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/GlobalAlias.h"
17 #include "llvm/GlobalVariable.h"
18 #include "llvm/Intrinsics.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Assembly/Writer.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/CodeGen/MachineBasicBlock.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineModuleInfo.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/Target/TargetRegisterInfo.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetInstrInfo.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/ADT/SmallPtrSet.h"
37 #include "llvm/ADT/SmallSet.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/ADT/StringExtras.h"
44 /// makeVTList - Return an instance of the SDVTList struct initialized with the
45 /// specified members.
46 static SDVTList makeVTList(const MVT *VTs, unsigned NumVTs) {
47 SDVTList Res = {VTs, NumVTs};
51 static const fltSemantics *MVTToAPFloatSemantics(MVT VT) {
52 switch (VT.getSimpleVT()) {
53 default: assert(0 && "Unknown FP format");
54 case MVT::f32: return &APFloat::IEEEsingle;
55 case MVT::f64: return &APFloat::IEEEdouble;
56 case MVT::f80: return &APFloat::x87DoubleExtended;
57 case MVT::f128: return &APFloat::IEEEquad;
58 case MVT::ppcf128: return &APFloat::PPCDoubleDouble;
62 SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {}
64 //===----------------------------------------------------------------------===//
65 // ConstantFPSDNode Class
66 //===----------------------------------------------------------------------===//
68 /// isExactlyValue - We don't rely on operator== working on double values, as
69 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
70 /// As such, this method can be used to do an exact bit-for-bit comparison of
71 /// two floating point values.
72 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
73 return getValueAPF().bitwiseIsEqual(V);
76 bool ConstantFPSDNode::isValueValidForType(MVT VT,
78 assert(VT.isFloatingPoint() && "Can only convert between FP types");
80 // PPC long double cannot be converted to any other type.
81 if (VT == MVT::ppcf128 ||
82 &Val.getSemantics() == &APFloat::PPCDoubleDouble)
85 // convert modifies in place, so make a copy.
86 APFloat Val2 = APFloat(Val);
88 (void) Val2.convert(*MVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
93 //===----------------------------------------------------------------------===//
95 //===----------------------------------------------------------------------===//
97 /// isBuildVectorAllOnes - Return true if the specified node is a
98 /// BUILD_VECTOR where all of the elements are ~0 or undef.
99 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
100 // Look through a bit convert.
101 if (N->getOpcode() == ISD::BIT_CONVERT)
102 N = N->getOperand(0).getNode();
104 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
106 unsigned i = 0, e = N->getNumOperands();
108 // Skip over all of the undef values.
109 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
112 // Do not accept an all-undef vector.
113 if (i == e) return false;
115 // Do not accept build_vectors that aren't all constants or which have non-~0
117 SDValue NotZero = N->getOperand(i);
118 if (isa<ConstantSDNode>(NotZero)) {
119 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue())
121 } else if (isa<ConstantFPSDNode>(NotZero)) {
122 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF().
123 bitcastToAPInt().isAllOnesValue())
128 // Okay, we have at least one ~0 value, check to see if the rest match or are
130 for (++i; i != e; ++i)
131 if (N->getOperand(i) != NotZero &&
132 N->getOperand(i).getOpcode() != ISD::UNDEF)
138 /// isBuildVectorAllZeros - Return true if the specified node is a
139 /// BUILD_VECTOR where all of the elements are 0 or undef.
140 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
141 // Look through a bit convert.
142 if (N->getOpcode() == ISD::BIT_CONVERT)
143 N = N->getOperand(0).getNode();
145 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
147 unsigned i = 0, e = N->getNumOperands();
149 // Skip over all of the undef values.
150 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF)
153 // Do not accept an all-undef vector.
154 if (i == e) return false;
156 // Do not accept build_vectors that aren't all constants or which have non-~0
158 SDValue Zero = N->getOperand(i);
159 if (isa<ConstantSDNode>(Zero)) {
160 if (!cast<ConstantSDNode>(Zero)->isNullValue())
162 } else if (isa<ConstantFPSDNode>(Zero)) {
163 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero())
168 // Okay, we have at least one ~0 value, check to see if the rest match or are
170 for (++i; i != e; ++i)
171 if (N->getOperand(i) != Zero &&
172 N->getOperand(i).getOpcode() != ISD::UNDEF)
177 /// isScalarToVector - Return true if the specified node is a
178 /// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low
179 /// element is not an undef.
180 bool ISD::isScalarToVector(const SDNode *N) {
181 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR)
184 if (N->getOpcode() != ISD::BUILD_VECTOR)
186 if (N->getOperand(0).getOpcode() == ISD::UNDEF)
188 unsigned NumElems = N->getNumOperands();
189 for (unsigned i = 1; i < NumElems; ++i) {
190 SDValue V = N->getOperand(i);
191 if (V.getOpcode() != ISD::UNDEF)
198 /// isDebugLabel - Return true if the specified node represents a debug
199 /// label (i.e. ISD::DBG_LABEL or TargetInstrInfo::DBG_LABEL node).
200 bool ISD::isDebugLabel(const SDNode *N) {
202 if (N->getOpcode() == ISD::DBG_LABEL)
204 if (N->isMachineOpcode() &&
205 N->getMachineOpcode() == TargetInstrInfo::DBG_LABEL)
210 /// getSetCCSwappedOperands - Return the operation corresponding to (Y op X)
211 /// when given the operation for (X op Y).
212 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
213 // To perform this operation, we just need to swap the L and G bits of the
215 unsigned OldL = (Operation >> 2) & 1;
216 unsigned OldG = (Operation >> 1) & 1;
217 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
218 (OldL << 1) | // New G bit
219 (OldG << 2)); // New L bit.
222 /// getSetCCInverse - Return the operation corresponding to !(X op Y), where
223 /// 'op' is a valid SetCC operation.
224 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) {
225 unsigned Operation = Op;
227 Operation ^= 7; // Flip L, G, E bits, but not U.
229 Operation ^= 15; // Flip all of the condition bits.
231 if (Operation > ISD::SETTRUE2)
232 Operation &= ~8; // Don't let N and U bits get set.
234 return ISD::CondCode(Operation);
238 /// isSignedOp - For an integer comparison, return 1 if the comparison is a
239 /// signed operation and 2 if the result is an unsigned comparison. Return zero
240 /// if the operation does not depend on the sign of the input (setne and seteq).
241 static int isSignedOp(ISD::CondCode Opcode) {
243 default: assert(0 && "Illegal integer setcc operation!");
245 case ISD::SETNE: return 0;
249 case ISD::SETGE: return 1;
253 case ISD::SETUGE: return 2;
257 /// getSetCCOrOperation - Return the result of a logical OR between different
258 /// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function
259 /// returns SETCC_INVALID if it is not possible to represent the resultant
261 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
263 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
264 // Cannot fold a signed integer setcc with an unsigned integer setcc.
265 return ISD::SETCC_INVALID;
267 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
269 // If the N and U bits get set then the resultant comparison DOES suddenly
270 // care about orderedness, and is true when ordered.
271 if (Op > ISD::SETTRUE2)
272 Op &= ~16; // Clear the U bit if the N bit is set.
274 // Canonicalize illegal integer setcc's.
275 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
278 return ISD::CondCode(Op);
281 /// getSetCCAndOperation - Return the result of a logical AND between different
282 /// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This
283 /// function returns zero if it is not possible to represent the resultant
285 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
287 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
288 // Cannot fold a signed setcc with an unsigned setcc.
289 return ISD::SETCC_INVALID;
291 // Combine all of the condition bits.
292 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
294 // Canonicalize illegal integer setcc's.
298 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
299 case ISD::SETOEQ: // SETEQ & SETU[LG]E
300 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
301 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
302 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
309 const TargetMachine &SelectionDAG::getTarget() const {
310 return MF->getTarget();
313 //===----------------------------------------------------------------------===//
314 // SDNode Profile Support
315 //===----------------------------------------------------------------------===//
317 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
319 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
323 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
324 /// solely with their pointer.
325 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
326 ID.AddPointer(VTList.VTs);
329 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
331 static void AddNodeIDOperands(FoldingSetNodeID &ID,
332 const SDValue *Ops, unsigned NumOps) {
333 for (; NumOps; --NumOps, ++Ops) {
334 ID.AddPointer(Ops->getNode());
335 ID.AddInteger(Ops->getResNo());
339 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
341 static void AddNodeIDOperands(FoldingSetNodeID &ID,
342 const SDUse *Ops, unsigned NumOps) {
343 for (; NumOps; --NumOps, ++Ops) {
344 ID.AddPointer(Ops->getVal());
345 ID.AddInteger(Ops->getSDValue().getResNo());
349 static void AddNodeIDNode(FoldingSetNodeID &ID,
350 unsigned short OpC, SDVTList VTList,
351 const SDValue *OpList, unsigned N) {
352 AddNodeIDOpcode(ID, OpC);
353 AddNodeIDValueTypes(ID, VTList);
354 AddNodeIDOperands(ID, OpList, N);
357 /// AddNodeIDCustom - If this is an SDNode with special info, add this info to
359 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
360 switch (N->getOpcode()) {
361 default: break; // Normal nodes don't need extra info.
363 ID.AddInteger(cast<ARG_FLAGSSDNode>(N)->getArgFlags().getRawBits());
365 case ISD::TargetConstant:
367 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
369 case ISD::TargetConstantFP:
370 case ISD::ConstantFP: {
371 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
374 case ISD::TargetGlobalAddress:
375 case ISD::GlobalAddress:
376 case ISD::TargetGlobalTLSAddress:
377 case ISD::GlobalTLSAddress: {
378 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
379 ID.AddPointer(GA->getGlobal());
380 ID.AddInteger(GA->getOffset());
383 case ISD::BasicBlock:
384 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
387 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
389 case ISD::DBG_STOPPOINT: {
390 const DbgStopPointSDNode *DSP = cast<DbgStopPointSDNode>(N);
391 ID.AddInteger(DSP->getLine());
392 ID.AddInteger(DSP->getColumn());
393 ID.AddPointer(DSP->getCompileUnit());
397 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
399 case ISD::MEMOPERAND: {
400 const MachineMemOperand &MO = cast<MemOperandSDNode>(N)->MO;
404 case ISD::FrameIndex:
405 case ISD::TargetFrameIndex:
406 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
409 case ISD::TargetJumpTable:
410 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
412 case ISD::ConstantPool:
413 case ISD::TargetConstantPool: {
414 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
415 ID.AddInteger(CP->getAlignment());
416 ID.AddInteger(CP->getOffset());
417 if (CP->isMachineConstantPoolEntry())
418 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID);
420 ID.AddPointer(CP->getConstVal());
424 const CallSDNode *Call = cast<CallSDNode>(N);
425 ID.AddInteger(Call->getCallingConv());
426 ID.AddInteger(Call->isVarArg());
430 const LoadSDNode *LD = cast<LoadSDNode>(N);
431 ID.AddInteger(LD->getAddressingMode());
432 ID.AddInteger(LD->getExtensionType());
433 ID.AddInteger(LD->getMemoryVT().getRawBits());
434 ID.AddInteger(LD->getRawFlags());
438 const StoreSDNode *ST = cast<StoreSDNode>(N);
439 ID.AddInteger(ST->getAddressingMode());
440 ID.AddInteger(ST->isTruncatingStore());
441 ID.AddInteger(ST->getMemoryVT().getRawBits());
442 ID.AddInteger(ST->getRawFlags());
445 case ISD::ATOMIC_CMP_SWAP_8:
446 case ISD::ATOMIC_SWAP_8:
447 case ISD::ATOMIC_LOAD_ADD_8:
448 case ISD::ATOMIC_LOAD_SUB_8:
449 case ISD::ATOMIC_LOAD_AND_8:
450 case ISD::ATOMIC_LOAD_OR_8:
451 case ISD::ATOMIC_LOAD_XOR_8:
452 case ISD::ATOMIC_LOAD_NAND_8:
453 case ISD::ATOMIC_LOAD_MIN_8:
454 case ISD::ATOMIC_LOAD_MAX_8:
455 case ISD::ATOMIC_LOAD_UMIN_8:
456 case ISD::ATOMIC_LOAD_UMAX_8:
457 case ISD::ATOMIC_CMP_SWAP_16:
458 case ISD::ATOMIC_SWAP_16:
459 case ISD::ATOMIC_LOAD_ADD_16:
460 case ISD::ATOMIC_LOAD_SUB_16:
461 case ISD::ATOMIC_LOAD_AND_16:
462 case ISD::ATOMIC_LOAD_OR_16:
463 case ISD::ATOMIC_LOAD_XOR_16:
464 case ISD::ATOMIC_LOAD_NAND_16:
465 case ISD::ATOMIC_LOAD_MIN_16:
466 case ISD::ATOMIC_LOAD_MAX_16:
467 case ISD::ATOMIC_LOAD_UMIN_16:
468 case ISD::ATOMIC_LOAD_UMAX_16:
469 case ISD::ATOMIC_CMP_SWAP_32:
470 case ISD::ATOMIC_SWAP_32:
471 case ISD::ATOMIC_LOAD_ADD_32:
472 case ISD::ATOMIC_LOAD_SUB_32:
473 case ISD::ATOMIC_LOAD_AND_32:
474 case ISD::ATOMIC_LOAD_OR_32:
475 case ISD::ATOMIC_LOAD_XOR_32:
476 case ISD::ATOMIC_LOAD_NAND_32:
477 case ISD::ATOMIC_LOAD_MIN_32:
478 case ISD::ATOMIC_LOAD_MAX_32:
479 case ISD::ATOMIC_LOAD_UMIN_32:
480 case ISD::ATOMIC_LOAD_UMAX_32:
481 case ISD::ATOMIC_CMP_SWAP_64:
482 case ISD::ATOMIC_SWAP_64:
483 case ISD::ATOMIC_LOAD_ADD_64:
484 case ISD::ATOMIC_LOAD_SUB_64:
485 case ISD::ATOMIC_LOAD_AND_64:
486 case ISD::ATOMIC_LOAD_OR_64:
487 case ISD::ATOMIC_LOAD_XOR_64:
488 case ISD::ATOMIC_LOAD_NAND_64:
489 case ISD::ATOMIC_LOAD_MIN_64:
490 case ISD::ATOMIC_LOAD_MAX_64:
491 case ISD::ATOMIC_LOAD_UMIN_64:
492 case ISD::ATOMIC_LOAD_UMAX_64: {
493 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
494 ID.AddInteger(AT->getRawFlags());
497 } // end switch (N->getOpcode())
500 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
502 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
503 AddNodeIDOpcode(ID, N->getOpcode());
504 // Add the return value info.
505 AddNodeIDValueTypes(ID, N->getVTList());
506 // Add the operand info.
507 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands());
509 // Handle SDNode leafs with special info.
510 AddNodeIDCustom(ID, N);
513 /// encodeMemSDNodeFlags - Generic routine for computing a value for use in
514 /// the CSE map that carries both alignment and volatility information.
516 static inline unsigned
517 encodeMemSDNodeFlags(bool isVolatile, unsigned Alignment) {
518 return isVolatile | ((Log2_32(Alignment) + 1) << 1);
521 //===----------------------------------------------------------------------===//
522 // SelectionDAG Class
523 //===----------------------------------------------------------------------===//
525 /// doNotCSE - Return true if CSE should not be performed for this node.
526 static bool doNotCSE(SDNode *N) {
527 if (N->getValueType(0) == MVT::Flag)
528 return true; // Never CSE anything that produces a flag.
530 switch (N->getOpcode()) {
532 case ISD::HANDLENODE:
534 case ISD::DBG_STOPPOINT:
537 return true; // Never CSE these nodes.
540 // Check that remaining values produced are not flags.
541 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
542 if (N->getValueType(i) == MVT::Flag)
543 return true; // Never CSE anything that produces a flag.
548 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
550 void SelectionDAG::RemoveDeadNodes() {
551 // Create a dummy node (which is not added to allnodes), that adds a reference
552 // to the root node, preventing it from being deleted.
553 HandleSDNode Dummy(getRoot());
555 SmallVector<SDNode*, 128> DeadNodes;
557 // Add all obviously-dead nodes to the DeadNodes worklist.
558 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I)
560 DeadNodes.push_back(I);
562 RemoveDeadNodes(DeadNodes);
564 // If the root changed (e.g. it was a dead load, update the root).
565 setRoot(Dummy.getValue());
568 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
569 /// given list, and any nodes that become unreachable as a result.
570 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes,
571 DAGUpdateListener *UpdateListener) {
573 // Process the worklist, deleting the nodes and adding their uses to the
575 while (!DeadNodes.empty()) {
576 SDNode *N = DeadNodes.back();
577 DeadNodes.pop_back();
580 UpdateListener->NodeDeleted(N, 0);
582 // Take the node out of the appropriate CSE map.
583 RemoveNodeFromCSEMaps(N);
585 // Next, brutally remove the operand list. This is safe to do, as there are
586 // no cycles in the graph.
587 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) {
588 SDNode *Operand = I->getVal();
589 Operand->removeUser(std::distance(N->op_begin(), I), N);
591 // Now that we removed this operand, see if there are no uses of it left.
592 if (Operand->use_empty())
593 DeadNodes.push_back(Operand);
596 if (N->OperandsNeedDelete)
597 delete[] N->OperandList;
602 // Finally, remove N itself.
603 NodeAllocator.Deallocate(AllNodes.remove(N));
607 void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){
608 SmallVector<SDNode*, 16> DeadNodes(1, N);
609 RemoveDeadNodes(DeadNodes, UpdateListener);
612 void SelectionDAG::DeleteNode(SDNode *N) {
613 assert(N->use_empty() && "Cannot delete a node that is not dead!");
615 // First take this out of the appropriate CSE map.
616 RemoveNodeFromCSEMaps(N);
618 // Finally, remove uses due to operands of this node, remove from the
619 // AllNodes list, and delete the node.
620 DeleteNodeNotInCSEMaps(N);
623 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
624 // Drop all of the operands and decrement used node's use counts.
625 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
626 I->getVal()->removeUser(std::distance(N->op_begin(), I), N);
628 if (N->OperandsNeedDelete) {
629 delete[] N->OperandList;
633 assert(N != AllNodes.begin());
634 NodeAllocator.Deallocate(AllNodes.remove(N));
637 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
638 /// correspond to it. This is useful when we're about to delete or repurpose
639 /// the node. We don't want future request for structurally identical nodes
640 /// to return N anymore.
641 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
643 switch (N->getOpcode()) {
644 case ISD::EntryToken:
645 assert(0 && "EntryToken should not be in CSEMaps!");
647 case ISD::HANDLENODE: return false; // noop.
649 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
650 "Cond code doesn't exist!");
651 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0;
652 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0;
654 case ISD::ExternalSymbol:
655 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
657 case ISD::TargetExternalSymbol:
659 TargetExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
661 case ISD::VALUETYPE: {
662 MVT VT = cast<VTSDNode>(N)->getVT();
663 if (VT.isExtended()) {
664 Erased = ExtendedValueTypeNodes.erase(VT);
666 Erased = ValueTypeNodes[VT.getSimpleVT()] != 0;
667 ValueTypeNodes[VT.getSimpleVT()] = 0;
672 // Remove it from the CSE Map.
673 Erased = CSEMap.RemoveNode(N);
677 // Verify that the node was actually in one of the CSE maps, unless it has a
678 // flag result (which cannot be CSE'd) or is one of the special cases that are
679 // not subject to CSE.
680 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
681 !N->isMachineOpcode() && !doNotCSE(N)) {
684 assert(0 && "Node is not in map!");
690 /// AddNonLeafNodeToCSEMaps - Add the specified node back to the CSE maps. It
691 /// has been taken out and modified in some way. If the specified node already
692 /// exists in the CSE maps, do not modify the maps, but return the existing node
693 /// instead. If it doesn't exist, add it and return null.
695 SDNode *SelectionDAG::AddNonLeafNodeToCSEMaps(SDNode *N) {
696 assert(N->getNumOperands() && "This is a leaf node!");
701 SDNode *New = CSEMap.GetOrInsertNode(N);
702 if (New != N) return New; // Node already existed.
706 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
707 /// were replaced with those specified. If this node is never memoized,
708 /// return null, otherwise return a pointer to the slot it would take. If a
709 /// node already exists with these operands, the slot will be non-null.
710 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
715 SDValue Ops[] = { Op };
717 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1);
718 AddNodeIDCustom(ID, N);
719 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
722 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
723 /// were replaced with those specified. If this node is never memoized,
724 /// return null, otherwise return a pointer to the slot it would take. If a
725 /// node already exists with these operands, the slot will be non-null.
726 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
727 SDValue Op1, SDValue Op2,
732 SDValue Ops[] = { Op1, Op2 };
734 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2);
735 AddNodeIDCustom(ID, N);
736 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
740 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
741 /// were replaced with those specified. If this node is never memoized,
742 /// return null, otherwise return a pointer to the slot it would take. If a
743 /// node already exists with these operands, the slot will be non-null.
744 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
745 const SDValue *Ops,unsigned NumOps,
751 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps);
752 AddNodeIDCustom(ID, N);
753 return CSEMap.FindNodeOrInsertPos(ID, InsertPos);
756 /// VerifyNode - Sanity check the given node. Aborts if it is invalid.
757 void SelectionDAG::VerifyNode(SDNode *N) {
758 switch (N->getOpcode()) {
761 case ISD::BUILD_PAIR: {
762 MVT VT = N->getValueType(0);
763 assert(N->getNumValues() == 1 && "Too many results!");
764 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
765 "Wrong return type!");
766 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
767 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
768 "Mismatched operand types!");
769 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
770 "Wrong operand type!");
771 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
772 "Wrong return type size");
775 case ISD::BUILD_VECTOR: {
776 assert(N->getNumValues() == 1 && "Too many results!");
777 assert(N->getValueType(0).isVector() && "Wrong return type!");
778 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
779 "Wrong number of operands!");
780 // FIXME: Change vector_shuffle to a variadic node with mask elements being
781 // operands of the node. Currently the mask is a BUILD_VECTOR passed as an
782 // operand, and it is not always possible to legalize it. Turning off the
783 // following checks at least makes it possible to legalize most of the time.
