https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102700
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 #5)
> Vectorization can only maintain one of the locations:
>
> - [t.ii:9:8] _6 = foo.ptr;
> - [t.ii:9:8] [t.ii:9:8] D.2488.ptr = _6;
> + [t.ii:9:8] vect__6.9_12 = MEM <vector(2) int> [(int *)&foo];
> [t.ii:4:5] MEM[(struct shared_count *)[t.ii:20:3] &D.2488 + 4B] ={v}
> {CLOBBER};
> - [t.ii:4:13] _7 = [t.ii:4:13] MEM[(struct shared_count &)[t.ii:20:11] &foo
> + 4].pi;
> - [t.ii:4:7] [t.ii:4:7] MEM[(struct shared_count *)[t.ii:20:3] &D.2488 +
> 4B].pi = _7;
> + [t.ii:9:8] MEM <vector(2) int> [(int *)&D.2488] = vect__6.9_12;
>
> so we combine the reads at t.ii:9:8 and t.ii:4:7 (that's the uninitialized
> one)
> into a vector read at t.ii:9:8. I don't see how GCCs IL allows us to do
> better.
Well, if DWARF6 agrees on some way how to debug vectorized code, maybe we'll
need to figure out a way how to encode multiple locations.
