| File: | include/llvm/CodeGen/SelectionDAGNodes.h |
| Warning: | line 1159, column 10 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/SmallPtrSet.h" | |||
| 28 | #include "llvm/ADT/SmallSet.h" | |||
| 29 | #include "llvm/ADT/SmallVector.h" | |||
| 30 | #include "llvm/ADT/Statistic.h" | |||
| 31 | #include "llvm/Analysis/AliasAnalysis.h" | |||
| 32 | #include "llvm/Analysis/MemoryLocation.h" | |||
| 33 | #include "llvm/CodeGen/DAGCombine.h" | |||
| 34 | #include "llvm/CodeGen/ISDOpcodes.h" | |||
| 35 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
| 36 | #include "llvm/CodeGen/MachineFunction.h" | |||
| 37 | #include "llvm/CodeGen/MachineMemOperand.h" | |||
| 38 | #include "llvm/CodeGen/RuntimeLibcalls.h" | |||
| 39 | #include "llvm/CodeGen/SelectionDAG.h" | |||
| 40 | #include "llvm/CodeGen/SelectionDAGAddressAnalysis.h" | |||
| 41 | #include "llvm/CodeGen/SelectionDAGNodes.h" | |||
| 42 | #include "llvm/CodeGen/SelectionDAGTargetInfo.h" | |||
| 43 | #include "llvm/CodeGen/TargetLowering.h" | |||
| 44 | #include "llvm/CodeGen/TargetRegisterInfo.h" | |||
| 45 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
| 46 | #include "llvm/CodeGen/ValueTypes.h" | |||
| 47 | #include "llvm/IR/Attributes.h" | |||
| 48 | #include "llvm/IR/Constant.h" | |||
| 49 | #include "llvm/IR/DataLayout.h" | |||
| 50 | #include "llvm/IR/DerivedTypes.h" | |||
| 51 | #include "llvm/IR/Function.h" | |||
| 52 | #include "llvm/IR/LLVMContext.h" | |||
| 53 | #include "llvm/IR/Metadata.h" | |||
| 54 | #include "llvm/Support/Casting.h" | |||
| 55 | #include "llvm/Support/CodeGen.h" | |||
| 56 | #include "llvm/Support/CommandLine.h" | |||
| 57 | #include "llvm/Support/Compiler.h" | |||
| 58 | #include "llvm/Support/Debug.h" | |||
| 59 | #include "llvm/Support/ErrorHandling.h" | |||
| 60 | #include "llvm/Support/KnownBits.h" | |||
| 61 | #include "llvm/Support/MachineValueType.h" | |||
| 62 | #include "llvm/Support/MathExtras.h" | |||
| 63 | #include "llvm/Support/raw_ostream.h" | |||
| 64 | #include "llvm/Target/TargetMachine.h" | |||
| 65 | #include "llvm/Target/TargetOptions.h" | |||
| 66 | #include <algorithm> | |||
| 67 | #include <cassert> | |||
| 68 | #include <cstdint> | |||
| 69 | #include <functional> | |||
| 70 | #include <iterator> | |||
| 71 | #include <string> | |||
| 72 | #include <tuple> | |||
| 73 | #include <utility> | |||
| 74 | ||||
| 75 | using namespace llvm; | |||
| 76 | ||||
| 77 | #define DEBUG_TYPE"dagcombine" "dagcombine" | |||
| 78 | ||||
| 79 | STATISTIC(NodesCombined , "Number of dag nodes combined")static llvm::Statistic NodesCombined = {"dagcombine", "NodesCombined" , "Number of dag nodes combined"}; | |||
| 80 | STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created")static llvm::Statistic PreIndexedNodes = {"dagcombine", "PreIndexedNodes" , "Number of pre-indexed nodes created"}; | |||
| 81 | STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created")static llvm::Statistic PostIndexedNodes = {"dagcombine", "PostIndexedNodes" , "Number of post-indexed nodes created"}; | |||
| 82 | STATISTIC(OpsNarrowed , "Number of load/op/store narrowed")static llvm::Statistic OpsNarrowed = {"dagcombine", "OpsNarrowed" , "Number of load/op/store narrowed"}; | |||
| 83 | 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"}; | |||
| 84 | STATISTIC(SlicedLoads, "Number of load sliced")static llvm::Statistic SlicedLoads = {"dagcombine", "SlicedLoads" , "Number of load sliced"}; | |||
| 85 | STATISTIC(NumFPLogicOpsConv, "Number of logic ops converted to fp ops")static llvm::Statistic NumFPLogicOpsConv = {"dagcombine", "NumFPLogicOpsConv" , "Number of logic ops converted to fp ops"}; | |||
| 86 | ||||
| 87 | static cl::opt<bool> | |||
| 88 | CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, | |||
| 89 | cl::desc("Enable DAG combiner's use of IR alias analysis")); | |||
| 90 | ||||
| 91 | static cl::opt<bool> | |||
| 92 | UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true), | |||
| 93 | cl::desc("Enable DAG combiner's use of TBAA")); | |||
| 94 | ||||
| 95 | #ifndef NDEBUG | |||
| 96 | static cl::opt<std::string> | |||
| 97 | CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden, | |||
| 98 | cl::desc("Only use DAG-combiner alias analysis in this" | |||
| 99 | " function")); | |||
| 100 | #endif | |||
| 101 | ||||
| 102 | /// Hidden option to stress test load slicing, i.e., when this option | |||
| 103 | /// is enabled, load slicing bypasses most of its profitability guards. | |||
| 104 | static cl::opt<bool> | |||
| 105 | StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden, | |||
| 106 | cl::desc("Bypass the profitability model of load slicing"), | |||
| 107 | cl::init(false)); | |||
| 108 | ||||
| 109 | static cl::opt<bool> | |||
| 110 | MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true), | |||
| 111 | cl::desc("DAG combiner may split indexing from loads")); | |||
| 112 | ||||
| 113 | static cl::opt<bool> | |||
| 114 | EnableStoreMerging("combiner-store-merging", cl::Hidden, cl::init(true), | |||
| 115 | cl::desc("DAG combiner enable merging multiple stores " | |||
| 116 | "into a wider store")); | |||
| 117 | ||||
| 118 | static cl::opt<unsigned> TokenFactorInlineLimit( | |||
| 119 | "combiner-tokenfactor-inline-limit", cl::Hidden, cl::init(2048), | |||
| 120 | cl::desc("Limit the number of operands to inline for Token Factors")); | |||
| 121 | ||||
| 122 | static cl::opt<unsigned> StoreMergeDependenceLimit( | |||
| 123 | "combiner-store-merge-dependence-limit", cl::Hidden, cl::init(10), | |||
| 124 | cl::desc("Limit the number of times for the same StoreNode and RootNode " | |||
| 125 | "to bail out in store merging dependence check")); | |||
| 126 | ||||
| 127 | namespace { | |||
| 128 | ||||
| 129 | class DAGCombiner { | |||
| 130 | SelectionDAG &DAG; | |||
| 131 | const TargetLowering &TLI; | |||
| 132 | CombineLevel Level; | |||
| 133 | CodeGenOpt::Level OptLevel; | |||
| 134 | bool LegalOperations = false; | |||
| 135 | bool LegalTypes = false; | |||
| 136 | bool ForCodeSize; | |||
| 137 | ||||
| 138 | /// Worklist of all of the nodes that need to be simplified. | |||
| 139 | /// | |||
| 140 | /// This must behave as a stack -- new nodes to process are pushed onto the | |||
| 141 | /// back and when processing we pop off of the back. | |||
| 142 | /// | |||
| 143 | /// The worklist will not contain duplicates but may contain null entries | |||
| 144 | /// due to nodes being deleted from the underlying DAG. | |||
| 145 | SmallVector<SDNode *, 64> Worklist; | |||
| 146 | ||||
| 147 | /// Mapping from an SDNode to its position on the worklist. | |||
| 148 | /// | |||
| 149 | /// This is used to find and remove nodes from the worklist (by nulling | |||
| 150 | /// them) when they are deleted from the underlying DAG. It relies on | |||
| 151 | /// stable indices of nodes within the worklist. | |||
| 152 | DenseMap<SDNode *, unsigned> WorklistMap; | |||
| 153 | /// This records all nodes attempted to add to the worklist since we | |||
| 154 | /// considered a new worklist entry. As we keep do not add duplicate nodes | |||
| 155 | /// in the worklist, this is different from the tail of the worklist. | |||
| 156 | SmallSetVector<SDNode *, 32> PruningList; | |||
| 157 | ||||
| 158 | /// Set of nodes which have been combined (at least once). | |||
| 159 | /// | |||
| 160 | /// This is used to allow us to reliably add any operands of a DAG node | |||
| 161 | /// which have not yet been combined to the worklist. | |||
| 162 | SmallPtrSet<SDNode *, 32> CombinedNodes; | |||
| 163 | ||||
| 164 | /// Map from candidate StoreNode to the pair of RootNode and count. | |||
| 165 | /// The count is used to track how many times we have seen the StoreNode | |||
| 166 | /// with the same RootNode bail out in dependence check. If we have seen | |||
| 167 | /// the bail out for the same pair many times over a limit, we won't | |||
| 168 | /// consider the StoreNode with the same RootNode as store merging | |||
| 169 | /// candidate again. | |||
| 170 | DenseMap<SDNode *, std::pair<SDNode *, unsigned>> StoreRootCountMap; | |||
| 171 | ||||
| 172 | // AA - Used for DAG load/store alias analysis. | |||
| 173 | AliasAnalysis *AA; | |||
| 174 | ||||
| 175 | /// When an instruction is simplified, add all users of the instruction to | |||
| 176 | /// the work lists because they might get more simplified now. | |||
| 177 | void AddUsersToWorklist(SDNode *N) { | |||
| 178 | for (SDNode *Node : N->uses()) | |||
| 179 | AddToWorklist(Node); | |||
| 180 | } | |||
| 181 | ||||
| 182 | // Prune potentially dangling nodes. This is called after | |||
| 183 | // any visit to a node, but should also be called during a visit after any | |||
| 184 | // failed combine which may have created a DAG node. | |||
| 185 | void clearAddedDanglingWorklistEntries() { | |||
| 186 | // Check any nodes added to the worklist to see if they are prunable. | |||
| 187 | while (!PruningList.empty()) { | |||
| 188 | auto *N = PruningList.pop_back_val(); | |||
| 189 | if (N->use_empty()) | |||
| 190 | recursivelyDeleteUnusedNodes(N); | |||
| 191 | } | |||
| 192 | } | |||
| 193 | ||||
| 194 | SDNode *getNextWorklistEntry() { | |||
| 195 | // Before we do any work, remove nodes that are not in use. | |||
| 196 | clearAddedDanglingWorklistEntries(); | |||
| 197 | SDNode *N = nullptr; | |||
| 198 | // The Worklist holds the SDNodes in order, but it may contain null | |||
| 199 | // entries. | |||
| 200 | while (!N && !Worklist.empty()) { | |||
| 201 | N = Worklist.pop_back_val(); | |||
| 202 | } | |||
| 203 | ||||
| 204 | if (N) { | |||
| 205 | bool GoodWorklistEntry = WorklistMap.erase(N); | |||
| 206 | (void)GoodWorklistEntry; | |||
| 207 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 208, __PRETTY_FUNCTION__)) | |||
| 208 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 208, __PRETTY_FUNCTION__)); | |||
| 209 | } | |||
| 210 | return N; | |||
| 211 | } | |||
| 212 | ||||
| 213 | /// Call the node-specific routine that folds each particular type of node. | |||
| 214 | SDValue visit(SDNode *N); | |||
| 215 | ||||
| 216 | public: | |||
| 217 | DAGCombiner(SelectionDAG &D, AliasAnalysis *AA, CodeGenOpt::Level OL) | |||
| 218 | : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes), | |||
| 219 | OptLevel(OL), AA(AA) { | |||
| 220 | ForCodeSize = DAG.getMachineFunction().getFunction().hasOptSize(); | |||
| 221 | ||||
| 222 | MaximumLegalStoreInBits = 0; | |||
| 223 | for (MVT VT : MVT::all_valuetypes()) | |||
| 224 | if (EVT(VT).isSimple() && VT != MVT::Other && | |||
| 225 | TLI.isTypeLegal(EVT(VT)) && | |||
| 226 | VT.getSizeInBits() >= MaximumLegalStoreInBits) | |||
| 227 | MaximumLegalStoreInBits = VT.getSizeInBits(); | |||
| 228 | } | |||
| 229 | ||||
| 230 | void ConsiderForPruning(SDNode *N) { | |||
| 231 | // Mark this for potential pruning. | |||
| 232 | PruningList.insert(N); | |||
| 233 | } | |||
| 234 | ||||
| 235 | /// Add to the worklist making sure its instance is at the back (next to be | |||
| 236 | /// processed.) | |||
| 237 | void AddToWorklist(SDNode *N) { | |||
| 238 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 239, __PRETTY_FUNCTION__)) | |||
| 239 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 239, __PRETTY_FUNCTION__)); | |||
| 240 | ||||
| 241 | // Skip handle nodes as they can't usefully be combined and confuse the | |||
| 242 | // zero-use deletion strategy. | |||
| 243 | if (N->getOpcode() == ISD::HANDLENODE) | |||
| 244 | return; | |||
| 245 | ||||
| 246 | ConsiderForPruning(N); | |||
| 247 | ||||
| 248 | if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second) | |||
| 249 | Worklist.push_back(N); | |||
| 250 | } | |||
| 251 | ||||
| 252 | /// Remove all instances of N from the worklist. | |||
| 253 | void removeFromWorklist(SDNode *N) { | |||
| 254 | CombinedNodes.erase(N); | |||
| 255 | PruningList.remove(N); | |||
| 256 | StoreRootCountMap.erase(N); | |||
| 257 | ||||
| 258 | auto It = WorklistMap.find(N); | |||
| 259 | if (It == WorklistMap.end()) | |||
| 260 | return; // Not in the worklist. | |||
| 261 | ||||
| 262 | // Null out the entry rather than erasing it to avoid a linear operation. | |||
| 263 | Worklist[It->second] = nullptr; | |||
| 264 | WorklistMap.erase(It); | |||
| 265 | } | |||
| 266 | ||||
| 267 | void deleteAndRecombine(SDNode *N); | |||
| 268 | bool recursivelyDeleteUnusedNodes(SDNode *N); | |||
| 269 | ||||
| 270 | /// Replaces all uses of the results of one DAG node with new values. | |||
| 271 | SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, | |||
| 272 | bool AddTo = true); | |||
| 273 | ||||
| 274 | /// Replaces all uses of the results of one DAG node with new values. | |||
| 275 | SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) { | |||
| 276 | return CombineTo(N, &Res, 1, AddTo); | |||
| 277 | } | |||
| 278 | ||||
| 279 | /// Replaces all uses of the results of one DAG node with new values. | |||
| 280 | SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, | |||
| 281 | bool AddTo = true) { | |||
| 282 | SDValue To[] = { Res0, Res1 }; | |||
| 283 | return CombineTo(N, To, 2, AddTo); | |||
| 284 | } | |||
| 285 | ||||
| 286 | void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO); | |||
| 287 | ||||
| 288 | private: | |||
| 289 | unsigned MaximumLegalStoreInBits; | |||
| 290 | ||||
| 291 | /// Check the specified integer node value to see if it can be simplified or | |||
| 292 | /// if things it uses can be simplified by bit propagation. | |||
| 293 | /// If so, return true. | |||
| 294 | bool SimplifyDemandedBits(SDValue Op) { | |||
| 295 | unsigned BitWidth = Op.getScalarValueSizeInBits(); | |||
| 296 | APInt DemandedBits = APInt::getAllOnesValue(BitWidth); | |||
| 297 | return SimplifyDemandedBits(Op, DemandedBits); | |||
| 298 | } | |||
| 299 | ||||
| 300 | bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits) { | |||
| 301 | EVT VT = Op.getValueType(); | |||
| 302 | unsigned NumElts = VT.isVector() ? VT.getVectorNumElements() : 1; | |||
| 303 | APInt DemandedElts = APInt::getAllOnesValue(NumElts); | |||
| 304 | return SimplifyDemandedBits(Op, DemandedBits, DemandedElts); | |||
| 305 | } | |||
| 306 | ||||
| 307 | /// Check the specified vector node value to see if it can be simplified or | |||
| 308 | /// if things it uses can be simplified as it only uses some of the | |||
| 309 | /// elements. If so, return true. | |||
| 310 | bool SimplifyDemandedVectorElts(SDValue Op) { | |||
| 311 | unsigned NumElts = Op.getValueType().getVectorNumElements(); | |||
| 312 | APInt DemandedElts = APInt::getAllOnesValue(NumElts); | |||
| 313 | return SimplifyDemandedVectorElts(Op, DemandedElts); | |||
| 314 | } | |||
| 315 | ||||
| 316 | bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, | |||
| 317 | const APInt &DemandedElts); | |||
| 318 | bool SimplifyDemandedVectorElts(SDValue Op, const APInt &DemandedElts, | |||
| 319 | bool AssumeSingleUse = false); | |||
| 320 | ||||
| 321 | bool CombineToPreIndexedLoadStore(SDNode *N); | |||
| 322 | bool CombineToPostIndexedLoadStore(SDNode *N); | |||
| 323 | SDValue SplitIndexingFromLoad(LoadSDNode *LD); | |||
| 324 | bool SliceUpLoad(SDNode *N); | |||
| 325 | ||||
| 326 | // Scalars have size 0 to distinguish from singleton vectors. | |||
| 327 | SDValue ForwardStoreValueToDirectLoad(LoadSDNode *LD); | |||
| 328 | bool getTruncatedStoreValue(StoreSDNode *ST, SDValue &Val); | |||
| 329 | bool extendLoadedValueToExtension(LoadSDNode *LD, SDValue &Val); | |||
| 330 | ||||
| 331 | /// Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed | |||
| 332 | /// load. | |||
| 333 | /// | |||
| 334 | /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced. | |||
| 335 | /// \param InVecVT type of the input vector to EVE with bitcasts resolved. | |||
| 336 | /// \param EltNo index of the vector element to load. | |||
| 337 | /// \param OriginalLoad load that EVE came from to be replaced. | |||
| 338 | /// \returns EVE on success SDValue() on failure. | |||
| 339 | SDValue scalarizeExtractedVectorLoad(SDNode *EVE, EVT InVecVT, | |||
| 340 | SDValue EltNo, | |||
| 341 | LoadSDNode *OriginalLoad); | |||
| 342 | void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad); | |||
| 343 | SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace); | |||
| 344 | SDValue SExtPromoteOperand(SDValue Op, EVT PVT); | |||
| 345 | SDValue ZExtPromoteOperand(SDValue Op, EVT PVT); | |||
| 346 | SDValue PromoteIntBinOp(SDValue Op); | |||
| 347 | SDValue PromoteIntShiftOp(SDValue Op); | |||
| 348 | SDValue PromoteExtend(SDValue Op); | |||
| 349 | bool PromoteLoad(SDValue Op); | |||
| 350 | ||||
| 351 | /// Call the node-specific routine that knows how to fold each | |||
| 352 | /// particular type of node. If that doesn't do anything, try the | |||
| 353 | /// target-specific DAG combines. | |||
| 354 | SDValue combine(SDNode *N); | |||
| 355 | ||||
| 356 | // Visitation implementation - Implement dag node combining for different | |||
| 357 | // node types. The semantics are as follows: | |||
| 358 | // Return Value: | |||
| 359 | // SDValue.getNode() == 0 - No change was made | |||
| 360 | // SDValue.getNode() == N - N was replaced, is dead and has been handled. | |||
| 361 | // otherwise - N should be replaced by the returned Operand. | |||
| 362 | // | |||
| 363 | SDValue visitTokenFactor(SDNode *N); | |||
| 364 | SDValue visitMERGE_VALUES(SDNode *N); | |||
| 365 | SDValue visitADD(SDNode *N); | |||
| 366 | SDValue visitADDLike(SDNode *N); | |||
| 367 | SDValue visitADDLikeCommutative(SDValue N0, SDValue N1, SDNode *LocReference); | |||
| 368 | SDValue visitSUB(SDNode *N); | |||
| 369 | SDValue visitADDSAT(SDNode *N); | |||
| 370 | SDValue visitSUBSAT(SDNode *N); | |||
| 371 | SDValue visitADDC(SDNode *N); | |||
| 372 | SDValue visitADDO(SDNode *N); | |||
| 373 | SDValue visitUADDOLike(SDValue N0, SDValue N1, SDNode *N); | |||
| 374 | SDValue visitSUBC(SDNode *N); | |||
| 375 | SDValue visitSUBO(SDNode *N); | |||
| 376 | SDValue visitADDE(SDNode *N); | |||
| 377 | SDValue visitADDCARRY(SDNode *N); | |||
| 378 | SDValue visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, SDNode *N); | |||
| 379 | SDValue visitSUBE(SDNode *N); | |||
| 380 | SDValue visitSUBCARRY(SDNode *N); | |||
| 381 | SDValue visitMUL(SDNode *N); | |||
| 382 | SDValue visitMULFIX(SDNode *N); | |||
| 383 | SDValue useDivRem(SDNode *N); | |||
| 384 | SDValue visitSDIV(SDNode *N); | |||
| 385 | SDValue visitSDIVLike(SDValue N0, SDValue N1, SDNode *N); | |||
| 386 | SDValue visitUDIV(SDNode *N); | |||
| 387 | SDValue visitUDIVLike(SDValue N0, SDValue N1, SDNode *N); | |||
| 388 | SDValue visitREM(SDNode *N); | |||
| 389 | SDValue visitMULHU(SDNode *N); | |||
| 390 | SDValue visitMULHS(SDNode *N); | |||
| 391 | SDValue visitSMUL_LOHI(SDNode *N); | |||
| 392 | SDValue visitUMUL_LOHI(SDNode *N); | |||
| 393 | SDValue visitMULO(SDNode *N); | |||
| 394 | SDValue visitIMINMAX(SDNode *N); | |||
| 395 | SDValue visitAND(SDNode *N); | |||
| 396 | SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *N); | |||
| 397 | SDValue visitOR(SDNode *N); | |||
| 398 | SDValue visitORLike(SDValue N0, SDValue N1, SDNode *N); | |||
| 399 | SDValue visitXOR(SDNode *N); | |||
| 400 | SDValue SimplifyVBinOp(SDNode *N); | |||
| 401 | SDValue visitSHL(SDNode *N); | |||
| 402 | SDValue visitSRA(SDNode *N); | |||
| 403 | SDValue visitSRL(SDNode *N); | |||
| 404 | SDValue visitFunnelShift(SDNode *N); | |||
| 405 | SDValue visitRotate(SDNode *N); | |||
| 406 | SDValue visitABS(SDNode *N); | |||
| 407 | SDValue visitBSWAP(SDNode *N); | |||
| 408 | SDValue visitBITREVERSE(SDNode *N); | |||
| 409 | SDValue visitCTLZ(SDNode *N); | |||
| 410 | SDValue visitCTLZ_ZERO_UNDEF(SDNode *N); | |||
| 411 | SDValue visitCTTZ(SDNode *N); | |||
| 412 | SDValue visitCTTZ_ZERO_UNDEF(SDNode *N); | |||
| 413 | SDValue visitCTPOP(SDNode *N); | |||
| 414 | SDValue visitSELECT(SDNode *N); | |||
| 415 | SDValue visitVSELECT(SDNode *N); | |||
| 416 | SDValue visitSELECT_CC(SDNode *N); | |||
| 417 | SDValue visitSETCC(SDNode *N); | |||
| 418 | SDValue visitSETCCCARRY(SDNode *N); | |||
| 419 | SDValue visitSIGN_EXTEND(SDNode *N); | |||
| 420 | SDValue visitZERO_EXTEND(SDNode *N); | |||
| 421 | SDValue visitANY_EXTEND(SDNode *N); | |||
| 422 | SDValue visitAssertExt(SDNode *N); | |||
| 423 | SDValue visitSIGN_EXTEND_INREG(SDNode *N); | |||
| 424 | SDValue visitSIGN_EXTEND_VECTOR_INREG(SDNode *N); | |||
| 425 | SDValue visitZERO_EXTEND_VECTOR_INREG(SDNode *N); | |||
| 426 | SDValue visitTRUNCATE(SDNode *N); | |||
| 427 | SDValue visitBITCAST(SDNode *N); | |||
| 428 | SDValue visitBUILD_PAIR(SDNode *N); | |||
| 429 | SDValue visitFADD(SDNode *N); | |||
| 430 | SDValue visitFSUB(SDNode *N); | |||
| 431 | SDValue visitFMUL(SDNode *N); | |||
| 432 | SDValue visitFMA(SDNode *N); | |||
| 433 | SDValue visitFDIV(SDNode *N); | |||
| 434 | SDValue visitFREM(SDNode *N); | |||
| 435 | SDValue visitFSQRT(SDNode *N); | |||
| 436 | SDValue visitFCOPYSIGN(SDNode *N); | |||
| 437 | SDValue visitFPOW(SDNode *N); | |||
| 438 | SDValue visitSINT_TO_FP(SDNode *N); | |||
| 439 | SDValue visitUINT_TO_FP(SDNode *N); | |||
| 440 | SDValue visitFP_TO_SINT(SDNode *N); | |||
| 441 | SDValue visitFP_TO_UINT(SDNode *N); | |||
| 442 | SDValue visitFP_ROUND(SDNode *N); | |||
| 443 | SDValue visitFP_EXTEND(SDNode *N); | |||
| 444 | SDValue visitFNEG(SDNode *N); | |||
| 445 | SDValue visitFABS(SDNode *N); | |||
| 446 | SDValue visitFCEIL(SDNode *N); | |||
| 447 | SDValue visitFTRUNC(SDNode *N); | |||
| 448 | SDValue visitFFLOOR(SDNode *N); | |||
| 449 | SDValue visitFMINNUM(SDNode *N); | |||
| 450 | SDValue visitFMAXNUM(SDNode *N); | |||
| 451 | SDValue visitFMINIMUM(SDNode *N); | |||
| 452 | SDValue visitFMAXIMUM(SDNode *N); | |||
| 453 | SDValue visitBRCOND(SDNode *N); | |||
| 454 | SDValue visitBR_CC(SDNode *N); | |||
| 455 | SDValue visitLOAD(SDNode *N); | |||
| 456 | ||||
| 457 | SDValue replaceStoreChain(StoreSDNode *ST, SDValue BetterChain); | |||
| 458 | SDValue replaceStoreOfFPConstant(StoreSDNode *ST); | |||
| 459 | ||||
| 460 | SDValue visitSTORE(SDNode *N); | |||
| 461 | SDValue visitLIFETIME_END(SDNode *N); | |||
| 462 | SDValue visitINSERT_VECTOR_ELT(SDNode *N); | |||
| 463 | SDValue visitEXTRACT_VECTOR_ELT(SDNode *N); | |||
| 464 | SDValue visitBUILD_VECTOR(SDNode *N); | |||
| 465 | SDValue visitCONCAT_VECTORS(SDNode *N); | |||
| 466 | SDValue visitEXTRACT_SUBVECTOR(SDNode *N); | |||
| 467 | SDValue visitVECTOR_SHUFFLE(SDNode *N); | |||
| 468 | SDValue visitSCALAR_TO_VECTOR(SDNode *N); | |||
| 469 | SDValue visitINSERT_SUBVECTOR(SDNode *N); | |||
| 470 | SDValue visitMLOAD(SDNode *N); | |||
| 471 | SDValue visitMSTORE(SDNode *N); | |||
| 472 | SDValue visitMGATHER(SDNode *N); | |||
| 473 | SDValue visitMSCATTER(SDNode *N); | |||
| 474 | SDValue visitFP_TO_FP16(SDNode *N); | |||
| 475 | SDValue visitFP16_TO_FP(SDNode *N); | |||
| 476 | SDValue visitVECREDUCE(SDNode *N); | |||
| 477 | ||||
| 478 | SDValue visitFADDForFMACombine(SDNode *N); | |||
| 479 | SDValue visitFSUBForFMACombine(SDNode *N); | |||
| 480 | SDValue visitFMULForFMADistributiveCombine(SDNode *N); | |||
| 481 | ||||
| 482 | SDValue XformToShuffleWithZero(SDNode *N); | |||
| 483 | bool reassociationCanBreakAddressingModePattern(unsigned Opc, | |||
| 484 | const SDLoc &DL, SDValue N0, | |||
| 485 | SDValue N1); | |||
| 486 | SDValue reassociateOpsCommutative(unsigned Opc, const SDLoc &DL, SDValue N0, | |||
| 487 | SDValue N1); | |||
| 488 | SDValue reassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, | |||
| 489 | SDValue N1, SDNodeFlags Flags); | |||
| 490 | ||||
| 491 | SDValue visitShiftByConstant(SDNode *N); | |||
| 492 | ||||
| 493 | SDValue foldSelectOfConstants(SDNode *N); | |||
| 494 | SDValue foldVSelectOfConstants(SDNode *N); | |||
| 495 | SDValue foldBinOpIntoSelect(SDNode *BO); | |||
| 496 | bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS); | |||
| 497 | SDValue hoistLogicOpWithSameOpcodeHands(SDNode *N); | |||
| 498 | SDValue SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2); | |||
| 499 | SDValue SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1, | |||
| 500 | SDValue N2, SDValue N3, ISD::CondCode CC, | |||
| 501 | bool NotExtCompare = false); | |||
| 502 | SDValue convertSelectOfFPConstantsToLoadOffset( | |||
| 503 | const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, | |||
| 504 | ISD::CondCode CC); | |||
| 505 | SDValue foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, SDValue N1, | |||
| 506 | SDValue N2, SDValue N3, ISD::CondCode CC); | |||
| 507 | SDValue foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, | |||
| 508 | const SDLoc &DL); | |||
| 509 | SDValue unfoldMaskedMerge(SDNode *N); | |||
| 510 | SDValue unfoldExtremeBitClearingToShifts(SDNode *N); | |||
| 511 | SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, | |||
| 512 | const SDLoc &DL, bool foldBooleans); | |||
| 513 | SDValue rebuildSetCC(SDValue N); | |||
| 514 | ||||
| 515 | bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, | |||
| 516 | SDValue &CC) const; | |||
| 517 | bool isOneUseSetCC(SDValue N) const; | |||
| 518 | bool isCheaperToUseNegatedFPOps(SDValue X, SDValue Y); | |||
| 519 | ||||
| 520 | SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, | |||
| 521 | unsigned HiOp); | |||
| 522 | SDValue CombineConsecutiveLoads(SDNode *N, EVT VT); | |||
| 523 | SDValue CombineExtLoad(SDNode *N); | |||
| 524 | SDValue CombineZExtLogicopShiftLoad(SDNode *N); | |||
| 525 | SDValue combineRepeatedFPDivisors(SDNode *N); | |||
| 526 | SDValue combineInsertEltToShuffle(SDNode *N, unsigned InsIndex); | |||
| 527 | SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT); | |||
| 528 | SDValue BuildSDIV(SDNode *N); | |||
| 529 | SDValue BuildSDIVPow2(SDNode *N); | |||
| 530 | SDValue BuildUDIV(SDNode *N); | |||
| 531 | SDValue BuildLogBase2(SDValue V, const SDLoc &DL); | |||
| 532 | SDValue BuildDivEstimate(SDValue N, SDValue Op, SDNodeFlags Flags); | |||
| 533 | SDValue buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags); | |||
| 534 | SDValue buildSqrtEstimate(SDValue Op, SDNodeFlags Flags); | |||
| 535 | SDValue buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, bool Recip); | |||
| 536 | SDValue buildSqrtNROneConst(SDValue Arg, SDValue Est, unsigned Iterations, | |||
| 537 | SDNodeFlags Flags, bool Reciprocal); | |||
| 538 | SDValue buildSqrtNRTwoConst(SDValue Arg, SDValue Est, unsigned Iterations, | |||
| 539 | SDNodeFlags Flags, bool Reciprocal); | |||
| 540 | SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, | |||
| 541 | bool DemandHighBits = true); | |||
| 542 | SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1); | |||
| 543 | SDValue MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg, | |||
| 544 | SDValue InnerPos, SDValue InnerNeg, | |||
| 545 | unsigned PosOpcode, unsigned NegOpcode, | |||
| 546 | const SDLoc &DL); | |||
| 547 | SDValue MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL); | |||
| 548 | SDValue MatchLoadCombine(SDNode *N); | |||
| 549 | SDValue MatchStoreCombine(StoreSDNode *N); | |||
| 550 | SDValue ReduceLoadWidth(SDNode *N); | |||
| 551 | SDValue ReduceLoadOpStoreWidth(SDNode *N); | |||
| 552 | SDValue splitMergedValStore(StoreSDNode *ST); | |||
| 553 | SDValue TransformFPLoadStorePair(SDNode *N); | |||
| 554 | SDValue convertBuildVecZextToZext(SDNode *N); | |||
| 555 | SDValue reduceBuildVecExtToExtBuildVec(SDNode *N); | |||
| 556 | SDValue reduceBuildVecToShuffle(SDNode *N); | |||
| 557 | SDValue createBuildVecShuffle(const SDLoc &DL, SDNode *N, | |||
| 558 | ArrayRef<int> VectorMask, SDValue VecIn1, | |||
| 559 | SDValue VecIn2, unsigned LeftIdx, | |||
| 560 | bool DidSplitVec); | |||
| 561 | SDValue matchVSelectOpSizesWithSetCC(SDNode *Cast); | |||
| 562 | ||||
| 563 | /// Walk up chain skipping non-aliasing memory nodes, | |||
| 564 | /// looking for aliasing nodes and adding them to the Aliases vector. | |||
| 565 | void GatherAllAliases(SDNode *N, SDValue OriginalChain, | |||
| 566 | SmallVectorImpl<SDValue> &Aliases); | |||
| 567 | ||||
| 568 | /// Return true if there is any possibility that the two addresses overlap. | |||
| 569 | bool isAlias(SDNode *Op0, SDNode *Op1) const; | |||
| 570 | ||||
| 571 | /// Walk up chain skipping non-aliasing memory nodes, looking for a better | |||
| 572 | /// chain (aliasing node.) | |||
| 573 | SDValue FindBetterChain(SDNode *N, SDValue Chain); | |||
| 574 | ||||
| 575 | /// Try to replace a store and any possibly adjacent stores on | |||
| 576 | /// consecutive chains with better chains. Return true only if St is | |||
| 577 | /// replaced. | |||
| 578 | /// | |||
| 579 | /// Notice that other chains may still be replaced even if the function | |||
| 580 | /// returns false. | |||
| 581 | bool findBetterNeighborChains(StoreSDNode *St); | |||
| 582 | ||||
| 583 | // Helper for findBetterNeighborChains. Walk up store chain add additional | |||
| 584 | // chained stores that do not overlap and can be parallelized. | |||
| 585 | bool parallelizeChainedStores(StoreSDNode *St); | |||
| 586 | ||||
| 587 | /// Holds a pointer to an LSBaseSDNode as well as information on where it | |||
| 588 | /// is located in a sequence of memory operations connected by a chain. | |||
| 589 | struct MemOpLink { | |||
| 590 | // Ptr to the mem node. | |||
| 591 | LSBaseSDNode *MemNode; | |||
| 592 | ||||
| 593 | // Offset from the base ptr. | |||
| 594 | int64_t OffsetFromBase; | |||
| 595 | ||||
| 596 | MemOpLink(LSBaseSDNode *N, int64_t Offset) | |||
| 597 | : MemNode(N), OffsetFromBase(Offset) {} | |||
| 598 | }; | |||
| 599 | ||||
| 600 | /// This is a helper function for visitMUL to check the profitability | |||
| 601 | /// of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2). | |||
| 602 | /// MulNode is the original multiply, AddNode is (add x, c1), | |||
| 603 | /// and ConstNode is c2. | |||
| 604 | bool isMulAddWithConstProfitable(SDNode *MulNode, | |||
| 605 | SDValue &AddNode, | |||
| 606 | SDValue &ConstNode); | |||
| 607 | ||||
| 608 | /// This is a helper function for visitAND and visitZERO_EXTEND. Returns | |||
| 609 | /// true if the (and (load x) c) pattern matches an extload. ExtVT returns | |||
| 610 | /// the type of the loaded value to be extended. | |||
| 611 | bool isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, | |||
| 612 | EVT LoadResultTy, EVT &ExtVT); | |||
| 613 | ||||
| 614 | /// Helper function to calculate whether the given Load/Store can have its | |||
| 615 | /// width reduced to ExtVT. | |||
| 616 | bool isLegalNarrowLdSt(LSBaseSDNode *LDSTN, ISD::LoadExtType ExtType, | |||
| 617 | EVT &MemVT, unsigned ShAmt = 0); | |||
| 618 | ||||
| 619 | /// Used by BackwardsPropagateMask to find suitable loads. | |||
| 620 | bool SearchForAndLoads(SDNode *N, SmallVectorImpl<LoadSDNode*> &Loads, | |||
| 621 | SmallPtrSetImpl<SDNode*> &NodesWithConsts, | |||
| 622 | ConstantSDNode *Mask, SDNode *&NodeToMask); | |||
| 623 | /// Attempt to propagate a given AND node back to load leaves so that they | |||
| 624 | /// can be combined into narrow loads. | |||
| 625 | bool BackwardsPropagateMask(SDNode *N, SelectionDAG &DAG); | |||
| 626 | ||||
| 627 | /// Helper function for MergeConsecutiveStores which merges the | |||
| 628 | /// component store chains. | |||
| 629 | SDValue getMergeStoreChains(SmallVectorImpl<MemOpLink> &StoreNodes, | |||
| 630 | unsigned NumStores); | |||
| 631 | ||||
| 632 | /// This is a helper function for MergeConsecutiveStores. When the | |||
| 633 | /// source elements of the consecutive stores are all constants or | |||
| 634 | /// all extracted vector elements, try to merge them into one | |||
| 635 | /// larger store introducing bitcasts if necessary. \return True | |||
| 636 | /// if a merged store was created. | |||
| 637 | bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes, | |||
| 638 | EVT MemVT, unsigned NumStores, | |||
| 639 | bool IsConstantSrc, bool UseVector, | |||
| 640 | bool UseTrunc); | |||
| 641 | ||||
| 642 | /// This is a helper function for MergeConsecutiveStores. Stores | |||
| 643 | /// that potentially may be merged with St are placed in | |||
| 644 | /// StoreNodes. RootNode is a chain predecessor to all store | |||
| 645 | /// candidates. | |||
| 646 | void getStoreMergeCandidates(StoreSDNode *St, | |||
| 647 | SmallVectorImpl<MemOpLink> &StoreNodes, | |||
| 648 | SDNode *&Root); | |||
| 649 | ||||
| 650 | /// Helper function for MergeConsecutiveStores. Checks if | |||
| 651 | /// candidate stores have indirect dependency through their | |||
| 652 | /// operands. RootNode is the predecessor to all stores calculated | |||
| 653 | /// by getStoreMergeCandidates and is used to prune the dependency check. | |||
| 654 | /// \return True if safe to merge. | |||
| 655 | bool checkMergeStoreCandidatesForDependencies( | |||
| 656 | SmallVectorImpl<MemOpLink> &StoreNodes, unsigned NumStores, | |||
| 657 | SDNode *RootNode); | |||
| 658 | ||||
| 659 | /// Merge consecutive store operations into a wide store. | |||
| 660 | /// This optimization uses wide integers or vectors when possible. | |||
| 661 | /// \return number of stores that were merged into a merged store (the | |||
| 662 | /// affected nodes are stored as a prefix in \p StoreNodes). | |||
| 663 | bool MergeConsecutiveStores(StoreSDNode *St); | |||
| 664 | ||||
| 665 | /// Try to transform a truncation where C is a constant: | |||
| 666 | /// (trunc (and X, C)) -> (and (trunc X), (trunc C)) | |||
| 667 | /// | |||
| 668 | /// \p N needs to be a truncation and its first operand an AND. Other | |||
| 669 | /// requirements are checked by the function (e.g. that trunc is | |||
| 670 | /// single-use) and if missed an empty SDValue is returned. | |||
| 671 | SDValue distributeTruncateThroughAnd(SDNode *N); | |||
| 672 | ||||
| 673 | /// Helper function to determine whether the target supports operation | |||
| 674 | /// given by \p Opcode for type \p VT, that is, whether the operation | |||
| 675 | /// is legal or custom before legalizing operations, and whether is | |||
| 676 | /// legal (but not custom) after legalization. | |||
| 677 | bool hasOperation(unsigned Opcode, EVT VT) { | |||
| 678 | if (LegalOperations) | |||
| 679 | return TLI.isOperationLegal(Opcode, VT); | |||
| 680 | return TLI.isOperationLegalOrCustom(Opcode, VT); | |||
| 681 | } | |||
| 682 | ||||
| 683 | public: | |||
| 684 | /// Runs the dag combiner on all nodes in the work list | |||
| 685 | void Run(CombineLevel AtLevel); | |||
| 686 | ||||
| 687 | SelectionDAG &getDAG() const { return DAG; } | |||
| 688 | ||||
| 689 | /// Returns a type large enough to hold any valid shift amount - before type | |||
| 690 | /// legalization these can be huge. | |||
| 691 | EVT getShiftAmountTy(EVT LHSTy) { | |||
| 692 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 692, __PRETTY_FUNCTION__)); | |||
| 693 | return TLI.getShiftAmountTy(LHSTy, DAG.getDataLayout(), LegalTypes); | |||
| 694 | } | |||
| 695 | ||||
| 696 | /// This method returns true if we are running before type legalization or | |||
| 697 | /// if the specified VT is legal. | |||
| 698 | bool isTypeLegal(const EVT &VT) { | |||
| 699 | if (!LegalTypes) return true; | |||
| 700 | return TLI.isTypeLegal(VT); | |||
| 701 | } | |||
| 702 | ||||
| 703 | /// Convenience wrapper around TargetLowering::getSetCCResultType | |||
| 704 | EVT getSetCCResultType(EVT VT) const { | |||
| 705 | return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
| 706 | } | |||
| 707 | ||||
| 708 | void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, | |||
| 709 | SDValue OrigLoad, SDValue ExtLoad, | |||
| 710 | ISD::NodeType ExtType); | |||
| 711 | }; | |||
| 712 | ||||
| 713 | /// This class is a DAGUpdateListener that removes any deleted | |||
| 714 | /// nodes from the worklist. | |||
| 715 | class WorklistRemover : public SelectionDAG::DAGUpdateListener { | |||
| 716 | DAGCombiner &DC; | |||
| 717 | ||||
| 718 | public: | |||
| 719 | explicit WorklistRemover(DAGCombiner &dc) | |||
| 720 | : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} | |||
| 721 | ||||
| 722 | void NodeDeleted(SDNode *N, SDNode *E) override { | |||
| 723 | DC.removeFromWorklist(N); | |||
| 724 | } | |||
| 725 | }; | |||
| 726 | ||||
| 727 | class WorklistInserter : public SelectionDAG::DAGUpdateListener { | |||
| 728 | DAGCombiner &DC; | |||
| 729 | ||||
| 730 | public: | |||
| 731 | explicit WorklistInserter(DAGCombiner &dc) | |||
| 732 | : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} | |||
| 733 | ||||
| 734 | // FIXME: Ideally we could add N to the worklist, but this causes exponential | |||
