Re: glibc test time regression after gawk update

Sam James <[email protected]> Mon, 06 Jul 2026 00:19:03 +0100
Newsgroups	gmane.comp.gnu.gawk.bugs,gmane.comp.lib.glibc.alpha
Organization	Gentoo
Message-ID	<[email protected]>
[email protected] writes:

> Hi.

Hi Arnold,

>
> I suspect that the problem is line 34 in the script:
>
> 	value = gensub(/^(INF|NAN|inf|nan)/, "+\\1", 1, val)
>
> There was a bug in gensub with \N escapes that has been fixed.
> Can you try with https://www.skeeve.com/gawk/gawk-5.4.0d.tar.gz?
>
> FYI, feedback on this ASAP would be really good, as I am planning
> to release 5.4.1 this week.

Sorry for the delay, I'd tried to find time earlier this week but events
got the better of me.

It looks like it solves the problem:
```
$ hyperfine "make test t=stdio-common/tst-printf-format-vas-double-f AWK=/tmp/gawk-5.4.0/bin/gawk" "make test t=stdio-common/tst-printf-format-vas-double-f AWK=/tmp/gawk-5.4-stable/bin/gawk"
Benchmark 1: make test t=stdio-common/tst-printf-format-vas-double-f AWK=/tmp/gawk-5.4.0/bin/gawk
  Time (mean ± σ):      3.303 s ±  0.013 s    [User: 3.567 s, System: 0.116 s]
  Range (min … max):    3.281 s …  3.320 s    10 runs

Benchmark 2: make test t=stdio-common/tst-printf-format-vas-double-f AWK=/tmp/gawk-5.4-stable/bin/gawk
  Time (mean ± σ):      2.369 s ±  0.018 s    [User: 2.632 s, System: 0.114 s]
  Range (min … max):    2.352 s …  2.415 s    10 runs

Summary
  make test t=stdio-common/tst-printf-format-vas-double-f AWK=/tmp/gawk-5.4-stable/bin/gawk ran
    1.39 ± 0.01 times faster than make test t=stdio-common/tst-printf-format-vas-double-f AWK=/tmp/gawk-5.4.0/bin/gawk
```

Hopefully Florian can confirm as well. I didn't compare the full
testsuite time, just picked a sample of the tst-printf-format-* tests
(as for some reason I couldn't easily re-run just all of those with
`subdirs=`).

>
> Thanks,
>
> Arnold

cheers,
sam

>
> Sam James <[email protected]> wrote:
>
>> Sam James <[email protected]> writes:
>>
>> > Florian Weimer <[email protected]> writes:
>> >
>> >> It looks like we have a very substantial test time regression in glibc
>> >> after a gawk update.  Based on quick poking with GDB, it may be related
>> >> to the MinRX change in gawk 5.4.0.
>> >>
>> >> This is the script:
>> >>
>> >>   <https://sourceware.org/cgit/glibc/tree/stdio-common/tst-printf-format.awk>
>> >>
>> >> Is this expected?  The input files can be quite large, with very long
>> >> lines, too.
>> >>
>> >> I suppose we can switch back to the old matcher (will that remain
>> >> available?), or maybe run it under the C locale.
>> >
>> > Maciej, can you take a look? I suspect you're best placed to try reduce
>> > something here as you're the author of the tests.
>>
>> Resending.
>>
>> >
>> >>
>> >> Thanks,
>> >> Florian
signature.asc (application/pgp-signature, 418 B)