* [PATCH V2] VECT: Apply MASK_LEN_{LOAD_LANES, STORE_LANES} into vectorizer
@ 2023-08-15 12:29 Juzhe-Zhong
2023-08-15 12:35 ` Richard Biener
0 siblings, 1 reply; 3+ messages in thread
From: Juzhe-Zhong @ 2023-08-15 12:29 UTC (permalink / raw)
To: gcc-patches; +Cc: richard.sandiford, rguenther, Juzhe-Zhong
Hi, Richard and Richi.
This patch is adding MASK_LEN_{LOAD_LANES,STORE_LANES} support into vectorizer.
Consider this simple case:
void __attribute__ ((noinline, noclone))
foo (int *__restrict a, int *__restrict b, int *__restrict c,
int *__restrict d, int *__restrict e, int *__restrict f,
int *__restrict g, int *__restrict h, int *__restrict j, int n)
{
for (int i = 0; i < n; ++i)
{
a[i] = j[i * 8];
b[i] = j[i * 8 + 1];
c[i] = j[i * 8 + 2];
d[i] = j[i * 8 + 3];
e[i] = j[i * 8 + 4];
f[i] = j[i * 8 + 5];
g[i] = j[i * 8 + 6];
h[i] = j[i * 8 + 7];
}
}
RVV Gimple IR:
_79 = .SELECT_VL (ivtmp_81, POLY_INT_CST [4, 4]);
ivtmp_125 = _79 * 32;
vect_array.8 = .MASK_LEN_LOAD_LANES (vectp_j.6_124, 32B, { -1, ... }, _79, 0);
vect__8.9_122 = vect_array.8[0];
vect__8.10_121 = vect_array.8[1];
vect__8.11_120 = vect_array.8[2];
vect__8.12_119 = vect_array.8[3];
vect__8.13_118 = vect_array.8[4];
vect__8.14_117 = vect_array.8[5];
vect__8.15_116 = vect_array.8[6];
vect__8.16_115 = vect_array.8[7];
vect_array.8 ={v} {CLOBBER};
ivtmp_114 = _79 * 4;
.MASK_LEN_STORE (vectp_a.17_113, 32B, { -1, ... }, _79, 0, vect__8.9_122);
.MASK_LEN_STORE (vectp_b.19_109, 32B, { -1, ... }, _79, 0, vect__8.10_121);
.MASK_LEN_STORE (vectp_c.21_105, 32B, { -1, ... }, _79, 0, vect__8.11_120);
.MASK_LEN_STORE (vectp_d.23_101, 32B, { -1, ... }, _79, 0, vect__8.12_119);
.MASK_LEN_STORE (vectp_e.25_97, 32B, { -1, ... }, _79, 0, vect__8.13_118);
.MASK_LEN_STORE (vectp_f.27_93, 32B, { -1, ... }, _79, 0, vect__8.14_117);
.MASK_LEN_STORE (vectp_g.29_89, 32B, { -1, ... }, _79, 0, vect__8.15_116);
.MASK_LEN_STORE (vectp_h.31_85, 32B, { -1, ... }, _79, 0, vect__8.16_115);
ASM:
foo:
lw t4,8(sp)
ld t5,0(sp)
ble t4,zero,.L5
.L3:
vsetvli t1,t4,e8,mf4,ta,ma
vlseg8e32.v v8,(t5)
slli t3,t1,2
slli t6,t1,5
vse32.v v8,0(a0)
vse32.v v9,0(a1)
vse32.v v10,0(a2)
vse32.v v11,0(a3)
vse32.v v12,0(a4)
vse32.v v13,0(a5)
vse32.v v14,0(a6)
vse32.v v15,0(a7)
sub t4,t4,t1
add t5,t5,t6
add a0,a0,t3
add a1,a1,t3
add a2,a2,t3
add a3,a3,t3
add a4,a4,t3
add a5,a5,t3
add a6,a6,t3
add a7,a7,t3
bne t4,zero,.L3
.L5:
ret
The details of the approach:
Step 1 - Modifiy the LANES LOAD/STORE support function (vect_load_lanes_supported/vect_store_lanes_supported):
+/* Return FN if vec_{masked_,mask_len,}load_lanes is available for COUNT
+ vectors of type VECTYPE. MASKED_P says whether the masked form is needed. */
-bool
+internal_fn
vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
bool masked_p)
{
- if (masked_p)
- return vect_lanes_optab_supported_p ("vec_mask_load_lanes",
- vec_mask_load_lanes_optab,
- vectype, count);
+ if (vect_lanes_optab_supported_p ("vec_mask_len_load_lanes",
+ vec_mask_len_load_lanes_optab,
+ vectype, count))
+ return IFN_MASK_LEN_LOAD_LANES;
+ else if (masked_p)
+ {
+ if (vect_lanes_optab_supported_p ("vec_mask_load_lanes",
+ vec_mask_load_lanes_optab,
+ vectype, count))
+ return IFN_MASK_LOAD_LANES;
+ }
else
- return vect_lanes_optab_supported_p ("vec_load_lanes",
- vec_load_lanes_optab,
- vectype, count);
+ {
+ if (vect_lanes_optab_supported_p ("vec_load_lanes",
+ vec_load_lanes_optab,
+ vectype, count))
+ return IFN_LOAD_LANES;
+ }
+ return IFN_LAST;
}
Instead of returning TRUE or FALSE whether target support the LANES LOAD/STORE.
I change it into return internal_fn of the LANES LOAD/STORE that target support,
If target didn't support any LANE LOAD/STORE optabs, return IFN_LAST.
Step 2 - Compute IFN for LANES LOAD/STORE (Only compute once).
if (!STMT_VINFO_STRIDED_P (first_stmt_info)
&& (can_overrun_p || !would_overrun_p)
&& compare_step_with_zero (vinfo, stmt_info) > 0)
{
/* First cope with the degenerate case of a single-element
vector. */
if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
;
else
{
/* Otherwise try using LOAD/STORE_LANES. */
*lanes_ifn
= vls_type == VLS_LOAD
? vect_load_lanes_supported (vectype, group_size, masked_p)
: vect_store_lanes_supported (vectype, group_size,
masked_p);
if (*lanes_ifn != IFN_LAST)
{
*memory_access_type = VMAT_LOAD_STORE_LANES;
overrun_p = would_overrun_p;
}
/* If that fails, try using permuting loads. */
else if (vls_type == VLS_LOAD
? vect_grouped_load_supported (vectype,
single_element_p,
group_size)
: vect_grouped_store_supported (vectype, group_size))
{
*memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
overrun_p = would_overrun_p;
}
}
}
Step 3 - Build MASK_LEN_{LANES_LOAD,LANES_STORE} Gimple IR:
+ if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
+ {
+ if (loop_lens)
+ final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ ncopies, vectype, j, 1);
+ else
+ final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
+ signed char biasval
+ = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
+ bias = build_int_cst (intQI_type_node, biasval);
+ if (!final_mask)
+ {
+ mask_vectype = truth_type_for (vectype);
+ final_mask = build_minus_one_cst (mask_vectype);
+ }
+ }
+
gcall *call;
- if (final_mask)
+ if (final_len && final_mask)
+ {
+ /* Emit:
+ MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
+ LEN, BIAS, VEC_ARRAY). */
+ unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
+ tree alias_ptr = build_int_cst (ref_type, align);
+ call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
+ dataref_ptr, alias_ptr,
+ final_mask, final_len, bias,
+ vec_array);
+ }
+ else if (final_mask)
The LEN and MASK flow is totally the same as other MASK_LEN_* load/store.
gcc/ChangeLog:
* internal-fn.cc (internal_load_fn_p): Apply MASK_LEN_{LOAD_LANES,STORE_LANES} into vectorizer.
(internal_store_fn_p): Ditto.
(internal_fn_len_index): Ditto.
(internal_fn_mask_index): Ditto.
(internal_fn_stored_value_index): Ditto.
* tree-vect-data-refs.cc (vect_store_lanes_supported): Ditto.
(vect_load_lanes_supported): Ditto.
* tree-vect-loop.cc: Ditto.
* tree-vect-slp.cc (vect_slp_prefer_store_lanes_p): Ditto.
* tree-vect-stmts.cc (check_load_store_for_partial_vectors): Ditto.
(get_group_load_store_type): Ditto.
(vectorizable_store): Ditto.
(vectorizable_load): Ditto.
* tree-vectorizer.h (vect_store_lanes_supported): Ditto.
(vect_load_lanes_supported): Ditto.