784 // MVT EltVT = N->getValueType(0).getVectorElementType();
785 // for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I)
786 // assert(I->getSDValue().getValueType() == EltVT &&
787 // "Wrong operand type!");
793 /// getMVTAlignment - Compute the default alignment value for the
796 unsigned SelectionDAG::getMVTAlignment(MVT VT) const {
797 const Type *Ty = VT == MVT::iPTR ?
798 PointerType::get(Type::Int8Ty, 0) :
801 return TLI.getTargetData()->getABITypeAlignment(Ty);
804 SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli)
805 : TLI(tli), FLI(fli),
806 EntryNode(ISD::EntryToken, getVTList(MVT::Other)),
807 Root(getEntryNode()) {
808 AllNodes.push_back(&EntryNode);
811 void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi) {
816 SelectionDAG::~SelectionDAG() {
820 void SelectionDAG::allnodes_clear() {
821 assert(&*AllNodes.begin() == &EntryNode);
822 AllNodes.remove(AllNodes.begin());
823 while (!AllNodes.empty()) {
824 SDNode *N = AllNodes.remove(AllNodes.begin());
825 N->SetNextInBucket(0);
827 if (N->OperandsNeedDelete) {
828 delete [] N->OperandList;
832 NodeAllocator.Deallocate(N);
836 void SelectionDAG::clear() {
838 OperandAllocator.Reset();
841 ExtendedValueTypeNodes.clear();
842 ExternalSymbols.clear();
843 TargetExternalSymbols.clear();
844 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
845 static_cast<CondCodeSDNode*>(0));
846 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
847 static_cast<SDNode*>(0));
850 AllNodes.push_back(&EntryNode);
851 Root = getEntryNode();
854 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, MVT VT) {
855 if (Op.getValueType() == VT) return Op;
856 APInt Imm = APInt::getLowBitsSet(Op.getValueSizeInBits(),
858 return getNode(ISD::AND, Op.getValueType(), Op,
859 getConstant(Imm, Op.getValueType()));
862 SDValue SelectionDAG::getConstant(uint64_t Val, MVT VT, bool isT) {
863 MVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
864 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
867 SDValue SelectionDAG::getConstant(const APInt &Val, MVT VT, bool isT) {
868 return getConstant(*ConstantInt::get(Val), VT, isT);
871 SDValue SelectionDAG::getConstant(const ConstantInt &Val, MVT VT, bool isT) {
872 assert(VT.isInteger() && "Cannot create FP integer constant!");
874 MVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
875 assert(Val.getBitWidth() == EltVT.getSizeInBits() &&
876 "APInt size does not match type size!");
878 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
880 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
884 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
886 return SDValue(N, 0);
888 N = NodeAllocator.Allocate<ConstantSDNode>();
889 new (N) ConstantSDNode(isT, &Val, EltVT);
890 CSEMap.InsertNode(N, IP);
891 AllNodes.push_back(N);
894 SDValue Result(N, 0);
896 SmallVector<SDValue, 8> Ops;
897 Ops.assign(VT.getVectorNumElements(), Result);
898 Result = getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
903 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) {
904 return getConstant(Val, TLI.getPointerTy(), isTarget);
908 SDValue SelectionDAG::getConstantFP(const APFloat& V, MVT VT, bool isTarget) {
909 return getConstantFP(*ConstantFP::get(V), VT, isTarget);
912 SDValue SelectionDAG::getConstantFP(const ConstantFP& V, MVT VT, bool isTarget){
913 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
916 VT.isVector() ? VT.getVectorElementType() : VT;
918 // Do the map lookup using the actual bit pattern for the floating point
919 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
920 // we don't have issues with SNANs.
921 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
923 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
927 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
929 return SDValue(N, 0);
931 N = NodeAllocator.Allocate<ConstantFPSDNode>();
932 new (N) ConstantFPSDNode(isTarget, &V, EltVT);
933 CSEMap.InsertNode(N, IP);
934 AllNodes.push_back(N);
937 SDValue Result(N, 0);
939 SmallVector<SDValue, 8> Ops;
940 Ops.assign(VT.getVectorNumElements(), Result);
941 Result = getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
946 SDValue SelectionDAG::getConstantFP(double Val, MVT VT, bool isTarget) {
948 VT.isVector() ? VT.getVectorElementType() : VT;
950 return getConstantFP(APFloat((float)Val), VT, isTarget);
952 return getConstantFP(APFloat(Val), VT, isTarget);
955 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV,
956 MVT VT, int64_t Offset,
960 // Truncate (with sign-extension) the offset value to the pointer size.
961 unsigned BitWidth = TLI.getPointerTy().getSizeInBits();
963 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth));
965 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
967 // If GV is an alias then use the aliasee for determining thread-localness.
968 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
969 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false));
972 if (GVar && GVar->isThreadLocal())
973 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
975 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
978 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
980 ID.AddInteger(Offset);
982 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
983 return SDValue(E, 0);
984 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>();
985 new (N) GlobalAddressSDNode(isTargetGA, GV, VT, Offset);
986 CSEMap.InsertNode(N, IP);
987 AllNodes.push_back(N);
988 return SDValue(N, 0);
991 SDValue SelectionDAG::getFrameIndex(int FI, MVT VT, bool isTarget) {
992 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
994 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
997 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
998 return SDValue(E, 0);
999 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>();
1000 new (N) FrameIndexSDNode(FI, VT, isTarget);
1001 CSEMap.InsertNode(N, IP);
1002 AllNodes.push_back(N);
1003 return SDValue(N, 0);
1006 SDValue SelectionDAG::getJumpTable(int JTI, MVT VT, bool isTarget){
1007 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1008 FoldingSetNodeID ID;
1009 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1012 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1013 return SDValue(E, 0);
1014 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>();
1015 new (N) JumpTableSDNode(JTI, VT, isTarget);
1016 CSEMap.InsertNode(N, IP);
1017 AllNodes.push_back(N);
1018 return SDValue(N, 0);
1021 SDValue SelectionDAG::getConstantPool(Constant *C, MVT VT,
1022 unsigned Alignment, int Offset,
1026 TLI.getTargetData()->getPreferredTypeAlignmentShift(C->getType());
1027 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1028 FoldingSetNodeID ID;
1029 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1030 ID.AddInteger(Alignment);
1031 ID.AddInteger(Offset);
1034 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1035 return SDValue(E, 0);
1036 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1037 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment);
1038 CSEMap.InsertNode(N, IP);
1039 AllNodes.push_back(N);
1040 return SDValue(N, 0);
1044 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, MVT VT,
1045 unsigned Alignment, int Offset,
1049 TLI.getTargetData()->getPreferredTypeAlignmentShift(C->getType());
1050 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1051 FoldingSetNodeID ID;
1052 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0);
1053 ID.AddInteger(Alignment);
1054 ID.AddInteger(Offset);
1055 C->AddSelectionDAGCSEId(ID);
1057 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1058 return SDValue(E, 0);
1059 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>();
1060 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment);
1061 CSEMap.InsertNode(N, IP);
1062 AllNodes.push_back(N);
1063 return SDValue(N, 0);
1067 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1068 FoldingSetNodeID ID;
1069 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0);
1072 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1073 return SDValue(E, 0);
1074 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>();
1075 new (N) BasicBlockSDNode(MBB);
1076 CSEMap.InsertNode(N, IP);
1077 AllNodes.push_back(N);
1078 return SDValue(N, 0);
1081 SDValue SelectionDAG::getArgFlags(ISD::ArgFlagsTy Flags) {
1082 FoldingSetNodeID ID;
1083 AddNodeIDNode(ID, ISD::ARG_FLAGS, getVTList(MVT::Other), 0, 0);
1084 ID.AddInteger(Flags.getRawBits());
1086 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1087 return SDValue(E, 0);
1088 SDNode *N = NodeAllocator.Allocate<ARG_FLAGSSDNode>();
1089 new (N) ARG_FLAGSSDNode(Flags);
1090 CSEMap.InsertNode(N, IP);
1091 AllNodes.push_back(N);
1092 return SDValue(N, 0);
1095 SDValue SelectionDAG::getValueType(MVT VT) {
1096 if (VT.isSimple() && (unsigned)VT.getSimpleVT() >= ValueTypeNodes.size())
1097 ValueTypeNodes.resize(VT.getSimpleVT()+1);
1099 SDNode *&N = VT.isExtended() ?
1100 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT()];
1102 if (N) return SDValue(N, 0);
1103 N = NodeAllocator.Allocate<VTSDNode>();
1104 new (N) VTSDNode(VT);
1105 AllNodes.push_back(N);
1106 return SDValue(N, 0);
1109 SDValue SelectionDAG::getExternalSymbol(const char *Sym, MVT VT) {
1110 SDNode *&N = ExternalSymbols[Sym];
1111 if (N) return SDValue(N, 0);
1112 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1113 new (N) ExternalSymbolSDNode(false, Sym, VT);
1114 AllNodes.push_back(N);
1115 return SDValue(N, 0);
1118 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, MVT VT) {
1119 SDNode *&N = TargetExternalSymbols[Sym];
1120 if (N) return SDValue(N, 0);
1121 N = NodeAllocator.Allocate<ExternalSymbolSDNode>();
1122 new (N) ExternalSymbolSDNode(true, Sym, VT);
1123 AllNodes.push_back(N);
1124 return SDValue(N, 0);
1127 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1128 if ((unsigned)Cond >= CondCodeNodes.size())
1129 CondCodeNodes.resize(Cond+1);
1131 if (CondCodeNodes[Cond] == 0) {
1132 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>();
1133 new (N) CondCodeSDNode(Cond);
1134 CondCodeNodes[Cond] = N;
1135 AllNodes.push_back(N);
1137 return SDValue(CondCodeNodes[Cond], 0);
1140 SDValue SelectionDAG::getRegister(unsigned RegNo, MVT VT) {
1141 FoldingSetNodeID ID;
1142 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0);
1143 ID.AddInteger(RegNo);
1145 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1146 return SDValue(E, 0);
1147 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>();
1148 new (N) RegisterSDNode(RegNo, VT);
1149 CSEMap.InsertNode(N, IP);
1150 AllNodes.push_back(N);
1151 return SDValue(N, 0);
1154 SDValue SelectionDAG::getDbgStopPoint(SDValue Root,
1155 unsigned Line, unsigned Col,
1156 const CompileUnitDesc *CU) {
1157 SDNode *N = NodeAllocator.Allocate<DbgStopPointSDNode>();
1158 new (N) DbgStopPointSDNode(Root, Line, Col, CU);
1159 AllNodes.push_back(N);
1160 return SDValue(N, 0);
1163 SDValue SelectionDAG::getLabel(unsigned Opcode,
1166 FoldingSetNodeID ID;
1167 SDValue Ops[] = { Root };
1168 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), &Ops[0], 1);
1169 ID.AddInteger(LabelID);
1171 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1172 return SDValue(E, 0);
1173 SDNode *N = NodeAllocator.Allocate<LabelSDNode>();
1174 new (N) LabelSDNode(Opcode, Root, LabelID);
1175 CSEMap.InsertNode(N, IP);
1176 AllNodes.push_back(N);
1177 return SDValue(N, 0);
1180 SDValue SelectionDAG::getSrcValue(const Value *V) {
1181 assert((!V || isa<PointerType>(V->getType())) &&
1182 "SrcValue is not a pointer?");
1184 FoldingSetNodeID ID;
1185 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0);
1189 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1190 return SDValue(E, 0);
1192 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>();
1193 new (N) SrcValueSDNode(V);
1194 CSEMap.InsertNode(N, IP);
1195 AllNodes.push_back(N);
1196 return SDValue(N, 0);
1199 SDValue SelectionDAG::getMemOperand(const MachineMemOperand &MO) {
1200 const Value *v = MO.getValue();
1201 assert((!v || isa<PointerType>(v->getType())) &&
1202 "SrcValue is not a pointer?");
1204 FoldingSetNodeID ID;
1205 AddNodeIDNode(ID, ISD::MEMOPERAND, getVTList(MVT::Other), 0, 0);
1209 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
1210 return SDValue(E, 0);
1212 SDNode *N = NodeAllocator.Allocate<MemOperandSDNode>();
1213 new (N) MemOperandSDNode(MO);
1214 CSEMap.InsertNode(N, IP);
1215 AllNodes.push_back(N);
1216 return SDValue(N, 0);
1219 /// CreateStackTemporary - Create a stack temporary, suitable for holding the
1220 /// specified value type.
1221 SDValue SelectionDAG::CreateStackTemporary(MVT VT, unsigned minAlign) {
1222 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo();
1223 unsigned ByteSize = VT.getSizeInBits()/8;
1224 const Type *Ty = VT.getTypeForMVT();
1225 unsigned StackAlign =
1226 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign);
1228 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign);
1229 return getFrameIndex(FrameIdx, TLI.getPointerTy());
1232 SDValue SelectionDAG::FoldSetCC(MVT VT, SDValue N1,
1233 SDValue N2, ISD::CondCode Cond) {
1234 // These setcc operations always fold.
1238 case ISD::SETFALSE2: return getConstant(0, VT);
1240 case ISD::SETTRUE2: return getConstant(1, VT);
1252 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!");
1256 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) {
1257 const APInt &C2 = N2C->getAPIntValue();
1258 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1259 const APInt &C1 = N1C->getAPIntValue();
1262 default: assert(0 && "Unknown integer setcc!");
1263 case ISD::SETEQ: return getConstant(C1 == C2, VT);
1264 case ISD::SETNE: return getConstant(C1 != C2, VT);
1265 case ISD::SETULT: return getConstant(C1.ult(C2), VT);
1266 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT);
1267 case ISD::SETULE: return getConstant(C1.ule(C2), VT);
1268 case ISD::SETUGE: return getConstant(C1.uge(C2), VT);
1269 case ISD::SETLT: return getConstant(C1.slt(C2), VT);
1270 case ISD::SETGT: return getConstant(C1.sgt(C2), VT);
1271 case ISD::SETLE: return getConstant(C1.sle(C2), VT);
1272 case ISD::SETGE: return getConstant(C1.sge(C2), VT);
1276 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1277 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
1278 // No compile time operations on this type yet.
1279 if (N1C->getValueType(0) == MVT::ppcf128)
1282 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
1285 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
1286 return getNode(ISD::UNDEF, VT);
1288 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT);
1289 case ISD::SETNE: if (R==APFloat::cmpUnordered)
1290 return getNode(ISD::UNDEF, VT);
1292 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan ||
1293 R==APFloat::cmpLessThan, VT);
1294 case ISD::SETLT: if (R==APFloat::cmpUnordered)
1295 return getNode(ISD::UNDEF, VT);
1297 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT);
1298 case ISD::SETGT: if (R==APFloat::cmpUnordered)
1299 return getNode(ISD::UNDEF, VT);
1301 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT);
1302 case ISD::SETLE: if (R==APFloat::cmpUnordered)
1303 return getNode(ISD::UNDEF, VT);
1305 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan ||
1306 R==APFloat::cmpEqual, VT);
1307 case ISD::SETGE: if (R==APFloat::cmpUnordered)
1308 return getNode(ISD::UNDEF, VT);
1310 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan ||
1311 R==APFloat::cmpEqual, VT);
1312 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT);
1313 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT);
1314 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered ||
1315 R==APFloat::cmpEqual, VT);
1316 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT);
1317 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered ||
1318 R==APFloat::cmpLessThan, VT);
1319 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan ||
1320 R==APFloat::cmpUnordered, VT);
1321 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT);
1322 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT);
1325 // Ensure that the constant occurs on the RHS.
1326 return getSetCC(VT, N2, N1, ISD::getSetCCSwappedOperands(Cond));
1330 // Could not fold it.
1334 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
1335 /// use this predicate to simplify operations downstream.
1336 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
1337 unsigned BitWidth = Op.getValueSizeInBits();
1338 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth);
1341 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1342 /// this predicate to simplify operations downstream. Mask is known to be zero
1343 /// for bits that V cannot have.
1344 bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask,
1345 unsigned Depth) const {
1346 APInt KnownZero, KnownOne;
1347 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1348 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1349 return (KnownZero & Mask) == Mask;
1352 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
1353 /// known to be either zero or one and return them in the KnownZero/KnownOne
1354 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
1356 void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask,
1357 APInt &KnownZero, APInt &KnownOne,
1358 unsigned Depth) const {
1359 unsigned BitWidth = Mask.getBitWidth();
1360 assert(BitWidth == Op.getValueType().getSizeInBits() &&
1361 "Mask size mismatches value type size!");
1363 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
1364 if (Depth == 6 || Mask == 0)
1365 return; // Limit search depth.
1367 APInt KnownZero2, KnownOne2;
1369 switch (Op.getOpcode()) {
1371 // We know all of the bits for a constant!
1372 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask;
1373 KnownZero = ~KnownOne & Mask;
1376 // If either the LHS or the RHS are Zero, the result is zero.
1377 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1378 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero,
1379 KnownZero2, KnownOne2, Depth+1);
1380 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1381 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1383 // Output known-1 bits are only known if set in both the LHS & RHS.
1384 KnownOne &= KnownOne2;
1385 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1386 KnownZero |= KnownZero2;
1389 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1390 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne,
1391 KnownZero2, KnownOne2, Depth+1);
1392 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1393 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1395 // Output known-0 bits are only known if clear in both the LHS & RHS.
1396 KnownZero &= KnownZero2;
1397 // Output known-1 are known to be set if set in either the LHS | RHS.
1398 KnownOne |= KnownOne2;
1401 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1402 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
1403 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1404 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1406 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1407 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1408 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1409 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1410 KnownZero = KnownZeroOut;
1414 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1415 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
1416 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1417 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1418 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1420 // If low bits are zero in either operand, output low known-0 bits.
1421 // Also compute a conserative estimate for high known-0 bits.
1422 // More trickiness is possible, but this is sufficient for the
1423 // interesting case of alignment computation.
1425 unsigned TrailZ = KnownZero.countTrailingOnes() +
1426 KnownZero2.countTrailingOnes();
1427 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
1428 KnownZero2.countLeadingOnes(),
1429 BitWidth) - BitWidth;
1431 TrailZ = std::min(TrailZ, BitWidth);
1432 LeadZ = std::min(LeadZ, BitWidth);
1433 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
1434 APInt::getHighBitsSet(BitWidth, LeadZ);
1439 // For the purposes of computing leading zeros we can conservatively
1440 // treat a udiv as a logical right shift by the power of 2 known to
1441 // be less than the denominator.
1442 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1443 ComputeMaskedBits(Op.getOperand(0),
1444 AllOnes, KnownZero2, KnownOne2, Depth+1);
1445 unsigned LeadZ = KnownZero2.countLeadingOnes();
1449 ComputeMaskedBits(Op.getOperand(1),
1450 AllOnes, KnownZero2, KnownOne2, Depth+1);
1451 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1452 if (RHSUnknownLeadingOnes != BitWidth)
1453 LeadZ = std::min(BitWidth,
1454 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1456 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
1460 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
1461 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
1462 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1463 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1465 // Only known if known in both the LHS and RHS.
1466 KnownOne &= KnownOne2;
1467 KnownZero &= KnownZero2;
1469 case ISD::SELECT_CC:
1470 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
1471 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
1472 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1473 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1475 // Only known if known in both the LHS and RHS.
1476 KnownOne &= KnownOne2;
1477 KnownZero &= KnownZero2;
1480 // If we know the result of a setcc has the top bits zero, use this info.
1481 if (TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult &&
1483 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1486 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1487 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1488 unsigned ShAmt = SA->getZExtValue();
1490 // If the shift count is an invalid immediate, don't do anything.
1491 if (ShAmt >= BitWidth)
1494 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt),
1495 KnownZero, KnownOne, Depth+1);
1496 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1497 KnownZero <<= ShAmt;
1499 // low bits known zero.
1500 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt);
1504 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1505 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1506 unsigned ShAmt = SA->getZExtValue();
1508 // If the shift count is an invalid immediate, don't do anything.
1509 if (ShAmt >= BitWidth)
1512 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt),
1513 KnownZero, KnownOne, Depth+1);
1514 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1515 KnownZero = KnownZero.lshr(ShAmt);
1516 KnownOne = KnownOne.lshr(ShAmt);
1518 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1519 KnownZero |= HighBits; // High bits known zero.
1523 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1524 unsigned ShAmt = SA->getZExtValue();
1526 // If the shift count is an invalid immediate, don't do anything.
1527 if (ShAmt >= BitWidth)
1530 APInt InDemandedMask = (Mask << ShAmt);
1531 // If any of the demanded bits are produced by the sign extension, we also
1532 // demand the input sign bit.
1533 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask;
1534 if (HighBits.getBoolValue())
1535 InDemandedMask |= APInt::getSignBit(BitWidth);
1537 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne,
1539 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1540 KnownZero = KnownZero.lshr(ShAmt);
1541 KnownOne = KnownOne.lshr(ShAmt);
1543 // Handle the sign bits.
1544 APInt SignBit = APInt::getSignBit(BitWidth);
1545 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask.
1547 if (KnownZero.intersects(SignBit)) {
1548 KnownZero |= HighBits; // New bits are known zero.
1549 } else if (KnownOne.intersects(SignBit)) {
1550 KnownOne |= HighBits; // New bits are known one.
1554 case ISD::SIGN_EXTEND_INREG: {
1555 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1556 unsigned EBits = EVT.getSizeInBits();
1558 // Sign extension. Compute the demanded bits in the result that are not
1559 // present in the input.
1560 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask;
1562 APInt InSignBit = APInt::getSignBit(EBits);
1563 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits);
1565 // If the sign extended bits are demanded, we know that the sign
1567 InSignBit.zext(BitWidth);
1568 if (NewBits.getBoolValue())
1569 InputDemandedBits |= InSignBit;
1571 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
1572 KnownZero, KnownOne, Depth+1);
1573 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1575 // If the sign bit of the input is known set or clear, then we know the
1576 // top bits of the result.
