1 //===- CGSCCPassManager.cpp - Managing & running CGSCC passes -------------===//
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 #include "llvm/Analysis/CGSCCPassManager.h"
11 #include "llvm/IR/CallSite.h"
12 #include "llvm/IR/InstIterator.h"
16 // Explicit template instantiations and specialization defininitions for core
20 // Explicit instantiations for the core proxy templates.
21 template class AllAnalysesOn<LazyCallGraph::SCC>;
22 template class AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>;
23 template class PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager,
24 LazyCallGraph &, CGSCCUpdateResult &>;
25 template class InnerAnalysisManagerProxy<CGSCCAnalysisManager, Module>;
26 template class OuterAnalysisManagerProxy<ModuleAnalysisManager,
27 LazyCallGraph::SCC, LazyCallGraph &>;
28 template class OuterAnalysisManagerProxy<CGSCCAnalysisManager, Function>;
30 /// Explicitly specialize the pass manager run method to handle call graph
34 PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &,
35 CGSCCUpdateResult &>::run(LazyCallGraph::SCC &InitialC,
36 CGSCCAnalysisManager &AM,
37 LazyCallGraph &G, CGSCCUpdateResult &UR) {
38 PreservedAnalyses PA = PreservedAnalyses::all();
41 dbgs() << "Starting CGSCC pass manager run.\n";
43 // The SCC may be refined while we are running passes over it, so set up
44 // a pointer that we can update.
45 LazyCallGraph::SCC *C = &InitialC;
47 for (auto &Pass : Passes) {
49 dbgs() << "Running pass: " << Pass->name() << " on " << *C << "\n";
51 PreservedAnalyses PassPA = Pass->run(*C, AM, G, UR);
53 // Update the SCC if necessary.
54 C = UR.UpdatedC ? UR.UpdatedC : C;
56 // Check that we didn't miss any update scenario.
57 assert(!UR.InvalidatedSCCs.count(C) && "Processing an invalid SCC!");
58 assert(C->begin() != C->end() && "Cannot have an empty SCC!");
60 // Update the analysis manager as each pass runs and potentially
61 // invalidates analyses.
62 AM.invalidate(*C, PassPA);
64 // Finally, we intersect the final preserved analyses to compute the
65 // aggregate preserved set for this pass manager.
66 PA.intersect(std::move(PassPA));
68 // FIXME: Historically, the pass managers all called the LLVM context's
69 // yield function here. We don't have a generic way to acquire the
70 // context and it isn't yet clear what the right pattern is for yielding
71 // in the new pass manager so it is currently omitted.
72 // ...getContext().yield();
75 // Invaliadtion was handled after each pass in the above loop for the current
76 // SCC. Therefore, the remaining analysis results in the AnalysisManager are
77 // preserved. We mark this with a set so that we don't need to inspect each
79 PA.preserveSet<AllAnalysesOn<LazyCallGraph::SCC>>();
82 dbgs() << "Finished CGSCC pass manager run.\n";
87 bool CGSCCAnalysisManagerModuleProxy::Result::invalidate(
88 Module &M, const PreservedAnalyses &PA,
89 ModuleAnalysisManager::Invalidator &Inv) {
90 // If literally everything is preserved, we're done.
91 if (PA.areAllPreserved())
92 return false; // This is still a valid proxy.
94 // If this proxy or the call graph is going to be invalidated, we also need
95 // to clear all the keys coming from that analysis.
97 // We also directly invalidate the FAM's module proxy if necessary, and if
98 // that proxy isn't preserved we can't preserve this proxy either. We rely on
99 // it to handle module -> function analysis invalidation in the face of
100 // structural changes and so if it's unavailable we conservatively clear the
101 // entire SCC layer as well rather than trying to do invalidation ourselves.
102 auto PAC = PA.getChecker<CGSCCAnalysisManagerModuleProxy>();
103 if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Module>>()) ||
104 Inv.invalidate<LazyCallGraphAnalysis>(M, PA) ||
105 Inv.invalidate<FunctionAnalysisManagerModuleProxy>(M, PA)) {
108 // And the proxy itself should be marked as invalid so that we can observe
109 // the new call graph. This isn't strictly necessary because we cheat
110 // above, but is still useful.
