[Bug target/66822] Turn -mtune=iamcu for size
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66822 --- Comment #7 from Uroš Bizjak ubizjak at gmail dot com --- (In reply to Uroš Bizjak from comment #3) Please hold a bit with this change, I have a patch that improves generation of zero_extend substantially. Eh, it won't have such drastic effect on code size, so if size is priority, X86_TUNE_ZERO_EXTEND_WITH_AND should be disabled.
[Bug target/66822] Turn -mtune=iamcu for size
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66822
--- Comment #6 from H.J. Lu hjl.tools at gmail dot com ---
(In reply to H.J. Lu from comment #5)
X86_TUNE_USE_BT also reduces code size:
[hjl@gnu-tools-1 pr66822]$ cat y.i
void bar (void);
void
foo (int x)
{
if (x != 2 x != 3 x != 10 x != 11 x != 17 x != 18 x !=
23)
bar ();
}
[hjl@gnu-tools-1 pr66822]$ make y.o
/export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu
-mtune=iamcu -S -o y.s y.i
/export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu
-mtune=iamcu -c -o y.o y.s
rm y.s
[hjl@gnu-tools-1 pr66822]$ objdump -dwr y.o
y.o: file format elf32-i386
Disassembly of section .text:
foo:
0: 83 f8 17cmp$0x17,%eax
3: 76 0b jbe10 foo+0x10
5: e9 fc ff ff ff jmp6 foo+0x6 6: R_386_PC32 bar
a: 8d b6 00 00 00 00 lea0x0(%esi),%esi
10: ba 0c 0c 86 00 mov$0x860c0c,%edx
15: 88 c1 mov%al,%cl
17: d3 ea shr%cl,%edx
19: 83 e2 01and$0x1,%edx
1c: 74 e7 je 5 foo+0x5
1e: c3 ret
[hjl@gnu-tools-1 pr66822]$ /export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu
-mtune=iamcu -c y.i -mtune-ctrl=use_bt
[hjl@gnu-tools-1 pr66822]$ objdump -dwr y.o
y.o: file format elf32-i386
Disassembly of section .text:
foo:
0: 83 f8 17cmp$0x17,%eax
3: 76 0b jbe10 foo+0x10
5: e9 fc ff ff ff jmp6 foo+0x6 6: R_386_PC32 bar
a: 8d b6 00 00 00 00 lea0x0(%esi),%esi
10: ba 0c 0c 86 00 mov$0x860c0c,%edx
15: 0f a3 c2bt %eax,%edx
18: 73 eb jae5 foo+0x5
1a: c3 ret
[hjl@gnu-tools-1 pr66822]$
We can reduce size further by
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index bca1fee..4f08d0a 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -2556,7 +2556,7 @@ static const struct ptt
processor_target_table[PROCESSOR_max] =
{i386, i386_cost, 4, 3, 4, 3, 4},
{i486, i486_cost, 16, 15, 16, 15, 16},
{pentium, pentium_cost, 16, 7, 16, 7, 16},
- {iamcu, iamcu_cost, 16, 7, 16, 7, 16},
+ {iamcu, iamcu_cost, 0, 0, 0, 0, 0},
{pentiumpro, pentiumpro_cost, 16, 15, 16, 10, 16},
{pentium4, pentium4_cost, 0, 0, 0, 0, 0},
{nocona, nocona_cost, 0, 0, 0, 0, 0},
and get
foo:
0: 83 f8 17cmp$0x17,%eax
3: 76 05 jbea foo+0xa
5: e9 fc ff ff ff jmp6 foo+0x6 6: R_386_PC32 bar
a: ba 0c 0c 86 00 mov$0x860c0c,%edx
f: 0f a3 c2bt %eax,%edx
12: 73 f1 jae5 foo+0x5
14: c3 ret
[hjl@gnu-tools-1 pr66822]$
[Bug target/66822] Turn -mtune=iamcu for size
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66822
--- Comment #5 from H.J. Lu hjl.tools at gmail dot com ---
X86_TUNE_USE_BT also reduces code size:
[hjl@gnu-tools-1 pr66822]$ cat y.i
void bar (void);
void
foo (int x)
{
if (x != 2 x != 3 x != 10 x != 11 x != 17 x != 18 x != 23)
bar ();
}
[hjl@gnu-tools-1 pr66822]$ make y.o
/export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu -mtune=iamcu
-S -o y.s y.i
/export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu -mtune=iamcu
-c -o y.o y.s
rm y.s
[hjl@gnu-tools-1 pr66822]$ objdump -dwr y.o
y.o: file format elf32-i386
Disassembly of section .text:
foo:
0: 83 f8 17cmp$0x17,%eax
3: 76 0b jbe10 foo+0x10
5: e9 fc ff ff ff jmp6 foo+0x6 6: R_386_PC32 bar
a: 8d b6 00 00 00 00 lea0x0(%esi),%esi
10: ba 0c 0c 86 00 mov$0x860c0c,%edx
15: 88 c1 mov%al,%cl
17: d3 ea shr%cl,%edx
19: 83 e2 01and$0x1,%edx
1c: 74 e7 je 5 foo+0x5
1e: c3 ret
[hjl@gnu-tools-1 pr66822]$ /export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu -mtune=iamcu
-c y.i -mtune-ctrl=use_bt
[hjl@gnu-tools-1 pr66822]$ objdump -dwr y.o
y.o: file format elf32-i386
Disassembly of section .text:
foo:
0: 83 f8 17cmp$0x17,%eax
3: 76 0b jbe10 foo+0x10
5: e9 fc ff ff ff jmp6 foo+0x6 6: R_386_PC32 bar
a: 8d b6 00 00 00 00 lea0x0(%esi),%esi
10: ba 0c 0c 86 00 mov$0x860c0c,%edx
15: 0f a3 c2bt %eax,%edx
18: 73 eb jae5 foo+0x5
1a: c3 ret
[hjl@gnu-tools-1 pr66822]$
[Bug target/66822] Turn -mtune=iamcu for size
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66822
H.J. Lu hjl.tools at gmail dot com changed:
What|Removed |Added
Summary|Turn off|Turn -mtune=iamcu for size
|X86_TUNE_ZERO_EXTEND_WITH_A |
|ND for -miamcu |
--- Comment #4 from H.J. Lu hjl.tools at gmail dot com ---
-mtune=iamcu should be tuned for size. X86_TUNE_ZERO_EXTEND_WITH_AND
lead to bigger size:
[hjl@gnu-tools-1 pr66822]$ cat x.i
unsigned int
foo (unsigned short x)
{
return x;
}
[hjl@gnu-tools-1 pr66822]$ /export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu -mtune=iamcu
-c -o x.o x.i -mtune-ctrl=^zero_extend_with_and
[hjl@gnu-tools-1 pr66822]$ objdump -dwr x.o
x.o: file format elf32-i386
Disassembly of section .text:
foo:
0: 0f b7 c0movzwl %ax,%eax
3: c3 ret
[hjl@gnu-tools-1 pr66822]$ /export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O2 -m32 -miamcu -mtune=iamcu
-c -o x.o x.i -mtune-ctrl=zero_extend_with_and
[hjl@gnu-tools-1 pr66822]$ objdump -dwr x.o
x.o: file format elf32-i386
Disassembly of section .text:
foo:
0: 25 ff ff 00 00 and$0x,%eax
5: c3 ret
[hjl@gnu-tools-1 pr66822]$
