1 //===- llvm/CodeGen/SelectionDAG.h - InstSelection DAG ----------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file declares the SelectionDAG class, and transitively defines the 10 // SDNode class and subclasses. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_CODEGEN_SELECTIONDAG_H 15 #define LLVM_CODEGEN_SELECTIONDAG_H 16 17 #include "llvm/ADT/APFloat.h" 18 #include "llvm/ADT/APInt.h" 19 #include "llvm/ADT/ArrayRef.h" 20 #include "llvm/ADT/DenseMap.h" 21 #include "llvm/ADT/DenseSet.h" 22 #include "llvm/ADT/FoldingSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/StringMap.h" 25 #include "llvm/ADT/ilist.h" 26 #include "llvm/ADT/iterator.h" 27 #include "llvm/ADT/iterator_range.h" 28 #include "llvm/CodeGen/DAGCombine.h" 29 #include "llvm/CodeGen/ISDOpcodes.h" 30 #include "llvm/CodeGen/MachineFunction.h" 31 #include "llvm/CodeGen/MachineMemOperand.h" 32 #include "llvm/CodeGen/MachineValueType.h" 33 #include "llvm/CodeGen/SelectionDAGNodes.h" 34 #include "llvm/CodeGen/ValueTypes.h" 35 #include "llvm/IR/DebugLoc.h" 36 #include "llvm/IR/Metadata.h" 37 #include "llvm/Support/Allocator.h" 38 #include "llvm/Support/ArrayRecycler.h" 39 #include "llvm/Support/CodeGen.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/RecyclingAllocator.h" 42 #include <cassert> 43 #include <cstdint> 44 #include <functional> 45 #include <map> 46 #include <string> 47 #include <tuple> 48 #include <utility> 49 #include <vector> 50 51 namespace llvm { 52 53 class DIExpression; 54 class DILabel; 55 class DIVariable; 56 class Function; 57 class Pass; 58 class Type; 59 template <class GraphType> struct GraphTraits; 60 template <typename T, unsigned int N> class SmallSetVector; 61 template <typename T, typename Enable> struct FoldingSetTrait; 62 class AAResults; 63 class BlockAddress; 64 class BlockFrequencyInfo; 65 class Constant; 66 class ConstantFP; 67 class ConstantInt; 68 class DataLayout; 69 struct fltSemantics; 70 class FunctionLoweringInfo; 71 class FunctionVarLocs; 72 class GlobalValue; 73 struct KnownBits; 74 class LLVMContext; 75 class MachineBasicBlock; 76 class MachineConstantPoolValue; 77 class MCSymbol; 78 class OptimizationRemarkEmitter; 79 class ProfileSummaryInfo; 80 class SDDbgValue; 81 class SDDbgOperand; 82 class SDDbgLabel; 83 class SelectionDAG; 84 class SelectionDAGTargetInfo; 85 class TargetLibraryInfo; 86 class TargetLowering; 87 class TargetMachine; 88 class TargetSubtargetInfo; 89 class Value; 90 91 template <typename T> class GenericSSAContext; 92 using SSAContext = GenericSSAContext<Function>; 93 template <typename T> class GenericUniformityInfo; 94 using UniformityInfo = GenericUniformityInfo<SSAContext>; 95 96 class SDVTListNode : public FoldingSetNode { 97 friend struct FoldingSetTrait<SDVTListNode>; 98 99 /// A reference to an Interned FoldingSetNodeID for this node. 100 /// The Allocator in SelectionDAG holds the data. 101 /// SDVTList contains all types which are frequently accessed in SelectionDAG. 102 /// The size of this list is not expected to be big so it won't introduce 103 /// a memory penalty. 104 FoldingSetNodeIDRef FastID; 105 const EVT *VTs; 106 unsigned int NumVTs; 107 /// The hash value for SDVTList is fixed, so cache it to avoid 108 /// hash calculation. 109 unsigned HashValue; 110 111 public: 112 SDVTListNode(const FoldingSetNodeIDRef ID, const EVT *VT, unsigned int Num) : 113 FastID(ID), VTs(VT), NumVTs(Num) { 114 HashValue = ID.ComputeHash(); 115 } 116 117 SDVTList getSDVTList() { 118 SDVTList result = {VTs, NumVTs}; 119 return result; 120 } 121 }; 122 123 /// Specialize FoldingSetTrait for SDVTListNode 124 /// to avoid computing temp FoldingSetNodeID and hash value. 125 template<> struct FoldingSetTrait<SDVTListNode> : DefaultFoldingSetTrait<SDVTListNode> { 126 static void Profile(const SDVTListNode &X, FoldingSetNodeID& ID) { 127 ID = X.FastID; 128 } 129 130 static bool Equals(const SDVTListNode &X, const FoldingSetNodeID &ID, 131 unsigned IDHash, FoldingSetNodeID &TempID) { 132 if (X.HashValue != IDHash) 133 return false; 134 return ID == X.FastID; 135 } 136 137 static unsigned ComputeHash(const SDVTListNode &X, FoldingSetNodeID &TempID) { 138 return X.HashValue; 139 } 140 }; 141 142 template <> struct ilist_alloc_traits<SDNode> { 143 static void deleteNode(SDNode *) { 144 llvm_unreachable("ilist_traits<SDNode> shouldn't see a deleteNode call!"); 145 } 146 }; 147 148 /// Keeps track of dbg_value information through SDISel. We do 149 /// not build SDNodes for these so as not to perturb the generated code; 150 /// instead the info is kept off to the side in this structure. Each SDNode may 151 /// have one or more associated dbg_value entries. This information is kept in 152 /// DbgValMap. 153 /// Byval parameters are handled separately because they don't use alloca's, 154 /// which busts the normal mechanism. There is good reason for handling all 155 /// parameters separately: they may not have code generated for them, they 156 /// should always go at the beginning of the function regardless of other code 157 /// motion, and debug info for them is potentially useful even if the parameter 158 /// is unused. Right now only byval parameters are handled separately. 159 class SDDbgInfo { 160 BumpPtrAllocator Alloc; 161 SmallVector<SDDbgValue*, 32> DbgValues; 162 SmallVector<SDDbgValue*, 32> ByvalParmDbgValues; 163 SmallVector<SDDbgLabel*, 4> DbgLabels; 164 using DbgValMapType = DenseMap<const SDNode *, SmallVector<SDDbgValue *, 2>>; 165 DbgValMapType DbgValMap; 166 167 public: 168 SDDbgInfo() = default; 169 SDDbgInfo(const SDDbgInfo &) = delete; 170 SDDbgInfo &operator=(const SDDbgInfo &) = delete; 171 172 void add(SDDbgValue *V, bool isParameter); 173 174 void add(SDDbgLabel *L) { DbgLabels.push_back(L); } 175 176 /// Invalidate all DbgValues attached to the node and remove 177 /// it from the Node-to-DbgValues map. 178 void erase(const SDNode *Node); 179 180 void clear() { 181 DbgValMap.clear(); 182 DbgValues.clear(); 183 ByvalParmDbgValues.clear(); 184 DbgLabels.clear(); 185 Alloc.Reset(); 186 } 187 188 BumpPtrAllocator &getAlloc() { return Alloc; } 189 190 bool empty() const { 191 return DbgValues.empty() && ByvalParmDbgValues.empty() && DbgLabels.empty(); 192 } 193 194 ArrayRef<SDDbgValue*> getSDDbgValues(const SDNode *Node) const { 195 auto I = DbgValMap.find(Node); 196 if (I != DbgValMap.end()) 197 return I->second; 198 return ArrayRef<SDDbgValue*>(); 199 } 200 201 using DbgIterator = SmallVectorImpl<SDDbgValue*>::iterator; 202 using DbgLabelIterator = SmallVectorImpl<SDDbgLabel*>::iterator; 203 204 DbgIterator DbgBegin() { return DbgValues.begin(); } 205 DbgIterator DbgEnd() { return DbgValues.end(); } 206 DbgIterator ByvalParmDbgBegin() { return ByvalParmDbgValues.begin(); } 207 DbgIterator ByvalParmDbgEnd() { return ByvalParmDbgValues.end(); } 208 DbgLabelIterator DbgLabelBegin() { return DbgLabels.begin(); } 209 DbgLabelIterator DbgLabelEnd() { return DbgLabels.end(); } 210 }; 211 212 void checkForCycles(const SelectionDAG *DAG, bool force = false); 213 214 /// This is used to represent a portion of an LLVM function in a low-level 215 /// Data Dependence DAG representation suitable for instruction selection. 216 /// This DAG is constructed as the first step of instruction selection in order 217 /// to allow implementation of machine specific optimizations 218 /// and code simplifications. 219 /// 220 /// The representation used by the SelectionDAG is a target-independent 221 /// representation, which has some similarities to the GCC RTL representation, 222 /// but is significantly more simple, powerful, and is a graph form instead of a 223 /// linear form. 224 /// 225 class SelectionDAG { 226 const TargetMachine &TM; 227 const SelectionDAGTargetInfo *TSI = nullptr; 228 const TargetLowering *TLI = nullptr; 229 const TargetLibraryInfo *LibInfo = nullptr; 230 const FunctionVarLocs *FnVarLocs = nullptr; 231 MachineFunction *MF; 232 Pass *SDAGISelPass = nullptr; 233 LLVMContext *Context; 234 CodeGenOptLevel OptLevel; 235 236 UniformityInfo *UA = nullptr; 237 FunctionLoweringInfo * FLI = nullptr; 238 239 /// The function-level optimization remark emitter. Used to emit remarks 240 /// whenever manipulating the DAG. 241 OptimizationRemarkEmitter *ORE; 242 243 ProfileSummaryInfo *PSI = nullptr; 244 BlockFrequencyInfo *BFI = nullptr; 245 246 /// List of non-single value types. 247 FoldingSet<SDVTListNode> VTListMap; 248 249 /// Pool allocation for misc. objects that are created once per SelectionDAG. 250 BumpPtrAllocator Allocator; 251 252 /// The starting token. 253 SDNode EntryNode; 254 255 /// The root of the entire DAG. 256 SDValue Root; 257 258 /// A linked list of nodes in the current DAG. 259 ilist<SDNode> AllNodes; 260 261 /// The AllocatorType for allocating SDNodes. We use 262 /// pool allocation with recycling. 263 using NodeAllocatorType = RecyclingAllocator<BumpPtrAllocator, SDNode, 264 sizeof(LargestSDNode), 265 alignof(MostAlignedSDNode)>; 266 267 /// Pool allocation for nodes. 268 NodeAllocatorType NodeAllocator; 269 270 /// This structure is used to memoize nodes, automatically performing 271 /// CSE with existing nodes when a duplicate is requested. 272 FoldingSet<SDNode> CSEMap; 273 274 /// Pool allocation for machine-opcode SDNode operands. 275 BumpPtrAllocator OperandAllocator; 276 ArrayRecycler<SDUse> OperandRecycler; 277 278 /// Tracks dbg_value and dbg_label information through SDISel. 279 SDDbgInfo *DbgInfo; 280 281 using CallSiteInfo = MachineFunction::CallSiteInfo; 282 using CallSiteInfoImpl = MachineFunction::CallSiteInfoImpl; 283 284 struct NodeExtraInfo { 285 CallSiteInfo CSInfo; 286 MDNode *HeapAllocSite = nullptr; 287 MDNode *PCSections = nullptr; 288 bool NoMerge = false; 289 }; 290 /// Out-of-line extra information for SDNodes. 291 DenseMap<const SDNode *, NodeExtraInfo> SDEI; 292 293 /// PersistentId counter to be used when inserting the next 294 /// SDNode to this SelectionDAG. We do not place that under 295 /// `#if LLVM_ENABLE_ABI_BREAKING_CHECKS` intentionally because 296 /// it adds unneeded complexity without noticeable 297 /// benefits (see discussion with @thakis in D120714). 298 uint16_t NextPersistentId = 0; 299 300 public: 301 /// Clients of various APIs that cause global effects on 302 /// the DAG can optionally implement this interface. This allows the clients 303 /// to handle the various sorts of updates that happen. 304 /// 305 /// A DAGUpdateListener automatically registers itself with DAG when it is 306 /// constructed, and removes itself when destroyed in RAII fashion. 307 struct DAGUpdateListener { 308 DAGUpdateListener *const Next; 309 SelectionDAG &DAG; 310 311 explicit DAGUpdateListener(SelectionDAG &D) 312 : Next(D.UpdateListeners), DAG(D) { 313 DAG.UpdateListeners = this; 314 } 315 316 virtual ~DAGUpdateListener() { 317 assert(DAG.UpdateListeners == this && 318 "DAGUpdateListeners must be destroyed in LIFO order"); 319 DAG.UpdateListeners = Next; 320 } 321 322 /// The node N that was deleted and, if E is not null, an 323 /// equivalent node E that replaced it. 324 virtual void NodeDeleted(SDNode *N, SDNode *E); 325 326 /// The node N that was updated. 327 virtual void NodeUpdated(SDNode *N); 328 329 /// The node N that was inserted. 330 virtual void NodeInserted(SDNode *N); 331 }; 332 333 struct DAGNodeDeletedListener : public DAGUpdateListener { 334 std::function<void(SDNode *, SDNode *)> Callback; 335 336 DAGNodeDeletedListener(SelectionDAG &DAG, 337 std::function<void(SDNode *, SDNode *)> Callback) 338 : DAGUpdateListener(DAG), Callback(std::move(Callback)) {} 339 340 void NodeDeleted(SDNode *N, SDNode *E) override { Callback(N, E); } 341 342 private: 343 virtual void anchor(); 344 }; 345 346 struct DAGNodeInsertedListener : public DAGUpdateListener { 347 std::function<void(SDNode *)> Callback; 348 349 DAGNodeInsertedListener(SelectionDAG &DAG, 350 std::function<void(SDNode *)> Callback) 351 : DAGUpdateListener(DAG), Callback(std::move(Callback)) {} 352 353 void NodeInserted(SDNode *N) override { Callback(N); } 354 355 private: 356 virtual void anchor(); 357 }; 358 359 /// Help to insert SDNodeFlags automatically in transforming. Use 360 /// RAII to save and resume flags in current scope. 361 class FlagInserter { 362 SelectionDAG &DAG; 363 SDNodeFlags Flags; 364 FlagInserter *LastInserter; 365 366 public: 367 FlagInserter(SelectionDAG &SDAG, SDNodeFlags Flags) 368 : DAG(SDAG), Flags(Flags), 369 LastInserter(SDAG.getFlagInserter()) { 370 SDAG.setFlagInserter(this); 371 } 372 FlagInserter(SelectionDAG &SDAG, SDNode *N) 373 : FlagInserter(SDAG, N->getFlags()) {} 374 375 FlagInserter(const FlagInserter &) = delete; 376 FlagInserter &operator=(const FlagInserter &) = delete; 377 ~FlagInserter() { DAG.setFlagInserter(LastInserter); } 378 379 SDNodeFlags getFlags() const { return Flags; } 380 }; 381 382 /// When true, additional steps are taken to 383 /// ensure that getConstant() and similar functions return DAG nodes that 384 /// have legal types. This is important after type legalization since 385 /// any illegally typed nodes generated after this point will not experience 386 /// type legalization. 387 bool NewNodesMustHaveLegalTypes = false; 388 389 private: 390 /// DAGUpdateListener is a friend so it can manipulate the listener stack. 391 friend struct DAGUpdateListener; 392 393 /// Linked list of registered DAGUpdateListener instances. 394 /// This stack is maintained by DAGUpdateListener RAII. 395 DAGUpdateListener *UpdateListeners = nullptr; 396 397 /// Implementation of setSubgraphColor. 398 /// Return whether we had to truncate the search. 