* Re: [PATCH] autopar TLC
@ 2022-08-09 11:36 Richard Biener
0 siblings, 0 replies; 4+ messages in thread
From: Richard Biener @ 2022-08-09 11:36 UTC (permalink / raw)
To: gcc-patches; +Cc: Jakub Jelinek
On Tue, 2 Aug 2022, Richard Biener wrote:
> The following removes all excessive update_ssa calls from OMP
> expansion, thereby rewriting the atomic load and store cases to
> GIMPLE code generation. I don't think autopar ever exercises the
> atomics code though.
>
> There's not much test coverage overall so I've built SPEC 2k17
> with -floop-parallelize-all -ftree-parallelize-loops=2 with and
> without LTO (and otherwise -Ofast plus -march=haswell) without
> fallout.
>
> If there's any fallout it's not OK to update SSA form for
> each and every OMP stmt lowered.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu.
>
> Any objections?
I have pushed this now (and will deal with eventual fallout).
Richard.
> Thanks,
> Richard.
>
> * omp-expand.cc (expand_omp_atomic_load): Emit GIMPLE
> directly. Avoid update_ssa when in SSA form.
> (expand_omp_atomic_store): Likewise.
> (expand_omp_atomic_fetch_op): Avoid update_ssa when in SSA
> form.
> (expand_omp_atomic_pipeline): Likewise.
> (expand_omp_atomic_mutex): Likewise.
> * tree-parloops.cc (gen_parallel_loop): Use
> TODO_update_ssa_no_phi after loop_version.
> ---
> gcc/omp-expand.cc | 81 +++++++++++++++++++++++++++-----------------
> gcc/tree-parloops.cc | 2 +-
> 2 files changed, 50 insertions(+), 33 deletions(-)
>
> diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
> index 64e6308fc7b..48fbd157c6e 100644
> --- a/gcc/omp-expand.cc
> +++ b/gcc/omp-expand.cc
> @@ -8617,7 +8617,7 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
> basic_block store_bb;
> location_t loc;
> gimple *stmt;
> - tree decl, call, type, itype;
> + tree decl, type, itype;
>
> gsi = gsi_last_nondebug_bb (load_bb);
> stmt = gsi_stmt (gsi);
> @@ -8637,23 +8637,33 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
> itype = TREE_TYPE (TREE_TYPE (decl));
>
> enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
> - tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
> - call = build_call_expr_loc (loc, decl, 2, addr, mo);
> + tree mo = build_int_cst (integer_type_node,
> + omp_memory_order_to_memmodel (omo));
> + gcall *call = gimple_build_call (decl, 2, addr, mo);
> + gimple_set_location (call, loc);
> + gimple_set_vuse (call, gimple_vuse (stmt));
> + gimple *repl;
> if (!useless_type_conversion_p (type, itype))
> - call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
> - call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
> -
> - force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
> - gsi_remove (&gsi, true);
> + {
> + tree lhs = make_ssa_name (itype);
> + gimple_call_set_lhs (call, lhs);
> + gsi_insert_before (&gsi, call, GSI_SAME_STMT);
> + repl = gimple_build_assign (loaded_val,
> + build1 (VIEW_CONVERT_EXPR, type, lhs));
> + gimple_set_location (repl, loc);
> + }
> + else
> + {
> + gimple_call_set_lhs (call, loaded_val);
> + repl = call;
> + }
> + gsi_replace (&gsi, repl, true);
>
> store_bb = single_succ (load_bb);
> gsi = gsi_last_nondebug_bb (store_bb);
> gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE);
> gsi_remove (&gsi, true);
>
> - if (gimple_in_ssa_p (cfun))
> - update_ssa (TODO_update_ssa_no_phi);
> -
> return true;
> }
>
> @@ -8669,7 +8679,7 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
> basic_block store_bb = single_succ (load_bb);
> location_t loc;
> gimple *stmt;
> - tree decl, call, type, itype;
> + tree decl, type, itype;
> machine_mode imode;
> bool exchange;
>
> @@ -8710,25 +8720,36 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
> if (!useless_type_conversion_p (itype, type))
> stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val);
> enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
> - tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
> - call = build_call_expr_loc (loc, decl, 3, addr, stored_val, mo);
> + tree mo = build_int_cst (integer_type_node,
> + omp_memory_order_to_memmodel (omo));
> + stored_val = force_gimple_operand_gsi (&gsi, stored_val, true, NULL_TREE,
> + true, GSI_SAME_STMT);
> + gcall *call = gimple_build_call (decl, 3, addr, stored_val, mo);
> + gimple_set_location (call, loc);
> + gimple_set_vuse (call, gimple_vuse (stmt));
> + gimple_set_vdef (call, gimple_vdef (stmt));
> +
> + gimple *repl = call;
> if (exchange)
> {
> if (!useless_type_conversion_p (type, itype))
> - call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
> - call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
> + {
> + tree lhs = make_ssa_name (itype);
> + gimple_call_set_lhs (call, lhs);
> + gsi_insert_before (&gsi, call, GSI_SAME_STMT);
> + repl = gimple_build_assign (loaded_val,
> + build1 (VIEW_CONVERT_EXPR, type, lhs));
> + gimple_set_location (repl, loc);
> + }
> + else
> + gimple_call_set_lhs (call, loaded_val);
> }
> -
> - force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
> - gsi_remove (&gsi, true);
> + gsi_replace (&gsi, repl, true);
>
> /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */
> gsi = gsi_last_nondebug_bb (load_bb);
> gsi_remove (&gsi, true);
>
> - if (gimple_in_ssa_p (cfun))
> - update_ssa (TODO_update_ssa_no_phi);
> -
> return true;
> }
>
> @@ -8874,10 +8895,7 @@ expand_omp_atomic_fetch_op (basic_block load_bb,
> gsi_remove (&gsi, true);
>
> if (gimple_in_ssa_p (cfun))
> - {
> - release_defs (stmt);
> - update_ssa (TODO_update_ssa_no_phi);
> - }
> + release_defs (stmt);
>
> return true;
> }
> @@ -9333,16 +9351,16 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
> }
>
> /* Remove GIMPLE_OMP_ATOMIC_STORE. */
> + stmt = gsi_stmt (si);
> gsi_remove (&si, true);
> + if (gimple_in_ssa_p (cfun))
> + release_defs (stmt);
>
> class loop *loop = alloc_loop ();
> loop->header = loop_header;
> loop->latch = store_bb;
> add_loop (loop, loop_header->loop_father);
>
> - if (gimple_in_ssa_p (cfun))
> - update_ssa (TODO_update_ssa_no_phi);
> -
> return true;
> }
>
> @@ -9399,15 +9417,14 @@ expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb,
> gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE);
>
> stmt = gimple_build_assign (unshare_expr (mem), stored_val);
> + gimple_set_vuse (stmt, gimple_vuse (gsi_stmt (si)));
> + gimple_set_vdef (stmt, gimple_vdef (gsi_stmt (si)));
> gsi_insert_before (&si, stmt, GSI_SAME_STMT);
>
> t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END);
> t = build_call_expr (t, 0);
> force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT);
> gsi_remove (&si, true);
> -
> - if (gimple_in_ssa_p (cfun))
> - update_ssa (TODO_update_ssa_no_phi);
> return true;
> }
>
> diff --git a/gcc/tree-parloops.cc b/gcc/tree-parloops.cc
> index 2d3aa78cd24..b070527ee6e 100644
> --- a/gcc/tree-parloops.cc
> +++ b/gcc/tree-parloops.cc
> @@ -3082,7 +3082,7 @@ gen_parallel_loop (class loop *loop,
> profile_probability::unlikely (),
> profile_probability::likely (),
> profile_probability::unlikely (), true);
> - update_ssa (TODO_update_ssa);
> + update_ssa (TODO_update_ssa_no_phi);
> free_original_copy_tables ();
> }
>
>
--
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
^ permalink raw reply [flat|nested] 4+ messages in thread
* [PATCH] autopar TLC
@ 2022-08-02 11:51 Richard Biener
0 siblings, 0 replies; 4+ messages in thread
From: Richard Biener @ 2022-08-02 11:51 UTC (permalink / raw)
To: gcc-patches; +Cc: Jakub Jelinek
The following removes all excessive update_ssa calls from OMP
expansion, thereby rewriting the atomic load and store cases to
GIMPLE code generation. I don't think autopar ever exercises the
atomics code though.
