Pandas packing on Windows missing necessary .dlls

[bjones1]

This is a followup on #1580. On my Windows PC (see below), I see failures in tests/functional/test_libraries.py::test_pandas_extension[onedir] (and possibly onefile -- I'm solving one failure at a time). Investigating, running test test reports Intel MKL FATAL ERROR: Cannot load mkl_intel_thread.dll. Copying from site-packages\numpy\core\*.dll to the dist\ directory fixes this. (Just copying the requested .dll by itself doesn't.)

I'm assuming that PyInstaller can't analyze DLL dependencies -- @matysek, can you confirm or deny? Based on this, I can write a Windows-specific hook. Does anyone see this failure on other platforms (meaning hooks for other OSes)?

Version info:

__main__.py                 80 INFO     PyInstaller: 3.2.dev0+g5e80840.mod
__main__.py                 81 INFO     Python: 3.4.4
__main__.py                 82 INFO     Platform: Windows-10-10.0.10240

[bjones1]

Just looking at the log answers my question: yes, PyInstaller does walk DLL dependencies. For example:

bindepend.py               211 DEBUG    Analyzing c:\users\bjones\downloads\winpython-32bit-3.4.4.1qt5\python-3.4.4\lib\site-packages\numpy\core\mkl_rt.dll
bindepend.py               514 DEBUG    Skipping KERNEL32.dll dependency of mkl_rt.dll
bindepend.py               211 DEBUG    Analyzing c:\users\bjones\downloads\winpython-32bit-3.4.4.1qt5\python-3.4.4\Lib\site-packages\numpy\core\libmmd.dll
bindepend.py               514 DEBUG    Skipping KERNEL32.dll dependency of libmmd.dll

I'm not sure yet why it doesn't find all the dependencies. Still looking. depend.bindepend.getfullnameof at least generates the correct path on my PC.

[bjones1]

It looks like mkl_rt.dll, libmmd.dll, and libifcoremd.dll are found, but nothing else. Can someone recommend a good dependency walker tool? Neither dependency walker nor PE Explorer seemed to be able to find the additional dependencies. Or perhaps this needs to be added to a hook?

[codewarrior0]

Intel MKL FATAL ERROR: Cannot load mkl_intel_thread.dll

I'm certain this means that the DLL in question is being loaded at runtime via LoadLibrary because the error message is unique to MKL. If it were loaded at load time, you would see a Python ImportError with "The specified DLL could not be found".

We can't scan for DLL loads done via LoadLibrary, so the only way out is a hook.

[bjones1]

Good point -- thanks. I'll look at some hooking.

[bjones1]

Closed by 08020fd.