| File: | llvm/include/llvm/CodeGen/SelectionDAGNodes.h |
| Warning: | line 1332, column 12 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- DAGCombiner.cpp - Implement a DAG node combiner --------------------===// |
| 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 pass combines dag nodes to form fewer, simpler DAG nodes. It can be run |
| 10 | // both before and after the DAG is legalized. |
| 11 | // |
| 12 | // This pass is not a substitute for the LLVM IR instcombine pass. This pass is |
| 13 | // primarily intended to handle simplification opportunities that are implicit |
| 14 | // in the LLVM IR and exposed by the various codegen lowering phases. |
| 15 | // |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | |
| 18 | #include "llvm/ADT/APFloat.h" |
| 19 | #include "llvm/ADT/APInt.h" |
| 20 | #include "llvm/ADT/ArrayRef.h" |
| 21 | #include "llvm/ADT/DenseMap.h" |
| 22 | #include "llvm/ADT/IntervalMap.h" |
| 23 | #include "llvm/ADT/None.h" |
| 24 | #include "llvm/ADT/Optional.h" |
| 25 | #include "llvm/ADT/STLExtras.h" |
| 26 | #include "llvm/ADT/SetVector.h" |
| 27 | #include "llvm/ADT/SmallBitVector.h" |
| 28 | #include "llvm/ADT/SmallPtrSet.h" |
| 29 | #include "llvm/ADT/SmallSet.h" |
| 30 | #include "llvm/ADT/SmallVector.h" |
| 31 | #include "llvm/ADT/Statistic.h" |
| 32 | #include "llvm/Analysis/AliasAnalysis.h" |
| 33 | #include "llvm/Analysis/MemoryLocation.h" |
| 34 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 35 | #include "llvm/Analysis/VectorUtils.h" |
| 36 | #include "llvm/CodeGen/DAGCombine.h" |
| 37 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 38 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 39 | #include "llvm/CodeGen/MachineFunction.h" |
| 40 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 41 | #include "llvm/CodeGen/RuntimeLibcalls.h" |
| 42 | #include "llvm/CodeGen/SelectionDAG.h" |
| 43 | #include "llvm/CodeGen/SelectionDAGAddressAnalysis.h" |
| 44 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 45 | #include "llvm/CodeGen/SelectionDAGTargetInfo.h" |
| 46 | #include "llvm/CodeGen/TargetLowering.h" |
| 47 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 48 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 49 | #include "llvm/CodeGen/ValueTypes.h" |
| 50 | #include "llvm/IR/Attributes.h" |
| 51 | #include "llvm/IR/Constant.h" |
| 52 | #include "llvm/IR/DataLayout.h" |
| 53 | #include "llvm/IR/DerivedTypes.h" |
| 54 | #include "llvm/IR/Function.h" |
| 55 | #include "llvm/IR/LLVMContext.h" |
| 56 | #include "llvm/IR/Metadata.h" |
| 57 | #include "llvm/Support/Casting.h" |
| 58 | #include "llvm/Support/CodeGen.h" |
| 59 | #include "llvm/Support/CommandLine.h" |
| 60 | #include "llvm/Support/Compiler.h" |
| 61 | #include "llvm/Support/Debug.h" |
| 62 | #include "llvm/Support/ErrorHandling.h" |
| 63 | #include "llvm/Support/KnownBits.h" |
| 64 | #include "llvm/Support/MachineValueType.h" |
| 65 | #include "llvm/Support/MathExtras.h" |
| 66 | #include "llvm/Support/raw_ostream.h" |
| 67 | #include "llvm/Target/TargetMachine.h" |
| 68 | #include "llvm/Target/TargetOptions.h" |
| 69 | #include <algorithm> |
| 70 | #include <cassert> |
| 71 | #include <cstdint> |
| 72 | #include <functional> |
| 73 | #include <iterator> |
| 74 | #include <string> |
| 75 | #include <tuple> |
| 76 | #include <utility> |
| 77 | |
| 78 | using namespace llvm; |
| 79 | |
| 80 | #define DEBUG_TYPE"dagcombine" "dagcombine" |
| 81 | |
| 82 | STATISTIC(NodesCombined , "Number of dag nodes combined")static llvm::Statistic NodesCombined = {"dagcombine", "NodesCombined" , "Number of dag nodes combined"}; |
| 83 | STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created")static llvm::Statistic PreIndexedNodes = {"dagcombine", "PreIndexedNodes" , "Number of pre-indexed nodes created"}; |
| 84 | STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created")static llvm::Statistic PostIndexedNodes = {"dagcombine", "PostIndexedNodes" , "Number of post-indexed nodes created"}; |
| 85 | STATISTIC(OpsNarrowed , "Number of load/op/store narrowed")static llvm::Statistic OpsNarrowed = {"dagcombine", "OpsNarrowed" , "Number of load/op/store narrowed"}; |
| 86 | STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int")static llvm::Statistic LdStFP2Int = {"dagcombine", "LdStFP2Int" , "Number of fp load/store pairs transformed to int"}; |
| 87 | STATISTIC(SlicedLoads, "Number of load sliced")static llvm::Statistic SlicedLoads = {"dagcombine", "SlicedLoads" , "Number of load sliced"}; |
| 88 | STATISTIC(NumFPLogicOpsConv, "Number of logic ops converted to fp ops")static llvm::Statistic NumFPLogicOpsConv = {"dagcombine", "NumFPLogicOpsConv" , "Number of logic ops converted to fp ops"}; |
| 89 | |
| 90 | static cl::opt<bool> |
| 91 | CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, |
| 92 | cl::desc("Enable DAG combiner's use of IR alias analysis")); |
| 93 | |
| 94 | static cl::opt<bool> |
| 95 | UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true), |
| 96 | cl::desc("Enable DAG combiner's use of TBAA")); |
| 97 | |
| 98 | #ifndef NDEBUG |
| 99 | static cl::opt<std::string> |
| 100 | CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden, |
| 101 | cl::desc("Only use DAG-combiner alias analysis in this" |
| 102 | " function")); |
| 103 | #endif |
| 104 | |
| 105 | /// Hidden option to stress test load slicing, i.e., when this option |
| 106 | /// is enabled, load slicing bypasses most of its profitability guards. |
| 107 | static cl::opt<bool> |
| 108 | StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden, |
| 109 | cl::desc("Bypass the profitability model of load slicing"), |
| 110 | cl::init(false)); |
| 111 | |
| 112 | static cl::opt<bool> |
| 113 | MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true), |
| 114 | cl::desc("DAG combiner may split indexing from loads")); |
| 115 | |
| 116 | static cl::opt<bool> |
| 117 | EnableStoreMerging("combiner-store-merging", cl::Hidden, cl::init(true), |
| 118 | cl::desc("DAG combiner enable merging multiple stores " |
| 119 | "into a wider store")); |
| 120 | |
| 121 | static cl::opt<unsigned> TokenFactorInlineLimit( |
| 122 | "combiner-tokenfactor-inline-limit", cl::Hidden, cl::init(2048), |
| 123 | cl::desc("Limit the number of operands to inline for Token Factors")); |
| 124 | |
| 125 | static cl::opt<unsigned> StoreMergeDependenceLimit( |
| 126 | "combiner-store-merge-dependence-limit", cl::Hidden, cl::init(10), |
| 127 | cl::desc("Limit the number of times for the same StoreNode and RootNode " |
| 128 | "to bail out in store merging dependence check")); |
| 129 | |
| 130 | static cl::opt<bool> EnableReduceLoadOpStoreWidth( |
| 131 | "combiner-reduce-load-op-store-width", cl::Hidden, cl::init(true), |
| 132 | cl::desc("DAG cominber enable reducing the width of load/op/store " |
| 133 | "sequence")); |
| 134 | |
| 135 | static cl::opt<bool> EnableShrinkLoadReplaceStoreWithStore( |
| 136 | "combiner-shrink-load-replace-store-with-store", cl::Hidden, cl::init(true), |
| 137 | cl::desc("DAG cominber enable load/<replace bytes>/store with " |
| 138 | "a narrower store")); |
| 139 | |
| 140 | namespace { |
| 141 | |
| 142 | class DAGCombiner { |
| 143 | SelectionDAG &DAG; |
| 144 | const TargetLowering &TLI; |
| 145 | const SelectionDAGTargetInfo *STI; |
| 146 | CombineLevel Level; |
| 147 | CodeGenOpt::Level OptLevel; |
| 148 | bool LegalDAG = false; |
| 149 | bool LegalOperations = false; |
| 150 | bool LegalTypes = false; |
| 151 | bool ForCodeSize; |
| 152 | bool DisableGenericCombines; |
| 153 | |
| 154 | /// Worklist of all of the nodes that need to be simplified. |
| 155 | /// |
| 156 | /// This must behave as a stack -- new nodes to process are pushed onto the |
| 157 | /// back and when processing we pop off of the back. |
| 158 | /// |
| 159 | /// The worklist will not contain duplicates but may contain null entries |
| 160 | /// due to nodes being deleted from the underlying DAG. |
| 161 | SmallVector<SDNode *, 64> Worklist; |
| 162 | |
| 163 | /// Mapping from an SDNode to its position on the worklist. |
| 164 | /// |
| 165 | /// This is used to find and remove nodes from the worklist (by nulling |
| 166 | /// them) when they are deleted from the underlying DAG. It relies on |
| 167 | /// stable indices of nodes within the worklist. |
| 168 | DenseMap<SDNode *, unsigned> WorklistMap; |
| 169 | /// This records all nodes attempted to add to the worklist since we |
| 170 | /// considered a new worklist entry. As we keep do not add duplicate nodes |
| 171 | /// in the worklist, this is different from the tail of the worklist. |
| 172 | SmallSetVector<SDNode *, 32> PruningList; |
| 173 | |
| 174 | /// Set of nodes which have been combined (at least once). |
| 175 | /// |
| 176 | /// This is used to allow us to reliably add any operands of a DAG node |
| 177 | /// which have not yet been combined to the worklist. |
| 178 | SmallPtrSet<SDNode *, 32> CombinedNodes; |
| 179 | |
| 180 | /// Map from candidate StoreNode to the pair of RootNode and count. |
| 181 | /// The count is used to track how many times we have seen the StoreNode |
| 182 | /// with the same RootNode bail out in dependence check. If we have seen |
| 183 | /// the bail out for the same pair many times over a limit, we won't |
| 184 | /// consider the StoreNode with the same RootNode as store merging |
| 185 | /// candidate again. |
| 186 | DenseMap<SDNode *, std::pair<SDNode *, unsigned>> StoreRootCountMap; |
| 187 | |
| 188 | // AA - Used for DAG load/store alias analysis. |
| 189 | AliasAnalysis *AA; |
| 190 | |
| 191 | /// When an instruction is simplified, add all users of the instruction to |
| 192 | /// the work lists because they might get more simplified now. |
| 193 | void AddUsersToWorklist(SDNode *N) { |
| 194 | for (SDNode *Node : N->uses()) |
| 195 | AddToWorklist(Node); |
| 196 | } |
| 197 | |
| 198 | /// Convenient shorthand to add a node and all of its user to the worklist. |
| 199 | void AddToWorklistWithUsers(SDNode *N) { |
| 200 | AddUsersToWorklist(N); |
| 201 | AddToWorklist(N); |
| 202 | } |
| 203 | |
| 204 | // Prune potentially dangling nodes. This is called after |
| 205 | // any visit to a node, but should also be called during a visit after any |
| 206 | // failed combine which may have created a DAG node. |
| 207 | void clearAddedDanglingWorklistEntries() { |
| 208 | // Check any nodes added to the worklist to see if they are prunable. |
| 209 | while (!PruningList.empty()) { |
| 210 | auto *N = PruningList.pop_back_val(); |
| 211 | if (N->use_empty()) |
| 212 | recursivelyDeleteUnusedNodes(N); |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | SDNode *getNextWorklistEntry() { |
| 217 | // Before we do any work, remove nodes that are not in use. |
| 218 | clearAddedDanglingWorklistEntries(); |
| 219 | SDNode *N = nullptr; |
| 220 | // The Worklist holds the SDNodes in order, but it may contain null |
| 221 | // entries. |
| 222 | while (!N && !Worklist.empty()) { |
| 223 | N = Worklist.pop_back_val(); |
| 224 | } |
| 225 | |
| 226 | if (N) { |
| 227 | bool GoodWorklistEntry = WorklistMap.erase(N); |
| 228 | (void)GoodWorklistEntry; |
| 229 | assert(GoodWorklistEntry &&((GoodWorklistEntry && "Found a worklist entry without a corresponding map entry!" ) ? static_cast<void> (0) : __assert_fail ("GoodWorklistEntry && \"Found a worklist entry without a corresponding map entry!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 230, __PRETTY_FUNCTION__)) |
| 230 | "Found a worklist entry without a corresponding map entry!")((GoodWorklistEntry && "Found a worklist entry without a corresponding map entry!" ) ? static_cast<void> (0) : __assert_fail ("GoodWorklistEntry && \"Found a worklist entry without a corresponding map entry!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 230, __PRETTY_FUNCTION__)); |
| 231 | } |
| 232 | return N; |
| 233 | } |
| 234 | |
| 235 | /// Call the node-specific routine that folds each particular type of node. |
| 236 | SDValue visit(SDNode *N); |
| 237 | |
| 238 | public: |
| 239 | DAGCombiner(SelectionDAG &D, AliasAnalysis *AA, CodeGenOpt::Level OL) |
| 240 | : DAG(D), TLI(D.getTargetLoweringInfo()), |
| 241 | STI(D.getSubtarget().getSelectionDAGInfo()), |
| 242 | Level(BeforeLegalizeTypes), OptLevel(OL), AA(AA) { |
| 243 | ForCodeSize = DAG.shouldOptForSize(); |
| 244 | DisableGenericCombines = STI && STI->disableGenericCombines(OptLevel); |
| 245 | |
| 246 | MaximumLegalStoreInBits = 0; |
| 247 | // We use the minimum store size here, since that's all we can guarantee |
| 248 | // for the scalable vector types. |
| 249 | for (MVT VT : MVT::all_valuetypes()) |
| 250 | if (EVT(VT).isSimple() && VT != MVT::Other && |
| 251 | TLI.isTypeLegal(EVT(VT)) && |
| 252 | VT.getSizeInBits().getKnownMinSize() >= MaximumLegalStoreInBits) |
| 253 | MaximumLegalStoreInBits = VT.getSizeInBits().getKnownMinSize(); |
| 254 | } |
| 255 | |
| 256 | void ConsiderForPruning(SDNode *N) { |
| 257 | // Mark this for potential pruning. |
| 258 | PruningList.insert(N); |
| 259 | } |
| 260 | |
| 261 | /// Add to the worklist making sure its instance is at the back (next to be |
| 262 | /// processed.) |
| 263 | void AddToWorklist(SDNode *N) { |
| 264 | assert(N->getOpcode() != ISD::DELETED_NODE &&((N->getOpcode() != ISD::DELETED_NODE && "Deleted Node added to Worklist" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Deleted Node added to Worklist\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 265, __PRETTY_FUNCTION__)) |
| 265 | "Deleted Node added to Worklist")((N->getOpcode() != ISD::DELETED_NODE && "Deleted Node added to Worklist" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Deleted Node added to Worklist\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 265, __PRETTY_FUNCTION__)); |
| 266 | |
| 267 | // Skip handle nodes as they can't usefully be combined and confuse the |
| 268 | // zero-use deletion strategy. |
| 269 | if (N->getOpcode() == ISD::HANDLENODE) |
| 270 | return; |
| 271 | |
| 272 | ConsiderForPruning(N); |
| 273 | |
| 274 | if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second) |
| 275 | Worklist.push_back(N); |
| 276 | } |
| 277 | |
| 278 | /// Remove all instances of N from the worklist. |
| 279 | void removeFromWorklist(SDNode *N) { |
| 280 | CombinedNodes.erase(N); |
| 281 | PruningList.remove(N); |
| 282 | StoreRootCountMap.erase(N); |
| 283 | |
| 284 | auto It = WorklistMap.find(N); |
| 285 | if (It == WorklistMap.end()) |
| 286 | return; // Not in the worklist. |
| 287 | |
| 288 | // Null out the entry rather than erasing it to avoid a linear operation. |
| 289 | Worklist[It->second] = nullptr; |
| 290 | WorklistMap.erase(It); |
| 291 | } |
| 292 | |
| 293 | void deleteAndRecombine(SDNode *N); |
| 294 | bool recursivelyDeleteUnusedNodes(SDNode *N); |
| 295 | |
| 296 | /// Replaces all uses of the results of one DAG node with new values. |
| 297 | SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, |
| 298 | bool AddTo = true); |
| 299 | |
| 300 | /// Replaces all uses of the results of one DAG node with new values. |
| 301 | SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) { |
| 302 | return CombineTo(N, &Res, 1, AddTo); |
| 303 | } |
| 304 | |
| 305 | /// Replaces all uses of the results of one DAG node with new values. |
| 306 | SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, |
| 307 | bool AddTo = true) { |
| 308 | SDValue To[] = { Res0, Res1 }; |
| 309 | return CombineTo(N, To, 2, AddTo); |
| 310 | } |
| 311 | |
| 312 | void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO); |
| 313 | |
| 314 | private: |
| 315 | unsigned MaximumLegalStoreInBits; |
| 316 | |
| 317 | /// Check the specified integer node value to see if it can be simplified or |
| 318 | /// if things it uses can be simplified by bit propagation. |
| 319 | /// If so, return true. |
| 320 | bool SimplifyDemandedBits(SDValue Op) { |
| 321 | unsigned BitWidth = Op.getScalarValueSizeInBits(); |
| 322 | APInt DemandedBits = APInt::getAllOnesValue(BitWidth); |
| 323 | return SimplifyDemandedBits(Op, DemandedBits); |
| 324 | } |
| 325 | |
| 326 | bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits) { |
| 327 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); |
| 328 | KnownBits Known; |
| 329 | if (!TLI.SimplifyDemandedBits(Op, DemandedBits, Known, TLO, 0, false)) |
| 330 | return false; |
| 331 | |
| 332 | // Revisit the node. |
| 333 | AddToWorklist(Op.getNode()); |
| 334 | |
| 335 | CommitTargetLoweringOpt(TLO); |
| 336 | return true; |
| 337 | } |
| 338 | |
| 339 | /// Check the specified vector node value to see if it can be simplified or |
| 340 | /// if things it uses can be simplified as it only uses some of the |
| 341 | /// elements. If so, return true. |
| 342 | bool SimplifyDemandedVectorElts(SDValue Op) { |
| 343 | // TODO: For now just pretend it cannot be simplified. |
| 344 | if (Op.getValueType().isScalableVector()) |
| 345 | return false; |
| 346 | |
| 347 | unsigned NumElts = Op.getValueType().getVectorNumElements(); |
| 348 | APInt DemandedElts = APInt::getAllOnesValue(NumElts); |
| 349 | return SimplifyDemandedVectorElts(Op, DemandedElts); |
| 350 | } |
| 351 | |
| 352 | bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, |
| 353 | const APInt &DemandedElts, |
| 354 | bool AssumeSingleUse = false); |
| 355 | bool SimplifyDemandedVectorElts(SDValue Op, const APInt &DemandedElts, |
| 356 | bool AssumeSingleUse = false); |
| 357 | |
| 358 | bool CombineToPreIndexedLoadStore(SDNode *N); |
| 359 | bool CombineToPostIndexedLoadStore(SDNode *N); |
| 360 | SDValue SplitIndexingFromLoad(LoadSDNode *LD); |
| 361 | bool SliceUpLoad(SDNode *N); |
| 362 | |
| 363 | // Scalars have size 0 to distinguish from singleton vectors. |
| 364 | SDValue ForwardStoreValueToDirectLoad(LoadSDNode *LD); |
| 365 | bool getTruncatedStoreValue(StoreSDNode *ST, SDValue &Val); |
| 366 | bool extendLoadedValueToExtension(LoadSDNode *LD, SDValue &Val); |
| 367 | |
| 368 | /// Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed |
| 369 | /// load. |
| 370 | /// |
| 371 | /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced. |
| 372 | /// \param InVecVT type of the input vector to EVE with bitcasts resolved. |
| 373 | /// \param EltNo index of the vector element to load. |
| 374 | /// \param OriginalLoad load that EVE came from to be replaced. |
| 375 | /// \returns EVE on success SDValue() on failure. |
| 376 | SDValue scalarizeExtractedVectorLoad(SDNode *EVE, EVT InVecVT, |
| 377 | SDValue EltNo, |
| 378 | LoadSDNode *OriginalLoad); |
| 379 | void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad); |
| 380 | SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace); |
| 381 | SDValue SExtPromoteOperand(SDValue Op, EVT PVT); |
| 382 | SDValue ZExtPromoteOperand(SDValue Op, EVT PVT); |
| 383 | SDValue PromoteIntBinOp(SDValue Op); |
| 384 | SDValue PromoteIntShiftOp(SDValue Op); |
| 385 | SDValue PromoteExtend(SDValue Op); |
| 386 | bool PromoteLoad(SDValue Op); |
| 387 | |
| 388 | /// Call the node-specific routine that knows how to fold each |
| 389 | /// particular type of node. If that doesn't do anything, try the |
| 390 | /// target-specific DAG combines. |
| 391 | SDValue combine(SDNode *N); |
| 392 | |
| 393 | // Visitation implementation - Implement dag node combining for different |
| 394 | // node types. The semantics are as follows: |
| 395 | // Return Value: |
| 396 | // SDValue.getNode() == 0 - No change was made |
| 397 | // SDValue.getNode() == N - N was replaced, is dead and has been handled. |
| 398 | // otherwise - N should be replaced by the returned Operand. |
| 399 | // |
| 400 | SDValue visitTokenFactor(SDNode *N); |
| 401 | SDValue visitMERGE_VALUES(SDNode *N); |
| 402 | SDValue visitADD(SDNode *N); |
| 403 | SDValue visitADDLike(SDNode *N); |
| 404 | SDValue visitADDLikeCommutative(SDValue N0, SDValue N1, SDNode *LocReference); |
| 405 | SDValue visitSUB(SDNode *N); |
| 406 | SDValue visitADDSAT(SDNode *N); |
| 407 | SDValue visitSUBSAT(SDNode *N); |
| 408 | SDValue visitADDC(SDNode *N); |
| 409 | SDValue visitADDO(SDNode *N); |
| 410 | SDValue visitUADDOLike(SDValue N0, SDValue N1, SDNode *N); |
| 411 | SDValue visitSUBC(SDNode *N); |
| 412 | SDValue visitSUBO(SDNode *N); |
| 413 | SDValue visitADDE(SDNode *N); |
| 414 | SDValue visitADDCARRY(SDNode *N); |
| 415 | SDValue visitSADDO_CARRY(SDNode *N); |
| 416 | SDValue visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, SDNode *N); |
| 417 | SDValue visitSUBE(SDNode *N); |
| 418 | SDValue visitSUBCARRY(SDNode *N); |
| 419 | SDValue visitSSUBO_CARRY(SDNode *N); |
| 420 | SDValue visitMUL(SDNode *N); |
| 421 | SDValue visitMULFIX(SDNode *N); |
| 422 | SDValue useDivRem(SDNode *N); |
| 423 | SDValue visitSDIV(SDNode *N); |
| 424 | SDValue visitSDIVLike(SDValue N0, SDValue N1, SDNode *N); |
| 425 | SDValue visitUDIV(SDNode *N); |
| 426 | SDValue visitUDIVLike(SDValue N0, SDValue N1, SDNode *N); |
| 427 | SDValue visitREM(SDNode *N); |
| 428 | SDValue visitMULHU(SDNode *N); |
| 429 | SDValue visitMULHS(SDNode *N); |
| 430 | SDValue visitSMUL_LOHI(SDNode *N); |
| 431 | SDValue visitUMUL_LOHI(SDNode *N); |
| 432 | SDValue visitMULO(SDNode *N); |
| 433 | SDValue visitIMINMAX(SDNode *N); |
| 434 | SDValue visitAND(SDNode *N); |
| 435 | SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *N); |
| 436 | SDValue visitOR(SDNode *N); |
| 437 | SDValue visitORLike(SDValue N0, SDValue N1, SDNode *N); |
| 438 | SDValue visitXOR(SDNode *N); |
| 439 | SDValue SimplifyVBinOp(SDNode *N); |
| 440 | SDValue visitSHL(SDNode *N); |
| 441 | SDValue visitSRA(SDNode *N); |
| 442 | SDValue visitSRL(SDNode *N); |
| 443 | SDValue visitFunnelShift(SDNode *N); |
| 444 | SDValue visitRotate(SDNode *N); |
| 445 | SDValue visitABS(SDNode *N); |
| 446 | SDValue visitBSWAP(SDNode *N); |
| 447 | SDValue visitBITREVERSE(SDNode *N); |
| 448 | SDValue visitCTLZ(SDNode *N); |
| 449 | SDValue visitCTLZ_ZERO_UNDEF(SDNode *N); |
| 450 | SDValue visitCTTZ(SDNode *N); |
| 451 | SDValue visitCTTZ_ZERO_UNDEF(SDNode *N); |
| 452 | SDValue visitCTPOP(SDNode *N); |
| 453 | SDValue visitSELECT(SDNode *N); |
| 454 | SDValue visitVSELECT(SDNode *N); |
| 455 | SDValue visitSELECT_CC(SDNode *N); |
| 456 | SDValue visitSETCC(SDNode *N); |
| 457 | SDValue visitSETCCCARRY(SDNode *N); |
| 458 | SDValue visitSIGN_EXTEND(SDNode *N); |
| 459 | SDValue visitZERO_EXTEND(SDNode *N); |
| 460 | SDValue visitANY_EXTEND(SDNode *N); |
| 461 | SDValue visitAssertExt(SDNode *N); |
| 462 | SDValue visitAssertAlign(SDNode *N); |
| 463 | SDValue visitSIGN_EXTEND_INREG(SDNode *N); |
| 464 | SDValue visitSIGN_EXTEND_VECTOR_INREG(SDNode *N); |
| 465 | SDValue visitZERO_EXTEND_VECTOR_INREG(SDNode *N); |
| 466 | SDValue visitTRUNCATE(SDNode *N); |
| 467 | SDValue visitBITCAST(SDNode *N); |
| 468 | SDValue visitFREEZE(SDNode *N); |
| 469 | SDValue visitBUILD_PAIR(SDNode *N); |
| 470 | SDValue visitFADD(SDNode *N); |
| 471 | SDValue visitSTRICT_FADD(SDNode *N); |
| 472 | SDValue visitFSUB(SDNode *N); |
| 473 | SDValue visitFMUL(SDNode *N); |
| 474 | SDValue visitFMA(SDNode *N); |
| 475 | SDValue visitFDIV(SDNode *N); |
| 476 | SDValue visitFREM(SDNode *N); |
| 477 | SDValue visitFSQRT(SDNode *N); |
| 478 | SDValue visitFCOPYSIGN(SDNode *N); |
| 479 | SDValue visitFPOW(SDNode *N); |
| 480 | SDValue visitSINT_TO_FP(SDNode *N); |
| 481 | SDValue visitUINT_TO_FP(SDNode *N); |
| 482 | SDValue visitFP_TO_SINT(SDNode *N); |
| 483 | SDValue visitFP_TO_UINT(SDNode *N); |
| 484 | SDValue visitFP_ROUND(SDNode *N); |
| 485 | SDValue visitFP_EXTEND(SDNode *N); |
| 486 | SDValue visitFNEG(SDNode *N); |
| 487 | SDValue visitFABS(SDNode *N); |
| 488 | SDValue visitFCEIL(SDNode *N); |
| 489 | SDValue visitFTRUNC(SDNode *N); |
| 490 | SDValue visitFFLOOR(SDNode *N); |
| 491 | SDValue visitFMINNUM(SDNode *N); |
| 492 | SDValue visitFMAXNUM(SDNode *N); |
| 493 | SDValue visitFMINIMUM(SDNode *N); |
| 494 | SDValue visitFMAXIMUM(SDNode *N); |
| 495 | SDValue visitBRCOND(SDNode *N); |
| 496 | SDValue visitBR_CC(SDNode *N); |
| 497 | SDValue visitLOAD(SDNode *N); |
| 498 | |
| 499 | SDValue replaceStoreChain(StoreSDNode *ST, SDValue BetterChain); |
| 500 | SDValue replaceStoreOfFPConstant(StoreSDNode *ST); |
| 501 | |
| 502 | SDValue visitSTORE(SDNode *N); |
| 503 | SDValue visitLIFETIME_END(SDNode *N); |
| 504 | SDValue visitINSERT_VECTOR_ELT(SDNode *N); |
| 505 | SDValue visitEXTRACT_VECTOR_ELT(SDNode *N); |
| 506 | SDValue visitBUILD_VECTOR(SDNode *N); |
| 507 | SDValue visitCONCAT_VECTORS(SDNode *N); |
| 508 | SDValue visitEXTRACT_SUBVECTOR(SDNode *N); |
| 509 | SDValue visitVECTOR_SHUFFLE(SDNode *N); |
| 510 | SDValue visitSCALAR_TO_VECTOR(SDNode *N); |
| 511 | SDValue visitINSERT_SUBVECTOR(SDNode *N); |
| 512 | SDValue visitMLOAD(SDNode *N); |
| 513 | SDValue visitMSTORE(SDNode *N); |
| 514 | SDValue visitMGATHER(SDNode *N); |
| 515 | SDValue visitMSCATTER(SDNode *N); |
| 516 | SDValue visitFP_TO_FP16(SDNode *N); |
| 517 | SDValue visitFP16_TO_FP(SDNode *N); |
| 518 | SDValue visitVECREDUCE(SDNode *N); |
| 519 | |
| 520 | SDValue visitFADDForFMACombine(SDNode *N); |
| 521 | SDValue visitFSUBForFMACombine(SDNode *N); |
| 522 | SDValue visitFMULForFMADistributiveCombine(SDNode *N); |
| 523 | |
| 524 | SDValue XformToShuffleWithZero(SDNode *N); |
| 525 | bool reassociationCanBreakAddressingModePattern(unsigned Opc, |
| 526 | const SDLoc &DL, SDValue N0, |
| 527 | SDValue N1); |
| 528 | SDValue reassociateOpsCommutative(unsigned Opc, const SDLoc &DL, SDValue N0, |
| 529 | SDValue N1); |
| 530 | SDValue reassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, |
| 531 | SDValue N1, SDNodeFlags Flags); |
| 532 | |
| 533 | SDValue visitShiftByConstant(SDNode *N); |
| 534 | |
| 535 | SDValue foldSelectOfConstants(SDNode *N); |
| 536 | SDValue foldVSelectOfConstants(SDNode *N); |
| 537 | SDValue foldBinOpIntoSelect(SDNode *BO); |
| 538 | bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS); |
| 539 | SDValue hoistLogicOpWithSameOpcodeHands(SDNode *N); |
| 540 | SDValue SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2); |
| 541 | SDValue SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1, |
| 542 | SDValue N2, SDValue N3, ISD::CondCode CC, |
| 543 | bool NotExtCompare = false); |
| 544 | SDValue convertSelectOfFPConstantsToLoadOffset( |
| 545 | const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, |
| 546 | ISD::CondCode CC); |
| 547 | SDValue foldSignChangeInBitcast(SDNode *N); |
| 548 | SDValue foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, SDValue N1, |
| 549 | SDValue N2, SDValue N3, ISD::CondCode CC); |
| 550 | SDValue foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, |
| 551 | const SDLoc &DL); |
| 552 | SDValue foldSubToUSubSat(EVT DstVT, SDNode *N); |
| 553 | SDValue unfoldMaskedMerge(SDNode *N); |
| 554 | SDValue unfoldExtremeBitClearingToShifts(SDNode *N); |
| 555 | SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, |
| 556 | const SDLoc &DL, bool foldBooleans); |
| 557 | SDValue rebuildSetCC(SDValue N); |
| 558 | |
| 559 | bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, |
| 560 | SDValue &CC, bool MatchStrict = false) const; |
| 561 | bool isOneUseSetCC(SDValue N) const; |
| 562 | |
| 563 | SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, |
| 564 | unsigned HiOp); |
| 565 | SDValue CombineConsecutiveLoads(SDNode *N, EVT VT); |
| 566 | SDValue CombineExtLoad(SDNode *N); |
| 567 | SDValue CombineZExtLogicopShiftLoad(SDNode *N); |
| 568 | SDValue combineRepeatedFPDivisors(SDNode *N); |
| 569 | SDValue combineInsertEltToShuffle(SDNode *N, unsigned InsIndex); |
| 570 | SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT); |
| 571 | SDValue BuildSDIV(SDNode *N); |
| 572 | SDValue BuildSDIVPow2(SDNode *N); |
| 573 | SDValue BuildUDIV(SDNode *N); |
| 574 | SDValue BuildLogBase2(SDValue V, const SDLoc &DL); |
| 575 | SDValue BuildDivEstimate(SDValue N, SDValue Op, SDNodeFlags Flags); |
| 576 | SDValue buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags); |
| 577 | SDValue buildSqrtEstimate(SDValue Op, SDNodeFlags Flags); |
| 578 | SDValue buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, bool Recip); |
| 579 | SDValue buildSqrtNROneConst(SDValue Arg, SDValue Est, unsigned Iterations, |
| 580 | SDNodeFlags Flags, bool Reciprocal); |
| 581 | SDValue buildSqrtNRTwoConst(SDValue Arg, SDValue Est, unsigned Iterations, |
| 582 | SDNodeFlags Flags, bool Reciprocal); |
| 583 | SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, |
| 584 | bool DemandHighBits = true); |
| 585 | SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1); |
| 586 | SDValue MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg, |
| 587 | SDValue InnerPos, SDValue InnerNeg, |
| 588 | unsigned PosOpcode, unsigned NegOpcode, |
| 589 | const SDLoc &DL); |
| 590 | SDValue MatchFunnelPosNeg(SDValue N0, SDValue N1, SDValue Pos, SDValue Neg, |
| 591 | SDValue InnerPos, SDValue InnerNeg, |
| 592 | unsigned PosOpcode, unsigned NegOpcode, |
| 593 | const SDLoc &DL); |
| 594 | SDValue MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL); |
| 595 | SDValue MatchLoadCombine(SDNode *N); |
| 596 | SDValue mergeTruncStores(StoreSDNode *N); |
| 597 | SDValue ReduceLoadWidth(SDNode *N); |
| 598 | SDValue ReduceLoadOpStoreWidth(SDNode *N); |
| 599 | SDValue splitMergedValStore(StoreSDNode *ST); |
| 600 | SDValue TransformFPLoadStorePair(SDNode *N); |
| 601 | SDValue convertBuildVecZextToZext(SDNode *N); |
| 602 | SDValue reduceBuildVecExtToExtBuildVec(SDNode *N); |
| 603 | SDValue reduceBuildVecTruncToBitCast(SDNode *N); |
| 604 | SDValue reduceBuildVecToShuffle(SDNode *N); |
| 605 | SDValue createBuildVecShuffle(const SDLoc &DL, SDNode *N, |
| 606 | ArrayRef<int> VectorMask, SDValue VecIn1, |
| 607 | SDValue VecIn2, unsigned LeftIdx, |
| 608 | bool DidSplitVec); |
| 609 | SDValue matchVSelectOpSizesWithSetCC(SDNode *Cast); |
| 610 | |
| 611 | /// Walk up chain skipping non-aliasing memory nodes, |
| 612 | /// looking for aliasing nodes and adding them to the Aliases vector. |
| 613 | void GatherAllAliases(SDNode *N, SDValue OriginalChain, |
| 614 | SmallVectorImpl<SDValue> &Aliases); |
| 615 | |
| 616 | /// Return true if there is any possibility that the two addresses overlap. |
| 617 | bool isAlias(SDNode *Op0, SDNode *Op1) const; |
| 618 | |
| 619 | /// Walk up chain skipping non-aliasing memory nodes, looking for a better |
| 620 | /// chain (aliasing node.) |
| 621 | SDValue FindBetterChain(SDNode *N, SDValue Chain); |
| 622 | |
| 623 | /// Try to replace a store and any possibly adjacent stores on |
| 624 | /// consecutive chains with better chains. Return true only if St is |
| 625 | /// replaced. |
| 626 | /// |
| 627 | /// Notice that other chains may still be replaced even if the function |
| 628 | /// returns false. |
| 629 | bool findBetterNeighborChains(StoreSDNode *St); |
| 630 | |
| 631 | // Helper for findBetterNeighborChains. Walk up store chain add additional |
| 632 | // chained stores that do not overlap and can be parallelized. |
| 633 | bool parallelizeChainedStores(StoreSDNode *St); |
| 634 | |
| 635 | /// Holds a pointer to an LSBaseSDNode as well as information on where it |
| 636 | /// is located in a sequence of memory operations connected by a chain. |
| 637 | struct MemOpLink { |
| 638 | // Ptr to the mem node. |
| 639 | LSBaseSDNode *MemNode; |
| 640 | |
| 641 | // Offset from the base ptr. |
| 642 | int64_t OffsetFromBase; |
| 643 | |
| 644 | MemOpLink(LSBaseSDNode *N, int64_t Offset) |
| 645 | : MemNode(N), OffsetFromBase(Offset) {} |
| 646 | }; |
| 647 | |
| 648 | // Classify the origin of a stored value. |
| 649 | enum class StoreSource { Unknown, Constant, Extract, Load }; |
| 650 | StoreSource getStoreSource(SDValue StoreVal) { |
| 651 | switch (StoreVal.getOpcode()) { |
| 652 | case ISD::Constant: |
| 653 | case ISD::ConstantFP: |
| 654 | return StoreSource::Constant; |
| 655 | case ISD::EXTRACT_VECTOR_ELT: |
| 656 | case ISD::EXTRACT_SUBVECTOR: |
| 657 | return StoreSource::Extract; |
| 658 | case ISD::LOAD: |
| 659 | return StoreSource::Load; |
| 660 | default: |
| 661 | return StoreSource::Unknown; |
| 662 | } |
| 663 | } |
| 664 | |
| 665 | /// This is a helper function for visitMUL to check the profitability |
| 666 | /// of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2). |
| 667 | /// MulNode is the original multiply, AddNode is (add x, c1), |
| 668 | /// and ConstNode is c2. |
| 669 | bool isMulAddWithConstProfitable(SDNode *MulNode, |
| 670 | SDValue &AddNode, |
| 671 | SDValue &ConstNode); |
| 672 | |
| 673 | /// This is a helper function for visitAND and visitZERO_EXTEND. Returns |
| 674 | /// true if the (and (load x) c) pattern matches an extload. ExtVT returns |
| 675 | /// the type of the loaded value to be extended. |
| 676 | bool isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, |
| 677 | EVT LoadResultTy, EVT &ExtVT); |
| 678 | |
| 679 | /// Helper function to calculate whether the given Load/Store can have its |
| 680 | /// width reduced to ExtVT. |
| 681 | bool isLegalNarrowLdSt(LSBaseSDNode *LDSTN, ISD::LoadExtType ExtType, |
| 682 | EVT &MemVT, unsigned ShAmt = 0); |
| 683 | |
| 684 | /// Used by BackwardsPropagateMask to find suitable loads. |
| 685 | bool SearchForAndLoads(SDNode *N, SmallVectorImpl<LoadSDNode*> &Loads, |
| 686 | SmallPtrSetImpl<SDNode*> &NodesWithConsts, |
| 687 | ConstantSDNode *Mask, SDNode *&NodeToMask); |
| 688 | /// Attempt to propagate a given AND node back to load leaves so that they |
| 689 | /// can be combined into narrow loads. |
| 690 | bool BackwardsPropagateMask(SDNode *N); |
| 691 | |
| 692 | /// Helper function for mergeConsecutiveStores which merges the component |
| 693 | /// store chains. |
| 694 | SDValue getMergeStoreChains(SmallVectorImpl<MemOpLink> &StoreNodes, |
| 695 | unsigned NumStores); |
| 696 | |
| 697 | /// This is a helper function for mergeConsecutiveStores. When the source |
| 698 | /// elements of the consecutive stores are all constants or all extracted |
| 699 | /// vector elements, try to merge them into one larger store introducing |
| 700 | /// bitcasts if necessary. \return True if a merged store was created. |
| 701 | bool mergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes, |
| 702 | EVT MemVT, unsigned NumStores, |
| 703 | bool IsConstantSrc, bool UseVector, |
| 704 | bool UseTrunc); |
| 705 | |
| 706 | /// This is a helper function for mergeConsecutiveStores. Stores that |
| 707 | /// potentially may be merged with St are placed in StoreNodes. RootNode is |
| 708 | /// a chain predecessor to all store candidates. |
| 709 | void getStoreMergeCandidates(StoreSDNode *St, |
| 710 | SmallVectorImpl<MemOpLink> &StoreNodes, |
| 711 | SDNode *&Root); |
| 712 | |
| 713 | /// Helper function for mergeConsecutiveStores. Checks if candidate stores |
| 714 | /// have indirect dependency through their operands. RootNode is the |
| 715 | /// predecessor to all stores calculated by getStoreMergeCandidates and is |
| 716 | /// used to prune the dependency check. \return True if safe to merge. |
| 717 | bool checkMergeStoreCandidatesForDependencies( |
| 718 | SmallVectorImpl<MemOpLink> &StoreNodes, unsigned NumStores, |
| 719 | SDNode *RootNode); |
| 720 | |
| 721 | /// This is a helper function for mergeConsecutiveStores. Given a list of |
| 722 | /// store candidates, find the first N that are consecutive in memory. |
| 723 | /// Returns 0 if there are not at least 2 consecutive stores to try merging. |
