Re: [RFC/PATCH 1/5] gettext: fix bug in git-sh-i18n's eval_gettext() by using envsubst(1)

From: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Date: Nov 9, 2010, 11:57 UTC
Message-ID: <AANLkTimhbp3AngtJjBYhHpa173=D-XJOg9L2sd6YMCwA@mail.gmail.com>
In-Reply-To: <4CD933BF.6070105@viscovery.net>

On Tue, Nov 9, 2010 at 12:42, Johannes Sixt <j.sixt@viscovery.net> wrote:

Show 24 quoted lines

> Am 11/9/2010 11:52, schrieb Ævar Arnfjörð Bjarmason:
>> On Tue, Nov 9, 2010 at 11:36, Johannes Sixt <j.sixt@viscovery.net> wrote:
>>> $ time (for i in {1..100}; do git version; done) > /dev/null
>>>
>>> real    0m5.610s
> ...
>>> $ time (for i in {1..100}; do eval_gettext foobar; done) > /dev/null
>>>
>>> real    0m20.578s
> ...
>>> Note that there are only 100 iterations, so we are talking about 0.2
>>> seconds per eval_gettext call! That's an awful lot of time even for a
>>> single error message.
>>
>> Thanks for elaborating. But just so I understand you correctly it's a
>> cost of invoking *any* program in shellscripts on Windows? So e.g. 10
>> sed calls would cost the same as 10 git-sh-i18n--envsubst calls (but
>> of course 5 eval_gettext() calls, since it calls git-sh-i18n--envsubst
>> twice).
>
> An invocation of sed, cat, etc. (POSIX/MSYS tools) is cheaper by a factor
> of 2 than a git invocation for a reason that I do not understand. (Perhaps
> it has to do with the number of DLLs that are linked; git has 9 static
> dependencies, MSYS tools only 3.)

git-sh-i18n--envsubst doesn't need to be linked to anything but the C library.

If it is that's because I fat-fingered the Makefile part of the patch. So it should be cheaper than cat/sed etc. if anything.

    $ ldd -r git-sh-i18n--envsubst
        linux-vdso.so.1 =>  (0x00007fff163c4000)
        libz.so.1 => /usr/lib/libz.so.1 (0x00007f742dea6000)
        libcrypto.so.0.9.8 => /usr/lib/libcrypto.so.0.9.8 (0x00007f742db05000)
        libpthread.so.0 => /lib/libpthread.so.0 (0x00007f742d8e8000)
        libc.so.6 => /lib/libc.so.6 (0x00007f742d587000)
        libdl.so.2 => /lib/libdl.so.2 (0x00007f742d383000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f742e0cc000)

That's on Linux, it doesn't need libz, libcrypto etc. It just needs standard C library functions along with its own main().

Show 8 quoted lines

> eval_gettext involves 2 git invocation (git-sh-i18n--envsubst counts as
> much as git) and 1 subshell in practice (some subshells are optimized away).
>
>> So e.g. using eval_gettext once isn't a bigger problem than calling
>> some trivial sed substitution twice?
>
> Look at the timings: In my book, eval_gettext counts like about 8 sed
> substitutions.

How about if you just replace your tests with "cat". That should give a more accurate indication of what speed it *should* be operating at, once I fix those Makefile issues.

Anyway, if it's no more expensive than cat(1) (which it shouldn't be) it probably won't be a problem to use git-sh-i18n--envsubst.

Actually I'd rather keep the current calling semantics (using variables in the strings) than improve performance slightly on Windows for what are mostly non-critical error messages. But that's just my opinion.