---
gcc/internal-fn.cc | 7 ++
gcc/tree-vect-data-refs.cc | 61 ++++++++++------
gcc/tree-vect-loop.cc | 11 +--
gcc/tree-vect-slp.cc | 2 +-
gcc/tree-vect-stmts.cc | 141 ++++++++++++++++++++++++++++---------
gcc/tree-vectorizer.h | 4 +-
6 files changed, 163 insertions(+), 63 deletions(-)
diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
index 4f2b20a79e5..cc1ede58799 100644
--- a/gcc/internal-fn.cc
+++ b/gcc/internal-fn.cc
@@ -4578,6 +4578,7 @@ internal_load_fn_p (internal_fn fn)
case IFN_MASK_LOAD:
case IFN_LOAD_LANES:
case IFN_MASK_LOAD_LANES:
+ case IFN_MASK_LEN_LOAD_LANES:
case IFN_GATHER_LOAD:
case IFN_MASK_GATHER_LOAD:
case IFN_MASK_LEN_GATHER_LOAD:
@@ -4600,6 +4601,7 @@ internal_store_fn_p (internal_fn fn)
case IFN_MASK_STORE:
case IFN_STORE_LANES:
case IFN_MASK_STORE_LANES:
+ case IFN_MASK_LEN_STORE_LANES:
case IFN_SCATTER_STORE:
case IFN_MASK_SCATTER_STORE:
case IFN_MASK_LEN_SCATTER_STORE:
@@ -4672,6 +4674,8 @@ internal_fn_len_index (internal_fn fn)
case IFN_COND_LEN_NEG:
case IFN_MASK_LEN_LOAD:
case IFN_MASK_LEN_STORE:
+ case IFN_MASK_LEN_LOAD_LANES:
+ case IFN_MASK_LEN_STORE_LANES:
return 3;
default:
@@ -4689,8 +4693,10 @@ internal_fn_mask_index (internal_fn fn)
{
case IFN_MASK_LOAD:
case IFN_MASK_LOAD_LANES:
+ case IFN_MASK_LEN_LOAD_LANES:
case IFN_MASK_STORE:
case IFN_MASK_STORE_LANES:
+ case IFN_MASK_LEN_STORE_LANES:
case IFN_MASK_LEN_LOAD:
case IFN_MASK_LEN_STORE:
return 2;
@@ -4726,6 +4732,7 @@ internal_fn_stored_value_index (internal_fn fn)
return 4;
case IFN_MASK_LEN_STORE:
+ case IFN_MASK_LEN_STORE_LANES:
return 5;
default:
diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index a3570c45b52..3e9a284666c 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -5438,22 +5438,31 @@ vect_grouped_store_supported (tree vectype, unsigned HOST_WIDE_INT count)
return false;
}
+/* Return FN if vec_{mask_,mask_len_}store_lanes is available for COUNT vectors
+ of type VECTYPE. MASKED_P says whether the masked form is needed. */
-/* Return TRUE if vec_{mask_}store_lanes is available for COUNT vectors of
- type VECTYPE. MASKED_P says whether the masked form is needed. */
-
-bool
+internal_fn
vect_store_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
bool masked_p)
{
- if (masked_p)
- return vect_lanes_optab_supported_p ("vec_mask_store_lanes",
- vec_mask_store_lanes_optab,
- vectype, count);
+ if (vect_lanes_optab_supported_p ("vec_mask_len_store_lanes",
+ vec_mask_len_store_lanes_optab, vectype,
+ count))
+ return IFN_MASK_LEN_STORE_LANES;
+ else if (masked_p)
+ {
+ if (vect_lanes_optab_supported_p ("vec_mask_store_lanes",
+ vec_mask_store_lanes_optab, vectype,
+ count))
+ return IFN_MASK_STORE_LANES;
+ }
else
- return vect_lanes_optab_supported_p ("vec_store_lanes",
- vec_store_lanes_optab,
- vectype, count);
+ {
+ if (vect_lanes_optab_supported_p ("vec_store_lanes",
+ vec_store_lanes_optab, vectype, count))
+ return IFN_STORE_LANES;
+ }
+ return IFN_LAST;
}
@@ -6056,21 +6065,31 @@ vect_grouped_load_supported (tree vectype, bool single_element_p,
return false;
}
-/* Return TRUE if vec_{masked_}load_lanes is available for COUNT vectors of
- type VECTYPE. MASKED_P says whether the masked form is needed. */
+/* Return FN if vec_{masked_,mask_len_}load_lanes is available for COUNT vectors
+ of type VECTYPE. MASKED_P says whether the masked form is needed. */
-bool
+internal_fn
vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
bool masked_p)
{
- if (masked_p)
- return vect_lanes_optab_supported_p ("vec_mask_load_lanes",
- vec_mask_load_lanes_optab,
- vectype, count);
+ if (vect_lanes_optab_supported_p ("vec_mask_len_load_lanes",
+ vec_mask_len_load_lanes_optab, vectype,
+ count))
+ return IFN_MASK_LEN_LOAD_LANES;
+ else if (masked_p)
+ {
+ if (vect_lanes_optab_supported_p ("vec_mask_load_lanes",
+ vec_mask_load_lanes_optab, vectype,
+ count))
+ return IFN_MASK_LOAD_LANES;
+ }
else
- return vect_lanes_optab_supported_p ("vec_load_lanes",
- vec_load_lanes_optab,
- vectype, count);
+ {
+ if (vect_lanes_optab_supported_p ("vec_load_lanes", vec_load_lanes_optab,
+ vectype, count))
+ return IFN_LOAD_LANES;
+ }
+ return IFN_LAST;
}
/* Function vect_permute_load_chain.
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index bc3063c3615..1fcd8d07ea1 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -2839,7 +2839,8 @@ start_over:
instructions record it and move on to the next instance. */
if (loads_permuted
&& SLP_INSTANCE_KIND (instance) == slp_inst_kind_store
- && vect_store_lanes_supported (vectype, group_size, false))
+ && vect_store_lanes_supported (vectype, group_size, false)
+ != IFN_LAST)
{
FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (instance), i, load_node)
{
@@ -2848,9 +2849,9 @@ start_over:
/* Use SLP for strided accesses (or if we can't
load-lanes). */
if (STMT_VINFO_STRIDED_P (stmt_vinfo)
- || ! vect_load_lanes_supported
+ || vect_load_lanes_supported
(STMT_VINFO_VECTYPE (stmt_vinfo),
- DR_GROUP_SIZE (stmt_vinfo), false))
+ DR_GROUP_SIZE (stmt_vinfo), false) == IFN_LAST)
break;
}
@@ -3153,7 +3154,7 @@ again:
vinfo = DR_GROUP_FIRST_ELEMENT (vinfo);
unsigned int size = DR_GROUP_SIZE (vinfo);
tree vectype = STMT_VINFO_VECTYPE (vinfo);
- if (! vect_store_lanes_supported (vectype, size, false)
+ if (vect_store_lanes_supported (vectype, size, false) == IFN_LAST
&& ! known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)
&& ! vect_grouped_store_supported (vectype, size))
return opt_result::failure_at (vinfo->stmt,
@@ -3165,7 +3166,7 @@ again:
bool single_element_p = !DR_GROUP_NEXT_ELEMENT (vinfo);
size = DR_GROUP_SIZE (vinfo);
vectype = STMT_VINFO_VECTYPE (vinfo);
- if (! vect_load_lanes_supported (vectype, size, false)
+ if (vect_load_lanes_supported (vectype, size, false) == IFN_LAST
&& ! vect_grouped_load_supported (vectype, single_element_p,
size))
return opt_result::failure_at (vinfo->stmt,
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index cf91b21cf7d..9ad2634762e 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -3094,7 +3094,7 @@ vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
|| multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
return false;
- return vect_store_lanes_supported (vectype, group_size, false);
+ return vect_store_lanes_supported (vectype, group_size, false) != IFN_LAST;
}
/* Analyze an SLP instance starting from a group of grouped stores. Call
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 86d033aa60c..cd8e0a76374 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -1610,9 +1610,15 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
bool is_load = (vls_type == VLS_LOAD);
if (memory_access_type == VMAT_LOAD_STORE_LANES)
{
- if (is_load
- ? !vect_load_lanes_supported (vectype, group_size, true)
- : !vect_store_lanes_supported (vectype, group_size, true))
+ internal_fn ifn
+ = (is_load ? vect_load_lanes_supported (vectype, group_size, true)
+ : vect_store_lanes_supported (vectype, group_size, true));
+ if (ifn == IFN_MASK_LEN_LOAD_LANES || ifn == IFN_MASK_LEN_STORE_LANES)
+ vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1);
+ else if (ifn == IFN_MASK_LOAD_LANES || ifn == IFN_MASK_STORE_LANES)
+ vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
+ scalar_mask);
+ else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
@@ -1620,10 +1626,7 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
" the target doesn't have an appropriate"
" load/store-lanes instruction.\n");
LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
- return;
}
- vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
- scalar_mask);
return;
}
@@ -2074,7 +2077,8 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
poly_int64 *poffset,
dr_alignment_support *alignment_support_scheme,
int *misalignment,
- gather_scatter_info *gs_info)
+ gather_scatter_info *gs_info,
+ internal_fn *lanes_ifn)
{
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
@@ -2272,24 +2276,30 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
;
- /* Otherwise try using LOAD/STORE_LANES. */
- else if (vls_type == VLS_LOAD
- ? vect_load_lanes_supported (vectype, group_size, masked_p)
- : vect_store_lanes_supported (vectype, group_size,
- masked_p))
+ else
{
- *memory_access_type = VMAT_LOAD_STORE_LANES;
- overrun_p = would_overrun_p;
- }
+ /* Otherwise try using LOAD/STORE_LANES. */
+ *lanes_ifn
+ = vls_type == VLS_LOAD
+ ? vect_load_lanes_supported (vectype, group_size, masked_p)
+ : vect_store_lanes_supported (vectype, group_size,
+ masked_p);
+ if (*lanes_ifn != IFN_LAST)
+ {
+ *memory_access_type = VMAT_LOAD_STORE_LANES;
+ overrun_p = would_overrun_p;
+ }
- /* If that fails, try using permuting loads. */
- else if (vls_type == VLS_LOAD
- ? vect_grouped_load_supported (vectype, single_element_p,
- group_size)
- : vect_grouped_store_supported (vectype, group_size))
- {
- *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
- overrun_p = would_overrun_p;
+ /* If that fails, try using permuting loads. */
+ else if (vls_type == VLS_LOAD
+ ? vect_grouped_load_supported (vectype,
+ single_element_p,
+ group_size)
+ : vect_grouped_store_supported (vectype, group_size))
+ {
+ *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
+ overrun_p = would_overrun_p;
+ }
}
}
@@ -2378,7 +2388,8 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
poly_int64 *poffset,
dr_alignment_support *alignment_support_scheme,
int *misalignment,
- gather_scatter_info *gs_info)
+ gather_scatter_info *gs_info,
+ internal_fn *lanes_ifn)
{
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
@@ -2441,7 +2452,7 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
masked_p,
vls_type, memory_access_type, poffset,
alignment_support_scheme,
- misalignment, gs_info))
+ misalignment, gs_info, lanes_ifn))
return false;
}
else if (STMT_VINFO_STRIDED_P (stmt_info))
@@ -3087,11 +3098,8 @@ vect_get_loop_variant_data_ptr_increment (
loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
tree step = vect_dr_behavior (vinfo, dr_info)->step;
- /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
- IVs are updated by variable amount but we will support them in the future.