| 724 | unsigned getConsecutiveStores(SmallVectorImpl<MemOpLink> &StoreNodes, |
| 725 | int64_t ElementSizeBytes) const; |
| 726 | |
| 727 | /// This is a helper function for mergeConsecutiveStores. It is used for |
| 728 | /// store chains that are composed entirely of constant values. |
| 729 | bool tryStoreMergeOfConstants(SmallVectorImpl<MemOpLink> &StoreNodes, |
| 730 | unsigned NumConsecutiveStores, |
| 731 | EVT MemVT, SDNode *Root, bool AllowVectors); |
| 732 | |
| 733 | /// This is a helper function for mergeConsecutiveStores. It is used for |
| 734 | /// store chains that are composed entirely of extracted vector elements. |
| 735 | /// When extracting multiple vector elements, try to store them in one |
| 736 | /// vector store rather than a sequence of scalar stores. |
| 737 | bool tryStoreMergeOfExtracts(SmallVectorImpl<MemOpLink> &StoreNodes, |
| 738 | unsigned NumConsecutiveStores, EVT MemVT, |
| 739 | SDNode *Root); |
| 740 | |
| 741 | /// This is a helper function for mergeConsecutiveStores. It is used for |
| 742 | /// store chains that are composed entirely of loaded values. |
| 743 | bool tryStoreMergeOfLoads(SmallVectorImpl<MemOpLink> &StoreNodes, |
| 744 | unsigned NumConsecutiveStores, EVT MemVT, |
| 745 | SDNode *Root, bool AllowVectors, |
| 746 | bool IsNonTemporalStore, bool IsNonTemporalLoad); |
| 747 | |
| 748 | /// Merge consecutive store operations into a wide store. |
| 749 | /// This optimization uses wide integers or vectors when possible. |
| 750 | /// \return true if stores were merged. |
| 751 | bool mergeConsecutiveStores(StoreSDNode *St); |
| 752 | |
| 753 | /// Try to transform a truncation where C is a constant: |
| 754 | /// (trunc (and X, C)) -> (and (trunc X), (trunc C)) |
| 755 | /// |
| 756 | /// \p N needs to be a truncation and its first operand an AND. Other |
| 757 | /// requirements are checked by the function (e.g. that trunc is |
| 758 | /// single-use) and if missed an empty SDValue is returned. |
| 759 | SDValue distributeTruncateThroughAnd(SDNode *N); |
| 760 | |
| 761 | /// Helper function to determine whether the target supports operation |
| 762 | /// given by \p Opcode for type \p VT, that is, whether the operation |
| 763 | /// is legal or custom before legalizing operations, and whether is |
| 764 | /// legal (but not custom) after legalization. |
| 765 | bool hasOperation(unsigned Opcode, EVT VT) { |
| 766 | return TLI.isOperationLegalOrCustom(Opcode, VT, LegalOperations); |
| 767 | } |
| 768 | |
| 769 | public: |
| 770 | /// Runs the dag combiner on all nodes in the work list |
| 771 | void Run(CombineLevel AtLevel); |
| 772 | |
| 773 | SelectionDAG &getDAG() const { return DAG; } |
| 774 | |
| 775 | /// Returns a type large enough to hold any valid shift amount - before type |
| 776 | /// legalization these can be huge. |
| 777 | EVT getShiftAmountTy(EVT LHSTy) { |
| 778 | assert(LHSTy.isInteger() && "Shift amount is not an integer type!")((LHSTy.isInteger() && "Shift amount is not an integer type!" ) ? static_cast<void> (0) : __assert_fail ("LHSTy.isInteger() && \"Shift amount is not an integer type!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 778, __PRETTY_FUNCTION__)); |
| 779 | return TLI.getShiftAmountTy(LHSTy, DAG.getDataLayout(), LegalTypes); |
| 780 | } |
| 781 | |
| 782 | /// This method returns true if we are running before type legalization or |
| 783 | /// if the specified VT is legal. |
| 784 | bool isTypeLegal(const EVT &VT) { |
| 785 | if (!LegalTypes) return true; |
| 786 | return TLI.isTypeLegal(VT); |
| 787 | } |
| 788 | |
| 789 | /// Convenience wrapper around TargetLowering::getSetCCResultType |
| 790 | EVT getSetCCResultType(EVT VT) const { |
| 791 | return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); |
| 792 | } |
| 793 | |
| 794 | void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, |
| 795 | SDValue OrigLoad, SDValue ExtLoad, |
| 796 | ISD::NodeType ExtType); |
| 797 | }; |
| 798 | |
| 799 | /// This class is a DAGUpdateListener that removes any deleted |
| 800 | /// nodes from the worklist. |
| 801 | class WorklistRemover : public SelectionDAG::DAGUpdateListener { |
| 802 | DAGCombiner &DC; |
| 803 | |
| 804 | public: |
| 805 | explicit WorklistRemover(DAGCombiner &dc) |
| 806 | : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} |
| 807 | |
| 808 | void NodeDeleted(SDNode *N, SDNode *E) override { |
| 809 | DC.removeFromWorklist(N); |
| 810 | } |
| 811 | }; |
| 812 | |
| 813 | class WorklistInserter : public SelectionDAG::DAGUpdateListener { |
| 814 | DAGCombiner &DC; |
| 815 | |
| 816 | public: |
| 817 | explicit WorklistInserter(DAGCombiner &dc) |
| 818 | : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} |
| 819 | |
| 820 | // FIXME: Ideally we could add N to the worklist, but this causes exponential |
| 821 | // compile time costs in large DAGs, e.g. Halide. |
| 822 | void NodeInserted(SDNode *N) override { DC.ConsiderForPruning(N); } |
| 823 | }; |
| 824 | |
| 825 | } // end anonymous namespace |
| 826 | |
| 827 | //===----------------------------------------------------------------------===// |
| 828 | // TargetLowering::DAGCombinerInfo implementation |
| 829 | //===----------------------------------------------------------------------===// |
| 830 | |
| 831 | void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { |
| 832 | ((DAGCombiner*)DC)->AddToWorklist(N); |
| 833 | } |
| 834 | |
| 835 | SDValue TargetLowering::DAGCombinerInfo:: |
| 836 | CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) { |
| 837 | return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo); |
| 838 | } |
| 839 | |
| 840 | SDValue TargetLowering::DAGCombinerInfo:: |
| 841 | CombineTo(SDNode *N, SDValue Res, bool AddTo) { |
| 842 | return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo); |
| 843 | } |
| 844 | |
| 845 | SDValue TargetLowering::DAGCombinerInfo:: |
| 846 | CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) { |
| 847 | return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo); |
| 848 | } |
| 849 | |
| 850 | bool TargetLowering::DAGCombinerInfo:: |
| 851 | recursivelyDeleteUnusedNodes(SDNode *N) { |
| 852 | return ((DAGCombiner*)DC)->recursivelyDeleteUnusedNodes(N); |
| 853 | } |
| 854 | |
| 855 | void TargetLowering::DAGCombinerInfo:: |
| 856 | CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { |
| 857 | return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO); |
| 858 | } |
| 859 | |
| 860 | //===----------------------------------------------------------------------===// |
| 861 | // Helper Functions |
| 862 | //===----------------------------------------------------------------------===// |
| 863 | |
| 864 | void DAGCombiner::deleteAndRecombine(SDNode *N) { |
| 865 | removeFromWorklist(N); |
| 866 | |
| 867 | // If the operands of this node are only used by the node, they will now be |
| 868 | // dead. Make sure to re-visit them and recursively delete dead nodes. |
| 869 | for (const SDValue &Op : N->ops()) |
| 870 | // For an operand generating multiple values, one of the values may |
| 871 | // become dead allowing further simplification (e.g. split index |
| 872 | // arithmetic from an indexed load). |
| 873 | if (Op->hasOneUse() || Op->getNumValues() > 1) |
| 874 | AddToWorklist(Op.getNode()); |
| 875 | |
| 876 | DAG.DeleteNode(N); |
| 877 | } |
| 878 | |
| 879 | // APInts must be the same size for most operations, this helper |
| 880 | // function zero extends the shorter of the pair so that they match. |
| 881 | // We provide an Offset so that we can create bitwidths that won't overflow. |
| 882 | static void zeroExtendToMatch(APInt &LHS, APInt &RHS, unsigned Offset = 0) { |
| 883 | unsigned Bits = Offset + std::max(LHS.getBitWidth(), RHS.getBitWidth()); |
| 884 | LHS = LHS.zextOrSelf(Bits); |
| 885 | RHS = RHS.zextOrSelf(Bits); |
| 886 | } |
| 887 | |
| 888 | // Return true if this node is a setcc, or is a select_cc |
| 889 | // that selects between the target values used for true and false, making it |
| 890 | // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to |
| 891 | // the appropriate nodes based on the type of node we are checking. This |
| 892 | // simplifies life a bit for the callers. |
| 893 | bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, |
| 894 | SDValue &CC, bool MatchStrict) const { |
| 895 | if (N.getOpcode() == ISD::SETCC) { |
| 896 | LHS = N.getOperand(0); |
| 897 | RHS = N.getOperand(1); |
| 898 | CC = N.getOperand(2); |
| 899 | return true; |
| 900 | } |
| 901 | |
| 902 | if (MatchStrict && |
| 903 | (N.getOpcode() == ISD::STRICT_FSETCC || |
| 904 | N.getOpcode() == ISD::STRICT_FSETCCS)) { |
| 905 | LHS = N.getOperand(1); |
| 906 | RHS = N.getOperand(2); |
| 907 | CC = N.getOperand(3); |
| 908 | return true; |
| 909 | } |
| 910 | |
| 911 | if (N.getOpcode() != ISD::SELECT_CC || |
| 912 | !TLI.isConstTrueVal(N.getOperand(2).getNode()) || |
| 913 | !TLI.isConstFalseVal(N.getOperand(3).getNode())) |
| 914 | return false; |
| 915 | |
| 916 | if (TLI.getBooleanContents(N.getValueType()) == |
| 917 | TargetLowering::UndefinedBooleanContent) |
| 918 | return false; |
| 919 | |
| 920 | LHS = N.getOperand(0); |
| 921 | RHS = N.getOperand(1); |
| 922 | CC = N.getOperand(4); |
| 923 | return true; |
| 924 | } |
| 925 | |
| 926 | /// Return true if this is a SetCC-equivalent operation with only one use. |
| 927 | /// If this is true, it allows the users to invert the operation for free when |
| 928 | /// it is profitable to do so. |
| 929 | bool DAGCombiner::isOneUseSetCC(SDValue N) const { |
| 930 | SDValue N0, N1, N2; |
| 931 | if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse()) |
| 932 | return true; |
| 933 | return false; |
| 934 | } |
| 935 | |
| 936 | static bool isConstantSplatVectorMaskForType(SDNode *N, EVT ScalarTy) { |
| 937 | if (!ScalarTy.isSimple()) |
| 938 | return false; |
| 939 | |
| 940 | uint64_t MaskForTy = 0ULL; |
| 941 | switch (ScalarTy.getSimpleVT().SimpleTy) { |
| 942 | case MVT::i8: |
| 943 | MaskForTy = 0xFFULL; |
| 944 | break; |
| 945 | case MVT::i16: |
| 946 | MaskForTy = 0xFFFFULL; |
| 947 | break; |
| 948 | case MVT::i32: |
| 949 | MaskForTy = 0xFFFFFFFFULL; |
| 950 | break; |
| 951 | default: |
| 952 | return false; |
| 953 | break; |
| 954 | } |
| 955 | |
| 956 | APInt Val; |
| 957 | if (ISD::isConstantSplatVector(N, Val)) |
| 958 | return Val.getLimitedValue() == MaskForTy; |
| 959 | |
| 960 | return false; |
| 961 | } |
| 962 | |
| 963 | // Determines if it is a constant integer or a splat/build vector of constant |
| 964 | // integers (and undefs). |
| 965 | // Do not permit build vector implicit truncation. |
| 966 | static bool isConstantOrConstantVector(SDValue N, bool NoOpaques = false) { |
| 967 | if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N)) |
| 968 | return !(Const->isOpaque() && NoOpaques); |
| 969 | if (N.getOpcode() != ISD::BUILD_VECTOR && N.getOpcode() != ISD::SPLAT_VECTOR) |
| 970 | return false; |
| 971 | unsigned BitWidth = N.getScalarValueSizeInBits(); |
| 972 | for (const SDValue &Op : N->op_values()) { |
| 973 | if (Op.isUndef()) |
| 974 | continue; |
| 975 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Op); |
| 976 | if (!Const || Const->getAPIntValue().getBitWidth() != BitWidth || |
| 977 | (Const->isOpaque() && NoOpaques)) |
| 978 | return false; |
| 979 | } |
| 980 | return true; |
| 981 | } |
| 982 | |
| 983 | // Determines if a BUILD_VECTOR is composed of all-constants possibly mixed with |
| 984 | // undef's. |
| 985 | static bool isAnyConstantBuildVector(SDValue V, bool NoOpaques = false) { |
| 986 | if (V.getOpcode() != ISD::BUILD_VECTOR) |
| 987 | return false; |
| 988 | return isConstantOrConstantVector(V, NoOpaques) || |
| 989 | ISD::isBuildVectorOfConstantFPSDNodes(V.getNode()); |
| 990 | } |
| 991 | |
| 992 | // Determine if this an indexed load with an opaque target constant index. |
| 993 | static bool canSplitIdx(LoadSDNode *LD) { |
| 994 | return MaySplitLoadIndex && |
| 995 | (LD->getOperand(2).getOpcode() != ISD::TargetConstant || |
| 996 | !cast<ConstantSDNode>(LD->getOperand(2))->isOpaque()); |
| 997 | } |
| 998 | |
| 999 | bool DAGCombiner::reassociationCanBreakAddressingModePattern(unsigned Opc, |
| 1000 | const SDLoc &DL, |
| 1001 | SDValue N0, |
| 1002 | SDValue N1) { |
| 1003 | // Currently this only tries to ensure we don't undo the GEP splits done by |
| 1004 | // CodeGenPrepare when shouldConsiderGEPOffsetSplit is true. To ensure this, |
| 1005 | // we check if the following transformation would be problematic: |
| 1006 | // (load/store (add, (add, x, offset1), offset2)) -> |
| 1007 | // (load/store (add, x, offset1+offset2)). |
| 1008 | |
| 1009 | if (Opc != ISD::ADD || N0.getOpcode() != ISD::ADD) |
| 1010 | return false; |
| 1011 | |
| 1012 | if (N0.hasOneUse()) |
| 1013 | return false; |
| 1014 | |
| 1015 | auto *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1)); |
| 1016 | auto *C2 = dyn_cast<ConstantSDNode>(N1); |
| 1017 | if (!C1 || !C2) |
| 1018 | return false; |
| 1019 | |
| 1020 | const APInt &C1APIntVal = C1->getAPIntValue(); |
| 1021 | const APInt &C2APIntVal = C2->getAPIntValue(); |
| 1022 | if (C1APIntVal.getBitWidth() > 64 || C2APIntVal.getBitWidth() > 64) |
| 1023 | return false; |
| 1024 | |
| 1025 | const APInt CombinedValueIntVal = C1APIntVal + C2APIntVal; |
| 1026 | if (CombinedValueIntVal.getBitWidth() > 64) |
| 1027 | return false; |
| 1028 | const int64_t CombinedValue = CombinedValueIntVal.getSExtValue(); |
| 1029 | |
| 1030 | for (SDNode *Node : N0->uses()) { |
| 1031 | auto LoadStore = dyn_cast<MemSDNode>(Node); |
| 1032 | if (LoadStore) { |
| 1033 | // Is x[offset2] already not a legal addressing mode? If so then |
| 1034 | // reassociating the constants breaks nothing (we test offset2 because |
| 1035 | // that's the one we hope to fold into the load or store). |
| 1036 | TargetLoweringBase::AddrMode AM; |
| 1037 | AM.HasBaseReg = true; |
| 1038 | AM.BaseOffs = C2APIntVal.getSExtValue(); |
| 1039 | EVT VT = LoadStore->getMemoryVT(); |
| 1040 | unsigned AS = LoadStore->getAddressSpace(); |
| 1041 | Type *AccessTy = VT.getTypeForEVT(*DAG.getContext()); |
| 1042 | if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) |
| 1043 | continue; |
| 1044 | |
| 1045 | // Would x[offset1+offset2] still be a legal addressing mode? |
| 1046 | AM.BaseOffs = CombinedValue; |
| 1047 | if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) |
| 1048 | return true; |
| 1049 | } |
| 1050 | } |
| 1051 | |
| 1052 | return false; |
| 1053 | } |
| 1054 | |
| 1055 | // Helper for DAGCombiner::reassociateOps. Try to reassociate an expression |
| 1056 | // such as (Opc N0, N1), if \p N0 is the same kind of operation as \p Opc. |
| 1057 | SDValue DAGCombiner::reassociateOpsCommutative(unsigned Opc, const SDLoc &DL, |
| 1058 | SDValue N0, SDValue N1) { |
| 1059 | EVT VT = N0.getValueType(); |
| 1060 | |
| 1061 | if (N0.getOpcode() != Opc) |
| 1062 | return SDValue(); |
| 1063 | |
| 1064 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) { |
| 1065 | if (DAG.isConstantIntBuildVectorOrConstantInt(N1)) { |
| 1066 | // Reassociate: (op (op x, c1), c2) -> (op x, (op c1, c2)) |
| 1067 | if (SDValue OpNode = |
| 1068 | DAG.FoldConstantArithmetic(Opc, DL, VT, {N0.getOperand(1), N1})) |
| 1069 | return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode); |
| 1070 | return SDValue(); |
| 1071 | } |
| 1072 | if (N0.hasOneUse()) { |
| 1073 | // Reassociate: (op (op x, c1), y) -> (op (op x, y), c1) |
| 1074 | // iff (op x, c1) has one use |
| 1075 | SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1); |
| 1076 | if (!OpNode.getNode()) |
| 1077 | return SDValue(); |
| 1078 | return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1)); |
| 1079 | } |
| 1080 | } |
| 1081 | return SDValue(); |
| 1082 | } |
| 1083 | |
| 1084 | // Try to reassociate commutative binops. |
| 1085 | SDValue DAGCombiner::reassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, |
| 1086 | SDValue N1, SDNodeFlags Flags) { |
| 1087 | assert(TLI.isCommutativeBinOp(Opc) && "Operation not commutative.")((TLI.isCommutativeBinOp(Opc) && "Operation not commutative." ) ? static_cast<void> (0) : __assert_fail ("TLI.isCommutativeBinOp(Opc) && \"Operation not commutative.\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1087, __PRETTY_FUNCTION__)); |
| 1088 | |
| 1089 | // Floating-point reassociation is not allowed without loose FP math. |
| 1090 | if (N0.getValueType().isFloatingPoint() || |
| 1091 | N1.getValueType().isFloatingPoint()) |
| 1092 | if (!Flags.hasAllowReassociation() || !Flags.hasNoSignedZeros()) |
| 1093 | return SDValue(); |
| 1094 | |
| 1095 | if (SDValue Combined = reassociateOpsCommutative(Opc, DL, N0, N1)) |
| 1096 | return Combined; |
| 1097 | if (SDValue Combined = reassociateOpsCommutative(Opc, DL, N1, N0)) |
| 1098 | return Combined; |
| 1099 | return SDValue(); |
| 1100 | } |
| 1101 | |
| 1102 | SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, |
| 1103 | bool AddTo) { |
| 1104 | assert(N->getNumValues() == NumTo && "Broken CombineTo call!")((N->getNumValues() == NumTo && "Broken CombineTo call!" ) ? static_cast<void> (0) : __assert_fail ("N->getNumValues() == NumTo && \"Broken CombineTo call!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1104, __PRETTY_FUNCTION__)); |
| 1105 | ++NodesCombined; |
| 1106 | LLVM_DEBUG(dbgs() << "\nReplacing.1 "; N->dump(&DAG); dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo - 1 << " other values\n"; } } while (false) |
| 1107 | To[0].getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo - 1 << " other values\n"; } } while (false) |
| 1108 | dbgs() << " and " << NumTo - 1 << " other values\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo - 1 << " other values\n"; } } while (false); |
| 1109 | for (unsigned i = 0, e = NumTo; i != e; ++i) |
| 1110 | assert((!To[i].getNode() ||(((!To[i].getNode() || N->getValueType(i) == To[i].getValueType ()) && "Cannot combine value to value of different type!" ) ? static_cast<void> (0) : __assert_fail ("(!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && \"Cannot combine value to value of different type!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1112, __PRETTY_FUNCTION__)) |
| 1111 | N->getValueType(i) == To[i].getValueType()) &&(((!To[i].getNode() || N->getValueType(i) == To[i].getValueType ()) && "Cannot combine value to value of different type!" ) ? static_cast<void> (0) : __assert_fail ("(!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && \"Cannot combine value to value of different type!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1112, __PRETTY_FUNCTION__)) |
| 1112 | "Cannot combine value to value of different type!")(((!To[i].getNode() || N->getValueType(i) == To[i].getValueType ()) && "Cannot combine value to value of different type!" ) ? static_cast<void> (0) : __assert_fail ("(!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && \"Cannot combine value to value of different type!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1112, __PRETTY_FUNCTION__)); |
| 1113 | |
| 1114 | WorklistRemover DeadNodes(*this); |
| 1115 | DAG.ReplaceAllUsesWith(N, To); |
| 1116 | if (AddTo) { |
| 1117 | // Push the new nodes and any users onto the worklist |
| 1118 | for (unsigned i = 0, e = NumTo; i != e; ++i) { |
| 1119 | if (To[i].getNode()) { |
| 1120 | AddToWorklist(To[i].getNode()); |
| 1121 | AddUsersToWorklist(To[i].getNode()); |
| 1122 | } |
| 1123 | } |
| 1124 | } |
| 1125 | |
| 1126 | // Finally, if the node is now dead, remove it from the graph. The node |
| 1127 | // may not be dead if the replacement process recursively simplified to |
| 1128 | // something else needing this node. |
| 1129 | if (N->use_empty()) |
| 1130 | deleteAndRecombine(N); |
| 1131 | return SDValue(N, 0); |
| 1132 | } |
| 1133 | |
| 1134 | void DAGCombiner:: |
| 1135 | CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { |
| 1136 | // Replace the old value with the new one. |
| 1137 | ++NodesCombined; |
| 1138 | LLVM_DEBUG(dbgs() << "\nReplacing.2 "; TLO.Old.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) |
| 1139 | dbgs() << "\nWith: "; TLO.New.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) |
| 1140 | dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ); |
| 1141 | |
| 1142 | // Replace all uses. If any nodes become isomorphic to other nodes and |
| 1143 | // are deleted, make sure to remove them from our worklist. |
| 1144 | WorklistRemover DeadNodes(*this); |
| 1145 | DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New); |
| 1146 | |
| 1147 | // Push the new node and any (possibly new) users onto the worklist. |
| 1148 | AddToWorklistWithUsers(TLO.New.getNode()); |
| 1149 | |
| 1150 | // Finally, if the node is now dead, remove it from the graph. The node |
| 1151 | // may not be dead if the replacement process recursively simplified to |
| 1152 | // something else needing this node. |
| 1153 | if (TLO.Old.getNode()->use_empty()) |
| 1154 | deleteAndRecombine(TLO.Old.getNode()); |
| 1155 | } |
| 1156 | |
| 1157 | /// Check the specified integer node value to see if it can be simplified or if |
| 1158 | /// things it uses can be simplified by bit propagation. If so, return true. |
| 1159 | bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, |
| 1160 | const APInt &DemandedElts, |
| 1161 | bool AssumeSingleUse) { |
| 1162 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); |
| 1163 | KnownBits Known; |
| 1164 | if (!TLI.SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO, 0, |
| 1165 | AssumeSingleUse)) |
| 1166 | return false; |
| 1167 | |
| 1168 | // Revisit the node. |
| 1169 | AddToWorklist(Op.getNode()); |
| 1170 | |
| 1171 | CommitTargetLoweringOpt(TLO); |
| 1172 | return true; |
| 1173 | } |
| 1174 | |
| 1175 | /// Check the specified vector node value to see if it can be simplified or |
| 1176 | /// if things it uses can be simplified as it only uses some of the elements. |
| 1177 | /// If so, return true. |
| 1178 | bool DAGCombiner::SimplifyDemandedVectorElts(SDValue Op, |
| 1179 | const APInt &DemandedElts, |
| 1180 | bool AssumeSingleUse) { |
| 1181 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); |
| 1182 | APInt KnownUndef, KnownZero; |
| 1183 | if (!TLI.SimplifyDemandedVectorElts(Op, DemandedElts, KnownUndef, KnownZero, |
| 1184 | TLO, 0, AssumeSingleUse)) |
| 1185 | return false; |
| 1186 | |
| 1187 | // Revisit the node. |
| 1188 | AddToWorklist(Op.getNode()); |
| 1189 | |
| 1190 | CommitTargetLoweringOpt(TLO); |
| 1191 | return true; |
| 1192 | } |
| 1193 | |
| 1194 | void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) { |
| 1195 | SDLoc DL(Load); |
| 1196 | EVT VT = Load->getValueType(0); |
| 1197 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, SDValue(ExtLoad, 0)); |
| 1198 | |
| 1199 | LLVM_DEBUG(dbgs() << "\nReplacing.9 "; Load->dump(&DAG); dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) |
| 1200 | Trunc.getNode()->dump(&DAG); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false); |
| 1201 | WorklistRemover DeadNodes(*this); |
| 1202 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc); |
| 1203 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1)); |
| 1204 | deleteAndRecombine(Load); |
| 1205 | AddToWorklist(Trunc.getNode()); |
| 1206 | } |
| 1207 | |
| 1208 | SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) { |
| 1209 | Replace = false; |
| 1210 | SDLoc DL(Op); |
| 1211 | if (ISD::isUNINDEXEDLoad(Op.getNode())) { |
| 1212 | LoadSDNode *LD = cast<LoadSDNode>(Op); |
| 1213 | EVT MemVT = LD->getMemoryVT(); |
| 1214 | ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD |
| 1215 | : LD->getExtensionType(); |
| 1216 | Replace = true; |
| 1217 | return DAG.getExtLoad(ExtType, DL, PVT, |
| 1218 | LD->getChain(), LD->getBasePtr(), |
| 1219 | MemVT, LD->getMemOperand()); |
| 1220 | } |
| 1221 | |
| 1222 | unsigned Opc = Op.getOpcode(); |
| 1223 | switch (Opc) { |
| 1224 | default: break; |
| 1225 | case ISD::AssertSext: |
| 1226 | if (SDValue Op0 = SExtPromoteOperand(Op.getOperand(0), PVT)) |
| 1227 | return DAG.getNode(ISD::AssertSext, DL, PVT, Op0, Op.getOperand(1)); |
| 1228 | break; |
| 1229 | case ISD::AssertZext: |
| 1230 | if (SDValue Op0 = ZExtPromoteOperand(Op.getOperand(0), PVT)) |
| 1231 | return DAG.getNode(ISD::AssertZext, DL, PVT, Op0, Op.getOperand(1)); |
| 1232 | break; |
| 1233 | case ISD::Constant: { |
| 1234 | unsigned ExtOpc = |
| 1235 | Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
| 1236 | return DAG.getNode(ExtOpc, DL, PVT, Op); |
| 1237 | } |
| 1238 | } |
| 1239 | |
| 1240 | if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT)) |
| 1241 | return SDValue(); |
| 1242 | return DAG.getNode(ISD::ANY_EXTEND, DL, PVT, Op); |
| 1243 | } |
| 1244 | |
| 1245 | SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) { |
| 1246 | if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT)) |
| 1247 | return SDValue(); |
| 1248 | EVT OldVT = Op.getValueType(); |
| 1249 | SDLoc DL(Op); |
| 1250 | bool Replace = false; |
| 1251 | SDValue NewOp = PromoteOperand(Op, PVT, Replace); |
| 1252 | if (!NewOp.getNode()) |
| 1253 | return SDValue(); |
| 1254 | AddToWorklist(NewOp.getNode()); |
| 1255 | |
| 1256 | if (Replace) |
| 1257 | ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); |
| 1258 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, NewOp.getValueType(), NewOp, |
| 1259 | DAG.getValueType(OldVT)); |
| 1260 | } |
| 1261 | |
| 1262 | SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) { |
| 1263 | EVT OldVT = Op.getValueType(); |
| 1264 | SDLoc DL(Op); |
| 1265 | bool Replace = false; |
| 1266 | SDValue NewOp = PromoteOperand(Op, PVT, Replace); |
| 1267 | if (!NewOp.getNode()) |
| 1268 | return SDValue(); |
| 1269 | AddToWorklist(NewOp.getNode()); |
| 1270 | |
| 1271 | if (Replace) |
| 1272 | ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); |
| 1273 | return DAG.getZeroExtendInReg(NewOp, DL, OldVT); |
| 1274 | } |
| 1275 | |
| 1276 | /// Promote the specified integer binary operation if the target indicates it is |
| 1277 | /// beneficial. e.g. On x86, it's usually better to promote i16 operations to |
| 1278 | /// i32 since i16 instructions are longer. |
| 1279 | SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) { |
| 1280 | if (!LegalOperations) |
| 1281 | return SDValue(); |
| 1282 | |
| 1283 | EVT VT = Op.getValueType(); |
| 1284 | if (VT.isVector() || !VT.isInteger()) |
| 1285 | return SDValue(); |
| 1286 | |
| 1287 | // If operation type is 'undesirable', e.g. i16 on x86, consider |
| 1288 | // promoting it. |
| 1289 | unsigned Opc = Op.getOpcode(); |
| 1290 | if (TLI.isTypeDesirableForOp(Opc, VT)) |
| 1291 | return SDValue(); |
| 1292 | |
| 1293 | EVT PVT = VT; |
| 1294 | // Consult target whether it is a good idea to promote this operation and |
| 1295 | // what's the right type to promote it to. |
| 1296 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { |
| 1297 | assert(PVT != VT && "Don't know what type to promote to!")((PVT != VT && "Don't know what type to promote to!") ? static_cast<void> (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1297, __PRETTY_FUNCTION__)); |
| 1298 | |
| 1299 | LLVM_DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); |
| 1300 | |
| 1301 | bool Replace0 = false; |
| 1302 | SDValue N0 = Op.getOperand(0); |
| 1303 | SDValue NN0 = PromoteOperand(N0, PVT, Replace0); |
| 1304 | |
| 1305 | bool Replace1 = false; |
| 1306 | SDValue N1 = Op.getOperand(1); |
| 1307 | SDValue NN1 = PromoteOperand(N1, PVT, Replace1); |
| 1308 | SDLoc DL(Op); |
| 1309 | |
| 1310 | SDValue RV = |
| 1311 | DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, NN0, NN1)); |
| 1312 | |
| 1313 | // We are always replacing N0/N1's use in N and only need additional |
| 1314 | // replacements if there are additional uses. |
| 1315 | // Note: We are checking uses of the *nodes* (SDNode) rather than values |
| 1316 | // (SDValue) here because the node may reference multiple values |
| 1317 | // (for example, the chain value of a load node). |
| 1318 | Replace0 &= !N0->hasOneUse(); |
| 1319 | Replace1 &= (N0 != N1) && !N1->hasOneUse(); |
| 1320 | |
| 1321 | // Combine Op here so it is preserved past replacements. |
| 1322 | CombineTo(Op.getNode(), RV); |
| 1323 | |
| 1324 | // If operands have a use ordering, make sure we deal with |
| 1325 | // predecessor first. |
| 1326 | if (Replace0 && Replace1 && N0.getNode()->isPredecessorOf(N1.getNode())) { |
| 1327 | std::swap(N0, N1); |
| 1328 | std::swap(NN0, NN1); |
| 1329 | } |
| 1330 | |
| 1331 | if (Replace0) { |
| 1332 | AddToWorklist(NN0.getNode()); |
| 1333 | ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode()); |
| 1334 | } |
| 1335 | if (Replace1) { |
| 1336 | AddToWorklist(NN1.getNode()); |
| 1337 | ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode()); |
| 1338 | } |
| 1339 | return Op; |
| 1340 | } |
| 1341 | return SDValue(); |
| 1342 | } |
| 1343 | |
| 1344 | /// Promote the specified integer shift operation if the target indicates it is |
| 1345 | /// beneficial. e.g. On x86, it's usually better to promote i16 operations to |
| 1346 | /// i32 since i16 instructions are longer. |
| 1347 | SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) { |
| 1348 | if (!LegalOperations) |
| 1349 | return SDValue(); |
| 1350 | |
| 1351 | EVT VT = Op.getValueType(); |
| 1352 | if (VT.isVector() || !VT.isInteger()) |
| 1353 | return SDValue(); |
| 1354 | |
| 1355 | // If operation type is 'undesirable', e.g. i16 on x86, consider |
| 1356 | // promoting it. |
| 1357 | unsigned Opc = Op.getOpcode(); |
| 1358 | if (TLI.isTypeDesirableForOp(Opc, VT)) |
| 1359 | return SDValue(); |
| 1360 | |
| 1361 | EVT PVT = VT; |
| 1362 | // Consult target whether it is a good idea to promote this operation and |
| 1363 | // what's the right type to promote it to. |
| 1364 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { |
| 1365 | assert(PVT != VT && "Don't know what type to promote to!")((PVT != VT && "Don't know what type to promote to!") ? static_cast<void> (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1365, __PRETTY_FUNCTION__)); |
| 1366 | |
| 1367 | LLVM_DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); |
| 1368 | |
| 1369 | bool Replace = false; |
| 1370 | SDValue N0 = Op.getOperand(0); |
| 1371 | SDValue N1 = Op.getOperand(1); |
| 1372 | if (Opc == ISD::SRA) |
| 1373 | N0 = SExtPromoteOperand(N0, PVT); |
| 1374 | else if (Opc == ISD::SRL) |
| 1375 | N0 = ZExtPromoteOperand(N0, PVT); |
| 1376 | else |
| 1377 | N0 = PromoteOperand(N0, PVT, Replace); |
| 1378 | |
| 1379 | if (!N0.getNode()) |
| 1380 | return SDValue(); |
| 1381 | |
| 1382 | SDLoc DL(Op); |
| 1383 | SDValue RV = |
| 1384 | DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1)); |
| 1385 | |
| 1386 | if (Replace) |
| 1387 | ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode()); |
| 1388 | |
| 1389 | // Deal with Op being deleted. |
| 1390 | if (Op && Op.getOpcode() != ISD::DELETED_NODE) |
| 1391 | return RV; |
| 1392 | } |
| 1393 | return SDValue(); |
| 1394 | } |
| 1395 | |
| 1396 | SDValue DAGCombiner::PromoteExtend(SDValue Op) { |
| 1397 | if (!LegalOperations) |
| 1398 | return SDValue(); |
| 1399 | |
| 1400 | EVT VT = Op.getValueType(); |
| 1401 | if (VT.isVector() || !VT.isInteger()) |
| 1402 | return SDValue(); |
| 1403 | |
| 1404 | // If operation type is 'undesirable', e.g. i16 on x86, consider |
| 1405 | // promoting it. |
| 1406 | unsigned Opc = Op.getOpcode(); |
| 1407 | if (TLI.isTypeDesirableForOp(Opc, VT)) |
| 1408 | return SDValue(); |
| 1409 | |
| 1410 | EVT PVT = VT; |
| 1411 | // Consult target whether it is a good idea to promote this operation and |
| 1412 | // what's the right type to promote it to. |
| 1413 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { |
| 1414 | assert(PVT != VT && "Don't know what type to promote to!")((PVT != VT && "Don't know what type to promote to!") ? static_cast<void> (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1414, __PRETTY_FUNCTION__)); |
| 1415 | // fold (aext (aext x)) -> (aext x) |
| 1416 | // fold (aext (zext x)) -> (zext x) |
| 1417 | // fold (aext (sext x)) -> (sext x) |
| 1418 | LLVM_DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); |
| 1419 | return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0)); |
| 1420 | } |
| 1421 | return SDValue(); |
| 1422 | } |
| 1423 | |
| 1424 | bool DAGCombiner::PromoteLoad(SDValue Op) { |
| 1425 | if (!LegalOperations) |
| 1426 | return false; |
| 1427 | |
| 1428 | if (!ISD::isUNINDEXEDLoad(Op.getNode())) |
| 1429 | return false; |
| 1430 | |
| 1431 | EVT VT = Op.getValueType(); |
| 1432 | if (VT.isVector() || !VT.isInteger()) |
| 1433 | return false; |
| 1434 | |
| 1435 | // If operation type is 'undesirable', e.g. i16 on x86, consider |
| 1436 | // promoting it. |
| 1437 | unsigned Opc = Op.getOpcode(); |
| 1438 | if (TLI.isTypeDesirableForOp(Opc, VT)) |
| 1439 | return false; |
| 1440 | |
| 1441 | EVT PVT = VT; |
| 1442 | // Consult target whether it is a good idea to promote this operation and |
| 1443 | // what's the right type to promote it to. |
| 1444 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { |
| 1445 | assert(PVT != VT && "Don't know what type to promote to!")((PVT != VT && "Don't know what type to promote to!") ? static_cast<void> (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1445, __PRETTY_FUNCTION__)); |
| 1446 | |
| 1447 | SDLoc DL(Op); |
| 1448 | SDNode *N = Op.getNode(); |
| 1449 | LoadSDNode *LD = cast<LoadSDNode>(N); |
| 1450 | EVT MemVT = LD->getMemoryVT(); |
| 1451 | ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD |
| 1452 | : LD->getExtensionType(); |
| 1453 | SDValue NewLD = DAG.getExtLoad(ExtType, DL, PVT, |
| 1454 | LD->getChain(), LD->getBasePtr(), |
| 1455 | MemVT, LD->getMemOperand()); |
| 1456 | SDValue Result = DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD); |
| 1457 | |
| 1458 | LLVM_DEBUG(dbgs() << "\nPromoting "; N->dump(&DAG); dbgs() << "\nTo: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false) |
| 1459 | Result.getNode()->dump(&DAG); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false); |
| 1460 | WorklistRemover DeadNodes(*this); |
| 1461 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); |
| 1462 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1)); |
| 1463 | deleteAndRecombine(N); |
| 1464 | AddToWorklist(Result.getNode()); |
| 1465 | return true; |
| 1466 | } |
| 1467 | return false; |
| 1468 | } |
| 1469 | |
| 1470 | /// Recursively delete a node which has no uses and any operands for |
| 1471 | /// which it is the only use. |
| 1472 | /// |
| 1473 | /// Note that this both deletes the nodes and removes them from the worklist. |
| 1474 | /// It also adds any nodes who have had a user deleted to the worklist as they |
| 1475 | /// may now have only one use and subject to other combines. |
| 1476 | bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) { |
| 1477 | if (!N->use_empty()) |
| 1478 | return false; |
| 1479 | |
| 1480 | SmallSetVector<SDNode *, 16> Nodes; |
| 1481 | Nodes.insert(N); |
| 1482 | do { |
| 1483 | N = Nodes.pop_back_val(); |
| 1484 | if (!N) |
| 1485 | continue; |
| 1486 | |
| 1487 | if (N->use_empty()) { |
| 1488 | for (const SDValue &ChildN : N->op_values()) |
| 1489 | Nodes.insert(ChildN.getNode()); |
| 1490 | |
| 1491 | removeFromWorklist(N); |
| 1492 | DAG.DeleteNode(N); |
| 1493 | } else { |
| 1494 | AddToWorklist(N); |
