Messages in this thread | | | Date | Tue, 04 May 2010 11:24:02 -0700 | From | "H. Peter Anvin" <> | Subject | Re: [PATCH 1/2] x86: eliminate TS_XSAVE |
| |
On 05/04/2010 11:15 AM, Suresh Siddha wrote: > On Tue, 2010-05-04 at 00:41 -0700, Avi Kivity wrote: >> On 05/04/2010 12:45 AM, H. Peter Anvin wrote: >>> >>>> I was trying to avoid a performance regression relative to the current >>>> code, as it appears that some care was taken to avoid the memory reference. >>>> >>>> I agree that it's probably negligible compared to the save/restore >>>> code. If the x86 maintainers agree as well, I'll replace it with >>>> cpu_has_xsave. >>>> >>>> >>> I asked Suresh to comment on this, since he wrote the original code. He >>> did confirm that the intent was to avoid a global memory reference. >>> >>> >> >> Ok, so you're happy with the patch as is? > > As use_xsave() is in the hot context switch path, I would like to go > with Avi's proposal. >
I would tend to agree. Saving a likely cache miss in the hot context switch path is worthwhile.
I would like to request one change, however. I would like to see the alternatives code to be:
movb $0,reg movb $1,reg
... instead of using xor (which has to be padded with NOPs, which is of course pointless since the slot is a fixed size.) I would suggest using a byte-sized variable instead of a dword-size variable to save a few bytes, too.
Once the jump label framework is integrated and has matured, I think we should consider using it to save the mov/test/jump.
-hpa
| |