StackOverflowError in summarysize during misc test

[nalimilan]

I see new failures when building Fedora/RHEL RPM nightlies. These have bee, introduced since d47f24b (from Dec. 22).

Error in testset misc:
Error During Test
  Test threw an exception of type StackOverflowError
  Expression: $(Expr(:escape, :(summarysize(Core)))) $(Expr(:escape, :>)) $(Expr(:escape, :(summarysize(Core.Inference)))) $(Expr(:escape, :>)) $(Expr(:escape, :(Core.sizeof(Core))))
  StackOverflowError:
Error in testset misc:
Error During Test
  Test threw an exception of type StackOverflowError
  Expression: summarysize(Base) > 10000 * sizeof(Int)
  StackOverflowError:
Error in testset ambiguous:
Error During Test
  Expression evaluated to non-Boolean
  Expression: ((x->begin 
        isempty(x) || println(x)
    end))(detect_ambiguities(Core,Base; imported=true))
       Value: nothing

https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-25-x86_64/00492486-julia/build.log.gz

[colbec]

Also seen in openSUSE Leap 42.2.

[tkelman]

Does this happen in a default source build, or only when you're building against some system dependencies? Scroll up to where the ambiguity test ran and it should be printing the list of ambiguous methods.

[colbec]

Not sure what you mean by "list of ambiguous methods." In:

Julia Version 0.6.0-dev.1698
Commit 4793e88* (2016-12-25 07:38 UTC)
Platform Info:
  OS: Linux (x86_64-suse-linux)
  CPU: Intel(R) Core(TM) i5-4460  CPU @ 3.20GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, haswell)

terminal output that mentions "ambiguous" (selectively) is:

threads (2)               |   38.54  |  1.34  |  3.5 | 83.29      | 2158.79 
error (2)                 |    0.64  |  0.00  |  0.0 |  2.63      | 2158.79 
WARNING: Method definition ambig(Any, Integer) in module TestMain_ambiguous at /home/colin/Downloads/julia6/test/ambiguous.jl:7 overwritten at /home/colin/Downloads/julia6/test/ambiguous.jl:84.
boundscheck (5)           |    2.45  |  0.00  |  0.0 |  4.55      | 1654.04 
cartesian (5)             |    0.04  |  0.00  |  0.0 |  0.10      | 1656.18 
	From worker 2:	Skipping Base.<|
	From worker 2:	Skipping Base.>:
	From worker 2:	Skipping Base.active_repl
	From worker 2:	Skipping Base.active_repl_backend
asmvariant (5)            |    0.05  |  0.00  |  0.0 |  1.33      | 1658.32 
osutils (5)               |    0.02  |  0.00  |  0.0 |  0.24      | 1660.92 
ambiguous (2)             |    1.01  |  0.05  |  5.1 | 33.64      | 2158.79 
channels (5)              |    8.77  |  0.75  |  8.6 | 446.92     | 1877.34 
dsp (5)                   |    6.67  |  0.00  |  0.0 | 319.38     | 1877.34 
fft (2)                   |   16.24  |  0.79  |  4.9 | 767.73     | 2158.79 
examples (5)              |   12.14  |  0.00  |  0.0 | 679.32     | 2485.93 
parallel (3)              |   62.43  |  0.00  |  0.0 |  1.43      | 1100.48 
cmdlineargs (4)           |  102.76  |  0.05  |  0.1 | 50.24      | 1121.77 
	From worker 1:	Worker 3 failed running test misc:
Some tests did not pass: 1279452 passed, 0 failed, 2 errored, 0 broken.
 in finish at ./test.jl:524
 in macro expansion; at ./util.jl:288 [inlined]
 in macro expansion; at /home/colin/Downloads/julia6/test/testdefs.jl:0 [inlined]
 in anonymous at ./<missing>:?
 in eval at ./boot.jl:236
 in runtests at /home/colin/Downloads/julia6/test/testdefs.jl:16
 in runtests at /home/colin/Downloads/julia6/test/testdefs.jl:4
 in #586 at ./multi.jl:1455
 in run_work_thunk at ./multi.jl:1004
 in macro expansion at ./multi.jl:1455 [inlined]
 in #585 at ./event.jl:66misc: Error During Test
  Test threw an exception of type StackOverflowError
  Expression: $(Expr(:escape, :(summarysize(Core)))) $(Expr(:escape, :>)) $(Expr(:escape, :(summarysize(Core.Inference)))) $(Expr(:escape, :>)) $(Expr(:escape, :(Core.sizeof(Core))))
  StackOverflowError:
misc: Error During Test
  Test threw an exception of type StackOverflowError
  Expression: summarysize(Base) > 10000 * sizeof(Int)
  StackOverflowError:

