ARM64 Darwin Fixes

[ppoage]

To be edited~

• Code is formatted (go fmt ./...)
• Linter passes (golangci-lint run)
• All tests pass with race detector (go test -race ./...)
• Benchmarks don't regress (FFI overhead < 200ns)
• New code has tests (minimum 70% coverage, current: 89.1%)
• Platform-specific code tested on target OS
• Assembly changes validated on real hardware
• Documentation updated (if applicable)
• Commit messages follow conventions
• No sensitive data (credentials, tokens, etc.)

Note: I added some checks and panics as we were experiencing very hard to pin down issues with segfaults, silent failures that caused later corruptions, and race conditions. At one point, running with -race actually stopped the segfault.

[kolkov]

Hey @ppoage, thanks for the comprehensive work on ARM64!

Quick note: we merged v0.3.6 on Dec 29 with similar HFA/sret fixes, so there's a struct layout conflict now:

v0.3.6 layout:

• r8 (X8 sret) at offset 184 (end of struct)
• Separate fr1-fr4 for D0-D3 returns (offsets 152-176)

Your layout:

• sret at offset 72 (shifts all float offsets)
• Reuses f1-f4 for returns

What we'd like to adopt from your PR:

• r2 return (X1) — we're missing this for 9-16 byte structs
• darwin example update

Issues:

• .DS_Store file — should not be committed
• Inline HFA detection in call_unix.go — we use classification.go

Options:

1. Rebase on current main and adapt to v0.3.6 struct layout
2. Or we can cherry-pick the r2 handling into v0.3.7

Preferred: Option 1 — please rebase on main, adapt to our struct layout, and force-push. This keeps the fix properly attributed to you.

Note: This is preliminary feedback — I'm still out of town without my laptop. Will do a deeper review when I'm back.

Will review wgpu/naga/gogpu PRs next — the type-safe ObjC messaging looks critical and great!

[ppoage]

Ok, rebased.

I have some extra changes in this rebase, but most should be necessary. Once the initial fix was made with the objc cffi (HFA/etc), I had to fix nested struct handling because that was the root cause for why we couldn't send or receive window size. And then I was seeing some weird things between floats/ints so I added mixed support.

FYI BenchmarkGoffiStringOutput segfaults on mac. It looks like it calls with the string as a pointer. Didn't want to modify your benchmark, but it does run if that's fixed.

Edit: Also when trying to add textures to metal, discovered oversized structs (size >16) aren't being handled properly. I fixed that and added a test for coverage (which now passes). That will be coming in a future commit

[ppoage]

Ok, I fixed the oversized struct handling. Also added an assembly shim for use in testing to verify the whole chain through the ffi call doesn't mangle/drop something.

FYI, the current ffi doesn't accept the stack based extra arguments. I'm seeing if I can add the texture pipeline for metal, but the current ABI implementation isn't complete and I don't know what approach you prefer to take here.

[kolkov]

Thanks for the excellent rebase! I've done a deep review and the PR looks great.

Key improvements I verified:

• Struct layout compatible with v0.3.6 (r8@184, fr1-fr4@152-176)
• r2 (X1) return for 9-16 byte structs - this was missing in v0.3.6
• uint64 bit patterns instead of float64 - cleaner for mixed float32/float64
• Nested struct handling via placeStructRegisters()
• Mixed int/float struct support via countStructRegUsage()
• Comprehensive test coverage (darwin_objc_test.go, abi_capture_test.s)

One small fix needed for the benchmark:

In ffi/benchmark_test.go line ~152 (darwin case for BenchmarkGoffiStringOutput), the fix should be:

// args contains pointers to argument storage, not the values themselves
_ = CallFunction(cif, sym, unsafe.Pointer(&result), []unsafe.Pointer{unsafe.Pointer(&strPtr)})

Once that's fixed, I'll approve and merge. Great work on the ARM64 darwin support!

Re: stack-based extra arguments - that's a known limitation, we can address it in a future PR.

[ppoage]

Fix for the benchmark pushed!

Note: I haven't run any of the tests for other OS's, I assume they pass as ARM64 was the one getting changes

[kolkov]

Thanks for the quick fix! Benchmark fix looks correct.

Verified:

• strPtr now passed as unsafe.Pointer(&strPtr) - correct API usage
• 64 ns/op on M3 Pro - excellent performance

Approving. Will merge once CI passes.

[codecov-commenter]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

ℹ️ You can also turn on project coverage checks and project coverage reporting on Pull Request comment

Thanks for integrating Codecov - We've got you covered ☂️