1577 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear
1578 KnownZero |= NewBits;
1579 KnownOne &= ~NewBits;
1580 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1581 KnownOne |= NewBits;
1582 KnownZero &= ~NewBits;
1583 } else { // Input sign bit unknown
1584 KnownZero &= ~NewBits;
1585 KnownOne &= ~NewBits;
1592 unsigned LowBits = Log2_32(BitWidth)+1;
1593 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1598 if (ISD::isZEXTLoad(Op.getNode())) {
1599 LoadSDNode *LD = cast<LoadSDNode>(Op);
1600 MVT VT = LD->getMemoryVT();
1601 unsigned MemBits = VT.getSizeInBits();
1602 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask;
1606 case ISD::ZERO_EXTEND: {
1607 MVT InVT = Op.getOperand(0).getValueType();
1608 unsigned InBits = InVT.getSizeInBits();
1609 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1610 APInt InMask = Mask;
1611 InMask.trunc(InBits);
1612 KnownZero.trunc(InBits);
1613 KnownOne.trunc(InBits);
1614 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1615 KnownZero.zext(BitWidth);
1616 KnownOne.zext(BitWidth);
1617 KnownZero |= NewBits;
1620 case ISD::SIGN_EXTEND: {
1621 MVT InVT = Op.getOperand(0).getValueType();
1622 unsigned InBits = InVT.getSizeInBits();
1623 APInt InSignBit = APInt::getSignBit(InBits);
1624 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask;
1625 APInt InMask = Mask;
1626 InMask.trunc(InBits);
1628 // If any of the sign extended bits are demanded, we know that the sign
1629 // bit is demanded. Temporarily set this bit in the mask for our callee.
1630 if (NewBits.getBoolValue())
1631 InMask |= InSignBit;
1633 KnownZero.trunc(InBits);
1634 KnownOne.trunc(InBits);
1635 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1637 // Note if the sign bit is known to be zero or one.
1638 bool SignBitKnownZero = KnownZero.isNegative();
1639 bool SignBitKnownOne = KnownOne.isNegative();
1640 assert(!(SignBitKnownZero && SignBitKnownOne) &&
1641 "Sign bit can't be known to be both zero and one!");
1643 // If the sign bit wasn't actually demanded by our caller, we don't
1644 // want it set in the KnownZero and KnownOne result values. Reset the
1645 // mask and reapply it to the result values.
1647 InMask.trunc(InBits);
1648 KnownZero &= InMask;
1651 KnownZero.zext(BitWidth);
1652 KnownOne.zext(BitWidth);
1654 // If the sign bit is known zero or one, the top bits match.
1655 if (SignBitKnownZero)
1656 KnownZero |= NewBits;
1657 else if (SignBitKnownOne)
1658 KnownOne |= NewBits;
1661 case ISD::ANY_EXTEND: {
1662 MVT InVT = Op.getOperand(0).getValueType();
1663 unsigned InBits = InVT.getSizeInBits();
1664 APInt InMask = Mask;
1665 InMask.trunc(InBits);
1666 KnownZero.trunc(InBits);
1667 KnownOne.trunc(InBits);
1668 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1669 KnownZero.zext(BitWidth);
1670 KnownOne.zext(BitWidth);
1673 case ISD::TRUNCATE: {
1674 MVT InVT = Op.getOperand(0).getValueType();
1675 unsigned InBits = InVT.getSizeInBits();
1676 APInt InMask = Mask;
1677 InMask.zext(InBits);
1678 KnownZero.zext(InBits);
1679 KnownOne.zext(InBits);
1680 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1);
1681 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1682 KnownZero.trunc(BitWidth);
1683 KnownOne.trunc(BitWidth);
1686 case ISD::AssertZext: {
1687 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1688 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
1689 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
1691 KnownZero |= (~InMask) & Mask;
1695 // All bits are zero except the low bit.
1696 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1);
1700 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
1701 // We know that the top bits of C-X are clear if X contains less bits
1702 // than C (i.e. no wrap-around can happen). For example, 20-X is
1703 // positive if we can prove that X is >= 0 and < 16.
1704 if (CLHS->getAPIntValue().isNonNegative()) {
1705 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros();
1706 // NLZ can't be BitWidth with no sign bit
1707 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
1708 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2,
1711 // If all of the MaskV bits are known to be zero, then we know the
1712 // output top bits are zero, because we now know that the output is
1714 if ((KnownZero2 & MaskV) == MaskV) {
1715 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros();
1716 // Top bits known zero.
1717 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
1724 // Output known-0 bits are known if clear or set in both the low clear bits
1725 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
1726 // low 3 bits clear.
1727 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
1728 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
1729 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1730 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
1732 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
1733 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1734 KnownZeroOut = std::min(KnownZeroOut,
1735 KnownZero2.countTrailingOnes());
1737 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
1741 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1742 const APInt &RA = Rem->getAPIntValue();
1743 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1744 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1745 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1746 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
1748 // If the sign bit of the first operand is zero, the sign bit of
1749 // the result is zero. If the first operand has no one bits below
1750 // the second operand's single 1 bit, its sign will be zero.
1751 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1752 KnownZero2 |= ~LowBits;
1754 KnownZero |= KnownZero2 & Mask;
1756 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1761 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1762 const APInt &RA = Rem->getAPIntValue();
1763 if (RA.isPowerOf2()) {
1764 APInt LowBits = (RA - 1);
1765 APInt Mask2 = LowBits & Mask;
1766 KnownZero |= ~LowBits & Mask;
1767 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1768 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1773 // Since the result is less than or equal to either operand, any leading
1774 // zero bits in either operand must also exist in the result.
1775 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1776 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne,
1778 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2,
1781 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1782 KnownZero2.countLeadingOnes());
1784 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1788 // Allow the target to implement this method for its nodes.
1789 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
1790 case ISD::INTRINSIC_WO_CHAIN:
1791 case ISD::INTRINSIC_W_CHAIN:
1792 case ISD::INTRINSIC_VOID:
1793 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this);
1799 /// ComputeNumSignBits - Return the number of times the sign bit of the
1800 /// register is replicated into the other bits. We know that at least 1 bit
1801 /// is always equal to the sign bit (itself), but other cases can give us
1802 /// information. For example, immediately after an "SRA X, 2", we know that
1803 /// the top 3 bits are all equal to each other, so we return 3.
1804 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{
1805 MVT VT = Op.getValueType();
1806 assert(VT.isInteger() && "Invalid VT!");
1807 unsigned VTBits = VT.getSizeInBits();
1809 unsigned FirstAnswer = 1;
1812 return 1; // Limit search depth.
1814 switch (Op.getOpcode()) {
1816 case ISD::AssertSext:
1817 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
1818 return VTBits-Tmp+1;
1819 case ISD::AssertZext:
1820 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
1823 case ISD::Constant: {
1824 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue();
1825 // If negative, return # leading ones.
1826 if (Val.isNegative())
1827 return Val.countLeadingOnes();
1829 // Return # leading zeros.
1830 return Val.countLeadingZeros();
1833 case ISD::SIGN_EXTEND:
1834 Tmp = VTBits-Op.getOperand(0).getValueType().getSizeInBits();
1835 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
1837 case ISD::SIGN_EXTEND_INREG:
1838 // Max of the input and what this extends.
1839 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
1842 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1843 return std::max(Tmp, Tmp2);
1846 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1847 // SRA X, C -> adds C sign bits.
1848 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1849 Tmp += C->getZExtValue();
1850 if (Tmp > VTBits) Tmp = VTBits;
1854 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1855 // shl destroys sign bits.
1856 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1857 if (C->getZExtValue() >= VTBits || // Bad shift.
1858 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
1859 return Tmp - C->getZExtValue();
1864 case ISD::XOR: // NOT is handled here.
1865 // Logical binary ops preserve the number of sign bits at the worst.
1866 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1868 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
1869 FirstAnswer = std::min(Tmp, Tmp2);
1870 // We computed what we know about the sign bits as our first
1871 // answer. Now proceed to the generic code that uses
1872 // ComputeMaskedBits, and pick whichever answer is better.
1877 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
1878 if (Tmp == 1) return 1; // Early out.
1879 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1);
1880 return std::min(Tmp, Tmp2);
1883 // If setcc returns 0/-1, all bits are sign bits.
1884 if (TLI.getSetCCResultContents() ==
1885 TargetLowering::ZeroOrNegativeOneSetCCResult)
1890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1891 unsigned RotAmt = C->getZExtValue() & (VTBits-1);
1893 // Handle rotate right by N like a rotate left by 32-N.
1894 if (Op.getOpcode() == ISD::ROTR)
1895 RotAmt = (VTBits-RotAmt) & (VTBits-1);
1897 // If we aren't rotating out all of the known-in sign bits, return the
1898 // number that are left. This handles rotl(sext(x), 1) for example.
1899 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1900 if (Tmp > RotAmt+1) return Tmp-RotAmt;
1904 // Add can have at most one carry bit. Thus we know that the output
1905 // is, at worst, one more bit than the inputs.
1906 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1907 if (Tmp == 1) return 1; // Early out.
1909 // Special case decrementing a value (ADD X, -1):
1910 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
1911 if (CRHS->isAllOnesValue()) {
1912 APInt KnownZero, KnownOne;
1913 APInt Mask = APInt::getAllOnesValue(VTBits);
1914 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
1916 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1918 if ((KnownZero | APInt(VTBits, 1)) == Mask)
1921 // If we are subtracting one from a positive number, there is no carry
1922 // out of the result.
1923 if (KnownZero.isNegative())
1927 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
1928 if (Tmp2 == 1) return 1;
1929 return std::min(Tmp, Tmp2)-1;
1933 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
1934 if (Tmp2 == 1) return 1;
1937 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
1938 if (CLHS->isNullValue()) {
1939 APInt KnownZero, KnownOne;
1940 APInt Mask = APInt::getAllOnesValue(VTBits);
1941 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
1942 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1944 if ((KnownZero | APInt(VTBits, 1)) == Mask)
1947 // If the input is known to be positive (the sign bit is known clear),
1948 // the output of the NEG has the same number of sign bits as the input.
1949 if (KnownZero.isNegative())
1952 // Otherwise, we treat this like a SUB.
1955 // Sub can have at most one carry bit. Thus we know that the output
1956 // is, at worst, one more bit than the inputs.
1957 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
1958 if (Tmp == 1) return 1; // Early out.
1959 return std::min(Tmp, Tmp2)-1;
1962 // FIXME: it's tricky to do anything useful for this, but it is an important
1963 // case for targets like X86.
1967 // Handle LOADX separately here. EXTLOAD case will fallthrough.
1968 if (Op.getOpcode() == ISD::LOAD) {
1969 LoadSDNode *LD = cast<LoadSDNode>(Op);
1970 unsigned ExtType = LD->getExtensionType();
1973 case ISD::SEXTLOAD: // '17' bits known
1974 Tmp = LD->getMemoryVT().getSizeInBits();
1975 return VTBits-Tmp+1;
1976 case ISD::ZEXTLOAD: // '16' bits known
1977 Tmp = LD->getMemoryVT().getSizeInBits();
1982 // Allow the target to implement this method for its nodes.
1983 if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1984 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1985 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1986 Op.getOpcode() == ISD::INTRINSIC_VOID) {
1987 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
1988 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
1991 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1992 // use this information.
1993 APInt KnownZero, KnownOne;
1994 APInt Mask = APInt::getAllOnesValue(VTBits);
1995 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
1997 if (KnownZero.isNegative()) { // sign bit is 0
1999 } else if (KnownOne.isNegative()) { // sign bit is 1;
2006 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2007 // the number of identical bits in the top of the input value.
2009 Mask <<= Mask.getBitWidth()-VTBits;
2010 // Return # leading zeros. We use 'min' here in case Val was zero before
2011 // shifting. We don't want to return '64' as for an i32 "0".
2012 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
2016 bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
2017 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2018 if (!GA) return false;
2019 if (GA->getOffset() != 0) return false;
2020 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal());
2021 if (!GV) return false;
2022 MachineModuleInfo *MMI = getMachineModuleInfo();
2023 return MMI && MMI->hasDebugInfo() && MMI->isVerified(GV);
2027 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
2028 /// element of the result of the vector shuffle.
2029 SDValue SelectionDAG::getShuffleScalarElt(const SDNode *N, unsigned i) {
2030 MVT VT = N->getValueType(0);
2031 SDValue PermMask = N->getOperand(2);
2032 SDValue Idx = PermMask.getOperand(i);
2033 if (Idx.getOpcode() == ISD::UNDEF)
2034 return getNode(ISD::UNDEF, VT.getVectorElementType());
2035 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
2036 unsigned NumElems = PermMask.getNumOperands();
2037 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
2040 if (V.getOpcode() == ISD::BIT_CONVERT) {
2041 V = V.getOperand(0);
2042 if (V.getValueType().getVectorNumElements() != NumElems)
2045 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
2046 return (Index == 0) ? V.getOperand(0)
2047 : getNode(ISD::UNDEF, VT.getVectorElementType());
2048 if (V.getOpcode() == ISD::BUILD_VECTOR)
2049 return V.getOperand(Index);
2050 if (V.getOpcode() == ISD::VECTOR_SHUFFLE)
2051 return getShuffleScalarElt(V.getNode(), Index);
2056 /// getNode - Gets or creates the specified node.
2058 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT) {
2059 FoldingSetNodeID ID;
2060 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0);
2062 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2063 return SDValue(E, 0);
2064 SDNode *N = NodeAllocator.Allocate<SDNode>();
2065 new (N) SDNode(Opcode, SDNode::getSDVTList(VT));
2066 CSEMap.InsertNode(N, IP);
2068 AllNodes.push_back(N);
2072 return SDValue(N, 0);
2075 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT, SDValue Operand) {
2076 // Constant fold unary operations with an integer constant operand.
2077 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) {
2078 const APInt &Val = C->getAPIntValue();
2079 unsigned BitWidth = VT.getSizeInBits();
2082 case ISD::SIGN_EXTEND:
2083 return getConstant(APInt(Val).sextOrTrunc(BitWidth), VT);
2084 case ISD::ANY_EXTEND:
2085 case ISD::ZERO_EXTEND:
2087 return getConstant(APInt(Val).zextOrTrunc(BitWidth), VT);
2088 case ISD::UINT_TO_FP:
2089 case ISD::SINT_TO_FP: {
2090 const uint64_t zero[] = {0, 0};
2091 // No compile time operations on this type.
2092 if (VT==MVT::ppcf128)
2094 APFloat apf = APFloat(APInt(BitWidth, 2, zero));
2095 (void)apf.convertFromAPInt(Val,
2096 Opcode==ISD::SINT_TO_FP,
2097 APFloat::rmNearestTiesToEven);
2098 return getConstantFP(apf, VT);
2100 case ISD::BIT_CONVERT:
2101 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
2102 return getConstantFP(Val.bitsToFloat(), VT);
2103 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
2104 return getConstantFP(Val.bitsToDouble(), VT);
2107 return getConstant(Val.byteSwap(), VT);
2109 return getConstant(Val.countPopulation(), VT);
2111 return getConstant(Val.countLeadingZeros(), VT);
2113 return getConstant(Val.countTrailingZeros(), VT);
2117 // Constant fold unary operations with a floating point constant operand.
2118 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
2119 APFloat V = C->getValueAPF(); // make copy
2120 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
2124 return getConstantFP(V, VT);
2127 return getConstantFP(V, VT);
2129 case ISD::FP_EXTEND: {
2131 // This can return overflow, underflow, or inexact; we don't care.
2132 // FIXME need to be more flexible about rounding mode.
2133 (void)V.convert(*MVTToAPFloatSemantics(VT),
2134 APFloat::rmNearestTiesToEven, &ignored);
2135 return getConstantFP(V, VT);
2137 case ISD::FP_TO_SINT:
2138 case ISD::FP_TO_UINT: {
2141 assert(integerPartWidth >= 64);
2142 // FIXME need to be more flexible about rounding mode.
2143 APFloat::opStatus s = V.convertToInteger(&x, 64U,
2144 Opcode==ISD::FP_TO_SINT,
2145 APFloat::rmTowardZero, &ignored);
2146 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
2148 return getConstant(x, VT);
2150 case ISD::BIT_CONVERT:
2151 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
2152 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
2153 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
2154 return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
2160 unsigned OpOpcode = Operand.getNode()->getOpcode();
2162 case ISD::TokenFactor:
2163 case ISD::CONCAT_VECTORS:
2164 return Operand; // Factor or concat of one node? No need.
2165 case ISD::FP_ROUND: assert(0 && "Invalid method to make FP_ROUND node");
2166 case ISD::FP_EXTEND:
2167 assert(VT.isFloatingPoint() &&
2168 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
2169 if (Operand.getValueType() == VT) return Operand; // noop conversion.
2170 if (Operand.getOpcode() == ISD::UNDEF)
2171 return getNode(ISD::UNDEF, VT);
2173 case ISD::SIGN_EXTEND:
2174 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2175 "Invalid SIGN_EXTEND!");
2176 if (Operand.getValueType() == VT) return Operand; // noop extension
2177 assert(Operand.getValueType().bitsLT(VT)
2178 && "Invalid sext node, dst < src!");
2179 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
2180 return getNode(OpOpcode, VT, Operand.getNode()->getOperand(0));
2182 case ISD::ZERO_EXTEND:
2183 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2184 "Invalid ZERO_EXTEND!");
2185 if (Operand.getValueType() == VT) return Operand; // noop extension
2186 assert(Operand.getValueType().bitsLT(VT)
2187 && "Invalid zext node, dst < src!");
2188 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
2189 return getNode(ISD::ZERO_EXTEND, VT, Operand.getNode()->getOperand(0));
2191 case ISD::ANY_EXTEND:
2192 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2193 "Invalid ANY_EXTEND!");
2194 if (Operand.getValueType() == VT) return Operand; // noop extension
2195 assert(Operand.getValueType().bitsLT(VT)
2196 && "Invalid anyext node, dst < src!");
2197 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND)
2198 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
2199 return getNode(OpOpcode, VT, Operand.getNode()->getOperand(0));
2202 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
2203 "Invalid TRUNCATE!");
2204 if (Operand.getValueType() == VT) return Operand; // noop truncate
2205 assert(Operand.getValueType().bitsGT(VT)
2206 && "Invalid truncate node, src < dst!");
2207 if (OpOpcode == ISD::TRUNCATE)
2208 return getNode(ISD::TRUNCATE, VT, Operand.getNode()->getOperand(0));
2209 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
2210 OpOpcode == ISD::ANY_EXTEND) {
2211 // If the source is smaller than the dest, we still need an extend.
2212 if (Operand.getNode()->getOperand(0).getValueType().bitsLT(VT))
2213 return getNode(OpOpcode, VT, Operand.getNode()->getOperand(0));
2214 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT))
2215 return getNode(ISD::TRUNCATE, VT, Operand.getNode()->getOperand(0));
2217 return Operand.getNode()->getOperand(0);
2220 case ISD::BIT_CONVERT:
2221 // Basic sanity checking.
2222 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits()
2223 && "Cannot BIT_CONVERT between types of different sizes!");
2224 if (VT == Operand.getValueType()) return Operand; // noop conversion.
2225 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x)
2226 return getNode(ISD::BIT_CONVERT, VT, Operand.getOperand(0));
2227 if (OpOpcode == ISD::UNDEF)
2228 return getNode(ISD::UNDEF, VT);
2230 case ISD::SCALAR_TO_VECTOR:
2231 assert(VT.isVector() && !Operand.getValueType().isVector() &&
2232 VT.getVectorElementType() == Operand.getValueType() &&
2233 "Illegal SCALAR_TO_VECTOR node!");
2234 if (OpOpcode == ISD::UNDEF)
2235 return getNode(ISD::UNDEF, VT);
2236 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
2237 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
2238 isa<ConstantSDNode>(Operand.getOperand(1)) &&
2239 Operand.getConstantOperandVal(1) == 0 &&
2240 Operand.getOperand(0).getValueType() == VT)
2241 return Operand.getOperand(0);
2244 if (OpOpcode == ISD::FSUB) // -(X-Y) -> (Y-X)
2245 return getNode(ISD::FSUB, VT, Operand.getNode()->getOperand(1),
2246 Operand.getNode()->getOperand(0));
2247 if (OpOpcode == ISD::FNEG) // --X -> X
2248 return Operand.getNode()->getOperand(0);
2251 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
2252 return getNode(ISD::FABS, VT, Operand.getNode()->getOperand(0));
2257 SDVTList VTs = getVTList(VT);
2258 if (VT != MVT::Flag) { // Don't CSE flag producing nodes
2259 FoldingSetNodeID ID;
2260 SDValue Ops[1] = { Operand };
2261 AddNodeIDNode(ID, Opcode, VTs, Ops, 1);
2263 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2264 return SDValue(E, 0);
2265 N = NodeAllocator.Allocate<UnarySDNode>();
2266 new (N) UnarySDNode(Opcode, VTs, Operand);
2267 CSEMap.InsertNode(N, IP);
2269 N = NodeAllocator.Allocate<UnarySDNode>();
2270 new (N) UnarySDNode(Opcode, VTs, Operand);
2273 AllNodes.push_back(N);
2277 return SDValue(N, 0);
2280 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode,
2282 ConstantSDNode *Cst1,
2283 ConstantSDNode *Cst2) {
2284 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue();
2287 case ISD::ADD: return getConstant(C1 + C2, VT);
2288 case ISD::SUB: return getConstant(C1 - C2, VT);
2289 case ISD::MUL: return getConstant(C1 * C2, VT);
2291 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT);
2294 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT);
2297 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT);
2300 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT);
2302 case ISD::AND: return getConstant(C1 & C2, VT);
2303 case ISD::OR: return getConstant(C1 | C2, VT);
2304 case ISD::XOR: return getConstant(C1 ^ C2, VT);
2305 case ISD::SHL: return getConstant(C1 << C2, VT);
2306 case ISD::SRL: return getConstant(C1.lshr(C2), VT);
2307 case ISD::SRA: return getConstant(C1.ashr(C2), VT);
2308 case ISD::ROTL: return getConstant(C1.rotl(C2), VT);
2309 case ISD::ROTR: return getConstant(C1.rotr(C2), VT);
2316 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
2317 SDValue N1, SDValue N2) {
2318 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2319 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2322 case ISD::TokenFactor:
2323 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
2324 N2.getValueType() == MVT::Other && "Invalid token factor!");
2325 // Fold trivial token factors.
