {
4065 case 0: return getNode(Opcode, DL, VT);
4066 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4067 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4068 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4072 // Copy from an SDUse array into an SDValue array for use with
4073 // the regular getNode logic.
4074 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
4075 return getNode(Opcode, DL, VT, &NewOps[0], NumOps);
4078 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
4079 const SDValue *Ops, unsigned NumOps) {
4081 case 0: return getNode(Opcode, DL, VT);
4082 case 1: return getNode(Opcode, DL, VT, Ops[0]);
4083 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
4084 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
4090 case ISD::SELECT_CC: {
4091 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
4092 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
4093 "LHS and RHS of condition must have same type!");
4094 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4095 "True and False arms of SelectCC must have same type!");
4096 assert(Ops[2].getValueType() == VT &&
4097 "select_cc node must be of same type as true and false value!");
4101 assert(NumOps == 5 && "BR_CC takes 5 operands!");
4102 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4103 "LHS/RHS of comparison should match types!");
4110 SDVTList VTs = getVTList(VT);
4112 if (VT != MVT::Flag) {
4113 FoldingSetNodeID ID;
4114 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
4117 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4118 return SDValue(E, 0);
4120 N = NodeAllocator.Allocate<SDNode>();
4121 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
4122 CSEMap.InsertNode(N, IP);
4124 N = NodeAllocator.Allocate<SDNode>();
4125 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
4128 AllNodes.push_back(N);
4132 return SDValue(N, 0);
4135 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4136 const std::vector<EVT> &ResultTys,
4137 const SDValue *Ops, unsigned NumOps) {
4138 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()),
4142 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4143 const EVT *VTs, unsigned NumVTs,
4144 const SDValue *Ops, unsigned NumOps) {
4146 return getNode(Opcode, DL, VTs[0], Ops, NumOps);
4147 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps);
4150 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4151 const SDValue *Ops, unsigned NumOps) {
4152 if (VTList.NumVTs == 1)
4153 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps);
4157 // FIXME: figure out how to safely handle things like
4158 // int foo(int x) { return 1 << (x & 255); }
4159 // int bar() { return foo(256); }
4160 case ISD::SRA_PARTS:
4161 case ISD::SRL_PARTS:
4162 case ISD::SHL_PARTS:
4163 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4164 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
4165 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4166 else if (N3.getOpcode() == ISD::AND)
4167 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
4168 // If the and is only masking out bits that cannot effect the shift,
4169 // eliminate the and.
4170 unsigned NumBits = VT.getScalarType().getSizeInBits()*2;
4171 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
4172 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4178 // Memoize the node unless it returns a flag.
4180 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4181 FoldingSetNodeID ID;
4182 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4184 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4185 return SDValue(E, 0);
4188 N = NodeAllocator.Allocate<UnarySDNode>();
4189 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4190 } else if (NumOps == 2) {
4191 N = NodeAllocator.Allocate<BinarySDNode>();
4192 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4193 } else if (NumOps == 3) {
4194 N = NodeAllocator.Allocate<TernarySDNode>();
4195 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
4197 N = NodeAllocator.Allocate<SDNode>();
4198 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
4200 CSEMap.InsertNode(N, IP);
4203 N = NodeAllocator.Allocate<UnarySDNode>();
4204 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4205 } else if (NumOps == 2) {
4206 N = NodeAllocator.Allocate<BinarySDNode>();
4207 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4208 } else if (NumOps == 3) {
4209 N = NodeAllocator.Allocate<TernarySDNode>();
4210 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
4212 N = NodeAllocator.Allocate<SDNode>();
4213 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
4216 AllNodes.push_back(N);
4220 return SDValue(N, 0);
4223 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) {
4224 return getNode(Opcode, DL, VTList, 0, 0);
4227 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4229 SDValue Ops[] = { N1 };
4230 return getNode(Opcode, DL, VTList, Ops, 1);
4233 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4234 SDValue N1, SDValue N2) {
4235 SDValue Ops[] = { N1, N2 };
4236 return getNode(Opcode, DL, VTList, Ops, 2);
4239 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4240 SDValue N1, SDValue N2, SDValue N3) {
4241 SDValue Ops[] = { N1, N2, N3 };
4242 return getNode(Opcode, DL, VTList, Ops, 3);
4245 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4246 SDValue N1, SDValue N2, SDValue N3,
4248 SDValue Ops[] = { N1, N2, N3, N4 };
4249 return getNode(Opcode, DL, VTList, Ops, 4);
4252 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4253 SDValue N1, SDValue N2, SDValue N3,
4254 SDValue N4, SDValue N5) {
4255 SDValue Ops[] = { N1, N2, N3, N4, N5 };
4256 return getNode(Opcode, DL, VTList, Ops, 5);
4259 SDVTList SelectionDAG::getVTList(EVT VT) {
4260 return makeVTList(SDNode::getValueTypeList(VT), 1);
4263 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
4264 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4265 E = VTList.rend(); I != E; ++I)
4266 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
4269 EVT *Array = Allocator.Allocate<EVT>(2);
4272 SDVTList Result = makeVTList(Array, 2);
4273 VTList.push_back(Result);
4277 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
4278 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4279 E = VTList.rend(); I != E; ++I)
4280 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4284 EVT *Array = Allocator.Allocate<EVT>(3);
4288 SDVTList Result = makeVTList(Array, 3);
4289 VTList.push_back(Result);
4293 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
4294 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4295 E = VTList.rend(); I != E; ++I)
4296 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4297 I->VTs[2] == VT3 && I->VTs[3] == VT4)
4300 EVT *Array = Allocator.Allocate<EVT>(4);
4305 SDVTList Result = makeVTList(Array, 4);
4306 VTList.push_back(Result);
4310 SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) {
4312 case 0: llvm_unreachable("Cannot have nodes without results!");
4313 case 1: return getVTList(VTs[0]);
4314 case 2: return getVTList(VTs[0], VTs[1]);
4315 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
4316 case 4: return getVTList(VTs[0], VTs[1], VTs[2], VTs[3]);
4320 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4321 E = VTList.rend(); I != E; ++I) {
4322 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
4325 bool NoMatch = false;
4326 for (unsigned i = 2; i != NumVTs; ++i)
4327 if (VTs[i] != I->VTs[i]) {
4335 EVT *Array = Allocator.Allocate<EVT>(NumVTs);
4336 std::copy(VTs, VTs+NumVTs, Array);
4337 SDVTList Result = makeVTList(Array, NumVTs);
4338 VTList.push_back(Result);
4343 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
4344 /// specified operands. If the resultant node already exists in the DAG,
4345 /// this does not modify the specified node, instead it returns the node that
4346 /// already exists. If the resultant node does not exist in the DAG, the
4347 /// input node is returned. As a degenerate case, if you specify the same
4348 /// input operands as the node already has, the input node is returned.
4349 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
4350 SDNode *N = InN.getNode();
4351 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
4353 // Check to see if there is no change.
4354 if (Op == N->getOperand(0)) return InN;
4356 // See if the modified node already exists.
4357 void *InsertPos = 0;
4358 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
4359 return SDValue(Existing, InN.getResNo());
4361 // Nope it doesn't. Remove the node from its current place in the maps.
4363 if (!RemoveNodeFromCSEMaps(N))
4366 // Now we update the operands.
4367 N->OperandList[0].set(Op);
4369 // If this gets put into a CSE map, add it.
4370 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4374 SDValue SelectionDAG::
4375 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
4376 SDNode *N = InN.getNode();
4377 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
4379 // Check to see if there is no change.
4380 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
4381 return InN; // No operands changed, just return the input node.
4383 // See if the modified node already exists.
4384 void *InsertPos = 0;
4385 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
4386 return SDValue(Existing, InN.getResNo());
4388 // Nope it doesn't. Remove the node from its current place in the maps.
