On Mon, 10 Nov 2025, Victor Do Nascimento wrote:
> For multi-exit loops where the IV value used outside of the loop is
> may differ depending on which exit is taken, with it being updated
> prior to some exits, as per the example given below, we find that its
> value is not properly reset upon entering the epilog loop if the IV is
> updated prior to the exit due to the the following condition for
> setting the `early_break_first_element_p' predicate in the function
> `vectorizable_live_operation':
>
> STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def
>
> Consider, for example:
>
> while(c[i] != 0)
> {
> i += 1;
> if (c[i] == a)
> break;
> if (c[i] == b)
> break;
> d[i] = i;
> }
> return i;
>
> If we break out of the loop on while (c[i] != 0), the condition is
> satisfied.
>
> If, however, we break on any of the other 2 further conditions, the
> variable that is used outside of the loop is no longer classified as
> `vect_induction_def' but rather as the `vect_internal_def' argument
> passed to the phi-node via the loop latch.
>
> Irrespective of this, in resetting the loop values in preparation for
> the epilog loop, we still wish to rest its value to the zeroth value
> of the BITFIELD_REF of the the relevant vector.
>
> Therefore, we remove this condition from
> `early_break_first_element_p', leaving only `all_exits_as_early_p ||
> !main_exit_edge' as sufficient conditions for setting
> `early_break_first_element_p'.
This seems to be a fix independent of this series, so I wonder if
you have a testcase that is fixed by this? We'd likely want to
fix this on branches as well.
In fact I see this hunk in Tamars "support vectorization of early break
forced live IVs as scalar" patch as well, but without being mentioned
in the ChangeLog.
That said, can you two coordinate and split this out as a bug fix,
hopefully with test coverage?
Thanks,
Richard.
> ---
> gcc/tree-vect-loop.cc | 3 +--
> 1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 3b038169c95..35fa01548df 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -10337,8 +10337,7 @@ vectorizable_live_operation (vec_info *vinfo,
> stmt_vec_info stmt_info,
> to the latch then we're restarting the iteration in the
> scalar loop. So get the first live value. */
> bool early_break_first_element_p
> - = (all_exits_as_early_p || !main_exit_edge)
> - && STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def;
> + = (all_exits_as_early_p || !main_exit_edge);
> if (early_break_first_element_p)
> {
> tmp_vec_lhs = vec_lhs0;
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
