On Tuesday, 24 February 2015 at 16:05:06 UTC, Marc Schütz wrote:
On Tuesday, 24 February 2015 at 15:14:03 UTC, Ola Fosheim
Grøstad wrote:
On Tuesday, 24 February 2015 at 14:45:17 UTC, Marc Schütz
wrote:
It's possible to (ab)use the MMU as a write barrier.
Uhm...
The throughput for L/SFENCE is 5 cycles and SFENCE 33
(Typo, MFENCE is 33...)
cycles... The cost of a page miss is 1000 cycles + overhead
for copying.
And that's assuming that you have enough memory and that the
TLB isn't affected...
Yes, but if the program has a small working set, you pay that
price only a few times (once per page), and in any case, only
during a running background scan, rather than always.
But if you can control where the code is running when you run the
GC scan then you might as well have a separate code path for
concurrent GC too (i.e. two versions of the code). One with
fences and one without...
If you loose the TLB, then you also loose all caches too IIRC.