| 1495 | } |
| 1496 | } while (!Nodes.empty()); |
| 1497 | return true; |
| 1498 | } |
| 1499 | |
| 1500 | //===----------------------------------------------------------------------===// |
| 1501 | // Main DAG Combiner implementation |
| 1502 | //===----------------------------------------------------------------------===// |
| 1503 | |
| 1504 | void DAGCombiner::Run(CombineLevel AtLevel) { |
| 1505 | // set the instance variables, so that the various visit routines may use it. |
| 1506 | Level = AtLevel; |
| 1507 | LegalDAG = Level >= AfterLegalizeDAG; |
| 1508 | LegalOperations = Level >= AfterLegalizeVectorOps; |
| 1509 | LegalTypes = Level >= AfterLegalizeTypes; |
| 1510 | |
| 1511 | WorklistInserter AddNodes(*this); |
| 1512 | |
| 1513 | // Add all the dag nodes to the worklist. |
| 1514 | for (SDNode &Node : DAG.allnodes()) |
| 1515 | AddToWorklist(&Node); |
| 1516 | |
| 1517 | // Create a dummy node (which is not added to allnodes), that adds a reference |
| 1518 | // to the root node, preventing it from being deleted, and tracking any |
| 1519 | // changes of the root. |
| 1520 | HandleSDNode Dummy(DAG.getRoot()); |
| 1521 | |
| 1522 | // While we have a valid worklist entry node, try to combine it. |
| 1523 | while (SDNode *N = getNextWorklistEntry()) { |
| 1524 | // If N has no uses, it is dead. Make sure to revisit all N's operands once |
| 1525 | // N is deleted from the DAG, since they too may now be dead or may have a |
| 1526 | // reduced number of uses, allowing other xforms. |
| 1527 | if (recursivelyDeleteUnusedNodes(N)) |
| 1528 | continue; |
| 1529 | |
| 1530 | WorklistRemover DeadNodes(*this); |
| 1531 | |
| 1532 | // If this combine is running after legalizing the DAG, re-legalize any |
| 1533 | // nodes pulled off the worklist. |
| 1534 | if (LegalDAG) { |
| 1535 | SmallSetVector<SDNode *, 16> UpdatedNodes; |
| 1536 | bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes); |
| 1537 | |
| 1538 | for (SDNode *LN : UpdatedNodes) |
| 1539 | AddToWorklistWithUsers(LN); |
| 1540 | |
| 1541 | if (!NIsValid) |
| 1542 | continue; |
| 1543 | } |
| 1544 | |
| 1545 | LLVM_DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nCombining: "; N->dump (&DAG); } } while (false); |
| 1546 | |
| 1547 | // Add any operands of the new node which have not yet been combined to the |
| 1548 | // worklist as well. Because the worklist uniques things already, this |
| 1549 | // won't repeatedly process the same operand. |
| 1550 | CombinedNodes.insert(N); |
| 1551 | for (const SDValue &ChildN : N->op_values()) |
| 1552 | if (!CombinedNodes.count(ChildN.getNode())) |
| 1553 | AddToWorklist(ChildN.getNode()); |
| 1554 | |
| 1555 | SDValue RV = combine(N); |
| 1556 | |
| 1557 | if (!RV.getNode()) |
| 1558 | continue; |
| 1559 | |
| 1560 | ++NodesCombined; |
| 1561 | |
| 1562 | // If we get back the same node we passed in, rather than a new node or |
| 1563 | // zero, we know that the node must have defined multiple values and |
| 1564 | // CombineTo was used. Since CombineTo takes care of the worklist |
| 1565 | // mechanics for us, we have no work to do in this case. |
| 1566 | if (RV.getNode() == N) |
| 1567 | continue; |
| 1568 | |
| 1569 | assert(N->getOpcode() != ISD::DELETED_NODE &&((N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode () != ISD::DELETED_NODE && "Node was deleted but visit returned new node!" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned new node!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1571, __PRETTY_FUNCTION__)) |
| 1570 | RV.getOpcode() != ISD::DELETED_NODE &&((N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode () != ISD::DELETED_NODE && "Node was deleted but visit returned new node!" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned new node!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1571, __PRETTY_FUNCTION__)) |
| 1571 | "Node was deleted but visit returned new node!")((N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode () != ISD::DELETED_NODE && "Node was deleted but visit returned new node!" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned new node!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1571, __PRETTY_FUNCTION__)); |
| 1572 | |
| 1573 | LLVM_DEBUG(dbgs() << " ... into: "; RV.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << " ... into: "; RV.getNode() ->dump(&DAG); } } while (false); |
| 1574 | |
| 1575 | if (N->getNumValues() == RV.getNode()->getNumValues()) |
| 1576 | DAG.ReplaceAllUsesWith(N, RV.getNode()); |
| 1577 | else { |
| 1578 | assert(N->getValueType(0) == RV.getValueType() &&((N->getValueType(0) == RV.getValueType() && N-> getNumValues() == 1 && "Type mismatch") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && \"Type mismatch\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1579, __PRETTY_FUNCTION__)) |
| 1579 | N->getNumValues() == 1 && "Type mismatch")((N->getValueType(0) == RV.getValueType() && N-> getNumValues() == 1 && "Type mismatch") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && \"Type mismatch\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1579, __PRETTY_FUNCTION__)); |
| 1580 | DAG.ReplaceAllUsesWith(N, &RV); |
| 1581 | } |
| 1582 | |
| 1583 | // Push the new node and any users onto the worklist. Omit this if the |
| 1584 | // new node is the EntryToken (e.g. if a store managed to get optimized |
| 1585 | // out), because re-visiting the EntryToken and its users will not uncover |
| 1586 | // any additional opportunities, but there may be a large number of such |
| 1587 | // users, potentially causing compile time explosion. |
| 1588 | if (RV.getOpcode() != ISD::EntryToken) { |
| 1589 | AddToWorklist(RV.getNode()); |
| 1590 | AddUsersToWorklist(RV.getNode()); |
| 1591 | } |
| 1592 | |
| 1593 | // Finally, if the node is now dead, remove it from the graph. The node |
| 1594 | // may not be dead if the replacement process recursively simplified to |
| 1595 | // something else needing this node. This will also take care of adding any |
| 1596 | // operands which have lost a user to the worklist. |
| 1597 | recursivelyDeleteUnusedNodes(N); |
| 1598 | } |
| 1599 | |
| 1600 | // If the root changed (e.g. it was a dead load, update the root). |
| 1601 | DAG.setRoot(Dummy.getValue()); |
| 1602 | DAG.RemoveDeadNodes(); |
| 1603 | } |
| 1604 | |
| 1605 | SDValue DAGCombiner::visit(SDNode *N) { |
| 1606 | switch (N->getOpcode()) { |
| 1607 | default: break; |
| 1608 | case ISD::TokenFactor: return visitTokenFactor(N); |
| 1609 | case ISD::MERGE_VALUES: return visitMERGE_VALUES(N); |
| 1610 | case ISD::ADD: return visitADD(N); |
| 1611 | case ISD::SUB: return visitSUB(N); |
| 1612 | case ISD::SADDSAT: |
| 1613 | case ISD::UADDSAT: return visitADDSAT(N); |
| 1614 | case ISD::SSUBSAT: |
| 1615 | case ISD::USUBSAT: return visitSUBSAT(N); |
| 1616 | case ISD::ADDC: return visitADDC(N); |
| 1617 | case ISD::SADDO: |
| 1618 | case ISD::UADDO: return visitADDO(N); |
| 1619 | case ISD::SUBC: return visitSUBC(N); |
| 1620 | case ISD::SSUBO: |
| 1621 | case ISD::USUBO: return visitSUBO(N); |
| 1622 | case ISD::ADDE: return visitADDE(N); |
| 1623 | case ISD::ADDCARRY: return visitADDCARRY(N); |
| 1624 | case ISD::SADDO_CARRY: return visitSADDO_CARRY(N); |
| 1625 | case ISD::SUBE: return visitSUBE(N); |
| 1626 | case ISD::SUBCARRY: return visitSUBCARRY(N); |
| 1627 | case ISD::SSUBO_CARRY: return visitSSUBO_CARRY(N); |
| 1628 | case ISD::SMULFIX: |
| 1629 | case ISD::SMULFIXSAT: |
| 1630 | case ISD::UMULFIX: |
| 1631 | case ISD::UMULFIXSAT: return visitMULFIX(N); |
| 1632 | case ISD::MUL: return visitMUL(N); |
| 1633 | case ISD::SDIV: return visitSDIV(N); |
| 1634 | case ISD::UDIV: return visitUDIV(N); |
| 1635 | case ISD::SREM: |
| 1636 | case ISD::UREM: return visitREM(N); |
| 1637 | case ISD::MULHU: return visitMULHU(N); |
| 1638 | case ISD::MULHS: return visitMULHS(N); |
| 1639 | case ISD::SMUL_LOHI: return visitSMUL_LOHI(N); |
| 1640 | case ISD::UMUL_LOHI: return visitUMUL_LOHI(N); |
| 1641 | case ISD::SMULO: |
| 1642 | case ISD::UMULO: return visitMULO(N); |
| 1643 | case ISD::SMIN: |
| 1644 | case ISD::SMAX: |
| 1645 | case ISD::UMIN: |
| 1646 | case ISD::UMAX: return visitIMINMAX(N); |
| 1647 | case ISD::AND: return visitAND(N); |
| 1648 | case ISD::OR: return visitOR(N); |
| 1649 | case ISD::XOR: return visitXOR(N); |
| 1650 | case ISD::SHL: return visitSHL(N); |
| 1651 | case ISD::SRA: return visitSRA(N); |
| 1652 | case ISD::SRL: return visitSRL(N); |
| 1653 | case ISD::ROTR: |
| 1654 | case ISD::ROTL: return visitRotate(N); |
| 1655 | case ISD::FSHL: |
| 1656 | case ISD::FSHR: return visitFunnelShift(N); |
| 1657 | case ISD::ABS: return visitABS(N); |
| 1658 | case ISD::BSWAP: return visitBSWAP(N); |
| 1659 | case ISD::BITREVERSE: return visitBITREVERSE(N); |
| 1660 | case ISD::CTLZ: return visitCTLZ(N); |
| 1661 | case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N); |
| 1662 | case ISD::CTTZ: return visitCTTZ(N); |
| 1663 | case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N); |
| 1664 | case ISD::CTPOP: return visitCTPOP(N); |
| 1665 | case ISD::SELECT: return visitSELECT(N); |
| 1666 | case ISD::VSELECT: return visitVSELECT(N); |
| 1667 | case ISD::SELECT_CC: return visitSELECT_CC(N); |
| 1668 | case ISD::SETCC: return visitSETCC(N); |
| 1669 | case ISD::SETCCCARRY: return visitSETCCCARRY(N); |
| 1670 | case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); |
| 1671 | case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); |
| 1672 | case ISD::ANY_EXTEND: return visitANY_EXTEND(N); |
| 1673 | case ISD::AssertSext: |
| 1674 | case ISD::AssertZext: return visitAssertExt(N); |
| 1675 | case ISD::AssertAlign: return visitAssertAlign(N); |
| 1676 | case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); |
| 1677 | case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N); |
| 1678 | case ISD::ZERO_EXTEND_VECTOR_INREG: return visitZERO_EXTEND_VECTOR_INREG(N); |
| 1679 | case ISD::TRUNCATE: return visitTRUNCATE(N); |
| 1680 | case ISD::BITCAST: return visitBITCAST(N); |
| 1681 | case ISD::BUILD_PAIR: return visitBUILD_PAIR(N); |
| 1682 | case ISD::FADD: return visitFADD(N); |
| 1683 | case ISD::STRICT_FADD: return visitSTRICT_FADD(N); |
| 1684 | case ISD::FSUB: return visitFSUB(N); |
| 1685 | case ISD::FMUL: return visitFMUL(N); |
| 1686 | case ISD::FMA: return visitFMA(N); |
| 1687 | case ISD::FDIV: return visitFDIV(N); |
| 1688 | case ISD::FREM: return visitFREM(N); |
| 1689 | case ISD::FSQRT: return visitFSQRT(N); |
| 1690 | case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); |
| 1691 | case ISD::FPOW: return visitFPOW(N); |
| 1692 | case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); |
| 1693 | case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); |
| 1694 | case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); |
| 1695 | case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); |
| 1696 | case ISD::FP_ROUND: return visitFP_ROUND(N); |
| 1697 | case ISD::FP_EXTEND: return visitFP_EXTEND(N); |
| 1698 | case ISD::FNEG: return visitFNEG(N); |
| 1699 | case ISD::FABS: return visitFABS(N); |
| 1700 | case ISD::FFLOOR: return visitFFLOOR(N); |
| 1701 | case ISD::FMINNUM: return visitFMINNUM(N); |
| 1702 | case ISD::FMAXNUM: return visitFMAXNUM(N); |
| 1703 | case ISD::FMINIMUM: return visitFMINIMUM(N); |
| 1704 | case ISD::FMAXIMUM: return visitFMAXIMUM(N); |
| 1705 | case ISD::FCEIL: return visitFCEIL(N); |
| 1706 | case ISD::FTRUNC: return visitFTRUNC(N); |
| 1707 | case ISD::BRCOND: return visitBRCOND(N); |
| 1708 | case ISD::BR_CC: return visitBR_CC(N); |
| 1709 | case ISD::LOAD: return visitLOAD(N); |
| 1710 | case ISD::STORE: return visitSTORE(N); |
| 1711 | case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); |
| 1712 | case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N); |
| 1713 | case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N); |
| 1714 | case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N); |
| 1715 | case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N); |
| 1716 | case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); |
| 1717 | case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N); |
| 1718 | case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N); |
| 1719 | case ISD::MGATHER: return visitMGATHER(N); |
| 1720 | case ISD::MLOAD: return visitMLOAD(N); |
| 1721 | case ISD::MSCATTER: return visitMSCATTER(N); |
| 1722 | case ISD::MSTORE: return visitMSTORE(N); |
| 1723 | case ISD::LIFETIME_END: return visitLIFETIME_END(N); |
| 1724 | case ISD::FP_TO_FP16: return visitFP_TO_FP16(N); |
| 1725 | case ISD::FP16_TO_FP: return visitFP16_TO_FP(N); |
| 1726 | case ISD::FREEZE: return visitFREEZE(N); |
| 1727 | case ISD::VECREDUCE_FADD: |
| 1728 | case ISD::VECREDUCE_FMUL: |
| 1729 | case ISD::VECREDUCE_ADD: |
| 1730 | case ISD::VECREDUCE_MUL: |
| 1731 | case ISD::VECREDUCE_AND: |
| 1732 | case ISD::VECREDUCE_OR: |
| 1733 | case ISD::VECREDUCE_XOR: |
| 1734 | case ISD::VECREDUCE_SMAX: |
| 1735 | case ISD::VECREDUCE_SMIN: |
| 1736 | case ISD::VECREDUCE_UMAX: |
| 1737 | case ISD::VECREDUCE_UMIN: |
| 1738 | case ISD::VECREDUCE_FMAX: |
| 1739 | case ISD::VECREDUCE_FMIN: return visitVECREDUCE(N); |
| 1740 | } |
| 1741 | return SDValue(); |
| 1742 | } |
| 1743 | |
| 1744 | SDValue DAGCombiner::combine(SDNode *N) { |
| 1745 | SDValue RV; |
| 1746 | if (!DisableGenericCombines) |
| 1747 | RV = visit(N); |
| 1748 | |
| 1749 | // If nothing happened, try a target-specific DAG combine. |
| 1750 | if (!RV.getNode()) { |
| 1751 | assert(N->getOpcode() != ISD::DELETED_NODE &&((N->getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned NULL!" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned NULL!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1752, __PRETTY_FUNCTION__)) |
| 1752 | "Node was deleted but visit returned NULL!")((N->getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned NULL!" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned NULL!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1752, __PRETTY_FUNCTION__)); |
| 1753 | |
| 1754 | if (N->getOpcode() >= ISD::BUILTIN_OP_END || |
| 1755 | TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) { |
| 1756 | |
| 1757 | // Expose the DAG combiner to the target combiner impls. |
| 1758 | TargetLowering::DAGCombinerInfo |
| 1759 | DagCombineInfo(DAG, Level, false, this); |
| 1760 | |
| 1761 | RV = TLI.PerformDAGCombine(N, DagCombineInfo); |
| 1762 | } |
| 1763 | } |
| 1764 | |
| 1765 | // If nothing happened still, try promoting the operation. |
| 1766 | if (!RV.getNode()) { |
| 1767 | switch (N->getOpcode()) { |
| 1768 | default: break; |
| 1769 | case ISD::ADD: |
| 1770 | case ISD::SUB: |
| 1771 | case ISD::MUL: |
| 1772 | case ISD::AND: |
| 1773 | case ISD::OR: |
| 1774 | case ISD::XOR: |
| 1775 | RV = PromoteIntBinOp(SDValue(N, 0)); |
| 1776 | break; |
| 1777 | case ISD::SHL: |
| 1778 | case ISD::SRA: |
| 1779 | case ISD::SRL: |
| 1780 | RV = PromoteIntShiftOp(SDValue(N, 0)); |
| 1781 | break; |
| 1782 | case ISD::SIGN_EXTEND: |
| 1783 | case ISD::ZERO_EXTEND: |
| 1784 | case ISD::ANY_EXTEND: |
| 1785 | RV = PromoteExtend(SDValue(N, 0)); |
| 1786 | break; |
| 1787 | case ISD::LOAD: |
| 1788 | if (PromoteLoad(SDValue(N, 0))) |
| 1789 | RV = SDValue(N, 0); |
| 1790 | break; |
| 1791 | } |
| 1792 | } |
| 1793 | |
| 1794 | // If N is a commutative binary node, try to eliminate it if the commuted |
| 1795 | // version is already present in the DAG. |
| 1796 | if (!RV.getNode() && TLI.isCommutativeBinOp(N->getOpcode()) && |
| 1797 | N->getNumValues() == 1) { |
| 1798 | SDValue N0 = N->getOperand(0); |
| 1799 | SDValue N1 = N->getOperand(1); |
| 1800 | |
| 1801 | // Constant operands are canonicalized to RHS. |
| 1802 | if (N0 != N1 && (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1))) { |
| 1803 | SDValue Ops[] = {N1, N0}; |
| 1804 | SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops, |
| 1805 | N->getFlags()); |
| 1806 | if (CSENode) |
| 1807 | return SDValue(CSENode, 0); |
| 1808 | } |
| 1809 | } |
| 1810 | |
| 1811 | return RV; |
| 1812 | } |
| 1813 | |
| 1814 | /// Given a node, return its input chain if it has one, otherwise return a null |
| 1815 | /// sd operand. |
| 1816 | static SDValue getInputChainForNode(SDNode *N) { |
| 1817 | if (unsigned NumOps = N->getNumOperands()) { |
| 1818 | if (N->getOperand(0).getValueType() == MVT::Other) |
| 1819 | return N->getOperand(0); |
| 1820 | if (N->getOperand(NumOps-1).getValueType() == MVT::Other) |
| 1821 | return N->getOperand(NumOps-1); |
| 1822 | for (unsigned i = 1; i < NumOps-1; ++i) |
| 1823 | if (N->getOperand(i).getValueType() == MVT::Other) |
| 1824 | return N->getOperand(i); |
| 1825 | } |
| 1826 | return SDValue(); |
| 1827 | } |
| 1828 | |
| 1829 | SDValue DAGCombiner::visitTokenFactor(SDNode *N) { |
| 1830 | // If N has two operands, where one has an input chain equal to the other, |
| 1831 | // the 'other' chain is redundant. |
| 1832 | if (N->getNumOperands() == 2) { |
| 1833 | if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1)) |
| 1834 | return N->getOperand(0); |
| 1835 | if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0)) |
| 1836 | return N->getOperand(1); |
| 1837 | } |
| 1838 | |
| 1839 | // Don't simplify token factors if optnone. |
| 1840 | if (OptLevel == CodeGenOpt::None) |
| 1841 | return SDValue(); |
| 1842 | |
| 1843 | // Don't simplify the token factor if the node itself has too many operands. |
| 1844 | if (N->getNumOperands() > TokenFactorInlineLimit) |
| 1845 | return SDValue(); |
| 1846 | |
| 1847 | // If the sole user is a token factor, we should make sure we have a |
| 1848 | // chance to merge them together. This prevents TF chains from inhibiting |
| 1849 | // optimizations. |
| 1850 | if (N->hasOneUse() && N->use_begin()->getOpcode() == ISD::TokenFactor) |
| 1851 | AddToWorklist(*(N->use_begin())); |
| 1852 | |
| 1853 | SmallVector<SDNode *, 8> TFs; // List of token factors to visit. |
| 1854 | SmallVector<SDValue, 8> Ops; // Ops for replacing token factor. |
| 1855 | SmallPtrSet<SDNode*, 16> SeenOps; |
| 1856 | bool Changed = false; // If we should replace this token factor. |
| 1857 | |
| 1858 | // Start out with this token factor. |
| 1859 | TFs.push_back(N); |
| 1860 | |
| 1861 | // Iterate through token factors. The TFs grows when new token factors are |
| 1862 | // encountered. |
| 1863 | for (unsigned i = 0; i < TFs.size(); ++i) { |
| 1864 | // Limit number of nodes to inline, to avoid quadratic compile times. |
| 1865 | // We have to add the outstanding Token Factors to Ops, otherwise we might |
| 1866 | // drop Ops from the resulting Token Factors. |
| 1867 | if (Ops.size() > TokenFactorInlineLimit) { |
| 1868 | for (unsigned j = i; j < TFs.size(); j++) |
| 1869 | Ops.emplace_back(TFs[j], 0); |
| 1870 | // Drop unprocessed Token Factors from TFs, so we do not add them to the |
| 1871 | // combiner worklist later. |
| 1872 | TFs.resize(i); |
| 1873 | break; |
| 1874 | } |
| 1875 | |
| 1876 | SDNode *TF = TFs[i]; |
| 1877 | // Check each of the operands. |
| 1878 | for (const SDValue &Op : TF->op_values()) { |
| 1879 | switch (Op.getOpcode()) { |
| 1880 | case ISD::EntryToken: |
| 1881 | // Entry tokens don't need to be added to the list. They are |
| 1882 | // redundant. |
| 1883 | Changed = true; |
| 1884 | break; |
| 1885 | |
| 1886 | case ISD::TokenFactor: |
| 1887 | if (Op.hasOneUse() && !is_contained(TFs, Op.getNode())) { |
| 1888 | // Queue up for processing. |
| 1889 | TFs.push_back(Op.getNode()); |
| 1890 | Changed = true; |
| 1891 | break; |
| 1892 | } |
| 1893 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
| 1894 | |
| 1895 | default: |
| 1896 | // Only add if it isn't already in the list. |
| 1897 | if (SeenOps.insert(Op.getNode()).second) |
| 1898 | Ops.push_back(Op); |
| 1899 | else |
| 1900 | Changed = true; |
| 1901 | break; |
| 1902 | } |
| 1903 | } |
| 1904 | } |
| 1905 | |
| 1906 | // Re-visit inlined Token Factors, to clean them up in case they have been |
| 1907 | // removed. Skip the first Token Factor, as this is the current node. |
| 1908 | for (unsigned i = 1, e = TFs.size(); i < e; i++) |
| 1909 | AddToWorklist(TFs[i]); |
| 1910 | |
| 1911 | // Remove Nodes that are chained to another node in the list. Do so |
| 1912 | // by walking up chains breath-first stopping when we've seen |
| 1913 | // another operand. In general we must climb to the EntryNode, but we can exit |
| 1914 | // early if we find all remaining work is associated with just one operand as |
| 1915 | // no further pruning is possible. |
| 1916 | |
| 1917 | // List of nodes to search through and original Ops from which they originate. |
| 1918 | SmallVector<std::pair<SDNode *, unsigned>, 8> Worklist; |
| 1919 | SmallVector<unsigned, 8> OpWorkCount; // Count of work for each Op. |
| 1920 | SmallPtrSet<SDNode *, 16> SeenChains; |
| 1921 | bool DidPruneOps = false; |
| 1922 | |
| 1923 | unsigned NumLeftToConsider = 0; |
| 1924 | for (const SDValue &Op : Ops) { |
| 1925 | Worklist.push_back(std::make_pair(Op.getNode(), NumLeftToConsider++)); |
| 1926 | OpWorkCount.push_back(1); |
| 1927 | } |
| 1928 | |
| 1929 | auto AddToWorklist = [&](unsigned CurIdx, SDNode *Op, unsigned OpNumber) { |
| 1930 | // If this is an Op, we can remove the op from the list. Remark any |
| 1931 | // search associated with it as from the current OpNumber. |
| 1932 | if (SeenOps.contains(Op)) { |
| 1933 | Changed = true; |
| 1934 | DidPruneOps = true; |
| 1935 | unsigned OrigOpNumber = 0; |
| 1936 | while (OrigOpNumber < Ops.size() && Ops[OrigOpNumber].getNode() != Op) |
| 1937 | OrigOpNumber++; |
| 1938 | assert((OrigOpNumber != Ops.size()) &&(((OrigOpNumber != Ops.size()) && "expected to find TokenFactor Operand" ) ? static_cast<void> (0) : __assert_fail ("(OrigOpNumber != Ops.size()) && \"expected to find TokenFactor Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1939, __PRETTY_FUNCTION__)) |
| 1939 | "expected to find TokenFactor Operand")(((OrigOpNumber != Ops.size()) && "expected to find TokenFactor Operand" ) ? static_cast<void> (0) : __assert_fail ("(OrigOpNumber != Ops.size()) && \"expected to find TokenFactor Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1939, __PRETTY_FUNCTION__)); |
| 1940 | // Re-mark worklist from OrigOpNumber to OpNumber |
| 1941 | for (unsigned i = CurIdx + 1; i < Worklist.size(); ++i) { |
| 1942 | if (Worklist[i].second == OrigOpNumber) { |
| 1943 | Worklist[i].second = OpNumber; |
| 1944 | } |
| 1945 | } |
| 1946 | OpWorkCount[OpNumber] += OpWorkCount[OrigOpNumber]; |
| 1947 | OpWorkCount[OrigOpNumber] = 0; |
| 1948 | NumLeftToConsider--; |
| 1949 | } |
| 1950 | // Add if it's a new chain |
| 1951 | if (SeenChains.insert(Op).second) { |
| 1952 | OpWorkCount[OpNumber]++; |
| 1953 | Worklist.push_back(std::make_pair(Op, OpNumber)); |
| 1954 | } |
| 1955 | }; |
| 1956 | |
| 1957 | for (unsigned i = 0; i < Worklist.size() && i < 1024; ++i) { |
| 1958 | // We need at least be consider at least 2 Ops to prune. |
| 1959 | if (NumLeftToConsider <= 1) |
| 1960 | break; |
| 1961 | auto CurNode = Worklist[i].first; |
| 1962 | auto CurOpNumber = Worklist[i].second; |
| 1963 | assert((OpWorkCount[CurOpNumber] > 0) &&(((OpWorkCount[CurOpNumber] > 0) && "Node should not appear in worklist" ) ? static_cast<void> (0) : __assert_fail ("(OpWorkCount[CurOpNumber] > 0) && \"Node should not appear in worklist\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1964, __PRETTY_FUNCTION__)) |
| 1964 | "Node should not appear in worklist")(((OpWorkCount[CurOpNumber] > 0) && "Node should not appear in worklist" ) ? static_cast<void> (0) : __assert_fail ("(OpWorkCount[CurOpNumber] > 0) && \"Node should not appear in worklist\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1964, __PRETTY_FUNCTION__)); |
| 1965 | switch (CurNode->getOpcode()) { |
| 1966 | case ISD::EntryToken: |
| 1967 | // Hitting EntryToken is the only way for the search to terminate without |
| 1968 | // hitting |
| 1969 | // another operand's search. Prevent us from marking this operand |
| 1970 | // considered. |
| 1971 | NumLeftToConsider++; |
| 1972 | break; |
| 1973 | case ISD::TokenFactor: |
| 1974 | for (const SDValue &Op : CurNode->op_values()) |
| 1975 | AddToWorklist(i, Op.getNode(), CurOpNumber); |
| 1976 | break; |
| 1977 | case ISD::LIFETIME_START: |
| 1978 | case ISD::LIFETIME_END: |
| 1979 | case ISD::CopyFromReg: |
| 1980 | case ISD::CopyToReg: |
| 1981 | AddToWorklist(i, CurNode->getOperand(0).getNode(), CurOpNumber); |
| 1982 | break; |
| 1983 | default: |
| 1984 | if (auto *MemNode = dyn_cast<MemSDNode>(CurNode)) |
| 1985 | AddToWorklist(i, MemNode->getChain().getNode(), CurOpNumber); |
| 1986 | break; |
| 1987 | } |
| 1988 | OpWorkCount[CurOpNumber]--; |
| 1989 | if (OpWorkCount[CurOpNumber] == 0) |
| 1990 | NumLeftToConsider--; |
| 1991 | } |
| 1992 | |
| 1993 | // If we've changed things around then replace token factor. |
| 1994 | if (Changed) { |
| 1995 | SDValue Result; |
| 1996 | if (Ops.empty()) { |
| 1997 | // The entry token is the only possible outcome. |
| 1998 | Result = DAG.getEntryNode(); |
| 1999 | } else { |
| 2000 | if (DidPruneOps) { |
| 2001 | SmallVector<SDValue, 8> PrunedOps; |
| 2002 | // |
| 2003 | for (const SDValue &Op : Ops) { |
| 2004 | if (SeenChains.count(Op.getNode()) == 0) |
| 2005 | PrunedOps.push_back(Op); |
| 2006 | } |
| 2007 | Result = DAG.getTokenFactor(SDLoc(N), PrunedOps); |
| 2008 | } else { |
| 2009 | Result = DAG.getTokenFactor(SDLoc(N), Ops); |
| 2010 | } |
| 2011 | } |
| 2012 | return Result; |
| 2013 | } |
| 2014 | return SDValue(); |
| 2015 | } |
| 2016 | |
| 2017 | /// MERGE_VALUES can always be eliminated. |
| 2018 | SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) { |
| 2019 | WorklistRemover DeadNodes(*this); |
| 2020 | // Replacing results may cause a different MERGE_VALUES to suddenly |
| 2021 | // be CSE'd with N, and carry its uses with it. Iterate until no |
| 2022 | // uses remain, to ensure that the node can be safely deleted. |
| 2023 | // First add the users of this node to the work list so that they |
| 2024 | // can be tried again once they have new operands. |
| 2025 | AddUsersToWorklist(N); |
| 2026 | do { |
| 2027 | // Do as a single replacement to avoid rewalking use lists. |
| 2028 | SmallVector<SDValue, 8> Ops; |
| 2029 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) |
| 2030 | Ops.push_back(N->getOperand(i)); |
| 2031 | DAG.ReplaceAllUsesWith(N, Ops.data()); |
| 2032 | } while (!N->use_empty()); |
| 2033 | deleteAndRecombine(N); |
| 2034 | return SDValue(N, 0); // Return N so it doesn't get rechecked! |
| 2035 | } |
| 2036 | |
| 2037 | /// If \p N is a ConstantSDNode with isOpaque() == false return it casted to a |
| 2038 | /// ConstantSDNode pointer else nullptr. |
| 2039 | static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) { |
| 2040 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N); |
| 2041 | return Const != nullptr && !Const->isOpaque() ? Const : nullptr; |
| 2042 | } |
| 2043 | |
| 2044 | /// Return true if 'Use' is a load or a store that uses N as its base pointer |
| 2045 | /// and that N may be folded in the load / store addressing mode. |
| 2046 | static bool canFoldInAddressingMode(SDNode *N, SDNode *Use, SelectionDAG &DAG, |
| 2047 | const TargetLowering &TLI) { |
| 2048 | EVT VT; |
| 2049 | unsigned AS; |
| 2050 | |
| 2051 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) { |
| 2052 | if (LD->isIndexed() || LD->getBasePtr().getNode() != N) |
| 2053 | return false; |
| 2054 | VT = LD->getMemoryVT(); |
| 2055 | AS = LD->getAddressSpace(); |
| 2056 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) { |
| 2057 | if (ST->isIndexed() || ST->getBasePtr().getNode() != N) |
| 2058 | return false; |
| 2059 | VT = ST->getMemoryVT(); |
| 2060 | AS = ST->getAddressSpace(); |
| 2061 | } else if (MaskedLoadSDNode *LD = dyn_cast<MaskedLoadSDNode>(Use)) { |
| 2062 | if (LD->isIndexed() || LD->getBasePtr().getNode() != N) |
| 2063 | return false; |
| 2064 | VT = LD->getMemoryVT(); |
| 2065 | AS = LD->getAddressSpace(); |
| 2066 | } else if (MaskedStoreSDNode *ST = dyn_cast<MaskedStoreSDNode>(Use)) { |
| 2067 | if (ST->isIndexed() || ST->getBasePtr().getNode() != N) |
| 2068 | return false; |
| 2069 | VT = ST->getMemoryVT(); |
| 2070 | AS = ST->getAddressSpace(); |
| 2071 | } else |
| 2072 | return false; |
| 2073 | |
| 2074 | TargetLowering::AddrMode AM; |
| 2075 | if (N->getOpcode() == ISD::ADD) { |
| 2076 | AM.HasBaseReg = true; |
| 2077 | ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); |
| 2078 | if (Offset) |
| 2079 | // [reg +/- imm] |
| 2080 | AM.BaseOffs = Offset->getSExtValue(); |
| 2081 | else |
| 2082 | // [reg +/- reg] |
| 2083 | AM.Scale = 1; |
| 2084 | } else if (N->getOpcode() == ISD::SUB) { |
| 2085 | AM.HasBaseReg = true; |
| 2086 | ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); |
| 2087 | if (Offset) |
| 2088 | // [reg +/- imm] |
| 2089 | AM.BaseOffs = -Offset->getSExtValue(); |
| 2090 | else |
| 2091 | // [reg +/- reg] |
| 2092 | AM.Scale = 1; |
| 2093 | } else |
| 2094 | return false; |
| 2095 | |
| 2096 | return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, |
| 2097 | VT.getTypeForEVT(*DAG.getContext()), AS); |
| 2098 | } |
| 2099 | |
| 2100 | SDValue DAGCombiner::foldBinOpIntoSelect(SDNode *BO) { |
| 2101 | assert(TLI.isBinOp(BO->getOpcode()) && BO->getNumValues() == 1 &&((TLI.isBinOp(BO->getOpcode()) && BO->getNumValues () == 1 && "Unexpected binary operator") ? static_cast <void> (0) : __assert_fail ("TLI.isBinOp(BO->getOpcode()) && BO->getNumValues() == 1 && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2102, __PRETTY_FUNCTION__)) |
| 2102 | "Unexpected binary operator")((TLI.isBinOp(BO->getOpcode()) && BO->getNumValues () == 1 && "Unexpected binary operator") ? static_cast <void> (0) : __assert_fail ("TLI.isBinOp(BO->getOpcode()) && BO->getNumValues() == 1 && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2102, __PRETTY_FUNCTION__)); |
| 2103 | |
| 2104 | // Don't do this unless the old select is going away. We want to eliminate the |
| 2105 | // binary operator, not replace a binop with a select. |
| 2106 | // TODO: Handle ISD::SELECT_CC. |
| 2107 | unsigned SelOpNo = 0; |
| 2108 | SDValue Sel = BO->getOperand(0); |
| 2109 | if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) { |
| 2110 | SelOpNo = 1; |
| 2111 | Sel = BO->getOperand(1); |
| 2112 | } |
| 2113 | |
| 2114 | if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) |
| 2115 | return SDValue(); |
| 2116 | |
| 2117 | SDValue CT = Sel.getOperand(1); |
| 2118 | if (!isConstantOrConstantVector(CT, true) && |
| 2119 | !DAG.isConstantFPBuildVectorOrConstantFP(CT)) |
| 2120 | return SDValue(); |
| 2121 | |
| 2122 | SDValue CF = Sel.getOperand(2); |
| 2123 | if (!isConstantOrConstantVector(CF, true) && |
| 2124 | !DAG.isConstantFPBuildVectorOrConstantFP(CF)) |
| 2125 | return SDValue(); |
| 2126 | |
| 2127 | // Bail out if any constants are opaque because we can't constant fold those. |
| 2128 | // The exception is "and" and "or" with either 0 or -1 in which case we can |
| 2129 | // propagate non constant operands into select. I.e.: |
| 2130 | // and (select Cond, 0, -1), X --> select Cond, 0, X |
| 2131 | // or X, (select Cond, -1, 0) --> select Cond, -1, X |
| 2132 | auto BinOpcode = BO->getOpcode(); |
| 2133 | bool CanFoldNonConst = |
| 2134 | (BinOpcode == ISD::AND || BinOpcode == ISD::OR) && |
| 2135 | (isNullOrNullSplat(CT) || isAllOnesOrAllOnesSplat(CT)) && |
| 2136 | (isNullOrNullSplat(CF) || isAllOnesOrAllOnesSplat(CF)); |
| 2137 | |
| 2138 | SDValue CBO = BO->getOperand(SelOpNo ^ 1); |
| 2139 | if (!CanFoldNonConst && |
| 2140 | !isConstantOrConstantVector(CBO, true) && |
| 2141 | !DAG.isConstantFPBuildVectorOrConstantFP(CBO)) |
| 2142 | return SDValue(); |
| 2143 | |
| 2144 | EVT VT = BO->getValueType(0); |
| 2145 | |
| 2146 | // We have a select-of-constants followed by a binary operator with a |
| 2147 | // constant. Eliminate the binop by pulling the constant math into the select. |
| 2148 | // Example: add (select Cond, CT, CF), CBO --> select Cond, CT + CBO, CF + CBO |
| 2149 | SDLoc DL(Sel); |
| 2150 | SDValue NewCT = SelOpNo ? DAG.getNode(BinOpcode, DL, VT, CBO, CT) |
| 2151 | : DAG.getNode(BinOpcode, DL, VT, CT, CBO); |
| 2152 | if (!CanFoldNonConst && !NewCT.isUndef() && |
| 2153 | !isConstantOrConstantVector(NewCT, true) && |
| 2154 | !DAG.isConstantFPBuildVectorOrConstantFP(NewCT)) |
| 2155 | return SDValue(); |
| 2156 | |
| 2157 | SDValue NewCF = SelOpNo ? DAG.getNode(BinOpcode, DL, VT, CBO, CF) |
| 2158 | : DAG.getNode(BinOpcode, DL, VT, CF, CBO); |
| 2159 | if (!CanFoldNonConst && !NewCF.isUndef() && |
| 2160 | !isConstantOrConstantVector(NewCF, true) && |
| 2161 | !DAG.isConstantFPBuildVectorOrConstantFP(NewCF)) |
| 2162 | return SDValue(); |
| 2163 | |
| 2164 | SDValue SelectOp = DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF); |
| 2165 | SelectOp->setFlags(BO->getFlags()); |
| 2166 | return SelectOp; |
| 2167 | } |
| 2168 | |
| 2169 | static SDValue foldAddSubBoolOfMaskedVal(SDNode *N, SelectionDAG &DAG) { |
| 2170 | assert((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) &&(((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD:: SUB) && "Expecting add or sub") ? static_cast<void > (0) : __assert_fail ("(N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && \"Expecting add or sub\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2171, __PRETTY_FUNCTION__)) |
| 2171 | "Expecting add or sub")(((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD:: SUB) && "Expecting add or sub") ? static_cast<void > (0) : __assert_fail ("(N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && \"Expecting add or sub\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2171, __PRETTY_FUNCTION__)); |
| 2172 | |
| 2173 | // Match a constant operand and a zext operand for the math instruction: |
| 2174 | // add Z, C |
| 2175 | // sub C, Z |
| 2176 | bool IsAdd = N->getOpcode() == ISD::ADD; |
| 2177 | SDValue C = IsAdd ? N->getOperand(1) : N->getOperand(0); |
| 2178 | SDValue Z = IsAdd ? N->getOperand(0) : N->getOperand(1); |
| 2179 | auto *CN = dyn_cast<ConstantSDNode>(C); |
| 2180 | if (!CN || Z.getOpcode() != ISD::ZERO_EXTEND) |
| 2181 | return SDValue(); |
| 2182 | |
| 2183 | // Match the zext operand as a setcc of a boolean. |
| 2184 | if (Z.getOperand(0).getOpcode() != ISD::SETCC || |
| 2185 | Z.getOperand(0).getValueType() != MVT::i1) |
| 2186 | return SDValue(); |
| 2187 | |
| 2188 | // Match the compare as: setcc (X & 1), 0, eq. |
| 2189 | SDValue SetCC = Z.getOperand(0); |
| 2190 | ISD::CondCode CC = cast<CondCodeSDNode>(SetCC->getOperand(2))->get(); |
| 2191 | if (CC != ISD::SETEQ || !isNullConstant(SetCC.getOperand(1)) || |
| 2192 | SetCC.getOperand(0).getOpcode() != ISD::AND || |
| 2193 | !isOneConstant(SetCC.getOperand(0).getOperand(1))) |
| 2194 | return SDValue(); |
| 2195 | |
| 2196 | // We are adding/subtracting a constant and an inverted low bit. Turn that |
| 2197 | // into a subtract/add of the low bit with incremented/decremented constant: |
| 2198 | // add (zext i1 (seteq (X & 1), 0)), C --> sub C+1, (zext (X & 1)) |
| 2199 | // sub C, (zext i1 (seteq (X & 1), 0)) --> add C-1, (zext (X & 1)) |
| 2200 | EVT VT = C.getValueType(); |
| 2201 | SDLoc DL(N); |
| 2202 | SDValue LowBit = DAG.getZExtOrTrunc(SetCC.getOperand(0), DL, VT); |
| 2203 | SDValue C1 = IsAdd ? DAG.getConstant(CN->getAPIntValue() + 1, DL, VT) : |