- */
- gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
- && memory_access_type != VMAT_LOAD_STORE_LANES);
+ /* gather/scatter never reach here. */
+ gcc_assert (memory_access_type != VMAT_GATHER_SCATTER);
/* When we support SELECT_VL pattern, we dynamic adjust
the memory address by .SELECT_VL result.
@@ -8094,9 +8102,11 @@ vectorizable_store (vec_info *vinfo,
enum dr_alignment_support alignment_support_scheme;
int misalignment;
poly_int64 poffset;
+ internal_fn lanes_ifn;
if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask, vls_type,
ncopies, &memory_access_type, &poffset,
- &alignment_support_scheme, &misalignment, &gs_info))
+ &alignment_support_scheme, &misalignment, &gs_info,
+ &lanes_ifn))
return false;
if (mask)
@@ -8885,6 +8895,8 @@ vectorizable_store (vec_info *vinfo,
}
tree final_mask = NULL;
+ tree final_len = NULL;
+ tree bias = NULL;
if (loop_masks)
final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
ncopies, vectype, j);
@@ -8892,8 +8904,37 @@ vectorizable_store (vec_info *vinfo,
final_mask = prepare_vec_mask (loop_vinfo, mask_vectype,
final_mask, vec_mask, gsi);
+ if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
+ {
+ if (loop_lens)
+ final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ ncopies, vectype, j, 1);
+ else
+ final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
+ signed char biasval
+ = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
+ bias = build_int_cst (intQI_type_node, biasval);
+ if (!final_mask)
+ {
+ mask_vectype = truth_type_for (vectype);
+ final_mask = build_minus_one_cst (mask_vectype);
+ }
+ }
+
gcall *call;
- if (final_mask)
+ if (final_len && final_mask)
+ {
+ /* Emit:
+ MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
+ LEN, BIAS, VEC_ARRAY). */
+ unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
+ tree alias_ptr = build_int_cst (ref_type, align);
+ call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
+ dataref_ptr, alias_ptr,
+ final_mask, final_len, bias,
+ vec_array);
+ }
+ else if (final_mask)
{
/* Emit:
MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
@@ -9598,9 +9639,11 @@ vectorizable_load (vec_info *vinfo,
enum dr_alignment_support alignment_support_scheme;
int misalignment;
poly_int64 poffset;
+ internal_fn lanes_ifn;
if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask, VLS_LOAD,
ncopies, &memory_access_type, &poffset,
- &alignment_support_scheme, &misalignment, &gs_info))
+ &alignment_support_scheme, &misalignment, &gs_info,
+ &lanes_ifn))
return false;
if (mask)
@@ -10386,6 +10429,8 @@ vectorizable_load (vec_info *vinfo,
tree vec_array = create_vector_array (vectype, vec_num);
tree final_mask = NULL_TREE;
+ tree final_len = NULL_TREE;
+ tree bias = NULL_TREE;
if (loop_masks)
final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
ncopies, vectype, j);
@@ -10393,8 +10438,36 @@ vectorizable_load (vec_info *vinfo,
final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
vec_mask, gsi);
+ if (lanes_ifn == IFN_MASK_LEN_LOAD_LANES)
+ {
+ if (loop_lens)
+ final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ ncopies, vectype, j, 1);
+ else
+ final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
+ signed char biasval
+ = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
+ bias = build_int_cst (intQI_type_node, biasval);
+ if (!final_mask)
+ {
+ mask_vectype = truth_type_for (vectype);
+ final_mask = build_minus_one_cst (mask_vectype);
+ }
+ }
+
gcall *call;
- if (final_mask)
+ if (final_len && final_mask)
+ {
+ /* Emit:
+ VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
+ VEC_MASK, LEN, BIAS). */
+ unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
+ tree alias_ptr = build_int_cst (ref_type, align);
+ call = gimple_build_call_internal (IFN_MASK_LEN_LOAD_LANES, 5,
+ dataref_ptr, alias_ptr,
+ final_mask, final_len, bias);
+ }
+ else if (final_mask)
{
/* Emit:
VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 1de144988c8..53a3d78d545 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -2297,9 +2297,9 @@ extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *,
extern void vect_copy_ref_info (tree, tree);
extern tree vect_create_destination_var (tree, tree);
extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
-extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
+extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
-extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
+extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
extern void vect_permute_store_chain (vec_info *, vec<tree> &,
unsigned int, stmt_vec_info,
gimple_stmt_iterator *, vec<tree> *);
--
2.36.3
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH V2] VECT: Apply MASK_LEN_{LOAD_LANES, STORE_LANES} into vectorizer
2023-08-15 12:29 [PATCH V2] VECT: Apply MASK_LEN_{LOAD_LANES, STORE_LANES} into vectorizer Juzhe-Zhong
@ 2023-08-15 12:35 ` Richard Biener
2023-08-16 6:06 ` Li, Pan2
0 siblings, 1 reply; 3+ messages in thread
From: Richard Biener @ 2023-08-15 12:35 UTC (permalink / raw)
To: Juzhe-Zhong; +Cc: gcc-patches, richard.sandiford
On Tue, 15 Aug 2023, Juzhe-Zhong wrote:
> Hi, Richard and Richi.
>
> This patch is adding MASK_LEN_{LOAD_LANES,STORE_LANES} support into vectorizer.
>
> Consider this simple case:
>
> void __attribute__ ((noinline, noclone))
> foo (int *__restrict a, int *__restrict b, int *__restrict c,
> int *__restrict d, int *__restrict e, int *__restrict f,
> int *__restrict g, int *__restrict h, int *__restrict j, int n)
> {
> for (int i = 0; i < n; ++i)
> {
> a[i] = j[i * 8];
> b[i] = j[i * 8 + 1];
> c[i] = j[i * 8 + 2];
> d[i] = j[i * 8 + 3];
> e[i] = j[i * 8 + 4];
> f[i] = j[i * 8 + 5];
> g[i] = j[i * 8 + 6];
> h[i] = j[i * 8 + 7];
> }
> }
>
> RVV Gimple IR:
>
> _79 = .SELECT_VL (ivtmp_81, POLY_INT_CST [4, 4]);
> ivtmp_125 = _79 * 32;
> vect_array.8 = .MASK_LEN_LOAD_LANES (vectp_j.6_124, 32B, { -1, ... }, _79, 0);
> vect__8.9_122 = vect_array.8[0];
> vect__8.10_121 = vect_array.8[1];
> vect__8.11_120 = vect_array.8[2];
> vect__8.12_119 = vect_array.8[3];
> vect__8.13_118 = vect_array.8[4];
> vect__8.14_117 = vect_array.8[5];
> vect__8.15_116 = vect_array.8[6];
> vect__8.16_115 = vect_array.8[7];
> vect_array.8 ={v} {CLOBBER};
> ivtmp_114 = _79 * 4;
> .MASK_LEN_STORE (vectp_a.17_113, 32B, { -1, ... }, _79, 0, vect__8.9_122);
> .MASK_LEN_STORE (vectp_b.19_109, 32B, { -1, ... }, _79, 0, vect__8.10_121);
> .MASK_LEN_STORE (vectp_c.21_105, 32B, { -1, ... }, _79, 0, vect__8.11_120);
> .MASK_LEN_STORE (vectp_d.23_101, 32B, { -1, ... }, _79, 0, vect__8.12_119);
> .MASK_LEN_STORE (vectp_e.25_97, 32B, { -1, ... }, _79, 0, vect__8.13_118);
> .MASK_LEN_STORE (vectp_f.27_93, 32B, { -1, ... }, _79, 0, vect__8.14_117);
> .MASK_LEN_STORE (vectp_g.29_89, 32B, { -1, ... }, _79, 0, vect__8.15_116);
> .MASK_LEN_STORE (vectp_h.31_85, 32B, { -1, ... }, _79, 0, vect__8.16_115);
>
> ASM:
>
> foo:
> lw t4,8(sp)
> ld t5,0(sp)
> ble t4,zero,.L5
> .L3:
> vsetvli t1,t4,e8,mf4,ta,ma
> vlseg8e32.v v8,(t5)
> slli t3,t1,2
> slli t6,t1,5
> vse32.v v8,0(a0)
> vse32.v v9,0(a1)
> vse32.v v10,0(a2)
> vse32.v v11,0(a3)
> vse32.v v12,0(a4)
> vse32.v v13,0(a5)
> vse32.v v14,0(a6)
> vse32.v v15,0(a7)
> sub t4,t4,t1
> add t5,t5,t6
> add a0,a0,t3
> add a1,a1,t3
> add a2,a2,t3
> add a3,a3,t3
> add a4,a4,t3
> add a5,a5,t3
> add a6,a6,t3
> add a7,a7,t3
> bne t4,zero,.L3
> .L5:
> ret
>
> The details of the approach:
>
> Step 1 - Modifiy the LANES LOAD/STORE support function (vect_load_lanes_supported/vect_store_lanes_supported):
>
> +/* Return FN if vec_{masked_,mask_len,}load_lanes is available for COUNT
> + vectors of type VECTYPE. MASKED_P says whether the masked form is needed. */
>
> -bool
> +internal_fn
> vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
> bool masked_p)
> {
> - if (masked_p)
> - return vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> - vec_mask_load_lanes_optab,
> - vectype, count);
> + if (vect_lanes_optab_supported_p ("vec_mask_len_load_lanes",
> + vec_mask_len_load_lanes_optab,
> + vectype, count))
> + return IFN_MASK_LEN_LOAD_LANES;
> + else if (masked_p)
> + {
> + if (vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> + vec_mask_load_lanes_optab,
> + vectype, count))
> + return IFN_MASK_LOAD_LANES;
> + }
> else
> - return vect_lanes_optab_supported_p ("vec_load_lanes",
> - vec_load_lanes_optab,
> - vectype, count);
> + {
> + if (vect_lanes_optab_supported_p ("vec_load_lanes",
> + vec_load_lanes_optab,
> + vectype, count))
> + return IFN_LOAD_LANES;
> + }
> + return IFN_LAST;
> }
>
> Instead of returning TRUE or FALSE whether target support the LANES LOAD/STORE.