| 735 | // compile time costs in large DAGs, e.g. Halide. | |||
| 736 | void NodeInserted(SDNode *N) override { DC.ConsiderForPruning(N); } | |||
| 737 | }; | |||
| 738 | ||||
| 739 | } // end anonymous namespace | |||
| 740 | ||||
| 741 | //===----------------------------------------------------------------------===// | |||
| 742 | // TargetLowering::DAGCombinerInfo implementation | |||
| 743 | //===----------------------------------------------------------------------===// | |||
| 744 | ||||
| 745 | void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { | |||
| 746 | ((DAGCombiner*)DC)->AddToWorklist(N); | |||
| 747 | } | |||
| 748 | ||||
| 749 | SDValue TargetLowering::DAGCombinerInfo:: | |||
| 750 | CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) { | |||
| 751 | return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo); | |||
| 752 | } | |||
| 753 | ||||
| 754 | SDValue TargetLowering::DAGCombinerInfo:: | |||
| 755 | CombineTo(SDNode *N, SDValue Res, bool AddTo) { | |||
| 756 | return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo); | |||
| 757 | } | |||
| 758 | ||||
| 759 | SDValue TargetLowering::DAGCombinerInfo:: | |||
| 760 | CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) { | |||
| 761 | return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo); | |||
| 762 | } | |||
| 763 | ||||
| 764 | bool TargetLowering::DAGCombinerInfo:: | |||
| 765 | recursivelyDeleteUnusedNodes(SDNode *N) { | |||
| 766 | return ((DAGCombiner*)DC)->recursivelyDeleteUnusedNodes(N); | |||
| 767 | } | |||
| 768 | ||||
| 769 | void TargetLowering::DAGCombinerInfo:: | |||
| 770 | CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { | |||
| 771 | return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO); | |||
| 772 | } | |||
| 773 | ||||
| 774 | //===----------------------------------------------------------------------===// | |||
| 775 | // Helper Functions | |||
| 776 | //===----------------------------------------------------------------------===// | |||
| 777 | ||||
| 778 | void DAGCombiner::deleteAndRecombine(SDNode *N) { | |||
| 779 | removeFromWorklist(N); | |||
| 780 | ||||
| 781 | // If the operands of this node are only used by the node, they will now be | |||
| 782 | // dead. Make sure to re-visit them and recursively delete dead nodes. | |||
| 783 | for (const SDValue &Op : N->ops()) | |||
| 784 | // For an operand generating multiple values, one of the values may | |||
| 785 | // become dead allowing further simplification (e.g. split index | |||
| 786 | // arithmetic from an indexed load). | |||
| 787 | if (Op->hasOneUse() || Op->getNumValues() > 1) | |||
| 788 | AddToWorklist(Op.getNode()); | |||
| 789 | ||||
| 790 | DAG.DeleteNode(N); | |||
| 791 | } | |||
| 792 | ||||
| 793 | // APInts must be the same size for most operations, this helper | |||
| 794 | // function zero extends the shorter of the pair so that they match. | |||
| 795 | // We provide an Offset so that we can create bitwidths that won't overflow. | |||
| 796 | static void zeroExtendToMatch(APInt &LHS, APInt &RHS, unsigned Offset = 0) { | |||
| 797 | unsigned Bits = Offset + std::max(LHS.getBitWidth(), RHS.getBitWidth()); | |||
| 798 | LHS = LHS.zextOrSelf(Bits); | |||
| 799 | RHS = RHS.zextOrSelf(Bits); | |||
| 800 | } | |||
| 801 | ||||
| 802 | // Return true if this node is a setcc, or is a select_cc | |||
| 803 | // that selects between the target values used for true and false, making it | |||
| 804 | // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to | |||
| 805 | // the appropriate nodes based on the type of node we are checking. This | |||
| 806 | // simplifies life a bit for the callers. | |||
| 807 | bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, | |||
| 808 | SDValue &CC) const { | |||
| 809 | if (N.getOpcode() == ISD::SETCC) { | |||
| 810 | LHS = N.getOperand(0); | |||
| 811 | RHS = N.getOperand(1); | |||
| 812 | CC = N.getOperand(2); | |||
| 813 | return true; | |||
| 814 | } | |||
| 815 | ||||
| 816 | if (N.getOpcode() != ISD::SELECT_CC || | |||
| 817 | !TLI.isConstTrueVal(N.getOperand(2).getNode()) || | |||
| 818 | !TLI.isConstFalseVal(N.getOperand(3).getNode())) | |||
| 819 | return false; | |||
| 820 | ||||
| 821 | if (TLI.getBooleanContents(N.getValueType()) == | |||
| 822 | TargetLowering::UndefinedBooleanContent) | |||
| 823 | return false; | |||
| 824 | ||||
| 825 | LHS = N.getOperand(0); | |||
| 826 | RHS = N.getOperand(1); | |||
| 827 | CC = N.getOperand(4); | |||
| 828 | return true; | |||
| 829 | } | |||
| 830 | ||||
| 831 | /// Return true if this is a SetCC-equivalent operation with only one use. | |||
| 832 | /// If this is true, it allows the users to invert the operation for free when | |||
| 833 | /// it is profitable to do so. | |||
| 834 | bool DAGCombiner::isOneUseSetCC(SDValue N) const { | |||
| 835 | SDValue N0, N1, N2; | |||
| 836 | if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse()) | |||
| 837 | return true; | |||
| 838 | return false; | |||
| 839 | } | |||
| 840 | ||||
| 841 | // Returns the SDNode if it is a constant float BuildVector | |||
| 842 | // or constant float. | |||
| 843 | static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) { | |||
| 844 | if (isa<ConstantFPSDNode>(N)) | |||
| 845 | return N.getNode(); | |||
| 846 | if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode())) | |||
| 847 | return N.getNode(); | |||
| 848 | return nullptr; | |||
| 849 | } | |||
| 850 | ||||
| 851 | // Determines if it is a constant integer or a build vector of constant | |||
| 852 | // integers (and undefs). | |||
| 853 | // Do not permit build vector implicit truncation. | |||
| 854 | static bool isConstantOrConstantVector(SDValue N, bool NoOpaques = false) { | |||
| 855 | if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N)) | |||
| 856 | return !(Const->isOpaque() && NoOpaques); | |||
| 857 | if (N.getOpcode() != ISD::BUILD_VECTOR) | |||
| 858 | return false; | |||
| 859 | unsigned BitWidth = N.getScalarValueSizeInBits(); | |||
| 860 | for (const SDValue &Op : N->op_values()) { | |||
| 861 | if (Op.isUndef()) | |||
| 862 | continue; | |||
| 863 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Op); | |||
| 864 | if (!Const || Const->getAPIntValue().getBitWidth() != BitWidth || | |||
| 865 | (Const->isOpaque() && NoOpaques)) | |||
| 866 | return false; | |||
| 867 | } | |||
| 868 | return true; | |||
| 869 | } | |||
| 870 | ||||
| 871 | // Determines if a BUILD_VECTOR is composed of all-constants possibly mixed with | |||
| 872 | // undef's. | |||
| 873 | static bool isAnyConstantBuildVector(SDValue V, bool NoOpaques = false) { | |||
| 874 | if (V.getOpcode() != ISD::BUILD_VECTOR) | |||
| 875 | return false; | |||
| 876 | return isConstantOrConstantVector(V, NoOpaques) || | |||
| 877 | ISD::isBuildVectorOfConstantFPSDNodes(V.getNode()); | |||
| 878 | } | |||
| 879 | ||||
| 880 | bool DAGCombiner::reassociationCanBreakAddressingModePattern(unsigned Opc, | |||
| 881 | const SDLoc &DL, | |||
| 882 | SDValue N0, | |||
| 883 | SDValue N1) { | |||
| 884 | // Currently this only tries to ensure we don't undo the GEP splits done by | |||
| 885 | // CodeGenPrepare when shouldConsiderGEPOffsetSplit is true. To ensure this, | |||
| 886 | // we check if the following transformation would be problematic: | |||
| 887 | // (load/store (add, (add, x, offset1), offset2)) -> | |||
| 888 | // (load/store (add, x, offset1+offset2)). | |||
| 889 | ||||
| 890 | if (Opc != ISD::ADD || N0.getOpcode() != ISD::ADD) | |||
| 891 | return false; | |||
| 892 | ||||
| 893 | if (N0.hasOneUse()) | |||
| 894 | return false; | |||
| 895 | ||||
| 896 | auto *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
| 897 | auto *C2 = dyn_cast<ConstantSDNode>(N1); | |||
| 898 | if (!C1 || !C2) | |||
| 899 | return false; | |||
| 900 | ||||
| 901 | const APInt &C1APIntVal = C1->getAPIntValue(); | |||
| 902 | const APInt &C2APIntVal = C2->getAPIntValue(); | |||
| 903 | if (C1APIntVal.getBitWidth() > 64 || C2APIntVal.getBitWidth() > 64) | |||
| 904 | return false; | |||
| 905 | ||||
| 906 | const APInt CombinedValueIntVal = C1APIntVal + C2APIntVal; | |||
| 907 | if (CombinedValueIntVal.getBitWidth() > 64) | |||
| 908 | return false; | |||
| 909 | const int64_t CombinedValue = CombinedValueIntVal.getSExtValue(); | |||
| 910 | ||||
| 911 | for (SDNode *Node : N0->uses()) { | |||
| 912 | auto LoadStore = dyn_cast<MemSDNode>(Node); | |||
| 913 | if (LoadStore) { | |||
| 914 | // Is x[offset2] already not a legal addressing mode? If so then | |||
| 915 | // reassociating the constants breaks nothing (we test offset2 because | |||
| 916 | // that's the one we hope to fold into the load or store). | |||
| 917 | TargetLoweringBase::AddrMode AM; | |||
| 918 | AM.HasBaseReg = true; | |||
| 919 | AM.BaseOffs = C2APIntVal.getSExtValue(); | |||
| 920 | EVT VT = LoadStore->getMemoryVT(); | |||
| 921 | unsigned AS = LoadStore->getAddressSpace(); | |||
| 922 | Type *AccessTy = VT.getTypeForEVT(*DAG.getContext()); | |||
| 923 | if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) | |||
| 924 | continue; | |||
| 925 | ||||
| 926 | // Would x[offset1+offset2] still be a legal addressing mode? | |||
| 927 | AM.BaseOffs = CombinedValue; | |||
| 928 | if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) | |||
| 929 | return true; | |||
| 930 | } | |||
| 931 | } | |||
| 932 | ||||
| 933 | return false; | |||
| 934 | } | |||
| 935 | ||||
| 936 | // Helper for DAGCombiner::reassociateOps. Try to reassociate an expression | |||
| 937 | // such as (Opc N0, N1), if \p N0 is the same kind of operation as \p Opc. | |||
| 938 | SDValue DAGCombiner::reassociateOpsCommutative(unsigned Opc, const SDLoc &DL, | |||
| 939 | SDValue N0, SDValue N1) { | |||
| 940 | EVT VT = N0.getValueType(); | |||
| 941 | ||||
| 942 | if (N0.getOpcode() != Opc) | |||
| 943 | return SDValue(); | |||
| 944 | ||||
| 945 | // Don't reassociate reductions. | |||
| 946 | if (N0->getFlags().hasVectorReduction()) | |||
| 947 | return SDValue(); | |||
| 948 | ||||
| 949 | if (SDNode *C1 = DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) { | |||
| 950 | if (SDNode *C2 = DAG.isConstantIntBuildVectorOrConstantInt(N1)) { | |||
| 951 | // Reassociate: (op (op x, c1), c2) -> (op x, (op c1, c2)) | |||
| 952 | if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, C1, C2)) | |||
| 953 | return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode); | |||
| 954 | return SDValue(); | |||
| 955 | } | |||
| 956 | if (N0.hasOneUse()) { | |||
| 957 | // Reassociate: (op (op x, c1), y) -> (op (op x, y), c1) | |||
| 958 | // iff (op x, c1) has one use | |||
| 959 | SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1); | |||
| 960 | if (!OpNode.getNode()) | |||
| 961 | return SDValue(); | |||
| 962 | return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1)); | |||
| 963 | } | |||
| 964 | } | |||
| 965 | return SDValue(); | |||
| 966 | } | |||
| 967 | ||||
| 968 | // Try to reassociate commutative binops. | |||
| 969 | SDValue DAGCombiner::reassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, | |||
| 970 | SDValue N1, SDNodeFlags Flags) { | |||
| 971 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 971, __PRETTY_FUNCTION__)); | |||
| 972 | // Don't reassociate reductions. | |||
| 973 | if (Flags.hasVectorReduction()) | |||
| 974 | return SDValue(); | |||
| 975 | ||||
| 976 | // Floating-point reassociation is not allowed without loose FP math. | |||
| 977 | if (N0.getValueType().isFloatingPoint() || | |||
| 978 | N1.getValueType().isFloatingPoint()) | |||
| 979 | if (!Flags.hasAllowReassociation() || !Flags.hasNoSignedZeros()) | |||
| 980 | return SDValue(); | |||
| 981 | ||||
| 982 | if (SDValue Combined = reassociateOpsCommutative(Opc, DL, N0, N1)) | |||
| 983 | return Combined; | |||
| 984 | if (SDValue Combined = reassociateOpsCommutative(Opc, DL, N1, N0)) | |||
| 985 | return Combined; | |||
| 986 | return SDValue(); | |||
| 987 | } | |||
| 988 | ||||
| 989 | SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, | |||
| 990 | bool AddTo) { | |||
| 991 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 991, __PRETTY_FUNCTION__)); | |||
| 992 | ++NodesCombined; | |||
| 993 | 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) | |||
| 994 | 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) | |||
| 995 | 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); | |||
| 996 | for (unsigned i = 0, e = NumTo; i != e; ++i) | |||
| 997 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 999, __PRETTY_FUNCTION__)) | |||
| 998 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 999, __PRETTY_FUNCTION__)) | |||
| 999 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 999, __PRETTY_FUNCTION__)); | |||
| 1000 | ||||
| 1001 | WorklistRemover DeadNodes(*this); | |||
| 1002 | DAG.ReplaceAllUsesWith(N, To); | |||
| 1003 | if (AddTo) { | |||
| 1004 | // Push the new nodes and any users onto the worklist | |||
| 1005 | for (unsigned i = 0, e = NumTo; i != e; ++i) { | |||
| 1006 | if (To[i].getNode()) { | |||
| 1007 | AddToWorklist(To[i].getNode()); | |||
| 1008 | AddUsersToWorklist(To[i].getNode()); | |||
| 1009 | } | |||
| 1010 | } | |||
| 1011 | } | |||
| 1012 | ||||
| 1013 | // Finally, if the node is now dead, remove it from the graph. The node | |||
| 1014 | // may not be dead if the replacement process recursively simplified to | |||
| 1015 | // something else needing this node. | |||
| 1016 | if (N->use_empty()) | |||
| 1017 | deleteAndRecombine(N); | |||
| 1018 | return SDValue(N, 0); | |||
| 1019 | } | |||
| 1020 | ||||
| 1021 | void DAGCombiner:: | |||
| 1022 | CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { | |||
| 1023 | // Replace all uses. If any nodes become isomorphic to other nodes and | |||
| 1024 | // are deleted, make sure to remove them from our worklist. | |||
| 1025 | WorklistRemover DeadNodes(*this); | |||
| 1026 | DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New); | |||
| 1027 | ||||
| 1028 | // Push the new node and any (possibly new) users onto the worklist. | |||
| 1029 | AddToWorklist(TLO.New.getNode()); | |||
| 1030 | AddUsersToWorklist(TLO.New.getNode()); | |||
| 1031 | ||||
| 1032 | // Finally, if the node is now dead, remove it from the graph. The node | |||
| 1033 | // may not be dead if the replacement process recursively simplified to | |||
| 1034 | // something else needing this node. | |||
| 1035 | if (TLO.Old.getNode()->use_empty()) | |||
| 1036 | deleteAndRecombine(TLO.Old.getNode()); | |||
| 1037 | } | |||
| 1038 | ||||
| 1039 | /// Check the specified integer node value to see if it can be simplified or if | |||
| 1040 | /// things it uses can be simplified by bit propagation. If so, return true. | |||
| 1041 | bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, | |||
| 1042 | const APInt &DemandedElts) { | |||
| 1043 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); | |||
| 1044 | KnownBits Known; | |||
| 1045 | if (!TLI.SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO)) | |||
| 1046 | return false; | |||
| 1047 | ||||
| 1048 | // Revisit the node. | |||
| 1049 | AddToWorklist(Op.getNode()); | |||
| 1050 | ||||
| 1051 | // Replace the old value with the new one. | |||
| 1052 | ++NodesCombined; | |||
| 1053 | 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 ) | |||
| 1054 | 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 ) | |||
| 1055 | 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 ); | |||
| 1056 | ||||
| 1057 | CommitTargetLoweringOpt(TLO); | |||
| 1058 | return true; | |||
| 1059 | } | |||
| 1060 | ||||
| 1061 | /// Check the specified vector node value to see if it can be simplified or | |||
| 1062 | /// if things it uses can be simplified as it only uses some of the elements. | |||
| 1063 | /// If so, return true. | |||
| 1064 | bool DAGCombiner::SimplifyDemandedVectorElts(SDValue Op, | |||
| 1065 | const APInt &DemandedElts, | |||
| 1066 | bool AssumeSingleUse) { | |||
| 1067 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); | |||
| 1068 | APInt KnownUndef, KnownZero; | |||
| 1069 | if (!TLI.SimplifyDemandedVectorElts(Op, DemandedElts, KnownUndef, KnownZero, | |||
| 1070 | TLO, 0, AssumeSingleUse)) | |||
| 1071 | return false; | |||
| 1072 | ||||
| 1073 | // Revisit the node. | |||
| 1074 | AddToWorklist(Op.getNode()); | |||
| 1075 | ||||
| 1076 | // Replace the old value with the new one. | |||
| 1077 | ++NodesCombined; | |||
| 1078 | 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 ) | |||
| 1079 | 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 ) | |||
| 1080 | 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 ); | |||
| 1081 | ||||
| 1082 | CommitTargetLoweringOpt(TLO); | |||
| 1083 | return true; | |||
| 1084 | } | |||
| 1085 | ||||
| 1086 | void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) { | |||
| 1087 | SDLoc DL(Load); | |||
| 1088 | EVT VT = Load->getValueType(0); | |||
| 1089 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, SDValue(ExtLoad, 0)); | |||
| 1090 | ||||
| 1091 | 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) | |||
| 1092 | 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); | |||
| 1093 | WorklistRemover DeadNodes(*this); | |||
| 1094 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc); | |||
| 1095 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1)); | |||
| 1096 | deleteAndRecombine(Load); | |||
| 1097 | AddToWorklist(Trunc.getNode()); | |||
| 1098 | } | |||
| 1099 | ||||
| 1100 | SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) { | |||
| 1101 | Replace = false; | |||
| 1102 | SDLoc DL(Op); | |||
| 1103 | if (ISD::isUNINDEXEDLoad(Op.getNode())) { | |||
| 1104 | LoadSDNode *LD = cast<LoadSDNode>(Op); | |||
| 1105 | EVT MemVT = LD->getMemoryVT(); | |||
| 1106 | ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD | |||
| 1107 | : LD->getExtensionType(); | |||
| 1108 | Replace = true; | |||
| 1109 | return DAG.getExtLoad(ExtType, DL, PVT, | |||
| 1110 | LD->getChain(), LD->getBasePtr(), | |||
| 1111 | MemVT, LD->getMemOperand()); | |||
| 1112 | } | |||
| 1113 | ||||
| 1114 | unsigned Opc = Op.getOpcode(); | |||
| 1115 | switch (Opc) { | |||
| 1116 | default: break; | |||
| 1117 | case ISD::AssertSext: | |||
| 1118 | if (SDValue Op0 = SExtPromoteOperand(Op.getOperand(0), PVT)) | |||
| 1119 | return DAG.getNode(ISD::AssertSext, DL, PVT, Op0, Op.getOperand(1)); | |||
| 1120 | break; | |||
| 1121 | case ISD::AssertZext: | |||
| 1122 | if (SDValue Op0 = ZExtPromoteOperand(Op.getOperand(0), PVT)) | |||
| 1123 | return DAG.getNode(ISD::AssertZext, DL, PVT, Op0, Op.getOperand(1)); | |||
| 1124 | break; | |||
| 1125 | case ISD::Constant: { | |||
| 1126 | unsigned ExtOpc = | |||
| 1127 | Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
| 1128 | return DAG.getNode(ExtOpc, DL, PVT, Op); | |||
| 1129 | } | |||
| 1130 | } | |||
| 1131 | ||||
| 1132 | if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT)) | |||
| 1133 | return SDValue(); | |||
| 1134 | return DAG.getNode(ISD::ANY_EXTEND, DL, PVT, Op); | |||
| 1135 | } | |||
| 1136 | ||||
| 1137 | SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) { | |||
| 1138 | if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT)) | |||
| 1139 | return SDValue(); | |||
| 1140 | EVT OldVT = Op.getValueType(); | |||
| 1141 | SDLoc DL(Op); | |||
| 1142 | bool Replace = false; | |||
| 1143 | SDValue NewOp = PromoteOperand(Op, PVT, Replace); | |||
| 1144 | if (!NewOp.getNode()) | |||
| 1145 | return SDValue(); | |||
| 1146 | AddToWorklist(NewOp.getNode()); | |||
| 1147 | ||||
| 1148 | if (Replace) | |||
| 1149 | ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); | |||
| 1150 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, NewOp.getValueType(), NewOp, | |||
| 1151 | DAG.getValueType(OldVT)); | |||
| 1152 | } | |||
| 1153 | ||||
| 1154 | SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) { | |||
| 1155 | EVT OldVT = Op.getValueType(); | |||
| 1156 | SDLoc DL(Op); | |||
| 1157 | bool Replace = false; | |||
| 1158 | SDValue NewOp = PromoteOperand(Op, PVT, Replace); | |||
| 1159 | if (!NewOp.getNode()) | |||
| 1160 | return SDValue(); | |||
| 1161 | AddToWorklist(NewOp.getNode()); | |||
| 1162 | ||||
| 1163 | if (Replace) | |||
| 1164 | ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); | |||
| 1165 | return DAG.getZeroExtendInReg(NewOp, DL, OldVT); | |||
| 1166 | } | |||
| 1167 | ||||
| 1168 | /// Promote the specified integer binary operation if the target indicates it is | |||
| 1169 | /// beneficial. e.g. On x86, it's usually better to promote i16 operations to | |||
| 1170 | /// i32 since i16 instructions are longer. | |||
| 1171 | SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) { | |||
| 1172 | if (!LegalOperations) | |||
| 1173 | return SDValue(); | |||
| 1174 | ||||
| 1175 | EVT VT = Op.getValueType(); | |||
| 1176 | if (VT.isVector() || !VT.isInteger()) | |||
| 1177 | return SDValue(); | |||
| 1178 | ||||
| 1179 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
| 1180 | // promoting it. | |||
| 1181 | unsigned Opc = Op.getOpcode(); | |||
| 1182 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
| 1183 | return SDValue(); | |||
| 1184 | ||||
| 1185 | EVT PVT = VT; | |||
| 1186 | // Consult target whether it is a good idea to promote this operation and | |||
| 1187 | // what's the right type to promote it to. | |||
| 1188 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
| 1189 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1189, __PRETTY_FUNCTION__)); | |||
| 1190 | ||||
| 1191 | LLVM_DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); | |||
| 1192 | ||||
| 1193 | bool Replace0 = false; | |||
| 1194 | SDValue N0 = Op.getOperand(0); | |||
| 1195 | SDValue NN0 = PromoteOperand(N0, PVT, Replace0); | |||
| 1196 | ||||
| 1197 | bool Replace1 = false; | |||
| 1198 | SDValue N1 = Op.getOperand(1); | |||
| 1199 | SDValue NN1 = PromoteOperand(N1, PVT, Replace1); | |||
| 1200 | SDLoc DL(Op); | |||
| 1201 | ||||
| 1202 | SDValue RV = | |||
| 1203 | DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, NN0, NN1)); | |||
| 1204 | ||||
| 1205 | // We are always replacing N0/N1's use in N and only need | |||
| 1206 | // additional replacements if there are additional uses. | |||
| 1207 | Replace0 &= !N0->hasOneUse(); | |||
| 1208 | Replace1 &= (N0 != N1) && !N1->hasOneUse(); | |||
| 1209 | ||||
| 1210 | // Combine Op here so it is preserved past replacements. | |||
| 1211 | CombineTo(Op.getNode(), RV); | |||
| 1212 | ||||
| 1213 | // If operands have a use ordering, make sure we deal with | |||
| 1214 | // predecessor first. | |||
| 1215 | if (Replace0 && Replace1 && N0.getNode()->isPredecessorOf(N1.getNode())) { | |||
| 1216 | std::swap(N0, N1); | |||
| 1217 | std::swap(NN0, NN1); | |||
| 1218 | } | |||
| 1219 | ||||
| 1220 | if (Replace0) { | |||
| 1221 | AddToWorklist(NN0.getNode()); | |||
| 1222 | ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode()); | |||
| 1223 | } | |||
| 1224 | if (Replace1) { | |||
| 1225 | AddToWorklist(NN1.getNode()); | |||
| 1226 | ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode()); | |||
| 1227 | } | |||
| 1228 | return Op; | |||
| 1229 | } | |||
| 1230 | return SDValue(); | |||
| 1231 | } | |||
| 1232 | ||||
| 1233 | /// Promote the specified integer shift operation if the target indicates it is | |||
| 1234 | /// beneficial. e.g. On x86, it's usually better to promote i16 operations to | |||
| 1235 | /// i32 since i16 instructions are longer. | |||
| 1236 | SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) { | |||
| 1237 | if (!LegalOperations) | |||
| 1238 | return SDValue(); | |||
| 1239 | ||||
| 1240 | EVT VT = Op.getValueType(); | |||
| 1241 | if (VT.isVector() || !VT.isInteger()) | |||
| 1242 | return SDValue(); | |||
| 1243 | ||||
| 1244 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
| 1245 | // promoting it. | |||
| 1246 | unsigned Opc = Op.getOpcode(); | |||
| 1247 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
| 1248 | return SDValue(); | |||
| 1249 | ||||
| 1250 | EVT PVT = VT; | |||
| 1251 | // Consult target whether it is a good idea to promote this operation and | |||
| 1252 | // what's the right type to promote it to. | |||
| 1253 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
| 1254 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1254, __PRETTY_FUNCTION__)); | |||
| 1255 | ||||
| 1256 | LLVM_DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); | |||
| 1257 | ||||
| 1258 | bool Replace = false; | |||
| 1259 | SDValue N0 = Op.getOperand(0); | |||
| 1260 | SDValue N1 = Op.getOperand(1); | |||
| 1261 | if (Opc == ISD::SRA) | |||
| 1262 | N0 = SExtPromoteOperand(N0, PVT); | |||
| 1263 | else if (Opc == ISD::SRL) | |||
| 1264 | N0 = ZExtPromoteOperand(N0, PVT); | |||
| 1265 | else | |||
| 1266 | N0 = PromoteOperand(N0, PVT, Replace); | |||
| 1267 | ||||
| 1268 | if (!N0.getNode()) | |||
| 1269 | return SDValue(); | |||
| 1270 | ||||
| 1271 | SDLoc DL(Op); | |||
| 1272 | SDValue RV = | |||
| 1273 | DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1)); | |||
| 1274 | ||||
| 1275 | if (Replace) | |||
| 1276 | ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode()); | |||
| 1277 | ||||
| 1278 | // Deal with Op being deleted. | |||
| 1279 | if (Op && Op.getOpcode() != ISD::DELETED_NODE) | |||
| 1280 | return RV; | |||
| 1281 | } | |||
| 1282 | return SDValue(); | |||
| 1283 | } | |||
| 1284 | ||||
| 1285 | SDValue DAGCombiner::PromoteExtend(SDValue Op) { | |||
| 1286 | if (!LegalOperations) | |||
| 1287 | return SDValue(); | |||
| 1288 | ||||
| 1289 | EVT VT = Op.getValueType(); | |||
| 1290 | if (VT.isVector() || !VT.isInteger()) | |||
| 1291 | return SDValue(); | |||
| 1292 | ||||
| 1293 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
| 1294 | // promoting it. | |||
| 1295 | unsigned Opc = Op.getOpcode(); | |||
| 1296 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
| 1297 | return SDValue(); | |||
| 1298 | ||||
| 1299 | EVT PVT = VT; | |||
| 1300 | // Consult target whether it is a good idea to promote this operation and | |||
| 1301 | // what's the right type to promote it to. | |||
| 1302 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
| 1303 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1303, __PRETTY_FUNCTION__)); | |||
| 1304 | // fold (aext (aext x)) -> (aext x) | |||
| 1305 | // fold (aext (zext x)) -> (zext x) | |||
| 1306 | // fold (aext (sext x)) -> (sext x) | |||
| 1307 | LLVM_DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); | |||
| 1308 | return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0)); | |||
| 1309 | } | |||
| 1310 | return SDValue(); | |||
| 1311 | } | |||
| 1312 | ||||
| 1313 | bool DAGCombiner::PromoteLoad(SDValue Op) { | |||
| 1314 | if (!LegalOperations) | |||
| 1315 | return false; | |||
| 1316 | ||||
| 1317 | if (!ISD::isUNINDEXEDLoad(Op.getNode())) | |||
| 1318 | return false; | |||
| 1319 | ||||
| 1320 | EVT VT = Op.getValueType(); | |||
| 1321 | if (VT.isVector() || !VT.isInteger()) | |||
| 1322 | return false; | |||
| 1323 | ||||
| 1324 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
| 1325 | // promoting it. | |||
| 1326 | unsigned Opc = Op.getOpcode(); | |||
| 1327 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
| 1328 | return false; | |||
| 1329 | ||||
| 1330 | EVT PVT = VT; | |||
| 1331 | // Consult target whether it is a good idea to promote this operation and | |||
| 1332 | // what's the right type to promote it to. | |||
| 1333 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
| 1334 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1334, __PRETTY_FUNCTION__)); | |||
| 1335 | ||||
| 1336 | SDLoc DL(Op); | |||
| 1337 | SDNode *N = Op.getNode(); | |||
| 1338 | LoadSDNode *LD = cast<LoadSDNode>(N); | |||
| 1339 | EVT MemVT = LD->getMemoryVT(); | |||
| 1340 | ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD | |||
| 1341 | : LD->getExtensionType(); | |||
| 1342 | SDValue NewLD = DAG.getExtLoad(ExtType, DL, PVT, | |||
| 1343 | LD->getChain(), LD->getBasePtr(), | |||
| 1344 | MemVT, LD->getMemOperand()); | |||
| 1345 | SDValue Result = DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD); | |||
| 1346 | ||||
| 1347 | 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) | |||
| 1348 | 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); | |||
| 1349 | WorklistRemover DeadNodes(*this); | |||
| 1350 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); | |||
| 1351 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1)); | |||
| 1352 | deleteAndRecombine(N); | |||
| 1353 | AddToWorklist(Result.getNode()); | |||
| 1354 | return true; | |||
| 1355 | } | |||
| 1356 | return false; | |||
| 1357 | } | |||
| 1358 | ||||
| 1359 | /// Recursively delete a node which has no uses and any operands for | |||
| 1360 | /// which it is the only use. | |||
| 1361 | /// | |||
| 1362 | /// Note that this both deletes the nodes and removes them from the worklist. | |||
| 1363 | /// It also adds any nodes who have had a user deleted to the worklist as they | |||
| 1364 | /// may now have only one use and subject to other combines. | |||
| 1365 | bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) { | |||
| 1366 | if (!N->use_empty()) | |||
| 1367 | return false; | |||
| 1368 | ||||
| 1369 | SmallSetVector<SDNode *, 16> Nodes; | |||
| 1370 | Nodes.insert(N); | |||
| 1371 | do { | |||
| 1372 | N = Nodes.pop_back_val(); | |||
| 1373 | if (!N) | |||
| 1374 | continue; | |||
| 1375 | ||||
| 1376 | if (N->use_empty()) { | |||
| 1377 | for (const SDValue &ChildN : N->op_values()) | |||
| 1378 | Nodes.insert(ChildN.getNode()); | |||
| 1379 | ||||
| 1380 | removeFromWorklist(N); | |||
| 1381 | DAG.DeleteNode(N); | |||
| 1382 | } else { | |||
| 1383 | AddToWorklist(N); | |||
| 1384 | } | |||
| 1385 | } while (!Nodes.empty()); | |||
| 1386 | return true; | |||
| 1387 | } | |||
| 1388 | ||||
| 1389 | //===----------------------------------------------------------------------===// | |||
| 1390 | // Main DAG Combiner implementation | |||
| 1391 | //===----------------------------------------------------------------------===// | |||
| 1392 | ||||
| 1393 | void DAGCombiner::Run(CombineLevel AtLevel) { | |||
| 1394 | // set the instance variables, so that the various visit routines may use it. | |||
| 1395 | Level = AtLevel; | |||
| 1396 | LegalOperations = Level >= AfterLegalizeVectorOps; | |||
| 1397 | LegalTypes = Level >= AfterLegalizeTypes; | |||
| 1398 | ||||
| 1399 | WorklistInserter AddNodes(*this); | |||
| 1400 | ||||
| 1401 | // Add all the dag nodes to the worklist. | |||
| 1402 | for (SDNode &Node : DAG.allnodes()) | |||
| 1403 | AddToWorklist(&Node); | |||
| 1404 | ||||
| 1405 | // Create a dummy node (which is not added to allnodes), that adds a reference | |||
| 1406 | // to the root node, preventing it from being deleted, and tracking any | |||
| 1407 | // changes of the root. | |||
| 1408 | HandleSDNode Dummy(DAG.getRoot()); | |||
| 1409 | ||||
| 1410 | // While we have a valid worklist entry node, try to combine it. | |||
| 1411 | while (SDNode *N = getNextWorklistEntry()) { | |||
| 1412 | // If N has no uses, it is dead. Make sure to revisit all N's operands once | |||
| 1413 | // N is deleted from the DAG, since they too may now be dead or may have a | |||
| 1414 | // reduced number of uses, allowing other xforms. | |||
| 1415 | if (recursivelyDeleteUnusedNodes(N)) | |||
| 1416 | continue; | |||
| 1417 | ||||
| 1418 | WorklistRemover DeadNodes(*this); | |||
| 1419 | ||||
| 1420 | // If this combine is running after legalizing the DAG, re-legalize any | |||
| 1421 | // nodes pulled off the worklist. | |||
| 1422 | if (Level == AfterLegalizeDAG) { | |||
| 1423 | SmallSetVector<SDNode *, 16> UpdatedNodes; | |||
| 1424 | bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes); | |||
| 1425 | ||||
| 1426 | for (SDNode *LN : UpdatedNodes) { | |||
| 1427 | AddUsersToWorklist(LN); | |||
| 1428 | AddToWorklist(LN); | |||
| 1429 | } | |||
| 1430 | if (!NIsValid) | |||
| 1431 | continue; | |||
| 1432 | } | |||
| 1433 | ||||
| 1434 | LLVM_DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nCombining: "; N->dump (&DAG); } } while (false); | |||
| 1435 | ||||
| 1436 | // Add any operands of the new node which have not yet been combined to the | |||
| 1437 | // worklist as well. Because the worklist uniques things already, this | |||
| 1438 | // won't repeatedly process the same operand. | |||
| 1439 | CombinedNodes.insert(N); | |||
| 1440 | for (const SDValue &ChildN : N->op_values()) | |||
| 1441 | if (!CombinedNodes.count(ChildN.getNode())) | |||
| 1442 | AddToWorklist(ChildN.getNode()); | |||
| 1443 | ||||
| 1444 | SDValue RV = combine(N); | |||
| 1445 | ||||
| 1446 | if (!RV.getNode()) | |||
| 1447 | continue; | |||
| 1448 | ||||
| 1449 | ++NodesCombined; | |||
| 1450 | ||||
| 1451 | // If we get back the same node we passed in, rather than a new node or | |||
| 1452 | // zero, we know that the node must have defined multiple values and | |||
| 1453 | // CombineTo was used. Since CombineTo takes care of the worklist | |||
| 1454 | // mechanics for us, we have no work to do in this case. | |||
| 1455 | if (RV.getNode() == N) | |||
| 1456 | continue; | |||
| 1457 | ||||
| 1458 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1460, __PRETTY_FUNCTION__)) | |||
| 1459 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1460, __PRETTY_FUNCTION__)) | |||
| 1460 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1460, __PRETTY_FUNCTION__)); | |||
| 1461 | ||||
| 1462 | LLVM_DEBUG(dbgs() << " ... into: "; RV.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << " ... into: "; RV.getNode() ->dump(&DAG); } } while (false); | |||
| 1463 | ||||
| 1464 | if (N->getNumValues() == RV.getNode()->getNumValues()) | |||
| 1465 | DAG.ReplaceAllUsesWith(N, RV.getNode()); | |||
| 1466 | else { | |||
| 1467 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1468, __PRETTY_FUNCTION__)) | |||
| 1468 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1468, __PRETTY_FUNCTION__)); | |||
| 1469 | DAG.ReplaceAllUsesWith(N, &RV); | |||
| 1470 | } | |||
| 1471 | ||||
| 1472 | // Push the new node and any users onto the worklist | |||
| 1473 | AddToWorklist(RV.getNode()); | |||
| 1474 | AddUsersToWorklist(RV.getNode()); | |||
| 1475 | ||||
| 1476 | // Finally, if the node is now dead, remove it from the graph. The node | |||
| 1477 | // may not be dead if the replacement process recursively simplified to | |||
| 1478 | // something else needing this node. This will also take care of adding any | |||
| 1479 | // operands which have lost a user to the worklist. | |||
| 1480 | recursivelyDeleteUnusedNodes(N); | |||
| 1481 | } | |||
| 1482 | ||||
| 1483 | // If the root changed (e.g. it was a dead load, update the root). | |||
| 1484 | DAG.setRoot(Dummy.getValue()); | |||
| 1485 | DAG.RemoveDeadNodes(); | |||
| 1486 | } | |||
| 1487 | ||||
| 1488 | SDValue DAGCombiner::visit(SDNode *N) { | |||
| 1489 | switch (N->getOpcode()) { | |||
| 1490 | default: break; | |||
| 1491 | case ISD::TokenFactor: return visitTokenFactor(N); | |||
| 1492 | case ISD::MERGE_VALUES: return visitMERGE_VALUES(N); | |||
| 1493 | case ISD::ADD: return visitADD(N); | |||
| 1494 | case ISD::SUB: return visitSUB(N); | |||
| 1495 | case ISD::SADDSAT: | |||
| 1496 | case ISD::UADDSAT: return visitADDSAT(N); | |||
| 1497 | case ISD::SSUBSAT: | |||
| 1498 | case ISD::USUBSAT: return visitSUBSAT(N); | |||
| 1499 | case ISD::ADDC: return visitADDC(N); | |||
| 1500 | case ISD::SADDO: | |||
| 1501 | case ISD::UADDO: return visitADDO(N); | |||
| 1502 | case ISD::SUBC: return visitSUBC(N); | |||
| 1503 | case ISD::SSUBO: | |||
| 1504 | case ISD::USUBO: return visitSUBO(N); | |||
| 1505 | case ISD::ADDE: return visitADDE(N); | |||
| 1506 | case ISD::ADDCARRY: return visitADDCARRY(N); | |||
| 1507 | case ISD::SUBE: return visitSUBE(N); | |||
| 1508 | case ISD::SUBCARRY: return visitSUBCARRY(N); | |||
| 1509 | case ISD::SMULFIX: | |||
| 1510 | case ISD::SMULFIXSAT: | |||
| 1511 | case ISD::UMULFIX: | |||
| 1512 | case ISD::UMULFIXSAT: return visitMULFIX(N); | |||
| 1513 | case ISD::MUL: return visitMUL(N); | |||
| 1514 | case ISD::SDIV: return visitSDIV(N); | |||
| 1515 | case ISD::UDIV: return visitUDIV(N); | |||
| 1516 | case ISD::SREM: | |||
| 1517 | case ISD::UREM: return visitREM(N); | |||
| 1518 | case ISD::MULHU: return visitMULHU(N); | |||
| 1519 | case ISD::MULHS: return visitMULHS(N); | |||
| 1520 | case ISD::SMUL_LOHI: return visitSMUL_LOHI(N); | |||
| 1521 | case ISD::UMUL_LOHI: return visitUMUL_LOHI(N); | |||
| 1522 | case ISD::SMULO: | |||
| 1523 | case ISD::UMULO: return visitMULO(N); | |||
