[PATCH] Fix #pragma omp atomic with floating point types (PR middle-end/36106)

[Zdenek Dvorak]

Hi,

> Fixed by doing the loading in the integral type, using DECL_NO_TBAA_P on
> the pointer, and VCEing afterwards.
> CCing Diego/Zdenek, in case I've messed up the gimple_in_ssa_p (cfun)
> aka -ftree-parallelize-loops=N cases (I've tested on attached autopar-1.c)
> or if DECL_NO_TBAA_P is too risky (doing memcpy into a temporary would be
> slower though, and __sync_val_compare_and_swap stores/reads using alias set
> 0 anyway).

I do not see a problem with this (but I am not really confident about
the alias analysis consequences, I'd suggest to wait for Diego's
opinion),

Zdenek