1 //===- BreakCriticalEdges.cpp - Critical Edge Elimination Pass ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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 // forward dominator (set, immediate dominators, tree, and frontier)
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "break-crit-edges"
20 #include "llvm/Transforms/Scalar.h"
21 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
22 #include "llvm/Analysis/DominanceFrontier.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/ProfileInfo.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Type.h"
28 #include "llvm/Support/CFG.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
34 STATISTIC(NumBroken, "Number of blocks inserted");
37 struct BreakCriticalEdges : public FunctionPass {
38 static char ID; // Pass identification, replacement for typeid
39 BreakCriticalEdges() : FunctionPass(ID) {
40 initializeBreakCriticalEdgesPass(*PassRegistry::getPassRegistry());
43 virtual bool runOnFunction(Function &F);
45 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
46 AU.addPreserved<DominatorTree>();
47 AU.addPreserved<DominanceFrontier>();
48 AU.addPreserved<LoopInfo>();
49 AU.addPreserved<ProfileInfo>();
51 // No loop canonicalization guarantees are broken by this pass.
52 AU.addPreservedID(LoopSimplifyID);
57 char BreakCriticalEdges::ID = 0;
58 INITIALIZE_PASS(BreakCriticalEdges, "break-crit-edges",
59 "Break critical edges in CFG", false, false)
61 // Publically exposed interface to pass...
62 char &llvm::BreakCriticalEdgesID = BreakCriticalEdges::ID;
63 FunctionPass *llvm::createBreakCriticalEdgesPass() {
64 return new BreakCriticalEdges();
67 // runOnFunction - Loop over all of the edges in the CFG, breaking critical
68 // edges as they are found.
70 bool BreakCriticalEdges::runOnFunction(Function &F) {
72 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
73 TerminatorInst *TI = I->getTerminator();
74 if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI))
75 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
76 if (SplitCriticalEdge(TI, i, this)) {
85 //===----------------------------------------------------------------------===//
86 // Implementation of the external critical edge manipulation functions
87 //===----------------------------------------------------------------------===//
89 // isCriticalEdge - Return true if the specified edge is a critical edge.
90 // Critical edges are edges from a block with multiple successors to a block
91 // with multiple predecessors.
93 bool llvm::isCriticalEdge(const TerminatorInst *TI, unsigned SuccNum,
94 bool AllowIdenticalEdges) {
95 assert(SuccNum < TI->getNumSuccessors() && "Illegal edge specification!");
96 if (TI->getNumSuccessors() == 1) return false;
98 const BasicBlock *Dest = TI->getSuccessor(SuccNum);
99 const_pred_iterator I = pred_begin(Dest), E = pred_end(Dest);
101 // If there is more than one predecessor, this is a critical edge...
102 assert(I != E && "No preds, but we have an edge to the block?");
103 const BasicBlock *FirstPred = *I;
104 ++I; // Skip one edge due to the incoming arc from TI.
105 if (!AllowIdenticalEdges)
108 // If AllowIdenticalEdges is true, then we allow this edge to be considered
109 // non-critical iff all preds come from TI's block.
111 const BasicBlock *P = *I;
114 // Note: leave this as is until no one ever compiles with either gcc 4.0.1
115 // or Xcode 2. This seems to work around the pred_iterator assert in PR 2207
122 /// CreatePHIsForSplitLoopExit - When a loop exit edge is split, LCSSA form
123 /// may require new PHIs in the new exit block. This function inserts the
124 /// new PHIs, as needed. Preds is a list of preds inside the loop, SplitBB
125 /// is the new loop exit block, and DestBB is the old loop exit, now the
126 /// successor of SplitBB.
127 static void CreatePHIsForSplitLoopExit(SmallVectorImpl<BasicBlock *> &Preds,
129 BasicBlock *DestBB) {
130 // SplitBB shouldn't have anything non-trivial in it yet.
131 assert(SplitBB->getFirstNonPHI() == SplitBB->getTerminator() &&
132 "SplitBB has non-PHI nodes!");
134 // For each PHI in the destination block...