4390 if (!RemoveNodeFromCSEMaps(N))
4393 // Now we update the operands.
4394 if (N->OperandList[0] != Op1)
4395 N->OperandList[0].set(Op1);
4396 if (N->OperandList[1] != Op2)
4397 N->OperandList[1].set(Op2);
4399 // If this gets put into a CSE map, add it.
4400 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4404 SDValue SelectionDAG::
4405 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
4406 SDValue Ops[] = { Op1, Op2, Op3 };
4407 return UpdateNodeOperands(N, Ops, 3);
4410 SDValue SelectionDAG::
4411 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4412 SDValue Op3, SDValue Op4) {
4413 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
4414 return UpdateNodeOperands(N, Ops, 4);
4417 SDValue SelectionDAG::
4418 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4419 SDValue Op3, SDValue Op4, SDValue Op5) {
4420 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
4421 return UpdateNodeOperands(N, Ops, 5);
4424 SDValue SelectionDAG::
4425 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
4426 SDNode *N = InN.getNode();
4427 assert(N->getNumOperands() == NumOps &&
4428 "Update with wrong number of operands");
4430 // Check to see if there is no change.
4431 bool AnyChange = false;
4432 for (unsigned i = 0; i != NumOps; ++i) {
4433 if (Ops[i] != N->getOperand(i)) {
4439 // No operands changed, just return the input node.
4440 if (!AnyChange) return InN;
4442 // See if the modified node already exists.
4443 void *InsertPos = 0;
4444 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
4445 return SDValue(Existing, InN.getResNo());
4447 // Nope it doesn't. Remove the node from its current place in the maps.
4449 if (!RemoveNodeFromCSEMaps(N))
4452 // Now we update the operands.
4453 for (unsigned i = 0; i != NumOps; ++i)
4454 if (N->OperandList[i] != Ops[i])
4455 N->OperandList[i].set(Ops[i]);
4457 // If this gets put into a CSE map, add it.
4458 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4462 /// DropOperands - Release the operands and set this node to have
4464 void SDNode::DropOperands() {
4465 // Unlike the code in MorphNodeTo that does this, we don't need to
4466 // watch for dead nodes here.
4467 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
4473 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
4476 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4478 SDVTList VTs = getVTList(VT);
4479 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4482 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4483 EVT VT, SDValue Op1) {
4484 SDVTList VTs = getVTList(VT);
4485 SDValue Ops[] = { Op1 };
4486 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4489 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4490 EVT VT, SDValue Op1,
4492 SDVTList VTs = getVTList(VT);
4493 SDValue Ops[] = { Op1, Op2 };
4494 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4497 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4498 EVT VT, SDValue Op1,
4499 SDValue Op2, SDValue Op3) {
4500 SDVTList VTs = getVTList(VT);
4501 SDValue Ops[] = { Op1, Op2, Op3 };
4502 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4505 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4506 EVT VT, const SDValue *Ops,
4508 SDVTList VTs = getVTList(VT);
4509 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4512 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4513 EVT VT1, EVT VT2, const SDValue *Ops,
4515 SDVTList VTs = getVTList(VT1, VT2);
4516 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4519 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4521 SDVTList VTs = getVTList(VT1, VT2);
4522 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4525 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4526 EVT VT1, EVT VT2, EVT VT3,
4527 const SDValue *Ops, unsigned NumOps) {
4528 SDVTList VTs = getVTList(VT1, VT2, VT3);
4529 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4532 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4533 EVT VT1, EVT VT2, EVT VT3, EVT VT4,
4534 const SDValue *Ops, unsigned NumOps) {
4535 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4536 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4539 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4542 SDVTList VTs = getVTList(VT1, VT2);
4543 SDValue Ops[] = { Op1 };
4544 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4547 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4549 SDValue Op1, SDValue Op2) {
4550 SDVTList VTs = getVTList(VT1, VT2);
4551 SDValue Ops[] = { Op1, Op2 };
4552 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4555 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4557 SDValue Op1, SDValue Op2,
4559 SDVTList VTs = getVTList(VT1, VT2);
4560 SDValue Ops[] = { Op1, Op2, Op3 };
4561 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4564 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4565 EVT VT1, EVT VT2, EVT VT3,
4566 SDValue Op1, SDValue Op2,
4568 SDVTList VTs = getVTList(VT1, VT2, VT3);
4569 SDValue Ops[] = { Op1, Op2, Op3 };
4570 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4573 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4574 SDVTList VTs, const SDValue *Ops,
4576 N = MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4577 // Reset the NodeID to -1.
4582 /// MorphNodeTo - This *mutates* the specified node to have the specified
4583 /// return type, opcode, and operands.
4585 /// Note that MorphNodeTo returns the resultant node. If there is already a
4586 /// node of the specified opcode and operands, it returns that node instead of
4587 /// the current one. Note that the DebugLoc need not be the same.
4589 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4590 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4591 /// node, and because it doesn't require CSE recalculation for any of
4592 /// the node's users.
4594 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4595 SDVTList VTs, const SDValue *Ops,
4597 // If an identical node already exists, use it.
4599 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4600 FoldingSetNodeID ID;
4601 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4602 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4606 if (!RemoveNodeFromCSEMaps(N))
4609 // Start the morphing.
4611 N->ValueList = VTs.VTs;
4612 N->NumValues = VTs.NumVTs;
4614 // Clear the operands list, updating used nodes to remove this from their
4615 // use list. Keep track of any operands that become dead as a result.
4616 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4617 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
4619 SDNode *Used = Use.getNode();
4621 if (Used->use_empty())
4622 DeadNodeSet.insert(Used);
4625 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) {
4626 // Initialize the memory references information.
4627 MN->setMemRefs(0, 0);
4628 // If NumOps is larger than the # of operands we can have in a
4629 // MachineSDNode, reallocate the operand list.
4630 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) {
4631 if (MN->OperandsNeedDelete)
4632 delete[] MN->OperandList;
4633 if (NumOps > array_lengthof(MN->LocalOperands))
4634 // We're creating a final node that will live unmorphed for the
4635 // remainder of the current SelectionDAG iteration, so we can allocate
4636 // the operands directly out of a pool with no recycling metadata.
4637 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4640 MN->InitOperands(MN->LocalOperands, Ops, NumOps);
4641 MN->OperandsNeedDelete = false;
4643 MN->InitOperands(MN->OperandList, Ops, NumOps);
4645 // If NumOps is larger than the # of operands we currently have, reallocate
4646 // the operand list.
4647 if (NumOps > N->NumOperands) {
4648 if (N->OperandsNeedDelete)
4649 delete[] N->OperandList;
4650 N->InitOperands(new SDUse[NumOps], Ops, NumOps);
4651 N->OperandsNeedDelete = true;
4653 N->InitOperands(N->OperandList, Ops, NumOps);
4656 // Delete any nodes that are still dead after adding the uses for the
4658 if (!DeadNodeSet.empty()) {
4659 SmallVector<SDNode *, 16> DeadNodes;
4660 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4661 E = DeadNodeSet.end(); I != E; ++I)
4662 if ((*I)->use_empty())
4663 DeadNodes.push_back(*I);
4664 RemoveDeadNodes(DeadNodes);
4668 CSEMap.InsertNode(N, IP); // Memoize the new node.
4673 /// getMachineNode - These are used for target selectors to create a new node
4674 /// with specified return type(s), MachineInstr opcode, and operands.
4676 /// Note that getMachineNode returns the resultant node. If there is already a
4677 /// node of the specified opcode and operands, it returns that node instead of
4678 /// the current one.
