| 1 | //===- HexagonCommonGEP.cpp -----------------------------------------------===// |
| 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 | #include "Hexagon.h" |
| 10 | |
| 11 | #include "llvm/ADT/ArrayRef.h" |
| 12 | #include "llvm/ADT/FoldingSet.h" |
| 13 | #include "llvm/ADT/GraphTraits.h" |
| 14 | #include "llvm/ADT/STLExtras.h" |
| 15 | #include "llvm/ADT/SetVector.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/ADT/StringRef.h" |
| 18 | #include "llvm/Analysis/LoopInfo.h" |
| 19 | #include "llvm/Analysis/PostDominators.h" |
| 20 | #include "llvm/IR/BasicBlock.h" |
| 21 | #include "llvm/IR/Constant.h" |
| 22 | #include "llvm/IR/Constants.h" |
| 23 | #include "llvm/IR/DerivedTypes.h" |
| 24 | #include "llvm/IR/Dominators.h" |
| 25 | #include "llvm/IR/Function.h" |
| 26 | #include "llvm/IR/Instruction.h" |
| 27 | #include "llvm/IR/Instructions.h" |
| 28 | #include "llvm/IR/Type.h" |
| 29 | #include "llvm/IR/Use.h" |
| 30 | #include "llvm/IR/User.h" |
| 31 | #include "llvm/IR/Value.h" |
| 32 | #include "llvm/IR/Verifier.h" |
| 33 | #include "llvm/InitializePasses.h" |
| 34 | #include "llvm/Pass.h" |
| 35 | #include "llvm/Support/Allocator.h" |
| 36 | #include "llvm/Support/Casting.h" |
| 37 | #include "llvm/Support/CommandLine.h" |
| 38 | #include "llvm/Support/Compiler.h" |
| 39 | #include "llvm/Support/Debug.h" |
| 40 | #include "llvm/Support/raw_ostream.h" |
| 41 | #include "llvm/Transforms/Utils/Local.h" |
| 42 | #include <cassert> |
| 43 | #include <cstddef> |
| 44 | #include <cstdint> |
| 45 | #include <iterator> |
| 46 | #include <map> |
| 47 | #include <set> |
| 48 | #include <utility> |
| 49 | #include <vector> |
| 50 | |
| 51 | #define DEBUG_TYPE "commgep" |
| 52 | |
| 53 | using namespace llvm; |
| 54 | |
| 55 | static cl::opt<bool> OptSpeculate("commgep-speculate" , cl::init(Val: true), |
| 56 | cl::Hidden); |
| 57 | |
| 58 | static cl::opt<bool> OptEnableInv("commgep-inv" , cl::init(Val: true), cl::Hidden); |
| 59 | |
| 60 | static cl::opt<bool> OptEnableConst("commgep-const" , cl::init(Val: true), |
| 61 | cl::Hidden); |
| 62 | |
| 63 | namespace { |
| 64 | |
| 65 | struct GepNode; |
| 66 | using NodeSet = std::set<GepNode *>; |
| 67 | using NodeToValueMap = std::map<GepNode *, Value *>; |
| 68 | using NodeVect = std::vector<GepNode *>; |
| 69 | using NodeChildrenMap = std::map<GepNode *, NodeVect>; |
| 70 | using UseSet = SetVector<Use *>; |
| 71 | using NodeToUsesMap = std::map<GepNode *, UseSet>; |
| 72 | |
| 73 | // Numbering map for gep nodes. Used to keep track of ordering for |
| 74 | // gep nodes. |
| 75 | struct NodeOrdering { |
| 76 | NodeOrdering() = default; |
| 77 | |
| 78 | void insert(const GepNode *N) { Map.insert(x: std::make_pair(x&: N, y&: ++LastNum)); } |
| 79 | void clear() { Map.clear(); } |
| 80 | |
| 81 | bool operator()(const GepNode *N1, const GepNode *N2) const { |
| 82 | auto F1 = Map.find(x: N1), F2 = Map.find(x: N2); |
| 83 | assert(F1 != Map.end() && F2 != Map.end()); |
| 84 | return F1->second < F2->second; |
| 85 | } |
| 86 | |
| 87 | private: |
| 88 | std::map<const GepNode *, unsigned> Map; |
| 89 | unsigned LastNum = 0; |
| 90 | }; |
| 91 | |
| 92 | class HexagonCommonGEP : public FunctionPass { |
| 93 | public: |
| 94 | static char ID; |
| 95 | |
| 96 | HexagonCommonGEP() : FunctionPass(ID) {} |
| 97 | |
| 98 | bool runOnFunction(Function &F) override; |
| 99 | StringRef getPassName() const override { return "Hexagon Common GEP" ; } |
| 100 | |
| 101 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 102 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 103 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 104 | AU.addRequired<PostDominatorTreeWrapperPass>(); |
| 105 | AU.addPreserved<PostDominatorTreeWrapperPass>(); |
| 106 | AU.addRequired<LoopInfoWrapperPass>(); |
| 107 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 108 | FunctionPass::getAnalysisUsage(AU); |
| 109 | } |
| 110 | |
| 111 | private: |
| 112 | using ValueToNodeMap = std::map<Value *, GepNode *>; |
| 113 | using ValueVect = std::vector<Value *>; |
| 114 | using NodeToValuesMap = std::map<GepNode *, ValueVect>; |
| 115 | |
| 116 | void getBlockTraversalOrder(BasicBlock *Root, ValueVect &Order); |
| 117 | bool isHandledGepForm(GetElementPtrInst *GepI); |
| 118 | void processGepInst(GetElementPtrInst *GepI, ValueToNodeMap &NM); |
| 119 | void collect(); |
| 120 | void common(); |
| 121 | |
| 122 | BasicBlock *recalculatePlacement(GepNode *Node, NodeChildrenMap &NCM, |
| 123 | NodeToValueMap &Loc); |
| 124 | BasicBlock *recalculatePlacementRec(GepNode *Node, NodeChildrenMap &NCM, |
| 125 | NodeToValueMap &Loc); |
| 126 | bool isInvariantIn(Value *Val, Loop *L); |
| 127 | bool isInvariantIn(GepNode *Node, Loop *L); |
| 128 | bool isInMainPath(BasicBlock *B, Loop *L); |
| 129 | BasicBlock *adjustForInvariance(GepNode *Node, NodeChildrenMap &NCM, |
| 130 | NodeToValueMap &Loc); |
| 131 | void separateChainForNode(GepNode *Node, Use *U, NodeToValueMap &Loc); |
| 132 | void separateConstantChains(GepNode *Node, NodeChildrenMap &NCM, |
| 133 | NodeToValueMap &Loc); |
| 134 | void computeNodePlacement(NodeToValueMap &Loc); |
| 135 | |
| 136 | Value *fabricateGEP(NodeVect &NA, BasicBlock::iterator At, |
| 137 | BasicBlock *LocB); |
| 138 | void getAllUsersForNode(GepNode *Node, ValueVect &Values, |
| 139 | NodeChildrenMap &NCM); |
| 140 | void materialize(NodeToValueMap &Loc); |
| 141 | |
| 142 | void removeDeadCode(); |
| 143 | |
| 144 | NodeVect Nodes; |
| 145 | NodeToUsesMap Uses; |
| 146 | NodeOrdering NodeOrder; // Node ordering, for deterministic behavior. |
| 147 | SpecificBumpPtrAllocator<GepNode> *Mem; |
| 148 | LLVMContext *Ctx; |
| 149 | LoopInfo *LI; |
| 150 | DominatorTree *DT; |
| 151 | PostDominatorTree *PDT; |
| 152 | Function *Fn; |
| 153 | }; |
| 154 | |
| 155 | } // end anonymous namespace |
| 156 | |
| 157 | char HexagonCommonGEP::ID = 0; |
| 158 | |
| 159 | INITIALIZE_PASS_BEGIN(HexagonCommonGEP, "hcommgep" , "Hexagon Common GEP" , |
| 160 | false, false) |
| 161 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 162 | INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass) |
| 163 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) |