| 2204 | DAG.getConstant(CN->getAPIntValue() - 1, DL, VT); |
| 2205 | return DAG.getNode(IsAdd ? ISD::SUB : ISD::ADD, DL, VT, C1, LowBit); |
| 2206 | } |
| 2207 | |
| 2208 | /// Try to fold a 'not' shifted sign-bit with add/sub with constant operand into |
| 2209 | /// a shift and add with a different constant. |
| 2210 | static SDValue foldAddSubOfSignBit(SDNode *N, SelectionDAG &DAG) { |
| 2211 | assert((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) &&(((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD:: SUB) && "Expecting add or sub") ? static_cast<void > (0) : __assert_fail ("(N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && \"Expecting add or sub\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2212, __PRETTY_FUNCTION__)) |
| 2212 | "Expecting add or sub")(((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD:: SUB) && "Expecting add or sub") ? static_cast<void > (0) : __assert_fail ("(N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && \"Expecting add or sub\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2212, __PRETTY_FUNCTION__)); |
| 2213 | |
| 2214 | // We need a constant operand for the add/sub, and the other operand is a |
| 2215 | // logical shift right: add (srl), C or sub C, (srl). |
| 2216 | bool IsAdd = N->getOpcode() == ISD::ADD; |
| 2217 | SDValue ConstantOp = IsAdd ? N->getOperand(1) : N->getOperand(0); |
| 2218 | SDValue ShiftOp = IsAdd ? N->getOperand(0) : N->getOperand(1); |
| 2219 | if (!DAG.isConstantIntBuildVectorOrConstantInt(ConstantOp) || |
| 2220 | ShiftOp.getOpcode() != ISD::SRL) |
| 2221 | return SDValue(); |
| 2222 | |
| 2223 | // The shift must be of a 'not' value. |
| 2224 | SDValue Not = ShiftOp.getOperand(0); |
| 2225 | if (!Not.hasOneUse() || !isBitwiseNot(Not)) |
| 2226 | return SDValue(); |
| 2227 | |
| 2228 | // The shift must be moving the sign bit to the least-significant-bit. |
| 2229 | EVT VT = ShiftOp.getValueType(); |
| 2230 | SDValue ShAmt = ShiftOp.getOperand(1); |
| 2231 | ConstantSDNode *ShAmtC = isConstOrConstSplat(ShAmt); |
| 2232 | if (!ShAmtC || ShAmtC->getAPIntValue() != (VT.getScalarSizeInBits() - 1)) |
| 2233 | return SDValue(); |
| 2234 | |
| 2235 | // Eliminate the 'not' by adjusting the shift and add/sub constant: |
| 2236 | // add (srl (not X), 31), C --> add (sra X, 31), (C + 1) |
| 2237 | // sub C, (srl (not X), 31) --> add (srl X, 31), (C - 1) |
| 2238 | SDLoc DL(N); |
| 2239 | auto ShOpcode = IsAdd ? ISD::SRA : ISD::SRL; |
| 2240 | SDValue NewShift = DAG.getNode(ShOpcode, DL, VT, Not.getOperand(0), ShAmt); |
| 2241 | if (SDValue NewC = |
| 2242 | DAG.FoldConstantArithmetic(IsAdd ? ISD::ADD : ISD::SUB, DL, VT, |
| 2243 | {ConstantOp, DAG.getConstant(1, DL, VT)})) |
| 2244 | return DAG.getNode(ISD::ADD, DL, VT, NewShift, NewC); |
| 2245 | return SDValue(); |
| 2246 | } |
| 2247 | |
| 2248 | /// Try to fold a node that behaves like an ADD (note that N isn't necessarily |
| 2249 | /// an ISD::ADD here, it could for example be an ISD::OR if we know that there |
| 2250 | /// are no common bits set in the operands). |
| 2251 | SDValue DAGCombiner::visitADDLike(SDNode *N) { |
| 2252 | SDValue N0 = N->getOperand(0); |
| 2253 | SDValue N1 = N->getOperand(1); |
| 2254 | EVT VT = N0.getValueType(); |
| 2255 | SDLoc DL(N); |
| 2256 | |
| 2257 | // fold vector ops |
| 2258 | if (VT.isVector()) { |
| 2259 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 2260 | return FoldedVOp; |
| 2261 | |
| 2262 | // fold (add x, 0) -> x, vector edition |
| 2263 | if (ISD::isBuildVectorAllZeros(N1.getNode())) |
| 2264 | return N0; |
| 2265 | if (ISD::isBuildVectorAllZeros(N0.getNode())) |
| 2266 | return N1; |
| 2267 | } |
| 2268 | |
| 2269 | // fold (add x, undef) -> undef |
| 2270 | if (N0.isUndef()) |
| 2271 | return N0; |
| 2272 | |
| 2273 | if (N1.isUndef()) |
| 2274 | return N1; |
| 2275 | |
| 2276 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { |
| 2277 | // canonicalize constant to RHS |
| 2278 | if (!DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 2279 | return DAG.getNode(ISD::ADD, DL, VT, N1, N0); |
| 2280 | // fold (add c1, c2) -> c1+c2 |
| 2281 | return DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N0, N1}); |
| 2282 | } |
| 2283 | |
| 2284 | // fold (add x, 0) -> x |
| 2285 | if (isNullConstant(N1)) |
| 2286 | return N0; |
| 2287 | |
| 2288 | if (isConstantOrConstantVector(N1, /* NoOpaque */ true)) { |
| 2289 | // fold ((A-c1)+c2) -> (A+(c2-c1)) |
| 2290 | if (N0.getOpcode() == ISD::SUB && |
| 2291 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaque */ true)) { |
| 2292 | SDValue Sub = |
| 2293 | DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N1, N0.getOperand(1)}); |
| 2294 | assert(Sub && "Constant folding failed")((Sub && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("Sub && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2294, __PRETTY_FUNCTION__)); |
| 2295 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Sub); |
| 2296 | } |
| 2297 | |
| 2298 | // fold ((c1-A)+c2) -> (c1+c2)-A |
| 2299 | if (N0.getOpcode() == ISD::SUB && |
| 2300 | isConstantOrConstantVector(N0.getOperand(0), /* NoOpaque */ true)) { |
| 2301 | SDValue Add = |
| 2302 | DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N1, N0.getOperand(0)}); |
| 2303 | assert(Add && "Constant folding failed")((Add && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("Add && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2303, __PRETTY_FUNCTION__)); |
| 2304 | return DAG.getNode(ISD::SUB, DL, VT, Add, N0.getOperand(1)); |
| 2305 | } |
| 2306 | |
| 2307 | // add (sext i1 X), 1 -> zext (not i1 X) |
| 2308 | // We don't transform this pattern: |
| 2309 | // add (zext i1 X), -1 -> sext (not i1 X) |
| 2310 | // because most (?) targets generate better code for the zext form. |
| 2311 | if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.hasOneUse() && |
| 2312 | isOneOrOneSplat(N1)) { |
| 2313 | SDValue X = N0.getOperand(0); |
| 2314 | if ((!LegalOperations || |
| 2315 | (TLI.isOperationLegal(ISD::XOR, X.getValueType()) && |
| 2316 | TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) && |
| 2317 | X.getScalarValueSizeInBits() == 1) { |
| 2318 | SDValue Not = DAG.getNOT(DL, X, X.getValueType()); |
| 2319 | return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Not); |
| 2320 | } |
| 2321 | } |
| 2322 | |
| 2323 | // Fold (add (or x, c0), c1) -> (add x, (c0 + c1)) if (or x, c0) is |
| 2324 | // equivalent to (add x, c0). |
| 2325 | if (N0.getOpcode() == ISD::OR && |
| 2326 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaque */ true) && |
| 2327 | DAG.haveNoCommonBitsSet(N0.getOperand(0), N0.getOperand(1))) { |
| 2328 | if (SDValue Add0 = DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, |
| 2329 | {N1, N0.getOperand(1)})) |
| 2330 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Add0); |
| 2331 | } |
| 2332 | } |
| 2333 | |
| 2334 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 2335 | return NewSel; |
| 2336 | |
| 2337 | // reassociate add |
| 2338 | if (!reassociationCanBreakAddressingModePattern(ISD::ADD, DL, N0, N1)) { |
| 2339 | if (SDValue RADD = reassociateOps(ISD::ADD, DL, N0, N1, N->getFlags())) |
| 2340 | return RADD; |
| 2341 | } |
| 2342 | // fold ((0-A) + B) -> B-A |
| 2343 | if (N0.getOpcode() == ISD::SUB && isNullOrNullSplat(N0.getOperand(0))) |
| 2344 | return DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); |
| 2345 | |
| 2346 | // fold (A + (0-B)) -> A-B |
| 2347 | if (N1.getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0))) |
| 2348 | return DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(1)); |
| 2349 | |
| 2350 | // fold (A+(B-A)) -> B |
| 2351 | if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) |
| 2352 | return N1.getOperand(0); |
| 2353 | |
| 2354 | // fold ((B-A)+A) -> B |
| 2355 | if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1)) |
| 2356 | return N0.getOperand(0); |
| 2357 | |
| 2358 | // fold ((A-B)+(C-A)) -> (C-B) |
| 2359 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && |
| 2360 | N0.getOperand(0) == N1.getOperand(1)) |
| 2361 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), |
| 2362 | N0.getOperand(1)); |
| 2363 | |
| 2364 | // fold ((A-B)+(B-C)) -> (A-C) |
| 2365 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && |
| 2366 | N0.getOperand(1) == N1.getOperand(0)) |
| 2367 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), |
| 2368 | N1.getOperand(1)); |
| 2369 | |
| 2370 | // fold (A+(B-(A+C))) to (B-C) |
| 2371 | if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && |
| 2372 | N0 == N1.getOperand(1).getOperand(0)) |
| 2373 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), |
| 2374 | N1.getOperand(1).getOperand(1)); |
| 2375 | |
| 2376 | // fold (A+(B-(C+A))) to (B-C) |
| 2377 | if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && |
| 2378 | N0 == N1.getOperand(1).getOperand(1)) |
| 2379 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), |
| 2380 | N1.getOperand(1).getOperand(0)); |
| 2381 | |
| 2382 | // fold (A+((B-A)+or-C)) to (B+or-C) |
| 2383 | if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) && |
| 2384 | N1.getOperand(0).getOpcode() == ISD::SUB && |
| 2385 | N0 == N1.getOperand(0).getOperand(1)) |
| 2386 | return DAG.getNode(N1.getOpcode(), DL, VT, N1.getOperand(0).getOperand(0), |
| 2387 | N1.getOperand(1)); |
| 2388 | |
| 2389 | // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant |
| 2390 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) { |
| 2391 | SDValue N00 = N0.getOperand(0); |
| 2392 | SDValue N01 = N0.getOperand(1); |
| 2393 | SDValue N10 = N1.getOperand(0); |
| 2394 | SDValue N11 = N1.getOperand(1); |
| 2395 | |
| 2396 | if (isConstantOrConstantVector(N00) || isConstantOrConstantVector(N10)) |
| 2397 | return DAG.getNode(ISD::SUB, DL, VT, |
| 2398 | DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10), |
| 2399 | DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11)); |
| 2400 | } |
| 2401 | |
| 2402 | // fold (add (umax X, C), -C) --> (usubsat X, C) |
| 2403 | if (N0.getOpcode() == ISD::UMAX && hasOperation(ISD::USUBSAT, VT)) { |
| 2404 | auto MatchUSUBSAT = [](ConstantSDNode *Max, ConstantSDNode *Op) { |
| 2405 | return (!Max && !Op) || |
| 2406 | (Max && Op && Max->getAPIntValue() == (-Op->getAPIntValue())); |
| 2407 | }; |
| 2408 | if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchUSUBSAT, |
| 2409 | /*AllowUndefs*/ true)) |
| 2410 | return DAG.getNode(ISD::USUBSAT, DL, VT, N0.getOperand(0), |
| 2411 | N0.getOperand(1)); |
| 2412 | } |
| 2413 | |
| 2414 | if (SimplifyDemandedBits(SDValue(N, 0))) |
| 2415 | return SDValue(N, 0); |
| 2416 | |
| 2417 | if (isOneOrOneSplat(N1)) { |
| 2418 | // fold (add (xor a, -1), 1) -> (sub 0, a) |
| 2419 | if (isBitwiseNot(N0)) |
| 2420 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), |
| 2421 | N0.getOperand(0)); |
| 2422 | |
| 2423 | // fold (add (add (xor a, -1), b), 1) -> (sub b, a) |
| 2424 | if (N0.getOpcode() == ISD::ADD || |
| 2425 | N0.getOpcode() == ISD::UADDO || |
| 2426 | N0.getOpcode() == ISD::SADDO) { |
| 2427 | SDValue A, Xor; |
| 2428 | |
| 2429 | if (isBitwiseNot(N0.getOperand(0))) { |
| 2430 | A = N0.getOperand(1); |
| 2431 | Xor = N0.getOperand(0); |
| 2432 | } else if (isBitwiseNot(N0.getOperand(1))) { |
| 2433 | A = N0.getOperand(0); |
| 2434 | Xor = N0.getOperand(1); |
| 2435 | } |
| 2436 | |
| 2437 | if (Xor) |
| 2438 | return DAG.getNode(ISD::SUB, DL, VT, A, Xor.getOperand(0)); |
| 2439 | } |
| 2440 | |
| 2441 | // Look for: |
| 2442 | // add (add x, y), 1 |
| 2443 | // And if the target does not like this form then turn into: |
| 2444 | // sub y, (xor x, -1) |
| 2445 | if (!TLI.preferIncOfAddToSubOfNot(VT) && N0.hasOneUse() && |
| 2446 | N0.getOpcode() == ISD::ADD) { |
| 2447 | SDValue Not = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0), |
| 2448 | DAG.getAllOnesConstant(DL, VT)); |
| 2449 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(1), Not); |
| 2450 | } |
| 2451 | } |
| 2452 | |
| 2453 | // (x - y) + -1 -> add (xor y, -1), x |
| 2454 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && |
| 2455 | isAllOnesOrAllOnesSplat(N1)) { |
| 2456 | SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1), N1); |
| 2457 | return DAG.getNode(ISD::ADD, DL, VT, Xor, N0.getOperand(0)); |
| 2458 | } |
| 2459 | |
| 2460 | if (SDValue Combined = visitADDLikeCommutative(N0, N1, N)) |
| 2461 | return Combined; |
| 2462 | |
| 2463 | if (SDValue Combined = visitADDLikeCommutative(N1, N0, N)) |
| 2464 | return Combined; |
| 2465 | |
| 2466 | return SDValue(); |
| 2467 | } |
| 2468 | |
| 2469 | SDValue DAGCombiner::visitADD(SDNode *N) { |
| 2470 | SDValue N0 = N->getOperand(0); |
| 2471 | SDValue N1 = N->getOperand(1); |
| 2472 | EVT VT = N0.getValueType(); |
| 2473 | SDLoc DL(N); |
| 2474 | |
| 2475 | if (SDValue Combined = visitADDLike(N)) |
| 2476 | return Combined; |
| 2477 | |
| 2478 | if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG)) |
| 2479 | return V; |
| 2480 | |
| 2481 | if (SDValue V = foldAddSubOfSignBit(N, DAG)) |
| 2482 | return V; |
| 2483 | |
| 2484 | // fold (a+b) -> (a|b) iff a and b share no bits. |
| 2485 | if ((!LegalOperations || TLI.isOperationLegal(ISD::OR, VT)) && |
| 2486 | DAG.haveNoCommonBitsSet(N0, N1)) |
| 2487 | return DAG.getNode(ISD::OR, DL, VT, N0, N1); |
| 2488 | |
| 2489 | // Fold (add (vscale * C0), (vscale * C1)) to (vscale * (C0 + C1)). |
| 2490 | if (N0.getOpcode() == ISD::VSCALE && N1.getOpcode() == ISD::VSCALE) { |
| 2491 | const APInt &C0 = N0->getConstantOperandAPInt(0); |
| 2492 | const APInt &C1 = N1->getConstantOperandAPInt(0); |
| 2493 | return DAG.getVScale(DL, VT, C0 + C1); |
| 2494 | } |
| 2495 | |
| 2496 | // fold a+vscale(c1)+vscale(c2) -> a+vscale(c1+c2) |
| 2497 | if ((N0.getOpcode() == ISD::ADD) && |
| 2498 | (N0.getOperand(1).getOpcode() == ISD::VSCALE) && |
| 2499 | (N1.getOpcode() == ISD::VSCALE)) { |
| 2500 | const APInt &VS0 = N0.getOperand(1)->getConstantOperandAPInt(0); |
| 2501 | const APInt &VS1 = N1->getConstantOperandAPInt(0); |
| 2502 | SDValue VS = DAG.getVScale(DL, VT, VS0 + VS1); |
| 2503 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), VS); |
| 2504 | } |
| 2505 | |
| 2506 | return SDValue(); |
| 2507 | } |
| 2508 | |
| 2509 | SDValue DAGCombiner::visitADDSAT(SDNode *N) { |
| 2510 | unsigned Opcode = N->getOpcode(); |
| 2511 | SDValue N0 = N->getOperand(0); |
| 2512 | SDValue N1 = N->getOperand(1); |
| 2513 | EVT VT = N0.getValueType(); |
| 2514 | SDLoc DL(N); |
| 2515 | |
| 2516 | // fold vector ops |
| 2517 | if (VT.isVector()) { |
| 2518 | // TODO SimplifyVBinOp |
| 2519 | |
| 2520 | // fold (add_sat x, 0) -> x, vector edition |
| 2521 | if (ISD::isBuildVectorAllZeros(N1.getNode())) |
| 2522 | return N0; |
| 2523 | if (ISD::isBuildVectorAllZeros(N0.getNode())) |
| 2524 | return N1; |
| 2525 | } |
| 2526 | |
| 2527 | // fold (add_sat x, undef) -> -1 |
| 2528 | if (N0.isUndef() || N1.isUndef()) |
| 2529 | return DAG.getAllOnesConstant(DL, VT); |
| 2530 | |
| 2531 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { |
| 2532 | // canonicalize constant to RHS |
| 2533 | if (!DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 2534 | return DAG.getNode(Opcode, DL, VT, N1, N0); |
| 2535 | // fold (add_sat c1, c2) -> c3 |
| 2536 | return DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1}); |
| 2537 | } |
| 2538 | |
| 2539 | // fold (add_sat x, 0) -> x |
| 2540 | if (isNullConstant(N1)) |
| 2541 | return N0; |
| 2542 | |
| 2543 | // If it cannot overflow, transform into an add. |
| 2544 | if (Opcode == ISD::UADDSAT) |
| 2545 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) |
| 2546 | return DAG.getNode(ISD::ADD, DL, VT, N0, N1); |
| 2547 | |
| 2548 | return SDValue(); |
| 2549 | } |
| 2550 | |
| 2551 | static SDValue getAsCarry(const TargetLowering &TLI, SDValue V) { |
| 2552 | bool Masked = false; |
| 2553 | |
| 2554 | // First, peel away TRUNCATE/ZERO_EXTEND/AND nodes due to legalization. |
| 2555 | while (true) { |
| 2556 | if (V.getOpcode() == ISD::TRUNCATE || V.getOpcode() == ISD::ZERO_EXTEND) { |
| 2557 | V = V.getOperand(0); |
| 2558 | continue; |
| 2559 | } |
| 2560 | |
| 2561 | if (V.getOpcode() == ISD::AND && isOneConstant(V.getOperand(1))) { |
| 2562 | Masked = true; |
| 2563 | V = V.getOperand(0); |
| 2564 | continue; |
| 2565 | } |
| 2566 | |
| 2567 | break; |
| 2568 | } |
| 2569 | |
| 2570 | // If this is not a carry, return. |
| 2571 | if (V.getResNo() != 1) |
| 2572 | return SDValue(); |
| 2573 | |
| 2574 | if (V.getOpcode() != ISD::ADDCARRY && V.getOpcode() != ISD::SUBCARRY && |
| 2575 | V.getOpcode() != ISD::UADDO && V.getOpcode() != ISD::USUBO) |
| 2576 | return SDValue(); |
| 2577 | |
| 2578 | EVT VT = V.getNode()->getValueType(0); |
| 2579 | if (!TLI.isOperationLegalOrCustom(V.getOpcode(), VT)) |
| 2580 | return SDValue(); |
| 2581 | |
| 2582 | // If the result is masked, then no matter what kind of bool it is we can |
| 2583 | // return. If it isn't, then we need to make sure the bool type is either 0 or |
| 2584 | // 1 and not other values. |
| 2585 | if (Masked || |
| 2586 | TLI.getBooleanContents(V.getValueType()) == |
| 2587 | TargetLoweringBase::ZeroOrOneBooleanContent) |
| 2588 | return V; |
| 2589 | |
| 2590 | return SDValue(); |
| 2591 | } |
| 2592 | |
| 2593 | /// Given the operands of an add/sub operation, see if the 2nd operand is a |
| 2594 | /// masked 0/1 whose source operand is actually known to be 0/-1. If so, invert |
| 2595 | /// the opcode and bypass the mask operation. |
| 2596 | static SDValue foldAddSubMasked1(bool IsAdd, SDValue N0, SDValue N1, |
| 2597 | SelectionDAG &DAG, const SDLoc &DL) { |
| 2598 | if (N1.getOpcode() != ISD::AND || !isOneOrOneSplat(N1->getOperand(1))) |
| 2599 | return SDValue(); |
| 2600 | |
| 2601 | EVT VT = N0.getValueType(); |
| 2602 | if (DAG.ComputeNumSignBits(N1.getOperand(0)) != VT.getScalarSizeInBits()) |
| 2603 | return SDValue(); |
| 2604 | |
| 2605 | // add N0, (and (AssertSext X, i1), 1) --> sub N0, X |
| 2606 | // sub N0, (and (AssertSext X, i1), 1) --> add N0, X |
| 2607 | return DAG.getNode(IsAdd ? ISD::SUB : ISD::ADD, DL, VT, N0, N1.getOperand(0)); |
| 2608 | } |
| 2609 | |
| 2610 | /// Helper for doing combines based on N0 and N1 being added to each other. |
| 2611 | SDValue DAGCombiner::visitADDLikeCommutative(SDValue N0, SDValue N1, |
| 2612 | SDNode *LocReference) { |
| 2613 | EVT VT = N0.getValueType(); |
| 2614 | SDLoc DL(LocReference); |
| 2615 | |
| 2616 | // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)) |
| 2617 | if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB && |
| 2618 | isNullOrNullSplat(N1.getOperand(0).getOperand(0))) |
| 2619 | return DAG.getNode(ISD::SUB, DL, VT, N0, |
| 2620 | DAG.getNode(ISD::SHL, DL, VT, |
| 2621 | N1.getOperand(0).getOperand(1), |
| 2622 | N1.getOperand(1))); |
| 2623 | |
| 2624 | if (SDValue V = foldAddSubMasked1(true, N0, N1, DAG, DL)) |
| 2625 | return V; |
| 2626 | |
| 2627 | // Look for: |
| 2628 | // add (add x, 1), y |
| 2629 | // And if the target does not like this form then turn into: |
| 2630 | // sub y, (xor x, -1) |
| 2631 | if (!TLI.preferIncOfAddToSubOfNot(VT) && N0.hasOneUse() && |
| 2632 | N0.getOpcode() == ISD::ADD && isOneOrOneSplat(N0.getOperand(1))) { |
| 2633 | SDValue Not = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0), |
| 2634 | DAG.getAllOnesConstant(DL, VT)); |
| 2635 | return DAG.getNode(ISD::SUB, DL, VT, N1, Not); |
| 2636 | } |
| 2637 | |
| 2638 | // Hoist one-use subtraction by non-opaque constant: |
| 2639 | // (x - C) + y -> (x + y) - C |
| 2640 | // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors. |
| 2641 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && |
| 2642 | isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { |
| 2643 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), N1); |
| 2644 | return DAG.getNode(ISD::SUB, DL, VT, Add, N0.getOperand(1)); |
| 2645 | } |
| 2646 | // Hoist one-use subtraction from non-opaque constant: |
| 2647 | // (C - x) + y -> (y - x) + C |
| 2648 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && |
| 2649 | isConstantOrConstantVector(N0.getOperand(0), /*NoOpaques=*/true)) { |
| 2650 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); |
| 2651 | return DAG.getNode(ISD::ADD, DL, VT, Sub, N0.getOperand(0)); |
| 2652 | } |
| 2653 | |
| 2654 | // If the target's bool is represented as 0/1, prefer to make this 'sub 0/1' |
| 2655 | // rather than 'add 0/-1' (the zext should get folded). |
| 2656 | // add (sext i1 Y), X --> sub X, (zext i1 Y) |
| 2657 | if (N0.getOpcode() == ISD::SIGN_EXTEND && |
| 2658 | N0.getOperand(0).getScalarValueSizeInBits() == 1 && |
| 2659 | TLI.getBooleanContents(VT) == TargetLowering::ZeroOrOneBooleanContent) { |
| 2660 | SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); |
| 2661 | return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt); |
| 2662 | } |
| 2663 | |
| 2664 | // add X, (sextinreg Y i1) -> sub X, (and Y 1) |
| 2665 | if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { |
| 2666 | VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1)); |
| 2667 | if (TN->getVT() == MVT::i1) { |
| 2668 | SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), |
| 2669 | DAG.getConstant(1, DL, VT)); |
| 2670 | return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt); |
| 2671 | } |
| 2672 | } |
| 2673 | |
| 2674 | // (add X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) |
| 2675 | if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1)) && |
| 2676 | N1.getResNo() == 0) |
| 2677 | return DAG.getNode(ISD::ADDCARRY, DL, N1->getVTList(), |
| 2678 | N0, N1.getOperand(0), N1.getOperand(2)); |
| 2679 | |
| 2680 | // (add X, Carry) -> (addcarry X, 0, Carry) |
| 2681 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) |
| 2682 | if (SDValue Carry = getAsCarry(TLI, N1)) |
| 2683 | return DAG.getNode(ISD::ADDCARRY, DL, |
| 2684 | DAG.getVTList(VT, Carry.getValueType()), N0, |
| 2685 | DAG.getConstant(0, DL, VT), Carry); |
| 2686 | |
| 2687 | return SDValue(); |
| 2688 | } |
| 2689 | |
| 2690 | SDValue DAGCombiner::visitADDC(SDNode *N) { |
| 2691 | SDValue N0 = N->getOperand(0); |
| 2692 | SDValue N1 = N->getOperand(1); |
| 2693 | EVT VT = N0.getValueType(); |
| 2694 | SDLoc DL(N); |
| 2695 | |
| 2696 | // If the flag result is dead, turn this into an ADD. |
| 2697 | if (!N->hasAnyUseOfValue(1)) |
| 2698 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), |
| 2699 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); |
| 2700 | |
| 2701 | // canonicalize constant to RHS. |
| 2702 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); |
| 2703 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); |
| 2704 | if (N0C && !N1C) |
| 2705 | return DAG.getNode(ISD::ADDC, DL, N->getVTList(), N1, N0); |
| 2706 | |
| 2707 | // fold (addc x, 0) -> x + no carry out |
| 2708 | if (isNullConstant(N1)) |
| 2709 | return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, |
| 2710 | DL, MVT::Glue)); |
| 2711 | |
| 2712 | // If it cannot overflow, transform into an add. |
| 2713 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) |
| 2714 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), |
| 2715 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); |
| 2716 | |
| 2717 | return SDValue(); |
| 2718 | } |
| 2719 | |
| 2720 | /** |
| 2721 | * Flips a boolean if it is cheaper to compute. If the Force parameters is set, |
| 2722 | * then the flip also occurs if computing the inverse is the same cost. |
| 2723 | * This function returns an empty SDValue in case it cannot flip the boolean |
| 2724 | * without increasing the cost of the computation. If you want to flip a boolean |
| 2725 | * no matter what, use DAG.getLogicalNOT. |
| 2726 | */ |
| 2727 | static SDValue extractBooleanFlip(SDValue V, SelectionDAG &DAG, |
| 2728 | const TargetLowering &TLI, |
| 2729 | bool Force) { |
| 2730 | if (Force && isa<ConstantSDNode>(V)) |
| 2731 | return DAG.getLogicalNOT(SDLoc(V), V, V.getValueType()); |
| 2732 | |
| 2733 | if (V.getOpcode() != ISD::XOR) |
| 2734 | return SDValue(); |
| 2735 | |
| 2736 | ConstantSDNode *Const = isConstOrConstSplat(V.getOperand(1), false); |
| 2737 | if (!Const) |
| 2738 | return SDValue(); |
| 2739 | |
| 2740 | EVT VT = V.getValueType(); |
| 2741 | |
| 2742 | bool IsFlip = false; |
| 2743 | switch(TLI.getBooleanContents(VT)) { |
| 2744 | case TargetLowering::ZeroOrOneBooleanContent: |
| 2745 | IsFlip = Const->isOne(); |
| 2746 | break; |
| 2747 | case TargetLowering::ZeroOrNegativeOneBooleanContent: |
| 2748 | IsFlip = Const->isAllOnesValue(); |
| 2749 | break; |
| 2750 | case TargetLowering::UndefinedBooleanContent: |
| 2751 | IsFlip = (Const->getAPIntValue() & 0x01) == 1; |
| 2752 | break; |
| 2753 | } |
| 2754 | |
| 2755 | if (IsFlip) |
| 2756 | return V.getOperand(0); |
| 2757 | if (Force) |
| 2758 | return DAG.getLogicalNOT(SDLoc(V), V, V.getValueType()); |
| 2759 | return SDValue(); |
| 2760 | } |
| 2761 | |
| 2762 | SDValue DAGCombiner::visitADDO(SDNode *N) { |
| 2763 | SDValue N0 = N->getOperand(0); |
| 2764 | SDValue N1 = N->getOperand(1); |
| 2765 | EVT VT = N0.getValueType(); |
| 2766 | bool IsSigned = (ISD::SADDO == N->getOpcode()); |
| 2767 | |
| 2768 | EVT CarryVT = N->getValueType(1); |
| 2769 | SDLoc DL(N); |
| 2770 | |
| 2771 | // If the flag result is dead, turn this into an ADD. |
| 2772 | if (!N->hasAnyUseOfValue(1)) |
| 2773 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), |
| 2774 | DAG.getUNDEF(CarryVT)); |
| 2775 | |
| 2776 | // canonicalize constant to RHS. |
| 2777 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && |
| 2778 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 2779 | return DAG.getNode(N->getOpcode(), DL, N->getVTList(), N1, N0); |
| 2780 | |
| 2781 | // fold (addo x, 0) -> x + no carry out |
| 2782 | if (isNullOrNullSplat(N1)) |
| 2783 | return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); |
| 2784 | |
| 2785 | if (!IsSigned) { |
| 2786 | // If it cannot overflow, transform into an add. |
| 2787 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) |
| 2788 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), |
| 2789 | DAG.getConstant(0, DL, CarryVT)); |
| 2790 | |
| 2791 | // fold (uaddo (xor a, -1), 1) -> (usub 0, a) and flip carry. |
| 2792 | if (isBitwiseNot(N0) && isOneOrOneSplat(N1)) { |
| 2793 | SDValue Sub = DAG.getNode(ISD::USUBO, DL, N->getVTList(), |
| 2794 | DAG.getConstant(0, DL, VT), N0.getOperand(0)); |
| 2795 | return CombineTo( |
| 2796 | N, Sub, DAG.getLogicalNOT(DL, Sub.getValue(1), Sub->getValueType(1))); |
| 2797 | } |
| 2798 | |
| 2799 | if (SDValue Combined = visitUADDOLike(N0, N1, N)) |
| 2800 | return Combined; |
| 2801 | |
| 2802 | if (SDValue Combined = visitUADDOLike(N1, N0, N)) |
| 2803 | return Combined; |
| 2804 | } |
| 2805 | |
| 2806 | return SDValue(); |
| 2807 | } |
| 2808 | |
| 2809 | SDValue DAGCombiner::visitUADDOLike(SDValue N0, SDValue N1, SDNode *N) { |
| 2810 | EVT VT = N0.getValueType(); |
| 2811 | if (VT.isVector()) |
| 2812 | return SDValue(); |
| 2813 | |
| 2814 | // (uaddo X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) |
| 2815 | // If Y + 1 cannot overflow. |
| 2816 | if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1))) { |
| 2817 | SDValue Y = N1.getOperand(0); |
| 2818 | SDValue One = DAG.getConstant(1, SDLoc(N), Y.getValueType()); |
| 2819 | if (DAG.computeOverflowKind(Y, One) == SelectionDAG::OFK_Never) |
| 2820 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, Y, |
| 2821 | N1.getOperand(2)); |
| 2822 | } |
| 2823 | |
| 2824 | // (uaddo X, Carry) -> (addcarry X, 0, Carry) |
| 2825 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) |
| 2826 | if (SDValue Carry = getAsCarry(TLI, N1)) |
| 2827 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, |
| 2828 | DAG.getConstant(0, SDLoc(N), VT), Carry); |
| 2829 | |
| 2830 | return SDValue(); |
| 2831 | } |
| 2832 | |
| 2833 | SDValue DAGCombiner::visitADDE(SDNode *N) { |
| 2834 | SDValue N0 = N->getOperand(0); |
| 2835 | SDValue N1 = N->getOperand(1); |
| 2836 | SDValue CarryIn = N->getOperand(2); |
| 2837 | |
| 2838 | // canonicalize constant to RHS |
| 2839 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); |
| 2840 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); |
| 2841 | if (N0C && !N1C) |
| 2842 | return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(), |
| 2843 | N1, N0, CarryIn); |
| 2844 | |
| 2845 | // fold (adde x, y, false) -> (addc x, y) |
| 2846 | if (CarryIn.getOpcode() == ISD::CARRY_FALSE) |
| 2847 | return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1); |
| 2848 | |
| 2849 | return SDValue(); |
| 2850 | } |
| 2851 | |
| 2852 | SDValue DAGCombiner::visitADDCARRY(SDNode *N) { |
| 2853 | SDValue N0 = N->getOperand(0); |
| 2854 | SDValue N1 = N->getOperand(1); |
| 2855 | SDValue CarryIn = N->getOperand(2); |
| 2856 | SDLoc DL(N); |
| 2857 | |
| 2858 | // canonicalize constant to RHS |
| 2859 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); |
| 2860 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); |
| 2861 | if (N0C && !N1C) |
| 2862 | return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), N1, N0, CarryIn); |
| 2863 | |
| 2864 | // fold (addcarry x, y, false) -> (uaddo x, y) |
| 2865 | if (isNullConstant(CarryIn)) { |
| 2866 | if (!LegalOperations || |
| 2867 | TLI.isOperationLegalOrCustom(ISD::UADDO, N->getValueType(0))) |
| 2868 | return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N0, N1); |
| 2869 | } |
| 2870 | |
| 2871 | // fold (addcarry 0, 0, X) -> (and (ext/trunc X), 1) and no carry. |
| 2872 | if (isNullConstant(N0) && isNullConstant(N1)) { |
| 2873 | EVT VT = N0.getValueType(); |
| 2874 | EVT CarryVT = CarryIn.getValueType(); |
| 2875 | SDValue CarryExt = DAG.getBoolExtOrTrunc(CarryIn, DL, VT, CarryVT); |
| 2876 | AddToWorklist(CarryExt.getNode()); |
| 2877 | return CombineTo(N, DAG.getNode(ISD::AND, DL, VT, CarryExt, |
| 2878 | DAG.getConstant(1, DL, VT)), |
| 2879 | DAG.getConstant(0, DL, CarryVT)); |
| 2880 | } |
| 2881 | |
| 2882 | if (SDValue Combined = visitADDCARRYLike(N0, N1, CarryIn, N)) |
| 2883 | return Combined; |
| 2884 | |
| 2885 | if (SDValue Combined = visitADDCARRYLike(N1, N0, CarryIn, N)) |
| 2886 | return Combined; |
| 2887 | |
| 2888 | return SDValue(); |
| 2889 | } |
| 2890 | |
| 2891 | SDValue DAGCombiner::visitSADDO_CARRY(SDNode *N) { |
| 2892 | SDValue N0 = N->getOperand(0); |
| 2893 | SDValue N1 = N->getOperand(1); |
| 2894 | SDValue CarryIn = N->getOperand(2); |
| 2895 | SDLoc DL(N); |
| 2896 | |
| 2897 | // canonicalize constant to RHS |
| 2898 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); |
| 2899 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); |
| 2900 | if (N0C && !N1C) |
| 2901 | return DAG.getNode(ISD::SADDO_CARRY, DL, N->getVTList(), N1, N0, CarryIn); |
| 2902 | |
| 2903 | // fold (saddo_carry x, y, false) -> (saddo x, y) |
| 2904 | if (isNullConstant(CarryIn)) { |
| 2905 | if (!LegalOperations || |
| 2906 | TLI.isOperationLegalOrCustom(ISD::SADDO, N->getValueType(0))) |
| 2907 | return DAG.getNode(ISD::SADDO, DL, N->getVTList(), N0, N1); |
| 2908 | } |
| 2909 | |
| 2910 | return SDValue(); |
| 2911 | } |
| 2912 | |
| 2913 | /** |
| 2914 | * If we are facing some sort of diamond carry propapagtion pattern try to |
| 2915 | * break it up to generate something like: |
| 2916 | * (addcarry X, 0, (addcarry A, B, Z):Carry) |
| 2917 | * |
| 2918 | * The end result is usually an increase in operation required, but because the |
| 2919 | * carry is now linearized, other tranforms can kick in and optimize the DAG. |
| 2920 | * |
| 2921 | * Patterns typically look something like |
| 2922 | * (uaddo A, B) |
| 2923 | * / \ |
| 2924 | * Carry Sum |
| 2925 | * | \ |
| 2926 | * | (addcarry *, 0, Z) |
| 2927 | * | / |
| 2928 | * \ Carry |
| 2929 | * | / |
| 2930 | * (addcarry X, *, *) |
| 2931 | * |
| 2932 | * But numerous variation exist. Our goal is to identify A, B, X and Z and |
| 2933 | * produce a combine with a single path for carry propagation. |
| 2934 | */ |
| 2935 | static SDValue combineADDCARRYDiamond(DAGCombiner &Combiner, SelectionDAG &DAG, |
| 2936 | SDValue X, SDValue Carry0, SDValue Carry1, |
| 2937 | SDNode *N) { |
| 2938 | if (Carry1.getResNo() != 1 || Carry0.getResNo() != 1) |
| 2939 | return SDValue(); |
| 2940 | if (Carry1.getOpcode() != ISD::UADDO) |
| 2941 | return SDValue(); |
| 2942 | |
| 2943 | SDValue Z; |
| 2944 | |
| 2945 | /** |
| 2946 | * First look for a suitable Z. It will present itself in the form of |
| 2947 | * (addcarry Y, 0, Z) or its equivalent (uaddo Y, 1) for Z=true |
| 2948 | */ |
| 2949 | if (Carry0.getOpcode() == ISD::ADDCARRY && |
| 2950 | isNullConstant(Carry0.getOperand(1))) { |
| 2951 | Z = Carry0.getOperand(2); |
| 2952 | } else if (Carry0.getOpcode() == ISD::UADDO && |
| 2953 | isOneConstant(Carry0.getOperand(1))) { |
| 2954 | EVT VT = Combiner.getSetCCResultType(Carry0.getValueType()); |
| 2955 | Z = DAG.getConstant(1, SDLoc(Carry0.getOperand(1)), VT); |
| 2956 | } else { |
| 2957 | // We couldn't find a suitable Z. |
| 2958 | return SDValue(); |
| 2959 | } |
| 2960 | |
| 2961 | |
| 2962 | auto cancelDiamond = [&](SDValue A,SDValue B) { |
| 2963 | SDLoc DL(N); |
| 2964 | SDValue NewY = DAG.getNode(ISD::ADDCARRY, DL, Carry0->getVTList(), A, B, Z); |
| 2965 | Combiner.AddToWorklist(NewY.getNode()); |
| 2966 | return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), X, |
| 2967 | DAG.getConstant(0, DL, X.getValueType()), |
| 2968 | NewY.getValue(1)); |
| 2969 | }; |
| 2970 | |
| 2971 | /** |
| 2972 | * (uaddo A, B) |
| 2973 | * | |
| 2974 | * Sum |
| 2975 | * | |
| 2976 | * (addcarry *, 0, Z) |
| 2977 | */ |
| 2978 | if (Carry0.getOperand(0) == Carry1.getValue(0)) { |
| 2979 | return cancelDiamond(Carry1.getOperand(0), Carry1.getOperand(1)); |
| 2980 | } |
| 2981 | |
| 2982 | /** |
| 2983 | * (addcarry A, 0, Z) |
| 2984 | * | |
| 2985 | * Sum |
| 2986 | * | |
| 2987 | * (uaddo *, B) |
| 2988 | */ |
| 2989 | if (Carry1.getOperand(0) == Carry0.getValue(0)) { |
| 2990 | return cancelDiamond(Carry0.getOperand(0), Carry1.getOperand(1)); |
| 2991 | } |
| 2992 | |
| 2993 | if (Carry1.getOperand(1) == Carry0.getValue(0)) { |
| 2994 | return cancelDiamond(Carry1.getOperand(0), Carry0.getOperand(0)); |
| 2995 | } |
| 2996 | |
| 2997 | return SDValue(); |
| 2998 | } |
| 2999 | |
| 3000 | // If we are facing some sort of diamond carry/borrow in/out pattern try to |
| 3001 | // match patterns like: |
| 3002 | // |
| 3003 | // (uaddo A, B) CarryIn |
| 3004 | // | \ | |
| 3005 | // | \ | |
| 3006 | // PartialSum PartialCarryOutX / |
| 3007 | // | | / |
| 3008 | // | ____|____________/ |
| 3009 | // | / | |
| 3010 | // (uaddo *, *) \________ |
| 3011 | // | \ \ |
| 3012 | // | \ | |
| 3013 | // | PartialCarryOutY | |
| 3014 | // | \ | |
| 3015 | // | \ / |
| 3016 | // AddCarrySum | ______/ |
| 3017 | // | / |
| 3018 | // CarryOut = (or *, *) |
| 3019 | // |
| 3020 | // And generate ADDCARRY (or SUBCARRY) with two result values: |
| 3021 | // |
| 3022 | // {AddCarrySum, CarryOut} = (addcarry A, B, CarryIn) |
| 3023 | // |
| 3024 | // Our goal is to identify A, B, and CarryIn and produce ADDCARRY/SUBCARRY with |
| 3025 | // a single path for carry/borrow out propagation: |
| 3026 | static SDValue combineCarryDiamond(DAGCombiner &Combiner, SelectionDAG &DAG, |
| 3027 | const TargetLowering &TLI, SDValue Carry0, |
| 3028 | SDValue Carry1, SDNode *N) { |
| 3029 | if (Carry0.getResNo() != 1 || Carry1.getResNo() != 1) |
