Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 1 | //===--- HexagonCommonGEP.cpp ---------------------------------------------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | |
| 10 | #define DEBUG_TYPE "commgep" |
| 11 | |
| 12 | #include "llvm/Pass.h" |
| 13 | #include "llvm/ADT/FoldingSet.h" |
| 14 | #include "llvm/ADT/STLExtras.h" |
| 15 | #include "llvm/Analysis/LoopInfo.h" |
| 16 | #include "llvm/Analysis/PostDominators.h" |
| 17 | #include "llvm/CodeGen/MachineFunctionAnalysis.h" |
| 18 | #include "llvm/IR/Constants.h" |
| 19 | #include "llvm/IR/Dominators.h" |
| 20 | #include "llvm/IR/Function.h" |
| 21 | #include "llvm/IR/Instructions.h" |
| 22 | #include "llvm/IR/Verifier.h" |
| 23 | #include "llvm/Support/Allocator.h" |
| 24 | #include "llvm/Support/CommandLine.h" |
| 25 | #include "llvm/Support/Debug.h" |
| 26 | #include "llvm/Support/raw_ostream.h" |
| 27 | #include "llvm/Transforms/Scalar.h" |
| 28 | #include "llvm/Transforms/Utils/Local.h" |
| 29 | |
| 30 | #include <map> |
| 31 | #include <set> |
| 32 | #include <vector> |
| 33 | |
| 34 | #include "HexagonTargetMachine.h" |
| 35 | |
| 36 | using namespace llvm; |
| 37 | |
| 38 | static cl::opt<bool> OptSpeculate("commgep-speculate", cl::init(true), |
| 39 | cl::Hidden, cl::ZeroOrMore); |
| 40 | |
| 41 | static cl::opt<bool> OptEnableInv("commgep-inv", cl::init(true), cl::Hidden, |
| 42 | cl::ZeroOrMore); |
| 43 | |
| 44 | static cl::opt<bool> OptEnableConst("commgep-const", cl::init(true), |
| 45 | cl::Hidden, cl::ZeroOrMore); |
| 46 | |
| 47 | namespace llvm { |
| 48 | void initializeHexagonCommonGEPPass(PassRegistry&); |
| 49 | } |
| 50 | |
| 51 | namespace { |
| 52 | struct GepNode; |
| 53 | typedef std::set<GepNode*> NodeSet; |
| 54 | typedef std::map<GepNode*,Value*> NodeToValueMap; |
| 55 | typedef std::vector<GepNode*> NodeVect; |
| 56 | typedef std::map<GepNode*,NodeVect> NodeChildrenMap; |
| 57 | typedef std::set<Use*> UseSet; |
| 58 | typedef std::map<GepNode*,UseSet> NodeToUsesMap; |
| 59 | |
| 60 | // Numbering map for gep nodes. Used to keep track of ordering for |
| 61 | // gep nodes. |
Benjamin Kramer | 9a5d788 | 2015-07-18 17:43:23 +0000 | [diff] [blame] | 62 | struct NodeOrdering { |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 63 | NodeOrdering() : LastNum(0) {} |
Benjamin Kramer | 9a5d788 | 2015-07-18 17:43:23 +0000 | [diff] [blame] | 64 | |
| 65 | void insert(const GepNode *N) { Map.insert(std::make_pair(N, ++LastNum)); } |
| 66 | void clear() { Map.clear(); } |
| 67 | |
| 68 | bool operator()(const GepNode *N1, const GepNode *N2) const { |
| 69 | auto F1 = Map.find(N1), F2 = Map.find(N2); |
| 70 | assert(F1 != Map.end() && F2 != Map.end()); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 71 | return F1->second < F2->second; |
| 72 | } |
Benjamin Kramer | 9a5d788 | 2015-07-18 17:43:23 +0000 | [diff] [blame] | 73 | |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 74 | private: |
Benjamin Kramer | 9a5d788 | 2015-07-18 17:43:23 +0000 | [diff] [blame] | 75 | std::map<const GepNode *, unsigned> Map; |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 76 | unsigned LastNum; |
| 77 | }; |
| 78 | |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 79 | class HexagonCommonGEP : public FunctionPass { |
| 80 | public: |
| 81 | static char ID; |
| 82 | HexagonCommonGEP() : FunctionPass(ID) { |
| 83 | initializeHexagonCommonGEPPass(*PassRegistry::getPassRegistry()); |
| 84 | } |
| 85 | virtual bool runOnFunction(Function &F); |
| 86 | virtual const char *getPassName() const { |
| 87 | return "Hexagon Common GEP"; |
| 88 | } |
| 89 | |
| 90 | virtual void getAnalysisUsage(AnalysisUsage &AU) const { |
| 91 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 92 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 93 | AU.addRequired<PostDominatorTree>(); |
| 94 | AU.addPreserved<PostDominatorTree>(); |
| 95 | AU.addRequired<LoopInfoWrapperPass>(); |
| 96 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 97 | FunctionPass::getAnalysisUsage(AU); |
| 98 | } |
| 99 | |
| 100 | private: |
| 101 | typedef std::map<Value*,GepNode*> ValueToNodeMap; |
| 102 | typedef std::vector<Value*> ValueVect; |
| 103 | typedef std::map<GepNode*,ValueVect> NodeToValuesMap; |
| 104 | |
| 105 | void getBlockTraversalOrder(BasicBlock *Root, ValueVect &Order); |
| 106 | bool isHandledGepForm(GetElementPtrInst *GepI); |
| 107 | void processGepInst(GetElementPtrInst *GepI, ValueToNodeMap &NM); |
| 108 | void collect(); |
| 109 | void common(); |
| 110 | |
| 111 | BasicBlock *recalculatePlacement(GepNode *Node, NodeChildrenMap &NCM, |
| 112 | NodeToValueMap &Loc); |
| 113 | BasicBlock *recalculatePlacementRec(GepNode *Node, NodeChildrenMap &NCM, |
| 114 | NodeToValueMap &Loc); |
| 115 | bool isInvariantIn(Value *Val, Loop *L); |
| 116 | bool isInvariantIn(GepNode *Node, Loop *L); |
| 117 | bool isInMainPath(BasicBlock *B, Loop *L); |
| 118 | BasicBlock *adjustForInvariance(GepNode *Node, NodeChildrenMap &NCM, |
| 119 | NodeToValueMap &Loc); |
| 120 | void separateChainForNode(GepNode *Node, Use *U, NodeToValueMap &Loc); |
| 121 | void separateConstantChains(GepNode *Node, NodeChildrenMap &NCM, |
| 122 | NodeToValueMap &Loc); |
| 123 | void computeNodePlacement(NodeToValueMap &Loc); |
| 124 | |
| 125 | Value *fabricateGEP(NodeVect &NA, BasicBlock::iterator At, |
| 126 | BasicBlock *LocB); |
| 127 | void getAllUsersForNode(GepNode *Node, ValueVect &Values, |
| 128 | NodeChildrenMap &NCM); |
| 129 | void materialize(NodeToValueMap &Loc); |
| 130 | |
| 131 | void removeDeadCode(); |
| 132 | |
| 133 | NodeVect Nodes; |
| 134 | NodeToUsesMap Uses; |
| 135 | NodeOrdering NodeOrder; // Node ordering, for deterministic behavior. |
| 136 | SpecificBumpPtrAllocator<GepNode> *Mem; |
| 137 | LLVMContext *Ctx; |
| 138 | LoopInfo *LI; |
| 139 | DominatorTree *DT; |
| 140 | PostDominatorTree *PDT; |
| 141 | Function *Fn; |
| 142 | }; |
| 143 | } |
| 144 | |
| 145 | |
| 146 | char HexagonCommonGEP::ID = 0; |
| 147 | INITIALIZE_PASS_BEGIN(HexagonCommonGEP, "hcommgep", "Hexagon Common GEP", |
| 148 | false, false) |
| 149 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 150 | INITIALIZE_PASS_DEPENDENCY(PostDominatorTree) |
| 151 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) |
| 152 | INITIALIZE_PASS_END(HexagonCommonGEP, "hcommgep", "Hexagon Common GEP", |
