Yes, this has been measured. Have a look at the two papers on HDTrans (Shapiro, Sridhar, and Bungale). The only time you need an I-cache flush is when the translation buffer gets flushed entirely. So with a 4M translation buffer, you flush after every 4M of translated code (not to be confused with *executed* code).
On Sun, Oct 20, 2013 at 6:04 PM, Sandro Magi <[email protected]> wrote: > Re: dynamic translation for read barriers, my initial impression is > skepticism that the improved branch prediction a dynamically swapped read > barrier would outweigh the cost of i-cache flushes. Has this sort of design > actually been measured? > > Sandro
_______________________________________________ bitc-dev mailing list [email protected] http://www.coyotos.org/mailman/listinfo/bitc-dev