| 3030 | return SDValue(); |
| 3031 | unsigned Opcode = Carry0.getOpcode(); |
| 3032 | if (Opcode != Carry1.getOpcode()) |
| 3033 | return SDValue(); |
| 3034 | if (Opcode != ISD::UADDO && Opcode != ISD::USUBO) |
| 3035 | return SDValue(); |
| 3036 | |
| 3037 | // Canonicalize the add/sub of A and B as Carry0 and the add/sub of the |
| 3038 | // carry/borrow in as Carry1. (The top and middle uaddo nodes respectively in |
| 3039 | // the above ASCII art.) |
| 3040 | if (Carry1.getOperand(0) != Carry0.getValue(0) && |
| 3041 | Carry1.getOperand(1) != Carry0.getValue(0)) |
| 3042 | std::swap(Carry0, Carry1); |
| 3043 | if (Carry1.getOperand(0) != Carry0.getValue(0) && |
| 3044 | Carry1.getOperand(1) != Carry0.getValue(0)) |
| 3045 | return SDValue(); |
| 3046 | |
| 3047 | // The carry in value must be on the righthand side for subtraction. |
| 3048 | unsigned CarryInOperandNum = |
| 3049 | Carry1.getOperand(0) == Carry0.getValue(0) ? 1 : 0; |
| 3050 | if (Opcode == ISD::USUBO && CarryInOperandNum != 1) |
| 3051 | return SDValue(); |
| 3052 | SDValue CarryIn = Carry1.getOperand(CarryInOperandNum); |
| 3053 | |
| 3054 | unsigned NewOp = Opcode == ISD::UADDO ? ISD::ADDCARRY : ISD::SUBCARRY; |
| 3055 | if (!TLI.isOperationLegalOrCustom(NewOp, Carry0.getValue(0).getValueType())) |
| 3056 | return SDValue(); |
| 3057 | |
| 3058 | // Verify that the carry/borrow in is plausibly a carry/borrow bit. |
| 3059 | // TODO: make getAsCarry() aware of how partial carries are merged. |
| 3060 | if (CarryIn.getOpcode() != ISD::ZERO_EXTEND) |
| 3061 | return SDValue(); |
| 3062 | CarryIn = CarryIn.getOperand(0); |
| 3063 | if (CarryIn.getValueType() != MVT::i1) |
| 3064 | return SDValue(); |
| 3065 | |
| 3066 | SDLoc DL(N); |
| 3067 | SDValue Merged = |
| 3068 | DAG.getNode(NewOp, DL, Carry1->getVTList(), Carry0.getOperand(0), |
| 3069 | Carry0.getOperand(1), CarryIn); |
| 3070 | |
| 3071 | // Please note that because we have proven that the result of the UADDO/USUBO |
| 3072 | // of A and B feeds into the UADDO/USUBO that does the carry/borrow in, we can |
| 3073 | // therefore prove that if the first UADDO/USUBO overflows, the second |
| 3074 | // UADDO/USUBO cannot. For example consider 8-bit numbers where 0xFF is the |
| 3075 | // maximum value. |
| 3076 | // |
| 3077 | // 0xFF + 0xFF == 0xFE with carry but 0xFE + 1 does not carry |
| 3078 | // 0x00 - 0xFF == 1 with a carry/borrow but 1 - 1 == 0 (no carry/borrow) |
| 3079 | // |
| 3080 | // This is important because it means that OR and XOR can be used to merge |
| 3081 | // carry flags; and that AND can return a constant zero. |
| 3082 | // |
| 3083 | // TODO: match other operations that can merge flags (ADD, etc) |
| 3084 | DAG.ReplaceAllUsesOfValueWith(Carry1.getValue(0), Merged.getValue(0)); |
| 3085 | if (N->getOpcode() == ISD::AND) |
| 3086 | return DAG.getConstant(0, DL, MVT::i1); |
| 3087 | return Merged.getValue(1); |
| 3088 | } |
| 3089 | |
| 3090 | SDValue DAGCombiner::visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, |
| 3091 | SDNode *N) { |
| 3092 | // fold (addcarry (xor a, -1), b, c) -> (subcarry b, a, !c) and flip carry. |
| 3093 | if (isBitwiseNot(N0)) |
| 3094 | if (SDValue NotC = extractBooleanFlip(CarryIn, DAG, TLI, true)) { |
| 3095 | SDLoc DL(N); |
| 3096 | SDValue Sub = DAG.getNode(ISD::SUBCARRY, DL, N->getVTList(), N1, |
| 3097 | N0.getOperand(0), NotC); |
| 3098 | return CombineTo( |
| 3099 | N, Sub, DAG.getLogicalNOT(DL, Sub.getValue(1), Sub->getValueType(1))); |
| 3100 | } |
| 3101 | |
| 3102 | // Iff the flag result is dead: |
| 3103 | // (addcarry (add|uaddo X, Y), 0, Carry) -> (addcarry X, Y, Carry) |
| 3104 | // Don't do this if the Carry comes from the uaddo. It won't remove the uaddo |
| 3105 | // or the dependency between the instructions. |
| 3106 | if ((N0.getOpcode() == ISD::ADD || |
| 3107 | (N0.getOpcode() == ISD::UADDO && N0.getResNo() == 0 && |
| 3108 | N0.getValue(1) != CarryIn)) && |
| 3109 | isNullConstant(N1) && !N->hasAnyUseOfValue(1)) |
| 3110 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), |
| 3111 | N0.getOperand(0), N0.getOperand(1), CarryIn); |
| 3112 | |
| 3113 | /** |
| 3114 | * When one of the addcarry argument is itself a carry, we may be facing |
| 3115 | * a diamond carry propagation. In which case we try to transform the DAG |
| 3116 | * to ensure linear carry propagation if that is possible. |
| 3117 | */ |
| 3118 | if (auto Y = getAsCarry(TLI, N1)) { |
| 3119 | // Because both are carries, Y and Z can be swapped. |
| 3120 | if (auto R = combineADDCARRYDiamond(*this, DAG, N0, Y, CarryIn, N)) |
| 3121 | return R; |
| 3122 | if (auto R = combineADDCARRYDiamond(*this, DAG, N0, CarryIn, Y, N)) |
| 3123 | return R; |
| 3124 | } |
| 3125 | |
| 3126 | return SDValue(); |
| 3127 | } |
| 3128 | |
| 3129 | // Attempt to create a USUBSAT(LHS, RHS) node with DstVT, performing a |
| 3130 | // clamp/truncation if necessary. |
| 3131 | static SDValue getTruncatedUSUBSAT(EVT DstVT, EVT SrcVT, SDValue LHS, |
| 3132 | SDValue RHS, SelectionDAG &DAG, |
| 3133 | const SDLoc &DL) { |
| 3134 | assert(DstVT.getScalarSizeInBits() <= SrcVT.getScalarSizeInBits() &&((DstVT.getScalarSizeInBits() <= SrcVT.getScalarSizeInBits () && "Illegal truncation") ? static_cast<void> (0) : __assert_fail ("DstVT.getScalarSizeInBits() <= SrcVT.getScalarSizeInBits() && \"Illegal truncation\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3135, __PRETTY_FUNCTION__)) |
| 3135 | "Illegal truncation")((DstVT.getScalarSizeInBits() <= SrcVT.getScalarSizeInBits () && "Illegal truncation") ? static_cast<void> (0) : __assert_fail ("DstVT.getScalarSizeInBits() <= SrcVT.getScalarSizeInBits() && \"Illegal truncation\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3135, __PRETTY_FUNCTION__)); |
| 3136 | |
| 3137 | if (DstVT == SrcVT) |
| 3138 | return DAG.getNode(ISD::USUBSAT, DL, DstVT, LHS, RHS); |
| 3139 | |
| 3140 | // If the LHS is zero-extended then we can perform the USUBSAT as DstVT by |
| 3141 | // clamping RHS. |
| 3142 | APInt UpperBits = APInt::getBitsSetFrom(SrcVT.getScalarSizeInBits(), |
| 3143 | DstVT.getScalarSizeInBits()); |
| 3144 | if (!DAG.MaskedValueIsZero(LHS, UpperBits)) |
| 3145 | return SDValue(); |
| 3146 | |
| 3147 | SDValue SatLimit = |
| 3148 | DAG.getConstant(APInt::getLowBitsSet(SrcVT.getScalarSizeInBits(), |
| 3149 | DstVT.getScalarSizeInBits()), |
| 3150 | DL, SrcVT); |
| 3151 | RHS = DAG.getNode(ISD::UMIN, DL, SrcVT, RHS, SatLimit); |
| 3152 | RHS = DAG.getNode(ISD::TRUNCATE, DL, DstVT, RHS); |
| 3153 | LHS = DAG.getNode(ISD::TRUNCATE, DL, DstVT, LHS); |
| 3154 | return DAG.getNode(ISD::USUBSAT, DL, DstVT, LHS, RHS); |
| 3155 | } |
| 3156 | |
| 3157 | // Try to find umax(a,b) - b or a - umin(a,b) patterns that may be converted to |
| 3158 | // usubsat(a,b), optionally as a truncated type. |
| 3159 | SDValue DAGCombiner::foldSubToUSubSat(EVT DstVT, SDNode *N) { |
| 3160 | if (N->getOpcode() != ISD::SUB || |
| 3161 | !(!LegalOperations || hasOperation(ISD::USUBSAT, DstVT))) |
| 3162 | return SDValue(); |
| 3163 | |
| 3164 | EVT SubVT = N->getValueType(0); |
| 3165 | SDValue Op0 = N->getOperand(0); |
| 3166 | SDValue Op1 = N->getOperand(1); |
| 3167 | |
| 3168 | // Try to find umax(a,b) - b or a - umin(a,b) patterns |
| 3169 | // they may be converted to usubsat(a,b). |
| 3170 | if (Op0.getOpcode() == ISD::UMAX) { |
| 3171 | SDValue MaxLHS = Op0.getOperand(0); |
| 3172 | SDValue MaxRHS = Op0.getOperand(1); |
| 3173 | if (MaxLHS == Op1) |
| 3174 | return getTruncatedUSUBSAT(DstVT, SubVT, MaxRHS, Op1, DAG, SDLoc(N)); |
| 3175 | if (MaxRHS == Op1) |
| 3176 | return getTruncatedUSUBSAT(DstVT, SubVT, MaxLHS, Op1, DAG, SDLoc(N)); |
| 3177 | } |
| 3178 | |
| 3179 | if (Op1.getOpcode() == ISD::UMIN) { |
| 3180 | SDValue MinLHS = Op1.getOperand(0); |
| 3181 | SDValue MinRHS = Op1.getOperand(1); |
| 3182 | if (MinLHS == Op0) |
| 3183 | return getTruncatedUSUBSAT(DstVT, SubVT, Op0, MinRHS, DAG, SDLoc(N)); |
| 3184 | if (MinRHS == Op0) |
| 3185 | return getTruncatedUSUBSAT(DstVT, SubVT, Op0, MinLHS, DAG, SDLoc(N)); |
| 3186 | } |
| 3187 | |
| 3188 | // sub(a,trunc(umin(zext(a),b))) -> usubsat(a,trunc(umin(b,SatLimit))) |
| 3189 | if (Op1.getOpcode() == ISD::TRUNCATE && |
| 3190 | Op1.getOperand(0).getOpcode() == ISD::UMIN) { |
| 3191 | SDValue MinLHS = Op1.getOperand(0).getOperand(0); |
| 3192 | SDValue MinRHS = Op1.getOperand(0).getOperand(1); |
| 3193 | if (MinLHS.getOpcode() == ISD::ZERO_EXTEND && MinLHS.getOperand(0) == Op0) |
| 3194 | return getTruncatedUSUBSAT(DstVT, MinLHS.getValueType(), MinLHS, MinRHS, |
| 3195 | DAG, SDLoc(N)); |
| 3196 | if (MinRHS.getOpcode() == ISD::ZERO_EXTEND && MinRHS.getOperand(0) == Op0) |
| 3197 | return getTruncatedUSUBSAT(DstVT, MinLHS.getValueType(), MinRHS, MinLHS, |
| 3198 | DAG, SDLoc(N)); |
| 3199 | } |
| 3200 | |
| 3201 | return SDValue(); |
| 3202 | } |
| 3203 | |
| 3204 | // Since it may not be valid to emit a fold to zero for vector initializers |
| 3205 | // check if we can before folding. |
| 3206 | static SDValue tryFoldToZero(const SDLoc &DL, const TargetLowering &TLI, EVT VT, |
| 3207 | SelectionDAG &DAG, bool LegalOperations) { |
| 3208 | if (!VT.isVector()) |
| 3209 | return DAG.getConstant(0, DL, VT); |
| 3210 | if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) |
| 3211 | return DAG.getConstant(0, DL, VT); |
| 3212 | return SDValue(); |
| 3213 | } |
| 3214 | |
| 3215 | SDValue DAGCombiner::visitSUB(SDNode *N) { |
| 3216 | SDValue N0 = N->getOperand(0); |
| 3217 | SDValue N1 = N->getOperand(1); |
| 3218 | EVT VT = N0.getValueType(); |
| 3219 | SDLoc DL(N); |
| 3220 | |
| 3221 | // fold vector ops |
| 3222 | if (VT.isVector()) { |
| 3223 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 3224 | return FoldedVOp; |
| 3225 | |
| 3226 | // fold (sub x, 0) -> x, vector edition |
| 3227 | if (ISD::isBuildVectorAllZeros(N1.getNode())) |
| 3228 | return N0; |
| 3229 | } |
| 3230 | |
| 3231 | // fold (sub x, x) -> 0 |
| 3232 | // FIXME: Refactor this and xor and other similar operations together. |
| 3233 | if (N0 == N1) |
| 3234 | return tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); |
| 3235 | |
| 3236 | // fold (sub c1, c2) -> c3 |
| 3237 | if (SDValue C = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N0, N1})) |
| 3238 | return C; |
| 3239 | |
| 3240 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 3241 | return NewSel; |
| 3242 | |
| 3243 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); |
| 3244 | |
| 3245 | // fold (sub x, c) -> (add x, -c) |
| 3246 | if (N1C) { |
| 3247 | return DAG.getNode(ISD::ADD, DL, VT, N0, |
| 3248 | DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); |
| 3249 | } |
| 3250 | |
| 3251 | if (isNullOrNullSplat(N0)) { |
| 3252 | unsigned BitWidth = VT.getScalarSizeInBits(); |
| 3253 | // Right-shifting everything out but the sign bit followed by negation is |
| 3254 | // the same as flipping arithmetic/logical shift type without the negation: |
| 3255 | // -(X >>u 31) -> (X >>s 31) |
| 3256 | // -(X >>s 31) -> (X >>u 31) |
| 3257 | if (N1->getOpcode() == ISD::SRA || N1->getOpcode() == ISD::SRL) { |
| 3258 | ConstantSDNode *ShiftAmt = isConstOrConstSplat(N1.getOperand(1)); |
| 3259 | if (ShiftAmt && ShiftAmt->getAPIntValue() == (BitWidth - 1)) { |
| 3260 | auto NewSh = N1->getOpcode() == ISD::SRA ? ISD::SRL : ISD::SRA; |
| 3261 | if (!LegalOperations || TLI.isOperationLegal(NewSh, VT)) |
| 3262 | return DAG.getNode(NewSh, DL, VT, N1.getOperand(0), N1.getOperand(1)); |
| 3263 | } |
| 3264 | } |
| 3265 | |
| 3266 | // 0 - X --> 0 if the sub is NUW. |
| 3267 | if (N->getFlags().hasNoUnsignedWrap()) |
| 3268 | return N0; |
| 3269 | |
| 3270 | if (DAG.MaskedValueIsZero(N1, ~APInt::getSignMask(BitWidth))) { |
| 3271 | // N1 is either 0 or the minimum signed value. If the sub is NSW, then |
| 3272 | // N1 must be 0 because negating the minimum signed value is undefined. |
| 3273 | if (N->getFlags().hasNoSignedWrap()) |
| 3274 | return N0; |
| 3275 | |
| 3276 | // 0 - X --> X if X is 0 or the minimum signed value. |
| 3277 | return N1; |
| 3278 | } |
| 3279 | |
| 3280 | // Convert 0 - abs(x). |
| 3281 | SDValue Result; |
| 3282 | if (N1->getOpcode() == ISD::ABS && |
| 3283 | !TLI.isOperationLegalOrCustom(ISD::ABS, VT) && |
| 3284 | TLI.expandABS(N1.getNode(), Result, DAG, true)) |
| 3285 | return Result; |
| 3286 | } |
| 3287 | |
| 3288 | // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) |
| 3289 | if (isAllOnesOrAllOnesSplat(N0)) |
| 3290 | return DAG.getNode(ISD::XOR, DL, VT, N1, N0); |
| 3291 | |
| 3292 | // fold (A - (0-B)) -> A+B |
| 3293 | if (N1.getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0))) |
| 3294 | return DAG.getNode(ISD::ADD, DL, VT, N0, N1.getOperand(1)); |
| 3295 | |
| 3296 | // fold A-(A-B) -> B |
| 3297 | if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0)) |
| 3298 | return N1.getOperand(1); |
| 3299 | |
| 3300 | // fold (A+B)-A -> B |
| 3301 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) |
| 3302 | return N0.getOperand(1); |
| 3303 | |
| 3304 | // fold (A+B)-B -> A |
| 3305 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) |
| 3306 | return N0.getOperand(0); |
| 3307 | |
| 3308 | // fold (A+C1)-C2 -> A+(C1-C2) |
| 3309 | if (N0.getOpcode() == ISD::ADD && |
| 3310 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && |
| 3311 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { |
| 3312 | SDValue NewC = |
| 3313 | DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N0.getOperand(1), N1}); |
| 3314 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3314, __PRETTY_FUNCTION__)); |
| 3315 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), NewC); |
| 3316 | } |
| 3317 | |
| 3318 | // fold C2-(A+C1) -> (C2-C1)-A |
| 3319 | if (N1.getOpcode() == ISD::ADD) { |
| 3320 | SDValue N11 = N1.getOperand(1); |
| 3321 | if (isConstantOrConstantVector(N0, /* NoOpaques */ true) && |
| 3322 | isConstantOrConstantVector(N11, /* NoOpaques */ true)) { |
| 3323 | SDValue NewC = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N0, N11}); |
| 3324 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3324, __PRETTY_FUNCTION__)); |
| 3325 | return DAG.getNode(ISD::SUB, DL, VT, NewC, N1.getOperand(0)); |
| 3326 | } |
| 3327 | } |
| 3328 | |
| 3329 | // fold (A-C1)-C2 -> A-(C1+C2) |
| 3330 | if (N0.getOpcode() == ISD::SUB && |
| 3331 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && |
| 3332 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { |
| 3333 | SDValue NewC = |
| 3334 | DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N0.getOperand(1), N1}); |
| 3335 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3335, __PRETTY_FUNCTION__)); |
| 3336 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), NewC); |
| 3337 | } |
| 3338 | |
| 3339 | // fold (c1-A)-c2 -> (c1-c2)-A |
| 3340 | if (N0.getOpcode() == ISD::SUB && |
| 3341 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && |
| 3342 | isConstantOrConstantVector(N0.getOperand(0), /* NoOpaques */ true)) { |
| 3343 | SDValue NewC = |
| 3344 | DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N0.getOperand(0), N1}); |
| 3345 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3345, __PRETTY_FUNCTION__)); |
| 3346 | return DAG.getNode(ISD::SUB, DL, VT, NewC, N0.getOperand(1)); |
| 3347 | } |
| 3348 | |
| 3349 | // fold ((A+(B+or-C))-B) -> A+or-C |
| 3350 | if (N0.getOpcode() == ISD::ADD && |
| 3351 | (N0.getOperand(1).getOpcode() == ISD::SUB || |
| 3352 | N0.getOperand(1).getOpcode() == ISD::ADD) && |
| 3353 | N0.getOperand(1).getOperand(0) == N1) |
| 3354 | return DAG.getNode(N0.getOperand(1).getOpcode(), DL, VT, N0.getOperand(0), |
| 3355 | N0.getOperand(1).getOperand(1)); |
| 3356 | |
| 3357 | // fold ((A+(C+B))-B) -> A+C |
| 3358 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1).getOpcode() == ISD::ADD && |
| 3359 | N0.getOperand(1).getOperand(1) == N1) |
| 3360 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), |
| 3361 | N0.getOperand(1).getOperand(0)); |
| 3362 | |
| 3363 | // fold ((A-(B-C))-C) -> A-B |
| 3364 | if (N0.getOpcode() == ISD::SUB && N0.getOperand(1).getOpcode() == ISD::SUB && |
| 3365 | N0.getOperand(1).getOperand(1) == N1) |
| 3366 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), |
| 3367 | N0.getOperand(1).getOperand(0)); |
| 3368 | |
| 3369 | // fold (A-(B-C)) -> A+(C-B) |
| 3370 | if (N1.getOpcode() == ISD::SUB && N1.hasOneUse()) |
| 3371 | return DAG.getNode(ISD::ADD, DL, VT, N0, |
| 3372 | DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(1), |
| 3373 | N1.getOperand(0))); |
| 3374 | |
| 3375 | // A - (A & B) -> A & (~B) |
| 3376 | if (N1.getOpcode() == ISD::AND) { |
| 3377 | SDValue A = N1.getOperand(0); |
| 3378 | SDValue B = N1.getOperand(1); |
| 3379 | if (A != N0) |
| 3380 | std::swap(A, B); |
| 3381 | if (A == N0 && |
| 3382 | (N1.hasOneUse() || isConstantOrConstantVector(B, /*NoOpaques=*/true))) { |
| 3383 | SDValue InvB = |
| 3384 | DAG.getNode(ISD::XOR, DL, VT, B, DAG.getAllOnesConstant(DL, VT)); |
| 3385 | return DAG.getNode(ISD::AND, DL, VT, A, InvB); |
| 3386 | } |
| 3387 | } |
| 3388 | |
| 3389 | // fold (X - (-Y * Z)) -> (X + (Y * Z)) |
| 3390 | if (N1.getOpcode() == ISD::MUL && N1.hasOneUse()) { |
| 3391 | if (N1.getOperand(0).getOpcode() == ISD::SUB && |
| 3392 | isNullOrNullSplat(N1.getOperand(0).getOperand(0))) { |
| 3393 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, |
| 3394 | N1.getOperand(0).getOperand(1), |
| 3395 | N1.getOperand(1)); |
| 3396 | return DAG.getNode(ISD::ADD, DL, VT, N0, Mul); |
| 3397 | } |
| 3398 | if (N1.getOperand(1).getOpcode() == ISD::SUB && |
| 3399 | isNullOrNullSplat(N1.getOperand(1).getOperand(0))) { |
| 3400 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, |
| 3401 | N1.getOperand(0), |
| 3402 | N1.getOperand(1).getOperand(1)); |
| 3403 | return DAG.getNode(ISD::ADD, DL, VT, N0, Mul); |
| 3404 | } |
| 3405 | } |
| 3406 | |
| 3407 | // If either operand of a sub is undef, the result is undef |
| 3408 | if (N0.isUndef()) |
| 3409 | return N0; |
| 3410 | if (N1.isUndef()) |
| 3411 | return N1; |
| 3412 | |
| 3413 | if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG)) |
| 3414 | return V; |
| 3415 | |
| 3416 | if (SDValue V = foldAddSubOfSignBit(N, DAG)) |
| 3417 | return V; |
| 3418 | |
| 3419 | if (SDValue V = foldAddSubMasked1(false, N0, N1, DAG, SDLoc(N))) |
| 3420 | return V; |
| 3421 | |
| 3422 | if (SDValue V = foldSubToUSubSat(VT, N)) |
| 3423 | return V; |
| 3424 | |
| 3425 | // (x - y) - 1 -> add (xor y, -1), x |
| 3426 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && isOneOrOneSplat(N1)) { |
| 3427 | SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1), |
| 3428 | DAG.getAllOnesConstant(DL, VT)); |
| 3429 | return DAG.getNode(ISD::ADD, DL, VT, Xor, N0.getOperand(0)); |
| 3430 | } |
| 3431 | |
| 3432 | // Look for: |
| 3433 | // sub y, (xor x, -1) |
| 3434 | // And if the target does not like this form then turn into: |
| 3435 | // add (add x, y), 1 |
| 3436 | if (TLI.preferIncOfAddToSubOfNot(VT) && N1.hasOneUse() && isBitwiseNot(N1)) { |
| 3437 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, N1.getOperand(0)); |
| 3438 | return DAG.getNode(ISD::ADD, DL, VT, Add, DAG.getConstant(1, DL, VT)); |
| 3439 | } |
| 3440 | |
| 3441 | // Hoist one-use addition by non-opaque constant: |
| 3442 | // (x + C) - y -> (x - y) + C |
| 3443 | if (N0.hasOneUse() && N0.getOpcode() == ISD::ADD && |
| 3444 | isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { |
| 3445 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1); |
| 3446 | return DAG.getNode(ISD::ADD, DL, VT, Sub, N0.getOperand(1)); |
| 3447 | } |
| 3448 | // y - (x + C) -> (y - x) - C |
| 3449 | if (N1.hasOneUse() && N1.getOpcode() == ISD::ADD && |
| 3450 | isConstantOrConstantVector(N1.getOperand(1), /*NoOpaques=*/true)) { |
| 3451 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(0)); |
| 3452 | return DAG.getNode(ISD::SUB, DL, VT, Sub, N1.getOperand(1)); |
| 3453 | } |
| 3454 | // (x - C) - y -> (x - y) - C |
| 3455 | // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors. |
| 3456 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && |
| 3457 | isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { |
| 3458 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1); |
| 3459 | return DAG.getNode(ISD::SUB, DL, VT, Sub, N0.getOperand(1)); |
| 3460 | } |
| 3461 | // (C - x) - y -> C - (x + y) |
| 3462 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && |
| 3463 | isConstantOrConstantVector(N0.getOperand(0), /*NoOpaques=*/true)) { |
| 3464 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(1), N1); |
| 3465 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), Add); |
| 3466 | } |
| 3467 | |
| 3468 | // If the target's bool is represented as 0/-1, prefer to make this 'add 0/-1' |
| 3469 | // rather than 'sub 0/1' (the sext should get folded). |
| 3470 | // sub X, (zext i1 Y) --> add X, (sext i1 Y) |
| 3471 | if (N1.getOpcode() == ISD::ZERO_EXTEND && |
| 3472 | N1.getOperand(0).getScalarValueSizeInBits() == 1 && |
| 3473 | TLI.getBooleanContents(VT) == |
| 3474 | TargetLowering::ZeroOrNegativeOneBooleanContent) { |
| 3475 | SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, N1.getOperand(0)); |
| 3476 | return DAG.getNode(ISD::ADD, DL, VT, N0, SExt); |
| 3477 | } |
| 3478 | |
| 3479 | // fold Y = sra (X, size(X)-1); sub (xor (X, Y), Y) -> (abs X) |
| 3480 | if (TLI.isOperationLegalOrCustom(ISD::ABS, VT)) { |
| 3481 | if (N0.getOpcode() == ISD::XOR && N1.getOpcode() == ISD::SRA) { |
| 3482 | SDValue X0 = N0.getOperand(0), X1 = N0.getOperand(1); |
| 3483 | SDValue S0 = N1.getOperand(0); |
| 3484 | if ((X0 == S0 && X1 == N1) || (X0 == N1 && X1 == S0)) |
| 3485 | if (ConstantSDNode *C = isConstOrConstSplat(N1.getOperand(1))) |
| 3486 | if (C->getAPIntValue() == (VT.getScalarSizeInBits() - 1)) |
| 3487 | return DAG.getNode(ISD::ABS, SDLoc(N), VT, S0); |
| 3488 | } |
| 3489 | } |
| 3490 | |
| 3491 | // If the relocation model supports it, consider symbol offsets. |
| 3492 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) |
| 3493 | if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) { |
| 3494 | // fold (sub Sym, c) -> Sym-c |
| 3495 | if (N1C && GA->getOpcode() == ISD::GlobalAddress) |
| 3496 | return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, |
| 3497 | GA->getOffset() - |
| 3498 | (uint64_t)N1C->getSExtValue()); |
| 3499 | // fold (sub Sym+c1, Sym+c2) -> c1-c2 |
| 3500 | if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1)) |
| 3501 | if (GA->getGlobal() == GB->getGlobal()) |
| 3502 | return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(), |
| 3503 | DL, VT); |
| 3504 | } |
| 3505 | |
| 3506 | // sub X, (sextinreg Y i1) -> add X, (and Y 1) |
| 3507 | if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { |
| 3508 | VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1)); |
| 3509 | if (TN->getVT() == MVT::i1) { |
| 3510 | SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), |
| 3511 | DAG.getConstant(1, DL, VT)); |
| 3512 | return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt); |
| 3513 | } |
| 3514 | } |
| 3515 | |
| 3516 | // canonicalize (sub X, (vscale * C)) to (add X, (vscale * -C)) |
| 3517 | if (N1.getOpcode() == ISD::VSCALE) { |
| 3518 | const APInt &IntVal = N1.getConstantOperandAPInt(0); |
| 3519 | return DAG.getNode(ISD::ADD, DL, VT, N0, DAG.getVScale(DL, VT, -IntVal)); |
| 3520 | } |
| 3521 | |
| 3522 | // Prefer an add for more folding potential and possibly better codegen: |
| 3523 | // sub N0, (lshr N10, width-1) --> add N0, (ashr N10, width-1) |
| 3524 | if (!LegalOperations && N1.getOpcode() == ISD::SRL && N1.hasOneUse()) { |
| 3525 | SDValue ShAmt = N1.getOperand(1); |
| 3526 | ConstantSDNode *ShAmtC = isConstOrConstSplat(ShAmt); |
| 3527 | if (ShAmtC && |
| 3528 | ShAmtC->getAPIntValue() == (N1.getScalarValueSizeInBits() - 1)) { |
| 3529 | SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, N1.getOperand(0), ShAmt); |
| 3530 | return DAG.getNode(ISD::ADD, DL, VT, N0, SRA); |
| 3531 | } |
| 3532 | } |
| 3533 | |
| 3534 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) { |
| 3535 | // (sub Carry, X) -> (addcarry (sub 0, X), 0, Carry) |
| 3536 | if (SDValue Carry = getAsCarry(TLI, N0)) { |
| 3537 | SDValue X = N1; |
| 3538 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 3539 | SDValue NegX = DAG.getNode(ISD::SUB, DL, VT, Zero, X); |
| 3540 | return DAG.getNode(ISD::ADDCARRY, DL, |
| 3541 | DAG.getVTList(VT, Carry.getValueType()), NegX, Zero, |
| 3542 | Carry); |
| 3543 | } |
| 3544 | } |
| 3545 | |
| 3546 | return SDValue(); |
| 3547 | } |
| 3548 | |
| 3549 | SDValue DAGCombiner::visitSUBSAT(SDNode *N) { |
| 3550 | SDValue N0 = N->getOperand(0); |
| 3551 | SDValue N1 = N->getOperand(1); |
| 3552 | EVT VT = N0.getValueType(); |
| 3553 | SDLoc DL(N); |
| 3554 | |
| 3555 | // fold vector ops |
| 3556 | if (VT.isVector()) { |
| 3557 | // TODO SimplifyVBinOp |
| 3558 | |
| 3559 | // fold (sub_sat x, 0) -> x, vector edition |
| 3560 | if (ISD::isBuildVectorAllZeros(N1.getNode())) |
| 3561 | return N0; |
| 3562 | } |
| 3563 | |
| 3564 | // fold (sub_sat x, undef) -> 0 |
| 3565 | if (N0.isUndef() || N1.isUndef()) |
| 3566 | return DAG.getConstant(0, DL, VT); |
| 3567 | |
| 3568 | // fold (sub_sat x, x) -> 0 |
| 3569 | if (N0 == N1) |
| 3570 | return DAG.getConstant(0, DL, VT); |
| 3571 | |
| 3572 | // fold (sub_sat c1, c2) -> c3 |
| 3573 | if (SDValue C = DAG.FoldConstantArithmetic(N->getOpcode(), DL, VT, {N0, N1})) |
| 3574 | return C; |
| 3575 | |
| 3576 | // fold (sub_sat x, 0) -> x |
| 3577 | if (isNullConstant(N1)) |
| 3578 | return N0; |
| 3579 | |
| 3580 | return SDValue(); |
| 3581 | } |
| 3582 | |
| 3583 | SDValue DAGCombiner::visitSUBC(SDNode *N) { |
| 3584 | SDValue N0 = N->getOperand(0); |
| 3585 | SDValue N1 = N->getOperand(1); |
| 3586 | EVT VT = N0.getValueType(); |
| 3587 | SDLoc DL(N); |
| 3588 | |
| 3589 | // If the flag result is dead, turn this into an SUB. |
| 3590 | if (!N->hasAnyUseOfValue(1)) |
| 3591 | return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), |
| 3592 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); |
| 3593 | |
| 3594 | // fold (subc x, x) -> 0 + no borrow |
| 3595 | if (N0 == N1) |
| 3596 | return CombineTo(N, DAG.getConstant(0, DL, VT), |
| 3597 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); |
| 3598 | |
| 3599 | // fold (subc x, 0) -> x + no borrow |
| 3600 | if (isNullConstant(N1)) |
| 3601 | return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); |
| 3602 | |
| 3603 | // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow |
| 3604 | if (isAllOnesConstant(N0)) |
| 3605 | return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), |
| 3606 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); |
| 3607 | |
| 3608 | return SDValue(); |
| 3609 | } |
| 3610 | |
| 3611 | SDValue DAGCombiner::visitSUBO(SDNode *N) { |
| 3612 | SDValue N0 = N->getOperand(0); |
| 3613 | SDValue N1 = N->getOperand(1); |
| 3614 | EVT VT = N0.getValueType(); |
| 3615 | bool IsSigned = (ISD::SSUBO == N->getOpcode()); |
| 3616 | |
| 3617 | EVT CarryVT = N->getValueType(1); |
| 3618 | SDLoc DL(N); |
| 3619 | |
| 3620 | // If the flag result is dead, turn this into an SUB. |
| 3621 | if (!N->hasAnyUseOfValue(1)) |
| 3622 | return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), |
| 3623 | DAG.getUNDEF(CarryVT)); |
| 3624 | |
| 3625 | // fold (subo x, x) -> 0 + no borrow |
| 3626 | if (N0 == N1) |
| 3627 | return CombineTo(N, DAG.getConstant(0, DL, VT), |
| 3628 | DAG.getConstant(0, DL, CarryVT)); |
| 3629 | |
| 3630 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); |
| 3631 | |
| 3632 | // fold (subox, c) -> (addo x, -c) |
| 3633 | if (IsSigned && N1C && !N1C->getAPIntValue().isMinSignedValue()) { |
| 3634 | return DAG.getNode(ISD::SADDO, DL, N->getVTList(), N0, |
| 3635 | DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); |
| 3636 | } |
| 3637 | |
| 3638 | // fold (subo x, 0) -> x + no borrow |
| 3639 | if (isNullOrNullSplat(N1)) |
| 3640 | return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); |
| 3641 | |
| 3642 | // Canonicalize (usubo -1, x) -> ~x, i.e. (xor x, -1) + no borrow |
| 3643 | if (!IsSigned && isAllOnesOrAllOnesSplat(N0)) |
| 3644 | return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), |
| 3645 | DAG.getConstant(0, DL, CarryVT)); |
| 3646 | |
| 3647 | return SDValue(); |
| 3648 | } |
| 3649 | |
| 3650 | SDValue DAGCombiner::visitSUBE(SDNode *N) { |
| 3651 | SDValue N0 = N->getOperand(0); |
| 3652 | SDValue N1 = N->getOperand(1); |
| 3653 | SDValue CarryIn = N->getOperand(2); |
| 3654 | |
| 3655 | // fold (sube x, y, false) -> (subc x, y) |
| 3656 | if (CarryIn.getOpcode() == ISD::CARRY_FALSE) |
| 3657 | return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1); |
| 3658 | |
| 3659 | return SDValue(); |
| 3660 | } |
| 3661 | |
| 3662 | SDValue DAGCombiner::visitSUBCARRY(SDNode *N) { |
| 3663 | SDValue N0 = N->getOperand(0); |
| 3664 | SDValue N1 = N->getOperand(1); |
| 3665 | SDValue CarryIn = N->getOperand(2); |
| 3666 | |
| 3667 | // fold (subcarry x, y, false) -> (usubo x, y) |
| 3668 | if (isNullConstant(CarryIn)) { |
| 3669 | if (!LegalOperations || |
| 3670 | TLI.isOperationLegalOrCustom(ISD::USUBO, N->getValueType(0))) |
| 3671 | return DAG.getNode(ISD::USUBO, SDLoc(N), N->getVTList(), N0, N1); |
| 3672 | } |
| 3673 | |
| 3674 | return SDValue(); |
| 3675 | } |
| 3676 | |
| 3677 | SDValue DAGCombiner::visitSSUBO_CARRY(SDNode *N) { |
| 3678 | SDValue N0 = N->getOperand(0); |
| 3679 | SDValue N1 = N->getOperand(1); |
| 3680 | SDValue CarryIn = N->getOperand(2); |
| 3681 | |
| 3682 | // fold (ssubo_carry x, y, false) -> (ssubo x, y) |
| 3683 | if (isNullConstant(CarryIn)) { |
| 3684 | if (!LegalOperations || |
| 3685 | TLI.isOperationLegalOrCustom(ISD::SSUBO, N->getValueType(0))) |
| 3686 | return DAG.getNode(ISD::SSUBO, SDLoc(N), N->getVTList(), N0, N1); |
| 3687 | } |
| 3688 | |
| 3689 | return SDValue(); |
| 3690 | } |
| 3691 | |
| 3692 | // Notice that "mulfix" can be any of SMULFIX, SMULFIXSAT, UMULFIX and |
| 3693 | // UMULFIXSAT here. |
| 3694 | SDValue DAGCombiner::visitMULFIX(SDNode *N) { |
| 3695 | SDValue N0 = N->getOperand(0); |
| 3696 | SDValue N1 = N->getOperand(1); |
| 3697 | SDValue Scale = N->getOperand(2); |
| 3698 | EVT VT = N0.getValueType(); |
| 3699 | |
| 3700 | // fold (mulfix x, undef, scale) -> 0 |
| 3701 | if (N0.isUndef() || N1.isUndef()) |
| 3702 | return DAG.getConstant(0, SDLoc(N), VT); |
| 3703 | |
| 3704 | // Canonicalize constant to RHS (vector doesn't have to splat) |
| 3705 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && |
| 3706 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 3707 | return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0, Scale); |
| 3708 | |
| 3709 | // fold (mulfix x, 0, scale) -> 0 |
| 3710 | if (isNullConstant(N1)) |
| 3711 | return DAG.getConstant(0, SDLoc(N), VT); |
| 3712 | |
| 3713 | return SDValue(); |
| 3714 | } |
| 3715 | |
| 3716 | SDValue DAGCombiner::visitMUL(SDNode *N) { |
| 3717 | SDValue N0 = N->getOperand(0); |
| 3718 | SDValue N1 = N->getOperand(1); |
| 3719 | EVT VT = N0.getValueType(); |
| 3720 | |
| 3721 | // fold (mul x, undef) -> 0 |
| 3722 | if (N0.isUndef() || N1.isUndef()) |
| 3723 | return DAG.getConstant(0, SDLoc(N), VT); |
| 3724 | |
| 3725 | bool N1IsConst = false; |
| 3726 | bool N1IsOpaqueConst = false; |
| 3727 | APInt ConstValue1; |
| 3728 | |
| 3729 | // fold vector ops |
| 3730 | if (VT.isVector()) { |
| 3731 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 3732 | return FoldedVOp; |
| 3733 | |
| 3734 | N1IsConst = ISD::isConstantSplatVector(N1.getNode(), ConstValue1); |
| 3735 | assert((!N1IsConst ||(((!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits ()) && "Splat APInt should be element width") ? static_cast <void> (0) : __assert_fail ("(!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3737, __PRETTY_FUNCTION__)) |
| 3736 | ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) &&(((!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits ()) && "Splat APInt should be element width") ? static_cast <void> (0) : __assert_fail ("(!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3737, __PRETTY_FUNCTION__)) |
| 3737 | "Splat APInt should be element width")(((!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits ()) && "Splat APInt should be element width") ? static_cast <void> (0) : __assert_fail ("(!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3737, __PRETTY_FUNCTION__)); |
| 3738 | } else { |
| 3739 | N1IsConst = isa<ConstantSDNode>(N1); |
| 3740 | if (N1IsConst) { |
| 3741 | ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue(); |
| 3742 | N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque(); |
| 3743 | } |
| 3744 | } |
| 3745 | |
| 3746 | // fold (mul c1, c2) -> c1*c2 |
| 3747 | if (SDValue C = DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT, {N0, N1})) |
| 3748 | return C; |
| 3749 | |
| 3750 | // canonicalize constant to RHS (vector doesn't have to splat) |
| 3751 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && |
| 3752 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 3753 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0); |
| 3754 | |
| 3755 | // fold (mul x, 0) -> 0 |
| 3756 | if (N1IsConst && ConstValue1.isNullValue()) |
| 3757 | return N1; |
| 3758 | |
| 3759 | // fold (mul x, 1) -> x |
| 3760 | if (N1IsConst && ConstValue1.isOneValue()) |
| 3761 | return N0; |
| 3762 | |
| 3763 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 3764 | return NewSel; |
| 3765 | |
| 3766 | // fold (mul x, -1) -> 0-x |
| 3767 | if (N1IsConst && ConstValue1.isAllOnesValue()) { |
| 3768 | SDLoc DL(N); |
| 3769 | return DAG.getNode(ISD::SUB, DL, VT, |
| 3770 | DAG.getConstant(0, DL, VT), N0); |
| 3771 | } |
| 3772 | |
| 3773 | // fold (mul x, (1 << c)) -> x << c |
| 3774 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && |
| 3775 | DAG.isKnownToBeAPowerOfTwo(N1) && |
| 3776 | (!VT.isVector() || Level <= AfterLegalizeVectorOps)) { |
| 3777 | SDLoc DL(N); |
| 3778 | SDValue LogBase2 = BuildLogBase2(N1, DL); |
| 3779 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); |
| 3780 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); |
| 3781 | return DAG.getNode(ISD::SHL, DL, VT, N0, Trunc); |
| 3782 | } |
| 3783 | |
| 3784 | // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c |
| 3785 | if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2()) { |
| 3786 | unsigned Log2Val = (-ConstValue1).logBase2(); |
| 3787 | SDLoc DL(N); |
| 3788 | // FIXME: If the input is something that is easily negated (e.g. a |
| 3789 | // single-use add), we should put the negate there. |
| 3790 | return DAG.getNode(ISD::SUB, DL, VT, |
| 3791 | DAG.getConstant(0, DL, VT), |
| 3792 | DAG.getNode(ISD::SHL, DL, VT, N0, |
| 3793 | DAG.getConstant(Log2Val, DL, |
| 3794 | getShiftAmountTy(N0.getValueType())))); |
| 3795 | } |
| 3796 | |
| 3797 | // Try to transform: |
| 3798 | // (1) multiply-by-(power-of-2 +/- 1) into shift and add/sub. |
| 3799 | // mul x, (2^N + 1) --> add (shl x, N), x |
| 3800 | // mul x, (2^N - 1) --> sub (shl x, N), x |
| 3801 | // Examples: x * 33 --> (x << 5) + x |
| 3802 | // x * 15 --> (x << 4) - x |
| 3803 | // x * -33 --> -((x << 5) + x) |
| 3804 | // x * -15 --> -((x << 4) - x) ; this reduces --> x - (x << 4) |
| 3805 | // (2) multiply-by-(power-of-2 +/- power-of-2) into shifts and add/sub. |
