Change nanobind build options to remove -Os

[jorgensd]

With gcc, -Os (recommended by nanobind) can lead to significantly slower binaries than -O2. DOLFINx has considerable templated code, which gets compiled by the nanobind wrappers. This makes performance of the wrapper library sensitive to the compiler options.

This improves the run-time of #2891 to be better than in 0.7.1

Fixes #2891.

[IgorBaratta]

I confirm that the changes fix #2891 for me .

But:

Size optimizations can be disabled by specifying the optional NOMINSIZE argument, though doing so is not recommended.

from https://nanobind.readthedocs.io/en/latest/api_cmake.html

Is it then a bug in nanobind?

[chrisrichardson]

Do we need MODULE?

[chrisrichardson]

@IgorBaratta - not a bug in nanobind, but a feature of the compiler. Each compiler may treat -Os differently.

[jorgensd]

Do we need MODULE?

Reading the docs Igor linked to,

nanobind_add_module() performs the following steps to produce bindings.

It creates a CMake library via add_library(target_name MODULE ...) and enables the use of C++17 features during compilation.

[garth-wells]

With macOS/clang, the difference between -Os and -O3 is negligible for the code in #2891 (comment).

[jorgensd]

With macOS/clang, the difference between -Os and -O3 is negligible for the code in #2891 (comment).

I clearly has a large effect on linux with gcc.

[chrisrichardson]

add_library needs MODULE - I don't think nanobind_add_module does.

[jorgensd]

add_library needs MODULE - I don't think nanobind_add_module does.

Feel free to test it. I'm currently rewriting dolfinx_mpc to support nanobind