* [PATCH] More efficiently canoncialize operands in fold_stmt
@ 2011-07-14 13:37 Richard Guenther
0 siblings, 0 replies; only message in thread
From: Richard Guenther @ 2011-07-14 13:37 UTC (permalink / raw)
To: gcc-patches
We currently go through a tree expression generation when canonicalizing
operand order. That's quite wasteful. The following patch re-orders
things in a way to avoid this and remove some duplicate code.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Richard.
2011-07-14 Richard Guenther <rguenther@suse.de>
* gimple-fold.c (fold_gimple_assign): Remove operand swapping.
(fold_stmt_1): Do it here directly on gimple and as a first thing.
Index: gcc/gimple-fold.c
===================================================================
*** gcc/gimple-fold.c (revision 176267)
--- gcc/gimple-fold.c (working copy)
*************** fold_gimple_assign (gimple_stmt_iterator
*** 817,842 ****
if (!result)
result = fold_binary_loc (loc, subcode,
! TREE_TYPE (gimple_assign_lhs (stmt)),
! gimple_assign_rhs1 (stmt),
! gimple_assign_rhs2 (stmt));
if (result)
{
STRIP_USELESS_TYPE_CONVERSION (result);
if (valid_gimple_rhs_p (result))
return result;
-
- /* Fold might have produced non-GIMPLE, so if we trust it blindly
- we lose canonicalization opportunities. Do not go again
- through fold here though, or the same non-GIMPLE will be
- produced. */
- if (commutative_tree_code (subcode)
- && tree_swap_operands_p (gimple_assign_rhs1 (stmt),
- gimple_assign_rhs2 (stmt), false))
- return build2 (subcode, TREE_TYPE (gimple_assign_lhs (stmt)),
- gimple_assign_rhs2 (stmt),
- gimple_assign_rhs1 (stmt));
}
break;
--- 817,831 ----
if (!result)
result = fold_binary_loc (loc, subcode,
! TREE_TYPE (gimple_assign_lhs (stmt)),
! gimple_assign_rhs1 (stmt),
! gimple_assign_rhs2 (stmt));
if (result)
{
STRIP_USELESS_TYPE_CONVERSION (result);
if (valid_gimple_rhs_p (result))
return result;
}
break;
*************** fold_gimple_assign (gimple_stmt_iterator
*** 852,869 ****
STRIP_USELESS_TYPE_CONVERSION (result);
if (valid_gimple_rhs_p (result))
return result;
-
- /* Fold might have produced non-GIMPLE, so if we trust it blindly
- we lose canonicalization opportunities. Do not go again
- through fold here though, or the same non-GIMPLE will be
- produced. */
- if (commutative_ternary_tree_code (subcode)
- && tree_swap_operands_p (gimple_assign_rhs1 (stmt),
- gimple_assign_rhs2 (stmt), false))
- return build3 (subcode, TREE_TYPE (gimple_assign_lhs (stmt)),
- gimple_assign_rhs2 (stmt),
- gimple_assign_rhs1 (stmt),
- gimple_assign_rhs3 (stmt));
}
break;
--- 841,846 ----
*************** fold_stmt_1 (gimple_stmt_iterator *gsi,
*** 1576,1583 ****
case GIMPLE_ASSIGN:
{
unsigned old_num_ops = gimple_num_ops (stmt);
! tree new_rhs = fold_gimple_assign (gsi);
tree lhs = gimple_assign_lhs (stmt);
if (new_rhs
&& !useless_type_conversion_p (TREE_TYPE (lhs),
TREE_TYPE (new_rhs)))
--- 1553,1574 ----
case GIMPLE_ASSIGN:
{
unsigned old_num_ops = gimple_num_ops (stmt);
! enum tree_code subcode = gimple_assign_rhs_code (stmt);
tree lhs = gimple_assign_lhs (stmt);
+ tree new_rhs;
+ /* First canonicalize operand order. This avoids building new
+ trees if this is the only thing fold would later do. */
+ if ((commutative_tree_code (subcode)
+ || commutative_ternary_tree_code (subcode))
+ && tree_swap_operands_p (gimple_assign_rhs1 (stmt),
+ gimple_assign_rhs2 (stmt), false))
+ {
+ tree tem = gimple_assign_rhs1 (stmt);
+ gimple_assign_set_rhs1 (stmt, gimple_assign_rhs2 (stmt));
+ gimple_assign_set_rhs2 (stmt, tem);
+ changed = true;
+ }
+ new_rhs = fold_gimple_assign (gsi);
if (new_rhs
&& !useless_type_conversion_p (TREE_TYPE (lhs),
TREE_TYPE (new_rhs)))
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2011-07-14 13:29 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2011-07-14 13:37 [PATCH] More efficiently canoncialize operands in fold_stmt Richard Guenther
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).