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/Dominators.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) {}
41 virtual bool runOnFunction(Function &F);
43 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
44 AU.addPreserved<DominatorTree>();
45 AU.addPreserved<DominanceFrontier>();
46 AU.addPreserved<LoopInfo>();
47 AU.addPreserved<ProfileInfo>();
49 // No loop canonicalization guarantees are broken by this pass.
50 AU.addPreservedID(LoopSimplifyID);
55 char BreakCriticalEdges::ID = 0;
56 static RegisterPass<BreakCriticalEdges>
57 X("break-crit-edges", "Break critical edges in CFG");
59 // Publically exposed interface to pass...
60 const PassInfo *const llvm::BreakCriticalEdgesID = &X;
61 FunctionPass *llvm::createBreakCriticalEdgesPass() {
62 return new BreakCriticalEdges();
65 // runOnFunction - Loop over all of the edges in the CFG, breaking critical
66 // edges as they are found.
68 bool BreakCriticalEdges::runOnFunction(Function &F) {
70 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
71 TerminatorInst *TI = I->getTerminator();
72 if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI))
73 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
74 if (SplitCriticalEdge(TI, i, this)) {
83 //===----------------------------------------------------------------------===//
84 // Implementation of the external critical edge manipulation functions
85 //===----------------------------------------------------------------------===//
87 // isCriticalEdge - Return true if the specified edge is a critical edge.
88 // Critical edges are edges from a block with multiple successors to a block
89 // with multiple predecessors.
91 bool llvm::isCriticalEdge(const TerminatorInst *TI, unsigned SuccNum,
92 bool AllowIdenticalEdges) {
93 assert(SuccNum < TI->getNumSuccessors() && "Illegal edge specification!");
94 if (TI->getNumSuccessors() == 1) return false;
96 const BasicBlock *Dest = TI->getSuccessor(SuccNum);
97 const_pred_iterator I = pred_begin(Dest), E = pred_end(Dest);
99 // If there is more than one predecessor, this is a critical edge...
100 assert(I != E && "No preds, but we have an edge to the block?");
101 const BasicBlock *FirstPred = *I;
102 ++I; // Skip one edge due to the incoming arc from TI.
103 if (!AllowIdenticalEdges)
106 // If AllowIdenticalEdges is true, then we allow this edge to be considered
107 // non-critical iff all preds come from TI's block.
111 // Note: leave this as is until no one ever compiles with either gcc 4.0.1
112 // or Xcode 2. This seems to work around the pred_iterator assert in PR 2207
119 /// CreatePHIsForSplitLoopExit - When a loop exit edge is split, LCSSA form
120 /// may require new PHIs in the new exit block. This function inserts the
121 /// new PHIs, as needed. Preds is a list of preds inside the loop, SplitBB
122 /// is the new loop exit block, and DestBB is the old loop exit, now the
123 /// successor of SplitBB.
124 static void CreatePHIsForSplitLoopExit(SmallVectorImpl<BasicBlock *> &Preds,
126 BasicBlock *DestBB) {
127 // SplitBB shouldn't have anything non-trivial in it yet.
128 assert(SplitBB->getFirstNonPHI() == SplitBB->getTerminator() &&
129 "SplitBB has non-PHI nodes!");
131 // For each PHI in the destination block...
132 for (BasicBlock::iterator I = DestBB->begin();
133 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
134 unsigned Idx = PN->getBasicBlockIndex(SplitBB);
135 Value *V = PN->getIncomingValue(Idx);
136 // If the input is a PHI which already satisfies LCSSA, don't create
138 if (const PHINode *VP = dyn_cast<PHINode>(V))
139 if (VP->getParent() == SplitBB)
141 // Otherwise a new PHI is needed. Create one and populate it.
142 PHINode *NewPN = PHINode::Create(PN->getType(), "split",
143 SplitBB->getTerminator());
144 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
145 NewPN->addIncoming(V, Preds[i]);
146 // Update the original PHI.