| 3806 | // mul x, (2^N + 2^M) --> (add (shl x, N), (shl x, M)) |
| 3807 | // mul x, (2^N - 2^M) --> (sub (shl x, N), (shl x, M)) |
| 3808 | // Examples: x * 0x8800 --> (x << 15) + (x << 11) |
| 3809 | // x * 0xf800 --> (x << 16) - (x << 11) |
| 3810 | // x * -0x8800 --> -((x << 15) + (x << 11)) |
| 3811 | // x * -0xf800 --> -((x << 16) - (x << 11)) ; (x << 11) - (x << 16) |
| 3812 | if (N1IsConst && TLI.decomposeMulByConstant(*DAG.getContext(), VT, N1)) { |
| 3813 | // TODO: We could handle more general decomposition of any constant by |
| 3814 | // having the target set a limit on number of ops and making a |
| 3815 | // callback to determine that sequence (similar to sqrt expansion). |
| 3816 | unsigned MathOp = ISD::DELETED_NODE; |
| 3817 | APInt MulC = ConstValue1.abs(); |
| 3818 | // The constant `2` should be treated as (2^0 + 1). |
| 3819 | unsigned TZeros = MulC == 2 ? 0 : MulC.countTrailingZeros(); |
| 3820 | MulC.lshrInPlace(TZeros); |
| 3821 | if ((MulC - 1).isPowerOf2()) |
| 3822 | MathOp = ISD::ADD; |
| 3823 | else if ((MulC + 1).isPowerOf2()) |
| 3824 | MathOp = ISD::SUB; |
| 3825 | |
| 3826 | if (MathOp != ISD::DELETED_NODE) { |
| 3827 | unsigned ShAmt = |
| 3828 | MathOp == ISD::ADD ? (MulC - 1).logBase2() : (MulC + 1).logBase2(); |
| 3829 | ShAmt += TZeros; |
| 3830 | assert(ShAmt < VT.getScalarSizeInBits() &&((ShAmt < VT.getScalarSizeInBits() && "multiply-by-constant generated out of bounds shift" ) ? static_cast<void> (0) : __assert_fail ("ShAmt < VT.getScalarSizeInBits() && \"multiply-by-constant generated out of bounds shift\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3831, __PRETTY_FUNCTION__)) |
| 3831 | "multiply-by-constant generated out of bounds shift")((ShAmt < VT.getScalarSizeInBits() && "multiply-by-constant generated out of bounds shift" ) ? static_cast<void> (0) : __assert_fail ("ShAmt < VT.getScalarSizeInBits() && \"multiply-by-constant generated out of bounds shift\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3831, __PRETTY_FUNCTION__)); |
| 3832 | SDLoc DL(N); |
| 3833 | SDValue Shl = |
| 3834 | DAG.getNode(ISD::SHL, DL, VT, N0, DAG.getConstant(ShAmt, DL, VT)); |
| 3835 | SDValue R = |
| 3836 | TZeros ? DAG.getNode(MathOp, DL, VT, Shl, |
| 3837 | DAG.getNode(ISD::SHL, DL, VT, N0, |
| 3838 | DAG.getConstant(TZeros, DL, VT))) |
| 3839 | : DAG.getNode(MathOp, DL, VT, Shl, N0); |
| 3840 | if (ConstValue1.isNegative()) |
| 3841 | R = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), R); |
| 3842 | return R; |
| 3843 | } |
| 3844 | } |
| 3845 | |
| 3846 | // (mul (shl X, c1), c2) -> (mul X, c2 << c1) |
| 3847 | if (N0.getOpcode() == ISD::SHL && |
| 3848 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && |
| 3849 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { |
| 3850 | SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT, N1, N0.getOperand(1)); |
| 3851 | if (isConstantOrConstantVector(C3)) |
| 3852 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), C3); |
| 3853 | } |
| 3854 | |
| 3855 | // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one |
| 3856 | // use. |
| 3857 | { |
| 3858 | SDValue Sh(nullptr, 0), Y(nullptr, 0); |
| 3859 | |
| 3860 | // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). |
| 3861 | if (N0.getOpcode() == ISD::SHL && |
| 3862 | isConstantOrConstantVector(N0.getOperand(1)) && |
| 3863 | N0.getNode()->hasOneUse()) { |
| 3864 | Sh = N0; Y = N1; |
| 3865 | } else if (N1.getOpcode() == ISD::SHL && |
| 3866 | isConstantOrConstantVector(N1.getOperand(1)) && |
| 3867 | N1.getNode()->hasOneUse()) { |
| 3868 | Sh = N1; Y = N0; |
| 3869 | } |
| 3870 | |
| 3871 | if (Sh.getNode()) { |
| 3872 | SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, Sh.getOperand(0), Y); |
| 3873 | return DAG.getNode(ISD::SHL, SDLoc(N), VT, Mul, Sh.getOperand(1)); |
| 3874 | } |
| 3875 | } |
| 3876 | |
| 3877 | // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) |
| 3878 | if (DAG.isConstantIntBuildVectorOrConstantInt(N1) && |
| 3879 | N0.getOpcode() == ISD::ADD && |
| 3880 | DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)) && |
| 3881 | isMulAddWithConstProfitable(N, N0, N1)) |
| 3882 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, |
| 3883 | DAG.getNode(ISD::MUL, SDLoc(N0), VT, |
| 3884 | N0.getOperand(0), N1), |
| 3885 | DAG.getNode(ISD::MUL, SDLoc(N1), VT, |
| 3886 | N0.getOperand(1), N1)); |
| 3887 | |
| 3888 | // Fold (mul (vscale * C0), C1) to (vscale * (C0 * C1)). |
| 3889 | if (N0.getOpcode() == ISD::VSCALE) |
| 3890 | if (ConstantSDNode *NC1 = isConstOrConstSplat(N1)) { |
| 3891 | const APInt &C0 = N0.getConstantOperandAPInt(0); |
| 3892 | const APInt &C1 = NC1->getAPIntValue(); |
| 3893 | return DAG.getVScale(SDLoc(N), VT, C0 * C1); |
| 3894 | } |
| 3895 | |
| 3896 | // Fold ((mul x, 0/undef) -> 0, |
| 3897 | // (mul x, 1) -> x) -> x) |
| 3898 | // -> and(x, mask) |
| 3899 | // We can replace vectors with '0' and '1' factors with a clearing mask. |
| 3900 | if (VT.isFixedLengthVector()) { |
| 3901 | unsigned NumElts = VT.getVectorNumElements(); |
| 3902 | SmallBitVector ClearMask; |
| 3903 | ClearMask.reserve(NumElts); |
| 3904 | auto IsClearMask = [&ClearMask](ConstantSDNode *V) { |
| 3905 | if (!V || V->isNullValue()) { |
| 3906 | ClearMask.push_back(true); |
| 3907 | return true; |
| 3908 | } |
| 3909 | ClearMask.push_back(false); |
| 3910 | return V->isOne(); |
| 3911 | }; |
| 3912 | if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::AND, VT)) && |
| 3913 | ISD::matchUnaryPredicate(N1, IsClearMask, /*AllowUndefs*/ true)) { |
| 3914 | assert(N1.getOpcode() == ISD::BUILD_VECTOR && "Unknown constant vector")((N1.getOpcode() == ISD::BUILD_VECTOR && "Unknown constant vector" ) ? static_cast<void> (0) : __assert_fail ("N1.getOpcode() == ISD::BUILD_VECTOR && \"Unknown constant vector\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3914, __PRETTY_FUNCTION__)); |
| 3915 | SDLoc DL(N); |
| 3916 | EVT LegalSVT = N1.getOperand(0).getValueType(); |
| 3917 | SDValue Zero = DAG.getConstant(0, DL, LegalSVT); |
| 3918 | SDValue AllOnes = DAG.getAllOnesConstant(DL, LegalSVT); |
| 3919 | SmallVector<SDValue, 16> Mask(NumElts, AllOnes); |
| 3920 | for (unsigned I = 0; I != NumElts; ++I) |
| 3921 | if (ClearMask[I]) |
| 3922 | Mask[I] = Zero; |
| 3923 | return DAG.getNode(ISD::AND, DL, VT, N0, DAG.getBuildVector(VT, DL, Mask)); |
| 3924 | } |
| 3925 | } |
| 3926 | |
| 3927 | // reassociate mul |
| 3928 | if (SDValue RMUL = reassociateOps(ISD::MUL, SDLoc(N), N0, N1, N->getFlags())) |
| 3929 | return RMUL; |
| 3930 | |
| 3931 | return SDValue(); |
| 3932 | } |
| 3933 | |
| 3934 | /// Return true if divmod libcall is available. |
| 3935 | static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned, |
| 3936 | const TargetLowering &TLI) { |
| 3937 | RTLIB::Libcall LC; |
| 3938 | EVT NodeType = Node->getValueType(0); |
| 3939 | if (!NodeType.isSimple()) |
| 3940 | return false; |
| 3941 | switch (NodeType.getSimpleVT().SimpleTy) { |
| 3942 | default: return false; // No libcall for vector types. |
| 3943 | case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break; |
| 3944 | case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break; |
| 3945 | case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break; |
| 3946 | case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break; |
| 3947 | case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break; |
| 3948 | } |
| 3949 | |
| 3950 | return TLI.getLibcallName(LC) != nullptr; |
| 3951 | } |
| 3952 | |
| 3953 | /// Issue divrem if both quotient and remainder are needed. |
| 3954 | SDValue DAGCombiner::useDivRem(SDNode *Node) { |
| 3955 | if (Node->use_empty()) |
| 3956 | return SDValue(); // This is a dead node, leave it alone. |
| 3957 | |
| 3958 | unsigned Opcode = Node->getOpcode(); |
| 3959 | bool isSigned = (Opcode == ISD::SDIV) || (Opcode == ISD::SREM); |
| 3960 | unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; |
| 3961 | |
| 3962 | // DivMod lib calls can still work on non-legal types if using lib-calls. |
| 3963 | EVT VT = Node->getValueType(0); |
| 3964 | if (VT.isVector() || !VT.isInteger()) |
| 3965 | return SDValue(); |
| 3966 | |
| 3967 | if (!TLI.isTypeLegal(VT) && !TLI.isOperationCustom(DivRemOpc, VT)) |
| 3968 | return SDValue(); |
| 3969 | |
| 3970 | // If DIVREM is going to get expanded into a libcall, |
| 3971 | // but there is no libcall available, then don't combine. |
| 3972 | if (!TLI.isOperationLegalOrCustom(DivRemOpc, VT) && |
| 3973 | !isDivRemLibcallAvailable(Node, isSigned, TLI)) |
| 3974 | return SDValue(); |
| 3975 | |
| 3976 | // If div is legal, it's better to do the normal expansion |
| 3977 | unsigned OtherOpcode = 0; |
| 3978 | if ((Opcode == ISD::SDIV) || (Opcode == ISD::UDIV)) { |
| 3979 | OtherOpcode = isSigned ? ISD::SREM : ISD::UREM; |
| 3980 | if (TLI.isOperationLegalOrCustom(Opcode, VT)) |
| 3981 | return SDValue(); |
| 3982 | } else { |
| 3983 | OtherOpcode = isSigned ? ISD::SDIV : ISD::UDIV; |
| 3984 | if (TLI.isOperationLegalOrCustom(OtherOpcode, VT)) |
| 3985 | return SDValue(); |
| 3986 | } |
| 3987 | |
| 3988 | SDValue Op0 = Node->getOperand(0); |
| 3989 | SDValue Op1 = Node->getOperand(1); |
| 3990 | SDValue combined; |
| 3991 | for (SDNode::use_iterator UI = Op0.getNode()->use_begin(), |
| 3992 | UE = Op0.getNode()->use_end(); UI != UE; ++UI) { |
| 3993 | SDNode *User = *UI; |
| 3994 | if (User == Node || User->getOpcode() == ISD::DELETED_NODE || |
| 3995 | User->use_empty()) |
| 3996 | continue; |
| 3997 | // Convert the other matching node(s), too; |
| 3998 | // otherwise, the DIVREM may get target-legalized into something |
| 3999 | // target-specific that we won't be able to recognize. |
| 4000 | unsigned UserOpc = User->getOpcode(); |
| 4001 | if ((UserOpc == Opcode || UserOpc == OtherOpcode || UserOpc == DivRemOpc) && |
| 4002 | User->getOperand(0) == Op0 && |
| 4003 | User->getOperand(1) == Op1) { |
| 4004 | if (!combined) { |
| 4005 | if (UserOpc == OtherOpcode) { |
| 4006 | SDVTList VTs = DAG.getVTList(VT, VT); |
| 4007 | combined = DAG.getNode(DivRemOpc, SDLoc(Node), VTs, Op0, Op1); |
| 4008 | } else if (UserOpc == DivRemOpc) { |
| 4009 | combined = SDValue(User, 0); |
| 4010 | } else { |
| 4011 | assert(UserOpc == Opcode)((UserOpc == Opcode) ? static_cast<void> (0) : __assert_fail ("UserOpc == Opcode", "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4011, __PRETTY_FUNCTION__)); |
| 4012 | continue; |
| 4013 | } |
| 4014 | } |
| 4015 | if (UserOpc == ISD::SDIV || UserOpc == ISD::UDIV) |
| 4016 | CombineTo(User, combined); |
| 4017 | else if (UserOpc == ISD::SREM || UserOpc == ISD::UREM) |
| 4018 | CombineTo(User, combined.getValue(1)); |
| 4019 | } |
| 4020 | } |
| 4021 | return combined; |
| 4022 | } |
| 4023 | |
| 4024 | static SDValue simplifyDivRem(SDNode *N, SelectionDAG &DAG) { |
| 4025 | SDValue N0 = N->getOperand(0); |
| 4026 | SDValue N1 = N->getOperand(1); |
| 4027 | EVT VT = N->getValueType(0); |
| 4028 | SDLoc DL(N); |
| 4029 | |
| 4030 | unsigned Opc = N->getOpcode(); |
| 4031 | bool IsDiv = (ISD::SDIV == Opc) || (ISD::UDIV == Opc); |
| 4032 | ConstantSDNode *N1C = isConstOrConstSplat(N1); |
| 4033 | |
| 4034 | // X / undef -> undef |
| 4035 | // X % undef -> undef |
| 4036 | // X / 0 -> undef |
| 4037 | // X % 0 -> undef |
| 4038 | // NOTE: This includes vectors where any divisor element is zero/undef. |
| 4039 | if (DAG.isUndef(Opc, {N0, N1})) |
| 4040 | return DAG.getUNDEF(VT); |
| 4041 | |
| 4042 | // undef / X -> 0 |
| 4043 | // undef % X -> 0 |
| 4044 | if (N0.isUndef()) |
| 4045 | return DAG.getConstant(0, DL, VT); |
| 4046 | |
| 4047 | // 0 / X -> 0 |
| 4048 | // 0 % X -> 0 |
| 4049 | ConstantSDNode *N0C = isConstOrConstSplat(N0); |
| 4050 | if (N0C && N0C->isNullValue()) |
| 4051 | return N0; |
| 4052 | |
| 4053 | // X / X -> 1 |
| 4054 | // X % X -> 0 |
| 4055 | if (N0 == N1) |
| 4056 | return DAG.getConstant(IsDiv ? 1 : 0, DL, VT); |
| 4057 | |
| 4058 | // X / 1 -> X |
| 4059 | // X % 1 -> 0 |
| 4060 | // If this is a boolean op (single-bit element type), we can't have |
| 4061 | // division-by-zero or remainder-by-zero, so assume the divisor is 1. |
| 4062 | // TODO: Similarly, if we're zero-extending a boolean divisor, then assume |
| 4063 | // it's a 1. |
| 4064 | if ((N1C && N1C->isOne()) || (VT.getScalarType() == MVT::i1)) |
| 4065 | return IsDiv ? N0 : DAG.getConstant(0, DL, VT); |
| 4066 | |
| 4067 | return SDValue(); |
| 4068 | } |
| 4069 | |
| 4070 | SDValue DAGCombiner::visitSDIV(SDNode *N) { |
| 4071 | SDValue N0 = N->getOperand(0); |
| 4072 | SDValue N1 = N->getOperand(1); |
| 4073 | EVT VT = N->getValueType(0); |
| 4074 | EVT CCVT = getSetCCResultType(VT); |
| 4075 | |
| 4076 | // fold vector ops |
| 4077 | if (VT.isVector()) |
| 4078 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 4079 | return FoldedVOp; |
| 4080 | |
| 4081 | SDLoc DL(N); |
| 4082 | |
| 4083 | // fold (sdiv c1, c2) -> c1/c2 |
| 4084 | ConstantSDNode *N1C = isConstOrConstSplat(N1); |
| 4085 | if (SDValue C = DAG.FoldConstantArithmetic(ISD::SDIV, DL, VT, {N0, N1})) |
| 4086 | return C; |
| 4087 | |
| 4088 | // fold (sdiv X, -1) -> 0-X |
| 4089 | if (N1C && N1C->isAllOnesValue()) |
| 4090 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0); |
| 4091 | |
| 4092 | // fold (sdiv X, MIN_SIGNED) -> select(X == MIN_SIGNED, 1, 0) |
| 4093 | if (N1C && N1C->getAPIntValue().isMinSignedValue()) |
| 4094 | return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), |
| 4095 | DAG.getConstant(1, DL, VT), |
| 4096 | DAG.getConstant(0, DL, VT)); |
| 4097 | |
| 4098 | if (SDValue V = simplifyDivRem(N, DAG)) |
| 4099 | return V; |
| 4100 | |
| 4101 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 4102 | return NewSel; |
| 4103 | |
| 4104 | // If we know the sign bits of both operands are zero, strength reduce to a |
| 4105 | // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 |
| 4106 | if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) |
| 4107 | return DAG.getNode(ISD::UDIV, DL, N1.getValueType(), N0, N1); |
| 4108 | |
| 4109 | if (SDValue V = visitSDIVLike(N0, N1, N)) { |
| 4110 | // If the corresponding remainder node exists, update its users with |
| 4111 | // (Dividend - (Quotient * Divisor). |
| 4112 | if (SDNode *RemNode = DAG.getNodeIfExists(ISD::SREM, N->getVTList(), |
| 4113 | { N0, N1 })) { |
| 4114 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, V, N1); |
| 4115 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); |
| 4116 | AddToWorklist(Mul.getNode()); |
| 4117 | AddToWorklist(Sub.getNode()); |
| 4118 | CombineTo(RemNode, Sub); |
| 4119 | } |
| 4120 | return V; |
| 4121 | } |
| 4122 | |
| 4123 | // sdiv, srem -> sdivrem |
| 4124 | // If the divisor is constant, then return DIVREM only if isIntDivCheap() is |
| 4125 | // true. Otherwise, we break the simplification logic in visitREM(). |
| 4126 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); |
| 4127 | if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) |
| 4128 | if (SDValue DivRem = useDivRem(N)) |
| 4129 | return DivRem; |
| 4130 | |
| 4131 | return SDValue(); |
| 4132 | } |
| 4133 | |
| 4134 | SDValue DAGCombiner::visitSDIVLike(SDValue N0, SDValue N1, SDNode *N) { |
| 4135 | SDLoc DL(N); |
| 4136 | EVT VT = N->getValueType(0); |
| 4137 | EVT CCVT = getSetCCResultType(VT); |
| 4138 | unsigned BitWidth = VT.getScalarSizeInBits(); |
| 4139 | |
| 4140 | // Helper for determining whether a value is a power-2 constant scalar or a |
| 4141 | // vector of such elements. |
| 4142 | auto IsPowerOfTwo = [](ConstantSDNode *C) { |
| 4143 | if (C->isNullValue() || C->isOpaque()) |
| 4144 | return false; |
| 4145 | if (C->getAPIntValue().isPowerOf2()) |
| 4146 | return true; |
| 4147 | if ((-C->getAPIntValue()).isPowerOf2()) |
| 4148 | return true; |
| 4149 | return false; |
| 4150 | }; |
| 4151 | |
| 4152 | // fold (sdiv X, pow2) -> simple ops after legalize |
| 4153 | // FIXME: We check for the exact bit here because the generic lowering gives |
| 4154 | // better results in that case. The target-specific lowering should learn how |
| 4155 | // to handle exact sdivs efficiently. |
| 4156 | if (!N->getFlags().hasExact() && ISD::matchUnaryPredicate(N1, IsPowerOfTwo)) { |
| 4157 | // Target-specific implementation of sdiv x, pow2. |
| 4158 | if (SDValue Res = BuildSDIVPow2(N)) |
| 4159 | return Res; |
| 4160 | |
| 4161 | // Create constants that are functions of the shift amount value. |
| 4162 | EVT ShiftAmtTy = getShiftAmountTy(N0.getValueType()); |
| 4163 | SDValue Bits = DAG.getConstant(BitWidth, DL, ShiftAmtTy); |
| 4164 | SDValue C1 = DAG.getNode(ISD::CTTZ, DL, VT, N1); |
| 4165 | C1 = DAG.getZExtOrTrunc(C1, DL, ShiftAmtTy); |
| 4166 | SDValue Inexact = DAG.getNode(ISD::SUB, DL, ShiftAmtTy, Bits, C1); |
| 4167 | if (!isConstantOrConstantVector(Inexact)) |
| 4168 | return SDValue(); |
| 4169 | |
| 4170 | // Splat the sign bit into the register |
| 4171 | SDValue Sign = DAG.getNode(ISD::SRA, DL, VT, N0, |
| 4172 | DAG.getConstant(BitWidth - 1, DL, ShiftAmtTy)); |
| 4173 | AddToWorklist(Sign.getNode()); |
| 4174 | |
| 4175 | // Add (N0 < 0) ? abs2 - 1 : 0; |
| 4176 | SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, Sign, Inexact); |
| 4177 | AddToWorklist(Srl.getNode()); |
| 4178 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Srl); |
| 4179 | AddToWorklist(Add.getNode()); |
| 4180 | SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Add, C1); |
| 4181 | AddToWorklist(Sra.getNode()); |
| 4182 | |
| 4183 | // Special case: (sdiv X, 1) -> X |
| 4184 | // Special Case: (sdiv X, -1) -> 0-X |
| 4185 | SDValue One = DAG.getConstant(1, DL, VT); |
| 4186 | SDValue AllOnes = DAG.getAllOnesConstant(DL, VT); |
| 4187 | SDValue IsOne = DAG.getSetCC(DL, CCVT, N1, One, ISD::SETEQ); |
| 4188 | SDValue IsAllOnes = DAG.getSetCC(DL, CCVT, N1, AllOnes, ISD::SETEQ); |
| 4189 | SDValue IsOneOrAllOnes = DAG.getNode(ISD::OR, DL, CCVT, IsOne, IsAllOnes); |
| 4190 | Sra = DAG.getSelect(DL, VT, IsOneOrAllOnes, N0, Sra); |
| 4191 | |
| 4192 | // If dividing by a positive value, we're done. Otherwise, the result must |
| 4193 | // be negated. |
| 4194 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 4195 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, Zero, Sra); |
| 4196 | |
| 4197 | // FIXME: Use SELECT_CC once we improve SELECT_CC constant-folding. |
| 4198 | SDValue IsNeg = DAG.getSetCC(DL, CCVT, N1, Zero, ISD::SETLT); |
| 4199 | SDValue Res = DAG.getSelect(DL, VT, IsNeg, Sub, Sra); |
| 4200 | return Res; |
| 4201 | } |
| 4202 | |
| 4203 | // If integer divide is expensive and we satisfy the requirements, emit an |
| 4204 | // alternate sequence. Targets may check function attributes for size/speed |
| 4205 | // trade-offs. |
| 4206 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); |
| 4207 | if (isConstantOrConstantVector(N1) && |
| 4208 | !TLI.isIntDivCheap(N->getValueType(0), Attr)) |
| 4209 | if (SDValue Op = BuildSDIV(N)) |
| 4210 | return Op; |
| 4211 | |
| 4212 | return SDValue(); |
| 4213 | } |
| 4214 | |
| 4215 | SDValue DAGCombiner::visitUDIV(SDNode *N) { |
| 4216 | SDValue N0 = N->getOperand(0); |
| 4217 | SDValue N1 = N->getOperand(1); |
| 4218 | EVT VT = N->getValueType(0); |
| 4219 | EVT CCVT = getSetCCResultType(VT); |
| 4220 | |
| 4221 | // fold vector ops |
| 4222 | if (VT.isVector()) |
| 4223 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 4224 | return FoldedVOp; |
| 4225 | |
| 4226 | SDLoc DL(N); |
| 4227 | |
| 4228 | // fold (udiv c1, c2) -> c1/c2 |
| 4229 | ConstantSDNode *N1C = isConstOrConstSplat(N1); |
| 4230 | if (SDValue C = DAG.FoldConstantArithmetic(ISD::UDIV, DL, VT, {N0, N1})) |
| 4231 | return C; |
| 4232 | |
| 4233 | // fold (udiv X, -1) -> select(X == -1, 1, 0) |
| 4234 | if (N1C && N1C->getAPIntValue().isAllOnesValue()) |
| 4235 | return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), |
| 4236 | DAG.getConstant(1, DL, VT), |
| 4237 | DAG.getConstant(0, DL, VT)); |
| 4238 | |
| 4239 | if (SDValue V = simplifyDivRem(N, DAG)) |
| 4240 | return V; |
| 4241 | |
| 4242 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 4243 | return NewSel; |
| 4244 | |
| 4245 | if (SDValue V = visitUDIVLike(N0, N1, N)) { |
| 4246 | // If the corresponding remainder node exists, update its users with |
| 4247 | // (Dividend - (Quotient * Divisor). |
| 4248 | if (SDNode *RemNode = DAG.getNodeIfExists(ISD::UREM, N->getVTList(), |
| 4249 | { N0, N1 })) { |
| 4250 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, V, N1); |
| 4251 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); |
| 4252 | AddToWorklist(Mul.getNode()); |
| 4253 | AddToWorklist(Sub.getNode()); |
| 4254 | CombineTo(RemNode, Sub); |
| 4255 | } |
| 4256 | return V; |
| 4257 | } |
| 4258 | |
| 4259 | // sdiv, srem -> sdivrem |
| 4260 | // If the divisor is constant, then return DIVREM only if isIntDivCheap() is |
| 4261 | // true. Otherwise, we break the simplification logic in visitREM(). |
| 4262 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); |
| 4263 | if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) |
| 4264 | if (SDValue DivRem = useDivRem(N)) |
| 4265 | return DivRem; |
| 4266 | |
| 4267 | return SDValue(); |
| 4268 | } |
| 4269 | |
| 4270 | SDValue DAGCombiner::visitUDIVLike(SDValue N0, SDValue N1, SDNode *N) { |
| 4271 | SDLoc DL(N); |
| 4272 | EVT VT = N->getValueType(0); |
| 4273 | |
| 4274 | // fold (udiv x, (1 << c)) -> x >>u c |
| 4275 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && |
| 4276 | DAG.isKnownToBeAPowerOfTwo(N1)) { |
| 4277 | SDValue LogBase2 = BuildLogBase2(N1, DL); |
| 4278 | AddToWorklist(LogBase2.getNode()); |
| 4279 | |
| 4280 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); |
| 4281 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); |
| 4282 | AddToWorklist(Trunc.getNode()); |
| 4283 | return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); |
| 4284 | } |
| 4285 | |
| 4286 | // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 |
| 4287 | if (N1.getOpcode() == ISD::SHL) { |
| 4288 | SDValue N10 = N1.getOperand(0); |
| 4289 | if (isConstantOrConstantVector(N10, /*NoOpaques*/ true) && |
| 4290 | DAG.isKnownToBeAPowerOfTwo(N10)) { |
| 4291 | SDValue LogBase2 = BuildLogBase2(N10, DL); |
| 4292 | AddToWorklist(LogBase2.getNode()); |
| 4293 | |
| 4294 | EVT ADDVT = N1.getOperand(1).getValueType(); |
| 4295 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ADDVT); |
| 4296 | AddToWorklist(Trunc.getNode()); |
| 4297 | SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT, N1.getOperand(1), Trunc); |
| 4298 | AddToWorklist(Add.getNode()); |
| 4299 | return DAG.getNode(ISD::SRL, DL, VT, N0, Add); |
| 4300 | } |
| 4301 | } |
| 4302 | |
| 4303 | // fold (udiv x, c) -> alternate |
| 4304 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); |
| 4305 | if (isConstantOrConstantVector(N1) && |
| 4306 | !TLI.isIntDivCheap(N->getValueType(0), Attr)) |
| 4307 | if (SDValue Op = BuildUDIV(N)) |
| 4308 | return Op; |
| 4309 | |
| 4310 | return SDValue(); |
| 4311 | } |
| 4312 | |
| 4313 | // handles ISD::SREM and ISD::UREM |
| 4314 | SDValue DAGCombiner::visitREM(SDNode *N) { |
| 4315 | unsigned Opcode = N->getOpcode(); |
| 4316 | SDValue N0 = N->getOperand(0); |
| 4317 | SDValue N1 = N->getOperand(1); |
| 4318 | EVT VT = N->getValueType(0); |
| 4319 | EVT CCVT = getSetCCResultType(VT); |
| 4320 | |
| 4321 | bool isSigned = (Opcode == ISD::SREM); |
| 4322 | SDLoc DL(N); |
| 4323 | |
| 4324 | // fold (rem c1, c2) -> c1%c2 |
| 4325 | ConstantSDNode *N1C = isConstOrConstSplat(N1); |
| 4326 | if (SDValue C = DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1})) |
| 4327 | return C; |
| 4328 | |
| 4329 | // fold (urem X, -1) -> select(X == -1, 0, x) |
| 4330 | if (!isSigned && N1C && N1C->getAPIntValue().isAllOnesValue()) |
| 4331 | return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), |
| 4332 | DAG.getConstant(0, DL, VT), N0); |
| 4333 | |
| 4334 | if (SDValue V = simplifyDivRem(N, DAG)) |
| 4335 | return V; |
| 4336 | |
| 4337 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 4338 | return NewSel; |
| 4339 | |
| 4340 | if (isSigned) { |
| 4341 | // If we know the sign bits of both operands are zero, strength reduce to a |
| 4342 | // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 |
| 4343 | if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) |
| 4344 | return DAG.getNode(ISD::UREM, DL, VT, N0, N1); |
| 4345 | } else { |
| 4346 | if (DAG.isKnownToBeAPowerOfTwo(N1)) { |
| 4347 | // fold (urem x, pow2) -> (and x, pow2-1) |
| 4348 | SDValue NegOne = DAG.getAllOnesConstant(DL, VT); |
| 4349 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); |
| 4350 | AddToWorklist(Add.getNode()); |
| 4351 | return DAG.getNode(ISD::AND, DL, VT, N0, Add); |
| 4352 | } |
| 4353 | if (N1.getOpcode() == ISD::SHL && |
| 4354 | DAG.isKnownToBeAPowerOfTwo(N1.getOperand(0))) { |
| 4355 | // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) |
| 4356 | SDValue NegOne = DAG.getAllOnesConstant(DL, VT); |
| 4357 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); |
| 4358 | AddToWorklist(Add.getNode()); |
| 4359 | return DAG.getNode(ISD::AND, DL, VT, N0, Add); |
| 4360 | } |
| 4361 | } |
| 4362 | |
| 4363 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); |
| 4364 | |
| 4365 | // If X/C can be simplified by the division-by-constant logic, lower |
| 4366 | // X%C to the equivalent of X-X/C*C. |
| 4367 | // Reuse the SDIVLike/UDIVLike combines - to avoid mangling nodes, the |
| 4368 | // speculative DIV must not cause a DIVREM conversion. We guard against this |
| 4369 | // by skipping the simplification if isIntDivCheap(). When div is not cheap, |
| 4370 | // combine will not return a DIVREM. Regardless, checking cheapness here |
| 4371 | // makes sense since the simplification results in fatter code. |
| 4372 | if (DAG.isKnownNeverZero(N1) && !TLI.isIntDivCheap(VT, Attr)) { |
| 4373 | SDValue OptimizedDiv = |
| 4374 | isSigned ? visitSDIVLike(N0, N1, N) : visitUDIVLike(N0, N1, N); |
| 4375 | if (OptimizedDiv.getNode()) { |
| 4376 | // If the equivalent Div node also exists, update its users. |
| 4377 | unsigned DivOpcode = isSigned ? ISD::SDIV : ISD::UDIV; |
| 4378 | if (SDNode *DivNode = DAG.getNodeIfExists(DivOpcode, N->getVTList(), |
| 4379 | { N0, N1 })) |
| 4380 | CombineTo(DivNode, OptimizedDiv); |
| 4381 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, OptimizedDiv, N1); |
| 4382 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); |
| 4383 | AddToWorklist(OptimizedDiv.getNode()); |
| 4384 | AddToWorklist(Mul.getNode()); |
| 4385 | return Sub; |
| 4386 | } |
| 4387 | } |
| 4388 | |
| 4389 | // sdiv, srem -> sdivrem |
| 4390 | if (SDValue DivRem = useDivRem(N)) |
| 4391 | return DivRem.getValue(1); |
| 4392 | |
| 4393 | return SDValue(); |
| 4394 | } |
| 4395 | |
| 4396 | SDValue DAGCombiner::visitMULHS(SDNode *N) { |
| 4397 | SDValue N0 = N->getOperand(0); |
| 4398 | SDValue N1 = N->getOperand(1); |
| 4399 | EVT VT = N->getValueType(0); |
| 4400 | SDLoc DL(N); |
| 4401 | |
| 4402 | if (VT.isVector()) { |
| 4403 | // fold (mulhs x, 0) -> 0 |
| 4404 | // do not return N0/N1, because undef node may exist. |
| 4405 | if (ISD::isBuildVectorAllZeros(N0.getNode()) || |
| 4406 | ISD::isBuildVectorAllZeros(N1.getNode())) |
| 4407 | return DAG.getConstant(0, DL, VT); |
| 4408 | } |
| 4409 | |
| 4410 | // fold (mulhs x, 0) -> 0 |
| 4411 | if (isNullConstant(N1)) |
| 4412 | return N1; |
| 4413 | // fold (mulhs x, 1) -> (sra x, size(x)-1) |
| 4414 | if (isOneConstant(N1)) |
| 4415 | return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0, |
| 4416 | DAG.getConstant(N0.getScalarValueSizeInBits() - 1, DL, |
| 4417 | getShiftAmountTy(N0.getValueType()))); |
| 4418 | |
| 4419 | // fold (mulhs x, undef) -> 0 |
| 4420 | if (N0.isUndef() || N1.isUndef()) |
| 4421 | return DAG.getConstant(0, DL, VT); |
| 4422 | |
| 4423 | // If the type twice as wide is legal, transform the mulhs to a wider multiply |
| 4424 | // plus a shift. |
| 4425 | if (!TLI.isOperationLegalOrCustom(ISD::MULHS, VT) && VT.isSimple() && |
| 4426 | !VT.isVector()) { |
| 4427 | MVT Simple = VT.getSimpleVT(); |
| 4428 | unsigned SimpleSize = Simple.getSizeInBits(); |
| 4429 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); |
| 4430 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { |
| 4431 | N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); |
| 4432 | N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); |
| 4433 | N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); |
| 4434 | N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, |
| 4435 | DAG.getConstant(SimpleSize, DL, |
| 4436 | getShiftAmountTy(N1.getValueType()))); |
| 4437 | return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); |
| 4438 | } |
| 4439 | } |
| 4440 | |
| 4441 | return SDValue(); |
| 4442 | } |
| 4443 | |
| 4444 | SDValue DAGCombiner::visitMULHU(SDNode *N) { |
| 4445 | SDValue N0 = N->getOperand(0); |
| 4446 | SDValue N1 = N->getOperand(1); |
| 4447 | EVT VT = N->getValueType(0); |
| 4448 | SDLoc DL(N); |
| 4449 | |
| 4450 | if (VT.isVector()) { |
| 4451 | // fold (mulhu x, 0) -> 0 |
| 4452 | // do not return N0/N1, because undef node may exist. |
| 4453 | if (ISD::isBuildVectorAllZeros(N0.getNode()) || |
| 4454 | ISD::isBuildVectorAllZeros(N1.getNode())) |
| 4455 | return DAG.getConstant(0, DL, VT); |
| 4456 | } |
| 4457 | |
| 4458 | // fold (mulhu x, 0) -> 0 |
| 4459 | if (isNullConstant(N1)) |
| 4460 | return N1; |
| 4461 | // fold (mulhu x, 1) -> 0 |
| 4462 | if (isOneConstant(N1)) |
| 4463 | return DAG.getConstant(0, DL, N0.getValueType()); |
| 4464 | // fold (mulhu x, undef) -> 0 |
| 4465 | if (N0.isUndef() || N1.isUndef()) |
| 4466 | return DAG.getConstant(0, DL, VT); |
| 4467 | |
| 4468 | // fold (mulhu x, (1 << c)) -> x >> (bitwidth - c) |
| 4469 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && |
| 4470 | DAG.isKnownToBeAPowerOfTwo(N1) && hasOperation(ISD::SRL, VT)) { |
| 4471 | unsigned NumEltBits = VT.getScalarSizeInBits(); |
| 4472 | SDValue LogBase2 = BuildLogBase2(N1, DL); |
| 4473 | SDValue SRLAmt = DAG.getNode( |
| 4474 | ISD::SUB, DL, VT, DAG.getConstant(NumEltBits, DL, VT), LogBase2); |
| 4475 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); |
| 4476 | SDValue Trunc = DAG.getZExtOrTrunc(SRLAmt, DL, ShiftVT); |
| 4477 | return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); |
| 4478 | } |
| 4479 | |
| 4480 | // If the type twice as wide is legal, transform the mulhu to a wider multiply |
| 4481 | // plus a shift. |
| 4482 | if (!TLI.isOperationLegalOrCustom(ISD::MULHU, VT) && VT.isSimple() && |
| 4483 | !VT.isVector()) { |
| 4484 | MVT Simple = VT.getSimpleVT(); |
| 4485 | unsigned SimpleSize = Simple.getSizeInBits(); |
| 4486 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); |
| 4487 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { |
| 4488 | N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); |
| 4489 | N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); |
| 4490 | N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); |
| 4491 | N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, |
| 4492 | DAG.getConstant(SimpleSize, DL, |
| 4493 | getShiftAmountTy(N1.getValueType()))); |
| 4494 | return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); |
| 4495 | } |
| 4496 | } |
| 4497 | |
| 4498 | return SDValue(); |
| 4499 | } |
| 4500 | |
| 4501 | /// Perform optimizations common to nodes that compute two values. LoOp and HiOp |
| 4502 | /// give the opcodes for the two computations that are being performed. Return |
| 4503 | /// true if a simplification was made. |
| 4504 | SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, |
| 4505 | unsigned HiOp) { |
| 4506 | // If the high half is not needed, just compute the low half. |
| 4507 | bool HiExists = N->hasAnyUseOfValue(1); |
| 4508 | if (!HiExists && (!LegalOperations || |
| 4509 | TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) { |
| 4510 | SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); |
| 4511 | return CombineTo(N, Res, Res); |
| 4512 | } |
| 4513 | |
| 4514 | // If the low half is not needed, just compute the high half. |
| 4515 | bool LoExists = N->hasAnyUseOfValue(0); |
| 4516 | if (!LoExists && (!LegalOperations || |
| 4517 | TLI.isOperationLegalOrCustom(HiOp, N->getValueType(1)))) { |
| 4518 | SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); |
| 4519 | return CombineTo(N, Res, Res); |
| 4520 | } |
| 4521 | |
| 4522 | // If both halves are used, return as it is. |
| 4523 | if (LoExists && HiExists) |
| 4524 | return SDValue(); |
| 4525 | |
| 4526 | // If the two computed results can be simplified separately, separate them. |
| 4527 | if (LoExists) { |
| 4528 | SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); |
| 4529 | AddToWorklist(Lo.getNode()); |
| 4530 | SDValue LoOpt = combine(Lo.getNode()); |
| 4531 | if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() && |
| 4532 | (!LegalOperations || |
| 4533 | TLI.isOperationLegalOrCustom(LoOpt.getOpcode(), LoOpt.getValueType()))) |
| 4534 | return CombineTo(N, LoOpt, LoOpt); |
| 4535 | } |
| 4536 | |
| 4537 | if (HiExists) { |
| 4538 | SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); |
| 4539 | AddToWorklist(Hi.getNode()); |
| 4540 | SDValue HiOpt = combine(Hi.getNode()); |
| 4541 | if (HiOpt.getNode() && HiOpt != Hi && |
| 4542 | (!LegalOperations || |
| 4543 | TLI.isOperationLegalOrCustom(HiOpt.getOpcode(), HiOpt.getValueType()))) |
| 4544 | return CombineTo(N, HiOpt, HiOpt); |
| 4545 | } |
| 4546 | |
| 4547 | return SDValue(); |
| 4548 | } |
| 4549 | |
| 4550 | SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) { |
| 4551 | if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS)) |
| 4552 | return Res; |
| 4553 | |
| 4554 | EVT VT = N->getValueType(0); |
| 4555 | SDLoc DL(N); |
| 4556 | |
| 4557 | // If the type is twice as wide is legal, transform the mulhu to a wider |
| 4558 | // multiply plus a shift. |
| 4559 | if (VT.isSimple() && !VT.isVector()) { |
| 4560 | MVT Simple = VT.getSimpleVT(); |
| 4561 | unsigned SimpleSize = Simple.getSizeInBits(); |
| 4562 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); |
| 4563 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { |
| 4564 | SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0)); |
| 4565 | SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1)); |
| 4566 | Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); |
| 4567 | // Compute the high part as N1. |
| 4568 | Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, |
| 4569 | DAG.getConstant(SimpleSize, DL, |
| 4570 | getShiftAmountTy(Lo.getValueType()))); |
| 4571 | Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); |
| 4572 | // Compute the low part as N0. |
| 4573 | Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); |
| 4574 | return CombineTo(N, Lo, Hi); |
| 4575 | } |
| 4576 | } |
| 4577 | |
| 4578 | return SDValue(); |
| 4579 | } |
| 4580 | |
| 4581 | SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) { |
| 4582 | if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU)) |
| 4583 | return Res; |
| 4584 | |
| 4585 | EVT VT = N->getValueType(0); |
| 4586 | SDLoc DL(N); |
| 4587 | |
| 4588 | // (umul_lohi N0, 0) -> (0, 0) |
| 4589 | if (isNullConstant(N->getOperand(1))) { |
| 4590 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 4591 | return CombineTo(N, Zero, Zero); |
| 4592 | } |
| 4593 | |
| 4594 | // (umul_lohi N0, 1) -> (N0, 0) |
| 4595 | if (isOneConstant(N->getOperand(1))) { |
