Re: Avoidable floating point save/restore?

[Phillip Susi <psusi@cfl.rr.com>]

The FP registers only need to be saved/restored if one or both of the 
current task and the task being switched to have made use of them ( i.e. 
they ever do FP or MMX math ).  Initially FP access is disabled and the 
first time a task tries to use the FPU, a fault is triggered and the 
kernel enables the FPU for that task and sets a flag so it remembers it 
needs to save/restore the state when switching in/out of that task.


John Richard Moser wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> I found this from comp.os.minix (actually part of a MINIX FAQ):
>
> =====CUT=====
> From: kjb=733301@cs.vu.nl (Kees J Bot)
> Subject: Re: MMX/3DNow support was RE: MINIX Development?
> Date: Wed, 23 Jul 2003 20:15:03 +0200
>
> This is really a hardware floating point issue, because the MMX
> registers share the FP registers. This was done so that MMX unaware OSen
> can still support MMX programs, because when they save and restore the
> FP registers then the MMX state is also saved and restored if that
> happens to be what the FP registers are used for. This saving and
> restoring is what Minix doesn't do. So if two processes use FP/MMX then
> a context switch from one to the other will clobber the FP state of
> both. What is needed to make this work is a trap handler that reacts to
> the use of FP, so that Minix can save the FP state of the process that
> last used FP and load the FP state of the current process. On a context
> switch Minix merely sets the "don't use FP" bit in some register. Costs?
> One FP interrupt handler, some FP save/restore/setup code, some memory
> per process to store the FP state into, and some memory to store the FP
> state when a user process catches a signal. (Not sure about the signal
> business, much check with Philip.) This isn't much work, we can simply
> take Minix-vmd's code, but I haven't seen any need yet. Minix has to use
> software FP as distributed, or it won't run on your old 386, so Minix
> itself doesn't need it. Anyone here who wants to use Minix for some
> heavy number crunching? If so then I could be persuaded to add an
> ENABLE_FPU to the next release, by default off. I don't care about MMX,
> that's way too exotic for Minix.
> =====CUT=====
>
> I'm trying to make heads or tails about what in the heck is going on
> here.  It looks like they're saying you don't need to save/restore FP
> registers between context switches unless one process uses FP and the
> other uses MMX; but that doesn't make ANY sense at all.  If
> gnome-session divides 3.14/2.28 and then gimp divides 3.33/2.22 and then
> we switch BACK to gnome-session and it wants to divide the result by
> 1.92, wouldn't we need the FPU registers back in the exact state they
> were at before switching away?
>
> - --
>     We will enslave their women, eat their children and rape their
>     cattle!
>                   -- Bosc, Evil alien overlord from the fifth dimension
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.3 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
>
> iQIVAwUBRUYgNAs1xW0HCTEFAQJnqBAAgMjqTSA0OC6ThwUSoG4lO0uFCdUERJKl
> O0Mxtq9XqKOU4W7XnCtIkBF0Gy1HdMm8/L1MXLRM0FMv6jz17nNb3TGvLRC/4KKh
> vgl4KN/2D6tYYdKyh+v0n/zgVYGJ0lZn3sNQdBhxXet1Zjm6n7REFvJJD07PzIcm
> XK5vciaVHYfNb0N3SdSL7JBJdMHzrbq2DmqUD5OsVOok37e1F3S4+T3EBZy+mk/i
> D6lcx2KJzp8vIWmoskN+WLBox8fihSoNtxOK/oLComHwyJ3z1FEuILTvYqEjGUGg
> k1TdEQYhTrFCjkpLgK5lhn1hD8dZRldN4bVEt9uWU5WiRCGiykPog3wlVz7T6bQr
> SOAVKA4157kfDvG3KY55pNTxy0j8TgBWQhPyh/dTcRlzYbZuqccRcHQbYyBEGByZ
> Of7tVUnHzgNd8kuzWqVh4ULDJ7YyNj+INFegoluLi/3JcQFL/EhoUYX/QNbX1VK1
> vkrwGOneAC2KttjaH/cZHUuYfwZ1hkMmDjbkHlGmb4z56TRKOtUt03ziMDDhO4Fv
> WWjVn+CdAiOTkhuVgIQ+EpnNCSP/yRUTmxDrI3+KJaIW6zlaAZyS29KoWA+EH75K
> mJldV7amMcD+6KGVP7tHuTOpGXKYa2tCCfwtGI65Z7rWvGD6RUzz+waiT4kPav1G
> 1lObgtnarqY=
> =fw5+
> -----END PGP SIGNATURE-----
> -

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/