Re: [PATCH] x86: Align jump targets to 1 byte boundaries

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, Apr 10, 2015 at 7:53 AM, Denys Vlasenko <dvlasenk@redhat.com> wrote:
>
> There are people who experimentally researched this.
> According to this guy:
>
> http://www.agner.org/optimize/microarchitecture.pdf
>
> Intel CPUs can decode only up to 16 bytes at a time

Indeed.

For intel decoding, the old "4-1-1-1" decode patterns are almost
entirely immaterial these days. Even the "single uop" (the "1"s int he
4-1-1-1) cover the vast majority of cases.

So for Intel decoders, the biggest limit - especially for x86-64
instructions - tends to be the 16-byte decode window. The problem with
x86 decoding isn't that individual instructions are complicated, but
the fact that when you try to decode multiple instructions at once,
finding the start of each instruction is somewhat painful. What I
*think* Intel does is have this rather complex net of logic that
basically decodes 16 bytes in parallel, but has this rippling thing
that just disables the incorrect decodes.

That said, the fetch boundary from L2 is probably an issue too,
especially if the front-end hasn't had time to run ahead of the
execution engine. That's likely where the "32 byte alignment" comes
from.

                        Linus