Re: 8.0 performance issue when running build.sh?

[Maxime Villard <max%m00nbsd.net@localhost>]

Le 06/07/2018 à 16:48, Martin Husemann a écrit :
> On Fri, Jul 06, 2018 at 04:40:48PM +0200, Maxime Villard wrote:
> > > This are all successfull builds of HEAD for alpha that happened after June 1:
> >
> > What does that mean, are you building something *on* an Alpha CPU, or are
> > you building the Alpha port on another CPU?
>
> It is all about the releng auto build cluster, which is running amd64.

Then it is likely caused by two things:

 * in EagerFPU I made a mistake initially, and it caused the FPU state to
   be restored when the kernel was leaving a softint. I sent a pullup
   already for netbsd-8, but it hasn't yet been applied. The fix removes a
   save+restore, which improves performance.

 * the XSTATE_BV bitmap is not zeroed at execve time in NetBSD-8. It
   probably causes some performance loss, because XRSTOR always restores
   each FPU state instead of just the ones we used. In recent CPUs there
   are many FPU states and we use only a few in the base system, so the
   extra restoring costs us. Even more so with EagerFPU, I guess, because
   we do save+restore unconditionally, rather than on-demand. I fixed that
   in October 2017 in -current, but it didn't make it to -8. I guess it
   will have to, now.

Beyond that we need to use XSAVEOPT, too.

Maxime