Re: [PATCH v4 2/2] powerpc/irq: inline call_do_irq() and call_do_softirq()

From: Segher Boessenkool
Date: Fri Nov 29 2019 - 13:48:58 EST


Hi!

On Wed, Nov 27, 2019 at 04:15:15PM +0100, Christophe Leroy wrote:
> Le 27/11/2019 à 15:59, Segher Boessenkool a écrit :
> >On Wed, Nov 27, 2019 at 02:50:30PM +0100, Christophe Leroy wrote:
> >>So what do we do ? We just drop the "r2" clobber ?
> >
> >You have to make sure your asm code works for all ABIs. This is quite
> >involved if you do a call to an external function. The compiler does
> >*not* see this call, so you will have to make sure that all that the
> >compiler and linker do will work, or prevent some of those things (say,
> >inlining of the function containing the call).
>
> But the whole purpose of the patch is to inline the call to __do_irq()
> in order to avoid the trampoline function.

Yes, so you call __do_irq. You have to make sure that what you tell the
compiler -- and what you *don't tell the compiler -- works with what the
ABIs require, and what the called function expects and provides.

> >That does not fix everything. The called function requires a specific
> >value in r2 on entry.
>
> Euh ... but there is nothing like that when using existing
> call_do_irq().

> How does GCC know that call_do_irq() has same TOC as __do_irq() ?

The existing call_do_irq isn't C code. It doesn't do anything with r2,
as far as I can see; __do_irq just gets whatever the caller of call_do_irq
has.

So I guess all the callers of call_do_irq have the correct r2 value always
already? In that case everything Just Works.

> >So all this needs verification. Hopefully you can get away with just
> >not clobbering r2 (and not adding a nop after the bl), sure. But this
> >needs to be checked.
> >
> >Changing control flow inside inline assembler always is problematic.
> >Another problem in this case (on all ABIs) is that the compiler does
> >not see you call __do_irq. Again, you can probably get away with that
> >too, but :-)
>
> Anyway it sees I reference it, as it is in input arguments. Isn't it
> enough ?

It is enough for some things, sure. But not all.


Segher