SimplifyCFGPass.cpp revision f4f10e37791cef519a057d10d12f688333f554a7
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" 38using namespace llvm; 39 40STATISTIC(NumSimpl, "Number of blocks simplified"); 41 42namespace { 43 struct CFGSimplifyPass : public FunctionPass { 44 static char ID; // Pass identification, replacement for typeid 45 CFGSimplifyPass() : FunctionPass(&ID) {} 46 47 virtual bool runOnFunction(Function &F); 48 }; 49} 50 51char CFGSimplifyPass::ID = 0; 52static RegisterPass<CFGSimplifyPass> X("simplifycfg", "Simplify the CFG"); 53 54// Public interface to the CFGSimplification pass 55FunctionPass *llvm::createCFGSimplificationPass() { 56 return new CFGSimplifyPass(); 57} 58 59/// ChangeToUnreachable - Insert an unreachable instruction before the specified 60/// instruction, making it and the rest of the code in the block dead. 61static void ChangeToUnreachable(Instruction *I) { 62 BasicBlock *BB = I->getParent(); 63 // Loop over all of the successors, removing BB's entry from any PHI 64 // nodes. 65 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 66 (*SI)->removePredecessor(BB); 67 68 new UnreachableInst(I->getContext(), I); 69 70 // All instructions after this are dead. 71 BasicBlock::iterator BBI = I, BBE = BB->end(); 72 while (BBI != BBE) { 73 if (!BBI->use_empty()) 74 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 75 BB->getInstList().erase(BBI++); 76 } 77} 78 79/// ChangeToCall - Convert the specified invoke into a normal call. 80static void ChangeToCall(InvokeInst *II) { 81 BasicBlock *BB = II->getParent(); 82 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 83 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args.begin(), 84 Args.end(), "", II); 85 NewCall->takeName(II); 86 NewCall->setCallingConv(II->getCallingConv()); 87 NewCall->setAttributes(II->getAttributes()); 88 II->replaceAllUsesWith(NewCall); 89 90 // Follow the call by a branch to the normal destination. 91 BranchInst::Create(II->getNormalDest(), II); 92 93 // Update PHI nodes in the unwind destination 94 II->getUnwindDest()->removePredecessor(BB); 95 BB->getInstList().erase(II); 96} 97 98static bool MarkAliveBlocks(BasicBlock *BB, 99 SmallPtrSet<BasicBlock*, 128> &Reachable) { 100 101 SmallVector<BasicBlock*, 128> Worklist; 102 Worklist.push_back(BB); 103 bool Changed = false; 104 do { 105 BB = Worklist.pop_back_val(); 106 107 if (!Reachable.insert(BB)) 108 continue; 109 110 // Do a quick scan of the basic block, turning any obviously unreachable 111 // instructions into LLVM unreachable insts. The instruction combining pass 112 // canonicalizes unreachable insts into stores to null or undef. 113 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){ 114 if (CallInst *CI = dyn_cast<CallInst>(BBI)) { 115 if (CI->doesNotReturn()) { 116 // If we found a call to a no-return function, insert an unreachable 117 // instruction after it. Make sure there isn't *already* one there 118 // though. 119 ++BBI; 120 if (!isa<UnreachableInst>(BBI)) { 121 ChangeToUnreachable(BBI); 122 Changed = true; 123 } 124 break; 125 } 126 } 127 128 // Store to undef and store to null are undefined and used to signal that 129 // they should be changed to unreachable by passes that can't modify the 130 // CFG. 131 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 132 Value *Ptr = SI->getOperand(1); 133 134 if (isa<UndefValue>(Ptr) || 135 (isa<ConstantPointerNull>(Ptr) && 136 SI->getPointerAddressSpace() == 0)) { 137 ChangeToUnreachable(SI); 138 Changed = true; 139 break; 140 } 141 } 142 } 143 144 // Turn invokes that call 'nounwind' functions into ordinary calls. 145 if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) 146 if (II->doesNotThrow()) { 147 ChangeToCall(II); 148 Changed = true; 149 } 150 151 Changed |= ConstantFoldTerminator(BB); 152 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 153 Worklist.push_back(*SI); 154 } while (!Worklist.empty()); 155 return Changed; 156} 157 158/// RemoveUnreachableBlocksFromFn - Remove blocks that are not reachable, even 159/// if they are in a dead cycle. Return true if a change was made, false 160/// otherwise. 161static bool RemoveUnreachableBlocksFromFn(Function &F) { 162 SmallPtrSet<BasicBlock*, 128> Reachable; 163 bool Changed = MarkAliveBlocks(F.begin(), Reachable); 164 165 // If there are unreachable blocks in the CFG... 166 if (Reachable.size() == F.size()) 167 return Changed; 168 169 assert(Reachable.size() < F.size()); 170 NumSimpl += F.size()-Reachable.size(); 171 172 // Loop over all of the basic blocks that are not reachable, dropping all of 173 // their internal references... 