There's not much test coverage overall so I've built SPEC 2k17
with -floop-parallelize-all -ftree-parallelize-loops=2 with and
without LTO (and otherwise -Ofast plus -march=haswell) without
fallout.
If there's any fallout it's not OK to update SSA form for
each and every OMP stmt lowered.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Any objections?
Thanks,
Richard.
* omp-expand.cc (expand_omp_atomic_load): Emit GIMPLE
directly. Avoid update_ssa when in SSA form.
(expand_omp_atomic_store): Likewise.
(expand_omp_atomic_fetch_op): Avoid update_ssa when in SSA
form.
(expand_omp_atomic_pipeline): Likewise.
(expand_omp_atomic_mutex): Likewise.
* tree-parloops.cc (gen_parallel_loop): Use
TODO_update_ssa_no_phi after loop_version.
---
gcc/omp-expand.cc | 81 +++++++++++++++++++++++++++-----------------
gcc/tree-parloops.cc | 2 +-
2 files changed, 50 insertions(+), 33 deletions(-)
diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
index 64e6308fc7b..48fbd157c6e 100644
--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -8617,7 +8617,7 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
basic_block store_bb;
location_t loc;
gimple *stmt;
- tree decl, call, type, itype;
+ tree decl, type, itype;
gsi = gsi_last_nondebug_bb (load_bb);
stmt = gsi_stmt (gsi);
@@ -8637,23 +8637,33 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
itype = TREE_TYPE (TREE_TYPE (decl));
enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
- tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
- call = build_call_expr_loc (loc, decl, 2, addr, mo);
+ tree mo = build_int_cst (integer_type_node,
+ omp_memory_order_to_memmodel (omo));
+ gcall *call = gimple_build_call (decl, 2, addr, mo);
+ gimple_set_location (call, loc);
+ gimple_set_vuse (call, gimple_vuse (stmt));
+ gimple *repl;
if (!useless_type_conversion_p (type, itype))
- call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
- call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
-
- force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
- gsi_remove (&gsi, true);
+ {
+ tree lhs = make_ssa_name (itype);
+ gimple_call_set_lhs (call, lhs);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ repl = gimple_build_assign (loaded_val,
+ build1 (VIEW_CONVERT_EXPR, type, lhs));
+ gimple_set_location (repl, loc);
+ }
+ else
+ {
+ gimple_call_set_lhs (call, loaded_val);
+ repl = call;
+ }
+ gsi_replace (&gsi, repl, true);
store_bb = single_succ (load_bb);
gsi = gsi_last_nondebug_bb (store_bb);
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE);
gsi_remove (&gsi, true);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -8669,7 +8679,7 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
basic_block store_bb = single_succ (load_bb);
location_t loc;
gimple *stmt;
- tree decl, call, type, itype;
+ tree decl, type, itype;
machine_mode imode;
bool exchange;
@@ -8710,25 +8720,36 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
if (!useless_type_conversion_p (itype, type))
stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val);
enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
- tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
- call = build_call_expr_loc (loc, decl, 3, addr, stored_val, mo);
+ tree mo = build_int_cst (integer_type_node,
+ omp_memory_order_to_memmodel (omo));
+ stored_val = force_gimple_operand_gsi (&gsi, stored_val, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ gcall *call = gimple_build_call (decl, 3, addr, stored_val, mo);
+ gimple_set_location (call, loc);
+ gimple_set_vuse (call, gimple_vuse (stmt));
+ gimple_set_vdef (call, gimple_vdef (stmt));
+
