Re: [PATCH] Poison gettext with the Ook language

From: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Date: Oct 23, 2018, 16:45 UTC
Message-ID: <871s8gd32p.fsf@evledraar.gmail.com>
In-Reply-To: <CACsJy8CX78EbANbv8a354djJaO6dKRpXshHhHJTspJvOSewgpA@mail.gmail.com>

On Tue, Oct 23 2018, Duy Nguyen wrote:

Show 37 quoted lines

> On Tue, Oct 23, 2018 at 12:17 PM Ævar Arnfjörð Bjarmason
> <avarab@gmail.com> wrote:
>>
>>
>> On Tue, Oct 23 2018, Johannes Schindelin wrote:
>>
>> > Hi Ævar,
>> >
>> > On Mon, 22 Oct 2018, Ævar Arnfjörð Bjarmason wrote:
>> >
>> >> So I think the only reason to keep it [GETTEXT_POISON] compile-time is
>> >> performance, but I don't think that matters. It's not like we're
>> >> printing gigabytes of _() formatted output. Everything where formatting
>> >> matters is plumbing which doesn't use this API. These messages are
>> >> always for human consumption.
>> >
>> > Well, let's make sure that your impression is correct before going too
>> > far. I, too, had the impression that gettext cannot possibly be expensive,
>> > especifally in Git for Windows' settings, where we do not even ship
>> > translations. Yet see the commit message of cc5e1bf99247 (gettext: avoid
>> > initialization if the locale dir is not present, 2018-04-21):
>> >
>> >       The runtime of a simple `git.exe version` call on Windows is
>> >       currently dominated by the gettext setup, adding a whopping ~150ms
>> >       to the ~210ms total.
>> >
>> > I would be in favor of your change to make this a runtime option, of
>> > course, as long as it does not affect performance greatly (in particular
>> > on Windows, where we fight an uphill battle to make Git faster).
>>
>> How expensive gettext() may or may not be isn't relevant to the
>> GETTEXT_POISON compile-time option.
>
> If a user requests NO_GETTEXT, they should have zero (or near zero)
> cost related to gettext. Which is true until now (the inline _()
> version is expanded to pretty much no-op with the exception of NULL
> string)

I'm assuming by "until now" you mean until my RFC patch in https://public-inbox.org/git/875zxtd59e.fsf@evledraar.gmail.com/

Yeah it's more expensive, but I don't see how it matters. We'd be nano-optimizing a thing that's never going to become a bottleneck. I.e. the patch is basically taking the commented-out codepath in this test program:

    #include <stdio.h>
    #include <stdlib.h>
    #include <string.h>

    int have_flag(void)
    {
        return 0;
        /*
        static int flag = -1;
        if (flag == -1)
            flag = strlen(getenv("GIT_TEST_FOO"));
        return flag;
        */
    }

    int main(void)
    {
        while (1) {
            const char *out = have_flag() ? "foo" : "bar";
            puts(out);
        }
    }

When I test that on my laptop with gcc 8.2.0 I get ~230MB/s of output on both versions, i.e. where have_flag() can be trivially constant folded, and where we need to call getenv() for both of:

    GIT_TEST_FOO= ./main | pv >/dev/null
    GIT_TEST_FOO=1 ./main | pv >/dev/null

We don't have any codepaths where we're calling gettext() to output more than a screenfull or two of output, so optimizing this doesn't make sense.

Show 10 quoted lines

>> The effect of what I'm suggesting here, and which my WIP patch in
>> <875zxtd59e.fsf@evledraar.gmail.com> implements is that we'd do a
>> one-time getenv() for each process that prints a _() message that we
>> aren't doing now, and for each message call a function that would check
>> a boolean "are we in poison mode" static global variable.
>
> Just don't do the getenv() inside _() even if you just do it one time.
> getenv() may invalidate whatever value from the previous call. I would
> not want to find out my code broke because of an getenv() inside some
> innocent _()...

How would any code break? We have various getenv("GIT_TEST_*...") already that work the same way. Unless you set that in the environment the function is idempotent, and I don't see how anyone would do that accidentally.

> And we should still respect NO_GETTEXT, not doing any extra work when
> it's defined.

This is not how any of our NO_* defines work. *Sometimes* defining them guarantees you do no extra work, but in other cases we've decided that bending over backwards with #ifdef in various places isn't worth it.

E.g. grep_opt in grep.h is a good example of this. Even if you don't compile with PCRE you're pointlessly memzero-ing out members of the struct that'll never be used in your config, because it's easier to maintain the code that way (types always checked at compile-time).

(And no, despite my involvement in PCRE I didn't introduce that particular pattern)

For the reasons noted above I think NO_GETTEXT is another such case. Just like GIT_TEST_SPLIT_INDEX (added in your d6e3c181bc ("read-cache: force split index mode with GIT_TEST_SPLIT_INDEX", 2014-06-13)) etc.