[LoopDeletion] Clear block & loop dispo cache after breaking backedge.

[fhahn]

Summary

breakLoopBackedge may remove blocks and loops. Also clear block &
loop disposition to avoid the cache containing invalid blocks and loops.

[dblaikie]

While I guess this is a fix for crash/UB, was the problem covered by any tests or could we add test coverage? (or maybe the failure's so unlikely as to be impractical to test - some ABA allocation situation which rarely occurs?)

[fhahn]

In D134663#3816040, @dblaikie wrote:

While I guess this is a fix for crash/UB, was the problem covered by any tests or could we add test coverage? (or maybe the failure's so unlikely as to be impractical to test - some ABA allocation situation which rarely occurs?)

Without this change, multiple transform tests fail when LLVM is built with ASAN, so this should be covered by existing tests + buildbots running ASAN.

[dblaikie]

In D134663#3816053, @fhahn wrote:

In D134663#3816040, @dblaikie wrote:

While I guess this is a fix for crash/UB, was the problem covered by any tests or could we add test coverage? (or maybe the failure's so unlikely as to be impractical to test - some ABA allocation situation which rarely occurs?)

Without this change, multiple transform tests fail when LLVM is built with ASAN, so this should be covered by existing tests + buildbots running ASAN.

Ah, sounds good to me, then (though I'll leave it to other folks more familiar with the code to approve) - maybe a link to/quote from one of those failures would be good to include in the commit message.

[nikic]

LGTM

[fhahn]

In D134663#3816095, @dblaikie wrote:

In D134663#3816053, @fhahn wrote:

In D134663#3816040, @dblaikie wrote:

Ah, sounds good to me, then (though I'll leave it to other folks more familiar with the code to approve) - maybe a link to/quote from one of those failures would be good to include in the commit message.

I added a note about how this is covered to the commit message but didn't have a link to a specific failure in buildbot handy.

[uabelho]

Hello,

I wrote
https://github.com/llvm/llvm-project/issues/58136
about a failed asseretion with -verify-scev that starts happening with this patch.

[qianzhen]

Hello @fhahn ,

I open https://github.com/llvm/llvm-project/issues/58454 for a new failed assertion on SCEV verify in LoopUnrollAndJam, starting with this patch.
Thanks!

llvm/lib/Analysis/ScalarEvolution.cpp

#endif
 
   // Verify intergity of loop disposition cache.
   for (const auto &It : LoopDispositions) {
     const SCEV *S = It.first;
     auto &Values = It.second;
     for (auto &V : Values) {
       auto CachedDisposition = V.getInt();
       const auto *Loop = V.getPointer();
-      // TODO: Make sure LoopDispositions contains no invalid loops.
-      if (!ValidLoops.contains(Loop))
-        continue;
       const auto RecomputedDisposition = SE2.getLoopDisposition(S, Loop);
       if (CachedDisposition != RecomputedDisposition) {
         dbgs() << "Cached disposition of " << *S << " for loop " << *Loop
                << " is incorrect: cached "
                << loopDispositionToStr(CachedDisposition) << ", actual "
                << loopDispositionToStr(RecomputedDisposition) << "\n";
         std::abort();
       }

llvm/lib/Transforms/Utils/LoopUtils.cpp

 void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE,
                              LoopInfo &LI, MemorySSA *MSSA) {
   auto *Latch = L->getLoopLatch();
   assert(Latch && "multiple latches not yet supported");
   auto *Header = L->getHeader();
   Loop *OutermostLoop = L->getOutermostLoop();
 
   SE.forgetLoop(L);
+  SE.forgetBlockAndLoopDispositions();
 
   std::unique_ptr<MemorySSAUpdater> MSSAU;
   if (MSSA)
     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
 
   // Update the CFG and domtree.  We chose to special case a couple of
   // of common cases for code quality and test readability reasons.
   [&]() -> void {