| Message ID | 20230721100532.1342700-1-juzhe.zhong@rivai.ai |
|---|---|
| State | New |
| Headers | show |
| Series | [V2] VECT: Support floating-point in-order reduction for length loop control | expand |
On Fri, 21 Jul 2023, juzhe.zhong@rivai.ai wrote: > From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai> > > Hi, Richard and Richi. > > This patch support floating-point in-order reduction for loop length control. > > Consider this following case: > > float foo (float *__restrict a, int n) > { > float result = 1.0; > for (int i = 0; i < n; i++) > result += a[i]; > return result; > } > > When compile with **NO** -ffast-math on ARM SVE, we will end up with: > > loop_mask = WHILE_ULT > result = MASK_FOLD_LEFT_PLUS (...loop_mask...) > > For RVV, we don't use length loop control instead of mask: > > So, with this patch, we expect to see: > > loop_len = SELECT_VL > result = MASK_LEN_FOLD_LEFT_PLUS (...loop_len...) > > gcc/ChangeLog: > > * tree-vect-loop.cc (get_masked_reduction_fn): Add mask_len_fold_left. > (vectorize_fold_left_reduction): Ditto. > (vectorizable_reduction): Ditto. > (vect_transform_reduction): Ditto. > > --- > gcc/tree-vect-loop.cc | 41 ++++++++++++++++++++++++++++++++++++----- > 1 file changed, 36 insertions(+), 5 deletions(-) > > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc > index b44fb9c7712..59ab7879d55 100644 > --- a/gcc/tree-vect-loop.cc > +++ b/gcc/tree-vect-loop.cc > @@ -6800,11 +6800,13 @@ static internal_fn > get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) > { > internal_fn mask_reduc_fn; > + internal_fn mask_len_reduc_fn; > > switch (reduc_fn) > { > case IFN_FOLD_LEFT_PLUS: > mask_reduc_fn = IFN_MASK_FOLD_LEFT_PLUS; > + mask_len_reduc_fn = IFN_MASK_LEN_FOLD_LEFT_PLUS; > break; > > default: > @@ -6814,6 +6816,9 @@ get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) > if (direct_internal_fn_supported_p (mask_reduc_fn, vectype_in, > OPTIMIZE_FOR_SPEED)) > return mask_reduc_fn; > + if (direct_internal_fn_supported_p (mask_len_reduc_fn, vectype_in, > + OPTIMIZE_FOR_SPEED)) > + return mask_len_reduc_fn; > return IFN_LAST; > } > > @@ -6834,7 +6839,8 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, > gimple *reduc_def_stmt, > tree_code code, internal_fn reduc_fn, > tree ops[3], tree vectype_in, > - int reduc_index, vec_loop_masks *masks) > + int reduc_index, vec_loop_masks *masks, > + vec_loop_lens *lens) > { > class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); > tree vectype_out = STMT_VINFO_VECTYPE (stmt_info); > @@ -6896,8 +6902,18 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, > { > gimple *new_stmt; > tree mask = NULL_TREE; > + tree len = NULL_TREE; > + tree bias = NULL_TREE; > if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) > mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, i); > + if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) > + { > + len = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, vectype_in, > + i, 1); > + signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); > + bias = build_int_cst (intQI_type_node, biasval); > + mask = build_minus_one_cst (truth_type_for (vectype_in)); > + } > > /* Handle MINUS by adding the negative. */ > if (reduc_fn != IFN_LAST && code == MINUS_EXPR) > @@ -6917,7 +6933,10 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, > the preceding operation. */ > if (reduc_fn != IFN_LAST || (mask && mask_reduc_fn != IFN_LAST)) > { > - if (mask && mask_reduc_fn != IFN_LAST) > + if (len && mask && mask_reduc_fn != IFN_LAST) check mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS instead? > + new_stmt = gimple_build_call_internal (mask_reduc_fn, 5, reduc_var, > + def0, mask, len, bias); > + else if (mask && mask_reduc_fn != IFN_LAST) Likewise. Otherwise looks good to me. Richard. > new_stmt = gimple_build_call_internal (mask_reduc_fn, 3, reduc_var, > def0, mask); > else > @@ -7979,6 +7998,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo, > else if (loop_vinfo && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) > { > vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); > + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); > internal_fn cond_fn = get_conditional_internal_fn (op.code, op.type); > > if (reduction_type != FOLD_LEFT_REDUCTION > @@ -8006,8 +8026,17 @@ vectorizable_reduction (loop_vec_info loop_vinfo, > LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; > } > else > - vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, > - vectype_in, NULL); > + { > + internal_fn mask_reduc_fn > + = get_masked_reduction_fn (reduc_fn, vectype_in); > + > + if (mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS) > + vect_record_loop_len (loop_vinfo, lens, ncopies * vec_num, > + vectype_in, 1); > + else > + vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, > + vectype_in, NULL); > + } > } > return true; > } > @@ -8137,6 +8166,7 @@ vect_transform_reduction (loop_vec_info loop_vinfo, > code_helper code = canonicalize_code (op.code, op.type); > internal_fn cond_fn = get_conditional_internal_fn (code, op.type); > vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); > + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); > bool mask_by_cond_expr = use_mask_by_cond_expr_p (code, cond_fn, vectype_in); > > /* Transform. */ > @@ -8162,7 +8192,8 @@ vect_transform_reduction (loop_vec_info loop_vinfo, > gcc_assert (code.is_tree_code ()); > return vectorize_fold_left_reduction > (loop_vinfo, stmt_info, gsi, vec_stmt, slp_node, reduc_def_phi, > - tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks); > + tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks, > + lens); > } > > bool single_defuse_cycle = STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info); >
Thanks Richi, Address comment on V3: https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625167.html Bootstrap and regression is on the way. juzhe.zhong@rivai.ai From: Richard Biener Date: 2023-07-21 18:51 To: Ju-Zhe Zhong CC: gcc-patches; richard.sandiford Subject: Re: [PATCH V2] VECT: Support floating-point in-order reduction for length loop control On Fri, 21 Jul 2023, juzhe.zhong@rivai.ai wrote: > From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai> > > Hi, Richard and Richi. > > This patch support floating-point in-order reduction for loop length control. > > Consider this following case: > > float foo (float *__restrict a, int n) > { > float result = 1.0; > for (int i = 0; i < n; i++) > result += a[i]; > return result; > } > > When compile with **NO** -ffast-math on ARM SVE, we will end up with: > > loop_mask = WHILE_ULT > result = MASK_FOLD_LEFT_PLUS (...loop_mask...) > > For RVV, we don't use length loop control instead of mask: > > So, with this patch, we expect to see: > > loop_len = SELECT_VL > result = MASK_LEN_FOLD_LEFT_PLUS (...loop_len...) > > gcc/ChangeLog: > > * tree-vect-loop.cc (get_masked_reduction_fn): Add mask_len_fold_left. > (vectorize_fold_left_reduction): Ditto. > (vectorizable_reduction): Ditto. > (vect_transform_reduction): Ditto. > > --- > gcc/tree-vect-loop.cc | 41 ++++++++++++++++++++++++++++++++++++----- > 1 file changed, 36 insertions(+), 5 deletions(-) > > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc > index b44fb9c7712..59ab7879d55 100644 > --- a/gcc/tree-vect-loop.cc > +++ b/gcc/tree-vect-loop.cc > @@ -6800,11 +6800,13 @@ static internal_fn > get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) > { > internal_fn mask_reduc_fn; > + internal_fn mask_len_reduc_fn; > > switch (reduc_fn) > { > case IFN_FOLD_LEFT_PLUS: > mask_reduc_fn = IFN_MASK_FOLD_LEFT_PLUS; > + mask_len_reduc_fn = IFN_MASK_LEN_FOLD_LEFT_PLUS; > break; > > default: > @@ -6814,6 +6816,9 @@ get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) > if (direct_internal_fn_supported_p (mask_reduc_fn, vectype_in, > OPTIMIZE_FOR_SPEED)) > return mask_reduc_fn; > + if (direct_internal_fn_supported_p (mask_len_reduc_fn, vectype_in, > + OPTIMIZE_FOR_SPEED)) > + return mask_len_reduc_fn; > return IFN_LAST; > } > > @@ -6834,7 +6839,8 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, > gimple *reduc_def_stmt, > tree_code