[PATCH] D39356: [ThinLTO] Fix missing call graph edges for calls with bitcasts.

[Teresa Johnson via Phabricator via llvm-commits]

tejohnson added a comment.

In https://reviews.llvm.org/D39356#917074, @efriedma wrote:

> > The issue here is that we detected that this GV was a Callee, so we didn't add a reference edge.
>
> Are you sure that's what's happening?  The isCallee test is an exact equality comparison, so it should fail if we're calling a bitcasted global.


Hmm, I think that is a good question, looking at the code again. @vsapsai, can you see why the reference wasn't recorded in findRefEdges?

> 
> 
>> The crucial part is distinction between int ext(); and int ext(void);.
> 
> Yes, this is exactly right.

Got it, thanks.



================
Comment at: llvm/lib/Analysis/ModuleSummaryAnalysis.cpp:283
           continue;
-        // Skip direct calls.
-        if (!CS.getCalledValue() || isa<Constant>(CS.getCalledValue()))
-          continue;
+        assert(CalledValue && !isa<Constant>(CalledValue) &&
+               "Expected indirect call");
----------------
I'm a little worried that we may have a case with CalledValue==null (presumably it wasn't an accident that we were checking for a null CalledValue before). Anyone know if that is valid?


https://reviews.llvm.org/D39356