Hi!
The following patch implements what I've failed to do in time for 9.x,
in particular, if #pragma omp simd has if clause and the argument is
not constant zero (handled by the previous patch), or constant non-zero
(ignored, we want normal vectorization in that case as before), this patch
arrange
On Wed, 15 May 2019, Jakub Jelinek wrote:
> Hi!
>
> The following patch implements what I've failed to do in time for 9.x,
> in particular, if #pragma omp simd has if clause and the argument is
> not constant zero (handled by the previous patch), or constant non-zero
> (ignored, we want normal ve
On Thu, May 16, 2019 at 09:53:06AM +0200, Richard Biener wrote:
> > + if (nonconst_simd_if)
> > +{
> > + if (sctx.lane == NULL_TREE)
> > + {
> > + sctx.idx = create_tmp_var (unsigned_type_node);
> > + sctx.lane = create_tmp_var (unsigned_type_node);
> > + }
>
> Does forcing a
On Thu, 16 May 2019, Jakub Jelinek wrote:
> On Thu, May 16, 2019 at 09:53:06AM +0200, Richard Biener wrote:
> > > + if (nonconst_simd_if)
> > > +{
> > > + if (sctx.lane == NULL_TREE)
> > > + {
> > > + sctx.idx = create_tmp_var (unsigned_type_node);
> > > + sctx.lane = create_tmp_var
On Thu, May 16, 2019 at 11:30:38AM +0200, Richard Biener wrote:
> > note_simd_array_uses indeed does walk the IL and does look at the calls,
> > but I'd need some data structure where to store the argument; we don't have
> > loop_vinfo yet (we don't have it even before the loop over vector sizes),
On Thu, 16 May 2019, Jakub Jelinek wrote:
> On Thu, May 16, 2019 at 11:30:38AM +0200, Richard Biener wrote:
> > > note_simd_array_uses indeed does walk the IL and does look at the calls,
> > > but I'd need some data structure where to store the argument; we don't
> > > have
> > > loop_vinfo yet (
