Hi Mathias, On Mon, Dec 13, 2021 at 5:32 PM Mathias Krause <[email protected]> wrote: > > Am 13.12.21 um 15:37 schrieb Jason A. Donenfeld: > > On Mon, Dec 13, 2021 at 3:33 PM Mathias Krause <[email protected]> > > wrote: > >> yeah, don't worry. We can keep this change downstream in grsec. Dunno if > >> Jason wants to take it as well, his call. It was just a side observation > >> which came out of our need to read and understand the code to provide a > >> workaround for the gcc bug we were hitting. > > > > I suppose I can load it up in kbench9000 to see if it makes a > > difference. If it doesn't matter much, I'd prefer sticking with the > > formally verified stuff. But if there is a nice speedup, then I guess > > we can revisit more handwavy "obviously this is the same thing" > > arguments. > > Had to hack main.c and run.sh a little, but here are some numbers: > > root@box:~# ./run.sh > [+] Setting no-turbo to status 1 > [+] Setting non-boot CPUs to status 0 > [+] Inserting module to run tests > insmod: ERROR: could not insert module kbench9000.ko: Unknown symbol in > module > [+] Gathering results > ever64: 115100 cycles per call > ever64_out_r: 115080 cycles per call > ever64_out_rm: 113957 cycles per call > [+] Setting non-boot CPUs to status 1 > [+] Setting no-turbo to status 0 > > Slightly faster.
Huh, that's actually a pretty nice speedup for just changing register allocation... What CPU is this? Jason
