1 //===- SimplifyCFGPass.cpp - CFG Simplification Pass ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements dead code elimination and basic block merging, along 10 // with a collection of other peephole control flow optimizations. For example: 11 // 12 // * Removes basic blocks with no predecessors. 13 // * Merges a basic block into its predecessor if there is only one and the 14 // predecessor only has one successor. 15 // * Eliminates PHI nodes for basic blocks with a single predecessor. 16 // * Eliminates a basic block that only contains an unconditional branch. 17 // * Changes invoke instructions to nounwind functions to be calls. 18 // * Change things like "if (x) if (y)" into "if (x&y)". 19 // * etc.. 20 // 21 //===----------------------------------------------------------------------===// 22 23 #include "llvm/ADT/MapVector.h" 24 #include "llvm/ADT/SmallPtrSet.h" 25 #include "llvm/ADT/SmallVector.h" 26 #include "llvm/ADT/Statistic.h" 27 #include "llvm/Analysis/AssumptionCache.h" 28 #include "llvm/Analysis/CFG.h" 29 #include "llvm/Analysis/DomTreeUpdater.h" 30 #include "llvm/Analysis/GlobalsModRef.h" 31 #include "llvm/Analysis/TargetTransformInfo.h" 32 #include "llvm/IR/Attributes.h" 33 #include "llvm/IR/CFG.h" 34 #include "llvm/IR/Constants.h" 35 #include "llvm/IR/DataLayout.h" 36 #include "llvm/IR/Dominators.h" 37 #include "llvm/IR/Instructions.h" 38 #include "llvm/IR/IntrinsicInst.h" 39 #include "llvm/IR/Module.h" 40 #include "llvm/IR/ValueHandle.h" 41 #include "llvm/InitializePasses.h" 42 #include "llvm/Pass.h" 43 #include "llvm/Support/CommandLine.h" 44 #include "llvm/Transforms/Scalar.h" 45 #include "llvm/Transforms/Scalar/SimplifyCFG.h" 46 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 47 #include "llvm/Transforms/Utils/Local.h" 48 #include "llvm/Transforms/Utils/SimplifyCFGOptions.h" 49 #include <utility> 50 using namespace llvm; 51 52 #define DEBUG_TYPE "simplifycfg" 53 54 static cl::opt<unsigned> UserBonusInstThreshold( 55 "bonus-inst-threshold", cl::Hidden, cl::init(1), 56 cl::desc("Control the number of bonus instructions (default = 1)")); 57 58 static cl::opt<bool> UserKeepLoops( 59 "keep-loops", cl::Hidden, cl::init(true), 60 cl::desc("Preserve canonical loop structure (default = true)")); 61 62 static cl::opt<bool> UserSwitchToLookup( 63 "switch-to-lookup", cl::Hidden, cl::init(false), 64 cl::desc("Convert switches to lookup tables (default = false)")); 65 66 static cl::opt<bool> UserForwardSwitchCond( 67 "forward-switch-cond", cl::Hidden, cl::init(false), 68 cl::desc("Forward switch condition to phi ops (default = false)")); 69 70 static cl::opt<bool> UserHoistCommonInsts( 71 "hoist-common-insts", cl::Hidden, cl::init(false), 72 cl::desc("hoist common instructions (default = false)")); 73 74 static cl::opt<bool> UserSinkCommonInsts( 75 "sink-common-insts", cl::Hidden, cl::init(false), 76 cl::desc("Sink common instructions (default = false)")); 77 78 79 STATISTIC(NumSimpl, "Number of blocks simplified"); 80 81 static bool tailMergeBlocksWithSimilarFunctionTerminators(Function &F, 82 DomTreeUpdater *DTU) { 83 SmallMapVector<unsigned /*TerminatorOpcode*/, SmallVector<BasicBlock *, 2>, 4> 84 Structure; 85 86 // Scan all the blocks in the function, record the interesting-ones. 87 for (BasicBlock &BB : F) { 88 if (DTU && DTU->isBBPendingDeletion(&BB)) 89 continue; 90 91 // We are only interested in function-terminating blocks. 92 if (!succ_empty(&BB)) 93 continue; 94 95 auto *Term = BB.getTerminator(); 96 97 // Fow now only support `ret`/`resume` function terminators. 98 // FIXME: lift this restriction. 99 switch (Term->getOpcode()) { 100 case Instruction::Ret: 101 case Instruction::Resume: 102 break; 103 default: 104 continue; 105 } 106 107 // We can't tail-merge block that contains a musttail call. 108 if (BB.getTerminatingMustTailCall()) 109 continue; 110 111 // Calls to experimental_deoptimize must be followed by a return 112 // of the value computed by experimental_deoptimize. 