399 bool setSubgraphColorHelper(SDNode *N, const char *Color, 400 DenseSet<SDNode *> &visited, 401 int level, bool &printed); 402 403 template <typename SDNodeT, typename... ArgTypes> 404 SDNodeT *newSDNode(ArgTypes &&... Args) { 405 return new (NodeAllocator.template Allocate<SDNodeT>()) 406 SDNodeT(std::forward<ArgTypes>(Args)...); 407 } 408 409 /// Build a synthetic SDNodeT with the given args and extract its subclass 410 /// data as an integer (e.g. for use in a folding set). 411 /// 412 /// The args to this function are the same as the args to SDNodeT's 413 /// constructor, except the second arg (assumed to be a const DebugLoc&) is 414 /// omitted. 415 template <typename SDNodeT, typename... ArgTypes> 416 static uint16_t getSyntheticNodeSubclassData(unsigned IROrder, 417 ArgTypes &&... Args) { 418 // The compiler can reduce this expression to a constant iff we pass an 419 // empty DebugLoc. Thankfully, the debug location doesn't have any bearing 420 // on the subclass data. 421 return SDNodeT(IROrder, DebugLoc(), std::forward<ArgTypes>(Args)...) 422 .getRawSubclassData(); 423 } 424 425 template <typename SDNodeTy> 426 static uint16_t getSyntheticNodeSubclassData(unsigned Opc, unsigned Order, 427 SDVTList VTs, EVT MemoryVT, 428 MachineMemOperand *MMO) { 429 return SDNodeTy(Opc, Order, DebugLoc(), VTs, MemoryVT, MMO) 430 .getRawSubclassData(); 431 } 432 433 void createOperands(SDNode *Node, ArrayRef<SDValue> Vals); 434 435 void removeOperands(SDNode *Node) { 436 if (!Node->OperandList) 437 return; 438 OperandRecycler.deallocate( 439 ArrayRecycler<SDUse>::Capacity::get(Node->NumOperands), 440 Node->OperandList); 441 Node->NumOperands = 0; 442 Node->OperandList = nullptr; 443 } 444 void CreateTopologicalOrder(std::vector<SDNode*>& Order); 445 446 public: 447 // Maximum depth for recursive analysis such as computeKnownBits, etc. 448 static constexpr unsigned MaxRecursionDepth = 6; 449 450 explicit SelectionDAG(const TargetMachine &TM, CodeGenOptLevel); 451 SelectionDAG(const SelectionDAG &) = delete; 452 SelectionDAG &operator=(const SelectionDAG &) = delete; 453 ~SelectionDAG(); 454 455 /// Prepare this SelectionDAG to process code in the given MachineFunction. 456 void init(MachineFunction &NewMF, OptimizationRemarkEmitter &NewORE, 457 Pass *PassPtr, const TargetLibraryInfo *LibraryInfo, 458 UniformityInfo *UA, ProfileSummaryInfo *PSIin, 459 BlockFrequencyInfo *BFIin, FunctionVarLocs const *FnVarLocs); 460 461 void setFunctionLoweringInfo(FunctionLoweringInfo * FuncInfo) { 462 FLI = FuncInfo; 463 } 464 465 /// Clear state and free memory necessary to make this 466 /// SelectionDAG ready to process a new block. 467 void clear(); 468 469 MachineFunction &getMachineFunction() const { return *MF; } 470 const Pass *getPass() const { return SDAGISelPass; } 471 472 const DataLayout &getDataLayout() const { return MF->getDataLayout(); } 473 const TargetMachine &getTarget() const { return TM; } 474 const TargetSubtargetInfo &getSubtarget() const { return MF->getSubtarget(); } 475 template <typename STC> const STC &getSubtarget() const { 476 return MF->getSubtarget<STC>(); 477 } 478 const TargetLowering &getTargetLoweringInfo() const { return *TLI; } 479 const TargetLibraryInfo &getLibInfo() const { return *LibInfo; } 480 const SelectionDAGTargetInfo &getSelectionDAGInfo() const { return *TSI; } 481 const UniformityInfo *getUniformityInfo() const { return UA; } 482 /// Returns the result of the AssignmentTrackingAnalysis pass if it's 483 /// available, otherwise return nullptr. 484 const FunctionVarLocs *getFunctionVarLocs() const { return FnVarLocs; } 485 LLVMContext *getContext() const { return Context; } 486 OptimizationRemarkEmitter &getORE() const { return *ORE; } 487 ProfileSummaryInfo *getPSI() const { return PSI; } 488 BlockFrequencyInfo *getBFI() const { return BFI; } 489 490 FlagInserter *getFlagInserter() { return Inserter; } 491 void setFlagInserter(FlagInserter *FI) { Inserter = FI; } 492 493 /// Just dump dot graph to a user-provided path and title. 494 /// This doesn't open the dot viewer program and 495 /// helps visualization when outside debugging session. 496 /// FileName expects absolute path. If provided 497 /// without any path separators then the file 498 /// will be created in the current directory. 499 /// Error will be emitted if the path is insane. 500 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 501 LLVM_DUMP_METHOD void dumpDotGraph(const Twine &FileName, const Twine &Title); 502 #endif 503 504 /// Pop up a GraphViz/gv window with the DAG rendered using 'dot'. 505 void viewGraph(const std::string &Title); 506 void viewGraph(); 507 508 #if LLVM_ENABLE_ABI_BREAKING_CHECKS 509 std::map<const SDNode *, std::string> NodeGraphAttrs; 510 #endif 511 512 /// Clear all previously defined node graph attributes. 513 /// Intended to be used from a debugging tool (eg. gdb). 514 void clearGraphAttrs(); 515 516 /// Set graph attributes for a node. (eg. "color=red".) 517 void setGraphAttrs(const SDNode *N, const char *Attrs); 518 519 /// Get graph attributes for a node. (eg. "color=red".) 520 /// Used from getNodeAttributes. 521 std::string getGraphAttrs(const SDNode *N) const; 522 523 /// Convenience for setting node color attribute. 524 void setGraphColor(const SDNode *N, const char *Color); 525 526 /// Convenience for setting subgraph color attribute. 527 void setSubgraphColor(SDNode *N, const char *Color); 528 529 using allnodes_const_iterator = ilist<SDNode>::const_iterator; 530 531 allnodes_const_iterator allnodes_begin() const { return AllNodes.begin(); } 532 allnodes_const_iterator allnodes_end() const { return AllNodes.end(); } 533 534 using allnodes_iterator = ilist<SDNode>::iterator; 535 536 allnodes_iterator allnodes_begin() { return AllNodes.begin(); } 537 allnodes_iterator allnodes_end() { return AllNodes.end(); } 538 539 ilist<SDNode>::size_type allnodes_size() const { 540 return AllNodes.size(); 541 } 542 543 iterator_range<allnodes_iterator> allnodes() { 544 return make_range(allnodes_begin(), allnodes_end()); 545 } 546 iterator_range<allnodes_const_iterator> allnodes() const { 547 return make_range(allnodes_begin(), allnodes_end()); 548 } 549 550 /// Return the root tag of the SelectionDAG. 551 const SDValue &getRoot() const { return Root; } 552 553 /// Return the token chain corresponding to the entry of the function. 554 SDValue getEntryNode() const { 555 return SDValue(const_cast<SDNode *>(&EntryNode), 0); 556 } 557 558 /// Set the current root tag of the SelectionDAG. 559 /// 560 const SDValue &setRoot(SDValue N) { 561 assert((!N.getNode() || N.getValueType() == MVT::Other) && 562 "DAG root value is not a chain!"); 563 if (N.getNode()) 564 checkForCycles(N.getNode(), this); 565 Root = N; 566 if (N.getNode()) 567 checkForCycles(this); 568 return Root; 569 } 570 571 #ifndef NDEBUG 572 void VerifyDAGDivergence(); 573 #endif 574 575 /// This iterates over the nodes in the SelectionDAG, folding 576 /// certain types of nodes together, or eliminating superfluous nodes. The 577 /// Level argument controls whether Combine is allowed to produce nodes and 578 /// types that are illegal on the target. 579 void Combine(CombineLevel Level, AAResults *AA, CodeGenOptLevel OptLevel); 580 581 /// This transforms the SelectionDAG into a SelectionDAG that 582 /// only uses types natively supported by the target. 583 /// Returns "true" if it made any changes. 584 /// 585 /// Note that this is an involved process that may invalidate pointers into 586 /// the graph. 587 bool LegalizeTypes(); 588 589 /// This transforms the SelectionDAG into a SelectionDAG that is 590 /// compatible with the target instruction selector, as indicated by the 591 /// TargetLowering object. 592 /// 593 /// Note that this is an involved process that may invalidate pointers into 594 /// the graph. 595 void Legalize(); 596 597 /// Transforms a SelectionDAG node and any operands to it into a node 598 /// that is compatible with the target instruction selector, as indicated by 599 /// the TargetLowering object. 600 /// 601 /// \returns true if \c N is a valid, legal node after calling this. 602 /// 603 /// This essentially runs a single recursive walk of the \c Legalize process 604 /// over the given node (and its operands). This can be used to incrementally 605 /// legalize the DAG. All of the nodes which are directly replaced, 606 /// potentially including N, are added to the output parameter \c 607 /// UpdatedNodes so that the delta to the DAG can be understood by the 608 /// caller. 609 /// 610 /// When this returns false, N has been legalized in a way that make the 611 /// pointer passed in no longer valid. It may have even been deleted from the 612 /// DAG, and so it shouldn't be used further. When this returns true, the 613 /// N passed in is a legal node, and can be immediately processed as such. 614 /// This may still have done some work on the DAG, and will still populate 615 /// UpdatedNodes with any new nodes replacing those originally in the DAG. 616 bool LegalizeOp(SDNode *N, SmallSetVector<SDNode *, 16> &UpdatedNodes); 617 618 /// This transforms the SelectionDAG into a SelectionDAG 619 /// that only uses vector math operations supported by the target. This is 620 /// necessary as a separate step from Legalize because unrolling a vector 621 /// operation can introduce illegal types, which requires running 622 /// LegalizeTypes again. 623 /// 624 /// This returns true if it made any changes; in that case, LegalizeTypes 625 /// is called again before Legalize. 626 /// 627 /// Note that this is an involved process that may invalidate pointers into 628 /// the graph. 629 bool LegalizeVectors(); 630 631 /// This method deletes all unreachable nodes in the SelectionDAG. 632 void RemoveDeadNodes(); 633 634 /// Remove the specified node from the system. This node must 635 /// have no referrers. 636 void DeleteNode(SDNode *N); 637 638 /// Return an SDVTList that represents the list of values specified. 639 SDVTList getVTList(EVT VT); 640 SDVTList getVTList(EVT VT1, EVT VT2); 641 SDVTList getVTList(EVT VT1, EVT VT2, EVT VT3); 642 SDVTList getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4); 643 SDVTList getVTList(ArrayRef<EVT> VTs); 644 645 //===--------------------------------------------------------------------===// 646 // Node creation methods. 647 648 /// Create a ConstantSDNode wrapping a constant value. 649 /// If VT is a vector type, the constant is splatted into a BUILD_VECTOR. 650 /// 651 /// If only legal types can be produced, this does the necessary 652 /// transformations (e.g., if the vector element type is illegal). 653 /// @{ 654 SDValue getConstant(uint64_t Val, const SDLoc &DL, EVT VT, 655 bool isTarget = false, bool isOpaque = false); 656 SDValue getConstant(const APInt &Val, const SDLoc &DL, EVT VT, 657 bool isTarget = false, bool isOpaque = false); 658 659 SDValue getAllOnesConstant(const SDLoc &DL, EVT VT, bool IsTarget = false, 660 bool IsOpaque = false) { 661 return getConstant(APInt::getAllOnes(VT.getScalarSizeInBits()), DL, VT, 662 IsTarget, IsOpaque); 663 } 664 665 SDValue getConstant(const ConstantInt &Val, const SDLoc &DL, EVT VT, 666 bool isTarget = false, bool isOpaque = false); 667 SDValue getIntPtrConstant(uint64_t Val, const SDLoc &DL, 668 bool isTarget = false); 669 SDValue getShiftAmountConstant(uint64_t Val, EVT VT, const SDLoc &DL, 670 bool LegalTypes = true); 671 SDValue getVectorIdxConstant(uint64_t Val, const SDLoc &DL, 672 bool isTarget = false); 673 674 SDValue getTargetConstant(uint64_t Val, const SDLoc &DL, EVT VT, 675 bool isOpaque = false) { 676 return getConstant(Val, DL, VT, true, isOpaque); 677 } 678 SDValue getTargetConstant(const APInt &Val, const SDLoc &DL, EVT VT, 679 bool isOpaque = false) { 680 return getConstant(Val, DL, VT, true, isOpaque); 681 } 682 SDValue getTargetConstant(const ConstantInt &Val, const SDLoc &DL, EVT VT, 683 bool isOpaque = false) { 684 return getConstant(Val, DL, VT, true, isOpaque); 685 } 686 687 /// Create a true or false constant of type \p VT using the target's 688 /// BooleanContent for type \p OpVT. 689 SDValue getBoolConstant(bool V, const SDLoc &DL, EVT VT, EVT OpVT); 690 /// @} 691 692 /// Create a ConstantFPSDNode wrapping a constant value. 693 /// If VT is a vector type, the constant is splatted into a BUILD_VECTOR. 694 /// 695 /// If only legal types can be produced, this does the necessary 696 /// transformations (e.g., if the vector element type is illegal). 