1 //===- SimplifyCFGPass.cpp - CFG Simplification Pass ----------------------===// 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 // This file implements dead code elimination and basic block merging, along 11 // with a collection of other peephole control flow optimizations. For example: 12 // 13 // * Removes basic blocks with no predecessors. 14 // * Merges a basic block into its predecessor if there is only one and the 15 // predecessor only has one successor. 16 // * Eliminates PHI nodes for basic blocks with a single predecessor. 17 // * Eliminates a basic block that only contains an unconditional branch. 18 // * Changes invoke instructions to nounwind functions to be calls. 19 // * Change things like "if (x) if (y)" into "if (x&y)". 20 // * etc.. 21 // 22 //===----------------------------------------------------------------------===// 23 24 #define DEBUG_TYPE "simplifycfg" 25 #include "llvm/Transforms/Scalar.h" 26 #include "llvm/Transforms/Utils/Local.h" 27 #include "llvm/Constants.h" 28 #include "llvm/Instructions.h" 29 #include "llvm/IntrinsicInst.h" 30 #include "llvm/Module.h" 31 #include "llvm/Attributes.h" 32 #include "llvm/Support/CFG.h" 33 #include "llvm/Pass.h" 34 #include "llvm/Target/TargetData.h" 35 #include "llvm/ADT/SmallVector.h" 36 #include "llvm/ADT/SmallPtrSet.h" 37 #include "llvm/ADT/Statistic.h" 38 using namespace llvm; 39 40 STATISTIC(NumSimpl, "Number of blocks simplified"); 41 42 namespace { 43 struct CFGSimplifyPass : public FunctionPass { 44 static char ID; // Pass identification, replacement for typeid 45 CFGSimplifyPass() : FunctionPass(ID) { 46 initializeCFGSimplifyPassPass(*PassRegistry::getPassRegistry()); 47 } 48 49 virtual bool runOnFunction(Function &F); 50 }; 51 } 52 53 char CFGSimplifyPass::ID = 0; 54 INITIALIZE_PASS(CFGSimplifyPass, "simplifycfg", 55 "Simplify the CFG", false, false) 56 57 // Public interface to the CFGSimplification pass 58 FunctionPass *llvm::createCFGSimplificationPass() { 59 return new CFGSimplifyPass(); 60 } 61 62 /// ChangeToUnreachable - Insert an unreachable instruction before the specified 63 /// instruction, making it and the rest of the code in the block dead. 64 static void ChangeToUnreachable(Instruction *I, bool UseLLVMTrap) { 65 BasicBlock *BB = I->getParent(); 66 // Loop over all of the successors, removing BB's entry from any PHI 67 // nodes. 68 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 69 (*SI)->removePredecessor(BB); 70 71 // Insert a call to llvm.trap right before this. This turns the undefined 72 // behavior into a hard fail instead of falling through into random code. 73 if (UseLLVMTrap) { 74 Function *TrapFn = 75 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap); 76 CallInst *CallTrap = CallInst::Create(TrapFn, "", I); 77 CallTrap->setDebugLoc(I->getDebugLoc()); 78 } 79 new UnreachableInst(I->getContext(), I); 80 81 // All instructions after this are dead. 82 BasicBlock::iterator BBI = I, BBE = BB->end(); 83 while (BBI != BBE) { 84 if (!BBI->use_empty()) 85 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 86 BB->getInstList().erase(BBI++); 87 } 88 } 89 90 /// ChangeToCall - Convert the specified invoke into a normal call. 91 static void ChangeToCall(InvokeInst *II) { 92 BasicBlock *BB = II->getParent(); 93 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 94 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, "", II); 95 NewCall->takeName(II); 96 NewCall->setCallingConv(II->getCallingConv()); 97 NewCall->setAttributes(II->getAttributes()); 98 NewCall->setDebugLoc(II->getDebugLoc()); 99 II->replaceAllUsesWith(NewCall); 100 101 // Follow the call by a branch to the normal destination. 