This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: split_complex_* flaw
- From: David Edelsohn <dje at watson dot ibm dot com>
- To: Alan Modra <amodra at bigpond dot net dot au>, Richard Henderson <rth at redhat dot com>
- Cc: gcc at gcc dot gnu dot org, Aldy Hernandez <aldyh at redhat dot com>
- Date: Sat, 31 Jan 2004 22:43:45 -0500
- Subject: Re: split_complex_* flaw
- References: <20040131104133.GC5080@bubble.modra.org>
>>>>> Alan Modra writes:
> Consider _Complex float on powerpc64 linux.
> a) It is laid out in the parameter save area of memory as two adjacent
> 32 bit words, ie. in one doubleword.
>
> Two adjacent float args are laid out in the parameter save area in
> separate doublewords.
>>>>> Richard Henderson writes:
> I strongly
> suggest just fixing it so that the rule is simple: complex
> arguments are passed as two independent scalar arguments.
Which symptom are we concerned about? On AIX and PPC64 Linux, the
parameter save area is GPR register images. GPRs are doublewords for
PPC64, so all parameter save area slots are doublewords. Float and
_Complex float arguments only occupy half of the slot, but the slot still
is a doubleword.
A pointer to a complex float is a pointer to a pair of packed,
adjacent floats occupying a doubleword, but the stack image is two
doublewords.
I thought the SPLIT_COMPLEX_ARGS stuff meant that the parameter
passing machinery of GCC only saw two floats, so it should just treat it
as two, scalar arguments. If some part of GCC is storing a packed
_Complex float in the parameter save area, that would appear to be the
problem and wouldn't be honoring SPLIT_COMPLEX_ARGS.
David