+ gimple *repl = call;
if (exchange)
{
if (!useless_type_conversion_p (type, itype))
- call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
- call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
+ {
+ tree lhs = make_ssa_name (itype);
+ gimple_call_set_lhs (call, lhs);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ repl = gimple_build_assign (loaded_val,
+ build1 (VIEW_CONVERT_EXPR, type, lhs));
+ gimple_set_location (repl, loc);
+ }
+ else
+ gimple_call_set_lhs (call, loaded_val);
}
-
- force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
- gsi_remove (&gsi, true);
+ gsi_replace (&gsi, repl, true);
/* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */
gsi = gsi_last_nondebug_bb (load_bb);
gsi_remove (&gsi, true);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -8874,10 +8895,7 @@ expand_omp_atomic_fetch_op (basic_block load_bb,
gsi_remove (&gsi, true);
if (gimple_in_ssa_p (cfun))
- {
- release_defs (stmt);
- update_ssa (TODO_update_ssa_no_phi);
- }
+ release_defs (stmt);
return true;
}
@@ -9333,16 +9351,16 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
}
/* Remove GIMPLE_OMP_ATOMIC_STORE. */
+ stmt = gsi_stmt (si);
gsi_remove (&si, true);
+ if (gimple_in_ssa_p (cfun))
+ release_defs (stmt);
class loop *loop = alloc_loop ();
loop->header = loop_header;
loop->latch = store_bb;
add_loop (loop, loop_header->loop_father);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -9399,15 +9417,14 @@ expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb,
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE);
stmt = gimple_build_assign (unshare_expr (mem), stored_val);
+ gimple_set_vuse (stmt, gimple_vuse (gsi_stmt (si)));
+ gimple_set_vdef (stmt, gimple_vdef (gsi_stmt (si)));
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END);
t = build_call_expr (t, 0);
force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT);
gsi_remove (&si, true);
-
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
return true;
}
diff --git a/gcc/tree-parloops.cc b/gcc/tree-parloops.cc
index 2d3aa78cd24..b070527ee6e 100644
--- a/gcc/tree-parloops.cc
+++ b/gcc/tree-parloops.cc
@@ -3082,7 +3082,7 @@ gen_parallel_loop (class loop *loop,
profile_probability::unlikely (),
profile_probability::likely (),
profile_probability::unlikely (), true);
- update_ssa (TODO_update_ssa);
+ update_ssa (TODO_update_ssa_no_phi);
free_original_copy_tables ();
}
--
2.35.3
^ permalink raw reply [flat|nested] 4+ messages in thread
* [PATCH] autopar TLC
@ 2022-08-02 11:51 Richard Biener
0 siblings, 0 replies; 4+ messages in thread
From: Richard Biener @ 2022-08-02 11:51 UTC (permalink / raw)
To: gcc-patches; +Cc: Jakub Jelinek
The following removes all excessive update_ssa calls from OMP
expansion, thereby rewriting the atomic load and store cases to
GIMPLE code generation. I don't think autopar ever exercises the
atomics code though.
There's not much test coverage overall so I've built SPEC 2k17
with -floop-parallelize-all -ftree-parallelize-loops=2 with and
without LTO (and otherwise -Ofast plus -march=haswell) without
fallout.
If there's any fallout it's not OK to update SSA form for
each and every OMP stmt lowered.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Any objections?
Thanks,
Richard.
* omp-expand.cc (expand_omp_atomic_load): Emit GIMPLE
directly. Avoid update_ssa when in SSA form.
(expand_omp_atomic_store): Likewise.
(expand_omp_atomic_fetch_op): Avoid update_ssa when in SSA
form.
(expand_omp_atomic_pipeline): Likewise.
(expand_omp_atomic_mutex): Likewise.
* tree-parloops.cc (gen_parallel_loop): Use
TODO_update_ssa_no_phi after loop_version.
