Re: [PATCH v3] skip_prefix: rewrite so that prefix is scanned once
Duy Nguyen <pclouds@gmail.com> writes:
Show 33 quoted lines
> On Tue, Mar 04, 2014 at 01:09:39AM +0100, David Kastrup wrote:
>> Duy Nguyen <pclouds@gmail.com> writes:
>>
>> > On Tue, Mar 4, 2014 at 5:43 AM, Junio C Hamano <gitster@pobox.com> wrote:
>> >> diff --git a/git-compat-util.h b/git-compat-util.h
>> >> index cbd86c3..68ffaef 100644
>> >> --- a/git-compat-util.h
>> >> +++ b/git-compat-util.h
>> >> @@ -357,8 +357,14 @@ extern int suffixcmp(const char *str, const char *suffix);
>> >>
>> >> static inline const char *skip_prefix(const char *str, const char *prefix)
>> >> {
>> >> - size_t len = strlen(prefix);
>> >> - return strncmp(str, prefix, len) ? NULL : str + len;
>> >
>> > Just a note. gcc does optimize strlen("abcdef") to 6, and with that
>> > information at compile time built-in strncmp might do better.
>>
>> Indeed, most (but not all) of the calls have a constant string as
>> prefix. However, strncmp in each iteration checks for both *str as well
>> as *prefix to be different from '\0' independently (and it appears
>> unlikely to me that the optimizer will figure out that it's unnecessary
>> for either) _and_ compares them for equality so it's not likely to be
>> faster than the open-coded loop.
>>
>> One could, however, use memcmp instead of strncmp. I'm just not sure
>> whether memcmp is guaranteed not to peek beyond the first mismatching
>> byte even if the count would allow for more. It could lead to undefined
>> behavior if the first mismatching byte would be the ending NUL byte of
>> str.
>
> It turns out gcc does not generate a call to strncmp either. It
> inlines repz cmpsb instead.Oh wow. So it _does_ know that it's not necessary to check for a NUL byte when the length of one argument is already known. I am seriously impressed.
> I recall we had a discussion long ago about the inefficiency of repz
> and and open-coded loop is preferred,
I think that this mostly applies for Pentium I, possibly also the dead-ended Pentium Pro architecture (that sort-of translated the x86 opcodes into RISC instructions). I think that later processor variants (and AMD anyway) got those instructions back to usable shape.
One thing where there was a _lot_ of performance difference between open-coding and builtin was using memcpy (repz movb) for copying well-aligned data bytewise rather than copying, say, integer arrays element-wise.
But since that was egg-on-face material, the hardware got better at it.
And anyway, GCC should know what to pick here. So with GCC being as smart as that (using the equivalent of memcmp on its own initiative instead of the full strncmp), I don't think we have a reasonable chance to beat its performance with an open-coded variant.
Show 23 quoted lines
> produces this assembly with gcc -O2 (on x86, apparently)
>
> -- 8< --
> 00000000 <main>:
> 0: 55 push %ebp
> 1: b9 03 00 00 00 mov $0x3,%ecx
> 6: 89 e5 mov %esp,%ebp
> 8: 57 push %edi
> 9: bf 00 00 00 00 mov $0x0,%edi
> e: 56 push %esi
> f: 53 push %ebx
> 10: 83 e4 f0 and $0xfffffff0,%esp
> 13: 83 ec 10 sub $0x10,%esp
> 16: 8b 45 0c mov 0xc(%ebp),%eax
> 19: 8b 40 04 mov 0x4(%eax),%eax
> 1c: 89 c6 mov %eax,%esi
> 1e: f3 a6 repz cmpsb %es:(%edi),%ds:(%esi)
> 20: 0f 97 c3 seta %bl
> 23: 0f 92 c1 setb %cl
> 26: 83 c0 03 add $0x3,%eax
> 29: 31 d2 xor %edx,%edx
> 2b: 38 cb cmp %cl,%bl
> 2d: 0f 44 d0 cmove %eax,%edx
More like i686 than x86 here.
Show 9 quoted lines
> 30: 89 14 24 mov %edx,(%esp)
> 33: e8 fc ff ff ff call 34 <main+0x34>
> 38: 8d 65 f4 lea -0xc(%ebp),%esp
> 3b: 31 c0 xor %eax,%eax
> 3d: 5b pop %ebx
> 3e: 5e pop %esi
> 3f: 5f pop %edi
> 40: 5d pop %ebp
> 41: c3 ret
Well, we won't beat this here.
--
David Kastrup
Siddharth Goel