113 // I.e., we can not change `ret` to `br` for this block. 114 if (auto *CI = 115 dyn_cast_or_null<CallInst>(Term->getPrevNonDebugInstruction())) { 116 if (Function *F = CI->getCalledFunction()) 117 if (Intrinsic::ID ID = F->getIntrinsicID()) 118 if (ID == Intrinsic::experimental_deoptimize) 119 continue; 120 } 121 122 // PHI nodes cannot have token type, so if the terminator has an operand 123 // with token type, we can not tail-merge this kind of function terminators. 124 if (any_of(Term->operands(), 125 [](Value *Op) { return Op->getType()->isTokenTy(); })) 126 continue; 127 128 // Canonical blocks are uniqued based on the terminator type (opcode). 129 Structure[Term->getOpcode()].emplace_back(&BB); 130 } 131 132 bool Changed = false; 133 134 std::vector<DominatorTree::UpdateType> Updates; 135 136 for (ArrayRef<BasicBlock *> BBs : make_second_range(Structure)) { 137 SmallVector<PHINode *, 1> NewOps; 138 139 // We don't want to change IR just because we can. 140 // Only do that if there are at least two blocks we'll tail-merge. 141 if (BBs.size() < 2) 142 continue; 143 144 Changed = true; 145 146 if (DTU) 147 Updates.reserve(Updates.size() + BBs.size()); 148 149 BasicBlock *CanonicalBB; 150 Instruction *CanonicalTerm; 151 { 152 auto *Term = BBs[0]->getTerminator(); 153 154 // Create a canonical block for this function terminator type now, 155 // placing it *before* the first block that will branch to it. 156 CanonicalBB = BasicBlock::Create( 157 F.getContext(), Twine("common.") + Term->getOpcodeName(), &F, BBs[0]); 158 // We'll also need a PHI node per each operand of the terminator. 159 NewOps.resize(Term->getNumOperands()); 160 for (auto I : zip(Term->operands(), NewOps)) { 161 std::get<1>(I) = PHINode::Create(std::get<0>(I)->getType(), 162 /*NumReservedValues=*/BBs.size(), 163 CanonicalBB->getName() + ".op"); 164 CanonicalBB->getInstList().push_back(std::get<1>(I)); 165 } 166 // Make it so that this canonical block actually has the right 167 // terminator. 168 CanonicalTerm = Term->clone(); 169 CanonicalBB->getInstList().push_back(CanonicalTerm); 170 // If the canonical terminator has operands, rewrite it to take PHI's. 171 for (auto I : zip(NewOps, CanonicalTerm->operands())) 172 std::get<1>(I) = std::get<0>(I); 173 } 174 175 // Now, go through each block (with the current terminator type) 176 // we've recorded, and rewrite it to branch to the new common block. 177 const DILocation *CommonDebugLoc = nullptr; 178 for (BasicBlock *BB : BBs) { 179 auto *Term = BB->getTerminator(); 180 181 // Aha, found a new non-canonical function terminator. If it has operands, 182 // forward them to the PHI nodes in the canonical block. 183 for (auto I : zip(Term->operands(), NewOps)) 184 std::get<1>(I)->addIncoming(std::get<0>(I), BB); 185 186 // Compute the debug location common to all the original terminators. 187 if (!CommonDebugLoc) 188 CommonDebugLoc = Term->getDebugLoc(); 189 else 190 CommonDebugLoc = 191 DILocation::getMergedLocation(CommonDebugLoc, Term->getDebugLoc()); 192 193 // And turn BB into a block that just unconditionally branches 194 // to the canonical block. 195 Term->eraseFromParent(); 196 BranchInst::Create(CanonicalBB, BB); 197 if (DTU) 198 Updates.push_back({DominatorTree::Insert, BB, CanonicalBB}); 199 } 200 201 CanonicalTerm->setDebugLoc(CommonDebugLoc); 202 } 203 204 if (DTU) 205 DTU->applyUpdates(Updates); 206 207 return Changed; 208 } 209 210 /// Call SimplifyCFG on all the blocks in the function, 211 /// iterating until no more changes are made. 212 static bool iterativelySimplifyCFG(Function &F, const TargetTransformInfo &TTI, 213 DomTreeUpdater *DTU, 214 const SimplifyCFGOptions &Options) { 215 bool Changed = false; 216 bool LocalChange = true; 217 218 SmallVector<std::pair<const BasicBlock *, const BasicBlock *>, 32> Edges; 