https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97307
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #3)
> The patch will cause
>
> FAIL: gcc.dg/vect/pr65947-3.c scan-tree-dump-times vect "LOOP VECTORIZED" 2
>
> since the testcase has exactly such a pattern:
>
> unsigned int
> condition_reduction (unsigned int *a, unsigned int min_v, unsigned int *b)
> {
> unsigned int last = N + 65;
> unsigned int aval;
>
> for (unsigned int i = 0; i < N; i++)
> {
> aval = a[i];
> if (b[i] < min_v)
> last = aval;
> }
> return last;
> }
>
> and sinking the aval = a[i] load causes if-conversion to no longer happen
> because a[i] now appears to only conditionally trap. It will be vectorized
> with masked loads if those are available though, so I'm considering to
> XFAIL it for the moment.
Isn't that an important real-world case though? I mean, people who care a lot
about the vectorization often change code from the loads in the conditionals to
preloading the value before the condition, such that it can be vectorized and
now it won't (and masked loads are less common among architectures that can
vectorize such loops than non-masked ones).
So shouldn't we just sink that way calls before vectorization and after
vectorization everything? Or have some way to tell ifcvt it can undo the
sinking.
