BreakCriticalEdges.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===- BreakCriticalEdges.cpp - Critical Edge Elimination 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// BreakCriticalEdges pass - Break all of the critical edges in the CFG by 11// inserting a dummy basic block. This pass may be "required" by passes that 12// cannot deal with critical edges. For this usage, the structure type is 13// forward declared. This pass obviously invalidates the CFG, but can update 14// dominator trees. 15// 16//===----------------------------------------------------------------------===// 17 18#define DEBUG_TYPE "break-crit-edges" 19#include "llvm/Transforms/Scalar.h" 20#include "llvm/ADT/SmallVector.h" 21#include "llvm/ADT/Statistic.h" 22#include "llvm/Analysis/CFG.h" 23#include "llvm/Analysis/LoopInfo.h" 24#include "llvm/IR/CFG.h" 25#include "llvm/IR/Dominators.h" 26#include "llvm/IR/Function.h" 27#include "llvm/IR/Instructions.h" 28#include "llvm/IR/Type.h" 29#include "llvm/Support/ErrorHandling.h" 30#include "llvm/Transforms/Utils/BasicBlockUtils.h" 31using namespace llvm; 32 33STATISTIC(NumBroken, "Number of blocks inserted"); 34 35namespace { 36 struct BreakCriticalEdges : public FunctionPass { 37 static char ID; // Pass identification, replacement for typeid 38 BreakCriticalEdges() : FunctionPass(ID) { 39 initializeBreakCriticalEdgesPass(*PassRegistry::getPassRegistry()); 40 } 41 42 bool runOnFunction(Function &F) override; 43 44 void getAnalysisUsage(AnalysisUsage &AU) const override { 45 AU.addPreserved<DominatorTreeWrapperPass>(); 46 AU.addPreserved<LoopInfo>(); 47 48 // No loop canonicalization guarantees are broken by this pass. 49 AU.addPreservedID(LoopSimplifyID); 50 } 51 }; 52} 53 54char BreakCriticalEdges::ID = 0; 55INITIALIZE_PASS(BreakCriticalEdges, "break-crit-edges", 56 "Break critical edges in CFG", false, false) 57 58// Publicly exposed interface to pass... 59char &llvm::BreakCriticalEdgesID = BreakCriticalEdges::ID; 60FunctionPass *llvm::createBreakCriticalEdgesPass() { 61 return new BreakCriticalEdges(); 62} 63 64// runOnFunction - Loop over all of the edges in the CFG, breaking critical 65// edges as they are found. 66// 67bool BreakCriticalEdges::runOnFunction(Function &F) { 68 bool Changed = false; 69 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) { 70 TerminatorInst *TI = I->getTerminator(); 71 if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI)) 72 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 73 if (SplitCriticalEdge(TI, i, this)) { 74 ++NumBroken; 75 Changed = true; 76 } 77 } 78 79 return Changed; 80} 81 82//===----------------------------------------------------------------------===// 83// Implementation of the external critical edge manipulation functions 84//===----------------------------------------------------------------------===// 85 86/// createPHIsForSplitLoopExit - When a loop exit edge is split, LCSSA form 87/// may require new PHIs in the new exit block. This function inserts the 88/// new PHIs, as needed. Preds is a list of preds inside the loop, SplitBB 89/// is the new loop exit block, and DestBB is the old loop exit, now the 90/// successor of SplitBB. 91static void createPHIsForSplitLoopExit(ArrayRef<BasicBlock *> Preds, 92 BasicBlock *SplitBB, 93 BasicBlock *DestBB) { 94 // SplitBB shouldn't have anything non-trivial in it yet. 95 assert((SplitBB->getFirstNonPHI() == SplitBB->getTerminator() || 96 SplitBB->isLandingPad()) && "SplitBB has non-PHI nodes!"); 97 98 // For each PHI in the destination block. 99 for (BasicBlock::iterator I = DestBB->begin(); 100 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 101 unsigned Idx = PN->getBasicBlockIndex(SplitBB); 102 Value *V = PN->getIncomingValue(Idx); 103 104 // If the input is a PHI which already satisfies LCSSA, don't create 105 // a new one. 106 if (const PHINode *VP = dyn_cast<PHINode>(V)) 107 if (VP->getParent() == SplitBB) 108 continue; 109 110 // Otherwise a new PHI is needed. Create one and populate it. 111 PHINode *NewPN = 112 PHINode::Create(PN->getType(), Preds.size(), "split", 113 SplitBB->isLandingPad() ? 114 SplitBB->begin() : SplitBB->getTerminator()); 115 for (unsigned i = 0, e = Preds.size(); i != e; ++i) 116 NewPN->addIncoming(V, Preds[i]); 117 118 // Update the original PHI. 119 PN->setIncomingValue(Idx, NewPN); 120 } 121} 122 123/// SplitCriticalEdge - If this edge is a critical edge, insert a new node to 124/// split the critical edge. This will update DominatorTree information if it 125/// is available, thus calling this pass will not invalidate either of them. 126/// This returns the new block if the edge was split, null otherwise. 