1 //===-- SelectionDAG.cpp - Implement the SelectionDAG data structures -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAG class.
12 //===----------------------------------------------------------------------===//
13 #include "llvm/CodeGen/SelectionDAG.h"
14 #include "llvm/Constants.h"
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/Function.h"
17 #include "llvm/GlobalAlias.h"
18 #include "llvm/GlobalVariable.h"
19 #include "llvm/Intrinsics.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Assembly/Writer.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/PseudoSourceValue.h"
28 #include "llvm/Target/TargetRegisterInfo.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/Target/TargetInstrInfo.h"
33 #include "llvm/Target/TargetIntrinsicInfo.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/ManagedStatic.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/System/Mutex.h"
41 #include "llvm/ADT/SetVector.h"
42 #include "llvm/ADT/SmallPtrSet.h"
43 #include "llvm/ADT/SmallSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/StringExtras.h"
50 /// makeVTList - Return an instance of the SDVTList struct initialized with the
51 /// specified members.
52 static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
53 SDVTList Res = {VTs, NumVTs};
57 static const fltSemantics *EVTToAPFloatSemantics(EVT VT) {
58 switch (VT.getSimpleVT().SimpleTy) {
59 default: llvm_unreachable("Unknown FP format");
60 case MVT::f32: return &APFloat::IEEEsingle;
61 case MVT::f64: return &APFloat::IEEEdouble;
62 case MVT::f80: return &APFloat::x87DoubleExtended;
63 case MVT::f128: return &APFloat::IEEEquad;
64 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
68 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
70 //===----------------------------------------------------------------------===//
71 // ConstantFPSDNode Class
72 //===----------------------------------------------------------------------===//
74 /// isExactlyValue - We don't rely on operator== working on double values, as
75 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
76 /// As such, this method can be used to do an exact bit-for-bit comparison of
77 /// two floating point values.
78 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
79 return getValueAPF().bitwiseIsEqual(V);
82 bool ConstantFPSDNode::isValueValidForType(EVT VT,
84 assert(VT.isFloatingPoint() && "Can only convert between FP types");
86 // PPC long double cannot be converted to any other type.
87 if (VT == MVT::ppcf128 ||
88 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
91 // convert modifies in place, so make a copy.
92 APFloat Val2 = APFloat(Val);
94 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
99 //===----------------------------------------------------------------------===//
101 //===----------------------------------------------------------------------===//
103 /// isBuildVectorAllOnes - Return true if the specified node is a
104 /// BUILD_VECTOR where all of the elements are ~0 or undef.
105 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
106 // Look through a bit convert.
107 if (N->getOpcode() == ISD::BIT_CONVERT)
108 N = N->getOperand(0).getNode();
110 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
112 unsigned i = 0, e = N->getNumOperands();
114 // Skip over all of the undef values.
115 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
118 // Do not accept an all-undef vector.
119 if (i == e) return false;
121 // Do not accept build_vectors that aren't all constants or which have non-~0
123 SDValue NotZero = N->getOperand(i);
124 if (isa<ConstantSDNode>(NotZero)) {
125 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
127 } else if (isa<ConstantFPSDNode>(NotZero)) {
128 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
129 bitcastToAPInt().isAllOnesValue())
134 // Okay, we have at least one ~0 value, check to see if the rest match or are
136 for (++i; i != e; ++i)
137 if (N->getOperand(i) != NotZero &&
138 N->getOperand(i).getOpcode() != ISD::UNDEF)
144 /// isBuildVectorAllZeros - Return true if the specified node is a
145 /// BUILD_VECTOR where all of the elements are 0 or undef.
146 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
147 // Look through a bit convert.
148 if (N->getOpcode() == ISD::BIT_CONVERT)
149 N = N->getOperand(0).getNode();
151 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
153 unsigned i = 0, e = N->getNumOperands();
155 // Skip over all of the undef values.
156 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
159 // Do not accept an all-undef vector.
160 if (i == e) return false;
162 // Do not accept build_vectors that aren't all constants or which have non-0
164 SDValue Zero = N->getOperand(i);
165 if (isa<ConstantSDNode>(Zero)) {
166 if (!cast<ConstantSDNode>(Zero)->isNullValue())
168 } else if (isa<ConstantFPSDNode>(Zero)) {
169 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
174 // Okay, we have at least one 0 value, check to see if the rest match or are
176 for (++i; i != e; ++i)
177 if (N->getOperand(i) != Zero &&
178 N->getOperand(i).getOpcode() != ISD::UNDEF)
183 /// isScalarToVector - Return true if the specified node is a
184 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
185 /// element is not an undef.
186 bool ISD::isScalarToVector(const SDNode *N) {
187 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
190 if (N->getOpcode() != ISD::BUILD_VECTOR)
192 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
194 unsigned NumElems = N->getNumOperands();
195 for (unsigned i = 1; i < NumElems; ++i) {
196 SDValue V = N->getOperand(i);
197 if (V.getOpcode() != ISD::UNDEF)
204 /// isDebugLabel - Return true if the specified node represents a debug
205 /// label (i.e. ISD::DBG_LABEL or TargetInstrInfo::DBG_LABEL node).
206 bool ISD::isDebugLabel(const SDNode *N) {
208 if (N->getOpcode() == ISD::DBG_LABEL)
210 if (N->isMachineOpcode() &&
211 N->getMachineOpcode() == TargetInstrInfo::DBG_LABEL)
216 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
217 /// when given the operation for (X op Y).
218 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
219 // To perform this operation, we just need to swap the L and G bits of the
221 unsigned OldL = (Operation >> 2) & 1;
222 unsigned OldG = (Operation >> 1) & 1;
223 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
224 (OldL << 1) | // New G bit
225 (OldG << 2)); // New L bit.
228 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
229 /// 'op' is a valid SetCC operation.
230 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
231 unsigned Operation = Op;
233 Operation ^= 7; // Flip L, G, E bits, but not U.
235 Operation ^= 15; // Flip all of the condition bits.
237 if (Operation > ISD::SETTRUE2)
238 Operation &= ~8; // Don't let N and U bits get set.
240 return ISD::CondCode(Operation);
244 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
245 /// signed operation and 2 if the result is an unsigned comparison. Return zero
246 /// if the operation does not depend on the sign of the input (setne and seteq).
247 static int isSignedOp(ISD::CondCode Opcode) {
249 default: llvm_unreachable("Illegal integer setcc operation!");
251 case ISD::SETNE: return 0;
255 case ISD::SETGE: return 1;
259 case ISD::SETUGE: return 2;
263 /// getSetCCOrOperation - Return the result of a logical OR between different
264 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
265 /// returns SETCC_INVALID if it is not possible to represent the resultant
267 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
269 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
270 // Cannot fold a signed integer setcc with an unsigned integer setcc.
271 return ISD::SETCC_INVALID;
273 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
275 // If the N and U bits get set then the resultant comparison DOES suddenly
276 // care about orderedness, and is true when ordered.
277 if (Op > ISD::SETTRUE2)
278 Op &= ~16; // Clear the U bit if the N bit is set.
280 // Canonicalize illegal integer setcc's.
281 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
284 return ISD::CondCode(Op);
287 /// getSetCCAndOperation - Return the result of a logical AND between different
288 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
289 /// function returns zero if it is not possible to represent the resultant
291 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
293 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
294 // Cannot fold a signed setcc with an unsigned setcc.
295 return ISD::SETCC_INVALID;
297 // Combine all of the condition bits.
298 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
300 // Canonicalize illegal integer setcc's.
304 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
305 case ISD::SETOEQ: // SETEQ & SETU[LG]E
306 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
307 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
308 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
315 const TargetMachine &SelectionDAG::getTarget() const {
316 return MF->getTarget();
319 //===----------------------------------------------------------------------===//
320 // SDNode Profile Support
321 //===----------------------------------------------------------------------===//
323 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
325 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
329 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
330 /// solely with their pointer.
331 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
332 ID.AddPointer(VTList.VTs);
335 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
337 static void AddNodeIDOperands(FoldingSetNodeID &ID,
338 const SDValue *Ops, unsigned NumOps) {
339 for (; NumOps; --NumOps, ++Ops) {
340 ID.AddPointer(Ops->getNode());
341 ID.AddInteger(Ops->getResNo());
345 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
347 static void AddNodeIDOperands(FoldingSetNodeID &ID,
348 const SDUse *Ops, unsigned NumOps) {
349 for (; NumOps; --NumOps, ++Ops) {
350 ID.AddPointer(Ops->getNode());
351 ID.AddInteger(Ops->getResNo());
355 static void AddNodeIDNode(FoldingSetNodeID &ID,
356 unsigned short OpC, SDVTList VTList,
357 const SDValue *OpList, unsigned N) {
358 AddNodeIDOpcode(ID, OpC);
359 AddNodeIDValueTypes(ID, VTList);
360 AddNodeIDOperands(ID, OpList, N);
363 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
365 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
366 switch (N->getOpcode()) {
367 case ISD::TargetExternalSymbol:
368 case ISD::ExternalSymbol:
369 llvm_unreachable("Should only be used on nodes with operands");
370 default: break; // Normal nodes don't need extra info.
371 case ISD::TargetConstant:
373 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
375 case ISD::TargetConstantFP:
376 case ISD::ConstantFP: {
377 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
380 case ISD::TargetGlobalAddress:
381 case ISD::GlobalAddress:
382 case ISD::TargetGlobalTLSAddress:
383 case ISD::GlobalTLSAddress: {
384 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
385 ID.AddPointer(GA->getGlobal());
386 ID.AddInteger(GA->getOffset());
387 ID.AddInteger(GA->getTargetFlags());
390 case ISD::BasicBlock:
391 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
394 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
396 case ISD::DBG_STOPPOINT: {
397 const DbgStopPointSDNode *DSP = cast<DbgStopPointSDNode>(N);
398 ID.AddInteger(DSP->getLine());
399 ID.AddInteger(DSP->getColumn());
400 ID.AddPointer(DSP->getCompileUnit());
404 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
406 case ISD::FrameIndex:
407 case ISD::TargetFrameIndex:
408 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
411 case ISD::TargetJumpTable:
412 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
413 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
415 case ISD::ConstantPool:
416 case ISD::TargetConstantPool: {
417 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
418 ID.AddInteger(CP->getAlignment());
419 ID.AddInteger(CP->getOffset());
420 if (CP->isMachineConstantPoolEntry())
421 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
423 ID.AddPointer(CP->getConstVal());
424 ID.AddInteger(CP->getTargetFlags());
428 const LoadSDNode *LD = cast<LoadSDNode>(N);
429 ID.AddInteger(LD->getMemoryVT().getRawBits());
430 ID.AddInteger(LD->getRawSubclassData());
434 const StoreSDNode *ST = cast<StoreSDNode>(N);
435 ID.AddInteger(ST->getMemoryVT().getRawBits());
436 ID.AddInteger(ST->getRawSubclassData());
439 case ISD::ATOMIC_CMP_SWAP:
440 case ISD::ATOMIC_SWAP:
441 case ISD::ATOMIC_LOAD_ADD:
442 case ISD::ATOMIC_LOAD_SUB:
443 case ISD::ATOMIC_LOAD_AND:
444 case ISD::ATOMIC_LOAD_OR:
445 case ISD::ATOMIC_LOAD_XOR:
446 case ISD::ATOMIC_LOAD_NAND:
447 case ISD::ATOMIC_LOAD_MIN:
448 case ISD::ATOMIC_LOAD_MAX:
449 case ISD::ATOMIC_LOAD_UMIN:
450 case ISD::ATOMIC_LOAD_UMAX: {
451 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
452 ID.AddInteger(AT->getMemoryVT().getRawBits());
453 ID.AddInteger(AT->getRawSubclassData());
456 case ISD::VECTOR_SHUFFLE: {
457 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
458 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
460 ID.AddInteger(SVN->getMaskElt(i));
463 } // end switch (N->getOpcode())
466 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
468 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
469 AddNodeIDOpcode(ID, N->getOpcode());
470 // Add the return value info.
471 AddNodeIDValueTypes(ID, N->getVTList());
472 // Add the operand info.
473 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
475 // Handle SDNode leafs with special info.
476 AddNodeIDCustom(ID, N);
479 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
480 /// the CSE map that carries volatility, indexing mode, and
481 /// extension/truncation information.
483 static inline unsigned
484 encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile) {
485 assert((ConvType & 3) == ConvType &&
486 "ConvType may not require more than 2 bits!");
487 assert((AM & 7) == AM &&
488 "AM may not require more than 3 bits!");
494 //===----------------------------------------------------------------------===//
495 // SelectionDAG Class
496 //===----------------------------------------------------------------------===//
498 /// doNotCSE - Return true if CSE should not be performed for this node.
499 static bool doNotCSE(SDNode *N) {
500 if (N->getValueType(0) == MVT::Flag)
501 return true; // Never CSE anything that produces a flag.
503 switch (N->getOpcode()) {
505 case ISD::HANDLENODE:
507 case ISD::DBG_STOPPOINT:
509 return true; // Never CSE these nodes.
512 // Check that remaining values produced are not flags.
513 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
514 if (N->getValueType(i) == MVT::Flag)
515 return true; // Never CSE anything that produces a flag.
520 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
522 void SelectionDAG::RemoveDeadNodes() {
523 // Create a dummy node (which is not added to allnodes), that adds a reference
524 // to the root node, preventing it from being deleted.
525 HandleSDNode Dummy(getRoot());
527 SmallVector<SDNode*, 128> DeadNodes;
529 // Add all obviously-dead nodes to the DeadNodes worklist.
530 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
532 DeadNodes.push_back(I);
534 RemoveDeadNodes(DeadNodes);
536 // If the root changed (e.g. it was a dead load, update the root).
537 setRoot(Dummy.getValue());
540 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
541 /// given list, and any nodes that become unreachable as a result.
542 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
543 DAGUpdateListener *UpdateListener) {
545 // Process the worklist, deleting the nodes and adding their uses to the
547 while (!DeadNodes.empty()) {
548 SDNode *N = DeadNodes.pop_back_val();
551 UpdateListener->NodeDeleted(N, 0);
553 // Take the node out of the appropriate CSE map.
554 RemoveNodeFromCSEMaps(N);
556 // Next, brutally remove the operand list. This is safe to do, as there are
557 // no cycles in the graph.
558 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
560 SDNode *Operand = Use.getNode();
563 // Now that we removed this operand, see if there are no uses of it left.
564 if (Operand->use_empty())
565 DeadNodes.push_back(Operand);
572 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
573 SmallVector<SDNode*, 16> DeadNodes(1, N);
574 RemoveDeadNodes(DeadNodes, UpdateListener);
577 void SelectionDAG::DeleteNode(SDNode *N) {
578 // First take this out of the appropriate CSE map.
579 RemoveNodeFromCSEMaps(N);
581 // Finally, remove uses due to operands of this node, remove from the
582 // AllNodes list, and delete the node.
583 DeleteNodeNotInCSEMaps(N);
586 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
587 assert(N != AllNodes.begin() && "Cannot delete the entry node!");
588 assert(N->use_empty() && "Cannot delete a node that is not dead!");
590 // Drop all of the operands and decrement used node's use counts.
596 void SelectionDAG::DeallocateNode(SDNode *N) {
597 if (N->OperandsNeedDelete)
598 delete[] N->OperandList;
600 // Set the opcode to DELETED_NODE to help catch bugs when node
601 // memory is reallocated.
602 N->NodeType = ISD::DELETED_NODE;
604 NodeAllocator.Deallocate(AllNodes.remove(N));
607 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
608 /// correspond to it. This is useful when we're about to delete or repurpose
609 /// the node. We don't want future request for structurally identical nodes
610 /// to return N anymore.
611 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
613 switch (N->getOpcode()) {
614 case ISD::EntryToken:
615 llvm_unreachable("EntryToken should not be in CSEMaps!");
617 case ISD::HANDLENODE: return false; // noop.
619 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
620 "Cond code doesn't exist!");
621 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
622 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
624 case ISD::ExternalSymbol:
625 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
627 case ISD::TargetExternalSymbol: {
628 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
629 Erased = TargetExternalSymbols.erase(
630 std::pair<std::string,unsigned char>(ESN->getSymbol(),
631 ESN->getTargetFlags()));
634 case ISD::VALUETYPE: {
635 EVT VT = cast<VTSDNode>(N)->getVT();
636 if (VT.isExtended()) {
637 Erased = ExtendedValueTypeNodes.erase(VT);
639 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0;
640 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0;
645 // Remove it from the CSE Map.
646 Erased = CSEMap.RemoveNode(N);
650 // Verify that the node was actually in one of the CSE maps, unless it has a
651 // flag result (which cannot be CSE'd) or is one of the special cases that are
652 // not subject to CSE.
653 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
654 !N->isMachineOpcode() && !doNotCSE(N)) {
657 llvm_unreachable("Node is not in map!");
663 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
664 /// maps and modified in place. Add it back to the CSE maps, unless an identical
665 /// node already exists, in which case transfer all its users to the existing
666 /// node. This transfer can potentially trigger recursive merging.
669 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N,
670 DAGUpdateListener *UpdateListener) {
671 // For node types that aren't CSE'd, just act as if no identical node
674 SDNode *Existing = CSEMap.GetOrInsertNode(N);
676 // If there was already an existing matching node, use ReplaceAllUsesWith
677 // to replace the dead one with the existing one. This can cause
678 // recursive merging of other unrelated nodes down the line.
679 ReplaceAllUsesWith(N, Existing, UpdateListener);
681 // N is now dead. Inform the listener if it exists and delete it.
683 UpdateListener->NodeDeleted(N, Existing);
684 DeleteNodeNotInCSEMaps(N);
689 // If the node doesn't already exist, we updated it. Inform a listener if
692 UpdateListener->NodeUpdated(N);
695 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
696 /// were replaced with those specified. If this node is never memoized,
697 /// return null, otherwise return a pointer to the slot it would take. If a
698 /// node already exists with these operands, the slot will be non-null.
699 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
704 SDValue Ops[] = { Op };
706 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
707 AddNodeIDCustom(ID, N);
708 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
711 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
712 /// were replaced with those specified. If this node is never memoized,
713 /// return null, otherwise return a pointer to the slot it would take. If a
714 /// node already exists with these operands, the slot will be non-null.
715 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
716 SDValue Op1, SDValue Op2,
721 SDValue Ops[] = { Op1, Op2 };
723 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
724 AddNodeIDCustom(ID, N);
725 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
729 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
730 /// were replaced with those specified. If this node is never memoized,
731 /// return null, otherwise return a pointer to the slot it would take. If a
732 /// node already exists with these operands, the slot will be non-null.
733 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
734 const SDValue *Ops,unsigned NumOps,
740 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
741 AddNodeIDCustom(ID, N);
742 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
745 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
746 void SelectionDAG::VerifyNode(SDNode *N) {
747 switch (N->getOpcode()) {
750 case ISD::BUILD_PAIR: {
751 EVT VT = N->getValueType(0);
752 assert(N->getNumValues() == 1 && "Too many results!");
753 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
754 "Wrong return type!");
755 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
756 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
757 "Mismatched operand types!");
758 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
759 "Wrong operand type!");
760 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
761 "Wrong return type size");
764 case ISD::BUILD_VECTOR: {
765 assert(N->getNumValues() == 1 && "Too many results!");
766 assert(N->getValueType(0).isVector() && "Wrong return type!");
767 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
768 "Wrong number of operands!");
769 EVT EltVT = N->getValueType(0).getVectorElementType();
770 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
771 assert((I->getValueType() == EltVT ||
772 (EltVT.isInteger() && I->getValueType().isInteger() &&
773 EltVT.bitsLE(I->getValueType()))) &&
774 "Wrong operand type!");
780 /// getEVTAlignment - Compute the default alignment value for the
783 unsigned SelectionDAG::getEVTAlignment(EVT VT) const {
784 const Type *Ty = VT == MVT::iPTR ?
