So, finally looping back to the httpd vs. apr implementations of these
functions...
apr was testing null against equality and both *str1 null and *str2 null
characters - that's obviously wrong since str2 of null. It also performed
final unnecessary increments of str1 and str1, which I've
On Tue, Nov 24, 2015 at 10:07 AM, Mikhail T.
wrote:
> On 24.11.2015 10:08, William A Rowe Jr wrote:
>
> As long as this function is promoted for fast ASCII-specific token
> recognition and has no other unexpected equalities, it serves a useful
> purpose.
>
> Because of
On 24.11.2015 10:08, William A Rowe Jr wrote:
> As long as this function is promoted for fast ASCII-specific token
> recognition and has no other unexpected equalities, it serves a useful
> purpose.
Because of this, I'd suggest renaming it to something, that emphasizes
it being ASCII-only.
On Tue, Nov 24, 2015 at 4:12 AM, Mikhail T. wrote:
> On 23.11.2015 19:43, Yann Ylavic wrote:
>
>> That's expected (or at least no cared about in this test case). We simply
>> want res to not be optimized out, so print it before leaving, without any
>> particular
On Tue, Nov 24, 2015 at 6:10 AM, Mikhail T. wrote:
>
> Attached is the same program with one more pair of
> functions added (and an easy way to add more "candidates" to the
> main-driver). I changed the FOR-loop define to obtain repeatable results:
This test program
On Tue, Nov 24, 2015 at 6:40 AM, Jim Jagielski wrote:
> It really depends on the OS and the version of the OS. In
> my test cases on OSX and CentOS5 and centOS6, I see
> measurable improvements.
>
Part of the reason for your differences... on this console here I have;
$ set |
It really depends on the OS and the version of the OS. In
my test cases on OSX and CentOS5 and centOS6, I see
measurable improvements.
> On Nov 23, 2015, at 3:12 PM, Christophe JAILLET
> wrote:
>
> Hi Jim,
>
> Do you have done some benchmark and have results to
Hi Jim,
Do you have done some benchmark and have results to share?
I tried to do some but the benefit of the optimized version is not that
clear, at least on my system:
gcc 5.2.1
Linux linux 4.2.0-18-generic #22-Ubuntu SMP Fri Nov 6 18:25:50 UTC
2015 x86_64 x86_64 x86_64 GNU/Linux
Hi Christophe,
On Mon, Nov 23, 2015 at 9:12 PM, Christophe JAILLET
wrote:
>
> I tried to do some but the benefit of the optimized version is not that
> clear, at least on my system:
>gcc 5.2.1
>Linux linux 4.2.0-18-generic #22-Ubuntu SMP Fri Nov 6 18:25:50
I modified your test program a bit (to measure time from it, see
attached), tried with -O{2,3,s}, and except -Os I always have better
results with the "optimized" version, eg:
$ ./a-O3.out 0 15000 xcxcxcxcxcxcxcxcxcxcwwaa
xcxcxcxcxcxcxcxcxcxcwwaa 0
Please note that the changes in ap_str[n]casecmp(), ie:
++ps1;
++ps2;
was a first try/change which (obviously) did nothing.
You may ignore it.
On Mon, Nov 23, 2015 at 11:43 PM, Yann Ylavic wrote:
> with attachment...
>
> On Mon, Nov 23, 2015 at 11:42 PM,
with attachment...
On Mon, Nov 23, 2015 at 11:42 PM, Yann Ylavic wrote:
> I modified your test program a bit (to measure time from it, see
> attached), tried with -O{2,3,s}, and except -Os I always have better
> results with the "optimized" version, eg:
>
> $ ./a-O3.out 0
On 23.11.2015 23:14, William A Rowe Jr wrote:
> L1 cache and other direct effects of cpu internal optimization.
Just what I was thinking. Attached is the same program with one more
pair of functions added (and an easy way to add more "candidates" to the
main-driver). I changed the FOR-loop define
On Nov 23, 2015 21:12, "Mikhail T." wrote:
>
> On 23.11.2015 19:43, Yann Ylavic wrote:
>>
>> No measured difference in my tests, I guess it depends on likelyhood to
fail/succeed early in the string or not.
>
> ? I don't see, where it wins anything -- but I do see, where
On 23.11.2015 19:05, Yann Ylavic wrote:
> Here is the correct (new) test, along with the diff wrt the original
> (Christophe's) test.c.
BTW, if the program measures its own time, should it not use getrusage()
instead of gettimeofday()?
