On February 10, 2017 8:45:09 PM GMT+01:00, Jakub Jelinek <ja...@redhat.com>
wrote:
>Hi!
>
>UBSAN_*_OVERFLOW ifns on vector types prefer to just check in a loop
>for overflow, but perform the vector arithmetics on vectors. The
>advantage
>is that the partial results don't have to be inserted one by one into
>the
>result vector.
>This requires that there is an instruction to perform that. We were
>checking that, except that I forgot that the vector type could have an
>scalar integral mode (like TImode), at which point it actually would
>perform
>say TImode addition rather than V16QImode addition.
>
>Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok
>for
>trunk?
OK.
Richard.
>2017-02-10 Jakub Jelinek <ja...@redhat.com>
>
> PR middle-end/79454
> * internal-fn.c (expand_vector_ubsan_overflow): Use piece-wise
> result computation whenever lhs doesn't have vector mode, not
> just when it has BLKmode.
>
>--- gcc/internal-fn.c.jj 2017-02-09 14:55:34.000000000 +0100
>+++ gcc/internal-fn.c 2017-02-10 15:58:50.724312535 +0100
>@@ -1909,7 +1909,7 @@ expand_vector_ubsan_overflow (location_t
> {
> optab op;
> lhsr = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
>- if (GET_MODE (lhsr) == BLKmode
>+ if (!VECTOR_MODE_P (GET_MODE (lhsr))
> || (op = optab_for_tree_code (code, TREE_TYPE (arg0),
> optab_default)) == unknown_optab
> || (optab_handler (op, TYPE_MODE (TREE_TYPE (arg0)))
>
> Jakub
