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> UserSwitchRangeToICmp( 63 "switch-range-to-icmp", cl::Hidden, cl::init(false), 64 cl::desc( 65 "Convert switches into an integer range comparison (default = false)")); 66 67 static cl::opt<bool> UserSwitchToLookup( 68 "switch-to-lookup", cl::Hidden, cl::init(false), 69 cl::desc("Convert switches to lookup tables (default = false)")); 70 71 static cl::opt<bool> UserForwardSwitchCond( 72 "forward-switch-cond", cl::Hidden, cl::init(false), 73 cl::desc("Forward switch condition to phi ops (default = false)")); 74 75 static cl::opt<bool> UserHoistCommonInsts( 76 "hoist-common-insts", cl::Hidden, cl::init(false), 77 cl::desc("hoist common instructions (default = false)")); 78 79 static cl::opt<bool> UserSinkCommonInsts( 80 "sink-common-insts", cl::Hidden, cl::init(false), 81 cl::desc("Sink common instructions (default = false)")); 82 83 84 STATISTIC(NumSimpl, "Number of blocks simplified"); 85 86 static bool 87 performBlockTailMerging(Function &F, ArrayRef<BasicBlock *> BBs, 88 std::vector<DominatorTree::UpdateType> *Updates) { 89 SmallVector<PHINode *, 1> NewOps; 90 91 // We don't want to change IR just because we can. 92 // Only do that if there are at least two blocks we'll tail-merge. 93 if (BBs.size() < 2) 94 return false; 95 96 if (Updates) 97 Updates->reserve(Updates->size() + BBs.size()); 98 99 BasicBlock *CanonicalBB; 100 Instruction *CanonicalTerm; 101 { 102 auto *Term = BBs[0]->getTerminator(); 103 104 // Create a canonical block for this function terminator type now, 105 // placing it *before* the first block that will branch to it. 106 CanonicalBB = BasicBlock::Create( 107 F.getContext(), Twine("common.") + Term->getOpcodeName(), &F, BBs[0]); 108 // We'll also need a PHI node per each operand of the terminator. 109 NewOps.resize(Term->getNumOperands()); 110 for (auto I : zip(Term->operands(), NewOps)) { 111 std::get<1>(I) = PHINode::Create(std::get<0>(I)->getType(), 112 /*NumReservedValues=*/BBs.size(), 113 CanonicalBB->getName() + ".op"); 114 CanonicalBB->getInstList().push_back(std::get<1>(I)); 115 } 116 // Make it so that this canonical block actually has the right 117 // terminator. 118 CanonicalTerm = Term->clone(); 119 CanonicalBB->getInstList().push_back(CanonicalTerm); 120 // If the canonical terminator has operands, rewrite it to take PHI's. 121 for (auto I : zip(NewOps, CanonicalTerm->operands())) 122 std::get<1>(I) = std::get<0>(I); 123 } 124 125 // Now, go through each block (with the current terminator type) 126 // we've recorded, and rewrite it to branch to the new common block. 127 const DILocation *CommonDebugLoc = nullptr; 128 for (BasicBlock *BB : BBs) { 129 auto *Term = BB->getTerminator(); 130 assert(Term->getOpcode() == CanonicalTerm->getOpcode() && 131 "All blocks to be tail-merged must be the same " 132 "(function-terminating) terminator type."); 133 134 // Aha, found a new non-canonical function terminator. If it has operands, 135 // forward them to the PHI nodes in the canonical block. 136 for (auto I : zip(Term->operands(), NewOps)) 137 std::get<1>(I)->addIncoming(std::get<0>(I), BB); 138 139 // Compute the debug location common to all the original terminators. 140 if (!CommonDebugLoc) 141 CommonDebugLoc = Term->getDebugLoc(); 142 else 143 CommonDebugLoc = 144 DILocation::getMergedLocation(CommonDebugLoc, Term->getDebugLoc()); 145 146 // And turn BB into a block that just unconditionally branches 147 // to the canonical block. 