| 1524 | case ISD::SMIN: | |||
| 1525 | case ISD::SMAX: | |||
| 1526 | case ISD::UMIN: | |||
| 1527 | case ISD::UMAX: return visitIMINMAX(N); | |||
| 1528 | case ISD::AND: return visitAND(N); | |||
| 1529 | case ISD::OR: return visitOR(N); | |||
| 1530 | case ISD::XOR: return visitXOR(N); | |||
| 1531 | case ISD::SHL: return visitSHL(N); | |||
| 1532 | case ISD::SRA: return visitSRA(N); | |||
| 1533 | case ISD::SRL: return visitSRL(N); | |||
| 1534 | case ISD::ROTR: | |||
| 1535 | case ISD::ROTL: return visitRotate(N); | |||
| 1536 | case ISD::FSHL: | |||
| 1537 | case ISD::FSHR: return visitFunnelShift(N); | |||
| 1538 | case ISD::ABS: return visitABS(N); | |||
| 1539 | case ISD::BSWAP: return visitBSWAP(N); | |||
| 1540 | case ISD::BITREVERSE: return visitBITREVERSE(N); | |||
| 1541 | case ISD::CTLZ: return visitCTLZ(N); | |||
| 1542 | case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N); | |||
| 1543 | case ISD::CTTZ: return visitCTTZ(N); | |||
| 1544 | case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N); | |||
| 1545 | case ISD::CTPOP: return visitCTPOP(N); | |||
| 1546 | case ISD::SELECT: return visitSELECT(N); | |||
| 1547 | case ISD::VSELECT: return visitVSELECT(N); | |||
| 1548 | case ISD::SELECT_CC: return visitSELECT_CC(N); | |||
| 1549 | case ISD::SETCC: return visitSETCC(N); | |||
| 1550 | case ISD::SETCCCARRY: return visitSETCCCARRY(N); | |||
| 1551 | case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); | |||
| 1552 | case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); | |||
| 1553 | case ISD::ANY_EXTEND: return visitANY_EXTEND(N); | |||
| 1554 | case ISD::AssertSext: | |||
| 1555 | case ISD::AssertZext: return visitAssertExt(N); | |||
| 1556 | case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); | |||
| 1557 | case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N); | |||
| 1558 | case ISD::ZERO_EXTEND_VECTOR_INREG: return visitZERO_EXTEND_VECTOR_INREG(N); | |||
| 1559 | case ISD::TRUNCATE: return visitTRUNCATE(N); | |||
| 1560 | case ISD::BITCAST: return visitBITCAST(N); | |||
| 1561 | case ISD::BUILD_PAIR: return visitBUILD_PAIR(N); | |||
| 1562 | case ISD::FADD: return visitFADD(N); | |||
| 1563 | case ISD::FSUB: return visitFSUB(N); | |||
| 1564 | case ISD::FMUL: return visitFMUL(N); | |||
| 1565 | case ISD::FMA: return visitFMA(N); | |||
| 1566 | case ISD::FDIV: return visitFDIV(N); | |||
| 1567 | case ISD::FREM: return visitFREM(N); | |||
| 1568 | case ISD::FSQRT: return visitFSQRT(N); | |||
| 1569 | case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); | |||
| 1570 | case ISD::FPOW: return visitFPOW(N); | |||
| 1571 | case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); | |||
| 1572 | case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); | |||
| 1573 | case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); | |||
| 1574 | case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); | |||
| 1575 | case ISD::FP_ROUND: return visitFP_ROUND(N); | |||
| 1576 | case ISD::FP_EXTEND: return visitFP_EXTEND(N); | |||
| 1577 | case ISD::FNEG: return visitFNEG(N); | |||
| 1578 | case ISD::FABS: return visitFABS(N); | |||
| 1579 | case ISD::FFLOOR: return visitFFLOOR(N); | |||
| 1580 | case ISD::FMINNUM: return visitFMINNUM(N); | |||
| 1581 | case ISD::FMAXNUM: return visitFMAXNUM(N); | |||
| 1582 | case ISD::FMINIMUM: return visitFMINIMUM(N); | |||
| 1583 | case ISD::FMAXIMUM: return visitFMAXIMUM(N); | |||
| 1584 | case ISD::FCEIL: return visitFCEIL(N); | |||
| 1585 | case ISD::FTRUNC: return visitFTRUNC(N); | |||
| 1586 | case ISD::BRCOND: return visitBRCOND(N); | |||
| 1587 | case ISD::BR_CC: return visitBR_CC(N); | |||
| 1588 | case ISD::LOAD: return visitLOAD(N); | |||
| 1589 | case ISD::STORE: return visitSTORE(N); | |||
| 1590 | case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); | |||
| 1591 | case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N); | |||
| 1592 | case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N); | |||
| 1593 | case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N); | |||
| 1594 | case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N); | |||
| 1595 | case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); | |||
| 1596 | case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N); | |||
| 1597 | case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N); | |||
| 1598 | case ISD::MGATHER: return visitMGATHER(N); | |||
| 1599 | case ISD::MLOAD: return visitMLOAD(N); | |||
| 1600 | case ISD::MSCATTER: return visitMSCATTER(N); | |||
| 1601 | case ISD::MSTORE: return visitMSTORE(N); | |||
| 1602 | case ISD::LIFETIME_END: return visitLIFETIME_END(N); | |||
| 1603 | case ISD::FP_TO_FP16: return visitFP_TO_FP16(N); | |||
| 1604 | case ISD::FP16_TO_FP: return visitFP16_TO_FP(N); | |||
| 1605 | case ISD::VECREDUCE_FADD: | |||
| 1606 | case ISD::VECREDUCE_FMUL: | |||
| 1607 | case ISD::VECREDUCE_ADD: | |||
| 1608 | case ISD::VECREDUCE_MUL: | |||
| 1609 | case ISD::VECREDUCE_AND: | |||
| 1610 | case ISD::VECREDUCE_OR: | |||
| 1611 | case ISD::VECREDUCE_XOR: | |||
| 1612 | case ISD::VECREDUCE_SMAX: | |||
| 1613 | case ISD::VECREDUCE_SMIN: | |||
| 1614 | case ISD::VECREDUCE_UMAX: | |||
| 1615 | case ISD::VECREDUCE_UMIN: | |||
| 1616 | case ISD::VECREDUCE_FMAX: | |||
| 1617 | case ISD::VECREDUCE_FMIN: return visitVECREDUCE(N); | |||
| 1618 | } | |||
| 1619 | return SDValue(); | |||
| 1620 | } | |||
| 1621 | ||||
| 1622 | SDValue DAGCombiner::combine(SDNode *N) { | |||
| 1623 | SDValue RV = visit(N); | |||
| 1624 | ||||
| 1625 | // If nothing happened, try a target-specific DAG combine. | |||
| 1626 | if (!RV.getNode()) { | |||
| 1627 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1628, __PRETTY_FUNCTION__)) | |||
| 1628 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1628, __PRETTY_FUNCTION__)); | |||
| 1629 | ||||
| 1630 | if (N->getOpcode() >= ISD::BUILTIN_OP_END || | |||
| 1631 | TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) { | |||
| 1632 | ||||
| 1633 | // Expose the DAG combiner to the target combiner impls. | |||
| 1634 | TargetLowering::DAGCombinerInfo | |||
| 1635 | DagCombineInfo(DAG, Level, false, this); | |||
| 1636 | ||||
| 1637 | RV = TLI.PerformDAGCombine(N, DagCombineInfo); | |||
| 1638 | } | |||
| 1639 | } | |||
| 1640 | ||||
| 1641 | // If nothing happened still, try promoting the operation. | |||
| 1642 | if (!RV.getNode()) { | |||
| 1643 | switch (N->getOpcode()) { | |||
| 1644 | default: break; | |||
| 1645 | case ISD::ADD: | |||
| 1646 | case ISD::SUB: | |||
| 1647 | case ISD::MUL: | |||
| 1648 | case ISD::AND: | |||
| 1649 | case ISD::OR: | |||
| 1650 | case ISD::XOR: | |||
| 1651 | RV = PromoteIntBinOp(SDValue(N, 0)); | |||
| 1652 | break; | |||
| 1653 | case ISD::SHL: | |||
| 1654 | case ISD::SRA: | |||
| 1655 | case ISD::SRL: | |||
| 1656 | RV = PromoteIntShiftOp(SDValue(N, 0)); | |||
| 1657 | break; | |||
| 1658 | case ISD::SIGN_EXTEND: | |||
| 1659 | case ISD::ZERO_EXTEND: | |||
| 1660 | case ISD::ANY_EXTEND: | |||
| 1661 | RV = PromoteExtend(SDValue(N, 0)); | |||
| 1662 | break; | |||
| 1663 | case ISD::LOAD: | |||
| 1664 | if (PromoteLoad(SDValue(N, 0))) | |||
| 1665 | RV = SDValue(N, 0); | |||
| 1666 | break; | |||
| 1667 | } | |||
| 1668 | } | |||
| 1669 | ||||
| 1670 | // If N is a commutative binary node, try to eliminate it if the commuted | |||
| 1671 | // version is already present in the DAG. | |||
| 1672 | if (!RV.getNode() && TLI.isCommutativeBinOp(N->getOpcode()) && | |||
| 1673 | N->getNumValues() == 1) { | |||
| 1674 | SDValue N0 = N->getOperand(0); | |||
| 1675 | SDValue N1 = N->getOperand(1); | |||
| 1676 | ||||
| 1677 | // Constant operands are canonicalized to RHS. | |||
| 1678 | if (N0 != N1 && (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1))) { | |||
| 1679 | SDValue Ops[] = {N1, N0}; | |||
| 1680 | SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops, | |||
| 1681 | N->getFlags()); | |||
| 1682 | if (CSENode) | |||
| 1683 | return SDValue(CSENode, 0); | |||
| 1684 | } | |||
| 1685 | } | |||
| 1686 | ||||
| 1687 | return RV; | |||
| 1688 | } | |||
| 1689 | ||||
| 1690 | /// Given a node, return its input chain if it has one, otherwise return a null | |||
| 1691 | /// sd operand. | |||
| 1692 | static SDValue getInputChainForNode(SDNode *N) { | |||
| 1693 | if (unsigned NumOps = N->getNumOperands()) { | |||
| 1694 | if (N->getOperand(0).getValueType() == MVT::Other) | |||
| 1695 | return N->getOperand(0); | |||
| 1696 | if (N->getOperand(NumOps-1).getValueType() == MVT::Other) | |||
| 1697 | return N->getOperand(NumOps-1); | |||
| 1698 | for (unsigned i = 1; i < NumOps-1; ++i) | |||
| 1699 | if (N->getOperand(i).getValueType() == MVT::Other) | |||
| 1700 | return N->getOperand(i); | |||
| 1701 | } | |||
| 1702 | return SDValue(); | |||
| 1703 | } | |||
| 1704 | ||||
| 1705 | SDValue DAGCombiner::visitTokenFactor(SDNode *N) { | |||
| 1706 | // If N has two operands, where one has an input chain equal to the other, | |||
| 1707 | // the 'other' chain is redundant. | |||
| 1708 | if (N->getNumOperands() == 2) { | |||
| 1709 | if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1)) | |||
| 1710 | return N->getOperand(0); | |||
| 1711 | if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0)) | |||
| 1712 | return N->getOperand(1); | |||
| 1713 | } | |||
| 1714 | ||||
| 1715 | // Don't simplify token factors if optnone. | |||
| 1716 | if (OptLevel == CodeGenOpt::None) | |||
| 1717 | return SDValue(); | |||
| 1718 | ||||
| 1719 | // If the sole user is a token factor, we should make sure we have a | |||
| 1720 | // chance to merge them together. This prevents TF chains from inhibiting | |||
| 1721 | // optimizations. | |||
| 1722 | if (N->hasOneUse() && N->use_begin()->getOpcode() == ISD::TokenFactor) | |||
| 1723 | AddToWorklist(*(N->use_begin())); | |||
| 1724 | ||||
| 1725 | SmallVector<SDNode *, 8> TFs; // List of token factors to visit. | |||
| 1726 | SmallVector<SDValue, 8> Ops; // Ops for replacing token factor. | |||
| 1727 | SmallPtrSet<SDNode*, 16> SeenOps; | |||
| 1728 | bool Changed = false; // If we should replace this token factor. | |||
| 1729 | ||||
| 1730 | // Start out with this token factor. | |||
| 1731 | TFs.push_back(N); | |||
| 1732 | ||||
| 1733 | // Iterate through token factors. The TFs grows when new token factors are | |||
| 1734 | // encountered. | |||
| 1735 | for (unsigned i = 0; i < TFs.size(); ++i) { | |||
| 1736 | // Limit number of nodes to inline, to avoid quadratic compile times. | |||
| 1737 | // We have to add the outstanding Token Factors to Ops, otherwise we might | |||
| 1738 | // drop Ops from the resulting Token Factors. | |||
| 1739 | if (Ops.size() > TokenFactorInlineLimit) { | |||
| 1740 | for (unsigned j = i; j < TFs.size(); j++) | |||
| 1741 | Ops.emplace_back(TFs[j], 0); | |||
| 1742 | // Drop unprocessed Token Factors from TFs, so we do not add them to the | |||
| 1743 | // combiner worklist later. | |||
| 1744 | TFs.resize(i); | |||
| 1745 | break; | |||
| 1746 | } | |||
| 1747 | ||||
| 1748 | SDNode *TF = TFs[i]; | |||
| 1749 | // Check each of the operands. | |||
| 1750 | for (const SDValue &Op : TF->op_values()) { | |||
| 1751 | switch (Op.getOpcode()) { | |||
| 1752 | case ISD::EntryToken: | |||
| 1753 | // Entry tokens don't need to be added to the list. They are | |||
| 1754 | // redundant. | |||
| 1755 | Changed = true; | |||
| 1756 | break; | |||
| 1757 | ||||
| 1758 | case ISD::TokenFactor: | |||
| 1759 | if (Op.hasOneUse() && !is_contained(TFs, Op.getNode())) { | |||
| 1760 | // Queue up for processing. | |||
| 1761 | TFs.push_back(Op.getNode()); | |||
| 1762 | Changed = true; | |||
| 1763 | break; | |||
| 1764 | } | |||
| 1765 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 1766 | ||||
| 1767 | default: | |||
| 1768 | // Only add if it isn't already in the list. | |||
| 1769 | if (SeenOps.insert(Op.getNode()).second) | |||
| 1770 | Ops.push_back(Op); | |||
| 1771 | else | |||
| 1772 | Changed = true; | |||
| 1773 | break; | |||
| 1774 | } | |||
| 1775 | } | |||
| 1776 | } | |||
| 1777 | ||||
| 1778 | // Re-visit inlined Token Factors, to clean them up in case they have been | |||
| 1779 | // removed. Skip the first Token Factor, as this is the current node. | |||
| 1780 | for (unsigned i = 1, e = TFs.size(); i < e; i++) | |||
| 1781 | AddToWorklist(TFs[i]); | |||
| 1782 | ||||
| 1783 | // Remove Nodes that are chained to another node in the list. Do so | |||
| 1784 | // by walking up chains breath-first stopping when we've seen | |||
| 1785 | // another operand. In general we must climb to the EntryNode, but we can exit | |||
| 1786 | // early if we find all remaining work is associated with just one operand as | |||
| 1787 | // no further pruning is possible. | |||
| 1788 | ||||
| 1789 | // List of nodes to search through and original Ops from which they originate. | |||
| 1790 | SmallVector<std::pair<SDNode *, unsigned>, 8> Worklist; | |||
| 1791 | SmallVector<unsigned, 8> OpWorkCount; // Count of work for each Op. | |||
| 1792 | SmallPtrSet<SDNode *, 16> SeenChains; | |||
| 1793 | bool DidPruneOps = false; | |||
| 1794 | ||||
| 1795 | unsigned NumLeftToConsider = 0; | |||
| 1796 | for (const SDValue &Op : Ops) { | |||
| 1797 | Worklist.push_back(std::make_pair(Op.getNode(), NumLeftToConsider++)); | |||
| 1798 | OpWorkCount.push_back(1); | |||
| 1799 | } | |||
| 1800 | ||||
| 1801 | auto AddToWorklist = [&](unsigned CurIdx, SDNode *Op, unsigned OpNumber) { | |||
| 1802 | // If this is an Op, we can remove the op from the list. Remark any | |||
| 1803 | // search associated with it as from the current OpNumber. | |||
| 1804 | if (SeenOps.count(Op) != 0) { | |||
| 1805 | Changed = true; | |||
| 1806 | DidPruneOps = true; | |||
| 1807 | unsigned OrigOpNumber = 0; | |||
| 1808 | while (OrigOpNumber < Ops.size() && Ops[OrigOpNumber].getNode() != Op) | |||
| 1809 | OrigOpNumber++; | |||
| 1810 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1811, __PRETTY_FUNCTION__)) | |||
| 1811 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1811, __PRETTY_FUNCTION__)); | |||
| 1812 | // Re-mark worklist from OrigOpNumber to OpNumber | |||
| 1813 | for (unsigned i = CurIdx + 1; i < Worklist.size(); ++i) { | |||
| 1814 | if (Worklist[i].second == OrigOpNumber) { | |||
| 1815 | Worklist[i].second = OpNumber; | |||
| 1816 | } | |||
| 1817 | } | |||
| 1818 | OpWorkCount[OpNumber] += OpWorkCount[OrigOpNumber]; | |||
| 1819 | OpWorkCount[OrigOpNumber] = 0; | |||
| 1820 | NumLeftToConsider--; | |||
| 1821 | } | |||
| 1822 | // Add if it's a new chain | |||
| 1823 | if (SeenChains.insert(Op).second) { | |||
| 1824 | OpWorkCount[OpNumber]++; | |||
| 1825 | Worklist.push_back(std::make_pair(Op, OpNumber)); | |||
| 1826 | } | |||
| 1827 | }; | |||
| 1828 | ||||
| 1829 | for (unsigned i = 0; i < Worklist.size() && i < 1024; ++i) { | |||
| 1830 | // We need at least be consider at least 2 Ops to prune. | |||
| 1831 | if (NumLeftToConsider <= 1) | |||
| 1832 | break; | |||
| 1833 | auto CurNode = Worklist[i].first; | |||
| 1834 | auto CurOpNumber = Worklist[i].second; | |||
| 1835 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1836, __PRETTY_FUNCTION__)) | |||
| 1836 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1836, __PRETTY_FUNCTION__)); | |||
| 1837 | switch (CurNode->getOpcode()) { | |||
| 1838 | case ISD::EntryToken: | |||
| 1839 | // Hitting EntryToken is the only way for the search to terminate without | |||
| 1840 | // hitting | |||
| 1841 | // another operand's search. Prevent us from marking this operand | |||
| 1842 | // considered. | |||
| 1843 | NumLeftToConsider++; | |||
| 1844 | break; | |||
| 1845 | case ISD::TokenFactor: | |||
| 1846 | for (const SDValue &Op : CurNode->op_values()) | |||
| 1847 | AddToWorklist(i, Op.getNode(), CurOpNumber); | |||
| 1848 | break; | |||
| 1849 | case ISD::LIFETIME_START: | |||
| 1850 | case ISD::LIFETIME_END: | |||
| 1851 | case ISD::CopyFromReg: | |||
| 1852 | case ISD::CopyToReg: | |||
| 1853 | AddToWorklist(i, CurNode->getOperand(0).getNode(), CurOpNumber); | |||
| 1854 | break; | |||
| 1855 | default: | |||
| 1856 | if (auto *MemNode = dyn_cast<MemSDNode>(CurNode)) | |||
| 1857 | AddToWorklist(i, MemNode->getChain().getNode(), CurOpNumber); | |||
| 1858 | break; | |||
| 1859 | } | |||
| 1860 | OpWorkCount[CurOpNumber]--; | |||
| 1861 | if (OpWorkCount[CurOpNumber] == 0) | |||
| 1862 | NumLeftToConsider--; | |||
| 1863 | } | |||
| 1864 | ||||
| 1865 | // If we've changed things around then replace token factor. | |||
| 1866 | if (Changed) { | |||
| 1867 | SDValue Result; | |||
| 1868 | if (Ops.empty()) { | |||
| 1869 | // The entry token is the only possible outcome. | |||
| 1870 | Result = DAG.getEntryNode(); | |||
| 1871 | } else { | |||
| 1872 | if (DidPruneOps) { | |||
| 1873 | SmallVector<SDValue, 8> PrunedOps; | |||
| 1874 | // | |||
| 1875 | for (const SDValue &Op : Ops) { | |||
| 1876 | if (SeenChains.count(Op.getNode()) == 0) | |||
| 1877 | PrunedOps.push_back(Op); | |||
| 1878 | } | |||
| 1879 | Result = DAG.getTokenFactor(SDLoc(N), PrunedOps); | |||
| 1880 | } else { | |||
| 1881 | Result = DAG.getTokenFactor(SDLoc(N), Ops); | |||
| 1882 | } | |||
| 1883 | } | |||
| 1884 | return Result; | |||
| 1885 | } | |||
| 1886 | return SDValue(); | |||
| 1887 | } | |||
| 1888 | ||||
| 1889 | /// MERGE_VALUES can always be eliminated. | |||
| 1890 | SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) { | |||
| 1891 | WorklistRemover DeadNodes(*this); | |||
| 1892 | // Replacing results may cause a different MERGE_VALUES to suddenly | |||
| 1893 | // be CSE'd with N, and carry its uses with it. Iterate until no | |||
| 1894 | // uses remain, to ensure that the node can be safely deleted. | |||
| 1895 | // First add the users of this node to the work list so that they | |||
| 1896 | // can be tried again once they have new operands. | |||
| 1897 | AddUsersToWorklist(N); | |||
| 1898 | do { | |||
| 1899 | // Do as a single replacement to avoid rewalking use lists. | |||
| 1900 | SmallVector<SDValue, 8> Ops; | |||
| 1901 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) | |||
| 1902 | Ops.push_back(N->getOperand(i)); | |||
| 1903 | DAG.ReplaceAllUsesWith(N, Ops.data()); | |||
| 1904 | } while (!N->use_empty()); | |||
| 1905 | deleteAndRecombine(N); | |||
| 1906 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
| 1907 | } | |||
| 1908 | ||||
| 1909 | /// If \p N is a ConstantSDNode with isOpaque() == false return it casted to a | |||
| 1910 | /// ConstantSDNode pointer else nullptr. | |||
| 1911 | static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) { | |||
| 1912 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N); | |||
| 1913 | return Const != nullptr && !Const->isOpaque() ? Const : nullptr; | |||
| 1914 | } | |||
| 1915 | ||||
| 1916 | SDValue DAGCombiner::foldBinOpIntoSelect(SDNode *BO) { | |||
| 1917 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1918, __PRETTY_FUNCTION__)) | |||
| 1918 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1918, __PRETTY_FUNCTION__)); | |||
| 1919 | ||||
| 1920 | // Don't do this unless the old select is going away. We want to eliminate the | |||
| 1921 | // binary operator, not replace a binop with a select. | |||
| 1922 | // TODO: Handle ISD::SELECT_CC. | |||
| 1923 | unsigned SelOpNo = 0; | |||
| 1924 | SDValue Sel = BO->getOperand(0); | |||
| 1925 | if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) { | |||
| 1926 | SelOpNo = 1; | |||
| 1927 | Sel = BO->getOperand(1); | |||
| 1928 | } | |||
| 1929 | ||||
| 1930 | if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) | |||
| 1931 | return SDValue(); | |||
| 1932 | ||||
| 1933 | SDValue CT = Sel.getOperand(1); | |||
| 1934 | if (!isConstantOrConstantVector(CT, true) && | |||
| 1935 | !isConstantFPBuildVectorOrConstantFP(CT)) | |||
| 1936 | return SDValue(); | |||
| 1937 | ||||
| 1938 | SDValue CF = Sel.getOperand(2); | |||
| 1939 | if (!isConstantOrConstantVector(CF, true) && | |||
| 1940 | !isConstantFPBuildVectorOrConstantFP(CF)) | |||
| 1941 | return SDValue(); | |||
| 1942 | ||||
| 1943 | // Bail out if any constants are opaque because we can't constant fold those. | |||
| 1944 | // The exception is "and" and "or" with either 0 or -1 in which case we can | |||
| 1945 | // propagate non constant operands into select. I.e.: | |||
| 1946 | // and (select Cond, 0, -1), X --> select Cond, 0, X | |||
| 1947 | // or X, (select Cond, -1, 0) --> select Cond, -1, X | |||
| 1948 | auto BinOpcode = BO->getOpcode(); | |||
| 1949 | bool CanFoldNonConst = | |||
| 1950 | (BinOpcode == ISD::AND || BinOpcode == ISD::OR) && | |||
| 1951 | (isNullOrNullSplat(CT) || isAllOnesOrAllOnesSplat(CT)) && | |||
| 1952 | (isNullOrNullSplat(CF) || isAllOnesOrAllOnesSplat(CF)); | |||
| 1953 | ||||
| 1954 | SDValue CBO = BO->getOperand(SelOpNo ^ 1); | |||
| 1955 | if (!CanFoldNonConst && | |||
| 1956 | !isConstantOrConstantVector(CBO, true) && | |||
| 1957 | !isConstantFPBuildVectorOrConstantFP(CBO)) | |||
| 1958 | return SDValue(); | |||
| 1959 | ||||
| 1960 | EVT VT = Sel.getValueType(); | |||
| 1961 | ||||
| 1962 | // In case of shift value and shift amount may have different VT. For instance | |||
| 1963 | // on x86 shift amount is i8 regardles of LHS type. Bail out if we have | |||
| 1964 | // swapped operands and value types do not match. NB: x86 is fine if operands | |||
| 1965 | // are not swapped with shift amount VT being not bigger than shifted value. | |||
| 1966 | // TODO: that is possible to check for a shift operation, correct VTs and | |||
| 1967 | // still perform optimization on x86 if needed. | |||
| 1968 | if (SelOpNo && VT != CBO.getValueType()) | |||
| 1969 | return SDValue(); | |||
| 1970 | ||||
| 1971 | // We have a select-of-constants followed by a binary operator with a | |||
| 1972 | // constant. Eliminate the binop by pulling the constant math into the select. | |||
| 1973 | // Example: add (select Cond, CT, CF), CBO --> select Cond, CT + CBO, CF + CBO | |||
| 1974 | SDLoc DL(Sel); | |||
| 1975 | SDValue NewCT = SelOpNo ? DAG.getNode(BinOpcode, DL, VT, CBO, CT) | |||
| 1976 | : DAG.getNode(BinOpcode, DL, VT, CT, CBO); | |||
| 1977 | if (!CanFoldNonConst && !NewCT.isUndef() && | |||
| 1978 | !isConstantOrConstantVector(NewCT, true) && | |||
| 1979 | !isConstantFPBuildVectorOrConstantFP(NewCT)) | |||
| 1980 | return SDValue(); | |||
| 1981 | ||||
| 1982 | SDValue NewCF = SelOpNo ? DAG.getNode(BinOpcode, DL, VT, CBO, CF) | |||
| 1983 | : DAG.getNode(BinOpcode, DL, VT, CF, CBO); | |||
| 1984 | if (!CanFoldNonConst && !NewCF.isUndef() && | |||
| 1985 | !isConstantOrConstantVector(NewCF, true) && | |||
| 1986 | !isConstantFPBuildVectorOrConstantFP(NewCF)) | |||
| 1987 | return SDValue(); | |||
| 1988 | ||||
| 1989 | SDValue SelectOp = DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF); | |||
| 1990 | SelectOp->setFlags(BO->getFlags()); | |||
| 1991 | return SelectOp; | |||
| 1992 | } | |||
| 1993 | ||||
| 1994 | static SDValue foldAddSubBoolOfMaskedVal(SDNode *N, SelectionDAG &DAG) { | |||
| 1995 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1996, __PRETTY_FUNCTION__)) | |||
| 1996 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1996, __PRETTY_FUNCTION__)); | |||
| 1997 | ||||
| 1998 | // Match a constant operand and a zext operand for the math instruction: | |||
| 1999 | // add Z, C | |||
| 2000 | // sub C, Z | |||
| 2001 | bool IsAdd = N->getOpcode() == ISD::ADD; | |||
| 2002 | SDValue C = IsAdd ? N->getOperand(1) : N->getOperand(0); | |||
| 2003 | SDValue Z = IsAdd ? N->getOperand(0) : N->getOperand(1); | |||
| 2004 | auto *CN = dyn_cast<ConstantSDNode>(C); | |||
| 2005 | if (!CN || Z.getOpcode() != ISD::ZERO_EXTEND) | |||
| 2006 | return SDValue(); | |||
| 2007 | ||||
| 2008 | // Match the zext operand as a setcc of a boolean. | |||
| 2009 | if (Z.getOperand(0).getOpcode() != ISD::SETCC || | |||
| 2010 | Z.getOperand(0).getValueType() != MVT::i1) | |||
| 2011 | return SDValue(); | |||
| 2012 | ||||
| 2013 | // Match the compare as: setcc (X & 1), 0, eq. | |||
| 2014 | SDValue SetCC = Z.getOperand(0); | |||
| 2015 | ISD::CondCode CC = cast<CondCodeSDNode>(SetCC->getOperand(2))->get(); | |||
| 2016 | if (CC != ISD::SETEQ || !isNullConstant(SetCC.getOperand(1)) || | |||
| 2017 | SetCC.getOperand(0).getOpcode() != ISD::AND || | |||
| 2018 | !isOneConstant(SetCC.getOperand(0).getOperand(1))) | |||
| 2019 | return SDValue(); | |||
| 2020 | ||||
| 2021 | // We are adding/subtracting a constant and an inverted low bit. Turn that | |||
| 2022 | // into a subtract/add of the low bit with incremented/decremented constant: | |||
| 2023 | // add (zext i1 (seteq (X & 1), 0)), C --> sub C+1, (zext (X & 1)) | |||
| 2024 | // sub C, (zext i1 (seteq (X & 1), 0)) --> add C-1, (zext (X & 1)) | |||
| 2025 | EVT VT = C.getValueType(); | |||
| 2026 | SDLoc DL(N); | |||
| 2027 | SDValue LowBit = DAG.getZExtOrTrunc(SetCC.getOperand(0), DL, VT); | |||
| 2028 | SDValue C1 = IsAdd ? DAG.getConstant(CN->getAPIntValue() + 1, DL, VT) : | |||
| 2029 | DAG.getConstant(CN->getAPIntValue() - 1, DL, VT); | |||
| 2030 | return DAG.getNode(IsAdd ? ISD::SUB : ISD::ADD, DL, VT, C1, LowBit); | |||
| 2031 | } | |||
| 2032 | ||||
| 2033 | /// Try to fold a 'not' shifted sign-bit with add/sub with constant operand into | |||
| 2034 | /// a shift and add with a different constant. | |||
| 2035 | static SDValue foldAddSubOfSignBit(SDNode *N, SelectionDAG &DAG) { | |||
| 2036 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2037, __PRETTY_FUNCTION__)) | |||
| 2037 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2037, __PRETTY_FUNCTION__)); | |||
| 2038 | ||||
| 2039 | // We need a constant operand for the add/sub, and the other operand is a | |||
| 2040 | // logical shift right: add (srl), C or sub C, (srl). | |||
| 2041 | // TODO - support non-uniform vector amounts. | |||
| 2042 | bool IsAdd = N->getOpcode() == ISD::ADD; | |||
| 2043 | SDValue ConstantOp = IsAdd ? N->getOperand(1) : N->getOperand(0); | |||
| 2044 | SDValue ShiftOp = IsAdd ? N->getOperand(0) : N->getOperand(1); | |||
| 2045 | ConstantSDNode *C = isConstOrConstSplat(ConstantOp); | |||
| 2046 | if (!C || ShiftOp.getOpcode() != ISD::SRL) | |||
| 2047 | return SDValue(); | |||
| 2048 | ||||
| 2049 | // The shift must be of a 'not' value. | |||
| 2050 | SDValue Not = ShiftOp.getOperand(0); | |||
| 2051 | if (!Not.hasOneUse() || !isBitwiseNot(Not)) | |||
| 2052 | return SDValue(); | |||
| 2053 | ||||
| 2054 | // The shift must be moving the sign bit to the least-significant-bit. | |||
| 2055 | EVT VT = ShiftOp.getValueType(); | |||
| 2056 | SDValue ShAmt = ShiftOp.getOperand(1); | |||
| 2057 | ConstantSDNode *ShAmtC = isConstOrConstSplat(ShAmt); | |||
| 2058 | if (!ShAmtC || ShAmtC->getAPIntValue() != (VT.getScalarSizeInBits() - 1)) | |||
| 2059 | return SDValue(); | |||
| 2060 | ||||
| 2061 | // Eliminate the 'not' by adjusting the shift and add/sub constant: | |||
| 2062 | // add (srl (not X), 31), C --> add (sra X, 31), (C + 1) | |||
| 2063 | // sub C, (srl (not X), 31) --> add (srl X, 31), (C - 1) | |||
| 2064 | SDLoc DL(N); | |||
| 2065 | auto ShOpcode = IsAdd ? ISD::SRA : ISD::SRL; | |||
| 2066 | SDValue NewShift = DAG.getNode(ShOpcode, DL, VT, Not.getOperand(0), ShAmt); | |||
| 2067 | APInt NewC = IsAdd ? C->getAPIntValue() + 1 : C->getAPIntValue() - 1; | |||
| 2068 | return DAG.getNode(ISD::ADD, DL, VT, NewShift, DAG.getConstant(NewC, DL, VT)); | |||
| 2069 | } | |||
| 2070 | ||||
| 2071 | /// Try to fold a node that behaves like an ADD (note that N isn't necessarily | |||
| 2072 | /// an ISD::ADD here, it could for example be an ISD::OR if we know that there | |||
| 2073 | /// are no common bits set in the operands). | |||
| 2074 | SDValue DAGCombiner::visitADDLike(SDNode *N) { | |||
| 2075 | SDValue N0 = N->getOperand(0); | |||
| 2076 | SDValue N1 = N->getOperand(1); | |||
| 2077 | EVT VT = N0.getValueType(); | |||
| 2078 | SDLoc DL(N); | |||
| 2079 | ||||
| 2080 | // fold vector ops | |||
| 2081 | if (VT.isVector()) { | |||
| 2082 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 2083 | return FoldedVOp; | |||
| 2084 | ||||
| 2085 | // fold (add x, 0) -> x, vector edition | |||
| 2086 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 2087 | return N0; | |||
| 2088 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
| 2089 | return N1; | |||
| 2090 | } | |||
| 2091 | ||||
| 2092 | // fold (add x, undef) -> undef | |||
| 2093 | if (N0.isUndef()) | |||
| 2094 | return N0; | |||
| 2095 | ||||
| 2096 | if (N1.isUndef()) | |||
| 2097 | return N1; | |||
| 2098 | ||||
| 2099 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { | |||
| 2100 | // canonicalize constant to RHS | |||
| 2101 | if (!DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 2102 | return DAG.getNode(ISD::ADD, DL, VT, N1, N0); | |||
| 2103 | // fold (add c1, c2) -> c1+c2 | |||
| 2104 | return DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, N0.getNode(), | |||
| 2105 | N1.getNode()); | |||
| 2106 | } | |||
| 2107 | ||||
| 2108 | // fold (add x, 0) -> x | |||
| 2109 | if (isNullConstant(N1)) | |||
| 2110 | return N0; | |||
| 2111 | ||||
| 2112 | if (isConstantOrConstantVector(N1, /* NoOpaque */ true)) { | |||
| 2113 | // fold ((A-c1)+c2) -> (A+(c2-c1)) | |||
| 2114 | if (N0.getOpcode() == ISD::SUB && | |||
| 2115 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaque */ true)) { | |||
| 2116 | SDValue Sub = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, N1.getNode(), | |||
| 2117 | N0.getOperand(1).getNode()); | |||
| 2118 | assert(Sub && "Constant folding failed")((Sub && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("Sub && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2118, __PRETTY_FUNCTION__)); | |||
| 2119 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Sub); | |||
| 2120 | } | |||
| 2121 | ||||
| 2122 | // fold ((c1-A)+c2) -> (c1+c2)-A | |||
| 2123 | if (N0.getOpcode() == ISD::SUB && | |||
| 2124 | isConstantOrConstantVector(N0.getOperand(0), /* NoOpaque */ true)) { | |||
| 2125 | SDValue Add = DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, N1.getNode(), | |||
| 2126 | N0.getOperand(0).getNode()); | |||
| 2127 | assert(Add && "Constant folding failed")((Add && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("Add && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2127, __PRETTY_FUNCTION__)); | |||
| 2128 | return DAG.getNode(ISD::SUB, DL, VT, Add, N0.getOperand(1)); | |||
| 2129 | } | |||
| 2130 | ||||
| 2131 | // add (sext i1 X), 1 -> zext (not i1 X) | |||
| 2132 | // We don't transform this pattern: | |||
| 2133 | // add (zext i1 X), -1 -> sext (not i1 X) | |||
| 2134 | // because most (?) targets generate better code for the zext form. | |||
| 2135 | if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.hasOneUse() && | |||
| 2136 | isOneOrOneSplat(N1)) { | |||
| 2137 | SDValue X = N0.getOperand(0); | |||
| 2138 | if ((!LegalOperations || | |||
| 2139 | (TLI.isOperationLegal(ISD::XOR, X.getValueType()) && | |||
| 2140 | TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) && | |||
| 2141 | X.getScalarValueSizeInBits() == 1) { | |||
| 2142 | SDValue Not = DAG.getNOT(DL, X, X.getValueType()); | |||
| 2143 | return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Not); | |||
| 2144 | } | |||
| 2145 | } | |||
| 2146 | ||||
| 2147 | // Undo the add -> or combine to merge constant offsets from a frame index. | |||
| 2148 | if (N0.getOpcode() == ISD::OR && | |||
| 2149 | isa<FrameIndexSDNode>(N0.getOperand(0)) && | |||
| 2150 | isa<ConstantSDNode>(N0.getOperand(1)) && | |||
| 2151 | DAG.haveNoCommonBitsSet(N0.getOperand(0), N0.getOperand(1))) { | |||
| 2152 | SDValue Add0 = DAG.getNode(ISD::ADD, DL, VT, N1, N0.getOperand(1)); | |||
| 2153 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Add0); | |||
| 2154 | } | |||
| 2155 | } | |||
| 2156 | ||||
| 2157 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 2158 | return NewSel; | |||
| 2159 | ||||
| 2160 | // reassociate add | |||
| 2161 | if (!reassociationCanBreakAddressingModePattern(ISD::ADD, DL, N0, N1)) { | |||
| 2162 | if (SDValue RADD = reassociateOps(ISD::ADD, DL, N0, N1, N->getFlags())) | |||
| 2163 | return RADD; | |||
| 2164 | } | |||
| 2165 | // fold ((0-A) + B) -> B-A | |||
| 2166 | if (N0.getOpcode() == ISD::SUB && isNullOrNullSplat(N0.getOperand(0))) | |||
| 2167 | return DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); | |||
| 2168 | ||||
| 2169 | // fold (A + (0-B)) -> A-B | |||
| 2170 | if (N1.getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0))) | |||
| 2171 | return DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(1)); | |||
| 2172 | ||||
| 2173 | // fold (A+(B-A)) -> B | |||
| 2174 | if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) | |||
| 2175 | return N1.getOperand(0); | |||
| 2176 | ||||
| 2177 | // fold ((B-A)+A) -> B | |||
| 2178 | if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1)) | |||
| 2179 | return N0.getOperand(0); | |||
| 2180 | ||||
| 2181 | // fold ((A-B)+(C-A)) -> (C-B) | |||
| 2182 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && | |||
| 2183 | N0.getOperand(0) == N1.getOperand(1)) | |||
| 2184 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), | |||
| 2185 | N0.getOperand(1)); | |||
| 2186 | ||||
| 2187 | // fold ((A-B)+(B-C)) -> (A-C) | |||
| 2188 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && | |||
| 2189 | N0.getOperand(1) == N1.getOperand(0)) | |||
| 2190 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), | |||
| 2191 | N1.getOperand(1)); | |||
| 2192 | ||||
| 2193 | // fold (A+(B-(A+C))) to (B-C) | |||
| 2194 | if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && | |||
| 2195 | N0 == N1.getOperand(1).getOperand(0)) | |||
| 2196 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), | |||
| 2197 | N1.getOperand(1).getOperand(1)); | |||
| 2198 | ||||
| 2199 | // fold (A+(B-(C+A))) to (B-C) | |||
| 2200 | if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && | |||
| 2201 | N0 == N1.getOperand(1).getOperand(1)) | |||
| 2202 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), | |||
| 2203 | N1.getOperand(1).getOperand(0)); | |||
| 2204 | ||||
| 2205 | // fold (A+((B-A)+or-C)) to (B+or-C) | |||
| 2206 | if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) && | |||
| 2207 | N1.getOperand(0).getOpcode() == ISD::SUB && | |||
| 2208 | N0 == N1.getOperand(0).getOperand(1)) | |||
| 2209 | return DAG.getNode(N1.getOpcode(), DL, VT, N1.getOperand(0).getOperand(0), | |||
| 2210 | N1.getOperand(1)); | |||
| 2211 | ||||
| 2212 | // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant | |||
| 2213 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) { | |||
| 2214 | SDValue N00 = N0.getOperand(0); | |||
| 2215 | SDValue N01 = N0.getOperand(1); | |||
| 2216 | SDValue N10 = N1.getOperand(0); | |||
| 2217 | SDValue N11 = N1.getOperand(1); | |||
| 2218 | ||||
| 2219 | if (isConstantOrConstantVector(N00) || isConstantOrConstantVector(N10)) | |||
| 2220 | return DAG.getNode(ISD::SUB, DL, VT, | |||
| 2221 | DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10), | |||
| 2222 | DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11)); | |||
| 2223 | } | |||
| 2224 | ||||
| 2225 | // fold (add (umax X, C), -C) --> (usubsat X, C) | |||
| 2226 | if (N0.getOpcode() == ISD::UMAX && hasOperation(ISD::USUBSAT, VT)) { | |||
| 2227 | auto MatchUSUBSAT = [](ConstantSDNode *Max, ConstantSDNode *Op) { | |||
| 2228 | return (!Max && !Op) || | |||
| 2229 | (Max && Op && Max->getAPIntValue() == (-Op->getAPIntValue())); | |||
| 2230 | }; | |||
| 2231 | if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchUSUBSAT, | |||
| 2232 | /*AllowUndefs*/ true)) | |||
| 2233 | return DAG.getNode(ISD::USUBSAT, DL, VT, N0.getOperand(0), | |||
| 2234 | N0.getOperand(1)); | |||
| 2235 | } | |||
| 2236 | ||||
| 2237 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
| 2238 | return SDValue(N, 0); | |||
| 2239 | ||||
| 2240 | if (isOneOrOneSplat(N1)) { | |||
| 2241 | // fold (add (xor a, -1), 1) -> (sub 0, a) | |||
| 2242 | if (isBitwiseNot(N0)) | |||
| 2243 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), | |||
| 2244 | N0.getOperand(0)); | |||
| 2245 | ||||
| 2246 | // fold (add (add (xor a, -1), b), 1) -> (sub b, a) | |||
| 2247 | if (N0.getOpcode() == ISD::ADD || | |||
| 2248 | N0.getOpcode() == ISD::UADDO || | |||
| 2249 | N0.getOpcode() == ISD::SADDO) { | |||
| 2250 | SDValue A, Xor; | |||
| 2251 | ||||
| 2252 | if (isBitwiseNot(N0.getOperand(0))) { | |||
| 2253 | A = N0.getOperand(1); | |||
| 2254 | Xor = N0.getOperand(0); | |||
| 2255 | } else if (isBitwiseNot(N0.getOperand(1))) { | |||