697 /// The forms that take a double should only be used for simple constants 698 /// that can be exactly represented in VT. No checks are made. 699 /// @{ 700 SDValue getConstantFP(double Val, const SDLoc &DL, EVT VT, 701 bool isTarget = false); 702 SDValue getConstantFP(const APFloat &Val, const SDLoc &DL, EVT VT, 703 bool isTarget = false); 704 SDValue getConstantFP(const ConstantFP &V, const SDLoc &DL, EVT VT, 705 bool isTarget = false); 706 SDValue getTargetConstantFP(double Val, const SDLoc &DL, EVT VT) { 707 return getConstantFP(Val, DL, VT, true); 708 } 709 SDValue getTargetConstantFP(const APFloat &Val, const SDLoc &DL, EVT VT) { 710 return getConstantFP(Val, DL, VT, true); 711 } 712 SDValue getTargetConstantFP(const ConstantFP &Val, const SDLoc &DL, EVT VT) { 713 return getConstantFP(Val, DL, VT, true); 714 } 715 /// @} 716 717 SDValue getGlobalAddress(const GlobalValue *GV, const SDLoc &DL, EVT VT, 718 int64_t offset = 0, bool isTargetGA = false, 719 unsigned TargetFlags = 0); 720 SDValue getTargetGlobalAddress(const GlobalValue *GV, const SDLoc &DL, EVT VT, 721 int64_t offset = 0, unsigned TargetFlags = 0) { 722 return getGlobalAddress(GV, DL, VT, offset, true, TargetFlags); 723 } 724 SDValue getFrameIndex(int FI, EVT VT, bool isTarget = false); 725 SDValue getTargetFrameIndex(int FI, EVT VT) { 726 return getFrameIndex(FI, VT, true); 727 } 728 SDValue getJumpTable(int JTI, EVT VT, bool isTarget = false, 729 unsigned TargetFlags = 0); 730 SDValue getTargetJumpTable(int JTI, EVT VT, unsigned TargetFlags = 0) { 731 return getJumpTable(JTI, VT, true, TargetFlags); 732 } 733 SDValue getJumpTableDebugInfo(int JTI, SDValue Chain, const SDLoc &DL); 734 SDValue getConstantPool(const Constant *C, EVT VT, 735 MaybeAlign Align = std::nullopt, int Offs = 0, 736 bool isT = false, unsigned TargetFlags = 0); 737 SDValue getTargetConstantPool(const Constant *C, EVT VT, 738 MaybeAlign Align = std::nullopt, int Offset = 0, 739 unsigned TargetFlags = 0) { 740 return getConstantPool(C, VT, Align, Offset, true, TargetFlags); 741 } 742 SDValue getConstantPool(MachineConstantPoolValue *C, EVT VT, 743 MaybeAlign Align = std::nullopt, int Offs = 0, 744 bool isT = false, unsigned TargetFlags = 0); 745 SDValue getTargetConstantPool(MachineConstantPoolValue *C, EVT VT, 746 MaybeAlign Align = std::nullopt, int Offset = 0, 747 unsigned TargetFlags = 0) { 748 return getConstantPool(C, VT, Align, Offset, true, TargetFlags); 749 } 750 // When generating a branch to a BB, we don't in general know enough 751 // to provide debug info for the BB at that time, so keep this one around. 752 SDValue getBasicBlock(MachineBasicBlock *MBB); 753 SDValue getExternalSymbol(const char *Sym, EVT VT); 754 SDValue getTargetExternalSymbol(const char *Sym, EVT VT, 755 unsigned TargetFlags = 0); 756 SDValue getMCSymbol(MCSymbol *Sym, EVT VT); 757 758 SDValue getValueType(EVT); 759 SDValue getRegister(unsigned Reg, EVT VT); 760 SDValue getRegisterMask(const uint32_t *RegMask); 761 SDValue getEHLabel(const SDLoc &dl, SDValue Root, MCSymbol *Label); 762 SDValue getLabelNode(unsigned Opcode, const SDLoc &dl, SDValue Root, 763 MCSymbol *Label); 764 SDValue getBlockAddress(const BlockAddress *BA, EVT VT, int64_t Offset = 0, 765 bool isTarget = false, unsigned TargetFlags = 0); 766 SDValue getTargetBlockAddress(const BlockAddress *BA, EVT VT, 767 int64_t Offset = 0, unsigned TargetFlags = 0) { 768 return getBlockAddress(BA, VT, Offset, true, TargetFlags); 769 } 770 771 SDValue getCopyToReg(SDValue Chain, const SDLoc &dl, unsigned Reg, 772 SDValue N) { 773 return getNode(ISD::CopyToReg, dl, MVT::Other, Chain, 774 getRegister(Reg, N.getValueType()), N); 775 } 776 777 // This version of the getCopyToReg method takes an extra operand, which 778 // indicates that there is potentially an incoming glue value (if Glue is not 779 // null) and that there should be a glue result. 780 SDValue getCopyToReg(SDValue Chain, const SDLoc &dl, unsigned Reg, SDValue N, 781 SDValue Glue) { 782 SDVTList VTs = getVTList(MVT::Other, MVT::Glue); 783 SDValue Ops[] = { Chain, getRegister(Reg, N.getValueType()), N, Glue }; 784 return getNode(ISD::CopyToReg, dl, VTs, 785 ArrayRef(Ops, Glue.getNode() ? 4 : 3)); 786 } 787 788 // Similar to last getCopyToReg() except parameter Reg is a SDValue 789 SDValue getCopyToReg(SDValue Chain, const SDLoc &dl, SDValue Reg, SDValue N, 790 SDValue Glue) { 791 SDVTList VTs = getVTList(MVT::Other, MVT::Glue); 792 SDValue Ops[] = { Chain, Reg, N, Glue }; 793 return getNode(ISD::CopyToReg, dl, VTs, 794 ArrayRef(Ops, Glue.getNode() ? 4 : 3)); 795 } 796 797 SDValue getCopyFromReg(SDValue Chain, const SDLoc &dl, unsigned Reg, EVT VT) { 798 SDVTList VTs = getVTList(VT, MVT::Other); 799 SDValue Ops[] = { Chain, getRegister(Reg, VT) }; 800 return getNode(ISD::CopyFromReg, dl, VTs, Ops); 801 } 802 803 // This version of the getCopyFromReg method takes an extra operand, which 804 // indicates that there is potentially an incoming glue value (if Glue is not 805 // null) and that there should be a glue result. 806 SDValue getCopyFromReg(SDValue Chain, const SDLoc &dl, unsigned Reg, EVT VT, 807 SDValue Glue) { 808 SDVTList VTs = getVTList(VT, MVT::Other, MVT::Glue); 809 SDValue Ops[] = { Chain, getRegister(Reg, VT), Glue }; 810 return getNode(ISD::CopyFromReg, dl, VTs, 811 ArrayRef(Ops, Glue.getNode() ? 3 : 2)); 812 } 813 814 SDValue getCondCode(ISD::CondCode Cond); 815 816 /// Return an ISD::VECTOR_SHUFFLE node. The number of elements in VT, 817 /// which must be a vector type, must match the number of mask elements 818 /// NumElts. An integer mask element equal to -1 is treated as undefined. 819 SDValue getVectorShuffle(EVT VT, const SDLoc &dl, SDValue N1, SDValue N2, 820 ArrayRef<int> Mask); 821 822 /// Return an ISD::BUILD_VECTOR node. The number of elements in VT, 823 /// which must be a vector type, must match the number of operands in Ops. 824 /// The operands must have the same type as (or, for integers, a type wider 825 /// than) VT's element type. 826 SDValue getBuildVector(EVT VT, const SDLoc &DL, ArrayRef<SDValue> Ops) { 827 // VerifySDNode (via InsertNode) checks BUILD_VECTOR later. 828 return getNode(ISD::BUILD_VECTOR, DL, VT, Ops); 829 } 830 831 /// Return an ISD::BUILD_VECTOR node. The number of elements in VT, 832 /// which must be a vector type, must match the number of operands in Ops. 833 /// The operands must have the same type as (or, for integers, a type wider 834 /// than) VT's element type. 835 SDValue getBuildVector(EVT VT, const SDLoc &DL, ArrayRef<SDUse> Ops) { 836 // VerifySDNode (via InsertNode) checks BUILD_VECTOR later. 837 return getNode(ISD::BUILD_VECTOR, DL, VT, Ops); 838 } 839 840 /// Return a splat ISD::BUILD_VECTOR node, consisting of Op splatted to all 841 /// elements. VT must be a vector type. Op's type must be the same as (or, 842 /// for integers, a type wider than) VT's element type. 843 SDValue getSplatBuildVector(EVT VT, const SDLoc &DL, SDValue Op) { 844 // VerifySDNode (via InsertNode) checks BUILD_VECTOR later. 845 if (Op.getOpcode() == ISD::UNDEF) { 846 assert((VT.getVectorElementType() == Op.getValueType() || 847 (VT.isInteger() && 848 VT.getVectorElementType().bitsLE(Op.getValueType()))) && 849 "A splatted value must have a width equal or (for integers) " 850 "greater than the vector element type!"); 851 return getNode(ISD::UNDEF, SDLoc(), VT); 852 } 853 854 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Op); 855 return getNode(ISD::BUILD_VECTOR, DL, VT, Ops); 856 } 857 858 // Return a splat ISD::SPLAT_VECTOR node, consisting of Op splatted to all 859 // elements. 860 SDValue getSplatVector(EVT VT, const SDLoc &DL, SDValue Op) { 861 if (Op.getOpcode() == ISD::UNDEF) { 862 assert((VT.getVectorElementType() == Op.getValueType() || 863 (VT.isInteger() && 864 VT.getVectorElementType().bitsLE(Op.getValueType()))) && 865 "A splatted value must have a width equal or (for integers) " 866 "greater than the vector element type!"); 867 return getNode(ISD::UNDEF, SDLoc(), VT); 868 } 869 return getNode(ISD::SPLAT_VECTOR, DL, VT, Op); 870 } 871 872 /// Returns a node representing a splat of one value into all lanes 873 /// of the provided vector type. This is a utility which returns 874 /// either a BUILD_VECTOR or SPLAT_VECTOR depending on the 875 /// scalability of the desired vector type. 876 SDValue getSplat(EVT VT, const SDLoc &DL, SDValue Op) { 877 assert(VT.isVector() && "Can't splat to non-vector type"); 878 return VT.isScalableVector() ? 879 getSplatVector(VT, DL, Op) : getSplatBuildVector(VT, DL, Op); 880 } 881 882 /// Returns a vector of type ResVT whose elements contain the linear sequence 883 /// <0, Step, Step * 2, Step * 3, ...> 884 SDValue getStepVector(const SDLoc &DL, EVT ResVT, APInt StepVal); 885 886 /// Returns a vector of type ResVT whose elements contain the linear sequence 887 /// <0, 1, 2, 3, ...> 888 SDValue getStepVector(const SDLoc &DL, EVT ResVT); 889 890 /// Returns an ISD::VECTOR_SHUFFLE node semantically equivalent to 891 /// the shuffle node in input but with swapped operands. 892 /// 893 /// Example: shuffle A, B, <0,5,2,7> -> shuffle B, A, <4,1,6,3> 894 SDValue getCommutedVectorShuffle(const ShuffleVectorSDNode &SV); 895 896 /// Convert Op, which must be of float type, to the 897 /// float type VT, by either extending or rounding (by truncation). 898 SDValue getFPExtendOrRound(SDValue Op, const SDLoc &DL, EVT VT); 899 900 /// Convert Op, which must be a STRICT operation of float type, to the 901 /// float type VT, by either extending or rounding (by truncation). 902 std::pair<SDValue, SDValue> 903 getStrictFPExtendOrRound(SDValue Op, SDValue Chain, const SDLoc &DL, EVT VT); 904 905 /// Convert *_EXTEND_VECTOR_INREG to *_EXTEND opcode. 906 static unsigned getOpcode_EXTEND(unsigned Opcode) { 907 switch (Opcode) { 908 case ISD::ANY_EXTEND: 909 case ISD::ANY_EXTEND_VECTOR_INREG: 910 return ISD::ANY_EXTEND; 911 case ISD::ZERO_EXTEND: 912 case ISD::ZERO_EXTEND_VECTOR_INREG: 913 return ISD::ZERO_EXTEND; 914 case ISD::SIGN_EXTEND: 915 case ISD::SIGN_EXTEND_VECTOR_INREG: 916 return ISD::SIGN_EXTEND; 917 } 918 llvm_unreachable("Unknown opcode"); 919 } 920 921 /// Convert *_EXTEND to *_EXTEND_VECTOR_INREG opcode. 922 static unsigned getOpcode_EXTEND_VECTOR_INREG(unsigned Opcode) { 923 switch (Opcode) { 924 case ISD::ANY_EXTEND: 925 case ISD::ANY_EXTEND_VECTOR_INREG: 926 return ISD::ANY_EXTEND_VECTOR_INREG; 927 case ISD::ZERO_EXTEND: 928 case ISD::ZERO_EXTEND_VECTOR_INREG: 929 return ISD::ZERO_EXTEND_VECTOR_INREG; 930 case ISD::SIGN_EXTEND: 931 case ISD::SIGN_EXTEND_VECTOR_INREG: 932 return ISD::SIGN_EXTEND_VECTOR_INREG; 933 } 934 llvm_unreachable("Unknown opcode"); 935 } 936 937 /// Convert Op, which must be of integer type, to the 938 /// integer type VT, by either any-extending or truncating it. 939 SDValue getAnyExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 940 941 /// Convert Op, which must be of integer type, to the 942 /// integer type VT, by either sign-extending or truncating it. 943 SDValue getSExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 944 945 /// Convert Op, which must be of integer type, to the 946 /// integer type VT, by either zero-extending or truncating it. 947 SDValue getZExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 948 949 /// Convert Op, which must be of integer type, to the 950 /// integer type VT, by either any/sign/zero-extending (depending on IsAny / 951 /// IsSigned) or truncating it. 952 SDValue getExtOrTrunc(SDValue Op, const SDLoc &DL, 953 EVT VT, unsigned Opcode) { 954 switch(Opcode) { 955 case ISD::ANY_EXTEND: 956 return getAnyExtOrTrunc(Op, DL, VT); 957 case ISD::ZERO_EXTEND: 958 return getZExtOrTrunc(Op, DL, VT); 959 case ISD::SIGN_EXTEND: 960 return getSExtOrTrunc(Op, DL, VT); 961 } 962 llvm_unreachable("Unsupported opcode"); 963 } 964 965 /// Convert Op, which must be of integer type, to the 966 /// integer type VT, by either sign/zero-extending (depending on IsSigned) or 967 /// truncating it. 968 SDValue getExtOrTrunc(bool IsSigned, SDValue Op, const SDLoc &DL, EVT VT) { 969 return IsSigned ? getSExtOrTrunc(Op, DL, VT) : getZExtOrTrunc(Op, DL, VT); 970 } 971 972 /// Convert Op, which must be of integer type, to the 973 /// integer type VT, by first bitcasting (from potential vector) to 974 /// corresponding scalar type then either any-extending or truncating it. 975 SDValue getBitcastedAnyExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 976 977 /// Convert Op, which must be of integer type, to the 978 /// integer type VT, by first bitcasting (from potential vector) to 979 /// corresponding scalar type then either sign-extending or truncating it. 980 SDValue getBitcastedSExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 981 982 /// Convert Op, which must be of integer type, to the 983 /// integer type VT, by first bitcasting (from potential vector) to 984 /// corresponding scalar type then either zero-extending or truncating it. 985 SDValue getBitcastedZExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 986 987 /// Return the expression required to zero extend the Op 988 /// value assuming it was the smaller SrcTy value. 989 SDValue getZeroExtendInReg(SDValue Op, const SDLoc &DL, EVT VT); 990 991 /// Convert Op, which must be of integer type, to the integer type VT, by 992 /// either truncating it or performing either zero or sign extension as 993 /// appropriate extension for the pointer's semantics. 994 SDValue getPtrExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT); 995 996 /// Return the expression required to extend the Op as a pointer value 997 /// assuming it was the smaller SrcTy value. This may be either a zero extend 998 /// or a sign extend. 999 SDValue getPtrExtendInReg(SDValue Op, const SDLoc &DL, EVT VT); 1000 1001 /// Convert Op, which must be of integer type, to the integer type VT, 1002 /// by using an extension appropriate for the target's 1003 /// BooleanContent for type OpVT or truncating it. 1004 SDValue getBoolExtOrTrunc(SDValue Op, const SDLoc &SL, EVT VT, EVT OpVT); 1005 1006 /// Create negative operation as (SUB 0, Val). 1007 SDValue getNegative(SDValue Val, const SDLoc &DL, EVT VT); 1008 1009 /// Create a bitwise NOT operation as (XOR Val, -1). 1010 SDValue getNOT(const SDLoc &DL, SDValue Val, EVT VT); 1011 1012 /// Create a logical NOT operation as (XOR Val, BooleanOne). 1013 SDValue getLogicalNOT(const SDLoc &DL, SDValue Val, EVT VT); 1014 1015 /// Create a vector-predicated logical NOT operation as (VP_XOR Val, 1016 /// BooleanOne, Mask, EVL). 1017 SDValue getVPLogicalNOT(const SDLoc &DL, SDValue Val, SDValue Mask, 1018 SDValue EVL, EVT VT); 1019 1020 /// Convert a vector-predicated Op, which must be an integer vector, to the 1021 /// vector-type VT, by performing either vector-predicated zext or truncating 1022 /// it. The Op will be returned as-is if Op and VT are vectors containing 1023 /// integer with same width. 