| 164 | INITIALIZE_PASS_END(HexagonCommonGEP, "hcommgep" , "Hexagon Common GEP" , |
| 165 | false, false) |
| 166 | |
| 167 | namespace { |
| 168 | |
| 169 | struct GepNode { |
| 170 | enum { |
| 171 | None = 0, |
| 172 | Root = 0x01, |
| 173 | Internal = 0x02, |
| 174 | Used = 0x04, |
| 175 | InBounds = 0x08, |
| 176 | Pointer = 0x10, // See note below. |
| 177 | }; |
| 178 | // Note: GEP indices generally traverse nested types, and so a GepNode |
| 179 | // (representing a single index) can be associated with some composite |
| 180 | // type. The exception is the GEP input, which is a pointer, and not |
| 181 | // a composite type (at least not in the sense of having sub-types). |
| 182 | // Also, the corresponding index plays a different role as well: it is |
| 183 | // simply added to the input pointer. Since pointer types are becoming |
| 184 | // opaque (i.e. are no longer going to include the pointee type), the |
| 185 | // two pieces of information (1) the fact that it's a pointer, and |
| 186 | // (2) the pointee type, need to be stored separately. The pointee type |
| 187 | // will be stored in the PTy member, while the fact that the node |
| 188 | // operates on a pointer will be reflected by the flag "Pointer". |
| 189 | |
| 190 | uint32_t Flags = 0; |
| 191 | union { |
| 192 | GepNode *Parent; |
| 193 | Value *BaseVal; |
| 194 | }; |
| 195 | Value *Idx = nullptr; |
| 196 | Type *PTy = nullptr; // Type indexed by this node. For pointer nodes |
| 197 | // this is the "pointee" type, and indexing a |
| 198 | // pointer does not change the type. |
| 199 | |
| 200 | GepNode() : Parent(nullptr) {} |
| 201 | GepNode(const GepNode *N) : Flags(N->Flags), Idx(N->Idx), PTy(N->PTy) { |
| 202 | if (Flags & Root) |
| 203 | BaseVal = N->BaseVal; |
| 204 | else |
| 205 | Parent = N->Parent; |
| 206 | } |
| 207 | |
| 208 | friend raw_ostream &operator<< (raw_ostream &OS, const GepNode &GN); |
| 209 | }; |
| 210 | |
| 211 | raw_ostream &operator<< (raw_ostream &OS, const GepNode &GN) { |
| 212 | OS << "{ {" ; |
| 213 | bool Comma = false; |
| 214 | if (GN.Flags & GepNode::Root) { |
| 215 | OS << "root" ; |
| 216 | Comma = true; |
| 217 | } |
| 218 | if (GN.Flags & GepNode::Internal) { |
| 219 | if (Comma) |
| 220 | OS << ','; |
| 221 | OS << "internal" ; |
| 222 | Comma = true; |
| 223 | } |
| 224 | if (GN.Flags & GepNode::Used) { |
| 225 | if (Comma) |
| 226 | OS << ','; |
| 227 | OS << "used" ; |
| 228 | } |
| 229 | if (GN.Flags & GepNode::InBounds) { |
| 230 | if (Comma) |
| 231 | OS << ','; |
| 232 | OS << "inbounds" ; |
| 233 | } |
| 234 | if (GN.Flags & GepNode::Pointer) { |
| 235 | if (Comma) |
| 236 | OS << ','; |
| 237 | OS << "pointer" ; |
| 238 | } |
| 239 | OS << "} " ; |
| 240 | if (GN.Flags & GepNode::Root) |
| 241 | OS << "BaseVal:" << GN.BaseVal->getName() << '(' << GN.BaseVal << ')'; |
| 242 | else |
| 243 | OS << "Parent:" << GN.Parent; |
| 244 | |
| 245 | OS << " Idx:" ; |
| 246 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: GN.Idx)) |
| 247 | OS << CI->getValue().getSExtValue(); |
| 248 | else if (GN.Idx->hasName()) |
| 249 | OS << GN.Idx->getName(); |
| 250 | else |
| 251 | OS << "<anon> =" << *GN.Idx; |
| 252 | |
| 253 | OS << " PTy:" ; |
| 254 | if (GN.PTy->isStructTy()) { |
| 255 | StructType *STy = cast<StructType>(Val: GN.PTy); |
| 256 | if (!STy->isLiteral()) |
| 257 | OS << GN.PTy->getStructName(); |
| 258 | else |
| 259 | OS << "<anon-struct>:" << *STy; |
| 260 | } |
| 261 | else |
| 262 | OS << *GN.PTy; |
| 263 | OS << " }" ; |
| 264 | return OS; |
| 265 | } |
| 266 | |
| 267 | template <typename NodeContainer> |
| 268 | void dump_node_container(raw_ostream &OS, const NodeContainer &S) { |
| 269 | using const_iterator = typename NodeContainer::const_iterator; |
| 270 | |
| 271 | for (const_iterator I = S.begin(), E = S.end(); I != E; ++I) |
| 272 | OS << *I << ' ' << **I << '\n'; |
| 273 | } |
| 274 | |
| 275 | raw_ostream &operator<< (raw_ostream &OS, |
| 276 | const NodeVect &S) LLVM_ATTRIBUTE_UNUSED; |
| 277 | raw_ostream &operator<< (raw_ostream &OS, const NodeVect &S) { |
| 278 | dump_node_container(OS, S); |
| 279 | return OS; |
| 280 | } |
| 281 | |
| 282 | raw_ostream &operator<< (raw_ostream &OS, |
| 283 | const NodeToUsesMap &M) LLVM_ATTRIBUTE_UNUSED; |
| 284 | raw_ostream &operator<< (raw_ostream &OS, const NodeToUsesMap &M){ |
| 285 | for (const auto &I : M) { |
| 286 | const UseSet &Us = I.second; |
| 287 | OS << I.first << " -> #" << Us.size() << '{'; |
| 288 | for (const Use *U : Us) { |
| 289 | User *R = U->getUser(); |
| 290 | if (R->hasName()) |
| 291 | OS << ' ' << R->getName(); |
| 292 | else |
| 293 | OS << " <?>(" << *R << ')'; |
| 294 | } |
| 295 | OS << " }\n" ; |
| 296 | } |
| 297 | return OS; |
| 298 | } |
| 299 | |
| 300 | struct in_set { |
| 301 | in_set(const NodeSet &S) : NS(S) {} |
| 302 | |
| 303 | bool operator() (GepNode *N) const { |
| 304 | return NS.find(x: N) != NS.end(); |
| 305 | } |
| 306 | |
| 307 | private: |
| 308 | const NodeSet &NS; |
| 309 | }; |
| 310 | |
| 311 | } // end anonymous namespace |
| 312 | |
| 313 | inline void *operator new(size_t, SpecificBumpPtrAllocator<GepNode> &A) { |
| 314 | return A.Allocate(); |
| 315 | } |
| 316 | |
| 317 | void HexagonCommonGEP::getBlockTraversalOrder(BasicBlock *Root, |
| 318 | ValueVect &Order) { |
| 319 | // Compute block ordering for a typical DT-based traversal of the flow |
| 320 | // graph: "before visiting a block, all of its dominators must have been |
| 321 | // visited". |
| 322 | |
| 323 | Order.push_back(x: Root); |
| 324 | for (auto *DTN : children<DomTreeNode*>(G: DT->getNode(BB: Root))) |
| 325 | getBlockTraversalOrder(Root: DTN->getBlock(), Order); |
| 326 | } |
| 327 | |
| 328 | bool HexagonCommonGEP::isHandledGepForm(GetElementPtrInst *GepI) { |
| 329 | // No vector GEPs. |
| 330 | if (!GepI->getType()->isPointerTy()) |
| 331 | return false; |
| 332 | // No GEPs without any indices. (Is this possible?) |
| 333 | if (GepI->idx_begin() == GepI->idx_end()) |
| 334 | return false; |
| 335 | return true; |
| 336 | } |
| 337 | |
| 338 | void HexagonCommonGEP::processGepInst(GetElementPtrInst *GepI, |
| 339 | ValueToNodeMap &NM) { |
| 340 | LLVM_DEBUG(dbgs() << "Visiting GEP: " << *GepI << '\n'); |
| 341 | GepNode *N = new (*Mem) GepNode; |