102 BranchInst::Create(II->getNormalDest(), II); 103 104 // Update PHI nodes in the unwind destination 105 II->getUnwindDest()->removePredecessor(BB); 106 BB->getInstList().erase(II); 107 } 108 109 static bool MarkAliveBlocks(BasicBlock *BB, 110 SmallPtrSet<BasicBlock*, 128> &Reachable) { 111 112 SmallVector<BasicBlock*, 128> Worklist; 113 Worklist.push_back(BB); 114 bool Changed = false; 115 do { 116 BB = Worklist.pop_back_val(); 117 118 if (!Reachable.insert(BB)) 119 continue; 120 121 // Do a quick scan of the basic block, turning any obviously unreachable 122 // instructions into LLVM unreachable insts. The instruction combining pass 123 // canonicalizes unreachable insts into stores to null or undef. 124 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){ 125 if (CallInst *CI = dyn_cast<CallInst>(BBI)) { 126 if (CI->doesNotReturn()) { 127 // If we found a call to a no-return function, insert an unreachable 128 // instruction after it. Make sure there isn't *already* one there 129 // though. 130 ++BBI; 131 if (!isa<UnreachableInst>(BBI)) { 132 // Don't insert a call to llvm.trap right before the unreachable. 133 ChangeToUnreachable(BBI, false); 134 Changed = true; 135 } 136 break; 137 } 138 } 139 140 // Store to undef and store to null are undefined and used to signal that 141 // they should be changed to unreachable by passes that can't modify the 142 // CFG. 143 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 144 // Don't touch volatile stores. 145 if (SI->isVolatile()) continue; 146 147 Value *Ptr = SI->getOperand(1); 148 149 if (isa<UndefValue>(Ptr) || 150 (isa<ConstantPointerNull>(Ptr) && 151 SI->getPointerAddressSpace() == 0)) { 152 ChangeToUnreachable(SI, true); 153 Changed = true; 154 break; 155 } 156 } 157 } 158 159 // Turn invokes that call 'nounwind' functions into ordinary calls. 160 if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) 161 if (II->doesNotThrow()) { 162 ChangeToCall(II); 163 Changed = true; 164 } 165 166 Changed |= ConstantFoldTerminator(BB, true); 167 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 168 Worklist.push_back(*SI); 169 } while (!Worklist.empty()); 170 return Changed; 171 } 172 173 /// RemoveUnreachableBlocksFromFn - Remove blocks that are not reachable, even 174 /// if they are in a dead cycle. Return true if a change was made, false 175 /// otherwise. 176 static bool RemoveUnreachableBlocksFromFn(Function &F) { 177 SmallPtrSet<BasicBlock*, 128> Reachable; 178 bool Changed = MarkAliveBlocks(F.begin(), Reachable); 179 180 // If there are unreachable blocks in the CFG... 181 if (Reachable.size() == F.size()) 182 return Changed; 183 184 assert(Reachable.size() < F.size()); 185 NumSimpl += F.size()-Reachable.size(); 186 187 // Loop over all of the basic blocks that are not reachable, dropping all of 188 // their internal references... 189 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) { 190 if (Reachable.count(BB)) 191 continue; 192 193 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 194 if (Reachable.count(*SI)) 195 (*SI)->removePredecessor(BB); 196 BB->dropAllReferences(); 197 } 198 199 for (Function::iterator I = ++F.begin(); I != F.end();) 200 if (!Reachable.count(I)) 201 I = F.getBasicBlockList().erase(I); 202 else 203 ++I; 204 205 return true; 206 } 207 208 /// MergeEmptyReturnBlocks - If we have more than one empty (other than phi 209 /// node) return blocks, merge them together to promote recursive block merging. 210 static bool MergeEmptyReturnBlocks(Function &F) { 211 bool Changed = false; 212 213 BasicBlock *RetBlock = 0; 214 215 // Scan all the blocks in the function, looking for empty return blocks. 216 for (Function::iterator BBI = F.begin(), E = F.end(); BBI != E; ) { 217 BasicBlock &BB = *BBI++; 218 219 // Only look at return blocks. 220 ReturnInst *Ret = dyn_cast<ReturnInst>(BB.getTerminator()); 221 if (Ret == 0) continue; 222 223 // Only look at the block if it is empty or the only other thing in it is a 224 // single PHI node that is the operand to the return. 