Search
Monday, 11th of June 2018, 5:07:24 UTC
11:07:25
corci
Project sbcl-master-windows build #2674: FIXED in 37 min: http://ci.cor-lab.de/job/sbcl-master-windows/2674/
11:31:34
corci
Project sbcl-master build #3463: FIXED in 1 hr 3 min: http://ci.cor-lab.de/job/sbcl-master/3463/
11:37:41
corci
Project sbcl-master-windows build #2675: FAILURE in 30 min: http://ci.cor-lab.de/job/sbcl-master-windows/2675/
11:38:18
stassats`
scymtym: something's wrong with SB-REGALLOC::VERTICES-BEST-COLOR/GENERAL on x86
11:38:48
stassats`
it doesn't do any work on some input, returning NIL
11:40:24
scymtym
stassats`: do you have a way of reproducing the issue? not just a self-build i assume?
11:40:41
stassats`
trying to reduce it further
11:44:24
stassats`
scymtym: (defun f (a) (declare ((simple-array (complex double-float)) a)) (* (aref a 0) (conjugate (aref a 0))))
11:44:34
scymtym
stassats`: the only way i see for this to happen is COLORS being 0. could you add an ASSERT or AVER guarding against that? (assuming you have a way to quickly rebuild - i'm currently spinning up a vm)
11:46:03
stassats`
but do-sc-locations never runs
11:48:14
scymtym
can you paste the macroexpansion?
11:48:54
stassats`
would it be different from yours?
11:50:54
stassats`
170 has all oddly set bits
11:51:03
scymtym
potentially, for example, it uses sb!vm:finite-sc-offset-limit at macroexpansion time
11:51:06
stassats`
so, logbitp never trues
11:56:08
stassats`
SB-VM::COMPLEX-DOUBLE-REG>
11:56:23
stassats`
that's for the tn of the first vertex
11:59:28
stassats`
but vertex domain says 85
12:00:40
stassats`
vertex-target-vertices have 170
12:00:46
stassats`
find-vertex-color has 85
12:01:15
scymtym
sorry, i have to go to a meeting now. will be back later
12:02:43
stassats`
oh, vertex-domain is computed
12:05:57
stassats`
find-vertex-color has element size 1, vertex-domain sets odd bits
12:20:46
stassats`
COMPLEX-DOUBLE-REG being 2 is a bit strangge
12:21:33
stassats`
or is vertex-element-size something different?
12:23:52
stassats`
SB-VM::COMPLEX-DOUBLE-REG indeed has 2 element-size:
12:24:48
stassats`
i guess that's something about x87
12:34:25
stassats`
it's trying to move between double-reg and complex-double-reg
12:34:35
stassats`
which on x86 have different element sizes
12:34:59
stassats`
or not move, but interact in some way
12:38:14
stassats`
18147707b54ef78c9fdbae2075705c0ea4a5558d appears to be the culprit
13:34:45
scymtym_
stassats`: maybe it can fall back to increment 1 in that situation
13:56:41
corci
Project sbcl-master build #3465: FIXED in 1 hr 2 min: http://ci.cor-lab.de/job/sbcl-master/3465/
15:35:16
stassats`
scymtym: i already fixed it
16:00:56
scymtym
stassats`: i saw. thanks. i'm starting to wonder whether the compilation speedup is worth it. despite this issue, there is the problem with SCs that specify locations in orders different from by offset, ascending
Monday, 11th of June 2018, 17:07:24 UTC