135 for (BasicBlock::iterator I = DestBB->begin();
136 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
137 unsigned Idx = PN->getBasicBlockIndex(SplitBB);
138 Value *V = PN->getIncomingValue(Idx);
139 // If the input is a PHI which already satisfies LCSSA, don't create
141 if (const PHINode *VP = dyn_cast<PHINode>(V))
142 if (VP->getParent() == SplitBB)
144 // Otherwise a new PHI is needed. Create one and populate it.
145 PHINode *NewPN = PHINode::Create(PN->getType(), "split",
146 SplitBB->getTerminator());
147 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
148 NewPN->addIncoming(V, Preds[i]);
149 // Update the original PHI.
150 PN->setIncomingValue(Idx, NewPN);
154 /// SplitCriticalEdge - If this edge is a critical edge, insert a new node to
155 /// split the critical edge. This will update DominatorTree and
156 /// DominatorFrontier information if it is available, thus calling this pass
157 /// will not invalidate either of them. This returns the new block if the edge
158 /// was split, null otherwise.
160 /// If MergeIdenticalEdges is true (not the default), *all* edges from TI to the
161 /// specified successor will be merged into the same critical edge block.
162 /// This is most commonly interesting with switch instructions, which may
163 /// have many edges to any one destination. This ensures that all edges to that
164 /// dest go to one block instead of each going to a different block, but isn't
165 /// the standard definition of a "critical edge".
167 /// It is invalid to call this function on a critical edge that starts at an
168 /// IndirectBrInst. Splitting these edges will almost always create an invalid
169 /// program because the address of the new block won't be the one that is jumped
172 BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum,
173 Pass *P, bool MergeIdenticalEdges) {
174 if (!isCriticalEdge(TI, SuccNum, MergeIdenticalEdges)) return 0;
176 assert(!isa<IndirectBrInst>(TI) &&
177 "Cannot split critical edge from IndirectBrInst");
179 BasicBlock *TIBB = TI->getParent();
180 BasicBlock *DestBB = TI->getSuccessor(SuccNum);
182 // Create a new basic block, linking it into the CFG.
183 BasicBlock *NewBB = BasicBlock::Create(TI->getContext(),
184 TIBB->getName() + "." + DestBB->getName() + "_crit_edge");
185 // Create our unconditional branch.
186 BranchInst::Create(DestBB, NewBB);
188 // Branch to the new block, breaking the edge.
189 TI->setSuccessor(SuccNum, NewBB);
191 // Insert the block into the function... right after the block TI lives in.
192 Function &F = *TIBB->getParent();
193 Function::iterator FBBI = TIBB;
194 F.getBasicBlockList().insert(++FBBI, NewBB);
196 // If there are any PHI nodes in DestBB, we need to update them so that they
197 // merge incoming values from NewBB instead of from TIBB.
198 if (PHINode *APHI = dyn_cast<PHINode>(DestBB->begin())) {
199 // This conceptually does:
200 // foreach (PHINode *PN in DestBB)
201 // PN->setIncomingBlock(PN->getIncomingBlock(TIBB), NewBB);
202 // but is optimized for two cases.
204 if (APHI->getNumIncomingValues() <= 8) { // Small # preds case.
206 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) {
207 // We no longer enter through TIBB, now we come in through NewBB.
208 // Revector exactly one entry in the PHI node that used to come from
209 // TIBB to come from NewBB.
210 PHINode *PN = cast<PHINode>(I);
212 // Reuse the previous value of BBIdx if it lines up. In cases where we
213 // have multiple phi nodes with *lots* of predecessors, this is a speed
214 // win because we don't have to scan the PHI looking for TIBB. This
215 // happens because the BB list of PHI nodes are usually in the same
217 if (PN->getIncomingBlock(BBIdx) != TIBB)
218 BBIdx = PN->getBasicBlockIndex(TIBB);
219 PN->setIncomingBlock(BBIdx, NewBB);
222 // However, the foreach loop is slow for blocks with lots of predecessors
223 // because PHINode::getIncomingBlock is O(n) in # preds. Instead, walk
224 // the user list of TIBB to find the PHI nodes.
