On Thu, Jul 11, 2002 at 10:49:55AM +0100, Jon Burgess wrote:
> I'm fairly sure i've seen comments that say that cache manipulation code
> should be run uncached. My current thought is that it is probably safe to
> manipulate the d-cache with cached code, but I-cache manipulation which
> could invalidate the cacheline containing the currently executing code
> really should be run uncached. I think the CPU probably then skips
> instructions until it gets to the next cacheline, what effect this has
> depends on how the instructions are aligned relative to the cacheline.
> Given how hit-and-miss this is I am suspicous that this problem could
> simply be lurking undiscovered.
>
> The patch below makes the I-Cache routines run via kseg1, it is a bit
> ugly but seems to work. I have not measured the performance impact of
> this patch.
Have you tried to insert a large number of nops instead? Or preferably,
how about replacing the __restore_flags() in your example with the
following piece of inline assembler:
__asm__ __volatile__("mtc0\t%0, $12" ::"r" (flags) : "memory");
Ralf