114 // Directly check if the relevant set is preserved so we can short circuit
115 // invalidating SCCs below.
116 bool AreSCCAnalysesPreserved =
117 PA.allAnalysesInSetPreserved<AllAnalysesOn<LazyCallGraph::SCC>>();
119 // Ok, we have a graph, so we can propagate the invalidation down into it.
121 for (auto &RC : G->postorder_ref_sccs())
123 Optional<PreservedAnalyses> InnerPA;
125 // Check to see whether the preserved set needs to be adjusted based on
126 // module-level analysis invalidation triggering deferred invalidation
128 if (auto *OuterProxy =
129 InnerAM->getCachedResult<ModuleAnalysisManagerCGSCCProxy>(C))
130 for (const auto &OuterInvalidationPair :
131 OuterProxy->getOuterInvalidations()) {
132 AnalysisKey *OuterAnalysisID = OuterInvalidationPair.first;
133 const auto &InnerAnalysisIDs = OuterInvalidationPair.second;
134 if (Inv.invalidate(OuterAnalysisID, M, PA)) {
137 for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs)
138 InnerPA->abandon(InnerAnalysisID);
142 // Check if we needed a custom PA set. If so we'll need to run the inner
145 InnerAM->invalidate(C, *InnerPA);
149 // Otherwise we only need to do invalidation if the original PA set didn't
150 // preserve all SCC analyses.
151 if (!AreSCCAnalysesPreserved)
152 InnerAM->invalidate(C, PA);
155 // Return false to indicate that this result is still a valid proxy.
160 CGSCCAnalysisManagerModuleProxy::Result
161 CGSCCAnalysisManagerModuleProxy::run(Module &M, ModuleAnalysisManager &AM) {
162 // Force the Function analysis manager to also be available so that it can
163 // be accessed in an SCC analysis and proxied onward to function passes.
164 // FIXME: It is pretty awkward to just drop the result here and assert that
165 // we can find it again later.
166 (void)AM.getResult<FunctionAnalysisManagerModuleProxy>(M);
168 return Result(*InnerAM, AM.getResult<LazyCallGraphAnalysis>(M));
171 AnalysisKey FunctionAnalysisManagerCGSCCProxy::Key;
173 FunctionAnalysisManagerCGSCCProxy::Result
174 FunctionAnalysisManagerCGSCCProxy::run(LazyCallGraph::SCC &C,
175 CGSCCAnalysisManager &AM,
177 // Collect the FunctionAnalysisManager from the Module layer and use that to
178 // build the proxy result.
180 // This allows us to rely on the FunctionAnalysisMangaerModuleProxy to
181 // invalidate the function analyses.
182 auto &MAM = AM.getResult<ModuleAnalysisManagerCGSCCProxy>(C, CG).getManager();
183 Module &M = *C.begin()->getFunction().getParent();
184 auto *FAMProxy = MAM.getCachedResult<FunctionAnalysisManagerModuleProxy>(M);
185 assert(FAMProxy && "The CGSCC pass manager requires that the FAM module "
186 "proxy is run on the module prior to entering the CGSCC "
189 // Note that we special-case invalidation handling of this proxy in the CGSCC
190 // analysis manager's Module proxy. This avoids the need to do anything
191 // special here to recompute all of this if ever the FAM's module proxy goes
193 return Result(FAMProxy->getManager());
196 bool FunctionAnalysisManagerCGSCCProxy::Result::invalidate(
197 LazyCallGraph::SCC &C, const PreservedAnalyses &PA,
198 CGSCCAnalysisManager::Invalidator &Inv) {
199 for (LazyCallGraph::Node &N : C)
200 FAM->invalidate(N.getFunction(), PA);
202 // This proxy doesn't need to handle invalidation itself. Instead, the
203 // module-level CGSCC proxy handles it above by ensuring that if the
204 // module-level FAM proxy becomes invalid the entire SCC layer, which
205 // includes this proxy, is cleared.
209 } // End llvm namespace
212 /// Helper function to update both the \c CGSCCAnalysisManager \p AM and the \c
213 /// CGSCCPassManager's \c CGSCCUpdateResult \p UR based on a range of newly
216 /// The range of new SCCs must be in postorder already. The SCC they were split
217 /// out of must be provided as \p C. The current node being mutated and
218 /// triggering updates must be passed as \p N.
