On Wed, 2001-10-03 at 05:44, Wojciech Puchar wrote:
> > Date: Tue, 02 Oct 2001 20:54:44 -0400
> > From: David Forrai <d.forrai@ieee.org>
> > To: NetBSD i386 Port <port-i386@netbsd.org>
> > Subject: DUMMY_NOPS
> >
> > The following option is available in the kernel config file:
> >
> > # eliminate delay no-ops in I/O; recommended on all but very old
> > machines
> > options         DUMMY_NOPS
> >
> > I interpret the comment as telling me I should enable (uncomment as
> > shown) this option in all but very old machines.  Am I correct about
> > this?
> yes.
> anyway i've never seen machines which don't work with this option.
> including 386sx.
> 
> shouldn't DUMMY_NOPS be default?

To quote an acquaintance of mine: the plural of "anecdote" is not
"data".  There are definitely machines on which this used to lose.

BTW, you'd also get some (slight, but larger) performance gain from
ICU_SPECIAL_MASK_MODE, which I added ~8 years ago.  In theory this
should work on all recent machines.

I've often thought that we should have a separate kernel configuration
for "Pentium or newer" machines.  We could do both of these, as well as
disable a lot of the 486 workarounds in copyout*(), etc, and get some
improvement.  Maybe even have it use different GCC optimizations
("-march=pentium"?).  Anyone care enough to do it?