| 342 | Value *PtrOp = GepI->getPointerOperand(); |
| 343 | uint32_t InBounds = GepI->isInBounds() ? GepNode::InBounds : 0; |
| 344 | ValueToNodeMap::iterator F = NM.find(x: PtrOp); |
| 345 | if (F == NM.end()) { |
| 346 | N->BaseVal = PtrOp; |
| 347 | N->Flags |= GepNode::Root | InBounds; |
| 348 | } else { |
| 349 | // If PtrOp was a GEP instruction, it must have already been processed. |
| 350 | // The ValueToNodeMap entry for it is the last gep node in the generated |
| 351 | // chain. Link to it here. |
| 352 | N->Parent = F->second; |
| 353 | } |
| 354 | N->PTy = GepI->getSourceElementType(); |
| 355 | N->Flags |= GepNode::Pointer; |
| 356 | N->Idx = *GepI->idx_begin(); |
| 357 | |
| 358 | // Collect the list of users of this GEP instruction. Will add it to the |
| 359 | // last node created for it. |
| 360 | UseSet Us; |
| 361 | for (Value::user_iterator UI = GepI->user_begin(), UE = GepI->user_end(); |
| 362 | UI != UE; ++UI) { |
| 363 | // Check if this gep is used by anything other than other geps that |
| 364 | // we will process. |
| 365 | if (isa<GetElementPtrInst>(Val: *UI)) { |
| 366 | GetElementPtrInst *UserG = cast<GetElementPtrInst>(Val: *UI); |
| 367 | if (isHandledGepForm(GepI: UserG)) |
| 368 | continue; |
| 369 | } |
| 370 | Us.insert(X: &UI.getUse()); |
| 371 | } |
| 372 | Nodes.push_back(x: N); |
| 373 | NodeOrder.insert(N); |
| 374 | |
| 375 | // Skip the first index operand, since it was already handled above. This |
| 376 | // dereferences the pointer operand. |
| 377 | GepNode *PN = N; |
| 378 | Type *PtrTy = GepI->getSourceElementType(); |
| 379 | for (Use &U : llvm::drop_begin(RangeOrContainer: GepI->indices())) { |
| 380 | Value *Op = U; |
| 381 | GepNode *Nx = new (*Mem) GepNode; |
| 382 | Nx->Parent = PN; // Link Nx to the previous node. |
| 383 | Nx->Flags |= GepNode::Internal | InBounds; |
| 384 | Nx->PTy = PtrTy; |
| 385 | Nx->Idx = Op; |
| 386 | Nodes.push_back(x: Nx); |
| 387 | NodeOrder.insert(N: Nx); |
| 388 | PN = Nx; |
| 389 | |
| 390 | PtrTy = GetElementPtrInst::getTypeAtIndex(Ty: PtrTy, Idx: Op); |
| 391 | } |
| 392 | |
| 393 | // After last node has been created, update the use information. |
| 394 | if (!Us.empty()) { |
| 395 | PN->Flags |= GepNode::Used; |
| 396 | Uses[PN].insert_range(R&: Us); |
| 397 | } |
| 398 | |
| 399 | // Link the last node with the originating GEP instruction. This is to |
| 400 | // help with linking chained GEP instructions. |
| 401 | NM.insert(x: std::make_pair(x&: GepI, y&: PN)); |
| 402 | } |
| 403 | |
| 404 | void HexagonCommonGEP::collect() { |
| 405 | // Establish depth-first traversal order of the dominator tree. |
| 406 | ValueVect BO; |
| 407 | getBlockTraversalOrder(Root: &Fn->front(), Order&: BO); |
| 408 | |
| 409 | // The creation of gep nodes requires DT-traversal. When processing a GEP |
| 410 | // instruction that uses another GEP instruction as the base pointer, the |
| 411 | // gep node for the base pointer should already exist. |
| 412 | ValueToNodeMap NM; |
| 413 | for (Value *I : BO) { |
| 414 | BasicBlock *B = cast<BasicBlock>(Val: I); |
| 415 | for (Instruction &J : *B) |
| 416 | if (auto *GepI = dyn_cast<GetElementPtrInst>(Val: &J)) |
| 417 | if (isHandledGepForm(GepI)) |
| 418 | processGepInst(GepI, NM); |
| 419 | } |
| 420 | |
| 421 | LLVM_DEBUG(dbgs() << "Gep nodes after initial collection:\n" << Nodes); |
| 422 | } |
| 423 | |
| 424 | static void invert_find_roots(const NodeVect &Nodes, NodeChildrenMap &NCM, |
| 425 | NodeVect &Roots) { |
| 426 | for (GepNode *N : Nodes) { |
| 427 | if (N->Flags & GepNode::Root) { |
| 428 | Roots.push_back(x: N); |
| 429 | continue; |
| 430 | } |
| 431 | GepNode *PN = N->Parent; |
| 432 | NCM[PN].push_back(x: N); |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | static void nodes_for_root(GepNode *Root, NodeChildrenMap &NCM, |
| 437 | NodeSet &Nodes) { |
| 438 | NodeVect Work; |
| 439 | Work.push_back(x: Root); |
| 440 | Nodes.insert(x: Root); |
| 441 | |
| 442 | while (!Work.empty()) { |
| 443 | NodeVect::iterator First = Work.begin(); |
| 444 | GepNode *N = *First; |
| 445 | Work.erase(position: First); |
| 446 | NodeChildrenMap::iterator CF = NCM.find(x: N); |
| 447 | if (CF != NCM.end()) { |
| 448 | llvm::append_range(C&: Work, R&: CF->second); |
| 449 | Nodes.insert(first: CF->second.begin(), last: CF->second.end()); |
| 450 | } |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | namespace { |
| 455 | |
| 456 | using NodeSymRel = std::set<NodeSet>; |
| 457 | using NodePair = std::pair<GepNode *, GepNode *>; |
| 458 | using NodePairSet = std::set<NodePair>; |
| 459 | |
| 460 | } // end anonymous namespace |
| 461 | |
| 462 | static const NodeSet *node_class(GepNode *N, NodeSymRel &Rel) { |
| 463 | for (const NodeSet &S : Rel) |
| 464 | if (S.count(x: N)) |
| 465 | return &S; |
| 466 | return nullptr; |
| 467 | } |
| 468 | |
| 469 | // Create an ordered pair of GepNode pointers. The pair will be used in |
| 470 | // determining equality. The only purpose of the ordering is to eliminate |
| 471 | // duplication due to the commutativity of equality/non-equality. |
| 472 | static NodePair node_pair(GepNode *N1, GepNode *N2) { |
| 473 | uintptr_t P1 = reinterpret_cast<uintptr_t>(N1); |
| 474 | uintptr_t P2 = reinterpret_cast<uintptr_t>(N2); |
| 475 | if (P1 <= P2) |
| 476 | return std::make_pair(x&: N1, y&: N2); |
| 477 | return std::make_pair(x&: N2, y&: N1); |
| 478 | } |
| 479 | |
| 480 | static unsigned node_hash(GepNode *N) { |
| 481 | // Include everything except flags and parent. |
| 482 | FoldingSetNodeID ID; |
| 483 | ID.AddPointer(Ptr: N->Idx); |
| 484 | ID.AddPointer(Ptr: N->PTy); |
| 485 | return ID.ComputeHash(); |
| 486 | } |
| 487 | |
| 488 | static bool node_eq(GepNode *N1, GepNode *N2, NodePairSet &Eq, |
| 489 | NodePairSet &Ne) { |
| 490 | // Don't cache the result for nodes with different hashes. The hash |
| 491 | // comparison is fast enough. |
| 492 | if (node_hash(N: N1) != node_hash(N: N2)) |
| 493 | return false; |
| 494 | |
| 495 | NodePair NP = node_pair(N1, N2); |
| 496 | NodePairSet::iterator FEq = Eq.find(x: NP); |
| 497 | if (FEq != Eq.end()) |
| 498 | return true; |
| 499 | NodePairSet::iterator FNe = Ne.find(x: NP); |
| 500 | if (FNe != Ne.end()) |
