Thanks Richi.
The following is the V2 patch:
Testing on X86 and aarch64 are running.
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 43d742e3c92..e7f7f976f11 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -760,7 +760,7 @@ vect_get_and_check_slp_defs (vec_info *vinfo, unsigned char
swap,
|| dt == vect_external_def)
&& !GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
&& (TREE_CODE (type) == BOOLEAN_TYPE
- || !can_duplicate_and_interleave_p (vinfo, stmts.length (),
+ && !can_duplicate_and_interleave_p (vinfo, stmts.length (),
type)))
{
if (dump_enabled_p ())
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 6ce4868d3e1..6c47121e158 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -9859,10 +9859,16 @@ vectorizable_load (vec_info *vinfo,
mask_index = internal_fn_mask_index (ifn);
if (mask_index >= 0 && slp_node)
mask_index = vect_slp_child_index_for_operand (call, mask_index);
+ slp_tree slp_op = NULL;
if (mask_index >= 0
&& !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index,
- &mask, NULL, &mask_dt, &mask_vectype))
+ &mask, &slp_op, &mask_dt, &mask_vectype))
return false;
+ /* MASK_LEN_GATHER_LOAD dummy mask -1 should always match the
+ MASK_VECTYPE. */
+ if (mask_index >= 0 && slp_node && mask_dt == vect_constant_def
+ && !vect_maybe_update_slp_op_vectype (slp_op, mask_vectype))
+ gcc_unreachable ();
}
juzhe.zh...@rivai.ai
From: Richard Biener
Date: 2023-11-02 19:11
To: Juzhe-Zhong
CC: gcc-patches; richard.sandiford
Subject: Re: [tree-optimization/111721] VECT: Support SLP for
MASK_LEN_GATHER_LOAD with dummy mask
On Thu, 2 Nov 2023, Juzhe-Zhong wrote:
> This patch fixes following FAILs for RVV:
> FAIL: gcc.dg/vect/vect-gather-1.c -flto -ffat-lto-objects scan-tree-dump
> vect "Loop contains only SLP stmts"
> FAIL: gcc.dg/vect/vect-gather-1.c scan-tree-dump vect "Loop contains only SLP
> stmts"
>
> Bootstrap on X86 and regtest passed.
>
> Tested on aarch64 passed.
>
> Ok for trunk ?
>
> PR tree-optimization/111721
>
> gcc/ChangeLog:
>
> * tree-vect-slp.cc (vect_get_and_check_slp_defs): Support SLP for
> dummy mask -1.
> * tree-vect-stmts.cc (vectorizable_load): Ditto.
>
> ---
> gcc/tree-vect-slp.cc | 14 ++++++++++++--
> gcc/tree-vect-stmts.cc | 8 +++++++-
> 2 files changed, 19 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index 43d742e3c92..23ca0318e31 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -756,8 +756,7 @@ vect_get_and_check_slp_defs (vec_info *vinfo, unsigned
> char swap,
> {
> tree type = TREE_TYPE (oprnd);
> dt = dts[i];
> - if ((dt == vect_constant_def
> - || dt == vect_external_def)
> + if (dt == vect_external_def
> && !GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
> && (TREE_CODE (type) == BOOLEAN_TYPE
> || !can_duplicate_and_interleave_p (vinfo, stmts.length (),
> @@ -769,6 +768,17 @@ vect_get_and_check_slp_defs (vec_info *vinfo, unsigned
> char swap,
> "for variable-length SLP %T\n", oprnd);
> return -1;
> }
> + if (dt == vect_constant_def
> + && !GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
> + && !can_duplicate_and_interleave_p (vinfo, stmts.length (), type))
> + {
> + if (dump_enabled_p ())
> + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> + "Build SLP failed: invalid type of def "
> + "for variable-length SLP %T\n",
> + oprnd);
> + return -1;
> + }
I don't think that's quite correct. can_duplicate_and_interleave_p
doesn't get enough info here and IIRC even materializing arbitrary
constants isn't possible with VLA vectors. The very first thing
the function does is
tree base_vector_type = get_vectype_for_scalar_type (vinfo, elt_type,
count);
if (!base_vector_type || !VECTOR_MODE_P (TYPE_MODE (base_vector_type)))
return false;
but for masks that's not going to get us the correct vector type.
While I don't understand why we have that 'BOOLEAN_TYPE' special
case (maybe the intent was to identify 'mask' operands that way?),
we might want to require that we can materialize both all-zero
and all-ones constant 'mask's. But then 'mask' operands should
be properly identified here.
Maybe we can also simply delay the check to the point we know
whether we're facing an uniform constant or not (note for 'first',
we cannot really special-case vect_constant_def as the second
SLP lane might demote that to vect_external_def). It's always
a balance of whether to reject sth at SLP build time (possibly
allowing operand swapping to do magic) or to delay checks
to stmt analysis time. That might also explain that you
do not see fallout of the "wrong" change (the later checking
will catch it anyway).
There's probably testsuite coverage for SVE here.
That said, a "correct" patch might be to simply change
&& (TREE_CODE (type) == BOOLEAN_TYPE
|| !can_duplicate_and_interleave_p (vinfo, stmts.length
(),
type)))
to
&& TREE_CODE (type) != BOOLEAN_TYPE
&& !can_duplicate_and_interleave_p (vinfo, stmts.length
(),
type)
thus delay 'mask' operand validation here.
Note I still think we should improve TREE_CODE (type) == BOOLEAN_TYPE
to identify internal function mask operands only.
Richard.
>
> /* For the swapping logic below force vect_reduction_def
> for the reduction op in a SLP reduction group. */
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 6ce4868d3e1..6c47121e158 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -9859,10 +9859,16 @@ vectorizable_load (vec_info *vinfo,
> mask_index = internal_fn_mask_index (ifn);
> if (mask_index >= 0 && slp_node)
> mask_index = vect_slp_child_index_for_operand (call, mask_index);
> + slp_tree slp_op = NULL;
> if (mask_index >= 0
> && !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index,
> - &mask, NULL, &mask_dt, &mask_vectype))
> + &mask, &slp_op, &mask_dt, &mask_vectype))
> return false;
> + /* MASK_LEN_GATHER_LOAD dummy mask -1 should always match the
> + MASK_VECTYPE. */
> + if (mask_index >= 0 && slp_node && mask_dt == vect_constant_def
> + && !vect_maybe_update_slp_op_vectype (slp_op, mask_vectype))
> + gcc_unreachable ();
> }
>
> tree vectype = STMT_VINFO_VECTYPE (stmt_info);
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
