Re: Structure Return Testcase

[Geoff Keating <geoffk at geoffk dot org>]

Mark Mitchell <mark@codesourcery.com> writes:

> >>>>> "Geoff" == Geoff Keating <geoffk@geoffk.org> writes:
> 
>     >> There's really no chioce here: the committee has spoken.
> 
>     Geoff> Great!  What did they say?  (And, come to think of it,
>     Geoff> which committee?)
> 
> Probably we're talking sideways.  I meant the ANSI/ISO C committee;
> they've said that you have to do something memmove-like if you don't
> know there's no overlap.  That implies that if you do the
> pass-a-pointer-to-the-result trick, then you need to either a)
> guarantee that you pass a temporary or b) use memmove.
> 
> That's all I meant; I don't know what each of the various ABIs says
> about this.

Aah, I see.

Yes, in general ISO C requires that the arguments to memcpy() must not overlap.

However, at least on PowerPC, the ABI specifies that when you write

extern struct xxx func(struct xxx *a);

then this gets turned into

extern void func(struct xxx *hidden, struct xxx *a);

and the `hidden' parameter points to a new temporary.  So there's an
implication that `hidden' can't overlap `a'.  So in this case,
memcpy() is safe; the bug is that we don't properly create a temporary
for `hidden'.

-- 
- Geoffrey Keating <geoffk@geoffk.org>