Date: Thu, 23 Aug 2001 14:03:44 -0500 (CDT) From: Kevin Day <toasty@temphost.dragondata.com> To: peter@wemm.org (Peter Wemm) Cc: dwmalone@maths.tcd.ie (David Malone), hackers@FreeBSD.ORG Subject: Re: XMM[0-7] preserved across context switch? Message-ID: <200108231903.f7NJ3i722007@temphost.dragondata.com> In-Reply-To: <20010823044124.12EE8380F@overcee.netplex.com.au> from "Peter Wemm" at Aug 22, 2001 09:41:24 PM
next in thread | previous in thread | raw e-mail | index | archive | help
> Yes, but the question was "how is it preserved"? The SSE stuff works the > same as the FPU stuff in that it is switched lazily. See npxsave() and > where it is called. If a process "attaches" to the fpu, its state is kept > in the fpu the whole time. It is not extracted at context switch time. > So, we can be running a different process while the fpu/xmm stuff is holding > the original process's context. If the new process tries to use the SSE/fpu > stuff, a trap happens, we save the original process's context in the original > pcb and then give ownership to the new process. > > And for SMP, we handle it differently again. :-/ Ok, that's what I expected. Has anyone looked at using MOVDQA/MOVDQU or MOVNTDQ for faster bcopy/bzero routines? Someone at Intel suggested they may be faster than using the FPU regs.... -- Kevin To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200108231903.f7NJ3i722007>