| 2256 | A = N0.getOperand(0); | |||
| 2257 | Xor = N0.getOperand(1); | |||
| 2258 | } | |||
| 2259 | ||||
| 2260 | if (Xor) | |||
| 2261 | return DAG.getNode(ISD::SUB, DL, VT, A, Xor.getOperand(0)); | |||
| 2262 | } | |||
| 2263 | ||||
| 2264 | // Look for: | |||
| 2265 | // add (add x, y), 1 | |||
| 2266 | // And if the target does not like this form then turn into: | |||
| 2267 | // sub y, (xor x, -1) | |||
| 2268 | if (!TLI.preferIncOfAddToSubOfNot(VT) && N0.hasOneUse() && | |||
| 2269 | N0.getOpcode() == ISD::ADD) { | |||
| 2270 | SDValue Not = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0), | |||
| 2271 | DAG.getAllOnesConstant(DL, VT)); | |||
| 2272 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(1), Not); | |||
| 2273 | } | |||
| 2274 | } | |||
| 2275 | ||||
| 2276 | // (x - y) + -1 -> add (xor y, -1), x | |||
| 2277 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && | |||
| 2278 | isAllOnesOrAllOnesSplat(N1)) { | |||
| 2279 | SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1), N1); | |||
| 2280 | return DAG.getNode(ISD::ADD, DL, VT, Xor, N0.getOperand(0)); | |||
| 2281 | } | |||
| 2282 | ||||
| 2283 | if (SDValue Combined = visitADDLikeCommutative(N0, N1, N)) | |||
| 2284 | return Combined; | |||
| 2285 | ||||
| 2286 | if (SDValue Combined = visitADDLikeCommutative(N1, N0, N)) | |||
| 2287 | return Combined; | |||
| 2288 | ||||
| 2289 | return SDValue(); | |||
| 2290 | } | |||
| 2291 | ||||
| 2292 | SDValue DAGCombiner::visitADD(SDNode *N) { | |||
| 2293 | SDValue N0 = N->getOperand(0); | |||
| 2294 | SDValue N1 = N->getOperand(1); | |||
| 2295 | EVT VT = N0.getValueType(); | |||
| 2296 | SDLoc DL(N); | |||
| 2297 | ||||
| 2298 | if (SDValue Combined = visitADDLike(N)) | |||
| 2299 | return Combined; | |||
| 2300 | ||||
| 2301 | if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG)) | |||
| 2302 | return V; | |||
| 2303 | ||||
| 2304 | if (SDValue V = foldAddSubOfSignBit(N, DAG)) | |||
| 2305 | return V; | |||
| 2306 | ||||
| 2307 | // fold (a+b) -> (a|b) iff a and b share no bits. | |||
| 2308 | if ((!LegalOperations || TLI.isOperationLegal(ISD::OR, VT)) && | |||
| 2309 | DAG.haveNoCommonBitsSet(N0, N1)) | |||
| 2310 | return DAG.getNode(ISD::OR, DL, VT, N0, N1); | |||
| 2311 | ||||
| 2312 | return SDValue(); | |||
| 2313 | } | |||
| 2314 | ||||
| 2315 | SDValue DAGCombiner::visitADDSAT(SDNode *N) { | |||
| 2316 | unsigned Opcode = N->getOpcode(); | |||
| 2317 | SDValue N0 = N->getOperand(0); | |||
| 2318 | SDValue N1 = N->getOperand(1); | |||
| 2319 | EVT VT = N0.getValueType(); | |||
| 2320 | SDLoc DL(N); | |||
| 2321 | ||||
| 2322 | // fold vector ops | |||
| 2323 | if (VT.isVector()) { | |||
| 2324 | // TODO SimplifyVBinOp | |||
| 2325 | ||||
| 2326 | // fold (add_sat x, 0) -> x, vector edition | |||
| 2327 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 2328 | return N0; | |||
| 2329 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
| 2330 | return N1; | |||
| 2331 | } | |||
| 2332 | ||||
| 2333 | // fold (add_sat x, undef) -> -1 | |||
| 2334 | if (N0.isUndef() || N1.isUndef()) | |||
| 2335 | return DAG.getAllOnesConstant(DL, VT); | |||
| 2336 | ||||
| 2337 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { | |||
| 2338 | // canonicalize constant to RHS | |||
| 2339 | if (!DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 2340 | return DAG.getNode(Opcode, DL, VT, N1, N0); | |||
| 2341 | // fold (add_sat c1, c2) -> c3 | |||
| 2342 | return DAG.FoldConstantArithmetic(Opcode, DL, VT, N0.getNode(), | |||
| 2343 | N1.getNode()); | |||
| 2344 | } | |||
| 2345 | ||||
| 2346 | // fold (add_sat x, 0) -> x | |||
| 2347 | if (isNullConstant(N1)) | |||
| 2348 | return N0; | |||
| 2349 | ||||
| 2350 | // If it cannot overflow, transform into an add. | |||
| 2351 | if (Opcode == ISD::UADDSAT) | |||
| 2352 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) | |||
| 2353 | return DAG.getNode(ISD::ADD, DL, VT, N0, N1); | |||
| 2354 | ||||
| 2355 | return SDValue(); | |||
| 2356 | } | |||
| 2357 | ||||
| 2358 | static SDValue getAsCarry(const TargetLowering &TLI, SDValue V) { | |||
| 2359 | bool Masked = false; | |||
| 2360 | ||||
| 2361 | // First, peel away TRUNCATE/ZERO_EXTEND/AND nodes due to legalization. | |||
| 2362 | while (true) { | |||
| 2363 | if (V.getOpcode() == ISD::TRUNCATE || V.getOpcode() == ISD::ZERO_EXTEND) { | |||
| 2364 | V = V.getOperand(0); | |||
| 2365 | continue; | |||
| 2366 | } | |||
| 2367 | ||||
| 2368 | if (V.getOpcode() == ISD::AND && isOneConstant(V.getOperand(1))) { | |||
| 2369 | Masked = true; | |||
| 2370 | V = V.getOperand(0); | |||
| 2371 | continue; | |||
| 2372 | } | |||
| 2373 | ||||
| 2374 | break; | |||
| 2375 | } | |||
| 2376 | ||||
| 2377 | // If this is not a carry, return. | |||
| 2378 | if (V.getResNo() != 1) | |||
| 2379 | return SDValue(); | |||
| 2380 | ||||
| 2381 | if (V.getOpcode() != ISD::ADDCARRY && V.getOpcode() != ISD::SUBCARRY && | |||
| 2382 | V.getOpcode() != ISD::UADDO && V.getOpcode() != ISD::USUBO) | |||
| 2383 | return SDValue(); | |||
| 2384 | ||||
| 2385 | EVT VT = V.getNode()->getValueType(0); | |||
| 2386 | if (!TLI.isOperationLegalOrCustom(V.getOpcode(), VT)) | |||
| 2387 | return SDValue(); | |||
| 2388 | ||||
| 2389 | // If the result is masked, then no matter what kind of bool it is we can | |||
| 2390 | // return. If it isn't, then we need to make sure the bool type is either 0 or | |||
| 2391 | // 1 and not other values. | |||
| 2392 | if (Masked || | |||
| 2393 | TLI.getBooleanContents(V.getValueType()) == | |||
| 2394 | TargetLoweringBase::ZeroOrOneBooleanContent) | |||
| 2395 | return V; | |||
| 2396 | ||||
| 2397 | return SDValue(); | |||
| 2398 | } | |||
| 2399 | ||||
| 2400 | /// Given the operands of an add/sub operation, see if the 2nd operand is a | |||
| 2401 | /// masked 0/1 whose source operand is actually known to be 0/-1. If so, invert | |||
| 2402 | /// the opcode and bypass the mask operation. | |||
| 2403 | static SDValue foldAddSubMasked1(bool IsAdd, SDValue N0, SDValue N1, | |||
| 2404 | SelectionDAG &DAG, const SDLoc &DL) { | |||
| 2405 | if (N1.getOpcode() != ISD::AND || !isOneOrOneSplat(N1->getOperand(1))) | |||
| 2406 | return SDValue(); | |||
| 2407 | ||||
| 2408 | EVT VT = N0.getValueType(); | |||
| 2409 | if (DAG.ComputeNumSignBits(N1.getOperand(0)) != VT.getScalarSizeInBits()) | |||
| 2410 | return SDValue(); | |||
| 2411 | ||||
| 2412 | // add N0, (and (AssertSext X, i1), 1) --> sub N0, X | |||
| 2413 | // sub N0, (and (AssertSext X, i1), 1) --> add N0, X | |||
| 2414 | return DAG.getNode(IsAdd ? ISD::SUB : ISD::ADD, DL, VT, N0, N1.getOperand(0)); | |||
| 2415 | } | |||
| 2416 | ||||
| 2417 | /// Helper for doing combines based on N0 and N1 being added to each other. | |||
| 2418 | SDValue DAGCombiner::visitADDLikeCommutative(SDValue N0, SDValue N1, | |||
| 2419 | SDNode *LocReference) { | |||
| 2420 | EVT VT = N0.getValueType(); | |||
| 2421 | SDLoc DL(LocReference); | |||
| 2422 | ||||
| 2423 | // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)) | |||
| 2424 | if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB && | |||
| 2425 | isNullOrNullSplat(N1.getOperand(0).getOperand(0))) | |||
| 2426 | return DAG.getNode(ISD::SUB, DL, VT, N0, | |||
| 2427 | DAG.getNode(ISD::SHL, DL, VT, | |||
| 2428 | N1.getOperand(0).getOperand(1), | |||
| 2429 | N1.getOperand(1))); | |||
| 2430 | ||||
| 2431 | if (SDValue V = foldAddSubMasked1(true, N0, N1, DAG, DL)) | |||
| 2432 | return V; | |||
| 2433 | ||||
| 2434 | // Look for: | |||
| 2435 | // add (add x, 1), y | |||
| 2436 | // And if the target does not like this form then turn into: | |||
| 2437 | // sub y, (xor x, -1) | |||
| 2438 | if (!TLI.preferIncOfAddToSubOfNot(VT) && N0.hasOneUse() && | |||
| 2439 | N0.getOpcode() == ISD::ADD && isOneOrOneSplat(N0.getOperand(1))) { | |||
| 2440 | SDValue Not = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0), | |||
| 2441 | DAG.getAllOnesConstant(DL, VT)); | |||
| 2442 | return DAG.getNode(ISD::SUB, DL, VT, N1, Not); | |||
| 2443 | } | |||
| 2444 | ||||
| 2445 | // Hoist one-use subtraction by non-opaque constant: | |||
| 2446 | // (x - C) + y -> (x + y) - C | |||
| 2447 | // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors. | |||
| 2448 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && | |||
| 2449 | isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { | |||
| 2450 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), N1); | |||
| 2451 | return DAG.getNode(ISD::SUB, DL, VT, Add, N0.getOperand(1)); | |||
| 2452 | } | |||
| 2453 | // Hoist one-use subtraction from non-opaque constant: | |||
| 2454 | // (C - x) + y -> (y - x) + C | |||
| 2455 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && | |||
| 2456 | isConstantOrConstantVector(N0.getOperand(0), /*NoOpaques=*/true)) { | |||
| 2457 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); | |||
| 2458 | return DAG.getNode(ISD::ADD, DL, VT, Sub, N0.getOperand(0)); | |||
| 2459 | } | |||
| 2460 | ||||
| 2461 | // If the target's bool is represented as 0/1, prefer to make this 'sub 0/1' | |||
| 2462 | // rather than 'add 0/-1' (the zext should get folded). | |||
| 2463 | // add (sext i1 Y), X --> sub X, (zext i1 Y) | |||
| 2464 | if (N0.getOpcode() == ISD::SIGN_EXTEND && | |||
| 2465 | N0.getOperand(0).getScalarValueSizeInBits() == 1 && | |||
| 2466 | TLI.getBooleanContents(VT) == TargetLowering::ZeroOrOneBooleanContent) { | |||
| 2467 | SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); | |||
| 2468 | return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt); | |||
| 2469 | } | |||
| 2470 | ||||
| 2471 | // add X, (sextinreg Y i1) -> sub X, (and Y 1) | |||
| 2472 | if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { | |||
| 2473 | VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1)); | |||
| 2474 | if (TN->getVT() == MVT::i1) { | |||
| 2475 | SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), | |||
| 2476 | DAG.getConstant(1, DL, VT)); | |||
| 2477 | return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt); | |||
| 2478 | } | |||
| 2479 | } | |||
| 2480 | ||||
| 2481 | // (add X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) | |||
| 2482 | if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1)) && | |||
| 2483 | N1.getResNo() == 0) | |||
| 2484 | return DAG.getNode(ISD::ADDCARRY, DL, N1->getVTList(), | |||
| 2485 | N0, N1.getOperand(0), N1.getOperand(2)); | |||
| 2486 | ||||
| 2487 | // (add X, Carry) -> (addcarry X, 0, Carry) | |||
| 2488 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) | |||
| 2489 | if (SDValue Carry = getAsCarry(TLI, N1)) | |||
| 2490 | return DAG.getNode(ISD::ADDCARRY, DL, | |||
| 2491 | DAG.getVTList(VT, Carry.getValueType()), N0, | |||
| 2492 | DAG.getConstant(0, DL, VT), Carry); | |||
| 2493 | ||||
| 2494 | return SDValue(); | |||
| 2495 | } | |||
| 2496 | ||||
| 2497 | SDValue DAGCombiner::visitADDC(SDNode *N) { | |||
| 2498 | SDValue N0 = N->getOperand(0); | |||
| 2499 | SDValue N1 = N->getOperand(1); | |||
| 2500 | EVT VT = N0.getValueType(); | |||
| 2501 | SDLoc DL(N); | |||
| 2502 | ||||
| 2503 | // If the flag result is dead, turn this into an ADD. | |||
| 2504 | if (!N->hasAnyUseOfValue(1)) | |||
| 2505 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
| 2506 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
| 2507 | ||||
| 2508 | // canonicalize constant to RHS. | |||
| 2509 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
| 2510 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
| 2511 | if (N0C && !N1C) | |||
| 2512 | return DAG.getNode(ISD::ADDC, DL, N->getVTList(), N1, N0); | |||
| 2513 | ||||
| 2514 | // fold (addc x, 0) -> x + no carry out | |||
| 2515 | if (isNullConstant(N1)) | |||
| 2516 | return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, | |||
| 2517 | DL, MVT::Glue)); | |||
| 2518 | ||||
| 2519 | // If it cannot overflow, transform into an add. | |||
| 2520 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) | |||
| 2521 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
| 2522 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
| 2523 | ||||
| 2524 | return SDValue(); | |||
| 2525 | } | |||
| 2526 | ||||
| 2527 | static SDValue flipBoolean(SDValue V, const SDLoc &DL, | |||
| 2528 | SelectionDAG &DAG, const TargetLowering &TLI) { | |||
| 2529 | EVT VT = V.getValueType(); | |||
| 2530 | ||||
| 2531 | SDValue Cst; | |||
| 2532 | switch (TLI.getBooleanContents(VT)) { | |||
| 2533 | case TargetLowering::ZeroOrOneBooleanContent: | |||
| 2534 | case TargetLowering::UndefinedBooleanContent: | |||
| 2535 | Cst = DAG.getConstant(1, DL, VT); | |||
| 2536 | break; | |||
| 2537 | case TargetLowering::ZeroOrNegativeOneBooleanContent: | |||
| 2538 | Cst = DAG.getAllOnesConstant(DL, VT); | |||
| 2539 | break; | |||
| 2540 | } | |||
| 2541 | ||||
| 2542 | return DAG.getNode(ISD::XOR, DL, VT, V, Cst); | |||
| 2543 | } | |||
| 2544 | ||||
| 2545 | /** | |||
| 2546 | * Flips a boolean if it is cheaper to compute. If the Force parameters is set, | |||
| 2547 | * then the flip also occurs if computing the inverse is the same cost. | |||
| 2548 | * This function returns an empty SDValue in case it cannot flip the boolean | |||
| 2549 | * without increasing the cost of the computation. If you want to flip a boolean | |||
| 2550 | * no matter what, use flipBoolean. | |||
| 2551 | */ | |||
| 2552 | static SDValue extractBooleanFlip(SDValue V, SelectionDAG &DAG, | |||
| 2553 | const TargetLowering &TLI, | |||
| 2554 | bool Force) { | |||
| 2555 | if (Force && isa<ConstantSDNode>(V)) | |||
| 2556 | return flipBoolean(V, SDLoc(V), DAG, TLI); | |||
| 2557 | ||||
| 2558 | if (V.getOpcode() != ISD::XOR) | |||
| 2559 | return SDValue(); | |||
| 2560 | ||||
| 2561 | ConstantSDNode *Const = isConstOrConstSplat(V.getOperand(1), false); | |||
| 2562 | if (!Const) | |||
| 2563 | return SDValue(); | |||
| 2564 | ||||
| 2565 | EVT VT = V.getValueType(); | |||
| 2566 | ||||
| 2567 | bool IsFlip = false; | |||
| 2568 | switch(TLI.getBooleanContents(VT)) { | |||
| 2569 | case TargetLowering::ZeroOrOneBooleanContent: | |||
| 2570 | IsFlip = Const->isOne(); | |||
| 2571 | break; | |||
| 2572 | case TargetLowering::ZeroOrNegativeOneBooleanContent: | |||
| 2573 | IsFlip = Const->isAllOnesValue(); | |||
| 2574 | break; | |||
| 2575 | case TargetLowering::UndefinedBooleanContent: | |||
| 2576 | IsFlip = (Const->getAPIntValue() & 0x01) == 1; | |||
| 2577 | break; | |||
| 2578 | } | |||
| 2579 | ||||
| 2580 | if (IsFlip) | |||
| 2581 | return V.getOperand(0); | |||
| 2582 | if (Force) | |||
| 2583 | return flipBoolean(V, SDLoc(V), DAG, TLI); | |||
| 2584 | return SDValue(); | |||
| 2585 | } | |||
| 2586 | ||||
| 2587 | SDValue DAGCombiner::visitADDO(SDNode *N) { | |||
| 2588 | SDValue N0 = N->getOperand(0); | |||
| 2589 | SDValue N1 = N->getOperand(1); | |||
| 2590 | EVT VT = N0.getValueType(); | |||
| 2591 | bool IsSigned = (ISD::SADDO == N->getOpcode()); | |||
| 2592 | ||||
| 2593 | EVT CarryVT = N->getValueType(1); | |||
| 2594 | SDLoc DL(N); | |||
| 2595 | ||||
| 2596 | // If the flag result is dead, turn this into an ADD. | |||
| 2597 | if (!N->hasAnyUseOfValue(1)) | |||
| 2598 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
| 2599 | DAG.getUNDEF(CarryVT)); | |||
| 2600 | ||||
| 2601 | // canonicalize constant to RHS. | |||
| 2602 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 2603 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 2604 | return DAG.getNode(N->getOpcode(), DL, N->getVTList(), N1, N0); | |||
| 2605 | ||||
| 2606 | // fold (addo x, 0) -> x + no carry out | |||
| 2607 | if (isNullOrNullSplat(N1)) | |||
| 2608 | return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); | |||
| 2609 | ||||
| 2610 | if (!IsSigned) { | |||
| 2611 | // If it cannot overflow, transform into an add. | |||
| 2612 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) | |||
| 2613 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
| 2614 | DAG.getConstant(0, DL, CarryVT)); | |||
| 2615 | ||||
| 2616 | // fold (uaddo (xor a, -1), 1) -> (usub 0, a) and flip carry. | |||
| 2617 | if (isBitwiseNot(N0) && isOneOrOneSplat(N1)) { | |||
| 2618 | SDValue Sub = DAG.getNode(ISD::USUBO, DL, N->getVTList(), | |||
| 2619 | DAG.getConstant(0, DL, VT), N0.getOperand(0)); | |||
| 2620 | return CombineTo(N, Sub, | |||
| 2621 | flipBoolean(Sub.getValue(1), DL, DAG, TLI)); | |||
| 2622 | } | |||
| 2623 | ||||
| 2624 | if (SDValue Combined = visitUADDOLike(N0, N1, N)) | |||
| 2625 | return Combined; | |||
| 2626 | ||||
| 2627 | if (SDValue Combined = visitUADDOLike(N1, N0, N)) | |||
| 2628 | return Combined; | |||
| 2629 | } | |||
| 2630 | ||||
| 2631 | return SDValue(); | |||
| 2632 | } | |||
| 2633 | ||||
| 2634 | SDValue DAGCombiner::visitUADDOLike(SDValue N0, SDValue N1, SDNode *N) { | |||
| 2635 | EVT VT = N0.getValueType(); | |||
| 2636 | if (VT.isVector()) | |||
| 2637 | return SDValue(); | |||
| 2638 | ||||
| 2639 | // (uaddo X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) | |||
| 2640 | // If Y + 1 cannot overflow. | |||
| 2641 | if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1))) { | |||
| 2642 | SDValue Y = N1.getOperand(0); | |||
| 2643 | SDValue One = DAG.getConstant(1, SDLoc(N), Y.getValueType()); | |||
| 2644 | if (DAG.computeOverflowKind(Y, One) == SelectionDAG::OFK_Never) | |||
| 2645 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, Y, | |||
| 2646 | N1.getOperand(2)); | |||
| 2647 | } | |||
| 2648 | ||||
| 2649 | // (uaddo X, Carry) -> (addcarry X, 0, Carry) | |||
| 2650 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) | |||
| 2651 | if (SDValue Carry = getAsCarry(TLI, N1)) | |||
| 2652 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, | |||
| 2653 | DAG.getConstant(0, SDLoc(N), VT), Carry); | |||
| 2654 | ||||
| 2655 | return SDValue(); | |||
| 2656 | } | |||
| 2657 | ||||
| 2658 | SDValue DAGCombiner::visitADDE(SDNode *N) { | |||
| 2659 | SDValue N0 = N->getOperand(0); | |||
| 2660 | SDValue N1 = N->getOperand(1); | |||
| 2661 | SDValue CarryIn = N->getOperand(2); | |||
| 2662 | ||||
| 2663 | // canonicalize constant to RHS | |||
| 2664 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
| 2665 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
| 2666 | if (N0C && !N1C) | |||
| 2667 | return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(), | |||
| 2668 | N1, N0, CarryIn); | |||
| 2669 | ||||
| 2670 | // fold (adde x, y, false) -> (addc x, y) | |||
| 2671 | if (CarryIn.getOpcode() == ISD::CARRY_FALSE) | |||
| 2672 | return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1); | |||
| 2673 | ||||
| 2674 | return SDValue(); | |||
| 2675 | } | |||
| 2676 | ||||
| 2677 | SDValue DAGCombiner::visitADDCARRY(SDNode *N) { | |||
| 2678 | SDValue N0 = N->getOperand(0); | |||
| 2679 | SDValue N1 = N->getOperand(1); | |||
| 2680 | SDValue CarryIn = N->getOperand(2); | |||
| 2681 | SDLoc DL(N); | |||
| 2682 | ||||
| 2683 | // canonicalize constant to RHS | |||
| 2684 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
| 2685 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
| 2686 | if (N0C && !N1C) | |||
| ||||
| 2687 | return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), N1, N0, CarryIn); | |||
| 2688 | ||||
| 2689 | // fold (addcarry x, y, false) -> (uaddo x, y) | |||
| 2690 | if (isNullConstant(CarryIn)) { | |||
| 2691 | if (!LegalOperations || | |||
| 2692 | TLI.isOperationLegalOrCustom(ISD::UADDO, N->getValueType(0))) | |||
| 2693 | return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N0, N1); | |||
| 2694 | } | |||
| 2695 | ||||
| 2696 | // fold (addcarry 0, 0, X) -> (and (ext/trunc X), 1) and no carry. | |||
| 2697 | if (isNullConstant(N0) && isNullConstant(N1)) { | |||
| 2698 | EVT VT = N0.getValueType(); | |||
| 2699 | EVT CarryVT = CarryIn.getValueType(); | |||
| 2700 | SDValue CarryExt = DAG.getBoolExtOrTrunc(CarryIn, DL, VT, CarryVT); | |||
| 2701 | AddToWorklist(CarryExt.getNode()); | |||
| 2702 | return CombineTo(N, DAG.getNode(ISD::AND, DL, VT, CarryExt, | |||
| 2703 | DAG.getConstant(1, DL, VT)), | |||
| 2704 | DAG.getConstant(0, DL, CarryVT)); | |||
| 2705 | } | |||
| 2706 | ||||
| 2707 | if (SDValue Combined = visitADDCARRYLike(N0, N1, CarryIn, N)) | |||
| 2708 | return Combined; | |||
| 2709 | ||||
| 2710 | if (SDValue Combined = visitADDCARRYLike(N1, N0, CarryIn, N)) | |||
| 2711 | return Combined; | |||
| 2712 | ||||
| 2713 | return SDValue(); | |||
| 2714 | } | |||
| 2715 | ||||
| 2716 | /** | |||
| 2717 | * If we are facing some sort of diamond carry propapagtion pattern try to | |||
| 2718 | * break it up to generate something like: | |||
| 2719 | * (addcarry X, 0, (addcarry A, B, Z):Carry) | |||
| 2720 | * | |||
| 2721 | * The end result is usually an increase in operation required, but because the | |||
| 2722 | * carry is now linearized, other tranforms can kick in and optimize the DAG. | |||
| 2723 | * | |||
| 2724 | * Patterns typically look something like | |||
| 2725 | * (uaddo A, B) | |||
| 2726 | * / \ | |||
| 2727 | * Carry Sum | |||
| 2728 | * | \ | |||
| 2729 | * | (addcarry *, 0, Z) | |||
| 2730 | * | / | |||
| 2731 | * \ Carry | |||
| 2732 | * | / | |||
| 2733 | * (addcarry X, *, *) | |||
| 2734 | * | |||
| 2735 | * But numerous variation exist. Our goal is to identify A, B, X and Z and | |||
| 2736 | * produce a combine with a single path for carry propagation. | |||
| 2737 | */ | |||
| 2738 | static SDValue combineADDCARRYDiamond(DAGCombiner &Combiner, SelectionDAG &DAG, | |||
| 2739 | SDValue X, SDValue Carry0, SDValue Carry1, | |||
| 2740 | SDNode *N) { | |||
| 2741 | if (Carry1.getResNo() != 1 || Carry0.getResNo() != 1) | |||
| 2742 | return SDValue(); | |||
| 2743 | if (Carry1.getOpcode() != ISD::UADDO) | |||
| 2744 | return SDValue(); | |||
| 2745 | ||||
| 2746 | SDValue Z; | |||
| 2747 | ||||
| 2748 | /** | |||
| 2749 | * First look for a suitable Z. It will present itself in the form of | |||
| 2750 | * (addcarry Y, 0, Z) or its equivalent (uaddo Y, 1) for Z=true | |||
| 2751 | */ | |||
| 2752 | if (Carry0.getOpcode() == ISD::ADDCARRY && | |||
| 2753 | isNullConstant(Carry0.getOperand(1))) { | |||
| 2754 | Z = Carry0.getOperand(2); | |||
| 2755 | } else if (Carry0.getOpcode() == ISD::UADDO && | |||
| 2756 | isOneConstant(Carry0.getOperand(1))) { | |||
| 2757 | EVT VT = Combiner.getSetCCResultType(Carry0.getValueType()); | |||
| 2758 | Z = DAG.getConstant(1, SDLoc(Carry0.getOperand(1)), VT); | |||
| 2759 | } else { | |||
| 2760 | // We couldn't find a suitable Z. | |||
| 2761 | return SDValue(); | |||
| 2762 | } | |||
| 2763 | ||||
| 2764 | ||||
| 2765 | auto cancelDiamond = [&](SDValue A,SDValue B) { | |||
| 2766 | SDLoc DL(N); | |||
| 2767 | SDValue NewY = DAG.getNode(ISD::ADDCARRY, DL, Carry0->getVTList(), A, B, Z); | |||
| 2768 | Combiner.AddToWorklist(NewY.getNode()); | |||
| 2769 | return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), X, | |||
| 2770 | DAG.getConstant(0, DL, X.getValueType()), | |||
| 2771 | NewY.getValue(1)); | |||
| 2772 | }; | |||
| 2773 | ||||
| 2774 | /** | |||
| 2775 | * (uaddo A, B) | |||
| 2776 | * | | |||
| 2777 | * Sum | |||
| 2778 | * | | |||
| 2779 | * (addcarry *, 0, Z) | |||
| 2780 | */ | |||
| 2781 | if (Carry0.getOperand(0) == Carry1.getValue(0)) { | |||
| 2782 | return cancelDiamond(Carry1.getOperand(0), Carry1.getOperand(1)); | |||
| 2783 | } | |||
| 2784 | ||||
| 2785 | /** | |||
| 2786 | * (addcarry A, 0, Z) | |||
| 2787 | * | | |||
| 2788 | * Sum | |||
| 2789 | * | | |||
| 2790 | * (uaddo *, B) | |||
| 2791 | */ | |||
| 2792 | if (Carry1.getOperand(0) == Carry0.getValue(0)) { | |||
| 2793 | return cancelDiamond(Carry0.getOperand(0), Carry1.getOperand(1)); | |||
| 2794 | } | |||
| 2795 | ||||
| 2796 | if (Carry1.getOperand(1) == Carry0.getValue(0)) { | |||
| 2797 | return cancelDiamond(Carry1.getOperand(0), Carry0.getOperand(0)); | |||
| 2798 | } | |||
| 2799 | ||||
| 2800 | return SDValue(); | |||
| 2801 | } | |||
| 2802 | ||||
| 2803 | SDValue DAGCombiner::visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, | |||
| 2804 | SDNode *N) { | |||
| 2805 | // fold (addcarry (xor a, -1), b, c) -> (subcarry b, a, !c) and flip carry. | |||
| 2806 | if (isBitwiseNot(N0)) | |||
| 2807 | if (SDValue NotC = extractBooleanFlip(CarryIn, DAG, TLI, true)) { | |||
| 2808 | SDLoc DL(N); | |||
| 2809 | SDValue Sub = DAG.getNode(ISD::SUBCARRY, DL, N->getVTList(), N1, | |||
| 2810 | N0.getOperand(0), NotC); | |||
| 2811 | return CombineTo(N, Sub, | |||
| 2812 | flipBoolean(Sub.getValue(1), DL, DAG, TLI)); | |||
| 2813 | } | |||
| 2814 | ||||
| 2815 | // Iff the flag result is dead: | |||
| 2816 | // (addcarry (add|uaddo X, Y), 0, Carry) -> (addcarry X, Y, Carry) | |||
| 2817 | // Don't do this if the Carry comes from the uaddo. It won't remove the uaddo | |||
| 2818 | // or the dependency between the instructions. | |||
| 2819 | if ((N0.getOpcode() == ISD::ADD || | |||
| 2820 | (N0.getOpcode() == ISD::UADDO && N0.getResNo() == 0 && | |||
| 2821 | N0.getValue(1) != CarryIn)) && | |||
| 2822 | isNullConstant(N1) && !N->hasAnyUseOfValue(1)) | |||
| 2823 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), | |||
| 2824 | N0.getOperand(0), N0.getOperand(1), CarryIn); | |||
| 2825 | ||||
| 2826 | /** | |||
| 2827 | * When one of the addcarry argument is itself a carry, we may be facing | |||
| 2828 | * a diamond carry propagation. In which case we try to transform the DAG | |||
| 2829 | * to ensure linear carry propagation if that is possible. | |||
| 2830 | */ | |||
| 2831 | if (auto Y = getAsCarry(TLI, N1)) { | |||
| 2832 | // Because both are carries, Y and Z can be swapped. | |||
| 2833 | if (auto R = combineADDCARRYDiamond(*this, DAG, N0, Y, CarryIn, N)) | |||
| 2834 | return R; | |||
| 2835 | if (auto R = combineADDCARRYDiamond(*this, DAG, N0, CarryIn, Y, N)) | |||
| 2836 | return R; | |||
| 2837 | } | |||
| 2838 | ||||
| 2839 | return SDValue(); | |||
| 2840 | } | |||
| 2841 | ||||
| 2842 | // Since it may not be valid to emit a fold to zero for vector initializers | |||
| 2843 | // check if we can before folding. | |||
| 2844 | static SDValue tryFoldToZero(const SDLoc &DL, const TargetLowering &TLI, EVT VT, | |||
| 2845 | SelectionDAG &DAG, bool LegalOperations) { | |||
| 2846 | if (!VT.isVector()) | |||
| 2847 | return DAG.getConstant(0, DL, VT); | |||
| 2848 | if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) | |||
| 2849 | return DAG.getConstant(0, DL, VT); | |||
| 2850 | return SDValue(); | |||
| 2851 | } | |||
| 2852 | ||||
| 2853 | SDValue DAGCombiner::visitSUB(SDNode *N) { | |||
| 2854 | SDValue N0 = N->getOperand(0); | |||
| 2855 | SDValue N1 = N->getOperand(1); | |||
| 2856 | EVT VT = N0.getValueType(); | |||
| 2857 | SDLoc DL(N); | |||
| 2858 | ||||
| 2859 | // fold vector ops | |||
| 2860 | if (VT.isVector()) { | |||
| 2861 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 2862 | return FoldedVOp; | |||
| 2863 | ||||
| 2864 | // fold (sub x, 0) -> x, vector edition | |||
| 2865 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 2866 | return N0; | |||
| 2867 | } | |||
| 2868 | ||||
| 2869 | // fold (sub x, x) -> 0 | |||
| 2870 | // FIXME: Refactor this and xor and other similar operations together. | |||
| 2871 | if (N0 == N1) | |||
| 2872 | return tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); | |||
| 2873 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 2874 | DAG.isConstantIntBuildVectorOrConstantInt(N1)) { | |||
| 2875 | // fold (sub c1, c2) -> c1-c2 | |||
| 2876 | return DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, N0.getNode(), | |||
| 2877 | N1.getNode()); | |||
| 2878 | } | |||
| 2879 | ||||
| 2880 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 2881 | return NewSel; | |||
| 2882 | ||||
| 2883 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); | |||
| 2884 | ||||
| 2885 | // fold (sub x, c) -> (add x, -c) | |||
| 2886 | if (N1C) { | |||
| 2887 | return DAG.getNode(ISD::ADD, DL, VT, N0, | |||
| 2888 | DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); | |||
| 2889 | } | |||
| 2890 | ||||
| 2891 | if (isNullOrNullSplat(N0)) { | |||
| 2892 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
| 2893 | // Right-shifting everything out but the sign bit followed by negation is | |||
| 2894 | // the same as flipping arithmetic/logical shift type without the negation: | |||
| 2895 | // -(X >>u 31) -> (X >>s 31) | |||
| 2896 | // -(X >>s 31) -> (X >>u 31) | |||
| 2897 | if (N1->getOpcode() == ISD::SRA || N1->getOpcode() == ISD::SRL) { | |||
| 2898 | ConstantSDNode *ShiftAmt = isConstOrConstSplat(N1.getOperand(1)); | |||
| 2899 | if (ShiftAmt && ShiftAmt->getAPIntValue() == (BitWidth - 1)) { | |||
| 2900 | auto NewSh = N1->getOpcode() == ISD::SRA ? ISD::SRL : ISD::SRA; | |||
| 2901 | if (!LegalOperations || TLI.isOperationLegal(NewSh, VT)) | |||
| 2902 | return DAG.getNode(NewSh, DL, VT, N1.getOperand(0), N1.getOperand(1)); | |||
| 2903 | } | |||
| 2904 | } | |||
| 2905 | ||||
| 2906 | // 0 - X --> 0 if the sub is NUW. | |||
| 2907 | if (N->getFlags().hasNoUnsignedWrap()) | |||
| 2908 | return N0; | |||
| 2909 | ||||
| 2910 | if (DAG.MaskedValueIsZero(N1, ~APInt::getSignMask(BitWidth))) { | |||
| 2911 | // N1 is either 0 or the minimum signed value. If the sub is NSW, then | |||
| 2912 | // N1 must be 0 because negating the minimum signed value is undefined. | |||
| 2913 | if (N->getFlags().hasNoSignedWrap()) | |||
| 2914 | return N0; | |||
| 2915 | ||||
| 2916 | // 0 - X --> X if X is 0 or the minimum signed value. | |||
| 2917 | return N1; | |||
| 2918 | } | |||
| 2919 | } | |||
| 2920 | ||||
| 2921 | // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) | |||
| 2922 | if (isAllOnesOrAllOnesSplat(N0)) | |||
| 2923 | return DAG.getNode(ISD::XOR, DL, VT, N1, N0); | |||
| 2924 | ||||
| 2925 | // fold (A - (0-B)) -> A+B | |||
| 2926 | if (N1.getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0))) | |||
| 2927 | return DAG.getNode(ISD::ADD, DL, VT, N0, N1.getOperand(1)); | |||
| 2928 | ||||
| 2929 | // fold A-(A-B) -> B | |||
| 2930 | if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0)) | |||
| 2931 | return N1.getOperand(1); | |||
| 2932 | ||||
| 2933 | // fold (A+B)-A -> B | |||
| 2934 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) | |||
| 2935 | return N0.getOperand(1); | |||
| 2936 | ||||
| 2937 | // fold (A+B)-B -> A | |||
| 2938 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) | |||
| 2939 | return N0.getOperand(0); | |||
| 2940 | ||||
| 2941 | // fold (A+C1)-C2 -> A+(C1-C2) | |||
| 2942 | if (N0.getOpcode() == ISD::ADD && | |||
| 2943 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && | |||
| 2944 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { | |||
| 2945 | SDValue NewC = DAG.FoldConstantArithmetic( | |||
| 2946 | ISD::SUB, DL, VT, N0.getOperand(1).getNode(), N1.getNode()); | |||
| 2947 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2947, __PRETTY_FUNCTION__)); | |||
| 2948 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), NewC); | |||
| 2949 | } | |||
| 2950 | ||||
| 2951 | // fold C2-(A+C1) -> (C2-C1)-A | |||
| 2952 | if (N1.getOpcode() == ISD::ADD) { | |||
| 2953 | SDValue N11 = N1.getOperand(1); | |||
| 2954 | if (isConstantOrConstantVector(N0, /* NoOpaques */ true) && | |||
| 2955 | isConstantOrConstantVector(N11, /* NoOpaques */ true)) { | |||
| 2956 | SDValue NewC = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, N0.getNode(), | |||
| 2957 | N11.getNode()); | |||
| 2958 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2958, __PRETTY_FUNCTION__)); | |||
| 2959 | return DAG.getNode(ISD::SUB, DL, VT, NewC, N1.getOperand(0)); | |||
| 2960 | } | |||
| 2961 | } | |||
| 2962 | ||||
| 2963 | // fold (A-C1)-C2 -> A-(C1+C2) | |||
| 2964 | if (N0.getOpcode() == ISD::SUB && | |||
| 2965 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && | |||
| 2966 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { | |||
| 2967 | SDValue NewC = DAG.FoldConstantArithmetic( | |||
| 2968 | ISD::ADD, DL, VT, N0.getOperand(1).getNode(), N1.getNode()); | |||
| 2969 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2969, __PRETTY_FUNCTION__)); | |||
| 2970 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), NewC); | |||
| 2971 | } | |||
| 2972 | ||||
| 2973 | // fold (c1-A)-c2 -> (c1-c2)-A | |||
| 2974 | if (N0.getOpcode() == ISD::SUB && | |||
| 2975 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && | |||
| 2976 | isConstantOrConstantVector(N0.getOperand(0), /* NoOpaques */ true)) { | |||
| 2977 | SDValue NewC = DAG.FoldConstantArithmetic( | |||
| 2978 | ISD::SUB, DL, VT, N0.getOperand(0).getNode(), N1.getNode()); | |||
| 2979 | assert(NewC && "Constant folding failed")((NewC && "Constant folding failed") ? static_cast< void> (0) : __assert_fail ("NewC && \"Constant folding failed\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2979, __PRETTY_FUNCTION__)); | |||
| 2980 | return DAG.getNode(ISD::SUB, DL, VT, NewC, N0.getOperand(1)); | |||
| 2981 | } | |||
| 2982 | ||||
| 2983 | // fold ((A+(B+or-C))-B) -> A+or-C | |||
| 2984 | if (N0.getOpcode() == ISD::ADD && | |||
| 2985 | (N0.getOperand(1).getOpcode() == ISD::SUB || | |||
| 2986 | N0.getOperand(1).getOpcode() == ISD::ADD) && | |||
| 2987 | N0.getOperand(1).getOperand(0) == N1) | |||
| 2988 | return DAG.getNode(N0.getOperand(1).getOpcode(), DL, VT, N0.getOperand(0), | |||
| 2989 | N0.getOperand(1).getOperand(1)); | |||
| 2990 | ||||
| 2991 | // fold ((A+(C+B))-B) -> A+C | |||
| 2992 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1).getOpcode() == ISD::ADD && | |||
| 2993 | N0.getOperand(1).getOperand(1) == N1) | |||
| 2994 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), | |||
| 2995 | N0.getOperand(1).getOperand(0)); | |||
| 2996 | ||||
| 2997 | // fold ((A-(B-C))-C) -> A-B | |||
| 2998 | if (N0.getOpcode() == ISD::SUB && N0.getOperand(1).getOpcode() == ISD::SUB && | |||
| 2999 | N0.getOperand(1).getOperand(1) == N1) | |||
| 3000 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), | |||
| 3001 | N0.getOperand(1).getOperand(0)); | |||
| 3002 | ||||
| 3003 | // fold (A-(B-C)) -> A+(C-B) | |||
| 3004 | if (N1.getOpcode() == ISD::SUB && N1.hasOneUse()) | |||
| 3005 | return DAG.getNode(ISD::ADD, DL, VT, N0, | |||
| 3006 | DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(1), | |||
| 3007 | N1.getOperand(0))); | |||
| 3008 | ||||
| 3009 | // fold (X - (-Y * Z)) -> (X + (Y * Z)) | |||
| 3010 | if (N1.getOpcode() == ISD::MUL && N1.hasOneUse()) { | |||
| 3011 | if (N1.getOperand(0).getOpcode() == ISD::SUB && | |||
| 3012 | isNullOrNullSplat(N1.getOperand(0).getOperand(0))) { | |||
| 3013 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, | |||
| 3014 | N1.getOperand(0).getOperand(1), | |||
| 3015 | N1.getOperand(1)); | |||
| 3016 | return DAG.getNode(ISD::ADD, DL, VT, N0, Mul); | |||
| 3017 | } | |||
| 3018 | if (N1.getOperand(1).getOpcode() == ISD::SUB && | |||
| 3019 | isNullOrNullSplat(N1.getOperand(1).getOperand(0))) { | |||
| 3020 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, | |||
| 3021 | N1.getOperand(0), | |||
| 3022 | N1.getOperand(1).getOperand(1)); | |||
| 3023 | return DAG.getNode(ISD::ADD, DL, VT, N0, Mul); | |||
| 3024 | } | |||
| 3025 | } | |||
| 3026 | ||||
| 3027 | // If either operand of a sub is undef, the result is undef | |||
| 3028 | if (N0.isUndef()) | |||
| 3029 | return N0; | |||
| 3030 | if (N1.isUndef()) | |||
| 3031 | return N1; | |||
| 3032 | ||||
| 3033 | if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG)) | |||
| 3034 | return V; | |||
| 3035 | ||||
| 3036 | if (SDValue V = foldAddSubOfSignBit(N, DAG)) | |||
| 3037 | return V; | |||
| 3038 | ||||
| 3039 | if (SDValue V = foldAddSubMasked1(false, N0, N1, DAG, SDLoc(N))) | |||
| 3040 | return V; | |||
| 3041 | ||||
| 3042 | // (x - y) - 1 -> add (xor y, -1), x | |||
| 3043 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && isOneOrOneSplat(N1)) { | |||
| 3044 | SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1), | |||
| 3045 | DAG.getAllOnesConstant(DL, VT)); | |||
| 3046 | return DAG.getNode(ISD::ADD, DL, VT, Xor, N0.getOperand(0)); | |||
| 3047 | } | |||
| 3048 | ||||
| 3049 | // Look for: | |||
| 3050 | // sub y, (xor x, -1) | |||
| 3051 | // And if the target does not like this form then turn into: | |||
| 3052 | // add (add x, y), 1 | |||
| 3053 | if (TLI.preferIncOfAddToSubOfNot(VT) && N1.hasOneUse() && isBitwiseNot(N1)) { | |||
| 3054 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, N1.getOperand(0)); | |||
| 3055 | return DAG.getNode(ISD::ADD, DL, VT, Add, DAG.getConstant(1, DL, VT)); | |||
| 3056 | } | |||
| 3057 | ||||
| 3058 | // Hoist one-use addition by non-opaque constant: | |||