code, internal_fn reduc_fn, > tree ops[3], tree vectype_in, > - int reduc_index, vec_loop_masks *masks) > + int reduc_index, vec_loop_masks *masks, > + vec_loop_lens *lens) > { > class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); > tree vectype_out = STMT_VINFO_VECTYPE (stmt_info); > @@ -6896,8 +6902,18 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, > { > gimple *new_stmt; > tree mask = NULL_TREE; > + tree len = NULL_TREE; > + tree bias = NULL_TREE; > if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) > mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, i); > + if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) > + { > + len = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, vectype_in, > + i, 1); > + signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); > + bias = build_int_cst (intQI_type_node, biasval); > + mask = build_minus_one_cst (truth_type_for (vectype_in)); > + } > > /* Handle MINUS by adding the negative. */ > if (reduc_fn != IFN_LAST && code == MINUS_EXPR) > @@ -6917,7 +6933,10 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, > the preceding operation. */ > if (reduc_fn != IFN_LAST || (mask && mask_reduc_fn != IFN_LAST)) > { > - if (mask && mask_reduc_fn != IFN_LAST) > + if (len && mask && mask_reduc_fn != IFN_LAST) check mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS instead? > + new_stmt = gimple_build_call_internal (mask_reduc_fn, 5, reduc_var, > + def0, mask, len, bias); > + else if (mask && mask_reduc_fn != IFN_LAST) Likewise. Otherwise looks good to me. Richard. > new_stmt = gimple_build_call_internal (mask_reduc_fn, 3, reduc_var, > def0, mask); > else > @@ -7979,6 +7998,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo, > else if (loop_vinfo && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) > { > vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); > + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); > internal_fn cond_fn = get_conditional_internal_fn (op.code, op.type); > > if (reduction_type != FOLD_LEFT_REDUCTION > @@ -8006,8 +8026,17 @@ vectorizable_reduction (loop_vec_info loop_vinfo, > LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; > } > else > - vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, > - vectype_in, NULL); > + { > + internal_fn mask_reduc_fn > + = get_masked_reduction_fn (reduc_fn, vectype_in); > + > + if (mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS) > + vect_record_loop_len (loop_vinfo, lens, ncopies * vec_num, > + vectype_in, 1); > + else > + vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, > + vectype_in, NULL); > + } > } > return true; > } > @@ -8137,6 +8166,7 @@ vect_transform_reduction (loop_vec_info loop_vinfo, > code_helper code = canonicalize_code (op.code, op.type); > internal_fn cond_fn = get_conditional_internal_fn (code, op.type); > vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); > + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); > bool mask_by_cond_expr = use_mask_by_cond_expr_p (code, cond_fn, vectype_in); > > /* Transform. */ > @@ -8162,7 +8192,8 @@ vect_transform_reduction (loop_vec_info loop_vinfo, > gcc_assert (code.is_tree_code ()); > return vectorize_fold_left_reduction > (loop_vinfo, stmt_info, gsi, vec_stmt, slp_node, reduc_def_phi, > - tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks); > + tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks, > + lens); > } > > bool single_defuse_cycle = STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info); >
Oh. Sorry for missing a fix, Now I fix as you suggested on V4
https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625169.html
Change it as follows:
if (mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS)
new_stmt = gimple_build_call_internal (mask_reduc_fn, 5, reduc_var,
def0, mask, len, bias);
else if (mask_reduc_fn == IFN_MASK_FOLD_LEFT_PLUS)
new_stmt = gimple_build_call_internal (mask_reduc_fn, 3, reduc_var,
def0, mask);
else
new_stmt = gimple_build_call_internal (reduc_fn, 2, reduc_var,
def0);
Sorry for that.