2326 if (N1.getOpcode() == ISD::EntryToken) return N2;
2327 if (N2.getOpcode() == ISD::EntryToken) return N1;
2328 if (N1 == N2) return N1;
2330 case ISD::CONCAT_VECTORS:
2331 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2332 // one big BUILD_VECTOR.
2333 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2334 N2.getOpcode() == ISD::BUILD_VECTOR) {
2335 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2336 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2337 return getNode(ISD::BUILD_VECTOR, VT, &Elts[0], Elts.size());
2341 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2342 N1.getValueType() == VT && "Binary operator types must match!");
2343 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
2344 // worth handling here.
2345 if (N2C && N2C->isNullValue())
2347 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X
2354 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() &&
2355 N1.getValueType() == VT && "Binary operator types must match!");
2356 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
2357 // it's worth handling here.
2358 if (N2C && N2C->isNullValue())
2365 assert(VT.isInteger() && "This operator does not apply to FP types!");
2375 assert(N1.getValueType() == N2.getValueType() &&
2376 N1.getValueType() == VT && "Binary operator types must match!");
2378 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
2379 assert(N1.getValueType() == VT &&
2380 N1.getValueType().isFloatingPoint() &&
2381 N2.getValueType().isFloatingPoint() &&
2382 "Invalid FCOPYSIGN!");
2389 assert(VT == N1.getValueType() &&
2390 "Shift operators return type must be the same as their first arg");
2391 assert(VT.isInteger() && N2.getValueType().isInteger() &&
2392 "Shifts only work on integers");
2393 assert(N2.getValueType() == TLI.getShiftAmountTy() &&
2394 "Wrong type for shift amount");
2396 // Always fold shifts of i1 values so the code generator doesn't need to
2397 // handle them. Since we know the size of the shift has to be less than the
2398 // size of the value, the shift/rotate count is guaranteed to be zero.
2402 case ISD::FP_ROUND_INREG: {
2403 MVT EVT = cast<VTSDNode>(N2)->getVT();
2404 assert(VT == N1.getValueType() && "Not an inreg round!");
2405 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() &&
2406 "Cannot FP_ROUND_INREG integer types");
2407 assert(EVT.bitsLE(VT) && "Not rounding down!");
2408 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding.
2412 assert(VT.isFloatingPoint() &&
2413 N1.getValueType().isFloatingPoint() &&
2414 VT.bitsLE(N1.getValueType()) &&
2415 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!");
2416 if (N1.getValueType() == VT) return N1; // noop conversion.
2418 case ISD::AssertSext:
2419 case ISD::AssertZext: {
2420 MVT EVT = cast<VTSDNode>(N2)->getVT();
2421 assert(VT == N1.getValueType() && "Not an inreg extend!");
2422 assert(VT.isInteger() && EVT.isInteger() &&
2423 "Cannot *_EXTEND_INREG FP types");
2424 assert(EVT.bitsLE(VT) && "Not extending!");
2425 if (VT == EVT) return N1; // noop assertion.
2428 case ISD::SIGN_EXTEND_INREG: {
2429 MVT EVT = cast<VTSDNode>(N2)->getVT();
2430 assert(VT == N1.getValueType() && "Not an inreg extend!");
2431 assert(VT.isInteger() && EVT.isInteger() &&
2432 "Cannot *_EXTEND_INREG FP types");
2433 assert(EVT.bitsLE(VT) && "Not extending!");
2434 if (EVT == VT) return N1; // Not actually extending
2437 APInt Val = N1C->getAPIntValue();
2438 unsigned FromBits = cast<VTSDNode>(N2)->getVT().getSizeInBits();
2439 Val <<= Val.getBitWidth()-FromBits;
2440 Val = Val.ashr(Val.getBitWidth()-FromBits);
2441 return getConstant(Val, VT);
2445 case ISD::EXTRACT_VECTOR_ELT:
2446 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
2447 if (N1.getOpcode() == ISD::UNDEF)
2448 return getNode(ISD::UNDEF, VT);
2450 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
2451 // expanding copies of large vectors from registers.
2453 N1.getOpcode() == ISD::CONCAT_VECTORS &&
2454 N1.getNumOperands() > 0) {
2456 N1.getOperand(0).getValueType().getVectorNumElements();
2457 return getNode(ISD::EXTRACT_VECTOR_ELT, VT,
2458 N1.getOperand(N2C->getZExtValue() / Factor),
2459 getConstant(N2C->getZExtValue() % Factor,
2460 N2.getValueType()));
2463 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
2464 // expanding large vector constants.
2465 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR)
2466 return N1.getOperand(N2C->getZExtValue());
2468 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
2469 // operations are lowered to scalars.
2470 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
2471 if (N1.getOperand(2) == N2)
2472 return N1.getOperand(1);
2474 return getNode(ISD::EXTRACT_VECTOR_ELT, VT, N1.getOperand(0), N2);
2477 case ISD::EXTRACT_ELEMENT:
2478 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
2479 assert(!N1.getValueType().isVector() && !VT.isVector() &&
2480 (N1.getValueType().isInteger() == VT.isInteger()) &&
2481 "Wrong types for EXTRACT_ELEMENT!");
2483 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
2484 // 64-bit integers into 32-bit parts. Instead of building the extract of
2485 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
2486 if (N1.getOpcode() == ISD::BUILD_PAIR)
2487 return N1.getOperand(N2C->getZExtValue());
2489 // EXTRACT_ELEMENT of a constant int is also very common.
2490 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2491 unsigned ElementSize = VT.getSizeInBits();
2492 unsigned Shift = ElementSize * N2C->getZExtValue();
2493 APInt ShiftedVal = C->getAPIntValue().lshr(Shift);
2494 return getConstant(ShiftedVal.trunc(ElementSize), VT);
2497 case ISD::EXTRACT_SUBVECTOR:
2498 if (N1.getValueType() == VT) // Trivial extraction.
2505 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C);
2506 if (SV.getNode()) return SV;
2507 } else { // Cannonicalize constant to RHS if commutative
2508 if (isCommutativeBinOp(Opcode)) {
2509 std::swap(N1C, N2C);
2515 // Constant fold FP operations.
2516 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
2517 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
2519 if (!N2CFP && isCommutativeBinOp(Opcode)) {
2520 // Cannonicalize constant to RHS if commutative
2521 std::swap(N1CFP, N2CFP);
2523 } else if (N2CFP && VT != MVT::ppcf128) {
2524 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
2525 APFloat::opStatus s;
2528 s = V1.add(V2, APFloat::rmNearestTiesToEven);
2529 if (s != APFloat::opInvalidOp)
2530 return getConstantFP(V1, VT);
2533 s = V1.subtract(V2, APFloat::rmNearestTiesToEven);
2534 if (s!=APFloat::opInvalidOp)
2535 return getConstantFP(V1, VT);
2538 s = V1.multiply(V2, APFloat::rmNearestTiesToEven);
2539 if (s!=APFloat::opInvalidOp)
2540 return getConstantFP(V1, VT);
2543 s = V1.divide(V2, APFloat::rmNearestTiesToEven);
2544 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2545 return getConstantFP(V1, VT);
2548 s = V1.mod(V2, APFloat::rmNearestTiesToEven);
2549 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero)
2550 return getConstantFP(V1, VT);
2552 case ISD::FCOPYSIGN:
2554 return getConstantFP(V1, VT);
2560 // Canonicalize an UNDEF to the RHS, even over a constant.
2561 if (N1.getOpcode() == ISD::UNDEF) {
2562 if (isCommutativeBinOp(Opcode)) {
2566 case ISD::FP_ROUND_INREG:
2567 case ISD::SIGN_EXTEND_INREG:
2573 return N1; // fold op(undef, arg2) -> undef
2581 return getConstant(0, VT); // fold op(undef, arg2) -> 0
2582 // For vectors, we can't easily build an all zero vector, just return
2589 // Fold a bunch of operators when the RHS is undef.
2590 if (N2.getOpcode() == ISD::UNDEF) {
2593 if (N1.getOpcode() == ISD::UNDEF)
2594 // Handle undef ^ undef -> 0 special case. This is a common
2596 return getConstant(0, VT);
2611 return N2; // fold op(arg1, undef) -> undef
2617 return getConstant(0, VT); // fold op(arg1, undef) -> 0
2618 // For vectors, we can't easily build an all zero vector, just return
2623 return getConstant(VT.getIntegerVTBitMask(), VT);
2624 // For vectors, we can't easily build an all one vector, just return
2632 // Memoize this node if possible.
2634 SDVTList VTs = getVTList(VT);
2635 if (VT != MVT::Flag) {
2636 SDValue Ops[] = { N1, N2 };
2637 FoldingSetNodeID ID;
2638 AddNodeIDNode(ID, Opcode, VTs, Ops, 2);
2640 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2641 return SDValue(E, 0);
2642 N = NodeAllocator.Allocate<BinarySDNode>();
2643 new (N) BinarySDNode(Opcode, VTs, N1, N2);
2644 CSEMap.InsertNode(N, IP);
2646 N = NodeAllocator.Allocate<BinarySDNode>();
2647 new (N) BinarySDNode(Opcode, VTs, N1, N2);
2650 AllNodes.push_back(N);
2654 return SDValue(N, 0);
2657 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
2658 SDValue N1, SDValue N2, SDValue N3) {
2659 // Perform various simplifications.
2660 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2661 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
2663 case ISD::CONCAT_VECTORS:
2664 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to
2665 // one big BUILD_VECTOR.
2666 if (N1.getOpcode() == ISD::BUILD_VECTOR &&
2667 N2.getOpcode() == ISD::BUILD_VECTOR &&
2668 N3.getOpcode() == ISD::BUILD_VECTOR) {
2669 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end());
2670 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end());
2671 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end());
2672 return getNode(ISD::BUILD_VECTOR, VT, &Elts[0], Elts.size());
2676 // Use FoldSetCC to simplify SETCC's.
2677 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get());
2678 if (Simp.getNode()) return Simp;
2683 if (N1C->getZExtValue())
2684 return N2; // select true, X, Y -> X
2686 return N3; // select false, X, Y -> Y
2689 if (N2 == N3) return N2; // select C, X, X -> X
2693 if (N2C->getZExtValue()) // Unconditional branch
2694 return getNode(ISD::BR, MVT::Other, N1, N3);
2696 return N1; // Never-taken branch
2699 case ISD::VECTOR_SHUFFLE:
2700 assert(VT == N1.getValueType() && VT == N2.getValueType() &&
2701 VT.isVector() && N3.getValueType().isVector() &&
2702 N3.getOpcode() == ISD::BUILD_VECTOR &&
2703 VT.getVectorNumElements() == N3.getNumOperands() &&
2704 "Illegal VECTOR_SHUFFLE node!");
2706 case ISD::BIT_CONVERT:
2707 // Fold bit_convert nodes from a type to themselves.
2708 if (N1.getValueType() == VT)
2713 // Memoize node if it doesn't produce a flag.
2715 SDVTList VTs = getVTList(VT);
2716 if (VT != MVT::Flag) {
2717 SDValue Ops[] = { N1, N2, N3 };
2718 FoldingSetNodeID ID;
2719 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
2721 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
2722 return SDValue(E, 0);
2723 N = NodeAllocator.Allocate<TernarySDNode>();
2724 new (N) TernarySDNode(Opcode, VTs, N1, N2, N3);
2725 CSEMap.InsertNode(N, IP);
2727 N = NodeAllocator.Allocate<TernarySDNode>();
2728 new (N) TernarySDNode(Opcode, VTs, N1, N2, N3);
2730 AllNodes.push_back(N);
2734 return SDValue(N, 0);
2737 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
2738 SDValue N1, SDValue N2, SDValue N3,
2740 SDValue Ops[] = { N1, N2, N3, N4 };
2741 return getNode(Opcode, VT, Ops, 4);
2744 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
2745 SDValue N1, SDValue N2, SDValue N3,
2746 SDValue N4, SDValue N5) {
2747 SDValue Ops[] = { N1, N2, N3, N4, N5 };
2748 return getNode(Opcode, VT, Ops, 5);
2751 /// getMemsetValue - Vectorized representation of the memset value
2753 static SDValue getMemsetValue(SDValue Value, MVT VT, SelectionDAG &DAG) {
2754 unsigned NumBits = VT.isVector() ?
2755 VT.getVectorElementType().getSizeInBits() : VT.getSizeInBits();
2756 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
2757 APInt Val = APInt(NumBits, C->getZExtValue() & 255);
2759 for (unsigned i = NumBits; i > 8; i >>= 1) {
2760 Val = (Val << Shift) | Val;
2764 return DAG.getConstant(Val, VT);
2765 return DAG.getConstantFP(APFloat(Val), VT);
2768 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2769 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value);
2771 for (unsigned i = NumBits; i > 8; i >>= 1) {
2772 Value = DAG.getNode(ISD::OR, VT,
2773 DAG.getNode(ISD::SHL, VT, Value,
2774 DAG.getConstant(Shift,
2775 TLI.getShiftAmountTy())),
2783 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
2784 /// used when a memcpy is turned into a memset when the source is a constant
2786 static SDValue getMemsetStringVal(MVT VT, SelectionDAG &DAG,
2787 const TargetLowering &TLI,
2788 std::string &Str, unsigned Offset) {
2789 // Handle vector with all elements zero.
2792 return DAG.getConstant(0, VT);
2793 unsigned NumElts = VT.getVectorNumElements();
2794 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
2795 return DAG.getNode(ISD::BIT_CONVERT, VT,
2796 DAG.getConstant(0, MVT::getVectorVT(EltVT, NumElts)));
2799 assert(!VT.isVector() && "Can't handle vector type here!");
2800 unsigned NumBits = VT.getSizeInBits();
2801 unsigned MSB = NumBits / 8;
2803 if (TLI.isLittleEndian())
2804 Offset = Offset + MSB - 1;
2805 for (unsigned i = 0; i != MSB; ++i) {
2806 Val = (Val << 8) | (unsigned char)Str[Offset];
2807 Offset += TLI.isLittleEndian() ? -1 : 1;
2809 return DAG.getConstant(Val, VT);
2812 /// getMemBasePlusOffset - Returns base and offset node for the
2814 static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset,
2815 SelectionDAG &DAG) {
2816 MVT VT = Base.getValueType();
2817 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT));
2820 /// isMemSrcFromString - Returns true if memcpy source is a string constant.
2822 static bool isMemSrcFromString(SDValue Src, std::string &Str) {
2823 unsigned SrcDelta = 0;
2824 GlobalAddressSDNode *G = NULL;
2825 if (Src.getOpcode() == ISD::GlobalAddress)
2826 G = cast<GlobalAddressSDNode>(Src);
2827 else if (Src.getOpcode() == ISD::ADD &&
2828 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
2829 Src.getOperand(1).getOpcode() == ISD::Constant) {
2830 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
2831 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
2836 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
2837 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false))
2843 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
2844 /// to replace the memset / memcpy is below the threshold. It also returns the
2845 /// types of the sequence of memory ops to perform memset / memcpy.
2847 bool MeetsMaxMemopRequirement(std::vector<MVT> &MemOps,
2848 SDValue Dst, SDValue Src,
2849 unsigned Limit, uint64_t Size, unsigned &Align,
2850 std::string &Str, bool &isSrcStr,
2852 const TargetLowering &TLI) {
2853 isSrcStr = isMemSrcFromString(Src, Str);
2854 bool isSrcConst = isa<ConstantSDNode>(Src);
2855 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses();
2856 MVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr);
2857 if (VT != MVT::iAny) {
2858 unsigned NewAlign = (unsigned)
2859 TLI.getTargetData()->getABITypeAlignment(VT.getTypeForMVT());
2860 // If source is a string constant, this will require an unaligned load.
2861 if (NewAlign > Align && (isSrcConst || AllowUnalign)) {
2862 if (Dst.getOpcode() != ISD::FrameIndex) {
2863 // Can't change destination alignment. It requires a unaligned store.
2867 int FI = cast<FrameIndexSDNode>(Dst)->getIndex();
2868 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2869 if (MFI->isFixedObjectIndex(FI)) {
2870 // Can't change destination alignment. It requires a unaligned store.
2874 // Give the stack frame object a larger alignment if needed.
2875 if (MFI->getObjectAlignment(FI) < NewAlign)
2876 MFI->setObjectAlignment(FI, NewAlign);
2883 if (VT == MVT::iAny) {
2887 switch (Align & 7) {
2888 case 0: VT = MVT::i64; break;
2889 case 4: VT = MVT::i32; break;
2890 case 2: VT = MVT::i16; break;
2891 default: VT = MVT::i8; break;
2896 while (!TLI.isTypeLegal(LVT))
2897 LVT = (MVT::SimpleValueType)(LVT.getSimpleVT() - 1);
2898 assert(LVT.isInteger());
2904 unsigned NumMemOps = 0;
2906 unsigned VTSize = VT.getSizeInBits() / 8;
2907 while (VTSize > Size) {
2908 // For now, only use non-vector load / store's for the left-over pieces.
2909 if (VT.isVector()) {
2911 while (!TLI.isTypeLegal(VT))
2912 VT = (MVT::SimpleValueType)(VT.getSimpleVT() - 1);
2913 VTSize = VT.getSizeInBits() / 8;
2915 VT = (MVT::SimpleValueType)(VT.getSimpleVT() - 1);
2920 if (++NumMemOps > Limit)
2922 MemOps.push_back(VT);
2929 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG,
2930 SDValue Chain, SDValue Dst,
2931 SDValue Src, uint64_t Size,
2932 unsigned Align, bool AlwaysInline,
2933 const Value *DstSV, uint64_t DstSVOff,
2934 const Value *SrcSV, uint64_t SrcSVOff){
2935 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2937 // Expand memcpy to a series of load and store ops if the size operand falls
2938 // below a certain threshold.
2939 std::vector<MVT> MemOps;
2940 uint64_t Limit = -1ULL;
2942 Limit = TLI.getMaxStoresPerMemcpy();
2943 unsigned DstAlign = Align; // Destination alignment can change.
2946 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
2947 Str, CopyFromStr, DAG, TLI))
2951 bool isZeroStr = CopyFromStr && Str.empty();
2952 SmallVector<SDValue, 8> OutChains;
2953 unsigned NumMemOps = MemOps.size();
2954 uint64_t SrcOff = 0, DstOff = 0;
2955 for (unsigned i = 0; i < NumMemOps; i++) {
2957 unsigned VTSize = VT.getSizeInBits() / 8;
2958 SDValue Value, Store;
2960 if (CopyFromStr && (isZeroStr || !VT.isVector())) {
2961 // It's unlikely a store of a vector immediate can be done in a single
2962 // instruction. It would require a load from a constantpool first.
2963 // We also handle store a vector with all zero's.
2964 // FIXME: Handle other cases where store of vector immediate is done in
2965 // a single instruction.
2966 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff);
2967 Store = DAG.getStore(Chain, Value,
2968 getMemBasePlusOffset(Dst, DstOff, DAG),
2969 DstSV, DstSVOff + DstOff, false, DstAlign);
2971 Value = DAG.getLoad(VT, Chain,
2972 getMemBasePlusOffset(Src, SrcOff, DAG),
2973 SrcSV, SrcSVOff + SrcOff, false, Align);
2974 Store = DAG.getStore(Chain, Value,
2975 getMemBasePlusOffset(Dst, DstOff, DAG),
2976 DstSV, DstSVOff + DstOff, false, DstAlign);
2978 OutChains.push_back(Store);
2983 return DAG.getNode(ISD::TokenFactor, MVT::Other,
2984 &OutChains[0], OutChains.size());
2987 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG,
2988 SDValue Chain, SDValue Dst,
2989 SDValue Src, uint64_t Size,
2990 unsigned Align, bool AlwaysInline,
2991 const Value *DstSV, uint64_t DstSVOff,
2992 const Value *SrcSV, uint64_t SrcSVOff){
2993 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2995 // Expand memmove to a series of load and store ops if the size operand falls
2996 // below a certain threshold.
2997 std::vector<MVT> MemOps;
2998 uint64_t Limit = -1ULL;
3000 Limit = TLI.getMaxStoresPerMemmove();
3001 unsigned DstAlign = Align; // Destination alignment can change.
3004 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
3005 Str, CopyFromStr, DAG, TLI))
3008 uint64_t SrcOff = 0, DstOff = 0;
3010 SmallVector<SDValue, 8> LoadValues;
3011 SmallVector<SDValue, 8> LoadChains;
3012 SmallVector<SDValue, 8> OutChains;
3013 unsigned NumMemOps = MemOps.size();
3014 for (unsigned i = 0; i < NumMemOps; i++) {
3016 unsigned VTSize = VT.getSizeInBits() / 8;
3017 SDValue Value, Store;
3019 Value = DAG.getLoad(VT, Chain,
3020 getMemBasePlusOffset(Src, SrcOff, DAG),
3021 SrcSV, SrcSVOff + SrcOff, false, Align);
3022 LoadValues.push_back(Value);
3023 LoadChains.push_back(Value.getValue(1));
3026 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
3027 &LoadChains[0], LoadChains.size());
3029 for (unsigned i = 0; i < NumMemOps; i++) {
3031 unsigned VTSize = VT.getSizeInBits() / 8;
3032 SDValue Value, Store;
3034 Store = DAG.getStore(Chain, LoadValues[i],
3035 getMemBasePlusOffset(Dst, DstOff, DAG),
3036 DstSV, DstSVOff + DstOff, false, DstAlign);
3037 OutChains.push_back(Store);
3041 return DAG.getNode(ISD::TokenFactor, MVT::Other,
3042 &OutChains[0], OutChains.size());
3045 static SDValue getMemsetStores(SelectionDAG &DAG,
3046 SDValue Chain, SDValue Dst,
3047 SDValue Src, uint64_t Size,
3049 const Value *DstSV, uint64_t DstSVOff) {
3050 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3052 // Expand memset to a series of load/store ops if the size operand
3053 // falls below a certain threshold.
