Hi,

On Fri, 2008-01-11 at 10:57 -0500, Daniel Jacobowitz wrote:
> On Fri, Jan 11, 2008 at 01:55:20PM -0200, Luis Machado wrote:
> > > How about an extra flag inside the gdbarch_tdep structure that
> > > you'll be using inside the push dummy call?  It would be a shame
> > > duplicate this function when most of it is common.
> > 
> > Yes. This would work, but putting linux specific code inside the
> > ppc-sysv-tdep.c file doesn't look quite right. But if there isn't a
> > better way of doing this, should be OK.
> 
> For practical reasons, this is how it is usually done.

Actually, we won't need to go through the flag. I've checked and the
PPC/Linux ABI conforms with the PPC/SysV ABI regarding Calling Sequences
and the Stack layout.

Considering we have a function with a great number of float parameters,
exceeding the number of available FP registers, we need to store those
parameters in the stack space, and then there's a problem with the
ABI/GDB.

The SysV ABI says we should promote exceeding float parameters (the ones
that should go to the stack) to double, and store then with 8 bytes
alignment in the stack.

Trying to call a function like that, from within GDB, made it clear that
some parameters were getting trashed, and this is due to the fact that
GCC (XLC as well) doesn't promote floats to doubles and does not align
them to 8 bytes in the stack. Actually, it just aligns the prototyped
float parameters to 4 bytes in the stack.

This is something that needs to be improved in the ABI text. I've
attached a note next to the portion of code responsible for handling
float parameters.

The patch follows, with a minor change in the ABI handling for PPC32 and
a testcase Thiago wrote to verify the problem.

Comments?

Regards,

-- 
Luis Machado
Software Engineer 
IBM Linux Technology Center