1024 SDValue getVPZExtOrTrunc(const SDLoc &DL, EVT VT, SDValue Op, SDValue Mask, 1025 SDValue EVL); 1026 1027 /// Convert a vector-predicated Op, which must be of integer type, to the 1028 /// vector-type integer type VT, by either truncating it or performing either 1029 /// vector-predicated zero or sign extension as appropriate extension for the 1030 /// pointer's semantics. This function just redirects to getVPZExtOrTrunc 1031 /// right now. 1032 SDValue getVPPtrExtOrTrunc(const SDLoc &DL, EVT VT, SDValue Op, SDValue Mask, 1033 SDValue EVL); 1034 1035 /// Returns sum of the base pointer and offset. 1036 /// Unlike getObjectPtrOffset this does not set NoUnsignedWrap by default. 1037 SDValue getMemBasePlusOffset(SDValue Base, TypeSize Offset, const SDLoc &DL, 1038 const SDNodeFlags Flags = SDNodeFlags()); 1039 SDValue getMemBasePlusOffset(SDValue Base, SDValue Offset, const SDLoc &DL, 1040 const SDNodeFlags Flags = SDNodeFlags()); 1041 1042 /// Create an add instruction with appropriate flags when used for 1043 /// addressing some offset of an object. i.e. if a load is split into multiple 1044 /// components, create an add nuw from the base pointer to the offset. 1045 SDValue getObjectPtrOffset(const SDLoc &SL, SDValue Ptr, TypeSize Offset) { 1046 SDNodeFlags Flags; 1047 Flags.setNoUnsignedWrap(true); 1048 return getMemBasePlusOffset(Ptr, Offset, SL, Flags); 1049 } 1050 1051 SDValue getObjectPtrOffset(const SDLoc &SL, SDValue Ptr, SDValue Offset) { 1052 // The object itself can't wrap around the address space, so it shouldn't be 1053 // possible for the adds of the offsets to the split parts to overflow. 1054 SDNodeFlags Flags; 1055 Flags.setNoUnsignedWrap(true); 1056 return getMemBasePlusOffset(Ptr, Offset, SL, Flags); 1057 } 1058 1059 /// Return a new CALLSEQ_START node, that starts new call frame, in which 1060 /// InSize bytes are set up inside CALLSEQ_START..CALLSEQ_END sequence and 1061 /// OutSize specifies part of the frame set up prior to the sequence. 1062 SDValue getCALLSEQ_START(SDValue Chain, uint64_t InSize, uint64_t OutSize, 1063 const SDLoc &DL) { 1064 SDVTList VTs = getVTList(MVT::Other, MVT::Glue); 1065 SDValue Ops[] = { Chain, 1066 getIntPtrConstant(InSize, DL, true), 1067 getIntPtrConstant(OutSize, DL, true) }; 1068 return getNode(ISD::CALLSEQ_START, DL, VTs, Ops); 1069 } 1070 1071 /// Return a new CALLSEQ_END node, which always must have a 1072 /// glue result (to ensure it's not CSE'd). 1073 /// CALLSEQ_END does not have a useful SDLoc. 1074 SDValue getCALLSEQ_END(SDValue Chain, SDValue Op1, SDValue Op2, 1075 SDValue InGlue, const SDLoc &DL) { 1076 SDVTList NodeTys = getVTList(MVT::Other, MVT::Glue); 1077 SmallVector<SDValue, 4> Ops; 1078 Ops.push_back(Chain); 1079 Ops.push_back(Op1); 1080 Ops.push_back(Op2); 1081 if (InGlue.getNode()) 1082 Ops.push_back(InGlue); 1083 return getNode(ISD::CALLSEQ_END, DL, NodeTys, Ops); 1084 } 1085 1086 SDValue getCALLSEQ_END(SDValue Chain, uint64_t Size1, uint64_t Size2, 1087 SDValue Glue, const SDLoc &DL) { 1088 return getCALLSEQ_END( 1089 Chain, getIntPtrConstant(Size1, DL, /*isTarget=*/true), 1090 getIntPtrConstant(Size2, DL, /*isTarget=*/true), Glue, DL); 1091 } 1092 1093 /// Return true if the result of this operation is always undefined. 1094 bool isUndef(unsigned Opcode, ArrayRef<SDValue> Ops); 1095 1096 /// Return an UNDEF node. UNDEF does not have a useful SDLoc. 1097 SDValue getUNDEF(EVT VT) { 1098 return getNode(ISD::UNDEF, SDLoc(), VT); 1099 } 1100 1101 /// Return a node that represents the runtime scaling 'MulImm * RuntimeVL'. 1102 SDValue getVScale(const SDLoc &DL, EVT VT, APInt MulImm, 1103 bool ConstantFold = true); 1104 1105 SDValue getElementCount(const SDLoc &DL, EVT VT, ElementCount EC, 1106 bool ConstantFold = true); 1107 1108 /// Return a GLOBAL_OFFSET_TABLE node. This does not have a useful SDLoc. 1109 SDValue getGLOBAL_OFFSET_TABLE(EVT VT) { 1110 return getNode(ISD::GLOBAL_OFFSET_TABLE, SDLoc(), VT); 1111 } 1112 1113 /// Gets or creates the specified node. 1114 /// 1115 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, 1116 ArrayRef<SDUse> Ops); 1117 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, 1118 ArrayRef<SDValue> Ops, const SDNodeFlags Flags); 1119 SDValue getNode(unsigned Opcode, const SDLoc &DL, ArrayRef<EVT> ResultTys, 1120 ArrayRef<SDValue> Ops); 1121 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, 1122 ArrayRef<SDValue> Ops, const SDNodeFlags Flags); 1123 1124 // Use flags from current flag inserter. 1125 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, 1126 ArrayRef<SDValue> Ops); 1127 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, 1128 ArrayRef<SDValue> Ops); 1129 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue Operand); 1130 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue N1, 1131 SDValue N2); 1132 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue N1, 1133 SDValue N2, SDValue N3); 1134 1135 // Specialize based on number of operands. 1136 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT); 1137 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue Operand, 1138 const SDNodeFlags Flags); 1139 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue N1, 1140 SDValue N2, const SDNodeFlags Flags); 1141 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue N1, 1142 SDValue N2, SDValue N3, const SDNodeFlags Flags); 1143 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue N1, 1144 SDValue N2, SDValue N3, SDValue N4); 1145 SDValue getNode(unsigned Opcode, const SDLoc &DL, EVT VT, SDValue N1, 1146 SDValue N2, SDValue N3, SDValue N4, SDValue N5); 1147 1148 // Specialize again based on number of operands for nodes with a VTList 1149 // rather than a single VT. 1150 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList); 1151 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, SDValue N); 1152 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, SDValue N1, 1153 SDValue N2); 1154 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, SDValue N1, 1155 SDValue N2, SDValue N3); 1156 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, SDValue N1, 1157 SDValue N2, SDValue N3, SDValue N4); 1158 SDValue getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, SDValue N1, 1159 SDValue N2, SDValue N3, SDValue N4, SDValue N5); 1160 1161 /// Compute a TokenFactor to force all the incoming stack arguments to be 1162 /// loaded from the stack. This is used in tail call lowering to protect 1163 /// stack arguments from being clobbered. 1164 SDValue getStackArgumentTokenFactor(SDValue Chain); 1165 1166 SDValue getMemcpy(SDValue Chain, const SDLoc &dl, SDValue Dst, SDValue Src, 1167 SDValue Size, Align Alignment, bool isVol, 1168 bool AlwaysInline, bool isTailCall, 1169 MachinePointerInfo DstPtrInfo, 1170 MachinePointerInfo SrcPtrInfo, 1171 const AAMDNodes &AAInfo = AAMDNodes(), 1172 AAResults *AA = nullptr); 1173 1174 SDValue getMemmove(SDValue Chain, const SDLoc &dl, SDValue Dst, SDValue Src, 1175 SDValue Size, Align Alignment, bool isVol, bool isTailCall, 1176 MachinePointerInfo DstPtrInfo, 1177 MachinePointerInfo SrcPtrInfo, 1178 const AAMDNodes &AAInfo = AAMDNodes(), 1179 AAResults *AA = nullptr); 1180 1181 SDValue getMemset(SDValue Chain, const SDLoc &dl, SDValue Dst, SDValue Src, 1182 SDValue Size, Align Alignment, bool isVol, 1183 bool AlwaysInline, bool isTailCall, 1184 MachinePointerInfo DstPtrInfo, 1185 const AAMDNodes &AAInfo = AAMDNodes()); 1186 1187 SDValue getAtomicMemcpy(SDValue Chain, const SDLoc &dl, SDValue Dst, 1188 SDValue Src, SDValue Size, Type *SizeTy, 1189 unsigned ElemSz, bool isTailCall, 1190 MachinePointerInfo DstPtrInfo, 1191 MachinePointerInfo SrcPtrInfo); 1192 1193 SDValue getAtomicMemmove(SDValue Chain, const SDLoc &dl, SDValue Dst, 1194 SDValue Src, SDValue Size, Type *SizeTy, 1195 unsigned ElemSz, bool isTailCall, 1196 MachinePointerInfo DstPtrInfo, 1197 MachinePointerInfo SrcPtrInfo); 1198 1199 SDValue getAtomicMemset(SDValue Chain, const SDLoc &dl, SDValue Dst, 1200 SDValue Value, SDValue Size, Type *SizeTy, 1201 unsigned ElemSz, bool isTailCall, 1202 MachinePointerInfo DstPtrInfo); 1203 1204 /// Helper function to make it easier to build SetCC's if you just have an 1205 /// ISD::CondCode instead of an SDValue. 1206 SDValue getSetCC(const SDLoc &DL, EVT VT, SDValue LHS, SDValue RHS, 1207 ISD::CondCode Cond, SDValue Chain = SDValue(), 1208 bool IsSignaling = false) { 1209 assert(LHS.getValueType().isVector() == RHS.getValueType().isVector() && 1210 "Vector/scalar operand type mismatch for setcc"); 1211 assert(LHS.getValueType().isVector() == VT.isVector() && 1212 "Vector/scalar result type mismatch for setcc"); 1213 assert(Cond != ISD::SETCC_INVALID && 1214 "Cannot create a setCC of an invalid node."); 1215 if (Chain) 1216 return getNode(IsSignaling ? ISD::STRICT_FSETCCS : ISD::STRICT_FSETCC, DL, 1217 {VT, MVT::Other}, {Chain, LHS, RHS, getCondCode(Cond)}); 1218 return getNode(ISD::SETCC, DL, VT, LHS, RHS, getCondCode(Cond)); 1219 } 1220 1221 /// Helper function to make it easier to build VP_SETCCs if you just have an 1222 /// ISD::CondCode instead of an SDValue. 1223 SDValue getSetCCVP(const SDLoc &DL, EVT VT, SDValue LHS, SDValue RHS, 1224 ISD::CondCode Cond, SDValue Mask, SDValue EVL) { 1225 assert(LHS.getValueType().isVector() && RHS.getValueType().isVector() && 1226 "Cannot compare scalars"); 1227 assert(Cond != ISD::SETCC_INVALID && 1228 "Cannot create a setCC of an invalid node."); 1229 return getNode(ISD::VP_SETCC, DL, VT, LHS, RHS, getCondCode(Cond), Mask, 1230 EVL); 1231 } 1232 1233 /// Helper function to make it easier to build Select's if you just have 1234 /// operands and don't want to check for vector. 1235 SDValue getSelect(const SDLoc &DL, EVT VT, SDValue Cond, SDValue LHS, 1236 SDValue RHS) { 1237 assert(LHS.getValueType() == VT && RHS.getValueType() == VT && 1238 "Cannot use select on differing types"); 1239 auto Opcode = Cond.getValueType().isVector() ? ISD::VSELECT : ISD::SELECT; 1240 return getNode(Opcode, DL, VT, Cond, LHS, RHS); 1241 } 1242 1243 /// Helper function to make it easier to build SelectCC's if you just have an 1244 /// ISD::CondCode instead of an SDValue. 1245 SDValue getSelectCC(const SDLoc &DL, SDValue LHS, SDValue RHS, SDValue True, 1246 SDValue False, ISD::CondCode Cond) { 1247 return getNode(ISD::SELECT_CC, DL, True.getValueType(), LHS, RHS, True, 1248 False, getCondCode(Cond)); 1249 } 1250 1251 /// Try to simplify a select/vselect into 1 of its operands or a constant. 1252 SDValue simplifySelect(SDValue Cond, SDValue TVal, SDValue FVal); 1253 1254 /// Try to simplify a shift into 1 of its operands or a constant. 1255 SDValue simplifyShift(SDValue X, SDValue Y); 1256 1257 /// Try to simplify a floating-point binary operation into 1 of its operands 1258 /// or a constant. 1259 SDValue simplifyFPBinop(unsigned Opcode, SDValue X, SDValue Y, 1260 SDNodeFlags Flags); 1261 1262 /// VAArg produces a result and token chain, and takes a pointer 1263 /// and a source value as input. 1264 SDValue getVAArg(EVT VT, const SDLoc &dl, SDValue Chain, SDValue Ptr, 1265 SDValue SV, unsigned Align); 1266 1267 /// Gets a node for an atomic cmpxchg op. There are two 1268 /// valid Opcodes. ISD::ATOMIC_CMO_SWAP produces the value loaded and a 1269 /// chain result. ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS produces the value loaded, 1270 /// a success flag (initially i1), and a chain. 1271 SDValue getAtomicCmpSwap(unsigned Opcode, const SDLoc &dl, EVT MemVT, 1272 SDVTList VTs, SDValue Chain, SDValue Ptr, 1273 SDValue Cmp, SDValue Swp, MachineMemOperand *MMO); 1274 1275 /// Gets a node for an atomic op, produces result (if relevant) 1276 /// and chain and takes 2 operands. 1277 SDValue getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT, SDValue Chain, 1278 SDValue Ptr, SDValue Val, MachineMemOperand *MMO); 1279 1280 /// Gets a node for an atomic op, produces result and chain and 1281 /// takes 1 operand. 1282 SDValue getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT, EVT VT, 1283 SDValue Chain, SDValue Ptr, MachineMemOperand *MMO); 1284 1285 /// Gets a node for an atomic op, produces result and chain and takes N 1286 /// operands. 1287 SDValue getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT, 1288 SDVTList VTList, ArrayRef<SDValue> Ops, 1289 MachineMemOperand *MMO); 1290 1291 /// Creates a MemIntrinsicNode that may produce a 1292 /// result and takes a list of operands. Opcode may be INTRINSIC_VOID, 1293 /// INTRINSIC_W_CHAIN, or a target-specific opcode with a value not 1294 /// less than FIRST_TARGET_MEMORY_OPCODE. 1295 SDValue getMemIntrinsicNode( 1296 unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops, 1297 EVT MemVT, MachinePointerInfo PtrInfo, Align Alignment, 1298 MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad | 1299 MachineMemOperand::MOStore, 1300 uint64_t Size = 0, const AAMDNodes &AAInfo = AAMDNodes()); 1301 1302 inline SDValue getMemIntrinsicNode( 1303 unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops, 1304 EVT MemVT, MachinePointerInfo PtrInfo, 1305 MaybeAlign Alignment = std::nullopt, 1306 MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad | 1307 MachineMemOperand::MOStore, 1308 uint64_t Size = 0, const AAMDNodes &AAInfo = AAMDNodes()) { 1309 // Ensure that codegen never sees alignment 0 1310 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, MemVT, PtrInfo, 1311 Alignment.value_or(getEVTAlign(MemVT)), Flags, 1312 Size, AAInfo); 1313 } 1314 1315 SDValue getMemIntrinsicNode(unsigned Opcode, const SDLoc &dl, SDVTList VTList, 1316 ArrayRef<SDValue> Ops, EVT MemVT, 1317 MachineMemOperand *MMO); 1318 1319 /// Creates a LifetimeSDNode that starts (`IsStart==true`) or ends 1320 /// (`IsStart==false`) the lifetime of the portion of `FrameIndex` between 1321 /// offsets `Offset` and `Offset + Size`. 