148 Term->eraseFromParent(); 149 BranchInst::Create(CanonicalBB, BB); 150 if (Updates) 151 Updates->push_back({DominatorTree::Insert, BB, CanonicalBB}); 152 } 153 154 CanonicalTerm->setDebugLoc(CommonDebugLoc); 155 156 return true; 157 } 158 159 static bool tailMergeBlocksWithSimilarFunctionTerminators(Function &F, 160 DomTreeUpdater *DTU) { 161 SmallMapVector<unsigned /*TerminatorOpcode*/, SmallVector<BasicBlock *, 2>, 4> 162 Structure; 163 164 // Scan all the blocks in the function, record the interesting-ones. 165 for (BasicBlock &BB : F) { 166 if (DTU && DTU->isBBPendingDeletion(&BB)) 167 continue; 168 169 // We are only interested in function-terminating blocks. 170 if (!succ_empty(&BB)) 171 continue; 172 173 auto *Term = BB.getTerminator(); 174 175 // Fow now only support `ret`/`resume` function terminators. 176 // FIXME: lift this restriction. 177 switch (Term->getOpcode()) { 178 case Instruction::Ret: 179 case Instruction::Resume: 180 break; 181 default: 182 continue; 183 } 184 185 // We can't tail-merge block that contains a musttail call. 186 if (BB.getTerminatingMustTailCall()) 187 continue; 188 189 // Calls to experimental_deoptimize must be followed by a return 190 // of the value computed by experimental_deoptimize. 191 // I.e., we can not change `ret` to `br` for this block. 192 if (auto *CI = 193 dyn_cast_or_null<CallInst>(Term->getPrevNonDebugInstruction())) { 194 if (Function *F = CI->getCalledFunction()) 195 if (Intrinsic::ID ID = F->getIntrinsicID()) 196 if (ID == Intrinsic::experimental_deoptimize) 197 continue; 198 } 199 200 // PHI nodes cannot have token type, so if the terminator has an operand 201 // with token type, we can not tail-merge this kind of function terminators. 202 if (any_of(Term->operands(), 203 [](Value *Op) { return Op->getType()->isTokenTy(); })) 204 continue; 205 206 // Canonical blocks are uniqued based on the terminator type (opcode). 207 Structure[Term->getOpcode()].emplace_back(&BB); 208 } 209 210 bool Changed = false; 211 212 std::vector<DominatorTree::UpdateType> Updates; 213 214 for (ArrayRef<BasicBlock *> BBs : make_second_range(Structure)) 215 Changed |= performBlockTailMerging(F, BBs, DTU ? &Updates : nullptr); 216 217 if (DTU) 218 DTU->applyUpdates(Updates); 219 220 return Changed; 221 } 222 223 /// Call SimplifyCFG on all the blocks in the function, 224 /// iterating until no more changes are made. 225 static bool iterativelySimplifyCFG(Function &F, const TargetTransformInfo &TTI, 226 DomTreeUpdater *DTU, 227 const SimplifyCFGOptions &Options) { 228 bool Changed = false; 229 bool LocalChange = true; 230 231 SmallVector<std::pair<const BasicBlock *, const BasicBlock *>, 32> Edges; 232 FindFunctionBackedges(F, Edges); 233 SmallPtrSet<BasicBlock *, 16> UniqueLoopHeaders; 234 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 235 UniqueLoopHeaders.insert(const_cast<BasicBlock *>(Edges[i].second)); 236 237 SmallVector<WeakVH, 16> LoopHeaders(UniqueLoopHeaders.begin(), 238 UniqueLoopHeaders.end()); 239 240 unsigned IterCnt = 0; 241 (void)IterCnt; 242 while (LocalChange) { 243 assert(IterCnt++ < 1000 && "Iterative simplification didn't converge!"); 244 LocalChange = false; 245 246 // Loop over all of the basic blocks and remove them if they are unneeded. 247 for (Function::iterator BBIt = F.begin(); BBIt != F.end(); ) { 248 BasicBlock &BB = *BBIt++; 249 if (DTU) { 250 assert( 251 !DTU->isBBPendingDeletion(&BB) && 252 "Should not end up trying to simplify blocks marked for removal."); 253 // Make sure that the advanced iterator does not point at the blocks 254 // that are marked for removal, skip over all such blocks. 