2023-09-04 15:24:52

by David Laight

[permalink] [raw]
Subject: RE: [RFC PATCH v2 1/4] tools/nolibc: x86-64: Use `rep movsb` for `memcpy()` and `memmove()`

From: Ammar Faizi
> Sent: 02 September 2023 13:37
> On 2023/09/02 午後7:29, Alviro Iskandar Setiawan wrote:
> > On Sat, Sep 2, 2023 at 1:38 PM Ammar Faizi wrote:
> >> Ok, I'll do that.
> >
> > Another micro-optimization. Since the likely case is the forward copy,
> > make it the case that doesn't take the jump.
> >
> > Pseudo C:
> > if (unlikely(dst - src < n))
> > backward_copy();
> > else
> > forward_copy();
>
> Point taken.

I believe it makes almost no difference.
I'm sure I read that modern (Intel at least) cpu never do
static branch prediction.
So 'cold cache' there is 50% chance of the branch being taken.
Nothing the compiler can do will affect it.
OTOH having
if (likely(dst - src >= n))
forwards();
else
backwards();
(and in that order) probably makes the code a bit easier to read.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)