|
01:50
ilbot3 joined
04:56
jepeway joined
05:35
quester joined
07:16
FROGGS joined
09:04
FROGGS joined
09:46
FROGGS joined
10:56
vendethiel joined
13:20
camelia joined
|
|||
| dalek | arVM: 2be52fd | brrt++ | src/spesh/ (7 files): Change throw_adhoc to oops in spesh Spesh routines should probably not throw catchable exceptions. |
14:00 | |
|
14:59
AndChat|228864 joined
|
|||
| AndChat|228864 | brrt: they forgot to publish the news about your grant? | 15:00 | |
|
15:20
zakharyas joined
18:25
FROGGS joined
19:40
dalek joined
19:59
brrt joined
20:01
brrt` joined
20:02
brrt` left
20:07
brrt joined
|
|||
| brrt | AndChat|228864: I suspect they either expected me to publish it, or haven't gotten around to it yet | 20:08 | |
| also, would it maybe be a nice idea if I could do a talk at YAPC::EU | |||
| timotimo | +1 | 20:11 | |
| brrt: do you have a clue what might be wrong about my BBs and preds and succs in my branch? | |||
| in "union_goto_chains" | |||
| brrt | uhm..... | ||
| no, and not at a computer that can quickly compile and run tests | 20:12 | ||
| if you have a gist i could check it out? | |||
| timotimo | ah, ok | ||
| sure, i can run it and it'll readily explode | |||
| brrt | but it's a good example of more complex than it might look :-) | ||
| timotimo | it explodes very early in nqp compilation | ||
| brrt | hmm | ||
| timotimo | gist.github.com/timo/90ea1cd5a34f80b76764 | 20:13 | |
| do you see the lack of "yes" outputs? | 20:14 | ||
| brrt | opening :-) | ||
| yes, i see it | 20:17 | ||
| that is really weird... | 20:18 | ||
| timotimo | t.h8.lv/nqp_early_explosion_with_spesh.txt | ||
| here's some more spesh so that you can compare BB addresses with the spesh log | |||
| the ordering in the code is weird | 20:20 | ||
| in the output, that is | |||
|
20:26
brrt joined
|
|||
| brrt | \o | 20:26 | |
| emacs just died | |||
|
20:27
TimToady joined
|
|||
| brrt | timotimo: evidently these predecessor arrays are empty for some reason | 20:29 | |
| that could (potentially) be due to a bug in generating the spesh graph | 20:30 | ||
| but that seems somewhat unlikely though | |||
| in fact, the spesh graph output knows all about them | 20:31 | ||
| timotimo | that's what confuses the F out of me | 20:32 | |
| and the "number of preds" should be correct, too | |||
| and i've triple-checked that i always use the pidx in a predaray and have num_preds as its bound | 20:33 | ||
| brrt | yes, and it is | ||
| timotimo | and the pred arrays don't alias amongst themselves | 20:35 | |
| maybe something else is already removing those for some reason | 20:36 | ||
| i shall put some logging into the manipulate function that removes succs and stuff | |||
| brrt | hmm | ||
| it seems unlikely to me tbh | |||
| maybe step through it with gdb | |||
| to be completely honest, i'm not sure i understand your purpose there | 20:37 | ||
| in this branch | |||
| you want to optimize jumplist, how? | |||
| timotimo | t.h8.lv/nqp_early_explosion_with_spesh.txt - has a bit more info now | 20:40 | |
| jumplist currently creates a whole bunch of BBs that consist of only a single goto instruction | |||
| brrt | hmm, and? | ||
| timotimo | that increases code size | ||
| i don't actually know how big a goto is | |||
| if you think it's not worth much, i'll just drop it for the time being | 20:41 | ||
| but then i have no good idea what to do next j) | |||
| ;) | |||
| brrt | ok. wait a minute :-) | 20:42 | |
| you want to replace our goto's with what, exactly? | |||
| i'm not trying to be mean, i just don't understand it | 20:43 | ||
| anyway, you can say succ[sidx]->pred[succ[sidx]->num_pred] = NULL as succ[sidcx]->pred[k] = NULL; succ[sidx]->num_pred = k-1; | 20:54 | ||
| and i would say just take a pointer to the successor you're looking at :-) | 20:55 | ||
| ok, so your basic way of doing is this | 20:56 | ||
| take a jumplist | |||
| all those bbs that simply go to their direct predecessors can be removed | |||
| right? but that... will not necessarily ever work, because jumplist depends on a numeric argument, and you're not changing that number, are you? | 20:57 | ||
| integer | |||
|
20:59
brrt` joined
|
|||
| brrt` | uhm, how much of my last rambling actually came through :-) | 21:00 | |
| also note that the JIT compilation of the jumplist depends on all the basic blocks containing goto's | 21:01 | ||
| and that they have to be aligned, even | |||
| jumplist is very much a space for time tradeoff :-) | 21:02 | ||
| brrt | oh wow | 21:04 | |
| timotimo: if you're unsure what to do, i can tell you that deopt bridges would be really beneficial to a): JIT in case of a better code generator (my grant project) and b): spesh trace compilation | 21:08 | ||
| neither necessarily essential for any (no pressure), but both would benefit tremendously | 21:09 | ||
| i'm not entirely sure any of my irc messages are coming through.. :-) | 21:12 | ||
| timotimo | sorry | ||
| i got a bunch of your messages | |||
| want me to gist you what i got? | |||
| oooooooooooh; could it be succ[sidx]->pred[succ[sidx]->num_pred] = NULL was out of bounds anyway? | 21:13 | ||
| oh, if the jit compilation of jumplist requires a list of BBs that are all gotos, i'd break it with my "optimization" | 21:14 | ||
| that's not good | |||
| thanks for the hint | |||
| nwc10 | .tell jnthn is there a bug at line 377 of src/6model/containers.c -- should that HASH_ADD_KEYPTR() call be inside the if (!entry) block? | 22:22 | |
|
22:54
quester left
|
|||