| 3059 | // (x + C) - y -> (x - y) + C | |||
| 3060 | if (N0.hasOneUse() && N0.getOpcode() == ISD::ADD && | |||
| 3061 | isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { | |||
| 3062 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1); | |||
| 3063 | return DAG.getNode(ISD::ADD, DL, VT, Sub, N0.getOperand(1)); | |||
| 3064 | } | |||
| 3065 | // y - (x + C) -> (y - x) - C | |||
| 3066 | if (N1.hasOneUse() && N1.getOpcode() == ISD::ADD && | |||
| 3067 | isConstantOrConstantVector(N1.getOperand(1), /*NoOpaques=*/true)) { | |||
| 3068 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(0)); | |||
| 3069 | return DAG.getNode(ISD::SUB, DL, VT, Sub, N1.getOperand(1)); | |||
| 3070 | } | |||
| 3071 | // (x - C) - y -> (x - y) - C | |||
| 3072 | // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors. | |||
| 3073 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && | |||
| 3074 | isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { | |||
| 3075 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1); | |||
| 3076 | return DAG.getNode(ISD::SUB, DL, VT, Sub, N0.getOperand(1)); | |||
| 3077 | } | |||
| 3078 | // (C - x) - y -> C - (x + y) | |||
| 3079 | if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && | |||
| 3080 | isConstantOrConstantVector(N0.getOperand(0), /*NoOpaques=*/true)) { | |||
| 3081 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(1), N1); | |||
| 3082 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), Add); | |||
| 3083 | } | |||
| 3084 | ||||
| 3085 | // If the target's bool is represented as 0/-1, prefer to make this 'add 0/-1' | |||
| 3086 | // rather than 'sub 0/1' (the sext should get folded). | |||
| 3087 | // sub X, (zext i1 Y) --> add X, (sext i1 Y) | |||
| 3088 | if (N1.getOpcode() == ISD::ZERO_EXTEND && | |||
| 3089 | N1.getOperand(0).getScalarValueSizeInBits() == 1 && | |||
| 3090 | TLI.getBooleanContents(VT) == | |||
| 3091 | TargetLowering::ZeroOrNegativeOneBooleanContent) { | |||
| 3092 | SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, N1.getOperand(0)); | |||
| 3093 | return DAG.getNode(ISD::ADD, DL, VT, N0, SExt); | |||
| 3094 | } | |||
| 3095 | ||||
| 3096 | // fold Y = sra (X, size(X)-1); sub (xor (X, Y), Y) -> (abs X) | |||
| 3097 | if (TLI.isOperationLegalOrCustom(ISD::ABS, VT)) { | |||
| 3098 | if (N0.getOpcode() == ISD::XOR && N1.getOpcode() == ISD::SRA) { | |||
| 3099 | SDValue X0 = N0.getOperand(0), X1 = N0.getOperand(1); | |||
| 3100 | SDValue S0 = N1.getOperand(0); | |||
| 3101 | if ((X0 == S0 && X1 == N1) || (X0 == N1 && X1 == S0)) { | |||
| 3102 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); | |||
| 3103 | if (ConstantSDNode *C = isConstOrConstSplat(N1.getOperand(1))) | |||
| 3104 | if (C->getAPIntValue() == (OpSizeInBits - 1)) | |||
| 3105 | return DAG.getNode(ISD::ABS, SDLoc(N), VT, S0); | |||
| 3106 | } | |||
| 3107 | } | |||
| 3108 | } | |||
| 3109 | ||||
| 3110 | // If the relocation model supports it, consider symbol offsets. | |||
| 3111 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) | |||
| 3112 | if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) { | |||
| 3113 | // fold (sub Sym, c) -> Sym-c | |||
| 3114 | if (N1C && GA->getOpcode() == ISD::GlobalAddress) | |||
| 3115 | return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, | |||
| 3116 | GA->getOffset() - | |||
| 3117 | (uint64_t)N1C->getSExtValue()); | |||
| 3118 | // fold (sub Sym+c1, Sym+c2) -> c1-c2 | |||
| 3119 | if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1)) | |||
| 3120 | if (GA->getGlobal() == GB->getGlobal()) | |||
| 3121 | return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(), | |||
| 3122 | DL, VT); | |||
| 3123 | } | |||
| 3124 | ||||
| 3125 | // sub X, (sextinreg Y i1) -> add X, (and Y 1) | |||
| 3126 | if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { | |||
| 3127 | VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1)); | |||
| 3128 | if (TN->getVT() == MVT::i1) { | |||
| 3129 | SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), | |||
| 3130 | DAG.getConstant(1, DL, VT)); | |||
| 3131 | return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt); | |||
| 3132 | } | |||
| 3133 | } | |||
| 3134 | ||||
| 3135 | // Prefer an add for more folding potential and possibly better codegen: | |||
| 3136 | // sub N0, (lshr N10, width-1) --> add N0, (ashr N10, width-1) | |||
| 3137 | if (!LegalOperations && N1.getOpcode() == ISD::SRL && N1.hasOneUse()) { | |||
| 3138 | SDValue ShAmt = N1.getOperand(1); | |||
| 3139 | ConstantSDNode *ShAmtC = isConstOrConstSplat(ShAmt); | |||
| 3140 | if (ShAmtC && | |||
| 3141 | ShAmtC->getAPIntValue() == (N1.getScalarValueSizeInBits() - 1)) { | |||
| 3142 | SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, N1.getOperand(0), ShAmt); | |||
| 3143 | return DAG.getNode(ISD::ADD, DL, VT, N0, SRA); | |||
| 3144 | } | |||
| 3145 | } | |||
| 3146 | ||||
| 3147 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) { | |||
| 3148 | // (sub Carry, X) -> (addcarry (sub 0, X), 0, Carry) | |||
| 3149 | if (SDValue Carry = getAsCarry(TLI, N0)) { | |||
| 3150 | SDValue X = N1; | |||
| 3151 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
| 3152 | SDValue NegX = DAG.getNode(ISD::SUB, DL, VT, Zero, X); | |||
| 3153 | return DAG.getNode(ISD::ADDCARRY, DL, | |||
| 3154 | DAG.getVTList(VT, Carry.getValueType()), NegX, Zero, | |||
| 3155 | Carry); | |||
| 3156 | } | |||
| 3157 | } | |||
| 3158 | ||||
| 3159 | return SDValue(); | |||
| 3160 | } | |||
| 3161 | ||||
| 3162 | SDValue DAGCombiner::visitSUBSAT(SDNode *N) { | |||
| 3163 | SDValue N0 = N->getOperand(0); | |||
| 3164 | SDValue N1 = N->getOperand(1); | |||
| 3165 | EVT VT = N0.getValueType(); | |||
| 3166 | SDLoc DL(N); | |||
| 3167 | ||||
| 3168 | // fold vector ops | |||
| 3169 | if (VT.isVector()) { | |||
| 3170 | // TODO SimplifyVBinOp | |||
| 3171 | ||||
| 3172 | // fold (sub_sat x, 0) -> x, vector edition | |||
| 3173 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 3174 | return N0; | |||
| 3175 | } | |||
| 3176 | ||||
| 3177 | // fold (sub_sat x, undef) -> 0 | |||
| 3178 | if (N0.isUndef() || N1.isUndef()) | |||
| 3179 | return DAG.getConstant(0, DL, VT); | |||
| 3180 | ||||
| 3181 | // fold (sub_sat x, x) -> 0 | |||
| 3182 | if (N0 == N1) | |||
| 3183 | return DAG.getConstant(0, DL, VT); | |||
| 3184 | ||||
| 3185 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 3186 | DAG.isConstantIntBuildVectorOrConstantInt(N1)) { | |||
| 3187 | // fold (sub_sat c1, c2) -> c3 | |||
| 3188 | return DAG.FoldConstantArithmetic(N->getOpcode(), DL, VT, N0.getNode(), | |||
| 3189 | N1.getNode()); | |||
| 3190 | } | |||
| 3191 | ||||
| 3192 | // fold (sub_sat x, 0) -> x | |||
| 3193 | if (isNullConstant(N1)) | |||
| 3194 | return N0; | |||
| 3195 | ||||
| 3196 | return SDValue(); | |||
| 3197 | } | |||
| 3198 | ||||
| 3199 | SDValue DAGCombiner::visitSUBC(SDNode *N) { | |||
| 3200 | SDValue N0 = N->getOperand(0); | |||
| 3201 | SDValue N1 = N->getOperand(1); | |||
| 3202 | EVT VT = N0.getValueType(); | |||
| 3203 | SDLoc DL(N); | |||
| 3204 | ||||
| 3205 | // If the flag result is dead, turn this into an SUB. | |||
| 3206 | if (!N->hasAnyUseOfValue(1)) | |||
| 3207 | return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), | |||
| 3208 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
| 3209 | ||||
| 3210 | // fold (subc x, x) -> 0 + no borrow | |||
| 3211 | if (N0 == N1) | |||
| 3212 | return CombineTo(N, DAG.getConstant(0, DL, VT), | |||
| 3213 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
| 3214 | ||||
| 3215 | // fold (subc x, 0) -> x + no borrow | |||
| 3216 | if (isNullConstant(N1)) | |||
| 3217 | return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
| 3218 | ||||
| 3219 | // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow | |||
| 3220 | if (isAllOnesConstant(N0)) | |||
| 3221 | return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), | |||
| 3222 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
| 3223 | ||||
| 3224 | return SDValue(); | |||
| 3225 | } | |||
| 3226 | ||||
| 3227 | SDValue DAGCombiner::visitSUBO(SDNode *N) { | |||
| 3228 | SDValue N0 = N->getOperand(0); | |||
| 3229 | SDValue N1 = N->getOperand(1); | |||
| 3230 | EVT VT = N0.getValueType(); | |||
| 3231 | bool IsSigned = (ISD::SSUBO == N->getOpcode()); | |||
| 3232 | ||||
| 3233 | EVT CarryVT = N->getValueType(1); | |||
| 3234 | SDLoc DL(N); | |||
| 3235 | ||||
| 3236 | // If the flag result is dead, turn this into an SUB. | |||
| 3237 | if (!N->hasAnyUseOfValue(1)) | |||
| 3238 | return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), | |||
| 3239 | DAG.getUNDEF(CarryVT)); | |||
| 3240 | ||||
| 3241 | // fold (subo x, x) -> 0 + no borrow | |||
| 3242 | if (N0 == N1) | |||
| 3243 | return CombineTo(N, DAG.getConstant(0, DL, VT), | |||
| 3244 | DAG.getConstant(0, DL, CarryVT)); | |||
| 3245 | ||||
| 3246 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); | |||
| 3247 | ||||
| 3248 | // fold (subox, c) -> (addo x, -c) | |||
| 3249 | if (IsSigned && N1C && !N1C->getAPIntValue().isMinSignedValue()) { | |||
| 3250 | return DAG.getNode(ISD::SADDO, DL, N->getVTList(), N0, | |||
| 3251 | DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); | |||
| 3252 | } | |||
| 3253 | ||||
| 3254 | // fold (subo x, 0) -> x + no borrow | |||
| 3255 | if (isNullOrNullSplat(N1)) | |||
| 3256 | return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); | |||
| 3257 | ||||
| 3258 | // Canonicalize (usubo -1, x) -> ~x, i.e. (xor x, -1) + no borrow | |||
| 3259 | if (!IsSigned && isAllOnesOrAllOnesSplat(N0)) | |||
| 3260 | return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), | |||
| 3261 | DAG.getConstant(0, DL, CarryVT)); | |||
| 3262 | ||||
| 3263 | return SDValue(); | |||
| 3264 | } | |||
| 3265 | ||||
| 3266 | SDValue DAGCombiner::visitSUBE(SDNode *N) { | |||
| 3267 | SDValue N0 = N->getOperand(0); | |||
| 3268 | SDValue N1 = N->getOperand(1); | |||
| 3269 | SDValue CarryIn = N->getOperand(2); | |||
| 3270 | ||||
| 3271 | // fold (sube x, y, false) -> (subc x, y) | |||
| 3272 | if (CarryIn.getOpcode() == ISD::CARRY_FALSE) | |||
| 3273 | return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1); | |||
| 3274 | ||||
| 3275 | return SDValue(); | |||
| 3276 | } | |||
| 3277 | ||||
| 3278 | SDValue DAGCombiner::visitSUBCARRY(SDNode *N) { | |||
| 3279 | SDValue N0 = N->getOperand(0); | |||
| 3280 | SDValue N1 = N->getOperand(1); | |||
| 3281 | SDValue CarryIn = N->getOperand(2); | |||
| 3282 | ||||
| 3283 | // fold (subcarry x, y, false) -> (usubo x, y) | |||
| 3284 | if (isNullConstant(CarryIn)) { | |||
| 3285 | if (!LegalOperations || | |||
| 3286 | TLI.isOperationLegalOrCustom(ISD::USUBO, N->getValueType(0))) | |||
| 3287 | return DAG.getNode(ISD::USUBO, SDLoc(N), N->getVTList(), N0, N1); | |||
| 3288 | } | |||
| 3289 | ||||
| 3290 | return SDValue(); | |||
| 3291 | } | |||
| 3292 | ||||
| 3293 | // Notice that "mulfix" can be any of SMULFIX, SMULFIXSAT, UMULFIX and | |||
| 3294 | // UMULFIXSAT here. | |||
| 3295 | SDValue DAGCombiner::visitMULFIX(SDNode *N) { | |||
| 3296 | SDValue N0 = N->getOperand(0); | |||
| 3297 | SDValue N1 = N->getOperand(1); | |||
| 3298 | SDValue Scale = N->getOperand(2); | |||
| 3299 | EVT VT = N0.getValueType(); | |||
| 3300 | ||||
| 3301 | // fold (mulfix x, undef, scale) -> 0 | |||
| 3302 | if (N0.isUndef() || N1.isUndef()) | |||
| 3303 | return DAG.getConstant(0, SDLoc(N), VT); | |||
| 3304 | ||||
| 3305 | // Canonicalize constant to RHS (vector doesn't have to splat) | |||
| 3306 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 3307 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 3308 | return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0, Scale); | |||
| 3309 | ||||
| 3310 | // fold (mulfix x, 0, scale) -> 0 | |||
| 3311 | if (isNullConstant(N1)) | |||
| 3312 | return DAG.getConstant(0, SDLoc(N), VT); | |||
| 3313 | ||||
| 3314 | return SDValue(); | |||
| 3315 | } | |||
| 3316 | ||||
| 3317 | SDValue DAGCombiner::visitMUL(SDNode *N) { | |||
| 3318 | SDValue N0 = N->getOperand(0); | |||
| 3319 | SDValue N1 = N->getOperand(1); | |||
| 3320 | EVT VT = N0.getValueType(); | |||
| 3321 | ||||
| 3322 | // fold (mul x, undef) -> 0 | |||
| 3323 | if (N0.isUndef() || N1.isUndef()) | |||
| 3324 | return DAG.getConstant(0, SDLoc(N), VT); | |||
| 3325 | ||||
| 3326 | bool N0IsConst = false; | |||
| 3327 | bool N1IsConst = false; | |||
| 3328 | bool N1IsOpaqueConst = false; | |||
| 3329 | bool N0IsOpaqueConst = false; | |||
| 3330 | APInt ConstValue0, ConstValue1; | |||
| 3331 | // fold vector ops | |||
| 3332 | if (VT.isVector()) { | |||
| 3333 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 3334 | return FoldedVOp; | |||
| 3335 | ||||
| 3336 | N0IsConst = ISD::isConstantSplatVector(N0.getNode(), ConstValue0); | |||
| 3337 | N1IsConst = ISD::isConstantSplatVector(N1.getNode(), ConstValue1); | |||
| 3338 | assert((!N0IsConst ||(((!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits ()) && "Splat APInt should be element width") ? static_cast <void> (0) : __assert_fail ("(!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3340, __PRETTY_FUNCTION__)) | |||
| 3339 | ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) &&(((!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits ()) && "Splat APInt should be element width") ? static_cast <void> (0) : __assert_fail ("(!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3340, __PRETTY_FUNCTION__)) | |||
| 3340 | "Splat APInt should be element width")(((!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits ()) && "Splat APInt should be element width") ? static_cast <void> (0) : __assert_fail ("(!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3340, __PRETTY_FUNCTION__)); | |||
| 3341 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3343, __PRETTY_FUNCTION__)) | |||
| 3342 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3343, __PRETTY_FUNCTION__)) | |||
| 3343 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3343, __PRETTY_FUNCTION__)); | |||
| 3344 | } else { | |||
| 3345 | N0IsConst = isa<ConstantSDNode>(N0); | |||
| 3346 | if (N0IsConst) { | |||
| 3347 | ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue(); | |||
| 3348 | N0IsOpaqueConst = cast<ConstantSDNode>(N0)->isOpaque(); | |||
| 3349 | } | |||
| 3350 | N1IsConst = isa<ConstantSDNode>(N1); | |||
| 3351 | if (N1IsConst) { | |||
| 3352 | ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue(); | |||
| 3353 | N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque(); | |||
| 3354 | } | |||
| 3355 | } | |||
| 3356 | ||||
| 3357 | // fold (mul c1, c2) -> c1*c2 | |||
| 3358 | if (N0IsConst && N1IsConst && !N0IsOpaqueConst && !N1IsOpaqueConst) | |||
| 3359 | return DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT, | |||
| 3360 | N0.getNode(), N1.getNode()); | |||
| 3361 | ||||
| 3362 | // canonicalize constant to RHS (vector doesn't have to splat) | |||
| 3363 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 3364 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 3365 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0); | |||
| 3366 | // fold (mul x, 0) -> 0 | |||
| 3367 | if (N1IsConst && ConstValue1.isNullValue()) | |||
| 3368 | return N1; | |||
| 3369 | // fold (mul x, 1) -> x | |||
| 3370 | if (N1IsConst && ConstValue1.isOneValue()) | |||
| 3371 | return N0; | |||
| 3372 | ||||
| 3373 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 3374 | return NewSel; | |||
| 3375 | ||||
| 3376 | // fold (mul x, -1) -> 0-x | |||
| 3377 | if (N1IsConst && ConstValue1.isAllOnesValue()) { | |||
| 3378 | SDLoc DL(N); | |||
| 3379 | return DAG.getNode(ISD::SUB, DL, VT, | |||
| 3380 | DAG.getConstant(0, DL, VT), N0); | |||
| 3381 | } | |||
| 3382 | // fold (mul x, (1 << c)) -> x << c | |||
| 3383 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && | |||
| 3384 | DAG.isKnownToBeAPowerOfTwo(N1) && | |||
| 3385 | (!VT.isVector() || Level <= AfterLegalizeVectorOps)) { | |||
| 3386 | SDLoc DL(N); | |||
| 3387 | SDValue LogBase2 = BuildLogBase2(N1, DL); | |||
| 3388 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); | |||
| 3389 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); | |||
| 3390 | return DAG.getNode(ISD::SHL, DL, VT, N0, Trunc); | |||
| 3391 | } | |||
| 3392 | // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c | |||
| 3393 | if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2()) { | |||
| 3394 | unsigned Log2Val = (-ConstValue1).logBase2(); | |||
| 3395 | SDLoc DL(N); | |||
| 3396 | // FIXME: If the input is something that is easily negated (e.g. a | |||
| 3397 | // single-use add), we should put the negate there. | |||
| 3398 | return DAG.getNode(ISD::SUB, DL, VT, | |||
| 3399 | DAG.getConstant(0, DL, VT), | |||
| 3400 | DAG.getNode(ISD::SHL, DL, VT, N0, | |||
| 3401 | DAG.getConstant(Log2Val, DL, | |||
| 3402 | getShiftAmountTy(N0.getValueType())))); | |||
| 3403 | } | |||
| 3404 | ||||
| 3405 | // Try to transform multiply-by-(power-of-2 +/- 1) into shift and add/sub. | |||
| 3406 | // mul x, (2^N + 1) --> add (shl x, N), x | |||
| 3407 | // mul x, (2^N - 1) --> sub (shl x, N), x | |||
| 3408 | // Examples: x * 33 --> (x << 5) + x | |||
| 3409 | // x * 15 --> (x << 4) - x | |||
| 3410 | // x * -33 --> -((x << 5) + x) | |||
| 3411 | // x * -15 --> -((x << 4) - x) ; this reduces --> x - (x << 4) | |||
| 3412 | if (N1IsConst && TLI.decomposeMulByConstant(*DAG.getContext(), VT, N1)) { | |||
| 3413 | // TODO: We could handle more general decomposition of any constant by | |||
| 3414 | // having the target set a limit on number of ops and making a | |||
| 3415 | // callback to determine that sequence (similar to sqrt expansion). | |||
| 3416 | unsigned MathOp = ISD::DELETED_NODE; | |||
| 3417 | APInt MulC = ConstValue1.abs(); | |||
| 3418 | if ((MulC - 1).isPowerOf2()) | |||
| 3419 | MathOp = ISD::ADD; | |||
| 3420 | else if ((MulC + 1).isPowerOf2()) | |||
| 3421 | MathOp = ISD::SUB; | |||
| 3422 | ||||
| 3423 | if (MathOp != ISD::DELETED_NODE) { | |||
| 3424 | unsigned ShAmt = | |||
| 3425 | MathOp == ISD::ADD ? (MulC - 1).logBase2() : (MulC + 1).logBase2(); | |||
| 3426 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3427, __PRETTY_FUNCTION__)) | |||
| 3427 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3427, __PRETTY_FUNCTION__)); | |||
| 3428 | SDLoc DL(N); | |||
| 3429 | SDValue Shl = | |||
| 3430 | DAG.getNode(ISD::SHL, DL, VT, N0, DAG.getConstant(ShAmt, DL, VT)); | |||
| 3431 | SDValue R = DAG.getNode(MathOp, DL, VT, Shl, N0); | |||
| 3432 | if (ConstValue1.isNegative()) | |||
| 3433 | R = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), R); | |||
| 3434 | return R; | |||
| 3435 | } | |||
| 3436 | } | |||
| 3437 | ||||
| 3438 | // (mul (shl X, c1), c2) -> (mul X, c2 << c1) | |||
| 3439 | if (N0.getOpcode() == ISD::SHL && | |||
| 3440 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && | |||
| 3441 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { | |||
| 3442 | SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT, N1, N0.getOperand(1)); | |||
| 3443 | if (isConstantOrConstantVector(C3)) | |||
| 3444 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), C3); | |||
| 3445 | } | |||
| 3446 | ||||
| 3447 | // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one | |||
| 3448 | // use. | |||
| 3449 | { | |||
| 3450 | SDValue Sh(nullptr, 0), Y(nullptr, 0); | |||
| 3451 | ||||
| 3452 | // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). | |||
| 3453 | if (N0.getOpcode() == ISD::SHL && | |||
| 3454 | isConstantOrConstantVector(N0.getOperand(1)) && | |||
| 3455 | N0.getNode()->hasOneUse()) { | |||
| 3456 | Sh = N0; Y = N1; | |||
| 3457 | } else if (N1.getOpcode() == ISD::SHL && | |||
| 3458 | isConstantOrConstantVector(N1.getOperand(1)) && | |||
| 3459 | N1.getNode()->hasOneUse()) { | |||
| 3460 | Sh = N1; Y = N0; | |||
| 3461 | } | |||
| 3462 | ||||
| 3463 | if (Sh.getNode()) { | |||
| 3464 | SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, Sh.getOperand(0), Y); | |||
| 3465 | return DAG.getNode(ISD::SHL, SDLoc(N), VT, Mul, Sh.getOperand(1)); | |||
| 3466 | } | |||
| 3467 | } | |||
| 3468 | ||||
| 3469 | // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) | |||
| 3470 | if (DAG.isConstantIntBuildVectorOrConstantInt(N1) && | |||
| 3471 | N0.getOpcode() == ISD::ADD && | |||
| 3472 | DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)) && | |||
| 3473 | isMulAddWithConstProfitable(N, N0, N1)) | |||
| 3474 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, | |||
| 3475 | DAG.getNode(ISD::MUL, SDLoc(N0), VT, | |||
| 3476 | N0.getOperand(0), N1), | |||
| 3477 | DAG.getNode(ISD::MUL, SDLoc(N1), VT, | |||
| 3478 | N0.getOperand(1), N1)); | |||
| 3479 | ||||
| 3480 | // reassociate mul | |||
| 3481 | if (SDValue RMUL = reassociateOps(ISD::MUL, SDLoc(N), N0, N1, N->getFlags())) | |||
| 3482 | return RMUL; | |||
| 3483 | ||||
| 3484 | return SDValue(); | |||
| 3485 | } | |||
| 3486 | ||||
| 3487 | /// Return true if divmod libcall is available. | |||
| 3488 | static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned, | |||
| 3489 | const TargetLowering &TLI) { | |||
| 3490 | RTLIB::Libcall LC; | |||
| 3491 | EVT NodeType = Node->getValueType(0); | |||
| 3492 | if (!NodeType.isSimple()) | |||
| 3493 | return false; | |||
| 3494 | switch (NodeType.getSimpleVT().SimpleTy) { | |||
| 3495 | default: return false; // No libcall for vector types. | |||
| 3496 | case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break; | |||
| 3497 | case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break; | |||
| 3498 | case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break; | |||
| 3499 | case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break; | |||
| 3500 | case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break; | |||
| 3501 | } | |||
| 3502 | ||||
| 3503 | return TLI.getLibcallName(LC) != nullptr; | |||
| 3504 | } | |||
| 3505 | ||||
| 3506 | /// Issue divrem if both quotient and remainder are needed. | |||
| 3507 | SDValue DAGCombiner::useDivRem(SDNode *Node) { | |||
| 3508 | if (Node->use_empty()) | |||
| 3509 | return SDValue(); // This is a dead node, leave it alone. | |||
| 3510 | ||||
| 3511 | unsigned Opcode = Node->getOpcode(); | |||
| 3512 | bool isSigned = (Opcode == ISD::SDIV) || (Opcode == ISD::SREM); | |||
| 3513 | unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; | |||
| 3514 | ||||
| 3515 | // DivMod lib calls can still work on non-legal types if using lib-calls. | |||
| 3516 | EVT VT = Node->getValueType(0); | |||
| 3517 | if (VT.isVector() || !VT.isInteger()) | |||
| 3518 | return SDValue(); | |||
| 3519 | ||||
| 3520 | if (!TLI.isTypeLegal(VT) && !TLI.isOperationCustom(DivRemOpc, VT)) | |||
| 3521 | return SDValue(); | |||
| 3522 | ||||
| 3523 | // If DIVREM is going to get expanded into a libcall, | |||
| 3524 | // but there is no libcall available, then don't combine. | |||
| 3525 | if (!TLI.isOperationLegalOrCustom(DivRemOpc, VT) && | |||
| 3526 | !isDivRemLibcallAvailable(Node, isSigned, TLI)) | |||
| 3527 | return SDValue(); | |||
| 3528 | ||||
| 3529 | // If div is legal, it's better to do the normal expansion | |||
| 3530 | unsigned OtherOpcode = 0; | |||
| 3531 | if ((Opcode == ISD::SDIV) || (Opcode == ISD::UDIV)) { | |||
| 3532 | OtherOpcode = isSigned ? ISD::SREM : ISD::UREM; | |||
| 3533 | if (TLI.isOperationLegalOrCustom(Opcode, VT)) | |||
| 3534 | return SDValue(); | |||
| 3535 | } else { | |||
| 3536 | OtherOpcode = isSigned ? ISD::SDIV : ISD::UDIV; | |||
| 3537 | if (TLI.isOperationLegalOrCustom(OtherOpcode, VT)) | |||
| 3538 | return SDValue(); | |||
| 3539 | } | |||
| 3540 | ||||
| 3541 | SDValue Op0 = Node->getOperand(0); | |||
| 3542 | SDValue Op1 = Node->getOperand(1); | |||
| 3543 | SDValue combined; | |||
| 3544 | for (SDNode::use_iterator UI = Op0.getNode()->use_begin(), | |||
| 3545 | UE = Op0.getNode()->use_end(); UI != UE; ++UI) { | |||
| 3546 | SDNode *User = *UI; | |||
| 3547 | if (User == Node || User->getOpcode() == ISD::DELETED_NODE || | |||
| 3548 | User->use_empty()) | |||
| 3549 | continue; | |||
| 3550 | // Convert the other matching node(s), too; | |||
| 3551 | // otherwise, the DIVREM may get target-legalized into something | |||
| 3552 | // target-specific that we won't be able to recognize. | |||
| 3553 | unsigned UserOpc = User->getOpcode(); | |||
| 3554 | if ((UserOpc == Opcode || UserOpc == OtherOpcode || UserOpc == DivRemOpc) && | |||
| 3555 | User->getOperand(0) == Op0 && | |||
| 3556 | User->getOperand(1) == Op1) { | |||
| 3557 | if (!combined) { | |||
| 3558 | if (UserOpc == OtherOpcode) { | |||
| 3559 | SDVTList VTs = DAG.getVTList(VT, VT); | |||
| 3560 | combined = DAG.getNode(DivRemOpc, SDLoc(Node), VTs, Op0, Op1); | |||
| 3561 | } else if (UserOpc == DivRemOpc) { | |||
| 3562 | combined = SDValue(User, 0); | |||
| 3563 | } else { | |||
| 3564 | assert(UserOpc == Opcode)((UserOpc == Opcode) ? static_cast<void> (0) : __assert_fail ("UserOpc == Opcode", "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3564, __PRETTY_FUNCTION__)); | |||
| 3565 | continue; | |||
| 3566 | } | |||
| 3567 | } | |||
| 3568 | if (UserOpc == ISD::SDIV || UserOpc == ISD::UDIV) | |||
| 3569 | CombineTo(User, combined); | |||
| 3570 | else if (UserOpc == ISD::SREM || UserOpc == ISD::UREM) | |||
| 3571 | CombineTo(User, combined.getValue(1)); | |||
| 3572 | } | |||
| 3573 | } | |||
| 3574 | return combined; | |||
| 3575 | } | |||
| 3576 | ||||
| 3577 | static SDValue simplifyDivRem(SDNode *N, SelectionDAG &DAG) { | |||
| 3578 | SDValue N0 = N->getOperand(0); | |||
| 3579 | SDValue N1 = N->getOperand(1); | |||
| 3580 | EVT VT = N->getValueType(0); | |||
| 3581 | SDLoc DL(N); | |||
| 3582 | ||||
| 3583 | unsigned Opc = N->getOpcode(); | |||
| 3584 | bool IsDiv = (ISD::SDIV == Opc) || (ISD::UDIV == Opc); | |||
| 3585 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
| 3586 | ||||
| 3587 | // X / undef -> undef | |||
| 3588 | // X % undef -> undef | |||
| 3589 | // X / 0 -> undef | |||
| 3590 | // X % 0 -> undef | |||
| 3591 | // NOTE: This includes vectors where any divisor element is zero/undef. | |||
| 3592 | if (DAG.isUndef(Opc, {N0, N1})) | |||
| 3593 | return DAG.getUNDEF(VT); | |||
| 3594 | ||||
| 3595 | // undef / X -> 0 | |||
| 3596 | // undef % X -> 0 | |||
| 3597 | if (N0.isUndef()) | |||
| 3598 | return DAG.getConstant(0, DL, VT); | |||
| 3599 | ||||
| 3600 | // 0 / X -> 0 | |||
| 3601 | // 0 % X -> 0 | |||
| 3602 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
| 3603 | if (N0C && N0C->isNullValue()) | |||
| 3604 | return N0; | |||
| 3605 | ||||
| 3606 | // X / X -> 1 | |||
| 3607 | // X % X -> 0 | |||
| 3608 | if (N0 == N1) | |||
| 3609 | return DAG.getConstant(IsDiv ? 1 : 0, DL, VT); | |||
| 3610 | ||||
| 3611 | // X / 1 -> X | |||
| 3612 | // X % 1 -> 0 | |||
| 3613 | // If this is a boolean op (single-bit element type), we can't have | |||
| 3614 | // division-by-zero or remainder-by-zero, so assume the divisor is 1. | |||
| 3615 | // TODO: Similarly, if we're zero-extending a boolean divisor, then assume | |||
| 3616 | // it's a 1. | |||
| 3617 | if ((N1C && N1C->isOne()) || (VT.getScalarType() == MVT::i1)) | |||
| 3618 | return IsDiv ? N0 : DAG.getConstant(0, DL, VT); | |||
| 3619 | ||||
| 3620 | return SDValue(); | |||
| 3621 | } | |||
| 3622 | ||||
| 3623 | SDValue DAGCombiner::visitSDIV(SDNode *N) { | |||
| 3624 | SDValue N0 = N->getOperand(0); | |||
| 3625 | SDValue N1 = N->getOperand(1); | |||
| 3626 | EVT VT = N->getValueType(0); | |||
| 3627 | EVT CCVT = getSetCCResultType(VT); | |||
| 3628 | ||||
| 3629 | // fold vector ops | |||
| 3630 | if (VT.isVector()) | |||
| 3631 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 3632 | return FoldedVOp; | |||
| 3633 | ||||
| 3634 | SDLoc DL(N); | |||
| 3635 | ||||
| 3636 | // fold (sdiv c1, c2) -> c1/c2 | |||
| 3637 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
| 3638 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
| 3639 | if (N0C && N1C && !N0C->isOpaque() && !N1C->isOpaque()) | |||
| 3640 | return DAG.FoldConstantArithmetic(ISD::SDIV, DL, VT, N0C, N1C); | |||
| 3641 | // fold (sdiv X, -1) -> 0-X | |||
| 3642 | if (N1C && N1C->isAllOnesValue()) | |||
| 3643 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0); | |||
| 3644 | // fold (sdiv X, MIN_SIGNED) -> select(X == MIN_SIGNED, 1, 0) | |||
| 3645 | if (N1C && N1C->getAPIntValue().isMinSignedValue()) | |||
| 3646 | return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), | |||
| 3647 | DAG.getConstant(1, DL, VT), | |||
| 3648 | DAG.getConstant(0, DL, VT)); | |||
| 3649 | ||||
| 3650 | if (SDValue V = simplifyDivRem(N, DAG)) | |||
| 3651 | return V; | |||
| 3652 | ||||
| 3653 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 3654 | return NewSel; | |||
| 3655 | ||||
| 3656 | // If we know the sign bits of both operands are zero, strength reduce to a | |||
| 3657 | // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 | |||
| 3658 | if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) | |||
| 3659 | return DAG.getNode(ISD::UDIV, DL, N1.getValueType(), N0, N1); | |||
| 3660 | ||||
| 3661 | if (SDValue V = visitSDIVLike(N0, N1, N)) { | |||
| 3662 | // If the corresponding remainder node exists, update its users with | |||
| 3663 | // (Dividend - (Quotient * Divisor). | |||
| 3664 | if (SDNode *RemNode = DAG.getNodeIfExists(ISD::SREM, N->getVTList(), | |||
| 3665 | { N0, N1 })) { | |||
| 3666 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, V, N1); | |||
| 3667 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); | |||
| 3668 | AddToWorklist(Mul.getNode()); | |||
| 3669 | AddToWorklist(Sub.getNode()); | |||
| 3670 | CombineTo(RemNode, Sub); | |||
| 3671 | } | |||
| 3672 | return V; | |||
| 3673 | } | |||
| 3674 | ||||
| 3675 | // sdiv, srem -> sdivrem | |||
| 3676 | // If the divisor is constant, then return DIVREM only if isIntDivCheap() is | |||
| 3677 | // true. Otherwise, we break the simplification logic in visitREM(). | |||
| 3678 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
| 3679 | if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
| 3680 | if (SDValue DivRem = useDivRem(N)) | |||
| 3681 | return DivRem; | |||
| 3682 | ||||
| 3683 | return SDValue(); | |||
| 3684 | } | |||
| 3685 | ||||
| 3686 | SDValue DAGCombiner::visitSDIVLike(SDValue N0, SDValue N1, SDNode *N) { | |||
| 3687 | SDLoc DL(N); | |||
| 3688 | EVT VT = N->getValueType(0); | |||
| 3689 | EVT CCVT = getSetCCResultType(VT); | |||
| 3690 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
| 3691 | ||||
| 3692 | // Helper for determining whether a value is a power-2 constant scalar or a | |||
| 3693 | // vector of such elements. | |||
| 3694 | auto IsPowerOfTwo = [](ConstantSDNode *C) { | |||
| 3695 | if (C->isNullValue() || C->isOpaque()) | |||
| 3696 | return false; | |||
| 3697 | if (C->getAPIntValue().isPowerOf2()) | |||
| 3698 | return true; | |||
| 3699 | if ((-C->getAPIntValue()).isPowerOf2()) | |||
| 3700 | return true; | |||
| 3701 | return false; | |||
| 3702 | }; | |||
| 3703 | ||||
| 3704 | // fold (sdiv X, pow2) -> simple ops after legalize | |||
| 3705 | // FIXME: We check for the exact bit here because the generic lowering gives | |||
| 3706 | // better results in that case. The target-specific lowering should learn how | |||
| 3707 | // to handle exact sdivs efficiently. | |||
| 3708 | if (!N->getFlags().hasExact() && ISD::matchUnaryPredicate(N1, IsPowerOfTwo)) { | |||
| 3709 | // Target-specific implementation of sdiv x, pow2. | |||
| 3710 | if (SDValue Res = BuildSDIVPow2(N)) | |||
| 3711 | return Res; | |||
| 3712 | ||||
| 3713 | // Create constants that are functions of the shift amount value. | |||
| 3714 | EVT ShiftAmtTy = getShiftAmountTy(N0.getValueType()); | |||
| 3715 | SDValue Bits = DAG.getConstant(BitWidth, DL, ShiftAmtTy); | |||
| 3716 | SDValue C1 = DAG.getNode(ISD::CTTZ, DL, VT, N1); | |||
| 3717 | C1 = DAG.getZExtOrTrunc(C1, DL, ShiftAmtTy); | |||
| 3718 | SDValue Inexact = DAG.getNode(ISD::SUB, DL, ShiftAmtTy, Bits, C1); | |||
| 3719 | if (!isConstantOrConstantVector(Inexact)) | |||
| 3720 | return SDValue(); | |||
| 3721 | ||||
| 3722 | // Splat the sign bit into the register | |||
| 3723 | SDValue Sign = DAG.getNode(ISD::SRA, DL, VT, N0, | |||
| 3724 | DAG.getConstant(BitWidth - 1, DL, ShiftAmtTy)); | |||
| 3725 | AddToWorklist(Sign.getNode()); | |||
| 3726 | ||||
| 3727 | // Add (N0 < 0) ? abs2 - 1 : 0; | |||
| 3728 | SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, Sign, Inexact); | |||
| 3729 | AddToWorklist(Srl.getNode()); | |||
| 3730 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Srl); | |||
| 3731 | AddToWorklist(Add.getNode()); | |||
| 3732 | SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Add, C1); | |||
| 3733 | AddToWorklist(Sra.getNode()); | |||
| 3734 | ||||
| 3735 | // Special case: (sdiv X, 1) -> X | |||
| 3736 | // Special Case: (sdiv X, -1) -> 0-X | |||
| 3737 | SDValue One = DAG.getConstant(1, DL, VT); | |||
| 3738 | SDValue AllOnes = DAG.getAllOnesConstant(DL, VT); | |||
| 3739 | SDValue IsOne = DAG.getSetCC(DL, CCVT, N1, One, ISD::SETEQ); | |||
| 3740 | SDValue IsAllOnes = DAG.getSetCC(DL, CCVT, N1, AllOnes, ISD::SETEQ); | |||
| 3741 | SDValue IsOneOrAllOnes = DAG.getNode(ISD::OR, DL, CCVT, IsOne, IsAllOnes); | |||
| 3742 | Sra = DAG.getSelect(DL, VT, IsOneOrAllOnes, N0, Sra); | |||
| 3743 | ||||
| 3744 | // If dividing by a positive value, we're done. Otherwise, the result must | |||
| 3745 | // be negated. | |||
| 3746 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
| 3747 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, Zero, Sra); | |||
| 3748 | ||||
| 3749 | // FIXME: Use SELECT_CC once we improve SELECT_CC constant-folding. | |||
| 3750 | SDValue IsNeg = DAG.getSetCC(DL, CCVT, N1, Zero, ISD::SETLT); | |||
| 3751 | SDValue Res = DAG.getSelect(DL, VT, IsNeg, Sub, Sra); | |||
| 3752 | return Res; | |||
| 3753 | } | |||
| 3754 | ||||
| 3755 | // If integer divide is expensive and we satisfy the requirements, emit an | |||
| 3756 | // alternate sequence. Targets may check function attributes for size/speed | |||
| 3757 | // trade-offs. | |||
| 3758 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
| 3759 | if (isConstantOrConstantVector(N1) && | |||
| 3760 | !TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
| 3761 | if (SDValue Op = BuildSDIV(N)) | |||
| 3762 | return Op; | |||
| 3763 | ||||
| 3764 | return SDValue(); | |||
| 3765 | } | |||
| 3766 | ||||
| 3767 | SDValue DAGCombiner::visitUDIV(SDNode *N) { | |||
| 3768 | SDValue N0 = N->getOperand(0); | |||
| 3769 | SDValue N1 = N->getOperand(1); | |||
| 3770 | EVT VT = N->getValueType(0); | |||
| 3771 | EVT CCVT = getSetCCResultType(VT); | |||
| 3772 | ||||
| 3773 | // fold vector ops | |||
| 3774 | if (VT.isVector()) | |||
| 3775 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 3776 | return FoldedVOp; | |||
| 3777 | ||||
| 3778 | SDLoc DL(N); | |||
| 3779 | ||||
| 3780 | // fold (udiv c1, c2) -> c1/c2 | |||
| 3781 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
| 3782 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
| 3783 | if (N0C && N1C) | |||
| 3784 | if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UDIV, DL, VT, | |||
| 3785 | N0C, N1C)) | |||
| 3786 | return Folded; | |||
| 3787 | // fold (udiv X, -1) -> select(X == -1, 1, 0) | |||
| 3788 | if (N1C && N1C->getAPIntValue().isAllOnesValue()) | |||
| 3789 | return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), | |||
| 3790 | DAG.getConstant(1, DL, VT), | |||
| 3791 | DAG.getConstant(0, DL, VT)); | |||
| 3792 | ||||
| 3793 | if (SDValue V = simplifyDivRem(N, DAG)) | |||
| 3794 | return V; | |||
| 3795 | ||||
| 3796 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 3797 | return NewSel; | |||
| 3798 | ||||
| 3799 | if (SDValue V = visitUDIVLike(N0, N1, N)) { | |||
| 3800 | // If the corresponding remainder node exists, update its users with | |||
| 3801 | // (Dividend - (Quotient * Divisor). | |||
| 3802 | if (SDNode *RemNode = DAG.getNodeIfExists(ISD::UREM, N->getVTList(), | |||
| 3803 | { N0, N1 })) { | |||
| 3804 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, V, N1); | |||
| 3805 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); | |||
| 3806 | AddToWorklist(Mul.getNode()); | |||
| 3807 | AddToWorklist(Sub.getNode()); | |||
| 3808 | CombineTo(RemNode, Sub); | |||
| 3809 | } | |||
| 3810 | return V; | |||
| 3811 | } | |||
| 3812 | ||||
| 3813 | // sdiv, srem -> sdivrem | |||
| 3814 | // If the divisor is constant, then return DIVREM only if isIntDivCheap() is | |||
| 3815 | // true. Otherwise, we break the simplification logic in visitREM(). | |||
