Inital zig build support and integrating it with the CI

[Avokadoen]

This PR adds a build.zig + build.zig.zon. It also add jobs in the CI to build with zig for windows and linux

This means there is now support for building box2d as a lib and enabling any zig build users (including other C/C++ projects) to trivially call zig fetch --save git+https://github.com/erincatto/box2d and start using it!

TODOs:

• Building as lib for zig fetch
• Building unit tests
• option BOX2D_DISABLE_SIMD
• option BOX2D_AVX2
• option BOX2D_DOCS
• option BOX2D_PROFILE

Out of PR scope?
option BOX2D_SAMPLES
option BOX2D_BENCHMARKS

[Avokadoen]

Before I go ahead and do additional changes I would like some input from you @erincatto:

1. Are you still interested in hosting a build.zig for box2d?
2. If yes, would it require supporting building samples and tests, or simply serve as a utility for zig users to use box2d as a library?

[erincatto]

Update: given the problems with enkiTS and my lack of knowledge of zig, it is probably best if the zig script does the minimal thing that helps users while still compiling the core library in CI.

[Avokadoen]

Ok. I'll make this PR do the following when I have time:

1. Allow zig users fetch and link with box2d
2. Test building on windows and Linux

I have the unit tests compiling and passing with the exception of deterministic one. I'll raise an issue in enkiTS to add support for zig build as well. Fixing the remaining test is trivial at that point.

[Avokadoen]

Created enkiTS#142

[Avokadoen]

dougbinks/enkiTS#143

[Avokadoen]

Converting to draft until enkiTS PR is merged

[Avokadoen]

Seems like the built determinism tests fails on both:

ENSURE( data.sleepStep == EXPECTED_SLEEP_STEP );
ENSURE( data.hash == EXPECTED_HASH );

[Avokadoen]

@erincatto are you suuuuure the test is deterministic? 😅

Not quite sure how a build script can cause a race condition in these tests. Am I maybe missing some flags or macros?

[erincatto]

Does zig have some flags for fast math? If the compiler introduces contractions (e.g. fused multiply-add) then determinism will fail. This needs to pass single-threaded before considering multi-threaded determinism.

[Avokadoen]

Issue was that I was missing -ffp-contract=off for the shared lib! 🤦

[Avokadoen]

Not sure this is desired, or avx2 should always be explicitly enabled.
Current logic will always enabled avx2 if features is supported by compile target, it will always be false if not supported.

[Avokadoen]

@erincatto I think this is ready, before you take a look. These are the following options I did not implement:

• BOX2D_DOCS
• BOX2D_PROFILE
• BOX2D_SAMPLES
• BOX2D_BENCHMARKS

It seems like users can lean on cmake for these. I doesnt make sense to support everything for both build infrastructures in my mind:

I also changed the CI to utilize a matrix strategy to remove some of the boilerplate in the .github folder

[Avokadoen]

Sorry for the spam 😮‍💨 (was wrestling github ci config)

[erincatto]

Can the number of CI steps be reduced and still have reasonable confirmation of compatibility?

[Avokadoen]

Yes, each permutation gets its own run in the UI.
I have a test run on my fork which should be viewable to anyone: https://github.com/Avokadoen/box2d/actions/runs/18271466024

[erincatto]

Can we reduce the number of stages to 1 or 2?

Imagine 10 years from now I have Odin, Swift, MinGW, Zig, Jai, etc. If I have 100 stages this is going to be quite unpleasant.

[Avokadoen]

Sure I'll get to it later today. I'll reduce it to 1 windows + 1 Linux target

Imagine 10 years from now I have Odin, Swift, MinGW, Zig, Jai, etc. If I have 100 stages this is going to be quite unpleasant.

You forgot to mention rust!

[Avokadoen]

Reduced zig testing to two jobs: https://github.com/Avokadoen/box2d/actions/runs/18657086820