Messages in this thread | | | Subject | Re: [PATCH] [RFC] [2.5 i386] GCC 3.1 -march support, PPRO_FENCE reduction, prefetch fixes and other CPU-related changes | From | Luca Barbieri <> | Date | 04 Aug 2002 22:43:34 +0200 |
| |
On Sun, 2002-08-04 at 23:54, Alan Cox wrote: > On Sun, 2002-08-04 at 21:23, Luca Barbieri wrote: > > Added, with the exception that sfence is only used if CONFIG_X86_OOSTORE > > is not defined (currently never). > > Ok sorry I follow what you are doing. What I don't understand is why you > are generating unneeded sfence/mfence instructions in the other cases ? It was my fault: I explained it incorrectly. sfence is only used if both CONFIG_X86_OOSTORE and CONFIG_MMXEXT are set, which currently never happens with the existing processors.
> When we use MMX/SSE we need the view to be consistent anyway so the > various copying routines already handle this internally. That's why sfence is not used unless CONFIG_X86_OOSTORE (and CONFIG_X86_MMXEXT) is defined. mfence and lfence instead replace the "lock; addl $0,0(%%esp)". Is this wrong?
[unhandled content-type:application/pgp-signature] | |