| 3816 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
| 3817 | if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
| 3818 | if (SDValue DivRem = useDivRem(N)) | |||
| 3819 | return DivRem; | |||
| 3820 | ||||
| 3821 | return SDValue(); | |||
| 3822 | } | |||
| 3823 | ||||
| 3824 | SDValue DAGCombiner::visitUDIVLike(SDValue N0, SDValue N1, SDNode *N) { | |||
| 3825 | SDLoc DL(N); | |||
| 3826 | EVT VT = N->getValueType(0); | |||
| 3827 | ||||
| 3828 | // fold (udiv x, (1 << c)) -> x >>u c | |||
| 3829 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && | |||
| 3830 | DAG.isKnownToBeAPowerOfTwo(N1)) { | |||
| 3831 | SDValue LogBase2 = BuildLogBase2(N1, DL); | |||
| 3832 | AddToWorklist(LogBase2.getNode()); | |||
| 3833 | ||||
| 3834 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); | |||
| 3835 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); | |||
| 3836 | AddToWorklist(Trunc.getNode()); | |||
| 3837 | return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); | |||
| 3838 | } | |||
| 3839 | ||||
| 3840 | // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 | |||
| 3841 | if (N1.getOpcode() == ISD::SHL) { | |||
| 3842 | SDValue N10 = N1.getOperand(0); | |||
| 3843 | if (isConstantOrConstantVector(N10, /*NoOpaques*/ true) && | |||
| 3844 | DAG.isKnownToBeAPowerOfTwo(N10)) { | |||
| 3845 | SDValue LogBase2 = BuildLogBase2(N10, DL); | |||
| 3846 | AddToWorklist(LogBase2.getNode()); | |||
| 3847 | ||||
| 3848 | EVT ADDVT = N1.getOperand(1).getValueType(); | |||
| 3849 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ADDVT); | |||
| 3850 | AddToWorklist(Trunc.getNode()); | |||
| 3851 | SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT, N1.getOperand(1), Trunc); | |||
| 3852 | AddToWorklist(Add.getNode()); | |||
| 3853 | return DAG.getNode(ISD::SRL, DL, VT, N0, Add); | |||
| 3854 | } | |||
| 3855 | } | |||
| 3856 | ||||
| 3857 | // fold (udiv x, c) -> alternate | |||
| 3858 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
| 3859 | if (isConstantOrConstantVector(N1) && | |||
| 3860 | !TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
| 3861 | if (SDValue Op = BuildUDIV(N)) | |||
| 3862 | return Op; | |||
| 3863 | ||||
| 3864 | return SDValue(); | |||
| 3865 | } | |||
| 3866 | ||||
| 3867 | // handles ISD::SREM and ISD::UREM | |||
| 3868 | SDValue DAGCombiner::visitREM(SDNode *N) { | |||
| 3869 | unsigned Opcode = N->getOpcode(); | |||
| 3870 | SDValue N0 = N->getOperand(0); | |||
| 3871 | SDValue N1 = N->getOperand(1); | |||
| 3872 | EVT VT = N->getValueType(0); | |||
| 3873 | EVT CCVT = getSetCCResultType(VT); | |||
| 3874 | ||||
| 3875 | bool isSigned = (Opcode == ISD::SREM); | |||
| 3876 | SDLoc DL(N); | |||
| 3877 | ||||
| 3878 | // fold (rem c1, c2) -> c1%c2 | |||
| 3879 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
| 3880 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
| 3881 | if (N0C && N1C) | |||
| 3882 | if (SDValue Folded = DAG.FoldConstantArithmetic(Opcode, DL, VT, N0C, N1C)) | |||
| 3883 | return Folded; | |||
| 3884 | // fold (urem X, -1) -> select(X == -1, 0, x) | |||
| 3885 | if (!isSigned && N1C && N1C->getAPIntValue().isAllOnesValue()) | |||
| 3886 | return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), | |||
| 3887 | DAG.getConstant(0, DL, VT), N0); | |||
| 3888 | ||||
| 3889 | if (SDValue V = simplifyDivRem(N, DAG)) | |||
| 3890 | return V; | |||
| 3891 | ||||
| 3892 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 3893 | return NewSel; | |||
| 3894 | ||||
| 3895 | if (isSigned) { | |||
| 3896 | // If we know the sign bits of both operands are zero, strength reduce to a | |||
| 3897 | // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 | |||
| 3898 | if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) | |||
| 3899 | return DAG.getNode(ISD::UREM, DL, VT, N0, N1); | |||
| 3900 | } else { | |||
| 3901 | SDValue NegOne = DAG.getAllOnesConstant(DL, VT); | |||
| 3902 | if (DAG.isKnownToBeAPowerOfTwo(N1)) { | |||
| 3903 | // fold (urem x, pow2) -> (and x, pow2-1) | |||
| 3904 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); | |||
| 3905 | AddToWorklist(Add.getNode()); | |||
| 3906 | return DAG.getNode(ISD::AND, DL, VT, N0, Add); | |||
| 3907 | } | |||
| 3908 | if (N1.getOpcode() == ISD::SHL && | |||
| 3909 | DAG.isKnownToBeAPowerOfTwo(N1.getOperand(0))) { | |||
| 3910 | // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) | |||
| 3911 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); | |||
| 3912 | AddToWorklist(Add.getNode()); | |||
| 3913 | return DAG.getNode(ISD::AND, DL, VT, N0, Add); | |||
| 3914 | } | |||
| 3915 | } | |||
| 3916 | ||||
| 3917 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
| 3918 | ||||
| 3919 | // If X/C can be simplified by the division-by-constant logic, lower | |||
| 3920 | // X%C to the equivalent of X-X/C*C. | |||
| 3921 | // Reuse the SDIVLike/UDIVLike combines - to avoid mangling nodes, the | |||
| 3922 | // speculative DIV must not cause a DIVREM conversion. We guard against this | |||
| 3923 | // by skipping the simplification if isIntDivCheap(). When div is not cheap, | |||
| 3924 | // combine will not return a DIVREM. Regardless, checking cheapness here | |||
| 3925 | // makes sense since the simplification results in fatter code. | |||
| 3926 | if (DAG.isKnownNeverZero(N1) && !TLI.isIntDivCheap(VT, Attr)) { | |||
| 3927 | SDValue OptimizedDiv = | |||
| 3928 | isSigned ? visitSDIVLike(N0, N1, N) : visitUDIVLike(N0, N1, N); | |||
| 3929 | if (OptimizedDiv.getNode()) { | |||
| 3930 | // If the equivalent Div node also exists, update its users. | |||
| 3931 | unsigned DivOpcode = isSigned ? ISD::SDIV : ISD::UDIV; | |||
| 3932 | if (SDNode *DivNode = DAG.getNodeIfExists(DivOpcode, N->getVTList(), | |||
| 3933 | { N0, N1 })) | |||
| 3934 | CombineTo(DivNode, OptimizedDiv); | |||
| 3935 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, OptimizedDiv, N1); | |||
| 3936 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); | |||
| 3937 | AddToWorklist(OptimizedDiv.getNode()); | |||
| 3938 | AddToWorklist(Mul.getNode()); | |||
| 3939 | return Sub; | |||
| 3940 | } | |||
| 3941 | } | |||
| 3942 | ||||
| 3943 | // sdiv, srem -> sdivrem | |||
| 3944 | if (SDValue DivRem = useDivRem(N)) | |||
| 3945 | return DivRem.getValue(1); | |||
| 3946 | ||||
| 3947 | return SDValue(); | |||
| 3948 | } | |||
| 3949 | ||||
| 3950 | SDValue DAGCombiner::visitMULHS(SDNode *N) { | |||
| 3951 | SDValue N0 = N->getOperand(0); | |||
| 3952 | SDValue N1 = N->getOperand(1); | |||
| 3953 | EVT VT = N->getValueType(0); | |||
| 3954 | SDLoc DL(N); | |||
| 3955 | ||||
| 3956 | if (VT.isVector()) { | |||
| 3957 | // fold (mulhs x, 0) -> 0 | |||
| 3958 | // do not return N0/N1, because undef node may exist. | |||
| 3959 | if (ISD::isBuildVectorAllZeros(N0.getNode()) || | |||
| 3960 | ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 3961 | return DAG.getConstant(0, DL, VT); | |||
| 3962 | } | |||
| 3963 | ||||
| 3964 | // fold (mulhs x, 0) -> 0 | |||
| 3965 | if (isNullConstant(N1)) | |||
| 3966 | return N1; | |||
| 3967 | // fold (mulhs x, 1) -> (sra x, size(x)-1) | |||
| 3968 | if (isOneConstant(N1)) | |||
| 3969 | return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0, | |||
| 3970 | DAG.getConstant(N0.getScalarValueSizeInBits() - 1, DL, | |||
| 3971 | getShiftAmountTy(N0.getValueType()))); | |||
| 3972 | ||||
| 3973 | // fold (mulhs x, undef) -> 0 | |||
| 3974 | if (N0.isUndef() || N1.isUndef()) | |||
| 3975 | return DAG.getConstant(0, DL, VT); | |||
| 3976 | ||||
| 3977 | // If the type twice as wide is legal, transform the mulhs to a wider multiply | |||
| 3978 | // plus a shift. | |||
| 3979 | if (VT.isSimple() && !VT.isVector()) { | |||
| 3980 | MVT Simple = VT.getSimpleVT(); | |||
| 3981 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
| 3982 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
| 3983 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
| 3984 | N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); | |||
| 3985 | N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); | |||
| 3986 | N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); | |||
| 3987 | N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, | |||
| 3988 | DAG.getConstant(SimpleSize, DL, | |||
| 3989 | getShiftAmountTy(N1.getValueType()))); | |||
| 3990 | return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); | |||
| 3991 | } | |||
| 3992 | } | |||
| 3993 | ||||
| 3994 | return SDValue(); | |||
| 3995 | } | |||
| 3996 | ||||
| 3997 | SDValue DAGCombiner::visitMULHU(SDNode *N) { | |||
| 3998 | SDValue N0 = N->getOperand(0); | |||
| 3999 | SDValue N1 = N->getOperand(1); | |||
| 4000 | EVT VT = N->getValueType(0); | |||
| 4001 | SDLoc DL(N); | |||
| 4002 | ||||
| 4003 | if (VT.isVector()) { | |||
| 4004 | // fold (mulhu x, 0) -> 0 | |||
| 4005 | // do not return N0/N1, because undef node may exist. | |||
| 4006 | if (ISD::isBuildVectorAllZeros(N0.getNode()) || | |||
| 4007 | ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 4008 | return DAG.getConstant(0, DL, VT); | |||
| 4009 | } | |||
| 4010 | ||||
| 4011 | // fold (mulhu x, 0) -> 0 | |||
| 4012 | if (isNullConstant(N1)) | |||
| 4013 | return N1; | |||
| 4014 | // fold (mulhu x, 1) -> 0 | |||
| 4015 | if (isOneConstant(N1)) | |||
| 4016 | return DAG.getConstant(0, DL, N0.getValueType()); | |||
| 4017 | // fold (mulhu x, undef) -> 0 | |||
| 4018 | if (N0.isUndef() || N1.isUndef()) | |||
| 4019 | return DAG.getConstant(0, DL, VT); | |||
| 4020 | ||||
| 4021 | // fold (mulhu x, (1 << c)) -> x >> (bitwidth - c) | |||
| 4022 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && | |||
| 4023 | DAG.isKnownToBeAPowerOfTwo(N1) && hasOperation(ISD::SRL, VT)) { | |||
| 4024 | unsigned NumEltBits = VT.getScalarSizeInBits(); | |||
| 4025 | SDValue LogBase2 = BuildLogBase2(N1, DL); | |||
| 4026 | SDValue SRLAmt = DAG.getNode( | |||
| 4027 | ISD::SUB, DL, VT, DAG.getConstant(NumEltBits, DL, VT), LogBase2); | |||
| 4028 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); | |||
| 4029 | SDValue Trunc = DAG.getZExtOrTrunc(SRLAmt, DL, ShiftVT); | |||
| 4030 | return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); | |||
| 4031 | } | |||
| 4032 | ||||
| 4033 | // If the type twice as wide is legal, transform the mulhu to a wider multiply | |||
| 4034 | // plus a shift. | |||
| 4035 | if (VT.isSimple() && !VT.isVector()) { | |||
| 4036 | MVT Simple = VT.getSimpleVT(); | |||
| 4037 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
| 4038 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
| 4039 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
| 4040 | N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); | |||
| 4041 | N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); | |||
| 4042 | N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); | |||
| 4043 | N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, | |||
| 4044 | DAG.getConstant(SimpleSize, DL, | |||
| 4045 | getShiftAmountTy(N1.getValueType()))); | |||
| 4046 | return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); | |||
| 4047 | } | |||
| 4048 | } | |||
| 4049 | ||||
| 4050 | return SDValue(); | |||
| 4051 | } | |||
| 4052 | ||||
| 4053 | /// Perform optimizations common to nodes that compute two values. LoOp and HiOp | |||
| 4054 | /// give the opcodes for the two computations that are being performed. Return | |||
| 4055 | /// true if a simplification was made. | |||
| 4056 | SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, | |||
| 4057 | unsigned HiOp) { | |||
| 4058 | // If the high half is not needed, just compute the low half. | |||
| 4059 | bool HiExists = N->hasAnyUseOfValue(1); | |||
| 4060 | if (!HiExists && (!LegalOperations || | |||
| 4061 | TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) { | |||
| 4062 | SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); | |||
| 4063 | return CombineTo(N, Res, Res); | |||
| 4064 | } | |||
| 4065 | ||||
| 4066 | // If the low half is not needed, just compute the high half. | |||
| 4067 | bool LoExists = N->hasAnyUseOfValue(0); | |||
| 4068 | if (!LoExists && (!LegalOperations || | |||
| 4069 | TLI.isOperationLegalOrCustom(HiOp, N->getValueType(1)))) { | |||
| 4070 | SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); | |||
| 4071 | return CombineTo(N, Res, Res); | |||
| 4072 | } | |||
| 4073 | ||||
| 4074 | // If both halves are used, return as it is. | |||
| 4075 | if (LoExists && HiExists) | |||
| 4076 | return SDValue(); | |||
| 4077 | ||||
| 4078 | // If the two computed results can be simplified separately, separate them. | |||
| 4079 | if (LoExists) { | |||
| 4080 | SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); | |||
| 4081 | AddToWorklist(Lo.getNode()); | |||
| 4082 | SDValue LoOpt = combine(Lo.getNode()); | |||
| 4083 | if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() && | |||
| 4084 | (!LegalOperations || | |||
| 4085 | TLI.isOperationLegalOrCustom(LoOpt.getOpcode(), LoOpt.getValueType()))) | |||
| 4086 | return CombineTo(N, LoOpt, LoOpt); | |||
| 4087 | } | |||
| 4088 | ||||
| 4089 | if (HiExists) { | |||
| 4090 | SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); | |||
| 4091 | AddToWorklist(Hi.getNode()); | |||
| 4092 | SDValue HiOpt = combine(Hi.getNode()); | |||
| 4093 | if (HiOpt.getNode() && HiOpt != Hi && | |||
| 4094 | (!LegalOperations || | |||
| 4095 | TLI.isOperationLegalOrCustom(HiOpt.getOpcode(), HiOpt.getValueType()))) | |||
| 4096 | return CombineTo(N, HiOpt, HiOpt); | |||
| 4097 | } | |||
| 4098 | ||||
| 4099 | return SDValue(); | |||
| 4100 | } | |||
| 4101 | ||||
| 4102 | SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) { | |||
| 4103 | if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS)) | |||
| 4104 | return Res; | |||
| 4105 | ||||
| 4106 | EVT VT = N->getValueType(0); | |||
| 4107 | SDLoc DL(N); | |||
| 4108 | ||||
| 4109 | // If the type is twice as wide is legal, transform the mulhu to a wider | |||
| 4110 | // multiply plus a shift. | |||
| 4111 | if (VT.isSimple() && !VT.isVector()) { | |||
| 4112 | MVT Simple = VT.getSimpleVT(); | |||
| 4113 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
| 4114 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
| 4115 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
| 4116 | SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0)); | |||
| 4117 | SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1)); | |||
| 4118 | Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); | |||
| 4119 | // Compute the high part as N1. | |||
| 4120 | Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, | |||
| 4121 | DAG.getConstant(SimpleSize, DL, | |||
| 4122 | getShiftAmountTy(Lo.getValueType()))); | |||
| 4123 | Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); | |||
| 4124 | // Compute the low part as N0. | |||
| 4125 | Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); | |||
| 4126 | return CombineTo(N, Lo, Hi); | |||
| 4127 | } | |||
| 4128 | } | |||
| 4129 | ||||
| 4130 | return SDValue(); | |||
| 4131 | } | |||
| 4132 | ||||
| 4133 | SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) { | |||
| 4134 | if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU)) | |||
| 4135 | return Res; | |||
| 4136 | ||||
| 4137 | EVT VT = N->getValueType(0); | |||
| 4138 | SDLoc DL(N); | |||
| 4139 | ||||
| 4140 | // (umul_lohi N0, 0) -> (0, 0) | |||
| 4141 | if (isNullConstant(N->getOperand(1))) { | |||
| 4142 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
| 4143 | return CombineTo(N, Zero, Zero); | |||
| 4144 | } | |||
| 4145 | ||||
| 4146 | // (umul_lohi N0, 1) -> (N0, 0) | |||
| 4147 | if (isOneConstant(N->getOperand(1))) { | |||
| 4148 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
| 4149 | return CombineTo(N, N->getOperand(0), Zero); | |||
| 4150 | } | |||
| 4151 | ||||
| 4152 | // If the type is twice as wide is legal, transform the mulhu to a wider | |||
| 4153 | // multiply plus a shift. | |||
| 4154 | if (VT.isSimple() && !VT.isVector()) { | |||
| 4155 | MVT Simple = VT.getSimpleVT(); | |||
| 4156 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
| 4157 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
| 4158 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
| 4159 | SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0)); | |||
| 4160 | SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1)); | |||
| 4161 | Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); | |||
| 4162 | // Compute the high part as N1. | |||
| 4163 | Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, | |||
| 4164 | DAG.getConstant(SimpleSize, DL, | |||
| 4165 | getShiftAmountTy(Lo.getValueType()))); | |||
| 4166 | Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); | |||
| 4167 | // Compute the low part as N0. | |||
| 4168 | Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); | |||
| 4169 | return CombineTo(N, Lo, Hi); | |||
| 4170 | } | |||
| 4171 | } | |||
| 4172 | ||||
| 4173 | return SDValue(); | |||
| 4174 | } | |||
| 4175 | ||||
| 4176 | SDValue DAGCombiner::visitMULO(SDNode *N) { | |||
| 4177 | SDValue N0 = N->getOperand(0); | |||
| 4178 | SDValue N1 = N->getOperand(1); | |||
| 4179 | EVT VT = N0.getValueType(); | |||
| 4180 | bool IsSigned = (ISD::SMULO == N->getOpcode()); | |||
| 4181 | ||||
| 4182 | EVT CarryVT = N->getValueType(1); | |||
| 4183 | SDLoc DL(N); | |||
| 4184 | ||||
| 4185 | // canonicalize constant to RHS. | |||
| 4186 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 4187 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 4188 | return DAG.getNode(N->getOpcode(), DL, N->getVTList(), N1, N0); | |||
| 4189 | ||||
| 4190 | // fold (mulo x, 0) -> 0 + no carry out | |||
| 4191 | if (isNullOrNullSplat(N1)) | |||
| 4192 | return CombineTo(N, DAG.getConstant(0, DL, VT), | |||
| 4193 | DAG.getConstant(0, DL, CarryVT)); | |||
| 4194 | ||||
| 4195 | // (mulo x, 2) -> (addo x, x) | |||
| 4196 | if (ConstantSDNode *C2 = isConstOrConstSplat(N1)) | |||
| 4197 | if (C2->getAPIntValue() == 2) | |||
| 4198 | return DAG.getNode(IsSigned ? ISD::SADDO : ISD::UADDO, DL, | |||
| 4199 | N->getVTList(), N0, N0); | |||
| 4200 | ||||
| 4201 | return SDValue(); | |||
| 4202 | } | |||
| 4203 | ||||
| 4204 | SDValue DAGCombiner::visitIMINMAX(SDNode *N) { | |||
| 4205 | SDValue N0 = N->getOperand(0); | |||
| 4206 | SDValue N1 = N->getOperand(1); | |||
| 4207 | EVT VT = N0.getValueType(); | |||
| 4208 | ||||
| 4209 | // fold vector ops | |||
| 4210 | if (VT.isVector()) | |||
| 4211 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 4212 | return FoldedVOp; | |||
| 4213 | ||||
| 4214 | // fold operation with constant operands. | |||
| 4215 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
| 4216 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); | |||
| 4217 | if (N0C && N1C) | |||
| 4218 | return DAG.FoldConstantArithmetic(N->getOpcode(), SDLoc(N), VT, N0C, N1C); | |||
| 4219 | ||||
| 4220 | // canonicalize constant to RHS | |||
| 4221 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 4222 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 4223 | return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0); | |||
| 4224 | ||||
| 4225 | // Is sign bits are zero, flip between UMIN/UMAX and SMIN/SMAX. | |||
| 4226 | // Only do this if the current op isn't legal and the flipped is. | |||
| 4227 | unsigned Opcode = N->getOpcode(); | |||
| 4228 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 4229 | if (!TLI.isOperationLegal(Opcode, VT) && | |||
| 4230 | (N0.isUndef() || DAG.SignBitIsZero(N0)) && | |||
| 4231 | (N1.isUndef() || DAG.SignBitIsZero(N1))) { | |||
| 4232 | unsigned AltOpcode; | |||
| 4233 | switch (Opcode) { | |||
| 4234 | case ISD::SMIN: AltOpcode = ISD::UMIN; break; | |||
| 4235 | case ISD::SMAX: AltOpcode = ISD::UMAX; break; | |||
| 4236 | case ISD::UMIN: AltOpcode = ISD::SMIN; break; | |||
| 4237 | case ISD::UMAX: AltOpcode = ISD::SMAX; break; | |||
| 4238 | default: llvm_unreachable("Unknown MINMAX opcode")::llvm::llvm_unreachable_internal("Unknown MINMAX opcode", "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4238); | |||
| 4239 | } | |||
| 4240 | if (TLI.isOperationLegal(AltOpcode, VT)) | |||
| 4241 | return DAG.getNode(AltOpcode, SDLoc(N), VT, N0, N1); | |||
| 4242 | } | |||
| 4243 | ||||
| 4244 | return SDValue(); | |||
| 4245 | } | |||
| 4246 | ||||
| 4247 | /// If this is a bitwise logic instruction and both operands have the same | |||
| 4248 | /// opcode, try to sink the other opcode after the logic instruction. | |||
| 4249 | SDValue DAGCombiner::hoistLogicOpWithSameOpcodeHands(SDNode *N) { | |||
| 4250 | SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); | |||
| 4251 | EVT VT = N0.getValueType(); | |||
| 4252 | unsigned LogicOpcode = N->getOpcode(); | |||
| 4253 | unsigned HandOpcode = N0.getOpcode(); | |||
| 4254 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4255, __PRETTY_FUNCTION__)) | |||
| 4255 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4255, __PRETTY_FUNCTION__)); | |||
| 4256 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4256, __PRETTY_FUNCTION__)); | |||
| 4257 | ||||
| 4258 | // Bail early if none of these transforms apply. | |||
| 4259 | if (N0.getNumOperands() == 0) | |||
| 4260 | return SDValue(); | |||
| 4261 | ||||
| 4262 | // FIXME: We should check number of uses of the operands to not increase | |||
| 4263 | // the instruction count for all transforms. | |||
| 4264 | ||||
| 4265 | // Handle size-changing casts. | |||
| 4266 | SDValue X = N0.getOperand(0); | |||
| 4267 | SDValue Y = N1.getOperand(0); | |||
| 4268 | EVT XVT = X.getValueType(); | |||
| 4269 | SDLoc DL(N); | |||
| 4270 | if (HandOpcode == ISD::ANY_EXTEND || HandOpcode == ISD::ZERO_EXTEND || | |||
| 4271 | HandOpcode == ISD::SIGN_EXTEND) { | |||
| 4272 | // If both operands have other uses, this transform would create extra | |||
| 4273 | // instructions without eliminating anything. | |||
| 4274 | if (!N0.hasOneUse() && !N1.hasOneUse()) | |||
| 4275 | return SDValue(); | |||
| 4276 | // We need matching integer source types. | |||
| 4277 | if (XVT != Y.getValueType()) | |||
| 4278 | return SDValue(); | |||
| 4279 | // Don't create an illegal op during or after legalization. Don't ever | |||
| 4280 | // create an unsupported vector op. | |||
| 4281 | if ((VT.isVector() || LegalOperations) && | |||
| 4282 | !TLI.isOperationLegalOrCustom(LogicOpcode, XVT)) | |||
| 4283 | return SDValue(); | |||
| 4284 | // Avoid infinite looping with PromoteIntBinOp. | |||
| 4285 | // TODO: Should we apply desirable/legal constraints to all opcodes? | |||
| 4286 | if (HandOpcode == ISD::ANY_EXTEND && LegalTypes && | |||
| 4287 | !TLI.isTypeDesirableForOp(LogicOpcode, XVT)) | |||
| 4288 | return SDValue(); | |||
| 4289 | // logic_op (hand_op X), (hand_op Y) --> hand_op (logic_op X, Y) | |||
| 4290 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); | |||
| 4291 | return DAG.getNode(HandOpcode, DL, VT, Logic); | |||
| 4292 | } | |||
| 4293 | ||||
| 4294 | // logic_op (truncate x), (truncate y) --> truncate (logic_op x, y) | |||
| 4295 | if (HandOpcode == ISD::TRUNCATE) { | |||
| 4296 | // If both operands have other uses, this transform would create extra | |||
| 4297 | // instructions without eliminating anything. | |||
| 4298 | if (!N0.hasOneUse() && !N1.hasOneUse()) | |||
| 4299 | return SDValue(); | |||
| 4300 | // We need matching source types. | |||
| 4301 | if (XVT != Y.getValueType()) | |||
| 4302 | return SDValue(); | |||
| 4303 | // Don't create an illegal op during or after legalization. | |||
| 4304 | if (LegalOperations && !TLI.isOperationLegal(LogicOpcode, XVT)) | |||
| 4305 | return SDValue(); | |||
| 4306 | // Be extra careful sinking truncate. If it's free, there's no benefit in | |||
| 4307 | // widening a binop. Also, don't create a logic op on an illegal type. | |||
| 4308 | if (TLI.isZExtFree(VT, XVT) && TLI.isTruncateFree(XVT, VT)) | |||
| 4309 | return SDValue(); | |||
| 4310 | if (!TLI.isTypeLegal(XVT)) | |||
| 4311 | return SDValue(); | |||
| 4312 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); | |||
| 4313 | return DAG.getNode(HandOpcode, DL, VT, Logic); | |||
| 4314 | } | |||
| 4315 | ||||
| 4316 | // For binops SHL/SRL/SRA/AND: | |||
| 4317 | // logic_op (OP x, z), (OP y, z) --> OP (logic_op x, y), z | |||
| 4318 | if ((HandOpcode == ISD::SHL || HandOpcode == ISD::SRL || | |||
| 4319 | HandOpcode == ISD::SRA || HandOpcode == ISD::AND) && | |||
| 4320 | N0.getOperand(1) == N1.getOperand(1)) { | |||
| 4321 | // If either operand has other uses, this transform is not an improvement. | |||
| 4322 | if (!N0.hasOneUse() || !N1.hasOneUse()) | |||
| 4323 | return SDValue(); | |||
| 4324 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); | |||
| 4325 | return DAG.getNode(HandOpcode, DL, VT, Logic, N0.getOperand(1)); | |||
| 4326 | } | |||
| 4327 | ||||
| 4328 | // Unary ops: logic_op (bswap x), (bswap y) --> bswap (logic_op x, y) | |||
| 4329 | if (HandOpcode == ISD::BSWAP) { | |||
| 4330 | // If either operand has other uses, this transform is not an improvement. | |||
| 4331 | if (!N0.hasOneUse() || !N1.hasOneUse()) | |||
| 4332 | return SDValue(); | |||
| 4333 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); | |||
| 4334 | return DAG.getNode(HandOpcode, DL, VT, Logic); | |||
| 4335 | } | |||
| 4336 | ||||
| 4337 | // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) | |||
| 4338 | // Only perform this optimization up until type legalization, before | |||
| 4339 | // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by | |||
| 4340 | // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and | |||
| 4341 | // we don't want to undo this promotion. | |||
| 4342 | // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper | |||
| 4343 | // on scalars. | |||
| 4344 | if ((HandOpcode == ISD::BITCAST || HandOpcode == ISD::SCALAR_TO_VECTOR) && | |||
| 4345 | Level <= AfterLegalizeTypes) { | |||
| 4346 | // Input types must be integer and the same. | |||
| 4347 | if (XVT.isInteger() && XVT == Y.getValueType()) { | |||
| 4348 | SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); | |||
| 4349 | return DAG.getNode(HandOpcode, DL, VT, Logic); | |||
| 4350 | } | |||
| 4351 | } | |||
| 4352 | ||||
| 4353 | // Xor/and/or are indifferent to the swizzle operation (shuffle of one value). | |||
| 4354 | // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B)) | |||
| 4355 | // If both shuffles use the same mask, and both shuffle within a single | |||
| 4356 | // vector, then it is worthwhile to move the swizzle after the operation. | |||
| 4357 | // The type-legalizer generates this pattern when loading illegal | |||
| 4358 | // vector types from memory. In many cases this allows additional shuffle | |||
| 4359 | // optimizations. | |||
| 4360 | // There are other cases where moving the shuffle after the xor/and/or | |||
| 4361 | // is profitable even if shuffles don't perform a swizzle. | |||
| 4362 | // If both shuffles use the same mask, and both shuffles have the same first | |||
| 4363 | // or second operand, then it might still be profitable to move the shuffle | |||
| 4364 | // after the xor/and/or operation. | |||
| 4365 | if (HandOpcode == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) { | |||
| 4366 | auto *SVN0 = cast<ShuffleVectorSDNode>(N0); | |||
| 4367 | auto *SVN1 = cast<ShuffleVectorSDNode>(N1); | |||
| 4368 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4369, __PRETTY_FUNCTION__)) | |||
| 4369 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4369, __PRETTY_FUNCTION__)); | |||
| 4370 | ||||
| 4371 | // Check that both shuffles use the same mask. The masks are known to be of | |||
| 4372 | // the same length because the result vector type is the same. | |||
| 4373 | // Check also that shuffles have only one use to avoid introducing extra | |||
| 4374 | // instructions. | |||
| 4375 | if (!SVN0->hasOneUse() || !SVN1->hasOneUse() || | |||
| 4376 | !SVN0->getMask().equals(SVN1->getMask())) | |||
| 4377 | return SDValue(); | |||
| 4378 | ||||
| 4379 | // Don't try to fold this node if it requires introducing a | |||
| 4380 | // build vector of all zeros that might be illegal at this stage. | |||
| 4381 | SDValue ShOp = N0.getOperand(1); | |||
| 4382 | if (LogicOpcode == ISD::XOR && !ShOp.isUndef()) | |||
| 4383 | ShOp = tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); | |||
| 4384 | ||||
| 4385 | // (logic_op (shuf (A, C), shuf (B, C))) --> shuf (logic_op (A, B), C) | |||
| 4386 | if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) { | |||
| 4387 | SDValue Logic = DAG.getNode(LogicOpcode, DL, VT, | |||
| 4388 | N0.getOperand(0), N1.getOperand(0)); | |||
| 4389 | return DAG.getVectorShuffle(VT, DL, Logic, ShOp, SVN0->getMask()); | |||
| 4390 | } | |||
| 4391 | ||||
| 4392 | // Don't try to fold this node if it requires introducing a | |||
| 4393 | // build vector of all zeros that might be illegal at this stage. | |||
| 4394 | ShOp = N0.getOperand(0); | |||
| 4395 | if (LogicOpcode == ISD::XOR && !ShOp.isUndef()) | |||
| 4396 | ShOp = tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); | |||
| 4397 | ||||
| 4398 | // (logic_op (shuf (C, A), shuf (C, B))) --> shuf (C, logic_op (A, B)) | |||
| 4399 | if (N0.getOperand(0) == N1.getOperand(0) && ShOp.getNode()) { | |||
| 4400 | SDValue Logic = DAG.getNode(LogicOpcode, DL, VT, N0.getOperand(1), | |||
| 4401 | N1.getOperand(1)); | |||
| 4402 | return DAG.getVectorShuffle(VT, DL, ShOp, Logic, SVN0->getMask()); | |||
| 4403 | } | |||
| 4404 | } | |||
| 4405 | ||||
| 4406 | return SDValue(); | |||
| 4407 | } | |||
| 4408 | ||||
| 4409 | /// Try to make (and/or setcc (LL, LR), setcc (RL, RR)) more efficient. | |||
| 4410 | SDValue DAGCombiner::foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, | |||
| 4411 | const SDLoc &DL) { | |||
| 4412 | SDValue LL, LR, RL, RR, N0CC, N1CC; | |||
| 4413 | if (!isSetCCEquivalent(N0, LL, LR, N0CC) || | |||
| 4414 | !isSetCCEquivalent(N1, RL, RR, N1CC)) | |||
| 4415 | return SDValue(); | |||
| 4416 | ||||
| 4417 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4418, __PRETTY_FUNCTION__)) | |||
| 4418 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4418, __PRETTY_FUNCTION__)); | |||
| 4419 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4421, __PRETTY_FUNCTION__)) | |||
| 4420 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4421, __PRETTY_FUNCTION__)) | |||
| 4421 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4421, __PRETTY_FUNCTION__)); | |||
| 4422 | ||||
| 4423 | // If we're here post-legalization or the logic op type is not i1, the logic | |||
| 4424 | // op type must match a setcc result type. Also, all folds require new | |||
| 4425 | // operations on the left and right operands, so those types must match. | |||
| 4426 | EVT VT = N0.getValueType(); | |||
| 4427 | EVT OpVT = LL.getValueType(); | |||
| 4428 | if (LegalOperations || VT.getScalarType() != MVT::i1) | |||
| 4429 | if (VT != getSetCCResultType(OpVT)) | |||
| 4430 | return SDValue(); | |||
| 4431 | if (OpVT != RL.getValueType()) | |||
| 4432 | return SDValue(); | |||
| 4433 | ||||
| 4434 | ISD::CondCode CC0 = cast<CondCodeSDNode>(N0CC)->get(); | |||
| 4435 | ISD::CondCode CC1 = cast<CondCodeSDNode>(N1CC)->get(); | |||
| 4436 | bool IsInteger = OpVT.isInteger(); | |||
| 4437 | if (LR == RR && CC0 == CC1 && IsInteger) { | |||
| 4438 | bool IsZero = isNullOrNullSplat(LR); | |||
| 4439 | bool IsNeg1 = isAllOnesOrAllOnesSplat(LR); | |||
| 4440 | ||||
| 4441 | // All bits clear? | |||
| 4442 | bool AndEqZero = IsAnd && CC1 == ISD::SETEQ && IsZero; | |||
| 4443 | // All sign bits clear? | |||
| 4444 | bool AndGtNeg1 = IsAnd && CC1 == ISD::SETGT && IsNeg1; | |||
| 4445 | // Any bits set? | |||
| 4446 | bool OrNeZero = !IsAnd && CC1 == ISD::SETNE && IsZero; | |||
| 4447 | // Any sign bits set? | |||
| 4448 | bool OrLtZero = !IsAnd && CC1 == ISD::SETLT && IsZero; | |||
| 4449 | ||||
| 4450 | // (and (seteq X, 0), (seteq Y, 0)) --> (seteq (or X, Y), 0) | |||
| 4451 | // (and (setgt X, -1), (setgt Y, -1)) --> (setgt (or X, Y), -1) | |||
| 4452 | // (or (setne X, 0), (setne Y, 0)) --> (setne (or X, Y), 0) | |||
| 4453 | // (or (setlt X, 0), (setlt Y, 0)) --> (setlt (or X, Y), 0) | |||
| 4454 | if (AndEqZero || AndGtNeg1 || OrNeZero || OrLtZero) { | |||
| 4455 | SDValue Or = DAG.getNode(ISD::OR, SDLoc(N0), OpVT, LL, RL); | |||
| 4456 | AddToWorklist(Or.getNode()); | |||
| 4457 | return DAG.getSetCC(DL, VT, Or, LR, CC1); | |||
| 4458 | } | |||
| 4459 | ||||
| 4460 | // All bits set? | |||
| 4461 | bool AndEqNeg1 = IsAnd && CC1 == ISD::SETEQ && IsNeg1; | |||
| 4462 | // All sign bits set? | |||
| 4463 | bool AndLtZero = IsAnd && CC1 == ISD::SETLT && IsZero; | |||
| 4464 | // Any bits clear? | |||
| 4465 | bool OrNeNeg1 = !IsAnd && CC1 == ISD::SETNE && IsNeg1; | |||
| 4466 | // Any sign bits clear? | |||
| 4467 | bool OrGtNeg1 = !IsAnd && CC1 == ISD::SETGT && IsNeg1; | |||
| 4468 | ||||
| 4469 | // (and (seteq X, -1), (seteq Y, -1)) --> (seteq (and X, Y), -1) | |||
| 4470 | // (and (setlt X, 0), (setlt Y, 0)) --> (setlt (and X, Y), 0) | |||
| 4471 | // (or (setne X, -1), (setne Y, -1)) --> (setne (and X, Y), -1) | |||
| 4472 | // (or (setgt X, -1), (setgt Y -1)) --> (setgt (and X, Y), -1) | |||
| 4473 | if (AndEqNeg1 || AndLtZero || OrNeNeg1 || OrGtNeg1) { | |||
| 4474 | SDValue And = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, LL, RL); | |||
| 4475 | AddToWorklist(And.getNode()); | |||
| 4476 | return DAG.getSetCC(DL, VT, And, LR, CC1); | |||
| 4477 | } | |||
| 4478 | } | |||
| 4479 | ||||
| 4480 | // TODO: What is the 'or' equivalent of this fold? | |||
| 4481 | // (and (setne X, 0), (setne X, -1)) --> (setuge (add X, 1), 2) | |||
| 4482 | if (IsAnd && LL == RL && CC0 == CC1 && OpVT.getScalarSizeInBits() > 1 && | |||
| 4483 | IsInteger && CC0 == ISD::SETNE && | |||
| 4484 | ((isNullConstant(LR) && isAllOnesConstant(RR)) || | |||
| 4485 | (isAllOnesConstant(LR) && isNullConstant(RR)))) { | |||
| 4486 | SDValue One = DAG.getConstant(1, DL, OpVT); | |||
| 4487 | SDValue Two = DAG.getConstant(2, DL, OpVT); | |||
| 4488 | SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), OpVT, LL, One); | |||
| 4489 | AddToWorklist(Add.getNode()); | |||
| 4490 | return DAG.getSetCC(DL, VT, Add, Two, ISD::SETUGE); | |||
| 4491 | } | |||
| 4492 | ||||
| 4493 | // Try more general transforms if the predicates match and the only user of | |||
| 4494 | // the compares is the 'and' or 'or'. | |||
| 4495 | if (IsInteger && TLI.convertSetCCLogicToBitwiseLogic(OpVT) && CC0 == CC1 && | |||
| 4496 | N0.hasOneUse() && N1.hasOneUse()) { | |||
| 4497 | // and (seteq A, B), (seteq C, D) --> seteq (or (xor A, B), (xor C, D)), 0 | |||
| 4498 | // or (setne A, B), (setne C, D) --> setne (or (xor A, B), (xor C, D)), 0 | |||
| 4499 | if ((IsAnd && CC1 == ISD::SETEQ) || (!IsAnd && CC1 == ISD::SETNE)) { | |||
| 4500 | SDValue XorL = DAG.getNode(ISD::XOR, SDLoc(N0), OpVT, LL, LR); | |||
| 4501 | SDValue XorR = DAG.getNode(ISD::XOR, SDLoc(N1), OpVT, RL, RR); | |||
| 4502 | SDValue Or = DAG.getNode(ISD::OR, DL, OpVT, XorL, XorR); | |||
| 4503 | SDValue Zero = DAG.getConstant(0, DL, OpVT); | |||
| 4504 | return DAG.getSetCC(DL, VT, Or, Zero, CC1); | |||
| 4505 | } | |||
| 4506 | ||||
| 4507 | // Turn compare of constants whose difference is 1 bit into add+and+setcc. | |||
| 4508 | // TODO - support non-uniform vector amounts. | |||
| 4509 | if ((IsAnd && CC1 == ISD::SETNE) || (!IsAnd && CC1 == ISD::SETEQ)) { | |||
| 4510 | // Match a shared variable operand and 2 non-opaque constant operands. | |||
| 4511 | ConstantSDNode *C0 = isConstOrConstSplat(LR); | |||
| 4512 | ConstantSDNode *C1 = isConstOrConstSplat(RR); | |||
| 4513 | if (LL == RL && C0 && C1 && !C0->isOpaque() && !C1->isOpaque()) { | |||
| 4514 | // Canonicalize larger constant as C0. | |||
| 4515 | if (C1->getAPIntValue().ugt(C0->getAPIntValue())) | |||
| 4516 | std::swap(C0, C1); | |||
| 4517 | ||||
| 4518 | // The difference of the constants must be a single bit. | |||
| 4519 | const APInt &C0Val = C0->getAPIntValue(); | |||
| 4520 | const APInt &C1Val = C1->getAPIntValue(); | |||
| 4521 | if ((C0Val - C1Val).isPowerOf2()) { | |||
| 4522 | // and/or (setcc X, C0, ne), (setcc X, C1, ne/eq) --> | |||
| 4523 | // setcc ((add X, -C1), ~(C0 - C1)), 0, ne/eq | |||
| 4524 | SDValue OffsetC = DAG.getConstant(-C1Val, DL, OpVT); | |||
| 4525 | SDValue Add = DAG.getNode(ISD::ADD, DL, OpVT, LL, OffsetC); | |||
| 4526 | SDValue MaskC = DAG.getConstant(~(C0Val - C1Val), DL, OpVT); | |||
| 4527 | SDValue And = DAG.getNode(ISD::AND, DL, OpVT, Add, MaskC); | |||
