2016-01-11 20:13 GMT+03:00 Jakub Jelinek <ja...@redhat.com>:
> Hi!
>
> Based on discussions on IRC, I'm submitting following fix for a regression
> on aarch64 - partial reversion (the case where VCE works too, just I thought
> using NOP_EXPR would be nicer) and change in the assert - op better be
> some integral value if converting it to VECTOR_BOOLEAN_TYPE_P's element
> type.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2016-01-11 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/69207
> * tree-vect-slp.c (vect_get_constant_vectors): For
> VECTOR_BOOLEAN_TYPE_P, assert op has integral type instead of
> fold_convertible_p to vector_type's element type, and always
> use VCE for non-VECTOR_BOOLEAN_TYPE_P.
>
> --- gcc/tree-vect-slp.c.jj 2016-01-08 21:45:57.000000000 +0100
> +++ gcc/tree-vect-slp.c 2016-01-11 12:07:19.633366712 +0100
> @@ -2999,12 +2999,9 @@ vect_get_constant_vectors (tree op, slp_
> gimple *init_stmt;
> if (VECTOR_BOOLEAN_TYPE_P (vector_type))
> {
> - gcc_assert (fold_convertible_p (TREE_TYPE (vector_type),
> - op));
> + gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op)));
> init_stmt = gimple_build_assign (new_temp, NOP_EXPR,
> op);
In vect_init_vector we had to introduce COND_EXPR to choose between 0 and -1 for
boolean vectors. Shouldn't we do similar in SLP?
Thanks,
Ilya
> }
> - else if (fold_convertible_p (TREE_TYPE (vector_type), op))
> - init_stmt = gimple_build_assign (new_temp, NOP_EXPR, op);
> else
> {
> op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type),
>
> Jakub
