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