| 153 | false, false) |
| 154 | |
| 155 | namespace { |
| 156 | struct GepNode { |
| 157 | enum { |
| 158 | None = 0, |
| 159 | Root = 0x01, |
| 160 | Internal = 0x02, |
| 161 | Used = 0x04 |
| 162 | }; |
| 163 | |
| 164 | uint32_t Flags; |
| 165 | union { |
| 166 | GepNode *Parent; |
| 167 | Value *BaseVal; |
| 168 | }; |
| 169 | Value *Idx; |
| 170 | Type *PTy; // Type of the pointer operand. |
| 171 | |
| 172 | GepNode() : Flags(0), Parent(0), Idx(0), PTy(0) {} |
| 173 | GepNode(const GepNode *N) : Flags(N->Flags), Idx(N->Idx), PTy(N->PTy) { |
| 174 | if (Flags & Root) |
| 175 | BaseVal = N->BaseVal; |
| 176 | else |
| 177 | Parent = N->Parent; |
| 178 | } |
| 179 | friend raw_ostream &operator<< (raw_ostream &OS, const GepNode &GN); |
| 180 | }; |
| 181 | |
| 182 | |
| 183 | Type *next_type(Type *Ty, Value *Idx) { |
| 184 | // Advance the type. |
| 185 | if (!Ty->isStructTy()) { |
| 186 | Type *NexTy = cast<SequentialType>(Ty)->getElementType(); |
| 187 | return NexTy; |
| 188 | } |
| 189 | // Otherwise it is a struct type. |
| 190 | ConstantInt *CI = dyn_cast<ConstantInt>(Idx); |
| 191 | assert(CI && "Struct type with non-constant index"); |
| 192 | int64_t i = CI->getValue().getSExtValue(); |
| 193 | Type *NextTy = cast<StructType>(Ty)->getElementType(i); |
| 194 | return NextTy; |
| 195 | } |
| 196 | |
| 197 | |
| 198 | raw_ostream &operator<< (raw_ostream &OS, const GepNode &GN) { |
| 199 | OS << "{ {"; |
| 200 | bool Comma = false; |
| 201 | if (GN.Flags & GepNode::Root) { |
| 202 | OS << "root"; |
| 203 | Comma = true; |
| 204 | } |
| 205 | if (GN.Flags & GepNode::Internal) { |
| 206 | if (Comma) |
| 207 | OS << ','; |
| 208 | OS << "internal"; |
| 209 | Comma = true; |
| 210 | } |
| 211 | if (GN.Flags & GepNode::Used) { |
| 212 | if (Comma) |
| 213 | OS << ','; |
| 214 | OS << "used"; |
| 215 | Comma = true; |
| 216 | } |
| 217 | OS << "} "; |
| 218 | if (GN.Flags & GepNode::Root) |
| 219 | OS << "BaseVal:" << GN.BaseVal->getName() << '(' << GN.BaseVal << ')'; |
| 220 | else |
| 221 | OS << "Parent:" << GN.Parent; |
| 222 | |
| 223 | OS << " Idx:"; |
| 224 | if (ConstantInt *CI = dyn_cast<ConstantInt>(GN.Idx)) |
| 225 | OS << CI->getValue().getSExtValue(); |
| 226 | else if (GN.Idx->hasName()) |
| 227 | OS << GN.Idx->getName(); |
| 228 | else |
| 229 | OS << "<anon> =" << *GN.Idx; |
| 230 | |
| 231 | OS << " PTy:"; |
| 232 | if (GN.PTy->isStructTy()) { |
| 233 | StructType *STy = cast<StructType>(GN.PTy); |
| 234 | if (!STy->isLiteral()) |
| 235 | OS << GN.PTy->getStructName(); |
| 236 | else |
| 237 | OS << "<anon-struct>:" << *STy; |
| 238 | } |
| 239 | else |
| 240 | OS << *GN.PTy; |
| 241 | OS << " }"; |
| 242 | return OS; |
| 243 | } |
| 244 | |
| 245 | |
| 246 | template <typename NodeContainer> |
| 247 | void dump_node_container(raw_ostream &OS, const NodeContainer &S) { |
| 248 | typedef typename NodeContainer::const_iterator const_iterator; |
| 249 | for (const_iterator I = S.begin(), E = S.end(); I != E; ++I) |
| 250 | OS << *I << ' ' << **I << '\n'; |
| 251 | } |
| 252 | |
| 253 | raw_ostream &operator<< (raw_ostream &OS, |
| 254 | const NodeVect &S) LLVM_ATTRIBUTE_UNUSED; |
| 255 | raw_ostream &operator<< (raw_ostream &OS, const NodeVect &S) { |
| 256 | dump_node_container(OS, S); |
| 257 | return OS; |
| 258 | } |
| 259 | |
| 260 | |
| 261 | raw_ostream &operator<< (raw_ostream &OS, |
| 262 | const NodeToUsesMap &M) LLVM_ATTRIBUTE_UNUSED; |
| 263 | raw_ostream &operator<< (raw_ostream &OS, const NodeToUsesMap &M){ |
| 264 | typedef NodeToUsesMap::const_iterator const_iterator; |
| 265 | for (const_iterator I = M.begin(), E = M.end(); I != E; ++I) { |
| 266 | const UseSet &Us = I->second; |
| 267 | OS << I->first << " -> #" << Us.size() << '{'; |
| 268 | for (UseSet::const_iterator J = Us.begin(), F = Us.end(); J != F; ++J) { |
| 269 | User *R = (*J)->getUser(); |
| 270 | if (R->hasName()) |
| 271 | OS << ' ' << R->getName(); |
| 272 | else |
| 273 | OS << " <?>(" << *R << ')'; |
| 274 | } |
| 275 | OS << " }\n"; |
| 276 | } |
| 277 | return OS; |
| 278 | } |
| 279 | |
| 280 | |
| 281 | struct in_set { |
| 282 | in_set(const NodeSet &S) : NS(S) {} |
| 283 | bool operator() (GepNode *N) const { |
| 284 | return NS.find(N) != NS.end(); |
| 285 | } |
| 286 | private: |
| 287 | const NodeSet &NS; |
| 288 | }; |
| 289 | } |
| 290 | |
| 291 | |
| 292 | inline void *operator new(size_t, SpecificBumpPtrAllocator<GepNode> &A) { |
| 293 | return A.Allocate(); |
| 294 | } |
| 295 | |
| 296 | |
| 297 | void HexagonCommonGEP::getBlockTraversalOrder(BasicBlock *Root, |
| 298 | ValueVect &Order) { |
| 299 | // Compute block ordering for a typical DT-based traversal of the flow |
| 300 | // graph: "before visiting a block, all of its dominators must have been |
| 301 | // visited". |
| 302 | |
| 303 | Order.push_back(Root); |
| 304 | DomTreeNode *DTN = DT->getNode(Root); |
| 305 | typedef GraphTraits<DomTreeNode*> GTN; |
| 306 | typedef GTN::ChildIteratorType Iter; |
| 307 | for (Iter I = GTN::child_begin(DTN), E = GTN::child_end(DTN); I != E; ++I) |
| 308 | getBlockTraversalOrder((*I)->getBlock(), Order); |
| 309 | } |
| 310 | |
| 311 | |
| 312 | bool HexagonCommonGEP::isHandledGepForm(GetElementPtrInst *GepI) { |
| 313 | // No vector GEPs. |
| 314 | if (!GepI->getType()->isPointerTy()) |
| 315 | return false; |
| 316 | // No GEPs without any indices. (Is this possible?) |
| 317 | if (GepI->idx_begin() == GepI->idx_end()) |
| 318 | return false; |
| 319 | return true; |
| 320 | } |
| 321 | |
| 322 | |
| 323 | void HexagonCommonGEP::processGepInst(GetElementPtrInst *GepI, |
| 324 | ValueToNodeMap &NM) { |
| 325 | DEBUG(dbgs() << "Visiting GEP: " << *GepI << '\n'); |
| 326 | GepNode *N = new (*Mem) GepNode; |
| 327 | Value *PtrOp = GepI->getPointerOperand(); |
| 328 | ValueToNodeMap::iterator F = NM.find(PtrOp); |
| 329 | if (F == NM.end()) { |
| 330 | N->BaseVal = PtrOp; |
| 331 | N->Flags |= GepNode::Root; |
| 332 | } else { |
| 333 | // If PtrOp was a GEP instruction, it must have already been processed. |
| 334 | // The ValueToNodeMap entry for it is the last gep node in the generated |
| 335 | // chain. Link to it here. |
| 336 | N->Parent = F->second; |
| 337 | } |
| 338 | N->PTy = PtrOp->getType(); |
| 339 | N->Idx = *GepI->idx_begin(); |
| 340 | |
| 341 | // Collect the list of users of this GEP instruction. Will add it to the |