225 if (Ret != &BB.front()) { 226 // Check for something else in the block. 227 BasicBlock::iterator I = Ret; 228 --I; 229 // Skip over debug info. 230 while (isa<DbgInfoIntrinsic>(I) && I != BB.begin()) 231 --I; 232 if (!isa<DbgInfoIntrinsic>(I) && 233 (!isa<PHINode>(I) || I != BB.begin() || 234 Ret->getNumOperands() == 0 || 235 Ret->getOperand(0) != I)) 236 continue; 237 } 238 239 // If this is the first returning block, remember it and keep going. 240 if (RetBlock == 0) { 241 RetBlock = &BB; 242 continue; 243 } 244 245 // Otherwise, we found a duplicate return block. Merge the two. 246 Changed = true; 247 248 // Case when there is no input to the return or when the returned values 249 // agree is trivial. Note that they can't agree if there are phis in the 250 // blocks. 251 if (Ret->getNumOperands() == 0 || 252 Ret->getOperand(0) == 253 cast<ReturnInst>(RetBlock->getTerminator())->getOperand(0)) { 254 BB.replaceAllUsesWith(RetBlock); 255 BB.eraseFromParent(); 256 continue; 257 } 258 259 // If the canonical return block has no PHI node, create one now. 260 PHINode *RetBlockPHI = dyn_cast<PHINode>(RetBlock->begin()); 261 if (RetBlockPHI == 0) { 262 Value *InVal = cast<ReturnInst>(RetBlock->getTerminator())->getOperand(0); 263 pred_iterator PB = pred_begin(RetBlock), PE = pred_end(RetBlock); 264 RetBlockPHI = PHINode::Create(Ret->getOperand(0)->getType(), 265 std::distance(PB, PE), "merge", 266 &RetBlock->front()); 267 268 for (pred_iterator PI = PB; PI != PE; ++PI) 269 RetBlockPHI->addIncoming(InVal, *PI); 270 RetBlock->getTerminator()->setOperand(0, RetBlockPHI); 271 } 272 273 // Turn BB into a block that just unconditionally branches to the return 274 // block. This handles the case when the two return blocks have a common 275 // predecessor but that return different things. 276 RetBlockPHI->addIncoming(Ret->getOperand(0), &BB); 277 BB.getTerminator()->eraseFromParent(); 278 BranchInst::Create(RetBlock, &BB); 279 } 280 281 return Changed; 282 } 283 284 /// IterativeSimplifyCFG - Call SimplifyCFG on all the blocks in the function, 285 /// iterating until no more changes are made. 286 static bool IterativeSimplifyCFG(Function &F, const TargetData *TD) { 287 bool Changed = false; 288 bool LocalChange = true; 289 while (LocalChange) { 290 LocalChange = false; 291 292 // Loop over all of the basic blocks and remove them if they are unneeded... 293 // 294 for (Function::iterator BBIt = F.begin(); BBIt != F.end(); ) { 295 if (SimplifyCFG(BBIt++, TD)) { 296 LocalChange = true; 297 ++NumSimpl; 298 } 299 } 300 Changed |= LocalChange; 301 } 302 return Changed; 303 } 304 305 // It is possible that we may require multiple passes over the code to fully 306 // simplify the CFG. 307 // 308 bool CFGSimplifyPass::runOnFunction(Function &F) { 309 const TargetData *TD = getAnalysisIfAvailable<TargetData>(); 310 bool EverChanged = RemoveUnreachableBlocksFromFn(F); 311 EverChanged |= MergeEmptyReturnBlocks(F); 312 EverChanged |= IterativeSimplifyCFG(F, TD); 313 314 // If neither pass changed anything, we're done. 315 if (!EverChanged) return false; 316 317 // IterativeSimplifyCFG can (rarely) make some loops dead. If this happens, 318 // RemoveUnreachableBlocksFromFn is needed to nuke them, which means we should 319 // iterate between the two optimizations. We structure the code like this to 320 // avoid reruning IterativeSimplifyCFG if the second pass of 321 // RemoveUnreachableBlocksFromFn doesn't do anything. 322 if (!RemoveUnreachableBlocksFromFn(F)) 323 return true; 324 325 do { 326 EverChanged = IterativeSimplifyCFG(F, TD); 327 EverChanged |= RemoveUnreachableBlocksFromFn(F); 328 } while (EverChanged); 329 330 return true; 331 } 332