220 /// This function returns the SCC containing \p N. This will be either \p C if
221 /// no new SCCs have been split out, or it will be the new SCC containing \p N.
222 template <typename SCCRangeT>
224 incorporateNewSCCRange(const SCCRangeT &NewSCCRange, LazyCallGraph &G,
225 LazyCallGraph::Node &N, LazyCallGraph::SCC *C,
226 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR,
227 bool DebugLogging = false) {
228 typedef LazyCallGraph::SCC SCC;
230 if (NewSCCRange.begin() == NewSCCRange.end())
233 // Add the current SCC to the worklist as its shape has changed.
234 UR.CWorklist.insert(C);
236 dbgs() << "Enqueuing the existing SCC in the worklist:" << *C << "\n";
241 // Update the current SCC. Note that if we have new SCCs, this must actually
243 assert(C != &*NewSCCRange.begin() &&
244 "Cannot insert new SCCs without changing current SCC!");
245 C = &*NewSCCRange.begin();
246 assert(G.lookupSCC(N) == C && "Failed to update current SCC!");
249 reverse(make_range(std::next(NewSCCRange.begin()), NewSCCRange.end()))) {
250 assert(C != &NewC && "No need to re-visit the current SCC!");
251 assert(OldC != &NewC && "Already handled the original SCC!");
252 UR.CWorklist.insert(&NewC);
254 dbgs() << "Enqueuing a newly formed SCC:" << NewC << "\n";
260 LazyCallGraph::SCC &llvm::updateCGAndAnalysisManagerForFunctionPass(
261 LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N,
262 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR, bool DebugLogging) {
263 typedef LazyCallGraph::Node Node;
264 typedef LazyCallGraph::Edge Edge;
265 typedef LazyCallGraph::SCC SCC;
266 typedef LazyCallGraph::RefSCC RefSCC;
268 RefSCC &InitialRC = InitialC.getOuterRefSCC();
270 RefSCC *RC = &InitialRC;
271 Function &F = N.getFunction();
273 // Walk the function body and build up the set of retained, promoted, and
275 SmallVector<Constant *, 16> Worklist;
276 SmallPtrSet<Constant *, 16> Visited;
277 SmallPtrSet<Node *, 16> RetainedEdges;
278 SmallSetVector<Node *, 4> PromotedRefTargets;
279 SmallSetVector<Node *, 4> DemotedCallTargets;
281 // First walk the function and handle all called functions. We do this first
282 // because if there is a single call edge, whether there are ref edges is
284 for (Instruction &I : instructions(F))
285 if (auto CS = CallSite(&I))
286 if (Function *Callee = CS.getCalledFunction())
287 if (Visited.insert(Callee).second && !Callee->isDeclaration()) {
288 Node &CalleeN = *G.lookup(*Callee);
289 Edge *E = N->lookup(CalleeN);
290 // FIXME: We should really handle adding new calls. While it will
291 // make downstream usage more complex, there is no fundamental
292 // limitation and it will allow passes within the CGSCC to be a bit
293 // more flexible in what transforms they can do. Until then, we
294 // verify that new calls haven't been introduced.
295 assert(E && "No function transformations should introduce *new* "
296 "call edges! Any new calls should be modeled as "
297 "promoted existing ref edges!");
298 RetainedEdges.insert(&CalleeN);
300 PromotedRefTargets.insert(&CalleeN);
303 // Now walk all references.
304 for (Instruction &I : instructions(F))
305 for (Value *Op : I.operand_values())
306 if (Constant *C = dyn_cast<Constant>(Op))
307 if (Visited.insert(C).second)
308 Worklist.push_back(C);
310 LazyCallGraph::visitReferences(Worklist, Visited, [&](Function &Referee) {
311 Node &RefereeN = *G.lookup(Referee);
312 Edge *E = N->lookup(RefereeN);
313 // FIXME: Similarly to new calls, we also currently preclude
314 // introducing new references. See above for details.
