Re: Deduplicate io_*_prep calls?

[Andres Freund <andres@anarazel.de>]

Hi,

On 2020-02-24 08:40:16 -0700, Jens Axboe wrote:
> Agree that the first patch looks fine, though I don't quite see why
> you want to pass in opcode as a separate argument as it's always
> req->opcode. Seeing it separate makes me a bit nervous, thinking that
> someone is reading it again from the sqe, or maybe not passing in
> the right opcode for the given request. So that seems fragile and it
> should go away.

Without extracting it into an argument the compiler can't know that
io_kiocb->opcode doesn't change between the two switches - and therefore
is unable to merge the switches.

To my knowledge there's no easy and general way to avoid that in C,
unfortunately. const pointers etc aren't generally a workaround, even
they were applicable here - due to the potential for other pointers
existing, the compiler can't assume values don't change.  With
sufficient annotations of pointers with restrict, pure, etc. one can get
it there sometimes.

Another possibility is having a const copy of the struct on the stack,
because then the compiler often is able to deduce that the value
changing would be undefined behaviour.


I'm not sure that means it's worth going for the separate argument - I
was doing that mostly to address your concern about the duplicated
switch cost.

Greetings,

Andres Freund