[patch, fortran] Load scalar intent-in variables at the beginning of procedures

[Janne Blomqvist]

On Thu, Nov 21, 2019 at 12:31 AM Janne Blomqvist
<blomqvist.janne@gmail.com> wrote:
>
> On Thu, Nov 21, 2019 at 12:18 AM Janne Blomqvist
> <blomqvist.janne@gmail.com> wrote:
> >
> > On Wed, Nov 20, 2019 at 11:35 PM Thomas König <tk@tkoenig.net> wrote:
> > > (Why do we zero %eax
> > > before each call? It should not be a variadic call right?)
> >
> > Not sure. Maybe some belt and suspenders thing? I guess someone better
> > versed in ABI minutiae knows better. It's not Fortran-specific though,
> > the C frontend does the same when calling a void function.
>
> Ah, scratch that, it is some varargs-thing, I had forgot that a C
> function with no arguments or lacking a prototype is considered a
> varargs. The code
>
> void foo();
> void bar(void);
>
> void testfoo()
> {
> foo();
> }
>
> void testbar()
> {
> bar();
> }
>
> void testunprototyped()
> {
> baz();
> }
>
>
> generates code (elided scaffolding):
>
> testfoo:
>     xorl %eax, %eax
>     jmp foo
> testbar:
>     jmp bar
> testunprototyped:
>     xorl %eax, %eax
>     jmp baz
>
>
> So probably this is due to the Fortran procedures lacking an interface
> being considered varargs by the caller.  Starts to smell like some
> leftover from PR 87689?

Thinking some more about it, I'm thinking we should leave it as is,
even though it's a (small) code bloat. So Fortran calling a procedure
with an implicit interface most closely resembles the C unprototyped
function call. The problem is that we don't know much about the
callee. What if a Fortran procedure calls a C varargs procedure? In
the Fortran caller, we don't know whether the callee is varargs or
not, but if we don't zero %eax all hell could break loose if it
happened to be a varargs function. Yes, we could hide behind Fortran
not supporting varargs, and it's all the users fault, but is it really
worth it? I'd say no.

(in some cases zeroing a register is used to tell the OoO hw to kill a
dependency, so it can be beneficial for performance even if not
strictly required for correctness)

-- 
Janne Blomqvist