This is the mail archive of the
mailing list for the GCC project.
Re: floating point sse & sse2
- From: Christoph Hellwig <hch at caldera dot de>
- To: George dot R dot Goffe at seagate dot com
- Cc: rkm at usol dot com, gcc at gcc dot gnu dot org
- Date: Fri, 1 Feb 2002 14:58:00 +0100
- Subject: Re: floating point sse & sse2
- References: <OF852DA5AA.D342D65A-ON88256B53.004A2B61@notes.seagate.com>
On Fri, Feb 01, 2002 at 05:31:17AM -0800, George.R.Goffe@seagate.com wrote:
> As I understand it, the kernel would have to save the contents (state) of
> processes during a context switch out and restore them after a return
> context switch. If the kernel doesn't "know" about the extra regs then they
> WILL NOT BE restored at "switch in" time which is NOT GOOD. I've seen this
> in the mainframe world with new processor architectures. The result is caos
> since every process would "see" every other processes regs. IBM calls this
> "unpredictible results" and they're right.
The kernel _does_ save SSE registers. This is the main reason why
current 2.2 and 2.4 kernel do task switches much slower than early 2.2.
Of course it doesn't work. We've performed a software upgrade.