4680 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) {
4681 SDVTList VTs = getVTList(VT);
4682 return getMachineNode(Opcode, dl, VTs, 0, 0);
4686 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) {
4687 SDVTList VTs = getVTList(VT);
4688 SDValue Ops[] = { Op1 };
4689 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4693 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4694 SDValue Op1, SDValue Op2) {
4695 SDVTList VTs = getVTList(VT);
4696 SDValue Ops[] = { Op1, Op2 };
4697 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4701 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4702 SDValue Op1, SDValue Op2, SDValue Op3) {
4703 SDVTList VTs = getVTList(VT);
4704 SDValue Ops[] = { Op1, Op2, Op3 };
4705 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4709 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4710 const SDValue *Ops, unsigned NumOps) {
4711 SDVTList VTs = getVTList(VT);
4712 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4716 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) {
4717 SDVTList VTs = getVTList(VT1, VT2);
4718 return getMachineNode(Opcode, dl, VTs, 0, 0);
4722 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4723 EVT VT1, EVT VT2, SDValue Op1) {
4724 SDVTList VTs = getVTList(VT1, VT2);
4725 SDValue Ops[] = { Op1 };
4726 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4730 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4731 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) {
4732 SDVTList VTs = getVTList(VT1, VT2);
4733 SDValue Ops[] = { Op1, Op2 };
4734 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4738 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4739 EVT VT1, EVT VT2, SDValue Op1,
4740 SDValue Op2, SDValue Op3) {
4741 SDVTList VTs = getVTList(VT1, VT2);
4742 SDValue Ops[] = { Op1, Op2, Op3 };
4743 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4747 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4749 const SDValue *Ops, unsigned NumOps) {
4750 SDVTList VTs = getVTList(VT1, VT2);
4751 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4755 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4756 EVT VT1, EVT VT2, EVT VT3,
4757 SDValue Op1, SDValue Op2) {
4758 SDVTList VTs = getVTList(VT1, VT2, VT3);
4759 SDValue Ops[] = { Op1, Op2 };
4760 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4764 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4765 EVT VT1, EVT VT2, EVT VT3,
4766 SDValue Op1, SDValue Op2, SDValue Op3) {
4767 SDVTList VTs = getVTList(VT1, VT2, VT3);
4768 SDValue Ops[] = { Op1, Op2, Op3 };
4769 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4773 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4774 EVT VT1, EVT VT2, EVT VT3,
4775 const SDValue *Ops, unsigned NumOps) {
4776 SDVTList VTs = getVTList(VT1, VT2, VT3);
4777 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4781 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1,
4782 EVT VT2, EVT VT3, EVT VT4,
4783 const SDValue *Ops, unsigned NumOps) {
4784 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4785 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4789 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4790 const std::vector<EVT> &ResultTys,
4791 const SDValue *Ops, unsigned NumOps) {
4792 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size());
4793 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4797 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs,
4798 const SDValue *Ops, unsigned NumOps) {
4799 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag;
4804 FoldingSetNodeID ID;
4805 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps);
4807 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4808 return cast<MachineSDNode>(E);
4811 // Allocate a new MachineSDNode.
4812 N = NodeAllocator.Allocate<MachineSDNode>();
4813 new (N) MachineSDNode(~Opcode, DL, VTs);
4815 // Initialize the operands list.
4816 if (NumOps > array_lengthof(N->LocalOperands))
4817 // We're creating a final node that will live unmorphed for the
4818 // remainder of the current SelectionDAG iteration, so we can allocate
4819 // the operands directly out of a pool with no recycling metadata.
4820 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4823 N->InitOperands(N->LocalOperands, Ops, NumOps);
4824 N->OperandsNeedDelete = false;
4827 CSEMap.InsertNode(N, IP);
4829 AllNodes.push_back(N);
4836 /// getTargetExtractSubreg - A convenience function for creating
4837 /// TargetOpcode::EXTRACT_SUBREG nodes.
4839 SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT,
4841 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4842 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
4843 VT, Operand, SRIdxVal);
4844 return SDValue(Subreg, 0);
4847 /// getTargetInsertSubreg - A convenience function for creating
4848 /// TargetOpcode::INSERT_SUBREG nodes.
4850 SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT,
4851 SDValue Operand, SDValue Subreg) {
4852 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4853 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
4854 VT, Operand, Subreg, SRIdxVal);
4855 return SDValue(Result, 0);
4858 /// getNodeIfExists - Get the specified node if it's already available, or
4859 /// else return NULL.
4860 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4861 const SDValue *Ops, unsigned NumOps) {
4862 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4863 FoldingSetNodeID ID;
4864 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4866 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4872 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4873 /// This can cause recursive merging of nodes in the DAG.
4875 /// This version assumes From has a single result value.
4877 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
4878 DAGUpdateListener *UpdateListener) {
4879 SDNode *From = FromN.getNode();
4880 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
4881 "Cannot replace with this method!");
4882 assert(From != To.getNode() && "Cannot replace uses of with self");
4884 // Iterate over all the existing uses of From. New uses will be added
4885 // to the beginning of the use list, which we avoid visiting.
4886 // This specifically avoids visiting uses of From that arise while the
4887 // replacement is happening, because any such uses would be the result
4888 // of CSE: If an existing node looks like From after one of its operands
4889 // is replaced by To, we don't want to replace of all its users with To
4890 // too. See PR3018 for more info.
4891 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4895 // This node is about to morph, remove its old self from the CSE maps.
4896 RemoveNodeFromCSEMaps(User);
4898 // A user can appear in a use list multiple times, and when this
4899 // happens the uses are usually next to each other in the list.
4900 // To help reduce the number of CSE recomputations, process all
4901 // the uses of this user that we can find this way.
4903 SDUse &Use = UI.getUse();
4906 } while (UI != UE && *UI == User);
4908 // Now that we have modified User, add it back to the CSE maps. If it
4909 // already exists there, recursively merge the results together.
4910 AddModifiedNodeToCSEMaps(User, UpdateListener);
4914 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4915 /// This can cause recursive merging of nodes in the DAG.
4917 /// This version assumes that for each value of From, there is a
4918 /// corresponding value in To in the same position with the same type.
4920 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
4921 DAGUpdateListener *UpdateListener) {
4923 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
4924 assert((!From->hasAnyUseOfValue(i) ||
4925 From->getValueType(i) == To->getValueType(i)) &&
4926 "Cannot use this version of ReplaceAllUsesWith!");
4929 // Handle the trivial case.
4933 // Iterate over just the existing users of From. See the comments in
4934 // the ReplaceAllUsesWith above.
4935 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4939 // This node is about to morph, remove its old self from the CSE maps.
4940 RemoveNodeFromCSEMaps(User);
4942 // A user can appear in a use list multiple times, and when this
4943 // happens the uses are usually next to each other in the list.
4944 // To help reduce the number of CSE recomputations, process all
4945 // the uses of this user that we can find this way.
4947 SDUse &Use = UI.getUse();
4950 } while (UI != UE && *UI == User);
4952 // Now that we have modified User, add it back to the CSE maps. If it
4953 // already exists there, recursively merge the results together.
4954 AddModifiedNodeToCSEMaps(User, UpdateListener);
4958 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4959 /// This can cause recursive merging of nodes in the DAG.
4961 /// This version can replace From with any result values. To must match the
4962 /// number and types of values returned by From.
4963 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
4965 DAGUpdateListener *UpdateListener) {
4966 if (From->getNumValues() == 1) // Handle the simple case efficiently.
4967 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
4969 // Iterate over just the existing users of From. See the comments in
4970 // the ReplaceAllUsesWith above.
4971 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4975 // This node is about to morph, remove its old self from the CSE maps.
4976 RemoveNodeFromCSEMaps(User);
4978 // A user can appear in a use list multiple times, and when this
4979 // happens the uses are usually next to each other in the list.
4980 // To help reduce the number of CSE recomputations, process all
4981 // the uses of this user that we can find this way.