785 PointerType::get(Type::getInt8Ty(*getContext()), 0) :
786 VT.getTypeForEVT(*getContext());
788 return TLI.getTargetData()->getABITypeAlignment(Ty);
791 // EntryNode could meaningfully have debug info if we can find it...
792 SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli)
793 : TLI(tli), FLI(fli), DW(0),
794 EntryNode(ISD::EntryToken, DebugLoc::getUnknownLoc(),
795 getVTList(MVT::Other)), Root(getEntryNode()) {
796 AllNodes.push_back(&EntryNode);
799 void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi,
804 Context = &mf.getFunction()->getContext();
807 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();
836 SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
837 return VT.bitsGT(Op.getValueType()) ?
838 getNode(ISD::SIGN_EXTEND, DL, VT, Op) :
839 getNode(ISD::TRUNCATE, DL, VT, Op);
842 SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) {
843 return VT.bitsGT(Op.getValueType()) ?
844 getNode(ISD::ZERO_EXTEND, DL, VT, Op) :
845 getNode(ISD::TRUNCATE, DL, VT, Op);
848 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) {
849 if (Op.getValueType() == VT) return Op;
850 APInt Imm = APInt::getLowBitsSet(Op.getValueSizeInBits(),
852 return getNode(ISD::AND, DL, Op.getValueType(), Op,
853 getConstant(Imm, Op.getValueType()));
856 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
858 SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) {
859 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
861 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
862 return getNode(ISD::XOR, DL, VT, Val, NegOne);
865 SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) {
866 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
867 assert((EltVT.getSizeInBits() >= 64 ||
868 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
869 "getConstant with a uint64_t value that doesn't fit in the type!");
870 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
873 SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) {
874 return getConstant(*ConstantInt::get(*Context, Val), VT, isT);
877 SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
878 assert(VT.isInteger() && "Cannot create FP integer constant!");
880 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
881 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
882 "APInt size does not match type size!");
884 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
886 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
890 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
892 return SDValue(N, 0);
894 N = NodeAllocator.Allocate<ConstantSDNode>();
895 new (N) ConstantSDNode(isT, &Val, EltVT);
896 CSEMap.InsertNode(N, IP);
897 AllNodes.push_back(N);
900 SDValue Result(N, 0);
902 SmallVector<SDValue, 8> Ops;
903 Ops.assign(VT.getVectorNumElements(), Result);
904 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
905 VT, &Ops[0], Ops.size());
910 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
911 return getConstant(Val, TLI.getPointerTy(), isTarget);
915 SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) {
916 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget);
919 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){
920 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
923 VT.isVector() ? VT.getVectorElementType() : VT;
925 // Do the map lookup using the actual bit pattern for the floating point
926 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
927 // we don't have issues with SNANs.
928 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
930 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
934 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
936 return SDValue(N, 0);
938 N = NodeAllocator.Allocate<ConstantFPSDNode>();
939 new (N) ConstantFPSDNode(isTarget, &V, EltVT);
940 CSEMap.InsertNode(N, IP);
941 AllNodes.push_back(N);
944 SDValue Result(N, 0);
946 SmallVector<SDValue, 8> Ops;
947 Ops.assign(VT.getVectorNumElements(), Result);
948 // FIXME DebugLoc info might be appropriate here
949 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(),
950 VT, &Ops[0], Ops.size());
955 SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) {
957 VT.isVector() ? VT.getVectorElementType() : VT;
959 return getConstantFP(APFloat((float)Val), VT, isTarget);
961 return getConstantFP(APFloat(Val), VT, isTarget);
964 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
965 EVT VT, int64_t Offset,
967 unsigned char TargetFlags) {
968 assert((TargetFlags == 0 || isTargetGA) &&
969 "Cannot set target flags on target-independent globals");
971 // Truncate (with sign-extension) the offset value to the pointer size.
972 EVT PTy = TLI.getPointerTy();
973 unsigned BitWidth = PTy.getSizeInBits();
975 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
977 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
979 // If GV is an alias then use the aliasee for determining thread-localness.
980 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
981 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
985 if (GVar && GVar->isThreadLocal())
986 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
988 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
991 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
993 ID.AddInteger(Offset);
994 ID.AddInteger(TargetFlags);
996 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
997 return SDValue(E, 0);
998 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>();
999 new (N) GlobalAddressSDNode(Opc, GV, VT, Offset, TargetFlags);
1000 CSEMap.InsertNode(N, IP);
1001 AllNodes.push_back(N);
1002 return SDValue(N, 0);
1005 SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1006 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1007 FoldingSetNodeID ID;
1008 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1011 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1012 return SDValue(E, 0);
1013 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>();
1014 new (N) FrameIndexSDNode(FI, VT, isTarget);
1015 CSEMap.InsertNode(N, IP);
1016 AllNodes.push_back(N);
1017 return SDValue(N, 0);
1020 SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1021 unsigned char TargetFlags) {
1022 assert((TargetFlags == 0 || isTarget) &&
1023 "Cannot set target flags on target-independent jump tables");
1024 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1025 FoldingSetNodeID ID;
1026 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1028 ID.AddInteger(TargetFlags);
1030 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1031 return SDValue(E, 0);
1032 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>();
1033 new (N) JumpTableSDNode(JTI, VT, isTarget, TargetFlags);
1034 CSEMap.InsertNode(N, IP);
1035 AllNodes.push_back(N);
1036 return SDValue(N, 0);
1039 SDValue SelectionDAG::getConstantPool(Constant *C, EVT VT,
1040 unsigned Alignment, int Offset,
1042 unsigned char TargetFlags) {
1043 assert((TargetFlags == 0 || isTarget) &&
1044 "Cannot set target flags on target-independent globals");
1046 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1047 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1048 FoldingSetNodeID ID;
1049 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1050 ID.AddInteger(Alignment);
1051 ID.AddInteger(Offset);
1053 ID.AddInteger(TargetFlags);
1055 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1056 return SDValue(E, 0);
1057 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1058 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1059 CSEMap.InsertNode(N, IP);
1060 AllNodes.push_back(N);
1061 return SDValue(N, 0);
1065 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1066 unsigned Alignment, int Offset,
1068 unsigned char TargetFlags) {
1069 assert((TargetFlags == 0 || isTarget) &&
1070 "Cannot set target flags on target-independent globals");
1072 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType());
1073 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1074 FoldingSetNodeID ID;
1075 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1076 ID.AddInteger(Alignment);
1077 ID.AddInteger(Offset);
1078 C->AddSelectionDAGCSEId(ID);
1079 ID.AddInteger(TargetFlags);
1081 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1082 return SDValue(E, 0);
1083 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1084 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags);
1085 CSEMap.InsertNode(N, IP);
1086 AllNodes.push_back(N);
1087 return SDValue(N, 0);
1090 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1091 FoldingSetNodeID ID;
1092 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1095 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1096 return SDValue(E, 0);
1097 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>();
1098 new (N) BasicBlockSDNode(MBB);
1099 CSEMap.InsertNode(N, IP);
1100 AllNodes.push_back(N);
1101 return SDValue(N, 0);
1104 SDValue SelectionDAG::getValueType(EVT VT) {
1105 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
1106 ValueTypeNodes.size())
1107 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1);
1109 SDNode *&N = VT.isExtended() ?
1110 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
1112 if (N) return SDValue(N, 0);
1113 N = NodeAllocator.Allocate<VTSDNode>();
1114 new (N) VTSDNode(VT);
1115 AllNodes.push_back(N);
1116 return SDValue(N, 0);
1119 SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
1120 SDNode *&N = ExternalSymbols[Sym];
1121 if (N) return SDValue(N, 0);
1122 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1123 new (N) ExternalSymbolSDNode(false, Sym, 0, VT);
1124 AllNodes.push_back(N);
1125 return SDValue(N, 0);
1128 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
1129 unsigned char TargetFlags) {
1131 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym,
1133 if (N) return SDValue(N, 0);
1134 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1135 new (N) ExternalSymbolSDNode(true, Sym, TargetFlags, VT);
1136 AllNodes.push_back(N);
1137 return SDValue(N, 0);
1140 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1141 if ((unsigned)Cond >= CondCodeNodes.size())
1142 CondCodeNodes.resize(Cond+1);
1144 if (CondCodeNodes[Cond] == 0) {
1145 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>();
1146 new (N) CondCodeSDNode(Cond);
1147 CondCodeNodes[Cond] = N;
1148 AllNodes.push_back(N);
1150 return SDValue(CondCodeNodes[Cond], 0);
1153 // commuteShuffle - swaps the values of N1 and N2, and swaps all indices in
1154 // the shuffle mask M that point at N1 to point at N2, and indices that point
1155 // N2 to point at N1.
1156 static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) {
1158 int NElts = M.size();
1159 for (int i = 0; i != NElts; ++i) {
1167 SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1,
1168 SDValue N2, const int *Mask) {
1169 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE");
1170 assert(VT.isVector() && N1.getValueType().isVector() &&
1171 "Vector Shuffle VTs must be a vectors");
1172 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType()
1173 && "Vector Shuffle VTs must have same element type");
1175 // Canonicalize shuffle undef, undef -> undef
1176 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF)
1177 return getUNDEF(VT);
1179 // Validate that all indices in Mask are within the range of the elements
1180 // input to the shuffle.
1181 unsigned NElts = VT.getVectorNumElements();
1182 SmallVector<int, 8> MaskVec;
1183 for (unsigned i = 0; i != NElts; ++i) {
1184 assert(Mask[i] < (int)(NElts * 2) && "Index out of range");
1185 MaskVec.push_back(Mask[i]);
1188 // Canonicalize shuffle v, v -> v, undef
1191 for (unsigned i = 0; i != NElts; ++i)
1192 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts;
1195 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1196 if (N1.getOpcode() == ISD::UNDEF)
1197 commuteShuffle(N1, N2, MaskVec);
1199 // Canonicalize all index into lhs, -> shuffle lhs, undef
1200 // Canonicalize all index into rhs, -> shuffle rhs, undef
1201 bool AllLHS = true, AllRHS = true;
1202 bool N2Undef = N2.getOpcode() == ISD::UNDEF;
1203 for (unsigned i = 0; i != NElts; ++i) {
1204 if (MaskVec[i] >= (int)NElts) {
1209 } else if (MaskVec[i] >= 0) {
1213 if (AllLHS && AllRHS)
1214 return getUNDEF(VT);
1215 if (AllLHS && !N2Undef)
1219 commuteShuffle(N1, N2, MaskVec);
1222 // If Identity shuffle, or all shuffle in to undef, return that node.
1223 bool AllUndef = true;
1224 bool Identity = true;
1225 for (unsigned i = 0; i != NElts; ++i) {
1226 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false;
1227 if (MaskVec[i] >= 0) AllUndef = false;
1229 if (Identity && NElts == N1.getValueType().getVectorNumElements())
1232 return getUNDEF(VT);
1234 FoldingSetNodeID ID;
1235 SDValue Ops[2] = { N1, N2 };
1236 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2);
1237 for (unsigned i = 0; i != NElts; ++i)
1238 ID.AddInteger(MaskVec[i]);
1241 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1242 return SDValue(E, 0);
1244 // Allocate the mask array for the node out of the BumpPtrAllocator, since
1245 // SDNode doesn't have access to it. This memory will be "leaked" when
1246 // the node is deallocated, but recovered when the NodeAllocator is released.
1247 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
1248 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int));
1250 ShuffleVectorSDNode *N = NodeAllocator.Allocate<ShuffleVectorSDNode>();
1251 new (N) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc);
1252 CSEMap.InsertNode(N, IP);
1253 AllNodes.push_back(N);
1254 return SDValue(N, 0);
1257 SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl,
1258 SDValue Val, SDValue DTy,
1259 SDValue STy, SDValue Rnd, SDValue Sat,
1260 ISD::CvtCode Code) {
1261 // If the src and dest types are the same and the conversion is between
1262 // integer types of the same sign or two floats, no conversion is necessary.
1264 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF))
1267 FoldingSetNodeID ID;
1269 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1270 return SDValue(E, 0);
1271 CvtRndSatSDNode *N = NodeAllocator.Allocate<CvtRndSatSDNode>();
1272 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat };
1273 new (N) CvtRndSatSDNode(VT, dl, Ops, 5, Code);
1274 CSEMap.InsertNode(N, IP);
1275 AllNodes.push_back(N);
1276 return SDValue(N, 0);
1279 SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) {
1280 FoldingSetNodeID ID;
1281 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1282 ID.AddInteger(RegNo);
1284 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1285 return SDValue(E, 0);
1286 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>();
1287 new (N) RegisterSDNode(RegNo, VT);
1288 CSEMap.InsertNode(N, IP);
1289 AllNodes.push_back(N);
1290 return SDValue(N, 0);
1293 SDValue SelectionDAG::getDbgStopPoint(DebugLoc DL, SDValue Root,
1294 unsigned Line, unsigned Col,
1296 SDNode *N = NodeAllocator.Allocate<DbgStopPointSDNode>();
1297 new (N) DbgStopPointSDNode(Root, Line, Col, CU);
1299 AllNodes.push_back(N);
1300 return SDValue(N, 0);
1303 SDValue SelectionDAG::getLabel(unsigned Opcode, DebugLoc dl,
1306 FoldingSetNodeID ID;
1307 SDValue Ops[] = { Root };
1308 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), &Ops[0], 1);
1309 ID.AddInteger(LabelID);
1311 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1312 return SDValue(E, 0);
1313 SDNode *N = NodeAllocator.Allocate<LabelSDNode>();
1314 new (N) LabelSDNode(Opcode, dl, Root, LabelID);
1315 CSEMap.InsertNode(N, IP);
1316 AllNodes.push_back(N);
1317 return SDValue(N, 0);
1320 SDValue SelectionDAG::getSrcValue(const Value *V) {
1321 assert((!V || isa<PointerType>(V->getType())) &&
1322 "SrcValue is not a pointer?");
1324 FoldingSetNodeID ID;
1325 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1329 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1330 return SDValue(E, 0);
1332 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>();
1333 new (N) SrcValueSDNode(V);
1334 CSEMap.InsertNode(N, IP);
1335 AllNodes.push_back(N);
1336 return SDValue(N, 0);
1339 /// getShiftAmountOperand - Return the specified value casted to
1340 /// the target's desired shift amount type.
1341 SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) {
1342 EVT OpTy = Op.getValueType();
1343 MVT ShTy = TLI.getShiftAmountTy();
1344 if (OpTy == ShTy || OpTy.isVector()) return Op;
1346 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
1347 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op);
1350 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1351 /// specified value type.
1352 SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
1353 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1354 unsigned ByteSize = VT.getStoreSize();
1355 const Type *Ty = VT.getTypeForEVT(*getContext());
1356 unsigned StackAlign =
1357 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1359 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign);
1360 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1363 /// CreateStackTemporary - Create a stack temporary suitable for holding
1364 /// either of the specified value types.
1365 SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
1366 unsigned Bytes = std::max(VT1.getStoreSizeInBits(),
1367 VT2.getStoreSizeInBits())/8;
1368 const Type *Ty1 = VT1.getTypeForEVT(*getContext());
1369 const Type *Ty2 = VT2.getTypeForEVT(*getContext());
1370 const TargetData *TD = TLI.getTargetData();
1371 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1),
1372 TD->getPrefTypeAlignment(Ty2));
1374 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1375 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align);
1376 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1379 SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1,
1380 SDValue N2, ISD::CondCode Cond, DebugLoc dl) {
1381 // These setcc operations always fold.
1385 case ISD::SETFALSE2: return getConstant(0, VT);
1387 case ISD::SETTRUE2: return getConstant(1, VT);
1399 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1403 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1404 const APInt &C2 = N2C->getAPIntValue();
1405 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1406 const APInt &C1 = N1C->getAPIntValue();
1409 default: llvm_unreachable("Unknown integer setcc!");
1410 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1411 case ISD::SETNE: return getConstant(C1 != C2, VT);
1412 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1413 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1414 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1415 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1416 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1417 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1418 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1419 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1423 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1424 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1425 // No compile time operations on this type yet.
1426 if (N1C->getValueType(0) == MVT::ppcf128)
1429 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1432 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1433 return getUNDEF(VT);
1435 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1436 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1437 return getUNDEF(VT);
1439 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1440 R==APFloat::cmpLessThan, VT);
1441 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1442 return getUNDEF(VT);
1444 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1445 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1446 return getUNDEF(VT);
1448 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1449 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1450 return getUNDEF(VT);
1452 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1453 R==APFloat::cmpEqual, VT);
1454 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1455 return getUNDEF(VT);
1457 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1458 R==APFloat::cmpEqual, VT);
1459 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1460 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1461 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1462 R==APFloat::cmpEqual, VT);
1463 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1464 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1465 R==APFloat::cmpLessThan, VT);
1466 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1467 R==APFloat::cmpUnordered, VT);
1468 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1469 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1472 // Ensure that the constant occurs on the RHS.
1473 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1477 // Could not fold it.
1481 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1482 /// use this predicate to simplify operations downstream.
1483 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1484 // This predicate is not safe for vector operations.
1485 if (Op.getValueType().isVector())
1488 unsigned BitWidth = Op.getValueSizeInBits();
1489 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1492 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1493 /// this predicate to simplify operations downstream. Mask is known to be zero
1494 /// for bits that V cannot have.
1495 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1496 unsigned Depth) const {
1497 APInt KnownZero, KnownOne;
1498 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1499 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1500 return (KnownZero & Mask) == Mask;
1503 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1504 /// known to be either zero or one and return them in the KnownZero/KnownOne
1505 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1507 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1508 APInt &KnownZero, APInt &KnownOne,
1509 unsigned Depth) const {
1510 unsigned BitWidth = Mask.getBitWidth();
1511 assert(BitWidth == Op.getValueType().getSizeInBits() &&
1512 "Mask size mismatches value type size!");
1514 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1515 if (Depth == 6 || Mask == 0)
1516 return; // Limit search depth.
1518 APInt KnownZero2, KnownOne2;
1520 switch (Op.getOpcode()) {
1522 // We know all of the bits for a constant!
1523 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1524 KnownZero = ~KnownOne & Mask;
1527 // If either the LHS or the RHS are Zero, the result is zero.
1528 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1529 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1530 KnownZero2, KnownOne2, Depth+1);
1531 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1532 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1534 // Output known-1 bits are only known if set in both the LHS & RHS.
1535 KnownOne &= KnownOne2;
1536 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1537 KnownZero |= KnownZero2;
1540 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1541 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1542 KnownZero2, KnownOne2, Depth+1);
1543 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1544 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1546 // Output known-0 bits are only known if clear in both the LHS & RHS.
1547 KnownZero &= KnownZero2;
1548 // Output known-1 are known to be set if set in either the LHS | RHS.
1549 KnownOne |= KnownOne2;
1552 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1553 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1554 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1555 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1557 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1558 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1559 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1560 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1561 KnownZero = KnownZeroOut;
1565 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1566 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1567 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1568 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1569 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1571 // If low bits are zero in either operand, output low known-0 bits.
1572 // Also compute a conserative estimate for high known-0 bits.
1573 // More trickiness is possible, but this is sufficient for the
1574 // interesting case of alignment computation.
1576 unsigned TrailZ = KnownZero.countTrailingOnes() +
1577 KnownZero2.countTrailingOnes();
1578 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1579 KnownZero2.countLeadingOnes(),
1580 BitWidth) - BitWidth;
1582 TrailZ = std::min(TrailZ, BitWidth);
1583 LeadZ = std::min(LeadZ, BitWidth);
1584 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1585 APInt::getHighBitsSet(BitWidth, LeadZ);
1590 // For the purposes of computing leading zeros we can conservatively
1591 // treat a udiv as a logical right shift by the power of 2 known to
1592 // be less than the denominator.
