[cfe-commits] [libcxxabi] r148750 - /libcxxabi/trunk/src/cxa_exception_storage.cpp

[Howard Hinnant]

On Jan 25, 2012, at 12:43 PM, Howard Hinnant wrote:

> On Jan 25, 2012, at 12:28 PM, John McCall wrote:
> 
>> On Jan 23, 2012, at 3:55 PM, Howard Hinnant wrote:
>>> Author: hhinnant
>>> Date: Mon Jan 23 17:55:58 2012
>>> New Revision: 148750
>>> 
>>> URL: http://llvm.org/viewvc/llvm-project?rev=148750&view=rev
>>> Log:
>>> A lot of the code in cxa_exception.cpp depends on __cxa_get_globals_fast() returning null if __cxa_get_globals() hasn't been called yet.  However it doesn't reliably do that, at least on OS X if __cxa_get_globals_fast() is called prior to pthread_key_create() running.  Our choice is to either limit our use of __cxa_get_globals_fast() more than we have, or to have __cxa_get_globals_fast() initialize with pthread_key_create() if necessary.  I chose the latter, and replaced pthread_once with a C++11 local static (which should do the same thing).
> 
> The use of so many pronouns is making me a little nervous, so I'm going to try to confirm the intended use of each pronoun below.  Please let me know if I get it wrong.
> 
>> Are you certain this function is not used in the implementation of _cxa_guard_acquire?
> 
> I am certain that neither __cxa_get_globals_fast() nor __cxa_get_globals() is used in the implementation of __cxa_guard_acquire().  I just checked and cxa_guard.cpp does not currently even include the prototypes for __cxa_get_globals_fast() and __cxa_get_globals().
> 
>> If so, please at least document in the guard-variables code that it is no longer allowed to call this.
> 
> I'm not certain why the guard-variables code should be prohibited from calling __cxa_get_globals_fast() or __cxa_get_globals(), though I don't currently see the need either.
> 
> My commit notes could probably have been clearer:  This commit made calling __cxa_get_globals_fast() safer, but slower, as opposed to the other way around.  You can now call __cxa_get_globals_fast() even if __cxa_get_globals() has not previously been called.  And if you do, you are now guaranteed that __cxa_get_globals_fast() will return NULL.  Previous to this commit __cxa_get_globals_fast() would return a random non-NULL pointer if called prior to __cxa_get_globals().
> 
> The previous state (returning a random non-NULL pointer) was conforming to the Itanium ABI spec, but not consistent with our current use of __cxa_get_globals_fast() (which expects a NULL pointer).
> 
> If we're talking past each other, just let me know, and I'll try to clarify.  Thanks for the review.

As I hit send, I just realized your point:

There is now a dependency of __cxa_get_globals_fast() on __cxa_guard_acquire() because of the use of the local static, and we don't want a circular dependency of __cxa_guard_acquire() on __cxa_get_globals_fast().  Good point.  I'll comment cxa_guard.cpp.

Howard