---
gcc/omp-expand.cc | 81 +++++++++++++++++++++++++++-----------------
gcc/tree-parloops.cc | 2 +-
2 files changed, 50 insertions(+), 33 deletions(-)
diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
index 64e6308fc7b..48fbd157c6e 100644
--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -8617,7 +8617,7 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
basic_block store_bb;
location_t loc;
gimple *stmt;
- tree decl, call, type, itype;
+ tree decl, type, itype;
gsi = gsi_last_nondebug_bb (load_bb);
stmt = gsi_stmt (gsi);
@@ -8637,23 +8637,33 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
itype = TREE_TYPE (TREE_TYPE (decl));
enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
- tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
- call = build_call_expr_loc (loc, decl, 2, addr, mo);
+ tree mo = build_int_cst (integer_type_node,
+ omp_memory_order_to_memmodel (omo));
+ gcall *call = gimple_build_call (decl, 2, addr, mo);
+ gimple_set_location (call, loc);
+ gimple_set_vuse (call, gimple_vuse (stmt));
+ gimple *repl;
if (!useless_type_conversion_p (type, itype))
- call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
- call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
-
- force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
- gsi_remove (&gsi, true);
+ {
+ tree lhs = make_ssa_name (itype);
+ gimple_call_set_lhs (call, lhs);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ repl = gimple_build_assign (loaded_val,
+ build1 (VIEW_CONVERT_EXPR, type, lhs));
+ gimple_set_location (repl, loc);
+ }
+ else
+ {
+ gimple_call_set_lhs (call, loaded_val);
+ repl = call;
+ }
+ gsi_replace (&gsi, repl, true);
store_bb = single_succ (load_bb);
gsi = gsi_last_nondebug_bb (store_bb);
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE);
gsi_remove (&gsi, true);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -8669,7 +8679,7 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
basic_block store_bb = single_succ (load_bb);
location_t loc;
gimple *stmt;
- tree decl, call, type, itype;
+ tree decl, type, itype;
machine_mode imode;
bool exchange;
@@ -8710,25 +8720,36 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
if (!useless_type_conversion_p (itype, type))
stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val);
enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
- tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
- call = build_call_expr_loc (loc, decl, 3, addr, stored_val, mo);
+ tree mo = build_int_cst (integer_type_node,
+ omp_memory_order_to_memmodel (omo));
+ stored_val = force_gimple_operand_gsi (&gsi, stored_val, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ gcall *call = gimple_build_call (decl, 3, addr, stored_val, mo);
+ gimple_set_location (call, loc);
+ gimple_set_vuse (call, gimple_vuse (stmt));
+ gimple_set_vdef (call, gimple_vdef (stmt));
+
+ gimple *repl = call;
if (exchange)
{
if (!useless_type_conversion_p (type, itype))
- call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
- call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
+ {
+ tree lhs = make_ssa_name (itype);
+ gimple_call_set_lhs (call, lhs);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ repl = gimple_build_assign (loaded_val,
+ build1 (VIEW_CONVERT_EXPR, type, lhs));
+ gimple_set_location (repl, loc);
+ }
+ else
+ gimple_call_set_lhs (call, loaded_val);
}
-
- force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
- gsi_remove (&gsi, true);
+ gsi_replace (&gsi, repl, true);
/* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */
gsi = gsi_last_nondebug_bb (load_bb);
gsi_remove (&gsi, true);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -8874,10 +8895,7 @@ expand_omp_atomic_fetch_op (basic_block load_bb,
gsi_remove (&gsi, true);
if (gimple_in_ssa_p (cfun))
- {
- release_defs (stmt);
- update_ssa (TODO_update_ssa_no_phi);
- }
+ release_defs (stmt);
return true;
}
@@ -9333,16 +9351,16 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
}
/* Remove GIMPLE_OMP_ATOMIC_STORE. */
+ stmt = gsi_stmt (si);
gsi_remove (&si, true);