| 501 | return false; |
| 502 | // Not previously compared. |
| 503 | bool Root1 = N1->Flags & GepNode::Root; |
| 504 | uint32_t CmpFlags = GepNode::Root | GepNode::Pointer; |
| 505 | bool Different = (N1->Flags & CmpFlags) != (N2->Flags & CmpFlags); |
| 506 | NodePair P = node_pair(N1, N2); |
| 507 | // If the root/pointer flags have different values, the nodes are |
| 508 | // different. |
| 509 | // If both nodes are root nodes, but their base pointers differ, |
| 510 | // they are different. |
| 511 | if (Different || (Root1 && N1->BaseVal != N2->BaseVal)) { |
| 512 | Ne.insert(x: P); |
| 513 | return false; |
| 514 | } |
| 515 | // Here the root/pointer flags are identical, and for root nodes the |
| 516 | // base pointers are equal, so the root nodes are equal. |
| 517 | // For non-root nodes, compare their parent nodes. |
| 518 | if (Root1 || node_eq(N1: N1->Parent, N2: N2->Parent, Eq, Ne)) { |
| 519 | Eq.insert(x: P); |
| 520 | return true; |
| 521 | } |
| 522 | return false; |
| 523 | } |
| 524 | |
| 525 | void HexagonCommonGEP::common() { |
| 526 | // The essence of this commoning is finding gep nodes that are equal. |
| 527 | // To do this we need to compare all pairs of nodes. To save time, |
| 528 | // first, partition the set of all nodes into sets of potentially equal |
| 529 | // nodes, and then compare pairs from within each partition. |
| 530 | using NodeSetMap = std::map<unsigned, NodeSet>; |
| 531 | NodeSetMap MaybeEq; |
| 532 | |
| 533 | for (GepNode *N : Nodes) { |
| 534 | unsigned H = node_hash(N); |
| 535 | MaybeEq[H].insert(x: N); |
| 536 | } |
| 537 | |
| 538 | // Compute the equivalence relation for the gep nodes. Use two caches, |
| 539 | // one for equality and the other for non-equality. |
| 540 | NodeSymRel EqRel; // Equality relation (as set of equivalence classes). |
| 541 | NodePairSet Eq, Ne; // Caches. |
| 542 | for (auto &I : MaybeEq) { |
| 543 | NodeSet &S = I.second; |
| 544 | for (NodeSet::iterator NI = S.begin(), NE = S.end(); NI != NE; ++NI) { |
| 545 | GepNode *N = *NI; |
| 546 | // If node already has a class, then the class must have been created |
| 547 | // in a prior iteration of this loop. Since equality is transitive, |
| 548 | // nothing more will be added to that class, so skip it. |
| 549 | if (node_class(N, Rel&: EqRel)) |
| 550 | continue; |
| 551 | |
| 552 | // Create a new class candidate now. |
| 553 | NodeSet C; |
| 554 | for (NodeSet::iterator NJ = std::next(x: NI); NJ != NE; ++NJ) |
| 555 | if (node_eq(N1: N, N2: *NJ, Eq, Ne)) |
| 556 | C.insert(x: *NJ); |
| 557 | // If Tmp is empty, N would be the only element in it. Don't bother |
| 558 | // creating a class for it then. |
| 559 | if (!C.empty()) { |
| 560 | C.insert(x: N); // Finalize the set before adding it to the relation. |
| 561 | std::pair<NodeSymRel::iterator, bool> Ins = EqRel.insert(x: C); |
| 562 | (void)Ins; |
| 563 | assert(Ins.second && "Cannot add a class" ); |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | |
| 568 | LLVM_DEBUG({ |
| 569 | dbgs() << "Gep node equality:\n" ; |
| 570 | for (NodePairSet::iterator I = Eq.begin(), E = Eq.end(); I != E; ++I) |
| 571 | dbgs() << "{ " << I->first << ", " << I->second << " }\n" ; |
| 572 | |
| 573 | dbgs() << "Gep equivalence classes:\n" ; |
| 574 | for (const NodeSet &S : EqRel) { |
| 575 | dbgs() << '{'; |
| 576 | for (NodeSet::const_iterator J = S.begin(), F = S.end(); J != F; ++J) { |
| 577 | if (J != S.begin()) |
| 578 | dbgs() << ','; |
| 579 | dbgs() << ' ' << *J; |
| 580 | } |
| 581 | dbgs() << " }\n" ; |
| 582 | } |
| 583 | }); |
| 584 | |
| 585 | // Create a projection from a NodeSet to the minimal element in it. |
| 586 | using ProjMap = std::map<const NodeSet *, GepNode *>; |
| 587 | ProjMap PM; |
| 588 | for (const NodeSet &S : EqRel) { |
| 589 | GepNode *Min = *llvm::min_element(Range: S, C: NodeOrder); |
| 590 | std::pair<ProjMap::iterator,bool> Ins = PM.insert(x: std::make_pair(x: &S, y&: Min)); |
| 591 | (void)Ins; |
| 592 | assert(Ins.second && "Cannot add minimal element" ); |
| 593 | |
| 594 | // Update the min element's flags, and user list. |
| 595 | uint32_t Flags = 0; |
| 596 | UseSet &MinUs = Uses[Min]; |
| 597 | for (GepNode *N : S) { |
| 598 | uint32_t NF = N->Flags; |
| 599 | // If N is used, append all original values of N to the list of |
| 600 | // original values of Min. |
| 601 | if (NF & GepNode::Used) { |
| 602 | auto &U = Uses[N]; |
| 603 | MinUs.insert_range(R&: U); |
| 604 | } |
| 605 | Flags |= NF; |
| 606 | } |
| 607 | if (MinUs.empty()) |
| 608 | Uses.erase(x: Min); |
| 609 | |
| 610 | // The collected flags should include all the flags from the min element. |
| 611 | assert((Min->Flags & Flags) == Min->Flags); |
| 612 | Min->Flags = Flags; |
| 613 | } |
| 614 | |
| 615 | // Commoning: for each non-root gep node, replace "Parent" with the |
| 616 | // selected (minimum) node from the corresponding equivalence class. |
| 617 | // If a given parent does not have an equivalence class, leave it |
| 618 | // unchanged (it means that it's the only element in its class). |
| 619 | for (GepNode *N : Nodes) { |
| 620 | if (N->Flags & GepNode::Root) |
| 621 | continue; |
| 622 | const NodeSet *PC = node_class(N: N->Parent, Rel&: EqRel); |
| 623 | if (!PC) |
| 624 | continue; |
| 625 | ProjMap::iterator F = PM.find(x: PC); |
| 626 | if (F == PM.end()) |
| 627 | continue; |
| 628 | // Found a replacement, use it. |
| 629 | GepNode *Rep = F->second; |
| 630 | N->Parent = Rep; |
| 631 | } |
| 632 | |
| 633 | LLVM_DEBUG(dbgs() << "Gep nodes after commoning:\n" << Nodes); |
| 634 | |
| 635 | // Finally, erase the nodes that are no longer used. |
| 636 | NodeSet Erase; |
| 637 | for (GepNode *N : Nodes) { |
| 638 | const NodeSet *PC = node_class(N, Rel&: EqRel); |
| 639 | if (!PC) |
| 640 | continue; |
| 641 | ProjMap::iterator F = PM.find(x: PC); |
| 642 | if (F == PM.end()) |
| 643 | continue; |
| 644 | if (N == F->second) |
| 645 | continue; |
| 646 | // Node for removal. |
| 647 | Erase.insert(x: N); |
| 648 | } |
| 649 | erase_if(C&: Nodes, P: in_set(Erase)); |
| 650 | |
| 651 | LLVM_DEBUG(dbgs() << "Gep nodes after post-commoning cleanup:\n" << Nodes); |
| 652 | } |
| 653 | |
| 654 | template <typename T> |