| 342 | // last node created for it. |
| 343 | UseSet Us; |
| 344 | for (Value::user_iterator UI = GepI->user_begin(), UE = GepI->user_end(); |
| 345 | UI != UE; ++UI) { |
| 346 | // Check if this gep is used by anything other than other geps that |
| 347 | // we will process. |
| 348 | if (isa<GetElementPtrInst>(*UI)) { |
| 349 | GetElementPtrInst *UserG = cast<GetElementPtrInst>(*UI); |
| 350 | if (isHandledGepForm(UserG)) |
| 351 | continue; |
| 352 | } |
| 353 | Us.insert(&UI.getUse()); |
| 354 | } |
| 355 | Nodes.push_back(N); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 356 | NodeOrder.insert(N); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 357 | |
| 358 | // Skip the first index operand, since we only handle 0. This dereferences |
| 359 | // the pointer operand. |
| 360 | GepNode *PN = N; |
| 361 | Type *PtrTy = cast<PointerType>(PtrOp->getType())->getElementType(); |
| 362 | for (User::op_iterator OI = GepI->idx_begin()+1, OE = GepI->idx_end(); |
| 363 | OI != OE; ++OI) { |
| 364 | Value *Op = *OI; |
| 365 | GepNode *Nx = new (*Mem) GepNode; |
| 366 | Nx->Parent = PN; // Link Nx to the previous node. |
| 367 | Nx->Flags |= GepNode::Internal; |
| 368 | Nx->PTy = PtrTy; |
| 369 | Nx->Idx = Op; |
| 370 | Nodes.push_back(Nx); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 371 | NodeOrder.insert(Nx); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 372 | PN = Nx; |
| 373 | |
| 374 | PtrTy = next_type(PtrTy, Op); |
| 375 | } |
| 376 | |
| 377 | // After last node has been created, update the use information. |
| 378 | if (!Us.empty()) { |
| 379 | PN->Flags |= GepNode::Used; |
| 380 | Uses[PN].insert(Us.begin(), Us.end()); |
| 381 | } |
| 382 | |
| 383 | // Link the last node with the originating GEP instruction. This is to |
| 384 | // help with linking chained GEP instructions. |
| 385 | NM.insert(std::make_pair(GepI, PN)); |
| 386 | } |
| 387 | |
| 388 | |
| 389 | void HexagonCommonGEP::collect() { |
| 390 | // Establish depth-first traversal order of the dominator tree. |
| 391 | ValueVect BO; |
Duncan P. N. Exon Smith | a72c6e2 | 2015-10-20 00:46:39 +0000 | [diff] [blame] | 392 | getBlockTraversalOrder(&Fn->front(), BO); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 393 | |
| 394 | // The creation of gep nodes requires DT-traversal. When processing a GEP |
| 395 | // instruction that uses another GEP instruction as the base pointer, the |
| 396 | // gep node for the base pointer should already exist. |
| 397 | ValueToNodeMap NM; |
| 398 | for (ValueVect::iterator I = BO.begin(), E = BO.end(); I != E; ++I) { |
| 399 | BasicBlock *B = cast<BasicBlock>(*I); |
| 400 | for (BasicBlock::iterator J = B->begin(), F = B->end(); J != F; ++J) { |
| 401 | if (!isa<GetElementPtrInst>(J)) |
| 402 | continue; |
| 403 | GetElementPtrInst *GepI = cast<GetElementPtrInst>(J); |
| 404 | if (isHandledGepForm(GepI)) |
| 405 | processGepInst(GepI, NM); |
| 406 | } |
| 407 | } |
| 408 | |
| 409 | DEBUG(dbgs() << "Gep nodes after initial collection:\n" << Nodes); |
| 410 | } |
| 411 | |
| 412 | |
| 413 | namespace { |
| 414 | void invert_find_roots(const NodeVect &Nodes, NodeChildrenMap &NCM, |
| 415 | NodeVect &Roots) { |
| 416 | typedef NodeVect::const_iterator const_iterator; |
| 417 | for (const_iterator I = Nodes.begin(), E = Nodes.end(); I != E; ++I) { |
| 418 | GepNode *N = *I; |
| 419 | if (N->Flags & GepNode::Root) { |
| 420 | Roots.push_back(N); |
| 421 | continue; |
| 422 | } |
| 423 | GepNode *PN = N->Parent; |
| 424 | NCM[PN].push_back(N); |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | void nodes_for_root(GepNode *Root, NodeChildrenMap &NCM, NodeSet &Nodes) { |
| 429 | NodeVect Work; |
| 430 | Work.push_back(Root); |
| 431 | Nodes.insert(Root); |
| 432 | |
| 433 | while (!Work.empty()) { |
| 434 | NodeVect::iterator First = Work.begin(); |
| 435 | GepNode *N = *First; |
| 436 | Work.erase(First); |
| 437 | NodeChildrenMap::iterator CF = NCM.find(N); |
| 438 | if (CF != NCM.end()) { |
| 439 | Work.insert(Work.end(), CF->second.begin(), CF->second.end()); |
| 440 | Nodes.insert(CF->second.begin(), CF->second.end()); |
| 441 | } |
| 442 | } |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | |
| 447 | namespace { |
| 448 | typedef std::set<NodeSet> NodeSymRel; |
| 449 | typedef std::pair<GepNode*,GepNode*> NodePair; |
| 450 | typedef std::set<NodePair> NodePairSet; |
| 451 | |
| 452 | const NodeSet *node_class(GepNode *N, NodeSymRel &Rel) { |
| 453 | for (NodeSymRel::iterator I = Rel.begin(), E = Rel.end(); I != E; ++I) |
| 454 | if (I->count(N)) |
| 455 | return &*I; |
| 456 | return 0; |
| 457 | } |
| 458 | |
| 459 | // Create an ordered pair of GepNode pointers. The pair will be used in |
| 460 | // determining equality. The only purpose of the ordering is to eliminate |
| 461 | // duplication due to the commutativity of equality/non-equality. |
| 462 | NodePair node_pair(GepNode *N1, GepNode *N2) { |
| 463 | uintptr_t P1 = uintptr_t(N1), P2 = uintptr_t(N2); |
| 464 | if (P1 <= P2) |
| 465 | return std::make_pair(N1, N2); |
| 466 | return std::make_pair(N2, N1); |
| 467 | } |
| 468 | |
| 469 | unsigned node_hash(GepNode *N) { |
| 470 | // Include everything except flags and parent. |
| 471 | FoldingSetNodeID ID; |
| 472 | ID.AddPointer(N->Idx); |
| 473 | ID.AddPointer(N->PTy); |
| 474 | return ID.ComputeHash(); |
| 475 | } |
| 476 | |
| 477 | bool node_eq(GepNode *N1, GepNode *N2, NodePairSet &Eq, NodePairSet &Ne) { |
| 478 | // Don't cache the result for nodes with different hashes. The hash |
| 479 | // comparison is fast enough. |
| 480 | if (node_hash(N1) != node_hash(N2)) |
| 481 | return false; |
| 482 | |
| 483 | NodePair NP = node_pair(N1, N2); |
| 484 | NodePairSet::iterator FEq = Eq.find(NP); |
| 485 | if (FEq != Eq.end()) |
| 486 | return true; |
| 487 | NodePairSet::iterator FNe = Ne.find(NP); |
| 488 | if (FNe != Ne.end()) |
| 489 | return false; |
| 490 | // Not previously compared. |
| 491 | bool Root1 = N1->Flags & GepNode::Root; |
| 492 | bool Root2 = N2->Flags & GepNode::Root; |
| 493 | NodePair P = node_pair(N1, N2); |
| 494 | // If the Root flag has different values, the nodes are different. |
| 495 | // If both nodes are root nodes, but their base pointers differ, |
| 496 | // they are different. |
| 497 | if (Root1 != Root2 || (Root1 && N1->BaseVal != N2->BaseVal)) { |
| 498 | Ne.insert(P); |