+ if (gimple_in_ssa_p (cfun))
+ release_defs (stmt);
class loop *loop = alloc_loop ();
loop->header = loop_header;
loop->latch = store_bb;
add_loop (loop, loop_header->loop_father);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -9399,15 +9417,14 @@ expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb,
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE);
stmt = gimple_build_assign (unshare_expr (mem), stored_val);
+ gimple_set_vuse (stmt, gimple_vuse (gsi_stmt (si)));
+ gimple_set_vdef (stmt, gimple_vdef (gsi_stmt (si)));
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END);
t = build_call_expr (t, 0);
force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT);
gsi_remove (&si, true);
-
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
return true;
}
diff --git a/gcc/tree-parloops.cc b/gcc/tree-parloops.cc
index 2d3aa78cd24..b070527ee6e 100644
--- a/gcc/tree-parloops.cc
+++ b/gcc/tree-parloops.cc
@@ -3082,7 +3082,7 @@ gen_parallel_loop (class loop *loop,
profile_probability::unlikely (),
profile_probability::likely (),
profile_probability::unlikely (), true);
- update_ssa (TODO_update_ssa);
+ update_ssa (TODO_update_ssa_no_phi);
free_original_copy_tables ();
}
--
2.35.3
^ permalink raw reply [flat|nested] 4+ messages in thread
* [PATCH] autopar TLC
@ 2022-08-02 11:51 Richard Biener
0 siblings, 0 replies; 4+ messages in thread
From: Richard Biener @ 2022-08-02 11:51 UTC (permalink / raw)
To: gcc-patches; +Cc: Jakub Jelinek
The following removes all excessive update_ssa calls from OMP
expansion, thereby rewriting the atomic load and store cases to
GIMPLE code generation. I don't think autopar ever exercises the
atomics code though.
There's not much test coverage overall so I've built SPEC 2k17
with -floop-parallelize-all -ftree-parallelize-loops=2 with and
without LTO (and otherwise -Ofast plus -march=haswell) without
fallout.
If there's any fallout it's not OK to update SSA form for
each and every OMP stmt lowered.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Any objections?
Thanks,
Richard.
* omp-expand.cc (expand_omp_atomic_load): Emit GIMPLE
directly. Avoid update_ssa when in SSA form.
(expand_omp_atomic_store): Likewise.
(expand_omp_atomic_fetch_op): Avoid update_ssa when in SSA
form.
(expand_omp_atomic_pipeline): Likewise.
(expand_omp_atomic_mutex): Likewise.
* tree-parloops.cc (gen_parallel_loop): Use
TODO_update_ssa_no_phi after loop_version.
---
gcc/omp-expand.cc | 81 +++++++++++++++++++++++++++-----------------
gcc/tree-parloops.cc | 2 +-
2 files changed, 50 insertions(+), 33 deletions(-)
diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
index 64e6308fc7b..48fbd157c6e 100644
--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -8617,7 +8617,7 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
basic_block store_bb;
location_t loc;
gimple *stmt;
- tree decl, call, type, itype;
+ tree decl, type, itype;
gsi = gsi_last_nondebug_bb (load_bb);
stmt = gsi_stmt (gsi);
@@ -8637,23 +8637,33 @@ expand_omp_atomic_load (basic_block load_bb, tree addr,
itype = TREE_TYPE (TREE_TYPE (decl));
enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
- tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
- call = build_call_expr_loc (loc, decl, 2, addr, mo);
+ tree mo = build_int_cst (integer_type_node,
+ omp_memory_order_to_memmodel (omo));
+ gcall *call = gimple_build_call (decl, 2, addr, mo);
+ gimple_set_location (call, loc);
+ gimple_set_vuse (call, gimple_vuse (stmt));
+ gimple *repl;
if (!useless_type_conversion_p (type, itype))
- call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
- call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
-
- force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
- gsi_remove (&gsi, true);
+ {
+ tree lhs = make_ssa_name (itype);
+ gimple_call_set_lhs (call, lhs);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ repl = gimple_build_assign (loaded_val,
+ build1 (VIEW_CONVERT_EXPR, type, lhs));
+ gimple_set_location (repl, loc);
+ }
+ else
+ {
+ gimple_call_set_lhs (call, loaded_val);
+ repl = call;
+ }
