Re: hppa lazy FPU switching and QEMU bugs

[Thor Lancelot Simon <tls%panix.com@localhost>]

On Wed, Aug 05, 2026 at 10:34:13AM +0200, Martin Husemann wrote:
> 
> Is the "assisted emulation" just an optimization and we could fall back
> to completely disabling FPU if running under qemu?

No, we disable the FPU after context switch as a way to implement lazy
FPU switching _on hardware_, too.  When the next process tries to use the
FPU, we get a trap and save the FPU state.  Obviously if the architecture
didn't consider any use of the FP regs to be "use of the FPU" this would be
totally unsafe, but it does.

Unfortunately QEMU lies - it claims to support coprocessor disable but it
never checks and never generates the trap.  So processes get their FP regs
switched with each other, and since *integer* multiply uses the FP regs
on PA, that's even worse than you'd expect.

The right fix is to eagerly switch the FP regs.  I have a 150 line change
with a bunch of comments that probes to see if the trap works, and if not,
always saves/restores the regs on context switch.  But it is Claude-authored
and our policy does not presently allow committing such code, so, I need to
strip it down to the comments and rewrite.  It will take me some time to do
that, particularly as using the FP regs in the kernel on HPPA involves some
trickery with GCC as we tell it we are not using the FPU when we compile.
If someone else wants this sooner, I can strip the code, leaving the comments,
and you can do the black-box exercise I would have done.

-- 
Thor Lancelot Simon	                                     tls%panix.com@localhost
  "Indeed, someone who is a citizen in a democracy is often not one in
   an oligarchy."
					- Aristotle, Politics III.1