3054 std::vector<MVT> MemOps;
3057 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(),
3058 Size, Align, Str, CopyFromStr, DAG, TLI))
3061 SmallVector<SDValue, 8> OutChains;
3062 uint64_t DstOff = 0;
3064 unsigned NumMemOps = MemOps.size();
3065 for (unsigned i = 0; i < NumMemOps; i++) {
3067 unsigned VTSize = VT.getSizeInBits() / 8;
3068 SDValue Value = getMemsetValue(Src, VT, DAG);
3069 SDValue Store = DAG.getStore(Chain, Value,
3070 getMemBasePlusOffset(Dst, DstOff, DAG),
3071 DstSV, DstSVOff + DstOff);
3072 OutChains.push_back(Store);
3076 return DAG.getNode(ISD::TokenFactor, MVT::Other,
3077 &OutChains[0], OutChains.size());
3080 SDValue SelectionDAG::getMemcpy(SDValue Chain, SDValue Dst,
3081 SDValue Src, SDValue Size,
3082 unsigned Align, bool AlwaysInline,
3083 const Value *DstSV, uint64_t DstSVOff,
3084 const Value *SrcSV, uint64_t SrcSVOff) {
3086 // Check to see if we should lower the memcpy to loads and stores first.
3087 // For cases within the target-specified limits, this is the best choice.
3088 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3090 // Memcpy with size zero? Just return the original chain.
3091 if (ConstantSize->isNullValue())
3095 getMemcpyLoadsAndStores(*this, Chain, Dst, Src,
3096 ConstantSize->getZExtValue(),
3097 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3098 if (Result.getNode())
3102 // Then check to see if we should lower the memcpy with target-specific
3103 // code. If the target chooses to do this, this is the next best.
3105 TLI.EmitTargetCodeForMemcpy(*this, Chain, Dst, Src, Size, Align,
3107 DstSV, DstSVOff, SrcSV, SrcSVOff);
3108 if (Result.getNode())
3111 // If we really need inline code and the target declined to provide it,
3112 // use a (potentially long) sequence of loads and stores.
3114 assert(ConstantSize && "AlwaysInline requires a constant size!");
3115 return getMemcpyLoadsAndStores(*this, Chain, Dst, Src,
3116 ConstantSize->getZExtValue(), Align, true,
3117 DstSV, DstSVOff, SrcSV, SrcSVOff);
3120 // Emit a library call.
3121 TargetLowering::ArgListTy Args;
3122 TargetLowering::ArgListEntry Entry;
3123 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3124 Entry.Node = Dst; Args.push_back(Entry);
3125 Entry.Node = Src; Args.push_back(Entry);
3126 Entry.Node = Size; Args.push_back(Entry);
3127 std::pair<SDValue,SDValue> CallResult =
3128 TLI.LowerCallTo(Chain, Type::VoidTy,
3129 false, false, false, false, CallingConv::C, false,
3130 getExternalSymbol("memcpy", TLI.getPointerTy()),
3132 return CallResult.second;
3135 SDValue SelectionDAG::getMemmove(SDValue Chain, SDValue Dst,
3136 SDValue Src, SDValue Size,
3138 const Value *DstSV, uint64_t DstSVOff,
3139 const Value *SrcSV, uint64_t SrcSVOff) {
3141 // Check to see if we should lower the memmove to loads and stores first.
3142 // For cases within the target-specified limits, this is the best choice.
3143 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3145 // Memmove with size zero? Just return the original chain.
3146 if (ConstantSize->isNullValue())
3150 getMemmoveLoadsAndStores(*this, Chain, Dst, Src,
3151 ConstantSize->getZExtValue(),
3152 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
3153 if (Result.getNode())
3157 // Then check to see if we should lower the memmove with target-specific
3158 // code. If the target chooses to do this, this is the next best.
3160 TLI.EmitTargetCodeForMemmove(*this, Chain, Dst, Src, Size, Align,
3161 DstSV, DstSVOff, SrcSV, SrcSVOff);
3162 if (Result.getNode())
3165 // Emit a library call.
3166 TargetLowering::ArgListTy Args;
3167 TargetLowering::ArgListEntry Entry;
3168 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3169 Entry.Node = Dst; Args.push_back(Entry);
3170 Entry.Node = Src; Args.push_back(Entry);
3171 Entry.Node = Size; Args.push_back(Entry);
3172 std::pair<SDValue,SDValue> CallResult =
3173 TLI.LowerCallTo(Chain, Type::VoidTy,
3174 false, false, false, false, CallingConv::C, false,
3175 getExternalSymbol("memmove", TLI.getPointerTy()),
3177 return CallResult.second;
3180 SDValue SelectionDAG::getMemset(SDValue Chain, SDValue Dst,
3181 SDValue Src, SDValue Size,
3183 const Value *DstSV, uint64_t DstSVOff) {
3185 // Check to see if we should lower the memset to stores first.
3186 // For cases within the target-specified limits, this is the best choice.
3187 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
3189 // Memset with size zero? Just return the original chain.
3190 if (ConstantSize->isNullValue())
3194 getMemsetStores(*this, Chain, Dst, Src, ConstantSize->getZExtValue(),
3195 Align, DstSV, DstSVOff);
3196 if (Result.getNode())
3200 // Then check to see if we should lower the memset with target-specific
3201 // code. If the target chooses to do this, this is the next best.
3203 TLI.EmitTargetCodeForMemset(*this, Chain, Dst, Src, Size, Align,
3205 if (Result.getNode())
3208 // Emit a library call.
3209 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType();
3210 TargetLowering::ArgListTy Args;
3211 TargetLowering::ArgListEntry Entry;
3212 Entry.Node = Dst; Entry.Ty = IntPtrTy;
3213 Args.push_back(Entry);
3214 // Extend or truncate the argument to be an i32 value for the call.
3215 if (Src.getValueType().bitsGT(MVT::i32))
3216 Src = getNode(ISD::TRUNCATE, MVT::i32, Src);
3218 Src = getNode(ISD::ZERO_EXTEND, MVT::i32, Src);
3219 Entry.Node = Src; Entry.Ty = Type::Int32Ty; Entry.isSExt = true;
3220 Args.push_back(Entry);
3221 Entry.Node = Size; Entry.Ty = IntPtrTy; Entry.isSExt = false;
3222 Args.push_back(Entry);
3223 std::pair<SDValue,SDValue> CallResult =
3224 TLI.LowerCallTo(Chain, Type::VoidTy,
3225 false, false, false, false, CallingConv::C, false,
3226 getExternalSymbol("memset", TLI.getPointerTy()),
3228 return CallResult.second;
3231 SDValue SelectionDAG::getAtomic(unsigned Opcode, SDValue Chain,
3232 SDValue Ptr, SDValue Cmp,
3233 SDValue Swp, const Value* PtrVal,
3234 unsigned Alignment) {
3235 assert((Opcode == ISD::ATOMIC_CMP_SWAP_8 ||
3236 Opcode == ISD::ATOMIC_CMP_SWAP_16 ||
3237 Opcode == ISD::ATOMIC_CMP_SWAP_32 ||
3238 Opcode == ISD::ATOMIC_CMP_SWAP_64) && "Invalid Atomic Op");
3239 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
3241 MVT VT = Cmp.getValueType();
3243 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3244 Alignment = getMVTAlignment(VT);
3246 SDVTList VTs = getVTList(VT, MVT::Other);
3247 FoldingSetNodeID ID;
3248 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
3249 AddNodeIDNode(ID, Opcode, VTs, Ops, 4);
3251 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3252 return SDValue(E, 0);
3253 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3254 new (N) AtomicSDNode(Opcode, VTs, Chain, Ptr, Cmp, Swp, PtrVal, Alignment);
3255 CSEMap.InsertNode(N, IP);
3256 AllNodes.push_back(N);
3257 return SDValue(N, 0);
3260 SDValue SelectionDAG::getAtomic(unsigned Opcode, SDValue Chain,
3261 SDValue Ptr, SDValue Val,
3262 const Value* PtrVal,
3263 unsigned Alignment) {
3264 assert((Opcode == ISD::ATOMIC_LOAD_ADD_8 ||
3265 Opcode == ISD::ATOMIC_LOAD_SUB_8 ||
3266 Opcode == ISD::ATOMIC_LOAD_AND_8 ||
3267 Opcode == ISD::ATOMIC_LOAD_OR_8 ||
3268 Opcode == ISD::ATOMIC_LOAD_XOR_8 ||
3269 Opcode == ISD::ATOMIC_LOAD_NAND_8 ||
3270 Opcode == ISD::ATOMIC_LOAD_MIN_8 ||
3271 Opcode == ISD::ATOMIC_LOAD_MAX_8 ||
3272 Opcode == ISD::ATOMIC_LOAD_UMIN_8 ||
3273 Opcode == ISD::ATOMIC_LOAD_UMAX_8 ||
3274 Opcode == ISD::ATOMIC_SWAP_8 ||
3275 Opcode == ISD::ATOMIC_LOAD_ADD_16 ||
3276 Opcode == ISD::ATOMIC_LOAD_SUB_16 ||
3277 Opcode == ISD::ATOMIC_LOAD_AND_16 ||
3278 Opcode == ISD::ATOMIC_LOAD_OR_16 ||
3279 Opcode == ISD::ATOMIC_LOAD_XOR_16 ||
3280 Opcode == ISD::ATOMIC_LOAD_NAND_16 ||
3281 Opcode == ISD::ATOMIC_LOAD_MIN_16 ||
3282 Opcode == ISD::ATOMIC_LOAD_MAX_16 ||
3283 Opcode == ISD::ATOMIC_LOAD_UMIN_16 ||
3284 Opcode == ISD::ATOMIC_LOAD_UMAX_16 ||
3285 Opcode == ISD::ATOMIC_SWAP_16 ||
3286 Opcode == ISD::ATOMIC_LOAD_ADD_32 ||
3287 Opcode == ISD::ATOMIC_LOAD_SUB_32 ||
3288 Opcode == ISD::ATOMIC_LOAD_AND_32 ||
3289 Opcode == ISD::ATOMIC_LOAD_OR_32 ||
3290 Opcode == ISD::ATOMIC_LOAD_XOR_32 ||
3291 Opcode == ISD::ATOMIC_LOAD_NAND_32 ||
3292 Opcode == ISD::ATOMIC_LOAD_MIN_32 ||
3293 Opcode == ISD::ATOMIC_LOAD_MAX_32 ||
3294 Opcode == ISD::ATOMIC_LOAD_UMIN_32 ||
3295 Opcode == ISD::ATOMIC_LOAD_UMAX_32 ||
3296 Opcode == ISD::ATOMIC_SWAP_32 ||
3297 Opcode == ISD::ATOMIC_LOAD_ADD_64 ||
3298 Opcode == ISD::ATOMIC_LOAD_SUB_64 ||
3299 Opcode == ISD::ATOMIC_LOAD_AND_64 ||
3300 Opcode == ISD::ATOMIC_LOAD_OR_64 ||
3301 Opcode == ISD::ATOMIC_LOAD_XOR_64 ||
3302 Opcode == ISD::ATOMIC_LOAD_NAND_64 ||
3303 Opcode == ISD::ATOMIC_LOAD_MIN_64 ||
3304 Opcode == ISD::ATOMIC_LOAD_MAX_64 ||
3305 Opcode == ISD::ATOMIC_LOAD_UMIN_64 ||
3306 Opcode == ISD::ATOMIC_LOAD_UMAX_64 ||
3307 Opcode == ISD::ATOMIC_SWAP_64) && "Invalid Atomic Op");
3309 MVT VT = Val.getValueType();
3311 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3312 Alignment = getMVTAlignment(VT);
3314 SDVTList VTs = getVTList(VT, MVT::Other);
3315 FoldingSetNodeID ID;
3316 SDValue Ops[] = {Chain, Ptr, Val};
3317 AddNodeIDNode(ID, Opcode, VTs, Ops, 3);
3319 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3320 return SDValue(E, 0);
3321 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>();
3322 new (N) AtomicSDNode(Opcode, VTs, Chain, Ptr, Val, PtrVal, Alignment);
3323 CSEMap.InsertNode(N, IP);
3324 AllNodes.push_back(N);
3325 return SDValue(N, 0);
3328 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
3329 /// Allowed to return something different (and simpler) if Simplify is true.
3330 SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps,
3332 if (Simplify && NumOps == 1)
3335 SmallVector<MVT, 4> VTs;
3336 VTs.reserve(NumOps);
3337 for (unsigned i = 0; i < NumOps; ++i)
3338 VTs.push_back(Ops[i].getValueType());
3339 return getNode(ISD::MERGE_VALUES, getVTList(&VTs[0], NumOps), Ops, NumOps);
3343 SelectionDAG::getMemIntrinsicNode(unsigned Opcode,
3344 const MVT *VTs, unsigned NumVTs,
3345 const SDValue *Ops, unsigned NumOps,
3346 MVT MemVT, const Value *srcValue, int SVOff,
3347 unsigned Align, bool Vol,
3348 bool ReadMem, bool WriteMem) {
3349 return getMemIntrinsicNode(Opcode, makeVTList(VTs, NumVTs), Ops, NumOps,
3350 MemVT, srcValue, SVOff, Align, Vol,
3355 SelectionDAG::getMemIntrinsicNode(unsigned Opcode, SDVTList VTList,
3356 const SDValue *Ops, unsigned NumOps,
3357 MVT MemVT, const Value *srcValue, int SVOff,
3358 unsigned Align, bool Vol,
3359 bool ReadMem, bool WriteMem) {
3360 // Memoize the node unless it returns a flag.
3361 MemIntrinsicSDNode *N;
3362 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3363 FoldingSetNodeID ID;
3364 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3366 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3367 return SDValue(E, 0);
3369 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3370 new (N) MemIntrinsicSDNode(Opcode, VTList, Ops, NumOps, MemVT,
3371 srcValue, SVOff, Align, Vol, ReadMem, WriteMem);
3372 CSEMap.InsertNode(N, IP);
3374 N = NodeAllocator.Allocate<MemIntrinsicSDNode>();
3375 new (N) MemIntrinsicSDNode(Opcode, VTList, Ops, NumOps, MemVT,
3376 srcValue, SVOff, Align, Vol, ReadMem, WriteMem);
3378 AllNodes.push_back(N);
3379 return SDValue(N, 0);
3383 SelectionDAG::getCall(unsigned CallingConv, bool IsVarArgs, bool IsTailCall,
3384 bool IsInreg, SDVTList VTs,
3385 const SDValue *Operands, unsigned NumOperands) {
3386 // Do not include isTailCall in the folding set profile.
3387 FoldingSetNodeID ID;
3388 AddNodeIDNode(ID, ISD::CALL, VTs, Operands, NumOperands);
3389 ID.AddInteger(CallingConv);
3390 ID.AddInteger(IsVarArgs);
3392 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
3393 // Instead of including isTailCall in the folding set, we just
3394 // set the flag of the existing node.
3396 cast<CallSDNode>(E)->setNotTailCall();
3397 return SDValue(E, 0);
3399 SDNode *N = NodeAllocator.Allocate<CallSDNode>();
3400 new (N) CallSDNode(CallingConv, IsVarArgs, IsTailCall, IsInreg,
3401 VTs, Operands, NumOperands);
3402 CSEMap.InsertNode(N, IP);
3403 AllNodes.push_back(N);
3404 return SDValue(N, 0);
3408 SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType,
3409 MVT VT, SDValue Chain,
3410 SDValue Ptr, SDValue Offset,
3411 const Value *SV, int SVOffset, MVT EVT,
3412 bool isVolatile, unsigned Alignment) {
3413 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3414 Alignment = getMVTAlignment(VT);
3417 ExtType = ISD::NON_EXTLOAD;
3418 } else if (ExtType == ISD::NON_EXTLOAD) {
3419 assert(VT == EVT && "Non-extending load from different memory type!");
3423 assert(EVT.getVectorNumElements() == VT.getVectorNumElements() &&
3424 "Invalid vector extload!");
3426 assert(EVT.bitsLT(VT) &&
3427 "Should only be an extending load, not truncating!");
3428 assert((ExtType == ISD::EXTLOAD || VT.isInteger()) &&
3429 "Cannot sign/zero extend a FP/Vector load!");
3430 assert(VT.isInteger() == EVT.isInteger() &&
3431 "Cannot convert from FP to Int or Int -> FP!");
3434 bool Indexed = AM != ISD::UNINDEXED;
3435 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
3436 "Unindexed load with an offset!");
3438 SDVTList VTs = Indexed ?
3439 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
3440 SDValue Ops[] = { Chain, Ptr, Offset };
3441 FoldingSetNodeID ID;
3442 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3);
3444 ID.AddInteger(ExtType);
3445 ID.AddInteger(EVT.getRawBits());
3446 ID.AddInteger(encodeMemSDNodeFlags(isVolatile, Alignment));
3448 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3449 return SDValue(E, 0);
3450 SDNode *N = NodeAllocator.Allocate<LoadSDNode>();
3451 new (N) LoadSDNode(Ops, VTs, AM, ExtType, EVT, SV, SVOffset,
3452 Alignment, isVolatile);
3453 CSEMap.InsertNode(N, IP);
3454 AllNodes.push_back(N);
3455 return SDValue(N, 0);
3458 SDValue SelectionDAG::getLoad(MVT VT,
3459 SDValue Chain, SDValue Ptr,
3460 const Value *SV, int SVOffset,
3461 bool isVolatile, unsigned Alignment) {
3462 SDValue Undef = getNode(ISD::UNDEF, Ptr.getValueType());
3463 return getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef,
3464 SV, SVOffset, VT, isVolatile, Alignment);
3467 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, MVT VT,
3468 SDValue Chain, SDValue Ptr,
3470 int SVOffset, MVT EVT,
3471 bool isVolatile, unsigned Alignment) {
3472 SDValue Undef = getNode(ISD::UNDEF, Ptr.getValueType());
3473 return getLoad(ISD::UNINDEXED, ExtType, VT, Chain, Ptr, Undef,
3474 SV, SVOffset, EVT, isVolatile, Alignment);
3478 SelectionDAG::getIndexedLoad(SDValue OrigLoad, SDValue Base,
3479 SDValue Offset, ISD::MemIndexedMode AM) {
3480 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
3481 assert(LD->getOffset().getOpcode() == ISD::UNDEF &&
3482 "Load is already a indexed load!");
3483 return getLoad(AM, LD->getExtensionType(), OrigLoad.getValueType(),
3484 LD->getChain(), Base, Offset, LD->getSrcValue(),
3485 LD->getSrcValueOffset(), LD->getMemoryVT(),
3486 LD->isVolatile(), LD->getAlignment());
3489 SDValue SelectionDAG::getStore(SDValue Chain, SDValue Val,
3490 SDValue Ptr, const Value *SV, int SVOffset,
3491 bool isVolatile, unsigned Alignment) {
3492 MVT VT = Val.getValueType();
3494 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3495 Alignment = getMVTAlignment(VT);
3497 SDVTList VTs = getVTList(MVT::Other);
3498 SDValue Undef = getNode(ISD::UNDEF, Ptr.getValueType());
3499 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3500 FoldingSetNodeID ID;
3501 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3502 ID.AddInteger(ISD::UNINDEXED);
3503 ID.AddInteger(false);
3504 ID.AddInteger(VT.getRawBits());
3505 ID.AddInteger(encodeMemSDNodeFlags(isVolatile, Alignment));
3507 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3508 return SDValue(E, 0);
3509 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3510 new (N) StoreSDNode(Ops, VTs, ISD::UNINDEXED, false,
3511 VT, SV, SVOffset, Alignment, isVolatile);
3512 CSEMap.InsertNode(N, IP);
3513 AllNodes.push_back(N);
3514 return SDValue(N, 0);
3517 SDValue SelectionDAG::getTruncStore(SDValue Chain, SDValue Val,
3518 SDValue Ptr, const Value *SV,
3519 int SVOffset, MVT SVT,
3520 bool isVolatile, unsigned Alignment) {
3521 MVT VT = Val.getValueType();
3524 return getStore(Chain, Val, Ptr, SV, SVOffset, isVolatile, Alignment);
3526 assert(VT.bitsGT(SVT) && "Not a truncation?");
3527 assert(VT.isInteger() == SVT.isInteger() &&
3528 "Can't do FP-INT conversion!");
3530 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3531 Alignment = getMVTAlignment(VT);
3533 SDVTList VTs = getVTList(MVT::Other);
3534 SDValue Undef = getNode(ISD::UNDEF, Ptr.getValueType());
3535 SDValue Ops[] = { Chain, Val, Ptr, Undef };
3536 FoldingSetNodeID ID;
3537 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3538 ID.AddInteger(ISD::UNINDEXED);
3540 ID.AddInteger(SVT.getRawBits());
3541 ID.AddInteger(encodeMemSDNodeFlags(isVolatile, Alignment));
3543 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3544 return SDValue(E, 0);
3545 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3546 new (N) StoreSDNode(Ops, VTs, ISD::UNINDEXED, true,
3547 SVT, SV, SVOffset, Alignment, isVolatile);
3548 CSEMap.InsertNode(N, IP);
3549 AllNodes.push_back(N);
3550 return SDValue(N, 0);
3554 SelectionDAG::getIndexedStore(SDValue OrigStore, SDValue Base,
3555 SDValue Offset, ISD::MemIndexedMode AM) {
3556 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
3557 assert(ST->getOffset().getOpcode() == ISD::UNDEF &&
3558 "Store is already a indexed store!");
3559 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
3560 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
3561 FoldingSetNodeID ID;
3562 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4);
3564 ID.AddInteger(ST->isTruncatingStore());
3565 ID.AddInteger(ST->getMemoryVT().getRawBits());
3566 ID.AddInteger(ST->getRawFlags());
3568 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3569 return SDValue(E, 0);
3570 SDNode *N = NodeAllocator.Allocate<StoreSDNode>();
3571 new (N) StoreSDNode(Ops, VTs, AM,
3572 ST->isTruncatingStore(), ST->getMemoryVT(),
3573 ST->getSrcValue(), ST->getSrcValueOffset(),
3574 ST->getAlignment(), ST->isVolatile());
3575 CSEMap.InsertNode(N, IP);
3576 AllNodes.push_back(N);
3577 return SDValue(N, 0);
3580 SDValue SelectionDAG::getVAArg(MVT VT,
3581 SDValue Chain, SDValue Ptr,
3583 SDValue Ops[] = { Chain, Ptr, SV };
3584 return getNode(ISD::VAARG, getVTList(VT, MVT::Other), Ops, 3);
3587 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
3588 const SDUse *Ops, unsigned NumOps) {
3590 case 0: return getNode(Opcode, VT);
3591 case 1: return getNode(Opcode, VT, Ops[0]);
3592 case 2: return getNode(Opcode, VT, Ops[0], Ops[1]);
3593 case 3: return getNode(Opcode, VT, Ops[0], Ops[1], Ops[2]);
3597 // Copy from an SDUse array into an SDValue array for use with
3598 // the regular getNode logic.