1322 SDValue getLifetimeNode(bool IsStart, const SDLoc &dl, SDValue Chain, 1323 int FrameIndex, int64_t Size, int64_t Offset = -1); 1324 1325 /// Creates a PseudoProbeSDNode with function GUID `Guid` and 1326 /// the index of the block `Index` it is probing, as well as the attributes 1327 /// `attr` of the probe. 1328 SDValue getPseudoProbeNode(const SDLoc &Dl, SDValue Chain, uint64_t Guid, 1329 uint64_t Index, uint32_t Attr); 1330 1331 /// Create a MERGE_VALUES node from the given operands. 1332 SDValue getMergeValues(ArrayRef<SDValue> Ops, const SDLoc &dl); 1333 1334 /// Loads are not normal binary operators: their result type is not 1335 /// determined by their operands, and they produce a value AND a token chain. 1336 /// 1337 /// This function will set the MOLoad flag on MMOFlags, but you can set it if 1338 /// you want. The MOStore flag must not be set. 1339 SDValue getLoad(EVT VT, const SDLoc &dl, SDValue Chain, SDValue Ptr, 1340 MachinePointerInfo PtrInfo, 1341 MaybeAlign Alignment = MaybeAlign(), 1342 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1343 const AAMDNodes &AAInfo = AAMDNodes(), 1344 const MDNode *Ranges = nullptr); 1345 SDValue getLoad(EVT VT, const SDLoc &dl, SDValue Chain, SDValue Ptr, 1346 MachineMemOperand *MMO); 1347 SDValue 1348 getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl, EVT VT, SDValue Chain, 1349 SDValue Ptr, MachinePointerInfo PtrInfo, EVT MemVT, 1350 MaybeAlign Alignment = MaybeAlign(), 1351 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1352 const AAMDNodes &AAInfo = AAMDNodes()); 1353 SDValue getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl, EVT VT, 1354 SDValue Chain, SDValue Ptr, EVT MemVT, 1355 MachineMemOperand *MMO); 1356 SDValue getIndexedLoad(SDValue OrigLoad, const SDLoc &dl, SDValue Base, 1357 SDValue Offset, ISD::MemIndexedMode AM); 1358 SDValue getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, 1359 const SDLoc &dl, SDValue Chain, SDValue Ptr, SDValue Offset, 1360 MachinePointerInfo PtrInfo, EVT MemVT, Align Alignment, 1361 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1362 const AAMDNodes &AAInfo = AAMDNodes(), 1363 const MDNode *Ranges = nullptr); 1364 inline SDValue getLoad( 1365 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &dl, 1366 SDValue Chain, SDValue Ptr, SDValue Offset, MachinePointerInfo PtrInfo, 1367 EVT MemVT, MaybeAlign Alignment = MaybeAlign(), 1368 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1369 const AAMDNodes &AAInfo = AAMDNodes(), const MDNode *Ranges = nullptr) { 1370 // Ensures that codegen never sees a None Alignment. 1371 return getLoad(AM, ExtType, VT, dl, Chain, Ptr, Offset, PtrInfo, MemVT, 1372 Alignment.value_or(getEVTAlign(MemVT)), MMOFlags, AAInfo, 1373 Ranges); 1374 } 1375 SDValue getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, 1376 const SDLoc &dl, SDValue Chain, SDValue Ptr, SDValue Offset, 1377 EVT MemVT, MachineMemOperand *MMO); 1378 1379 /// Helper function to build ISD::STORE nodes. 1380 /// 1381 /// This function will set the MOStore flag on MMOFlags, but you can set it if 1382 /// you want. The MOLoad and MOInvariant flags must not be set. 1383 1384 SDValue 1385 getStore(SDValue Chain, const SDLoc &dl, SDValue Val, SDValue Ptr, 1386 MachinePointerInfo PtrInfo, Align Alignment, 1387 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1388 const AAMDNodes &AAInfo = AAMDNodes()); 1389 inline SDValue 1390 getStore(SDValue Chain, const SDLoc &dl, SDValue Val, SDValue Ptr, 1391 MachinePointerInfo PtrInfo, MaybeAlign Alignment = MaybeAlign(), 1392 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1393 const AAMDNodes &AAInfo = AAMDNodes()) { 1394 return getStore(Chain, dl, Val, Ptr, PtrInfo, 1395 Alignment.value_or(getEVTAlign(Val.getValueType())), 1396 MMOFlags, AAInfo); 1397 } 1398 SDValue getStore(SDValue Chain, const SDLoc &dl, SDValue Val, SDValue Ptr, 1399 MachineMemOperand *MMO); 1400 SDValue 1401 getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val, SDValue Ptr, 1402 MachinePointerInfo PtrInfo, EVT SVT, Align Alignment, 1403 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1404 const AAMDNodes &AAInfo = AAMDNodes()); 1405 inline SDValue 1406 getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val, SDValue Ptr, 1407 MachinePointerInfo PtrInfo, EVT SVT, 1408 MaybeAlign Alignment = MaybeAlign(), 1409 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1410 const AAMDNodes &AAInfo = AAMDNodes()) { 1411 return getTruncStore(Chain, dl, Val, Ptr, PtrInfo, SVT, 1412 Alignment.value_or(getEVTAlign(SVT)), MMOFlags, 1413 AAInfo); 1414 } 1415 SDValue getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val, 1416 SDValue Ptr, EVT SVT, MachineMemOperand *MMO); 1417 SDValue getIndexedStore(SDValue OrigStore, const SDLoc &dl, SDValue Base, 1418 SDValue Offset, ISD::MemIndexedMode AM); 1419 1420 SDValue getLoadVP(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, 1421 const SDLoc &dl, SDValue Chain, SDValue Ptr, SDValue Offset, 1422 SDValue Mask, SDValue EVL, MachinePointerInfo PtrInfo, 1423 EVT MemVT, Align Alignment, 1424 MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo, 1425 const MDNode *Ranges = nullptr, bool IsExpanding = false); 1426 inline SDValue 1427 getLoadVP(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, 1428 const SDLoc &dl, SDValue Chain, SDValue Ptr, SDValue Offset, 1429 SDValue Mask, SDValue EVL, MachinePointerInfo PtrInfo, EVT MemVT, 1430 MaybeAlign Alignment = MaybeAlign(), 1431 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1432 const AAMDNodes &AAInfo = AAMDNodes(), 1433 const MDNode *Ranges = nullptr, bool IsExpanding = false) { 1434 // Ensures that codegen never sees a None Alignment. 1435 return getLoadVP(AM, ExtType, VT, dl, Chain, Ptr, Offset, Mask, EVL, 1436 PtrInfo, MemVT, Alignment.value_or(getEVTAlign(MemVT)), 1437 MMOFlags, AAInfo, Ranges, IsExpanding); 1438 } 1439 SDValue getLoadVP(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, 1440 const SDLoc &dl, SDValue Chain, SDValue Ptr, SDValue Offset, 1441 SDValue Mask, SDValue EVL, EVT MemVT, 1442 MachineMemOperand *MMO, bool IsExpanding = false); 1443 SDValue getLoadVP(EVT VT, const SDLoc &dl, SDValue Chain, SDValue Ptr, 1444 SDValue Mask, SDValue EVL, MachinePointerInfo PtrInfo, 1445 MaybeAlign Alignment, MachineMemOperand::Flags MMOFlags, 1446 const AAMDNodes &AAInfo, const MDNode *Ranges = nullptr, 1447 bool IsExpanding = false); 1448 SDValue getLoadVP(EVT VT, const SDLoc &dl, SDValue Chain, SDValue Ptr, 1449 SDValue Mask, SDValue EVL, MachineMemOperand *MMO, 1450 bool IsExpanding = false); 1451 SDValue getExtLoadVP(ISD::LoadExtType ExtType, const SDLoc &dl, EVT VT, 1452 SDValue Chain, SDValue Ptr, SDValue Mask, SDValue EVL, 1453 MachinePointerInfo PtrInfo, EVT MemVT, 1454 MaybeAlign Alignment, MachineMemOperand::Flags MMOFlags, 1455 const AAMDNodes &AAInfo, bool IsExpanding = false); 1456 SDValue getExtLoadVP(ISD::LoadExtType ExtType, const SDLoc &dl, EVT VT, 1457 SDValue Chain, SDValue Ptr, SDValue Mask, SDValue EVL, 1458 EVT MemVT, MachineMemOperand *MMO, 1459 bool IsExpanding = false); 1460 SDValue getIndexedLoadVP(SDValue OrigLoad, const SDLoc &dl, SDValue Base, 1461 SDValue Offset, ISD::MemIndexedMode AM); 1462 SDValue getStoreVP(SDValue Chain, const SDLoc &dl, SDValue Val, SDValue Ptr, 1463 SDValue Offset, SDValue Mask, SDValue EVL, EVT MemVT, 1464 MachineMemOperand *MMO, ISD::MemIndexedMode AM, 1465 bool IsTruncating = false, bool IsCompressing = false); 1466 SDValue getTruncStoreVP(SDValue Chain, const SDLoc &dl, SDValue Val, 1467 SDValue Ptr, SDValue Mask, SDValue EVL, 1468 MachinePointerInfo PtrInfo, EVT SVT, Align Alignment, 1469 MachineMemOperand::Flags MMOFlags, 1470 const AAMDNodes &AAInfo, bool IsCompressing = false); 1471 SDValue getTruncStoreVP(SDValue Chain, const SDLoc &dl, SDValue Val, 1472 SDValue Ptr, SDValue Mask, SDValue EVL, EVT SVT, 1473 MachineMemOperand *MMO, bool IsCompressing = false); 1474 SDValue getIndexedStoreVP(SDValue OrigStore, const SDLoc &dl, SDValue Base, 1475 SDValue Offset, ISD::MemIndexedMode AM); 1476 1477 SDValue getStridedLoadVP(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, 1478 EVT VT, const SDLoc &DL, SDValue Chain, SDValue Ptr, 1479 SDValue Offset, SDValue Stride, SDValue Mask, 1480 SDValue EVL, MachinePointerInfo PtrInfo, EVT MemVT, 1481 Align Alignment, MachineMemOperand::Flags MMOFlags, 1482 const AAMDNodes &AAInfo, 1483 const MDNode *Ranges = nullptr, 1484 bool IsExpanding = false); 1485 inline SDValue getStridedLoadVP( 1486 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &DL, 1487 SDValue Chain, SDValue Ptr, SDValue Offset, SDValue Stride, SDValue Mask, 1488 SDValue EVL, MachinePointerInfo PtrInfo, EVT MemVT, 1489 MaybeAlign Alignment = MaybeAlign(), 1490 MachineMemOperand::Flags MMOFlags = MachineMemOperand::MONone, 1491 const AAMDNodes &AAInfo = AAMDNodes(), const MDNode *Ranges = nullptr, 1492 bool IsExpanding = false) { 1493 // Ensures that codegen never sees a None Alignment. 1494 return getStridedLoadVP(AM, ExtType, VT, DL, Chain, Ptr, Offset, Stride, 1495 Mask, EVL, PtrInfo, MemVT, 1496 Alignment.value_or(getEVTAlign(MemVT)), MMOFlags, 1497 AAInfo, Ranges, IsExpanding); 1498 } 1499 SDValue getStridedLoadVP(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, 1500 EVT VT, const SDLoc &DL, SDValue Chain, SDValue Ptr, 1501 SDValue Offset, SDValue Stride, SDValue Mask, 1502 SDValue EVL, EVT MemVT, MachineMemOperand *MMO, 1503 bool IsExpanding = false); 1504 SDValue getStridedLoadVP(EVT VT, const SDLoc &DL, SDValue Chain, SDValue Ptr, 1505 SDValue Stride, SDValue Mask, SDValue EVL, 1506 MachinePointerInfo PtrInfo, MaybeAlign Alignment, 1507 MachineMemOperand::Flags MMOFlags, 1508 const AAMDNodes &AAInfo, 1509 const MDNode *Ranges = nullptr, 1510 bool IsExpanding = false); 1511 SDValue getStridedLoadVP(EVT VT, const SDLoc &DL, SDValue Chain, SDValue Ptr, 1512 SDValue Stride, SDValue Mask, SDValue EVL, 1513 MachineMemOperand *MMO, bool IsExpanding = false); 1514 SDValue 1515 getExtStridedLoadVP(ISD::LoadExtType ExtType, const SDLoc &DL, EVT VT, 1516 SDValue Chain, SDValue Ptr, SDValue Stride, SDValue Mask, 1517 SDValue EVL, MachinePointerInfo PtrInfo, EVT MemVT, 1518 MaybeAlign Alignment, MachineMemOperand::Flags MMOFlags, 1519 const AAMDNodes &AAInfo, bool IsExpanding = false); 1520 SDValue getExtStridedLoadVP(ISD::LoadExtType ExtType, const SDLoc &DL, EVT VT, 1521 SDValue Chain, SDValue Ptr, SDValue Stride, 1522 SDValue Mask, SDValue EVL, EVT MemVT, 1523 MachineMemOperand *MMO, bool IsExpanding = false); 1524 SDValue getIndexedStridedLoadVP(SDValue OrigLoad, const SDLoc &DL, 1525 SDValue Base, SDValue Offset, 1526 ISD::MemIndexedMode AM); 1527 SDValue getStridedStoreVP(SDValue Chain, const SDLoc &DL, SDValue Val, 1528 SDValue Ptr, SDValue Offset, SDValue Stride, 1529 SDValue Mask, SDValue EVL, EVT MemVT, 1530 MachineMemOperand *MMO, ISD::MemIndexedMode AM, 1531 bool IsTruncating = false, 1532 bool IsCompressing = false); 1533 SDValue getTruncStridedStoreVP(SDValue Chain, const SDLoc &DL, SDValue Val, 1534 SDValue Ptr, SDValue Stride, SDValue Mask, 1535 SDValue EVL, MachinePointerInfo PtrInfo, 1536 EVT SVT, Align Alignment, 1537 MachineMemOperand::Flags MMOFlags, 1538 const AAMDNodes &AAInfo, 1539 bool IsCompressing = false); 1540 SDValue getTruncStridedStoreVP(SDValue Chain, const SDLoc &DL, SDValue Val, 1541 SDValue Ptr, SDValue Stride, SDValue Mask, 1542 SDValue EVL, EVT SVT, MachineMemOperand *MMO, 1543 bool IsCompressing = false); 1544 SDValue getIndexedStridedStoreVP(SDValue OrigStore, const SDLoc &DL, 1545 SDValue Base, SDValue Offset, 1546 ISD::MemIndexedMode AM); 1547 1548 SDValue getGatherVP(SDVTList VTs, EVT VT, const SDLoc &dl, 1549 ArrayRef<SDValue> Ops, MachineMemOperand *MMO, 1550 ISD::MemIndexType IndexType); 1551 SDValue getScatterVP(SDVTList VTs, EVT VT, const SDLoc &dl, 1552 ArrayRef<SDValue> Ops, MachineMemOperand *MMO, 1553 ISD::MemIndexType IndexType); 1554 1555 SDValue getMaskedLoad(EVT VT, const SDLoc &dl, SDValue Chain, SDValue Base, 1556 SDValue Offset, SDValue Mask, SDValue Src0, EVT MemVT, 1557 MachineMemOperand *MMO, ISD::MemIndexedMode AM, 1558 ISD::LoadExtType, bool IsExpanding = false); 1559 SDValue getIndexedMaskedLoad(SDValue OrigLoad, const SDLoc &dl, SDValue Base, 1560 SDValue Offset, ISD::MemIndexedMode AM); 1561 SDValue getMaskedStore(SDValue Chain, const SDLoc &dl, SDValue Val, 1562 SDValue Base, SDValue Offset, SDValue Mask, EVT MemVT, 1563 MachineMemOperand *MMO, ISD::MemIndexedMode AM, 1564 bool IsTruncating = false, bool IsCompressing = false); 1565 SDValue getIndexedMaskedStore(SDValue OrigStore, const SDLoc &dl, 1566 SDValue Base, SDValue Offset, 1567 ISD::MemIndexedMode AM); 1568 SDValue getMaskedGather(SDVTList VTs, EVT MemVT, const SDLoc &dl, 1569 ArrayRef<SDValue> Ops, MachineMemOperand *MMO, 1570 ISD::MemIndexType IndexType, ISD::LoadExtType ExtTy); 1571 SDValue getMaskedScatter(SDVTList VTs, EVT MemVT, const SDLoc &dl, 1572 ArrayRef<SDValue> Ops, MachineMemOperand *MMO, 1573 ISD::MemIndexType IndexType, 1574 bool IsTruncating = false); 1575 1576 SDValue getGetFPEnv(SDValue Chain, const SDLoc &dl, SDValue Ptr, EVT MemVT, 1577 MachineMemOperand *MMO); 1578 SDValue getSetFPEnv(SDValue Chain, const SDLoc &dl, SDValue Ptr, EVT MemVT, 1579 MachineMemOperand *MMO); 1580 1581 /// Construct a node to track a Value* through the backend. 