[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-kernel
Subject:    Re: [PATCH v3] tty: tty_io: remove hung_up_tty_fops
From:       "Paul E. McKenney" <paulmck () kernel ! org>
Date:       2024-05-04 18:18:43
Message-ID: 7aab574e-b6ba-4038-802f-2d64af09d802 () paulmck-laptop
[Download RAW message or body]

On Sat, May 04, 2024 at 10:50:29AM -0700, Linus Torvalds wrote:
> On Fri, 3 May 2024 at 22:08, Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > You are right, this is going to need some arch-specific code for a few
> > of the architectures.  Hey, I was hoping!!!
> >
> > The compilers do not currently optimize these things, but things appear
> > to me to be heading in that direction.
> 
> Ok, so it sounds like right now it makes no sense - presumably
> __atomic_load_n() doesn't actually generate better code than
> READ_ONCE() does as-is, and we have the issue with having to make it
> per-architecture anyway.
> 
> But maybe in a couple of years we can revisit this when / if it
> actually generates better code and is more widely applicable.

Completely agreed.

Here is my current thoughts for possible optimizations of non-volatile
memory_order_relaxed atomics:

o	Loads from the same variable that can legitimately be
	reordered to be adjacent to one another can be fused
	into a single load.

o	Stores to the same variable that can legitimately be
	reordered to be adjacent to one another can be replaced
	by the last store in the series.

o	Loads and stores may not be invented.

o	The only way that a computation based on the value from
	a given load can instead use some other load is if the
	two loads are fused into a single load.

Anything that I am missing?

							Thanx, Paul

[prev in list] [next in list] [prev in thread] [next in thread]