4983 SDUse &Use = UI.getUse();
4984 const SDValue &ToOp = To[Use.getResNo()];
4987 } while (UI != UE && *UI == User);
4989 // Now that we have modified User, add it back to the CSE maps. If it
4990 // already exists there, recursively merge the results together.
4991 AddModifiedNodeToCSEMaps(User, UpdateListener);
4995 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
4996 /// uses of other values produced by From.getNode() alone. The Deleted
4997 /// vector is handled the same way as for ReplaceAllUsesWith.
4998 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
4999 DAGUpdateListener *UpdateListener){
5000 // Handle the really simple, really trivial case efficiently.
5001 if (From == To) return;
5003 // Handle the simple, trivial, case efficiently.
5004 if (From.getNode()->getNumValues() == 1) {
5005 ReplaceAllUsesWith(From, To, UpdateListener);
5009 // Iterate over just the existing users of From. See the comments in
5010 // the ReplaceAllUsesWith above.
5011 SDNode::use_iterator UI = From.getNode()->use_begin(),
5012 UE = From.getNode()->use_end();
5015 bool UserRemovedFromCSEMaps = false;
5017 // A user can appear in a use list multiple times, and when this
5018 // happens the uses are usually next to each other in the list.
5019 // To help reduce the number of CSE recomputations, process all
5020 // the uses of this user that we can find this way.
5022 SDUse &Use = UI.getUse();
5024 // Skip uses of different values from the same node.
5025 if (Use.getResNo() != From.getResNo()) {
5030 // If this node hasn't been modified yet, it's still in the CSE maps,
5031 // so remove its old self from the CSE maps.
5032 if (!UserRemovedFromCSEMaps) {
5033 RemoveNodeFromCSEMaps(User);
5034 UserRemovedFromCSEMaps = true;
5039 } while (UI != UE && *UI == User);
5041 // We are iterating over all uses of the From node, so if a use
5042 // doesn't use the specific value, no changes are made.
5043 if (!UserRemovedFromCSEMaps)
5046 // Now that we have modified User, add it back to the CSE maps. If it
5047 // already exists there, recursively merge the results together.
5048 AddModifiedNodeToCSEMaps(User, UpdateListener);
5053 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
5054 /// to record information about a use.
5061 /// operator< - Sort Memos by User.
5062 bool operator<(const UseMemo &L, const UseMemo &R) {
5063 return (intptr_t)L.User < (intptr_t)R.User;
5067 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
5068 /// uses of other values produced by From.getNode() alone. The same value
5069 /// may appear in both the From and To list. The Deleted vector is
5070 /// handled the same way as for ReplaceAllUsesWith.
5071 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
5074 DAGUpdateListener *UpdateListener){
5075 // Handle the simple, trivial case efficiently.
5077 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
5079 // Read up all the uses and make records of them. This helps
5080 // processing new uses that are introduced during the
5081 // replacement process.
5082 SmallVector<UseMemo, 4> Uses;
5083 for (unsigned i = 0; i != Num; ++i) {
5084 unsigned FromResNo = From[i].getResNo();
5085 SDNode *FromNode = From[i].getNode();
5086 for (SDNode::use_iterator UI = FromNode->use_begin(),
5087 E = FromNode->use_end(); UI != E; ++UI) {
5088 SDUse &Use = UI.getUse();
5089 if (Use.getResNo() == FromResNo) {
5090 UseMemo Memo = { *UI, i, &Use };
5091 Uses.push_back(Memo);
5096 // Sort the uses, so that all the uses from a given User are together.
5097 std::sort(Uses.begin(), Uses.end());
5099 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
5100 UseIndex != UseIndexEnd; ) {
5101 // We know that this user uses some value of From. If it is the right
5102 // value, update it.
5103 SDNode *User = Uses[UseIndex].User;
5105 // This node is about to morph, remove its old self from the CSE maps.
5106 RemoveNodeFromCSEMaps(User);
5108 // The Uses array is sorted, so all the uses for a given User
5109 // are next to each other in the list.
5110 // To help reduce the number of CSE recomputations, process all
5111 // the uses of this user that we can find this way.
5113 unsigned i = Uses[UseIndex].Index;
5114 SDUse &Use = *Uses[UseIndex].Use;
5118 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == 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, UpdateListener);
5126 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
5127 /// based on their topological order. It returns the maximum id and a vector
5128 /// of the SDNodes* in assigned order by reference.
5129 unsigned SelectionDAG::AssignTopologicalOrder() {
5131 unsigned DAGSize = 0;
5133 // SortedPos tracks the progress of the algorithm. Nodes before it are
5134 // sorted, nodes after it are unsorted. When the algorithm completes
5135 // it is at the end of the list.
5136 allnodes_iterator SortedPos = allnodes_begin();
5138 // Visit all the nodes. Move nodes with no operands to the front of
5139 // the list immediately. Annotate nodes that do have operands with their
5140 // operand count. Before we do this, the Node Id fields of the nodes
5141 // may contain arbitrary values. After, the Node Id fields for nodes
5142 // before SortedPos will contain the topological sort index, and the
5143 // Node Id fields for nodes At SortedPos and after will contain the
5144 // count of outstanding operands.
5145 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
5148 unsigned Degree = N->getNumOperands();
5150 // A node with no uses, add it to the result array immediately.
5151 N->setNodeId(DAGSize++);
5152 allnodes_iterator Q = N;
5154 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
5155 assert(SortedPos != AllNodes.end() && "Overran node list");
5158 // Temporarily use the Node Id as scratch space for the degree count.
5159 N->setNodeId(Degree);
5163 // Visit all the nodes. As we iterate, moves nodes into sorted order,
5164 // such that by the time the end is reached all nodes will be sorted.
5165 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
5168 // N is in sorted position, so all its uses have one less operand
5169 // that needs to be sorted.
5170 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5173 unsigned Degree = P->getNodeId();
5174 assert(Degree != 0 && "Invalid node degree");
5177 // All of P's operands are sorted, so P may sorted now.
5178 P->setNodeId(DAGSize++);
5180 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
5181 assert(SortedPos != AllNodes.end() && "Overran node list");
5184 // Update P's outstanding operand count.
5185 P->setNodeId(Degree);
5188 if (I == SortedPos) {
5191 dbgs() << "Overran sorted position:\n";
5194 llvm_unreachable(0);
5198 assert(SortedPos == AllNodes.end() &&
5199 "Topological sort incomplete!");
5200 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
5201 "First node in topological sort is not the entry token!");
5202 assert(AllNodes.front().getNodeId() == 0 &&
5203 "First node in topological sort has non-zero id!");
5204 assert(AllNodes.front().getNumOperands() == 0 &&
5205 "First node in topological sort has operands!");
5206 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
5207 "Last node in topologic sort has unexpected id!");
5208 assert(AllNodes.back().use_empty() &&
5209 "Last node in topologic sort has users!");
5210 assert(DAGSize == allnodes_size() && "Node count mismatch!");
5214 /// AssignOrdering - Assign an order to the SDNode.
5215 void SelectionDAG::AssignOrdering(const SDNode *SD, unsigned Order) {
5216 assert(SD && "Trying to assign an order to a null node!");
5217 Ordering->add(SD, Order);
5220 /// GetOrdering - Get the order for the SDNode.
