[Bug rtl-optimization/68755] [4.9/5 Regression] wrong code at -O3 on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68755

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Markus Trippelsdorf from comment #7)
> --param=allow-store-data-races=1 fixes the issue.
> With --param=allow-store-data-races=0 even 4.9 generates wrong code.

That sounds backwards btw.

It's predictive commoning performing the code motion but the load
is still guarded by the header copy check (so it's a valid transform).

Only sth at the RTL optimization stage then hoists the load before the
if.  AFAIK it's ifcvt2 and -fno-if-conversion2 fixes it.