127/// 128/// If MergeIdenticalEdges is true (not the default), *all* edges from TI to the 129/// specified successor will be merged into the same critical edge block. 130/// This is most commonly interesting with switch instructions, which may 131/// have many edges to any one destination. This ensures that all edges to that 132/// dest go to one block instead of each going to a different block, but isn't 133/// the standard definition of a "critical edge". 134/// 135/// It is invalid to call this function on a critical edge that starts at an 136/// IndirectBrInst. Splitting these edges will almost always create an invalid 137/// program because the address of the new block won't be the one that is jumped 138/// to. 139/// 140BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, 141 Pass *P, bool MergeIdenticalEdges, 142 bool DontDeleteUselessPhis, 143 bool SplitLandingPads) { 144 if (!isCriticalEdge(TI, SuccNum, MergeIdenticalEdges)) return 0; 145 146 assert(!isa<IndirectBrInst>(TI) && 147 "Cannot split critical edge from IndirectBrInst"); 148 149 BasicBlock *TIBB = TI->getParent(); 150 BasicBlock *DestBB = TI->getSuccessor(SuccNum); 151 152 // Splitting the critical edge to a landing pad block is non-trivial. Don't do 153 // it in this generic function. 154 if (DestBB->isLandingPad()) return 0; 155 156 // Create a new basic block, linking it into the CFG. 157 BasicBlock *NewBB = BasicBlock::Create(TI->getContext(), 158 TIBB->getName() + "." + DestBB->getName() + "_crit_edge"); 159 // Create our unconditional branch. 160 BranchInst *NewBI = BranchInst::Create(DestBB, NewBB); 161 NewBI->setDebugLoc(TI->getDebugLoc()); 162 163 // Branch to the new block, breaking the edge. 164 TI->setSuccessor(SuccNum, NewBB); 165 166 // Insert the block into the function... right after the block TI lives in. 167 Function &F = *TIBB->getParent(); 168 Function::iterator FBBI = TIBB; 169 F.getBasicBlockList().insert(++FBBI, NewBB); 170 171 // If there are any PHI nodes in DestBB, we need to update them so that they 172 // merge incoming values from NewBB instead of from TIBB. 173 { 174 unsigned BBIdx = 0; 175 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) { 176 // We no longer enter through TIBB, now we come in through NewBB. 177 // Revector exactly one entry in the PHI node that used to come from 178 // TIBB to come from NewBB. 179 PHINode *PN = cast<PHINode>(I); 180 181 // Reuse the previous value of BBIdx if it lines up. In cases where we 182 // have multiple phi nodes with *lots* of predecessors, this is a speed 183 // win because we don't have to scan the PHI looking for TIBB. This 184 // happens because the BB list of PHI nodes are usually in the same 185 // order. 186 if (PN->getIncomingBlock(BBIdx) != TIBB) 187 BBIdx = PN->getBasicBlockIndex(TIBB); 188 PN->setIncomingBlock(BBIdx, NewBB); 189 } 190 } 191 192 // If there are any other edges from TIBB to DestBB, update those to go 193 // through the split block, making those edges non-critical as well (and 194 // reducing the number of phi entries in the DestBB if relevant). 195 if (MergeIdenticalEdges) { 196 for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) { 197 if (TI->getSuccessor(i) != DestBB) continue; 198 199 // Remove an entry for TIBB from DestBB phi nodes. 200 DestBB->removePredecessor(TIBB, DontDeleteUselessPhis); 201 202 // We found another edge to DestBB, go to NewBB instead. 203 TI->setSuccessor(i, NewBB); 204 } 205 } 206 207 208 209 // If we don't have a pass object, we can't update anything... 210 if (P == 0) return NewBB; 211 212 DominatorTreeWrapperPass *DTWP = 213 P->getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 214 DominatorTree *DT = DTWP ? &DTWP->getDomTree() : 0; 215 LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>(); 216 217 // If we have nothing to update, just return. 218 if (DT == 0 && LI == 0) 219 return NewBB; 220 221 // Now update analysis information. Since the only predecessor of NewBB is 222 // the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate 223 // anything, as there are other successors of DestBB. However, if all other 224 // predecessors of DestBB are already dominated by DestBB (e.g. DestBB is a 225 // loop header) then NewBB dominates DestBB. 226 SmallVector<BasicBlock*, 8> OtherPreds; 227 228 // If there is a PHI in the block, loop over predecessors with it, which is 229 // faster than iterating pred_begin/end. 230 if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 231 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 232 if (PN->getIncomingBlock(i) != NewBB) 233 OtherPreds.push_back(PN->getIncomingBlock(i)); 234 } else { 235 for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB); 236 I != E; ++I) { 237 BasicBlock *P = *I; 238 if (P != NewBB) 239 OtherPreds.push_back(P); 240 } 241 } 242 243 bool NewBBDominatesDestBB = true; 244 245 // Should we update DominatorTree information? 