1593 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1594 ComputeMaskedBits(Op.getOperand(0),
1595 AllOnes, KnownZero2, KnownOne2, Depth+1);
1596 unsigned LeadZ = KnownZero2.countLeadingOnes();
1600 ComputeMaskedBits(Op.getOperand(1),
1601 AllOnes, KnownZero2, KnownOne2, Depth+1);
1602 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1603 if (RHSUnknownLeadingOnes != BitWidth)
1604 LeadZ = std::min(BitWidth,
1605 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1607 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1611 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1612 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1613 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1614 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1616 // Only known if known in both the LHS and RHS.
1617 KnownOne &= KnownOne2;
1618 KnownZero &= KnownZero2;
1620 case ISD::SELECT_CC:
1621 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1622 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1623 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1624 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1626 // Only known if known in both the LHS and RHS.
1627 KnownOne &= KnownOne2;
1628 KnownZero &= KnownZero2;
1636 if (Op.getResNo() != 1)
1638 // The boolean result conforms to getBooleanContents. Fall through.
1640 // If we know the result of a setcc has the top bits zero, use this info.
1641 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent &&
1643 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1646 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1647 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1648 unsigned ShAmt = SA->getZExtValue();
1650 // If the shift count is an invalid immediate, don't do anything.
1651 if (ShAmt >= BitWidth)
1654 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt),
1655 KnownZero, KnownOne, Depth+1);
1656 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1657 KnownZero <<= ShAmt;
1659 // low bits known zero.
1660 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1664 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1665 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1666 unsigned ShAmt = SA->getZExtValue();
1668 // If the shift count is an invalid immediate, don't do anything.
1669 if (ShAmt >= BitWidth)
1672 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1673 KnownZero, KnownOne, Depth+1);
1674 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1675 KnownZero = KnownZero.lshr(ShAmt);
1676 KnownOne = KnownOne.lshr(ShAmt);
1678 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1679 KnownZero |= HighBits; // High bits known zero.
1683 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1684 unsigned ShAmt = SA->getZExtValue();
1686 // If the shift count is an invalid immediate, don't do anything.
1687 if (ShAmt >= BitWidth)
1690 APInt InDemandedMask = (Mask << ShAmt);
1691 // If any of the demanded bits are produced by the sign extension, we also
1692 // demand the input sign bit.
1693 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1694 if (HighBits.getBoolValue())
1695 InDemandedMask |= APInt::getSignBit(BitWidth);
1697 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1699 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1700 KnownZero = KnownZero.lshr(ShAmt);
1701 KnownOne = KnownOne.lshr(ShAmt);
1703 // Handle the sign bits.
1704 APInt SignBit = APInt::getSignBit(BitWidth);
1705 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1707 if (KnownZero.intersects(SignBit)) {
1708 KnownZero |= HighBits; // New bits are known zero.
1709 } else if (KnownOne.intersects(SignBit)) {
1710 KnownOne |= HighBits; // New bits are known one.
1714 case ISD::SIGN_EXTEND_INREG: {
1715 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1716 unsigned EBits = EVT.getSizeInBits();
1718 // Sign extension. Compute the demanded bits in the result that are not
1719 // present in the input.
1720 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1722 APInt InSignBit = APInt::getSignBit(EBits);
1723 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1725 // If the sign extended bits are demanded, we know that the sign
1727 InSignBit.zext(BitWidth);
1728 if (NewBits.getBoolValue())
1729 InputDemandedBits |= InSignBit;
1731 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1732 KnownZero, KnownOne, Depth+1);
1733 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1735 // If the sign bit of the input is known set or clear, then we know the
1736 // top bits of the result.
1737 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1738 KnownZero |= NewBits;
1739 KnownOne &= ~NewBits;
1740 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1741 KnownOne |= NewBits;
1742 KnownZero &= ~NewBits;
1743 } else { // Input sign bit unknown
1744 KnownZero &= ~NewBits;
1745 KnownOne &= ~NewBits;
1752 unsigned LowBits = Log2_32(BitWidth)+1;
1753 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1758 if (ISD::isZEXTLoad(Op.getNode())) {
1759 LoadSDNode *LD = cast<LoadSDNode>(Op);
1760 EVT VT = LD->getMemoryVT();
1761 unsigned MemBits = VT.getSizeInBits();
1762 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1766 case ISD::ZERO_EXTEND: {
1767 EVT InVT = Op.getOperand(0).getValueType();
1768 unsigned InBits = InVT.getSizeInBits();
1769 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1770 APInt InMask = Mask;
1771 InMask.trunc(InBits);
1772 KnownZero.trunc(InBits);
1773 KnownOne.trunc(InBits);
1774 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1775 KnownZero.zext(BitWidth);
1776 KnownOne.zext(BitWidth);
1777 KnownZero |= NewBits;
1780 case ISD::SIGN_EXTEND: {
1781 EVT InVT = Op.getOperand(0).getValueType();
1782 unsigned InBits = InVT.getSizeInBits();
1783 APInt InSignBit = APInt::getSignBit(InBits);
1784 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1785 APInt InMask = Mask;
1786 InMask.trunc(InBits);
1788 // If any of the sign extended bits are demanded, we know that the sign
1789 // bit is demanded. Temporarily set this bit in the mask for our callee.
1790 if (NewBits.getBoolValue())
1791 InMask |= InSignBit;
1793 KnownZero.trunc(InBits);
1794 KnownOne.trunc(InBits);
1795 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1797 // Note if the sign bit is known to be zero or one.
1798 bool SignBitKnownZero = KnownZero.isNegative();
1799 bool SignBitKnownOne = KnownOne.isNegative();
1800 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1801 "Sign bit can't be known to be both zero and one!");
1803 // If the sign bit wasn't actually demanded by our caller, we don't
1804 // want it set in the KnownZero and KnownOne result values. Reset the
1805 // mask and reapply it to the result values.
1807 InMask.trunc(InBits);
1808 KnownZero &= InMask;
1811 KnownZero.zext(BitWidth);
1812 KnownOne.zext(BitWidth);
1814 // If the sign bit is known zero or one, the top bits match.
1815 if (SignBitKnownZero)
1816 KnownZero |= NewBits;
1817 else if (SignBitKnownOne)
1818 KnownOne |= NewBits;
1821 case ISD::ANY_EXTEND: {
1822 EVT InVT = Op.getOperand(0).getValueType();
1823 unsigned InBits = InVT.getSizeInBits();
1824 APInt InMask = Mask;
1825 InMask.trunc(InBits);
1826 KnownZero.trunc(InBits);
1827 KnownOne.trunc(InBits);
1828 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1829 KnownZero.zext(BitWidth);
1830 KnownOne.zext(BitWidth);
1833 case ISD::TRUNCATE: {
1834 EVT InVT = Op.getOperand(0).getValueType();
1835 unsigned InBits = InVT.getSizeInBits();
1836 APInt InMask = Mask;
1837 InMask.zext(InBits);
1838 KnownZero.zext(InBits);
1839 KnownOne.zext(InBits);
1840 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1841 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1842 KnownZero.trunc(BitWidth);
1843 KnownOne.trunc(BitWidth);
1846 case ISD::AssertZext: {
1847 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1848 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1849 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1851 KnownZero |= (~InMask) & Mask;
1855 // All bits are zero except the low bit.
1856 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1860 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1861 // We know that the top bits of C-X are clear if X contains less bits
1862 // than C (i.e. no wrap-around can happen). For example, 20-X is
1863 // positive if we can prove that X is >= 0 and < 16.
1864 if (CLHS->getAPIntValue().isNonNegative()) {
1865 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1866 // NLZ can't be BitWidth with no sign bit
1867 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1868 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1871 // If all of the MaskV bits are known to be zero, then we know the
1872 // output top bits are zero, because we now know that the output is
1874 if ((KnownZero2 & MaskV) == MaskV) {
1875 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1876 // Top bits known zero.
1877 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1884 // Output known-0 bits are known if clear or set in both the low clear bits
1885 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1886 // low 3 bits clear.
1887 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
1888 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1889 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1890 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1892 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1893 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1894 KnownZeroOut = std::min(KnownZeroOut,
1895 KnownZero2.countTrailingOnes());
1897 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1901 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1902 const APInt &RA = Rem->getAPIntValue();
1903 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1904 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1905 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1906 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1908 // If the sign bit of the first operand is zero, the sign bit of
1909 // the result is zero. If the first operand has no one bits below
1910 // the second operand's single 1 bit, its sign will be zero.
1911 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1912 KnownZero2 |= ~LowBits;
1914 KnownZero |= KnownZero2 & Mask;
1916 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1921 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1922 const APInt &RA = Rem->getAPIntValue();
1923 if (RA.isPowerOf2()) {
1924 APInt LowBits = (RA - 1);
1925 APInt Mask2 = LowBits & Mask;
1926 KnownZero |= ~LowBits & Mask;
1927 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1928 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1933 // Since the result is less than or equal to either operand, any leading
1934 // zero bits in either operand must also exist in the result.
1935 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1936 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne,
1938 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2,
1941 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1942 KnownZero2.countLeadingOnes());
1944 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1948 // Allow the target to implement this method for its nodes.
1949 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
1950 case ISD::INTRINSIC_WO_CHAIN:
1951 case ISD::INTRINSIC_W_CHAIN:
1952 case ISD::INTRINSIC_VOID:
1953 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this,
1960 /// ComputeNumSignBits - Return the number of times the sign bit of the
1961 /// register is replicated into the other bits. We know that at least 1 bit
1962 /// is always equal to the sign bit (itself), but other cases can give us
1963 /// information. For example, immediately after an "SRA X, 2", we know that
1964 /// the top 3 bits are all equal to each other, so we return 3.
1965 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
1966 EVT VT = Op.getValueType();
1967 assert(VT.isInteger() && "Invalid VT!");
1968 unsigned VTBits = VT.getSizeInBits();
1970 unsigned FirstAnswer = 1;
1973 return 1; // Limit search depth.
1975 switch (Op.getOpcode()) {
1977 case ISD::AssertSext:
1978 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
1979 return VTBits-Tmp+1;
1980 case ISD::AssertZext:
1981 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
1984 case ISD::Constant: {
1985 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
1986 // If negative, return # leading ones.
1987 if (Val.isNegative())
1988 return Val.countLeadingOnes();
1990 // Return # leading zeros.
1991 return Val.countLeadingZeros();
1994 case ISD::SIGN_EXTEND:
1995 Tmp = VTBits-Op.getOperand(0).getValueType().getSizeInBits();
1996 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
1998 case ISD::SIGN_EXTEND_INREG:
1999 // Max of the input and what this extends.
2000 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
2003 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2004 return std::max(Tmp, Tmp2);
2007 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2008 // SRA X, C -> adds C sign bits.
2009 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2010 Tmp += C->getZExtValue();
2011 if (Tmp > VTBits) Tmp = VTBits;
2015 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2016 // shl destroys sign bits.
2017 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2018 if (C->getZExtValue() >= VTBits || // Bad shift.
2019 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2020 return Tmp - C->getZExtValue();
2025 case ISD::XOR: // NOT is handled here.
2026 // Logical binary ops preserve the number of sign bits at the worst.
2027 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2029 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2030 FirstAnswer = std::min(Tmp, Tmp2);
2031 // We computed what we know about the sign bits as our first
2032 // answer. Now proceed to the generic code that uses
2033 // ComputeMaskedBits, and pick whichever answer is better.
2038 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2039 if (Tmp == 1) return 1; // Early out.
2040 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
2041 return std::min(Tmp, Tmp2);
2049 if (Op.getResNo() != 1)
2051 // The boolean result conforms to getBooleanContents. Fall through.
2053 // If setcc returns 0/-1, all bits are sign bits.
2054 if (TLI.getBooleanContents() ==
2055 TargetLowering::ZeroOrNegativeOneBooleanContent)
2060 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2061 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
2063 // Handle rotate right by N like a rotate left by 32-N.
2064 if (Op.getOpcode() == ISD::ROTR)
2065 RotAmt = (VTBits-RotAmt) & (VTBits-1);
2067 // If we aren't rotating out all of the known-in sign bits, return the
2068 // number that are left. This handles rotl(sext(x), 1) for example.
2069 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2070 if (Tmp > RotAmt+1) return Tmp-RotAmt;
2074 // Add can have at most one carry bit. Thus we know that the output
2075 // is, at worst, one more bit than the inputs.
2076 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2077 if (Tmp == 1) return 1; // Early out.
2079 // Special case decrementing a value (ADD X, -1):
2080 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
2081 if (CRHS->isAllOnesValue()) {
2082 APInt KnownZero, KnownOne;
2083 APInt Mask = APInt::getAllOnesValue(VTBits);
2084 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2086 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2088 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2091 // If we are subtracting one from a positive number, there is no carry
2092 // out of the result.
2093 if (KnownZero.isNegative())
2097 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2098 if (Tmp2 == 1) return 1;
2099 return std::min(Tmp, Tmp2)-1;
2103 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
2104 if (Tmp2 == 1) return 1;
2107 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
2108 if (CLHS->isNullValue()) {
2109 APInt KnownZero, KnownOne;
2110 APInt Mask = APInt::getAllOnesValue(VTBits);
2111 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2112 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2114 if ((KnownZero | APInt(VTBits, 1)) == Mask)
2117 // If the input is known to be positive (the sign bit is known clear),
2118 // the output of the NEG has the same number of sign bits as the input.
2119 if (KnownZero.isNegative())
2122 // Otherwise, we treat this like a SUB.
2125 // Sub can have at most one carry bit. Thus we know that the output
2126 // is, at worst, one more bit than the inputs.
2127 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
2128 if (Tmp == 1) return 1; // Early out.
2129 return std::min(Tmp, Tmp2)-1;
2132 // FIXME: it's tricky to do anything useful for this, but it is an important
2133 // case for targets like X86.
2137 // Handle LOADX separately here. EXTLOAD case will fallthrough.
2138 if (Op.getOpcode() == ISD::LOAD) {
2139 LoadSDNode *LD = cast<LoadSDNode>(Op);
2140 unsigned ExtType = LD->getExtensionType();
2143 case ISD::SEXTLOAD: // '17' bits known
2144 Tmp = LD->getMemoryVT().getSizeInBits();
2145 return VTBits-Tmp+1;
2146 case ISD::ZEXTLOAD: // '16' bits known
2147 Tmp = LD->getMemoryVT().getSizeInBits();
2152 // Allow the target to implement this method for its nodes.
2153 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
2154 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
2155 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
2156 Op.getOpcode() == ISD::INTRINSIC_VOID) {
2157 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
2158 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
2161 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2162 // use this information.
2163 APInt KnownZero, KnownOne;
2164 APInt Mask = APInt::getAllOnesValue(VTBits);
2165 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
2167 if (KnownZero.isNegative()) { // sign bit is 0
2169 } else if (KnownOne.isNegative()) { // sign bit is 1;
2176 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2177 // the number of identical bits in the top of the input value.
2179 Mask <<= Mask.getBitWidth()-VTBits;
2180 // Return # leading zeros. We use 'min' here in case Val was zero before
2181 // shifting. We don't want to return '64' as for an i32 "0".
2182 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2185 bool SelectionDAG::isKnownNeverNaN(SDValue Op) const {
2186 // If we're told that NaNs won't happen, assume they won't.
2187 if (FiniteOnlyFPMath())
2190 // If the value is a constant, we can obviously see if it is a NaN or not.
2191 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
2192 return !C->getValueAPF().isNaN();
2194 // TODO: Recognize more cases here.
2199 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2200 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2201 if (!GA) return false;
2202 if (GA->getOffset() != 0) return false;
2203 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2204 if (!GV) return false;
2205 MachineModuleInfo *MMI = getMachineModuleInfo();
2206 return MMI && MMI->hasDebugInfo();
2210 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2211 /// element of the result of the vector shuffle.
2212 SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N,
2214 EVT VT = N->getValueType(0);
2215 DebugLoc dl = N->getDebugLoc();
2216 if (N->getMaskElt(i) < 0)
2217 return getUNDEF(VT.getVectorElementType());
2218 unsigned Index = N->getMaskElt(i);
2219 unsigned NumElems = VT.getVectorNumElements();
2220 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2223 if (V.getOpcode() == ISD::BIT_CONVERT) {
2224 V = V.getOperand(0);
2225 EVT VVT = V.getValueType();
2226 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems)
2229 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2230 return (Index == 0) ? V.getOperand(0)
2231 : getUNDEF(VT.getVectorElementType());
2232 if (V.getOpcode() == ISD::BUILD_VECTOR)
2233 return V.getOperand(Index);
2234 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V))
2235 return getShuffleScalarElt(SVN, Index);
2240 /// getNode - Gets or creates the specified node.
2242 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) {
2243 FoldingSetNodeID ID;
2244 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2246 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2247 return SDValue(E, 0);
2248 SDNode *N = NodeAllocator.Allocate<SDNode>();
2249 new (N) SDNode(Opcode, DL, getVTList(VT));
2250 CSEMap.InsertNode(N, IP);
2252 AllNodes.push_back(N);
2256 return SDValue(N, 0);
2259 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
2260 EVT VT, SDValue Operand) {
2261 // Constant fold unary operations with an integer constant operand.
2262 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2263 const APInt &Val = C->getAPIntValue();
2264 unsigned BitWidth = VT.getSizeInBits();
2267 case ISD::SIGN_EXTEND:
2268 return getConstant(APInt(Val).sextOrTrunc(BitWidth), VT);
2269 case ISD::ANY_EXTEND:
2270 case ISD::ZERO_EXTEND:
2272 return getConstant(APInt(Val).zextOrTrunc(BitWidth), VT);
2273 case ISD::UINT_TO_FP:
2274 case ISD::SINT_TO_FP: {
2275 const uint64_t zero[] = {0, 0};
2276 // No compile time operations on this type.
2277 if (VT==MVT::ppcf128)
2279 APFloat apf = APFloat(APInt(BitWidth, 2, zero));
2280 (void)apf.convertFromAPInt(Val,
2281 Opcode==ISD::SINT_TO_FP,
2282 APFloat::rmNearestTiesToEven);
2283 return getConstantFP(apf, VT);
2285 case ISD::BIT_CONVERT:
2286 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2287 return getConstantFP(Val.bitsToFloat(), VT);
2288 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2289 return getConstantFP(Val.bitsToDouble(), VT);
2292 return getConstant(Val.byteSwap(), VT);
2294 return getConstant(Val.countPopulation(), VT);
2296 return getConstant(Val.countLeadingZeros(), VT);
2298 return getConstant(Val.countTrailingZeros(), VT);
2302 // Constant fold unary operations with a floating point constant operand.
2303 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2304 APFloat V = C->getValueAPF(); // make copy
2305 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2309 return getConstantFP(V, VT);
2312 return getConstantFP(V, VT);
2314 case ISD::FP_EXTEND: {
2316 // This can return overflow, underflow, or inexact; we don't care.
2317 // FIXME need to be more flexible about rounding mode.
2318 (void)V.convert(*EVTToAPFloatSemantics(VT),
2319 APFloat::rmNearestTiesToEven, &ignored);
2320 return getConstantFP(V, VT);
2322 case ISD::FP_TO_SINT:
2323 case ISD::FP_TO_UINT: {
2326 assert(integerPartWidth >= 64);
2327 // FIXME need to be more flexible about rounding mode.
2328 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
2329 Opcode==ISD::FP_TO_SINT,
2330 APFloat::rmTowardZero, &ignored);
2331 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2333 APInt api(VT.getSizeInBits(), 2, x);
2334 return getConstant(api, VT);
2336 case ISD::BIT_CONVERT:
2337 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2338 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2339 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2340 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2346 unsigned OpOpcode = Operand.getNode()->getOpcode();
2348 case ISD::TokenFactor:
2349 case ISD::MERGE_VALUES:
2350 case ISD::CONCAT_VECTORS:
2351 return Operand; // Factor, merge or concat of one node? No need.