-mi
On Tue, Nov 24, 2015 at 1:07 AM, Mikhail T. wrote:
>
> BTW, if the program measures its own time, should it not use getrusage()
> instead of gettimeofday()?
Well, it measures the time spent in the relevant code, with a
monotonic clock, that should be fair enough.
We
On Mon, Nov 23, 2015 at 11:42 PM, Yann Ylavic wrote:
> except -Os I always have better
> results with the "optimized" version
To reach better performances with -Os, we could possibly use:
int ap_strcasecmp(const char *s1, const char *s2)
{
const unsigned char *ps1 =
On 23.11.2015 17:43, Yann Ylavic wrote:
> with attachment...
There is a mistake somewhere in the optimized version:
./o 1 1 aa1a 0
Optimized (nb=1, len=0)
time = 0.611311 : res = 0
The result should not be zero. Indeed, the string.h version is correct:
./o 0
On 23.11.2015 19:05, Yann Ylavic wrote:
> while (ucharmap[*ps1] == ucharmap[*ps2++]) {
> if (*ps1++ == '\0') {
> return (0);
> }
> }
> return (ucharmap[*ps1] - ucharmap[*--ps2]);
Is there really a gain in inc- and decrementing this way? Would not it
be
FWIW, a new version using clock_gettime() instead of gettimeofday().
Same/Comparable results for optimized vs string.h's (the former wins
but with -Os).
On Tue, Nov 24, 2015 at 1:43 AM, Yann Ylavic wrote:
> On Tue, Nov 24, 2015 at 1:24 AM, Mikhail T.
On Tue, Nov 24, 2015 at 12:46 AM, Mikhail T. wrote:
> On 23.11.2015 17:43, Yann Ylavic wrote:
>
> with attachment...
>
> There is a mistake somewhere in the optimized version:
My bad, I somehow corrupted the original ap_str[n]casecmp() functions.
Here is the correct
On Tue, Nov 24, 2015 at 1:24 AM, Mikhail T. wrote:
>
> Is there really a gain in inc- and decrementing this way? Would not it be
> easier to read with the explicit increments -- and, incidentally, no
> decrements at all?
No measured difference in my tests, I guess it
On 23.11.2015 19:43, Yann Ylavic wrote:
> No measured difference in my tests, I guess it depends on likelyhood to
> fail/succeed early in the string or not.
? I don't see, where it wins anything -- but I do see, where it loses a
little...
> That's expected (or at least no cared about in this test
I just made a small application which takes as command line parameters
the number of iteration to run, which version of the algorithm to use,
the 2 strings to compare and the length to compare (or 0 for the non 'n'
versions)
Compiled using
gcc -O3 test.c
Tested using
I'm +1 if you are suggesting an ascii-only implementation, and by all means
introduce both an ap_ and apr_ flavor at the same time.
E.g. apr[r]_ascii_str[n]casecmp().
All characters <'A' || >'Z' && <'a || >'z' should be compared by identity.
I'm not sure which OS/X implementation you are
Pay special attention to;
The *strncasecmp*() function shall compare, *while ignoring differences in
case*, not more than *n* bytes from the string pointed to by *s1* to the
string pointed to by *s2*.
In the POSIX locale, *strcasecmp*() and *strncasecmp*() shall *behave as if
the strings had
+1
Le 20/11/2015 18:17, Jim Jagielski a écrit :
We use str[n]casecmp quite a bit. The rub is that some
platforms use a sensible implementation (such as OSX) which
uses a upper-lowercase map and is v. fast, and others
use a brain-dead version which does an actual tolower() of
each char in the
+1
On Fri, Nov 20, 2015 at 6:17 PM, Jim Jagielski wrote:
> We use str[n]casecmp quite a bit. The rub is that some
> platforms use a sensible implementation (such as OSX) which
> uses a upper-lowercase map and is v. fast, and others
> use a brain-dead version which does an
Implemented in r1715401
If people want to nit-pick about naming and wish to
rename it to something else, be my guest.
> On Nov 20, 2015, at 1:03 PM, Yann Ylavic wrote:
>
> +1
>
> On Fri, Nov 20, 2015 at 6:17 PM, Jim Jagielski wrote:
>> We use
At first glance this seems to meet the POSIX definition in the POSIX
context. Just want to be careful about using this for non-posix
applications and your doxygen seems to cover this. Looks good to me.
Implemented in r1715401
If people want to nit-pick about naming and wish to
rename it to
Le 20/11/2015 18:17, Jim Jagielski a écrit :
Ideally, it would be in apr
+1
This could also be even more interesting, because of apr_table_ functions.
CJ
31 matches
Mail list logo