[patches] Re: ARM EABI mcount fix (second try)

[Daniel Jacobowitz <dan@xxxxxxxxxxxxxxxx>]

On Fri, Mar 28, 2008 at 12:06:03AM +0000, Paul Brook wrote:
> Attached is a second attempt at making gmon profiling work on 
> arm-linux-gnueabi. It comprises two parts (in a single file):
> 
> - Define a new __gnu_mcount_nc entry point, with an ABI that doesn't require 
> the compiler to generate crufty old APCS frames. I'm choosing to define a new 
> entry point to preserve backwards compatibility, and to allow for different 
> variants as necessary.  The "nc" comes from the fact that we do not need to 
> pass mcount the location of a counter.  Further justification can be found in 
> this thread:  http://gcc.gnu.org/ml/gcc-patches/2005-05/msg01001.html
> 
> - Rewrite the the old _mcount implementation to not rely on particular 
> compiler behavior when building glibc itself.
> 
> Tested on arm-linux-gnueabi.

Seems fine to me.

Richard commented (in above thread) that __gnu_mcount_nc in that
scheme can not clobber non-core registers.  I believe that means
mcount_internal and everything it calls must be compiled with no
VFP, iWMMXt, et cetera.  It should only matter if we supported
a hard-VFP / hard-iWMMXt ABI or interprocedural register allocation,
but at least a warning comment about the problem would be nice.

Fortunately mcount_internal only calls
catomic_compare_and_exchange_bool_acq, which will be inlined anyway,
no other functions.  So just that file needs special options.

-- 
Daniel Jacobowitz
CodeSourcery