https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98167
--- Comment #18 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Andrew Pinski from comment #17)
> (In reply to Hongtao.liu from comment #16)
> > typedef int v4si __attribute__ ((vector_size(16)));
> >
> > v4si f(v4si a, v4si b) {
> > v4si a1 = __builtin_shufflevector (a, a, 2, 3 ,1 ,0);
> > v4si b1 = __builtin_shufflevector (b, a, 2, 3 ,1 ,0);
> > return a1 * b1;
> > }
> >
> > gcc generate
> >
> > f:
> > vpshufd xmm1, xmm1, 30
> > vpshufd xmm0, xmm0, 30
> > vpmulld xmm0, xmm0, xmm1
> > ret
> >
> > llvm generate
> >
> > f: # @f
> > vpmulld xmm0, xmm1, xmm0
> > vpshufd xmm0, xmm0, 30 # xmm0 = xmm0[2,3,1,0]
> > ret
>
> For the above, this is safe for -ftrapping-math as all elements are still
> used. It is when elements that are not used it might not be safe ...
For vector integers it should be ok?
For vector floating point we can add condition flag_unsafe_math_optimizations
|| !flag_trapping_math for the optimization.
