From: David Howells
> Sent: 22 November 2020 13:33
>
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> > - I worry a bit about the indirect call overhead and spectre v2.
>
> I don't know enough about how spectre v2 works to say if this would be a
> problem for the ops-table approach, but wouldn't it also affect the chain of
> conditional branches that we currently use, since it's branch-prediction
> based?
The advantage of the 'chain of branches' is that it can be converted
into a 'tree of branches' because the values are all separate bits.
So as well as putting the (expected) common one first; you can do:
if (likely((a & (A | B))) {
if (a & A) {
code for A;
} else {
code for B;
} else ...
So get better control over the branch sequence.
(Hopefully the compiler doesn't change the logic.
I want a dumb compiler that (mostly) compiles what I write!)
Part of the difficulty is deciding the common case.
There'll always be a benchmark that exercises an uncommon case.
Adding an indirect call does let you do things like adding
ITER_IOVER_SINGLE and ITER_KVEC_SINGLE that are used in the
common case of a single buffer fragment.
That might be a measurable gain.
It is also possible to optimise the common case to a direct
call (or even inline code) and use an indirect call for
everything else.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