3599 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps);
3600 return getNode(Opcode, VT, &NewOps[0], NumOps);
3603 SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
3604 const SDValue *Ops, unsigned NumOps) {
3606 case 0: return getNode(Opcode, VT);
3607 case 1: return getNode(Opcode, VT, Ops[0]);
3608 case 2: return getNode(Opcode, VT, Ops[0], Ops[1]);
3609 case 3: return getNode(Opcode, VT, Ops[0], Ops[1], Ops[2]);
3615 case ISD::SELECT_CC: {
3616 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
3617 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
3618 "LHS and RHS of condition must have same type!");
3619 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
3620 "True and False arms of SelectCC must have same type!");
3621 assert(Ops[2].getValueType() == VT &&
3622 "select_cc node must be of same type as true and false value!");
3626 assert(NumOps == 5 && "BR_CC takes 5 operands!");
3627 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
3628 "LHS/RHS of comparison should match types!");
3635 SDVTList VTs = getVTList(VT);
3636 if (VT != MVT::Flag) {
3637 FoldingSetNodeID ID;
3638 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps);
3640 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3641 return SDValue(E, 0);
3642 N = NodeAllocator.Allocate<SDNode>();
3643 new (N) SDNode(Opcode, VTs, Ops, NumOps);
3644 CSEMap.InsertNode(N, IP);
3646 N = NodeAllocator.Allocate<SDNode>();
3647 new (N) SDNode(Opcode, VTs, Ops, NumOps);
3649 AllNodes.push_back(N);
3653 return SDValue(N, 0);
3656 SDValue SelectionDAG::getNode(unsigned Opcode,
3657 const std::vector<MVT> &ResultTys,
3658 const SDValue *Ops, unsigned NumOps) {
3659 return getNode(Opcode, getNodeValueTypes(ResultTys), ResultTys.size(),
3663 SDValue SelectionDAG::getNode(unsigned Opcode,
3664 const MVT *VTs, unsigned NumVTs,
3665 const SDValue *Ops, unsigned NumOps) {
3667 return getNode(Opcode, VTs[0], Ops, NumOps);
3668 return getNode(Opcode, makeVTList(VTs, NumVTs), Ops, NumOps);
3671 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList,
3672 const SDValue *Ops, unsigned NumOps) {
3673 if (VTList.NumVTs == 1)
3674 return getNode(Opcode, VTList.VTs[0], Ops, NumOps);
3677 // FIXME: figure out how to safely handle things like
3678 // int foo(int x) { return 1 << (x & 255); }
3679 // int bar() { return foo(256); }
3681 case ISD::SRA_PARTS:
3682 case ISD::SRL_PARTS:
3683 case ISD::SHL_PARTS:
3684 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3685 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
3686 return getNode(Opcode, VT, N1, N2, N3.getOperand(0));
3687 else if (N3.getOpcode() == ISD::AND)
3688 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
3689 // If the and is only masking out bits that cannot effect the shift,
3690 // eliminate the and.
3691 unsigned NumBits = VT.getSizeInBits()*2;
3692 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
3693 return getNode(Opcode, VT, N1, N2, N3.getOperand(0));
3699 // Memoize the node unless it returns a flag.
3701 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
3702 FoldingSetNodeID ID;
3703 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
3705 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
3706 return SDValue(E, 0);
3708 N = NodeAllocator.Allocate<UnarySDNode>();
3709 new (N) UnarySDNode(Opcode, VTList, Ops[0]);
3710 } else if (NumOps == 2) {
3711 N = NodeAllocator.Allocate<BinarySDNode>();
3712 new (N) BinarySDNode(Opcode, VTList, Ops[0], Ops[1]);
3713 } else if (NumOps == 3) {
3714 N = NodeAllocator.Allocate<TernarySDNode>();
3715 new (N) TernarySDNode(Opcode, VTList, Ops[0], Ops[1], Ops[2]);
3717 N = NodeAllocator.Allocate<SDNode>();
3718 new (N) SDNode(Opcode, VTList, Ops, NumOps);
3720 CSEMap.InsertNode(N, IP);
3723 N = NodeAllocator.Allocate<UnarySDNode>();
3724 new (N) UnarySDNode(Opcode, VTList, Ops[0]);
3725 } else if (NumOps == 2) {
3726 N = NodeAllocator.Allocate<BinarySDNode>();
3727 new (N) BinarySDNode(Opcode, VTList, Ops[0], Ops[1]);
3728 } else if (NumOps == 3) {
3729 N = NodeAllocator.Allocate<TernarySDNode>();
3730 new (N) TernarySDNode(Opcode, VTList, Ops[0], Ops[1], Ops[2]);
3732 N = NodeAllocator.Allocate<SDNode>();
3733 new (N) SDNode(Opcode, VTList, Ops, NumOps);
3736 AllNodes.push_back(N);
3740 return SDValue(N, 0);
3743 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList) {
3744 return getNode(Opcode, VTList, 0, 0);
3747 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList,
3749 SDValue Ops[] = { N1 };
3750 return getNode(Opcode, VTList, Ops, 1);
3753 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList,
3754 SDValue N1, SDValue N2) {
3755 SDValue Ops[] = { N1, N2 };
3756 return getNode(Opcode, VTList, Ops, 2);
3759 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList,
3760 SDValue N1, SDValue N2, SDValue N3) {
3761 SDValue Ops[] = { N1, N2, N3 };
3762 return getNode(Opcode, VTList, Ops, 3);
3765 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList,
3766 SDValue N1, SDValue N2, SDValue N3,
3768 SDValue Ops[] = { N1, N2, N3, N4 };
3769 return getNode(Opcode, VTList, Ops, 4);
3772 SDValue SelectionDAG::getNode(unsigned Opcode, SDVTList VTList,
3773 SDValue N1, SDValue N2, SDValue N3,
3774 SDValue N4, SDValue N5) {
3775 SDValue Ops[] = { N1, N2, N3, N4, N5 };
3776 return getNode(Opcode, VTList, Ops, 5);
3779 SDVTList SelectionDAG::getVTList(MVT VT) {
3780 return makeVTList(SDNode::getValueTypeList(VT), 1);
3783 SDVTList SelectionDAG::getVTList(MVT VT1, MVT VT2) {
3784 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
3785 E = VTList.rend(); I != E; ++I)
3786 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2)
3789 MVT *Array = Allocator.Allocate<MVT>(2);
3792 SDVTList Result = makeVTList(Array, 2);
3793 VTList.push_back(Result);
3797 SDVTList SelectionDAG::getVTList(MVT VT1, MVT VT2, MVT VT3) {
3798 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
3799 E = VTList.rend(); I != E; ++I)
3800 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 &&
3804 MVT *Array = Allocator.Allocate<MVT>(3);
3808 SDVTList Result = makeVTList(Array, 3);
3809 VTList.push_back(Result);
3813 SDVTList SelectionDAG::getVTList(const MVT *VTs, unsigned NumVTs) {
3815 case 0: assert(0 && "Cannot have nodes without results!");
3816 case 1: return getVTList(VTs[0]);
3817 case 2: return getVTList(VTs[0], VTs[1]);
3818 case 3: return getVTList(VTs[0], VTs[1], VTs[2]);
3822 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(),
3823 E = VTList.rend(); I != E; ++I) {
3824 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1])
3827 bool NoMatch = false;
3828 for (unsigned i = 2; i != NumVTs; ++i)
3829 if (VTs[i] != I->VTs[i]) {
3837 MVT *Array = Allocator.Allocate<MVT>(NumVTs);
3838 std::copy(VTs, VTs+NumVTs, Array);
3839 SDVTList Result = makeVTList(Array, NumVTs);
3840 VTList.push_back(Result);
3845 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
3846 /// specified operands. If the resultant node already exists in the DAG,
3847 /// this does not modify the specified node, instead it returns the node that
3848 /// already exists. If the resultant node does not exist in the DAG, the
3849 /// input node is returned. As a degenerate case, if you specify the same
3850 /// input operands as the node already has, the input node is returned.
3851 SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) {
3852 SDNode *N = InN.getNode();
3853 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
3855 // Check to see if there is no change.
3856 if (Op == N->getOperand(0)) return InN;
3858 // See if the modified node already exists.
3859 void *InsertPos = 0;
3860 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
3861 return SDValue(Existing, InN.getResNo());
3863 // Nope it doesn't. Remove the node from its current place in the maps.
3865 if (!RemoveNodeFromCSEMaps(N))
3868 // Now we update the operands.
3869 N->OperandList[0].getVal()->removeUser(0, N);
3870 N->OperandList[0] = Op;
3871 N->OperandList[0].setUser(N);
3872 Op.getNode()->addUser(0, N);
3874 // If this gets put into a CSE map, add it.
3875 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
3879 SDValue SelectionDAG::
3880 UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) {
3881 SDNode *N = InN.getNode();
3882 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
3884 // Check to see if there is no change.
3885 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
3886 return InN; // No operands changed, just return the input node.
3888 // See if the modified node already exists.
3889 void *InsertPos = 0;
3890 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
3891 return SDValue(Existing, InN.getResNo());
3893 // Nope it doesn't. Remove the node from its current place in the maps.
3895 if (!RemoveNodeFromCSEMaps(N))
3898 // Now we update the operands.
3899 if (N->OperandList[0] != Op1) {
3900 N->OperandList[0].getVal()->removeUser(0, N);
3901 N->OperandList[0] = Op1;
3902 N->OperandList[0].setUser(N);
3903 Op1.getNode()->addUser(0, N);
3905 if (N->OperandList[1] != Op2) {
3906 N->OperandList[1].getVal()->removeUser(1, N);
3907 N->OperandList[1] = Op2;
3908 N->OperandList[1].setUser(N);
3909 Op2.getNode()->addUser(1, N);
3912 // If this gets put into a CSE map, add it.
3913 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
3917 SDValue SelectionDAG::
3918 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) {
3919 SDValue Ops[] = { Op1, Op2, Op3 };
3920 return UpdateNodeOperands(N, Ops, 3);
3923 SDValue SelectionDAG::
3924 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
3925 SDValue Op3, SDValue Op4) {
3926 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
3927 return UpdateNodeOperands(N, Ops, 4);
3930 SDValue SelectionDAG::
3931 UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2,
3932 SDValue Op3, SDValue Op4, SDValue Op5) {
3933 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
3934 return UpdateNodeOperands(N, Ops, 5);
3937 SDValue SelectionDAG::
3938 UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) {
3939 SDNode *N = InN.getNode();
3940 assert(N->getNumOperands() == NumOps &&
3941 "Update with wrong number of operands");
3943 // Check to see if there is no change.
3944 bool AnyChange = false;
3945 for (unsigned i = 0; i != NumOps; ++i) {
3946 if (Ops[i] != N->getOperand(i)) {
3952 // No operands changed, just return the input node.
3953 if (!AnyChange) return InN;
3955 // See if the modified node already exists.
3956 void *InsertPos = 0;
3957 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos))
3958 return SDValue(Existing, InN.getResNo());
3960 // Nope it doesn't. Remove the node from its current place in the maps.
3962 if (!RemoveNodeFromCSEMaps(N))
3965 // Now we update the operands.
3966 for (unsigned i = 0; i != NumOps; ++i) {
3967 if (N->OperandList[i] != Ops[i]) {
3968 N->OperandList[i].getVal()->removeUser(i, N);
3969 N->OperandList[i] = Ops[i];
3970 N->OperandList[i].setUser(N);
3971 Ops[i].getNode()->addUser(i, N);
3975 // If this gets put into a CSE map, add it.
3976 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
3980 /// DropOperands - Release the operands and set this node to have
3982 void SDNode::DropOperands() {
3983 // Unlike the code in MorphNodeTo that does this, we don't need to
3984 // watch for dead nodes here.
3985 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
3986 I->getVal()->removeUser(std::distance(op_begin(), I), this);
3991 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
3994 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
3996 SDVTList VTs = getVTList(VT);
3997 return SelectNodeTo(N, MachineOpc, VTs, 0, 0);
4000 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4001 MVT VT, SDValue Op1) {
4002 SDVTList VTs = getVTList(VT);
4003 SDValue Ops[] = { Op1 };
4004 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4007 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4008 MVT VT, SDValue Op1,
4010 SDVTList VTs = getVTList(VT);
4011 SDValue Ops[] = { Op1, Op2 };
4012 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4015 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4016 MVT VT, SDValue Op1,
4017 SDValue Op2, SDValue Op3) {
4018 SDVTList VTs = getVTList(VT);
4019 SDValue Ops[] = { Op1, Op2, Op3 };
4020 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4023 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4024 MVT VT, const SDValue *Ops,
4026 SDVTList VTs = getVTList(VT);
4027 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4030 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4031 MVT VT1, MVT VT2, const SDValue *Ops,
4033 SDVTList VTs = getVTList(VT1, VT2);
4034 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4037 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4039 SDVTList VTs = getVTList(VT1, VT2);
4040 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0);
4043 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4044 MVT VT1, MVT VT2, MVT VT3,
4045 const SDValue *Ops, unsigned NumOps) {
4046 SDVTList VTs = getVTList(VT1, VT2, VT3);
4047 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps);
4050 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4053 SDVTList VTs = getVTList(VT1, VT2);
4054 SDValue Ops[] = { Op1 };
4055 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1);
4058 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4060 SDValue Op1, SDValue Op2) {
4061 SDVTList VTs = getVTList(VT1, VT2);
4062 SDValue Ops[] = { Op1, Op2 };
4063 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2);
4066 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4068 SDValue Op1, SDValue Op2,
4070 SDVTList VTs = getVTList(VT1, VT2);
4071 SDValue Ops[] = { Op1, Op2, Op3 };
4072 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3);
4075 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
4076 SDVTList VTs, const SDValue *Ops,
4078 return MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps);
4081 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4083 SDVTList VTs = getVTList(VT);
4084 return MorphNodeTo(N, Opc, VTs, 0, 0);
4087 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4088 MVT VT, SDValue Op1) {
4089 SDVTList VTs = getVTList(VT);
4090 SDValue Ops[] = { Op1 };
4091 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4094 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4095 MVT VT, SDValue Op1,
4097 SDVTList VTs = getVTList(VT);
4098 SDValue Ops[] = { Op1, Op2 };
4099 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4102 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4103 MVT VT, SDValue Op1,
4104 SDValue Op2, SDValue Op3) {
4105 SDVTList VTs = getVTList(VT);
4106 SDValue Ops[] = { Op1, Op2, Op3 };
4107 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4110 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4111 MVT VT, const SDValue *Ops,
4113 SDVTList VTs = getVTList(VT);
4114 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4117 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4118 MVT VT1, MVT VT2, const SDValue *Ops,
4120 SDVTList VTs = getVTList(VT1, VT2);
4121 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4124 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4126 SDVTList VTs = getVTList(VT1, VT2);
4127 return MorphNodeTo(N, Opc, VTs, (SDValue *)0, 0);
4130 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4131 MVT VT1, MVT VT2, MVT VT3,
4132 const SDValue *Ops, unsigned NumOps) {
4133 SDVTList VTs = getVTList(VT1, VT2, VT3);
4134 return MorphNodeTo(N, Opc, VTs, Ops, NumOps);
4137 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4140 SDVTList VTs = getVTList(VT1, VT2);
4141 SDValue Ops[] = { Op1 };
4142 return MorphNodeTo(N, Opc, VTs, Ops, 1);
4145 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4147 SDValue Op1, SDValue Op2) {
4148 SDVTList VTs = getVTList(VT1, VT2);
4149 SDValue Ops[] = { Op1, Op2 };
4150 return MorphNodeTo(N, Opc, VTs, Ops, 2);
4153 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4155 SDValue Op1, SDValue Op2,
4157 SDVTList VTs = getVTList(VT1, VT2);
4158 SDValue Ops[] = { Op1, Op2, Op3 };
4159 return MorphNodeTo(N, Opc, VTs, Ops, 3);
4162 /// MorphNodeTo - These *mutate* the specified node to have the specified
4163 /// return type, opcode, and operands.
4165 /// Note that MorphNodeTo returns the resultant node. If there is already a
4166 /// node of the specified opcode and operands, it returns that node instead of
4167 /// the current one.
4169 /// Using MorphNodeTo is faster than creating a new node and swapping it in
4170 /// with ReplaceAllUsesWith both because it often avoids allocating a new
4171 /// node, and because it doesn't require CSE recalculation for any of
4172 /// the node's users.
4174 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
4175 SDVTList VTs, const SDValue *Ops,
4177 // If an identical node already exists, use it.
4179 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) {
4180 FoldingSetNodeID ID;
4181 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps);
4182 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP))
4186 if (!RemoveNodeFromCSEMaps(N))
4189 // Start the morphing.
4191 N->ValueList = VTs.VTs;
4192 N->NumValues = VTs.NumVTs;
4194 // Clear the operands list, updating used nodes to remove this from their
4195 // use list. Keep track of any operands that become dead as a result.
4196 SmallPtrSet<SDNode*, 16> DeadNodeSet;
4197 for (SDNode::op_iterator B = N->op_begin(), I = B, E = N->op_end();
4199 SDNode *Used = I->getVal();
4200 Used->removeUser(std::distance(B, I), N);
4201 if (Used->use_empty())
4202 DeadNodeSet.insert(Used);
4205 // If NumOps is larger than the # of operands we currently have, reallocate
4206 // the operand list.
4207 if (NumOps > N->NumOperands) {
4208 if (N->OperandsNeedDelete)
4209 delete[] N->OperandList;
4211 if (N->isMachineOpcode()) {
4212 // We're creating a final node that will live unmorphed for the
4213 // remainder of the current SelectionDAG iteration, so we can allocate
4214 // the operands directly out of a pool with no recycling metadata.
4215 N->OperandList = OperandAllocator.Allocate<SDUse>(NumOps);
4216 N->OperandsNeedDelete = false;
4218 N->OperandList = new SDUse[NumOps];
4219 N->OperandsNeedDelete = true;
4223 // Assign the new operands.
4224 N->NumOperands = NumOps;
4225 for (unsigned i = 0, e = NumOps; i != e; ++i) {
4226 N->OperandList[i] = Ops[i];
4227 N->OperandList[i].setUser(N);
4228 SDNode *ToUse = N->OperandList[i].getVal();
4229 ToUse->addUser(i, N);
4232 // Delete any nodes that are still dead after adding the uses for the
4234 SmallVector<SDNode *, 16> DeadNodes;
4235 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(),
4236 E = DeadNodeSet.end(); I != E; ++I)
4237 if ((*I)->use_empty())
4238 DeadNodes.push_back(*I);
4239 RemoveDeadNodes(DeadNodes);
4242 CSEMap.InsertNode(N, IP); // Memoize the new node.
4247 /// getTargetNode - These are used for target selectors to create a new node
4248 /// with specified return type(s), target opcode, and operands.
4250 /// Note that getTargetNode returns the resultant node. If there is already a
4251 /// node of the specified opcode and operands, it returns that node instead of
4252 /// the current one.