5221 unsigned SelectionDAG::GetOrdering(const SDNode *SD) const {
5222 assert(SD && "Trying to get the order of a null node!");
5223 return Ordering->getOrder(SD);
5227 //===----------------------------------------------------------------------===//
5229 //===----------------------------------------------------------------------===//
5231 HandleSDNode::~HandleSDNode() {
5235 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA,
5236 EVT VT, int64_t o, unsigned char TF)
5237 : SDNode(Opc, DebugLoc::getUnknownLoc(), getSDVTList(VT)),
5238 Offset(o), TargetFlags(TF) {
5239 TheGlobal = const_cast<GlobalValue*>(GA);
5242 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt,
5243 MachineMemOperand *mmo)
5244 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) {
5245 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(),
5246 MMO->isNonTemporal());
5247 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5248 assert(isNonTemporal() == MMO->isNonTemporal() &&
5249 "Non-temporal encoding error!");
5250 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5253 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs,
5254 const SDValue *Ops, unsigned NumOps, EVT memvt,
5255 MachineMemOperand *mmo)
5256 : SDNode(Opc, dl, VTs, Ops, NumOps),
5257 MemoryVT(memvt), MMO(mmo) {
5258 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(),
5259 MMO->isNonTemporal());
5260 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5261 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5264 /// Profile - Gather unique data for the node.
5266 void SDNode::Profile(FoldingSetNodeID &ID) const {
5267 AddNodeIDNode(ID, this);
5272 std::vector<EVT> VTs;
5275 VTs.reserve(MVT::LAST_VALUETYPE);
5276 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i)
5277 VTs.push_back(MVT((MVT::SimpleValueType)i));
5282 static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs;
5283 static ManagedStatic<EVTArray> SimpleVTArray;
5284 static ManagedStatic<sys::SmartMutex<true> > VTMutex;
5286 /// getValueTypeList - Return a pointer to the specified value type.
5288 const EVT *SDNode::getValueTypeList(EVT VT) {
5289 if (VT.isExtended()) {
5290 sys::SmartScopedLock<true> Lock(*VTMutex);
5291 return &(*EVTs->insert(VT).first);
5293 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy];
5297 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
5298 /// indicated value. This method ignores uses of other values defined by this
5300 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
5301 assert(Value < getNumValues() && "Bad value!");
5303 // TODO: Only iterate over uses of a given value of the node
5304 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
5305 if (UI.getUse().getResNo() == Value) {
5312 // Found exactly the right number of uses?
5317 /// hasAnyUseOfValue - Return true if there are any use of the indicated
5318 /// value. This method ignores uses of other values defined by this operation.
5319 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
5320 assert(Value < getNumValues() && "Bad value!");
5322 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
5323 if (UI.getUse().getResNo() == Value)
5330 /// isOnlyUserOf - Return true if this node is the only use of N.
5332 bool SDNode::isOnlyUserOf(SDNode *N) const {
5334 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
5345 /// isOperand - Return true if this node is an operand of N.
5347 bool SDValue::isOperandOf(SDNode *N) const {
5348 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5349 if (*this == N->getOperand(i))
5354 bool SDNode::isOperandOf(SDNode *N) const {
5355 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
5356 if (this == N->OperandList[i].getNode())
5361 /// reachesChainWithoutSideEffects - Return true if this operand (which must
5362 /// be a chain) reaches the specified operand without crossing any
5363 /// side-effecting instructions. In practice, this looks through token
5364 /// factors and non-volatile loads. In order to remain efficient, this only
5365 /// looks a couple of nodes in, it does not do an exhaustive search.
5366 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
5367 unsigned Depth) const {
5368 if (*this == Dest) return true;
5370 // Don't search too deeply, we just want to be able to see through
5371 // TokenFactor's etc.
5372 if (Depth == 0) return false;
5374 // If this is a token factor, all inputs to the TF happen in parallel. If any
5375 // of the operands of the TF reach dest, then we can do the xform.
5376 if (getOpcode() == ISD::TokenFactor) {
5377 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5378 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
5383 // Loads don't have side effects, look through them.
5384 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
5385 if (!Ld->isVolatile())
5386 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
5391 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
5392 /// is either an operand of N or it can be reached by traversing up the operands.
5393 /// NOTE: this is an expensive method. Use it carefully.
5394 bool SDNode::isPredecessorOf(SDNode *N) const {
5395 SmallPtrSet<SDNode *, 32> Visited;
5396 SmallVector<SDNode *, 16> Worklist;
5397 Worklist.push_back(N);
5400 N = Worklist.pop_back_val();
5401 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5402 SDNode *Op = N->getOperand(i).getNode();
5405 if (Visited.insert(Op))
5406 Worklist.push_back(Op);
5408 } while (!Worklist.empty());
5413 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
5414 assert(Num < NumOperands && "Invalid child # of SDNode!");
5415 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
5418 std::string SDNode::getOperationName(const SelectionDAG *G) const {
5419 switch (getOpcode()) {
5421 if (getOpcode() < ISD::BUILTIN_OP_END)
5422 return "<<Unknown DAG Node>>";
5423 if (isMachineOpcode()) {
5425 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
5426 if (getMachineOpcode() < TII->getNumOpcodes())
5427 return TII->get(getMachineOpcode()).getName();
5428 return "<<Unknown Machine Node #" + utostr(getOpcode()) + ">>";
5431 const TargetLowering &TLI = G->getTargetLoweringInfo();
5432 const char *Name = TLI.getTargetNodeName(getOpcode());
5433 if (Name) return Name;
5434 return "<<Unknown Target Node #" + utostr(getOpcode()) + ">>";
5436 return "<<Unknown Node #" + utostr(getOpcode()) + ">>";
5439 case ISD::DELETED_NODE:
5440 return "<<Deleted Node!>>";
5442 case ISD::PREFETCH: return "Prefetch";
5443 case ISD::MEMBARRIER: return "MemBarrier";
5444 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap";
5445 case ISD::ATOMIC_SWAP: return "AtomicSwap";
5446 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd";
5447 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub";
5448 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd";
5449 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr";
5450 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor";
5451 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand";
5452 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin";
5453 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax";
5454 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin";
5455 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax";
5456 case ISD::PCMARKER: return "PCMarker";
5457 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5458 case ISD::SRCVALUE: return "SrcValue";
5459 case ISD::EntryToken: return "EntryToken";
5460 case ISD::TokenFactor: return "TokenFactor";
5461 case ISD::AssertSext: return "AssertSext";
5462 case ISD::AssertZext: return "AssertZext";
5464 case ISD::BasicBlock: return "BasicBlock";
5465 case ISD::VALUETYPE: return "ValueType";
5466 case ISD::Register: return "Register";
5468 case ISD::Constant: return "Constant";
5469 case ISD::ConstantFP: return "ConstantFP";
5470 case ISD::GlobalAddress: return "GlobalAddress";
5471 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5472 case ISD::FrameIndex: return "FrameIndex";
