Debugging code generation when testing with package manager

m-j-w · February 6, 2019, 11:52pm

I’m the developer of the package CpuId. I’m currently facing an issue which appears to only occur when testing the package through the package manager, viz. typing ]test CpuId. In that case, the core functions containing an llvmcall are somehow differently inlined and yield a segfault. This does not seem to occur in, say, regular operations. I have theories on why this happens, but I can’t reproduce the issue outside the package manager testing environment.

My question: How do I get a view on the generated code when testing through the package manager? (eg. @testset, @test etc.).
Thanks!

tkf · February 7, 2019, 2:41am

Not directly answering your question, but did you try to launch a Julia REPL with the same command line flags used in Pkg.jl?: Pkg.jl/Operations.jl at 853b3f1fd9895db32b402d89e9dee153b66b2316 · JuliaLang/Pkg.jl · GitHub

Also, since @test wraps code into a try block, it may help if you do so when generating the code.

Tamas_Papp · February 7, 2019, 8:17am

Pkg.test just spawns a new Julia process and runs test/runtests.jl, so you should be able to reproduce everything that way.

m-j-w · February 7, 2019, 8:48am

Thanks, @tkf and @Tamas_Papp. Found the additional option --check-bounds=yes for julia when using Pkg.test. These additional bounds checks are in fact triggering the segfaults.