> I change it into return internal_fn of the LANES LOAD/STORE that target support,
> If target didn't support any LANE LOAD/STORE optabs, return IFN_LAST.
>
> Step 2 - Compute IFN for LANES LOAD/STORE (Only compute once).
>
> if (!STMT_VINFO_STRIDED_P (first_stmt_info)
> && (can_overrun_p || !would_overrun_p)
> && compare_step_with_zero (vinfo, stmt_info) > 0)
> {
> /* First cope with the degenerate case of a single-element
> vector. */
> if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> ;
>
> else
> {
> /* Otherwise try using LOAD/STORE_LANES. */
> *lanes_ifn
> = vls_type == VLS_LOAD
> ? vect_load_lanes_supported (vectype, group_size, masked_p)
> : vect_store_lanes_supported (vectype, group_size,
> masked_p);
> if (*lanes_ifn != IFN_LAST)
> {
> *memory_access_type = VMAT_LOAD_STORE_LANES;
> overrun_p = would_overrun_p;
> }
>
> /* If that fails, try using permuting loads. */
> else if (vls_type == VLS_LOAD
> ? vect_grouped_load_supported (vectype,
> single_element_p,
> group_size)
> : vect_grouped_store_supported (vectype, group_size))
> {
> *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> overrun_p = would_overrun_p;
> }
> }
> }
>
> Step 3 - Build MASK_LEN_{LANES_LOAD,LANES_STORE} Gimple IR:
>
> + if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
> + {
> + if (loop_lens)
> + final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + ncopies, vectype, j, 1);
> + else
> + final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
> + signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + if (!final_mask)
> + {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> + }
> + }
> +
> gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> + {
> + /* Emit:
> + MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
> + LEN, BIAS, VEC_ARRAY). */
> + unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
> + tree alias_ptr = build_int_cst (ref_type, align);
> + call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
> + dataref_ptr, alias_ptr,
> + final_mask, final_len, bias,
> + vec_array);
> + }
> + else if (final_mask)
>
> The LEN and MASK flow is totally the same as other MASK_LEN_* load/store.
OK.
Thanks,
Richard.
> gcc/ChangeLog:
>
> * internal-fn.cc (internal_load_fn_p): Apply MASK_LEN_{LOAD_LANES,STORE_LANES} into vectorizer.
> (internal_store_fn_p): Ditto.
> (internal_fn_len_index): Ditto.
> (internal_fn_mask_index): Ditto.
> (internal_fn_stored_value_index): Ditto.
> * tree-vect-data-refs.cc (vect_store_lanes_supported): Ditto.
> (vect_load_lanes_supported): Ditto.
> * tree-vect-loop.cc: Ditto.
> * tree-vect-slp.cc (vect_slp_prefer_store_lanes_p): Ditto.
> * tree-vect-stmts.cc (check_load_store_for_partial_vectors): Ditto.
> (get_group_load_store_type): Ditto.
> (vectorizable_store): Ditto.
> (vectorizable_load): Ditto.
> * tree-vectorizer.h (vect_store_lanes_supported): Ditto.
> (vect_load_lanes_supported): Ditto.
>
> ---
> gcc/internal-fn.cc | 7 ++
> gcc/tree-vect-data-refs.cc | 61 ++++++++++------
> gcc/tree-vect-loop.cc | 11 +--
> gcc/tree-vect-slp.cc | 2 +-
> gcc/tree-vect-stmts.cc | 141 ++++++++++++++++++++++++++++---------
> gcc/tree-vectorizer.h | 4 +-
> 6 files changed, 163 insertions(+), 63 deletions(-)
>
> diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> index 4f2b20a79e5..cc1ede58799 100644
> --- a/gcc/internal-fn.cc
> +++ b/gcc/internal-fn.cc
> @@ -4578,6 +4578,7 @@ internal_load_fn_p (internal_fn fn)
> case IFN_MASK_LOAD:
> case IFN_LOAD_LANES:
> case IFN_MASK_LOAD_LANES:
> + case IFN_MASK_LEN_LOAD_LANES:
> case IFN_GATHER_LOAD:
> case IFN_MASK_GATHER_LOAD:
> case IFN_MASK_LEN_GATHER_LOAD:
> @@ -4600,6 +4601,7 @@ internal_store_fn_p (internal_fn fn)
> case IFN_MASK_STORE:
> case IFN_STORE_LANES:
> case IFN_MASK_STORE_LANES:
> + case IFN_MASK_LEN_STORE_LANES:
> case IFN_SCATTER_STORE:
> case IFN_MASK_SCATTER_STORE:
> case IFN_MASK_LEN_SCATTER_STORE:
> @@ -4672,6 +4674,8 @@ internal_fn_len_index (internal_fn fn)
> case IFN_COND_LEN_NEG:
> case IFN_MASK_LEN_LOAD:
> case IFN_MASK_LEN_STORE:
> + case IFN_MASK_LEN_LOAD_LANES:
> + case IFN_MASK_LEN_STORE_LANES:
> return 3;
>
> default:
> @@ -4689,8 +4693,10 @@ internal_fn_mask_index (internal_fn fn)
> {
> case IFN_MASK_LOAD:
> case IFN_MASK_LOAD_LANES:
> + case IFN_MASK_LEN_LOAD_LANES:
> case IFN_MASK_STORE:
> case IFN_MASK_STORE_LANES:
> + case IFN_MASK_LEN_STORE_LANES:
> case IFN_MASK_LEN_LOAD:
> case IFN_MASK_LEN_STORE:
> return 2;
> @@ -4726,6 +4732,7 @@ internal_fn_stored_value_index (internal_fn fn)
> return 4;
>
> case IFN_MASK_LEN_STORE:
> + case IFN_MASK_LEN_STORE_LANES:
> return 5;
>
> default:
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index a3570c45b52..3e9a284666c 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5438,22 +5438,31 @@ vect_grouped_store_supported (tree vectype, unsigned HOST_WIDE_INT count)
> return false;
> }
>
> +/* Return FN if vec_{mask_,mask_len_}store_lanes is available for COUNT vectors
> + of type VECTYPE. MASKED_P says whether the masked form is needed. */
>
> -/* Return TRUE if vec_{mask_}store_lanes is available for COUNT vectors of
> - type VECTYPE. MASKED_P says whether the masked form is needed. */
> -
> -bool
> +internal_fn
> vect_store_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
> bool masked_p)
> {
> - if (masked_p)
> - return vect_lanes_optab_supported_p ("vec_mask_store_lanes",
> - vec_mask_store_lanes_optab,
> - vectype, count);
> + if (vect_lanes_optab_supported_p ("vec_mask_len_store_lanes",
> + vec_mask_len_store_lanes_optab, vectype,
> + count))
> + return IFN_MASK_LEN_STORE_LANES;
> + else if (masked_p)
> + {
> + if (vect_lanes_optab_supported_p ("vec_mask_store_lanes",
> + vec_mask_store_lanes_optab, vectype,
> + count))
> + return IFN_MASK_STORE_LANES;
> + }
> else
> - return vect_lanes_optab_supported_p ("vec_store_lanes",
> - vec_store_lanes_optab,
> - vectype, count);
> + {
> + if (vect_lanes_optab_supported_p ("vec_store_lanes",
> + vec_store_lanes_optab, vectype, count))
> + return IFN_STORE_LANES;
> + }
> + return IFN_LAST;
> }
>
>
> @@ -6056,21 +6065,31 @@ vect_grouped_load_supported (tree vectype, bool single_element_p,
> return false;
> }
>
> -/* Return TRUE if vec_{masked_}load_lanes is available for COUNT vectors of
> - type VECTYPE. MASKED_P says whether the masked form is needed. */
> +/* Return FN if vec_{masked_,mask_len_}load_lanes is available for COUNT vectors
> + of type VECTYPE. MASKED_P says whether the masked form is needed. */
>
> -bool
> +internal_fn
> vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
> bool masked_p)
> {
> - if (masked_p)
> - return vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> - vec_mask_load_lanes_optab,
> - vectype, count);
> + if (vect_lanes_optab_supported_p ("vec_mask_len_load_lanes",
> + vec_mask_len_load_lanes_optab, vectype,
> + count))
> + return IFN_MASK_LEN_LOAD_LANES;
> + else if (masked_p)
> + {
> + if (vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> + vec_mask_load_lanes_optab, vectype,
> + count))
> + return IFN_MASK_LOAD_LANES;
> + }
> else
> - return vect_lanes_optab_supported_p ("vec_load_lanes",
> - vec_load_lanes_optab,
> - vectype, count);
> + {
> + if (vect_lanes_optab_supported_p ("vec_load_lanes", vec_load_lanes_optab,
> + vectype, count))
> + return IFN_LOAD_LANES;
> + }
> + return IFN_LAST;
> }
>
> /* Function vect_permute_load_chain.