5473 case ISD::JumpTable: return "JumpTable";
5474 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5475 case ISD::RETURNADDR: return "RETURNADDR";
5476 case ISD::FRAMEADDR: return "FRAMEADDR";
5477 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5478 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5479 case ISD::LSDAADDR: return "LSDAADDR";
5480 case ISD::EHSELECTION: return "EHSELECTION";
5481 case ISD::EH_RETURN: return "EH_RETURN";
5482 case ISD::ConstantPool: return "ConstantPool";
5483 case ISD::ExternalSymbol: return "ExternalSymbol";
5484 case ISD::BlockAddress: return "BlockAddress";
5485 case ISD::INTRINSIC_WO_CHAIN:
5486 case ISD::INTRINSIC_VOID:
5487 case ISD::INTRINSIC_W_CHAIN: {
5488 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1;
5489 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue();
5490 if (IID < Intrinsic::num_intrinsics)
5491 return Intrinsic::getName((Intrinsic::ID)IID);
5492 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo())
5493 return TII->getName(IID);
5494 llvm_unreachable("Invalid intrinsic ID");
5497 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5498 case ISD::TargetConstant: return "TargetConstant";
5499 case ISD::TargetConstantFP:return "TargetConstantFP";
5500 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5501 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5502 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5503 case ISD::TargetJumpTable: return "TargetJumpTable";
5504 case ISD::TargetConstantPool: return "TargetConstantPool";
5505 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5506 case ISD::TargetBlockAddress: return "TargetBlockAddress";
5508 case ISD::CopyToReg: return "CopyToReg";
5509 case ISD::CopyFromReg: return "CopyFromReg";
5510 case ISD::UNDEF: return "undef";
5511 case ISD::MERGE_VALUES: return "merge_values";
5512 case ISD::INLINEASM: return "inlineasm";
5513 case ISD::EH_LABEL: return "eh_label";
5514 case ISD::HANDLENODE: return "handlenode";
5517 case ISD::FABS: return "fabs";
5518 case ISD::FNEG: return "fneg";
5519 case ISD::FSQRT: return "fsqrt";
5520 case ISD::FSIN: return "fsin";
5521 case ISD::FCOS: return "fcos";
5522 case ISD::FPOWI: return "fpowi";
5523 case ISD::FPOW: return "fpow";
5524 case ISD::FTRUNC: return "ftrunc";
5525 case ISD::FFLOOR: return "ffloor";
5526 case ISD::FCEIL: return "fceil";
5527 case ISD::FRINT: return "frint";
5528 case ISD::FNEARBYINT: return "fnearbyint";
5531 case ISD::ADD: return "add";
5532 case ISD::SUB: return "sub";
5533 case ISD::MUL: return "mul";
5534 case ISD::MULHU: return "mulhu";
5535 case ISD::MULHS: return "mulhs";
5536 case ISD::SDIV: return "sdiv";
5537 case ISD::UDIV: return "udiv";
5538 case ISD::SREM: return "srem";
5539 case ISD::UREM: return "urem";
5540 case ISD::SMUL_LOHI: return "smul_lohi";
5541 case ISD::UMUL_LOHI: return "umul_lohi";
5542 case ISD::SDIVREM: return "sdivrem";
5543 case ISD::UDIVREM: return "udivrem";
5544 case ISD::AND: return "and";
5545 case ISD::OR: return "or";
5546 case ISD::XOR: return "xor";
5547 case ISD::SHL: return "shl";
5548 case ISD::SRA: return "sra";
5549 case ISD::SRL: return "srl";
5550 case ISD::ROTL: return "rotl";
5551 case ISD::ROTR: return "rotr";
5552 case ISD::FADD: return "fadd";
5553 case ISD::FSUB: return "fsub";
5554 case ISD::FMUL: return "fmul";
5555 case ISD::FDIV: return "fdiv";
5556 case ISD::FREM: return "frem";
5557 case ISD::FCOPYSIGN: return "fcopysign";
5558 case ISD::FGETSIGN: return "fgetsign";
5560 case ISD::SETCC: return "setcc";
5561 case ISD::VSETCC: return "vsetcc";
5562 case ISD::SELECT: return "select";
5563 case ISD::SELECT_CC: return "select_cc";
5564 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5565 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5566 case ISD::CONCAT_VECTORS: return "concat_vectors";
5567 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5568 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5569 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5570 case ISD::CARRY_FALSE: return "carry_false";
5571 case ISD::ADDC: return "addc";
5572 case ISD::ADDE: return "adde";
5573 case ISD::SADDO: return "saddo";
5574 case ISD::UADDO: return "uaddo";
5575 case ISD::SSUBO: return "ssubo";
5576 case ISD::USUBO: return "usubo";
5577 case ISD::SMULO: return "smulo";
5578 case ISD::UMULO: return "umulo";
5579 case ISD::SUBC: return "subc";
5580 case ISD::SUBE: return "sube";
5581 case ISD::SHL_PARTS: return "shl_parts";
5582 case ISD::SRA_PARTS: return "sra_parts";
5583 case ISD::SRL_PARTS: return "srl_parts";
5585 // Conversion operators.
5586 case ISD::SIGN_EXTEND: return "sign_extend";
5587 case ISD::ZERO_EXTEND: return "zero_extend";
5588 case ISD::ANY_EXTEND: return "any_extend";
5589 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5590 case ISD::TRUNCATE: return "truncate";
5591 case ISD::FP_ROUND: return "fp_round";
5592 case ISD::FLT_ROUNDS_: return "flt_rounds";
5593 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5594 case ISD::FP_EXTEND: return "fp_extend";
5596 case ISD::SINT_TO_FP: return "sint_to_fp";
5597 case ISD::UINT_TO_FP: return "uint_to_fp";
5598 case ISD::FP_TO_SINT: return "fp_to_sint";
5599 case ISD::FP_TO_UINT: return "fp_to_uint";
5600 case ISD::BIT_CONVERT: return "bit_convert";
5602 case ISD::CONVERT_RNDSAT: {
5603 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) {
5604 default: llvm_unreachable("Unknown cvt code!");
5605 case ISD::CVT_FF: return "cvt_ff";
5606 case ISD::CVT_FS: return "cvt_fs";
5607 case ISD::CVT_FU: return "cvt_fu";
5608 case ISD::CVT_SF: return "cvt_sf";
5609 case ISD::CVT_UF: return "cvt_uf";
5610 case ISD::CVT_SS: return "cvt_ss";
5611 case ISD::CVT_SU: return "cvt_su";
5612 case ISD::CVT_US: return "cvt_us";
5613 case ISD::CVT_UU: return "cvt_uu";
5617 // Control flow instructions
5618 case ISD::BR: return "br";
5619 case ISD::BRIND: return "brind";
5620 case ISD::BR_JT: return "br_jt";
5621 case ISD::BRCOND: return "brcond";
5622 case ISD::BR_CC: return "br_cc";
5623 case ISD::CALLSEQ_START: return "callseq_start";
5624 case ISD::CALLSEQ_END: return "callseq_end";
5627 case ISD::LOAD: return "load";
5628 case ISD::STORE: return "store";
5629 case ISD::VAARG: return "vaarg";
5630 case ISD::VACOPY: return "vacopy";
5631 case ISD::VAEND: return "vaend";
5632 case ISD::VASTART: return "vastart";
5633 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5634 case ISD::EXTRACT_ELEMENT: return "extract_element";
5635 case ISD::BUILD_PAIR: return "build_pair";
5636 case ISD::STACKSAVE: return "stacksave";
5637 case ISD::STACKRESTORE: return "stackrestore";
5638 case ISD::TRAP: return "trap";
5641 case ISD::BSWAP: return "bswap";
5642 case ISD::CTPOP: return "ctpop";
5643 case ISD::CTTZ: return "cttz";
5644 case ISD::CTLZ: return "ctlz";
5647 case ISD::TRAMPOLINE: return "trampoline";
5650 switch (cast<CondCodeSDNode>(this)->get()) {
5651 default: llvm_unreachable("Unknown setcc condition!");
5652 case ISD::SETOEQ: return "setoeq";
5653 case ISD::SETOGT: return "setogt";
5654 case ISD::SETOGE: return "setoge";
5655 case ISD::SETOLT: return "setolt";
5656 case ISD::SETOLE: return "setole";
5657 case ISD::SETONE: return "setone";
5659 case ISD::SETO: return "seto";
5660 case ISD::SETUO: return "setuo";
5661 case ISD::SETUEQ: return "setue";
5662 case ISD::SETUGT: return "setugt";
5663 case ISD::SETUGE: return "setuge";
5664 case ISD::SETULT: return "setult";
5665 case ISD::SETULE: return "setule";
5666 case ISD::SETUNE: return "setune";
5668 case ISD::SETEQ: return "seteq";
5669 case ISD::SETGT: return "setgt";
5670 case ISD::SETGE: return "setge";
5671 case ISD::SETLT: return "setlt";
5672 case ISD::SETLE: return "setle";
5673 case ISD::SETNE: return "setne";
5678 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5687 return "<post-inc>";
5689 return "<post-dec>";
