> -----Original Message-----
> From: pan2...@intel.com <pan2...@intel.com>
> Sent: Monday, May 13, 2024 3:54 PM
> To: gcc-patches@gcc.gnu.org
> Cc: juzhe.zh...@rivai.ai; kito.ch...@gmail.com; richard.guent...@gmail.com;
> Tamar Christina <tamar.christ...@arm.com>; Richard Sandiford
> <richard.sandif...@arm.com>; Pan Li <pan2...@intel.com>
> Subject: [PATCH v1 1/3] Vect: Support loop len in vectorizable early exit
>
> From: Pan Li <pan2...@intel.com>
>
> This patch adds early break auto-vectorization support for target which
> use length on partial vectorization. Consider this following example:
>
> unsigned vect_a[802];
> unsigned vect_b[802];
>
> void test (unsigned x, int n)
> {
> for (int i = 0; i < n; i++)
> {
> vect_b[i] = x + i;
>
> if (vect_a[i] > x)
> break;
>
> vect_a[i] = x;
> }
> }
>
> We use VCOND_MASK_LEN to simulate the generate (mask && i < len + bias).
> And then the IR of RVV looks like below:
>
> ...
> _87 = .SELECT_VL (ivtmp_85, POLY_INT_CST [32, 32]);
> _55 = (int) _87;
> ...
> mask_patt_6.13_69 = vect_cst__62 < vect__3.12_67;
> vec_len_mask_72 = .VCOND_MASK_LEN (mask_patt_6.13_69, { -1, ... }, \
> {0, ... }, _87, 0);
> if (vec_len_mask_72 != { 0, ... })
> goto <bb 6>; [5.50%]
> else
> goto <bb 7>; [94.50%]
>
> The below tests are passed for this patch:
> 1. The riscv fully regression tests.
> 2. The aarch64 fully regression tests.
> 3. The x86 bootstrap tests.
> 4. The x86 fully regression tests.
>
> gcc/ChangeLog:
>
> * tree-vect-stmts.cc (vectorizable_early_exit): Add loop len
> handling for one or multiple stmt.
>
> Signed-off-by: Pan Li <pan2...@intel.com>
> ---
> gcc/tree-vect-stmts.cc | 47 ++++++++++++++++++++++++++++++++++++++++-
> -
> 1 file changed, 45 insertions(+), 2 deletions(-)
>
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 21e8fe98e44..bfd9d66568f 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -12896,7 +12896,9 @@ vectorizable_early_exit (vec_info *vinfo,
> stmt_vec_info stmt_info,
> ncopies = vect_get_num_copies (loop_vinfo, vectype);
>
> vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
> + vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
> bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> + bool len_loop_p = LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo);
>
> /* Now build the new conditional. Pattern gimple_conds get dropped during
> codegen so we must replace the original insn. */
> @@ -12960,12 +12962,11 @@ vectorizable_early_exit (vec_info *vinfo,
> stmt_vec_info stmt_info,
> {
> if (direct_internal_fn_supported_p (IFN_VCOND_MASK_LEN, vectype,
> OPTIMIZE_FOR_SPEED))
> - return false;
> + vect_record_loop_len (loop_vinfo, lens, ncopies, vectype, 1);
> else
> vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype, NULL);
> }
>
> -
> return true;
> }
>
> @@ -13018,6 +13019,25 @@ vectorizable_early_exit (vec_info *vinfo,
> stmt_vec_info stmt_info,
> stmts[i], &cond_gsi);
> workset.quick_push (stmt_mask);
> }
> + else if (len_loop_p)
> + for (unsigned i = 0; i < stmts.length (); i++)
> + {
> + tree all_ones_mask = build_all_ones_cst (vectype);
> + tree all_zero_mask = build_zero_cst (vectype);
> + tree len = vect_get_loop_len (loop_vinfo, gsi, lens, ncopies,
> + vectype, i, 1);
> + signed char cst = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS
> (loop_vinfo);
> + tree bias = build_int_cst (intQI_type_node, cst);
> + tree len_mask
> + = make_temp_ssa_name (TREE_TYPE (stmts[i]), NULL,
> "vec_len_mask");
> + gcall *call = gimple_build_call_internal (IFN_VCOND_MASK_LEN, 5,
> + stmts[i], all_ones_mask,
> + all_zero_mask, len, bias);
> + gimple_call_set_lhs (call, len_mask);
> + gsi_insert_before (&cond_gsi, call, GSI_SAME_STMT);
> +
> + workset.quick_push (len_mask);
> + }
> else
> workset.splice (stmts);
>
> @@ -13042,6 +13062,29 @@ vectorizable_early_exit (vec_info *vinfo,
> stmt_vec_info stmt_info,
> new_temp = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), mask,
> new_temp, &cond_gsi);
> }
> + else if (len_loop_p)
> + {
> + /* len_mask = VCOND_MASK_LEN (compare_mask, ones, zero, len, bias)
> +
> + which is equivalent to:
> +
> + len_mask = compare_mask mask && i < len ? 1 : 0
> + */
> + tree all_ones_mask = build_all_ones_cst (vectype);
> + tree all_zero_mask = build_zero_cst (vectype);
> + tree len
> + = vect_get_loop_len (loop_vinfo, gsi, lens, ncopies, vectype, 0, 1);
> + signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS
> (loop_vinfo);
> + tree bias = build_int_cst (intQI_type_node, biasval);
> + tree len_mask
> + = make_temp_ssa_name (TREE_TYPE (new_temp), NULL,
> "vec_len_mask");
> + gcall *call = gimple_build_call_internal (IFN_VCOND_MASK_LEN, 5,
> + new_temp, all_ones_mask,
> + all_zero_mask, len, bias);
> + gimple_call_set_lhs (call, len_mask);
> + gsi_insert_before (&cond_gsi, call, GSI_SAME_STMT);
> + new_temp = len_mask;
> + }
Could we perhaps factor these out to a helper? Something similar to what we
have with
prepare_vec_mask. It looks like the only difference between these two block of
code is the index being passed to
vect_get_loop_len.
It would be nice to have the shape the same way as the masked_loop_p case, i.e.
keep 1 call to get the loop_len and one to
build the mask using a helper.
Thanks,
Tamar
> }
>
> gcc_assert (new_temp);
> --
> 2.34.1
