cvs commit: src/sys/i386/isa npx.c
davidxu at freebsd.org
Sun May 28 16:21:05 PDT 2006
On Sunday 28 May 2006 21:58, Bruce Evans wrote:
> OK if it is correct to not copy the environment.
I will revert changes, I think I should keep it safe as before.
> BTW, your other patch that masks the top 16 bits in mxcsr needs to be
> extended to sometimes mask 0x40 (DAZ). The amd64 arch manual says to
> mask with mxcsr_mask which is in the fxsave image. Cleared bits in
> mxcsr_mask indicate reserved bits in a future-proof way. Unfortunately,
> this isn't past-proof so the following complications are needed: if
> the CPU stores 0 in mxcsr_mask, it means that mxcsr_mask is not
> supported; old CPUs (at least amd64 ones) that don't support mxcsr_mask
> also don't support DAZ, so the correct mask for them is 0xFFBF. These
> complications are implemented in Linux's i386 so I suppose they are
> no-ops at worst for non-amd64 CPUs. Linux stores the mask in a global
> variable. I think this breaks the asymmetric SMP case more than
process migrates from one CPU to another, I don't think you can use such
CPUs in same system if they are different in this case.
> FreeBSD's amd64 uses mxcsr_mask without fixing up the case
> where it is 0. So does Linux-2.6.10's x86_64 (it Initializes the
> global variable but doesn't use it).
> Behaviour of mxcsr and DAZ on some CPUs:
> AXP2600 (Barton): mxcsr_mask = 0; attempting to set DAZ traps
> P3-800 (freefall): same as AXP2600
> A64 (Id0xf48 stepping 8): mxcsr_mask = 0xFFFF; setting DAZ works
> > } else
> > +#endif
I haven't tried DAZ bit, will you fix it ?
> vm_machdep.c is missing a DEV_NPX ifdef around the call here. Of course,
> lots of things would break if the option to not use npx were exercised.
> CPU_ENABLE_SSE is a silly option too.
> > state->sv_87.sv_env.en_cw = cw;
> > newtd->td_pcb->pcb_flags |= PCB_NPXINITDONE;
> > }
More information about the cvs-src