| 4528 | SDValue Zero = DAG.getConstant(0, DL, OpVT); | |||
| 4529 | return DAG.getSetCC(DL, VT, And, Zero, CC0); | |||
| 4530 | } | |||
| 4531 | } | |||
| 4532 | } | |||
| 4533 | } | |||
| 4534 | ||||
| 4535 | // Canonicalize equivalent operands to LL == RL. | |||
| 4536 | if (LL == RR && LR == RL) { | |||
| 4537 | CC1 = ISD::getSetCCSwappedOperands(CC1); | |||
| 4538 | std::swap(RL, RR); | |||
| 4539 | } | |||
| 4540 | ||||
| 4541 | // (and (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) | |||
| 4542 | // (or (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) | |||
| 4543 | if (LL == RL && LR == RR) { | |||
| 4544 | ISD::CondCode NewCC = IsAnd ? ISD::getSetCCAndOperation(CC0, CC1, IsInteger) | |||
| 4545 | : ISD::getSetCCOrOperation(CC0, CC1, IsInteger); | |||
| 4546 | if (NewCC != ISD::SETCC_INVALID && | |||
| 4547 | (!LegalOperations || | |||
| 4548 | (TLI.isCondCodeLegal(NewCC, LL.getSimpleValueType()) && | |||
| 4549 | TLI.isOperationLegal(ISD::SETCC, OpVT)))) | |||
| 4550 | return DAG.getSetCC(DL, VT, LL, LR, NewCC); | |||
| 4551 | } | |||
| 4552 | ||||
| 4553 | return SDValue(); | |||
| 4554 | } | |||
| 4555 | ||||
| 4556 | /// This contains all DAGCombine rules which reduce two values combined by | |||
| 4557 | /// an And operation to a single value. This makes them reusable in the context | |||
| 4558 | /// of visitSELECT(). Rules involving constants are not included as | |||
| 4559 | /// visitSELECT() already handles those cases. | |||
| 4560 | SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1, SDNode *N) { | |||
| 4561 | EVT VT = N1.getValueType(); | |||
| 4562 | SDLoc DL(N); | |||
| 4563 | ||||
| 4564 | // fold (and x, undef) -> 0 | |||
| 4565 | if (N0.isUndef() || N1.isUndef()) | |||
| 4566 | return DAG.getConstant(0, DL, VT); | |||
| 4567 | ||||
| 4568 | if (SDValue V = foldLogicOfSetCCs(true, N0, N1, DL)) | |||
| 4569 | return V; | |||
| 4570 | ||||
| 4571 | if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL && | |||
| 4572 | VT.getSizeInBits() <= 64) { | |||
| 4573 | if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | |||
| 4574 | if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) { | |||
| 4575 | // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal | |||
| 4576 | // immediate for an add, but it is legal if its top c2 bits are set, | |||
| 4577 | // transform the ADD so the immediate doesn't need to be materialized | |||
| 4578 | // in a register. | |||
| 4579 | APInt ADDC = ADDI->getAPIntValue(); | |||
| 4580 | APInt SRLC = SRLI->getAPIntValue(); | |||
| 4581 | if (ADDC.getMinSignedBits() <= 64 && | |||
| 4582 | SRLC.ult(VT.getSizeInBits()) && | |||
| 4583 | !TLI.isLegalAddImmediate(ADDC.getSExtValue())) { | |||
| 4584 | APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), | |||
| 4585 | SRLC.getZExtValue()); | |||
| 4586 | if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) { | |||
| 4587 | ADDC |= Mask; | |||
| 4588 | if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) { | |||
| 4589 | SDLoc DL0(N0); | |||
| 4590 | SDValue NewAdd = | |||
| 4591 | DAG.getNode(ISD::ADD, DL0, VT, | |||
| 4592 | N0.getOperand(0), DAG.getConstant(ADDC, DL, VT)); | |||
| 4593 | CombineTo(N0.getNode(), NewAdd); | |||
| 4594 | // Return N so it doesn't get rechecked! | |||
| 4595 | return SDValue(N, 0); | |||
| 4596 | } | |||
| 4597 | } | |||
| 4598 | } | |||
| 4599 | } | |||
| 4600 | } | |||
| 4601 | } | |||
| 4602 | ||||
| 4603 | // Reduce bit extract of low half of an integer to the narrower type. | |||
| 4604 | // (and (srl i64:x, K), KMask) -> | |||
| 4605 | // (i64 zero_extend (and (srl (i32 (trunc i64:x)), K)), KMask) | |||
| 4606 | if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { | |||
| 4607 | if (ConstantSDNode *CAnd = dyn_cast<ConstantSDNode>(N1)) { | |||
| 4608 | if (ConstantSDNode *CShift = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | |||
| 4609 | unsigned Size = VT.getSizeInBits(); | |||
| 4610 | const APInt &AndMask = CAnd->getAPIntValue(); | |||
| 4611 | unsigned ShiftBits = CShift->getZExtValue(); | |||
| 4612 | ||||
| 4613 | // Bail out, this node will probably disappear anyway. | |||
| 4614 | if (ShiftBits == 0) | |||
| 4615 | return SDValue(); | |||
| 4616 | ||||
| 4617 | unsigned MaskBits = AndMask.countTrailingOnes(); | |||
| 4618 | EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), Size / 2); | |||
| 4619 | ||||
| 4620 | if (AndMask.isMask() && | |||
| 4621 | // Required bits must not span the two halves of the integer and | |||
| 4622 | // must fit in the half size type. | |||
| 4623 | (ShiftBits + MaskBits <= Size / 2) && | |||
| 4624 | TLI.isNarrowingProfitable(VT, HalfVT) && | |||
| 4625 | TLI.isTypeDesirableForOp(ISD::AND, HalfVT) && | |||
| 4626 | TLI.isTypeDesirableForOp(ISD::SRL, HalfVT) && | |||
| 4627 | TLI.isTruncateFree(VT, HalfVT) && | |||
| 4628 | TLI.isZExtFree(HalfVT, VT)) { | |||
| 4629 | // The isNarrowingProfitable is to avoid regressions on PPC and | |||
| 4630 | // AArch64 which match a few 64-bit bit insert / bit extract patterns | |||
| 4631 | // on downstream users of this. Those patterns could probably be | |||
| 4632 | // extended to handle extensions mixed in. | |||
| 4633 | ||||
| 4634 | SDValue SL(N0); | |||
| 4635 | assert(MaskBits <= Size)((MaskBits <= Size) ? static_cast<void> (0) : __assert_fail ("MaskBits <= Size", "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4635, __PRETTY_FUNCTION__)); | |||
| 4636 | ||||
| 4637 | // Extracting the highest bit of the low half. | |||
| 4638 | EVT ShiftVT = TLI.getShiftAmountTy(HalfVT, DAG.getDataLayout()); | |||
| 4639 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, HalfVT, | |||
| 4640 | N0.getOperand(0)); | |||
| 4641 | ||||
| 4642 | SDValue NewMask = DAG.getConstant(AndMask.trunc(Size / 2), SL, HalfVT); | |||
| 4643 | SDValue ShiftK = DAG.getConstant(ShiftBits, SL, ShiftVT); | |||
| 4644 | SDValue Shift = DAG.getNode(ISD::SRL, SL, HalfVT, Trunc, ShiftK); | |||
| 4645 | SDValue And = DAG.getNode(ISD::AND, SL, HalfVT, Shift, NewMask); | |||
| 4646 | return DAG.getNode(ISD::ZERO_EXTEND, SL, VT, And); | |||
| 4647 | } | |||
| 4648 | } | |||
| 4649 | } | |||
| 4650 | } | |||
| 4651 | ||||
| 4652 | return SDValue(); | |||
| 4653 | } | |||
| 4654 | ||||
| 4655 | bool DAGCombiner::isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, | |||
| 4656 | EVT LoadResultTy, EVT &ExtVT) { | |||
| 4657 | if (!AndC->getAPIntValue().isMask()) | |||
| 4658 | return false; | |||
| 4659 | ||||
| 4660 | unsigned ActiveBits = AndC->getAPIntValue().countTrailingOnes(); | |||
| 4661 | ||||
| 4662 | ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); | |||
| 4663 | EVT LoadedVT = LoadN->getMemoryVT(); | |||
| 4664 | ||||
| 4665 | if (ExtVT == LoadedVT && | |||
| 4666 | (!LegalOperations || | |||
| 4667 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT))) { | |||
| 4668 | // ZEXTLOAD will match without needing to change the size of the value being | |||
| 4669 | // loaded. | |||
| 4670 | return true; | |||
| 4671 | } | |||
| 4672 | ||||
| 4673 | // Do not change the width of a volatile or atomic loads. | |||
| 4674 | if (!LoadN->isSimple()) | |||
| 4675 | return false; | |||
| 4676 | ||||
| 4677 | // Do not generate loads of non-round integer types since these can | |||
| 4678 | // be expensive (and would be wrong if the type is not byte sized). | |||
| 4679 | if (!LoadedVT.bitsGT(ExtVT) || !ExtVT.isRound()) | |||
| 4680 | return false; | |||
| 4681 | ||||
| 4682 | if (LegalOperations && | |||
| 4683 | !TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT)) | |||
| 4684 | return false; | |||
| 4685 | ||||
| 4686 | if (!TLI.shouldReduceLoadWidth(LoadN, ISD::ZEXTLOAD, ExtVT)) | |||
| 4687 | return false; | |||
| 4688 | ||||
| 4689 | return true; | |||
| 4690 | } | |||
| 4691 | ||||
| 4692 | bool DAGCombiner::isLegalNarrowLdSt(LSBaseSDNode *LDST, | |||
| 4693 | ISD::LoadExtType ExtType, EVT &MemVT, | |||
| 4694 | unsigned ShAmt) { | |||
| 4695 | if (!LDST) | |||
| 4696 | return false; | |||
| 4697 | // Only allow byte offsets. | |||
| 4698 | if (ShAmt % 8) | |||
| 4699 | return false; | |||
| 4700 | ||||
| 4701 | // Do not generate loads of non-round integer types since these can | |||
| 4702 | // be expensive (and would be wrong if the type is not byte sized). | |||
| 4703 | if (!MemVT.isRound()) | |||
| 4704 | return false; | |||
| 4705 | ||||
| 4706 | // Don't change the width of a volatile or atomic loads. | |||
| 4707 | if (!LDST->isSimple()) | |||
| 4708 | return false; | |||
| 4709 | ||||
| 4710 | // Verify that we are actually reducing a load width here. | |||
| 4711 | if (LDST->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits()) | |||
| 4712 | return false; | |||
| 4713 | ||||
| 4714 | // Ensure that this isn't going to produce an unsupported memory access. | |||
| 4715 | if (ShAmt && | |||
| 4716 | !TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), MemVT, | |||
| 4717 | LDST->getAddressSpace(), ShAmt / 8, | |||
| 4718 | LDST->getMemOperand()->getFlags())) | |||
| 4719 | return false; | |||
| 4720 | ||||
| 4721 | // It's not possible to generate a constant of extended or untyped type. | |||
| 4722 | EVT PtrType = LDST->getBasePtr().getValueType(); | |||
| 4723 | if (PtrType == MVT::Untyped || PtrType.isExtended()) | |||
| 4724 | return false; | |||
| 4725 | ||||
| 4726 | if (isa<LoadSDNode>(LDST)) { | |||
| 4727 | LoadSDNode *Load = cast<LoadSDNode>(LDST); | |||
| 4728 | // Don't transform one with multiple uses, this would require adding a new | |||
| 4729 | // load. | |||
| 4730 | if (!SDValue(Load, 0).hasOneUse()) | |||
| 4731 | return false; | |||
| 4732 | ||||
| 4733 | if (LegalOperations && | |||
| 4734 | !TLI.isLoadExtLegal(ExtType, Load->getValueType(0), MemVT)) | |||
| 4735 | return false; | |||
| 4736 | ||||
| 4737 | // For the transform to be legal, the load must produce only two values | |||
| 4738 | // (the value loaded and the chain). Don't transform a pre-increment | |||
| 4739 | // load, for example, which produces an extra value. Otherwise the | |||
| 4740 | // transformation is not equivalent, and the downstream logic to replace | |||
| 4741 | // uses gets things wrong. | |||
| 4742 | if (Load->getNumValues() > 2) | |||
| 4743 | return false; | |||
| 4744 | ||||
| 4745 | // If the load that we're shrinking is an extload and we're not just | |||
| 4746 | // discarding the extension we can't simply shrink the load. Bail. | |||
| 4747 | // TODO: It would be possible to merge the extensions in some cases. | |||
| 4748 | if (Load->getExtensionType() != ISD::NON_EXTLOAD && | |||
| 4749 | Load->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits() + ShAmt) | |||
| 4750 | return false; | |||
| 4751 | ||||
| 4752 | if (!TLI.shouldReduceLoadWidth(Load, ExtType, MemVT)) | |||
| 4753 | return false; | |||
| 4754 | } else { | |||
| 4755 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4755, __PRETTY_FUNCTION__)); | |||
| 4756 | StoreSDNode *Store = cast<StoreSDNode>(LDST); | |||
| 4757 | // Can't write outside the original store | |||
| 4758 | if (Store->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits() + ShAmt) | |||
| 4759 | return false; | |||
| 4760 | ||||
| 4761 | if (LegalOperations && | |||
| 4762 | !TLI.isTruncStoreLegal(Store->getValue().getValueType(), MemVT)) | |||
| 4763 | return false; | |||
| 4764 | } | |||
| 4765 | return true; | |||
| 4766 | } | |||
| 4767 | ||||
| 4768 | bool DAGCombiner::SearchForAndLoads(SDNode *N, | |||
| 4769 | SmallVectorImpl<LoadSDNode*> &Loads, | |||
| 4770 | SmallPtrSetImpl<SDNode*> &NodesWithConsts, | |||
| 4771 | ConstantSDNode *Mask, | |||
| 4772 | SDNode *&NodeToMask) { | |||
| 4773 | // Recursively search for the operands, looking for loads which can be | |||
| 4774 | // narrowed. | |||
| 4775 | for (SDValue Op : N->op_values()) { | |||
| 4776 | if (Op.getValueType().isVector()) | |||
| 4777 | return false; | |||
| 4778 | ||||
| 4779 | // Some constants may need fixing up later if they are too large. | |||
| 4780 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { | |||
| 4781 | if ((N->getOpcode() == ISD::OR || N->getOpcode() == ISD::XOR) && | |||
| 4782 | (Mask->getAPIntValue() & C->getAPIntValue()) != C->getAPIntValue()) | |||
| 4783 | NodesWithConsts.insert(N); | |||
| 4784 | continue; | |||
| 4785 | } | |||
| 4786 | ||||
| 4787 | if (!Op.hasOneUse()) | |||
| 4788 | return false; | |||
| 4789 | ||||
| 4790 | switch(Op.getOpcode()) { | |||
| 4791 | case ISD::LOAD: { | |||
| 4792 | auto *Load = cast<LoadSDNode>(Op); | |||
| 4793 | EVT ExtVT; | |||
| 4794 | if (isAndLoadExtLoad(Mask, Load, Load->getValueType(0), ExtVT) && | |||
| 4795 | isLegalNarrowLdSt(Load, ISD::ZEXTLOAD, ExtVT)) { | |||
| 4796 | ||||
| 4797 | // ZEXTLOAD is already small enough. | |||
| 4798 | if (Load->getExtensionType() == ISD::ZEXTLOAD && | |||
| 4799 | ExtVT.bitsGE(Load->getMemoryVT())) | |||
| 4800 | continue; | |||
| 4801 | ||||
| 4802 | // Use LE to convert equal sized loads to zext. | |||
| 4803 | if (ExtVT.bitsLE(Load->getMemoryVT())) | |||
| 4804 | Loads.push_back(Load); | |||
| 4805 | ||||
| 4806 | continue; | |||
| 4807 | } | |||
| 4808 | return false; | |||
| 4809 | } | |||
| 4810 | case ISD::ZERO_EXTEND: | |||
| 4811 | case ISD::AssertZext: { | |||
| 4812 | unsigned ActiveBits = Mask->getAPIntValue().countTrailingOnes(); | |||
| 4813 | EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); | |||
| 4814 | EVT VT = Op.getOpcode() == ISD::AssertZext ? | |||
| 4815 | cast<VTSDNode>(Op.getOperand(1))->getVT() : | |||
| 4816 | Op.getOperand(0).getValueType(); | |||
| 4817 | ||||
| 4818 | // We can accept extending nodes if the mask is wider or an equal | |||
| 4819 | // width to the original type. | |||
| 4820 | if (ExtVT.bitsGE(VT)) | |||
| 4821 | continue; | |||
| 4822 | break; | |||
| 4823 | } | |||
| 4824 | case ISD::OR: | |||
| 4825 | case ISD::XOR: | |||
| 4826 | case ISD::AND: | |||
| 4827 | if (!SearchForAndLoads(Op.getNode(), Loads, NodesWithConsts, Mask, | |||
| 4828 | NodeToMask)) | |||
| 4829 | return false; | |||
| 4830 | continue; | |||
| 4831 | } | |||
| 4832 | ||||
| 4833 | // Allow one node which will masked along with any loads found. | |||
| 4834 | if (NodeToMask) | |||
| 4835 | return false; | |||
| 4836 | ||||
| 4837 | // Also ensure that the node to be masked only produces one data result. | |||
| 4838 | NodeToMask = Op.getNode(); | |||
| 4839 | if (NodeToMask->getNumValues() > 1) { | |||
| 4840 | bool HasValue = false; | |||
| 4841 | for (unsigned i = 0, e = NodeToMask->getNumValues(); i < e; ++i) { | |||
| 4842 | MVT VT = SDValue(NodeToMask, i).getSimpleValueType(); | |||
| 4843 | if (VT != MVT::Glue && VT != MVT::Other) { | |||
| 4844 | if (HasValue) { | |||
| 4845 | NodeToMask = nullptr; | |||
| 4846 | return false; | |||
| 4847 | } | |||
| 4848 | HasValue = true; | |||
| 4849 | } | |||
| 4850 | } | |||
| 4851 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4851, __PRETTY_FUNCTION__)); | |||
| 4852 | } | |||
| 4853 | } | |||
| 4854 | return true; | |||
| 4855 | } | |||
| 4856 | ||||
| 4857 | bool DAGCombiner::BackwardsPropagateMask(SDNode *N, SelectionDAG &DAG) { | |||
| 4858 | auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
| 4859 | if (!Mask) | |||
| 4860 | return false; | |||
| 4861 | ||||
| 4862 | if (!Mask->getAPIntValue().isMask()) | |||
| 4863 | return false; | |||
| 4864 | ||||
| 4865 | // No need to do anything if the and directly uses a load. | |||
| 4866 | if (isa<LoadSDNode>(N->getOperand(0))) | |||
| 4867 | return false; | |||
| 4868 | ||||
| 4869 | SmallVector<LoadSDNode*, 8> Loads; | |||
| 4870 | SmallPtrSet<SDNode*, 2> NodesWithConsts; | |||
| 4871 | SDNode *FixupNode = nullptr; | |||
| 4872 | if (SearchForAndLoads(N, Loads, NodesWithConsts, Mask, FixupNode)) { | |||
| 4873 | if (Loads.size() == 0) | |||
| 4874 | return false; | |||
| 4875 | ||||
| 4876 | LLVM_DEBUG(dbgs() << "Backwards propagate AND: "; N->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "Backwards propagate AND: " ; N->dump(); } } while (false); | |||
| 4877 | SDValue MaskOp = N->getOperand(1); | |||
| 4878 | ||||
| 4879 | // If it exists, fixup the single node we allow in the tree that needs | |||
| 4880 | // masking. | |||
| 4881 | if (FixupNode) { | |||
| 4882 | 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); | |||
| 4883 | SDValue And = DAG.getNode(ISD::AND, SDLoc(FixupNode), | |||
| 4884 | FixupNode->getValueType(0), | |||
| 4885 | SDValue(FixupNode, 0), MaskOp); | |||
| 4886 | DAG.ReplaceAllUsesOfValueWith(SDValue(FixupNode, 0), And); | |||
| 4887 | if (And.getOpcode() == ISD ::AND) | |||
| 4888 | DAG.UpdateNodeOperands(And.getNode(), SDValue(FixupNode, 0), MaskOp); | |||
| 4889 | } | |||
| 4890 | ||||
| 4891 | // Narrow any constants that need it. | |||
| 4892 | for (auto *LogicN : NodesWithConsts) { | |||
| 4893 | SDValue Op0 = LogicN->getOperand(0); | |||
| 4894 | SDValue Op1 = LogicN->getOperand(1); | |||
| 4895 | ||||
| 4896 | if (isa<ConstantSDNode>(Op0)) | |||
| 4897 | std::swap(Op0, Op1); | |||
| 4898 | ||||
| 4899 | SDValue And = DAG.getNode(ISD::AND, SDLoc(Op1), Op1.getValueType(), | |||
| 4900 | Op1, MaskOp); | |||
| 4901 | ||||
| 4902 | DAG.UpdateNodeOperands(LogicN, Op0, And); | |||
| 4903 | } | |||
| 4904 | ||||
| 4905 | // Create narrow loads. | |||
| 4906 | for (auto *Load : Loads) { | |||
| 4907 | 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); | |||
| 4908 | SDValue And = DAG.getNode(ISD::AND, SDLoc(Load), Load->getValueType(0), | |||
| 4909 | SDValue(Load, 0), MaskOp); | |||
| 4910 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), And); | |||
| 4911 | if (And.getOpcode() == ISD ::AND) | |||
| 4912 | And = SDValue( | |||
| 4913 | DAG.UpdateNodeOperands(And.getNode(), SDValue(Load, 0), MaskOp), 0); | |||
| 4914 | SDValue NewLoad = ReduceLoadWidth(And.getNode()); | |||
| 4915 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4916, __PRETTY_FUNCTION__)) | |||
| 4916 | "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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4916, __PRETTY_FUNCTION__)); | |||
| 4917 | CombineTo(Load, NewLoad, NewLoad.getValue(1)); | |||
| 4918 | } | |||
| 4919 | DAG.ReplaceAllUsesWith(N, N->getOperand(0).getNode()); | |||
| 4920 | return true; | |||
| 4921 | } | |||
| 4922 | return false; | |||
| 4923 | } | |||
| 4924 | ||||
| 4925 | // Unfold | |||
| 4926 | // x & (-1 'logical shift' y) | |||
| 4927 | // To | |||
| 4928 | // (x 'opposite logical shift' y) 'logical shift' y | |||
| 4929 | // if it is better for performance. | |||
| 4930 | SDValue DAGCombiner::unfoldExtremeBitClearingToShifts(SDNode *N) { | |||
| 4931 | assert(N->getOpcode() == ISD::AND)((N->getOpcode() == ISD::AND) ? static_cast<void> (0 ) : __assert_fail ("N->getOpcode() == ISD::AND", "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4931, __PRETTY_FUNCTION__)); | |||
| 4932 | ||||
| 4933 | SDValue N0 = N->getOperand(0); | |||
| 4934 | SDValue N1 = N->getOperand(1); | |||
| 4935 | ||||
| 4936 | // Do we actually prefer shifts over mask? | |||
| 4937 | if (!TLI.shouldFoldMaskToVariableShiftPair(N0)) | |||
| 4938 | return SDValue(); | |||
| 4939 | ||||
| 4940 | // Try to match (-1 '[outer] logical shift' y) | |||
| 4941 | unsigned OuterShift; | |||
| 4942 | unsigned InnerShift; // The opposite direction to the OuterShift. | |||
| 4943 | SDValue Y; // Shift amount. | |||
| 4944 | auto matchMask = [&OuterShift, &InnerShift, &Y](SDValue M) -> bool { | |||
| 4945 | if (!M.hasOneUse()) | |||
| 4946 | return false; | |||
| 4947 | OuterShift = M->getOpcode(); | |||
| 4948 | if (OuterShift == ISD::SHL) | |||
| 4949 | InnerShift = ISD::SRL; | |||
| 4950 | else if (OuterShift == ISD::SRL) | |||
| 4951 | InnerShift = ISD::SHL; | |||
| 4952 | else | |||
| 4953 | return false; | |||
| 4954 | if (!isAllOnesConstant(M->getOperand(0))) | |||
| 4955 | return false; | |||
| 4956 | Y = M->getOperand(1); | |||
| 4957 | return true; | |||
| 4958 | }; | |||
| 4959 | ||||
| 4960 | SDValue X; | |||
| 4961 | if (matchMask(N1)) | |||
| 4962 | X = N0; | |||
| 4963 | else if (matchMask(N0)) | |||
| 4964 | X = N1; | |||
| 4965 | else | |||
| 4966 | return SDValue(); | |||
| 4967 | ||||
| 4968 | SDLoc DL(N); | |||
| 4969 | EVT VT = N->getValueType(0); | |||
| 4970 | ||||
| 4971 | // tmp = x 'opposite logical shift' y | |||
| 4972 | SDValue T0 = DAG.getNode(InnerShift, DL, VT, X, Y); | |||
| 4973 | // ret = tmp 'logical shift' y | |||
| 4974 | SDValue T1 = DAG.getNode(OuterShift, DL, VT, T0, Y); | |||
| 4975 | ||||
| 4976 | return T1; | |||
| 4977 | } | |||
| 4978 | ||||
| 4979 | /// Try to replace shift/logic that tests if a bit is clear with mask + setcc. | |||
| 4980 | /// For a target with a bit test, this is expected to become test + set and save | |||
| 4981 | /// at least 1 instruction. | |||
| 4982 | static SDValue combineShiftAnd1ToBitTest(SDNode *And, SelectionDAG &DAG) { | |||
| 4983 | 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-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4983, __PRETTY_FUNCTION__)); | |||
| 4984 | ||||
| 4985 | // This is probably not worthwhile without a supported type. | |||
| 4986 | EVT VT = And->getValueType(0); | |||
| 4987 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 4988 | if (!TLI.isTypeLegal(VT)) | |||
| 4989 | return SDValue(); | |||
| 4990 | ||||
| 4991 | // Look through an optional extension and find a 'not'. | |||
| 4992 | // TODO: Should we favor test+set even without the 'not' op? | |||
| 4993 | SDValue Not = And->getOperand(0), And1 = And->getOperand(1); | |||
| 4994 | if (Not.getOpcode() == ISD::ANY_EXTEND) | |||
| 4995 | Not = Not.getOperand(0); | |||
| 4996 | if (!isBitwiseNot(Not) || !Not.hasOneUse() || !isOneConstant(And1)) | |||
| 4997 | return SDValue(); | |||
| 4998 | ||||
| 4999 | // Look though an optional truncation. The source operand may not be the same | |||
| 5000 | // type as the original 'and', but that is ok because we are masking off | |||
| 5001 | // everything but the low bit. | |||
| 5002 | SDValue Srl = Not.getOperand(0); | |||
| 5003 | if (Srl.getOpcode() == ISD::TRUNCATE) | |||
| 5004 | Srl = Srl.getOperand(0); | |||
| 5005 | ||||
| 5006 | // Match a shift-right by constant. | |||
| 5007 | if (Srl.getOpcode() != ISD::SRL || !Srl.hasOneUse() || | |||
| 5008 | !isa<ConstantSDNode>(Srl.getOperand(1))) | |||
| 5009 | return SDValue(); | |||
| 5010 | ||||
| 5011 | // We might have looked through casts that make this transform invalid. | |||
| 5012 | // TODO: If the source type is wider than the result type, do the mask and | |||
| 5013 | // compare in the source type. | |||
| 5014 | const APInt &ShiftAmt = Srl.getConstantOperandAPInt(1); | |||
| 5015 | unsigned VTBitWidth = VT.getSizeInBits(); | |||
| 5016 | if (ShiftAmt.uge(VTBitWidth)) | |||
| 5017 | return SDValue(); | |||
| 5018 | ||||
| 5019 | // Turn this into a bit-test pattern using mask op + setcc: | |||
| 5020 | // and (not (srl X, C)), 1 --> (and X, 1<<C) == 0 | |||
| 5021 | SDLoc DL(And); | |||
| 5022 | SDValue X = DAG.getZExtOrTrunc(Srl.getOperand(0), DL, VT); | |||
| 5023 | EVT CCVT = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
| 5024 | SDValue Mask = DAG.getConstant( | |||
| 5025 | APInt::getOneBitSet(VTBitWidth, ShiftAmt.getZExtValue()), DL, VT); | |||
| 5026 | SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, Mask); | |||
| 5027 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
| 5028 | SDValue Setcc = DAG.getSetCC(DL, CCVT, NewAnd, Zero, ISD::SETEQ); | |||
| 5029 | return DAG.getZExtOrTrunc(Setcc, DL, VT); | |||
| 5030 | } | |||
| 5031 | ||||
| 5032 | SDValue DAGCombiner::visitAND(SDNode *N) { | |||
| 5033 | SDValue N0 = N->getOperand(0); | |||
| 5034 | SDValue N1 = N->getOperand(1); | |||
| 5035 | EVT VT = N1.getValueType(); | |||
| 5036 | ||||
| 5037 | // x & x --> x | |||
| 5038 | if (N0 == N1) | |||
| 5039 | return N0; | |||
| 5040 | ||||
| 5041 | // fold vector ops | |||
| 5042 | if (VT.isVector()) { | |||
| 5043 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 5044 | return FoldedVOp; | |||
| 5045 | ||||
| 5046 | // fold (and x, 0) -> 0, vector edition | |||
| 5047 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
| 5048 | // do not return N0, because undef node may exist in N0 | |||
| 5049 | return DAG.getConstant(APInt::getNullValue(N0.getScalarValueSizeInBits()), | |||
| 5050 | SDLoc(N), N0.getValueType()); | |||
| 5051 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 5052 | // do not return N1, because undef node may exist in N1 | |||
| 5053 | return DAG.getConstant(APInt::getNullValue(N1.getScalarValueSizeInBits()), | |||
| 5054 | SDLoc(N), N1.getValueType()); | |||
| 5055 | ||||
| 5056 | // fold (and x, -1) -> x, vector edition | |||
| 5057 | if (ISD::isBuildVectorAllOnes(N0.getNode())) | |||
| 5058 | return N1; | |||
| 5059 | if (ISD::isBuildVectorAllOnes(N1.getNode())) | |||
| 5060 | return N0; | |||
| 5061 | } | |||
| 5062 | ||||
| 5063 | // fold (and c1, c2) -> c1&c2 | |||
| 5064 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
| 5065 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
| 5066 | if (N0C && N1C && !N1C->isOpaque()) | |||
| 5067 | return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C); | |||
| 5068 | // canonicalize constant to RHS | |||
| 5069 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 5070 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 5071 | return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0); | |||
| 5072 | // fold (and x, -1) -> x | |||
| 5073 | if (isAllOnesConstant(N1)) | |||
| 5074 | return N0; | |||
| 5075 | // if (and x, c) is known to be zero, return 0 | |||
| 5076 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
| 5077 | if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), | |||
| 5078 | APInt::getAllOnesValue(BitWidth))) | |||
| 5079 | return DAG.getConstant(0, SDLoc(N), VT); | |||
| 5080 | ||||
| 5081 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 5082 | return NewSel; | |||
| 5083 | ||||
| 5084 | // reassociate and | |||
| 5085 | if (SDValue RAND = reassociateOps(ISD::AND, SDLoc(N), N0, N1, N->getFlags())) | |||
| 5086 | return RAND; | |||
| 5087 | ||||
| 5088 | // Try to convert a constant mask AND into a shuffle clear mask. | |||
| 5089 | if (VT.isVector()) | |||
| 5090 | if (SDValue Shuffle = XformToShuffleWithZero(N)) | |||
| 5091 | return Shuffle; | |||
| 5092 | ||||
| 5093 | // fold (and (or x, C), D) -> D if (C & D) == D | |||
| 5094 | auto MatchSubset = [](ConstantSDNode *LHS, ConstantSDNode *RHS) { | |||
| 5095 | return RHS->getAPIntValue().isSubsetOf(LHS->getAPIntValue()); | |||
| 5096 | }; | |||
| 5097 | if (N0.getOpcode() == ISD::OR && | |||
| 5098 | ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchSubset)) | |||
| 5099 | return N1; | |||
| 5100 | // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. | |||
| 5101 | if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { | |||
| 5102 | SDValue N0Op0 = N0.getOperand(0); | |||
| 5103 | APInt Mask = ~N1C->getAPIntValue(); | |||
| 5104 | Mask = Mask.trunc(N0Op0.getScalarValueSizeInBits()); | |||
| 5105 | if (DAG.MaskedValueIsZero(N0Op0, Mask)) { | |||
| 5106 | SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), | |||
| 5107 | N0.getValueType(), N0Op0); | |||
| 5108 | ||||
| 5109 | // Replace uses of the AND with uses of the Zero extend node. | |||
| 5110 | CombineTo(N, Zext); | |||
| 5111 | ||||
| 5112 | // We actually want to replace all uses of the any_extend with the | |||
| 5113 | // zero_extend, to avoid duplicating things. This will later cause this | |||
| 5114 | // AND to be folded. | |||
| 5115 | CombineTo(N0.getNode(), Zext); | |||
| 5116 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
| 5117 | } | |||
| 5118 | } | |||
| 5119 | ||||
| 5120 | // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) -> | |||
| 5121 | // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must | |||
| 5122 | // already be zero by virtue of the width of the base type of the load. | |||
| 5123 | // | |||
| 5124 | // the 'X' node here can either be nothing or an extract_vector_elt to catch | |||
| 5125 | // more cases. | |||
| 5126 | if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
| 5127 | N0.getValueSizeInBits() == N0.getOperand(0).getScalarValueSizeInBits() && | |||
| 5128 | N0.getOperand(0).getOpcode() == ISD::LOAD && | |||
| 5129 | N0.getOperand(0).getResNo() == 0) || | |||
| 5130 | (N0.getOpcode() == ISD::LOAD && N0.getResNo() == 0)) { | |||
| 5131 | LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ? | |||
| 5132 | N0 : N0.getOperand(0) ); | |||
| 5133 | ||||
| 5134 | // Get the constant (if applicable) the zero'th operand is being ANDed with. | |||
| 5135 | // This can be a pure constant or a vector splat, in which case we treat the | |||
| 5136 | // vector as a scalar and use the splat value. | |||
| 5137 | APInt Constant = APInt::getNullValue(1); | |||
| 5138 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { | |||
| 5139 | Constant = C->getAPIntValue(); | |||
| 5140 | } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) { | |||
| 5141 | APInt SplatValue, SplatUndef; | |||
| 5142 | unsigned SplatBitSize; | |||
| 5143 | bool HasAnyUndefs; | |||
| 5144 | bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef, | |||
| 5145 | SplatBitSize, HasAnyUndefs); | |||
| 5146 | if (IsSplat) { | |||
| 5147 | // Undef bits can contribute to a possible optimisation if set, so | |||
| 5148 | // set them. | |||
| 5149 | SplatValue |= SplatUndef; | |||
| 5150 | ||||
| 5151 | // The splat value may be something like "0x00FFFFFF", which means 0 for | |||
| 5152 | // the first vector value and FF for the rest, repeating. We need a mask | |||
| 5153 | // that will apply equally to all members of the vector, so AND all the | |||
| 5154 | // lanes of the constant together. | |||
| 5155 | unsigned EltBitWidth = Vector->getValueType(0).getScalarSizeInBits(); | |||
| 5156 | ||||
| 5157 | // If the splat value has been compressed to a bitlength lower | |||
| 5158 | // than the size of the vector lane, we need to re-expand it to | |||
| 5159 | // the lane size. | |||
| 5160 | if (EltBitWidth > SplatBitSize) | |||
| 5161 | for (SplatValue = SplatValue.zextOrTrunc(EltBitWidth); | |||
| 5162 | SplatBitSize < EltBitWidth; SplatBitSize = SplatBitSize * 2) | |||
| 5163 | SplatValue |= SplatValue.shl(SplatBitSize); | |||
| 5164 | ||||
| 5165 | // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a | |||
| 5166 | // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value. | |||
| 5167 | if ((SplatBitSize % EltBitWidth) == 0) { | |||
| 5168 | Constant = APInt::getAllOnesValue(EltBitWidth); | |||
| 5169 | for (unsigned i = 0, n = (SplatBitSize / EltBitWidth); i < n; ++i) | |||
| 5170 | Constant &= SplatValue.extractBits(EltBitWidth, i * EltBitWidth); | |||
| 5171 | } | |||
| 5172 | } | |||
| 5173 | } | |||
| 5174 | ||||
| 5175 | // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is | |||
| 5176 | // actually legal and isn't going to get expanded, else this is a false | |||
| 5177 | // optimisation. | |||
| 5178 | bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD, | |||
| 5179 | Load->getValueType(0), | |||
| 5180 | Load->getMemoryVT()); | |||
| 5181 | ||||
| 5182 | // Resize the constant to the same size as the original memory access before | |||
| 5183 | // extension. If it is still the AllOnesValue then this AND is completely | |||
| 5184 | // unneeded. | |||
| 5185 | Constant = Constant.zextOrTrunc(Load->getMemoryVT().getScalarSizeInBits()); | |||
| 5186 | ||||
| 5187 | bool B; | |||
| 5188 | switch (Load->getExtensionType()) { | |||
| 5189 | default: B = false; break; | |||
| 5190 | case ISD::EXTLOAD: B = CanZextLoadProfitably; break; | |||
| 5191 | case ISD::ZEXTLOAD: | |||
| 5192 | case ISD::NON_EXTLOAD: B = true; break; | |||
| 5193 | } | |||
| 5194 | ||||
| 5195 | if (B && Constant.isAllOnesValue()) { | |||
| 5196 | // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to | |||
| 5197 | // preserve semantics once we get rid of the AND. | |||
| 5198 | SDValue NewLoad(Load, 0); | |||
| 5199 | ||||
| 5200 | // Fold the AND away. NewLoad may get replaced immediately. | |||
| 5201 | CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0); | |||
| 5202 | ||||
| 5203 | if (Load->getExtensionType() == ISD::EXTLOAD) { | |||
| 5204 | NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD, | |||
| 5205 | Load->getValueType(0), SDLoc(Load), | |||
| 5206 | Load->getChain(), Load->getBasePtr(), | |||
| 5207 | Load->getOffset(), Load->getMemoryVT(), | |||
| 5208 | Load->getMemOperand()); | |||
| 5209 | // Replace uses of the EXTLOAD with the new ZEXTLOAD. | |||
| 5210 | if (Load->getNumValues() == 3) { | |||
| 5211 | // PRE/POST_INC loads have 3 values. | |||
| 5212 | SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1), | |||
| 5213 | NewLoad.getValue(2) }; | |||
| 5214 | CombineTo(Load, To, 3, true); | |||
| 5215 | } else { | |||
| 5216 | CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1)); | |||
| 5217 | } | |||
| 5218 | } | |||
| 5219 | ||||
| 5220 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
| 5221 | } | |||
| 5222 | } | |||
| 5223 | ||||
| 5224 | // fold (and (load x), 255) -> (zextload x, i8) | |||
| 5225 | // fold (and (extload x, i16), 255) -> (zextload x, i8) | |||
| 5226 | // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8) | |||
| 5227 | if (!VT.isVector() && N1C && (N0.getOpcode() == ISD::LOAD || | |||
| 5228 | (N0.getOpcode() == ISD::ANY_EXTEND && | |||
| 5229 | N0.getOperand(0).getOpcode() == ISD::LOAD))) { | |||
| 5230 | if (SDValue Res = ReduceLoadWidth(N)) { | |||
| 5231 | LoadSDNode *LN0 = N0->getOpcode() == ISD::ANY_EXTEND | |||
| 5232 | ? cast<LoadSDNode>(N0.getOperand(0)) : cast<LoadSDNode>(N0); | |||
| 5233 | AddToWorklist(N); | |||
| 5234 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 0), Res); | |||
| 5235 | return SDValue(N, 0); | |||
| 5236 | } | |||
| 5237 | } | |||
| 5238 | ||||
| 5239 | if (Level >= AfterLegalizeTypes) { | |||
| 5240 | // Attempt to propagate the AND back up to the leaves which, if they're | |||
| 5241 | // loads, can be combined to narrow loads and the AND node can be removed. | |||
| 5242 | // Perform after legalization so that extend nodes will already be | |||
| 5243 | // combined into the loads. | |||
| 5244 | if (BackwardsPropagateMask(N, DAG)) { | |||
| 5245 | return SDValue(N, 0); | |||
| 5246 | } | |||
| 5247 | } | |||
| 5248 | ||||
| 5249 | if (SDValue Combined = visitANDLike(N0, N1, N)) | |||
| 5250 | return Combined; | |||
| 5251 | ||||
| 5252 | // Simplify: (and (op x...), (op y...)) -> (op (and x, y)) | |||
| 5253 | if (N0.getOpcode() == N1.getOpcode()) | |||
| 5254 | if (SDValue V = hoistLogicOpWithSameOpcodeHands(N)) | |||
| 5255 | return V; | |||
| 5256 | ||||
| 5257 | // Masking the negated extension of a boolean is just the zero-extended | |||
| 5258 | // boolean: | |||
| 5259 | // and (sub 0, zext(bool X)), 1 --> zext(bool X) | |||
| 5260 | // and (sub 0, sext(bool X)), 1 --> zext(bool X) | |||
| 5261 | // | |||
| 5262 | // Note: the SimplifyDemandedBits fold below can make an information-losing | |||
| 5263 | // transform, and then we have no way to find this better fold. | |||
| 5264 | if (N1C && N1C->isOne() && N0.getOpcode() == ISD::SUB) { | |||
| 5265 | if (isNullOrNullSplat(N0.getOperand(0))) { | |||
| 5266 | SDValue SubRHS = N0.getOperand(1); | |||
| 5267 | if (SubRHS.getOpcode() == ISD::ZERO_EXTEND && | |||
| 5268 | SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) | |||
| 5269 | return SubRHS; | |||
| 5270 | if (SubRHS.getOpcode() == ISD::SIGN_EXTEND && | |||
| 5271 | SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) | |||
| 5272 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, SubRHS.getOperand(0)); | |||
| 5273 | } | |||
| 5274 | } | |||
| 5275 | ||||
| 5276 | // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) | |||
| 5277 | // fold (and (sra)) -> (and (srl)) when possible. | |||
| 5278 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
| 5279 | return SDValue(N, 0); | |||
| 5280 | ||||
| 5281 | // fold (zext_inreg (extload x)) -> (zextload x) | |||
| 5282 | // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use | |||
| 5283 | if (ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
| 5284 | (ISD::isEXTLoad(N0.getNode()) || | |||
| 5285 | (ISD::isSEXTLoad(N0.getNode()) && N0.hasOneUse()))) { | |||
| 5286 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
| 5287 | EVT MemVT = LN0->getMemoryVT(); | |||
| 5288 | // If we zero all the possible extended bits, then we can turn this into | |||
| 5289 | // a zextload if we are running before legalize or the operation is legal. | |||
