Re: [RFA] Fix float argument passing in inferior function calls for ppc64

[Thiago Jung Bauermann <bauerman@br.ibm.com>]

On Tue, 2008-01-15 at 15:43 +0100, Mark Kettenis wrote:
> > The 64-bit PowerPC ELF ABI version 1.9 says that "single precision
> > floating point values are mapped to the second word in a single
> > doubleword". The ppc64_sysv_abi_push_dummy_call function in
> > ppc-sysv-tdep.c, however, implements version 1.7 of the ABI which says
> > that they should go in the first doubleword.
> 
> Aren't ABI changes fun?

Heh. :-)

> Is the first word used for anything in the new ABI?  If not, you could
> support both ABIs by copying the value into both the first and the
> second word.

Great idea. I will resend the patch using that approach then.

> > Because of this, if you are calling a function with many arguments and
> > some need to be passed on the stack, GDB will get it wrong for 32-bit
> > floats. This is why in Linux/ppc64 GDB fails the "Call function with
> > many float arguments" test posted here:
> > 
> > http://sourceware.org/ml/gdb-patches/2008-01/msg00291.html
> > 
> > This patch fixes the test. It makes GDB pass 32-bit floats in the second
> > word when passing them in the stack as stated in the current ABI.
> > 
> > I didn't touch the code which writes the float to a general register in
> > the first word because I'm not sure how to test it. It is probably
> > related to soft-float, I guess.
> 
> Actually, it may be related to varargs.  Or perhaps it is a leftover
> from the 32-bit ABI code that was copied.

I'll play with varargs and see if I can trigger that code and create a
test for it.

> > Tested on Linux/ppc64 with no regressions. I believe this will also fix
> > other operating systems supported by GDB on ppc64 since they use the
> > same push_dummy_call implementation (assuming they also follow the
> > current SysV ABI), but I don't have the means to test them either.
> > 
> > Maybe someone could test the patch and the testcase in *BSD? Luis' patch
> > which has the testcases would also need this testing.
> 
> There's no 64-bit OpenBSD/powerpc port; we only run on Apple hardware
> and we run the G5 machines in 32-bit mode.
> 
> It doesn't look like NetBSD and FreeBSD have a working 64-bit powerpc
> port either.

Ok, good to know that.

> I'm fairly certain Linux/ppc64 used the old ABI for a while.  Are you
> certain GDB doesn't need to support any people running those older
> installs anymore?

No, I'm not actually...
Writing in both words should resolve this.
-- 
[]'s
Thiago Jung Bauermann
Software Engineer
IBM Linux Technology Center