2352 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
2353 case ISD::FP_EXTEND:
2354 assert(VT.isFloatingPoint() &&
2355 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2356 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2357 if (Operand.getOpcode() == ISD::UNDEF)
2358 return getUNDEF(VT);
2360 case ISD::SIGN_EXTEND:
2361 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2362 "Invalid SIGN_EXTEND!");
2363 if (Operand.getValueType() == VT) return Operand; // noop extension
2364 assert(Operand.getValueType().bitsLT(VT)
2365 && "Invalid sext node, dst < src!");
2366 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2367 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2369 case ISD::ZERO_EXTEND:
2370 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2371 "Invalid ZERO_EXTEND!");
2372 if (Operand.getValueType() == VT) return Operand; // noop extension
2373 assert(Operand.getValueType().bitsLT(VT)
2374 && "Invalid zext node, dst < src!");
2375 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2376 return getNode(ISD::ZERO_EXTEND, DL, VT,
2377 Operand.getNode()->getOperand(0));
2379 case ISD::ANY_EXTEND:
2380 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2381 "Invalid ANY_EXTEND!");
2382 if (Operand.getValueType() == VT) return Operand; // noop extension
2383 assert(Operand.getValueType().bitsLT(VT)
2384 && "Invalid anyext node, dst < src!");
2385 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND)
2386 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2387 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2390 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2391 "Invalid TRUNCATE!");
2392 if (Operand.getValueType() == VT) return Operand; // noop truncate
2393 assert(Operand.getValueType().bitsGT(VT)
2394 && "Invalid truncate node, src < dst!");
2395 if (OpOpcode == ISD::TRUNCATE)
2396 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2397 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2398 OpOpcode == ISD::ANY_EXTEND) {
2399 // If the source is smaller than the dest, we still need an extend.
2400 if (Operand.getNode()->getOperand(0).getValueType().bitsLT(VT))
2401 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
2402 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2403 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0));
2405 return Operand.getNode()->getOperand(0);
2408 case ISD::BIT_CONVERT:
2409 // Basic sanity checking.
2410 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2411 && "Cannot BIT_CONVERT between types of different sizes!");
2412 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2413 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2414 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0));
2415 if (OpOpcode == ISD::UNDEF)
2416 return getUNDEF(VT);
2418 case ISD::SCALAR_TO_VECTOR:
2419 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2420 (VT.getVectorElementType() == Operand.getValueType() ||
2421 (VT.getVectorElementType().isInteger() &&
2422 Operand.getValueType().isInteger() &&
2423 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
2424 "Illegal SCALAR_TO_VECTOR node!");
2425 if (OpOpcode == ISD::UNDEF)
2426 return getUNDEF(VT);
2427 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2428 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2429 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2430 Operand.getConstantOperandVal(1) == 0 &&
2431 Operand.getOperand(0).getValueType() == VT)
2432 return Operand.getOperand(0);
2435 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0
2436 if (UnsafeFPMath && OpOpcode == ISD::FSUB)
2437 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1),
2438 Operand.getNode()->getOperand(0));
2439 if (OpOpcode == ISD::FNEG) // --X -> X
2440 return Operand.getNode()->getOperand(0);
2443 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2444 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0));
2449 SDVTList VTs = getVTList(VT);
2450 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2451 FoldingSetNodeID ID;
2452 SDValue Ops[1] = { Operand };
2453 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2455 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2456 return SDValue(E, 0);
2457 N = NodeAllocator.Allocate<UnarySDNode>();
2458 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2459 CSEMap.InsertNode(N, IP);
2461 N = NodeAllocator.Allocate<UnarySDNode>();
2462 new (N) UnarySDNode(Opcode, DL, VTs, Operand);
2465 AllNodes.push_back(N);
2469 return SDValue(N, 0);
2472 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2474 ConstantSDNode *Cst1,
2475 ConstantSDNode *Cst2) {
2476 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2479 case ISD::ADD: return getConstant(C1 + C2, VT);
2480 case ISD::SUB: return getConstant(C1 - C2, VT);
2481 case ISD::MUL: return getConstant(C1 * C2, VT);
2483 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2486 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2489 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2492 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2494 case ISD::AND: return getConstant(C1 & C2, VT);
2495 case ISD::OR: return getConstant(C1 | C2, VT);
2496 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2497 case ISD::SHL: return getConstant(C1 << C2, VT);
2498 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2499 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2500 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2501 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2508 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2509 SDValue N1, SDValue N2) {
2510 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2511 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2514 case ISD::TokenFactor:
2515 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2516 N2.getValueType() == MVT::Other && "Invalid token factor!");
2517 // Fold trivial token factors.
2518 if (N1.getOpcode() == ISD::EntryToken) return N2;
2519 if (N2.getOpcode() == ISD::EntryToken) return N1;
2520 if (N1 == N2) return N1;
2522 case ISD::CONCAT_VECTORS:
2523 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2524 // one big BUILD_VECTOR.
2525 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2526 N2.getOpcode() == ISD::BUILD_VECTOR) {
2527 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2528 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2529 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2533 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2534 N1.getValueType() == VT && "Binary operator types must match!");
2535 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2536 // worth handling here.
2537 if (N2C && N2C->isNullValue())
2539 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2546 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2547 N1.getValueType() == VT && "Binary operator types must match!");
2548 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2549 // it's worth handling here.
2550 if (N2C && N2C->isNullValue())
2560 assert(VT.isInteger() && "This operator does not apply to FP types!");
2568 if (Opcode == ISD::FADD) {
2570 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1))
2571 if (CFP->getValueAPF().isZero())
2574 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2575 if (CFP->getValueAPF().isZero())
2577 } else if (Opcode == ISD::FSUB) {
2579 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2))
2580 if (CFP->getValueAPF().isZero())
2584 assert(N1.getValueType() == N2.getValueType() &&
2585 N1.getValueType() == VT && "Binary operator types must match!");
2587 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2588 assert(N1.getValueType() == VT &&
2589 N1.getValueType().isFloatingPoint() &&
2590 N2.getValueType().isFloatingPoint() &&
2591 "Invalid FCOPYSIGN!");
2598 assert(VT == N1.getValueType() &&
2599 "Shift operators return type must be the same as their first arg");
2600 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2601 "Shifts only work on integers");
2603 // Always fold shifts of i1 values so the code generator doesn't need to
2604 // handle them. Since we know the size of the shift has to be less than the
2605 // size of the value, the shift/rotate count is guaranteed to be zero.
2609 case ISD::FP_ROUND_INREG: {
2610 EVT EVT = cast<VTSDNode>(N2)->getVT();
2611 assert(VT == N1.getValueType() && "Not an inreg round!");
2612 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2613 "Cannot FP_ROUND_INREG integer types");
2614 assert(EVT.bitsLE(VT) && "Not rounding down!");
2615 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2619 assert(VT.isFloatingPoint() &&
2620 N1.getValueType().isFloatingPoint() &&
2621 VT.bitsLE(N1.getValueType()) &&
2622 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2623 if (N1.getValueType() == VT) return N1; // noop conversion.
2625 case ISD::AssertSext:
2626 case ISD::AssertZext: {
2627 EVT EVT = cast<VTSDNode>(N2)->getVT();
2628 assert(VT == N1.getValueType() && "Not an inreg extend!");
2629 assert(VT.isInteger() && EVT.isInteger() &&
2630 "Cannot *_EXTEND_INREG FP types");
2631 assert(EVT.bitsLE(VT) && "Not extending!");
2632 if (VT == EVT) return N1; // noop assertion.
2635 case ISD::SIGN_EXTEND_INREG: {
2636 EVT EVT = cast<VTSDNode>(N2)->getVT();
2637 assert(VT == N1.getValueType() && "Not an inreg extend!");
2638 assert(VT.isInteger() && EVT.isInteger() &&
2639 "Cannot *_EXTEND_INREG FP types");
2640 assert(EVT.bitsLE(VT) && "Not extending!");
2641 if (EVT == VT) return N1; // Not actually extending
2644 APInt Val = N1C->getAPIntValue();
2645 unsigned FromBits = cast<VTSDNode>(N2)->getVT().getSizeInBits();
2646 Val <<= Val.getBitWidth()-FromBits;
2647 Val = Val.ashr(Val.getBitWidth()-FromBits);
2648 return getConstant(Val, VT);
2652 case ISD::EXTRACT_VECTOR_ELT:
2653 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2654 if (N1.getOpcode() == ISD::UNDEF)
2655 return getUNDEF(VT);
2657 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2658 // expanding copies of large vectors from registers.
2660 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2661 N1.getNumOperands() > 0) {
2663 N1.getOperand(0).getValueType().getVectorNumElements();
2664 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
2665 N1.getOperand(N2C->getZExtValue() / Factor),
2666 getConstant(N2C->getZExtValue() % Factor,
2667 N2.getValueType()));
2670 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2671 // expanding large vector constants.
2672 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) {
2673 SDValue Elt = N1.getOperand(N2C->getZExtValue());
2674 EVT VEltTy = N1.getValueType().getVectorElementType();
2675 if (Elt.getValueType() != VEltTy) {
2676 // If the vector element type is not legal, the BUILD_VECTOR operands
2677 // are promoted and implicitly truncated. Make that explicit here.
2678 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt);
2681 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT
2682 // result is implicitly extended.
2683 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt);
2688 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2689 // operations are lowered to scalars.
2690 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2691 // If the indices are the same, return the inserted element.
2692 if (N1.getOperand(2) == N2)
2693 return N1.getOperand(1);
2694 // If the indices are known different, extract the element from
2695 // the original vector.
2696 else if (isa<ConstantSDNode>(N1.getOperand(2)) &&
2697 isa<ConstantSDNode>(N2))
2698 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
2701 case ISD::EXTRACT_ELEMENT:
2702 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2703 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2704 (N1.getValueType().isInteger() == VT.isInteger()) &&
2705 "Wrong types for EXTRACT_ELEMENT!");
2707 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2708 // 64-bit integers into 32-bit parts. Instead of building the extract of
2709 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2710 if (N1.getOpcode() == ISD::BUILD_PAIR)
2711 return N1.getOperand(N2C->getZExtValue());
2713 // EXTRACT_ELEMENT of a constant int is also very common.
2714 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2715 unsigned ElementSize = VT.getSizeInBits();
2716 unsigned Shift = ElementSize * N2C->getZExtValue();
2717 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2718 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2721 case ISD::EXTRACT_SUBVECTOR:
2722 if (N1.getValueType() == VT) // Trivial extraction.
2729 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2730 if (SV.getNode()) return SV;
2731 } else { // Cannonicalize constant to RHS if commutative
2732 if (isCommutativeBinOp(Opcode)) {
2733 std::swap(N1C, N2C);
2739 // Constant fold FP operations.
2740 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2741 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2743 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2744 // Cannonicalize constant to RHS if commutative
2745 std::swap(N1CFP, N2CFP);
2747 } else if (N2CFP && VT != MVT::ppcf128) {
2748 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2749 APFloat::opStatus s;
2752 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2753 if (s != APFloat::opInvalidOp)
2754 return getConstantFP(V1, VT);
2757 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2758 if (s!=APFloat::opInvalidOp)
2759 return getConstantFP(V1, VT);
2762 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2763 if (s!=APFloat::opInvalidOp)
2764 return getConstantFP(V1, VT);
2767 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2768 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2769 return getConstantFP(V1, VT);
2772 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2773 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2774 return getConstantFP(V1, VT);
2776 case ISD::FCOPYSIGN:
2778 return getConstantFP(V1, VT);
2784 // Canonicalize an UNDEF to the RHS, even over a constant.
2785 if (N1.getOpcode() == ISD::UNDEF) {
2786 if (isCommutativeBinOp(Opcode)) {
2790 case ISD::FP_ROUND_INREG:
2791 case ISD::SIGN_EXTEND_INREG:
2797 return N1; // fold op(undef, arg2) -> undef
2805 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2806 // For vectors, we can't easily build an all zero vector, just return
2813 // Fold a bunch of operators when the RHS is undef.
2814 if (N2.getOpcode() == ISD::UNDEF) {
2817 if (N1.getOpcode() == ISD::UNDEF)
2818 // Handle undef ^ undef -> 0 special case. This is a common
2820 return getConstant(0, VT);
2830 return N2; // fold op(arg1, undef) -> undef
2844 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2845 // For vectors, we can't easily build an all zero vector, just return
2850 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
2851 // For vectors, we can't easily build an all one vector, just return
2859 // Memoize this node if possible.
2861 SDVTList VTs = getVTList(VT);
2862 if (VT != MVT::Flag) {
2863 SDValue Ops[] = { N1, N2 };
2864 FoldingSetNodeID ID;
2865 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2867 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2868 return SDValue(E, 0);
2869 N = NodeAllocator.Allocate<BinarySDNode>();
2870 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2871 CSEMap.InsertNode(N, IP);
2873 N = NodeAllocator.Allocate<BinarySDNode>();
2874 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2);
2877 AllNodes.push_back(N);
2881 return SDValue(N, 0);
2884 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2885 SDValue N1, SDValue N2, SDValue N3) {
2886 // Perform various simplifications.
2887 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2888 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2890 case ISD::CONCAT_VECTORS:
2891 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2892 // one big BUILD_VECTOR.
2893 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2894 N2.getOpcode() == ISD::BUILD_VECTOR &&
2895 N3.getOpcode() == ISD::BUILD_VECTOR) {
2896 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2897 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2898 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
2899 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size());
2903 // Use FoldSetCC to simplify SETCC's.
2904 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL);
2905 if (Simp.getNode()) return Simp;
2910 if (N1C->getZExtValue())
2911 return N2; // select true, X, Y -> X
2913 return N3; // select false, X, Y -> Y
2916 if (N2 == N3) return N2; // select C, X, X -> X
2920 if (N2C->getZExtValue()) // Unconditional branch
2921 return getNode(ISD::BR, DL, MVT::Other, N1, N3);
2923 return N1; // Never-taken branch
2926 case ISD::VECTOR_SHUFFLE:
2927 llvm_unreachable("should use getVectorShuffle constructor!");
2929 case ISD::BIT_CONVERT:
2930 // Fold bit_convert nodes from a type to themselves.
2931 if (N1.getValueType() == VT)
2936 // Memoize node if it doesn't produce a flag.
2938 SDVTList VTs = getVTList(VT);
2939 if (VT != MVT::Flag) {
2940 SDValue Ops[] = { N1, N2, N3 };
2941 FoldingSetNodeID ID;
2942 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
2944 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2945 return SDValue(E, 0);
2946 N = NodeAllocator.Allocate<TernarySDNode>();
2947 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
2948 CSEMap.InsertNode(N, IP);
2950 N = NodeAllocator.Allocate<TernarySDNode>();
2951 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3);
2953 AllNodes.push_back(N);
2957 return SDValue(N, 0);
2960 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2961 SDValue N1, SDValue N2, SDValue N3,
2963 SDValue Ops[] = { N1, N2, N3, N4 };
2964 return getNode(Opcode, DL, VT, Ops, 4);
2967 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
2968 SDValue N1, SDValue N2, SDValue N3,
2969 SDValue N4, SDValue N5) {
2970 SDValue Ops[] = { N1, N2, N3, N4, N5 };
2971 return getNode(Opcode, DL, VT, Ops, 5);
2974 /// getStackArgumentTokenFactor - Compute a TokenFactor to force all
2975 /// the incoming stack arguments to be loaded from the stack.
2976 SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
2977 SmallVector<SDValue, 8> ArgChains;
2979 // Include the original chain at the beginning of the list. When this is
2980 // used by target LowerCall hooks, this helps legalize find the
2981 // CALLSEQ_BEGIN node.
2982 ArgChains.push_back(Chain);
2984 // Add a chain value for each stack argument.
2985 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(),
2986 UE = getEntryNode().getNode()->use_end(); U != UE; ++U)
2987 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2988 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2989 if (FI->getIndex() < 0)
2990 ArgChains.push_back(SDValue(L, 1));
2992 // Build a tokenfactor for all the chains.
2993 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
2994 &ArgChains[0], ArgChains.size());
2997 /// getMemsetValue - Vectorized representation of the memset value
2999 static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
3001 unsigned NumBits = VT.isVector() ?
3002 VT.getVectorElementType().getSizeInBits() : VT.getSizeInBits();
3003 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3004 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
3006 for (unsigned i = NumBits; i > 8; i >>= 1) {
3007 Val = (Val << Shift) | Val;
3011 return DAG.getConstant(Val, VT);
3012 return DAG.getConstantFP(APFloat(Val), VT);
3015 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3016 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
3018 for (unsigned i = NumBits; i > 8; i >>= 1) {
3019 Value = DAG.getNode(ISD::OR, dl, VT,
3020 DAG.getNode(ISD::SHL, dl, VT, Value,
3021 DAG.getConstant(Shift,
3022 TLI.getShiftAmountTy())),
3030 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3031 /// used when a memcpy is turned into a memset when the source is a constant
3033 static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG,
3034 const TargetLowering &TLI,
3035 std::string &Str, unsigned Offset) {
3036 // Handle vector with all elements zero.
3039 return DAG.getConstant(0, VT);
3040 unsigned NumElts = VT.getVectorNumElements();
3041 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3042 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3044 EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts)));
3047 assert(!VT.isVector() && "Can't handle vector type here!");
3048 unsigned NumBits = VT.getSizeInBits();
3049 unsigned MSB = NumBits / 8;
3051 if (TLI.isLittleEndian())
3052 Offset = Offset + MSB - 1;
3053 for (unsigned i = 0; i != MSB; ++i) {
3054 Val = (Val << 8) | (unsigned char)Str[Offset];
3055 Offset += TLI.isLittleEndian() ? -1 : 1;
3057 return DAG.getConstant(Val, VT);
3060 /// getMemBasePlusOffset - Returns base and offset node for the
3062 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
3063 SelectionDAG &DAG) {
3064 EVT VT = Base.getValueType();
3065 return DAG.getNode(ISD::ADD, Base.getDebugLoc(),
3066 VT, Base, DAG.getConstant(Offset, VT));
3069 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
3071 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
3072 unsigned SrcDelta = 0;
3073 GlobalAddressSDNode *G = NULL;
3074 if (Src.getOpcode() == ISD::GlobalAddress)
3075 G = cast<GlobalAddressSDNode>(Src);
3076 else if (Src.getOpcode() == ISD::ADD &&
3077 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3078 Src.getOperand(1).getOpcode() == ISD::Constant) {
3079 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
3080 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
3085 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3086 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
3092 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
3093 /// to replace the memset / memcpy is below the threshold. It also returns the
3094 /// types of the sequence of memory ops to perform memset / memcpy.
3096 bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps,
3097 SDValue Dst, SDValue Src,
3098 unsigned Limit, uint64_t Size, unsigned &Align,
3099 std::string &Str, bool &isSrcStr,
3101 const TargetLowering &TLI) {
3102 isSrcStr = isMemSrcFromString(Src, Str);
3103 bool isSrcConst = isa<ConstantSDNode>(Src);
3104 EVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG);
3105 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses(VT);
3106 if (VT != MVT::iAny) {
3107 const Type *Ty = VT.getTypeForEVT(*DAG.getContext());
3108 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty);
3109 // If source is a string constant, this will require an unaligned load.
3110 if (NewAlign > Align && (isSrcConst || AllowUnalign)) {
3111 if (Dst.getOpcode() != ISD::FrameIndex) {
3112 // Can't change destination alignment. It requires a unaligned store.
3116 int FI = cast<FrameIndexSDNode>(Dst)->getIndex();
3117 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3118 if (MFI->isFixedObjectIndex(FI)) {
3119 // Can't change destination alignment. It requires a unaligned store.
3123 // Give the stack frame object a larger alignment if needed.
3124 if (MFI->getObjectAlignment(FI) < NewAlign)
3125 MFI->setObjectAlignment(FI, NewAlign);
3132 if (VT == MVT::iAny) {
3133 if (TLI.allowsUnalignedMemoryAccesses(MVT::i64)) {
3136 switch (Align & 7) {
3137 case 0: VT = MVT::i64; break;
3138 case 4: VT = MVT::i32; break;
3139 case 2: VT = MVT::i16; break;
3140 default: VT = MVT::i8; break;
3145 while (!TLI.isTypeLegal(LVT))
3146 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1);
3147 assert(LVT.isInteger());
3153 unsigned NumMemOps = 0;
3155 unsigned VTSize = VT.getSizeInBits() / 8;
3156 while (VTSize > Size) {
3157 // For now, only use non-vector load / store's for the left-over pieces.
