The overhead of loading import files is noticable

[sjamaan]

As reported by CaptainRant via IRC, the overhead in loading a .import.scm is quite a lot. This is an issue for CaptainRant in a project that involves loading plugins into a running binary.

Attached is a simple test case that compares loading a full module versus a simple "unit" style dynamic object.

The biggest overhead is in the call to "eval" in the generated import file. This can be ripped out with no consequence to the loading of the plugin; this may be due to the simplicity of the plugin or really due to the fact that there's no code running at toplevel which requires the evaled import. In any case, this should be improved.

[sjamaan]

If we're going to change import stuff for CHICKEN 5, might as well take this into consideration

[sjamaan]

Can be done without user-visible changes, I think, so no reason to delay 5.0 for it.

[sjamaan]

module benchmark, chicken 5 version

[sjamaan]

This could be caused by merge-se

[sjamaan]

I think this is fixed with the patch for #1457:

Without the patch:

$ ./benchmark-loader
>>> loading 'plugin.so'
0.052s CPU time, 3197/867 mutations (total/tracked), 3/85 GCs (major/minor), maximum live heap: 639.83 KiB
>>> loading 'unit-plugin.so'
0s CPU time, 61/22 mutations (total/tracked), 0/1 GCs (major/minor), maximum live heap: 640.27 KiB

With the patch:

>>> loading 'plugin.so'
0s CPU time, 85/44 mutations (total/tracked), 0/1 GCs (major/minor), maximum live heap: 301.5 KiB
>>> loading 'unit-plugin.so'
0.004s CPU time, 61/22 mutations (total/tracked), 0/1 GCs (major/minor), maximum live heap: 301.94 KiB

So we can close this when accepted.

[sjamaan]

Fixed with 36be0fc