Re: [PATCH] Improve dead code elimination with -fsanitize=address (PR84307)

[Paolo Bonzini <bonzini at gnu dot org>]

On 13/02/2018 14:00, Jakub Jelinek wrote:
>> Certainly, for now I'll revert.
> Reversion is not the right thing, the "fn spec" attributes were clearly
> incorrect. So, we should change them to something more conservative that
> will work.

That would only be "all dots", that is no fnspec at all.  Martin
suggested removing EAF_DIRECT, but I don't think I agree with his
reasoning.  Besides, aliasing doesn't see the shadow memory at all (see
call_may_clobber_ref_p_1), so it's okay to ignore it for the sake of
fnspecs.

>> But can you expand on why it's too early?  Indeed I suppose it may
>> affect inlining decisions, on the other hand it seems dangerous to apply
>> instrumentation after pretty much any optimization pass.
>
> It will prevent pretty much all optimizations.  We don't want -O2
> -fsanitize=address to be unusably slow, if people want to catch everything,
> they can always use -O0 -fsanitize=address.  The current placement of the
> passes has been a result of long discussions if I remember well.

I'm not sure it will be that bad, together with the fnspec.  Consider
that PR84340 is latent in current GCC; the testcases work because GCC
thinks that the &x pointer escaped, and thus treated the stores as not
dead.  In other words, -fsanitize=address -O2 _currently_ lacks an awful
lot of aliasing-based optimizations such as DSE, because all variables
are marked as escaping after the initial ASAN_MARK(UNPOISON, &var, sz).

With some luck (that we can ascertain between now and stage 1) the
negative effects of pass placement balance with the positive effects of
the fnspec.  But I agree that it requires some discussion and benchmarking.

Paolo