Mindless ABI adherence

[Linus Torvalds]

On Fri, 19 Oct 2001, Florian Weimer wrote:
>
> > 	add:
> > 		addl %edx,%eax
> > 		ret;
> >
> > 	stub$add:
> > 		movl 4(%esp),eax
> > 		movl 8(%esp),edx
> > 		jmp add;
>
> I think you could even avoid the jump, just by reordering the code.

In this case, yes. However, it was meant to be an example of:

 - not needing a stub at all when not needed (ie generating it later only
   if required - without having any global analysis)

 - the generic case, which can NOT be done with a re-order, because the
   calling convention change might require saving/restoring of registers
   (this doesn't happen with the x86 regcall optimization, but it could
   happen with other calling convention changes: return values in
   registers vs stack etc)

But my example was obviously too limited to show what the stub approach
_really_ can do.

(NOTE! The whole point of a stub approach is to not have to use it. If the
stub ends up being common, then that just means that the "optimization"
was a bad one. So the common case _should_ be that the stub doesn't exist
at all, or if it exists it should be at least dynamically the rare case.
And I realize that you cannot guarantee _that_ without a global pass, but
in many cases simple local heuristics are likely to be good for 99% of all
cases)

		Linus