* [PATCH] Early "SSA" prerequesite - make SSA def stmt update cheaper
@ 2016-01-21 13:57 Richard Biener
2016-04-19 12:14 ` Richard Biener
0 siblings, 1 reply; 2+ messages in thread
From: Richard Biener @ 2016-01-21 13:57 UTC (permalink / raw)
To: gcc-patches
This makes the SSA def stmt update during inlining cheaper by adjusting
it after remapping a SSA def instead of via an extra walk over all stmt
defs (which incidentially is not possible with FOR_EACH_SSA_* during
"early SSA" as we don't have SSA operands there).
I've tested this independently of the
[RFC] Delayed folding, match-and-simplify and early GIMPLE
patch.
This exposes that the walk_gimple_* stuff is somewhat awkward and
needs some refactoring (can't re-construct wi->gsi as gsi_for_stmt
only works for stmts in a BB and thus when we have a CFG). Need to
think about sth (simplest is require a gsi for walk_gimple_op, like
we do for walk_gimple_stmt).
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Queued for GCC 7.
Richard.
2016-01-21 Richard Biener <rguenther@suse.de>
* gimple-walk.h (struct walk_stmt_info): Add stmt member.
* gimple-walk.c (walk_gimple_op): Initialize it.
(walk_gimple_asm): Set wi->is_lhs before each callback invocation.
* tree-inline.c (remap_gimple_op_r): Set SSA_NAME_DEF_STMT when
remapping SSA names of defs.
(copy_bb): Remove walk over all SSA defs and SSA_NAME_DEF_STMT
adjustment.
Index: gcc/gimple-walk.c
===================================================================
*** gcc/gimple-walk.c (revision 232670)
--- gcc/gimple-walk.c (working copy)
*************** walk_gimple_asm (gasm *stmt, walk_tree_f
*** 100,108 ****
noutputs = gimple_asm_noutputs (stmt);
oconstraints = (const char **) alloca ((noutputs) * sizeof (const char *));
- if (wi)
- wi->is_lhs = true;
-
for (i = 0; i < noutputs; i++)
{
op = gimple_asm_output_op (stmt, i);
--- 100,105 ----
*************** walk_gimple_asm (gasm *stmt, walk_tree_f
*** 114,119 ****
--- 111,118 ----
&allows_reg, &is_inout))
wi->val_only = (allows_reg || !allows_mem);
}
+ if (wi)
+ wi->is_lhs = true;
ret = walk_tree (&TREE_VALUE (op), callback_op, wi, NULL);
if (ret)
return ret;
*************** walk_gimple_op (gimple *stmt, walk_tree_
*** 182,187 ****
--- 181,189 ----
unsigned i;
tree ret = NULL_TREE;
+ if (wi)
+ wi->stmt = stmt;
+
switch (gimple_code (stmt))
{
case GIMPLE_ASSIGN:
Index: gcc/gimple-walk.h
===================================================================
*** gcc/gimple-walk.h (revision 232670)
--- gcc/gimple-walk.h (working copy)
*************** struct walk_stmt_info
*** 28,33 ****
--- 28,34 ----
{
/* Points to the current statement being walked. */
gimple_stmt_iterator gsi;
+ gimple *stmt;
/* Additional data that the callback functions may want to carry
through the recursion. */
Index: gcc/tree-inline.c
===================================================================
*** gcc/tree-inline.c (revision 232670)
--- gcc/tree-inline.c (working copy)
*************** remap_gimple_op_r (tree *tp, int *walk_s
*** 862,871 ****
--- 862,877 ----
copy_body_data *id = (copy_body_data *) wi_p->info;
tree fn = id->src_fn;
+ /* For recursive invocations this is no longer the LHS itself. */
+ bool is_lhs = wi_p->is_lhs;
+ wi_p->is_lhs = false;
+
if (TREE_CODE (*tp) == SSA_NAME)
{
*tp = remap_ssa_name (*tp, id);
*walk_subtrees = 0;
+ if (is_lhs)
+ SSA_NAME_DEF_STMT (*tp) = wi_p->stmt;
return NULL;
}
else if (auto_var_in_fn_p (*tp, fn))
*************** copy_bb (copy_body_data *id, basic_block
*** 2089,2104 ****
maybe_duplicate_eh_stmt_fn (cfun, stmt, id->src_cfun, orig_stmt,
id->eh_map, id->eh_lp_nr);
- if (gimple_in_ssa_p (cfun) && !is_gimple_debug (stmt))
- {
- ssa_op_iter i;
- tree def;
-
- FOR_EACH_SSA_TREE_OPERAND (def, stmt, i, SSA_OP_DEF)
- if (TREE_CODE (def) == SSA_NAME)
- SSA_NAME_DEF_STMT (def) = stmt;
- }
-
gsi_next (©_gsi);
}
while (!gsi_end_p (copy_gsi));
--- 2095,2100 ----
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Early "SSA" prerequesite - make SSA def stmt update cheaper
2016-01-21 13:57 [PATCH] Early "SSA" prerequesite - make SSA def stmt update cheaper Richard Biener
@ 2016-04-19 12:14 ` Richard Biener
0 siblings, 0 replies; 2+ messages in thread
From: Richard Biener @ 2016-04-19 12:14 UTC (permalink / raw)
To: GCC Patches
On Thu, Jan 21, 2016 at 2:57 PM, Richard Biener <rguenther@suse.de> wrote:
>
> This makes the SSA def stmt update during inlining cheaper by adjusting
> it after remapping a SSA def instead of via an extra walk over all stmt
> defs (which incidentially is not possible with FOR_EACH_SSA_* during
> "early SSA" as we don't have SSA operands there).