4253 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT) {
4254 return getNode(~Opcode, VT).getNode();
4256 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT, SDValue Op1) {
4257 return getNode(~Opcode, VT, Op1).getNode();
4259 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT,
4260 SDValue Op1, SDValue Op2) {
4261 return getNode(~Opcode, VT, Op1, Op2).getNode();
4263 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT,
4264 SDValue Op1, SDValue Op2,
4266 return getNode(~Opcode, VT, Op1, Op2, Op3).getNode();
4268 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT,
4269 const SDValue *Ops, unsigned NumOps) {
4270 return getNode(~Opcode, VT, Ops, NumOps).getNode();
4272 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1, MVT VT2) {
4273 const MVT *VTs = getNodeValueTypes(VT1, VT2);
4275 return getNode(~Opcode, VTs, 2, &Op, 0).getNode();
4277 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1,
4278 MVT VT2, SDValue Op1) {
4279 const MVT *VTs = getNodeValueTypes(VT1, VT2);
4280 return getNode(~Opcode, VTs, 2, &Op1, 1).getNode();
4282 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1,
4283 MVT VT2, SDValue Op1,
4285 const MVT *VTs = getNodeValueTypes(VT1, VT2);
4286 SDValue Ops[] = { Op1, Op2 };
4287 return getNode(~Opcode, VTs, 2, Ops, 2).getNode();
4289 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1,
4290 MVT VT2, SDValue Op1,
4291 SDValue Op2, SDValue Op3) {
4292 const MVT *VTs = getNodeValueTypes(VT1, VT2);
4293 SDValue Ops[] = { Op1, Op2, Op3 };
4294 return getNode(~Opcode, VTs, 2, Ops, 3).getNode();
4296 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1, MVT VT2,
4297 const SDValue *Ops, unsigned NumOps) {
4298 const MVT *VTs = getNodeValueTypes(VT1, VT2);
4299 return getNode(~Opcode, VTs, 2, Ops, NumOps).getNode();
4301 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1, MVT VT2, MVT VT3,
4302 SDValue Op1, SDValue Op2) {
4303 const MVT *VTs = getNodeValueTypes(VT1, VT2, VT3);
4304 SDValue Ops[] = { Op1, Op2 };
4305 return getNode(~Opcode, VTs, 3, Ops, 2).getNode();
4307 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1, MVT VT2, MVT VT3,
4308 SDValue Op1, SDValue Op2,
4310 const MVT *VTs = getNodeValueTypes(VT1, VT2, VT3);
4311 SDValue Ops[] = { Op1, Op2, Op3 };
4312 return getNode(~Opcode, VTs, 3, Ops, 3).getNode();
4314 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1, MVT VT2, MVT VT3,
4315 const SDValue *Ops, unsigned NumOps) {
4316 const MVT *VTs = getNodeValueTypes(VT1, VT2, VT3);
4317 return getNode(~Opcode, VTs, 3, Ops, NumOps).getNode();
4319 SDNode *SelectionDAG::getTargetNode(unsigned Opcode, MVT VT1,
4320 MVT VT2, MVT VT3, MVT VT4,
4321 const SDValue *Ops, unsigned NumOps) {
4322 std::vector<MVT> VTList;
4323 VTList.push_back(VT1);
4324 VTList.push_back(VT2);
4325 VTList.push_back(VT3);
4326 VTList.push_back(VT4);
4327 const MVT *VTs = getNodeValueTypes(VTList);
4328 return getNode(~Opcode, VTs, 4, Ops, NumOps).getNode();
4330 SDNode *SelectionDAG::getTargetNode(unsigned Opcode,
4331 const std::vector<MVT> &ResultTys,
4332 const SDValue *Ops, unsigned NumOps) {
4333 const MVT *VTs = getNodeValueTypes(ResultTys);
4334 return getNode(~Opcode, VTs, ResultTys.size(),
4335 Ops, NumOps).getNode();
4338 /// getNodeIfExists - Get the specified node if it's already available, or
4339 /// else return NULL.
4340 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
4341 const SDValue *Ops, unsigned NumOps) {
4342 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) {
4343 FoldingSetNodeID ID;
4344 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps);
4346 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP))
4353 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4354 /// This can cause recursive merging of nodes in the DAG.
4356 /// This version assumes From has a single result value.
4358 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To,
4359 DAGUpdateListener *UpdateListener) {
4360 SDNode *From = FromN.getNode();
4361 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
4362 "Cannot replace with this method!");
4363 assert(From != To.getNode() && "Cannot replace uses of with self");
4365 while (!From->use_empty()) {
4366 SDNode::use_iterator UI = From->use_begin();
4369 // This node is about to morph, remove its old self from the CSE maps.
4370 RemoveNodeFromCSEMaps(U);
4372 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end();
4373 I != E; ++I, ++operandNum)
4374 if (I->getVal() == From) {
4375 From->removeUser(operandNum, U);
4378 To.getNode()->addUser(operandNum, U);
4381 // Now that we have modified U, add it back to the CSE maps. If it already
4382 // exists there, recursively merge the results together.
4383 if (SDNode *Existing = AddNonLeafNodeToCSEMaps(U)) {
4384 ReplaceAllUsesWith(U, Existing, UpdateListener);
4385 // U is now dead. Inform the listener if it exists and delete it.
4387 UpdateListener->NodeDeleted(U, Existing);
4388 DeleteNodeNotInCSEMaps(U);
4390 // If the node doesn't already exist, we updated it. Inform a listener if
4393 UpdateListener->NodeUpdated(U);
4398 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4399 /// This can cause recursive merging of nodes in the DAG.
4401 /// This version assumes From/To have matching types and numbers of result
4404 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To,
4405 DAGUpdateListener *UpdateListener) {
4406 assert(From->getVTList().VTs == To->getVTList().VTs &&
4407 From->getNumValues() == To->getNumValues() &&
4408 "Cannot use this version of ReplaceAllUsesWith!");
4410 // Handle the trivial case.
4414 while (!From->use_empty()) {
4415 SDNode::use_iterator UI = From->use_begin();
4418 // This node is about to morph, remove its old self from the CSE maps.
4419 RemoveNodeFromCSEMaps(U);
4421 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end();
4422 I != E; ++I, ++operandNum)
4423 if (I->getVal() == From) {
4424 From->removeUser(operandNum, U);
4425 I->getSDValue().setNode(To);
4426 To->addUser(operandNum, U);
4429 // Now that we have modified U, add it back to the CSE maps. If it already
4430 // exists there, recursively merge the results together.
4431 if (SDNode *Existing = AddNonLeafNodeToCSEMaps(U)) {
4432 ReplaceAllUsesWith(U, Existing, UpdateListener);
4433 // U is now dead. Inform the listener if it exists and delete it.
4435 UpdateListener->NodeDeleted(U, Existing);
4436 DeleteNodeNotInCSEMaps(U);
4438 // If the node doesn't already exist, we updated it. Inform a listener if
4441 UpdateListener->NodeUpdated(U);
4446 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
4447 /// This can cause recursive merging of nodes in the DAG.
4449 /// This version can replace From with any result values. To must match the
4450 /// number and types of values returned by From.
4451 void SelectionDAG::ReplaceAllUsesWith(SDNode *From,
4453 DAGUpdateListener *UpdateListener) {
4454 if (From->getNumValues() == 1) // Handle the simple case efficiently.
4455 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener);
4457 while (!From->use_empty()) {
4458 SDNode::use_iterator UI = From->use_begin();
4461 // This node is about to morph, remove its old self from the CSE maps.
4462 RemoveNodeFromCSEMaps(U);
4464 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end();
4465 I != E; ++I, ++operandNum)
4466 if (I->getVal() == From) {
4467 const SDValue &ToOp = To[I->getSDValue().getResNo()];
4468 From->removeUser(operandNum, U);
4471 ToOp.getNode()->addUser(operandNum, U);
4474 // Now that we have modified U, add it back to the CSE maps. If it already
4475 // exists there, recursively merge the results together.
4476 if (SDNode *Existing = AddNonLeafNodeToCSEMaps(U)) {
4477 ReplaceAllUsesWith(U, Existing, UpdateListener);
4478 // U is now dead. Inform the listener if it exists and delete it.
4480 UpdateListener->NodeDeleted(U, Existing);
4481 DeleteNodeNotInCSEMaps(U);
4483 // If the node doesn't already exist, we updated it. Inform a listener if
4486 UpdateListener->NodeUpdated(U);
4491 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
4492 /// uses of other values produced by From.getVal() alone. The Deleted vector is
4493 /// handled the same way as for ReplaceAllUsesWith.
4494 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To,
4495 DAGUpdateListener *UpdateListener){
4496 // Handle the really simple, really trivial case efficiently.
4497 if (From == To) return;
4499 // Handle the simple, trivial, case efficiently.
4500 if (From.getNode()->getNumValues() == 1) {
4501 ReplaceAllUsesWith(From, To, UpdateListener);
4505 // Get all of the users of From.getNode(). We want these in a nice,
4506 // deterministically ordered and uniqued set, so we use a SmallSetVector.
4507 SmallSetVector<SDNode*, 16> Users(From.getNode()->use_begin(), From.getNode()->use_end());
4509 while (!Users.empty()) {
4510 // We know that this user uses some value of From. If it is the right
4511 // value, update it.
4512 SDNode *User = Users.back();
4515 // Scan for an operand that matches From.
4516 SDNode::op_iterator Op = User->op_begin(), E = User->op_end();
4517 for (; Op != E; ++Op)
4518 if (*Op == From) break;
4520 // If there are no matches, the user must use some other result of From.
4521 if (Op == E) continue;
4523 // Okay, we know this user needs to be updated. Remove its old self
4524 // from the CSE maps.
4525 RemoveNodeFromCSEMaps(User);
4527 // Update all operands that match "From" in case there are multiple uses.
4528 for (; Op != E; ++Op) {
4530 From.getNode()->removeUser(Op-User->op_begin(), User);
4533 To.getNode()->addUser(Op-User->op_begin(), User);
4537 // Now that we have modified User, add it back to the CSE maps. If it
4538 // already exists there, recursively merge the results together.
4539 SDNode *Existing = AddNonLeafNodeToCSEMaps(User);
4541 if (UpdateListener) UpdateListener->NodeUpdated(User);
4542 continue; // Continue on to next user.
4545 // If there was already an existing matching node, use ReplaceAllUsesWith
4546 // to replace the dead one with the existing one. This can cause
4547 // recursive merging of other unrelated nodes down the line.
4548 ReplaceAllUsesWith(User, Existing, UpdateListener);
4550 // User is now dead. Notify a listener if present.
4551 if (UpdateListener) UpdateListener->NodeDeleted(User, Existing);
4552 DeleteNodeNotInCSEMaps(User);
4556 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
4557 /// uses of other values produced by From.getVal() alone. The same value may
4558 /// appear in both the From and To list. The Deleted vector is
4559 /// handled the same way as for ReplaceAllUsesWith.
4560 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
4563 DAGUpdateListener *UpdateListener){
4564 // Handle the simple, trivial case efficiently.
4566 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener);
4568 SmallVector<std::pair<SDNode *, unsigned>, 16> Users;
4569 for (unsigned i = 0; i != Num; ++i)
4570 for (SDNode::use_iterator UI = From[i].getNode()->use_begin(),
4571 E = From[i].getNode()->use_end(); UI != E; ++UI)
4572 Users.push_back(std::make_pair(*UI, i));
4574 while (!Users.empty()) {
4575 // We know that this user uses some value of From. If it is the right
4576 // value, update it.
4577 SDNode *User = Users.back().first;
4578 unsigned i = Users.back().second;
4581 // Scan for an operand that matches From.
4582 SDNode::op_iterator Op = User->op_begin(), E = User->op_end();
4583 for (; Op != E; ++Op)
4584 if (*Op == From[i]) break;
4586 // If there are no matches, the user must use some other result of From.
4587 if (Op == E) continue;
4589 // Okay, we know this user needs to be updated. Remove its old self
4590 // from the CSE maps.
4591 RemoveNodeFromCSEMaps(User);
4593 // Update all operands that match "From" in case there are multiple uses.
4594 for (; Op != E; ++Op) {
4595 if (*Op == From[i]) {
4596 From[i].getNode()->removeUser(Op-User->op_begin(), User);
4599 To[i].getNode()->addUser(Op-User->op_begin(), User);
4603 // Now that we have modified User, add it back to the CSE maps. If it
4604 // already exists there, recursively merge the results together.
4605 SDNode *Existing = AddNonLeafNodeToCSEMaps(User);
4607 if (UpdateListener) UpdateListener->NodeUpdated(User);
4608 continue; // Continue on to next user.
4611 // If there was already an existing matching node, use ReplaceAllUsesWith
4612 // to replace the dead one with the existing one. This can cause
4613 // recursive merging of other unrelated nodes down the line.
4614 ReplaceAllUsesWith(User, Existing, UpdateListener);
4616 // User is now dead. Notify a listener if present.
4617 if (UpdateListener) UpdateListener->NodeDeleted(User, Existing);
4618 DeleteNodeNotInCSEMaps(User);
4622 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
4623 /// based on their topological order. It returns the maximum id and a vector
4624 /// of the SDNodes* in assigned order by reference.
4625 unsigned SelectionDAG::AssignTopologicalOrder() {
4627 unsigned DAGSize = 0;
4629 // SortedPos tracks the progress of the algorithm. Nodes before it are
4630 // sorted, nodes after it are unsorted. When the algorithm completes
4631 // it is at the end of the list.
4632 allnodes_iterator SortedPos = allnodes_begin();
4634 // Visit all the nodes. Add nodes with no operands to the TopOrder result
4635 // array immediately. Annotate nodes that do have operands with their
4636 // operand count. Before we do this, the Node Id fields of the nodes
4637 // may contain arbitrary values. After, the Node Id fields for nodes
4638 // before SortedPos will contain the topological sort index, and the
4639 // Node Id fields for nodes At SortedPos and after will contain the
4640 // count of outstanding operands.
4641 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) {
4643 unsigned Degree = N->getNumOperands();
4645 // A node with no uses, add it to the result array immediately.
4646 N->setNodeId(DAGSize++);
4647 allnodes_iterator Q = N;
4649 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
4652 // Temporarily use the Node Id as scratch space for the degree count.
4653 N->setNodeId(Degree);
4657 // Visit all the nodes. As we iterate, moves nodes into sorted order,
4658 // such that by the time the end is reached all nodes will be sorted.
4659 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) {
4661 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4664 unsigned Degree = P->getNodeId();
4667 // All of P's operands are sorted, so P may sorted now.
4668 P->setNodeId(DAGSize++);
4670 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
4673 // Update P's outstanding operand count.
4674 P->setNodeId(Degree);
4679 assert(SortedPos == AllNodes.end() &&
4680 "Topological sort incomplete!");
4681 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
4682 "First node in topological sort is not the entry token!");
4683 assert(AllNodes.front().getNodeId() == 0 &&
4684 "First node in topological sort has non-zero id!");
4685 assert(AllNodes.front().getNumOperands() == 0 &&
4686 "First node in topological sort has operands!");
4687 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
4688 "Last node in topologic sort has unexpected id!");
4689 assert(AllNodes.back().use_empty() &&
4690 "Last node in topologic sort has users!");
4691 assert(DAGSize == allnodes_size() && "TopOrder result count mismatch!");
4697 //===----------------------------------------------------------------------===//
4699 //===----------------------------------------------------------------------===//
4701 // Out-of-line virtual method to give class a home.
4702 void SDNode::ANCHOR() {}
4703 void UnarySDNode::ANCHOR() {}
4704 void BinarySDNode::ANCHOR() {}
4705 void TernarySDNode::ANCHOR() {}
4706 void HandleSDNode::ANCHOR() {}
4707 void ConstantSDNode::ANCHOR() {}
4708 void ConstantFPSDNode::ANCHOR() {}
4709 void GlobalAddressSDNode::ANCHOR() {}
4710 void FrameIndexSDNode::ANCHOR() {}
4711 void JumpTableSDNode::ANCHOR() {}
4712 void ConstantPoolSDNode::ANCHOR() {}
4713 void BasicBlockSDNode::ANCHOR() {}
4714 void SrcValueSDNode::ANCHOR() {}
4715 void MemOperandSDNode::ANCHOR() {}
4716 void RegisterSDNode::ANCHOR() {}
4717 void DbgStopPointSDNode::ANCHOR() {}
4718 void LabelSDNode::ANCHOR() {}
4719 void ExternalSymbolSDNode::ANCHOR() {}
4720 void CondCodeSDNode::ANCHOR() {}
4721 void ARG_FLAGSSDNode::ANCHOR() {}
4722 void VTSDNode::ANCHOR() {}
4723 void MemSDNode::ANCHOR() {}
4724 void LoadSDNode::ANCHOR() {}
4725 void StoreSDNode::ANCHOR() {}
4726 void AtomicSDNode::ANCHOR() {}
4727 void MemIntrinsicSDNode::ANCHOR() {}
4728 void CallSDNode::ANCHOR() {}
4730 HandleSDNode::~HandleSDNode() {
4734 GlobalAddressSDNode::GlobalAddressSDNode(bool isTarget, const GlobalValue *GA,
4736 : SDNode(isa<GlobalVariable>(GA) &&
4737 cast<GlobalVariable>(GA)->isThreadLocal() ?
4739 (isTarget ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress) :
4741 (isTarget ? ISD::TargetGlobalAddress : ISD::GlobalAddress),
4742 getSDVTList(VT)), Offset(o) {
4743 TheGlobal = const_cast<GlobalValue*>(GA);
4746 MemSDNode::MemSDNode(unsigned Opc, SDVTList VTs, MVT memvt,
4747 const Value *srcValue, int SVO,
4748 unsigned alignment, bool vol)
4749 : SDNode(Opc, VTs), MemoryVT(memvt), SrcValue(srcValue), SVOffset(SVO),
4750 Flags(encodeMemSDNodeFlags(vol, alignment)) {
4752 assert(isPowerOf2_32(alignment) && "Alignment is not a power of 2!");
4753 assert(getAlignment() == alignment && "Alignment representation error!");
4754 assert(isVolatile() == vol && "Volatile representation error!");
4757 MemSDNode::MemSDNode(unsigned Opc, SDVTList VTs, const SDValue *Ops,
4758 unsigned NumOps, MVT memvt, const Value *srcValue,
4759 int SVO, unsigned alignment, bool vol)
4760 : SDNode(Opc, VTs, Ops, NumOps),
4761 MemoryVT(memvt), SrcValue(srcValue), SVOffset(SVO),
4762 Flags(vol | ((Log2_32(alignment) + 1) << 1)) {
4763 assert(isPowerOf2_32(alignment) && "Alignment is not a power of 2!");
4764 assert(getAlignment() == alignment && "Alignment representation error!");
4765 assert(isVolatile() == vol && "Volatile representation error!");
4768 /// getMemOperand - Return a MachineMemOperand object describing the memory
4769 /// reference performed by this memory reference.
4770 MachineMemOperand MemSDNode::getMemOperand() const {
4772 if (isa<LoadSDNode>(this))
4773 Flags = MachineMemOperand::MOLoad;
4774 else if (isa<StoreSDNode>(this))
4775 Flags = MachineMemOperand::MOStore;
4776 else if (isa<AtomicSDNode>(this)) {
4777 Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
4780 const MemIntrinsicSDNode* MemIntrinNode = dyn_cast<MemIntrinsicSDNode>(this);
4781 assert(MemIntrinNode && "Unknown MemSDNode opcode!");
4782 if (MemIntrinNode->readMem()) Flags |= MachineMemOperand::MOLoad;
4783 if (MemIntrinNode->writeMem()) Flags |= MachineMemOperand::MOStore;
4786 int Size = (getMemoryVT().getSizeInBits() + 7) >> 3;
4787 if (isVolatile()) Flags |= MachineMemOperand::MOVolatile;
4789 // Check if the memory reference references a frame index
4790 const FrameIndexSDNode *FI =
4791 dyn_cast<const FrameIndexSDNode>(getBasePtr().getNode());
4792 if (!getSrcValue() && FI)
4793 return MachineMemOperand(PseudoSourceValue::getFixedStack(FI->getIndex()),
4794 Flags, 0, Size, getAlignment());
4796 return MachineMemOperand(getSrcValue(), Flags, getSrcValueOffset(),
4797 Size, getAlignment());
4800 /// Profile - Gather unique data for the node.
4802 void SDNode::Profile(FoldingSetNodeID &ID) const {
4803 AddNodeIDNode(ID, this);
4806 /// getValueTypeList - Return a pointer to the specified value type.
4808 const MVT *SDNode::getValueTypeList(MVT VT) {
4809 if (VT.isExtended()) {
4810 static std::set<MVT, MVT::compareRawBits> EVTs;
4811 return &(*EVTs.insert(VT).first);
4813 static MVT VTs[MVT::LAST_VALUETYPE];
4814 VTs[VT.getSimpleVT()] = VT;
4815 return &VTs[VT.getSimpleVT()];
4819 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
4820 /// indicated value. This method ignores uses of other values defined by this
4822 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
4823 assert(Value < getNumValues() && "Bad value!");
4825 // TODO: Only iterate over uses of a given value of the node
4826 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
4827 if (UI.getUse().getSDValue().getResNo() == Value) {
4834 // Found exactly the right number of uses?
4839 /// hasAnyUseOfValue - Return true if there are any use of the indicated
4840 /// value. This method ignores uses of other values defined by this operation.
4841 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
4842 assert(Value < getNumValues() && "Bad value!");
4844 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
4845 if (UI.getUse().getSDValue().getResNo() == Value)
4852 /// isOnlyUserOf - Return true if this node is the only use of N.
4854 bool SDNode::isOnlyUserOf(SDNode *N) const {
4856 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
4867 /// isOperand - Return true if this node is an operand of N.
4869 bool SDValue::isOperandOf(SDNode *N) const {
4870 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4871 if (*this == N->getOperand(i))
4876 bool SDNode::isOperandOf(SDNode *N) const {
4877 for (unsigned i = 0, e = N->NumOperands; i != e; ++i)
4878 if (this == N->OperandList[i].getVal())
4883 /// reachesChainWithoutSideEffects - Return true if this operand (which must
4884 /// be a chain) reaches the specified operand without crossing any
4885 /// side-effecting instructions. In practice, this looks through token
4886 /// factors and non-volatile loads. In order to remain efficient, this only
4887 /// looks a couple of nodes in, it does not do an exhaustive search.
4888 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
4889 unsigned Depth) const {
4890 if (*this == Dest) return true;
4892 // Don't search too deeply, we just want to be able to see through
4893 // TokenFactor's etc.
4894 if (Depth == 0) return false;
4896 // If this is a token factor, all inputs to the TF happen in parallel. If any
4897 // of the operands of the TF reach dest, then we can do the xform.
4898 if (getOpcode() == ISD::TokenFactor) {
4899 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
4900 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1))
4905 // Loads don't have side effects, look through them.
4906 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
4907 if (!Ld->isVolatile())
4908 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
4914 static void findPredecessor(SDNode *N, const SDNode *P, bool &found,
4915 SmallPtrSet<SDNode *, 32> &Visited) {
4916 if (found || !Visited.insert(N))
4919 for (unsigned i = 0, e = N->getNumOperands(); !found && i != e; ++i) {
4920 SDNode *Op = N->getOperand(i).getNode();
4925 findPredecessor(Op, P, found, Visited);
4929 /// isPredecessorOf - Return true if this node is a predecessor of N. This node
4930 /// is either an operand of N or it can be reached by recursively traversing
4931 /// up the operands.