Test Summary:           |     Pass  Error  Broken     Total
  Overall               | 20030733      2      29  20030764

followed by:

osutils               |       21                       21
ambiguous             |       42                       42
channels              |       32                       32

Let me know where I should be looking.

[tkelman]

Looks like you only get the stack overflow then, not the Expression evaluated to non-Boolean from the ambiguity test. Does the stack overflow happen if you just run make test-misc on its own?

[colbec]

I have seen the non-Boolean error as well.
Both make test-misc and make test-ambiguous run successfully without error when run by themselves.

[tkelman]

Ah, the ambiguity is reproducible when running math and ambiguous on the same worker, it was caused by #19627 and probably just needs a *(::MockQuantity, ::Bool) method to be defined. I can't reproduce the misc failure on ubuntu with make testall1 so we'll need to keep debugging that. What does Base.summarysize(Base) return for you from a normal REPL? If you add @show summarysize(Base) around line 189 of test/misc.jl, what value does it output for test-misc by itself vs a failing run?

[colbec]

julia> Base.summarysize(Base)
41035484

I inserted the @show statement after line with import Base.summarysize

> make test-misc
    JULIA test/misc
Test (Worker) | Time (s) | GC (s) | GC % | Alloc (MB) | RSS (MB)
summarysize(Base) = 41465530
misc (1)      |    8.42  |  0.86  | 10.3 | 1177.03    | 363.44  

Test Summary: |    Pass    Total
  Overall     | 1279454  1279454
    SUCCESS

Strangely, I can't see the string "summarysize" anywhere in the output from make testall which consistently fails with the same output.

[tkelman]

probably because the show statement is triggering the stack overflow too, but only if some subset of tests have run previously on the same worker. running JULIA_CPU_CORES=1 ./julia test/runtests.jl foo bar etc with the list of tests that ran previously on the same worker as misc should reproduce it too? And possibly allow some reducing of the test case to a smaller subset?

[colbec]

OK here is my progress: thinking that the error might have been triggered by something recent in the misc worker I ran

> JULIA_CPU_CORES=1 julia6 test/runtests.jl cmdlineargs enums markdown nullable grisu llvmcall repl show broadcast goto reflection mpfr pollfd resolve version read spawn math copy blas vecelement simdloop intfuncs random offsetarray staged iobuffer hashing dict triplequote char printf sparse/sparse

which, disappointingly, run successfully. I may be shooting myself in the foot by testing in roughly the reverse order of testall, I don't know. I'll keep expanding the list.

[tkelman]

it should include misc at the end to reproduce the stack overflow, and may be order dependent

[colbec]

Right, thanks for noticing the missing misc - rather important. Adding the full complement of tests which ran on the same worker and misc at the end triggers the failure. Order does not seem to be important since my tests ran in reverse order. I will now go back through previous tests and report back.

[colbec]

Ok this is getting complicated.
> JULIA_CPU_CORES=1 julia6 test/runtests.jl math copy blas vecelement simdloop intfuncs random offsetarray staged iobuffer hashing dict triplequote char printf sparse/sparse linalg/matmul linalg/svd linalg/tridiag linalg/pinv linalg/givens linalg/arnoldi linalg/cholesky linalg/hessenberg misc
produces the error but
> JULIA_CPU_CORES=1 julia6 test/runtests.jl cmdlineargs enums markdown nullable grisu llvmcall repl show broadcast goto reflection mpfr pollfd resolve version read spawn math misc
does not.
So clearly I have to split the first (which failed), and I did this on the sparse test. Then ran both sides and both ran to success.
Please correct my thinking, which is that maybe there are two or more tests, at least one from each of both sides of the sparse test, which in combination with misc triggers the failure.