Re: -foptimize-sibling-calls and passing structs

[Gabriel Paubert <paubert at iram dot es>]

On Sun, 4 Jun 2000, Jamie Lokier wrote:

> Artur Skawina wrote:
> > r() gets compiled to:
> > 
> >   <r>:
> > 	sub    $0xc,%esp
> > 	lea    0x10(%esp,1),%eax
> > 	mov    %eax,0x10(%esp,1)
> > 	add    $0xc,%esp
> > 	jmp    <p>
> > 
> > which does seem to be a bit aggressive...
> 
> I assume you mean not aggressive enough.  It should be
> 
>     lea    4(%esp,1),%eax
>     mov    %eax,4(%esp,1)
>     jmp    <p>

It seems that you missed the key point: the generated code and the one you
suggest overwrite the first element of the structure with a pointer to the
structure. The result is likely to be hmmm... interesting :-(

Sibling calls should check for case like this and disable the
optimizationg but this is API dependant: for example this should not be
disabled on PPC for example since structure are always passed by reference
and r would become a simple "b p" instruction in this trivial example.

I have no idea on how to solve the problem (besides disabling sibling
calls optimization altogether), but it might be worth adding Arthur's code
to the testsuite.

	Gabriel.