patch for broken thunk on sparc

[Eric Botcazou]

> on sparc the thunk (support for C++ multiple inheritance)
> is created incorrectly if flag_pic or -mcmodel=*medany -m64 are present.
>
> The problem is in code (for -m32):
>         sub     %sp, 8, %sp
>         st      %o7, [%sp]
>         st      %l7, [%sp+4]
> which incorrectly uses window save space.
> If OS needs to save window after that sequence the restored registers
> will contain some junk.

Thanks for catching that!  This part of sparc_output_mi_thunk really is 
over-engineered and I should probably have refrained from writing it in the 
first place.  But let's make it correct now that it is there.

> For some reason -O and above optimize that thunk sequence away, so users
> at least have sort of workaround.

Because flag_delayed_branch is true.  The whole stuff is aimed to emit a thunk 
(i.e. essentially a sibling call) without using delay slots...

> Would be nice to patch all 4.x branches.

Indeed, I'll install the fix everywhere.

> Here is the proposed patch:
>
>         * config/sparc/sparc.c (emit_and_preserve): Use
>         correct slot offset to preserve registers

I think you'll be clobbering the previous frame.  We probably need to build a 
dummy frame of SPARC_STACK_ALIGN (FIRST_PARM_OFFSET(0) + 2*UNITS_PER_WORD) 
bytes.  And I also think that the offset can simply be STACK_POINTER_OFFSET.

-- 
Eric Botcazou