32-bit Windows test failure

[stevengj]

See the Appveyor test, and the discussion in JuliaIO/HDF5.jl#179 ; cc @tkelman

The 64-bit Windows c-blosc binary was built with x86_64-w64-mingw32-gcc (GCC) 4.6.3 using x86_64-w64-mingw32-gcc -shared -O3 -msse2 -I. *.c -o libblosc64-1.5.0.dll. And the 32-bit binary was built with i686-w64-mingw32-gcc (GCC) 4.6.3 using i686-w64-mingw32-gcc -shared -O3 -msse2 -I. *.c -o libblosc32-1.5.0.dll.

[tkelman]

Sorry, got distracted by something.

See http://github-multi-status.herokuapp.com/ for how to enable the multi-status indicators on PR's, if you haven't yet. The way you have Blosc.jl's deps/build.jl right now I need to manually delete libblosc.dll if I want to switch between 32 and 64 bit versions on the same computer, which is a bit annoying. But yes, I can reproduce this locally. I'll try building a dll with (edit: up-to-date) mingw-w64 and see if anything changes.

[tkelman]

Okay, no significant difference with a libblosc.dll made with i686-w64-mingw32-gcc 4.8.2 (https://dl.dropboxusercontent.com/u/8244638/libblosc32.dll), sorry for the wrong guess.

The failure is at @test roundtrip(a) for ty == Float64, i == 32. At the repl, a=rand(Float64,32); Blosc.decompress(Float64, Blosc.compress(a)) == a works fine. But compiled into the roundtrip function it segfaults.

[tkelman]

here's some @code_llvm output, if it helps https://gist.github.com/tkelman/556fb9c10e9048e426bf

[stevengj]

That's weird; does it work in Julia 0.4?

[tkelman]

Nope, crashes there too

[tkelman]

If you want to run a bigger matrix of stable and nightly you can do

environment:
  matrix:
# Releases
  - JULIAVERSION: "stable/win32"
  - JULIAVERSION: "stable/win64"
# Nightlies
  - JULIAVERSION: "download/win32"
  - JULIAVERSION: "download/win64"

install:
# Download most recent Julia Windows binary
  - ps: (new-object net.webclient).DownloadFile($("http://status.julialang.org/"+$env:JULIAVERSION), "C:\projects\julia-binary.exe")

in appveyor.yml

[tkelman]

So what's the best way to determine whether this is a problem with blosc or a problem with base Julia?

[stevengj]

Is there a valgrind-like thing that you can run on Win32?

@vtjnash, any ideas on why a function would work in the REPL but not when compiled on Win32?

[vtjnash]

that would typically imply the JIT code had an error in the platform calling convention. i would have to walk it manually in a debugger to see what the program state was when it failed

[tkelman]

It looks like this should work now with 0.3.4, and presumably also latest nightlies. Was probably one of the stack alignment issues that @vtjnash caught. Try starting a new appveyor build.

[jakebolewski]

@vtjnash the build is passing now on Windows, thanks for all the hard work!

[stevengj]

Great!