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index bc3063c3615..1fcd8d07ea1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -2839,7 +2839,8 @@ start_over:
> instructions record it and move on to the next instance. */
> if (loads_permuted
> && SLP_INSTANCE_KIND (instance) == slp_inst_kind_store
> - && vect_store_lanes_supported (vectype, group_size, false))
> + && vect_store_lanes_supported (vectype, group_size, false)
> + != IFN_LAST)
> {
> FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (instance), i, load_node)
> {
> @@ -2848,9 +2849,9 @@ start_over:
> /* Use SLP for strided accesses (or if we can't
> load-lanes). */
> if (STMT_VINFO_STRIDED_P (stmt_vinfo)
> - || ! vect_load_lanes_supported
> + || vect_load_lanes_supported
> (STMT_VINFO_VECTYPE (stmt_vinfo),
> - DR_GROUP_SIZE (stmt_vinfo), false))
> + DR_GROUP_SIZE (stmt_vinfo), false) == IFN_LAST)
> break;
> }
>
> @@ -3153,7 +3154,7 @@ again:
> vinfo = DR_GROUP_FIRST_ELEMENT (vinfo);
> unsigned int size = DR_GROUP_SIZE (vinfo);
> tree vectype = STMT_VINFO_VECTYPE (vinfo);
> - if (! vect_store_lanes_supported (vectype, size, false)
> + if (vect_store_lanes_supported (vectype, size, false) == IFN_LAST
> && ! known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)
> && ! vect_grouped_store_supported (vectype, size))
> return opt_result::failure_at (vinfo->stmt,
> @@ -3165,7 +3166,7 @@ again:
> bool single_element_p = !DR_GROUP_NEXT_ELEMENT (vinfo);
> size = DR_GROUP_SIZE (vinfo);
> vectype = STMT_VINFO_VECTYPE (vinfo);
> - if (! vect_load_lanes_supported (vectype, size, false)
> + if (vect_load_lanes_supported (vectype, size, false) == IFN_LAST
> && ! vect_grouped_load_supported (vectype, single_element_p,
> size))
> return opt_result::failure_at (vinfo->stmt,
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index cf91b21cf7d..9ad2634762e 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -3094,7 +3094,7 @@ vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
> if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
> || multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
> return false;
> - return vect_store_lanes_supported (vectype, group_size, false);
> + return vect_store_lanes_supported (vectype, group_size, false) != IFN_LAST;
> }
>
> /* Analyze an SLP instance starting from a group of grouped stores. Call
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 86d033aa60c..cd8e0a76374 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -1610,9 +1610,15 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
> bool is_load = (vls_type == VLS_LOAD);
> if (memory_access_type == VMAT_LOAD_STORE_LANES)
> {
> - if (is_load
> - ? !vect_load_lanes_supported (vectype, group_size, true)
> - : !vect_store_lanes_supported (vectype, group_size, true))
> + internal_fn ifn
> + = (is_load ? vect_load_lanes_supported (vectype, group_size, true)
> + : vect_store_lanes_supported (vectype, group_size, true));
> + if (ifn == IFN_MASK_LEN_LOAD_LANES || ifn == IFN_MASK_LEN_STORE_LANES)
> + vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1);
> + else if (ifn == IFN_MASK_LOAD_LANES || ifn == IFN_MASK_STORE_LANES)
> + vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
> + scalar_mask);
> + else
> {
> if (dump_enabled_p ())
> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> @@ -1620,10 +1626,7 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
> " the target doesn't have an appropriate"
> " load/store-lanes instruction.\n");
> LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
> - return;
> }
> - vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
> - scalar_mask);
> return;
> }
>
> @@ -2074,7 +2077,8 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> poly_int64 *poffset,
> dr_alignment_support *alignment_support_scheme,
> int *misalignment,
> - gather_scatter_info *gs_info)
> + gather_scatter_info *gs_info,
> + internal_fn *lanes_ifn)
> {
> loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
> class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
> @@ -2272,24 +2276,30 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> ;
>
> - /* Otherwise try using LOAD/STORE_LANES. */
> - else if (vls_type == VLS_LOAD
> - ? vect_load_lanes_supported (vectype, group_size, masked_p)
> - : vect_store_lanes_supported (vectype, group_size,
> - masked_p))
> + else
> {
> - *memory_access_type = VMAT_LOAD_STORE_LANES;
> - overrun_p = would_overrun_p;
> - }
> + /* Otherwise try using LOAD/STORE_LANES. */
> + *lanes_ifn
> + = vls_type == VLS_LOAD
> + ? vect_load_lanes_supported (vectype, group_size, masked_p)
> + : vect_store_lanes_supported (vectype, group_size,
> + masked_p);
> + if (*lanes_ifn != IFN_LAST)
> + {
> + *memory_access_type = VMAT_LOAD_STORE_LANES;
> + overrun_p = would_overrun_p;
> + }
>
> - /* If that fails, try using permuting loads. */
> - else if (vls_type == VLS_LOAD
> - ? vect_grouped_load_supported (vectype, single_element_p,
> - group_size)
> - : vect_grouped_store_supported (vectype, group_size))
> - {
> - *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> - overrun_p = would_overrun_p;
> + /* If that fails, try using permuting loads. */
> + else if (vls_type == VLS_LOAD
> + ? vect_grouped_load_supported (vectype,
> + single_element_p,
> + group_size)
> + : vect_grouped_store_supported (vectype, group_size))
> + {
> + *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> + overrun_p = would_overrun_p;
> + }
> }
> }
>
> @@ -2378,7 +2388,8 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> poly_int64 *poffset,
> dr_alignment_support *alignment_support_scheme,
> int *misalignment,
> - gather_scatter_info *gs_info)
> + gather_scatter_info *gs_info,
> + internal_fn *lanes_ifn)
> {
> loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
> poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> @@ -2441,7 +2452,7 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> masked_p,
> vls_type, memory_access_type, poffset,
> alignment_support_scheme,
> - misalignment, gs_info))
> + misalignment, gs_info, lanes_ifn))
> return false;
> }
> else if (STMT_VINFO_STRIDED_P (stmt_info))
> @@ -3087,11 +3098,8 @@ vect_get_loop_variant_data_ptr_increment (
> loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> tree step = vect_dr_behavior (vinfo, dr_info)->step;
>
> - /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> - IVs are updated by variable amount but we will support them in the future.
> - */
> - gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> - && memory_access_type != VMAT_LOAD_STORE_LANES);
> + /* gather/scatter never reach here. */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER);
>
> /* When we support SELECT_VL pattern, we dynamic adjust
> the memory address by .SELECT_VL result.