Bootstrap && Regression on running.
juzhe.zhong@rivai.ai
From: Richard Biener
Date: 2023-07-21 18:51
To: Ju-Zhe Zhong
CC: gcc-patches; richard.sandiford
Subject: Re: [PATCH V2] VECT: Support floating-point in-order reduction for length loop control
On Fri, 21 Jul 2023, juzhe.zhong@rivai.ai wrote:
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> Hi, Richard and Richi.
>
> This patch support floating-point in-order reduction for loop length control.
>
> Consider this following case:
>
> float foo (float *__restrict a, int n)
> {
> float result = 1.0;
> for (int i = 0; i < n; i++)
> result += a[i];
> return result;
> }
>
> When compile with **NO** -ffast-math on ARM SVE, we will end up with:
>
> loop_mask = WHILE_ULT
> result = MASK_FOLD_LEFT_PLUS (...loop_mask...)
>
> For RVV, we don't use length loop control instead of mask:
>
> So, with this patch, we expect to see:
>
> loop_len = SELECT_VL
> result = MASK_LEN_FOLD_LEFT_PLUS (...loop_len...)
>
> gcc/ChangeLog:
>
> * tree-vect-loop.cc (get_masked_reduction_fn): Add mask_len_fold_left.
> (vectorize_fold_left_reduction): Ditto.
> (vectorizable_reduction): Ditto.
> (vect_transform_reduction): Ditto.
>
> ---
> gcc/tree-vect-loop.cc | 41 ++++++++++++++++++++++++++++++++++++-----
> 1 file changed, 36 insertions(+), 5 deletions(-)
>
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index b44fb9c7712..59ab7879d55 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -6800,11 +6800,13 @@ static internal_fn
> get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in)
> {
> internal_fn mask_reduc_fn;
> + internal_fn mask_len_reduc_fn;
>
> switch (reduc_fn)
> {
> case IFN_FOLD_LEFT_PLUS:
> mask_reduc_fn = IFN_MASK_FOLD_LEFT_PLUS;
> + mask_len_reduc_fn = IFN_MASK_LEN_FOLD_LEFT_PLUS;
> break;
>
> default:
> @@ -6814,6 +6816,9 @@ get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in)
> if (direct_internal_fn_supported_p (mask_reduc_fn, vectype_in,
> OPTIMIZE_FOR_SPEED))
> return mask_reduc_fn;
> + if (direct_internal_fn_supported_p (mask_len_reduc_fn, vectype_in,
> + OPTIMIZE_FOR_SPEED))
> + return mask_len_reduc_fn;
> return IFN_LAST;
> }
>
> @@ -6834,7 +6839,8 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo,
> gimple *reduc_def_stmt,
> tree_code code, internal_fn reduc_fn,
> tree ops[3], tree vectype_in,
> - int reduc_index, vec_loop_masks *masks)
> + int reduc_index, vec_loop_masks *masks,
> + vec_loop_lens *lens)
> {
> class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
> tree vectype_out = STMT_VINFO_VECTYPE (stmt_info);
> @@ -6896,8 +6902,18 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo,
> {
> gimple *new_stmt;
> tree mask = NULL_TREE;
> + tree len = NULL_TREE;
> + tree bias = NULL_TREE;
> if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
> mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, i);
> + if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
> + {
> + len = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, vectype_in,
> + i, 1);
> + signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + mask = build_minus_one_cst (truth_type_for (vectype_in));
> + }
>
> /* Handle MINUS by adding the negative. */
> if (reduc_fn != IFN_LAST && code == MINUS_EXPR)
> @@ -6917,7 +6933,10 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo,
> the preceding operation. */
> if (reduc_fn != IFN_LAST || (mask && mask_reduc_fn != IFN_LAST))
> {
> - if (mask && mask_reduc_fn != IFN_LAST)
> + if (len && mask && mask_reduc_fn != IFN_LAST)
check mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS instead?