| 499 | return false; |
| 500 | } |
| 501 | // Here the root flags are identical, and for root nodes the |
| 502 | // base pointers are equal, so the root nodes are equal. |
| 503 | // For non-root nodes, compare their parent nodes. |
| 504 | if (Root1 || node_eq(N1->Parent, N2->Parent, Eq, Ne)) { |
| 505 | Eq.insert(P); |
| 506 | return true; |
| 507 | } |
| 508 | return false; |
| 509 | } |
| 510 | } |
| 511 | |
| 512 | |
| 513 | void HexagonCommonGEP::common() { |
| 514 | // The essence of this commoning is finding gep nodes that are equal. |
| 515 | // To do this we need to compare all pairs of nodes. To save time, |
| 516 | // first, partition the set of all nodes into sets of potentially equal |
| 517 | // nodes, and then compare pairs from within each partition. |
| 518 | typedef std::map<unsigned,NodeSet> NodeSetMap; |
| 519 | NodeSetMap MaybeEq; |
| 520 | |
| 521 | for (NodeVect::iterator I = Nodes.begin(), E = Nodes.end(); I != E; ++I) { |
| 522 | GepNode *N = *I; |
| 523 | unsigned H = node_hash(N); |
| 524 | MaybeEq[H].insert(N); |
| 525 | } |
| 526 | |
| 527 | // Compute the equivalence relation for the gep nodes. Use two caches, |
| 528 | // one for equality and the other for non-equality. |
| 529 | NodeSymRel EqRel; // Equality relation (as set of equivalence classes). |
| 530 | NodePairSet Eq, Ne; // Caches. |
| 531 | for (NodeSetMap::iterator I = MaybeEq.begin(), E = MaybeEq.end(); |
| 532 | I != E; ++I) { |
| 533 | NodeSet &S = I->second; |
| 534 | for (NodeSet::iterator NI = S.begin(), NE = S.end(); NI != NE; ++NI) { |
| 535 | GepNode *N = *NI; |
| 536 | // If node already has a class, then the class must have been created |
| 537 | // in a prior iteration of this loop. Since equality is transitive, |
| 538 | // nothing more will be added to that class, so skip it. |
| 539 | if (node_class(N, EqRel)) |
| 540 | continue; |
| 541 | |
| 542 | // Create a new class candidate now. |
| 543 | NodeSet C; |
| 544 | for (NodeSet::iterator NJ = std::next(NI); NJ != NE; ++NJ) |
| 545 | if (node_eq(N, *NJ, Eq, Ne)) |
| 546 | C.insert(*NJ); |
| 547 | // If Tmp is empty, N would be the only element in it. Don't bother |
| 548 | // creating a class for it then. |
| 549 | if (!C.empty()) { |
| 550 | C.insert(N); // Finalize the set before adding it to the relation. |
| 551 | std::pair<NodeSymRel::iterator, bool> Ins = EqRel.insert(C); |
| 552 | (void)Ins; |
| 553 | assert(Ins.second && "Cannot add a class"); |
| 554 | } |
| 555 | } |
| 556 | } |
| 557 | |
| 558 | DEBUG({ |
| 559 | dbgs() << "Gep node equality:\n"; |
| 560 | for (NodePairSet::iterator I = Eq.begin(), E = Eq.end(); I != E; ++I) |
| 561 | dbgs() << "{ " << I->first << ", " << I->second << " }\n"; |
| 562 | |
| 563 | dbgs() << "Gep equivalence classes:\n"; |
| 564 | for (NodeSymRel::iterator I = EqRel.begin(), E = EqRel.end(); I != E; ++I) { |
| 565 | dbgs() << '{'; |
| 566 | const NodeSet &S = *I; |
| 567 | for (NodeSet::const_iterator J = S.begin(), F = S.end(); J != F; ++J) { |
| 568 | if (J != S.begin()) |
| 569 | dbgs() << ','; |
| 570 | dbgs() << ' ' << *J; |
| 571 | } |
| 572 | dbgs() << " }\n"; |
| 573 | } |
| 574 | }); |
| 575 | |
| 576 | |
| 577 | // Create a projection from a NodeSet to the minimal element in it. |
| 578 | typedef std::map<const NodeSet*,GepNode*> ProjMap; |
| 579 | ProjMap PM; |
| 580 | for (NodeSymRel::iterator I = EqRel.begin(), E = EqRel.end(); I != E; ++I) { |
| 581 | const NodeSet &S = *I; |
| 582 | GepNode *Min = *std::min_element(S.begin(), S.end(), NodeOrder); |
| 583 | std::pair<ProjMap::iterator,bool> Ins = PM.insert(std::make_pair(&S, Min)); |
| 584 | (void)Ins; |
| 585 | assert(Ins.second && "Cannot add minimal element"); |
| 586 | |
| 587 | // Update the min element's flags, and user list. |
| 588 | uint32_t Flags = 0; |
| 589 | UseSet &MinUs = Uses[Min]; |
| 590 | for (NodeSet::iterator J = S.begin(), F = S.end(); J != F; ++J) { |
| 591 | GepNode *N = *J; |
| 592 | uint32_t NF = N->Flags; |
| 593 | // If N is used, append all original values of N to the list of |
| 594 | // original values of Min. |
| 595 | if (NF & GepNode::Used) |
| 596 | MinUs.insert(Uses[N].begin(), Uses[N].end()); |
| 597 | Flags |= NF; |
| 598 | } |
| 599 | if (MinUs.empty()) |
| 600 | Uses.erase(Min); |
| 601 | |
| 602 | // The collected flags should include all the flags from the min element. |
| 603 | assert((Min->Flags & Flags) == Min->Flags); |
| 604 | Min->Flags = Flags; |
| 605 | } |
| 606 | |
| 607 | // Commoning: for each non-root gep node, replace "Parent" with the |
| 608 | // selected (minimum) node from the corresponding equivalence class. |
| 609 | // If a given parent does not have an equivalence class, leave it |
| 610 | // unchanged (it means that it's the only element in its class). |
| 611 | for (NodeVect::iterator I = Nodes.begin(), E = Nodes.end(); I != E; ++I) { |
| 612 | GepNode *N = *I; |
| 613 | if (N->Flags & GepNode::Root) |
| 614 | continue; |
| 615 | const NodeSet *PC = node_class(N->Parent, EqRel); |
| 616 | if (!PC) |
| 617 | continue; |
| 618 | ProjMap::iterator F = PM.find(PC); |
| 619 | if (F == PM.end()) |
| 620 | continue; |
| 621 | // Found a replacement, use it. |
| 622 | GepNode *Rep = F->second; |
| 623 | N->Parent = Rep; |
| 624 | } |
| 625 | |
| 626 | DEBUG(dbgs() << "Gep nodes after commoning:\n" << Nodes); |
| 627 | |
| 628 | // Finally, erase the nodes that are no longer used. |
| 629 | NodeSet Erase; |
| 630 | for (NodeVect::iterator I = Nodes.begin(), E = Nodes.end(); I != E; ++I) { |
| 631 | GepNode *N = *I; |
| 632 | const NodeSet *PC = node_class(N, EqRel); |
| 633 | if (!PC) |
| 634 | continue; |
| 635 | ProjMap::iterator F = PM.find(PC); |
| 636 | if (F == PM.end()) |
| 637 | continue; |
| 638 | if (N == F->second) |
| 639 | continue; |
| 640 | // Node for removal. |
| 641 | Erase.insert(*I); |
| 642 | } |
| 643 | NodeVect::iterator NewE = std::remove_if(Nodes.begin(), Nodes.end(), |
| 644 | in_set(Erase)); |
| 645 | Nodes.resize(std::distance(Nodes.begin(), NewE)); |
| 646 | |
| 647 | DEBUG(dbgs() << "Gep nodes after post-commoning cleanup:\n" << Nodes); |
| 648 | } |
| 649 | |
| 650 | |
| 651 | namespace { |
| 652 | template <typename T> |
| 653 | BasicBlock *nearest_common_dominator(DominatorTree *DT, T &Blocks) { |
| 654 | DEBUG({ |
| 655 | dbgs() << "NCD of {"; |
| 656 | for (typename T::iterator I = Blocks.begin(), E = Blocks.end(); |