> @@ -8094,9 +8102,11 @@ vectorizable_store (vec_info *vinfo,
> enum dr_alignment_support alignment_support_scheme;
> int misalignment;
> poly_int64 poffset;
> + internal_fn lanes_ifn;
> if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask, vls_type,
> ncopies, &memory_access_type, &poffset,
> - &alignment_support_scheme, &misalignment, &gs_info))
> + &alignment_support_scheme, &misalignment, &gs_info,
> + &lanes_ifn))
> return false;
>
> if (mask)
> @@ -8885,6 +8895,8 @@ vectorizable_store (vec_info *vinfo,
> }
>
> tree final_mask = NULL;
> + tree final_len = NULL;
> + tree bias = NULL;
> if (loop_masks)
> final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> ncopies, vectype, j);
> @@ -8892,8 +8904,37 @@ vectorizable_store (vec_info *vinfo,
> final_mask = prepare_vec_mask (loop_vinfo, mask_vectype,
> final_mask, vec_mask, gsi);
>
> + if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
> + {
> + if (loop_lens)
> + final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + ncopies, vectype, j, 1);
> + else
> + final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
> + signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + if (!final_mask)
> + {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> + }
> + }
> +
> gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> + {
> + /* Emit:
> + MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
> + LEN, BIAS, VEC_ARRAY). */
> + unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
> + tree alias_ptr = build_int_cst (ref_type, align);
> + call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
> + dataref_ptr, alias_ptr,
> + final_mask, final_len, bias,
> + vec_array);
> + }
> + else if (final_mask)
> {
> /* Emit:
> MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
> @@ -9598,9 +9639,11 @@ vectorizable_load (vec_info *vinfo,
> enum dr_alignment_support alignment_support_scheme;
> int misalignment;
> poly_int64 poffset;
> + internal_fn lanes_ifn;
> if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask, VLS_LOAD,
> ncopies, &memory_access_type, &poffset,
> - &alignment_support_scheme, &misalignment, &gs_info))
> + &alignment_support_scheme, &misalignment, &gs_info,
> + &lanes_ifn))
> return false;
>
> if (mask)
> @@ -10386,6 +10429,8 @@ vectorizable_load (vec_info *vinfo,
> tree vec_array = create_vector_array (vectype, vec_num);
>
> tree final_mask = NULL_TREE;
> + tree final_len = NULL_TREE;
> + tree bias = NULL_TREE;
> if (loop_masks)
> final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> ncopies, vectype, j);
> @@ -10393,8 +10438,36 @@ vectorizable_load (vec_info *vinfo,
> final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
> vec_mask, gsi);
>
> + if (lanes_ifn == IFN_MASK_LEN_LOAD_LANES)
> + {
> + if (loop_lens)
> + final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + ncopies, vectype, j, 1);
> + else
> + final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
> + signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + if (!final_mask)
> + {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> + }
> + }
> +
> gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> + {
> + /* Emit:
> + VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
> + VEC_MASK, LEN, BIAS). */
> + unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
> + tree alias_ptr = build_int_cst (ref_type, align);
> + call = gimple_build_call_internal (IFN_MASK_LEN_LOAD_LANES, 5,
> + dataref_ptr, alias_ptr,
> + final_mask, final_len, bias);
> + }
> + else if (final_mask)
> {
> /* Emit:
> VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 1de144988c8..53a3d78d545 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2297,9 +2297,9 @@ extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *,
> extern void vect_copy_ref_info (tree, tree);
> extern tree vect_create_destination_var (tree, tree);
> extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
> -extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> +extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
> -extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> +extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> extern void vect_permute_store_chain (vec_info *, vec<tree> &,
> unsigned int, stmt_vec_info,
> gimple_stmt_iterator *, vec<tree> *);
>
--
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
^ permalink raw reply [flat|nested] 3+ messages in thread
* RE: [PATCH V2] VECT: Apply MASK_LEN_{LOAD_LANES, STORE_LANES} into vectorizer
2023-08-15 12:35 ` Richard Biener
@ 2023-08-16 6:06 ` Li, Pan2
0 siblings, 0 replies; 3+ messages in thread
From: Li, Pan2 @ 2023-08-16 6:06 UTC (permalink / raw)
To: Richard Biener, Juzhe-Zhong; +Cc: gcc-patches, richard.sandiford
Committed, thanks Richard.
Pan
-----Original Message-----
From: Gcc-patches <gcc-patches-bounces+pan2.li=intel.com@gcc.gnu.org> On Behalf Of Richard Biener via Gcc-patches
Sent: Tuesday, August 15, 2023 8:35 PM
To: Juzhe-Zhong <juzhe.zhong@rivai.ai>
Cc: gcc-patches@gcc.gnu.org; richard.sandiford@arm.com
Subject: Re: [PATCH V2] VECT: Apply MASK_LEN_{LOAD_LANES, STORE_LANES} into vectorizer
On Tue, 15 Aug 2023, Juzhe-Zhong wrote:
> Hi, Richard and Richi.
>
> This patch is adding MASK_LEN_{LOAD_LANES,STORE_LANES} support into vectorizer.
>
> Consider this simple case:
>
> void __attribute__ ((noinline, noclone))
> foo (int *__restrict a, int *__restrict b, int *__restrict c,
> int *__restrict d, int *__restrict e, int *__restrict f,
> int *__restrict g, int *__restrict h, int *__restrict j, int n)
> {
> for (int i = 0; i < n; ++i)
> {
> a[i] = j[i * 8];
> b[i] = j[i * 8 + 1];
> c[i] = j[i * 8 + 2];
> d[i] = j[i * 8 + 3];
> e[i] = j[i * 8 + 4];
> f[i] = j[i * 8 + 5];
> g[i] = j[i * 8 + 6];
> h[i] = j[i * 8 + 7];
> }
> }
>
> RVV Gimple IR:
>
> _79 = .SELECT_VL (ivtmp_81, POLY_INT_CST [4, 4]);
> ivtmp_125 = _79 * 32;
> vect_array.8 = .MASK_LEN_LOAD_LANES (vectp_j.6_124, 32B, { -1, ... }, _79, 0);
> vect__8.9_122 = vect_array.8[0];
> vect__8.10_121 = vect_array.8[1];
> vect__8.11_120 = vect_array.8[2];
> vect__8.12_119 = vect_array.8[3];
> vect__8.13_118 = vect_array.8[4];
> vect__8.14_117 = vect_array.8[5];
> vect__8.15_116 = vect_array.8[6];
> vect__8.16_115 = vect_array.8[7];
> vect_array.8 ={v} {CLOBBER};
> ivtmp_114 = _79 * 4;
> .MASK_LEN_STORE (vectp_a.17_113, 32B, { -1, ... }, _79, 0, vect__8.9_122);
> .MASK_LEN_STORE (vectp_b.19_109, 32B, { -1, ... }, _79, 0, vect__8.10_121);
> .MASK_LEN_STORE (vectp_c.21_105, 32B, { -1, ... }, _79, 0, vect__8.11_120);
> .MASK_LEN_STORE (vectp_d.23_101, 32B, { -1, ... }, _79, 0, vect__8.12_119);
> .MASK_LEN_STORE (vectp_e.25_97, 32B, { -1, ... }, _79, 0, vect__8.13_118);
> .MASK_LEN_STORE (vectp_f.27_93, 32B, { -1, ... }, _79, 0, vect__8.14_117);
> .MASK_LEN_STORE (vectp_g.29_89, 32B, { -1, ... }, _79, 0, vect__8.15_116);
> .MASK_LEN_STORE (vectp_h.31_85, 32B, { -1, ... }, _79, 0, vect__8.16_115);
>
> ASM:
>
> foo:
> lw t4,8(sp)
> ld t5,0(sp)
> ble t4,zero,.L5
> .L3:
> vsetvli t1,t4,e8,mf4,ta,ma
> vlseg8e32.v v8,(t5)
> slli t3,t1,2
> slli t6,t1,5
> vse32.v v8,0(a0)
> vse32.v v9,0(a1)
> vse32.v v10,0(a2)
> vse32.v v11,0(a3)
> vse32.v v12,0(a4)
> vse32.v v13,0(a5)
> vse32.v v14,0(a6)
> vse32.v v15,0(a7)
> sub t4,t4,t1
> add t5,t5,t6
> add a0,a0,t3
> add a1,a1,t3
> add a2,a2,t3
> add a3,a3,t3
> add a4,a4,t3
> add a5,a5,t3
> add a6,a6,t3
> add a7,a7,t3
> bne t4,zero,.L3
> .L5:
> ret
>
> The details of the approach:
>
> Step 1 - Modifiy the LANES LOAD/STORE support function (vect_load_lanes_supported/vect_store_lanes_supported):
>
> +/* Return FN if vec_{masked_,mask_len,}load_lanes is available for COUNT
> + vectors of type VECTYPE. MASKED_P says whether the masked form is needed. */
>
> -bool
> +internal_fn
> vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
> bool masked_p)
> {
> - if (masked_p)
> - return vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> - vec_mask_load_lanes_optab,
> - vectype, count);
> + if (vect_lanes_optab_supported_p ("vec_mask_len_load_lanes",
> + vec_mask_len_load_lanes_optab,
> + vectype, count))
> + return IFN_MASK_LEN_LOAD_LANES;
> + else if (masked_p)
> + {
> + if (vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> + vec_mask_load_lanes_optab,
> + vectype, count))
> + return IFN_MASK_LOAD_LANES;
> + }
> else
> - return vect_lanes_optab_supported_p ("vec_load_lanes",
> - vec_load_lanes_optab,
> - vectype, count);
> + {
> + if (vect_lanes_optab_supported_p ("vec_load_lanes",
> + vec_load_lanes_optab,
> + vectype, count))
> + return IFN_LOAD_LANES;
> + }
> + return IFN_LAST;
> }
>
> Instead of returning TRUE or FALSE whether target support the LANES LOAD/STORE.
> I change it into return internal_fn of the LANES LOAD/STORE that target support,
> If target didn't support any LANE LOAD/STORE optabs, return IFN_LAST.
>
> Step 2 - Compute IFN for LANES LOAD/STORE (Only compute once).
