Fix cross-platform determinism on ARM64 NEON

[halx99]

CHANGES

1. Add windows arm64 ci
2. Fix samples build fail on windows arm64
3. Fixed DeterminismTest.CrossPlatformTest by disabling NEON FMA on native MSVC. This avoids codegen rounding differences between MSVC and other toolchains

Relative issues

• Changes to support arm64EC on MSVC. #1039
• '__asm__': undeclared identifier with VS 2022 / ARM64 #798

[erincatto]

I will only support a platform that runs with GitHub actions and passes unit tests. Just building it is not enough.

[halx99]

Finally, I fixed the DeterminismTest by disabling NEON intrinsics in the Windows ARM64 build, see ci check: https://github.com/halx99/box2d/actions/runs/23525079273/job/68476658553?pr=3

[halx99]

Finally+1, don't use NEON fma instructions solve the DeterminismTest issue

[halx99]

Why Linux ARM64 works with FMA instruction?

A: On linux arm64, the compiler translate vmlaq_f32 to FMUL + FADD, but on Windows Arm64, the msvc compiler translate to as-is aka FMLA

Verified by disassembling the same C source.

C source with NEON FMA instruction:

b2FloatW b2MulAddW( b2FloatW a, b2FloatW b, b2FloatW c )
{
	return vmlaq_f32( a, b, c );
}

Windows ARM64:

; Exported entry 249. b2MulAddW

EXPORT b2MulAddW
b2MulAddW
FMLA            V0.4S, V1.4S, V2.4S
RET
; End of function b2MulAddW

Linux ARM64:

EXPORT b2MulAddW
b2MulAddW
FMUL            V1.4S, V1.4S, V2.4S
FADD            V0.4S, V0.4S, V1.4S
RET
; End of function b2MulAddW

Android ARM64

EXPORT b2MulAddW
b2MulAddW
FMUL            V1.4S, V1.4S, V2.4S
FADD            V0.4S, V1.4S, V0.4S
RET
; End of function b2MulAddW

macOS ARM64

_b2MulAddW
FMUL            V1.4S, V1.4S, V2.4S
FADD            V0.4S, V1.4S, V0.4S
RET
; End of function _b2MulAddW

iOS ARM64

_b2MulAddW
FMUL            V1.4S, V1.4S, V2.4S
FADD            V0.4S, V1.4S, V0.4S
RET
; End of function _b2MulAddW

Tested Android ARM64, when remove -ffp-contact=off, then generated code is:

EXPORT b2MulAddW
b2MulAddW
FMLA            V0.4S, V2.4S, V1.4S
RET
; End of function b2MulAddW

[halx99]

Finally:

Root Cause of CrossPlatformTest Failure

• On GCC/Clang (Linux/Android/iOS/macOS), the compiler behavior can be controlled with
  -ffp-contract=off. This forces vmlaq_f32/vmlsq_f32 to be lowered into
  separate FMUL + FADD instructions, matching the execution order used on
  Intel SSE/AVX.

• On MSVC (Windows ARM64), the compiler always emits a fused FMLA instruction
  for these intrinsics. Even with /fp:strict, /fp:precise, or /fp:contract-,
  there is no way to disable contraction.

Conclusion:
The determinism test fails because non‑MSVC compilers can disable FMA contraction
to align with Intel’s implementation, while MSVC cannot. This difference in
compiler behavior leads to cross‑platform result mismatches.

[halx99]

I have no further changes on this PR, it’s ready for review.

[erincatto]

It is very cool you got this working. Thanks for looking into the determinism problem. I'll change the Neon code regardless of other factors.

You added 3 CI jobs for a platform I'm not sure anyone uses with Box2D. Can you justify this?

What is the real world use case for Box2D supporting this platform?

For this platform I would only want to support one CI job and it would be for unit tests, not samples. Samples take longer to build and are lower priority than unit tests.

[halx99]

You added 3 CI jobs for a platform I'm not sure anyone uses with Box2D. Can you justify this?

done

What is the real world use case for Box2D supporting this platform?

For example: Microsoft Surface Pro

[erincatto]

Implemented in #1045. Unit tests are passing.