Re: [PATCH 3/3, newlib] Allow locking routine to be retargeted

[Thomas Preudhomme <thomas.preudhomme@foss.arm.com>]

On 01/02/17 09:06, Freddie Chopin wrote:
> On Tue, 2017-01-31 at 17:19 +0000, Thomas Preudhomme wrote:
>>  > 1. Why is the lock used by at_quick_exit() not required - I have
>> no
>>  > idea. The linking error (when I do not provide storage for this
>> lock)
>>  > about multiple definitions is visible only when I actually use the
>>  > at_quick_exit() function.
>>
>> That makes sense. If you don't use at_quick_exit then probably no
>> symbol in the
>> corresponding object file is referenced and therefore the file is not
>> linked in.
>> That's exactly why the "doubly defined error on missing symbol" work
>> as
>> explained in the cover letter: file with dummy symbol is only linked
>> in if it
>> solves an undefined reference.
>
> This case seems a bit different, as it applies only to the
> at_quick_exit() lock. In my code I have no references to "env" or "tz"
> locks either, but I cannot drop these objects without the "multiple
> definitions" error. Maybe my test is flawed, but browsing the assembly
> output does not show any uses of __tz_lock(), __tz_unlock(),
> __env_lock() or __env_unlock() functions (which are not present in the
> output file), and these are the only places of newlib which use this
> particular locks. I'll try later with a simpler project.

Your code don't need to reference the env or tz lock directly for them to be 
linked in. The linker does a transitive closure of files until all references 
are resolved. So you code could reference a function a() present in a file A.o 
with that file containing references to function b() in file B.o which contains 
reference to tz_lock in file tz_lock.o which also contains the tz_mutex lock.

If you show me a small testcase where tz_lock or the env lock cannot be removed 
I can tell you why does the error happens.

Best regards,

Thomas