315 assert(E && "No function transformations should introduce *new* ref "
316 "edges! Any new ref edges would require IPO which "
317 "function passes aren't allowed to do!");
318 RetainedEdges.insert(&RefereeN);
320 DemotedCallTargets.insert(&RefereeN);
323 // First remove all of the edges that are no longer present in this function.
324 // We have to build a list of dead targets first and then remove them as the
325 // data structures will all be invalidated by removing them.
326 SmallVector<PointerIntPair<Node *, 1, Edge::Kind>, 4> DeadTargets;
328 if (!RetainedEdges.count(&E.getNode()))
329 DeadTargets.push_back({&E.getNode(), E.getKind()});
330 for (auto DeadTarget : DeadTargets) {
331 Node &TargetN = *DeadTarget.getPointer();
332 bool IsCall = DeadTarget.getInt() == Edge::Call;
333 SCC &TargetC = *G.lookupSCC(TargetN);
334 RefSCC &TargetRC = TargetC.getOuterRefSCC();
336 if (&TargetRC != RC) {
337 RC->removeOutgoingEdge(N, TargetN);
339 dbgs() << "Deleting outgoing edge from '" << N << "' to '" << TargetN
344 dbgs() << "Deleting internal " << (IsCall ? "call" : "ref")
345 << " edge from '" << N << "' to '" << TargetN << "'\n";
349 // For separate SCCs this is trivial.
350 RC->switchTrivialInternalEdgeToRef(N, TargetN);
352 // Otherwise we may end up re-structuring the call graph. First,
353 // invalidate any SCC analyses. We have to do this before we split
354 // functions into new SCCs and lose track of where their analyses are
356 // FIXME: We should accept a more precise preserved set here. For
357 // example, it might be possible to preserve some function analyses
358 // even as the SCC structure is changed.
359 AM.invalidate(*C, PreservedAnalyses::none());
360 // Now update the call graph.
361 C = incorporateNewSCCRange(RC->switchInternalEdgeToRef(N, TargetN), G,
362 N, C, AM, UR, DebugLogging);
366 auto NewRefSCCs = RC->removeInternalRefEdge(N, TargetN);
367 if (!NewRefSCCs.empty()) {
368 // Note that we don't bother to invalidate analyses as ref-edge
369 // connectivity is not really observable in any way and is intended
370 // exclusively to be used for ordering of transforms rather than for
371 // analysis conclusions.
373 // The RC worklist is in reverse postorder, so we first enqueue the
374 // current RefSCC as it will remain the parent of all split RefSCCs, then
375 // we enqueue the new ones in RPO except for the one which contains the
376 // source node as that is the "bottom" we will continue processing in the
378 UR.RCWorklist.insert(RC);
380 dbgs() << "Enqueuing the existing RefSCC in the update worklist: "
382 // Update the RC to the "bottom".
383 assert(G.lookupSCC(N) == C && "Changed the SCC when splitting RefSCCs!");
384 RC = &C->getOuterRefSCC();
385 assert(G.lookupRefSCC(N) == RC && "Failed to update current RefSCC!");
386 assert(NewRefSCCs.front() == RC &&
387 "New current RefSCC not first in the returned list!");
388 for (RefSCC *NewRC : reverse(
389 make_range(std::next(NewRefSCCs.begin()), NewRefSCCs.end()))) {
390 assert(NewRC != RC && "Should not encounter the current RefSCC further "
391 "in the postorder list of new RefSCCs.");
392 UR.RCWorklist.insert(NewRC);
394 dbgs() << "Enqueuing a new RefSCC in the update worklist: " << *NewRC
400 // Next demote all the call edges that are now ref edges. This helps make
401 // the SCCs small which should minimize the work below as we don't want to
402 // form cycles that this would break.
403 for (Node *RefTarget : DemotedCallTargets) {
404 SCC &TargetC = *G.lookupSCC(*RefTarget);
405 RefSCC &TargetRC = TargetC.getOuterRefSCC();
407 // The easy case is when the target RefSCC is not this RefSCC. This is
408 // only supported when the target RefSCC is a child of this RefSCC.
409 if (&TargetRC != RC) {
410 assert(RC->isAncestorOf(TargetRC) &&
411 "Cannot potentially form RefSCC cycles here!");
412 RC->switchOutgoingEdgeToRef(N, *RefTarget);
414 dbgs() << "Switch outgoing call edge to a ref edge from '" << N
415 << "' to '" << *RefTarget << "'\n";
419 // We are switching an internal call edge to a ref edge. This may split up
422 // For separate SCCs this is trivial.