3158 if (VT.isVector()) {
3160 while (!TLI.isTypeLegal(VT))
3161 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3162 VTSize = VT.getSizeInBits() / 8;
3164 // This can result in a type that is not legal on the target, e.g.
3165 // 1 or 2 bytes on PPC.
3166 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
3171 if (++NumMemOps > Limit)
3173 MemOps.push_back(VT);
3180 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3181 SDValue Chain, SDValue Dst,
3182 SDValue Src, uint64_t Size,
3183 unsigned Align, bool AlwaysInline,
3184 const Value *DstSV, uint64_t DstSVOff,
3185 const Value *SrcSV, uint64_t SrcSVOff){
3186 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3188 // Expand memcpy to a series of load and store ops if the size operand falls
3189 // below a certain threshold.
3190 std::vector<EVT> MemOps;
3191 uint64_t Limit = -1ULL;
3193 Limit = TLI.getMaxStoresPerMemcpy();
3194 unsigned DstAlign = Align; // Destination alignment can change.
3197 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3198 Str, CopyFromStr, DAG, TLI))
3202 bool isZeroStr = CopyFromStr && Str.empty();
3203 SmallVector<SDValue, 8> OutChains;
3204 unsigned NumMemOps = MemOps.size();
3205 uint64_t SrcOff = 0, DstOff = 0;
3206 for (unsigned i = 0; i != NumMemOps; ++i) {
3208 unsigned VTSize = VT.getSizeInBits() / 8;
3209 SDValue Value, Store;
3211 if (CopyFromStr && (isZeroStr || !VT.isVector())) {
3212 // It's unlikely a store of a vector immediate can be done in a single
3213 // instruction. It would require a load from a constantpool first.
3214 // We also handle store a vector with all zero's.
3215 // FIXME: Handle other cases where store of vector immediate is done in
3216 // a single instruction.
3217 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff);
3218 Store = DAG.getStore(Chain, dl, Value,
3219 getMemBasePlusOffset(Dst, DstOff, DAG),
3220 DstSV, DstSVOff + DstOff, false, DstAlign);
3222 // The type might not be legal for the target. This should only happen
3223 // if the type is smaller than a legal type, as on PPC, so the right
3224 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
3225 // to Load/Store if NVT==VT.
3226 // FIXME does the case above also need this?
3227 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
3228 assert(NVT.bitsGE(VT));
3229 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain,
3230 getMemBasePlusOffset(Src, SrcOff, DAG),
3231 SrcSV, SrcSVOff + SrcOff, VT, false, Align);
3232 Store = DAG.getTruncStore(Chain, dl, Value,
3233 getMemBasePlusOffset(Dst, DstOff, DAG),
3234 DstSV, DstSVOff + DstOff, VT, false, DstAlign);
3236 OutChains.push_back(Store);
3241 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3242 &OutChains[0], OutChains.size());
3245 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl,
3246 SDValue Chain, SDValue Dst,
3247 SDValue Src, uint64_t Size,
3248 unsigned Align, bool AlwaysInline,
3249 const Value *DstSV, uint64_t DstSVOff,
3250 const Value *SrcSV, uint64_t SrcSVOff){
3251 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3253 // Expand memmove to a series of load and store ops if the size operand falls
3254 // below a certain threshold.
3255 std::vector<EVT> MemOps;
3256 uint64_t Limit = -1ULL;
3258 Limit = TLI.getMaxStoresPerMemmove();
3259 unsigned DstAlign = Align; // Destination alignment can change.
3262 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3263 Str, CopyFromStr, DAG, TLI))
3266 uint64_t SrcOff = 0, DstOff = 0;
3268 SmallVector<SDValue, 8> LoadValues;
3269 SmallVector<SDValue, 8> LoadChains;
3270 SmallVector<SDValue, 8> OutChains;
3271 unsigned NumMemOps = MemOps.size();
3272 for (unsigned i = 0; i < NumMemOps; i++) {
3274 unsigned VTSize = VT.getSizeInBits() / 8;
3275 SDValue Value, Store;
3277 Value = DAG.getLoad(VT, dl, Chain,
3278 getMemBasePlusOffset(Src, SrcOff, DAG),
3279 SrcSV, SrcSVOff + SrcOff, false, Align);
3280 LoadValues.push_back(Value);
3281 LoadChains.push_back(Value.getValue(1));
3284 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3285 &LoadChains[0], LoadChains.size());
3287 for (unsigned i = 0; i < NumMemOps; i++) {
3289 unsigned VTSize = VT.getSizeInBits() / 8;
3290 SDValue Value, Store;
3292 Store = DAG.getStore(Chain, dl, LoadValues[i],
3293 getMemBasePlusOffset(Dst, DstOff, DAG),
3294 DstSV, DstSVOff + DstOff, false, DstAlign);
3295 OutChains.push_back(Store);
3299 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3300 &OutChains[0], OutChains.size());
3303 static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl,
3304 SDValue Chain, SDValue Dst,
3305 SDValue Src, uint64_t Size,
3307 const Value *DstSV, uint64_t DstSVOff) {
3308 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3310 // Expand memset to a series of load/store ops if the size operand
3311 // falls below a certain threshold.
3312 std::vector<EVT> MemOps;
3315 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(),
3316 Size, Align, Str, CopyFromStr, DAG, TLI))
3319 SmallVector<SDValue, 8> OutChains;
3320 uint64_t DstOff = 0;
3322 unsigned NumMemOps = MemOps.size();
3323 for (unsigned i = 0; i < NumMemOps; i++) {
3325 unsigned VTSize = VT.getSizeInBits() / 8;
3326 SDValue Value = getMemsetValue(Src, VT, DAG, dl);
3327 SDValue Store = DAG.getStore(Chain, dl, Value,
3328 getMemBasePlusOffset(Dst, DstOff, DAG),
3329 DstSV, DstSVOff + DstOff);
3330 OutChains.push_back(Store);
3334 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3335 &OutChains[0], OutChains.size());
3338 SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst,
3339 SDValue Src, SDValue Size,
3340 unsigned Align, bool AlwaysInline,
3341 const Value *DstSV, uint64_t DstSVOff,
3342 const Value *SrcSV, uint64_t SrcSVOff) {
3344 // Check to see if we should lower the memcpy to loads and stores first.
3345 // For cases within the target-specified limits, this is the best choice.
3346 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3348 // Memcpy with size zero? Just return the original chain.
3349 if (ConstantSize->isNullValue())
3353 getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3354 ConstantSize->getZExtValue(),
3355 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3356 if (Result.getNode())
3360 // Then check to see if we should lower the memcpy with target-specific
3361 // code. If the target chooses to do this, this is the next best.
3363 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align,
3365 DstSV, DstSVOff, SrcSV, SrcSVOff);
3366 if (Result.getNode())
3369 // If we really need inline code and the target declined to provide it,
3370 // use a (potentially long) sequence of loads and stores.
3372 assert(ConstantSize && "AlwaysInline requires a constant size!");
3373 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src,
3374 ConstantSize->getZExtValue(), Align, true,
3375 DstSV, DstSVOff, SrcSV, SrcSVOff);
3378 // Emit a library call.
3379 TargetLowering::ArgListTy Args;
3380 TargetLowering::ArgListEntry Entry;
3381 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3382 Entry.Node = Dst; Args.push_back(Entry);
3383 Entry.Node = Src; Args.push_back(Entry);
3384 Entry.Node = Size; Args.push_back(Entry);
3385 // FIXME: pass in DebugLoc
3386 std::pair<SDValue,SDValue> CallResult =
3387 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3388 false, false, false, false, 0,
3389 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false,
3390 /*isReturnValueUsed=*/false,
3391 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY),
3392 TLI.getPointerTy()),
3394 return CallResult.second;
3397 SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst,
3398 SDValue Src, SDValue Size,
3400 const Value *DstSV, uint64_t DstSVOff,
3401 const Value *SrcSV, uint64_t SrcSVOff) {
3403 // Check to see if we should lower the memmove to loads and stores first.
3404 // For cases within the target-specified limits, this is the best choice.
3405 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3407 // Memmove with size zero? Just return the original chain.
3408 if (ConstantSize->isNullValue())
3412 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src,
3413 ConstantSize->getZExtValue(),
3414 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3415 if (Result.getNode())
3419 // Then check to see if we should lower the memmove with target-specific
3420 // code. If the target chooses to do this, this is the next best.
3422 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align,
3423 DstSV, DstSVOff, SrcSV, SrcSVOff);
3424 if (Result.getNode())
3427 // Emit a library call.
3428 TargetLowering::ArgListTy Args;
3429 TargetLowering::ArgListEntry Entry;
3430 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext());
3431 Entry.Node = Dst; Args.push_back(Entry);
3432 Entry.Node = Src; Args.push_back(Entry);
3433 Entry.Node = Size; Args.push_back(Entry);
3434 // FIXME: pass in DebugLoc
3435 std::pair<SDValue,SDValue> CallResult =
3436 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3437 false, false, false, false, 0,
3438 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false,
3439 /*isReturnValueUsed=*/false,
3440 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE),
3441 TLI.getPointerTy()),
3443 return CallResult.second;
3446 SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst,
3447 SDValue Src, SDValue Size,
3449 const Value *DstSV, uint64_t DstSVOff) {
3451 // Check to see if we should lower the memset to stores first.
3452 // For cases within the target-specified limits, this is the best choice.
3453 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3455 // Memset with size zero? Just return the original chain.
3456 if (ConstantSize->isNullValue())
3460 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
3461 Align, DstSV, DstSVOff);
3462 if (Result.getNode())
3466 // Then check to see if we should lower the memset with target-specific
3467 // code. If the target chooses to do this, this is the next best.
3469 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align,
3471 if (Result.getNode())
3474 // Emit a library call.
3475 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext());
3476 TargetLowering::ArgListTy Args;
3477 TargetLowering::ArgListEntry Entry;
3478 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3479 Args.push_back(Entry);
3480 // Extend or truncate the argument to be an i32 value for the call.
3481 if (Src.getValueType().bitsGT(MVT::i32))
3482 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
3484 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
3486 Entry.Ty = Type::getInt32Ty(*getContext());
3487 Entry.isSExt = true;
3488 Args.push_back(Entry);
3490 Entry.Ty = IntPtrTy;
3491 Entry.isSExt = false;
3492 Args.push_back(Entry);
3493 // FIXME: pass in DebugLoc
3494 std::pair<SDValue,SDValue> CallResult =
3495 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()),
3496 false, false, false, false, 0,
3497 TLI.getLibcallCallingConv(RTLIB::MEMSET), false,
3498 /*isReturnValueUsed=*/false,
3499 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET),
3500 TLI.getPointerTy()),
3502 return CallResult.second;
3505 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3507 SDValue Ptr, SDValue Cmp,
3508 SDValue Swp, const Value* PtrVal,
3509 unsigned Alignment) {
3510 MachineFunction &MF = getMachineFunction();
3511 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3513 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3514 Alignment = getEVTAlignment(MemVT);
3516 // Check if the memory reference references a frame index
3518 if (const FrameIndexSDNode *FI =
3519 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) {
3520 if (FrameInfo->isFixedObjectIndex(FI->getIndex()))
3521 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3523 PtrVal = PseudoSourceValue::getStack();
3526 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3528 // For now, atomics are considered to be volatile always.
3529 Flags |= MachineMemOperand::MOVolatile;
3531 MachineMemOperand *MMO =
3532 MF.getMachineMemOperand(PtrVal, Flags, 0,
3533 MemVT.getStoreSize(), Alignment);
3535 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3538 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3540 SDValue Ptr, SDValue Cmp,
3541 SDValue Swp, MachineMemOperand *MMO) {
3542 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op");
3543 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3545 EVT VT = Cmp.getValueType();
3547 SDVTList VTs = getVTList(VT, MVT::Other);
3548 FoldingSetNodeID ID;
3549 ID.AddInteger(MemVT.getRawBits());
3550 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3551 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3553 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3554 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3555 return SDValue(E, 0);
3557 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3558 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Cmp, Swp, MMO);
3559 CSEMap.InsertNode(N, IP);
3560 AllNodes.push_back(N);
3561 return SDValue(N, 0);
3564 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3566 SDValue Ptr, SDValue Val,
3567 const Value* PtrVal,
3568 unsigned Alignment) {
3569 MachineFunction &MF = getMachineFunction();
3570 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3572 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3573 Alignment = getEVTAlignment(MemVT);
3575 // Check if the memory reference references a frame index
3577 if (const FrameIndexSDNode *FI =
3578 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3579 if (FrameInfo->isFixedObjectIndex(FI->getIndex()))
3580 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex());
3582 PtrVal = PseudoSourceValue::getStack();
3584 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
3586 // For now, atomics are considered to be volatile always.
3587 Flags |= MachineMemOperand::MOVolatile;
3589 MachineMemOperand *MMO =
3590 MF.getMachineMemOperand(PtrVal, Flags, 0,
3591 MemVT.getStoreSize(), Alignment);
3593 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO);
3596 SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT,
3598 SDValue Ptr, SDValue Val,
3599 MachineMemOperand *MMO) {
3600 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
3601 Opcode == ISD::ATOMIC_LOAD_SUB ||
3602 Opcode == ISD::ATOMIC_LOAD_AND ||
3603 Opcode == ISD::ATOMIC_LOAD_OR ||
3604 Opcode == ISD::ATOMIC_LOAD_XOR ||
3605 Opcode == ISD::ATOMIC_LOAD_NAND ||
3606 Opcode == ISD::ATOMIC_LOAD_MIN ||
3607 Opcode == ISD::ATOMIC_LOAD_MAX ||
3608 Opcode == ISD::ATOMIC_LOAD_UMIN ||
3609 Opcode == ISD::ATOMIC_LOAD_UMAX ||
3610 Opcode == ISD::ATOMIC_SWAP) &&
3611 "Invalid Atomic Op");
3613 EVT VT = Val.getValueType();
3615 SDVTList VTs = getVTList(VT, MVT::Other);
3616 FoldingSetNodeID ID;
3617 ID.AddInteger(MemVT.getRawBits());
3618 SDValue Ops[] = {Chain, Ptr, Val};
3619 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3621 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3622 cast<AtomicSDNode>(E)->refineAlignment(MMO);
3623 return SDValue(E, 0);
3625 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3626 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Val, MMO);
3627 CSEMap.InsertNode(N, IP);
3628 AllNodes.push_back(N);
3629 return SDValue(N, 0);
3632 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3633 /// Allowed to return something different (and simpler) if Simplify is true.
3634 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3639 SmallVector<EVT, 4> VTs;
3640 VTs.reserve(NumOps);
3641 for (unsigned i = 0; i < NumOps; ++i)
3642 VTs.push_back(Ops[i].getValueType());
3643 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps),
3648 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl,
3649 const EVT *VTs, unsigned NumVTs,
3650 const SDValue *Ops, unsigned NumOps,
3651 EVT MemVT, const Value *srcValue, int SVOff,
3652 unsigned Align, bool Vol,
3653 bool ReadMem, bool WriteMem) {
3654 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps,
3655 MemVT, srcValue, SVOff, Align, Vol,
3660 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3661 const SDValue *Ops, unsigned NumOps,
3662 EVT MemVT, const Value *srcValue, int SVOff,
3663 unsigned Align, bool Vol,
3664 bool ReadMem, bool WriteMem) {
3665 if (Align == 0) // Ensure that codegen never sees alignment 0
3666 Align = getEVTAlignment(MemVT);
3668 MachineFunction &MF = getMachineFunction();
3671 Flags |= MachineMemOperand::MOStore;
3673 Flags |= MachineMemOperand::MOLoad;
3675 Flags |= MachineMemOperand::MOVolatile;
3676 MachineMemOperand *MMO =
3677 MF.getMachineMemOperand(srcValue, Flags, SVOff,
3678 MemVT.getStoreSize(), Align);
3680 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3684 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList,
3685 const SDValue *Ops, unsigned NumOps,
3686 EVT MemVT, MachineMemOperand *MMO) {
3687 assert((Opcode == ISD::INTRINSIC_VOID ||
3688 Opcode == ISD::INTRINSIC_W_CHAIN ||
3689 (Opcode <= INT_MAX &&
3690 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) &&
3691 "Opcode is not a memory-accessing opcode!");
3693 // Memoize the node unless it returns a flag.
3694 MemIntrinsicSDNode *N;
3695 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3696 FoldingSetNodeID ID;
3697 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3699 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3700 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO);
3701 return SDValue(E, 0);
3704 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3705 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3706 CSEMap.InsertNode(N, IP);
3708 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3709 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO);
3711 AllNodes.push_back(N);
3712 return SDValue(N, 0);
3716 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3717 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3718 SDValue Ptr, SDValue Offset,
3719 const Value *SV, int SVOffset, EVT MemVT,
3720 bool isVolatile, unsigned Alignment) {
3721 MachineFunction &MF = getMachineFunction();
3722 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3724 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3725 Alignment = getEVTAlignment(VT);
3727 // Check if the memory reference references a frame index
3729 if (const FrameIndexSDNode *FI =
3730 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3731 if (FrameInfo->isFixedObjectIndex(FI->getIndex()))
3732 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3734 SV = PseudoSourceValue::getStack();
3736 unsigned Flags = MachineMemOperand::MOLoad;
3738 Flags |= MachineMemOperand::MOVolatile;
3739 MachineMemOperand *MMO =
3740 MF.getMachineMemOperand(SV, Flags, SVOffset,
3741 MemVT.getStoreSize(), Alignment);
3742 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO);
3746 SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl,
3747 ISD::LoadExtType ExtType, EVT VT, SDValue Chain,
3748 SDValue Ptr, SDValue Offset, EVT MemVT,
3749 MachineMemOperand *MMO) {
3751 ExtType = ISD::NON_EXTLOAD;
3752 } else if (ExtType == ISD::NON_EXTLOAD) {
3753 assert(VT == MemVT && "Non-extending load from different memory type!");
3757 assert(MemVT.getVectorNumElements() == VT.getVectorNumElements() &&
3758 "Invalid vector extload!");
3760 assert(MemVT.bitsLT(VT) &&
3761 "Should only be an extending load, not truncating!");
3762 assert((ExtType == ISD::EXTLOAD || VT.isInteger()) &&
3763 "Cannot sign/zero extend a FP/Vector load!");
3764 assert(VT.isInteger() == MemVT.isInteger() &&
3765 "Cannot convert from FP to Int or Int -> FP!");
3768 bool Indexed = AM != ISD::UNINDEXED;
3769 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3770 "Unindexed load with an offset!");