147 PN->setIncomingValue(Idx, NewPN);
151 /// SplitCriticalEdge - If this edge is a critical edge, insert a new node to
152 /// split the critical edge. This will update DominatorTree and
153 /// DominatorFrontier information if it is available, thus calling this pass
154 /// will not invalidate either of them. This returns the new block if the edge
155 /// was split, null otherwise.
157 /// If MergeIdenticalEdges is true (not the default), *all* edges from TI to the
158 /// specified successor will be merged into the same critical edge block.
159 /// This is most commonly interesting with switch instructions, which may
160 /// have many edges to any one destination. This ensures that all edges to that
161 /// dest go to one block instead of each going to a different block, but isn't
162 /// the standard definition of a "critical edge".
164 /// It is invalid to call this function on a critical edge that starts at an
165 /// IndirectBrInst. Splitting these edges will almost always create an invalid
166 /// program because the address of the new block won't be the one that is jumped
169 BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum,
170 Pass *P, bool MergeIdenticalEdges) {
171 if (!isCriticalEdge(TI, SuccNum, MergeIdenticalEdges)) return 0;
173 assert(!isa<IndirectBrInst>(TI) &&
174 "Cannot split critical edge from IndirectBrInst");
176 BasicBlock *TIBB = TI->getParent();
177 BasicBlock *DestBB = TI->getSuccessor(SuccNum);
179 // Create a new basic block, linking it into the CFG.
180 BasicBlock *NewBB = BasicBlock::Create(TI->getContext(),
181 TIBB->getName() + "." + DestBB->getName() + "_crit_edge");
182 // Create our unconditional branch.
183 BranchInst::Create(DestBB, NewBB);
185 // Branch to the new block, breaking the edge.
186 TI->setSuccessor(SuccNum, NewBB);
188 // Insert the block into the function... right after the block TI lives in.
189 Function &F = *TIBB->getParent();
190 Function::iterator FBBI = TIBB;
191 F.getBasicBlockList().insert(++FBBI, NewBB);
193 // If there are any PHI nodes in DestBB, we need to update them so that they
194 // merge incoming values from NewBB instead of from TIBB.
195 if (PHINode *APHI = dyn_cast<PHINode>(DestBB->begin())) {
196 // This conceptually does:
197 // foreach (PHINode *PN in DestBB)
198 // PN->setIncomingBlock(PN->getIncomingBlock(TIBB), NewBB);
199 // but is optimized for two cases.
201 if (APHI->getNumIncomingValues() <= 8) { // Small # preds case.
203 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) {
204 // We no longer enter through TIBB, now we come in through NewBB.
205 // Revector exactly one entry in the PHI node that used to come from
206 // TIBB to come from NewBB.
207 PHINode *PN = cast<PHINode>(I);
209 // Reuse the previous value of BBIdx if it lines up. In cases where we
210 // have multiple phi nodes with *lots* of predecessors, this is a speed
211 // win because we don't have to scan the PHI looking for TIBB. This
212 // happens because the BB list of PHI nodes are usually in the same
214 if (PN->getIncomingBlock(BBIdx) != TIBB)
215 BBIdx = PN->getBasicBlockIndex(TIBB);
216 PN->setIncomingBlock(BBIdx, NewBB);
219 // However, the foreach loop is slow for blocks with lots of predecessors
220 // because PHINode::getIncomingBlock is O(n) in # preds. Instead, walk
221 // the user list of TIBB to find the PHI nodes.
222 SmallPtrSet<PHINode*, 16> UpdatedPHIs;
224 for (Value::use_iterator UI = TIBB->use_begin(), E = TIBB->use_end();
226 Value::use_iterator Use = UI++;
227 if (PHINode *PN = dyn_cast<PHINode>(Use)) {
228 // Remove one entry from each PHI.