| 4596 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 4597 | return CombineTo(N, N->getOperand(0), Zero); |
| 4598 | } |
| 4599 | |
| 4600 | // If the type is twice as wide is legal, transform the mulhu to a wider |
| 4601 | // multiply plus a shift. |
| 4602 | if (VT.isSimple() && !VT.isVector()) { |
| 4603 | MVT Simple = VT.getSimpleVT(); |
| 4604 | unsigned SimpleSize = Simple.getSizeInBits(); |
| 4605 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); |
| 4606 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { |
| 4607 | SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0)); |
| 4608 | SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1)); |
| 4609 | Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); |
| 4610 | // Compute the high part as N1. |
| 4611 | Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, |
| 4612 | DAG.getConstant(SimpleSize, DL, |
| 4613 | getShiftAmountTy(Lo.getValueType()))); |
| 4614 | Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); |
| 4615 | // Compute the low part as N0. |
| 4616 | Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); |
| 4617 | return CombineTo(N, Lo, Hi); |
| 4618 | } |
| 4619 | } |
| 4620 | |
| 4621 | return SDValue(); |
| 4622 | } |
| 4623 | |
| 4624 | SDValue DAGCombiner::visitMULO(SDNode *N) { |
| 4625 | SDValue N0 = N->getOperand(0); |
| 4626 | SDValue N1 = N->getOperand(1); |
| 4627 | EVT VT = N0.getValueType(); |
| 4628 | bool IsSigned = (ISD::SMULO == N->getOpcode()); |
| 4629 | |
| 4630 | EVT CarryVT = N->getValueType(1); |
| 4631 | SDLoc DL(N); |
| 4632 | |
| 4633 | ConstantSDNode *N0C = isConstOrConstSplat(N0); |
| 4634 | ConstantSDNode *N1C = isConstOrConstSplat(N1); |
| 4635 | |
| 4636 | // fold operation with constant operands. |
| 4637 | // TODO: Move this to FoldConstantArithmetic when it supports nodes with |
| 4638 | // multiple results. |
| 4639 | if (N0C && N1C) { |
| 4640 | bool Overflow; |
| 4641 | APInt Result = |
| 4642 | IsSigned ? N0C->getAPIntValue().smul_ov(N1C->getAPIntValue(), Overflow) |
| 4643 | : N0C->getAPIntValue().umul_ov(N1C->getAPIntValue(), Overflow); |
| 4644 | return CombineTo(N, DAG.getConstant(Result, DL, VT), |
| 4645 | DAG.getBoolConstant(Overflow, DL, CarryVT, CarryVT)); |
| 4646 | } |
| 4647 | |
| 4648 | // canonicalize constant to RHS. |
| 4649 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && |
| 4650 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 4651 | return DAG.getNode(N->getOpcode(), DL, N->getVTList(), N1, N0); |
| 4652 | |
| 4653 | // fold (mulo x, 0) -> 0 + no carry out |
| 4654 | if (isNullOrNullSplat(N1)) |
| 4655 | return CombineTo(N, DAG.getConstant(0, DL, VT), |
| 4656 | DAG.getConstant(0, DL, CarryVT)); |
| 4657 | |
| 4658 | // (mulo x, 2) -> (addo x, x) |
| 4659 | if (N1C && N1C->getAPIntValue() == 2) |
| 4660 | return DAG.getNode(IsSigned ? ISD::SADDO : ISD::UADDO, DL, |
| 4661 | N->getVTList(), N0, N0); |
| 4662 | |
| 4663 | if (IsSigned) { |
| 4664 | // A 1 bit SMULO overflows if both inputs are 1. |
| 4665 | if (VT.getScalarSizeInBits() == 1) { |
| 4666 | SDValue And = DAG.getNode(ISD::AND, DL, VT, N0, N1); |
| 4667 | return CombineTo(N, And, |
| 4668 | DAG.getSetCC(DL, CarryVT, And, |
| 4669 | DAG.getConstant(0, DL, VT), ISD::SETNE)); |
| 4670 | } |
| 4671 | |
| 4672 | // Multiplying n * m significant bits yields a result of n + m significant |
| 4673 | // bits. If the total number of significant bits does not exceed the |
| 4674 | // result bit width (minus 1), there is no overflow. |
| 4675 | unsigned SignBits = DAG.ComputeNumSignBits(N0); |
| 4676 | if (SignBits > 1) |
| 4677 | SignBits += DAG.ComputeNumSignBits(N1); |
| 4678 | if (SignBits > VT.getScalarSizeInBits() + 1) |
| 4679 | return CombineTo(N, DAG.getNode(ISD::MUL, DL, VT, N0, N1), |
| 4680 | DAG.getConstant(0, DL, CarryVT)); |
| 4681 | } else { |
| 4682 | KnownBits N1Known = DAG.computeKnownBits(N1); |
| 4683 | KnownBits N0Known = DAG.computeKnownBits(N0); |
| 4684 | bool Overflow; |
| 4685 | (void)N0Known.getMaxValue().umul_ov(N1Known.getMaxValue(), Overflow); |
| 4686 | if (!Overflow) |
| 4687 | return CombineTo(N, DAG.getNode(ISD::MUL, DL, VT, N0, N1), |
| 4688 | DAG.getConstant(0, DL, CarryVT)); |
| 4689 | } |
| 4690 | |
| 4691 | return SDValue(); |
| 4692 | } |
| 4693 | |
| 4694 | SDValue DAGCombiner::visitIMINMAX(SDNode *N) { |
| 4695 | SDValue N0 = N->getOperand(0); |
| 4696 | SDValue N1 = N->getOperand(1); |
| 4697 | EVT VT = N0.getValueType(); |
| 4698 | unsigned Opcode = N->getOpcode(); |
| 4699 | |
| 4700 | // fold vector ops |
| 4701 | if (VT.isVector()) |
| 4702 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 4703 | return FoldedVOp; |
| 4704 | |
| 4705 | // fold operation with constant operands. |
| 4706 | if (SDValue C = DAG.FoldConstantArithmetic(Opcode, SDLoc(N), VT, {N0, N1})) |
| 4707 | return C; |
| 4708 | |
| 4709 | // canonicalize constant to RHS |
| 4710 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && |
| 4711 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 4712 | return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0); |
| 4713 | |
| 4714 | // Is sign bits are zero, flip between UMIN/UMAX and SMIN/SMAX. |
| 4715 | // Only do this if the current op isn't legal and the flipped is. |
| 4716 | if (!TLI.isOperationLegal(Opcode, VT) && |
| 4717 | (N0.isUndef() || DAG.SignBitIsZero(N0)) && |
| 4718 | (N1.isUndef() || DAG.SignBitIsZero(N1))) { |
| 4719 | unsigned AltOpcode; |
| 4720 | switch (Opcode) { |
| 4721 | case ISD::SMIN: AltOpcode = ISD::UMIN; break; |
| 4722 | case ISD::SMAX: AltOpcode = ISD::UMAX; break; |
| 4723 | case ISD::UMIN: AltOpcode = ISD::SMIN; break; |
| 4724 | case ISD::UMAX: AltOpcode = ISD::SMAX; break; |
| 4725 | default: llvm_unreachable("Unknown MINMAX opcode")::llvm::llvm_unreachable_internal("Unknown MINMAX opcode", "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4725); |
| 4726 | } |
| 4727 | if (TLI.isOperationLegal(AltOpcode, VT)) |
| 4728 | return DAG.getNode(AltOpcode, SDLoc(N), VT, N0, N1); |
| 4729 | } |
| 4730 | |
| 4731 | // Simplify the operands using demanded-bits information. |
| 4732 | if (SimplifyDemandedBits(SDValue(N, 0))) |
| 4733 | return SDValue(N, 0); |
| 4734 | |
| 4735 | return SDValue(); |
| 4736 | } |
| 4737 | |
| 4738 | /// If this is a bitwise logic instruction and both operands have the same |
| 4739 | /// opcode, try to sink the other opcode after the logic instruction. |
| 4740 | SDValue DAGCombiner::hoistLogicOpWithSameOpcodeHands(SDNode *N) { |
| 4741 | SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); |
| 4742 | EVT VT = N0.getValueType(); |
| 4743 | unsigned LogicOpcode = N->getOpcode(); |
| 4744 | unsigned HandOpcode = N0.getOpcode(); |
| 4745 | assert((LogicOpcode == ISD::AND || LogicOpcode == ISD::OR ||(((LogicOpcode == ISD::AND || LogicOpcode == ISD::OR || LogicOpcode == ISD::XOR) && "Expected logic opcode") ? static_cast <void> (0) : __assert_fail ("(LogicOpcode == ISD::AND || LogicOpcode == ISD::OR || LogicOpcode == ISD::XOR) && \"Expected logic opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4746, __PRETTY_FUNCTION__)) |
| 4746 | LogicOpcode == ISD::XOR) && "Expected logic opcode")(((LogicOpcode == ISD::AND || LogicOpcode == ISD::OR || LogicOpcode == ISD::XOR) && "Expected logic opcode") ? static_cast <void> (0) : __assert_fail ("(LogicOpcode == ISD::AND || LogicOpcode == ISD::OR || LogicOpcode == ISD::XOR) && \"Expected logic opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4746, __PRETTY_FUNCTION__)); |
| 4747 | assert(HandOpcode == N1.getOpcode() && "Bad input!")((HandOpcode == N1.getOpcode() && "Bad input!") ? static_cast <void> (0) : __assert_fail ("HandOpcode == N1.getOpcode() && \"Bad input!\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4747, __PRETTY_FUNCTION__)); |
| 4748 | |
| 4749 | // Bail early if none of these transforms apply. |
| 4750 | if (N0.getNumOperands() == 0) |
| 4751 | return SDValue(); |
| 4752 | |
| 4753 | // FIXME: We should check number of uses of the operands to not increase |
| 4754 | // the instruction count for all transforms. |
| 4755 | |
| 4756 | // Handle size-changing casts. |
| 4757 | SDValue X = N0.getOperand(0); |
| 4758 | SDValue Y = N1.getOperand(0); |
| 4759 | EVT XVT = X.getValueType(); |
| 4760 | SDLoc DL(N); |
| 4761 | if (HandOpcode == ISD::ANY_EXTEND || HandOpcode == ISD::ZERO_EXTEND || |
| 4762 | HandOpcode == ISD::SIGN_EXTEND) { |
| 4763 | // If both operands have other uses, this transform would create extra |
| 4764 | // instructions without eliminating anything. |
| 4765 | if (!N0.hasOneUse() && !N1.hasOneUse()) |
| 4766 | return SDValue(); |
| 4767 | // We need matching integer source types. |
| 4768 | if (XVT != Y.getValueType()) |
| 4769 | return SDValue(); |
| 4770 | // Don't create an illegal op during or after legalization. Don't ever |
| 4771 | // create an unsupported vector op. |
| 4772 | if ((VT.isVector() || LegalOperations) && |
| 4773 | !TLI.isOperationLegalOrCustom(LogicOpcode, XVT)) |
| 4774 | return SDValue(); |
| 4775 | // Avoid infinite looping with PromoteIntBinOp. |
| 4776 | // TODO: Should we apply desirable/legal constraints to all opcodes? |
| 4777 | if (HandOpcode == ISD::ANY_EXTEND && LegalTypes && |
| 4778 | !TLI.isTypeDesirableForOp(LogicOpcode, XVT)) |
| 4779 | return SDValue(); |
| 4780 | // logic_op (hand_op X), (hand_op Y) --> hand_op (logic_op X, Y) |
| 4781 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); |
| 4782 | return DAG.getNode(HandOpcode, DL, VT, Logic); |
| 4783 | } |
| 4784 | |
| 4785 | // logic_op (truncate x), (truncate y) --> truncate (logic_op x, y) |
| 4786 | if (HandOpcode == ISD::TRUNCATE) { |
| 4787 | // If both operands have other uses, this transform would create extra |
| 4788 | // instructions without eliminating anything. |
| 4789 | if (!N0.hasOneUse() && !N1.hasOneUse()) |
| 4790 | return SDValue(); |
| 4791 | // We need matching source types. |
| 4792 | if (XVT != Y.getValueType()) |
| 4793 | return SDValue(); |
| 4794 | // Don't create an illegal op during or after legalization. |
| 4795 | if (LegalOperations && !TLI.isOperationLegal(LogicOpcode, XVT)) |
| 4796 | return SDValue(); |
| 4797 | // Be extra careful sinking truncate. If it's free, there's no benefit in |
| 4798 | // widening a binop. Also, don't create a logic op on an illegal type. |
| 4799 | if (TLI.isZExtFree(VT, XVT) && TLI.isTruncateFree(XVT, VT)) |
| 4800 | return SDValue(); |
| 4801 | if (!TLI.isTypeLegal(XVT)) |
| 4802 | return SDValue(); |
| 4803 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); |
| 4804 | return DAG.getNode(HandOpcode, DL, VT, Logic); |
| 4805 | } |
| 4806 | |
| 4807 | // For binops SHL/SRL/SRA/AND: |
| 4808 | // logic_op (OP x, z), (OP y, z) --> OP (logic_op x, y), z |
| 4809 | if ((HandOpcode == ISD::SHL || HandOpcode == ISD::SRL || |
| 4810 | HandOpcode == ISD::SRA || HandOpcode == ISD::AND) && |
| 4811 | N0.getOperand(1) == N1.getOperand(1)) { |
| 4812 | // If either operand has other uses, this transform is not an improvement. |
| 4813 | if (!N0.hasOneUse() || !N1.hasOneUse()) |
| 4814 | return SDValue(); |
| 4815 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); |
| 4816 | return DAG.getNode(HandOpcode, DL, VT, Logic, N0.getOperand(1)); |
| 4817 | } |
| 4818 | |
| 4819 | // Unary ops: logic_op (bswap x), (bswap y) --> bswap (logic_op x, y) |
| 4820 | if (HandOpcode == ISD::BSWAP) { |
| 4821 | // If either operand has other uses, this transform is not an improvement. |
| 4822 | if (!N0.hasOneUse() || !N1.hasOneUse()) |
| 4823 | return SDValue(); |
| 4824 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); |
| 4825 | return DAG.getNode(HandOpcode, DL, VT, Logic); |
| 4826 | } |
| 4827 | |
| 4828 | // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) |
| 4829 | // Only perform this optimization up until type legalization, before |
| 4830 | // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by |
| 4831 | // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and |
| 4832 | // we don't want to undo this promotion. |
| 4833 | // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper |
| 4834 | // on scalars. |
| 4835 | if ((HandOpcode == ISD::BITCAST || HandOpcode == ISD::SCALAR_TO_VECTOR) && |
| 4836 | Level <= AfterLegalizeTypes) { |
| 4837 | // Input types must be integer and the same. |
| 4838 | if (XVT.isInteger() && XVT == Y.getValueType() && |
| 4839 | !(VT.isVector() && TLI.isTypeLegal(VT) && |
| 4840 | !XVT.isVector() && !TLI.isTypeLegal(XVT))) { |
| 4841 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); |
| 4842 | return DAG.getNode(HandOpcode, DL, VT, Logic); |
| 4843 | } |
| 4844 | } |
| 4845 | |
| 4846 | // Xor/and/or are indifferent to the swizzle operation (shuffle of one value). |
| 4847 | // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B)) |
| 4848 | // If both shuffles use the same mask, and both shuffle within a single |
| 4849 | // vector, then it is worthwhile to move the swizzle after the operation. |
| 4850 | // The type-legalizer generates this pattern when loading illegal |
| 4851 | // vector types from memory. In many cases this allows additional shuffle |
| 4852 | // optimizations. |
| 4853 | // There are other cases where moving the shuffle after the xor/and/or |
| 4854 | // is profitable even if shuffles don't perform a swizzle. |
| 4855 | // If both shuffles use the same mask, and both shuffles have the same first |
| 4856 | // or second operand, then it might still be profitable to move the shuffle |
| 4857 | // after the xor/and/or operation. |
| 4858 | if (HandOpcode == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) { |
| 4859 | auto *SVN0 = cast<ShuffleVectorSDNode>(N0); |
| 4860 | auto *SVN1 = cast<ShuffleVectorSDNode>(N1); |
| 4861 | assert(X.getValueType() == Y.getValueType() &&((X.getValueType() == Y.getValueType() && "Inputs to shuffles are not the same type" ) ? static_cast<void> (0) : __assert_fail ("X.getValueType() == Y.getValueType() && \"Inputs to shuffles are not the same type\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4862, __PRETTY_FUNCTION__)) |
| 4862 | "Inputs to shuffles are not the same type")((X.getValueType() == Y.getValueType() && "Inputs to shuffles are not the same type" ) ? static_cast<void> (0) : __assert_fail ("X.getValueType() == Y.getValueType() && \"Inputs to shuffles are not the same type\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4862, __PRETTY_FUNCTION__)); |
| 4863 | |
| 4864 | // Check that both shuffles use the same mask. The masks are known to be of |
| 4865 | // the same length because the result vector type is the same. |
| 4866 | // Check also that shuffles have only one use to avoid introducing extra |
| 4867 | // instructions. |
| 4868 | if (!SVN0->hasOneUse() || !SVN1->hasOneUse() || |
| 4869 | !SVN0->getMask().equals(SVN1->getMask())) |
| 4870 | return SDValue(); |
| 4871 | |
| 4872 | // Don't try to fold this node if it requires introducing a |
| 4873 | // build vector of all zeros that might be illegal at this stage. |
| 4874 | SDValue ShOp = N0.getOperand(1); |
| 4875 | if (LogicOpcode == ISD::XOR && !ShOp.isUndef()) |
| 4876 | ShOp = tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); |
| 4877 | |
| 4878 | // (logic_op (shuf (A, C), shuf (B, C))) --> shuf (logic_op (A, B), C) |
| 4879 | if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) { |
| 4880 | SDValue Logic = DAG.getNode(LogicOpcode, DL, VT, |
| 4881 | N0.getOperand(0), N1.getOperand(0)); |
| 4882 | return DAG.getVectorShuffle(VT, DL, Logic, ShOp, SVN0->getMask()); |
| 4883 | } |
| 4884 | |
| 4885 | // Don't try to fold this node if it requires introducing a |
| 4886 | // build vector of all zeros that might be illegal at this stage. |
| 4887 | ShOp = N0.getOperand(0); |
| 4888 | if (LogicOpcode == ISD::XOR && !ShOp.isUndef()) |
| 4889 | ShOp = tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); |
| 4890 | |
| 4891 | // (logic_op (shuf (C, A), shuf (C, B))) --> shuf (C, logic_op (A, B)) |
| 4892 | if (N0.getOperand(0) == N1.getOperand(0) && ShOp.getNode()) { |
| 4893 | SDValue Logic = DAG.getNode(LogicOpcode, DL, VT, N0.getOperand(1), |
| 4894 | N1.getOperand(1)); |
| 4895 | return DAG.getVectorShuffle(VT, DL, ShOp, Logic, SVN0->getMask()); |
| 4896 | } |
| 4897 | } |
| 4898 | |
| 4899 | return SDValue(); |
| 4900 | } |
| 4901 | |
| 4902 | /// Try to make (and/or setcc (LL, LR), setcc (RL, RR)) more efficient. |
| 4903 | SDValue DAGCombiner::foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, |
| 4904 | const SDLoc &DL) { |
| 4905 | SDValue LL, LR, RL, RR, N0CC, N1CC; |
| 4906 | if (!isSetCCEquivalent(N0, LL, LR, N0CC) || |
| 4907 | !isSetCCEquivalent(N1, RL, RR, N1CC)) |
| 4908 | return SDValue(); |
| 4909 | |
| 4910 | assert(N0.getValueType() == N1.getValueType() &&((N0.getValueType() == N1.getValueType() && "Unexpected operand types for bitwise logic op" ) ? static_cast<void> (0) : __assert_fail ("N0.getValueType() == N1.getValueType() && \"Unexpected operand types for bitwise logic op\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4911, __PRETTY_FUNCTION__)) |
| 4911 | "Unexpected operand types for bitwise logic op")((N0.getValueType() == N1.getValueType() && "Unexpected operand types for bitwise logic op" ) ? static_cast<void> (0) : __assert_fail ("N0.getValueType() == N1.getValueType() && \"Unexpected operand types for bitwise logic op\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4911, __PRETTY_FUNCTION__)); |
| 4912 | assert(LL.getValueType() == LR.getValueType() &&((LL.getValueType() == LR.getValueType() && RL.getValueType () == RR.getValueType() && "Unexpected operand types for setcc" ) ? static_cast<void> (0) : __assert_fail ("LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && \"Unexpected operand types for setcc\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4914, __PRETTY_FUNCTION__)) |
| 4913 | RL.getValueType() == RR.getValueType() &&((LL.getValueType() == LR.getValueType() && RL.getValueType () == RR.getValueType() && "Unexpected operand types for setcc" ) ? static_cast<void> (0) : __assert_fail ("LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && \"Unexpected operand types for setcc\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4914, __PRETTY_FUNCTION__)) |
| 4914 | "Unexpected operand types for setcc")((LL.getValueType() == LR.getValueType() && RL.getValueType () == RR.getValueType() && "Unexpected operand types for setcc" ) ? static_cast<void> (0) : __assert_fail ("LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && \"Unexpected operand types for setcc\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4914, __PRETTY_FUNCTION__)); |
| 4915 | |
| 4916 | // If we're here post-legalization or the logic op type is not i1, the logic |
| 4917 | // op type must match a setcc result type. Also, all folds require new |
| 4918 | // operations on the left and right operands, so those types must match. |
| 4919 | EVT VT = N0.getValueType(); |
| 4920 | EVT OpVT = LL.getValueType(); |
| 4921 | if (LegalOperations || VT.getScalarType() != MVT::i1) |
| 4922 | if (VT != getSetCCResultType(OpVT)) |
| 4923 | return SDValue(); |
| 4924 | if (OpVT != RL.getValueType()) |
| 4925 | return SDValue(); |
| 4926 | |
| 4927 | ISD::CondCode CC0 = cast<CondCodeSDNode>(N0CC)->get(); |
| 4928 | ISD::CondCode CC1 = cast<CondCodeSDNode>(N1CC)->get(); |
| 4929 | bool IsInteger = OpVT.isInteger(); |
| 4930 | if (LR == RR && CC0 == CC1 && IsInteger) { |
| 4931 | bool IsZero = isNullOrNullSplat(LR); |
| 4932 | bool IsNeg1 = isAllOnesOrAllOnesSplat(LR); |
| 4933 | |
| 4934 | // All bits clear? |
| 4935 | bool AndEqZero = IsAnd && CC1 == ISD::SETEQ && IsZero; |
| 4936 | // All sign bits clear? |
| 4937 | bool AndGtNeg1 = IsAnd && CC1 == ISD::SETGT && IsNeg1; |
| 4938 | // Any bits set? |
| 4939 | bool OrNeZero = !IsAnd && CC1 == ISD::SETNE && IsZero; |
| 4940 | // Any sign bits set? |
| 4941 | bool OrLtZero = !IsAnd && CC1 == ISD::SETLT && IsZero; |
| 4942 | |
| 4943 | // (and (seteq X, 0), (seteq Y, 0)) --> (seteq (or X, Y), 0) |
| 4944 | // (and (setgt X, -1), (setgt Y, -1)) --> (setgt (or X, Y), -1) |
| 4945 | // (or (setne X, 0), (setne Y, 0)) --> (setne (or X, Y), 0) |
| 4946 | // (or (setlt X, 0), (setlt Y, 0)) --> (setlt (or X, Y), 0) |
| 4947 | if (AndEqZero || AndGtNeg1 || OrNeZero || OrLtZero) { |
| 4948 | SDValue Or = DAG.getNode(ISD::OR, SDLoc(N0), OpVT, LL, RL); |
| 4949 | AddToWorklist(Or.getNode()); |
| 4950 | return DAG.getSetCC(DL, VT, Or, LR, CC1); |
| 4951 | } |
| 4952 | |
| 4953 | // All bits set? |
| 4954 | bool AndEqNeg1 = IsAnd && CC1 == ISD::SETEQ && IsNeg1; |
| 4955 | // All sign bits set? |
| 4956 | bool AndLtZero = IsAnd && CC1 == ISD::SETLT && IsZero; |
| 4957 | // Any bits clear? |
| 4958 | bool OrNeNeg1 = !IsAnd && CC1 == ISD::SETNE && IsNeg1; |
| 4959 | // Any sign bits clear? |
| 4960 | bool OrGtNeg1 = !IsAnd && CC1 == ISD::SETGT && IsNeg1; |
| 4961 | |
| 4962 | // (and (seteq X, -1), (seteq Y, -1)) --> (seteq (and X, Y), -1) |
| 4963 | // (and (setlt X, 0), (setlt Y, 0)) --> (setlt (and X, Y), 0) |
| 4964 | // (or (setne X, -1), (setne Y, -1)) --> (setne (and X, Y), -1) |
| 4965 | // (or (setgt X, -1), (setgt Y -1)) --> (setgt (and X, Y), -1) |
| 4966 | if (AndEqNeg1 || AndLtZero || OrNeNeg1 || OrGtNeg1) { |
| 4967 | SDValue And = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, LL, RL); |
| 4968 | AddToWorklist(And.getNode()); |
| 4969 | return DAG.getSetCC(DL, VT, And, LR, CC1); |
| 4970 | } |
| 4971 | } |
| 4972 | |
| 4973 | // TODO: What is the 'or' equivalent of this fold? |
| 4974 | // (and (setne X, 0), (setne X, -1)) --> (setuge (add X, 1), 2) |
| 4975 | if (IsAnd && LL == RL && CC0 == CC1 && OpVT.getScalarSizeInBits() > 1 && |
| 4976 | IsInteger && CC0 == ISD::SETNE && |
| 4977 | ((isNullConstant(LR) && isAllOnesConstant(RR)) || |
| 4978 | (isAllOnesConstant(LR) && isNullConstant(RR)))) { |
| 4979 | SDValue One = DAG.getConstant(1, DL, OpVT); |
| 4980 | SDValue Two = DAG.getConstant(2, DL, OpVT); |
| 4981 | SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), OpVT, LL, One); |
| 4982 | AddToWorklist(Add.getNode()); |
| 4983 | return DAG.getSetCC(DL, VT, Add, Two, ISD::SETUGE); |
| 4984 | } |
| 4985 | |
| 4986 | // Try more general transforms if the predicates match and the only user of |
| 4987 | // the compares is the 'and' or 'or'. |
| 4988 | if (IsInteger && TLI.convertSetCCLogicToBitwiseLogic(OpVT) && CC0 == CC1 && |
| 4989 | N0.hasOneUse() && N1.hasOneUse()) { |
| 4990 | // and (seteq A, B), (seteq C, D) --> seteq (or (xor A, B), (xor C, D)), 0 |
| 4991 | // or (setne A, B), (setne C, D) --> setne (or (xor A, B), (xor C, D)), 0 |
| 4992 | if ((IsAnd && CC1 == ISD::SETEQ) || (!IsAnd && CC1 == ISD::SETNE)) { |
| 4993 | SDValue XorL = DAG.getNode(ISD::XOR, SDLoc(N0), OpVT, LL, LR); |
| 4994 | SDValue XorR = DAG.getNode(ISD::XOR, SDLoc(N1), OpVT, RL, RR); |
| 4995 | SDValue Or = DAG.getNode(ISD::OR, DL, OpVT, XorL, XorR); |
| 4996 | SDValue Zero = DAG.getConstant(0, DL, OpVT); |
| 4997 | return DAG.getSetCC(DL, VT, Or, Zero, CC1); |
| 4998 | } |
| 4999 | |
| 5000 | // Turn compare of constants whose difference is 1 bit into add+and+setcc. |
| 5001 | // TODO - support non-uniform vector amounts. |
| 5002 | if ((IsAnd && CC1 == ISD::SETNE) || (!IsAnd && CC1 == ISD::SETEQ)) { |
| 5003 | // Match a shared variable operand and 2 non-opaque constant operands. |
| 5004 | ConstantSDNode *C0 = isConstOrConstSplat(LR); |
| 5005 | ConstantSDNode *C1 = isConstOrConstSplat(RR); |
| 5006 | if (LL == RL && C0 && C1 && !C0->isOpaque() && !C1->isOpaque()) { |
| 5007 | const APInt &CMax = |
| 5008 | APIntOps::umax(C0->getAPIntValue(), C1->getAPIntValue()); |
| 5009 | const APInt &CMin = |
| 5010 | APIntOps::umin(C0->getAPIntValue(), C1->getAPIntValue()); |
| 5011 | // The difference of the constants must be a single bit. |
| 5012 | if ((CMax - CMin).isPowerOf2()) { |
| 5013 | // and/or (setcc X, CMax, ne), (setcc X, CMin, ne/eq) --> |
| 5014 | // setcc ((sub X, CMin), ~(CMax - CMin)), 0, ne/eq |
| 5015 | SDValue Max = DAG.getNode(ISD::UMAX, DL, OpVT, LR, RR); |
| 5016 | SDValue Min = DAG.getNode(ISD::UMIN, DL, OpVT, LR, RR); |
| 5017 | SDValue Offset = DAG.getNode(ISD::SUB, DL, OpVT, LL, Min); |
| 5018 | SDValue Diff = DAG.getNode(ISD::SUB, DL, OpVT, Max, Min); |
| 5019 | SDValue Mask = DAG.getNOT(DL, Diff, OpVT); |
| 5020 | SDValue And = DAG.getNode(ISD::AND, DL, OpVT, Offset, Mask); |
| 5021 | SDValue Zero = DAG.getConstant(0, DL, OpVT); |
| 5022 | return DAG.getSetCC(DL, VT, And, Zero, CC0); |
| 5023 | } |
| 5024 | } |
| 5025 | } |
| 5026 | } |
| 5027 | |
| 5028 | // Canonicalize equivalent operands to LL == RL. |
| 5029 | if (LL == RR && LR == RL) { |
| 5030 | CC1 = ISD::getSetCCSwappedOperands(CC1); |
| 5031 | std::swap(RL, RR); |
| 5032 | } |
| 5033 | |
| 5034 | // (and (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) |
| 5035 | // (or (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) |
| 5036 | if (LL == RL && LR == RR) { |
| 5037 | ISD::CondCode NewCC = IsAnd ? ISD::getSetCCAndOperation(CC0, CC1, OpVT) |
| 5038 | : ISD::getSetCCOrOperation(CC0, CC1, OpVT); |
| 5039 | if (NewCC != ISD::SETCC_INVALID && |
| 5040 | (!LegalOperations || |
| 5041 | (TLI.isCondCodeLegal(NewCC, LL.getSimpleValueType()) && |
| 5042 | TLI.isOperationLegal(ISD::SETCC, OpVT)))) |
| 5043 | return DAG.getSetCC(DL, VT, LL, LR, NewCC); |
| 5044 | } |
| 5045 | |
| 5046 | return SDValue(); |
| 5047 | } |
| 5048 | |
| 5049 | /// This contains all DAGCombine rules which reduce two values combined by |
| 5050 | /// an And operation to a single value. This makes them reusable in the context |
| 5051 | /// of visitSELECT(). Rules involving constants are not included as |
| 5052 | /// visitSELECT() already handles those cases. |
| 5053 | SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1, SDNode *N) { |
| 5054 | EVT VT = N1.getValueType(); |
| 5055 | SDLoc DL(N); |
| 5056 | |
| 5057 | // fold (and x, undef) -> 0 |
| 5058 | if (N0.isUndef() || N1.isUndef()) |
| 5059 | return DAG.getConstant(0, DL, VT); |
| 5060 | |
| 5061 | if (SDValue V = foldLogicOfSetCCs(true, N0, N1, DL)) |
| 5062 | return V; |
| 5063 | |
| 5064 | if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL && |
| 5065 | VT.getSizeInBits() <= 64) { |
| 5066 | if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { |
| 5067 | if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) { |
| 5068 | // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal |
| 5069 | // immediate for an add, but it is legal if its top c2 bits are set, |
| 5070 | // transform the ADD so the immediate doesn't need to be materialized |
| 5071 | // in a register. |
| 5072 | APInt ADDC = ADDI->getAPIntValue(); |
| 5073 | APInt SRLC = SRLI->getAPIntValue(); |
| 5074 | if (ADDC.getMinSignedBits() <= 64 && |
| 5075 | SRLC.ult(VT.getSizeInBits()) && |
| 5076 | !TLI.isLegalAddImmediate(ADDC.getSExtValue())) { |
| 5077 | APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), |
| 5078 | SRLC.getZExtValue()); |
| 5079 | if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) { |
| 5080 | ADDC |= Mask; |
| 5081 | if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) { |
| 5082 | SDLoc DL0(N0); |
| 5083 | SDValue NewAdd = |
| 5084 | DAG.getNode(ISD::ADD, DL0, VT, |
| 5085 | N0.getOperand(0), DAG.getConstant(ADDC, DL, VT)); |
| 5086 | CombineTo(N0.getNode(), NewAdd); |
| 5087 | // Return N so it doesn't get rechecked! |
| 5088 | return SDValue(N, 0); |
| 5089 | } |
| 5090 | } |
| 5091 | } |
| 5092 | } |
| 5093 | } |
| 5094 | } |
| 5095 | |
| 5096 | // Reduce bit extract of low half of an integer to the narrower type. |
| 5097 | // (and (srl i64:x, K), KMask) -> |
| 5098 | // (i64 zero_extend (and (srl (i32 (trunc i64:x)), K)), KMask) |
| 5099 | if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { |
| 5100 | if (ConstantSDNode *CAnd = dyn_cast<ConstantSDNode>(N1)) { |
| 5101 | if (ConstantSDNode *CShift = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { |
| 5102 | unsigned Size = VT.getSizeInBits(); |
| 5103 | const APInt &AndMask = CAnd->getAPIntValue(); |
| 5104 | unsigned ShiftBits = CShift->getZExtValue(); |
| 5105 | |
| 5106 | // Bail out, this node will probably disappear anyway. |
| 5107 | if (ShiftBits == 0) |
| 5108 | return SDValue(); |
| 5109 | |
| 5110 | unsigned MaskBits = AndMask.countTrailingOnes(); |
| 5111 | EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), Size / 2); |
| 5112 | |
| 5113 | if (AndMask.isMask() && |
| 5114 | // Required bits must not span the two halves of the integer and |
| 5115 | // must fit in the half size type. |
| 5116 | (ShiftBits + MaskBits <= Size / 2) && |
| 5117 | TLI.isNarrowingProfitable(VT, HalfVT) && |
| 5118 | TLI.isTypeDesirableForOp(ISD::AND, HalfVT) && |
| 5119 | TLI.isTypeDesirableForOp(ISD::SRL, HalfVT) && |
| 5120 | TLI.isTruncateFree(VT, HalfVT) && |
| 5121 | TLI.isZExtFree(HalfVT, VT)) { |
| 5122 | // The isNarrowingProfitable is to avoid regressions on PPC and |
| 5123 | // AArch64 which match a few 64-bit bit insert / bit extract patterns |
| 5124 | // on downstream users of this. Those patterns could probably be |
| 5125 | // extended to handle extensions mixed in. |
| 5126 | |
| 5127 | SDValue SL(N0); |
| 5128 | assert(MaskBits <= Size)((MaskBits <= Size) ? static_cast<void> (0) : __assert_fail ("MaskBits <= Size", "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5128, __PRETTY_FUNCTION__)); |
| 5129 | |
| 5130 | // Extracting the highest bit of the low half. |
| 5131 | EVT ShiftVT = TLI.getShiftAmountTy(HalfVT, DAG.getDataLayout()); |
| 5132 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, HalfVT, |
| 5133 | N0.getOperand(0)); |
| 5134 | |
| 5135 | SDValue NewMask = DAG.getConstant(AndMask.trunc(Size / 2), SL, HalfVT); |
| 5136 | SDValue ShiftK = DAG.getConstant(ShiftBits, SL, ShiftVT); |
| 5137 | SDValue Shift = DAG.getNode(ISD::SRL, SL, HalfVT, Trunc, ShiftK); |
| 5138 | SDValue And = DAG.getNode(ISD::AND, SL, HalfVT, Shift, NewMask); |
| 5139 | return DAG.getNode(ISD::ZERO_EXTEND, SL, VT, And); |
| 5140 | } |
| 5141 | } |
| 5142 | } |
| 5143 | } |
| 5144 | |
| 5145 | return SDValue(); |
| 5146 | } |
| 5147 | |
| 5148 | bool DAGCombiner::isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, |
| 5149 | EVT LoadResultTy, EVT &ExtVT) { |
| 5150 | if (!AndC->getAPIntValue().isMask()) |
| 5151 | return false; |
| 5152 | |
| 5153 | unsigned ActiveBits = AndC->getAPIntValue().countTrailingOnes(); |
| 5154 | |
| 5155 | ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); |
| 5156 | EVT LoadedVT = LoadN->getMemoryVT(); |
| 5157 | |
| 5158 | if (ExtVT == LoadedVT && |
| 5159 | (!LegalOperations || |
| 5160 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT))) { |
| 5161 | // ZEXTLOAD will match without needing to change the size of the value being |
| 5162 | // loaded. |
| 5163 | return true; |
| 5164 | } |
| 5165 | |
| 5166 | // Do not change the width of a volatile or atomic loads. |
| 5167 | if (!LoadN->isSimple()) |
| 5168 | return false; |
| 5169 | |
| 5170 | // Do not generate loads of non-round integer types since these can |
| 5171 | // be expensive (and would be wrong if the type is not byte sized). |
| 5172 | if (!LoadedVT.bitsGT(ExtVT) || !ExtVT.isRound()) |
| 5173 | return false; |
| 5174 | |
| 5175 | if (LegalOperations && |
| 5176 | !TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT)) |
| 5177 | return false; |
| 5178 | |
| 5179 | if (!TLI.shouldReduceLoadWidth(LoadN, ISD::ZEXTLOAD, ExtVT)) |
| 5180 | return false; |
| 5181 | |
| 5182 | return true; |
| 5183 | } |
| 5184 | |
| 5185 | bool DAGCombiner::isLegalNarrowLdSt(LSBaseSDNode *LDST, |
| 5186 | ISD::LoadExtType ExtType, EVT &MemVT, |
| 5187 | unsigned ShAmt) { |
| 5188 | if (!LDST) |
| 5189 | return false; |
| 5190 | // Only allow byte offsets. |
| 5191 | if (ShAmt % 8) |
| 5192 | return false; |
| 5193 | |
| 5194 | // Do not generate loads of non-round integer types since these can |
| 5195 | // be expensive (and would be wrong if the type is not byte sized). |
| 5196 | if (!MemVT.isRound()) |
| 5197 | return false; |
| 5198 | |
| 5199 | // Don't change the width of a volatile or atomic loads. |
| 5200 | if (!LDST->isSimple()) |
| 5201 | return false; |
| 5202 | |
| 5203 | EVT LdStMemVT = LDST->getMemoryVT(); |
| 5204 | |
| 5205 | // Bail out when changing the scalable property, since we can't be sure that |
| 5206 | // we're actually narrowing here. |
| 5207 | if (LdStMemVT.isScalableVector() != MemVT.isScalableVector()) |
| 5208 | return false; |
| 5209 | |
| 5210 | // Verify that we are actually reducing a load width here. |
| 5211 | if (LdStMemVT.bitsLT(MemVT)) |
| 5212 | return false; |
| 5213 | |
| 5214 | // Ensure that this isn't going to produce an unsupported memory access. |
| 5215 | if (ShAmt) { |
| 5216 | assert(ShAmt % 8 == 0 && "ShAmt is byte offset")((ShAmt % 8 == 0 && "ShAmt is byte offset") ? static_cast <void> (0) : __assert_fail ("ShAmt % 8 == 0 && \"ShAmt is byte offset\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5216, __PRETTY_FUNCTION__)); |
| 5217 | const unsigned ByteShAmt = ShAmt / 8; |
| 5218 | const Align LDSTAlign = LDST->getAlign(); |
| 5219 | const Align NarrowAlign = commonAlignment(LDSTAlign, ByteShAmt); |
| 5220 | if (!TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), MemVT, |
| 5221 | LDST->getAddressSpace(), NarrowAlign, |
| 5222 | LDST->getMemOperand()->getFlags())) |
| 5223 | return false; |
| 5224 | } |
| 5225 | |
| 5226 | // It's not possible to generate a constant of extended or untyped type. |
| 5227 | EVT PtrType = LDST->getBasePtr().getValueType(); |
| 5228 | if (PtrType == MVT::Untyped || PtrType.isExtended()) |
| 5229 | return false; |
| 5230 | |
| 5231 | if (isa<LoadSDNode>(LDST)) { |
| 5232 | LoadSDNode *Load = cast<LoadSDNode>(LDST); |
| 5233 | // Don't transform one with multiple uses, this would require adding a new |
| 5234 | // load. |
| 5235 | if (!SDValue(Load, 0).hasOneUse()) |
| 5236 | return false; |
| 5237 | |
| 5238 | if (LegalOperations && |
| 5239 | !TLI.isLoadExtLegal(ExtType, Load->getValueType(0), MemVT)) |
| 5240 | return false; |
| 5241 | |
| 5242 | // For the transform to be legal, the load must produce only two values |
| 5243 | // (the value loaded and the chain). Don't transform a pre-increment |
| 5244 | // load, for example, which produces an extra value. Otherwise the |
| 5245 | // transformation is not equivalent, and the downstream logic to replace |
| 5246 | // uses gets things wrong. |
| 5247 | if (Load->getNumValues() > 2) |
| 5248 | return false; |
| 5249 | |
| 5250 | // If the load that we're shrinking is an extload and we're not just |
| 5251 | // discarding the extension we can't simply shrink the load. Bail. |
| 5252 | // TODO: It would be possible to merge the extensions in some cases. |
| 5253 | if (Load->getExtensionType() != ISD::NON_EXTLOAD && |
| 5254 | Load->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits() + ShAmt) |
| 5255 | return false; |
| 5256 | |
| 5257 | if (!TLI.shouldReduceLoadWidth(Load, ExtType, MemVT)) |
| 5258 | return false; |
| 5259 | } else { |
| 5260 | assert(isa<StoreSDNode>(LDST) && "It is not a Load nor a Store SDNode")((isa<StoreSDNode>(LDST) && "It is not a Load nor a Store SDNode" ) ? static_cast<void> (0) : __assert_fail ("isa<StoreSDNode>(LDST) && \"It is not a Load nor a Store SDNode\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5260, __PRETTY_FUNCTION__)); |
| 5261 | StoreSDNode *Store = cast<StoreSDNode>(LDST); |
