crosstool-ng/patches/gcc/4.3.2/160-flatten-switch-stmt-00.patch
Yann E. MORIN" 664c2d0100 Add latest gcc-4.3.2 as EXPERIMENTAL, using the 4.3.1 re-based patchset.
/trunk/config/cc/gcc.in                                          |    7     7     0     0 ++
 /trunk/patches/gcc/4.3.2/280-c99-complex-ugly-hack.patch         |    8     4     4     0 +-
 /trunk/patches/gcc/4.3.2/250-sh-pr24836.patch                    |   16     8     8     0 ++--
 /trunk/patches/gcc/4.3.2/340-make-mno-spe-work-as-expected.patch |   43    19    24     0 ++++++--------
 /trunk/patches/gcc/4.3.2/130-cross-compile.patch                 |   20    10    10     0 +++---
 5 files changed, 48 insertions(+), 46 deletions(-)
2008-09-25 16:06:19 +00:00

77 lines
3.0 KiB
Diff

Original patch from gentoo: gentoo/src/patchsets/gcc/4.3.1/gentoo/18_all_904-flatten-switch-stmt-00.patch
http://gcc.gnu.org/ml/gcc-patches/2007-04/msg00927.html
Hi,
The attached patch makes sure that we create smaller object code for
simple switch statements. We just make sure to flatten the switch
statement into an if-else chain, basically.
This fixes a size-regression as compared to gcc-3.4, as can be seen
below.
2007-04-15 Bernhard Fischer <..>
* stmt.c (expand_case): Do not create a complex binary tree when
optimizing for size but rather use the simple ordered list.
(emit_case_nodes): do not emit jumps to the default_label when
optimizing for size.
Not regtested so far.
Comments?
Attached is the test switch.c mentioned below.
$ for i in 2.95 3.3 3.4 4.0 4.1 4.2.orig-HEAD 4.3.orig-HEAD 4.3-HEAD;do
gcc-$i -DCHAIN -Os -o switch-CHAIN-$i.o -c switch.c ;done
$ for i in 2.95 3.3 3.4 4.0 4.1 4.2.orig-HEAD 4.3.orig-HEAD 4.3-HEAD;do
gcc-$i -UCHAIN -Os -o switch-$i.o -c switch.c ;done
$ size switch-*.o
text data bss dec hex filename
169 0 0 169 a9 switch-2.95.o
115 0 0 115 73 switch-3.3.o
103 0 0 103 67 switch-3.4.o
124 0 0 124 7c switch-4.0.o
124 0 0 124 7c switch-4.1.o
124 0 0 124 7c switch-4.2.orig-HEAD.o
95 0 0 95 5f switch-4.3-HEAD.o
124 0 0 124 7c switch-4.3.orig-HEAD.o
166 0 0 166 a6 switch-CHAIN-2.95.o
111 0 0 111 6f switch-CHAIN-3.3.o
95 0 0 95 5f switch-CHAIN-3.4.o
95 0 0 95 5f switch-CHAIN-4.0.o
95 0 0 95 5f switch-CHAIN-4.1.o
95 0 0 95 5f switch-CHAIN-4.2.orig-HEAD.o
95 0 0 95 5f switch-CHAIN-4.3-HEAD.o
95 0 0 95 5f switch-CHAIN-4.3.orig-HEAD.o
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment; filename="gcc-4.3.gcc-flatten-switch-stmt.00.diff"
diff -durN gcc-4.3.1.orig/gcc/stmt.c gcc-4.3.1/gcc/stmt.c
--- gcc-4.3.1.orig/gcc/stmt.c 2008-05-09 20:12:13.000000000 +0200
+++ gcc-4.3.1/gcc/stmt.c 2008-06-10 14:57:58.000000000 +0200
@@ -2509,7 +2509,11 @@
use_cost_table
= (TREE_CODE (orig_type) != ENUMERAL_TYPE
&& estimate_case_costs (case_list));
- balance_case_nodes (&case_list, NULL);
+ /* When optimizing for size, we want a straight list to avoid
+ jumps as much as possible. This basically creates an if-else
+ chain. */
+ if (!optimize_size)
+ balance_case_nodes (&case_list, NULL);
emit_case_nodes (index, case_list, default_label, index_type);
emit_jump (default_label);
}
@@ -3067,6 +3071,7 @@
{
if (!node_has_low_bound (node, index_type))
{
+ if (!optimize_size) /* don't jl to the .default_label. */
emit_cmp_and_jump_insns (index,
convert_modes
(mode, imode,