Re: sunffb

On Tue, 31 May 2005, Michael wrote:

>>> We finally got the sunffb driver to work on NetBSD/sparc64,
>>> including acceleration. Turns out there was a problem in the VIS
>>> assembler code:
>>> 	 wr			%g0, 0, %fprs
>>> this will turn off the FPU so any subsequent FPU instruction will
>>> cause a trap. On NetBSD this led to huge numbers of traps which
>>> rendered the Xserver unusable when acceleration was enabled.
>>> So for NetBSD I changed this to
>>> 	 wr			%g0, 4, %fprs
>>> to leave the FPU active and clean only the dirty bits. No more
>>> excessive traps and things work.

>> I'm inclined to simply change both occurrences.

> At least for NetBSD we'll definitely do so in our repository.

>>> Since I highly doubt this is an oversight - what's the purpose of
>>> turning off the FPU after using VIS? Apparently the author expected
>>> the OS to do something which NetBSD doesn't do, but what? And why?
>>> Provoking a trap seems a rather expensive way to reset the FPU ( if
>>> that's the purpose... )

>>> PS: yes, I mailed the author too

>> Let us know if you get a response.

> He said it's supposed to keep the scheduler from storing/saving FPU
> context until someone really uses it again ( which causes a trap ), at
> least on Linux. I'm not sure this helps much in this code - the ffb/afb
> doesn't have block transfer hardware ( they dropped it because VIS is
> faster ) so the VIS code will run a lot.
> Maybe writing 0 into %fprs should be Linux-specific.

I'm not a big fan of this idea, because module binaries are supposed to be 
portable across OS'es.  In this case, there are also SunOS and the other 
*BSD's to consider.

What does NetBSD do with the trap?  Does it export it as a signal to its 
user-land?  If so, which signal?  Also, is it "precise", or can the signal 
occur a few instructions later?

Another option would be to export something like xf86{En,Dis}ableFPU() from 
os-support/ and have OS-specific implementations there.

Marc.

+----------------------------------+-----------------------------------+
|  Marc Aurele La France           |  work:   1-780-492-9310           |
|  Computing and Network Services  |  fax:    1-780-492-1729           |
|  352 General Services Building   |  email:  [email protected]          |
|  University of Alberta           +-----------------------------------+
|  Edmonton, Alberta               |                                   |
|  T6G 2H1                         |     Standard disclaimers apply    |
|  CANADA                          |                                   |
+----------------------------------+-----------------------------------+
XFree86 developer and VP.  ATI driver and X server internals.