3772 SDVTList VTs = Indexed ?
3773 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3774 SDValue Ops[] = { Chain, Ptr, Offset };
3775 FoldingSetNodeID ID;
3776 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3777 ID.AddInteger(MemVT.getRawBits());
3778 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile()));
3780 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3781 cast<LoadSDNode>(E)->refineAlignment(MMO);
3782 return SDValue(E, 0);
3784 SDNode *N = NodeAllocator.Allocate<LoadSDNode>();
3785 new (N) LoadSDNode(Ops, dl, VTs, AM, ExtType, MemVT, MMO);
3786 CSEMap.InsertNode(N, IP);
3787 AllNodes.push_back(N);
3788 return SDValue(N, 0);
3791 SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl,
3792 SDValue Chain, SDValue Ptr,
3793 const Value *SV, int SVOffset,
3794 bool isVolatile, unsigned Alignment) {
3795 SDValue Undef = getUNDEF(Ptr.getValueType());
3796 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3797 SV, SVOffset, VT, isVolatile, Alignment);
3800 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT,
3801 SDValue Chain, SDValue Ptr,
3803 int SVOffset, EVT MemVT,
3804 bool isVolatile, unsigned Alignment) {
3805 SDValue Undef = getUNDEF(Ptr.getValueType());
3806 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef,
3807 SV, SVOffset, MemVT, isVolatile, Alignment);
3811 SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base,
3812 SDValue Offset, ISD::MemIndexedMode AM) {
3813 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
3814 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
3815 "Load is already a indexed load!");
3816 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(),
3817 LD->getChain(), Base, Offset, LD->getSrcValue(),
3818 LD->getSrcValueOffset(), LD->getMemoryVT(),
3819 LD->isVolatile(), LD->getAlignment());
3822 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3823 SDValue Ptr, const Value *SV, int SVOffset,
3824 bool isVolatile, unsigned Alignment) {
3825 MachineFunction &MF = getMachineFunction();
3826 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3828 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3829 Alignment = getEVTAlignment(Val.getValueType());
3831 // Check if the memory reference references a frame index
3833 if (const FrameIndexSDNode *FI =
3834 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3835 if (FrameInfo->isFixedObjectIndex(FI->getIndex()))
3836 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3838 SV = PseudoSourceValue::getStack();
3840 unsigned Flags = MachineMemOperand::MOStore;
3842 Flags |= MachineMemOperand::MOVolatile;
3843 MachineMemOperand *MMO =
3844 MF.getMachineMemOperand(SV, Flags, SVOffset,
3845 Val.getValueType().getStoreSize(), Alignment);
3847 return getStore(Chain, dl, Val, Ptr, MMO);
3850 SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val,
3851 SDValue Ptr, MachineMemOperand *MMO) {
3852 EVT VT = Val.getValueType();
3853 SDVTList VTs = getVTList(MVT::Other);
3854 SDValue Undef = getUNDEF(Ptr.getValueType());
3855 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3856 FoldingSetNodeID ID;
3857 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3858 ID.AddInteger(VT.getRawBits());
3859 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile()));
3861 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3862 cast<StoreSDNode>(E)->refineAlignment(MMO);
3863 return SDValue(E, 0);
3865 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3866 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, false, VT, MMO);
3867 CSEMap.InsertNode(N, IP);
3868 AllNodes.push_back(N);
3869 return SDValue(N, 0);
3872 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3873 SDValue Ptr, const Value *SV,
3874 int SVOffset, EVT SVT,
3875 bool isVolatile, unsigned Alignment) {
3876 MachineFunction &MF = getMachineFunction();
3877 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3879 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3880 Alignment = getEVTAlignment(SVT);
3882 // Check if the memory reference references a frame index
3884 if (const FrameIndexSDNode *FI =
3885 dyn_cast<const FrameIndexSDNode>(Ptr.getNode()))
3886 if (FrameInfo->isFixedObjectIndex(FI->getIndex()))
3887 SV = PseudoSourceValue::getFixedStack(FI->getIndex());
3889 SV = PseudoSourceValue::getStack();
3891 unsigned Flags = MachineMemOperand::MOStore;
3893 Flags |= MachineMemOperand::MOVolatile;
3894 MachineMemOperand *MMO =
3895 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment);
3897 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
3900 SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val,
3901 SDValue Ptr, EVT SVT,
3902 MachineMemOperand *MMO) {
3903 EVT VT = Val.getValueType();
3906 return getStore(Chain, dl, Val, Ptr, MMO);
3908 assert(VT.bitsGT(SVT) && "Not a truncation?");
3909 assert(VT.isInteger() == SVT.isInteger() &&
3910 "Can't do FP-INT conversion!");
3913 SDVTList VTs = getVTList(MVT::Other);
3914 SDValue Undef = getUNDEF(Ptr.getValueType());
3915 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3916 FoldingSetNodeID ID;
3917 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3918 ID.AddInteger(SVT.getRawBits());
3919 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile()));
3921 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3922 cast<StoreSDNode>(E)->refineAlignment(MMO);
3923 return SDValue(E, 0);
3925 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3926 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, true, SVT, MMO);
3927 CSEMap.InsertNode(N, IP);
3928 AllNodes.push_back(N);
3929 return SDValue(N, 0);
3933 SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base,
3934 SDValue Offset, ISD::MemIndexedMode AM) {
3935 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
3936 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
3937 "Store is already a indexed store!");
3938 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
3939 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
3940 FoldingSetNodeID ID;
3941 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3942 ID.AddInteger(ST->getMemoryVT().getRawBits());
3943 ID.AddInteger(ST->getRawSubclassData());
3945 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3946 return SDValue(E, 0);
3947 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3948 new (N) StoreSDNode(Ops, dl, VTs, AM,
3949 ST->isTruncatingStore(), ST->getMemoryVT(),
3950 ST->getMemOperand());
3951 CSEMap.InsertNode(N, IP);
3952 AllNodes.push_back(N);
3953 return SDValue(N, 0);
3956 SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl,
3957 SDValue Chain, SDValue Ptr,
3959 SDValue Ops[] = { Chain, Ptr, SV };
3960 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3);
3963 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3964 const SDUse *Ops, unsigned NumOps) {
3966 case 0: return getNode(Opcode, DL, VT);
3967 case 1: return getNode(Opcode, DL, VT, Ops[0]);
3968 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
3969 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
3973 // Copy from an SDUse array into an SDValue array for use with
3974 // the regular getNode logic.
3975 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
3976 return getNode(Opcode, DL, VT, &NewOps[0], NumOps);
3979 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT,
3980 const SDValue *Ops, unsigned NumOps) {
3982 case 0: return getNode(Opcode, DL, VT);
3983 case 1: return getNode(Opcode, DL, VT, Ops[0]);
3984 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
3985 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
3991 case ISD::SELECT_CC: {
3992 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
3993 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
3994 "LHS and RHS of condition must have same type!");
3995 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
3996 "True and False arms of SelectCC must have same type!");
3997 assert(Ops[2].getValueType() == VT &&
3998 "select_cc node must be of same type as true and false value!");
4002 assert(NumOps == 5 && "BR_CC takes 5 operands!");
4003 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
4004 "LHS/RHS of comparison should match types!");
4011 SDVTList VTs = getVTList(VT);
4013 if (VT != MVT::Flag) {
4014 FoldingSetNodeID ID;
4015 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
4018 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4019 return SDValue(E, 0);
4021 N = NodeAllocator.Allocate<SDNode>();
4022 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
4023 CSEMap.InsertNode(N, IP);
4025 N = NodeAllocator.Allocate<SDNode>();
4026 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps);
4029 AllNodes.push_back(N);
4033 return SDValue(N, 0);
4036 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4037 const std::vector<EVT> &ResultTys,
4038 const SDValue *Ops, unsigned NumOps) {
4039 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()),
4043 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL,
4044 const EVT *VTs, unsigned NumVTs,
4045 const SDValue *Ops, unsigned NumOps) {
4047 return getNode(Opcode, DL, VTs[0], Ops, NumOps);
4048 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps);
4051 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4052 const SDValue *Ops, unsigned NumOps) {
4053 if (VTList.NumVTs == 1)
4054 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps);
4058 // FIXME: figure out how to safely handle things like
4059 // int foo(int x) { return 1 << (x & 255); }
4060 // int bar() { return foo(256); }
4061 case ISD::SRA_PARTS:
4062 case ISD::SRL_PARTS:
4063 case ISD::SHL_PARTS:
4064 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4065 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
4066 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4067 else if (N3.getOpcode() == ISD::AND)
4068 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
4069 // If the and is only masking out bits that cannot effect the shift,
4070 // eliminate the and.
4071 unsigned NumBits = VT.getSizeInBits()*2;
4072 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
4073 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
4079 // Memoize the node unless it returns a flag.
4081 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4082 FoldingSetNodeID ID;
4083 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4085 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4086 return SDValue(E, 0);
4088 N = NodeAllocator.Allocate<UnarySDNode>();
4089 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4090 } else if (NumOps == 2) {
4091 N = NodeAllocator.Allocate<BinarySDNode>();
4092 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4093 } else if (NumOps == 3) {
4094 N = NodeAllocator.Allocate<TernarySDNode>();
4095 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
4097 N = NodeAllocator.Allocate<SDNode>();
4098 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
4100 CSEMap.InsertNode(N, IP);
4103 N = NodeAllocator.Allocate<UnarySDNode>();
4104 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]);
4105 } else if (NumOps == 2) {
4106 N = NodeAllocator.Allocate<BinarySDNode>();
4107 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]);
4108 } else if (NumOps == 3) {
4109 N = NodeAllocator.Allocate<TernarySDNode>();
4110 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]);
4112 N = NodeAllocator.Allocate<SDNode>();
4113 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps);
4116 AllNodes.push_back(N);
4120 return SDValue(N, 0);
4123 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) {
4124 return getNode(Opcode, DL, VTList, 0, 0);
4127 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4129 SDValue Ops[] = { N1 };
4130 return getNode(Opcode, DL, VTList, Ops, 1);
4133 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4134 SDValue N1, SDValue N2) {
4135 SDValue Ops[] = { N1, N2 };
4136 return getNode(Opcode, DL, VTList, Ops, 2);
4139 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4140 SDValue N1, SDValue N2, SDValue N3) {
4141 SDValue Ops[] = { N1, N2, N3 };
4142 return getNode(Opcode, DL, VTList, Ops, 3);
4145 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4146 SDValue N1, SDValue N2, SDValue N3,
4148 SDValue Ops[] = { N1, N2, N3, N4 };
4149 return getNode(Opcode, DL, VTList, Ops, 4);
4152 SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList,
4153 SDValue N1, SDValue N2, SDValue N3,
4154 SDValue N4, SDValue N5) {
4155 SDValue Ops[] = { N1, N2, N3, N4, N5 };
4156 return getNode(Opcode, DL, VTList, Ops, 5);
4159 SDVTList SelectionDAG::getVTList(EVT VT) {
4160 return makeVTList(SDNode::getValueTypeList(VT), 1);
4163 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
4164 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4165 E = VTList.rend(); I != E; ++I)
4166 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
4169 EVT *Array = Allocator.Allocate<EVT>(2);
4172 SDVTList Result = makeVTList(Array, 2);
4173 VTList.push_back(Result);
4177 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
4178 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4179 E = VTList.rend(); I != E; ++I)
4180 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4184 EVT *Array = Allocator.Allocate<EVT>(3);
4188 SDVTList Result = makeVTList(Array, 3);
4189 VTList.push_back(Result);
4193 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
4194 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4195 E = VTList.rend(); I != E; ++I)
4196 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
4197 I->VTs[2] == VT3 && I->VTs[3] == VT4)
4200 EVT *Array = Allocator.Allocate<EVT>(3);
4205 SDVTList Result = makeVTList(Array, 4);
4206 VTList.push_back(Result);
4210 SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) {
4212 case 0: llvm_unreachable("Cannot have nodes without results!");
4213 case 1: return getVTList(VTs[0]);
4214 case 2: return getVTList(VTs[0], VTs[1]);
4215 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
4219 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
4220 E = VTList.rend(); I != E; ++I) {
4221 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
4224 bool NoMatch = false;
4225 for (unsigned i = 2; i != NumVTs; ++i)
4226 if (VTs[i] != I->VTs[i]) {
4234 EVT *Array = Allocator.Allocate<EVT>(NumVTs);
4235 std::copy(VTs, VTs+NumVTs, Array);
4236 SDVTList Result = makeVTList(Array, NumVTs);
4237 VTList.push_back(Result);
4242 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
4243 /// specified operands. If the resultant node already exists in the DAG,
4244 /// this does not modify the specified node, instead it returns the node that
4245 /// already exists. If the resultant node does not exist in the DAG, the
4246 /// input node is returned. As a degenerate case, if you specify the same
4247 /// input operands as the node already has, the input node is returned.
4248 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
4249 SDNode *N = InN.getNode();
4250 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
4252 // Check to see if there is no change.
4253 if (Op == N->getOperand(0)) return InN;
4255 // See if the modified node already exists.
4256 void *InsertPos = 0;
4257 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
4258 return SDValue(Existing, InN.getResNo());
4260 // Nope it doesn't. Remove the node from its current place in the maps.
4262 if (!RemoveNodeFromCSEMaps(N))
4265 // Now we update the operands.
4266 N->OperandList[0].set(Op);
4268 // If this gets put into a CSE map, add it.
4269 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4273 SDValue SelectionDAG::
4274 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
4275 SDNode *N = InN.getNode();
4276 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
4278 // Check to see if there is no change.
4279 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
4280 return InN; // No operands changed, just return the input node.
4282 // See if the modified node already exists.
4283 void *InsertPos = 0;
4284 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
4285 return SDValue(Existing, InN.getResNo());
4287 // Nope it doesn't. Remove the node from its current place in the maps.
4289 if (!RemoveNodeFromCSEMaps(N))
4292 // Now we update the operands.
4293 if (N->OperandList[0] != Op1)
4294 N->OperandList[0].set(Op1);
4295 if (N->OperandList[1] != Op2)
4296 N->OperandList[1].set(Op2);
4298 // If this gets put into a CSE map, add it.
4299 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4303 SDValue SelectionDAG::
4304 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
4305 SDValue Ops[] = { Op1, Op2, Op3 };
4306 return UpdateNodeOperands(N, Ops, 3);
4309 SDValue SelectionDAG::
4310 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4311 SDValue Op3, SDValue Op4) {
4312 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
4313 return UpdateNodeOperands(N, Ops, 4);
4316 SDValue SelectionDAG::
4317 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
4318 SDValue Op3, SDValue Op4, SDValue Op5) {
4319 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
4320 return UpdateNodeOperands(N, Ops, 5);
4323 SDValue SelectionDAG::
4324 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
4325 SDNode *N = InN.getNode();
4326 assert(N->getNumOperands() == NumOps &&
4327 "Update with wrong number of operands");
4329 // Check to see if there is no change.
4330 bool AnyChange = false;
4331 for (unsigned i = 0; i != NumOps; ++i) {
4332 if (Ops[i] != N->getOperand(i)) {
4338 // No operands changed, just return the input node.
4339 if (!AnyChange) return InN;
4341 // See if the modified node already exists.
4342 void *InsertPos = 0;
4343 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
4344 return SDValue(Existing, InN.getResNo());
4346 // Nope it doesn't. Remove the node from its current place in the maps.
4348 if (!RemoveNodeFromCSEMaps(N))
4351 // Now we update the operands.
4352 for (unsigned i = 0; i != NumOps; ++i)
4353 if (N->OperandList[i] != Ops[i])
4354 N->OperandList[i].set(Ops[i]);
4356 // If this gets put into a CSE map, add it.
4357 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
4361 /// DropOperands - Release the operands and set this node to have
4363 void SDNode::DropOperands() {
4364 // Unlike the code in MorphNodeTo that does this, we don't need to
4365 // watch for dead nodes here.
4366 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
4372 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
4375 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4377 SDVTList VTs = getVTList(VT);
4378 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4381 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4382 EVT VT, SDValue Op1) {
4383 SDVTList VTs = getVTList(VT);
4384 SDValue Ops[] = { Op1 };
4385 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4388 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4389 EVT VT, SDValue Op1,
4391 SDVTList VTs = getVTList(VT);
4392 SDValue Ops[] = { Op1, Op2 };
4393 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4396 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4397 EVT VT, SDValue Op1,
4398 SDValue Op2, SDValue Op3) {
4399 SDVTList VTs = getVTList(VT);
4400 SDValue Ops[] = { Op1, Op2, Op3 };
4401 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4404 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4405 EVT VT, const SDValue *Ops,
4407 SDVTList VTs = getVTList(VT);
4408 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4411 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4412 EVT VT1, EVT VT2, const SDValue *Ops,
4414 SDVTList VTs = getVTList(VT1, VT2);
4415 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4418 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4420 SDVTList VTs = getVTList(VT1, VT2);
4421 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4424 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4425 EVT VT1, EVT VT2, EVT VT3,
4426 const SDValue *Ops, unsigned NumOps) {
4427 SDVTList VTs = getVTList(VT1, VT2, VT3);
4428 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4431 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4432 EVT VT1, EVT VT2, EVT VT3, EVT VT4,
4433 const SDValue *Ops, unsigned NumOps) {
4434 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4435 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4438 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4441 SDVTList VTs = getVTList(VT1, VT2);
4442 SDValue Ops[] = { Op1 };
4443 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4446 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4448 SDValue Op1, SDValue Op2) {
4449 SDVTList VTs = getVTList(VT1, VT2);
4450 SDValue Ops[] = { Op1, Op2 };
4451 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4454 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4456 SDValue Op1, SDValue Op2,
4458 SDVTList VTs = getVTList(VT1, VT2);
4459 SDValue Ops[] = { Op1, Op2, Op3 };
4460 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4463 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4464 EVT VT1, EVT VT2, EVT VT3,
4465 SDValue Op1, SDValue Op2,
4467 SDVTList VTs = getVTList(VT1, VT2, VT3);
4468 SDValue Ops[] = { Op1, Op2, Op3 };
4469 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4472 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4473 SDVTList VTs, const SDValue *Ops,
4475 return MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4478 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4480 SDVTList VTs = getVTList(VT);
4481 return MorphNodeTo(N, Opc, VTs, 0, 0);
4484 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4485 EVT VT, SDValue Op1) {
4486 SDVTList VTs = getVTList(VT);
4487 SDValue Ops[] = { Op1 };
4488 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4491 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4492 EVT VT, SDValue Op1,
4494 SDVTList VTs = getVTList(VT);
4495 SDValue Ops[] = { Op1, Op2 };
4496 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4499 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4500 EVT VT, SDValue Op1,
4501 SDValue Op2, SDValue Op3) {
4502 SDVTList VTs = getVTList(VT);
4503 SDValue Ops[] = { Op1, Op2, Op3 };
4504 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4507 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4508 EVT VT, const SDValue *Ops,
4510 SDVTList VTs = getVTList(VT);
4511 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4514 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4515 EVT VT1, EVT VT2, const SDValue *Ops,
4517 SDVTList VTs = getVTList(VT1, VT2);
4518 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4521 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4523 SDVTList VTs = getVTList(VT1, VT2);
4524 return MorphNodeTo(N, Opc, VTs, (SDValue *)0, 0);
4527 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4528 EVT VT1, EVT VT2, EVT VT3,
4529 const SDValue *Ops, unsigned NumOps) {
4530 SDVTList VTs = getVTList(VT1, VT2, VT3);
4531 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4534 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4537 SDVTList VTs = getVTList(VT1, VT2);
4538 SDValue Ops[] = { Op1 };
4539 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4542 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4544 SDValue Op1, SDValue Op2) {
4545 SDVTList VTs = getVTList(VT1, VT2);
4546 SDValue Ops[] = { Op1, Op2 };
4547 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4550 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4552 SDValue Op1, SDValue Op2,
4554 SDVTList VTs = getVTList(VT1, VT2);
4555 SDValue Ops[] = { Op1, Op2, Op3 };
4556 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4559 /// MorphNodeTo - These *mutate* the specified node to have the specified
4560 /// return type, opcode, and operands.
4562 /// Note that MorphNodeTo returns the resultant node. If there is already a
4563 /// node of the specified opcode and operands, it returns that node instead of
4564 /// the current one. Note that the DebugLoc need not be the same.
4566 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4567 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4568 /// node, and because it doesn't require CSE recalculation for any of
4569 /// the node's users.
4571 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4572 SDVTList VTs, const SDValue *Ops,
4574 // If an identical node already exists, use it.
4576 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4577 FoldingSetNodeID ID;
4578 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4579 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4583 if (!RemoveNodeFromCSEMaps(N))
4586 // Start the morphing.
4588 N->ValueList = VTs.VTs;
4589 N->NumValues = VTs.NumVTs;
4591 // Clear the operands list, updating used nodes to remove this from their
4592 // use list. Keep track of any operands that become dead as a result.
4593 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4594 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
4596 SDNode *Used = Use.getNode();
4598 if (Used->use_empty())
4599 DeadNodeSet.insert(Used);
4602 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) {
4603 // Initialize the memory references information.
4604 MN->setMemRefs(0, 0);
4605 // If NumOps is larger than the # of operands we can have in a
4606 // MachineSDNode, reallocate the operand list.