1582 SDValue getSrcValue(const Value *v); 1583 1584 /// Return an MDNodeSDNode which holds an MDNode. 1585 SDValue getMDNode(const MDNode *MD); 1586 1587 /// Return a bitcast using the SDLoc of the value operand, and casting to the 1588 /// provided type. Use getNode to set a custom SDLoc. 1589 SDValue getBitcast(EVT VT, SDValue V); 1590 1591 /// Return an AddrSpaceCastSDNode. 1592 SDValue getAddrSpaceCast(const SDLoc &dl, EVT VT, SDValue Ptr, unsigned SrcAS, 1593 unsigned DestAS); 1594 1595 /// Return a freeze using the SDLoc of the value operand. 1596 SDValue getFreeze(SDValue V); 1597 1598 /// Return an AssertAlignSDNode. 1599 SDValue getAssertAlign(const SDLoc &DL, SDValue V, Align A); 1600 1601 /// Swap N1 and N2 if Opcode is a commutative binary opcode 1602 /// and the canonical form expects the opposite order. 1603 void canonicalizeCommutativeBinop(unsigned Opcode, SDValue &N1, 1604 SDValue &N2) const; 1605 1606 /// Return the specified value casted to 1607 /// the target's desired shift amount type. 1608 SDValue getShiftAmountOperand(EVT LHSTy, SDValue Op); 1609 1610 /// Expand the specified \c ISD::VAARG node as the Legalize pass would. 1611 SDValue expandVAArg(SDNode *Node); 1612 1613 /// Expand the specified \c ISD::VACOPY node as the Legalize pass would. 1614 SDValue expandVACopy(SDNode *Node); 1615 1616 /// Returs an GlobalAddress of the function from the current module with 1617 /// name matching the given ExternalSymbol. Additionally can provide the 1618 /// matched function. 1619 /// Panics the function doesn't exists. 1620 SDValue getSymbolFunctionGlobalAddress(SDValue Op, 1621 Function **TargetFunction = nullptr); 1622 1623 /// *Mutate* the specified node in-place to have the 1624 /// specified operands. If the resultant node already exists in the DAG, 1625 /// this does not modify the specified node, instead it returns the node that 1626 /// already exists. If the resultant node does not exist in the DAG, the 1627 /// input node is returned. As a degenerate case, if you specify the same 1628 /// input operands as the node already has, the input node is returned. 1629 SDNode *UpdateNodeOperands(SDNode *N, SDValue Op); 1630 SDNode *UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2); 1631 SDNode *UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, 1632 SDValue Op3); 1633 SDNode *UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, 1634 SDValue Op3, SDValue Op4); 1635 SDNode *UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, 1636 SDValue Op3, SDValue Op4, SDValue Op5); 1637 SDNode *UpdateNodeOperands(SDNode *N, ArrayRef<SDValue> Ops); 1638 1639 /// Creates a new TokenFactor containing \p Vals. If \p Vals contains 64k 1640 /// values or more, move values into new TokenFactors in 64k-1 blocks, until 1641 /// the final TokenFactor has less than 64k operands. 1642 SDValue getTokenFactor(const SDLoc &DL, SmallVectorImpl<SDValue> &Vals); 1643 1644 /// *Mutate* the specified machine node's memory references to the provided 1645 /// list. 1646 void setNodeMemRefs(MachineSDNode *N, 1647 ArrayRef<MachineMemOperand *> NewMemRefs); 1648 1649 // Calculate divergence of node \p N based on its operands. 1650 bool calculateDivergence(SDNode *N); 1651 1652 // Propagates the change in divergence to users 1653 void updateDivergence(SDNode * N); 1654 1655 /// These are used for target selectors to *mutate* the 1656 /// specified node to have the specified return type, Target opcode, and 1657 /// operands. Note that target opcodes are stored as 1658 /// ~TargetOpcode in the node opcode field. The resultant node is returned. 1659 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT); 1660 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT, SDValue Op1); 1661 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT, 1662 SDValue Op1, SDValue Op2); 1663 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT, 1664 SDValue Op1, SDValue Op2, SDValue Op3); 1665 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT, 1666 ArrayRef<SDValue> Ops); 1667 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT1, EVT VT2); 1668 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT1, 1669 EVT VT2, ArrayRef<SDValue> Ops); 1670 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT1, 1671 EVT VT2, EVT VT3, ArrayRef<SDValue> Ops); 1672 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, EVT VT1, 1673 EVT VT2, SDValue Op1, SDValue Op2); 1674 SDNode *SelectNodeTo(SDNode *N, unsigned MachineOpc, SDVTList VTs, 1675 ArrayRef<SDValue> Ops); 1676 1677 /// This *mutates* the specified node to have the specified 1678 /// return type, opcode, and operands. 1679 SDNode *MorphNodeTo(SDNode *N, unsigned Opc, SDVTList VTs, 1680 ArrayRef<SDValue> Ops); 1681 1682 /// Mutate the specified strict FP node to its non-strict equivalent, 1683 /// unlinking the node from its chain and dropping the metadata arguments. 1684 /// The node must be a strict FP node. 1685 SDNode *mutateStrictFPToFP(SDNode *Node); 1686 1687 /// These are used for target selectors to create a new node 1688 /// with specified return type(s), MachineInstr opcode, and operands. 1689 /// 1690 /// Note that getMachineNode returns the resultant node. If there is already 1691 /// a node of the specified opcode and operands, it returns that node instead 1692 /// of the current one. 1693 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT); 1694 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT, 1695 SDValue Op1); 1696 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT, 1697 SDValue Op1, SDValue Op2); 1698 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT, 1699 SDValue Op1, SDValue Op2, SDValue Op3); 1700 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT, 1701 ArrayRef<SDValue> Ops); 1702 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT1, 1703 EVT VT2, SDValue Op1, SDValue Op2); 1704 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT1, 1705 EVT VT2, SDValue Op1, SDValue Op2, SDValue Op3); 1706 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT1, 1707 EVT VT2, ArrayRef<SDValue> Ops); 1708 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT1, 1709 EVT VT2, EVT VT3, SDValue Op1, SDValue Op2); 1710 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT1, 1711 EVT VT2, EVT VT3, SDValue Op1, SDValue Op2, 1712 SDValue Op3); 1713 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, EVT VT1, 1714 EVT VT2, EVT VT3, ArrayRef<SDValue> Ops); 1715 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, 1716 ArrayRef<EVT> ResultTys, ArrayRef<SDValue> Ops); 1717 MachineSDNode *getMachineNode(unsigned Opcode, const SDLoc &dl, SDVTList VTs, 1718 ArrayRef<SDValue> Ops); 1719 1720 /// A convenience function for creating TargetInstrInfo::EXTRACT_SUBREG nodes. 1721 SDValue getTargetExtractSubreg(int SRIdx, const SDLoc &DL, EVT VT, 1722 SDValue Operand); 1723 1724 /// A convenience function for creating TargetInstrInfo::INSERT_SUBREG nodes. 1725 SDValue getTargetInsertSubreg(int SRIdx, const SDLoc &DL, EVT VT, 1726 SDValue Operand, SDValue Subreg); 1727 1728 /// Get the specified node if it's already available, or else return NULL. 1729 SDNode *getNodeIfExists(unsigned Opcode, SDVTList VTList, 1730 ArrayRef<SDValue> Ops, const SDNodeFlags Flags); 1731 SDNode *getNodeIfExists(unsigned Opcode, SDVTList VTList, 1732 ArrayRef<SDValue> Ops); 1733 1734 /// Check if a node exists without modifying its flags. 1735 bool doesNodeExist(unsigned Opcode, SDVTList VTList, ArrayRef<SDValue> Ops); 1736 1737 /// Creates a SDDbgValue node. 1738 SDDbgValue *getDbgValue(DIVariable *Var, DIExpression *Expr, SDNode *N, 1739 unsigned R, bool IsIndirect, const DebugLoc &DL, 1740 unsigned O); 1741 1742 /// Creates a constant SDDbgValue node. 1743 SDDbgValue *getConstantDbgValue(DIVariable *Var, DIExpression *Expr, 1744 const Value *C, const DebugLoc &DL, 1745 unsigned O); 1746 1747 /// Creates a FrameIndex SDDbgValue node. 1748 SDDbgValue *getFrameIndexDbgValue(DIVariable *Var, DIExpression *Expr, 1749 unsigned FI, bool IsIndirect, 1750 const DebugLoc &DL, unsigned O); 1751 1752 /// Creates a FrameIndex SDDbgValue node. 1753 SDDbgValue *getFrameIndexDbgValue(DIVariable *Var, DIExpression *Expr, 1754 unsigned FI, 1755 ArrayRef<SDNode *> Dependencies, 1756 bool IsIndirect, const DebugLoc &DL, 1757 unsigned O); 1758 1759 /// Creates a VReg SDDbgValue node. 1760 SDDbgValue *getVRegDbgValue(DIVariable *Var, DIExpression *Expr, 1761 unsigned VReg, bool IsIndirect, 1762 const DebugLoc &DL, unsigned O); 1763 1764 /// Creates a SDDbgValue node from a list of locations. 1765 SDDbgValue *getDbgValueList(DIVariable *Var, DIExpression *Expr, 1766 ArrayRef<SDDbgOperand> Locs, 1767 ArrayRef<SDNode *> Dependencies, bool IsIndirect, 1768 const DebugLoc &DL, unsigned O, bool IsVariadic); 1769 1770 /// Creates a SDDbgLabel node. 1771 SDDbgLabel *getDbgLabel(DILabel *Label, const DebugLoc &DL, unsigned O); 1772 1773 /// Transfer debug values from one node to another, while optionally 1774 /// generating fragment expressions for split-up values. If \p InvalidateDbg 1775 /// is set, debug values are invalidated after they are transferred. 1776 void transferDbgValues(SDValue From, SDValue To, unsigned OffsetInBits = 0, 1777 unsigned SizeInBits = 0, bool InvalidateDbg = true); 1778 1779 /// Remove the specified node from the system. If any of its 1780 /// operands then becomes dead, remove them as well. Inform UpdateListener 1781 /// for each node deleted. 1782 void RemoveDeadNode(SDNode *N); 1783 1784 /// This method deletes the unreachable nodes in the 1785 /// given list, and any nodes that become unreachable as a result. 1786 void RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes); 1787 1788 /// Modify anything using 'From' to use 'To' instead. 1789 /// This can cause recursive merging of nodes in the DAG. Use the first 1790 /// version if 'From' is known to have a single result, use the second 1791 /// if you have two nodes with identical results (or if 'To' has a superset 1792 /// of the results of 'From'), use the third otherwise. 1793 /// 1794 /// These methods all take an optional UpdateListener, which (if not null) is 1795 /// informed about nodes that are deleted and modified due to recursive 1796 /// changes in the dag. 1797 /// 1798 /// These functions only replace all existing uses. It's possible that as 1799 /// these replacements are being performed, CSE may cause the From node 1800 /// to be given new uses. These new uses of From are left in place, and 1801 /// not automatically transferred to To. 1802 /// 1803 void ReplaceAllUsesWith(SDValue From, SDValue To); 1804 void ReplaceAllUsesWith(SDNode *From, SDNode *To); 1805 void ReplaceAllUsesWith(SDNode *From, const SDValue *To); 1806 1807 /// Replace any uses of From with To, leaving 1808 /// uses of other values produced by From.getNode() alone. 1809 void ReplaceAllUsesOfValueWith(SDValue From, SDValue To); 1810 1811 /// Like ReplaceAllUsesOfValueWith, but for multiple values at once. 1812 /// This correctly handles the case where 1813 /// there is an overlap between the From values and the To values. 1814 void ReplaceAllUsesOfValuesWith(const SDValue *From, const SDValue *To, 1815 unsigned Num); 1816 1817 /// If an existing load has uses of its chain, create a token factor node with 1818 /// that chain and the new memory node's chain and update users of the old 1819 /// chain to the token factor. This ensures that the new memory node will have 1820 /// the same relative memory dependency position as the old load. Returns the 1821 /// new merged load chain. 1822 SDValue makeEquivalentMemoryOrdering(SDValue OldChain, SDValue NewMemOpChain); 1823 1824 /// If an existing load has uses of its chain, create a token factor node with 1825 /// that chain and the new memory node's chain and update users of the old 1826 /// chain to the token factor. This ensures that the new memory node will have 1827 /// the same relative memory dependency position as the old load. Returns the 1828 /// new merged load chain. 1829 SDValue makeEquivalentMemoryOrdering(LoadSDNode *OldLoad, SDValue NewMemOp); 1830 1831 /// Topological-sort the AllNodes list and a 1832 /// assign a unique node id for each node in the DAG based on their 1833 /// topological order. Returns the number of nodes. 1834 unsigned AssignTopologicalOrder(); 1835 1836 /// Move node N in the AllNodes list to be immediately 1837 /// before the given iterator Position. This may be used to update the 1838 /// topological ordering when the list of nodes is modified. 1839 void RepositionNode(allnodes_iterator Position, SDNode *N) { 1840 AllNodes.insert(Position, AllNodes.remove(N)); 1841 } 1842 1843 /// Returns an APFloat semantics tag appropriate for the given type. If VT is 1844 /// a vector type, the element semantics are returned. 1845 static const fltSemantics &EVTToAPFloatSemantics(EVT VT) { 1846 switch (VT.getScalarType().getSimpleVT().SimpleTy) { 1847 default: llvm_unreachable("Unknown FP format"); 1848 case MVT::f16: return APFloat::IEEEhalf(); 1849 case MVT::bf16: return APFloat::BFloat(); 1850 case MVT::f32: return APFloat::IEEEsingle(); 1851 case MVT::f64: return APFloat::IEEEdouble(); 1852 case MVT::f80: return APFloat::x87DoubleExtended(); 1853 case MVT::f128: return APFloat::IEEEquad(); 1854 case MVT::ppcf128: return APFloat::PPCDoubleDouble(); 1855 } 1856 } 1857 1858 /// Add a dbg_value SDNode. If SD is non-null that means the 1859 /// value is produced by SD. 1860 void AddDbgValue(SDDbgValue *DB, bool isParameter); 1861 1862 /// Add a dbg_label SDNode. 