+ gsi_replace (&gsi, repl, true);
store_bb = single_succ (load_bb);
gsi = gsi_last_nondebug_bb (store_bb);
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE);
gsi_remove (&gsi, true);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -8669,7 +8679,7 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
basic_block store_bb = single_succ (load_bb);
location_t loc;
gimple *stmt;
- tree decl, call, type, itype;
+ tree decl, type, itype;
machine_mode imode;
bool exchange;
@@ -8710,25 +8720,36 @@ expand_omp_atomic_store (basic_block load_bb, tree addr,
if (!useless_type_conversion_p (itype, type))
stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val);
enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt);
- tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo));
- call = build_call_expr_loc (loc, decl, 3, addr, stored_val, mo);
+ tree mo = build_int_cst (integer_type_node,
+ omp_memory_order_to_memmodel (omo));
+ stored_val = force_gimple_operand_gsi (&gsi, stored_val, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ gcall *call = gimple_build_call (decl, 3, addr, stored_val, mo);
+ gimple_set_location (call, loc);
+ gimple_set_vuse (call, gimple_vuse (stmt));
+ gimple_set_vdef (call, gimple_vdef (stmt));
+
+ gimple *repl = call;
if (exchange)
{
if (!useless_type_conversion_p (type, itype))
- call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
- call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
+ {
+ tree lhs = make_ssa_name (itype);
+ gimple_call_set_lhs (call, lhs);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ repl = gimple_build_assign (loaded_val,
+ build1 (VIEW_CONVERT_EXPR, type, lhs));
+ gimple_set_location (repl, loc);
+ }
+ else
+ gimple_call_set_lhs (call, loaded_val);
}
-
- force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
- gsi_remove (&gsi, true);
+ gsi_replace (&gsi, repl, true);
/* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */
gsi = gsi_last_nondebug_bb (load_bb);
gsi_remove (&gsi, true);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -8874,10 +8895,7 @@ expand_omp_atomic_fetch_op (basic_block load_bb,
gsi_remove (&gsi, true);
if (gimple_in_ssa_p (cfun))
- {
- release_defs (stmt);
- update_ssa (TODO_update_ssa_no_phi);
- }
+ release_defs (stmt);
return true;
}
@@ -9333,16 +9351,16 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
}
/* Remove GIMPLE_OMP_ATOMIC_STORE. */
+ stmt = gsi_stmt (si);
gsi_remove (&si, true);
+ if (gimple_in_ssa_p (cfun))
+ release_defs (stmt);
class loop *loop = alloc_loop ();
loop->header = loop_header;
loop->latch = store_bb;
add_loop (loop, loop_header->loop_father);
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
-
return true;
}
@@ -9399,15 +9417,14 @@ expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb,
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE);
stmt = gimple_build_assign (unshare_expr (mem), stored_val);
+ gimple_set_vuse (stmt, gimple_vuse (gsi_stmt (si)));
+ gimple_set_vdef (stmt, gimple_vdef (gsi_stmt (si)));
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END);
t = build_call_expr (t, 0);
force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT);
gsi_remove (&si, true);
-
- if (gimple_in_ssa_p (cfun))
- update_ssa (TODO_update_ssa_no_phi);
return true;
}
diff --git a/gcc/tree-parloops.cc b/gcc/tree-parloops.cc
index 2d3aa78cd24..b070527ee6e 100644
--- a/gcc/tree-parloops.cc
+++ b/gcc/tree-parloops.cc
@@ -3082,7 +3082,7 @@ gen_parallel_loop (class loop *loop,
profile_probability::unlikely (),
profile_probability::likely (),
profile_probability::unlikely (), true);
- update_ssa (TODO_update_ssa);
+ update_ssa (TODO_update_ssa_no_phi);
free_original_copy_tables ();
}
--
2.35.3
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2022-08-09 11:36 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-08-09 11:36 [PATCH] autopar TLC Richard Biener
-- strict thread matches above, loose matches on Subject: below --
2022-08-02 11:51 Richard Biener
2022-08-02 11:51 Richard Biener
2022-08-02 11:51 Richard Biener
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).