5693 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5694 std::string S = "< ";
5708 if (getByValAlign())
5709 S += "byval-align:" + utostr(getByValAlign()) + " ";
5711 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5713 S += "byval-size:" + utostr(getByValSize()) + " ";
5717 void SDNode::dump() const { dump(0); }
5718 void SDNode::dump(const SelectionDAG *G) const {
5722 void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const {
5723 OS << (void*)this << ": ";
5725 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5727 if (getValueType(i) == MVT::Other)
5730 OS << getValueType(i).getEVTString();
5732 OS << " = " << getOperationName(G);
5735 void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const {
5736 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) {
5737 if (!MN->memoperands_empty()) {
5740 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(),
5741 e = MN->memoperands_end(); i != e; ++i) {
5748 } else if (const ShuffleVectorSDNode *SVN =
5749 dyn_cast<ShuffleVectorSDNode>(this)) {
5751 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) {
5752 int Idx = SVN->getMaskElt(i);
5760 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5761 OS << '<' << CSDN->getAPIntValue() << '>';
5762 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5763 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5764 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5765 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5766 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5769 CSDN->getValueAPF().bitcastToAPInt().dump();
5772 } else if (const GlobalAddressSDNode *GADN =
5773 dyn_cast<GlobalAddressSDNode>(this)) {
5774 int64_t offset = GADN->getOffset();
5776 WriteAsOperand(OS, GADN->getGlobal());
5779 OS << " + " << offset;
5781 OS << " " << offset;
5782 if (unsigned int TF = GADN->getTargetFlags())
5783 OS << " [TF=" << TF << ']';
5784 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5785 OS << "<" << FIDN->getIndex() << ">";
5786 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5787 OS << "<" << JTDN->getIndex() << ">";
5788 if (unsigned int TF = JTDN->getTargetFlags())
5789 OS << " [TF=" << TF << ']';
5790 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5791 int offset = CP->getOffset();
5792 if (CP->isMachineConstantPoolEntry())
5793 OS << "<" << *CP->getMachineCPVal() << ">";
5795 OS << "<" << *CP->getConstVal() << ">";
5797 OS << " + " << offset;
5799 OS << " " << offset;
5800 if (unsigned int TF = CP->getTargetFlags())
5801 OS << " [TF=" << TF << ']';
5802 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5804 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5806 OS << LBB->getName() << " ";
5807 OS << (const void*)BBDN->getBasicBlock() << ">";
5808 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5809 if (G && R->getReg() &&
5810 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5811 OS << " %" << G->getTarget().getRegisterInfo()->getName(R->getReg());
5813 OS << " %reg" << R->getReg();
5815 } else if (const ExternalSymbolSDNode *ES =
5816 dyn_cast<ExternalSymbolSDNode>(this)) {
5817 OS << "'" << ES->getSymbol() << "'";
5818 if (unsigned int TF = ES->getTargetFlags())
5819 OS << " [TF=" << TF << ']';
5820 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5822 OS << "<" << M->getValue() << ">";
5825 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5826 OS << ":" << N->getVT().getEVTString();
5828 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5829 OS << "<" << *LD->getMemOperand();
5832 switch (LD->getExtensionType()) {
5833 default: doExt = false; break;
5834 case ISD::EXTLOAD: OS << ", anyext"; break;
5835 case ISD::SEXTLOAD: OS << ", sext"; break;
5836 case ISD::ZEXTLOAD: OS << ", zext"; break;
5839 OS << " from " << LD->getMemoryVT().getEVTString();
5841 const char *AM = getIndexedModeName(LD->getAddressingMode());
5846 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
5847 OS << "<" << *ST->getMemOperand();
5849 if (ST->isTruncatingStore())
5850 OS << ", trunc to " << ST->getMemoryVT().getEVTString();
5852 const char *AM = getIndexedModeName(ST->getAddressingMode());
5857 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) {
5858 OS << "<" << *M->getMemOperand() << ">";
5859 } else if (const BlockAddressSDNode *BA =
5860 dyn_cast<BlockAddressSDNode>(this)) {
5862 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false);
5864 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false);
5866 if (unsigned int TF = BA->getTargetFlags())
5867 OS << " [TF=" << TF << ']';
5871 if (unsigned Order = G->GetOrdering(this))
5872 OS << " [ORD=" << Order << ']';
5874 if (getNodeId() != -1)
5875 OS << " [ID=" << getNodeId() << ']';
5878 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
5880 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
5881 if (i) OS << ", "; else OS << " ";
5882 OS << (void*)getOperand(i).getNode();
5883 if (unsigned RN = getOperand(i).getResNo())
5886 print_details(OS, G);
5889 static void printrWithDepthHelper(raw_ostream &OS, const SDNode *N,
5890 const SelectionDAG *G, unsigned depth,
5903 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5905 printrWithDepthHelper(OS, N->getOperand(i).getNode(), G, depth-1, indent+2);
5909 void SDNode::printrWithDepth(raw_ostream &OS, const SelectionDAG *G,
5910 unsigned depth) const {
5911 printrWithDepthHelper(OS, this, G, depth, 0);
5914 void SDNode::printrFull(raw_ostream &OS, const SelectionDAG *G) const {
5915 // Don't print impossibly deep things.
5916 printrWithDepth(OS, G, 100);
5919 void SDNode::dumprWithDepth(const SelectionDAG *G, unsigned depth) const {
5920 printrWithDepth(dbgs(), G, depth);
5923 void SDNode::dumprFull(const SelectionDAG *G) const {
5924 // Don't print impossibly deep things.
5925 dumprWithDepth(G, 100);
5928 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
5929 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5930 if (N->getOperand(i).getNode()->hasOneUse())
5931 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
5933 dbgs() << "\n" << std::string(indent+2, ' ')
5934 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
5938 dbgs().indent(indent);
5942 SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) {
5943 assert(N->getNumValues() == 1 &&
5944 "Can't unroll a vector with multiple results!");
5946 EVT VT = N->getValueType(0);
5947 unsigned NE = VT.getVectorNumElements();
5948 EVT EltVT = VT.getVectorElementType();
5949 DebugLoc dl = N->getDebugLoc();
5951 SmallVector<SDValue, 8> Scalars;
5952 SmallVector<SDValue, 4> Operands(N->getNumOperands());
5954 // If ResNE is 0, fully unroll the vector op.
5957 else if (NE > ResNE)
5961 for (i= 0; i != NE; ++i) {
5962 for (unsigned j = 0; j != N->getNumOperands(); ++j) {
5963 SDValue Operand = N->getOperand(j);
5964 EVT OperandVT = Operand.getValueType();
5965 if (OperandVT.isVector()) {
5966 // A vector operand; extract a single element.
5967 EVT OperandEltVT = OperandVT.getVectorElementType();
5968 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl,
5971 getConstant(i, MVT::i32));
5973 // A scalar operand; just use it as is.
5974 Operands[j] = Operand;
5978 switch (N->getOpcode()) {
5980 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
5981 &Operands[0], Operands.size()));
5988 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0],
5989 getShiftAmountOperand(Operands[1])));
5991 case ISD::SIGN_EXTEND_INREG:
5992 case ISD::FP_ROUND_INREG: {
5993 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType();
5994 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
5996 getValueType(ExtVT)));
6001 for (; i < ResNE; ++i)
6002 Scalars.push_back(getUNDEF(EltVT));
6004 return getNode(ISD::BUILD_VECTOR, dl,
6005 EVT::getVectorVT(*getContext(), EltVT, ResNE),
6006 &Scalars[0], Scalars.size());
6010 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
6011 /// location that is 'Dist' units away from the location that the 'Base' load
6012 /// is loading from.