>
> if (!STMT_VINFO_STRIDED_P (first_stmt_info)
> && (can_overrun_p || !would_overrun_p)
> && compare_step_with_zero (vinfo, stmt_info) > 0)
> {
> /* First cope with the degenerate case of a single-element
> vector. */
> if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> ;
>
> else
> {
> /* Otherwise try using LOAD/STORE_LANES. */
> *lanes_ifn
> = vls_type == VLS_LOAD
> ? vect_load_lanes_supported (vectype, group_size, masked_p)
> : vect_store_lanes_supported (vectype, group_size,
> masked_p);
> if (*lanes_ifn != IFN_LAST)
> {
> *memory_access_type = VMAT_LOAD_STORE_LANES;
> overrun_p = would_overrun_p;
> }
>
> /* If that fails, try using permuting loads. */
> else if (vls_type == VLS_LOAD
> ? vect_grouped_load_supported (vectype,
> single_element_p,
> group_size)
> : vect_grouped_store_supported (vectype, group_size))
> {
> *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> overrun_p = would_overrun_p;
> }
> }
> }
>
> Step 3 - Build MASK_LEN_{LANES_LOAD,LANES_STORE} Gimple IR:
>
> + if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
> + {
> + if (loop_lens)
> + final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + ncopies, vectype, j, 1);
> + else
> + final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
> + signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + if (!final_mask)
> + {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> + }
> + }
> +
> gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> + {
> + /* Emit:
> + MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
> + LEN, BIAS, VEC_ARRAY). */
> + unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
> + tree alias_ptr = build_int_cst (ref_type, align);
> + call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
> + dataref_ptr, alias_ptr,
> + final_mask, final_len, bias,
> + vec_array);
> + }
> + else if (final_mask)
>
> The LEN and MASK flow is totally the same as other MASK_LEN_* load/store.
OK.
Thanks,
Richard.
> gcc/ChangeLog:
>
> * internal-fn.cc (internal_load_fn_p): Apply MASK_LEN_{LOAD_LANES,STORE_LANES} into vectorizer.
> (internal_store_fn_p): Ditto.
> (internal_fn_len_index): Ditto.
> (internal_fn_mask_index): Ditto.
> (internal_fn_stored_value_index): Ditto.
> * tree-vect-data-refs.cc (vect_store_lanes_supported): Ditto.
> (vect_load_lanes_supported): Ditto.
> * tree-vect-loop.cc: Ditto.
> * tree-vect-slp.cc (vect_slp_prefer_store_lanes_p): Ditto.
> * tree-vect-stmts.cc (check_load_store_for_partial_vectors): Ditto.
> (get_group_load_store_type): Ditto.
> (vectorizable_store): Ditto.
> (vectorizable_load): Ditto.
> * tree-vectorizer.h (vect_store_lanes_supported): Ditto.
> (vect_load_lanes_supported): Ditto.
>
> ---
> gcc/internal-fn.cc | 7 ++
> gcc/tree-vect-data-refs.cc | 61 ++++++++++------
> gcc/tree-vect-loop.cc | 11 +--
> gcc/tree-vect-slp.cc | 2 +-
> gcc/tree-vect-stmts.cc | 141 ++++++++++++++++++++++++++++---------
> gcc/tree-vectorizer.h | 4 +-
> 6 files changed, 163 insertions(+), 63 deletions(-)
>
> diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> index 4f2b20a79e5..cc1ede58799 100644
> --- a/gcc/internal-fn.cc
> +++ b/gcc/internal-fn.cc
> @@ -4578,6 +4578,7 @@ internal_load_fn_p (internal_fn fn)
> case IFN_MASK_LOAD:
> case IFN_LOAD_LANES:
> case IFN_MASK_LOAD_LANES:
> + case IFN_MASK_LEN_LOAD_LANES:
> case IFN_GATHER_LOAD:
> case IFN_MASK_GATHER_LOAD:
> case IFN_MASK_LEN_GATHER_LOAD:
> @@ -4600,6 +4601,7 @@ internal_store_fn_p (internal_fn fn)
> case IFN_MASK_STORE:
> case IFN_STORE_LANES:
> case IFN_MASK_STORE_LANES:
> + case IFN_MASK_LEN_STORE_LANES:
> case IFN_SCATTER_STORE:
> case IFN_MASK_SCATTER_STORE:
> case IFN_MASK_LEN_SCATTER_STORE:
> @@ -4672,6 +4674,8 @@ internal_fn_len_index (internal_fn fn)
> case IFN_COND_LEN_NEG:
> case IFN_MASK_LEN_LOAD:
> case IFN_MASK_LEN_STORE:
> + case IFN_MASK_LEN_LOAD_LANES:
> + case IFN_MASK_LEN_STORE_LANES:
> return 3;
>
> default:
> @@ -4689,8 +4693,10 @@ internal_fn_mask_index (internal_fn fn)
> {
> case IFN_MASK_LOAD:
> case IFN_MASK_LOAD_LANES:
> + case IFN_MASK_LEN_LOAD_LANES:
> case IFN_MASK_STORE:
> case IFN_MASK_STORE_LANES:
> + case IFN_MASK_LEN_STORE_LANES:
> case IFN_MASK_LEN_LOAD:
> case IFN_MASK_LEN_STORE:
> return 2;
> @@ -4726,6 +4732,7 @@ internal_fn_stored_value_index (internal_fn fn)
> return 4;
>
> case IFN_MASK_LEN_STORE:
> + case IFN_MASK_LEN_STORE_LANES:
> return 5;
>
> default:
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index a3570c45b52..3e9a284666c 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5438,22 +5438,31 @@ vect_grouped_store_supported (tree vectype, unsigned HOST_WIDE_INT count)
> return false;
> }
>
> +/* Return FN if vec_{mask_,mask_len_}store_lanes is available for COUNT vectors
> + of type VECTYPE. MASKED_P says whether the masked form is needed. */
>
> -/* Return TRUE if vec_{mask_}store_lanes is available for COUNT vectors of
> - type VECTYPE. MASKED_P says whether the masked form is needed. */
> -
> -bool
> +internal_fn
> vect_store_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
> bool masked_p)
> {
> - if (masked_p)
> - return vect_lanes_optab_supported_p ("vec_mask_store_lanes",
> - vec_mask_store_lanes_optab,
> - vectype, count);
> + if (vect_lanes_optab_supported_p ("vec_mask_len_store_lanes",
> + vec_mask_len_store_lanes_optab, vectype,
> + count))
> + return IFN_MASK_LEN_STORE_LANES;
> + else if (masked_p)
> + {
> + if (vect_lanes_optab_supported_p ("vec_mask_store_lanes",
> + vec_mask_store_lanes_optab, vectype,
> + count))
> + return IFN_MASK_STORE_LANES;
> + }
> else
> - return vect_lanes_optab_supported_p ("vec_store_lanes",
> - vec_store_lanes_optab,
> - vectype, count);
> + {
> + if (vect_lanes_optab_supported_p ("vec_store_lanes",
> + vec_store_lanes_optab, vectype, count))
> + return IFN_STORE_LANES;
> + }
> + return IFN_LAST;
> }
>
>
> @@ -6056,21 +6065,31 @@ vect_grouped_load_supported (tree vectype, bool single_element_p,
> return false;
> }
>
> -/* Return TRUE if vec_{masked_}load_lanes is available for COUNT vectors of
> - type VECTYPE. MASKED_P says whether the masked form is needed. */
> +/* Return FN if vec_{masked_,mask_len_}load_lanes is available for COUNT vectors
> + of type VECTYPE. MASKED_P says whether the masked form is needed. */
>
> -bool
> +internal_fn
> vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count,
> bool masked_p)
> {
> - if (masked_p)
> - return vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> - vec_mask_load_lanes_optab,
> - vectype, count);
> + if (vect_lanes_optab_supported_p ("vec_mask_len_load_lanes",
> + vec_mask_len_load_lanes_optab, vectype,
> + count))
> + return IFN_MASK_LEN_LOAD_LANES;
> + else if (masked_p)
> + {
> + if (vect_lanes_optab_supported_p ("vec_mask_load_lanes",
> + vec_mask_load_lanes_optab, vectype,
> + count))
> + return IFN_MASK_LOAD_LANES;
> + }
> else
> - return vect_lanes_optab_supported_p ("vec_load_lanes",
> - vec_load_lanes_optab,
> - vectype, count);
> + {
> + if (vect_lanes_optab_supported_p ("vec_load_lanes", vec_load_lanes_optab,
> + vectype, count))
> + return IFN_LOAD_LANES;
> + }
> + return IFN_LAST;
> }
>
> /* Function vect_permute_load_chain.
