Re: PATCH libatomic

[christos%astron.com@localhost (Christos Zoulas)]

In article <20200509045010.GB4933%mail.duskware.de@localhost>,
Martin Husemann  <martin%duskware.de@localhost> wrote:
>On Fri, May 08, 2020 at 10:24:43PM +0000, maya%NetBSD.org@localhost wrote:
>> The indirection only applies to the first call. The magic is within
>> rtld.
>
>You are comparing PLT calls with ifunc (where even normal PLT calls have
>initial resolution overhead, but very tiny - while ifuncs may have
>arbitrary first time overhead).
>
>Thor is talking about something like the LOCK prefix in x86 asm (that
>is: real atomic ops).
>
>Kamil pointed out that the difference is known to the program at compile
>time or run time.
>
>Problem is that the is-not-lockless case will be rarely tested (if at all).
>
>Kamil said we should leave that trouble to the programmers and not make 
>decisions for them.
>
>Everyone is kind of right here, it is a matter of balancing "costs" and
>benefits.
>
>What I still not get is why it would be hard (or bad) to just have a
>pkgsrc version (or two, one gcc one llvm) and what costs that
>would bring (and I don't buy Kamil's "it is an integrated part of
>the toolchain" argument).

I am with Martin here. This belongs in pkgsrc and not in base. There is
an overhead using libatomic and we should not be penalizing everyone.
There are very few cases where applications should be using raw atomics,
and in these situations the applications should be making concious
architectural choices.

christos