> -----Original Message-----
> From: Alfie Richards <[email protected]>
> Sent: 20 July 2026 11:18
> To: [email protected]
> Cc: [email protected]; Tamar Christina <[email protected]>;
> [email protected]; [email protected]; [email protected]; Alfie
> Richards <[email protected]>
> Subject: [PATCH] vect: Fix vect_set_loop_controls_directly when
> LOOP_VINFO_EARLY_BREAKS_VECT_PEELED [PR 126301]
>
> When the partial loop control check is within the loop body, use the
> current control instead of the next loop control for the latch.
>
> gcc/ChangeLog:
>
> PR tree-optimization/126301
>
> * tree-vect-loop-manip.cc (vect_set_loop_controls_directly):
> Change to use current iteration control when
> LOOP_VINFO_EARLY_BREAKS_VECT_PEELED.
> (vect_set_loop_condition_partial_vectors_avx512):
> Change to use current iteration control when
> LOOP_VINFO_EARLY_BREAKS_VECT_PEELED.
>
> -- >8 --
>
> Fixes the linked PR but has the undesirable side effect of separating
> the check and branch from the mask definitions via branches so they can
> no longer be combined.
>
I've discussed with Alfie offline in that I think both problems can be solved
by doing
this differently and instead rewrite the loop control rather than just changing
the
feeding statement.
The loop being vectorized is essentially a do-while loop, but when we vectorize
we
insert an unneeded initial check because we don't make use of the fact that
it's a
PEELED loop to generate the loop control.
I'll send a patch for this once regressions finish.
Thanks,
Tamar
> I am working on a pass for that separately.
>
> Reg tested and bootstrapped on aarch64, x86_64, armhf
>
> KR,
> Alfie
>
> ---
> gcc/tree-vect-loop-manip.cc | 22 ++++++++++++++++++++--
> 1 file changed, 20 insertions(+), 2 deletions(-)
>
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index ec9dec1b713..c69aea41ad3 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -511,6 +511,10 @@ vect_set_loop_controls_directly (class loop *loop,
> loop_vec_info loop_vinfo,
> tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
>
> + /* Check if this control will be the main exit for the loop. */
> + bool main_exit_p = single_pred (LOOP_VINFO_LOOP (loop_vinfo)->latch)
> + == LOOP_VINFO_MAIN_EXIT (loop_vinfo)->src;
> +
> tree ctrl_type = rgc->type;
> unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
> poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc-
> >factor;
> @@ -714,7 +718,8 @@ vect_set_loop_controls_directly (class loop *loop,
> loop_vec_info loop_vinfo,
>
> /* Provide a definition of each control in the group. */
> tree next_ctrl = NULL_TREE;
> - tree ctrl;
> + tree ctrl = NULL_TREE;
> + tree exit_ctrl = NULL_TREE;
> unsigned int i;
> FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
> {
> @@ -822,6 +827,10 @@ vect_set_loop_controls_directly (class loop *loop,
> loop_vec_info loop_vinfo,
> gsi_insert_seq_before (test_gsi, seq, GSI_SAME_STMT);
> }
>
> + /* If this is the main exit for the loop, then this will be in the
> header
> + of the loop and so we should check next_ctrl. Otherwise, this is
> + in the body of the loop, and so we should use ctrl. */
> + exit_ctrl = main_exit_p ? next_ctrl : ctrl;
> vect_set_loop_control (loop, ctrl, init_ctrl, next_ctrl);
> }
>
> @@ -837,7 +846,7 @@ vect_set_loop_controls_directly (class loop *loop,
> loop_vec_info loop_vinfo,
> gimple_seq_add_stmt (header_seq, minus);
> }
>
> - return next_ctrl;
> + return exit_ctrl;
> }
>
> /* Set up the iteration condition and rgroup controls for LOOP, given
> @@ -1023,6 +1032,10 @@ vect_set_loop_condition_partial_vectors_avx512
> (class loop *loop,
> tree orig_niters = niters;
> gimple_seq preheader_seq = NULL;
>
> + /* Check if this control will be the main exit for the loop. */
> + bool main_exit_p = single_pred (LOOP_VINFO_LOOP (loop_vinfo)->latch)
> + == LOOP_VINFO_MAIN_EXIT (loop_vinfo)->src;
> +
> /* Create an IV that counts down from niters and whose step
> is the number of iterations processed in the current iteration.
> Produce the controls with compares like the following.
> @@ -1113,6 +1126,7 @@ vect_set_loop_condition_partial_vectors_avx512
> (class loop *loop,
> tree next_ctrl = NULL_TREE;
> tree first_rem = NULL_TREE;
> tree ctrl;
> + tree exit_ctrl = NULL_TREE;
> unsigned int i;
> FOR_EACH_VEC_ELT_REVERSE (rgc.controls, i, ctrl)
> {
> @@ -1201,6 +1215,10 @@ vect_set_loop_condition_partial_vectors_avx512
> (class loop *loop,
> UNKNOWN_LOCATION,
> LT_EXPR, ctrl_type, cmp_series, rem);
>
> + /* If this is the main exit for the loop, then this will be in the
> + header of the loop and so we should use next_ctrl. Otherwise,
> + this is in the body of the loop, and so we should use ctrl. */
> + exit_ctrl = main_exit_p ? next_ctrl : ctrl;
> vect_set_loop_control (loop, ctrl, init_ctrl, next_ctrl);
> }
> }
> --
> 2.34.1