| 655 | static BasicBlock *nearest_common_dominator(DominatorTree *DT, T &Blocks) { |
| 656 | LLVM_DEBUG({ |
| 657 | dbgs() << "NCD of {" ; |
| 658 | for (typename T::iterator I = Blocks.begin(), E = Blocks.end(); I != E; |
| 659 | ++I) { |
| 660 | if (!*I) |
| 661 | continue; |
| 662 | BasicBlock *B = cast<BasicBlock>(*I); |
| 663 | dbgs() << ' ' << B->getName(); |
| 664 | } |
| 665 | dbgs() << " }\n" ; |
| 666 | }); |
| 667 | |
| 668 | // Allow null basic blocks in Blocks. In such cases, return nullptr. |
| 669 | typename T::iterator I = Blocks.begin(), E = Blocks.end(); |
| 670 | if (I == E || !*I) |
| 671 | return nullptr; |
| 672 | BasicBlock *Dom = cast<BasicBlock>(*I); |
| 673 | while (++I != E) { |
| 674 | BasicBlock *B = cast_or_null<BasicBlock>(*I); |
| 675 | Dom = B ? DT->findNearestCommonDominator(A: Dom, B) : nullptr; |
| 676 | if (!Dom) |
| 677 | return nullptr; |
| 678 | } |
| 679 | LLVM_DEBUG(dbgs() << "computed:" << Dom->getName() << '\n'); |
| 680 | return Dom; |
| 681 | } |
| 682 | |
| 683 | template <typename T> |
| 684 | static BasicBlock *nearest_common_dominatee(DominatorTree *DT, T &Blocks) { |
| 685 | // If two blocks, A and B, dominate a block C, then A dominates B, |
| 686 | // or B dominates A. |
| 687 | typename T::iterator I = Blocks.begin(), E = Blocks.end(); |
| 688 | // Find the first non-null block. |
| 689 | while (I != E && !*I) |
| 690 | ++I; |
| 691 | if (I == E) |
| 692 | return DT->getRoot(); |
| 693 | BasicBlock *DomB = cast<BasicBlock>(*I); |
| 694 | while (++I != E) { |
| 695 | if (!*I) |
| 696 | continue; |
| 697 | BasicBlock *B = cast<BasicBlock>(*I); |
| 698 | if (DT->dominates(A: B, B: DomB)) |
| 699 | continue; |
| 700 | if (!DT->dominates(A: DomB, B)) |
| 701 | return nullptr; |
| 702 | DomB = B; |
| 703 | } |
| 704 | return DomB; |
| 705 | } |
| 706 | |
| 707 | // Find the first use in B of any value from Values. If no such use, |
| 708 | // return B->end(). |
| 709 | template <typename T> |
| 710 | static BasicBlock::iterator first_use_of_in_block(T &Values, BasicBlock *B) { |
| 711 | BasicBlock::iterator FirstUse = B->end(), BEnd = B->end(); |
| 712 | |
| 713 | using iterator = typename T::iterator; |
| 714 | |
| 715 | for (iterator I = Values.begin(), E = Values.end(); I != E; ++I) { |
| 716 | Value *V = *I; |
| 717 | // If V is used in a PHI node, the use belongs to the incoming block, |
| 718 | // not the block with the PHI node. In the incoming block, the use |
| 719 | // would be considered as being at the end of it, so it cannot |
| 720 | // influence the position of the first use (which is assumed to be |
| 721 | // at the end to start with). |
| 722 | if (isa<PHINode>(Val: V)) |
| 723 | continue; |
| 724 | if (!isa<Instruction>(Val: V)) |
| 725 | continue; |
| 726 | Instruction *In = cast<Instruction>(Val: V); |
| 727 | if (In->getParent() != B) |
| 728 | continue; |
| 729 | BasicBlock::iterator It = In->getIterator(); |
| 730 | if (std::distance(first: FirstUse, last: BEnd) < std::distance(first: It, last: BEnd)) |
| 731 | FirstUse = It; |
| 732 | } |
| 733 | return FirstUse; |
| 734 | } |
| 735 | |
| 736 | static bool is_empty(const BasicBlock *B) { |
| 737 | return B->empty() || (&*B->begin() == B->getTerminator()); |
| 738 | } |
| 739 | |
| 740 | BasicBlock *HexagonCommonGEP::recalculatePlacement(GepNode *Node, |
| 741 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 742 | LLVM_DEBUG(dbgs() << "Loc for node:" << Node << '\n'); |
| 743 | // Recalculate the placement for Node, assuming that the locations of |
| 744 | // its children in Loc are valid. |
| 745 | // Return nullptr if there is no valid placement for Node (for example, it |
| 746 | // uses an index value that is not available at the location required |
| 747 | // to dominate all children, etc.). |
| 748 | |
| 749 | // Find the nearest common dominator for: |
| 750 | // - all users, if the node is used, and |
| 751 | // - all children. |
| 752 | ValueVect Bs; |
| 753 | if (Node->Flags & GepNode::Used) { |
| 754 | // Append all blocks with uses of the original values to the |
| 755 | // block vector Bs. |
| 756 | NodeToUsesMap::iterator UF = Uses.find(x: Node); |
| 757 | assert(UF != Uses.end() && "Used node with no use information" ); |
| 758 | UseSet &Us = UF->second; |
| 759 | for (Use *U : Us) { |
| 760 | User *R = U->getUser(); |
| 761 | if (!isa<Instruction>(Val: R)) |
| 762 | continue; |
| 763 | BasicBlock *PB = isa<PHINode>(Val: R) |
| 764 | ? cast<PHINode>(Val: R)->getIncomingBlock(U: *U) |
| 765 | : cast<Instruction>(Val: R)->getParent(); |
| 766 | Bs.push_back(x: PB); |
| 767 | } |
| 768 | } |
| 769 | // Append the location of each child. |
| 770 | NodeChildrenMap::iterator CF = NCM.find(x: Node); |
| 771 | if (CF != NCM.end()) { |
| 772 | NodeVect &Cs = CF->second; |
| 773 | for (GepNode *CN : Cs) { |
| 774 | NodeToValueMap::iterator LF = Loc.find(x: CN); |
| 775 | // If the child is only used in GEP instructions (i.e. is not used in |
| 776 | // non-GEP instructions), the nearest dominator computed for it may |
| 777 | // have been null. In such case it won't have a location available. |
| 778 | if (LF == Loc.end()) |
| 779 | continue; |
| 780 | Bs.push_back(x: LF->second); |
| 781 | } |
| 782 | } |
| 783 | |
| 784 | BasicBlock *DomB = nearest_common_dominator(DT, Blocks&: Bs); |
| 785 | if (!DomB) |
| 786 | return nullptr; |
| 787 | // Check if the index used by Node dominates the computed dominator. |
| 788 | Instruction *IdxI = dyn_cast<Instruction>(Val: Node->Idx); |
| 789 | if (IdxI && !DT->dominates(A: IdxI->getParent(), B: DomB)) |
| 790 | return nullptr; |
| 791 | |
| 792 | // Avoid putting nodes into empty blocks. |
| 793 | while (is_empty(B: DomB)) { |
| 794 | DomTreeNode *N = (*DT)[DomB]->getIDom(); |
| 795 | if (!N) |
| 796 | break; |
| 797 | DomB = N->getBlock(); |
| 798 | } |
| 799 | |
| 800 | // Otherwise, DomB is fine. Update the location map. |
| 801 | Loc[Node] = DomB; |
| 802 | return DomB; |
| 803 | } |
| 804 | |
| 805 | BasicBlock *HexagonCommonGEP::recalculatePlacementRec(GepNode *Node, |
| 806 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 807 | LLVM_DEBUG(dbgs() << "LocRec begin for node:" << Node << '\n'); |
| 808 | // Recalculate the placement of Node, after recursively recalculating the |
