[CallGraphUpdater][FIX] Do not attempt to update unreachable functions

[jdoerfert]

Summary

In the sqlite3 build, under certain conditions, we ended up with an CG
update order in which an internal function ended up being unreachable
before CallGraphUpdate::reanalyzeFunction was called on it. That
resulted in a segfault as the function did not belong to any SCC
anymore. We now detect this and avoid to reanalyze such unreachable
functions.

I failed to simplify the 1.9MB sqlite3.bc file as conversion into .ll,
e.g., as performed by llvm-reduce, causes the problem to disappear.

[uenoku]

LGTM

[aeubanks]

Could you give an example stack trace? I'm not 100% sure this is the right fix.

[jdoerfert]

In D89188#2327975, @aeubanks wrote:

Could you give an example stack trace? I'm not 100% sure this is the right fix.

Will do. Need to go back to triaging all the known Attributor bugs, though more likely next week. Anyone that can reduce Attributor bugs should feel freeto do so.

[lebedev.ri]

This review seems to be stuck/dead, consider abandoning if no longer relevant.

llvm/lib/Transforms/Utils/CallGraphUpdater.cpp

 void CallGraphUpdater::reanalyzeFunction(Function &Fn) {
   if (CG) {
     CallGraphNode *OldCGN = CG->getOrInsertFunction(&Fn);
     OldCGN->removeAllCalledFunctions();
     CG->populateCallGraphNode(OldCGN);
   } else if (LCG) {
     LazyCallGraph::Node &N = LCG->get(Fn);
     LazyCallGraph::SCC *C = LCG->lookupSCC(N);
+    // If this function ended up outside of an SCC, through prior updates, we
+    // cannot reanalyze it, nor is it necessary.
+    if (C)
       updateCGAndAnalysisManagerForCGSCCPass(*LCG, *C, N, *AM, *UR, *FAM);
   }
 }
 
 void CallGraphUpdater::registerOutlinedFunction(Function &NewFn) {
   if (CG)
     CG->addToCallGraph(&NewFn);
 }