[LLVMdev] incorrect DSCallGraph for simple indirect call with vtable nearby

[Ben Liblit]

Will Dietz wrote:
> This is actually the expected behavior for EQTD :).

Expected by you, maybe.  :-D

> If you switch to TD you'll get better alias-analysis information, and
> in this example the correct result.

OK, I have switched to TDDataStructures as well, and I am also seeing 
much better (for my purposes) results in simple tests.  I'll keep poking 
at this some more and let you folks know if I hit any other surprises.

There is no mention of TDDataStructures in 
"poolalloc/dsa-manual/manual.tex", which is why I did not consider using 
it earlier.  Can you give me a high-level idea of how TDDataStructures 
and EQTDDataStructures differ?  Clearly they differ in the "gives the 
answers Ben expected" dimension, but that's not a very useful 
description more generally.  What should I know about the key 
differences between these two?

> FWIW at the moment DSA doesn't give good results for vtable-heavy
> code, marking all such callsites as incomplete and cannot resolve
> them.

Understood.  My plan for vtable-based calls is to pattern-match on the 
specific code sequences emitted by clang or whatever front end we are 
using.  Likewise I'll be exploiting C++ domain knowledge to recognize 
global variables that are vtables.  Once the clang type-system-rewrite 
work is released, that will also help me a lot as I'll have better 
information about the static types at each vtable call site.  So I think 
I'm in reasonable shape for vtables.  [I was prepared to be pleasantly 
surprised if DSA figured these out for me, but I certainly was not 
expecting or assuming that.]

Thanks again, Will, John, and Andrew.  You've been most helpful.  It's 
great to see that the high quality of LLVM's code is matched by the high 
quality of support found on this mailing list.

Regards,
Ben