229 if (PN->getParent() == DestBB && UpdatedPHIs.insert(PN))
230 PN->setOperand(Use.getOperandNo(), NewBB);
236 // If there are any other edges from TIBB to DestBB, update those to go
237 // through the split block, making those edges non-critical as well (and
238 // reducing the number of phi entries in the DestBB if relevant).
239 if (MergeIdenticalEdges) {
240 for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) {
241 if (TI->getSuccessor(i) != DestBB) continue;
243 // Remove an entry for TIBB from DestBB phi nodes.
244 DestBB->removePredecessor(TIBB);
246 // We found another edge to DestBB, go to NewBB instead.
247 TI->setSuccessor(i, NewBB);
253 // If we don't have a pass object, we can't update anything...
254 if (P == 0) return NewBB;
256 DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>();
257 DominanceFrontier *DF = P->getAnalysisIfAvailable<DominanceFrontier>();
258 LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>();
259 ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>();
261 // If we have nothing to update, just return.
262 if (DT == 0 && DF == 0 && LI == 0 && PI == 0)
265 // Now update analysis information. Since the only predecessor of NewBB is
266 // the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate
267 // anything, as there are other successors of DestBB. However, if all other
268 // predecessors of DestBB are already dominated by DestBB (e.g. DestBB is a
269 // loop header) then NewBB dominates DestBB.
270 SmallVector<BasicBlock*, 8> OtherPreds;
272 // If there is a PHI in the block, loop over predecessors with it, which is
273 // faster than iterating pred_begin/end.
274 if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
275 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
276 if (PN->getIncomingBlock(i) != NewBB)
277 OtherPreds.push_back(PN->getIncomingBlock(i));
279 for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB);
283 OtherPreds.push_back(P);
287 bool NewBBDominatesDestBB = true;
289 // Should we update DominatorTree information?
291 DomTreeNode *TINode = DT->getNode(TIBB);
293 // The new block is not the immediate dominator for any other nodes, but
294 // TINode is the immediate dominator for the new node.
296 if (TINode) { // Don't break unreachable code!
297 DomTreeNode *NewBBNode = DT->addNewBlock(NewBB, TIBB);
298 DomTreeNode *DestBBNode = 0;
300 // If NewBBDominatesDestBB hasn't been computed yet, do so with DT.
301 if (!OtherPreds.empty()) {
302 DestBBNode = DT->getNode(DestBB);
303 while (!OtherPreds.empty() && NewBBDominatesDestBB) {
304 if (DomTreeNode *OPNode = DT->getNode(OtherPreds.back()))
305 NewBBDominatesDestBB = DT->dominates(DestBBNode, OPNode);
306 OtherPreds.pop_back();
311 // If NewBBDominatesDestBB, then NewBB dominates DestBB, otherwise it
312 // doesn't dominate anything.
313 if (NewBBDominatesDestBB) {
314 if (!DestBBNode) DestBBNode = DT->getNode(DestBB);
315 DT->changeImmediateDominator(DestBBNode, NewBBNode);
320 // Should we update DominanceFrontier information?
322 // If NewBBDominatesDestBB hasn't been computed yet, do so with DF.
323 if (!OtherPreds.empty()) {
324 // FIXME: IMPLEMENT THIS!
325 llvm_unreachable("Requiring domfrontiers but not idom/domtree/domset."
326 " not implemented yet!");
329 // Since the new block is dominated by its only predecessor TIBB,
330 // it cannot be in any block's dominance frontier. If NewBB dominates
331 // DestBB, its dominance frontier is the same as DestBB's, otherwise it is
333 DominanceFrontier::DomSetType NewDFSet;
334 if (NewBBDominatesDestBB) {
335 DominanceFrontier::iterator I = DF->find(DestBB);
336 if (I != DF->end()) {
337 DF->addBasicBlock(NewBB, I->second);
339 if (I->second.count(DestBB)) {
340 // However NewBB's frontier does not include DestBB.