225 SmallPtrSet<PHINode*, 16> UpdatedPHIs;
227 for (Value::use_iterator UI = TIBB->use_begin(), E = TIBB->use_end();
229 Value::use_iterator Use = UI++;
230 if (PHINode *PN = dyn_cast<PHINode>(*Use)) {
231 // Remove one entry from each PHI.
232 if (PN->getParent() == DestBB && UpdatedPHIs.insert(PN))
233 PN->setOperand(Use.getOperandNo(), NewBB);
239 // If there are any other edges from TIBB to DestBB, update those to go
240 // through the split block, making those edges non-critical as well (and
241 // reducing the number of phi entries in the DestBB if relevant).
242 if (MergeIdenticalEdges) {
243 for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) {
244 if (TI->getSuccessor(i) != DestBB) continue;
246 // Remove an entry for TIBB from DestBB phi nodes.
247 DestBB->removePredecessor(TIBB);
249 // We found another edge to DestBB, go to NewBB instead.
250 TI->setSuccessor(i, NewBB);
256 // If we don't have a pass object, we can't update anything...
257 if (P == 0) return NewBB;
259 DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>();
260 DominanceFrontier *DF = P->getAnalysisIfAvailable<DominanceFrontier>();
261 LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>();
262 ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>();
264 // If we have nothing to update, just return.
265 if (DT == 0 && DF == 0 && LI == 0 && PI == 0)
268 // Now update analysis information. Since the only predecessor of NewBB is
269 // the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate
270 // anything, as there are other successors of DestBB. However, if all other
271 // predecessors of DestBB are already dominated by DestBB (e.g. DestBB is a
272 // loop header) then NewBB dominates DestBB.
273 SmallVector<BasicBlock*, 8> OtherPreds;
275 // If there is a PHI in the block, loop over predecessors with it, which is
276 // faster than iterating pred_begin/end.
277 if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
278 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
279 if (PN->getIncomingBlock(i) != NewBB)
280 OtherPreds.push_back(PN->getIncomingBlock(i));
282 for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB);
286 OtherPreds.push_back(P);
290 bool NewBBDominatesDestBB = true;
292 // Should we update DominatorTree information?
294 DomTreeNode *TINode = DT->getNode(TIBB);
296 // The new block is not the immediate dominator for any other nodes, but
297 // TINode is the immediate dominator for the new node.
299 if (TINode) { // Don't break unreachable code!
300 DomTreeNode *NewBBNode = DT->addNewBlock(NewBB, TIBB);
301 DomTreeNode *DestBBNode = 0;
303 // If NewBBDominatesDestBB hasn't been computed yet, do so with DT.
304 if (!OtherPreds.empty()) {
305 DestBBNode = DT->getNode(DestBB);
306 while (!OtherPreds.empty() && NewBBDominatesDestBB) {
307 if (DomTreeNode *OPNode = DT->getNode(OtherPreds.back()))
308 NewBBDominatesDestBB = DT->dominates(DestBBNode, OPNode);
309 OtherPreds.pop_back();
314 // If NewBBDominatesDestBB, then NewBB dominates DestBB, otherwise it
315 // doesn't dominate anything.
316 if (NewBBDominatesDestBB) {
317 if (!DestBBNode) DestBBNode = DT->getNode(DestBB);
318 DT->changeImmediateDominator(DestBBNode, NewBBNode);
323 // Should we update DominanceFrontier information?
325 // If NewBBDominatesDestBB hasn't been computed yet, do so with DF.
326 if (!OtherPreds.empty()) {
327 // FIXME: IMPLEMENT THIS!
328 llvm_unreachable("Requiring domfrontiers but not idom/domtree/domset."
329 " not implemented yet!");
332 // Since the new block is dominated by its only predecessor TIBB,
333 // it cannot be in any block's dominance frontier. If NewBB dominates
334 // DestBB, its dominance frontier is the same as DestBB's, otherwise it is
336 DominanceFrontier::DomSetType NewDFSet;
337 if (NewBBDominatesDestBB) {
338 DominanceFrontier::iterator I = DF->find(DestBB);
339 if (I != DF->end()) {
340 DF->addBasicBlock(NewBB, I->second);
342 if (I->second.count(DestBB)) {
343 // However NewBB's frontier does not include DestBB.
