Longint fixes

[gpwright]

Between version 1.9.178 and 1.9.222 of HDF5 the type hid_t was changed from int to long int. h5py now handles this appropriately by modifying it's own hid_t cython declaration based on the HDF5 version number.

Note: this changeset fixes issue #599 and issue #616

[gpwright]

Btw I'm pushing this to master but can we cherry-pick it onto the 2.5 branch since that's what I'm actually developing on? Should I create a separate pull request or can it be done as part of this one?

[andrewcollette]

I have some lingering concerns about this. Specifically, as I recall we never relied on the exact type involved in the "ctypedef int hid_t" declaration; in fact, this bit from the Cython docs (number 3 in the second list) specifically states that the exact type should not matter:

http://docs.cython.org/src/userguide/external_C_code.html#referencing-c-header-files

In other words, "hid_t" should always be what shows up in the Cython output so the "int" or "long int" should not be visible. Then it will map to whatever int/long int is declared by HDF5.

Any idea of how to address this? One approach might be to run Cython with the "int" definition, and then the "long int" definition, and simply diff the output files. Then we could see where the exact type was "leaking" into the Cython output.

[gpwright]

I understand your point, particularly the section in the cython docs that says "will work okay whatever the actual size of a word is (provided the header file defines it correctly)".

Moreover I just ran the diff as you suggested on builds with and without the typedef change, and the diff showed no changes to the builds except for the typedef and some unknown changes to the binary data inside the egg file.

I just verified for my own sanity that by reverting back to typedef int and trying to "import h5py" I am back to the original problem "Value Error: Not a property list class".

I have little or no knowledge of cython but my explanation would be that at runtime, python is instantiating int objects rather than long objects based on the cython type. The python int object is then assigned a long and overflows.

[andrewcollette]

Interesting! So what happens if we simply use "ctypedef long int" on both HDF5 1.8 and 1.10?

[andrewcollette]

(ctypedef long int hid_t, that is)

[gpwright]

I built h5py against HDF5 1.8 with ctypedef long int, and it seems fine. I can import h5py, and run some simple tests of my own, and all the built-in tests are passing.

Version 1.10 I couldn't find online... has this been released yet?

So are you thinking to skip the version check and just always use long?

[andrewcollette]

Sorry, 1.9 is the branch that will become 1.10.

Since the choice of type evidently doesn't affect the Cython-generated C files, I would suggest removing the version check and just using long int everywhere.

[gpwright]

I removed the version check

[andrewcollette]

Looks good. Thanks again for tracking this down!