On 14 Oct 2011, at 10:29, Mike Stump wrote:

> On Oct 14, 2011, at 2:05 AM, Iain Sandoe wrote:
>> This implements their use and also the GPRs - the latter makes an  
>> appreciable reduction in code size,
>
>> OK for trunk?
>
> Ok.  Watch for problems with async stack walking (hitting sample in  
> Activity Monitor, or the walking done by CrashReporter)...  that's  
> the only thing I can think of that might be strange.

This has taken some time to apply because of various bootstrap issues  
(version applied is attached)

In answer to your observation;
I didn't expect problems with FPR saves because the vendor's tools  
implement those.

To test what you suggested I built some code that dropped down a few  
stack levels (with saves of FPR/GPR) and then either aborts or spins  
on a sleep.

The crashlogs from the abort() and the instrumentation samples from  
the sleep were OK.

====

During doing this (and checking crosses to aix and eabisim) I noticed  
the following in rs6000/sysv4.h:

/* And similarly for general purpose registers.  */
#define GP_SAVE_INLINE(FIRST_REG) ((FIRST_REG) < 32     \
                                    && !optimize_size)

which gives rise to code (with -Os) like:

main:
         mr 11,1  #,
         stwu 1,-504(1)   #,,
         mflr 0   #,
         bl _savegpr_31   #
         lis 31,.LANCHOR0@ha      # tmp137,

which I doubt is what was intended ....

... copying David in case he feels that should be amended.

cheers
Iain