Hi,
On Thu, Apr 14 2011, Nicolas Pitre wrote:
> Did you disassemble the resulting binary to make sure this is actually
> as performant?
>
> I'm asking because gcc used to do a horrible dumb job with such patterns
> where it would allocate two instances of the structure on the stack i.e.
> one for the named variable and one for the initializer, then fill the
> later with zeroes, and then call memcpy() to copy the initializer over
> to the named instance.
Nothing horrible as far as I can tell; in fact, the code size on ARM
(stripped mmc_core.ko) decreases by 216 bytes after the patchset here.
Here's a sample disassembly of one function on ARM with gcc-4.6.0, and
diff -y. Explicit memset (which calls __memzero) on the left, and {0}
initializer (which calls memset) on the right:
00000698 <mmc_set_relative_addr>: | 000006c8
<mmc_set_relative_addr>:
698: e92d4010 push {r4, lr} | 6c8:
e92d4010 push {r4, lr}
69c: e2504000 subs r4, r0, #0 | 6cc:
e24dd030 sub sp, sp, #48 ; 0x30
6a0: e24dd030 sub sp, sp, #48 ; 0x30 | 6d0:
e1a04000 mov r4, r0
6a4: 059f0058 ldreq r0, [pc, #88] ; 704 <mmc_se | 6d4:
e3a01000 mov r1, #0
6a8: 03a010ce moveq r1, #206 ; 0xce | 6d8:
e1a0000d mov r0, sp
6ac: 0a000004 beq 6c4 <mmc_set_relative_addr+0x | 6dc:
e3a02030 mov r2, #48 ; 0x30
6b0: e5943000 ldr r3, [r4] | 6e0:
ebfffffe bl 0 <memset>
6b4: e3530000 cmp r3, #0 | 6e4:
e3540000 cmp r4, #0
6b8: 1a000002 bne 6c8 <mmc_set_relative_addr+0x | 6e8:
059f0048 ldreq r0, [pc, #72] ; 738 <mmc_se
6bc: e59f0040 ldr r0, [pc, #64] ; 704 <mmc_se | 6ec:
03a010c4 moveq r1, #196 ; 0xc4
6c0: e3a010cf mov r1, #207 ; 0xcf | 6f0:
0a000004 beq 708 <mmc_set_relative_addr+0x
6c4: ebfffffe bl 0 <__bug> | 6f4:
e5940000 ldr r0, [r4]
6c8: e1a0000d mov r0, sp | 6f8:
e3500000 cmp r0, #0
6cc: e3a01030 mov r1, #48 ; 0x30 | 6fc:
1a000002 bne 70c <mmc_set_relative_addr+0x
6d0: ebfffffe bl 0 <__memzero> | 700:
e59f0030 ldr r0, [pc, #48] ; 738 <mmc_se
6d4: e59430c8 ldr r3, [r4, #200] ; 0xc8 | 704:
e3a010c5 mov r1, #197 ; 0xc5
6d8: e3a02003 mov r2, #3 | 708:
ebfffffe bl 0 <__bug>
6dc: e1a03803 lsl r3, r3, #16 | 70c:
e59430c8 ldr r3, [r4, #200] ; 0xc8
6e0: e58d3004 str r3, [sp, #4] | 710:
e3a02003 mov r2, #3
6e4: e5940000 ldr r0, [r4] | 714:
e1a03803 lsl r3, r3, #16
6e8: e3a03015 mov r3, #21 | 718:
e58d3004 str r3, [sp, #4]
6ec: e1a0100d mov r1, sp | 71c:
e1a0100d mov r1, sp
6f0: e58d2000 str r2, [sp] | 720:
e3a03015 mov r3, #21
6f4: e58d3018 str r3, [sp, #24] | 724:
e58d2000 str r2, [sp]
6f8: ebfffffe bl 0 <mmc_wait_for_cmd> | 728:
e58d3018 str r3, [sp, #24]
6fc: e28dd030 add sp, sp, #48 ; 0x30 | 72c:
ebfffffe bl 0 <mmc_wait_for_cmd>
700: e8bd8010 pop {r4, pc} | 730:
e28dd030 add sp, sp, #48 ; 0x30
704: 0000001d .word 0x0000001d | 734:
e8bd8010 pop {r4, pc}
> 738:
0000001d .word 0x0000001d
Thanks,
- Chris.
--
Chris Ball <[email protected]> <http://printf.net/>
One Laptop Per Child
--
To unsubscribe from this list: send the line "unsubscribe linux-mmc" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