255 while (BBIt != F.end() && DTU->isBBPendingDeletion(&*BBIt)) 256 ++BBIt; 257 } 258 if (simplifyCFG(&BB, TTI, DTU, Options, LoopHeaders)) { 259 LocalChange = true; 260 ++NumSimpl; 261 } 262 } 263 Changed |= LocalChange; 264 } 265 return Changed; 266 } 267 268 static bool simplifyFunctionCFGImpl(Function &F, const TargetTransformInfo &TTI, 269 DominatorTree *DT, 270 const SimplifyCFGOptions &Options) { 271 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 272 273 bool EverChanged = removeUnreachableBlocks(F, DT ? &DTU : nullptr); 274 EverChanged |= 275 tailMergeBlocksWithSimilarFunctionTerminators(F, DT ? &DTU : nullptr); 276 EverChanged |= iterativelySimplifyCFG(F, TTI, DT ? &DTU : nullptr, Options); 277 278 // If neither pass changed anything, we're done. 279 if (!EverChanged) return false; 280 281 // iterativelySimplifyCFG can (rarely) make some loops dead. If this happens, 282 // removeUnreachableBlocks is needed to nuke them, which means we should 283 // iterate between the two optimizations. We structure the code like this to 284 // avoid rerunning iterativelySimplifyCFG if the second pass of 285 // removeUnreachableBlocks doesn't do anything. 286 if (!removeUnreachableBlocks(F, DT ? &DTU : nullptr)) 287 return true; 288 289 do { 290 EverChanged = iterativelySimplifyCFG(F, TTI, DT ? &DTU : nullptr, Options); 291 EverChanged |= removeUnreachableBlocks(F, DT ? &DTU : nullptr); 292 } while (EverChanged); 293 294 return true; 295 } 296 297 static bool simplifyFunctionCFG(Function &F, const TargetTransformInfo &TTI, 298 DominatorTree *DT, 299 const SimplifyCFGOptions &Options) { 300 assert((!RequireAndPreserveDomTree || 301 (DT && DT->verify(DominatorTree::VerificationLevel::Full))) && 302 "Original domtree is invalid?"); 303 304 bool Changed = simplifyFunctionCFGImpl(F, TTI, DT, Options); 305 306 assert((!RequireAndPreserveDomTree || 307 (DT && DT->verify(DominatorTree::VerificationLevel::Full))) && 308 "Failed to maintain validity of domtree!"); 309 310 return Changed; 311 } 312 313 // Command-line settings override compile-time settings. 314 static void applyCommandLineOverridesToOptions(SimplifyCFGOptions &Options) { 315 if (UserBonusInstThreshold.getNumOccurrences()) 316 Options.BonusInstThreshold = UserBonusInstThreshold; 317 if (UserForwardSwitchCond.getNumOccurrences()) 318 Options.ForwardSwitchCondToPhi = UserForwardSwitchCond; 319 if (UserSwitchRangeToICmp.getNumOccurrences()) 320 Options.ConvertSwitchRangeToICmp = UserSwitchRangeToICmp; 321 if (UserSwitchToLookup.getNumOccurrences()) 322 Options.ConvertSwitchToLookupTable = UserSwitchToLookup; 323 if (UserKeepLoops.getNumOccurrences()) 324 Options.NeedCanonicalLoop = UserKeepLoops; 325 if (UserHoistCommonInsts.getNumOccurrences()) 326 Options.HoistCommonInsts = UserHoistCommonInsts; 327 if (UserSinkCommonInsts.getNumOccurrences()) 328 Options.SinkCommonInsts = UserSinkCommonInsts; 329 } 330 331 SimplifyCFGPass::SimplifyCFGPass() { 332 applyCommandLineOverridesToOptions(Options); 333 } 334 335 SimplifyCFGPass::SimplifyCFGPass(const SimplifyCFGOptions &Opts) 336 : Options(Opts) { 337 applyCommandLineOverridesToOptions(Options); 338 } 339 340 void SimplifyCFGPass::printPipeline( 341 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 342 static_cast<PassInfoMixin<SimplifyCFGPass> *>(this)->printPipeline( 343 OS, MapClassName2PassName); 344 OS << "<"; 345 OS << "bonus-inst-threshold=" << Options.BonusInstThreshold << ";"; 346 OS << (Options.ForwardSwitchCondToPhi ? "" : "no-") << "forward-switch-cond;"; 347 OS << (Options.ConvertSwitchRangeToICmp ? "" : "no-") 348 << "switch-range-to-icmp;"; 349 OS << (Options.ConvertSwitchToLookupTable ? "" : "no-") 350 << "switch-to-lookup;"; 351 OS << (Options.NeedCanonicalLoop ? "" : "no-") << "keep-loops;"; 352 OS << (Options.HoistCommonInsts ? "" : "no-") << "hoist-common-insts;"; 353 OS << (Options.SinkCommonInsts ? "" : "no-") << "sink-common-insts"; 354 OS << ">"; 355 } 356 357 PreservedAnalyses SimplifyCFGPass::run(Function &F, 358 FunctionAnalysisManager &AM) { 359 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 360 Options.AC = &AM.getResult<AssumptionAnalysis>(F); 361 DominatorTree *DT = nullptr; 362 if (RequireAndPreserveDomTree) 363 DT = &AM.getResult<DominatorTreeAnalysis>(F); 364 if (F.hasFnAttribute(Attribute::OptForFuzzing)) { 365 Options.setSimplifyCondBranch(false).setFoldTwoEntryPHINode(false); 366 } else { 367 Options.setSimplifyCondBranch(true).setFoldTwoEntryPHINode(true); 368 } 369 if (!simplifyFunctionCFG(F, TTI, DT, Options)) 370 return PreservedAnalyses::all(); 371 PreservedAnalyses PA; 372 if (RequireAndPreserveDomTree) 373 PA.preserve<DominatorTreeAnalysis>(); 374 return PA; 375 } 376 377 namespace { 378 struct CFGSimplifyPass : public FunctionPass { 379 static char ID; 380 SimplifyCFGOptions Options; 381 std::function<bool(const Function &)> PredicateFtor; 382 383 CFGSimplifyPass(SimplifyCFGOptions Options_ = SimplifyCFGOptions(), 384 std::function<bool(const Function &)> Ftor = nullptr) 385 : FunctionPass(ID), Options(Options_), PredicateFtor(std::move(Ftor)) { 386 387 initializeCFGSimplifyPassPass(*PassRegistry::getPassRegistry()); 388 389 // Check for command-line overrides of options for debug/customization. 390 applyCommandLineOverridesToOptions(Options); 391 } 392 393 bool runOnFunction(Function &F) override { 394 if (skipFunction(F) || (PredicateFtor && !PredicateFtor(F))) 395 return false; 396 397 Options.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 398 DominatorTree *DT = nullptr; 399 if (RequireAndPreserveDomTree) 400 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 401 if (F.hasFnAttribute(Attribute::OptForFuzzing)) { 402 Options.setSimplifyCondBranch(false) 403 .setFoldTwoEntryPHINode(false); 404 } else { 405 Options.setSimplifyCondBranch(true) 406 .setFoldTwoEntryPHINode(true); 407 } 408 409 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 410 return simplifyFunctionCFG(F, TTI, DT, Options); 411 } 412 void getAnalysisUsage(AnalysisUsage &AU) const override { 413 AU.addRequired<AssumptionCacheTracker>(); 414 if (RequireAndPreserveDomTree) 415 AU.addRequired<DominatorTreeWrapperPass>(); 416 AU.addRequired<TargetTransformInfoWrapperPass>(); 417 if (RequireAndPreserveDomTree) 418 AU.addPreserved<DominatorTreeWrapperPass>(); 419 AU.addPreserved<GlobalsAAWrapperPass>(); 420 } 421 }; 422 } 423 424 char CFGSimplifyPass::ID = 0; 425 INITIALIZE_PASS_BEGIN(CFGSimplifyPass, "simplifycfg", "Simplify the CFG", false, 426 false) 427 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 428 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 429 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 430 INITIALIZE_PASS_END(CFGSimplifyPass, "simplifycfg", "Simplify the CFG", false, 431 false) 432 433 // Public interface to the CFGSimplification pass 434 FunctionPass * 435 llvm::createCFGSimplificationPass(SimplifyCFGOptions Options, 436 std::function<bool(const Function &)> Ftor) { 437 return new CFGSimplifyPass(Options, std::move(Ftor)); 438 } 439