4932 /// NOTE: this is an expensive method. Use it carefully.
4933 bool SDNode::isPredecessorOf(SDNode *N) const {
4934 SmallPtrSet<SDNode *, 32> Visited;
4936 findPredecessor(N, this, found, Visited);
4940 uint64_t SDNode::getConstantOperandVal(unsigned Num) const {
4941 assert(Num < NumOperands && "Invalid child # of SDNode!");
4942 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue();
4945 std::string SDNode::getOperationName(const SelectionDAG *G) const {
4946 switch (getOpcode()) {
4948 if (getOpcode() < ISD::BUILTIN_OP_END)
4949 return "<<Unknown DAG Node>>";
4950 if (isMachineOpcode()) {
4952 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo())
4953 if (getMachineOpcode() < TII->getNumOpcodes())
4954 return TII->get(getMachineOpcode()).getName();
4955 return "<<Unknown Machine Node>>";
4958 TargetLowering &TLI = G->getTargetLoweringInfo();
4959 const char *Name = TLI.getTargetNodeName(getOpcode());
4960 if (Name) return Name;
4961 return "<<Unknown Target Node>>";
4963 return "<<Unknown Node>>";
4966 case ISD::DELETED_NODE:
4967 return "<<Deleted Node!>>";
4969 case ISD::PREFETCH: return "Prefetch";
4970 case ISD::MEMBARRIER: return "MemBarrier";
4971 case ISD::ATOMIC_CMP_SWAP_8: return "AtomicCmpSwap8";
4972 case ISD::ATOMIC_SWAP_8: return "AtomicSwap8";
4973 case ISD::ATOMIC_LOAD_ADD_8: return "AtomicLoadAdd8";
4974 case ISD::ATOMIC_LOAD_SUB_8: return "AtomicLoadSub8";
4975 case ISD::ATOMIC_LOAD_AND_8: return "AtomicLoadAnd8";
4976 case ISD::ATOMIC_LOAD_OR_8: return "AtomicLoadOr8";
4977 case ISD::ATOMIC_LOAD_XOR_8: return "AtomicLoadXor8";
4978 case ISD::ATOMIC_LOAD_NAND_8: return "AtomicLoadNand8";
4979 case ISD::ATOMIC_LOAD_MIN_8: return "AtomicLoadMin8";
4980 case ISD::ATOMIC_LOAD_MAX_8: return "AtomicLoadMax8";
4981 case ISD::ATOMIC_LOAD_UMIN_8: return "AtomicLoadUMin8";
4982 case ISD::ATOMIC_LOAD_UMAX_8: return "AtomicLoadUMax8";
4983 case ISD::ATOMIC_CMP_SWAP_16: return "AtomicCmpSwap16";
4984 case ISD::ATOMIC_SWAP_16: return "AtomicSwap16";
4985 case ISD::ATOMIC_LOAD_ADD_16: return "AtomicLoadAdd16";
4986 case ISD::ATOMIC_LOAD_SUB_16: return "AtomicLoadSub16";
4987 case ISD::ATOMIC_LOAD_AND_16: return "AtomicLoadAnd16";
4988 case ISD::ATOMIC_LOAD_OR_16: return "AtomicLoadOr16";
4989 case ISD::ATOMIC_LOAD_XOR_16: return "AtomicLoadXor16";
4990 case ISD::ATOMIC_LOAD_NAND_16: return "AtomicLoadNand16";
4991 case ISD::ATOMIC_LOAD_MIN_16: return "AtomicLoadMin16";
4992 case ISD::ATOMIC_LOAD_MAX_16: return "AtomicLoadMax16";
4993 case ISD::ATOMIC_LOAD_UMIN_16: return "AtomicLoadUMin16";
4994 case ISD::ATOMIC_LOAD_UMAX_16: return "AtomicLoadUMax16";
4995 case ISD::ATOMIC_CMP_SWAP_32: return "AtomicCmpSwap32";
4996 case ISD::ATOMIC_SWAP_32: return "AtomicSwap32";
4997 case ISD::ATOMIC_LOAD_ADD_32: return "AtomicLoadAdd32";
4998 case ISD::ATOMIC_LOAD_SUB_32: return "AtomicLoadSub32";
4999 case ISD::ATOMIC_LOAD_AND_32: return "AtomicLoadAnd32";
5000 case ISD::ATOMIC_LOAD_OR_32: return "AtomicLoadOr32";
5001 case ISD::ATOMIC_LOAD_XOR_32: return "AtomicLoadXor32";
5002 case ISD::ATOMIC_LOAD_NAND_32: return "AtomicLoadNand32";
5003 case ISD::ATOMIC_LOAD_MIN_32: return "AtomicLoadMin32";
5004 case ISD::ATOMIC_LOAD_MAX_32: return "AtomicLoadMax32";
5005 case ISD::ATOMIC_LOAD_UMIN_32: return "AtomicLoadUMin32";
5006 case ISD::ATOMIC_LOAD_UMAX_32: return "AtomicLoadUMax32";
5007 case ISD::ATOMIC_CMP_SWAP_64: return "AtomicCmpSwap64";
5008 case ISD::ATOMIC_SWAP_64: return "AtomicSwap64";
5009 case ISD::ATOMIC_LOAD_ADD_64: return "AtomicLoadAdd64";
5010 case ISD::ATOMIC_LOAD_SUB_64: return "AtomicLoadSub64";
5011 case ISD::ATOMIC_LOAD_AND_64: return "AtomicLoadAnd64";
5012 case ISD::ATOMIC_LOAD_OR_64: return "AtomicLoadOr64";
5013 case ISD::ATOMIC_LOAD_XOR_64: return "AtomicLoadXor64";
5014 case ISD::ATOMIC_LOAD_NAND_64: return "AtomicLoadNand64";
5015 case ISD::ATOMIC_LOAD_MIN_64: return "AtomicLoadMin64";
5016 case ISD::ATOMIC_LOAD_MAX_64: return "AtomicLoadMax64";
5017 case ISD::ATOMIC_LOAD_UMIN_64: return "AtomicLoadUMin64";
5018 case ISD::ATOMIC_LOAD_UMAX_64: return "AtomicLoadUMax64";
5019 case ISD::PCMARKER: return "PCMarker";
5020 case ISD::READCYCLECOUNTER: return "ReadCycleCounter";
5021 case ISD::SRCVALUE: return "SrcValue";
5022 case ISD::MEMOPERAND: return "MemOperand";
5023 case ISD::EntryToken: return "EntryToken";
5024 case ISD::TokenFactor: return "TokenFactor";
5025 case ISD::AssertSext: return "AssertSext";
5026 case ISD::AssertZext: return "AssertZext";
5028 case ISD::BasicBlock: return "BasicBlock";
5029 case ISD::ARG_FLAGS: return "ArgFlags";
5030 case ISD::VALUETYPE: return "ValueType";
5031 case ISD::Register: return "Register";
5033 case ISD::Constant: return "Constant";
5034 case ISD::ConstantFP: return "ConstantFP";
5035 case ISD::GlobalAddress: return "GlobalAddress";
5036 case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
5037 case ISD::FrameIndex: return "FrameIndex";
5038 case ISD::JumpTable: return "JumpTable";
5039 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE";
5040 case ISD::RETURNADDR: return "RETURNADDR";
5041 case ISD::FRAMEADDR: return "FRAMEADDR";
5042 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET";
5043 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR";
5044 case ISD::EHSELECTION: return "EHSELECTION";
5045 case ISD::EH_RETURN: return "EH_RETURN";
5046 case ISD::ConstantPool: return "ConstantPool";
5047 case ISD::ExternalSymbol: return "ExternalSymbol";
5048 case ISD::INTRINSIC_WO_CHAIN: {
5049 unsigned IID = cast<ConstantSDNode>(getOperand(0))->getZExtValue();
5050 return Intrinsic::getName((Intrinsic::ID)IID);
5052 case ISD::INTRINSIC_VOID:
5053 case ISD::INTRINSIC_W_CHAIN: {
5054 unsigned IID = cast<ConstantSDNode>(getOperand(1))->getZExtValue();
5055 return Intrinsic::getName((Intrinsic::ID)IID);
5058 case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
5059 case ISD::TargetConstant: return "TargetConstant";
5060 case ISD::TargetConstantFP:return "TargetConstantFP";
5061 case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
5062 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
5063 case ISD::TargetFrameIndex: return "TargetFrameIndex";
5064 case ISD::TargetJumpTable: return "TargetJumpTable";
5065 case ISD::TargetConstantPool: return "TargetConstantPool";
5066 case ISD::TargetExternalSymbol: return "TargetExternalSymbol";
5068 case ISD::CopyToReg: return "CopyToReg";
5069 case ISD::CopyFromReg: return "CopyFromReg";
5070 case ISD::UNDEF: return "undef";
5071 case ISD::MERGE_VALUES: return "merge_values";
5072 case ISD::INLINEASM: return "inlineasm";
5073 case ISD::DBG_LABEL: return "dbg_label";
5074 case ISD::EH_LABEL: return "eh_label";
5075 case ISD::DECLARE: return "declare";
5076 case ISD::HANDLENODE: return "handlenode";
5077 case ISD::FORMAL_ARGUMENTS: return "formal_arguments";
5078 case ISD::CALL: return "call";
5081 case ISD::FABS: return "fabs";
5082 case ISD::FNEG: return "fneg";
5083 case ISD::FSQRT: return "fsqrt";
5084 case ISD::FSIN: return "fsin";
5085 case ISD::FCOS: return "fcos";
5086 case ISD::FPOWI: return "fpowi";
5087 case ISD::FPOW: return "fpow";
5088 case ISD::FTRUNC: return "ftrunc";
5089 case ISD::FFLOOR: return "ffloor";
5090 case ISD::FCEIL: return "fceil";
5091 case ISD::FRINT: return "frint";
5092 case ISD::FNEARBYINT: return "fnearbyint";
5095 case ISD::ADD: return "add";
5096 case ISD::SUB: return "sub";
5097 case ISD::MUL: return "mul";
5098 case ISD::MULHU: return "mulhu";
5099 case ISD::MULHS: return "mulhs";
5100 case ISD::SDIV: return "sdiv";
5101 case ISD::UDIV: return "udiv";
5102 case ISD::SREM: return "srem";
5103 case ISD::UREM: return "urem";
5104 case ISD::SMUL_LOHI: return "smul_lohi";
5105 case ISD::UMUL_LOHI: return "umul_lohi";
5106 case ISD::SDIVREM: return "sdivrem";
5107 case ISD::UDIVREM: return "udivrem";
5108 case ISD::AND: return "and";
5109 case ISD::OR: return "or";
5110 case ISD::XOR: return "xor";
5111 case ISD::SHL: return "shl";
5112 case ISD::SRA: return "sra";
5113 case ISD::SRL: return "srl";
5114 case ISD::ROTL: return "rotl";
5115 case ISD::ROTR: return "rotr";
5116 case ISD::FADD: return "fadd";
5117 case ISD::FSUB: return "fsub";
5118 case ISD::FMUL: return "fmul";
5119 case ISD::FDIV: return "fdiv";
5120 case ISD::FREM: return "frem";
5121 case ISD::FCOPYSIGN: return "fcopysign";
5122 case ISD::FGETSIGN: return "fgetsign";
5124 case ISD::SETCC: return "setcc";
5125 case ISD::VSETCC: return "vsetcc";
5126 case ISD::SELECT: return "select";
5127 case ISD::SELECT_CC: return "select_cc";
5128 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt";
5129 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt";
5130 case ISD::CONCAT_VECTORS: return "concat_vectors";
5131 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector";
5132 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector";
5133 case ISD::VECTOR_SHUFFLE: return "vector_shuffle";
5134 case ISD::CARRY_FALSE: return "carry_false";
5135 case ISD::ADDC: return "addc";
5136 case ISD::ADDE: return "adde";
5137 case ISD::SUBC: return "subc";
5138 case ISD::SUBE: return "sube";
5139 case ISD::SHL_PARTS: return "shl_parts";
5140 case ISD::SRA_PARTS: return "sra_parts";
5141 case ISD::SRL_PARTS: return "srl_parts";
5143 case ISD::EXTRACT_SUBREG: return "extract_subreg";
5144 case ISD::INSERT_SUBREG: return "insert_subreg";
5146 // Conversion operators.
5147 case ISD::SIGN_EXTEND: return "sign_extend";
5148 case ISD::ZERO_EXTEND: return "zero_extend";
5149 case ISD::ANY_EXTEND: return "any_extend";
5150 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg";
5151 case ISD::TRUNCATE: return "truncate";
5152 case ISD::FP_ROUND: return "fp_round";
5153 case ISD::FLT_ROUNDS_: return "flt_rounds";
5154 case ISD::FP_ROUND_INREG: return "fp_round_inreg";
5155 case ISD::FP_EXTEND: return "fp_extend";
5157 case ISD::SINT_TO_FP: return "sint_to_fp";
5158 case ISD::UINT_TO_FP: return "uint_to_fp";
5159 case ISD::FP_TO_SINT: return "fp_to_sint";
5160 case ISD::FP_TO_UINT: return "fp_to_uint";
5161 case ISD::BIT_CONVERT: return "bit_convert";
5163 // Control flow instructions
5164 case ISD::BR: return "br";
5165 case ISD::BRIND: return "brind";
5166 case ISD::BR_JT: return "br_jt";
5167 case ISD::BRCOND: return "brcond";
5168 case ISD::BR_CC: return "br_cc";
5169 case ISD::RET: return "ret";
5170 case ISD::CALLSEQ_START: return "callseq_start";
5171 case ISD::CALLSEQ_END: return "callseq_end";
5174 case ISD::LOAD: return "load";
5175 case ISD::STORE: return "store";
5176 case ISD::VAARG: return "vaarg";
5177 case ISD::VACOPY: return "vacopy";
5178 case ISD::VAEND: return "vaend";
5179 case ISD::VASTART: return "vastart";
5180 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc";
5181 case ISD::EXTRACT_ELEMENT: return "extract_element";
5182 case ISD::BUILD_PAIR: return "build_pair";
5183 case ISD::STACKSAVE: return "stacksave";
5184 case ISD::STACKRESTORE: return "stackrestore";
5185 case ISD::TRAP: return "trap";
5188 case ISD::BSWAP: return "bswap";
5189 case ISD::CTPOP: return "ctpop";
5190 case ISD::CTTZ: return "cttz";
5191 case ISD::CTLZ: return "ctlz";
5194 case ISD::DBG_STOPPOINT: return "dbg_stoppoint";
5195 case ISD::DEBUG_LOC: return "debug_loc";
5198 case ISD::TRAMPOLINE: return "trampoline";
5201 switch (cast<CondCodeSDNode>(this)->get()) {
5202 default: assert(0 && "Unknown setcc condition!");
5203 case ISD::SETOEQ: return "setoeq";
5204 case ISD::SETOGT: return "setogt";
5205 case ISD::SETOGE: return "setoge";
5206 case ISD::SETOLT: return "setolt";
5207 case ISD::SETOLE: return "setole";
5208 case ISD::SETONE: return "setone";
5210 case ISD::SETO: return "seto";
5211 case ISD::SETUO: return "setuo";
5212 case ISD::SETUEQ: return "setue";
5213 case ISD::SETUGT: return "setugt";
5214 case ISD::SETUGE: return "setuge";
5215 case ISD::SETULT: return "setult";
5216 case ISD::SETULE: return "setule";
5217 case ISD::SETUNE: return "setune";
5219 case ISD::SETEQ: return "seteq";
5220 case ISD::SETGT: return "setgt";
5221 case ISD::SETGE: return "setge";
5222 case ISD::SETLT: return "setlt";
5223 case ISD::SETLE: return "setle";
5224 case ISD::SETNE: return "setne";
5229 const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) {
5238 return "<post-inc>";
5240 return "<post-dec>";
5244 std::string ISD::ArgFlagsTy::getArgFlagsString() {
5245 std::string S = "< ";
5259 if (getByValAlign())
5260 S += "byval-align:" + utostr(getByValAlign()) + " ";
5262 S += "orig-align:" + utostr(getOrigAlign()) + " ";
5264 S += "byval-size:" + utostr(getByValSize()) + " ";
5268 void SDNode::dump() const { dump(0); }
5269 void SDNode::dump(const SelectionDAG *G) const {
5274 void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const {
5275 OS << (void*)this << ": ";
5277 for (unsigned i = 0, e = getNumValues(); i != e; ++i) {
5279 if (getValueType(i) == MVT::Other)
5282 OS << getValueType(i).getMVTString();
5284 OS << " = " << getOperationName(G);
5287 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
5289 OS << (void*)getOperand(i).getNode();
5290 if (unsigned RN = getOperand(i).getResNo())
5294 if (!isTargetOpcode() && getOpcode() == ISD::VECTOR_SHUFFLE) {
5295 SDNode *Mask = getOperand(2).getNode();
5297 for (unsigned i = 0, e = Mask->getNumOperands(); i != e; ++i) {
5299 if (Mask->getOperand(i).getOpcode() == ISD::UNDEF)
5302 OS << cast<ConstantSDNode>(Mask->getOperand(i))->getZExtValue();
5307 if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
5308 OS << '<' << CSDN->getAPIntValue() << '>';
5309 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) {
5310 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
5311 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>';
5312 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble)
5313 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>';
5316 CSDN->getValueAPF().bitcastToAPInt().dump();
5319 } else if (const GlobalAddressSDNode *GADN =
5320 dyn_cast<GlobalAddressSDNode>(this)) {
5321 int64_t offset = GADN->getOffset();
5323 WriteAsOperand(OS, GADN->getGlobal());
5326 OS << " + " << offset;
5328 OS << " " << offset;
5329 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) {
5330 OS << "<" << FIDN->getIndex() << ">";
5331 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) {
5332 OS << "<" << JTDN->getIndex() << ">";
5333 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){
5334 int offset = CP->getOffset();
5335 if (CP->isMachineConstantPoolEntry())
5336 OS << "<" << *CP->getMachineCPVal() << ">";
5338 OS << "<" << *CP->getConstVal() << ">";
5340 OS << " + " << offset;
5342 OS << " " << offset;
5343 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) {
5345 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock();
5347 OS << LBB->getName() << " ";
5348 OS << (const void*)BBDN->getBasicBlock() << ">";
5349 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) {
5350 if (G && R->getReg() &&
5351 TargetRegisterInfo::isPhysicalRegister(R->getReg())) {
5352 OS << " " << G->getTarget().getRegisterInfo()->getName(R->getReg());
5354 OS << " #" << R->getReg();
5356 } else if (const ExternalSymbolSDNode *ES =
5357 dyn_cast<ExternalSymbolSDNode>(this)) {
5358 OS << "'" << ES->getSymbol() << "'";
5359 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) {
5361 OS << "<" << M->getValue() << ">";
5364 } else if (const MemOperandSDNode *M = dyn_cast<MemOperandSDNode>(this)) {
5365 if (M->MO.getValue())
5366 OS << "<" << M->MO.getValue() << ":" << M->MO.getOffset() << ">";
5368 OS << "<null:" << M->MO.getOffset() << ">";
5369 } else if (const ARG_FLAGSSDNode *N = dyn_cast<ARG_FLAGSSDNode>(this)) {
5370 OS << N->getArgFlags().getArgFlagsString();
5371 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) {
5372 OS << ":" << N->getVT().getMVTString();
5374 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) {
5375 const Value *SrcValue = LD->getSrcValue();
5376 int SrcOffset = LD->getSrcValueOffset();
5382 OS << ":" << SrcOffset << ">";
5385 switch (LD->getExtensionType()) {
5386 default: doExt = false; break;
5387 case ISD::EXTLOAD: OS << " <anyext "; break;
5388 case ISD::SEXTLOAD: OS << " <sext "; break;
5389 case ISD::ZEXTLOAD: OS << " <zext "; break;
5392 OS << LD->getMemoryVT().getMVTString() << ">";
5394 const char *AM = getIndexedModeName(LD->getAddressingMode());
5397 if (LD->isVolatile())
5398 OS << " <volatile>";
5399 OS << " alignment=" << LD->getAlignment();
5400 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) {
5401 const Value *SrcValue = ST->getSrcValue();
5402 int SrcOffset = ST->getSrcValueOffset();
5408 OS << ":" << SrcOffset << ">";
5410 if (ST->isTruncatingStore())
5411 OS << " <trunc " << ST->getMemoryVT().getMVTString() << ">";
5413 const char *AM = getIndexedModeName(ST->getAddressingMode());
5416 if (ST->isVolatile())
5417 OS << " <volatile>";
5418 OS << " alignment=" << ST->getAlignment();
5419 } else if (const AtomicSDNode* AT = dyn_cast<AtomicSDNode>(this)) {
5420 const Value *SrcValue = AT->getSrcValue();
5421 int SrcOffset = AT->getSrcValueOffset();
5427 OS << ":" << SrcOffset << ">";
5428 if (AT->isVolatile())
5429 OS << " <volatile>";
5430 OS << " alignment=" << AT->getAlignment();
5434 static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) {
5435 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5436 if (N->getOperand(i).getNode()->hasOneUse())
5437 DumpNodes(N->getOperand(i).getNode(), indent+2, G);
5439 cerr << "\n" << std::string(indent+2, ' ')
5440 << (void*)N->getOperand(i).getNode() << ": <multiple use>";
5443 cerr << "\n" << std::string(indent, ' ');
5447 void SelectionDAG::dump() const {
5448 cerr << "SelectionDAG has " << AllNodes.size() << " nodes:";
5450 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end();
5452 const SDNode *N = I;
5453 if (!N->hasOneUse() && N != getRoot().getNode())
5454 DumpNodes(N, 2, this);
5457 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this);
5462 const Type *ConstantPoolSDNode::getType() const {
5463 if (isMachineConstantPoolEntry())
5464 return Val.MachineCPVal->getType();
5465 return Val.ConstVal->getType();