1863 void AddDbgLabel(SDDbgLabel *DB); 1864 1865 /// Get the debug values which reference the given SDNode. 1866 ArrayRef<SDDbgValue*> GetDbgValues(const SDNode* SD) const { 1867 return DbgInfo->getSDDbgValues(SD); 1868 } 1869 1870 public: 1871 /// Return true if there are any SDDbgValue nodes associated 1872 /// with this SelectionDAG. 1873 bool hasDebugValues() const { return !DbgInfo->empty(); } 1874 1875 SDDbgInfo::DbgIterator DbgBegin() const { return DbgInfo->DbgBegin(); } 1876 SDDbgInfo::DbgIterator DbgEnd() const { return DbgInfo->DbgEnd(); } 1877 1878 SDDbgInfo::DbgIterator ByvalParmDbgBegin() const { 1879 return DbgInfo->ByvalParmDbgBegin(); 1880 } 1881 SDDbgInfo::DbgIterator ByvalParmDbgEnd() const { 1882 return DbgInfo->ByvalParmDbgEnd(); 1883 } 1884 1885 SDDbgInfo::DbgLabelIterator DbgLabelBegin() const { 1886 return DbgInfo->DbgLabelBegin(); 1887 } 1888 SDDbgInfo::DbgLabelIterator DbgLabelEnd() const { 1889 return DbgInfo->DbgLabelEnd(); 1890 } 1891 1892 /// To be invoked on an SDNode that is slated to be erased. This 1893 /// function mirrors \c llvm::salvageDebugInfo. 1894 void salvageDebugInfo(SDNode &N); 1895 1896 void dump() const; 1897 1898 /// In most cases this function returns the ABI alignment for a given type, 1899 /// except for illegal vector types where the alignment exceeds that of the 1900 /// stack. In such cases we attempt to break the vector down to a legal type 1901 /// and return the ABI alignment for that instead. 1902 Align getReducedAlign(EVT VT, bool UseABI); 1903 1904 /// Create a stack temporary based on the size in bytes and the alignment 1905 SDValue CreateStackTemporary(TypeSize Bytes, Align Alignment); 1906 1907 /// Create a stack temporary, suitable for holding the specified value type. 1908 /// If minAlign is specified, the slot size will have at least that alignment. 1909 SDValue CreateStackTemporary(EVT VT, unsigned minAlign = 1); 1910 1911 /// Create a stack temporary suitable for holding either of the specified 1912 /// value types. 1913 SDValue CreateStackTemporary(EVT VT1, EVT VT2); 1914 1915 SDValue FoldSymbolOffset(unsigned Opcode, EVT VT, 1916 const GlobalAddressSDNode *GA, 1917 const SDNode *N2); 1918 1919 SDValue FoldConstantArithmetic(unsigned Opcode, const SDLoc &DL, EVT VT, 1920 ArrayRef<SDValue> Ops); 1921 1922 /// Fold floating-point operations when all operands are constants and/or 1923 /// undefined. 1924 SDValue foldConstantFPMath(unsigned Opcode, const SDLoc &DL, EVT VT, 1925 ArrayRef<SDValue> Ops); 1926 1927 /// Constant fold a setcc to true or false. 1928 SDValue FoldSetCC(EVT VT, SDValue N1, SDValue N2, ISD::CondCode Cond, 1929 const SDLoc &dl); 1930 1931 /// Return true if the sign bit of Op is known to be zero. 1932 /// We use this predicate to simplify operations downstream. 1933 bool SignBitIsZero(SDValue Op, unsigned Depth = 0) const; 1934 1935 /// Return true if 'Op & Mask' is known to be zero. We 1936 /// use this predicate to simplify operations downstream. Op and Mask are 1937 /// known to be the same type. 1938 bool MaskedValueIsZero(SDValue Op, const APInt &Mask, 1939 unsigned Depth = 0) const; 1940 1941 /// Return true if 'Op & Mask' is known to be zero in DemandedElts. We 1942 /// use this predicate to simplify operations downstream. Op and Mask are 1943 /// known to be the same type. 1944 bool MaskedValueIsZero(SDValue Op, const APInt &Mask, 1945 const APInt &DemandedElts, unsigned Depth = 0) const; 1946 1947 /// Return true if 'Op' is known to be zero in DemandedElts. We 1948 /// use this predicate to simplify operations downstream. 1949 bool MaskedVectorIsZero(SDValue Op, const APInt &DemandedElts, 1950 unsigned Depth = 0) const; 1951 1952 /// Return true if '(Op & Mask) == Mask'. 1953 /// Op and Mask are known to be the same type. 1954 bool MaskedValueIsAllOnes(SDValue Op, const APInt &Mask, 1955 unsigned Depth = 0) const; 1956 1957 /// For each demanded element of a vector, see if it is known to be zero. 1958 APInt computeVectorKnownZeroElements(SDValue Op, const APInt &DemandedElts, 1959 unsigned Depth = 0) const; 1960 1961 /// Determine which bits of Op are known to be either zero or one and return 1962 /// them in Known. For vectors, the known bits are those that are shared by 1963 /// every vector element. 1964 /// Targets can implement the computeKnownBitsForTargetNode method in the 1965 /// TargetLowering class to allow target nodes to be understood. 1966 KnownBits computeKnownBits(SDValue Op, unsigned Depth = 0) const; 1967 1968 /// Determine which bits of Op are known to be either zero or one and return 1969 /// them in Known. The DemandedElts argument allows us to only collect the 1970 /// known bits that are shared by the requested vector elements. 1971 /// Targets can implement the computeKnownBitsForTargetNode method in the 1972 /// TargetLowering class to allow target nodes to be understood. 1973 KnownBits computeKnownBits(SDValue Op, const APInt &DemandedElts, 1974 unsigned Depth = 0) const; 1975 1976 /// Used to represent the possible overflow behavior of an operation. 1977 /// Never: the operation cannot overflow. 1978 /// Always: the operation will always overflow. 1979 /// Sometime: the operation may or may not overflow. 1980 enum OverflowKind { 1981 OFK_Never, 1982 OFK_Sometime, 1983 OFK_Always, 1984 }; 1985 1986 /// Determine if the result of the signed addition of 2 nodes can overflow. 1987 OverflowKind computeOverflowForSignedAdd(SDValue N0, SDValue N1) const; 1988 1989 /// Determine if the result of the unsigned addition of 2 nodes can overflow. 1990 OverflowKind computeOverflowForUnsignedAdd(SDValue N0, SDValue N1) const; 1991 1992 /// Determine if the result of the addition of 2 nodes can overflow. 1993 OverflowKind computeOverflowForAdd(bool IsSigned, SDValue N0, 1994 SDValue N1) const { 1995 return IsSigned ? computeOverflowForSignedAdd(N0, N1) 1996 : computeOverflowForUnsignedAdd(N0, N1); 1997 } 1998 1999 /// Determine if the result of the addition of 2 nodes can never overflow. 2000 bool willNotOverflowAdd(bool IsSigned, SDValue N0, SDValue N1) const { 2001 return computeOverflowForAdd(IsSigned, N0, N1) == OFK_Never; 2002 } 2003 2004 /// Determine if the result of the signed sub of 2 nodes can overflow. 2005 OverflowKind computeOverflowForSignedSub(SDValue N0, SDValue N1) const; 2006 2007 /// Determine if the result of the unsigned sub of 2 nodes can overflow. 2008 OverflowKind computeOverflowForUnsignedSub(SDValue N0, SDValue N1) const; 2009 2010 /// Determine if the result of the sub of 2 nodes can overflow. 2011 OverflowKind computeOverflowForSub(bool IsSigned, SDValue N0, 2012 SDValue N1) const { 2013 return IsSigned ? computeOverflowForSignedSub(N0, N1) 2014 : computeOverflowForUnsignedSub(N0, N1); 2015 } 2016 2017 /// Determine if the result of the sub of 2 nodes can never overflow. 2018 bool willNotOverflowSub(bool IsSigned, SDValue N0, SDValue N1) const { 2019 return computeOverflowForSub(IsSigned, N0, N1) == OFK_Never; 2020 } 2021 2022 /// Determine if the result of the signed mul of 2 nodes can overflow. 2023 OverflowKind computeOverflowForSignedMul(SDValue N0, SDValue N1) const; 2024 2025 /// Determine if the result of the unsigned mul of 2 nodes can overflow. 2026 OverflowKind computeOverflowForUnsignedMul(SDValue N0, SDValue N1) const; 2027 2028 /// Determine if the result of the mul of 2 nodes can overflow. 2029 OverflowKind computeOverflowForMul(bool IsSigned, SDValue N0, 2030 SDValue N1) const { 2031 return IsSigned ? computeOverflowForSignedMul(N0, N1) 2032 : computeOverflowForUnsignedMul(N0, N1); 2033 } 2034 2035 /// Determine if the result of the mul of 2 nodes can never overflow. 2036 bool willNotOverflowMul(bool IsSigned, SDValue N0, SDValue N1) const { 2037 return computeOverflowForMul(IsSigned, N0, N1) == OFK_Never; 2038 } 2039 2040 /// Test if the given value is known to have exactly one bit set. This differs 2041 /// from computeKnownBits in that it doesn't necessarily determine which bit 2042 /// is set. 2043 bool isKnownToBeAPowerOfTwo(SDValue Val, unsigned Depth = 0) const; 2044 2045 /// Return the number of times the sign bit of the register is replicated into 2046 /// the other bits. We know that at least 1 bit is always equal to the sign 2047 /// bit (itself), but other cases can give us information. For example, 2048 /// immediately after an "SRA X, 2", we know that the top 3 bits are all equal 2049 /// to each other, so we return 3. Targets can implement the 2050 /// ComputeNumSignBitsForTarget method in the TargetLowering class to allow 2051 /// target nodes to be understood. 2052 unsigned ComputeNumSignBits(SDValue Op, unsigned Depth = 0) const; 2053 2054 /// Return the number of times the sign bit of the register is replicated into 2055 /// the other bits. We know that at least 1 bit is always equal to the sign 2056 /// bit (itself), but other cases can give us information. For example, 2057 /// immediately after an "SRA X, 2", we know that the top 3 bits are all equal 2058 /// to each other, so we return 3. The DemandedElts argument allows 2059 /// us to only collect the minimum sign bits of the requested vector elements. 2060 /// Targets can implement the ComputeNumSignBitsForTarget method in the 2061 /// TargetLowering class to allow target nodes to be understood. 2062 unsigned ComputeNumSignBits(SDValue Op, const APInt &DemandedElts, 2063 unsigned Depth = 0) const; 2064 2065 /// Get the upper bound on bit size for this Value \p Op as a signed integer. 2066 /// i.e. x == sext(trunc(x to MaxSignedBits) to bitwidth(x)). 2067 /// Similar to the APInt::getSignificantBits function. 2068 /// Helper wrapper to ComputeNumSignBits. 2069 unsigned ComputeMaxSignificantBits(SDValue Op, unsigned Depth = 0) const; 2070 2071 /// Get the upper bound on bit size for this Value \p Op as a signed integer. 2072 /// i.e. x == sext(trunc(x to MaxSignedBits) to bitwidth(x)). 2073 /// Similar to the APInt::getSignificantBits function. 2074 /// Helper wrapper to ComputeNumSignBits. 2075 unsigned ComputeMaxSignificantBits(SDValue Op, const APInt &DemandedElts, 2076 unsigned Depth = 0) const; 2077 2078 /// Return true if this function can prove that \p Op is never poison 2079 /// and, if \p PoisonOnly is false, does not have undef bits. 2080 bool isGuaranteedNotToBeUndefOrPoison(SDValue Op, bool PoisonOnly = false, 2081 unsigned Depth = 0) const; 2082 2083 /// Return true if this function can prove that \p Op is never poison 2084 /// and, if \p PoisonOnly is false, does not have undef bits. The DemandedElts 2085 /// argument limits the check to the requested vector elements. 2086 bool isGuaranteedNotToBeUndefOrPoison(SDValue Op, const APInt &DemandedElts, 2087 bool PoisonOnly = false, 2088 unsigned Depth = 0) const; 2089 2090 /// Return true if this function can prove that \p Op is never poison. 2091 bool isGuaranteedNotToBePoison(SDValue Op, unsigned Depth = 0) const { 2092 return isGuaranteedNotToBeUndefOrPoison(Op, /*PoisonOnly*/ true, Depth); 2093 } 2094 2095 /// Return true if this function can prove that \p Op is never poison. The 2096 /// DemandedElts argument limits the check to the requested vector elements. 2097 bool isGuaranteedNotToBePoison(SDValue Op, const APInt &DemandedElts, 2098 unsigned Depth = 0) const { 2099 return isGuaranteedNotToBeUndefOrPoison(Op, DemandedElts, 2100 /*PoisonOnly*/ true, Depth); 2101 } 2102 2103 /// Return true if Op can create undef or poison from non-undef & non-poison 2104 /// operands. The DemandedElts argument limits the check to the requested 2105 /// vector elements. 2106 /// 2107 /// \p ConsiderFlags controls whether poison producing flags on the 2108 /// instruction are considered. This can be used to see if the instruction 2109 /// could still introduce undef or poison even without poison generating flags 2110 /// which might be on the instruction. (i.e. could the result of 2111 /// Op->dropPoisonGeneratingFlags() still create poison or undef) 2112 bool canCreateUndefOrPoison(SDValue Op, const APInt &DemandedElts, 2113 bool PoisonOnly = false, 2114 bool ConsiderFlags = true, 2115 unsigned Depth = 0) const; 2116 2117 /// Return true if Op can create undef or poison from non-undef & non-poison 2118 /// operands. 2119 /// 2120 /// \p ConsiderFlags controls whether poison producing flags on the 2121 /// instruction are considered. This can be used to see if the instruction 2122 /// could still introduce undef or poison even without poison generating flags 2123 /// which might be on the instruction. (i.e. could the result of 2124 /// Op->dropPoisonGeneratingFlags() still create poison or undef) 2125 bool canCreateUndefOrPoison(SDValue Op, bool PoisonOnly = false, 2126 bool ConsiderFlags = true, 2127 unsigned Depth = 0) const; 2128 2129 /// Return true if the specified operand is an ISD::OR or ISD::XOR node 2130 /// that can be treated as an ISD::ADD node. 2131 /// or(x,y) == add(x,y) iff haveNoCommonBitsSet(x,y) 2132 /// xor(x,y) == add(x,y) iff isMinSignedConstant(y) 2133 bool isADDLike(SDValue Op) const; 2134 2135 /// Return true if the specified operand is an ISD::ADD with a ConstantSDNode 2136 /// on the right-hand side, or if it is an ISD::OR with a ConstantSDNode that 2137 /// is guaranteed to have the same semantics as an ADD. This handles the 2138 /// equivalence: 2139 /// X|Cst == X+Cst iff X&Cst = 0. 2140 bool isBaseWithConstantOffset(SDValue Op) const; 2141 2142 /// Test whether the given SDValue (or all elements of it, if it is a 2143 /// vector) is known to never be NaN. If \p SNaN is true, returns if \p Op is 2144 /// known to never be a signaling NaN (it may still be a qNaN). 