| 657 | I != E; ++I) { |
| 658 | if (!*I) |
| 659 | continue; |
| 660 | BasicBlock *B = cast<BasicBlock>(*I); |
| 661 | dbgs() << ' ' << B->getName(); |
| 662 | } |
| 663 | dbgs() << " }\n"; |
| 664 | }); |
| 665 | |
| 666 | // Allow null basic blocks in Blocks. In such cases, return 0. |
| 667 | typename T::iterator I = Blocks.begin(), E = Blocks.end(); |
| 668 | if (I == E || !*I) |
| 669 | return 0; |
| 670 | BasicBlock *Dom = cast<BasicBlock>(*I); |
| 671 | while (++I != E) { |
| 672 | BasicBlock *B = cast_or_null<BasicBlock>(*I); |
| 673 | Dom = B ? DT->findNearestCommonDominator(Dom, B) : 0; |
| 674 | if (!Dom) |
| 675 | return 0; |
| 676 | } |
| 677 | DEBUG(dbgs() << "computed:" << Dom->getName() << '\n'); |
| 678 | return Dom; |
| 679 | } |
| 680 | |
| 681 | template <typename T> |
| 682 | BasicBlock *nearest_common_dominatee(DominatorTree *DT, T &Blocks) { |
| 683 | // If two blocks, A and B, dominate a block C, then A dominates B, |
| 684 | // or B dominates A. |
| 685 | typename T::iterator I = Blocks.begin(), E = Blocks.end(); |
| 686 | // Find the first non-null block. |
| 687 | while (I != E && !*I) |
| 688 | ++I; |
| 689 | if (I == E) |
| 690 | return DT->getRoot(); |
| 691 | BasicBlock *DomB = cast<BasicBlock>(*I); |
| 692 | while (++I != E) { |
| 693 | if (!*I) |
| 694 | continue; |
| 695 | BasicBlock *B = cast<BasicBlock>(*I); |
| 696 | if (DT->dominates(B, DomB)) |
| 697 | continue; |
| 698 | if (!DT->dominates(DomB, B)) |
| 699 | return 0; |
| 700 | DomB = B; |
| 701 | } |
| 702 | return DomB; |
| 703 | } |
| 704 | |
| 705 | // Find the first use in B of any value from Values. If no such use, |
| 706 | // return B->end(). |
| 707 | template <typename T> |
| 708 | BasicBlock::iterator first_use_of_in_block(T &Values, BasicBlock *B) { |
| 709 | BasicBlock::iterator FirstUse = B->end(), BEnd = B->end(); |
| 710 | typedef typename T::iterator iterator; |
| 711 | for (iterator I = Values.begin(), E = Values.end(); I != E; ++I) { |
| 712 | Value *V = *I; |
| 713 | // If V is used in a PHI node, the use belongs to the incoming block, |
| 714 | // not the block with the PHI node. In the incoming block, the use |
| 715 | // would be considered as being at the end of it, so it cannot |
| 716 | // influence the position of the first use (which is assumed to be |
| 717 | // at the end to start with). |
| 718 | if (isa<PHINode>(V)) |
| 719 | continue; |
| 720 | if (!isa<Instruction>(V)) |
| 721 | continue; |
| 722 | Instruction *In = cast<Instruction>(V); |
| 723 | if (In->getParent() != B) |
| 724 | continue; |
Duncan P. N. Exon Smith | a72c6e2 | 2015-10-20 00:46:39 +0000 | [diff] [blame] | 725 | BasicBlock::iterator It = In->getIterator(); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 726 | if (std::distance(FirstUse, BEnd) < std::distance(It, BEnd)) |
| 727 | FirstUse = It; |
| 728 | } |
| 729 | return FirstUse; |
| 730 | } |
| 731 | |
| 732 | bool is_empty(const BasicBlock *B) { |
| 733 | return B->empty() || (&*B->begin() == B->getTerminator()); |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | |
| 738 | BasicBlock *HexagonCommonGEP::recalculatePlacement(GepNode *Node, |
| 739 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 740 | DEBUG(dbgs() << "Loc for node:" << Node << '\n'); |
| 741 | // Recalculate the placement for Node, assuming that the locations of |
| 742 | // its children in Loc are valid. |
| 743 | // Return 0 if there is no valid placement for Node (for example, it |
| 744 | // uses an index value that is not available at the location required |
| 745 | // to dominate all children, etc.). |
| 746 | |
| 747 | // Find the nearest common dominator for: |
| 748 | // - all users, if the node is used, and |
| 749 | // - all children. |
| 750 | ValueVect Bs; |
| 751 | if (Node->Flags & GepNode::Used) { |
| 752 | // Append all blocks with uses of the original values to the |
| 753 | // block vector Bs. |
| 754 | NodeToUsesMap::iterator UF = Uses.find(Node); |
| 755 | assert(UF != Uses.end() && "Used node with no use information"); |
| 756 | UseSet &Us = UF->second; |
| 757 | for (UseSet::iterator I = Us.begin(), E = Us.end(); I != E; ++I) { |
| 758 | Use *U = *I; |
| 759 | User *R = U->getUser(); |
| 760 | if (!isa<Instruction>(R)) |
| 761 | continue; |
| 762 | BasicBlock *PB = isa<PHINode>(R) |
| 763 | ? cast<PHINode>(R)->getIncomingBlock(*U) |
| 764 | : cast<Instruction>(R)->getParent(); |
| 765 | Bs.push_back(PB); |
| 766 | } |
| 767 | } |
| 768 | // Append the location of each child. |
| 769 | NodeChildrenMap::iterator CF = NCM.find(Node); |
| 770 | if (CF != NCM.end()) { |
| 771 | NodeVect &Cs = CF->second; |
| 772 | for (NodeVect::iterator I = Cs.begin(), E = Cs.end(); I != E; ++I) { |
| 773 | GepNode *CN = *I; |
| 774 | NodeToValueMap::iterator LF = Loc.find(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(LF->second); |
| 781 | } |
| 782 | } |
| 783 | |
| 784 | BasicBlock *DomB = nearest_common_dominator(DT, Bs); |
| 785 | if (!DomB) |
| 786 | return 0; |
| 787 | // Check if the index used by Node dominates the computed dominator. |
| 788 | Instruction *IdxI = dyn_cast<Instruction>(Node->Idx); |
| 789 | if (IdxI && !DT->dominates(IdxI->getParent(), DomB)) |
| 790 | return 0; |
| 791 | |
| 792 | // Avoid putting nodes into empty blocks. |
| 793 | while (is_empty(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 | |
| 806 | BasicBlock *HexagonCommonGEP::recalculatePlacementRec(GepNode *Node, |
| 807 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 808 | DEBUG(dbgs() << "LocRec begin for node:" << Node << '\n'); |
| 809 | // Recalculate the placement of Node, after recursively recalculating the |
| 810 | // placements of all its children. |
| 811 | NodeChildrenMap::iterator CF = NCM.find(Node); |
| 812 | if (CF != NCM.end()) { |
| 813 | NodeVect &Cs = CF->second; |
| 814 | for (NodeVect::iterator I = Cs.begin(), E = Cs.end(); I != E; ++I) |
| 815 | recalculatePlacementRec(*I, NCM, Loc); |
| 816 | } |
| 817 | BasicBlock *LB = recalculatePlacement(Node, NCM, Loc); |
| 818 | DEBUG(dbgs() << "LocRec end for node:" << Node << '\n'); |
| 819 | return LB; |
| 820 | } |
| 821 | |
| 822 | |
| 823 | bool HexagonCommonGEP::isInvariantIn(Value *Val, Loop *L) { |
| 824 | if (isa<Constant>(Val) || isa<Argument>(Val)) |
| 825 | return true; |
| 826 | Instruction *In = dyn_cast<Instruction>(Val); |