174 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) { 175 if (Reachable.count(BB)) 176 continue; 177 178 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 179 if (Reachable.count(*SI)) 180 (*SI)->removePredecessor(BB); 181 BB->dropAllReferences(); 182 } 183 184 for (Function::iterator I = ++F.begin(); I != F.end();) 185 if (!Reachable.count(I)) 186 I = F.getBasicBlockList().erase(I); 187 else 188 ++I; 189 190 return true; 191} 192 193/// MergeEmptyReturnBlocks - If we have more than one empty (other than phi 194/// node) return blocks, merge them together to promote recursive block merging. 195static bool MergeEmptyReturnBlocks(Function &F) { 196 bool Changed = false; 197 198 BasicBlock *RetBlock = 0; 199 200 // Scan all the blocks in the function, looking for empty return blocks. 201 for (Function::iterator BBI = F.begin(), E = F.end(); BBI != E; ) { 202 BasicBlock &BB = *BBI++; 203 204 // Only look at return blocks. 205 ReturnInst *Ret = dyn_cast<ReturnInst>(BB.getTerminator()); 206 if (Ret == 0) continue; 207 208 // Only look at the block if it is empty or the only other thing in it is a 209 // single PHI node that is the operand to the return. 210 if (Ret != &BB.front()) { 211 // Check for something else in the block. 212 BasicBlock::iterator I = Ret; 213 --I; 214 // Skip over debug info. 215 while (isa<DbgInfoIntrinsic>(I) && I != BB.begin()) 216 --I; 217 if (!isa<DbgInfoIntrinsic>(I) && 218 (!isa<PHINode>(I) || I != BB.begin() || 219 Ret->getNumOperands() == 0 || 220 Ret->getOperand(0) != I)) 221 continue; 222 } 223 224 // If this is the first returning block, remember it and keep going. 225 if (RetBlock == 0) { 226 RetBlock = &BB; 227 continue; 228 } 229 230 // Otherwise, we found a duplicate return block. Merge the two. 231 Changed = true; 232 233 // Case when there is no input to the return or when the returned values 234 // agree is trivial. Note that they can't agree if there are phis in the 235 // blocks. 236 if (Ret->getNumOperands() == 0 || 237 Ret->getOperand(0) == 238 cast<ReturnInst>(RetBlock->getTerminator())->getOperand(0)) { 239 BB.replaceAllUsesWith(RetBlock); 240 BB.eraseFromParent(); 241 continue; 242 } 243 244 // If the canonical return block has no PHI node, create one now. 245 PHINode *RetBlockPHI = dyn_cast<PHINode>(RetBlock->begin()); 246 if (RetBlockPHI == 0) { 247 Value *InVal = cast<ReturnInst>(RetBlock->getTerminator())->getOperand(0); 248 RetBlockPHI = PHINode::Create(Ret->getOperand(0)->getType(), "merge", 249 &RetBlock->front()); 250 251 for (pred_iterator PI = pred_begin(RetBlock), E = pred_end(RetBlock); 252 PI != E; ++PI) 253 RetBlockPHI->addIncoming(InVal, *PI); 254 RetBlock->getTerminator()->setOperand(0, RetBlockPHI); 255 } 256 257 // Turn BB into a block that just unconditionally branches to the return 258 // block. This handles the case when the two return blocks have a common 259 // predecessor but that return different things. 260 RetBlockPHI->addIncoming(Ret->getOperand(0), &BB); 261 BB.getTerminator()->eraseFromParent(); 262 BranchInst::Create(RetBlock, &BB); 263 } 264 265 return Changed; 266} 267 268/// IterativeSimplifyCFG - Call SimplifyCFG on all the blocks in the function, 269/// iterating until no more changes are made. 270static bool IterativeSimplifyCFG(Function &F, const TargetData *TD) { 271 bool Changed = false; 272 bool LocalChange = true; 273 while (LocalChange) { 274 LocalChange = false; 275 276 // Loop over all of the basic blocks (except the first one) and remove them 277 // if they are unneeded... 278 // 279 for (Function::iterator BBIt = ++F.begin(); BBIt != F.end(); ) { 280 if (SimplifyCFG(BBIt++, TD)) { 281 LocalChange = true; 282 ++NumSimpl; 283 } 284 } 285 Changed |= LocalChange; 286 } 287 return Changed; 288} 289 290// It is possible that we may require multiple passes over the code to fully 291// simplify the CFG. 292// 293bool CFGSimplifyPass::runOnFunction(Function &F) { 294 const TargetData *TD = getAnalysisIfAvailable<TargetData>(); 295 bool EverChanged = RemoveUnreachableBlocksFromFn(F); 296 EverChanged |= MergeEmptyReturnBlocks(F); 297 EverChanged |= IterativeSimplifyCFG(F, TD); 298 299 // If neither pass changed anything, we're done. 300 if (!EverChanged) return false; 301 302 // IterativeSimplifyCFG can (rarely) make some loops dead. If this happens, 303 // RemoveUnreachableBlocksFromFn is needed to nuke them, which means we should 304 // iterate between the two optimizations. We structure the code like this to 305 // avoid reruning IterativeSimplifyCFG if the second pass of 306 // RemoveUnreachableBlocksFromFn doesn't do anything. 307 if (!RemoveUnreachableBlocksFromFn(F)) 308 return true; 309 310 do { 311 EverChanged = IterativeSimplifyCFG(F, TD); 312 EverChanged |= RemoveUnreachableBlocksFromFn(F); 313 } while (EverChanged); 314 315 return true; 316} 317