| 809 | // placements of all its children. |
| 810 | NodeChildrenMap::iterator CF = NCM.find(x: Node); |
| 811 | if (CF != NCM.end()) { |
| 812 | NodeVect &Cs = CF->second; |
| 813 | for (GepNode *C : Cs) |
| 814 | recalculatePlacementRec(Node: C, NCM, Loc); |
| 815 | } |
| 816 | BasicBlock *LB = recalculatePlacement(Node, NCM, Loc); |
| 817 | LLVM_DEBUG(dbgs() << "LocRec end for node:" << Node << '\n'); |
| 818 | return LB; |
| 819 | } |
| 820 | |
| 821 | bool HexagonCommonGEP::isInvariantIn(Value *Val, Loop *L) { |
| 822 | if (isa<Constant>(Val) || isa<Argument>(Val)) |
| 823 | return true; |
| 824 | Instruction *In = dyn_cast<Instruction>(Val); |
| 825 | if (!In) |
| 826 | return false; |
| 827 | BasicBlock *HdrB = L->getHeader(), *DefB = In->getParent(); |
| 828 | return DT->properlyDominates(A: DefB, B: HdrB); |
| 829 | } |
| 830 | |
| 831 | bool HexagonCommonGEP::isInvariantIn(GepNode *Node, Loop *L) { |
| 832 | if (Node->Flags & GepNode::Root) |
| 833 | if (!isInvariantIn(Val: Node->BaseVal, L)) |
| 834 | return false; |
| 835 | return isInvariantIn(Val: Node->Idx, L); |
| 836 | } |
| 837 | |
| 838 | bool HexagonCommonGEP::isInMainPath(BasicBlock *B, Loop *L) { |
| 839 | BasicBlock *HB = L->getHeader(); |
| 840 | BasicBlock *LB = L->getLoopLatch(); |
| 841 | // B must post-dominate the loop header or dominate the loop latch. |
| 842 | if (PDT->dominates(A: B, B: HB)) |
| 843 | return true; |
| 844 | if (LB && DT->dominates(A: B, B: LB)) |
| 845 | return true; |
| 846 | return false; |
| 847 | } |
| 848 | |
| 849 | static BasicBlock *(DominatorTree *DT, Loop *L) { |
| 850 | if (BasicBlock *PH = L->getLoopPreheader()) |
| 851 | return PH; |
| 852 | if (!OptSpeculate) |
| 853 | return nullptr; |
| 854 | DomTreeNode *DN = DT->getNode(BB: L->getHeader()); |
| 855 | if (!DN) |
| 856 | return nullptr; |
| 857 | return DN->getIDom()->getBlock(); |
| 858 | } |
| 859 | |
| 860 | BasicBlock *HexagonCommonGEP::adjustForInvariance(GepNode *Node, |
| 861 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 862 | // Find the "topmost" location for Node: it must be dominated by both, |
| 863 | // its parent (or the BaseVal, if it's a root node), and by the index |
| 864 | // value. |
| 865 | ValueVect Bs; |
| 866 | if (Node->Flags & GepNode::Root) { |
| 867 | if (Instruction *PIn = dyn_cast<Instruction>(Val: Node->BaseVal)) |
| 868 | Bs.push_back(x: PIn->getParent()); |
| 869 | } else { |
| 870 | Bs.push_back(x: Loc[Node->Parent]); |
| 871 | } |
| 872 | if (Instruction *IIn = dyn_cast<Instruction>(Val: Node->Idx)) |
| 873 | Bs.push_back(x: IIn->getParent()); |
| 874 | BasicBlock *TopB = nearest_common_dominatee(DT, Blocks&: Bs); |
| 875 | |
| 876 | // Traverse the loop nest upwards until we find a loop in which Node |
| 877 | // is no longer invariant, or until we get to the upper limit of Node's |
| 878 | // placement. The traversal will also stop when a suitable "preheader" |
| 879 | // cannot be found for a given loop. The "preheader" may actually be |
| 880 | // a regular block outside of the loop (i.e. not guarded), in which case |
| 881 | // the Node will be speculated. |
| 882 | // For nodes that are not in the main path of the containing loop (i.e. |
| 883 | // are not executed in each iteration), do not move them out of the loop. |
| 884 | BasicBlock *LocB = cast_or_null<BasicBlock>(Val: Loc[Node]); |
| 885 | if (LocB) { |
| 886 | Loop *Lp = LI->getLoopFor(BB: LocB); |
| 887 | while (Lp) { |
| 888 | if (!isInvariantIn(Node, L: Lp) || !isInMainPath(B: LocB, L: Lp)) |
| 889 | break; |
| 890 | BasicBlock *NewLoc = preheader(DT, L: Lp); |
| 891 | if (!NewLoc || !DT->dominates(A: TopB, B: NewLoc)) |
| 892 | break; |
| 893 | Lp = Lp->getParentLoop(); |
| 894 | LocB = NewLoc; |
| 895 | } |
| 896 | } |
| 897 | Loc[Node] = LocB; |
| 898 | |
| 899 | // Recursively compute the locations of all children nodes. |
| 900 | NodeChildrenMap::iterator CF = NCM.find(x: Node); |
| 901 | if (CF != NCM.end()) { |
| 902 | NodeVect &Cs = CF->second; |
| 903 | for (GepNode *C : Cs) |
| 904 | adjustForInvariance(Node: C, NCM, Loc); |
| 905 | } |
| 906 | return LocB; |
| 907 | } |
| 908 | |
| 909 | namespace { |
| 910 | |
| 911 | struct LocationAsBlock { |
| 912 | LocationAsBlock(const NodeToValueMap &L) : Map(L) {} |
| 913 | |
| 914 | const NodeToValueMap ⤅ |
| 915 | }; |
| 916 | |
| 917 | raw_ostream &operator<< (raw_ostream &OS, |
| 918 | const LocationAsBlock &Loc) LLVM_ATTRIBUTE_UNUSED ; |
| 919 | raw_ostream &operator<< (raw_ostream &OS, const LocationAsBlock &Loc) { |
| 920 | for (const auto &I : Loc.Map) { |
| 921 | OS << I.first << " -> " ; |
| 922 | if (BasicBlock *B = cast_or_null<BasicBlock>(Val: I.second)) |
| 923 | OS << B->getName() << '(' << B << ')'; |
| 924 | else |
| 925 | OS << "<null-block>" ; |
| 926 | OS << '\n'; |
| 927 | } |
| 928 | return OS; |
| 929 | } |
| 930 | |
| 931 | inline bool is_constant(GepNode *N) { |
| 932 | return isa<ConstantInt>(Val: N->Idx); |
| 933 | } |
| 934 | |
| 935 | } // end anonymous namespace |
| 936 | |
| 937 | void HexagonCommonGEP::separateChainForNode(GepNode *Node, Use *U, |
| 938 | NodeToValueMap &Loc) { |
| 939 | User *R = U->getUser(); |
| 940 | LLVM_DEBUG(dbgs() << "Separating chain for node (" << Node << ") user: " << *R |
| 941 | << '\n'); |
| 942 | BasicBlock *PB = cast<Instruction>(Val: R)->getParent(); |
| 943 | |
| 944 | GepNode *N = Node; |
| 945 | GepNode *C = nullptr, *NewNode = nullptr; |
| 946 | while (is_constant(N) && !(N->Flags & GepNode::Root)) { |
| 947 | // XXX if (single-use) dont-replicate; |
| 948 | GepNode *NewN = new (*Mem) GepNode(N); |
| 949 | Nodes.push_back(x: NewN); |
| 950 | Loc[NewN] = PB; |
| 951 | |
| 952 | if (N == Node) |
| 953 | NewNode = NewN; |
| 954 | NewN->Flags &= ~GepNode::Used; |
| 955 | if (C) |
| 956 | C->Parent = NewN; |
| 957 | C = NewN; |
| 958 | N = N->Parent; |
| 959 | } |
| 960 | if (!NewNode) |
| 961 | return; |
| 962 | |
| 963 | // Move over all uses that share the same user as U from Node to NewNode. |
| 964 | NodeToUsesMap::iterator UF = Uses.find(x: Node); |
| 965 | assert(UF != Uses.end()); |
| 966 | UseSet &Us = UF->second; |
| 967 | UseSet NewUs; |
| 968 | for (Use *U : Us) { |
| 969 | if (U->getUser() == R) |
| 970 | NewUs.insert(X: U); |