246 if (DT) { 247 DomTreeNode *TINode = DT->getNode(TIBB); 248 249 // The new block is not the immediate dominator for any other nodes, but 250 // TINode is the immediate dominator for the new node. 251 // 252 if (TINode) { // Don't break unreachable code! 253 DomTreeNode *NewBBNode = DT->addNewBlock(NewBB, TIBB); 254 DomTreeNode *DestBBNode = 0; 255 256 // If NewBBDominatesDestBB hasn't been computed yet, do so with DT. 257 if (!OtherPreds.empty()) { 258 DestBBNode = DT->getNode(DestBB); 259 while (!OtherPreds.empty() && NewBBDominatesDestBB) { 260 if (DomTreeNode *OPNode = DT->getNode(OtherPreds.back())) 261 NewBBDominatesDestBB = DT->dominates(DestBBNode, OPNode); 262 OtherPreds.pop_back(); 263 } 264 OtherPreds.clear(); 265 } 266 267 // If NewBBDominatesDestBB, then NewBB dominates DestBB, otherwise it 268 // doesn't dominate anything. 269 if (NewBBDominatesDestBB) { 270 if (!DestBBNode) DestBBNode = DT->getNode(DestBB); 271 DT->changeImmediateDominator(DestBBNode, NewBBNode); 272 } 273 } 274 } 275 276 // Update LoopInfo if it is around. 277 if (LI) { 278 if (Loop *TIL = LI->getLoopFor(TIBB)) { 279 // If one or the other blocks were not in a loop, the new block is not 280 // either, and thus LI doesn't need to be updated. 281 if (Loop *DestLoop = LI->getLoopFor(DestBB)) { 282 if (TIL == DestLoop) { 283 // Both in the same loop, the NewBB joins loop. 284 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase()); 285 } else if (TIL->contains(DestLoop)) { 286 // Edge from an outer loop to an inner loop. Add to the outer loop. 287 TIL->addBasicBlockToLoop(NewBB, LI->getBase()); 288 } else if (DestLoop->contains(TIL)) { 289 // Edge from an inner loop to an outer loop. Add to the outer loop. 290 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase()); 291 } else { 292 // Edge from two loops with no containment relation. Because these 293 // are natural loops, we know that the destination block must be the 294 // header of its loop (adding a branch into a loop elsewhere would 295 // create an irreducible loop). 296 assert(DestLoop->getHeader() == DestBB && 297 "Should not create irreducible loops!"); 298 if (Loop *P = DestLoop->getParentLoop()) 299 P->addBasicBlockToLoop(NewBB, LI->getBase()); 300 } 301 } 302 // If TIBB is in a loop and DestBB is outside of that loop, we may need 303 // to update LoopSimplify form and LCSSA form. 304 if (!TIL->contains(DestBB) && 305 P->mustPreserveAnalysisID(LoopSimplifyID)) { 306 assert(!TIL->contains(NewBB) && 307 "Split point for loop exit is contained in loop!"); 308 309 // Update LCSSA form in the newly created exit block. 310 if (P->mustPreserveAnalysisID(LCSSAID)) 311 createPHIsForSplitLoopExit(TIBB, NewBB, DestBB); 312 313 // The only that we can break LoopSimplify form by splitting a critical 314 // edge is if after the split there exists some edge from TIL to DestBB 315 // *and* the only edge into DestBB from outside of TIL is that of 316 // NewBB. If the first isn't true, then LoopSimplify still holds, NewBB 317 // is the new exit block and it has no non-loop predecessors. If the 318 // second isn't true, then DestBB was not in LoopSimplify form prior to 319 // the split as it had a non-loop predecessor. In both of these cases, 320 // the predecessor must be directly in TIL, not in a subloop, or again 321 // LoopSimplify doesn't hold. 322 SmallVector<BasicBlock *, 4> LoopPreds; 323 for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB); I != E; 324 ++I) { 325 BasicBlock *P = *I; 326 if (P == NewBB) 327 continue; // The new block is known. 328 if (LI->getLoopFor(P) != TIL) { 329 // No need to re-simplify, it wasn't to start with. 330 LoopPreds.clear(); 331 break; 332 } 333 LoopPreds.push_back(P); 334 } 335 if (!LoopPreds.empty()) { 336 assert(!DestBB->isLandingPad() && 337 "We don't split edges to landing pads!"); 338 BasicBlock *NewExitBB = 339 SplitBlockPredecessors(DestBB, LoopPreds, "split", P); 340 if (P->mustPreserveAnalysisID(LCSSAID)) 341 createPHIsForSplitLoopExit(LoopPreds, NewExitBB, DestBB); 342 } 343 } 344 // LCSSA form was updated above for the case where LoopSimplify is 345 // available, which means that all predecessors of loop exit blocks 346 // are within the loop. Without LoopSimplify form, it would be 347 // necessary to insert a new phi. 348 assert((!P->mustPreserveAnalysisID(LCSSAID) || 349 P->mustPreserveAnalysisID(LoopSimplifyID)) && 350 "SplitCriticalEdge doesn't know how to update LCCSA form " 351 "without LoopSimplify!"); 352 } 353 } 354 355 return NewBB; 356} 357