| 827 | if (!In) |
| 828 | return false; |
| 829 | BasicBlock *HdrB = L->getHeader(), *DefB = In->getParent(); |
| 830 | return DT->properlyDominates(DefB, HdrB); |
| 831 | } |
| 832 | |
| 833 | |
| 834 | bool HexagonCommonGEP::isInvariantIn(GepNode *Node, Loop *L) { |
| 835 | if (Node->Flags & GepNode::Root) |
| 836 | if (!isInvariantIn(Node->BaseVal, L)) |
| 837 | return false; |
| 838 | return isInvariantIn(Node->Idx, L); |
| 839 | } |
| 840 | |
| 841 | |
| 842 | bool HexagonCommonGEP::isInMainPath(BasicBlock *B, Loop *L) { |
| 843 | BasicBlock *HB = L->getHeader(); |
| 844 | BasicBlock *LB = L->getLoopLatch(); |
| 845 | // B must post-dominate the loop header or dominate the loop latch. |
| 846 | if (PDT->dominates(B, HB)) |
| 847 | return true; |
| 848 | if (LB && DT->dominates(B, LB)) |
| 849 | return true; |
| 850 | return false; |
| 851 | } |
| 852 | |
| 853 | |
| 854 | namespace { |
| 855 | BasicBlock *preheader(DominatorTree *DT, Loop *L) { |
| 856 | if (BasicBlock *PH = L->getLoopPreheader()) |
| 857 | return PH; |
| 858 | if (!OptSpeculate) |
| 859 | return 0; |
| 860 | DomTreeNode *DN = DT->getNode(L->getHeader()); |
| 861 | if (!DN) |
| 862 | return 0; |
| 863 | return DN->getIDom()->getBlock(); |
| 864 | } |
| 865 | } |
| 866 | |
| 867 | |
| 868 | BasicBlock *HexagonCommonGEP::adjustForInvariance(GepNode *Node, |
| 869 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 870 | // Find the "topmost" location for Node: it must be dominated by both, |
| 871 | // its parent (or the BaseVal, if it's a root node), and by the index |
| 872 | // value. |
| 873 | ValueVect Bs; |
| 874 | if (Node->Flags & GepNode::Root) { |
| 875 | if (Instruction *PIn = dyn_cast<Instruction>(Node->BaseVal)) |
| 876 | Bs.push_back(PIn->getParent()); |
| 877 | } else { |
| 878 | Bs.push_back(Loc[Node->Parent]); |
| 879 | } |
| 880 | if (Instruction *IIn = dyn_cast<Instruction>(Node->Idx)) |
| 881 | Bs.push_back(IIn->getParent()); |
| 882 | BasicBlock *TopB = nearest_common_dominatee(DT, Bs); |
| 883 | |
| 884 | // Traverse the loop nest upwards until we find a loop in which Node |
| 885 | // is no longer invariant, or until we get to the upper limit of Node's |
| 886 | // placement. The traversal will also stop when a suitable "preheader" |
| 887 | // cannot be found for a given loop. The "preheader" may actually be |
| 888 | // a regular block outside of the loop (i.e. not guarded), in which case |
| 889 | // the Node will be speculated. |
| 890 | // For nodes that are not in the main path of the containing loop (i.e. |
| 891 | // are not executed in each iteration), do not move them out of the loop. |
| 892 | BasicBlock *LocB = cast_or_null<BasicBlock>(Loc[Node]); |
| 893 | if (LocB) { |
| 894 | Loop *Lp = LI->getLoopFor(LocB); |
| 895 | while (Lp) { |
| 896 | if (!isInvariantIn(Node, Lp) || !isInMainPath(LocB, Lp)) |
| 897 | break; |
| 898 | BasicBlock *NewLoc = preheader(DT, Lp); |
| 899 | if (!NewLoc || !DT->dominates(TopB, NewLoc)) |
| 900 | break; |
| 901 | Lp = Lp->getParentLoop(); |
| 902 | LocB = NewLoc; |
| 903 | } |
| 904 | } |
| 905 | Loc[Node] = LocB; |
| 906 | |
| 907 | // Recursively compute the locations of all children nodes. |
| 908 | NodeChildrenMap::iterator CF = NCM.find(Node); |
| 909 | if (CF != NCM.end()) { |
| 910 | NodeVect &Cs = CF->second; |
| 911 | for (NodeVect::iterator I = Cs.begin(), E = Cs.end(); I != E; ++I) |
| 912 | adjustForInvariance(*I, NCM, Loc); |
| 913 | } |
| 914 | return LocB; |
| 915 | } |
| 916 | |
| 917 | |
| 918 | namespace { |
| 919 | struct LocationAsBlock { |
| 920 | LocationAsBlock(const NodeToValueMap &L) : Map(L) {} |
| 921 | const NodeToValueMap ⤅ |
| 922 | }; |
| 923 | |
| 924 | raw_ostream &operator<< (raw_ostream &OS, |
| 925 | const LocationAsBlock &Loc) LLVM_ATTRIBUTE_UNUSED ; |
| 926 | raw_ostream &operator<< (raw_ostream &OS, const LocationAsBlock &Loc) { |
| 927 | for (NodeToValueMap::const_iterator I = Loc.Map.begin(), E = Loc.Map.end(); |
| 928 | I != E; ++I) { |
| 929 | OS << I->first << " -> "; |
| 930 | BasicBlock *B = cast<BasicBlock>(I->second); |
| 931 | OS << B->getName() << '(' << B << ')'; |
| 932 | OS << '\n'; |
| 933 | } |
| 934 | return OS; |
| 935 | } |
| 936 | |
| 937 | inline bool is_constant(GepNode *N) { |
| 938 | return isa<ConstantInt>(N->Idx); |
| 939 | } |
| 940 | } |
| 941 | |
| 942 | |
| 943 | void HexagonCommonGEP::separateChainForNode(GepNode *Node, Use *U, |
| 944 | NodeToValueMap &Loc) { |
| 945 | User *R = U->getUser(); |
| 946 | DEBUG(dbgs() << "Separating chain for node (" << Node << ") user: " |
| 947 | << *R << '\n'); |
| 948 | BasicBlock *PB = cast<Instruction>(R)->getParent(); |
| 949 | |
| 950 | GepNode *N = Node; |
| 951 | GepNode *C = 0, *NewNode = 0; |
| 952 | while (is_constant(N) && !(N->Flags & GepNode::Root)) { |
| 953 | // XXX if (single-use) dont-replicate; |
| 954 | GepNode *NewN = new (*Mem) GepNode(N); |
| 955 | Nodes.push_back(NewN); |
| 956 | Loc[NewN] = PB; |
| 957 | |
| 958 | if (N == Node) |
| 959 | NewNode = NewN; |
| 960 | NewN->Flags &= ~GepNode::Used; |
| 961 | if (C) |
| 962 | C->Parent = NewN; |
| 963 | C = NewN; |
| 964 | N = N->Parent; |
| 965 | } |
| 966 | if (!NewNode) |
| 967 | return; |
| 968 | |
| 969 | // Move over all uses that share the same user as U from Node to NewNode. |
| 970 | NodeToUsesMap::iterator UF = Uses.find(Node); |
| 971 | assert(UF != Uses.end()); |
| 972 | UseSet &Us = UF->second; |
| 973 | UseSet NewUs; |
| 974 | for (UseSet::iterator I = Us.begin(); I != Us.end(); ) { |
| 975 | User *S = (*I)->getUser(); |
| 976 | UseSet::iterator Nx = std::next(I); |
| 977 | if (S == R) { |
| 978 | NewUs.insert(*I); |
| 979 | Us.erase(I); |
| 980 | } |
| 981 | I = Nx; |
| 982 | } |
| 983 | if (Us.empty()) { |
| 984 | Node->Flags &= ~GepNode::Used; |
| 985 | Uses.erase(UF); |
| 986 | } |
| 987 | |
| 988 | // Should at least have U in NewUs. |
| 989 | NewNode->Flags |= GepNode::Used; |
| 990 | DEBUG(dbgs() << "new node: " << NewNode << " " << *NewNode << '\n'); |
| 991 | assert(!NewUs.empty()); |
| 992 | Uses[NewNode] = NewUs; |
| 993 | } |
| 994 | |
| 995 | |
| 996 | void HexagonCommonGEP::separateConstantChains(GepNode *Node, |
| 997 | NodeChildrenMap &NCM, NodeToValueMap &Loc) { |
| 998 | // First approximation: extract all chains. |
| 999 | NodeSet Ns; |
| 1000 | nodes_for_root(Node, NCM, Ns); |
| 1001 | |