2145 bool isKnownNeverNaN(SDValue Op, bool SNaN = false, unsigned Depth = 0) const; 2146 2147 /// \returns true if \p Op is known to never be a signaling NaN. 2148 bool isKnownNeverSNaN(SDValue Op, unsigned Depth = 0) const { 2149 return isKnownNeverNaN(Op, true, Depth); 2150 } 2151 2152 /// Test whether the given floating point SDValue is known to never be 2153 /// positive or negative zero. 2154 bool isKnownNeverZeroFloat(SDValue Op) const; 2155 2156 /// Test whether the given SDValue is known to contain non-zero value(s). 2157 bool isKnownNeverZero(SDValue Op, unsigned Depth = 0) const; 2158 2159 /// Test whether two SDValues are known to compare equal. This 2160 /// is true if they are the same value, or if one is negative zero and the 2161 /// other positive zero. 2162 bool isEqualTo(SDValue A, SDValue B) const; 2163 2164 /// Return true if A and B have no common bits set. As an example, this can 2165 /// allow an 'add' to be transformed into an 'or'. 2166 bool haveNoCommonBitsSet(SDValue A, SDValue B) const; 2167 2168 /// Test whether \p V has a splatted value for all the demanded elements. 2169 /// 2170 /// On success \p UndefElts will indicate the elements that have UNDEF 2171 /// values instead of the splat value, this is only guaranteed to be correct 2172 /// for \p DemandedElts. 2173 /// 2174 /// NOTE: The function will return true for a demanded splat of UNDEF values. 2175 bool isSplatValue(SDValue V, const APInt &DemandedElts, APInt &UndefElts, 2176 unsigned Depth = 0) const; 2177 2178 /// Test whether \p V has a splatted value. 2179 bool isSplatValue(SDValue V, bool AllowUndefs = false) const; 2180 2181 /// If V is a splatted value, return the source vector and its splat index. 2182 SDValue getSplatSourceVector(SDValue V, int &SplatIndex); 2183 2184 /// If V is a splat vector, return its scalar source operand by extracting 2185 /// that element from the source vector. If LegalTypes is true, this method 2186 /// may only return a legally-typed splat value. If it cannot legalize the 2187 /// splatted value it will return SDValue(). 2188 SDValue getSplatValue(SDValue V, bool LegalTypes = false); 2189 2190 /// If a SHL/SRA/SRL node \p V has a constant or splat constant shift amount 2191 /// that is less than the element bit-width of the shift node, return it. 2192 const APInt *getValidShiftAmountConstant(SDValue V, 2193 const APInt &DemandedElts) const; 2194 2195 /// If a SHL/SRA/SRL node \p V has constant shift amounts that are all less 2196 /// than the element bit-width of the shift node, return the minimum value. 2197 const APInt * 2198 getValidMinimumShiftAmountConstant(SDValue V, 2199 const APInt &DemandedElts) const; 2200 2201 /// If a SHL/SRA/SRL node \p V has constant shift amounts that are all less 2202 /// than the element bit-width of the shift node, return the maximum value. 2203 const APInt * 2204 getValidMaximumShiftAmountConstant(SDValue V, 2205 const APInt &DemandedElts) const; 2206 2207 /// Match a binop + shuffle pyramid that represents a horizontal reduction 2208 /// over the elements of a vector starting from the EXTRACT_VECTOR_ELT node /p 2209 /// Extract. The reduction must use one of the opcodes listed in /p 2210 /// CandidateBinOps and on success /p BinOp will contain the matching opcode. 2211 /// Returns the vector that is being reduced on, or SDValue() if a reduction 2212 /// was not matched. If \p AllowPartials is set then in the case of a 2213 /// reduction pattern that only matches the first few stages, the extracted 2214 /// subvector of the start of the reduction is returned. 2215 SDValue matchBinOpReduction(SDNode *Extract, ISD::NodeType &BinOp, 2216 ArrayRef<ISD::NodeType> CandidateBinOps, 2217 bool AllowPartials = false); 2218 2219 /// Utility function used by legalize and lowering to 2220 /// "unroll" a vector operation by splitting out the scalars and operating 2221 /// on each element individually. If the ResNE is 0, fully unroll the vector 2222 /// op. If ResNE is less than the width of the vector op, unroll up to ResNE. 2223 /// If the ResNE is greater than the width of the vector op, unroll the 2224 /// vector op and fill the end of the resulting vector with UNDEFS. 2225 SDValue UnrollVectorOp(SDNode *N, unsigned ResNE = 0); 2226 2227 /// Like UnrollVectorOp(), but for the [US](ADD|SUB|MUL)O family of opcodes. 2228 /// This is a separate function because those opcodes have two results. 2229 std::pair<SDValue, SDValue> UnrollVectorOverflowOp(SDNode *N, 2230 unsigned ResNE = 0); 2231 2232 /// Return true if loads are next to each other and can be 2233 /// merged. Check that both are nonvolatile and if LD is loading 2234 /// 'Bytes' bytes from a location that is 'Dist' units away from the 2235 /// location that the 'Base' load is loading from. 2236 bool areNonVolatileConsecutiveLoads(LoadSDNode *LD, LoadSDNode *Base, 2237 unsigned Bytes, int Dist) const; 2238 2239 /// Infer alignment of a load / store address. Return std::nullopt if it 2240 /// cannot be inferred. 2241 MaybeAlign InferPtrAlign(SDValue Ptr) const; 2242 2243 /// Split the scalar node with EXTRACT_ELEMENT using the provided VTs and 2244 /// return the low/high part. 2245 std::pair<SDValue, SDValue> SplitScalar(const SDValue &N, const SDLoc &DL, 2246 const EVT &LoVT, const EVT &HiVT); 2247 2248 /// Compute the VTs needed for the low/hi parts of a type 2249 /// which is split (or expanded) into two not necessarily identical pieces. 2250 std::pair<EVT, EVT> GetSplitDestVTs(const EVT &VT) const; 2251 2252 /// Compute the VTs needed for the low/hi parts of a type, dependent on an 2253 /// enveloping VT that has been split into two identical pieces. Sets the 2254 /// HisIsEmpty flag when hi type has zero storage size. 2255 std::pair<EVT, EVT> GetDependentSplitDestVTs(const EVT &VT, const EVT &EnvVT, 2256 bool *HiIsEmpty) const; 2257 2258 /// Split the vector with EXTRACT_SUBVECTOR using the provides 2259 /// VTs and return the low/high part. 2260 std::pair<SDValue, SDValue> SplitVector(const SDValue &N, const SDLoc &DL, 2261 const EVT &LoVT, const EVT &HiVT); 2262 2263 /// Split the vector with EXTRACT_SUBVECTOR and return the low/high part. 2264 std::pair<SDValue, SDValue> SplitVector(const SDValue &N, const SDLoc &DL) { 2265 EVT LoVT, HiVT; 2266 std::tie(LoVT, HiVT) = GetSplitDestVTs(N.getValueType()); 2267 return SplitVector(N, DL, LoVT, HiVT); 2268 } 2269 2270 /// Split the explicit vector length parameter of a VP operation. 2271 std::pair<SDValue, SDValue> SplitEVL(SDValue N, EVT VecVT, const SDLoc &DL); 2272 2273 /// Split the node's operand with EXTRACT_SUBVECTOR and 2274 /// return the low/high part. 2275 std::pair<SDValue, SDValue> SplitVectorOperand(const SDNode *N, unsigned OpNo) 2276 { 2277 return SplitVector(N->getOperand(OpNo), SDLoc(N)); 2278 } 2279 2280 /// Widen the vector up to the next power of two using INSERT_SUBVECTOR. 2281 SDValue WidenVector(const SDValue &N, const SDLoc &DL); 2282 2283 /// Append the extracted elements from Start to Count out of the vector Op in 2284 /// Args. If Count is 0, all of the elements will be extracted. The extracted 2285 /// elements will have type EVT if it is provided, and otherwise their type 2286 /// will be Op's element type. 2287 void ExtractVectorElements(SDValue Op, SmallVectorImpl<SDValue> &Args, 2288 unsigned Start = 0, unsigned Count = 0, 2289 EVT EltVT = EVT()); 2290 2291 /// Compute the default alignment value for the given type. 2292 Align getEVTAlign(EVT MemoryVT) const; 2293 2294 /// Test whether the given value is a constant int or similar node. 2295 SDNode *isConstantIntBuildVectorOrConstantInt(SDValue N) const; 2296 2297 /// Test whether the given value is a constant FP or similar node. 2298 SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) const ; 2299 2300 /// \returns true if \p N is any kind of constant or build_vector of 2301 /// constants, int or float. If a vector, it may not necessarily be a splat. 2302 inline bool isConstantValueOfAnyType(SDValue N) const { 2303 return isConstantIntBuildVectorOrConstantInt(N) || 2304 isConstantFPBuildVectorOrConstantFP(N); 2305 } 2306 2307 /// Set CallSiteInfo to be associated with Node. 2308 void addCallSiteInfo(const SDNode *Node, CallSiteInfoImpl &&CallInfo) { 2309 SDEI[Node].CSInfo = std::move(CallInfo); 2310 } 2311 /// Return CallSiteInfo associated with Node, or a default if none exists. 2312 CallSiteInfo getCallSiteInfo(const SDNode *Node) { 2313 auto I = SDEI.find(Node); 2314 return I != SDEI.end() ? std::move(I->second).CSInfo : CallSiteInfo(); 2315 } 2316 /// Set HeapAllocSite to be associated with Node. 2317 void addHeapAllocSite(const SDNode *Node, MDNode *MD) { 2318 SDEI[Node].HeapAllocSite = MD; 2319 } 2320 /// Return HeapAllocSite associated with Node, or nullptr if none exists. 2321 MDNode *getHeapAllocSite(const SDNode *Node) const { 2322 auto I = SDEI.find(Node); 2323 return I != SDEI.end() ? I->second.HeapAllocSite : nullptr; 2324 } 2325 /// Set PCSections to be associated with Node. 2326 void addPCSections(const SDNode *Node, MDNode *MD) { 2327 SDEI[Node].PCSections = MD; 2328 } 2329 /// Return PCSections associated with Node, or nullptr if none exists. 2330 MDNode *getPCSections(const SDNode *Node) const { 2331 auto It = SDEI.find(Node); 2332 return It != SDEI.end() ? It->second.PCSections : nullptr; 2333 } 2334 /// Set NoMergeSiteInfo to be associated with Node if NoMerge is true. 2335 void addNoMergeSiteInfo(const SDNode *Node, bool NoMerge) { 2336 if (NoMerge) 2337 SDEI[Node].NoMerge = NoMerge; 2338 } 2339 /// Return NoMerge info associated with Node. 2340 bool getNoMergeSiteInfo(const SDNode *Node) const { 2341 auto I = SDEI.find(Node); 2342 return I != SDEI.end() ? I->second.NoMerge : false; 2343 } 2344 2345 /// Copy extra info associated with one node to another. 2346 void copyExtraInfo(SDNode *From, SDNode *To); 2347 2348 /// Return the current function's default denormal handling kind for the given 2349 /// floating point type. 2350 DenormalMode getDenormalMode(EVT VT) const { 2351 return MF->getDenormalMode(EVTToAPFloatSemantics(VT)); 2352 } 2353 2354 bool shouldOptForSize() const; 2355 2356 /// Get the (commutative) neutral element for the given opcode, if it exists. 2357 SDValue getNeutralElement(unsigned Opcode, const SDLoc &DL, EVT VT, 2358 SDNodeFlags Flags); 2359 2360 /// Some opcodes may create immediate undefined behavior when used with some 2361 /// values (integer division-by-zero for example). Therefore, these operations 2362 /// are not generally safe to move around or change. 2363 bool isSafeToSpeculativelyExecute(unsigned Opcode) const { 2364 switch (Opcode) { 2365 case ISD::SDIV: 2366 case ISD::SREM: 2367 case ISD::SDIVREM: 2368 case ISD::UDIV: 2369 case ISD::UREM: 2370 case ISD::UDIVREM: 2371 return false; 2372 default: 2373 return true; 2374 } 2375 } 2376 2377 /// Check if the provided node is save to speculatively executed given its 2378 /// current arguments. So, while `udiv` the opcode is not safe to 2379 /// speculatively execute, a given `udiv` node may be if the denominator is 2380 /// known nonzero. 2381 bool isSafeToSpeculativelyExecuteNode(const SDNode *N) const { 2382 switch (N->getOpcode()) { 2383 case ISD::UDIV: 2384 return isKnownNeverZero(N->getOperand(1)); 2385 default: 2386 return isSafeToSpeculativelyExecute(N->getOpcode()); 2387 } 2388 } 2389 2390 SDValue makeStateFunctionCall(unsigned LibFunc, SDValue Ptr, SDValue InChain, 2391 const SDLoc &DLoc); 2392 2393 private: 2394 void InsertNode(SDNode *N); 2395 bool RemoveNodeFromCSEMaps(SDNode *N); 2396 void AddModifiedNodeToCSEMaps(SDNode *N); 2397 SDNode *FindModifiedNodeSlot(SDNode *N, SDValue Op, void *&InsertPos); 2398 SDNode *FindModifiedNodeSlot(SDNode *N, SDValue Op1, SDValue Op2, 2399 void *&InsertPos); 2400 SDNode *FindModifiedNodeSlot(SDNode *N, ArrayRef<SDValue> Ops, 2401 void *&InsertPos); 2402 SDNode *UpdateSDLocOnMergeSDNode(SDNode *N, const SDLoc &loc); 2403 2404 void DeleteNodeNotInCSEMaps(SDNode *N); 2405 void DeallocateNode(SDNode *N); 2406 2407 void allnodes_clear(); 2408 2409 /// Look up the node specified by ID in CSEMap. If it exists, return it. If 2410 /// not, return the insertion token that will make insertion faster. This 2411 /// overload is for nodes other than Constant or ConstantFP, use the other one 2412 /// for those. 2413 SDNode *FindNodeOrInsertPos(const FoldingSetNodeID &ID, void *&InsertPos); 2414 2415 /// Look up the node specified by ID in CSEMap. If it exists, return it. If 2416 /// not, return the insertion token that will make insertion faster. Performs 2417 /// additional processing for constant nodes. 2418 SDNode *FindNodeOrInsertPos(const FoldingSetNodeID &ID, const SDLoc &DL, 2419 void *&InsertPos); 2420 2421 /// Maps to auto-CSE operations. 2422 std::vector<CondCodeSDNode*> CondCodeNodes; 2423 2424 std::vector<SDNode*> ValueTypeNodes; 2425 std::map<EVT, SDNode*, EVT::compareRawBits> ExtendedValueTypeNodes; 2426 StringMap<SDNode*> ExternalSymbols; 2427 2428 std::map<std::pair<std::string, unsigned>, SDNode *> TargetExternalSymbols; 2429 DenseMap<MCSymbol *, SDNode *> MCSymbols; 2430 2431 FlagInserter *Inserter = nullptr; 2432 }; 2433 2434 template <> struct GraphTraits<SelectionDAG*> : public GraphTraits<SDNode*> { 2435 using nodes_iterator = pointer_iterator<SelectionDAG::allnodes_iterator>; 2436 2437 static nodes_iterator nodes_begin(SelectionDAG *G) { 2438 return nodes_iterator(G->allnodes_begin()); 2439 } 2440 2441 static nodes_iterator nodes_end(SelectionDAG *G) { 2442 return nodes_iterator(G->allnodes_end()); 2443 } 2444 }; 2445 2446 } // end namespace llvm 2447 2448 #endif // LLVM_CODEGEN_SELECTIONDAG_H 2449