| 5262 | // Can't write outside the original store |
| 5263 | if (Store->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits() + ShAmt) |
| 5264 | return false; |
| 5265 | |
| 5266 | if (LegalOperations && |
| 5267 | !TLI.isTruncStoreLegal(Store->getValue().getValueType(), MemVT)) |
| 5268 | return false; |
| 5269 | } |
| 5270 | return true; |
| 5271 | } |
| 5272 | |
| 5273 | bool DAGCombiner::SearchForAndLoads(SDNode *N, |
| 5274 | SmallVectorImpl<LoadSDNode*> &Loads, |
| 5275 | SmallPtrSetImpl<SDNode*> &NodesWithConsts, |
| 5276 | ConstantSDNode *Mask, |
| 5277 | SDNode *&NodeToMask) { |
| 5278 | // Recursively search for the operands, looking for loads which can be |
| 5279 | // narrowed. |
| 5280 | for (SDValue Op : N->op_values()) { |
| 5281 | if (Op.getValueType().isVector()) |
| 5282 | return false; |
| 5283 | |
| 5284 | // Some constants may need fixing up later if they are too large. |
| 5285 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { |
| 5286 | if ((N->getOpcode() == ISD::OR || N->getOpcode() == ISD::XOR) && |
| 5287 | (Mask->getAPIntValue() & C->getAPIntValue()) != C->getAPIntValue()) |
| 5288 | NodesWithConsts.insert(N); |
| 5289 | continue; |
| 5290 | } |
| 5291 | |
| 5292 | if (!Op.hasOneUse()) |
| 5293 | return false; |
| 5294 | |
| 5295 | switch(Op.getOpcode()) { |
| 5296 | case ISD::LOAD: { |
| 5297 | auto *Load = cast<LoadSDNode>(Op); |
| 5298 | EVT ExtVT; |
| 5299 | if (isAndLoadExtLoad(Mask, Load, Load->getValueType(0), ExtVT) && |
| 5300 | isLegalNarrowLdSt(Load, ISD::ZEXTLOAD, ExtVT)) { |
| 5301 | |
| 5302 | // ZEXTLOAD is already small enough. |
| 5303 | if (Load->getExtensionType() == ISD::ZEXTLOAD && |
| 5304 | ExtVT.bitsGE(Load->getMemoryVT())) |
| 5305 | continue; |
| 5306 | |
| 5307 | // Use LE to convert equal sized loads to zext. |
| 5308 | if (ExtVT.bitsLE(Load->getMemoryVT())) |
| 5309 | Loads.push_back(Load); |
| 5310 | |
| 5311 | continue; |
| 5312 | } |
| 5313 | return false; |
| 5314 | } |
| 5315 | case ISD::ZERO_EXTEND: |
| 5316 | case ISD::AssertZext: { |
| 5317 | unsigned ActiveBits = Mask->getAPIntValue().countTrailingOnes(); |
| 5318 | EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); |
| 5319 | EVT VT = Op.getOpcode() == ISD::AssertZext ? |
| 5320 | cast<VTSDNode>(Op.getOperand(1))->getVT() : |
| 5321 | Op.getOperand(0).getValueType(); |
| 5322 | |
| 5323 | // We can accept extending nodes if the mask is wider or an equal |
| 5324 | // width to the original type. |
| 5325 | if (ExtVT.bitsGE(VT)) |
| 5326 | continue; |
| 5327 | break; |
| 5328 | } |
| 5329 | case ISD::OR: |
| 5330 | case ISD::XOR: |
| 5331 | case ISD::AND: |
| 5332 | if (!SearchForAndLoads(Op.getNode(), Loads, NodesWithConsts, Mask, |
| 5333 | NodeToMask)) |
| 5334 | return false; |
| 5335 | continue; |
| 5336 | } |
| 5337 | |
| 5338 | // Allow one node which will masked along with any loads found. |
| 5339 | if (NodeToMask) |
| 5340 | return false; |
| 5341 | |
| 5342 | // Also ensure that the node to be masked only produces one data result. |
| 5343 | NodeToMask = Op.getNode(); |
| 5344 | if (NodeToMask->getNumValues() > 1) { |
| 5345 | bool HasValue = false; |
| 5346 | for (unsigned i = 0, e = NodeToMask->getNumValues(); i < e; ++i) { |
| 5347 | MVT VT = SDValue(NodeToMask, i).getSimpleValueType(); |
| 5348 | if (VT != MVT::Glue && VT != MVT::Other) { |
| 5349 | if (HasValue) { |
| 5350 | NodeToMask = nullptr; |
| 5351 | return false; |
| 5352 | } |
| 5353 | HasValue = true; |
| 5354 | } |
| 5355 | } |
| 5356 | assert(HasValue && "Node to be masked has no data result?")((HasValue && "Node to be masked has no data result?" ) ? static_cast<void> (0) : __assert_fail ("HasValue && \"Node to be masked has no data result?\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5356, __PRETTY_FUNCTION__)); |
| 5357 | } |
| 5358 | } |
| 5359 | return true; |
| 5360 | } |
| 5361 | |
| 5362 | bool DAGCombiner::BackwardsPropagateMask(SDNode *N) { |
| 5363 | auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1)); |
| 5364 | if (!Mask) |
| 5365 | return false; |
| 5366 | |
| 5367 | if (!Mask->getAPIntValue().isMask()) |
| 5368 | return false; |
| 5369 | |
| 5370 | // No need to do anything if the and directly uses a load. |
| 5371 | if (isa<LoadSDNode>(N->getOperand(0))) |
| 5372 | return false; |
| 5373 | |
| 5374 | SmallVector<LoadSDNode*, 8> Loads; |
| 5375 | SmallPtrSet<SDNode*, 2> NodesWithConsts; |
| 5376 | SDNode *FixupNode = nullptr; |
| 5377 | if (SearchForAndLoads(N, Loads, NodesWithConsts, Mask, FixupNode)) { |
| 5378 | if (Loads.size() == 0) |
| 5379 | return false; |
| 5380 | |
| 5381 | LLVM_DEBUG(dbgs() << "Backwards propagate AND: "; N->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "Backwards propagate AND: " ; N->dump(); } } while (false); |
| 5382 | SDValue MaskOp = N->getOperand(1); |
| 5383 | |
| 5384 | // If it exists, fixup the single node we allow in the tree that needs |
| 5385 | // masking. |
| 5386 | if (FixupNode) { |
| 5387 | LLVM_DEBUG(dbgs() << "First, need to fix up: "; FixupNode->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "First, need to fix up: "; FixupNode ->dump(); } } while (false); |
| 5388 | SDValue And = DAG.getNode(ISD::AND, SDLoc(FixupNode), |
| 5389 | FixupNode->getValueType(0), |
| 5390 | SDValue(FixupNode, 0), MaskOp); |
| 5391 | DAG.ReplaceAllUsesOfValueWith(SDValue(FixupNode, 0), And); |
| 5392 | if (And.getOpcode() == ISD ::AND) |
| 5393 | DAG.UpdateNodeOperands(And.getNode(), SDValue(FixupNode, 0), MaskOp); |
| 5394 | } |
| 5395 | |
| 5396 | // Narrow any constants that need it. |
| 5397 | for (auto *LogicN : NodesWithConsts) { |
| 5398 | SDValue Op0 = LogicN->getOperand(0); |
| 5399 | SDValue Op1 = LogicN->getOperand(1); |
| 5400 | |
| 5401 | if (isa<ConstantSDNode>(Op0)) |
| 5402 | std::swap(Op0, Op1); |
| 5403 | |
| 5404 | SDValue And = DAG.getNode(ISD::AND, SDLoc(Op1), Op1.getValueType(), |
| 5405 | Op1, MaskOp); |
| 5406 | |
| 5407 | DAG.UpdateNodeOperands(LogicN, Op0, And); |
| 5408 | } |
| 5409 | |
| 5410 | // Create narrow loads. |
| 5411 | for (auto *Load : Loads) { |
| 5412 | LLVM_DEBUG(dbgs() << "Propagate AND back to: "; Load->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "Propagate AND back to: "; Load ->dump(); } } while (false); |
| 5413 | SDValue And = DAG.getNode(ISD::AND, SDLoc(Load), Load->getValueType(0), |
| 5414 | SDValue(Load, 0), MaskOp); |
| 5415 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), And); |
| 5416 | if (And.getOpcode() == ISD ::AND) |
| 5417 | And = SDValue( |
| 5418 | DAG.UpdateNodeOperands(And.getNode(), SDValue(Load, 0), MaskOp), 0); |
| 5419 | SDValue NewLoad = ReduceLoadWidth(And.getNode()); |
| 5420 | assert(NewLoad &&((NewLoad && "Shouldn't be masking the load if it can't be narrowed" ) ? static_cast<void> (0) : __assert_fail ("NewLoad && \"Shouldn't be masking the load if it can't be narrowed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5421, __PRETTY_FUNCTION__)) |
| 5421 | "Shouldn't be masking the load if it can't be narrowed")((NewLoad && "Shouldn't be masking the load if it can't be narrowed" ) ? static_cast<void> (0) : __assert_fail ("NewLoad && \"Shouldn't be masking the load if it can't be narrowed\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5421, __PRETTY_FUNCTION__)); |
| 5422 | CombineTo(Load, NewLoad, NewLoad.getValue(1)); |
| 5423 | } |
| 5424 | DAG.ReplaceAllUsesWith(N, N->getOperand(0).getNode()); |
| 5425 | return true; |
| 5426 | } |
| 5427 | return false; |
| 5428 | } |
| 5429 | |
| 5430 | // Unfold |
| 5431 | // x & (-1 'logical shift' y) |
| 5432 | // To |
| 5433 | // (x 'opposite logical shift' y) 'logical shift' y |
| 5434 | // if it is better for performance. |
| 5435 | SDValue DAGCombiner::unfoldExtremeBitClearingToShifts(SDNode *N) { |
| 5436 | assert(N->getOpcode() == ISD::AND)((N->getOpcode() == ISD::AND) ? static_cast<void> (0 ) : __assert_fail ("N->getOpcode() == ISD::AND", "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5436, __PRETTY_FUNCTION__)); |
| 5437 | |
| 5438 | SDValue N0 = N->getOperand(0); |
| 5439 | SDValue N1 = N->getOperand(1); |
| 5440 | |
| 5441 | // Do we actually prefer shifts over mask? |
| 5442 | if (!TLI.shouldFoldMaskToVariableShiftPair(N0)) |
| 5443 | return SDValue(); |
| 5444 | |
| 5445 | // Try to match (-1 '[outer] logical shift' y) |
| 5446 | unsigned OuterShift; |
| 5447 | unsigned InnerShift; // The opposite direction to the OuterShift. |
| 5448 | SDValue Y; // Shift amount. |
| 5449 | auto matchMask = [&OuterShift, &InnerShift, &Y](SDValue M) -> bool { |
| 5450 | if (!M.hasOneUse()) |
| 5451 | return false; |
| 5452 | OuterShift = M->getOpcode(); |
| 5453 | if (OuterShift == ISD::SHL) |
| 5454 | InnerShift = ISD::SRL; |
| 5455 | else if (OuterShift == ISD::SRL) |
| 5456 | InnerShift = ISD::SHL; |
| 5457 | else |
| 5458 | return false; |
| 5459 | if (!isAllOnesConstant(M->getOperand(0))) |
| 5460 | return false; |
| 5461 | Y = M->getOperand(1); |
| 5462 | return true; |
| 5463 | }; |
| 5464 | |
| 5465 | SDValue X; |
| 5466 | if (matchMask(N1)) |
| 5467 | X = N0; |
| 5468 | else if (matchMask(N0)) |
| 5469 | X = N1; |
| 5470 | else |
| 5471 | return SDValue(); |
| 5472 | |
| 5473 | SDLoc DL(N); |
| 5474 | EVT VT = N->getValueType(0); |
| 5475 | |
| 5476 | // tmp = x 'opposite logical shift' y |
| 5477 | SDValue T0 = DAG.getNode(InnerShift, DL, VT, X, Y); |
| 5478 | // ret = tmp 'logical shift' y |
| 5479 | SDValue T1 = DAG.getNode(OuterShift, DL, VT, T0, Y); |
| 5480 | |
| 5481 | return T1; |
| 5482 | } |
| 5483 | |
| 5484 | /// Try to replace shift/logic that tests if a bit is clear with mask + setcc. |
| 5485 | /// For a target with a bit test, this is expected to become test + set and save |
| 5486 | /// at least 1 instruction. |
| 5487 | static SDValue combineShiftAnd1ToBitTest(SDNode *And, SelectionDAG &DAG) { |
| 5488 | assert(And->getOpcode() == ISD::AND && "Expected an 'and' op")((And->getOpcode() == ISD::AND && "Expected an 'and' op" ) ? static_cast<void> (0) : __assert_fail ("And->getOpcode() == ISD::AND && \"Expected an 'and' op\"" , "/build/llvm-toolchain-snapshot-13~++20210413100635+64c24f493e5f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5488, __PRETTY_FUNCTION__)); |
| 5489 | |
| 5490 | // This is probably not worthwhile without a supported type. |
| 5491 | EVT VT = And->getValueType(0); |
| 5492 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
| 5493 | if (!TLI.isTypeLegal(VT)) |
| 5494 | return SDValue(); |
| 5495 | |
| 5496 | // Look through an optional extension and find a 'not'. |
| 5497 | // TODO: Should we favor test+set even without the 'not' op? |
| 5498 | SDValue Not = And->getOperand(0), And1 = And->getOperand(1); |
| 5499 | if (Not.getOpcode() == ISD::ANY_EXTEND) |
| 5500 | Not = Not.getOperand(0); |
| 5501 | if (!isBitwiseNot(Not) || !Not.hasOneUse() || !isOneConstant(And1)) |
| 5502 | return SDValue(); |
| 5503 | |
| 5504 | // Look though an optional truncation. The source operand may not be the same |
| 5505 | // type as the original 'and', but that is ok because we are masking off |
| 5506 | // everything but the low bit. |
| 5507 | SDValue Srl = Not.getOperand(0); |
| 5508 | if (Srl.getOpcode() == ISD::TRUNCATE) |
| 5509 | Srl = Srl.getOperand(0); |
| 5510 | |
| 5511 | // Match a shift-right by constant. |
| 5512 | if (Srl.getOpcode() != ISD::SRL || !Srl.hasOneUse() || |
| 5513 | !isa<ConstantSDNode>(Srl.getOperand(1))) |
| 5514 | return SDValue(); |
| 5515 | |
| 5516 | // We might have looked through casts that make this transform invalid. |
| 5517 | // TODO: If the source type is wider than the result type, do the mask and |
| 5518 | // compare in the source type. |
| 5519 | const APInt &ShiftAmt = Srl.getConstantOperandAPInt(1); |
| 5520 | unsigned VTBitWidth = VT.getSizeInBits(); |
| 5521 | if (ShiftAmt.uge(VTBitWidth)) |
| 5522 | return SDValue(); |
| 5523 | |
| 5524 | // Turn this into a bit-test pattern using mask op + setcc: |
| 5525 | // and (not (srl X, C)), 1 --> (and X, 1<<C) == 0 |
| 5526 | SDLoc DL(And); |
| 5527 | SDValue X = DAG.getZExtOrTrunc(Srl.getOperand(0), DL, VT); |
| 5528 | EVT CCVT = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); |
| 5529 | SDValue Mask = DAG.getConstant( |
| 5530 | APInt::getOneBitSet(VTBitWidth, ShiftAmt.getZExtValue()), DL, VT); |
| 5531 | SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, Mask); |
| 5532 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 5533 | SDValue Setcc = DAG.getSetCC(DL, CCVT, NewAnd, Zero, ISD::SETEQ); |
| 5534 | return DAG.getZExtOrTrunc(Setcc, DL, VT); |
| 5535 | } |
| 5536 | |
| 5537 | SDValue DAGCombiner::visitAND(SDNode *N) { |
| 5538 | SDValue N0 = N->getOperand(0); |
| 5539 | SDValue N1 = N->getOperand(1); |
| 5540 | EVT VT = N1.getValueType(); |
| 5541 | |
| 5542 | // x & x --> x |
| 5543 | if (N0 == N1) |
| 5544 | return N0; |
| 5545 | |
| 5546 | // fold vector ops |
| 5547 | if (VT.isVector()) { |
| 5548 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) |
| 5549 | return FoldedVOp; |
| 5550 | |
| 5551 | // fold (and x, 0) -> 0, vector edition |
| 5552 | if (ISD::isBuildVectorAllZeros(N0.getNode())) |
| 5553 | // do not return N0, because undef node may exist in N0 |
| 5554 | return DAG.getConstant(APInt::getNullValue(N0.getScalarValueSizeInBits()), |
| 5555 | SDLoc(N), N0.getValueType()); |
| 5556 | if (ISD::isBuildVectorAllZeros(N1.getNode())) |
| 5557 | // do not return N1, because undef node may exist in N1 |
| 5558 | return DAG.getConstant(APInt::getNullValue(N1.getScalarValueSizeInBits()), |
| 5559 | SDLoc(N), N1.getValueType()); |
| 5560 | |
| 5561 | // fold (and x, -1) -> x, vector edition |
| 5562 | if (ISD::isBuildVectorAllOnes(N0.getNode())) |
| 5563 | return N1; |
| 5564 | if (ISD::isBuildVectorAllOnes(N1.getNode())) |
| 5565 | return N0; |
| 5566 | |
| 5567 | // fold (and (masked_load) (build_vec (x, ...))) to zext_masked_load |
| 5568 | auto *MLoad = dyn_cast<MaskedLoadSDNode>(N0); |
| 5569 | auto *BVec = dyn_cast<BuildVectorSDNode>(N1); |
| 5570 | if (MLoad && BVec && MLoad->getExtensionType() == ISD::EXTLOAD && |
| 5571 | N0.hasOneUse() && N1.hasOneUse()) { |
| 5572 | EVT LoadVT = MLoad->getMemoryVT(); |
| 5573 | EVT ExtVT = VT; |
| 5574 | if (TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT, LoadVT)) { |
| 5575 | // For this AND to be a zero extension of the masked load the elements |
| 5576 | // of the BuildVec must mask the bottom bits of the extended element |
| 5577 | // type |
| 5578 | if (ConstantSDNode *Splat = BVec->getConstantSplatNode()) { |
| 5579 | uint64_t ElementSize = |
| 5580 | LoadVT.getVectorElementType().getScalarSizeInBits(); |
| 5581 | if (Splat->getAPIntValue().isMask(ElementSize)) { |
| 5582 | return DAG.getMaskedLoad( |
| 5583 | ExtVT, SDLoc(N), MLoad->getChain(), MLoad->getBasePtr(), |
| 5584 | MLoad->getOffset(), MLoad->getMask(), MLoad->getPassThru(), |
| 5585 | LoadVT, MLoad->getMemOperand(), MLoad->getAddressingMode(), |
| 5586 | ISD::ZEXTLOAD, MLoad->isExpandingLoad()); |
| 5587 | } |
| 5588 | } |
| 5589 | } |
| 5590 | } |
| 5591 | } |
| 5592 | |
| 5593 | // fold (and c1, c2) -> c1&c2 |
| 5594 | ConstantSDNode *N1C = isConstOrConstSplat(N1); |
| 5595 | if (SDValue C = DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, {N0, N1})) |
| 5596 | return C; |
| 5597 | |
| 5598 | // canonicalize constant to RHS |
| 5599 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && |
| 5600 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) |
| 5601 | return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0); |
| 5602 | |
| 5603 | // fold (and x, -1) -> x |
| 5604 | if (isAllOnesConstant(N1)) |
| 5605 | return N0; |
| 5606 | |
| 5607 | // if (and x, c) is known to be zero, return 0 |
| 5608 | unsigned BitWidth = VT.getScalarSizeInBits(); |
| 5609 | if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), |
| 5610 | APInt::getAllOnesValue(BitWidth))) |
| 5611 | return DAG.getConstant(0, SDLoc(N), VT); |
| 5612 | |
| 5613 | if (SDValue NewSel = foldBinOpIntoSelect(N)) |
| 5614 | return NewSel; |
| 5615 | |
| 5616 | // reassociate and |
| 5617 | if (SDValue RAND = reassociateOps(ISD::AND, SDLoc(N), N0, N1, N->getFlags())) |
| 5618 | return RAND; |
| 5619 | |
| 5620 | // Try to convert a constant mask AND into a shuffle clear mask. |
| 5621 | if (VT.isVector()) |
| 5622 | if (SDValue Shuffle = XformToShuffleWithZero(N)) |
| 5623 | return Shuffle; |
| 5624 | |
| 5625 | if (SDValue Combined = combineCarryDiamond(*this, DAG, TLI, N0, N1, N)) |
| 5626 | return Combined; |
| 5627 | |
| 5628 | // fold (and (or x, C), D) -> D if (C & D) == D |
| 5629 | auto MatchSubset = [](ConstantSDNode *LHS, ConstantSDNode *RHS) { |
| 5630 | return RHS->getAPIntValue().isSubsetOf(LHS->getAPIntValue()); |
| 5631 | }; |
| 5632 | if (N0.getOpcode() == ISD::OR && |
| 5633 | ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchSubset)) |
| 5634 | return N1; |
| 5635 | // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. |
| 5636 | if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { |
| 5637 | SDValue N0Op0 = N0.getOperand(0); |
| 5638 | APInt Mask = ~N1C->getAPIntValue(); |
| 5639 | Mask = Mask.trunc(N0Op0.getScalarValueSizeInBits()); |
| 5640 | if (DAG.MaskedValueIsZero(N0Op0, Mask)) { |
| 5641 | SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), |
| 5642 | N0.getValueType(), N0Op0); |
| 5643 | |
| 5644 | // Replace uses of the AND with uses of the Zero extend node. |
| 5645 | CombineTo(N, Zext); |
| 5646 | |
| 5647 | // We actually want to replace all uses of the any_extend with the |
| 5648 | // zero_extend, to avoid duplicating things. This will later cause this |
| 5649 | // AND to be folded. |
| 5650 | CombineTo(N0.getNode(), Zext); |
| 5651 | return SDValue(N, 0); // Return N so it doesn't get rechecked! |
| 5652 | } |
| 5653 | } |
| 5654 | |
| 5655 | // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) -> |
| 5656 | // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must |
| 5657 | // already be zero by virtue of the width of the base type of the load. |
| 5658 | // |
| 5659 | // the 'X' node here can either be nothing or an extract_vector_elt to catch |
| 5660 | // more cases. |
| 5661 | if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && |
| 5662 | N0.getValueSizeInBits() == N0.getOperand(0).getScalarValueSizeInBits() && |
| 5663 | N0.getOperand(0).getOpcode() == ISD::LOAD && |
| 5664 | N0.getOperand(0).getResNo() == 0) || |
| 5665 | (N0.getOpcode() == ISD::LOAD && N0.getResNo() == 0)) { |
| 5666 | LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ? |
| 5667 | N0 : N0.getOperand(0) ); |
| 5668 | |
| 5669 | // Get the constant (if applicable) the zero'th operand is being ANDed with. |
| 5670 | // This can be a pure constant or a vector splat, in which case we treat the |
| 5671 | // vector as a scalar and use the splat value. |
| 5672 | APInt Constant = APInt::getNullValue(1); |
| 5673 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { |
| 5674 | Constant = C->getAPIntValue(); |
| 5675 | } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) { |
| 5676 | APInt SplatValue, SplatUndef; |
| 5677 | unsigned SplatBitSize; |
| 5678 | bool HasAnyUndefs; |
| 5679 | bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef, |
| 5680 | SplatBitSize, HasAnyUndefs); |
| 5681 | if (IsSplat) { |
| 5682 | // Undef bits can contribute to a possible optimisation if set, so |
| 5683 | // set them. |
| 5684 | SplatValue |= SplatUndef; |
| 5685 | |
| 5686 | // The splat value may be something like "0x00FFFFFF", which means 0 for |
| 5687 | // the first vector value and FF for the rest, repeating. We need a mask |
| 5688 | // that will apply equally to all members of the vector, so AND all the |
| 5689 | // lanes of the constant together. |
| 5690 | unsigned EltBitWidth = Vector->getValueType(0).getScalarSizeInBits(); |
| 5691 | |
| 5692 | // If the splat value has been compressed to a bitlength lower |
| 5693 | // than the size of the vector lane, we need to re-expand it to |
| 5694 | // the lane size. |
| 5695 | if (EltBitWidth > SplatBitSize) |
| 5696 | for (SplatValue = SplatValue.zextOrTrunc(EltBitWidth); |
| 5697 | SplatBitSize < EltBitWidth; SplatBitSize = SplatBitSize * 2) |
| 5698 | SplatValue |= SplatValue.shl(SplatBitSize); |
| 5699 | |
| 5700 | // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a |
| 5701 | // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value. |
| 5702 | if ((SplatBitSize % EltBitWidth) == 0) { |
| 5703 | Constant = APInt::getAllOnesValue(EltBitWidth); |
| 5704 | for (unsigned i = 0, n = (SplatBitSize / EltBitWidth); i < n; ++i) |
| 5705 | Constant &= SplatValue.extractBits(EltBitWidth, i * EltBitWidth); |
| 5706 | } |
| 5707 | } |
| 5708 | } |
| 5709 | |
| 5710 | // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is |
| 5711 | // actually legal and isn't going to get expanded, else this is a false |
| 5712 | // optimisation. |
| 5713 | bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD, |
| 5714 | Load->getValueType(0), |
| 5715 | Load->getMemoryVT()); |
| 5716 | |
| 5717 | // Resize the constant to the same size as the original memory access before |
| 5718 | // extension. If it is still the AllOnesValue then this AND is completely |
| 5719 | // unneeded. |
| 5720 | Constant = Constant.zextOrTrunc(Load->getMemoryVT().getScalarSizeInBits()); |
| 5721 | |
| 5722 | bool B; |
| 5723 | switch (Load->getExtensionType()) { |
| 5724 | default: B = false; break; |
| 5725 | case ISD::EXTLOAD: B = CanZextLoadProfitably; break; |
| 5726 | case ISD::ZEXTLOAD: |
| 5727 | case ISD::NON_EXTLOAD: B = true; break; |
| 5728 | } |
| 5729 | |
| 5730 | if (B && Constant.isAllOnesValue()) { |
| 5731 | // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to |
| 5732 | // preserve semantics once we get rid of the AND. |
| 5733 | SDValue NewLoad(Load, 0); |
| 5734 | |
| 5735 | // Fold the AND away. NewLoad may get replaced immediately. |
| 5736 | CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0); |
| 5737 | |
| 5738 | if (Load->getExtensionType() == ISD::EXTLOAD) { |
| 5739 | NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD, |
| 5740 | Load->getValueType(0), SDLoc(Load), |
| 5741 | Load->getChain(), Load->getBasePtr(), |
| 5742 | Load->getOffset(), Load->getMemoryVT(), |
| 5743 | Load->getMemOperand()); |
| 5744 | // Replace uses of the EXTLOAD with the new ZEXTLOAD. |
| 5745 | if (Load->getNumValues() == 3) { |
| 5746 | // PRE/POST_INC loads have 3 values. |
| 5747 | SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1), |
| 5748 | NewLoad.getValue(2) }; |
| 5749 | CombineTo(Load, To, 3, true); |
| 5750 | } else { |
| 5751 | CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1)); |
| 5752 | } |
| 5753 | } |
| 5754 | |
| 5755 | return SDValue(N, 0); // Return N so it doesn't get rechecked! |
| 5756 | } |
| 5757 | } |
| 5758 | |
| 5759 | // fold (and (masked_gather x)) -> (zext_masked_gather x) |
| 5760 | if (auto *GN0 = dyn_cast<MaskedGatherSDNode>(N0)) { |
| 5761 | EVT MemVT = GN0->getMemoryVT(); |
| 5762 | EVT ScalarVT = MemVT.getScalarType(); |
| 5763 | |
| 5764 | if (SDValue(GN0, 0).hasOneUse() && |
| 5765 | isConstantSplatVectorMaskForType(N1.getNode(), ScalarVT) && |
| 5766 | TLI.isVectorLoadExtDesirable(SDValue(SDValue(GN0, 0)))) { |
| 5767 | SDValue Ops[] = {GN0->getChain(), GN0->getPassThru(), GN0->getMask(), |
| 5768 | GN0->getBasePtr(), GN0->getIndex(), GN0->getScale()}; |
| 5769 | |
| 5770 | SDValue ZExtLoad = DAG.getMaskedGather( |
| 5771 | DAG.getVTList(VT, MVT::Other), MemVT, SDLoc(N), Ops, |
| 5772 | GN0->getMemOperand(), GN0->getIndexType(), ISD::ZEXTLOAD); |
| 5773 | |
| 5774 | CombineTo(N, ZExtLoad); |
| 5775 | AddToWorklist(ZExtLoad.getNode()); |
| 5776 | // Avoid recheck of N. |
| 5777 | return SDValue(N, 0); |
| 5778 | } |
| 5779 | } |
| 5780 | |
| 5781 | // fold (and (load x), 255) -> (zextload x, i8) |
| 5782 | // fold (and (extload x, i16), 255) -> (zextload x, i8) |
| 5783 | // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8) |
| 5784 | if (!VT.isVector() && N1C && (N0.getOpcode() == ISD::LOAD || |
| 5785 | (N0.getOpcode() == ISD::ANY_EXTEND && |
| 5786 | N0.getOperand(0).getOpcode() == ISD::LOAD))) { |
| 5787 | if (SDValue Res = ReduceLoadWidth(N)) { |
| 5788 | LoadSDNode *LN0 = N0->getOpcode() == ISD::ANY_EXTEND |
| 5789 | ? cast<LoadSDNode>(N0.getOperand(0)) : cast<LoadSDNode>(N0); |
| 5790 | AddToWorklist(N); |
| 5791 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 0), Res); |
| 5792 | return SDValue(N, 0); |
| 5793 | } |
| 5794 | } |
| 5795 | |
| 5796 | if (LegalTypes) { |
| 5797 | // Attempt to propagate the AND back up to the leaves which, if they're |
| 5798 | // loads, can be combined to narrow loads and the AND node can be removed. |
| 5799 | // Perform after legalization so that extend nodes will already be |
| 5800 | // combined into the loads. |
| 5801 | if (BackwardsPropagateMask(N)) |
| 5802 | return SDValue(N, 0); |
| 5803 | } |
| 5804 | |
| 5805 | if (SDValue Combined = visitANDLike(N0, N1, N)) |
| 5806 | return Combined; |
| 5807 | |
| 5808 | // Simplify: (and (op x...), (op y...)) -> (op (and x, y)) |
| 5809 | if (N0.getOpcode() == N1.getOpcode()) |
| 5810 | if (SDValue V = hoistLogicOpWithSameOpcodeHands(N)) |
| 5811 | return V; |
| 5812 | |
| 5813 | // Masking the negated extension of a boolean is just the zero-extended |
| 5814 | // boolean: |
| 5815 | // and (sub 0, zext(bool X)), 1 --> zext(bool X) |
| 5816 | // and (sub 0, sext(bool X)), 1 --> zext(bool X) |
| 5817 | // |
| 5818 | // Note: the SimplifyDemandedBits fold below can make an information-losing |
| 5819 | // transform, and then we have no way to find this better fold. |
| 5820 | if (N1C && N1C->isOne() && N0.getOpcode() == ISD::SUB) { |
| 5821 | if (isNullOrNullSplat(N0.getOperand(0))) { |
| 5822 | SDValue SubRHS = N0.getOperand(1); |
| 5823 | if (SubRHS.getOpcode() == ISD::ZERO_EXTEND && |
| 5824 | SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) |
| 5825 | return SubRHS; |
| 5826 | if (SubRHS.getOpcode() == ISD::SIGN_EXTEND && |
| 5827 | SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) |
| 5828 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, SubRHS.getOperand(0)); |
| 5829 | } |
| 5830 | } |
| 5831 | |
| 5832 | // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) |
| 5833 | // fold (and (sra)) -> (and (srl)) when possible. |
| 5834 | if (SimplifyDemandedBits(SDValue(N, 0))) |
| 5835 | return SDValue(N, 0); |
| 5836 | |
| 5837 | // fold (zext_inreg (extload x)) -> (zextload x) |
| 5838 | // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use |
| 5839 | if (ISD::isUNINDEXEDLoad(N0.getNode()) && |
| 5840 | (ISD::isEXTLoad(N0.getNode()) || |
| 5841 | (ISD::isSEXTLoad(N0.getNode()) && N0.hasOneUse()))) { |
| 5842 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); |
| 5843 | EVT MemVT = LN0->getMemoryVT(); |
| 5844 | // If we zero all the possible extended bits, then we can turn this into |
| 5845 | // a zextload if we are running before legalize or the operation is legal. |
| 5846 | unsigned ExtBitSize = N1.getScalarValueSizeInBits(); |
| 5847 | unsigned MemBitSize = MemVT.getScalarSizeInBits(); |
| 5848 | APInt ExtBits = APInt::getHighBitsSet(ExtBitSize, ExtBitSize - MemBitSize); |
| 5849 | if (DAG.MaskedValueIsZero(N1, ExtBits) && |
| 5850 | ((!LegalOperations && LN0->isSimple()) || |
| 5851 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) { |
| 5852 | SDValue ExtLoad = |
| 5853 | DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, LN0->getChain(), |
| 5854 | LN0->getBasePtr(), MemVT, LN0->getMemOperand()); |
| 5855 | AddToWorklist(N); |
| 5856 | CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); |
| 5857 | return SDValue(N, 0); // Return N so it doesn't get rechecked! |
| 5858 | } |
| 5859 | } |
| 5860 | |
| 5861 | // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const) |
| 5862 | if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) { |
| 5863 | if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), |
| 5864 | N0.getOperand(1), false)) |
| 5865 | return BSwap; |
| 5866 | } |
| 5867 | |
| 5868 | if (SDValue Shifts = unfoldExtremeBitClearingToShifts(N)) |
| 5869 | return Shifts; |
| 5870 | |
| 5871 | if (TLI.hasBitTest(N0, N1)) |
| 5872 | if (SDValue V = combineShiftAnd1ToBitTest(N, DAG)) |
| 5873 | return V; |
| 5874 | |
| 5875 | // Recognize the following pattern: |
| 5876 | // |
| 5877 | // AndVT = (and (sign_extend NarrowVT to AndVT) #bitmask) |
| 5878 | // |
| 5879 | // where bitmask is a mask that clears the upper bits of AndVT. The |
| 5880 | // number of bits in bitmask must be a power of two. |
| 5881 | auto IsAndZeroExtMask = [](SDValue LHS, SDValue RHS) { |
| 5882 | if (LHS->getOpcode() != ISD::SIGN_EXTEND) |
| 5883 | return false; |
| 5884 | |
| 5885 | auto *C = dyn_cast<ConstantSDNode>(RHS); |
| 5886 | if (!C) |
| 5887 | return false; |
| 5888 | |
| 5889 | if (!C->getAPIntValue().isMask( |
| 5890 | LHS.getOperand(0).getValueType().getFixedSizeInBits())) |
| 5891 | return false; |
| 5892 | |
| 5893 | return true; |
| 5894 | }; |
| 5895 | |
| 5896 | // Replace (and (sign_extend ...) #bitmask) with (zero_extend ...). |
| 5897 | if (IsAndZeroExtMask(N0, N1)) |
| 5898 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0.getOperand(0)); |
| 5899 | |
| 5900 | return SDValue(); |
| 5901 | } |
| 5902 | |
| 5903 | /// Match (a >> 8) | (a << 8) as (bswap a) >> 16. |
| 5904 | SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, |
| 5905 | bool DemandHighBits) { |
| 5906 | if (!LegalOperations) |
| 5907 | return SDValue(); |
| 5908 | |
| 5909 | EVT VT = N->getValueType(0); |
| 5910 | if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16) |
| 5911 | return SDValue(); |
| 5912 | if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) |
| 5913 | return SDValue(); |
| 5914 | |
| 5915 | // Recognize (and (shl a, 8), 0xff00), (and (srl a, 8), 0xff) |
| 5916 | bool LookPassAnd0 = false; |
| 5917 | bool LookPassAnd1 = false; |
| 5918 | if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL) |
| 5919 | std::swap(N0, N1); |
| 5920 | if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL) |
| 5921 | std::swap(N0, N1); |
| 5922 | if (N0.getOpcode() == ISD::AND) { |
| 5923 | if (!N0.getNode()->hasOneUse()) |
| 5924 | return SDValue(); |
| 5925 | ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); |
| 5926 | // Also handle 0xffff since the LHS is guaranteed to have zeros there. |
| 5927 | // This is needed for X86. |
| 5928 | if (!N01C || (N01C->getZExtValue() != 0xFF00 && |
| 5929 | N01C->getZExtValue() != 0xFFFF)) |
| 5930 | return SDValue(); |
| 5931 | N0 = N0.getOperand(0); |
| 5932 | LookPassAnd0 = true; |
| 5933 | } |
| 5934 | |
| 5935 | if (N1.getOpcode() == ISD::AND) { |
| 5936 | if (!N1.getNode()->hasOneUse()) |
| 5937 | return SDValue(); |
| 5938 | ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); |
| 5939 | if (!N11C || N11C->getZExtValue() != 0xFF) |
| 5940 | return SDValue(); |
| 5941 | N1 = N1.getOperand(0); |
| 5942 | LookPassAnd1 = true; |
| 5943 | } |
| 5944 | |
| 5945 | if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) |
| 5946 | std::swap(N0, N1); |
| 5947 | if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) |
| 5948 | return SDValue(); |
| 5949 | if (!N0.getNode()->hasOneUse() || !N1.getNode()->hasOneUse()) |
| 5950 | return SDValue(); |
| 5951 | |
| 5952 | ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); |
| 5953 | ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); |
| 5954 | if (!N01C || !N11C) |
| 5955 | return SDValue(); |
| 5956 | if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8) |
| 5957 | return SDValue(); |
| 5958 | |
| 5959 | // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8) |
| 5960 | SDValue N00 = N0->getOperand(0); |
| 5961 | if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) { |
| 5962 | if (!N00.getNode()->hasOneUse()) |
| 5963 | return SDValue(); |
| 5964 | ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1)); |
| 5965 | if (!N001C || N001C->getZExtValue() != 0xFF) |
| 5966 | return SDValue(); |
| 5967 | N00 = N00.getOperand(0); |
| 5968 | LookPassAnd0 = true; |
| 5969 | } |
| 5970 | |
| 5971 | SDValue N10 = N1->getOperand(0); |
| 5972 | if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) { |
| 5973 | if (!N10.getNode()->hasOneUse()) |
| 5974 | return SDValue(); |
| 5975 | ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1)); |
| 5976 | // Also allow 0xFFFF since the bits will be shifted out. This is needed |
| 5977 | // for X86. |
| 5978 | if (!N101C || (N101C->getZExtValue() != 0xFF00 && |
| 5979 | N101C->getZExtValue() != 0xFFFF)) |
| 5980 | return SDValue(); |
| 5981 | N10 = N10.getOperand(0); |
| 5982 | LookPassAnd1 = true; |
| 5983 | } |
| 5984 | |
| 5985 | if (N00 != N10) |
| 5986 | return SDValue(); |
| 5987 | |
| 5988 | // Make sure everything beyond the low halfword gets set to zero since the SRL |
| 5989 | // 16 will clear the top bits. |
| 5990 | unsigned OpSizeInBits = VT.getSizeInBits(); |
| 5991 | if (DemandHighBits && OpSizeInBits > 16) { |
| 5992 | // If the left-shift isn't masked out then the only way this is a bswap is |
| 5993 | // if all bits beyond the low 8 are 0. In that case the entire pattern |
| 5994 | // reduces to a left shift anyway: leave it for other parts of the combiner. |
| 5995 | if (!LookPassAnd0) |
| 5996 | return SDValue(); |
| 5997 | |
| 5998 | // However, if the right shift isn't masked out then it might be because |
| 5999 | // it's not needed. See if we can spot that too. |
| 6000 | if (!LookPassAnd1 && |
| 6001 | !DAG.MaskedValueIsZero( |
| 6002 | N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16))) |
| 6003 | return SDValue(); |
| 6004 | } |
| 6005 | |
| 6006 | SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); |
| 6007 | if (OpSizeInBits > 16) { |
| 6008 | SDLoc DL(N); |
| 6009 | Res = DAG.getNode(ISD::SRL, DL, VT, Res, |
| 6010 | DAG.getConstant(OpSizeInBits - 16, DL, |
| 6011 | getShiftAmountTy(VT))); |
| 6012 | } |
| 6013 | return Res; |
| 6014 | } |
| 6015 | |
| 6016 | /// Return true if the specified node is an element that makes up a 32-bit |
| 6017 | /// packed halfword byteswap. |
| 6018 | /// ((x & 0x000000ff) << 8) | |
| 6019 | /// ((x & 0x0000ff00) >> 8) | |
| 6020 | /// ((x & 0x00ff0000) << 8) | |
| 6021 | /// ((x & 0xff000000) >> 8) |
| 6022 | static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) { |
| 6023 | if (!N.getNode()->hasOneUse()) |
| 6024 | return false; |
| 6025 | |
| 6026 | unsigned Opc = N.getOpcode(); |
| 6027 | if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL) |
| 6028 | return false; |
| 6029 | |
| 6030 | SDValue N0 = N.getOperand(0); |
| 6031 | unsigned Opc0 = N0.getOpcode(); |
| 6032 | if (Opc0 != ISD::AND && Opc0 != ISD::SHL && Opc0 != ISD::SRL) |
| 6033 | return false; |
| 6034 | |
| 6035 | ConstantSDNode *N1C = nullptr; |
| 6036 | // SHL or SRL: look upstream for AND mask operand |
| 6037 | if (Opc == ISD::AND) |
| 6038 | N1C = dyn_cast<ConstantSDNode>(N.getOperand(1)); |
| 6039 | else if (Opc0 == ISD::AND) |
| 6040 | N1C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); |
| 6041 | if (!N1C) |
| 6042 | return false; |
| 6043 | |
| 6044 | unsigned MaskByteOffset; |
| 6045 | switch (N1C->getZExtValue()) { |
| 6046 | default: |
| 6047 | return false; |
| 6048 | case 0xFF: MaskByteOffset = 0; break; |
| 6049 | case 0xFF00: MaskByteOffset = 1; break; |
| 6050 | case 0xFFFF: |
| 6051 | // In case demanded bits didn't clear the bits that will be shifted out. |