| 1002 | DEBUG(dbgs() << "Separating constant chains for node: " << Node << '\n'); |
| 1003 | // Collect all used nodes together with the uses from loads and stores, |
| 1004 | // where the GEP node could be folded into the load/store instruction. |
| 1005 | NodeToUsesMap FNs; // Foldable nodes. |
| 1006 | for (NodeSet::iterator I = Ns.begin(), E = Ns.end(); I != E; ++I) { |
| 1007 | GepNode *N = *I; |
| 1008 | if (!(N->Flags & GepNode::Used)) |
| 1009 | continue; |
| 1010 | NodeToUsesMap::iterator UF = Uses.find(N); |
| 1011 | assert(UF != Uses.end()); |
| 1012 | UseSet &Us = UF->second; |
| 1013 | // Loads/stores that use the node N. |
| 1014 | UseSet LSs; |
| 1015 | for (UseSet::iterator J = Us.begin(), F = Us.end(); J != F; ++J) { |
| 1016 | Use *U = *J; |
| 1017 | User *R = U->getUser(); |
| 1018 | // We're interested in uses that provide the address. It can happen |
| 1019 | // that the value may also be provided via GEP, but we won't handle |
| 1020 | // those cases here for now. |
| 1021 | if (LoadInst *Ld = dyn_cast<LoadInst>(R)) { |
| 1022 | unsigned PtrX = LoadInst::getPointerOperandIndex(); |
| 1023 | if (&Ld->getOperandUse(PtrX) == U) |
| 1024 | LSs.insert(U); |
| 1025 | } else if (StoreInst *St = dyn_cast<StoreInst>(R)) { |
| 1026 | unsigned PtrX = StoreInst::getPointerOperandIndex(); |
| 1027 | if (&St->getOperandUse(PtrX) == U) |
| 1028 | LSs.insert(U); |
| 1029 | } |
| 1030 | } |
| 1031 | // Even if the total use count is 1, separating the chain may still be |
| 1032 | // beneficial, since the constant chain may be longer than the GEP alone |
| 1033 | // would be (e.g. if the parent node has a constant index and also has |
| 1034 | // other children). |
| 1035 | if (!LSs.empty()) |
| 1036 | FNs.insert(std::make_pair(N, LSs)); |
| 1037 | } |
| 1038 | |
| 1039 | DEBUG(dbgs() << "Nodes with foldable users:\n" << FNs); |
| 1040 | |
| 1041 | for (NodeToUsesMap::iterator I = FNs.begin(), E = FNs.end(); I != E; ++I) { |
| 1042 | GepNode *N = I->first; |
| 1043 | UseSet &Us = I->second; |
| 1044 | for (UseSet::iterator J = Us.begin(), F = Us.end(); J != F; ++J) |
| 1045 | separateChainForNode(N, *J, Loc); |
| 1046 | } |
| 1047 | } |
| 1048 | |
| 1049 | |
| 1050 | void HexagonCommonGEP::computeNodePlacement(NodeToValueMap &Loc) { |
| 1051 | // Compute the inverse of the Node.Parent links. Also, collect the set |
| 1052 | // of root nodes. |
| 1053 | NodeChildrenMap NCM; |
| 1054 | NodeVect Roots; |
| 1055 | invert_find_roots(Nodes, NCM, Roots); |
| 1056 | |
| 1057 | // Compute the initial placement determined by the users' locations, and |
| 1058 | // the locations of the child nodes. |
| 1059 | for (NodeVect::iterator I = Roots.begin(), E = Roots.end(); I != E; ++I) |
| 1060 | recalculatePlacementRec(*I, NCM, Loc); |
| 1061 | |
| 1062 | DEBUG(dbgs() << "Initial node placement:\n" << LocationAsBlock(Loc)); |
| 1063 | |
| 1064 | if (OptEnableInv) { |
| 1065 | for (NodeVect::iterator I = Roots.begin(), E = Roots.end(); I != E; ++I) |
| 1066 | adjustForInvariance(*I, NCM, Loc); |
| 1067 | |
| 1068 | DEBUG(dbgs() << "Node placement after adjustment for invariance:\n" |
| 1069 | << LocationAsBlock(Loc)); |
| 1070 | } |
| 1071 | if (OptEnableConst) { |
| 1072 | for (NodeVect::iterator I = Roots.begin(), E = Roots.end(); I != E; ++I) |
| 1073 | separateConstantChains(*I, NCM, Loc); |
| 1074 | } |
| 1075 | DEBUG(dbgs() << "Node use information:\n" << Uses); |
| 1076 | |
| 1077 | // At the moment, there is no further refinement of the initial placement. |
| 1078 | // Such a refinement could include splitting the nodes if they are placed |
| 1079 | // too far from some of its users. |
| 1080 | |
| 1081 | DEBUG(dbgs() << "Final node placement:\n" << LocationAsBlock(Loc)); |
| 1082 | } |
| 1083 | |
| 1084 | |
| 1085 | Value *HexagonCommonGEP::fabricateGEP(NodeVect &NA, BasicBlock::iterator At, |
| 1086 | BasicBlock *LocB) { |
| 1087 | DEBUG(dbgs() << "Fabricating GEP in " << LocB->getName() |
| 1088 | << " for nodes:\n" << NA); |
| 1089 | unsigned Num = NA.size(); |
| 1090 | GepNode *RN = NA[0]; |
| 1091 | assert((RN->Flags & GepNode::Root) && "Creating GEP for non-root"); |
| 1092 | |
| 1093 | Value *NewInst = 0; |
| 1094 | Value *Input = RN->BaseVal; |
| 1095 | Value **IdxList = new Value*[Num+1]; |
| 1096 | unsigned nax = 0; |
| 1097 | do { |
| 1098 | unsigned IdxC = 0; |
| 1099 | // If the type of the input of the first node is not a pointer, |
| 1100 | // we need to add an artificial i32 0 to the indices (because the |
| 1101 | // actual input in the IR will be a pointer). |
| 1102 | if (!NA[nax]->PTy->isPointerTy()) { |
| 1103 | Type *Int32Ty = Type::getInt32Ty(*Ctx); |
| 1104 | IdxList[IdxC++] = ConstantInt::get(Int32Ty, 0); |
| 1105 | } |
| 1106 | |
| 1107 | // Keep adding indices from NA until we have to stop and generate |
| 1108 | // an "intermediate" GEP. |
| 1109 | while (++nax <= Num) { |
| 1110 | GepNode *N = NA[nax-1]; |
| 1111 | IdxList[IdxC++] = N->Idx; |
| 1112 | if (nax < Num) { |
| 1113 | // We have to stop, if the expected type of the output of this node |
| 1114 | // is not the same as the input type of the next node. |
| 1115 | Type *NextTy = next_type(N->PTy, N->Idx); |
| 1116 | if (NextTy != NA[nax]->PTy) |
| 1117 | break; |
| 1118 | } |
| 1119 | } |
| 1120 | ArrayRef<Value*> A(IdxList, IdxC); |
| 1121 | Type *InpTy = Input->getType(); |
| 1122 | Type *ElTy = cast<PointerType>(InpTy->getScalarType())->getElementType(); |
Duncan P. N. Exon Smith | a72c6e2 | 2015-10-20 00:46:39 +0000 | [diff] [blame] | 1123 | NewInst = GetElementPtrInst::Create(ElTy, Input, A, "cgep", &*At); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 1124 | DEBUG(dbgs() << "new GEP: " << *NewInst << '\n'); |
| 1125 | Input = NewInst; |
| 1126 | } while (nax <= Num); |
| 1127 | |
| 1128 | delete[] IdxList; |
| 1129 | return NewInst; |
| 1130 | } |
| 1131 | |
| 1132 | |
| 1133 | void HexagonCommonGEP::getAllUsersForNode(GepNode *Node, ValueVect &Values, |
| 1134 | NodeChildrenMap &NCM) { |
| 1135 | NodeVect Work; |
| 1136 | Work.push_back(Node); |
| 1137 | |
| 1138 | while (!Work.empty()) { |
| 1139 | NodeVect::iterator First = Work.begin(); |
| 1140 | GepNode *N = *First; |
| 1141 | Work.erase(First); |
| 1142 | if (N->Flags & GepNode::Used) { |
| 1143 | NodeToUsesMap::iterator UF = Uses.find(N); |
| 1144 | assert(UF != Uses.end() && "No use information for used node"); |
