Search
Tuesday, 10th of October 2017, 11:22:18 UTC
11:57:41
stassats
apparently, fails in the top-level loop in braid.lisp
16:43:27
stassats
scymtym: regalloc: Faster hiding/showing of vertex neighbors seems to be breaking ppc
16:45:08
scymtym
stassats: that will probably be difficult to figure out. should i just revert for now?
16:45:30
stassats
it is really difficult to figure out indeed
16:46:23
scymtym
what breaks? miscompiled target image?
16:48:14
scymtym
do you know which functions gets miscompiled?
16:48:27
stassats
(do-all-symbols (s)) breaks
16:48:34
stassats
whether itself or something earlier
16:48:59
stassats
well, Memory fault at 0x187e5998 (pc=0x187e5998)
16:49:44
scymtym
i made a mistake in any case. i intended no change in behavior
16:50:27
stassats
i'll debug it for some time
16:51:14
scymtym
thanks, but don't waste too much time on it, if you don't want to
16:51:17
scymtym
i will take a look at the change, think hard and then revert unless i can figure it out
16:51:23
stassats
well, iterative is not on by default, so it must be something called by do-all-symbols
16:56:55
stassats
it does iterate over some symbols
16:57:27
stassats
(do-symbols (s :cl) (print s)) too
16:58:51
stassats
that's the last symbol
17:09:48
scymtym
stassats: can you try this http://paste.lisp.org/display/358231 ?
17:19:11
stassats
scymtym: that succeeds
17:19:38
scymtym
i see some scary differences on x86_64 as well
17:20:08
scymtym
it might have just spilled everything with the broken change
17:20:46
scymtym
i wish we had a way to track changes in compiler output
17:20:57
stassats
i have an idea how to do that
17:22:00
stassats
first prerequisite, serializing IR
17:22:18
scymtym
i'm going to test some more and then push the potential fix
17:22:29
stassats
then, load up the compiler, go through all functions in SB-C, and instrument
17:22:35
stassats
record the changes with each function call
17:22:36
scymtym
so we can run later phases on canned ir1 examples?
17:22:50
stassats
then a tool that visualizes everything (the hardest part)
17:23:21
stassats
scymtym: that could be arranged as well
17:24:14
scymtym
i had some success with mcclim for visualizing compilery things
17:24:27
stassats
first, a simple IR visualizer needs to be written (i'm thinking about doing one in JS), then a time-travelling one
17:25:54
scymtym
will the time-travelling one require non-destructive phases? or would you make snapshots?
17:26:26
stassats
on each function call go through the component and find changes to the IR
17:26:52
stassats
record them and the function that made them
17:27:17
stassats
with a call tree as well
17:27:26
stassats
so, basically, a profiler
17:29:14
scymtym
could also add tracing (in the model transformation sense): from which other ir objects has a given object been derived and by which operation. but you probably thought of that
19:25:48
stassats
looks like i'm able to fix a stack-analyze problem not by deleting an aver, but by doing better dead code elimination
19:26:23
stassats
(although semantically the aver is not right)
19:33:05
stassats
but optimizing and fixing a bug in one go is nicer
Tuesday, 10th of October 2017, 23:22:18 UTC