[Bug ipa/87843] [9 Regression] SPEC miscompilation of 403.gcc and 502.gcc_r benchmarks

[hubicka at ucw dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87843

--- Comment #7 from Jan Hubicka <hubicka at ucw dot cz> ---
> If we have less MEM_REFs then we probably strip them because we think they
> reference equal types.
> 
> I think I already told you that given that MEM_REFs use pointer types
> to carry alignment info _those_ may not become incomplete!  But I didn't
> expect that to cause wrong-code but missed optimizations.

We do not make them incomplete.  The problem actually seems to be in
early optimization where we optimize out the if conditional above.
Not sure why -ffat-lto-objects worked in this context.