[Mesa-dev] [PATCH] glsl: add a missing call to _mesa_locale_init

[Erik Faye-Lund]

On Thu, Jul 2, 2015 at 2:56 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> On Thu, Jul 2, 2015 at 5:54 PM, Matt Turner <mattst88 at gmail.com> wrote:
>> On Thu, Jul 2, 2015 at 2:22 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>>> Can this be done at dlopen/init time? For example what happens if you do
>>>
>>> static int foo = _mesa_locale_init()
>>>
>>> IIRC things like that are possible in C++, not sure about C.
>>
>> gcc has __attribute__((constructor)).
>>
>> But I don't think we really care... Erik's series converted the strtod
>> code from C++ to C (including moving locale init from being a static
>> constructor to being called in one_time_init) and fixing the memory
>> leak.
>
> Well, this is just going to happen over and over again, I was hoping
> there was an easy way to do static initializers in C. If not, then I
> guess we're stuck with this.

The good news is that the breakage was noticed real quick. I agree
that it'd be awesome to have this happen automatically, but AFAIK
there's no perfect solution for this:

* C++ static object initializer leads to libc++ dependencies.
* __attribute__((constructor)) is compiler specific.
* Naively mutex-protecting initialization leads to overhead in the common case.
* Lighter-weight double checked locks are tricky to implement, and
  also have some overhead
* pthread_once() is not available on Windows, and have some overhead

Generally, it leaves a bad taste in the mouth to have to know about
compiler-internals for being able to use the compiler. This could be
slightly mitigated by introducing something like _mesa_glsl_init(),
which means that call-sites only need to know about initializing the
compiler.

However, such a function would only call _mesa_locale_init() at this
point, so we need to know the same amount of things. I'd introduce
_mesa_glsl_init() only if it turns out we need to do more. Besides,
new call-sites for _mesa_strto{d,f} will crash visible and hard if
introduced without _mesa_locale_init.

So, while not being awesome, I think this is the lesser evil for now.

But maybe it would be more palatable to move this into
initialize_context_to_defaults() instead, which is a shared
initializer for stand-alone compilers? Perhaps it's more likely that
this function gets called by other non-driver call-sites... It would
also reduce the number of call-sites by one.