219 FindFunctionBackedges(F, Edges); 220 SmallPtrSet<BasicBlock *, 16> UniqueLoopHeaders; 221 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 222 UniqueLoopHeaders.insert(const_cast<BasicBlock *>(Edges[i].second)); 223 224 SmallVector<WeakVH, 16> LoopHeaders(UniqueLoopHeaders.begin(), 225 UniqueLoopHeaders.end()); 226 227 unsigned IterCnt = 0; 228 (void)IterCnt; 229 while (LocalChange) { 230 assert(IterCnt++ < 1000 && 231 "Sanity: iterative simplification didn't converge!"); 232 LocalChange = false; 233 234 // Loop over all of the basic blocks and remove them if they are unneeded. 235 for (Function::iterator BBIt = F.begin(); BBIt != F.end(); ) { 236 BasicBlock &BB = *BBIt++; 237 if (DTU) { 238 assert( 239 !DTU->isBBPendingDeletion(&BB) && 240 "Should not end up trying to simplify blocks marked for removal."); 241 // Make sure that the advanced iterator does not point at the blocks 242 // that are marked for removal, skip over all such blocks. 243 while (BBIt != F.end() && DTU->isBBPendingDeletion(&*BBIt)) 244 ++BBIt; 245 } 246 if (simplifyCFG(&BB, TTI, DTU, Options, LoopHeaders)) { 247 LocalChange = true; 248 ++NumSimpl; 249 } 250 } 251 Changed |= LocalChange; 252 } 253 return Changed; 254 } 255 256 static bool simplifyFunctionCFGImpl(Function &F, const TargetTransformInfo &TTI, 257 DominatorTree *DT, 258 const SimplifyCFGOptions &Options) { 259 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 260 261 bool EverChanged = removeUnreachableBlocks(F, DT ? &DTU : nullptr); 262 EverChanged |= 263 tailMergeBlocksWithSimilarFunctionTerminators(F, DT ? &DTU : nullptr); 264 EverChanged |= iterativelySimplifyCFG(F, TTI, DT ? &DTU : nullptr, Options); 265 266 // If neither pass changed anything, we're done. 267 if (!EverChanged) return false; 268 269 // iterativelySimplifyCFG can (rarely) make some loops dead. If this happens, 270 // removeUnreachableBlocks is needed to nuke them, which means we should 271 // iterate between the two optimizations. We structure the code like this to 272 // avoid rerunning iterativelySimplifyCFG if the second pass of 273 // removeUnreachableBlocks doesn't do anything. 274 if (!removeUnreachableBlocks(F, DT ? &DTU : nullptr)) 275 return true; 276 277 do { 278 EverChanged = iterativelySimplifyCFG(F, TTI, DT ? &DTU : nullptr, Options); 279 EverChanged |= removeUnreachableBlocks(F, DT ? &DTU : nullptr); 280 } while (EverChanged); 281 282 return true; 283 } 284 285 static bool simplifyFunctionCFG(Function &F, const TargetTransformInfo &TTI, 286 DominatorTree *DT, 287 const SimplifyCFGOptions &Options) { 288 assert((!RequireAndPreserveDomTree || 289 (DT && DT->verify(DominatorTree::VerificationLevel::Full))) && 290 "Original domtree is invalid?"); 291 292 bool Changed = simplifyFunctionCFGImpl(F, TTI, DT, Options); 293 294 assert((!RequireAndPreserveDomTree || 295 (DT && DT->verify(DominatorTree::VerificationLevel::Full))) && 296 "Failed to maintain validity of domtree!"); 297 298 return Changed; 299 } 300 301 // Command-line settings override compile-time settings. 302 static void applyCommandLineOverridesToOptions(SimplifyCFGOptions &Options) { 303 if (UserBonusInstThreshold.getNumOccurrences()) 304 Options.BonusInstThreshold = UserBonusInstThreshold; 305 if (UserForwardSwitchCond.getNumOccurrences()) 306 Options.ForwardSwitchCondToPhi = UserForwardSwitchCond; 307 if (UserSwitchToLookup.getNumOccurrences()) 308 Options.ConvertSwitchToLookupTable = UserSwitchToLookup; 309 if (UserKeepLoops.getNumOccurrences()) 310 Options.NeedCanonicalLoop = UserKeepLoops; 311 if (UserHoistCommonInsts.getNumOccurrences()) 312 Options.HoistCommonInsts = UserHoistCommonInsts; 313 if (UserSinkCommonInsts.getNumOccurrences()) 314 Options.SinkCommonInsts = UserSinkCommonInsts; 315 } 316 317 SimplifyCFGPass::SimplifyCFGPass() : Options() { 318 applyCommandLineOverridesToOptions(Options); 319 } 320 321 SimplifyCFGPass::SimplifyCFGPass(const