341 DominanceFrontier::iterator NF = DF->find(NewBB);
342 DF->removeFromFrontier(NF, DestBB);
346 DF->addBasicBlock(NewBB, DominanceFrontier::DomSetType());
348 DominanceFrontier::DomSetType NewDFSet;
349 NewDFSet.insert(DestBB);
350 DF->addBasicBlock(NewBB, NewDFSet);
354 // Update LoopInfo if it is around.
356 if (Loop *TIL = LI->getLoopFor(TIBB)) {
357 // If one or the other blocks were not in a loop, the new block is not
358 // either, and thus LI doesn't need to be updated.
359 if (Loop *DestLoop = LI->getLoopFor(DestBB)) {
360 if (TIL == DestLoop) {
361 // Both in the same loop, the NewBB joins loop.
362 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
363 } else if (TIL->contains(DestLoop)) {
364 // Edge from an outer loop to an inner loop. Add to the outer loop.
365 TIL->addBasicBlockToLoop(NewBB, LI->getBase());
366 } else if (DestLoop->contains(TIL)) {
367 // Edge from an inner loop to an outer loop. Add to the outer loop.
368 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
370 // Edge from two loops with no containment relation. Because these
371 // are natural loops, we know that the destination block must be the
372 // header of its loop (adding a branch into a loop elsewhere would
373 // create an irreducible loop).
374 assert(DestLoop->getHeader() == DestBB &&
375 "Should not create irreducible loops!");
376 if (Loop *P = DestLoop->getParentLoop())
377 P->addBasicBlockToLoop(NewBB, LI->getBase());
380 // If TIBB is in a loop and DestBB is outside of that loop, split the
381 // other exit blocks of the loop that also have predecessors outside
382 // the loop, to maintain a LoopSimplify guarantee.
383 if (!TIL->contains(DestBB) &&
384 P->mustPreserveAnalysisID(LoopSimplifyID)) {
385 assert(!TIL->contains(NewBB) &&
386 "Split point for loop exit is contained in loop!");
388 // Update LCSSA form in the newly created exit block.
389 if (P->mustPreserveAnalysisID(LCSSAID)) {
390 SmallVector<BasicBlock *, 1> OrigPred;
391 OrigPred.push_back(TIBB);
392 CreatePHIsForSplitLoopExit(OrigPred, NewBB, DestBB);
395 // For each unique exit block...
396 SmallVector<BasicBlock *, 4> ExitBlocks;
397 TIL->getExitBlocks(ExitBlocks);
398 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
399 // Collect all the preds that are inside the loop, and note
400 // whether there are any preds outside the loop.
401 SmallVector<BasicBlock *, 4> Preds;
402 bool HasPredOutsideOfLoop = false;
403 BasicBlock *Exit = ExitBlocks[i];
404 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit);
406 if (TIL->contains(*I))
409 HasPredOutsideOfLoop = true;
410 // If there are any preds not in the loop, we'll need to split
411 // the edges. The Preds.empty() check is needed because a block
412 // may appear multiple times in the list. We can't use
413 // getUniqueExitBlocks above because that depends on LoopSimplify
414 // form, which we're in the process of restoring!
415 if (!Preds.empty() && HasPredOutsideOfLoop) {
416 BasicBlock *NewExitBB =
417 SplitBlockPredecessors(Exit, Preds.data(), Preds.size(),
419 if (P->mustPreserveAnalysisID(LCSSAID))
420 CreatePHIsForSplitLoopExit(Preds, NewExitBB, Exit);
424 // LCSSA form was updated above for the case where LoopSimplify is
425 // available, which means that all predecessors of loop exit blocks
426 // are within the loop. Without LoopSimplify form, it would be
427 // necessary to insert a new phi.
428 assert((!P->mustPreserveAnalysisID(LCSSAID) ||
429 P->mustPreserveAnalysisID(LoopSimplifyID)) &&
430 "SplitCriticalEdge doesn't know how to update LCCSA form "
431 "without LoopSimplify!");
435 // Update ProfileInfo if it is around.
437 PI->splitEdge(TIBB, DestBB, NewBB, MergeIdenticalEdges);