4607 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) {
4608 if (MN->OperandsNeedDelete)
4609 delete[] MN->OperandList;
4610 if (NumOps > array_lengthof(MN->LocalOperands))
4611 // We're creating a final node that will live unmorphed for the
4612 // remainder of the current SelectionDAG iteration, so we can allocate
4613 // the operands directly out of a pool with no recycling metadata.
4614 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4617 MN->InitOperands(MN->LocalOperands, Ops, NumOps);
4618 MN->OperandsNeedDelete = false;
4620 MN->InitOperands(MN->OperandList, Ops, NumOps);
4622 // If NumOps is larger than the # of operands we currently have, reallocate
4623 // the operand list.
4624 if (NumOps > N->NumOperands) {
4625 if (N->OperandsNeedDelete)
4626 delete[] N->OperandList;
4627 N->InitOperands(new SDUse[NumOps], Ops, NumOps);
4628 N->OperandsNeedDelete = true;
4630 MN->InitOperands(MN->OperandList, Ops, NumOps);
4633 // Delete any nodes that are still dead after adding the uses for the
4635 SmallVector<SDNode *, 16> DeadNodes;
4636 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4637 E = DeadNodeSet.end(); I != E; ++I)
4638 if ((*I)->use_empty())
4639 DeadNodes.push_back(*I);
4640 RemoveDeadNodes(DeadNodes);
4643 CSEMap.InsertNode(N, IP); // Memoize the new node.
4648 /// getMachineNode - These are used for target selectors to create a new node
4649 /// with specified return type(s), MachineInstr opcode, and operands.
4651 /// Note that getMachineNode returns the resultant node. If there is already a
4652 /// node of the specified opcode and operands, it returns that node instead of
4653 /// the current one.
4655 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) {
4656 SDVTList VTs = getVTList(VT);
4657 return getMachineNode(Opcode, dl, VTs, 0, 0);
4661 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) {
4662 SDVTList VTs = getVTList(VT);
4663 SDValue Ops[] = { Op1 };
4664 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4668 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4669 SDValue Op1, SDValue Op2) {
4670 SDVTList VTs = getVTList(VT);
4671 SDValue Ops[] = { Op1, Op2 };
4672 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4676 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4677 SDValue Op1, SDValue Op2, SDValue Op3) {
4678 SDVTList VTs = getVTList(VT);
4679 SDValue Ops[] = { Op1, Op2, Op3 };
4680 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4684 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
4685 const SDValue *Ops, unsigned NumOps) {
4686 SDVTList VTs = getVTList(VT);
4687 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4691 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) {
4692 SDVTList VTs = getVTList(VT1, VT2);
4693 return getMachineNode(Opcode, dl, VTs, 0, 0);
4697 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4698 EVT VT1, EVT VT2, SDValue Op1) {
4699 SDVTList VTs = getVTList(VT1, VT2);
4700 SDValue Ops[] = { Op1 };
4701 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4705 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4706 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) {
4707 SDVTList VTs = getVTList(VT1, VT2);
4708 SDValue Ops[] = { Op1, Op2 };
4709 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4713 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4714 EVT VT1, EVT VT2, SDValue Op1,
4715 SDValue Op2, SDValue Op3) {
4716 SDVTList VTs = getVTList(VT1, VT2);
4717 SDValue Ops[] = { Op1, Op2, Op3 };
4718 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4722 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4724 const SDValue *Ops, unsigned NumOps) {
4725 SDVTList VTs = getVTList(VT1, VT2);
4726 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4730 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4731 EVT VT1, EVT VT2, EVT VT3,
4732 SDValue Op1, SDValue Op2) {
4733 SDVTList VTs = getVTList(VT1, VT2, VT3);
4734 SDValue Ops[] = { Op1, Op2 };
4735 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4739 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4740 EVT VT1, EVT VT2, EVT VT3,
4741 SDValue Op1, SDValue Op2, SDValue Op3) {
4742 SDVTList VTs = getVTList(VT1, VT2, VT3);
4743 SDValue Ops[] = { Op1, Op2, Op3 };
4744 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
4748 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4749 EVT VT1, EVT VT2, EVT VT3,
4750 const SDValue *Ops, unsigned NumOps) {
4751 SDVTList VTs = getVTList(VT1, VT2, VT3);
4752 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4756 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1,
4757 EVT VT2, EVT VT3, EVT VT4,
4758 const SDValue *Ops, unsigned NumOps) {
4759 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
4760 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4764 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
4765 const std::vector<EVT> &ResultTys,
4766 const SDValue *Ops, unsigned NumOps) {
4767 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size());
4768 return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
4772 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs,
4773 const SDValue *Ops, unsigned NumOps) {
4774 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag;
4779 FoldingSetNodeID ID;
4780 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps);
4782 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4783 return cast<MachineSDNode>(E);
4786 // Allocate a new MachineSDNode.
4787 N = NodeAllocator.Allocate<MachineSDNode>();
4788 new (N) MachineSDNode(~Opcode, DL, VTs);
4790 // Initialize the operands list.
4791 if (NumOps > array_lengthof(N->LocalOperands))
4792 // We're creating a final node that will live unmorphed for the
4793 // remainder of the current SelectionDAG iteration, so we can allocate
4794 // the operands directly out of a pool with no recycling metadata.
4795 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps),
4798 N->InitOperands(N->LocalOperands, Ops, NumOps);
4799 N->OperandsNeedDelete = false;
4802 CSEMap.InsertNode(N, IP);
4804 AllNodes.push_back(N);
4811 /// getTargetExtractSubreg - A convenience function for creating
4812 /// TargetInstrInfo::EXTRACT_SUBREG nodes.
4814 SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT,
4816 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4817 SDNode *Subreg = getMachineNode(TargetInstrInfo::EXTRACT_SUBREG, DL,
4818 VT, Operand, SRIdxVal);
4819 return SDValue(Subreg, 0);
4822 /// getTargetInsertSubreg - A convenience function for creating
4823 /// TargetInstrInfo::INSERT_SUBREG nodes.
4825 SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT,
4826 SDValue Operand, SDValue Subreg) {
4827 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32);
4828 SDNode *Result = getMachineNode(TargetInstrInfo::INSERT_SUBREG, DL,
4829 VT, Operand, Subreg, SRIdxVal);
4830 return SDValue(Result, 0);
4833 /// getNodeIfExists - Get the specified node if it's already available, or
4834 /// else return NULL.
4835 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4836 const SDValue *Ops, unsigned NumOps) {
4837 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4838 FoldingSetNodeID ID;
4839 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4841 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4847 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4848 /// This can cause recursive merging of nodes in the DAG.
4850 /// This version assumes From has a single result value.
4852 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
4853 DAGUpdateListener *UpdateListener) {
4854 SDNode *From = FromN.getNode();
4855 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
4856 "Cannot replace with this method!");
4857 assert(From != To.getNode() && "Cannot replace uses of with self");
4859 // Iterate over all the existing uses of From. New uses will be added
4860 // to the beginning of the use list, which we avoid visiting.
4861 // This specifically avoids visiting uses of From that arise while the
4862 // replacement is happening, because any such uses would be the result
4863 // of CSE: If an existing node looks like From after one of its operands
4864 // is replaced by To, we don't want to replace of all its users with To
4865 // too. See PR3018 for more info.
4866 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4870 // This node is about to morph, remove its old self from the CSE maps.
4871 RemoveNodeFromCSEMaps(User);
4873 // A user can appear in a use list multiple times, and when this
4874 // happens the uses are usually next to each other in the list.
4875 // To help reduce the number of CSE recomputations, process all
4876 // the uses of this user that we can find this way.
4878 SDUse &Use = UI.getUse();
4881 } while (UI != UE && *UI == User);
4883 // Now that we have modified User, add it back to the CSE maps. If it
4884 // already exists there, recursively merge the results together.
4885 AddModifiedNodeToCSEMaps(User, UpdateListener);
4889 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4890 /// This can cause recursive merging of nodes in the DAG.
4892 /// This version assumes that for each value of From, there is a
4893 /// corresponding value in To in the same position with the same type.
4895 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
4896 DAGUpdateListener *UpdateListener) {
4898 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
4899 assert((!From->hasAnyUseOfValue(i) ||
4900 From->getValueType(i) == To->getValueType(i)) &&
4901 "Cannot use this version of ReplaceAllUsesWith!");
4904 // Handle the trivial case.
4908 // Iterate over just the existing users of From. See the comments in
4909 // the ReplaceAllUsesWith above.
4910 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4914 // This node is about to morph, remove its old self from the CSE maps.
4915 RemoveNodeFromCSEMaps(User);
4917 // A user can appear in a use list multiple times, and when this
4918 // happens the uses are usually next to each other in the list.
4919 // To help reduce the number of CSE recomputations, process all
4920 // the uses of this user that we can find this way.
4922 SDUse &Use = UI.getUse();
4925 } while (UI != UE && *UI == User);
4927 // Now that we have modified User, add it back to the CSE maps. If it
4928 // already exists there, recursively merge the results together.
4929 AddModifiedNodeToCSEMaps(User, UpdateListener);
4933 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4934 /// This can cause recursive merging of nodes in the DAG.
4936 /// This version can replace From with any result values. To must match the
4937 /// number and types of values returned by From.
4938 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
4940 DAGUpdateListener *UpdateListener) {
4941 if (From->getNumValues() == 1) // Handle the simple case efficiently.
4942 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
4944 // Iterate over just the existing users of From. See the comments in
4945 // the ReplaceAllUsesWith above.
4946 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
4950 // This node is about to morph, remove its old self from the CSE maps.
4951 RemoveNodeFromCSEMaps(User);
4953 // A user can appear in a use list multiple times, and when this
4954 // happens the uses are usually next to each other in the list.
4955 // To help reduce the number of CSE recomputations, process all
4956 // the uses of this user that we can find this way.
4958 SDUse &Use = UI.getUse();
4959 const SDValue &ToOp = To[Use.getResNo()];
4962 } while (UI != UE && *UI == User);
4964 // Now that we have modified User, add it back to the CSE maps. If it
4965 // already exists there, recursively merge the results together.
4966 AddModifiedNodeToCSEMaps(User, UpdateListener);
4970 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
4971 /// uses of other values produced by From.getNode() alone. The Deleted
4972 /// vector is handled the same way as for ReplaceAllUsesWith.
4973 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
4974 DAGUpdateListener *UpdateListener){
4975 // Handle the really simple, really trivial case efficiently.
4976 if (From == To) return;
4978 // Handle the simple, trivial, case efficiently.
4979 if (From.getNode()->getNumValues() == 1) {
4980 ReplaceAllUsesWith(From, To, UpdateListener);
4984 // Iterate over just the existing users of From. See the comments in
4985 // the ReplaceAllUsesWith above.
4986 SDNode::use_iterator UI = From.getNode()->use_begin(),
4987 UE = From.getNode()->use_end();
4990 bool UserRemovedFromCSEMaps = false;
4992 // A user can appear in a use list multiple times, and when this
4993 // happens the uses are usually next to each other in the list.
4994 // To help reduce the number of CSE recomputations, process all
4995 // the uses of this user that we can find this way.
4997 SDUse &Use = UI.getUse();
4999 // Skip uses of different values from the same node.
5000 if (Use.getResNo() != From.getResNo()) {
5005 // If this node hasn't been modified yet, it's still in the CSE maps,
5006 // so remove its old self from the CSE maps.
5007 if (!UserRemovedFromCSEMaps) {
5008 RemoveNodeFromCSEMaps(User);
5009 UserRemovedFromCSEMaps = true;
5014 } while (UI != UE && *UI == User);
5016 // We are iterating over all uses of the From node, so if a use
5017 // doesn't use the specific value, no changes are made.
5018 if (!UserRemovedFromCSEMaps)
5021 // Now that we have modified User, add it back to the CSE maps. If it
5022 // already exists there, recursively merge the results together.
5023 AddModifiedNodeToCSEMaps(User, UpdateListener);
5028 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
5029 /// to record information about a use.
5036 /// operator< - Sort Memos by User.
5037 bool operator<(const UseMemo &L, const UseMemo &R) {
5038 return (intptr_t)L.User < (intptr_t)R.User;
5042 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
5043 /// uses of other values produced by From.getNode() alone. The same value
5044 /// may appear in both the From and To list. The Deleted vector is
5045 /// handled the same way as for ReplaceAllUsesWith.
5046 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
5049 DAGUpdateListener *UpdateListener){
5050 // Handle the simple, trivial case efficiently.
5052 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
5054 // Read up all the uses and make records of them. This helps
5055 // processing new uses that are introduced during the
5056 // replacement process.
5057 SmallVector<UseMemo, 4> Uses;
5058 for (unsigned i = 0; i != Num; ++i) {
5059 unsigned FromResNo = From[i].getResNo();
5060 SDNode *FromNode = From[i].getNode();
5061 for (SDNode::use_iterator UI = FromNode->use_begin(),
5062 E = FromNode->use_end(); UI != E; ++UI) {
5063 SDUse &Use = UI.getUse();
5064 if (Use.getResNo() == FromResNo) {
5065 UseMemo Memo = { *UI, i, &Use };
5066 Uses.push_back(Memo);
5071 // Sort the uses, so that all the uses from a given User are together.
5072 std::sort(Uses.begin(), Uses.end());
5074 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
5075 UseIndex != UseIndexEnd; ) {
5076 // We know that this user uses some value of From. If it is the right
5077 // value, update it.
5078 SDNode *User = Uses[UseIndex].User;
5080 // This node is about to morph, remove its old self from the CSE maps.
5081 RemoveNodeFromCSEMaps(User);
5083 // The Uses array is sorted, so all the uses for a given User
5084 // are next to each other in the list.
5085 // To help reduce the number of CSE recomputations, process all
5086 // the uses of this user that we can find this way.
5088 unsigned i = Uses[UseIndex].Index;
5089 SDUse &Use = *Uses[UseIndex].Use;
5093 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
5095 // Now that we have modified User, add it back to the CSE maps. If it
5096 // already exists there, recursively merge the results together.
5097 AddModifiedNodeToCSEMaps(User, UpdateListener);
5101 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
5102 /// based on their topological order. It returns the maximum id and a vector
5103 /// of the SDNodes* in assigned order by reference.
5104 unsigned SelectionDAG::AssignTopologicalOrder() {
5106 unsigned DAGSize = 0;
5108 // SortedPos tracks the progress of the algorithm. Nodes before it are
5109 // sorted, nodes after it are unsorted. When the algorithm completes
5110 // it is at the end of the list.
5111 allnodes_iterator SortedPos = allnodes_begin();
5113 // Visit all the nodes. Move nodes with no operands to the front of
5114 // the list immediately. Annotate nodes that do have operands with their
5115 // operand count. Before we do this, the Node Id fields of the nodes
5116 // may contain arbitrary values. After, the Node Id fields for nodes
5117 // before SortedPos will contain the topological sort index, and the
5118 // Node Id fields for nodes At SortedPos and after will contain the
5119 // count of outstanding operands.
5120 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
5122 unsigned Degree = N->getNumOperands();
5124 // A node with no uses, add it to the result array immediately.
5125 N->setNodeId(DAGSize++);
5126 allnodes_iterator Q = N;
5128 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
5131 // Temporarily use the Node Id as scratch space for the degree count.
5132 N->setNodeId(Degree);
5136 // Visit all the nodes. As we iterate, moves nodes into sorted order,
5137 // such that by the time the end is reached all nodes will be sorted.
5138 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
5140 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5143 unsigned Degree = P->getNodeId();
5146 // All of P's operands are sorted, so P may sorted now.
5147 P->setNodeId(DAGSize++);
5149 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
5152 // Update P's outstanding operand count.
5153 P->setNodeId(Degree);
5158 assert(SortedPos == AllNodes.end() &&
5159 "Topological sort incomplete!");
5160 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
5161 "First node in topological sort is not the entry token!");
5162 assert(AllNodes.front().getNodeId() == 0 &&
5163 "First node in topological sort has non-zero id!");
5164 assert(AllNodes.front().getNumOperands() == 0 &&
5165 "First node in topological sort has operands!");
5166 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
5167 "Last node in topologic sort has unexpected id!");
5168 assert(AllNodes.back().use_empty() &&
5169 "Last node in topologic sort has users!");
5170 assert(DAGSize == allnodes_size() && "Node count mismatch!");
5176 //===----------------------------------------------------------------------===//
5178 //===----------------------------------------------------------------------===//
5180 HandleSDNode::~HandleSDNode() {
5184 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA,
5185 EVT VT, int64_t o, unsigned char TF)
5186 : SDNode(Opc, DebugLoc::getUnknownLoc(), getSDVTList(VT)),
5187 Offset(o), TargetFlags(TF) {
5188 TheGlobal = const_cast<GlobalValue*>(GA);
5191 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt,
5192 MachineMemOperand *mmo)
5193 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) {
5194 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile());
5195 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5196 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5199 MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs,
5200 const SDValue *Ops, unsigned NumOps, EVT memvt,
5201 MachineMemOperand *mmo)
5202 : SDNode(Opc, dl, VTs, Ops, NumOps),
5203 MemoryVT(memvt), MMO(mmo) {
5204 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile());
5205 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!");
5206 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!");
5209 /// Profile - Gather unique data for the node.
5211 void SDNode::Profile(FoldingSetNodeID &ID) const {
5212 AddNodeIDNode(ID, this);
5217 std::vector<EVT> VTs;
5220 VTs.reserve(MVT::LAST_VALUETYPE);
5221 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i)
5222 VTs.push_back(MVT((MVT::SimpleValueType)i));
5227 static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs;
5228 static ManagedStatic<EVTArray> SimpleVTArray;
5229 static ManagedStatic<sys::SmartMutex<true> > VTMutex;
5231 /// getValueTypeList - Return a pointer to the specified value type.
5233 const EVT *SDNode::getValueTypeList(EVT VT) {
5234 if (VT.isExtended()) {
5235 sys::SmartScopedLock<true> Lock(*VTMutex);
5236 return &(*EVTs->insert(VT).first);
5238 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy];
5242 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
5243 /// indicated value. This method ignores uses of other values defined by this
5245 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
5246 assert(Value < getNumValues() && "Bad value!");
5248 // TODO: Only iterate over uses of a given value of the node
5249 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
5250 if (UI.getUse().getResNo() == Value) {
5257 // Found exactly the right number of uses?
5262 /// hasAnyUseOfValue - Return true if there are any use of the indicated
5263 /// value. This method ignores uses of other values defined by this operation.
5264 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
5265 assert(Value < getNumValues() && "Bad value!");
5267 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
5268 if (UI.getUse().getResNo() == Value)
5275 /// isOnlyUserOf - Return true if this node is the only use of N.
5277 bool SDNode::isOnlyUserOf(SDNode *N) const {
5279 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
5290 /// isOperand - Return true if this node is an operand of N.
5292 bool SDValue::isOperandOf(SDNode *N) const {
5293 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5294 if (*this == N->getOperand(i))
5299 bool SDNode::isOperandOf(SDNode *N) const {
5300 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
5301 if (this == N->OperandList[i].getNode())
5306 /// reachesChainWithoutSideEffects - Return true if this operand (which must
5307 /// be a chain) reaches the specified operand without crossing any
5308 /// side-effecting instructions. In practice, this looks through token
5309 /// factors and non-volatile loads. In order to remain efficient, this only
5310 /// looks a couple of nodes in, it does not do an exhaustive search.
5311 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
5312 unsigned Depth) const {
5313 if (*this == Dest) return true;
5315 // Don't search too deeply, we just want to be able to see through
5316 // TokenFactor's etc.
5317 if (Depth == 0) return false;
5319 // If this is a token factor, all inputs to the TF happen in parallel. If any
5320 // of the operands of the TF reach dest, then we can do the xform.
5321 if (getOpcode() == ISD::TokenFactor) {
5322 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5323 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
5328 // Loads don't have side effects, look through them.
