Avoidable floating point save/restore?
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
I found this from comp.os.minix (actually part of a MINIX FAQ):
=====CUT=====
From: [email protected] (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-----
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.
Quoted 65 lines · John Richard Moser
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1I found this from comp.os.minix (actually part of a MINIX FAQ):
=====CUT=====
From: [email protected] (Kees J Bot)
Subject: Re: MMX/3DNow support was RE: MINIX Development?
Date: Wed, 23 Jul 2003 20:15:03 +0200This 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_FPUto 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.orgiQIVAwUBRUYgNAs1xW0HCTEFAQJnqBAAgMjqTSA0OC6ThwUSoG4lO0uFCdUERJKl 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-----
-
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Ah, okay. That still leaves me wtf'ing at the stuff on comp.os.minix but maybe what you said is what they meant. ;)
Quoted 39 lines · Phillip Susi
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:
I found this from comp.os.minix (actually part of a MINIX FAQ):=====CUT=====
From: [email protected] (Kees J Bot)
Subject: Re: MMX/3DNow support was RE: MINIX Development?
Date: Wed, 23 Jul 2003 20:15:03 +0200This 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_FPUto the next release, by default off. I don't care about MMX, that's way too exotic for Minix.
=====CUT=====
Quoted 9 lines
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
iQIVAwUBRUZZRgs1xW0HCTEFAQJgeBAAlxMT3Sh4Bhq9Kp7bCEVLv/aGJSSBDVBe C/3KkU7u+CWX/Hbxhm1wiuIlWQ9PMKAFWArdZSM8TBUGa6No8e6HKhNOqfXDTUFA JkEpdQaewVcIK8PhPyS8cPu4DA+I7+3R3B0RKAhbDZTc8uyoG2jMwMNgmFMqGmGe +q50YuKVDP2iE7HuDCV5Ae1KLdL1VspXCZqOl0mmbPIhV89Lp2I2VZIbOhcq4oYx lWnUB7xACuWzdGP4kVIxGcZb9rqFOyJWTyaBwPuzcYQs7TVNFPRYCNoV++Rsv2BJ M576LILPOvxaJv6ticsia8r02HuEF4Yh1ZcArCuhO6wbJ3gjmCdynpea0dh46/Xx OzmAGnXeLgM+erNgP8ZT9m2SAtSpsYfx2gbbafcpQVk3UcnX0Z+l/DpfDDurdnHt zL2koT04M5HXSriicBB94BBtBZk09pXWvqQ26CJ3/vPHnNK1JDJwL6kW1kX2sWuc 7+rs2HRCTZIa2O+uaPgw8ZrlunuqCgV5oueZhpH1BHa2JjIExD0QrhhGUMMXIlyT xD54ghFFk/tndtAdwVR12S8mukQdkSUTxa8aHWqyoVS3XIe860bLuocyVrMGtdBs VYXNqz3dGFQhj5vpcYAaGPzKlTSDyzGQtubd9mWfqVjZ1HLZYcD+eozpXAtEIAIR lKdlVNqpqME=
=uZ51
-----END PGP SIGNATURE-----