423 RC->switchTrivialInternalEdgeToRef(N, *RefTarget);
427 // Otherwise we may end up re-structuring the call graph. First, invalidate
428 // any SCC analyses. We have to do this before we split functions into new
429 // SCCs and lose track of where their analyses are cached.
430 // FIXME: We should accept a more precise preserved set here. For example,
431 // it might be possible to preserve some function analyses even as the SCC
432 // structure is changed.
433 AM.invalidate(*C, PreservedAnalyses::none());
434 // Now update the call graph.
435 C = incorporateNewSCCRange(RC->switchInternalEdgeToRef(N, *RefTarget), G, N,
436 C, AM, UR, DebugLogging);
439 // Now promote ref edges into call edges.
440 for (Node *CallTarget : PromotedRefTargets) {
441 SCC &TargetC = *G.lookupSCC(*CallTarget);
442 RefSCC &TargetRC = TargetC.getOuterRefSCC();
444 // The easy case is when the target RefSCC is not this RefSCC. This is
445 // only supported when the target RefSCC is a child of this RefSCC.
446 if (&TargetRC != RC) {
447 assert(RC->isAncestorOf(TargetRC) &&
448 "Cannot potentially form RefSCC cycles here!");
449 RC->switchOutgoingEdgeToCall(N, *CallTarget);
451 dbgs() << "Switch outgoing ref edge to a call edge from '" << N
452 << "' to '" << *CallTarget << "'\n";
456 dbgs() << "Switch an internal ref edge to a call edge from '" << N
457 << "' to '" << *CallTarget << "'\n";
459 // Otherwise we are switching an internal ref edge to a call edge. This
460 // may merge away some SCCs, and we add those to the UpdateResult. We also
461 // need to make sure to update the worklist in the event SCCs have moved
462 // before the current one in the post-order sequence.
463 auto InitialSCCIndex = RC->find(*C) - RC->begin();
464 auto InvalidatedSCCs = RC->switchInternalEdgeToCall(N, *CallTarget);
465 if (!InvalidatedSCCs.empty()) {
467 assert(G.lookupSCC(N) == C && "Failed to update current SCC!");
469 // Any analyses cached for this SCC are no longer precise as the shape
470 // has changed by introducing this cycle.
471 AM.invalidate(*C, PreservedAnalyses::none());
473 for (SCC *InvalidatedC : InvalidatedSCCs) {
474 assert(InvalidatedC != C && "Cannot invalidate the current SCC!");
475 UR.InvalidatedSCCs.insert(InvalidatedC);
477 // Also clear any cached analyses for the SCCs that are dead. This
478 // isn't really necessary for correctness but can release memory.
479 AM.clear(*InvalidatedC);
482 auto NewSCCIndex = RC->find(*C) - RC->begin();
483 if (InitialSCCIndex < NewSCCIndex) {
484 // Put our current SCC back onto the worklist as we'll visit other SCCs
485 // that are now definitively ordered prior to the current one in the
486 // post-order sequence, and may end up observing more precise context to
487 // optimize the current SCC.
488 UR.CWorklist.insert(C);
490 dbgs() << "Enqueuing the existing SCC in the worklist: " << *C << "\n";
491 // Enqueue in reverse order as we pop off the back of the worklist.
492 for (SCC &MovedC : reverse(make_range(RC->begin() + InitialSCCIndex,
493 RC->begin() + NewSCCIndex))) {
494 UR.CWorklist.insert(&MovedC);
496 dbgs() << "Enqueuing a newly earlier in post-order SCC: " << MovedC
502 assert(!UR.InvalidatedSCCs.count(C) && "Invalidated the current SCC!");
503 assert(!UR.InvalidatedRefSCCs.count(RC) && "Invalidated the current RefSCC!");
504 assert(&C->getOuterRefSCC() == RC && "Current SCC not in current RefSCC!");
506 // Record the current RefSCC and SCC for higher layers of the CGSCC pass
507 // manager now that all the updates have been applied.
508 if (RC != &InitialRC)