SimplifyCFGOptions &Opts) 322 : Options(Opts) { 323 applyCommandLineOverridesToOptions(Options); 324 } 325 326 void SimplifyCFGPass::printPipeline( 327 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 328 static_cast<PassInfoMixin<SimplifyCFGPass> *>(this)->printPipeline( 329 OS, MapClassName2PassName); 330 OS << "<"; 331 OS << "bonus-inst-threshold=" << Options.BonusInstThreshold << ";"; 332 OS << (Options.ForwardSwitchCondToPhi ? "" : "no-") << "forward-switch-cond;"; 333 OS << (Options.ConvertSwitchToLookupTable ? "" : "no-") 334 << "switch-to-lookup;"; 335 OS << (Options.NeedCanonicalLoop ? "" : "no-") << "keep-loops;"; 336 OS << (Options.HoistCommonInsts ? "" : "no-") << "hoist-common-insts;"; 337 OS << (Options.SinkCommonInsts ? "" : "no-") << "sink-common-insts"; 338 OS << ">"; 339 } 340 341 PreservedAnalyses SimplifyCFGPass::run(Function &F, 342 FunctionAnalysisManager &AM) { 343 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 344 Options.AC = &AM.getResult<AssumptionAnalysis>(F); 345 DominatorTree *DT = nullptr; 346 if (RequireAndPreserveDomTree) 347 DT = &AM.getResult<DominatorTreeAnalysis>(F); 348 if (F.hasFnAttribute(Attribute::OptForFuzzing)) { 349 Options.setSimplifyCondBranch(false).setFoldTwoEntryPHINode(false); 350 } else { 351 Options.setSimplifyCondBranch(true).setFoldTwoEntryPHINode(true); 352 } 353 if (!simplifyFunctionCFG(F, TTI, DT, Options)) 354 return PreservedAnalyses::all(); 355 PreservedAnalyses PA; 356 if (RequireAndPreserveDomTree) 357 PA.preserve<DominatorTreeAnalysis>(); 358 return PA; 359 } 360 361 namespace { 362 struct CFGSimplifyPass : public FunctionPass { 363 static char ID; 364 SimplifyCFGOptions Options; 365 std::function<bool(const Function &)> PredicateFtor; 366 367 CFGSimplifyPass(SimplifyCFGOptions Options_ = SimplifyCFGOptions(), 368 std::function<bool(const Function &)> Ftor = nullptr) 369 : FunctionPass(ID), Options(Options_), PredicateFtor(std::move(Ftor)) { 370 371 initializeCFGSimplifyPassPass(*PassRegistry::getPassRegistry()); 372 373 // Check for command-line overrides of options for debug/customization. 374 applyCommandLineOverridesToOptions(Options); 375 } 376 377 bool runOnFunction(Function &F) override { 378 if (skipFunction(F) || (PredicateFtor && !PredicateFtor(F))) 379 return false; 380 381 Options.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 382 DominatorTree *DT = nullptr; 383 if (RequireAndPreserveDomTree) 384 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 385 if (F.hasFnAttribute(Attribute::OptForFuzzing)) { 386 Options.setSimplifyCondBranch(false) 387 .setFoldTwoEntryPHINode(false); 388 } else { 389 Options.setSimplifyCondBranch(true) 390 .setFoldTwoEntryPHINode(true); 391 } 392 393 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 394 return simplifyFunctionCFG(F, TTI, DT, Options); 395 } 396 void getAnalysisUsage(AnalysisUsage &AU) const override { 397 AU.addRequired<AssumptionCacheTracker>(); 398 if (RequireAndPreserveDomTree) 399 AU.addRequired<DominatorTreeWrapperPass>(); 400 AU.addRequired<TargetTransformInfoWrapperPass>(); 401 if (RequireAndPreserveDomTree) 402 AU.addPreserved<DominatorTreeWrapperPass>(); 403 AU.addPreserved<GlobalsAAWrapperPass>(); 404 } 405 }; 406 } 407 408 char CFGSimplifyPass::ID = 0; 409 INITIALIZE_PASS_BEGIN(CFGSimplifyPass, "simplifycfg", "Simplify the CFG", false, 410 false) 411 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 412 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 413 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 414 INITIALIZE_PASS_END(CFGSimplifyPass, "simplifycfg", "Simplify the CFG", false, 415 false) 416 417 // Public interface to the CFGSimplification pass 418 FunctionPass * 419 llvm::createCFGSimplificationPass(SimplifyCFGOptions Options, 420 std::function<bool(const Function &)> Ftor) { 421 return new CFGSimplifyPass(Options, std::move(Ftor)); 422 } 423