5329 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
5330 if (!Ld->isVolatile())
5331 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
5337 static void findPredecessor(SDNode *N, const SDNode *P, bool &found,
5338 SmallPtrSet<SDNode *, 32> &Visited) {
5339 if (found || !Visited.insert(N))
5342 for (unsigned i = 0, e = N->getNumOperands(); !found && i != e; ++i) {
5343 SDNode *Op = N->getOperand(i).getNode();
5348 findPredecessor(Op, P, found, Visited);
5352 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
5353 /// is either an operand of N or it can be reached by recursively traversing
5354 /// up the operands.
5355 /// NOTE: this is an expensive method. Use it carefully.
5356 bool SDNode::isPredecessorOf(SDNode *N) const {
5357 SmallPtrSet<SDNode *, 32> Visited;
5359 findPredecessor(N, this, found, Visited);
5363 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
5364 assert(Num < NumOperands && "Invalid child # of SDNode!");
5365 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
5368 std::string SDNode::getOperationName(const SelectionDAG *G) const {
5369 switch (getOpcode()) {
5371 if (getOpcode() < ISD::BUILTIN_OP_END)
5372 return "<<Unknown DAG Node>>";
5373 if (isMachineOpcode()) {
5375 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
5376 if (getMachineOpcode() < TII->getNumOpcodes())
5377 return TII->get(getMachineOpcode()).getName();
5378 return "<<Unknown Machine Node>>";
5381 const TargetLowering &TLI = G->getTargetLoweringInfo();
5382 const char *Name = TLI.getTargetNodeName(getOpcode());
5383 if (Name) return Name;
5384 return "<<Unknown Target Node>>";
5386 return "<<Unknown Node>>";
5389 case ISD::DELETED_NODE:
5390 return "<<Deleted Node!>>";
5392 case ISD::PREFETCH: return "Prefetch";
5393 case ISD::MEMBARRIER: return "MemBarrier";
5394 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap";
5395 case ISD::ATOMIC_SWAP: return "AtomicSwap";
5396 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd";
5397 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub";
5398 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd";
5399 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr";
5400 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor";
5401 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand";
5402 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin";
5403 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax";
5404 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin";
5405 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax";
5406 case ISD::PCMARKER: return "PCMarker";
5407 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5408 case ISD::SRCVALUE: return "SrcValue";
5409 case ISD::EntryToken: return "EntryToken";
5410 case ISD::TokenFactor: return "TokenFactor";
5411 case ISD::AssertSext: return "AssertSext";
5412 case ISD::AssertZext: return "AssertZext";
5414 case ISD::BasicBlock: return "BasicBlock";
5415 case ISD::VALUETYPE: return "ValueType";
5416 case ISD::Register: return "Register";
5418 case ISD::Constant: return "Constant";
5419 case ISD::ConstantFP: return "ConstantFP";
5420 case ISD::GlobalAddress: return "GlobalAddress";
5421 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5422 case ISD::FrameIndex: return "FrameIndex";
5423 case ISD::JumpTable: return "JumpTable";
5424 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5425 case ISD::RETURNADDR: return "RETURNADDR";
5426 case ISD::FRAMEADDR: return "FRAMEADDR";
5427 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5428 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5429 case ISD::LSDAADDR: return "LSDAADDR";
5430 case ISD::EHSELECTION: return "EHSELECTION";
5431 case ISD::EH_RETURN: return "EH_RETURN";
5432 case ISD::ConstantPool: return "ConstantPool";
5433 case ISD::ExternalSymbol: return "ExternalSymbol";
5434 case ISD::INTRINSIC_WO_CHAIN:
5435 case ISD::INTRINSIC_VOID:
5436 case ISD::INTRINSIC_W_CHAIN: {
5437 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1;
5438 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue();
5439 if (IID < Intrinsic::num_intrinsics)
5440 return Intrinsic::getName((Intrinsic::ID)IID);
5441 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo())
5442 return TII->getName(IID);
5443 llvm_unreachable("Invalid intrinsic ID");
5446 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5447 case ISD::TargetConstant: return "TargetConstant";
5448 case ISD::TargetConstantFP:return "TargetConstantFP";
5449 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5450 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5451 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5452 case ISD::TargetJumpTable: return "TargetJumpTable";
5453 case ISD::TargetConstantPool: return "TargetConstantPool";
5454 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5456 case ISD::CopyToReg: return "CopyToReg";
5457 case ISD::CopyFromReg: return "CopyFromReg";
5458 case ISD::UNDEF: return "undef";
5459 case ISD::MERGE_VALUES: return "merge_values";
5460 case ISD::INLINEASM: return "inlineasm";
5461 case ISD::DBG_LABEL: return "dbg_label";
5462 case ISD::EH_LABEL: return "eh_label";
5463 case ISD::HANDLENODE: return "handlenode";
5466 case ISD::FABS: return "fabs";
5467 case ISD::FNEG: return "fneg";
5468 case ISD::FSQRT: return "fsqrt";
5469 case ISD::FSIN: return "fsin";
5470 case ISD::FCOS: return "fcos";
5471 case ISD::FPOWI: return "fpowi";
5472 case ISD::FPOW: return "fpow";
5473 case ISD::FTRUNC: return "ftrunc";
5474 case ISD::FFLOOR: return "ffloor";
5475 case ISD::FCEIL: return "fceil";
5476 case ISD::FRINT: return "frint";
5477 case ISD::FNEARBYINT: return "fnearbyint";
5480 case ISD::ADD: return "add";
5481 case ISD::SUB: return "sub";
5482 case ISD::MUL: return "mul";
5483 case ISD::MULHU: return "mulhu";
5484 case ISD::MULHS: return "mulhs";
5485 case ISD::SDIV: return "sdiv";
5486 case ISD::UDIV: return "udiv";
5487 case ISD::SREM: return "srem";
5488 case ISD::UREM: return "urem";
5489 case ISD::SMUL_LOHI: return "smul_lohi";
5490 case ISD::UMUL_LOHI: return "umul_lohi";
5491 case ISD::SDIVREM: return "sdivrem";
5492 case ISD::UDIVREM: return "udivrem";
5493 case ISD::AND: return "and";
5494 case ISD::OR: return "or";
5495 case ISD::XOR: return "xor";
5496 case ISD::SHL: return "shl";
5497 case ISD::SRA: return "sra";
5498 case ISD::SRL: return "srl";
5499 case ISD::ROTL: return "rotl";
5500 case ISD::ROTR: return "rotr";
5501 case ISD::FADD: return "fadd";
5502 case ISD::FSUB: return "fsub";
5503 case ISD::FMUL: return "fmul";
5504 case ISD::FDIV: return "fdiv";
5505 case ISD::FREM: return "frem";
5506 case ISD::FCOPYSIGN: return "fcopysign";
5507 case ISD::FGETSIGN: return "fgetsign";
5509 case ISD::SETCC: return "setcc";
5510 case ISD::VSETCC: return "vsetcc";
5511 case ISD::SELECT: return "select";
5512 case ISD::SELECT_CC: return "select_cc";
5513 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5514 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5515 case ISD::CONCAT_VECTORS: return "concat_vectors";
5516 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5517 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5518 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5519 case ISD::CARRY_FALSE: return "carry_false";
5520 case ISD::ADDC: return "addc";
5521 case ISD::ADDE: return "adde";
5522 case ISD::SADDO: return "saddo";
5523 case ISD::UADDO: return "uaddo";
5524 case ISD::SSUBO: return "ssubo";
5525 case ISD::USUBO: return "usubo";
5526 case ISD::SMULO: return "smulo";
5527 case ISD::UMULO: return "umulo";
5528 case ISD::SUBC: return "subc";
5529 case ISD::SUBE: return "sube";
5530 case ISD::SHL_PARTS: return "shl_parts";
5531 case ISD::SRA_PARTS: return "sra_parts";
5532 case ISD::SRL_PARTS: return "srl_parts";
5534 // Conversion operators.
5535 case ISD::SIGN_EXTEND: return "sign_extend";
5536 case ISD::ZERO_EXTEND: return "zero_extend";
5537 case ISD::ANY_EXTEND: return "any_extend";
5538 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5539 case ISD::TRUNCATE: return "truncate";
5540 case ISD::FP_ROUND: return "fp_round";
5541 case ISD::FLT_ROUNDS_: return "flt_rounds";
5542 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5543 case ISD::FP_EXTEND: return "fp_extend";
5545 case ISD::SINT_TO_FP: return "sint_to_fp";
5546 case ISD::UINT_TO_FP: return "uint_to_fp";
5547 case ISD::FP_TO_SINT: return "fp_to_sint";
5548 case ISD::FP_TO_UINT: return "fp_to_uint";
5549 case ISD::BIT_CONVERT: return "bit_convert";
5551 case ISD::CONVERT_RNDSAT: {
5552 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) {
5553 default: llvm_unreachable("Unknown cvt code!");
5554 case ISD::CVT_FF: return "cvt_ff";
5555 case ISD::CVT_FS: return "cvt_fs";
5556 case ISD::CVT_FU: return "cvt_fu";
5557 case ISD::CVT_SF: return "cvt_sf";
5558 case ISD::CVT_UF: return "cvt_uf";
5559 case ISD::CVT_SS: return "cvt_ss";
5560 case ISD::CVT_SU: return "cvt_su";
5561 case ISD::CVT_US: return "cvt_us";
5562 case ISD::CVT_UU: return "cvt_uu";
5566 // Control flow instructions
5567 case ISD::BR: return "br";
5568 case ISD::BRIND: return "brind";
5569 case ISD::BR_JT: return "br_jt";
5570 case ISD::BRCOND: return "brcond";
5571 case ISD::BR_CC: return "br_cc";
5572 case ISD::CALLSEQ_START: return "callseq_start";
5573 case ISD::CALLSEQ_END: return "callseq_end";
5576 case ISD::LOAD: return "load";
5577 case ISD::STORE: return "store";
5578 case ISD::VAARG: return "vaarg";
5579 case ISD::VACOPY: return "vacopy";
5580 case ISD::VAEND: return "vaend";
5581 case ISD::VASTART: return "vastart";
5582 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5583 case ISD::EXTRACT_ELEMENT: return "extract_element";
5584 case ISD::BUILD_PAIR: return "build_pair";
5585 case ISD::STACKSAVE: return "stacksave";
5586 case ISD::STACKRESTORE: return "stackrestore";
5587 case ISD::TRAP: return "trap";
5590 case ISD::BSWAP: return "bswap";
5591 case ISD::CTPOP: return "ctpop";
5592 case ISD::CTTZ: return "cttz";
5593 case ISD::CTLZ: return "ctlz";
5596 case ISD::DBG_STOPPOINT: return "dbg_stoppoint";
5597 case ISD::DEBUG_LOC: return "debug_loc";
5600 case ISD::TRAMPOLINE: return "trampoline";
5603 switch (cast<CondCodeSDNode>(this)->get()) {
5604 default: llvm_unreachable("Unknown setcc condition!");
5605 case ISD::SETOEQ: return "setoeq";
5606 case ISD::SETOGT: return "setogt";
5607 case ISD::SETOGE: return "setoge";
5608 case ISD::SETOLT: return "setolt";
5609 case ISD::SETOLE: return "setole";
5610 case ISD::SETONE: return "setone";
5612 case ISD::SETO: return "seto";
5613 case ISD::SETUO: return "setuo";
5614 case ISD::SETUEQ: return "setue";
5615 case ISD::SETUGT: return "setugt";
5616 case ISD::SETUGE: return "setuge";
5617 case ISD::SETULT: return "setult";
5618 case ISD::SETULE: return "setule";
5619 case ISD::SETUNE: return "setune";
5621 case ISD::SETEQ: return "seteq";
5622 case ISD::SETGT: return "setgt";
5623 case ISD::SETGE: return "setge";
5624 case ISD::SETLT: return "setlt";
5625 case ISD::SETLE: return "setle";
5626 case ISD::SETNE: return "setne";
5631 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5640 return "<post-inc>";
5642 return "<post-dec>";
5646 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5647 std::string S = "< ";
5661 if (getByValAlign())
5662 S += "byval-align:" + utostr(getByValAlign()) + " ";
5664 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5666 S += "byval-size:" + utostr(getByValSize()) + " ";
5670 void SDNode::dump() const { dump(0); }
5671 void SDNode::dump(const SelectionDAG *G) const {
5675 void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const {
5676 OS << (void*)this << ": ";
5678 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5680 if (getValueType(i) == MVT::Other)
5683 OS << getValueType(i).getEVTString();
5685 OS << " = " << getOperationName(G);
5688 void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const {
5689 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) {
5690 if (!MN->memoperands_empty()) {
5693 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(),
5694 e = MN->memoperands_end(); i != e; ++i) {
5701 } else if (const ShuffleVectorSDNode *SVN =
5702 dyn_cast<ShuffleVectorSDNode>(this)) {
5704 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) {
5705 int Idx = SVN->getMaskElt(i);
5713 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5714 OS << '<' << CSDN->getAPIntValue() << '>';
5715 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5716 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5717 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5718 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5719 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5722 CSDN->getValueAPF().bitcastToAPInt().dump();
5725 } else if (const GlobalAddressSDNode *GADN =
5726 dyn_cast<GlobalAddressSDNode>(this)) {
5727 int64_t offset = GADN->getOffset();
5729 WriteAsOperand(OS, GADN->getGlobal());
5732 OS << " + " << offset;
5734 OS << " " << offset;
5735 if (unsigned int TF = GADN->getTargetFlags())
5736 OS << " [TF=" << TF << ']';
5737 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5738 OS << "<" << FIDN->getIndex() << ">";
5739 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5740 OS << "<" << JTDN->getIndex() << ">";
5741 if (unsigned int TF = JTDN->getTargetFlags())
5742 OS << " [TF=" << TF << ']';
5743 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5744 int offset = CP->getOffset();
5745 if (CP->isMachineConstantPoolEntry())
5746 OS << "<" << *CP->getMachineCPVal() << ">";
5748 OS << "<" << *CP->getConstVal() << ">";
5750 OS << " + " << offset;
5752 OS << " " << offset;
5753 if (unsigned int TF = CP->getTargetFlags())
5754 OS << " [TF=" << TF << ']';
5755 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5757 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5759 OS << LBB->getName() << " ";
5760 OS << (const void*)BBDN->getBasicBlock() << ">";
5761 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5762 if (G && R->getReg() &&
5763 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5764 OS << " " << G->getTarget().getRegisterInfo()->getName(R->getReg());
5766 OS << " #" << R->getReg();
5768 } else if (const ExternalSymbolSDNode *ES =
5769 dyn_cast<ExternalSymbolSDNode>(this)) {
5770 OS << "'" << ES->getSymbol() << "'";
5771 if (unsigned int TF = ES->getTargetFlags())
5772 OS << " [TF=" << TF << ']';
5773 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5775 OS << "<" << M->getValue() << ">";
5778 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5779 OS << ":" << N->getVT().getEVTString();
5781 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5782 OS << " <" << *LD->getMemOperand();
5785 switch (LD->getExtensionType()) {
5786 default: doExt = false; break;
5787 case ISD::EXTLOAD: OS << ", anyext"; break;
5788 case ISD::SEXTLOAD: OS << ", sext"; break;
5789 case ISD::ZEXTLOAD: OS << ", zext"; break;
5792 OS << " from " << LD->getMemoryVT().getEVTString();
5794 const char *AM = getIndexedModeName(LD->getAddressingMode());
5799 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
5800 OS << " <" << *ST->getMemOperand();
5802 if (ST->isTruncatingStore())
5803 OS << ", trunc to " << ST->getMemoryVT().getEVTString();
5805 const char *AM = getIndexedModeName(ST->getAddressingMode());
5810 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) {
5811 OS << " <" << *M->getMemOperand() << ">";
5815 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
5818 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
5820 OS << (void*)getOperand(i).getNode();
5821 if (unsigned RN = getOperand(i).getResNo())
5824 print_details(OS, G);
5827 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
5828 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5829 if (N->getOperand(i).getNode()->hasOneUse())
5830 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
5832 errs() << "\n" << std::string(indent+2, ' ')
5833 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
5837 errs().indent(indent);
5841 void SelectionDAG::dump() const {
5842 errs() << "SelectionDAG has " << AllNodes.size() << " nodes:";
5844 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
5846 const SDNode *N = I;
5847 if (!N->hasOneUse() && N != getRoot().getNode())
5848 DumpNodes(N, 2, this);
5851 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
5856 void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
5858 print_details(OS, G);
5861 typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet;
5862 static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent,
5863 const SelectionDAG *G, VisitedSDNodeSet &once) {
5864 if (!once.insert(N)) // If we've been here before, return now.
5866 // Dump the current SDNode, but don't end the line yet.
5867 OS << std::string(indent, ' ');
5869 // Having printed this SDNode, walk the children:
5870 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5871 const SDNode *child = N->getOperand(i).getNode();
5874 if (child->getNumOperands() == 0) {
5875 // This child has no grandchildren; print it inline right here.
5876 child->printr(OS, G);
5878 } else { // Just the address. FIXME: also print the child's opcode
5880 if (unsigned RN = N->getOperand(i).getResNo())
5885 // Dump children that have grandchildren on their own line(s).
5886 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5887 const SDNode *child = N->getOperand(i).getNode();
5888 DumpNodesr(OS, child, indent+2, G, once);
5892 void SDNode::dumpr() const {
5893 VisitedSDNodeSet once;
5894 DumpNodesr(errs(), this, 0, 0, once);
5897 void SDNode::dumpr(const SelectionDAG *G) const {
5898 VisitedSDNodeSet once;
5899 DumpNodesr(errs(), this, 0, G, once);
5903 // getAddressSpace - Return the address space this GlobalAddress belongs to.
5904 unsigned GlobalAddressSDNode::getAddressSpace() const {
5905 return getGlobal()->getType()->getAddressSpace();
5909 const Type *ConstantPoolSDNode::getType() const {
5910 if (isMachineConstantPoolEntry())
5911 return Val.MachineCPVal->getType();
5912 return Val.ConstVal->getType();
5915 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue,
5917 unsigned &SplatBitSize,
5919 unsigned MinSplatBits) {
5920 EVT VT = getValueType(0);
5921 assert(VT.isVector() && "Expected a vector type");
5922 unsigned sz = VT.getSizeInBits();
5923 if (MinSplatBits > sz)
5926 SplatValue = APInt(sz, 0);
5927 SplatUndef = APInt(sz, 0);
5929 // Get the bits. Bits with undefined values (when the corresponding element
5930 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
5931 // in SplatValue. If any of the values are not constant, give up and return
5933 unsigned int nOps = getNumOperands();
5934 assert(nOps > 0 && "isConstantSplat has 0-size build vector");
5935 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits();
5936 for (unsigned i = 0; i < nOps; ++i) {
5937 SDValue OpVal = getOperand(i);
5938 unsigned BitPos = i * EltBitSize;
5940 if (OpVal.getOpcode() == ISD::UNDEF)
5941 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos +EltBitSize);
5942 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal))
5943 SplatValue |= (APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize).
5944 zextOrTrunc(sz) << BitPos);
5945 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal))
5946 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos;
5951 // The build_vector is all constants or undefs. Find the smallest element
5952 // size that splats the vector.
5954 HasAnyUndefs = (SplatUndef != 0);
5957 unsigned HalfSize = sz / 2;
5958 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize);
5959 APInt LowValue = APInt(SplatValue).trunc(HalfSize);
5960 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize);
5961 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize);
5963 // If the two halves do not match (ignoring undef bits), stop here.
5964 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
5965 MinSplatBits > HalfSize)
5968 SplatValue = HighValue | LowValue;
5969 SplatUndef = HighUndef & LowUndef;
5978 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) {
5979 // Find the first non-undef value in the shuffle mask.
5981 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
5984 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!");
5986 // Make sure all remaining elements are either undef or the same as the first
5988 for (int Idx = Mask[i]; i != e; ++i)
5989 if (Mask[i] >= 0 && Mask[i] != Idx)