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index bc3063c3615..1fcd8d07ea1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -2839,7 +2839,8 @@ start_over:
> instructions record it and move on to the next instance. */
> if (loads_permuted
> && SLP_INSTANCE_KIND (instance) == slp_inst_kind_store
> - && vect_store_lanes_supported (vectype, group_size, false))
> + && vect_store_lanes_supported (vectype, group_size, false)
> + != IFN_LAST)
> {
> FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (instance), i, load_node)
> {
> @@ -2848,9 +2849,9 @@ start_over:
> /* Use SLP for strided accesses (or if we can't
> load-lanes). */
> if (STMT_VINFO_STRIDED_P (stmt_vinfo)
> - || ! vect_load_lanes_supported
> + || vect_load_lanes_supported
> (STMT_VINFO_VECTYPE (stmt_vinfo),
> - DR_GROUP_SIZE (stmt_vinfo), false))
> + DR_GROUP_SIZE (stmt_vinfo), false) == IFN_LAST)
> break;
> }
>
> @@ -3153,7 +3154,7 @@ again:
> vinfo = DR_GROUP_FIRST_ELEMENT (vinfo);
> unsigned int size = DR_GROUP_SIZE (vinfo);
> tree vectype = STMT_VINFO_VECTYPE (vinfo);
> - if (! vect_store_lanes_supported (vectype, size, false)
> + if (vect_store_lanes_supported (vectype, size, false) == IFN_LAST
> && ! known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)
> && ! vect_grouped_store_supported (vectype, size))
> return opt_result::failure_at (vinfo->stmt,
> @@ -3165,7 +3166,7 @@ again:
> bool single_element_p = !DR_GROUP_NEXT_ELEMENT (vinfo);
> size = DR_GROUP_SIZE (vinfo);
> vectype = STMT_VINFO_VECTYPE (vinfo);
> - if (! vect_load_lanes_supported (vectype, size, false)
> + if (vect_load_lanes_supported (vectype, size, false) == IFN_LAST
> && ! vect_grouped_load_supported (vectype, single_element_p,
> size))
> return opt_result::failure_at (vinfo->stmt,
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index cf91b21cf7d..9ad2634762e 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -3094,7 +3094,7 @@ vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
> if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
> || multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
> return false;
> - return vect_store_lanes_supported (vectype, group_size, false);
> + return vect_store_lanes_supported (vectype, group_size, false) != IFN_LAST;
> }
>
> /* Analyze an SLP instance starting from a group of grouped stores. Call
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 86d033aa60c..cd8e0a76374 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -1610,9 +1610,15 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
> bool is_load = (vls_type == VLS_LOAD);
> if (memory_access_type == VMAT_LOAD_STORE_LANES)
> {
> - if (is_load
> - ? !vect_load_lanes_supported (vectype, group_size, true)
> - : !vect_store_lanes_supported (vectype, group_size, true))
> + internal_fn ifn
> + = (is_load ? vect_load_lanes_supported (vectype, group_size, true)
> + : vect_store_lanes_supported (vectype, group_size, true));
> + if (ifn == IFN_MASK_LEN_LOAD_LANES || ifn == IFN_MASK_LEN_STORE_LANES)
> + vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1);
> + else if (ifn == IFN_MASK_LOAD_LANES || ifn == IFN_MASK_STORE_LANES)
> + vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
> + scalar_mask);
> + else
> {
> if (dump_enabled_p ())
> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> @@ -1620,10 +1626,7 @@ check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
> " the target doesn't have an appropriate"
> " load/store-lanes instruction.\n");
> LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
> - return;
> }
> - vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
> - scalar_mask);
> return;
> }
>
> @@ -2074,7 +2077,8 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> poly_int64 *poffset,
> dr_alignment_support *alignment_support_scheme,
> int *misalignment,
> - gather_scatter_info *gs_info)
> + gather_scatter_info *gs_info,
> + internal_fn *lanes_ifn)
> {
> loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
> class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
> @@ -2272,24 +2276,30 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> ;
>
> - /* Otherwise try using LOAD/STORE_LANES. */
> - else if (vls_type == VLS_LOAD
> - ? vect_load_lanes_supported (vectype, group_size, masked_p)
> - : vect_store_lanes_supported (vectype, group_size,
> - masked_p))
> + else
> {
> - *memory_access_type = VMAT_LOAD_STORE_LANES;
> - overrun_p = would_overrun_p;
> - }
> + /* Otherwise try using LOAD/STORE_LANES. */
> + *lanes_ifn
> + = vls_type == VLS_LOAD
> + ? vect_load_lanes_supported (vectype, group_size, masked_p)
> + : vect_store_lanes_supported (vectype, group_size,
> + masked_p);
> + if (*lanes_ifn != IFN_LAST)
> + {
> + *memory_access_type = VMAT_LOAD_STORE_LANES;
> + overrun_p = would_overrun_p;
> + }
>
> - /* If that fails, try using permuting loads. */
> - else if (vls_type == VLS_LOAD
> - ? vect_grouped_load_supported (vectype, single_element_p,
> - group_size)
> - : vect_grouped_store_supported (vectype, group_size))
> - {
> - *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> - overrun_p = would_overrun_p;
> + /* If that fails, try using permuting loads. */
> + else if (vls_type == VLS_LOAD
> + ? vect_grouped_load_supported (vectype,
> + single_element_p,
> + group_size)
> + : vect_grouped_store_supported (vectype, group_size))
> + {
> + *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> + overrun_p = would_overrun_p;
> + }
> }
> }
>
> @@ -2378,7 +2388,8 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> poly_int64 *poffset,
> dr_alignment_support *alignment_support_scheme,
> int *misalignment,
> - gather_scatter_info *gs_info)
> + gather_scatter_info *gs_info,
> + internal_fn *lanes_ifn)
> {
> loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
> poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> @@ -2441,7 +2452,7 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
> masked_p,
> vls_type, memory_access_type, poffset,
> alignment_support_scheme,
> - misalignment, gs_info))
> + misalignment, gs_info, lanes_ifn))
> return false;
> }
> else if (STMT_VINFO_STRIDED_P (stmt_info))
> @@ -3087,11 +3098,8 @@ vect_get_loop_variant_data_ptr_increment (
> loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> tree step = vect_dr_behavior (vinfo, dr_info)->step;
>
> - /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> - IVs are updated by variable amount but we will support them in the future.
> - */
> - gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> - && memory_access_type != VMAT_LOAD_STORE_LANES);
> + /* gather/scatter never reach here. */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER);
>
> /* When we support SELECT_VL pattern, we dynamic adjust
> the memory address by .SELECT_VL result.
> @@ -8094,9 +8102,11 @@ vectorizable_store (vec_info *vinfo,
> enum dr_alignment_support alignment_support_scheme;
> int misalignment;
> poly_int64 poffset;
> + internal_fn lanes_ifn;
> if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask, vls_type,
> ncopies, &memory_access_type, &poffset,
> - &alignment_support_scheme, &misalignment, &gs_info))
> + &alignment_support_scheme, &misalignment, &gs_info,
> + &lanes_ifn))
> return false;
>
> if (mask)
> @@ -8885,6 +8895,8 @@ vectorizable_store (vec_info *vinfo,
> }
>
> tree final_mask = NULL;
> + tree final_len = NULL;
> + tree bias = NULL;
> if (loop_masks)
> final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> ncopies, vectype, j);
> @@ -8892,8 +8904,37 @@ vectorizable_store (vec_info *vinfo,
> final_mask = prepare_vec_mask (loop_vinfo, mask_vectype,
> final_mask, vec_mask, gsi);
>
> + if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
> + {
> + if (loop_lens)
> + final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + ncopies, vectype, j, 1);
> + else
> + final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
> + signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + if (!final_mask)
> + {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> + }
> + }
> +
> gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> + {
> + /* Emit:
> + MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
> + LEN, BIAS, VEC_ARRAY). */
> + unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
> + tree alias_ptr = build_int_cst (ref_type, align);
> + call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
> + dataref_ptr, alias_ptr,
> + final_mask, final_len, bias,
> + vec_array);
> + }
> + else if (final_mask)
> {
> /* Emit:
> MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
> @@ -9598,9 +9639,11 @@ vectorizable_load (vec_info *vinfo,
> enum dr_alignment_support alignment_support_scheme;
> int misalignment;
> poly_int64 poffset;
> + internal_fn lanes_ifn;
> if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask, VLS_LOAD,
> ncopies, &memory_access_type, &poffset,
> - &alignment_support_scheme, &misalignment, &gs_info))
> + &alignment_support_scheme, &misalignment, &gs_info,
> + &lanes_ifn))
> return false;
>
> if (mask)
> @@ -10386,6 +10429,8 @@ vectorizable_load (vec_info *vinfo,
> tree vec_array = create_vector_array (vectype, vec_num);
>
> tree final_mask = NULL_TREE;
> + tree final_len = NULL_TREE;
> + tree bias = NULL_TREE;
> if (loop_masks)
> final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> ncopies, vectype, j);
> @@ -10393,8 +10438,36 @@ vectorizable_load (vec_info *vinfo,
> final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
> vec_mask, gsi);
>
> + if (lanes_ifn == IFN_MASK_LEN_LOAD_LANES)
> + {
> + if (loop_lens)
> + final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + ncopies, vectype, j, 1);
> + else
> + final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
> + signed char biasval
> + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
> + bias = build_int_cst (intQI_type_node, biasval);
> + if (!final_mask)
> + {
> + mask_vectype = truth_type_for (vectype);
> + final_mask = build_minus_one_cst (mask_vectype);
> + }
> + }
> +
> gcall *call;
> - if (final_mask)
> + if (final_len && final_mask)
> + {
> + /* Emit:
> + VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
> + VEC_MASK, LEN, BIAS). */
> + unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
> + tree alias_ptr = build_int_cst (ref_type, align);
> + call = gimple_build_call_internal (IFN_MASK_LEN_LOAD_LANES, 5,
> + dataref_ptr, alias_ptr,
> + final_mask, final_len, bias);
> + }
> + else if (final_mask)
> {
> /* Emit:
> VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 1de144988c8..53a3d78d545 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2297,9 +2297,9 @@ extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *,
> extern void vect_copy_ref_info (tree, tree);
> extern tree vect_create_destination_var (tree, tree);
> extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
> -extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> +extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
> -extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> +extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
> extern void vect_permute_store_chain (vec_info *, vec<tree> &,
> unsigned int, stmt_vec_info,
> gimple_stmt_iterator *, vec<tree> *);
>
--
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2023-08-16 6:06 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-08-15 12:29 [PATCH V2] VECT: Apply MASK_LEN_{LOAD_LANES, STORE_LANES} into vectorizer Juzhe-Zhong
2023-08-15 12:35 ` Richard Biener
2023-08-16 6:06 ` Li, Pan2
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).