freenode/#sbcl - IRC Chatlog
Search
17:18:47
White_Flame
I have a large ancient codebase that uses the equivalent of (defconstant *varname* ...)
17:19:53
White_Flame
I redefined defconstant with (locally (declare (sb-ext:muffle-conditions style-warning sb-kernel:asterisks-around-constant-variable-name)) (cl:defconstant ... yet I still get the STYLE-WARNINGS printing out into the repl buffer even if they're not reported as warnings at the end
17:23:36
pfdietz
SBCL build time on my machine is down by 15 seconds, from 1:30 to 1:15. I wonder what caused that.
17:24:46
scymtym
pfdietz: candidates are improved LZ-compression of debug information and improved pretty-printing of names within the compiler
18:08:29
stassats
well, just means you shouldn't ship it that way, but who cares about a declaration or two for development
18:08:37
joshe
would that only effect efficiency of compiled code, or are there correctness implications?
18:35:23
stassats
would be interesting to know how long does it take to build sbcl on one of the new amd chips, with 12 or 16, or even 32 cores
18:42:23
stassats
building sbcl is probably the only thing where i can benefit from more cores, but if it just goes from 50 seconds to 30 it's probably not worth it
18:44:44
stassats
should have better single-core performance too, but the next year should be even faster
18:46:38
stassats
(built with safety 0, with sb-fasteval to get make-host-1 to build in parallel too)
19:21:11
stassats
it would be more useful to time make-host-2 for scaling, although all the serial bits are important as well
19:22:32
joshe
I thought timing a parallel build under clisp would make the differences more obvious, but now I'm bisecting to find the last commit that broke clisp hosts
19:23:01
stassats
SBCL_MAKE_JOBS=-j160 => 120% cpu 5:42.97 total + SBCL_MAKE_PARALLEL=160 => 532% cpu 2:26.69 total
19:35:09
stassats
160 1222%/23.407, 80 1256%/24.769, 40 1276%/24.696, 20 1173%/25.796, 10 737%/ 33.087, 5 443%/49.413
19:40:50
stassats
so it does seem to scale to 12 cores, but since that's the only part of the build that scales well, the total times get little improvement