Search
Sunday, 17th of June 2018, 6:28:54 UTC
11:47:59
stassats`
playing with regalloc, ironclad build itme 18 seconds => 10 seconds
12:03:44
stassats`
now the hard part, writing comments and a commit message explaining it
12:26:44
stassats`
now i feel like that recoloring array is not needed and where the most time is spent now
12:28:54
stassats`
i could avoid computing it if there's no neighbor intersections
13:34:31
corci
Project sbcl-master build #3486: FAILURE in 1 hr 6 min: http://ci.cor-lab.de/job/sbcl-master/3486/
14:42:07
stassats`
avoiding computing the count map if there's no intersections makes building ironclad 9.4 seconds, i guess there's a lot of intersections
16:45:16
stassats`
now it looks like SSET-MEMBER is dominating, which is where i started, trying to handle collisions better
16:45:57
stassats`
but it looks more that find-dominators reaches 300 hops in SSET-ADJOIN when building ironclad
16:46:30
stassats`
now trying a vector + bit vector for regalloc incidence sets
17:22:37
stassats`
ok, iterating over that bit vector + universe is too slow
17:53:21
stassats`
looks like i'm out of ideas on how to optimize regalloc further
18:15:49
karlosz
is iterative pack as fast as default reg alloc now?
Sunday, 17th of June 2018, 18:28:54 UTC