Search
Saturday, 21st of September 2019, 17:55:07 UTC
20:02:24
stassats
i think it's call_into_c after all, not call_into_lisp
20:02:51
stassats
as (alien-funcall (extern-alien "funcall0" (function (values) long)) 0) goes haywire
20:08:51
stassats
or no, that's still call_into_lisp
20:22:47
stassats
(alien-funcall (sap-alien (sb-sys:int-sap 33) (function (values)))) doesn't work either
21:30:16
stassats
"segmentation fault gdb -p 91497"
21:30:23
stassats
things are bad when even gdb segfaults
22:27:38
stassats
ok, a double whammy of untagged reg-code and 3-bit fixnum tags
22:59:07
stassats
ok, now all tests pass ppc64le, even with sb-thread
22:59:44
stassats
sb-concurrency frlock is failing badly, though
3:15:14
saturn2
kudos to whoever added SB-EXT:SEARCH-ROOTS, very nice feature
3:37:59
karlosz
all the code in the runtime.S files annoy me. i could do what stassats suggested and move pthread_specific into C code, but it's hard to test i didn't break anything on each platform
3:38:27
karlosz
theres still so much duplicated code at least between the risc backends
4:33:01
stylewarning
karlosz: is SB-THREADS a typo in FEATURES
4:33:19
stylewarning
should :SB-THREAD or :SB-THREADS or both be in *features*
4:34:48
karlosz
stylewarning: what do you mean?
4:34:55
karlosz
i think its sb-thread
4:35:02
karlosz
sb-threads doesn't exist
4:35:10
stylewarning
karlosz: idk, on PPC, BT isn't working, and SB-THREADS is in my *features*
4:35:29
karlosz
how did you build it
4:35:37
karlosz
./make.sh -with-sb-thread?
4:35:49
stylewarning
i'm doing --with-.. now
4:37:16
karlosz
that's weird, sb-threads definitely should not be anywhere
4:39:13
stylewarning
karlosz: hm, I wonder what's going on
4:39:34
karlosz
i actually did ./make-config.sh --with-sb-thread
4:39:50
karlosz
and did the whole shebang with host-1, target-1, host-2, target-2
4:40:28
stylewarning
certainly the normal build with --with-sb-thread didn't add it to features
4:40:28
karlosz
stylewarning: i don't think --fancy will give you threads on ppc64 because it isn't marked as stable somewhere
4:40:44
stylewarning
karlosz: neither does --with- for make.sh
4:40:59
karlosz
it looks like stassats was able to build sb-thread fine
4:41:07
karlosz
you're building with master?
4:41:38
stylewarning
yes, pulled latest; trying again
4:44:10
karlosz
i'm building again as well...
4:45:34
karlosz
just a dumb guess, but you happen to be invoking the right executable right?
4:45:41
karlosz
like with ./run-sbcl.sh
4:46:02
karlosz
your machine has a nonthreaded sbcl which is the path default
4:46:25
stylewarning
karlosz: good question; will check
4:46:39
stylewarning
except sbcl takes 5 years to build
4:46:53
stylewarning
karlosz: i was not using the right sbcl
4:46:59
stylewarning
idk who globally installed it; shame on them
4:47:29
karlosz
its marked mgsk from a few days ago, i ran into the same problem when it mismatched executables and cores
4:47:42
karlosz
alos, self build is quite quick for me
4:48:12
karlosz
about the same time as cross compiling over ssh
4:48:43
karlosz
i actually didn't bother stress testing it at all because i don't have threaded workloads
4:48:57
karlosz
stassats mentioned that ppc64le actually passes all the threaded tests
4:49:17
stylewarning
it passes none of the quantum compiler tests
4:49:57
karlosz
yeah, i think i just tried to see if make-thread works
4:50:14
karlosz
can you describe the general nature of the problem? like, deadlocking or race conditions?
4:50:39
karlosz
could be that special variable access is really just hosed, but then again, the regression suite passes....
4:50:47
stylewarning
karlosz: for the quantum compiler? it has nothing to do with threads, just floating point numbers going into FORTRAN
4:51:12
karlosz
oh, so it's not a problem new with the threaded build
4:51:20
karlosz
i think thers still some ffi wackiness
4:51:33
karlosz
the c abi doesn't look fully updated for 64bit
4:51:39
karlosz
on the call_into_c side of things
4:51:40
stylewarning
cffi-libffi doesn't build by default
4:52:26
stylewarning
i hacked and slashed until it worked ("worked"), so who knows whose issue it is
4:53:38
karlosz
no clue - i've used the ppc64 port for like half a minute so i don't have a good idea what still remains
5:15:08
stylewarning
oof ppc allocates complex doubles like crazy
5:23:46
stylewarning
karlosz: ok, threading seems to work, cool
5:23:54
stylewarning
pegging 64 cores with a bunch of lisp
Sunday, 22nd of September 2019, 5:55:07 UTC