>
> I've tested this independently of the
> [RFC] Delayed folding, match-and-simplify and early GIMPLE
> patch.
>
> This exposes that the walk_gimple_* stuff is somewhat awkward and
> needs some refactoring (can't re-construct wi->gsi as gsi_for_stmt
> only works for stmts in a BB and thus when we have a CFG). Need to
> think about sth (simplest is require a gsi for walk_gimple_op, like
> we do for walk_gimple_stmt).
>
> Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
>
> Queued for GCC 7.
Re-bootstrapped and tested on x86_64-unknown-linux-gnu, applied
to trunk as r235190.
Richard.
> Richard.
>
> 2016-01-21 Richard Biener <rguenther@suse.de>
>
> * gimple-walk.h (struct walk_stmt_info): Add stmt member.
> * gimple-walk.c (walk_gimple_op): Initialize it.
> (walk_gimple_asm): Set wi->is_lhs before each callback invocation.
> * tree-inline.c (remap_gimple_op_r): Set SSA_NAME_DEF_STMT when
> remapping SSA names of defs.
> (copy_bb): Remove walk over all SSA defs and SSA_NAME_DEF_STMT
> adjustment.
>
> Index: gcc/gimple-walk.c
> ===================================================================
> *** gcc/gimple-walk.c (revision 232670)
> --- gcc/gimple-walk.c (working copy)
> *************** walk_gimple_asm (gasm *stmt, walk_tree_f
> *** 100,108 ****
> noutputs = gimple_asm_noutputs (stmt);
> oconstraints = (const char **) alloca ((noutputs) * sizeof (const char *));
>
> - if (wi)
> - wi->is_lhs = true;
> -
> for (i = 0; i < noutputs; i++)
> {
> op = gimple_asm_output_op (stmt, i);
> --- 100,105 ----
> *************** walk_gimple_asm (gasm *stmt, walk_tree_f
> *** 114,119 ****
> --- 111,118 ----
> &allows_reg, &is_inout))
> wi->val_only = (allows_reg || !allows_mem);
> }
> + if (wi)
> + wi->is_lhs = true;
> ret = walk_tree (&TREE_VALUE (op), callback_op, wi, NULL);
> if (ret)
> return ret;
> *************** walk_gimple_op (gimple *stmt, walk_tree_
> *** 182,187 ****
> --- 181,189 ----
> unsigned i;
> tree ret = NULL_TREE;
>
> + if (wi)
> + wi->stmt = stmt;
> +
> switch (gimple_code (stmt))
> {
> case GIMPLE_ASSIGN:
> Index: gcc/gimple-walk.h
> ===================================================================
> *** gcc/gimple-walk.h (revision 232670)
> --- gcc/gimple-walk.h (working copy)
> *************** struct walk_stmt_info
> *** 28,33 ****
> --- 28,34 ----
> {
> /* Points to the current statement being walked. */
> gimple_stmt_iterator gsi;
> + gimple *stmt;
>
> /* Additional data that the callback functions may want to carry
> through the recursion. */
> Index: gcc/tree-inline.c
> ===================================================================
> *** gcc/tree-inline.c (revision 232670)
> --- gcc/tree-inline.c (working copy)
> *************** remap_gimple_op_r (tree *tp, int *walk_s
> *** 862,871 ****
> --- 862,877 ----
> copy_body_data *id = (copy_body_data *) wi_p->info;
> tree fn = id->src_fn;
>
> + /* For recursive invocations this is no longer the LHS itself. */
> + bool is_lhs = wi_p->is_lhs;
> + wi_p->is_lhs = false;
> +
> if (TREE_CODE (*tp) == SSA_NAME)
> {
> *tp = remap_ssa_name (*tp, id);
> *walk_subtrees = 0;
> + if (is_lhs)
> + SSA_NAME_DEF_STMT (*tp) = wi_p->stmt;
> return NULL;
> }
> else if (auto_var_in_fn_p (*tp, fn))
> *************** copy_bb (copy_body_data *id, basic_block
> *** 2089,2104 ****
> maybe_duplicate_eh_stmt_fn (cfun, stmt, id->src_cfun, orig_stmt,
> id->eh_map, id->eh_lp_nr);
>
> - if (gimple_in_ssa_p (cfun) && !is_gimple_debug (stmt))
> - {
> - ssa_op_iter i;
> - tree def;
> -
> - FOR_EACH_SSA_TREE_OPERAND (def, stmt, i, SSA_OP_DEF)
> - if (TREE_CODE (def) == SSA_NAME)
> - SSA_NAME_DEF_STMT (def) = stmt;
> - }
> -
> gsi_next (©_gsi);
> }
> while (!gsi_end_p (copy_gsi));
> --- 2095,2100 ----
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2016-04-19 12:14 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-01-21 13:57 [PATCH] Early "SSA" prerequesite - make SSA def stmt update cheaper Richard Biener
2016-04-19 12:14 ` 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).