6013 bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
6014 unsigned Bytes, int Dist) const {
6015 if (LD->getChain() != Base->getChain())
6017 EVT VT = LD->getValueType(0);
6018 if (VT.getSizeInBits() / 8 != Bytes)
6021 SDValue Loc = LD->getOperand(1);
6022 SDValue BaseLoc = Base->getOperand(1);
6023 if (Loc.getOpcode() == ISD::FrameIndex) {
6024 if (BaseLoc.getOpcode() != ISD::FrameIndex)
6026 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo();
6027 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
6028 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
6029 int FS = MFI->getObjectSize(FI);
6030 int BFS = MFI->getObjectSize(BFI);
6031 if (FS != BFS || FS != (int)Bytes) return false;
6032 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
6034 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
6035 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
6036 if (V && (V->getSExtValue() == Dist*Bytes))
6040 GlobalValue *GV1 = NULL;
6041 GlobalValue *GV2 = NULL;
6042 int64_t Offset1 = 0;
6043 int64_t Offset2 = 0;
6044 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
6045 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
6046 if (isGA1 && isGA2 && GV1 == GV2)
6047 return Offset1 == (Offset2 + Dist*Bytes);
6052 /// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if
6053 /// it cannot be inferred.
6054 unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const {
6055 // If this is a GlobalAddress + cst, return the alignment.
6057 int64_t GVOffset = 0;
6058 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset))
6059 return MinAlign(GV->getAlignment(), GVOffset);
6061 // If this is a direct reference to a stack slot, use information about the
6062 // stack slot's alignment.
6063 int FrameIdx = 1 << 31;
6064 int64_t FrameOffset = 0;
6065 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
6066 FrameIdx = FI->getIndex();
6067 } else if (Ptr.getOpcode() == ISD::ADD &&
6068 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
6069 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
6070 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
6071 FrameOffset = Ptr.getConstantOperandVal(1);
6074 if (FrameIdx != (1 << 31)) {
6075 // FIXME: Handle FI+CST.
6076 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo();
6077 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
6079 if (MFI.isFixedObjectIndex(FrameIdx)) {
6080 int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx) + FrameOffset;
6082 // The alignment of the frame index can be determined from its offset from
6083 // the incoming frame position. If the frame object is at offset 32 and
6084 // the stack is guaranteed to be 16-byte aligned, then we know that the
6085 // object is 16-byte aligned.
6086 unsigned StackAlign = getTarget().getFrameInfo()->getStackAlignment();
6087 unsigned Align = MinAlign(ObjectOffset, StackAlign);
6089 // Finally, the frame object itself may have a known alignment. Factor
6090 // the alignment + offset into a new alignment. For example, if we know
6091 // the FI is 8 byte aligned, but the pointer is 4 off, we really have a
6092 // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte
6093 // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc.
6094 return std::max(Align, FIInfoAlign);
6102 void SelectionDAG::dump() const {
6103 dbgs() << "SelectionDAG has " << AllNodes.size() << " nodes:";
6105 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
6107 const SDNode *N = I;
6108 if (!N->hasOneUse() && N != getRoot().getNode())
6109 DumpNodes(N, 2, this);
6112 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
6117 void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
6119 print_details(OS, G);
6122 typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet;
6123 static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent,
6124 const SelectionDAG *G, VisitedSDNodeSet &once) {
6125 if (!once.insert(N)) // If we've been here before, return now.
6128 // Dump the current SDNode, but don't end the line yet.
6129 OS << std::string(indent, ' ');
6132 // Having printed this SDNode, walk the children:
6133 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6134 const SDNode *child = N->getOperand(i).getNode();
6139 if (child->getNumOperands() == 0) {
6140 // This child has no grandchildren; print it inline right here.
6141 child->printr(OS, G);
6143 } else { // Just the address. FIXME: also print the child's opcode.
6145 if (unsigned RN = N->getOperand(i).getResNo())
6152 // Dump children that have grandchildren on their own line(s).
6153 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6154 const SDNode *child = N->getOperand(i).getNode();
6155 DumpNodesr(OS, child, indent+2, G, once);
6159 void SDNode::dumpr() const {
6160 VisitedSDNodeSet once;
6161 DumpNodesr(dbgs(), this, 0, 0, once);
6164 void SDNode::dumpr(const SelectionDAG *G) const {
6165 VisitedSDNodeSet once;
6166 DumpNodesr(dbgs(), this, 0, G, once);
6170 // getAddressSpace - Return the address space this GlobalAddress belongs to.
6171 unsigned GlobalAddressSDNode::getAddressSpace() const {
6172 return getGlobal()->getType()->getAddressSpace();
6176 const Type *ConstantPoolSDNode::getType() const {
6177 if (isMachineConstantPoolEntry())
6178 return Val.MachineCPVal->getType();
6179 return Val.ConstVal->getType();
6182 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue,
6184 unsigned &SplatBitSize,
6186 unsigned MinSplatBits,
6188 EVT VT = getValueType(0);
6189 assert(VT.isVector() && "Expected a vector type");
6190 unsigned sz = VT.getSizeInBits();
6191 if (MinSplatBits > sz)
6194 SplatValue = APInt(sz, 0);
6195 SplatUndef = APInt(sz, 0);
6197 // Get the bits. Bits with undefined values (when the corresponding element
6198 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
6199 // in SplatValue. If any of the values are not constant, give up and return
6201 unsigned int nOps = getNumOperands();
6202 assert(nOps > 0 && "isConstantSplat has 0-size build vector");
6203 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits();
6205 for (unsigned j = 0; j < nOps; ++j) {
6206 unsigned i = isBigEndian ? nOps-1-j : j;
6207 SDValue OpVal = getOperand(i);
6208 unsigned BitPos = j * EltBitSize;
6210 if (OpVal.getOpcode() == ISD::UNDEF)
6211 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize);
6212 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal))
6213 SplatValue |= (APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize).
6214 zextOrTrunc(sz) << BitPos);
6215 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal))
6216 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos;
6221 // The build_vector is all constants or undefs. Find the smallest element
6222 // size that splats the vector.
6224 HasAnyUndefs = (SplatUndef != 0);
6227 unsigned HalfSize = sz / 2;
6228 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize);
6229 APInt LowValue = APInt(SplatValue).trunc(HalfSize);
6230 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize);
6231 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize);
6233 // If the two halves do not match (ignoring undef bits), stop here.
6234 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
6235 MinSplatBits > HalfSize)
6238 SplatValue = HighValue | LowValue;
6239 SplatUndef = HighUndef & LowUndef;
6248 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) {
6249 // Find the first non-undef value in the shuffle mask.
6251 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
6254 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!");
6256 // Make sure all remaining elements are either undef or the same as the first
6258 for (int Idx = Mask[i]; i != e; ++i)
6259 if (Mask[i] >= 0 && Mask[i] != Idx)
6265 static void checkForCyclesHelper(const SDNode *N,
6266 SmallPtrSet<const SDNode*, 32> &Visited,
6267 SmallPtrSet<const SDNode*, 32> &Checked) {
6268 // If this node has already been checked, don't check it again.
6269 if (Checked.count(N))
6272 // If a node has already been visited on this depth-first walk, reject it as
6274 if (!Visited.insert(N)) {
6275 dbgs() << "Offending node:\n";
6277 errs() << "Detected cycle in SelectionDAG\n";
6281 for(unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
6282 checkForCyclesHelper(N->getOperand(i).getNode(), Visited, Checked);
6289 void llvm::checkForCycles(const llvm::SDNode *N) {
6291 assert(N && "Checking nonexistant SDNode");
6292 SmallPtrSet<const SDNode*, 32> visited;
6293 SmallPtrSet<const SDNode*, 32> checked;
6294 checkForCyclesHelper(N, visited, checked);
6298 void llvm::checkForCycles(const llvm::SelectionDAG *DAG) {
6299 checkForCycles(DAG->getRoot().getNode());