> -----Original Message-----
> From: Alfie Richards <[email protected]>
> Sent: 17 September 2026 15:40
> To: [email protected]
> Cc: Alex Coplan <[email protected]>; Alice Carlotti
> <[email protected]>; Andrea Pinski
> <[email protected]>; Richard Biener <[email protected]>;
> Tamar Christina <[email protected]>; Zdenek Dvorak <[email protected]>;
> Robin Dapp <[email protected]>; Richard Sandiford
> <[email protected]>; Kyrylo Tkachov <[email protected]>;
> Alfie Richards <[email protected]>
> Subject: [PATCH v3 04/16] vect: Make vect_maybe_permute_loop_masks
> optional and retargetable
>
> Change vect_maybe_permute_loop_masks so that it can check if the
> permutation is possible without code generation.
>
> The original function is expanded to allow retargetting which source
> and destination mask arrays should be used. This allows it to be used
> for HSSR.
>
> The previous interface is available as an override where the controls
> are assumed to be from the rgroup_controls.
>
> gcc/ChangeLog:
>
> * tree-vect-loop-manip.cc (vect_maybe_permute_loop_masks):
> Make the function optional and retargetable.
> (vect_maybe_permute_loop_masks):
> New overload.
> ---
> gcc/tree-vect-loop-manip.cc | 41 +++++++++++++++++++++++++++++------
> --
> 1 file changed, 32 insertions(+), 9 deletions(-)
>
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 78445e1f9aa..103657245a8 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -324,18 +324,28 @@ interleave_supported_p (vec_perm_indices
> *indices, tree vectype,
> *indices);
> }
>
> -/* Try to use permutes to define the masks in DEST_RGM using the masks
> - in SRC_RGM, given that the former has twice as many masks as the
> - latter. Return true on success, adding any new statements to SEQ. */
> +/* Check if its possible to use permutes to define the masks in DEST_RGM
> using
> + the masks in SRC_RGM, given that the former has twice as many masks as
> the
> + latter. Return true on success.
> +
> + Optionally, if SEQ is non-null add new statements for the permutation to
> SEQ.
> + Additionally, if SRC_MASKS and DST_MASKS are defined, use these instead
> + of the masks from SRC_RGM and DST_RGM. */
>
> static bool
> -vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls
> *dest_rgm,
> - rgroup_controls *src_rgm)
> +vect_maybe_permute_loop_masks (gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm,
> + vec<tree> dst_controls,
> + vec<tree> src_controls)
> {
> tree src_masktype = src_rgm->type;
> tree dest_masktype = dest_rgm->type;
> + if (!src_masktype && !seq)
> + return false;
> machine_mode src_mode = TYPE_MODE (src_masktype);
> insn_code icode1, icode2;
> +
> if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter
I assume these two variables and the type are why you kept the rgroup_controls,
but..
> && (icode1 = optab_handler (vec_unpacku_hi_optab,
> src_mode)) != CODE_FOR_nothing
> @@ -347,10 +357,12 @@ vect_maybe_permute_loop_masks (gimple_seq
> *seq, rgroup_controls *dest_rgm,
> machine_mode dest_mode = insn_data[icode1].operand[0].mode;
> gcc_assert (dest_mode == insn_data[icode2].operand[0].mode);
> tree unpack_masktype = vect_halve_mask_nunits (src_masktype,
> dest_mode);
> + if (!seq)
> + return true;
> for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
Here and below don't you want to iterate over dst_controls.length()
and src_control.length()?
Otherwise I'd expect at least an assert here that the two are the same.
But I'd prefer the former.
So Patch is OK with the iteration lengths changed.
Thanks,
Tamar
> {
> - tree src = src_rgm->controls[i / 2];
> - tree dest = dest_rgm->controls[i];
> + tree src = src_controls[i / 2];
> + tree dest = dst_controls[i];
> tree_code code = ((i & 1) == (BYTES_BIG_ENDIAN ? 0 : 1)
> ? VEC_UNPACK_HI_EXPR
> : VEC_UNPACK_LO_EXPR);
> @@ -375,6 +387,8 @@ vect_maybe_permute_loop_masks (gimple_seq *seq,
> rgroup_controls *dest_rgm,
> && interleave_supported_p (&indices[0], src_masktype, 0)
> && interleave_supported_p (&indices[1], src_masktype, 1))
> {
> + if (!seq)
> + return true;
> /* The destination requires twice as many mask bits as the source, so
> we can use interleaving permutes to double up the number of bits. */
> tree masks[2];
> @@ -382,8 +396,8 @@ vect_maybe_permute_loop_masks (gimple_seq *seq,
> rgroup_controls *dest_rgm,
> masks[i] = vect_gen_perm_mask_checked (src_masktype, indices[i]);
> for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> {
> - tree src = src_rgm->controls[i / 2];
> - tree dest = dest_rgm->controls[i];
> + tree src = src_controls[i / 2];
> + tree dest = dst_controls[i];
> gimple *stmt = gimple_build_assign (dest, VEC_PERM_EXPR,
> src, src, masks[i & 1]);
> gimple_seq_add_stmt (seq, stmt);
> @@ -393,6 +407,15 @@ vect_maybe_permute_loop_masks (gimple_seq
> *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +static bool
> +vect_maybe_permute_loop_masks (gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + return vect_maybe_permute_loop_masks (seq, dest_rgm, src_rgm,
> + dest_rgm->controls, src_rgm-
> >controls);
> +}
> +
> /* Populate DEST_RGM->controls, given that they should add up to STEP.
>
> STEP = MIN_EXPR <ivtmp_34, VF>;
> --
> 2.43.0