| 1145 | UseSet &Us = UF->second; |
| 1146 | for (UseSet::iterator I = Us.begin(), E = Us.end(); I != E; ++I) |
| 1147 | Values.push_back((*I)->getUser()); |
| 1148 | } |
| 1149 | NodeChildrenMap::iterator CF = NCM.find(N); |
| 1150 | if (CF != NCM.end()) { |
| 1151 | NodeVect &Cs = CF->second; |
| 1152 | Work.insert(Work.end(), Cs.begin(), Cs.end()); |
| 1153 | } |
| 1154 | } |
| 1155 | } |
| 1156 | |
| 1157 | |
| 1158 | void HexagonCommonGEP::materialize(NodeToValueMap &Loc) { |
| 1159 | DEBUG(dbgs() << "Nodes before materialization:\n" << Nodes << '\n'); |
| 1160 | NodeChildrenMap NCM; |
| 1161 | NodeVect Roots; |
| 1162 | // Compute the inversion again, since computing placement could alter |
| 1163 | // "parent" relation between nodes. |
| 1164 | invert_find_roots(Nodes, NCM, Roots); |
| 1165 | |
| 1166 | while (!Roots.empty()) { |
| 1167 | NodeVect::iterator First = Roots.begin(); |
| 1168 | GepNode *Root = *First, *Last = *First; |
| 1169 | Roots.erase(First); |
| 1170 | |
| 1171 | NodeVect NA; // Nodes to assemble. |
| 1172 | // Append to NA all child nodes up to (and including) the first child |
| 1173 | // that: |
| 1174 | // (1) has more than 1 child, or |
| 1175 | // (2) is used, or |
| 1176 | // (3) has a child located in a different block. |
| 1177 | bool LastUsed = false; |
| 1178 | unsigned LastCN = 0; |
| 1179 | // The location may be null if the computation failed (it can legitimately |
| 1180 | // happen for nodes created from dead GEPs). |
| 1181 | Value *LocV = Loc[Last]; |
| 1182 | if (!LocV) |
| 1183 | continue; |
| 1184 | BasicBlock *LastB = cast<BasicBlock>(LocV); |
| 1185 | do { |
| 1186 | NA.push_back(Last); |
| 1187 | LastUsed = (Last->Flags & GepNode::Used); |
| 1188 | if (LastUsed) |
| 1189 | break; |
| 1190 | NodeChildrenMap::iterator CF = NCM.find(Last); |
| 1191 | LastCN = (CF != NCM.end()) ? CF->second.size() : 0; |
| 1192 | if (LastCN != 1) |
| 1193 | break; |
| 1194 | GepNode *Child = CF->second.front(); |
| 1195 | BasicBlock *ChildB = cast_or_null<BasicBlock>(Loc[Child]); |
| 1196 | if (ChildB != 0 && LastB != ChildB) |
| 1197 | break; |
| 1198 | Last = Child; |
| 1199 | } while (true); |
| 1200 | |
Duncan P. N. Exon Smith | a72c6e2 | 2015-10-20 00:46:39 +0000 | [diff] [blame] | 1201 | BasicBlock::iterator InsertAt = LastB->getTerminator()->getIterator(); |
Krzysztof Parzyszek | 79b2433 | 2015-07-08 19:22:28 +0000 | [diff] [blame] | 1202 | if (LastUsed || LastCN > 0) { |
| 1203 | ValueVect Urs; |
| 1204 | getAllUsersForNode(Root, Urs, NCM); |
| 1205 | BasicBlock::iterator FirstUse = first_use_of_in_block(Urs, LastB); |
| 1206 | if (FirstUse != LastB->end()) |
| 1207 | InsertAt = FirstUse; |
| 1208 | } |
| 1209 | |
| 1210 | // Generate a new instruction for NA. |
| 1211 | Value *NewInst = fabricateGEP(NA, InsertAt, LastB); |
| 1212 | |
| 1213 | // Convert all the children of Last node into roots, and append them |
| 1214 | // to the Roots list. |
| 1215 | if (LastCN > 0) { |
| 1216 | NodeVect &Cs = NCM[Last]; |
| 1217 | for (NodeVect::iterator I = Cs.begin(), E = Cs.end(); I != E; ++I) { |
| 1218 | GepNode *CN = *I; |
| 1219 | CN->Flags &= ~GepNode::Internal; |
| 1220 | CN->Flags |= GepNode::Root; |
| 1221 | CN->BaseVal = NewInst; |
| 1222 | Roots.push_back(CN); |
| 1223 | } |
| 1224 | } |
| 1225 | |
| 1226 | // Lastly, if the Last node was used, replace all uses with the new GEP. |
| 1227 | // The uses reference the original GEP values. |
| 1228 | if (LastUsed) { |
| 1229 | NodeToUsesMap::iterator UF = Uses.find(Last); |
| 1230 | assert(UF != Uses.end() && "No use information found"); |
| 1231 | UseSet &Us = UF->second; |
| 1232 | for (UseSet::iterator I = Us.begin(), E = Us.end(); I != E; ++I) { |
| 1233 | Use *U = *I; |
| 1234 | U->set(NewInst); |
| 1235 | } |
| 1236 | } |
| 1237 | } |
| 1238 | } |
| 1239 | |
| 1240 | |
| 1241 | void HexagonCommonGEP::removeDeadCode() { |
| 1242 | ValueVect BO; |
| 1243 | BO.push_back(&Fn->front()); |
| 1244 | |
| 1245 | for (unsigned i = 0; i < BO.size(); ++i) { |
| 1246 | BasicBlock *B = cast<BasicBlock>(BO[i]); |
| 1247 | DomTreeNode *N = DT->getNode(B); |
| 1248 | typedef GraphTraits<DomTreeNode*> GTN; |
| 1249 | typedef GTN::ChildIteratorType Iter; |
| 1250 | for (Iter I = GTN::child_begin(N), E = GTN::child_end(N); I != E; ++I) |
| 1251 | BO.push_back((*I)->getBlock()); |
| 1252 | } |
| 1253 | |
| 1254 | for (unsigned i = BO.size(); i > 0; --i) { |
| 1255 | BasicBlock *B = cast<BasicBlock>(BO[i-1]); |
| 1256 | BasicBlock::InstListType &IL = B->getInstList(); |
| 1257 | typedef BasicBlock::InstListType::reverse_iterator reverse_iterator; |
| 1258 | ValueVect Ins; |
| 1259 | for (reverse_iterator I = IL.rbegin(), E = IL.rend(); I != E; ++I) |
| 1260 | Ins.push_back(&*I); |
| 1261 | for (ValueVect::iterator I = Ins.begin(), E = Ins.end(); I != E; ++I) { |
| 1262 | Instruction *In = cast<Instruction>(*I); |
| 1263 | if (isInstructionTriviallyDead(In)) |
| 1264 | In->eraseFromParent(); |
| 1265 | } |
| 1266 | } |
| 1267 | } |
| 1268 | |
| 1269 | |
| 1270 | bool HexagonCommonGEP::runOnFunction(Function &F) { |
| 1271 | // For now bail out on C++ exception handling. |
| 1272 | for (Function::iterator A = F.begin(), Z = F.end(); A != Z; ++A) |
| 1273 | for (BasicBlock::iterator I = A->begin(), E = A->end(); I != E; ++I) |
| 1274 | if (isa<InvokeInst>(I) || isa<LandingPadInst>(I)) |
| 1275 | return false; |
| 1276 | |
| 1277 | Fn = &F; |
| 1278 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 1279 | PDT = &getAnalysis<PostDominatorTree>(); |
| 1280 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
| 1281 | Ctx = &F.getContext(); |
| 1282 | |
| 1283 | Nodes.clear(); |
| 1284 | Uses.clear(); |
| 1285 | NodeOrder.clear(); |
| 1286 | |
| 1287 | SpecificBumpPtrAllocator<GepNode> Allocator; |
| 1288 | Mem = &Allocator; |
| 1289 | |
| 1290 | collect(); |
| 1291 | common(); |
| 1292 | |
| 1293 | NodeToValueMap Loc; |
| 1294 | computeNodePlacement(Loc); |
| 1295 | materialize(Loc); |
| 1296 | removeDeadCode(); |
| 1297 | |
| 1298 | #ifdef XDEBUG |
| 1299 | // Run this only when expensive checks are enabled. |
| 1300 | verifyFunction(F); |
| 1301 | #endif |
| 1302 | return true; |
| 1303 | } |
| 1304 | |
| 1305 | |
| 1306 | namespace llvm { |
| 1307 | FunctionPass *createHexagonCommonGEP() { |
| 1308 | return new HexagonCommonGEP(); |
| 1309 | } |
| 1310 | } |