> + new_stmt = gimple_build_call_internal (mask_reduc_fn, 5, reduc_var,
> + def0, mask, len, bias);
> + else if (mask && mask_reduc_fn != IFN_LAST)
Likewise.
Otherwise looks good to me.
Richard.
> new_stmt = gimple_build_call_internal (mask_reduc_fn, 3, reduc_var,
> def0, mask);
> else
> @@ -7979,6 +7998,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
> else if (loop_vinfo && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
> {
> vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
> + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
> internal_fn cond_fn = get_conditional_internal_fn (op.code, op.type);
>
> if (reduction_type != FOLD_LEFT_REDUCTION
> @@ -8006,8 +8026,17 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
> LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
> }
> else
> - vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num,
> - vectype_in, NULL);
> + {
> + internal_fn mask_reduc_fn
> + = get_masked_reduction_fn (reduc_fn, vectype_in);
> +
> + if (mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS)
> + vect_record_loop_len (loop_vinfo, lens, ncopies * vec_num,
> + vectype_in, 1);
> + else
> + vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num,
> + vectype_in, NULL);
> + }
> }
> return true;
> }
> @@ -8137,6 +8166,7 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
> code_helper code = canonicalize_code (op.code, op.type);
> internal_fn cond_fn = get_conditional_internal_fn (code, op.type);
> vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
> + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
> bool mask_by_cond_expr = use_mask_by_cond_expr_p (code, cond_fn, vectype_in);
>
> /* Transform. */
> @@ -8162,7 +8192,8 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
> gcc_assert (code.is_tree_code ());
> return vectorize_fold_left_reduction
> (loop_vinfo, stmt_info, gsi, vec_stmt, slp_node, reduc_def_phi,
> - tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks);
> + tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks,
> + lens);
> }
>
> bool single_defuse_cycle = STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info);
>
Hi Richard, Bootstrap and regression are passed on X86 and no new testcases fail on AArch64 with V5 patch: https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625293.html V5 patch is ok for trunk? Best, Lehua
On Sun, 23 Jul 2023, Lehua Ding wrote: > Hi Richard, > > > Bootstrap and regression are passed on X86 and > no new testcases fail on AArch64 with V5 patch: > > > https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625293.html > > > V5 patch is ok for trunk? Yes.