| 5290 | unsigned ExtBitSize = N1.getScalarValueSizeInBits(); | |||
| 5291 | unsigned MemBitSize = MemVT.getScalarSizeInBits(); | |||
| 5292 | APInt ExtBits = APInt::getHighBitsSet(ExtBitSize, ExtBitSize - MemBitSize); | |||
| 5293 | if (DAG.MaskedValueIsZero(N1, ExtBits) && | |||
| 5294 | ((!LegalOperations && LN0->isSimple()) || | |||
| 5295 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) { | |||
| 5296 | SDValue ExtLoad = | |||
| 5297 | DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, LN0->getChain(), | |||
| 5298 | LN0->getBasePtr(), MemVT, LN0->getMemOperand()); | |||
| 5299 | AddToWorklist(N); | |||
| 5300 | CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); | |||
| 5301 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
| 5302 | } | |||
| 5303 | } | |||
| 5304 | ||||
| 5305 | // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const) | |||
| 5306 | if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) { | |||
| 5307 | if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), | |||
| 5308 | N0.getOperand(1), false)) | |||
| 5309 | return BSwap; | |||
| 5310 | } | |||
| 5311 | ||||
| 5312 | if (SDValue Shifts = unfoldExtremeBitClearingToShifts(N)) | |||
| 5313 | return Shifts; | |||
| 5314 | ||||
| 5315 | if (TLI.hasBitTest(N0, N1)) | |||
| 5316 | if (SDValue V = combineShiftAnd1ToBitTest(N, DAG)) | |||
| 5317 | return V; | |||
| 5318 | ||||
| 5319 | return SDValue(); | |||
| 5320 | } | |||
| 5321 | ||||
| 5322 | /// Match (a >> 8) | (a << 8) as (bswap a) >> 16. | |||
| 5323 | SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, | |||
| 5324 | bool DemandHighBits) { | |||
| 5325 | if (!LegalOperations) | |||
| 5326 | return SDValue(); | |||
| 5327 | ||||
| 5328 | EVT VT = N->getValueType(0); | |||
| 5329 | if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16) | |||
| 5330 | return SDValue(); | |||
| 5331 | if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) | |||
| 5332 | return SDValue(); | |||
| 5333 | ||||
| 5334 | // Recognize (and (shl a, 8), 0xff00), (and (srl a, 8), 0xff) | |||
| 5335 | bool LookPassAnd0 = false; | |||
| 5336 | bool LookPassAnd1 = false; | |||
| 5337 | if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL) | |||
| 5338 | std::swap(N0, N1); | |||
| 5339 | if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL) | |||
| 5340 | std::swap(N0, N1); | |||
| 5341 | if (N0.getOpcode() == ISD::AND) { | |||
| 5342 | if (!N0.getNode()->hasOneUse()) | |||
| 5343 | return SDValue(); | |||
| 5344 | ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
| 5345 | // Also handle 0xffff since the LHS is guaranteed to have zeros there. | |||
| 5346 | // This is needed for X86. | |||
| 5347 | if (!N01C || (N01C->getZExtValue() != 0xFF00 && | |||
| 5348 | N01C->getZExtValue() != 0xFFFF)) | |||
| 5349 | return SDValue(); | |||
| 5350 | N0 = N0.getOperand(0); | |||
| 5351 | LookPassAnd0 = true; | |||
| 5352 | } | |||
| 5353 | ||||
| 5354 | if (N1.getOpcode() == ISD::AND) { | |||
| 5355 | if (!N1.getNode()->hasOneUse()) | |||
| 5356 | return SDValue(); | |||
| 5357 | ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); | |||
| 5358 | if (!N11C || N11C->getZExtValue() != 0xFF) | |||
| 5359 | return SDValue(); | |||
| 5360 | N1 = N1.getOperand(0); | |||
| 5361 | LookPassAnd1 = true; | |||
| 5362 | } | |||
| 5363 | ||||
| 5364 | if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) | |||
| 5365 | std::swap(N0, N1); | |||
| 5366 | if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) | |||
| 5367 | return SDValue(); | |||
| 5368 | if (!N0.getNode()->hasOneUse() || !N1.getNode()->hasOneUse()) | |||
| 5369 | return SDValue(); | |||
| 5370 | ||||
| 5371 | ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
| 5372 | ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); | |||
| 5373 | if (!N01C || !N11C) | |||
| 5374 | return SDValue(); | |||
| 5375 | if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8) | |||
| 5376 | return SDValue(); | |||
| 5377 | ||||
| 5378 | // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8) | |||
| 5379 | SDValue N00 = N0->getOperand(0); | |||
| 5380 | if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) { | |||
| 5381 | if (!N00.getNode()->hasOneUse()) | |||
| 5382 | return SDValue(); | |||
| 5383 | ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1)); | |||
| 5384 | if (!N001C || N001C->getZExtValue() != 0xFF) | |||
| 5385 | return SDValue(); | |||
| 5386 | N00 = N00.getOperand(0); | |||
| 5387 | LookPassAnd0 = true; | |||
| 5388 | } | |||
| 5389 | ||||
| 5390 | SDValue N10 = N1->getOperand(0); | |||
| 5391 | if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) { | |||
| 5392 | if (!N10.getNode()->hasOneUse()) | |||
| 5393 | return SDValue(); | |||
| 5394 | ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1)); | |||
| 5395 | // Also allow 0xFFFF since the bits will be shifted out. This is needed | |||
| 5396 | // for X86. | |||
| 5397 | if (!N101C || (N101C->getZExtValue() != 0xFF00 && | |||
| 5398 | N101C->getZExtValue() != 0xFFFF)) | |||
| 5399 | return SDValue(); | |||
| 5400 | N10 = N10.getOperand(0); | |||
| 5401 | LookPassAnd1 = true; | |||
| 5402 | } | |||
| 5403 | ||||
| 5404 | if (N00 != N10) | |||
| 5405 | return SDValue(); | |||
| 5406 | ||||
| 5407 | // Make sure everything beyond the low halfword gets set to zero since the SRL | |||
| 5408 | // 16 will clear the top bits. | |||
| 5409 | unsigned OpSizeInBits = VT.getSizeInBits(); | |||
| 5410 | if (DemandHighBits && OpSizeInBits > 16) { | |||
| 5411 | // If the left-shift isn't masked out then the only way this is a bswap is | |||
| 5412 | // if all bits beyond the low 8 are 0. In that case the entire pattern | |||
| 5413 | // reduces to a left shift anyway: leave it for other parts of the combiner. | |||
| 5414 | if (!LookPassAnd0) | |||
| 5415 | return SDValue(); | |||
| 5416 | ||||
| 5417 | // However, if the right shift isn't masked out then it might be because | |||
| 5418 | // it's not needed. See if we can spot that too. | |||
| 5419 | if (!LookPassAnd1 && | |||
| 5420 | !DAG.MaskedValueIsZero( | |||
| 5421 | N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16))) | |||
| 5422 | return SDValue(); | |||
| 5423 | } | |||
| 5424 | ||||
| 5425 | SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); | |||
| 5426 | if (OpSizeInBits > 16) { | |||
| 5427 | SDLoc DL(N); | |||
| 5428 | Res = DAG.getNode(ISD::SRL, DL, VT, Res, | |||
| 5429 | DAG.getConstant(OpSizeInBits - 16, DL, | |||
| 5430 | getShiftAmountTy(VT))); | |||
| 5431 | } | |||
| 5432 | return Res; | |||
| 5433 | } | |||
| 5434 | ||||
| 5435 | /// Return true if the specified node is an element that makes up a 32-bit | |||
| 5436 | /// packed halfword byteswap. | |||
| 5437 | /// ((x & 0x000000ff) << 8) | | |||
| 5438 | /// ((x & 0x0000ff00) >> 8) | | |||
| 5439 | /// ((x & 0x00ff0000) << 8) | | |||
| 5440 | /// ((x & 0xff000000) >> 8) | |||
| 5441 | static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) { | |||
| 5442 | if (!N.getNode()->hasOneUse()) | |||
| 5443 | return false; | |||
| 5444 | ||||
| 5445 | unsigned Opc = N.getOpcode(); | |||
| 5446 | if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL) | |||
| 5447 | return false; | |||
| 5448 | ||||
| 5449 | SDValue N0 = N.getOperand(0); | |||
| 5450 | unsigned Opc0 = N0.getOpcode(); | |||
| 5451 | if (Opc0 != ISD::AND && Opc0 != ISD::SHL && Opc0 != ISD::SRL) | |||
| 5452 | return false; | |||
| 5453 | ||||
| 5454 | ConstantSDNode *N1C = nullptr; | |||
| 5455 | // SHL or SRL: look upstream for AND mask operand | |||
| 5456 | if (Opc == ISD::AND) | |||
| 5457 | N1C = dyn_cast<ConstantSDNode>(N.getOperand(1)); | |||
| 5458 | else if (Opc0 == ISD::AND) | |||
| 5459 | N1C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
| 5460 | if (!N1C) | |||
| 5461 | return false; | |||
| 5462 | ||||
| 5463 | unsigned MaskByteOffset; | |||
| 5464 | switch (N1C->getZExtValue()) { | |||
| 5465 | default: | |||
| 5466 | return false; | |||
| 5467 | case 0xFF: MaskByteOffset = 0; break; | |||
| 5468 | case 0xFF00: MaskByteOffset = 1; break; | |||
| 5469 | case 0xFFFF: | |||
| 5470 | // In case demanded bits didn't clear the bits that will be shifted out. | |||
| 5471 | // This is needed for X86. | |||
| 5472 | if (Opc == ISD::SRL || (Opc == ISD::AND && Opc0 == ISD::SHL)) { | |||
| 5473 | MaskByteOffset = 1; | |||
| 5474 | break; | |||
| 5475 | } | |||
| 5476 | return false; | |||
| 5477 | case 0xFF0000: MaskByteOffset = 2; break; | |||
| 5478 | case 0xFF000000: MaskByteOffset = 3; break; | |||
| 5479 | } | |||
| 5480 | ||||
| 5481 | // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00). | |||
| 5482 | if (Opc == ISD::AND) { | |||
| 5483 | if (MaskByteOffset == 0 || MaskByteOffset == 2) { | |||
| 5484 | // (x >> 8) & 0xff | |||
| 5485 | // (x >> 8) & 0xff0000 | |||
| 5486 | if (Opc0 != ISD::SRL) | |||
| 5487 | return false; | |||
| 5488 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
| 5489 | if (!C || C->getZExtValue() != 8) | |||
| 5490 | return false; | |||
| 5491 | } else { | |||
| 5492 | // (x << 8) & 0xff00 | |||
| 5493 | // (x << 8) & 0xff000000 | |||
| 5494 | if (Opc0 != ISD::SHL) | |||
| 5495 | return false; | |||
| 5496 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
| 5497 | if (!C || C->getZExtValue() != 8) | |||
| 5498 | return false; | |||
| 5499 | } | |||
| 5500 | } else if (Opc == ISD::SHL) { | |||
| 5501 | // (x & 0xff) << 8 | |||
| 5502 | // (x & 0xff0000) << 8 | |||
| 5503 | if (MaskByteOffset != 0 && MaskByteOffset != 2) | |||
| 5504 | return false; | |||
| 5505 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); | |||
| 5506 | if (!C || C->getZExtValue() != 8) | |||
| 5507 | return false; | |||
| 5508 | } else { // Opc == ISD::SRL | |||
| 5509 | // (x & 0xff00) >> 8 | |||
| 5510 | // (x & 0xff000000) >> 8 | |||
| 5511 | if (MaskByteOffset != 1 && MaskByteOffset != 3) | |||
| 5512 | return false; | |||
| 5513 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); | |||
| 5514 | if (!C || C->getZExtValue() != 8) | |||
| 5515 | return false; | |||
| 5516 | } | |||
| 5517 | ||||
| 5518 | if (Parts[MaskByteOffset]) | |||
| 5519 | return false; | |||
| 5520 | ||||
| 5521 | Parts[MaskByteOffset] = N0.getOperand(0).getNode(); | |||
| 5522 | return true; | |||
| 5523 | } | |||
| 5524 | ||||
| 5525 | // Match 2 elements of a packed halfword bswap. | |||
| 5526 | static bool isBSwapHWordPair(SDValue N, MutableArrayRef<SDNode *> Parts) { | |||
| 5527 | if (N.getOpcode() == ISD::OR) | |||
| 5528 | return isBSwapHWordElement(N.getOperand(0), Parts) && | |||
| 5529 | isBSwapHWordElement(N.getOperand(1), Parts); | |||
| 5530 | ||||
| 5531 | if (N.getOpcode() == ISD::SRL && N.getOperand(0).getOpcode() == ISD::BSWAP) { | |||
| 5532 | ConstantSDNode *C = isConstOrConstSplat(N.getOperand(1)); | |||
| 5533 | if (!C || C->getAPIntValue() != 16) | |||
| 5534 | return false; | |||
| 5535 | Parts[0] = Parts[1] = N.getOperand(0).getOperand(0).getNode(); | |||
| 5536 | return true; | |||
| 5537 | } | |||
| 5538 | ||||
| 5539 | return false; | |||
| 5540 | } | |||
| 5541 | ||||
| 5542 | /// Match a 32-bit packed halfword bswap. That is | |||
| 5543 | /// ((x & 0x000000ff) << 8) | | |||
| 5544 | /// ((x & 0x0000ff00) >> 8) | | |||
| 5545 | /// ((x & 0x00ff0000) << 8) | | |||
| 5546 | /// ((x & 0xff000000) >> 8) | |||
| 5547 | /// => (rotl (bswap x), 16) | |||
| 5548 | SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) { | |||
| 5549 | if (!LegalOperations) | |||
| 5550 | return SDValue(); | |||
| 5551 | ||||
| 5552 | EVT VT = N->getValueType(0); | |||
| 5553 | if (VT != MVT::i32) | |||
| 5554 | return SDValue(); | |||
| 5555 | if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) | |||
| 5556 | return SDValue(); | |||
| 5557 | ||||
| 5558 | // Look for either | |||
| 5559 | // (or (bswaphpair), (bswaphpair)) | |||
| 5560 | // (or (or (bswaphpair), (and)), (and)) | |||
| 5561 | // (or (or (and), (bswaphpair)), (and)) | |||
| 5562 | SDNode *Parts[4] = {}; | |||
| 5563 | ||||
| 5564 | if (isBSwapHWordPair(N0, Parts)) { | |||
| 5565 | // (or (or (and), (and)), (or (and), (and))) | |||
| 5566 | if (!isBSwapHWordPair(N1, Parts)) | |||
| 5567 | return SDValue(); | |||
| 5568 | } else if (N0.getOpcode() == ISD::OR) { | |||
| 5569 | // (or (or (or (and), (and)), (and)), (and)) | |||
| 5570 | if (!isBSwapHWordElement(N1, Parts)) | |||
| 5571 | return SDValue(); | |||
| 5572 | SDValue N00 = N0.getOperand(0); | |||
| 5573 | SDValue N01 = N0.getOperand(1); | |||
| 5574 | if (!(isBSwapHWordElement(N01, Parts) && isBSwapHWordPair(N00, Parts)) && | |||
| 5575 | !(isBSwapHWordElement(N00, Parts) && isBSwapHWordPair(N01, Parts))) | |||
| 5576 | return SDValue(); | |||
| 5577 | } else | |||
| 5578 | return SDValue(); | |||
| 5579 | ||||
| 5580 | // Make sure the parts are all coming from the same node. | |||
| 5581 | if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3]) | |||
| 5582 | return SDValue(); | |||
| 5583 | ||||
| 5584 | SDLoc DL(N); | |||
| 5585 | SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, | |||
| 5586 | SDValue(Parts[0], 0)); | |||
| 5587 | ||||
| 5588 | // Result of the bswap should be rotated by 16. If it's not legal, then | |||
| 5589 | // do (x << 16) | (x >> 16). | |||
| 5590 | SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT)); | |||
| 5591 | if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT)) | |||
| 5592 | return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt); | |||
| 5593 | if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT)) | |||
| 5594 | return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt); | |||
| 5595 | return DAG.getNode(ISD::OR, DL, VT, | |||
| 5596 | DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt), | |||
| 5597 | DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt)); | |||
| 5598 | } | |||
| 5599 | ||||
| 5600 | /// This contains all DAGCombine rules which reduce two values combined by | |||
| 5601 | /// an Or operation to a single value \see visitANDLike(). | |||
| 5602 | SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *N) { | |||
| 5603 | EVT VT = N1.getValueType(); | |||
| 5604 | SDLoc DL(N); | |||
| 5605 | ||||
| 5606 | // fold (or x, undef) -> -1 | |||
| 5607 | if (!LegalOperations && (N0.isUndef() || N1.isUndef())) | |||
| 5608 | return DAG.getAllOnesConstant(DL, VT); | |||
| 5609 | ||||
| 5610 | if (SDValue V = foldLogicOfSetCCs(false, N0, N1, DL)) | |||
| 5611 | return V; | |||
| 5612 | ||||
| 5613 | // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible. | |||
| 5614 | if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND && | |||
| 5615 | // Don't increase # computations. | |||
| 5616 | (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) { | |||
| 5617 | // We can only do this xform if we know that bits from X that are set in C2 | |||
| 5618 | // but not in C1 are already zero. Likewise for Y. | |||
| 5619 | if (const ConstantSDNode *N0O1C = | |||
| 5620 | getAsNonOpaqueConstant(N0.getOperand(1))) { | |||
| 5621 | if (const ConstantSDNode *N1O1C = | |||
| 5622 | getAsNonOpaqueConstant(N1.getOperand(1))) { | |||
| 5623 | // We can only do this xform if we know that bits from X that are set in | |||
| 5624 | // C2 but not in C1 are already zero. Likewise for Y. | |||
| 5625 | const APInt &LHSMask = N0O1C->getAPIntValue(); | |||
| 5626 | const APInt &RHSMask = N1O1C->getAPIntValue(); | |||
| 5627 | ||||
| 5628 | if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) && | |||
| 5629 | DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) { | |||
| 5630 | SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, | |||
| 5631 | N0.getOperand(0), N1.getOperand(0)); | |||
| 5632 | return DAG.getNode(ISD::AND, DL, VT, X, | |||
| 5633 | DAG.getConstant(LHSMask | RHSMask, DL, VT)); | |||
| 5634 | } | |||
| 5635 | } | |||
| 5636 | } | |||
| 5637 | } | |||
| 5638 | ||||
| 5639 | // (or (and X, M), (and X, N)) -> (and X, (or M, N)) | |||
| 5640 | if (N0.getOpcode() == ISD::AND && | |||
| 5641 | N1.getOpcode() == ISD::AND && | |||
| 5642 | N0.getOperand(0) == N1.getOperand(0) && | |||
| 5643 | // Don't increase # computations. | |||
| 5644 | (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) { | |||
| 5645 | SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, | |||
| 5646 | N0.getOperand(1), N1.getOperand(1)); | |||
| 5647 | return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), X); | |||
| 5648 | } | |||
| 5649 | ||||
| 5650 | return SDValue(); | |||
| 5651 | } | |||
| 5652 | ||||
| 5653 | /// OR combines for which the commuted variant will be tried as well. | |||
| 5654 | static SDValue visitORCommutative( | |||
| 5655 | SelectionDAG &DAG, SDValue N0, SDValue N1, SDNode *N) { | |||
| 5656 | EVT VT = N0.getValueType(); | |||
| 5657 | if (N0.getOpcode() == ISD::AND) { | |||
| 5658 | // fold (or (and X, (xor Y, -1)), Y) -> (or X, Y) | |||
| 5659 | if (isBitwiseNot(N0.getOperand(1)) && N0.getOperand(1).getOperand(0) == N1) | |||
| 5660 | return DAG.getNode(ISD::OR, SDLoc(N), VT, N0.getOperand(0), N1); | |||
| 5661 | ||||
| 5662 | // fold (or (and (xor Y, -1), X), Y) -> (or X, Y) | |||
| 5663 | if (isBitwiseNot(N0.getOperand(0)) && N0.getOperand(0).getOperand(0) == N1) | |||
| 5664 | return DAG.getNode(ISD::OR, SDLoc(N), VT, N0.getOperand(1), N1); | |||
| 5665 | } | |||
| 5666 | ||||
| 5667 | return SDValue(); | |||
| 5668 | } | |||
| 5669 | ||||
| 5670 | SDValue DAGCombiner::visitOR(SDNode *N) { | |||
| 5671 | SDValue N0 = N->getOperand(0); | |||
| 5672 | SDValue N1 = N->getOperand(1); | |||
| 5673 | EVT VT = N1.getValueType(); | |||
| 5674 | ||||
| 5675 | // x | x --> x | |||
| 5676 | if (N0 == N1) | |||
| 5677 | return N0; | |||
| 5678 | ||||
| 5679 | // fold vector ops | |||
| 5680 | if (VT.isVector()) { | |||
| 5681 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
| 5682 | return FoldedVOp; | |||
| 5683 | ||||
| 5684 | // fold (or x, 0) -> x, vector edition | |||
| 5685 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
| 5686 | return N1; | |||
| 5687 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
| 5688 | return N0; | |||
| 5689 | ||||
| 5690 | // fold (or x, -1) -> -1, vector edition | |||
| 5691 | if (ISD::isBuildVectorAllOnes(N0.getNode())) | |||
| 5692 | // do not return N0, because undef node may exist in N0 | |||
| 5693 | return DAG.getAllOnesConstant(SDLoc(N), N0.getValueType()); | |||
| 5694 | if (ISD::isBuildVectorAllOnes(N1.getNode())) | |||
| 5695 | // do not return N1, because undef node may exist in N1 | |||
| 5696 | return DAG.getAllOnesConstant(SDLoc(N), N1.getValueType()); | |||
| 5697 | ||||
| 5698 | // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask) | |||
| 5699 | // Do this only if the resulting shuffle is legal. | |||
| 5700 | if (isa<ShuffleVectorSDNode>(N0) && | |||
| 5701 | isa<ShuffleVectorSDNode>(N1) && | |||
| 5702 | // Avoid folding a node with illegal type. | |||
| 5703 | TLI.isTypeLegal(VT)) { | |||
| 5704 | bool ZeroN00 = ISD::isBuildVectorAllZeros(N0.getOperand(0).getNode()); | |||
| 5705 | bool ZeroN01 = ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode()); | |||
| 5706 | bool ZeroN10 = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); | |||
| 5707 | bool ZeroN11 = ISD::isBuildVectorAllZeros(N1.getOperand(1).getNode()); | |||
| 5708 | // Ensure both shuffles have a zero input. | |||
| 5709 | if ((ZeroN00 != ZeroN01) && (ZeroN10 != ZeroN11)) { | |||
| 5710 | assert((!ZeroN00 || !ZeroN01) && "Both inputs zero!")(((!ZeroN00 || !ZeroN01) && "Both inputs zero!") ? static_cast <void> (0) : __assert_fail ("(!ZeroN00 || !ZeroN01) && \"Both inputs zero!\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5710, __PRETTY_FUNCTION__)); | |||
| 5711 | assert((!ZeroN10 || !ZeroN11) && "Both inputs zero!")(((!ZeroN10 || !ZeroN11) && "Both inputs zero!") ? static_cast <void> (0) : __assert_fail ("(!ZeroN10 || !ZeroN11) && \"Both inputs zero!\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5711, __PRETTY_FUNCTION__)); | |||
| 5712 | const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0); | |||
| 5713 | const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1); | |||
| 5714 | bool CanFold = true; | |||
| 5715 | int NumElts = VT.getVectorNumElements(); | |||
| 5716 | SmallVector<int, 4> Mask(NumElts); | |||
| 5717 | ||||
| 5718 | for (int i = 0; i != NumElts; ++i) { | |||
| 5719 | int M0 = SV0->getMaskElt(i); | |||
| 5720 | int M1 = SV1->getMaskElt(i); | |||
| 5721 | ||||
| 5722 | // Determine if either index is pointing to a zero vector. | |||
| 5723 | bool M0Zero = M0 < 0 || (ZeroN00 == (M0 < NumElts)); | |||
| 5724 | bool M1Zero = M1 < 0 || (ZeroN10 == (M1 < NumElts)); | |||
| 5725 | ||||
| 5726 | // If one element is zero and the otherside is undef, keep undef. | |||
| 5727 | // This also handles the case that both are undef. | |||
| 5728 | if ((M0Zero && M1 < 0) || (M1Zero && M0 < 0)) { | |||
| 5729 | Mask[i] = -1; | |||
| 5730 | continue; | |||
| 5731 | } | |||
| 5732 | ||||
| 5733 | // Make sure only one of the elements is zero. | |||
| 5734 | if (M0Zero == M1Zero) { | |||
| 5735 | CanFold = false; | |||
| 5736 | break; | |||
| 5737 | } | |||
| 5738 | ||||
| 5739 | assert((M0 >= 0 || M1 >= 0) && "Undef index!")(((M0 >= 0 || M1 >= 0) && "Undef index!") ? static_cast <void> (0) : __assert_fail ("(M0 >= 0 || M1 >= 0) && \"Undef index!\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5739, __PRETTY_FUNCTION__)); | |||
| 5740 | ||||
| 5741 | // We have a zero and non-zero element. If the non-zero came from | |||
| 5742 | // SV0 make the index a LHS index. If it came from SV1, make it | |||
| 5743 | // a RHS index. We need to mod by NumElts because we don't care | |||
| 5744 | // which operand it came from in the original shuffles. | |||
| 5745 | Mask[i] = M1Zero ? M0 % NumElts : (M1 % NumElts) + NumElts; | |||
| 5746 | } | |||
| 5747 | ||||
| 5748 | if (CanFold) { | |||
| 5749 | SDValue NewLHS = ZeroN00 ? N0.getOperand(1) : N0.getOperand(0); | |||
| 5750 | SDValue NewRHS = ZeroN10 ? N1.getOperand(1) : N1.getOperand(0); | |||
| 5751 | ||||
| 5752 | SDValue LegalShuffle = | |||
| 5753 | TLI.buildLegalVectorShuffle(VT, SDLoc(N), NewLHS, NewRHS, | |||
| 5754 | Mask, DAG); | |||
| 5755 | if (LegalShuffle) | |||
| 5756 | return LegalShuffle; | |||
| 5757 | } | |||
| 5758 | } | |||
| 5759 | } | |||
| 5760 | } | |||
| 5761 | ||||
| 5762 | // fold (or c1, c2) -> c1|c2 | |||
| 5763 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
| 5764 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
| 5765 | if (N0C && N1C && !N1C->isOpaque()) | |||
| 5766 | return DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, N0C, N1C); | |||
| 5767 | // canonicalize constant to RHS | |||
| 5768 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
| 5769 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
| 5770 | return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0); | |||
| 5771 | // fold (or x, 0) -> x | |||
| 5772 | if (isNullConstant(N1)) | |||
| 5773 | return N0; | |||
| 5774 | // fold (or x, -1) -> -1 | |||
| 5775 | if (isAllOnesConstant(N1)) | |||
| 5776 | return N1; | |||
| 5777 | ||||
| 5778 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
| 5779 | return NewSel; | |||
| 5780 | ||||
| 5781 | // fold (or x, c) -> c iff (x & ~c) == 0 | |||
| 5782 | if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue())) | |||
| 5783 | return N1; | |||
| 5784 | ||||
| 5785 | if (SDValue Combined = visitORLike(N0, N1, N)) | |||
| 5786 | return Combined; | |||
| 5787 | ||||
| 5788 | // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16) | |||
| 5789 | if (SDValue BSwap = MatchBSwapHWord(N, N0, N1)) | |||
| 5790 | return BSwap; | |||
| 5791 | if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1)) | |||
| 5792 | return BSwap; | |||
| 5793 | ||||
| 5794 | // reassociate or | |||
| 5795 | if (SDValue ROR = reassociateOps(ISD::OR, SDLoc(N), N0, N1, N->getFlags())) | |||
| 5796 | return ROR; | |||
| 5797 | ||||
| 5798 | // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) | |||
| 5799 | // iff (c1 & c2) != 0 or c1/c2 are undef. | |||
| 5800 | auto MatchIntersect = [](ConstantSDNode *C1, ConstantSDNode *C2) { | |||
| 5801 | return !C1 || !C2 || C1->getAPIntValue().intersects(C2->getAPIntValue()); | |||
| 5802 | }; | |||
| 5803 | if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && | |||
| 5804 | ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchIntersect, true)) { | |||
| 5805 | if (SDValue COR = DAG.FoldConstantArithmetic( | |||
| 5806 | ISD::OR, SDLoc(N1), VT, N1.getNode(), N0.getOperand(1).getNode())) { | |||
| 5807 | SDValue IOR = DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1); | |||
| 5808 | AddToWorklist(IOR.getNode()); | |||
| 5809 | return DAG.getNode(ISD::AND, SDLoc(N), VT, COR, IOR); | |||
| 5810 | } | |||
| 5811 | } | |||
| 5812 | ||||
| 5813 | if (SDValue Combined = visitORCommutative(DAG, N0, N1, N)) | |||
| 5814 | return Combined; | |||
| 5815 | if (SDValue Combined = visitORCommutative(DAG, N1, N0, N)) | |||
| 5816 | return Combined; | |||
| 5817 | ||||
| 5818 | // Simplify: (or (op x...), (op y...)) -> (op (or x, y)) | |||
| 5819 | if (N0.getOpcode() == N1.getOpcode()) | |||
| 5820 | if (SDValue V = hoistLogicOpWithSameOpcodeHands(N)) | |||
| 5821 | return V; | |||
| 5822 | ||||
| 5823 | // See if this is some rotate idiom. | |||
| 5824 | if (SDValue Rot = MatchRotate(N0, N1, SDLoc(N))) | |||
| 5825 | return Rot; | |||
| 5826 | ||||
| 5827 | if (SDValue Load = MatchLoadCombine(N)) | |||
| 5828 | return Load; | |||
| 5829 | ||||
| 5830 | // Simplify the operands using demanded-bits information. | |||
| 5831 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
| 5832 | return SDValue(N, 0); | |||
| 5833 | ||||
| 5834 | // If OR can be rewritten into ADD, try combines based on ADD. | |||
| 5835 | if ((!LegalOperations || TLI.isOperationLegal(ISD::ADD, VT)) && | |||
| 5836 | DAG.haveNoCommonBitsSet(N0, N1)) | |||
| 5837 | if (SDValue Combined = visitADDLike(N)) | |||
| 5838 | return Combined; | |||
| 5839 | ||||
| 5840 | return SDValue(); | |||
| 5841 | } | |||
| 5842 | ||||
| 5843 | static SDValue stripConstantMask(SelectionDAG &DAG, SDValue Op, SDValue &Mask) { | |||
| 5844 | if (Op.getOpcode() == ISD::AND && | |||
| 5845 | DAG.isConstantIntBuildVectorOrConstantInt(Op.getOperand(1))) { | |||
| 5846 | Mask = Op.getOperand(1); | |||
| 5847 | return Op.getOperand(0); | |||
| 5848 | } | |||
| 5849 | return Op; | |||
| 5850 | } | |||
| 5851 | ||||
| 5852 | /// Match "(X shl/srl V1) & V2" where V2 may not be present. | |||
| 5853 | static bool matchRotateHalf(SelectionDAG &DAG, SDValue Op, SDValue &Shift, | |||
| 5854 | SDValue &Mask) { | |||
| 5855 | Op = stripConstantMask(DAG, Op, Mask); | |||
| 5856 | if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) { | |||
| 5857 | Shift = Op; | |||
| 5858 | return true; | |||
| 5859 | } | |||
| 5860 | return false; | |||
| 5861 | } | |||
| 5862 | ||||
| 5863 | /// Helper function for visitOR to extract the needed side of a rotate idiom | |||
| 5864 | /// from a shl/srl/mul/udiv. This is meant to handle cases where | |||
| 5865 | /// InstCombine merged some outside op with one of the shifts from | |||
| 5866 | /// the rotate pattern. | |||
| 5867 | /// \returns An empty \c SDValue if the needed shift couldn't be extracted. | |||
| 5868 | /// Otherwise, returns an expansion of \p ExtractFrom based on the following | |||
| 5869 | /// patterns: | |||
| 5870 | /// | |||
| 5871 | /// (or (add v v) (shrl v bitwidth-1)): | |||
| 5872 | /// expands (add v v) -> (shl v 1) | |||
| 5873 | /// | |||
| 5874 | /// (or (mul v c0) (shrl (mul v c1) c2)): | |||
| 5875 | /// expands (mul v c0) -> (shl (mul v c1) c3) | |||
| 5876 | /// | |||
| 5877 | /// (or (udiv v c0) (shl (udiv v c1) c2)): | |||
| 5878 | /// expands (udiv v c0) -> (shrl (udiv v c1) c3) | |||
| 5879 | /// | |||
| 5880 | /// (or (shl v c0) (shrl (shl v c1) c2)): | |||
| 5881 | /// expands (shl v c0) -> (shl (shl v c1) c3) | |||
| 5882 | /// | |||
| 5883 | /// (or (shrl v c0) (shl (shrl v c1) c2)): | |||
| 5884 | /// expands (shrl v c0) -> (shrl (shrl v c1) c3) | |||
| 5885 | /// | |||
| 5886 | /// Such that in all cases, c3+c2==bitwidth(op v c1). | |||
| 5887 | static SDValue extractShiftForRotate(SelectionDAG &DAG, SDValue OppShift, | |||
| 5888 | SDValue ExtractFrom, SDValue &Mask, | |||
| 5889 | const SDLoc &DL) { | |||
| 5890 | assert(OppShift && ExtractFrom && "Empty SDValue")((OppShift && ExtractFrom && "Empty SDValue") ? static_cast<void> (0) : __assert_fail ("OppShift && ExtractFrom && \"Empty SDValue\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5890, __PRETTY_FUNCTION__)); | |||
| 5891 | assert((((OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) && "Existing shift must be valid as a rotate half" ) ? static_cast<void> (0) : __assert_fail ("(OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) && \"Existing shift must be valid as a rotate half\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5893, __PRETTY_FUNCTION__)) | |||
| 5892 | (OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) &&(((OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) && "Existing shift must be valid as a rotate half" ) ? static_cast<void> (0) : __assert_fail ("(OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) && \"Existing shift must be valid as a rotate half\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5893, __PRETTY_FUNCTION__)) | |||
| 5893 | "Existing shift must be valid as a rotate half")(((OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) && "Existing shift must be valid as a rotate half" ) ? static_cast<void> (0) : __assert_fail ("(OppShift.getOpcode() == ISD::SHL || OppShift.getOpcode() == ISD::SRL) && \"Existing shift must be valid as a rotate half\"" , "/build/llvm-toolchain-snapshot-10~svn374877/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5893, __PRETTY_FUNCTION__)); | |||
| 5894 | ||||
| 5895 | ExtractFrom = stripConstantMask(DAG, ExtractFrom, Mask); | |||
| 5896 | ||||
| 5897 | // Value and Type of the shift. | |||
| 5898 | SDValue OppShiftLHS = OppShift.getOperand(0); | |||
| 5899 | EVT ShiftedVT = OppShiftLHS.getValueType(); | |||
| 5900 | ||||
| 5901 | // Amount of the existing shift. | |||
| 5902 | ConstantSDNode *OppShiftCst = isConstOrConstSplat(OppShift.getOperand(1)); | |||
| 5903 | ||||
| 5904 | // (add v v) -> (shl v 1) | |||
| 5905 | if (OppShift.getOpcode() == ISD::SRL && OppShiftCst && | |||
| 5906 | ExtractFrom.getOpcode() == ISD::ADD && | |||
| 5907 | ExtractFrom.getOperand(0) == ExtractFrom.getOperand(1) && | |||
| 5908 | ExtractFrom.getOperand(0) == OppShiftLHS && | |||
| 5909 | OppShiftCst->getAPIntValue() == ShiftedVT.getScalarSizeInBits() - 1) | |||
| 5910 | return DAG.getNode(ISD::SHL, DL, ShiftedVT, OppShiftLHS, | |||
| 5911 | DAG.getShiftAmountConstant(1, ShiftedVT, DL)); | |||
| 5912 | ||||
| 5913 | // Preconditions: | |||
| 5914 | // (or (op0 v c0) (shiftl/r (op0 v c1) c2)) | |||
| 5915 | // | |||
| 5916 | // Find opcode of the needed shift to be extracted from (op0 v c0). | |||
| 5917 | unsigned Opcode = ISD::DELETED_NODE; | |||
| 5918 | bool IsMulOrDiv = false; | |||
| 5919 | // Set Opcode and IsMulOrDiv if the extract opcode matches the needed shift | |||
| 5920 | // opcode or its arithmetic (mul or udiv) variant. | |||
| 5921 | auto SelectOpcode = [&](unsigned NeededShift, unsigned MulOrDivVariant) { | |||
| 5922 | IsMulOrDiv = ExtractFrom.getOpcode() == MulOrDivVariant; | |||
| 5923 | if (!IsMulOrDiv && ExtractFrom.getOpcode() != NeededShift) | |||
| 5924 | return false; | |||
| 5925 | Opcode = NeededShift; | |||
| 5926 | return true; | |||
| 5927 | }; | |||
| 5928 | // op0 must be either the needed shift opcode or the mul/udiv equivalent | |||
| 5929 | // that the needed shift can be extracted from. | |||
| 5930 | if ((OppShift.getOpcode() != ISD::SRL || !SelectOpcode(ISD::SHL, ISD::MUL)) && | |||
| 5931 | (OppShift.getOpcode() != ISD::SHL || !SelectOpcode(ISD::SRL, ISD::UDIV))) | |||
| 5932 | return SDValue(); | |||
| 5933 | ||||
| 5934 | // op0 must be the same opcode on both sides, have the same LHS argument, | |||
| 5935 | // and produce the same value type. | |||
| 5936 | if (OppShiftLHS.getOpcode() != ExtractFrom.getOpcode() || | |||
| 5937 | OppShiftLHS.getOperand(0) != ExtractFrom.getOperand(0) || | |||
| 5938 | ShiftedVT != ExtractFrom.getValueType()) | |||
| 5939 | return SDValue(); | |||
| 5940 | ||||
| 5941 | // Constant mul/udiv/shift amount from the RHS of the shift's LHS op. | |||
| 5942 | ConstantSDNode *OppLHSCst = isConstOrConstSplat(OppShiftLHS.getOperand(1)); | |||
| 5943 | // Constant mul/udiv/shift amount from the RHS of the ExtractFrom op. | |||
| 5944 | ConstantSDNode *ExtractFromCst = | |||
| 5945 | isConstOrConstSplat(ExtractFrom.getOperand(1)); | |||
| 5946 | // TODO: We should be able to handle non-uniform constant vectors for these values | |||
| 5947 | // Check that we have constant values. | |||
| 5948 | if (!OppShiftCst || !OppShiftCst->getAPIntValue() || | |||
| 5949 | !OppLHSCst || !OppLHSCst->getAPIntValue() || | |||
| 5950 | !ExtractFromCst || !ExtractFromCst->getAPIntValue()) | |||
| 5951 | return SDValue(); | |||
| 5952 | ||||
| 5953 | // Compute the shift amount we need to extract to complete the rotate. | |||
| 5954 | const unsigned VTWidth = ShiftedVT.getScalarSizeInBits(); | |||
| 5955 | if (OppShiftCst->getAPIntValue().ugt(VTWidth)) | |||
| 5956 | return SDValue(); | |||
| 5957 | APInt NeededShiftAmt = VTWidth - OppShiftCst->getAPIntValue(); | |||
| 5958 | // Normalize the bitwidth of the two mul/udiv/shift constant operands. | |||
| 5959 | APInt ExtractFromAmt = ExtractFromCst->getAPIntValue(); | |||
| 5960 | APInt OppLHSAmt = OppLHSCst->getAPIntValue(); | |||
| 5961 | zeroExtendToMatch(ExtractFromAmt, OppLHSAmt); | |||
| 5962 | ||||
| 5963 | // Now try extract the needed shift from the ExtractFrom op and see if the | |||
| 5964 | // result matches up with the existing shift's LHS op. | |||
| 5965 | if (IsMulOrDiv) { | |||
| 5966 | // Op to extract from is a mul or udiv by a constant. | |||
| 5967 | // Check: | |||
| 5968 | // c2 / (1 << (bitwidth(op0 v c0) - c1)) == c0 | |||
| 5969 | // c2 % (1 << (bitwidth(op0 v c0) - c1)) == 0 | |||
| 5970 | const APInt ExtractDiv = APInt::getOneBitSet(ExtractFromAmt.getBitWidth(), | |||
| 5971 | NeededShiftAmt.getZExtValue()); | |||
| 5972 | APInt ResultAmt; | |||
| 5973 | APInt Rem; | |||
| 5974 | APInt::udivrem(ExtractFromAmt, ExtractDiv, ResultAmt, Rem); | |||
| 5975 | if (Rem != 0 || ResultAmt != OppLHSAmt) | |||
| 5976 | return SDValue(); | |||
| 5977 | } else { | |||
| 5978 | // Op to extract from is a shift by a constant. | |||
| 5979 | // Check: | |||
| 5980 | // c2 - (bitwidth(op0 v c0) - c1) == c0 | |||
| 5981 | if (OppLHSAmt != ExtractFromAmt - NeededShiftAmt.zextOrTrunc( | |||
| 5982 | ExtractFromAmt.getBitWidth())) | |||
| 5983 | return SDValue(); | |||
| 5984 | } | |||
| 5985 | ||||
| 5986 | // Return the expanded shift op that should allow a rotate to be formed. | |||
| 5987 | EVT ShiftVT = OppShift.getOperand(1).getValueType(); | |||
| 5988 | EVT ResVT = ExtractFrom.getValueType(); | |||
| 5989 | SDValue NewShiftNode = DAG.getConstant(NeededShiftAmt, DL, ShiftVT); | |||
| 5990 | return DAG.getNode(Opcode, DL, ResVT, OppShiftLHS, NewShiftNode); | |||
| 5991 | } | |||
| 5992 | ||||
| 5993 | // Return true if we can prove that, whenever Neg and Pos are both in the | |||
| 5994 | // range [0, EltSize), Neg == (Pos == 0 ? 0 : EltSize - Pos). This means that | |||
| 5995 | // for two opposing shifts shift1 and shift2 and a value X with OpBits bits: | |||
| 5996 | // | |||
| 5997 | // (or (shift1 X, Neg), (shift2 X, Pos)) | |||
| 5998 | // | |||
| 5999 | // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate | |||
| 6000 | // in direction shift1 by Neg. The range [0, EltSize) means that we only need | |||
| 6001 | // to consider shift amounts with defined behavior. | |||
| 6002 | static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned EltSize, | |||
| 6003 | SelectionDAG &DAG) { | |||
| 6004 | // If EltSize is a power of 2 then: | |||
| 6005 | // | |||
| 6006 | // (a) (Pos == 0 ? 0 : EltSize - Pos) == (EltSize - Pos) & (EltSize - 1) | |||
| 6007 | // (b) Neg == Neg & (EltSize - 1) whenever Neg is in [0, EltSize). | |||
| 6008 | // | |||
| 6009 | // So if EltSize is a power of 2 and Neg is (and Neg', EltSize-1), we check | |||
| 6010 | // for the stronger condition: | |||
| 6011 | // | |||
| 6012 | // Neg & (EltSize - 1) == (EltSize - Pos) & (EltSize - 1) [A] | |||
| 6013 | // | |||
| 6014 | // for all Neg and Pos. Since Neg & (EltSize - 1) == Neg' & (EltSize - 1) | |||
| 6015 |