Add faster TocTree._toctree_copy method

[hofmandl1]

This was pair-programmed with @ax-lothas

As in the standalone html builder the navigation is flattened out for every single html page, the code needs to create a specialized toctree for every html page.

Previously this was done by deep-copying the complete navigation toctree and then stripping out the parts not needed on the particular page.

With this change the code only (deep)-copies the needed parts of the toctree avoiding unnecessary copying+throwing-away

The performance improvements seems to be smaller for smaller page counts and get bigger the more pages are involved:

+----------------------------------+-----------+-----------+-------------------+
|         830 Source Files         |   5.3.0   |  5.3.0mod |       Ratio       |
+----------------------------------+-----------+-----------+-------------------+
|          SingleThreaded          |  49.137 s |  43.517 s | 0.885625903087287 |
| Parallel (auto, 16 logical CPUs) |  13.334 s |  12.849 s | 0.963626818659067 |
+----------------------------------+-----------+-----------+-------------------+

+----------------------------------+-----------+-----------+-------------------+
|        6166 Source Files         |   5.3.0   |  5.3.0mod |       Ratio       |
+----------------------------------+-----------+-----------+-------------------+
|          SingleThreaded          | 907.497 s | 521.945 s | 0.575147906825036 |
| Parallel (auto, 16 logical CPUs) | 195.400 s | 150.139 s | 0.768367451381781 |
+----------------------------------+-----------+-----------+-------------------+

Relates

• slow running sphinx in large project #4576

[AA-Turner]

What do you mean by 'consistent' here?

A

[hofmandl1]

While _toctree_prune deletes unneeded stuff from a (previously done) deep copy of the toctree, _toctree_copy does the deep copy itself but only copies the parts, that _toctree_prune would not delete. So assuming, that _toctree_prune would be adjusted in order to e.g. not delete a particular node type, then _toctree_copy would also need to be adjusted to copy that particular node type.

TBH we didn't fully understand the different node types involved here. It would probably be best, to completely get rid of _toctree_prune in order not to have this consistency problem any more (maybe it is not that big of a problem if there aren't any adjustments to that code done in the foreseeable future). We tried getting rid of _toctree_prune but found that it is used in another place and didn't feel confident enough to do the necessary changes.

[AA-Turner]

Thanks!

A