Fail when reference counts are inconsistent.

[1uc]

In HighFive we have two types of invalid HIDs. The first category consists of:

• H5I_INVALID_HID which HighFive uses to indicate that the object is "empty",
• H5P_DEFAULT (aka 0) which is to "optimize" the default property.

The second category consists of anything else that H5Iis_valid reports as invalid, e.g. identifiers that have been closed, etc.

For the first category we must not increment or decrement the reference counters. However, for the second category, we'd like to get an error, because we're expecting either H5I_INVALID_HID, H5P_DEFAULT or a valid HID, but are given something else.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (61fa102) 84.74% compared to head (a738b47) 84.74%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #857   +/-   ##
=======================================
  Coverage   84.74%   84.74%           
=======================================
  Files          69       69           
  Lines        5053     5053           
=======================================
  Hits         4282     4282           
  Misses        771      771           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[1uc]

It can uncover bugs in user code that stem from something like:

HighFive::File* file = malloc(sizeof(HighFive::File));
// Let's make it explicit that the allocated space might not be
// zeroed out:
memset(file, 'f');

*file = File("foo.h5");

because the _hid in file (with quite high probability) is an invalid, positive HID the assignment used to not decrement the junk HID. However, the new version will error out. The bug is already present in such user code, because if they hit a valid HID by chance, then the reference counts will be off.

[1uc]

Reverting this back to draft to think more carefully about how to have a balance between not crashing, but still exposing existing bugs in user programs, see #857 (comment)

[1uc]

Since HDF5 will print an error message, that's essentially what we'd do as well. (We can't throw from the dtor for example.)