1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
|
Upstream-Status: Pending
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"
Index: gcc-4.6.0/gcc/stmt.c
===================================================================
--- gcc-4.6.0.orig/gcc/stmt.c
+++ gcc-4.6.0/gcc/stmt.c
@@ -2478,7 +2478,11 @@ expand_case (gimple stmt)
default code is emitted. */
use_cost_table = 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);
if (default_label)
emit_jump (default_label);
@@ -3046,6 +3050,7 @@ emit_case_nodes (rtx index, case_node_pt
{
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,
|
