On Fri, May 31, 2013 at 10:20:01AM +0200, Richard Biener wrote:
> The limit is there because a loop with more than one basic-block with code
> necessarily has to have conditionally executed BBs and eventually PHI nodes
> at merge points.
>
> Now, it may be that we properly determine if we can handle the PHIs in
> the non-header BB and that we properly bail out if we hit a conditional
> statement. But especially the latter would mean that we would not vectorize
> the loop.
Note that for OpenMP #pragma omp simd as well as Cilk+ #pragma simd we
actually want to be able to do more than that in the future, part of it
is something that can be done even for normal loops, but part of it would be
only if we detected that all iterations (or at least within vf) are
completely independent.
1) with better infrastructure we can do at least on some ISAs conditional
loads or conditional stores
2) calls to elemental functions (vector attribute in Cilk+, #pragma omp
declare simd) if the elemental function allows masking can be also
if-converted in some way
3) even if if-conversion of 1) and 2) above doesn't help, for the simd loops
we just want to fall back to just running vf scalar iterations of some bb
that couldn't be if-converted, where we can have say first and third bb
vectorized normally and second run in a loop from 0 to vf - 1, etc.
But none of this is anywhere close to the patchlets that have been posted.
Jakub