| 971 | } |
| 972 | for (Use *U : NewUs) |
| 973 | Us.remove(X: U); // erase takes an iterator. |
| 974 | |
| 975 | if (Us.empty()) { |
| 976 | Node->Flags &= ~GepNode::Used; |
| 977 | Uses.erase(position: UF); |
| 978 | } |
| 979 | |
| 980 | // Should at least have U in NewUs. |
| 981 | NewNode->Flags |= GepNode::Used; |
| 982 | LLVM_DEBUG(dbgs() << "new node: " << NewNode << " " << *NewNode << '\n'); |
| 983 | assert(!NewUs.empty()); |
| 984 | Uses[NewNode] = NewUs; |
| 985 | } |
| 986 | |
| 987 | void HexagonCommonGEP::separateConstantChains(GepNode *Node, |
| 988 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 989 | // First approximation: extract all chains. |
| 990 | NodeSet Ns; |
| 991 | nodes_for_root(Root: Node, NCM, Nodes&: Ns); |
| 992 | |
| 993 | LLVM_DEBUG(dbgs() << "Separating constant chains for node: " << Node << '\n'); |
| 994 | // Collect all used nodes together with the uses from loads and stores, |
| 995 | // where the GEP node could be folded into the load/store instruction. |
| 996 | NodeToUsesMap FNs; // Foldable nodes. |
| 997 | for (GepNode *N : Ns) { |
| 998 | if (!(N->Flags & GepNode::Used)) |
| 999 | continue; |
| 1000 | NodeToUsesMap::iterator UF = Uses.find(x: N); |
| 1001 | assert(UF != Uses.end()); |
| 1002 | UseSet &Us = UF->second; |
| 1003 | // Loads/stores that use the node N. |
| 1004 | UseSet LSs; |
| 1005 | for (Use *U : Us) { |
| 1006 | User *R = U->getUser(); |
| 1007 | // We're interested in uses that provide the address. It can happen |
| 1008 | // that the value may also be provided via GEP, but we won't handle |
| 1009 | // those cases here for now. |
| 1010 | if (LoadInst *Ld = dyn_cast<LoadInst>(Val: R)) { |
| 1011 | unsigned PtrX = LoadInst::getPointerOperandIndex(); |
| 1012 | if (&Ld->getOperandUse(i: PtrX) == U) |
| 1013 | LSs.insert(X: U); |
| 1014 | } else if (StoreInst *St = dyn_cast<StoreInst>(Val: R)) { |
| 1015 | unsigned PtrX = StoreInst::getPointerOperandIndex(); |
| 1016 | if (&St->getOperandUse(i: PtrX) == U) |
| 1017 | LSs.insert(X: U); |
| 1018 | } |
| 1019 | } |
| 1020 | // Even if the total use count is 1, separating the chain may still be |
| 1021 | // beneficial, since the constant chain may be longer than the GEP alone |
| 1022 | // would be (e.g. if the parent node has a constant index and also has |
| 1023 | // other children). |
| 1024 | if (!LSs.empty()) |
| 1025 | FNs.insert(x: std::make_pair(x&: N, y&: LSs)); |
| 1026 | } |
| 1027 | |
| 1028 | LLVM_DEBUG(dbgs() << "Nodes with foldable users:\n" << FNs); |
| 1029 | |
| 1030 | for (auto &FN : FNs) { |
| 1031 | GepNode *N = FN.first; |
| 1032 | UseSet &Us = FN.second; |
| 1033 | for (Use *U : Us) |
| 1034 | separateChainForNode(Node: N, U, Loc); |
| 1035 | } |
| 1036 | } |
| 1037 | |
| 1038 | void HexagonCommonGEP::computeNodePlacement(NodeToValueMap &Loc) { |
| 1039 | // Compute the inverse of the Node.Parent links. Also, collect the set |
| 1040 | // of root nodes. |
| 1041 | NodeChildrenMap NCM; |
| 1042 | NodeVect Roots; |
| 1043 | invert_find_roots(Nodes, NCM, Roots); |
| 1044 | |
| 1045 | // Compute the initial placement determined by the users' locations, and |
| 1046 | // the locations of the child nodes. |
| 1047 | for (GepNode *Root : Roots) |
| 1048 | recalculatePlacementRec(Node: Root, NCM, Loc); |
| 1049 | |
| 1050 | LLVM_DEBUG(dbgs() << "Initial node placement:\n" << LocationAsBlock(Loc)); |
| 1051 | |
| 1052 | if (OptEnableInv) { |
| 1053 | for (GepNode *Root : Roots) |
| 1054 | adjustForInvariance(Node: Root, NCM, Loc); |
| 1055 | |
| 1056 | LLVM_DEBUG(dbgs() << "Node placement after adjustment for invariance:\n" |
| 1057 | << LocationAsBlock(Loc)); |
| 1058 | } |
| 1059 | if (OptEnableConst) { |
| 1060 | for (GepNode *Root : Roots) |
| 1061 | separateConstantChains(Node: Root, NCM, Loc); |
| 1062 | } |
| 1063 | LLVM_DEBUG(dbgs() << "Node use information:\n" << Uses); |
| 1064 | |
| 1065 | // At the moment, there is no further refinement of the initial placement. |
| 1066 | // Such a refinement could include splitting the nodes if they are placed |
| 1067 | // too far from some of its users. |
| 1068 | |
| 1069 | LLVM_DEBUG(dbgs() << "Final node placement:\n" << LocationAsBlock(Loc)); |
| 1070 | } |
| 1071 | |
| 1072 | Value *HexagonCommonGEP::fabricateGEP(NodeVect &NA, BasicBlock::iterator At, |
| 1073 | BasicBlock *LocB) { |
| 1074 | LLVM_DEBUG(dbgs() << "Fabricating GEP in " << LocB->getName() |
| 1075 | << " for nodes:\n" |
| 1076 | << NA); |
| 1077 | unsigned Num = NA.size(); |
| 1078 | GepNode *RN = NA[0]; |
| 1079 | assert((RN->Flags & GepNode::Root) && "Creating GEP for non-root" ); |
| 1080 | |
| 1081 | GetElementPtrInst *NewInst = nullptr; |
| 1082 | Value *Input = RN->BaseVal; |
| 1083 | Type *InpTy = RN->PTy; |
| 1084 | |
| 1085 | unsigned Idx = 0; |
| 1086 | do { |
| 1087 | SmallVector<Value*, 4> IdxList; |
| 1088 | // If the type of the input of the first node is not a pointer, |
| 1089 | // we need to add an artificial i32 0 to the indices (because the |
| 1090 | // actual input in the IR will be a pointer). |
| 1091 | if (!(NA[Idx]->Flags & GepNode::Pointer)) { |
| 1092 | Type *Int32Ty = Type::getInt32Ty(C&: *Ctx); |
| 1093 | IdxList.push_back(Elt: ConstantInt::get(Ty: Int32Ty, V: 0)); |
| 1094 | } |
| 1095 | |
| 1096 | // Keep adding indices from NA until we have to stop and generate |
| 1097 | // an "intermediate" GEP. |
| 1098 | while (++Idx <= Num) { |
| 1099 | GepNode *N = NA[Idx-1]; |
| 1100 | IdxList.push_back(Elt: N->Idx); |
| 1101 | if (Idx < Num) { |
| 1102 | // We have to stop if we reach a pointer. |
| 1103 | if (NA[Idx]->Flags & GepNode::Pointer) |
| 1104 | break; |
| 1105 | } |
| 1106 | } |
| 1107 | NewInst = GetElementPtrInst::Create(PointeeType: InpTy, Ptr: Input, IdxList, NameStr: "cgep" , InsertBefore: At); |
| 1108 | NewInst->setIsInBounds(RN->Flags & GepNode::InBounds); |
| 1109 | LLVM_DEBUG(dbgs() << "new GEP: " << *NewInst << '\n'); |
| 1110 | if (Idx < Num) { |
| 1111 | Input = NewInst; |
| 1112 | InpTy = NA[Idx]->PTy; |
| 1113 | } |
| 1114 | } while (Idx <= Num); |
| 1115 | |
| 1116 | return NewInst; |
| 1117 | } |
| 1118 | |
| 1119 | void HexagonCommonGEP::getAllUsersForNode(GepNode *Node, ValueVect &Values, |
| 1120 | NodeChildrenMap &NCM) { |