344 DominanceFrontier::iterator NF = DF->find(NewBB);
345 DF->removeFromFrontier(NF, DestBB);
349 DF->addBasicBlock(NewBB, DominanceFrontier::DomSetType());
351 DominanceFrontier::DomSetType NewDFSet;
352 NewDFSet.insert(DestBB);
353 DF->addBasicBlock(NewBB, NewDFSet);
357 // Update LoopInfo if it is around.
359 if (Loop *TIL = LI->getLoopFor(TIBB)) {
360 // If one or the other blocks were not in a loop, the new block is not
361 // either, and thus LI doesn't need to be updated.
362 if (Loop *DestLoop = LI->getLoopFor(DestBB)) {
363 if (TIL == DestLoop) {
364 // Both in the same loop, the NewBB joins loop.
365 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
366 } else if (TIL->contains(DestLoop)) {
367 // Edge from an outer loop to an inner loop. Add to the outer loop.
368 TIL->addBasicBlockToLoop(NewBB, LI->getBase());
369 } else if (DestLoop->contains(TIL)) {
370 // Edge from an inner loop to an outer loop. Add to the outer loop.
371 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
373 // Edge from two loops with no containment relation. Because these
374 // are natural loops, we know that the destination block must be the
375 // header of its loop (adding a branch into a loop elsewhere would
376 // create an irreducible loop).
377 assert(DestLoop->getHeader() == DestBB &&
378 "Should not create irreducible loops!");
379 if (Loop *P = DestLoop->getParentLoop())
380 P->addBasicBlockToLoop(NewBB, LI->getBase());
383 // If TIBB is in a loop and DestBB is outside of that loop, split the
384 // other exit blocks of the loop that also have predecessors outside
385 // the loop, to maintain a LoopSimplify guarantee.
386 if (!TIL->contains(DestBB) &&
387 P->mustPreserveAnalysisID(LoopSimplifyID)) {
388 assert(!TIL->contains(NewBB) &&
389 "Split point for loop exit is contained in loop!");
391 // Update LCSSA form in the newly created exit block.
392 if (P->mustPreserveAnalysisID(LCSSAID)) {
393 SmallVector<BasicBlock *, 1> OrigPred;
394 OrigPred.push_back(TIBB);
395 CreatePHIsForSplitLoopExit(OrigPred, NewBB, DestBB);
398 // For each unique exit block...
399 SmallVector<BasicBlock *, 4> ExitBlocks;
400 TIL->getExitBlocks(ExitBlocks);
401 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
402 // Collect all the preds that are inside the loop, and note
403 // whether there are any preds outside the loop.
404 SmallVector<BasicBlock *, 4> Preds;
405 bool HasPredOutsideOfLoop = false;
406 BasicBlock *Exit = ExitBlocks[i];
407 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit);
410 if (TIL->contains(P))
413 HasPredOutsideOfLoop = true;
415 // If there are any preds not in the loop, we'll need to split
416 // the edges. The Preds.empty() check is needed because a block
417 // may appear multiple times in the list. We can't use
418 // getUniqueExitBlocks above because that depends on LoopSimplify
419 // form, which we're in the process of restoring!
420 if (!Preds.empty() && HasPredOutsideOfLoop) {
421 BasicBlock *NewExitBB =
422 SplitBlockPredecessors(Exit, Preds.data(), Preds.size(),
424 if (P->mustPreserveAnalysisID(LCSSAID))
425 CreatePHIsForSplitLoopExit(Preds, NewExitBB, Exit);
429 // LCSSA form was updated above for the case where LoopSimplify is
430 // available, which means that all predecessors of loop exit blocks
431 // are within the loop. Without LoopSimplify form, it would be
432 // necessary to insert a new phi.
433 assert((!P->mustPreserveAnalysisID(LCSSAID) ||
434 P->mustPreserveAnalysisID(LoopSimplifyID)) &&
435 "SplitCriticalEdge doesn't know how to update LCCSA form "
436 "without LoopSimplify!");
440 // Update ProfileInfo if it is around.
442 PI->splitEdge(TIBB, DestBB, NewBB, MergeIdenticalEdges);