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc index b44fb9c7712..59ab7879d55 100644 --- a/gcc/tree-vect-loop.cc +++ b/gcc/tree-vect-loop.cc @@ -6800,11 +6800,13 @@ static internal_fn get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) { internal_fn mask_reduc_fn; + internal_fn mask_len_reduc_fn; switch (reduc_fn) { case IFN_FOLD_LEFT_PLUS: mask_reduc_fn = IFN_MASK_FOLD_LEFT_PLUS; + mask_len_reduc_fn = IFN_MASK_LEN_FOLD_LEFT_PLUS; break; default: @@ -6814,6 +6816,9 @@ get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) if (direct_internal_fn_supported_p (mask_reduc_fn, vectype_in, OPTIMIZE_FOR_SPEED)) return mask_reduc_fn; + if (direct_internal_fn_supported_p (mask_len_reduc_fn, vectype_in, + OPTIMIZE_FOR_SPEED)) + return mask_len_reduc_fn; return IFN_LAST; } @@ -6834,7 +6839,8 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, gimple *reduc_def_stmt, tree_code code, internal_fn reduc_fn, tree ops[3], tree vectype_in, - int reduc_index, vec_loop_masks *masks) + int reduc_index, vec_loop_masks *masks, + vec_loop_lens *lens) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree vectype_out = STMT_VINFO_VECTYPE (stmt_info); @@ -6896,8 +6902,18 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, { gimple *new_stmt; tree mask = NULL_TREE; + tree len = NULL_TREE; + tree bias = NULL_TREE; if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, i); + if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) + { + len = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, vectype_in, + i, 1); + signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); + bias = build_int_cst (intQI_type_node, biasval); + mask = build_minus_one_cst (truth_type_for (vectype_in)); + } /* Handle MINUS by adding the negative. */ if (reduc_fn != IFN_LAST && code == MINUS_EXPR) @@ -6917,7 +6933,10 @@ vectorize_fold_left_reduction (loop_vec_info loop_vinfo, the preceding operation. */ if (reduc_fn != IFN_LAST || (mask && mask_reduc_fn != IFN_LAST)) { - if (mask && mask_reduc_fn != IFN_LAST) + if (len && mask && mask_reduc_fn != IFN_LAST) + new_stmt = gimple_build_call_internal (mask_reduc_fn, 5, reduc_var, + def0, mask, len, bias); + else if (mask && mask_reduc_fn != IFN_LAST) new_stmt = gimple_build_call_internal (mask_reduc_fn, 3, reduc_var, def0, mask); else @@ -7979,6 +7998,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo, else if (loop_vinfo && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) { vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); internal_fn cond_fn = get_conditional_internal_fn (op.code, op.type); if (reduction_type != FOLD_LEFT_REDUCTION @@ -8006,8 +8026,17 @@ vectorizable_reduction (loop_vec_info loop_vinfo, LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else - vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, - vectype_in, NULL); + { + internal_fn mask_reduc_fn + = get_masked_reduction_fn (reduc_fn, vectype_in); + + if (mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS) + vect_record_loop_len (loop_vinfo, lens, ncopies * vec_num, + vectype_in, 1); + else + vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, + vectype_in, NULL); + } } return true; } @@ -8137,6 +8166,7 @@ vect_transform_reduction (loop_vec_info loop_vinfo, code_helper code = canonicalize_code (op.code, op.type); internal_fn cond_fn = get_conditional_internal_fn (code, op.type); vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); bool mask_by_cond_expr = use_mask_by_cond_expr_p (code, cond_fn, vectype_in); /* Transform. */ @@ -8162,7 +8192,8 @@ vect_transform_reduction (loop_vec_info loop_vinfo, gcc_assert (code.is_tree_code ()); return vectorize_fold_left_reduction (loop_vinfo, stmt_info, gsi, vec_stmt, slp_node, reduc_def_phi, - tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks); + tree_code (code), reduc_fn, op.ops, vectype_in, reduc_index, masks, + lens); } bool single_defuse_cycle = STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info);
From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai> Hi, Richard and Richi. This patch support floating-point in-order reduction for loop length control. Consider this following case: float foo (float *__restrict a, int n) { float result = 1.0; for (int i = 0; i < n; i++) result += a[i]; return result; } When compile with **NO** -ffast-math on ARM SVE, we will end up with: loop_mask = WHILE_ULT result = MASK_FOLD_LEFT_PLUS (...loop_mask...) For RVV, we don't use length loop control instead of mask: So, with this patch, we expect to see: loop_len = SELECT_VL result = MASK_LEN_FOLD_LEFT_PLUS (...loop_len...) gcc/ChangeLog: * tree-vect-loop.cc (get_masked_reduction_fn): Add mask_len_fold_left. (vectorize_fold_left_reduction): Ditto. (vectorizable_reduction): Ditto. (vect_transform_reduction): Ditto. --- gcc/tree-vect-loop.cc | 41 ++++++++++++++++++++++++++++++++++++----- 1 file changed, 36 insertions(+), 5 deletions(-)