| 1121 | NodeVect Work; |
| 1122 | Work.push_back(x: Node); |
| 1123 | |
| 1124 | while (!Work.empty()) { |
| 1125 | NodeVect::iterator First = Work.begin(); |
| 1126 | GepNode *N = *First; |
| 1127 | Work.erase(position: First); |
| 1128 | if (N->Flags & GepNode::Used) { |
| 1129 | NodeToUsesMap::iterator UF = Uses.find(x: N); |
| 1130 | assert(UF != Uses.end() && "No use information for used node" ); |
| 1131 | UseSet &Us = UF->second; |
| 1132 | for (const auto &U : Us) |
| 1133 | Values.push_back(x: U->getUser()); |
| 1134 | } |
| 1135 | NodeChildrenMap::iterator CF = NCM.find(x: N); |
| 1136 | if (CF != NCM.end()) { |
| 1137 | NodeVect &Cs = CF->second; |
| 1138 | llvm::append_range(C&: Work, R&: Cs); |
| 1139 | } |
| 1140 | } |
| 1141 | } |
| 1142 | |
| 1143 | void HexagonCommonGEP::materialize(NodeToValueMap &Loc) { |
| 1144 | LLVM_DEBUG(dbgs() << "Nodes before materialization:\n" << Nodes << '\n'); |
| 1145 | NodeChildrenMap NCM; |
| 1146 | NodeVect Roots; |
| 1147 | // Compute the inversion again, since computing placement could alter |
| 1148 | // "parent" relation between nodes. |
| 1149 | invert_find_roots(Nodes, NCM, Roots); |
| 1150 | |
| 1151 | while (!Roots.empty()) { |
| 1152 | NodeVect::iterator First = Roots.begin(); |
| 1153 | GepNode *Root = *First, *Last = *First; |
| 1154 | Roots.erase(position: First); |
| 1155 | |
| 1156 | NodeVect NA; // Nodes to assemble. |
| 1157 | // Append to NA all child nodes up to (and including) the first child |
| 1158 | // that: |
| 1159 | // (1) has more than 1 child, or |
| 1160 | // (2) is used, or |
| 1161 | // (3) has a child located in a different block. |
| 1162 | bool LastUsed = false; |
| 1163 | unsigned LastCN = 0; |
| 1164 | // The location may be null if the computation failed (it can legitimately |
| 1165 | // happen for nodes created from dead GEPs). |
| 1166 | Value *LocV = Loc[Last]; |
| 1167 | if (!LocV) |
| 1168 | continue; |
| 1169 | BasicBlock *LastB = cast<BasicBlock>(Val: LocV); |
| 1170 | do { |
| 1171 | NA.push_back(x: Last); |
| 1172 | LastUsed = (Last->Flags & GepNode::Used); |
| 1173 | if (LastUsed) |
| 1174 | break; |
| 1175 | NodeChildrenMap::iterator CF = NCM.find(x: Last); |
| 1176 | LastCN = (CF != NCM.end()) ? CF->second.size() : 0; |
| 1177 | if (LastCN != 1) |
| 1178 | break; |
| 1179 | GepNode *Child = CF->second.front(); |
| 1180 | BasicBlock *ChildB = cast_or_null<BasicBlock>(Val: Loc[Child]); |
| 1181 | if (ChildB != nullptr && LastB != ChildB) |
| 1182 | break; |
| 1183 | Last = Child; |
| 1184 | } while (true); |
| 1185 | |
| 1186 | BasicBlock::iterator InsertAt = LastB->getTerminator()->getIterator(); |
| 1187 | if (LastUsed || LastCN > 0) { |
| 1188 | ValueVect Urs; |
| 1189 | getAllUsersForNode(Node: Root, Values&: Urs, NCM); |
| 1190 | BasicBlock::iterator FirstUse = first_use_of_in_block(Values&: Urs, B: LastB); |
| 1191 | if (FirstUse != LastB->end()) |
| 1192 | InsertAt = FirstUse; |
| 1193 | } |
| 1194 | |
| 1195 | // Generate a new instruction for NA. |
| 1196 | Value *NewInst = fabricateGEP(NA, At: InsertAt, LocB: LastB); |
| 1197 | |
| 1198 | // Convert all the children of Last node into roots, and append them |
| 1199 | // to the Roots list. |
| 1200 | if (LastCN > 0) { |
| 1201 | NodeVect &Cs = NCM[Last]; |
| 1202 | for (GepNode *CN : Cs) { |
| 1203 | CN->Flags &= ~GepNode::Internal; |
| 1204 | CN->Flags |= GepNode::Root; |
| 1205 | CN->BaseVal = NewInst; |
| 1206 | Roots.push_back(x: CN); |
| 1207 | } |
| 1208 | } |
| 1209 | |
| 1210 | // Lastly, if the Last node was used, replace all uses with the new GEP. |
| 1211 | // The uses reference the original GEP values. |
| 1212 | if (LastUsed) { |
| 1213 | NodeToUsesMap::iterator UF = Uses.find(x: Last); |
| 1214 | assert(UF != Uses.end() && "No use information found" ); |
| 1215 | UseSet &Us = UF->second; |
| 1216 | for (Use *U : Us) |
| 1217 | U->set(NewInst); |
| 1218 | } |
| 1219 | } |
| 1220 | } |
| 1221 | |
| 1222 | void HexagonCommonGEP::removeDeadCode() { |
| 1223 | ValueVect BO; |
| 1224 | BO.push_back(x: &Fn->front()); |
| 1225 | |
| 1226 | for (unsigned i = 0; i < BO.size(); ++i) { |
| 1227 | BasicBlock *B = cast<BasicBlock>(Val: BO[i]); |
| 1228 | for (auto *DTN : children<DomTreeNode *>(G: DT->getNode(BB: B))) |
| 1229 | BO.push_back(x: DTN->getBlock()); |
| 1230 | } |
| 1231 | |
| 1232 | for (Value *V : llvm::reverse(C&: BO)) { |
| 1233 | BasicBlock *B = cast<BasicBlock>(Val: V); |
| 1234 | ValueVect Ins; |
| 1235 | for (Instruction &I : llvm::reverse(C&: *B)) |
| 1236 | Ins.push_back(x: &I); |
| 1237 | for (Value *I : Ins) { |
| 1238 | Instruction *In = cast<Instruction>(Val: I); |
| 1239 | if (isInstructionTriviallyDead(I: In)) |
| 1240 | In->eraseFromParent(); |
| 1241 | } |
| 1242 | } |
| 1243 | } |
| 1244 | |
| 1245 | bool HexagonCommonGEP::runOnFunction(Function &F) { |
| 1246 | if (skipFunction(F)) |
| 1247 | return false; |
| 1248 | |
| 1249 | // For now bail out on C++ exception handling. |
| 1250 | for (const BasicBlock &BB : F) |
| 1251 | for (const Instruction &I : BB) |
| 1252 | if (isa<InvokeInst>(Val: I) || isa<LandingPadInst>(Val: I)) |
| 1253 | return false; |
| 1254 | |
| 1255 | Fn = &F; |
| 1256 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 1257 | PDT = &getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree(); |
| 1258 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
| 1259 | Ctx = &F.getContext(); |
| 1260 | |
| 1261 | Nodes.clear(); |
| 1262 | Uses.clear(); |
| 1263 | NodeOrder.clear(); |
| 1264 | |
| 1265 | SpecificBumpPtrAllocator<GepNode> Allocator; |
| 1266 | Mem = &Allocator; |
| 1267 | |
| 1268 | collect(); |
| 1269 | common(); |
| 1270 | |
| 1271 | NodeToValueMap Loc; |
| 1272 | computeNodePlacement(Loc); |
| 1273 | materialize(Loc); |
| 1274 | removeDeadCode(); |
| 1275 | |
| 1276 | #ifdef EXPENSIVE_CHECKS |
| 1277 | // Run this only when expensive checks are enabled. |
| 1278 | if (verifyFunction(F, &dbgs())) |
| 1279 | report_fatal_error("Broken function" ); |
| 1280 | #endif |
| 1281 | return true; |
| 1282 | } |
| 1283 | |
| 1284 | namespace llvm { |
| 1285 | |
| 1286 | FunctionPass *createHexagonCommonGEP() { |
| 1287 | return new HexagonCommonGEP(); |
| 1288 | } |
| 1289 | |
| 1290 | } // end namespace llvm |
| 1291 | |