bitshifter / glam-rs / 32471653433

99%

Build Type

push

github

Committed by web-flow

Commit Message

feat(vec3a): add explicit padding to scalar Vec3A for bytemuck Pod/Zeroable and zerocopy IntoBytes (#786)

Fixes #122

## Problem

`Vec3A` is a 16-byte aligned type containing three `f32`s, so its scalar (`scalar-math`, and the no-SIMD fallback) representation has 4 bytes of trailing padding. That padding is uninitialized, so bytemuck could only
derive `AnyBitPattern` for scalar `Vec3A` (and `Mat3A`/`Affine3A`), not `Pod` + `Zeroable`. The SIMD backends don't have this problem because their fourth lane is a real value (`new` stores `[x, y, z, z]`).

## Change

Add an explicit `_w: f32` padding field to scalar `Vec3A`, initialized to `z` to mirror the SIMD backends' `[x, y, z, z]` storage. This means every byte is now initialized, which unlocks:

- `bytemuck::Pod` + `Zeroable` for scalar `Vec3A`, `Mat3A`, and `Affine3A` (`NoUninit` and `AnyBitPattern` come along automatically via bytemuck's blanket impls).
- `zerocopy::IntoBytes` for the same types (previously only `FromBytes`/`Immutable`/`KnownLayout`).

Scalar `Vec3A` now implements `PartialEq` manually so `_w` is ignored.

## Performance

The only added work in the materializing path is a single store that defines the padding lane. When the result is inlined and only `x`/`y`/`z` are read, that store is dead-code eliminated and the generated code is unchanged.

Probe:

```rust
#![allow(improper_ctypes_definitions)]
use glam::Vec3A;

#[no_mangle]
pub extern "C" fn probe_vec3a_add(a: Vec3A, b: Vec3A) -> Vec3A { a + b }

#[no_mangle]
pub extern "C" fn probe_vec3a_add_sum(a: Vec3A, b: Vec3A) -> f32 {
    let r = a + b;
    r.x + r.y + r.z
}
```

`probe_vec3a_add` before (`a867ca86`):

```asm
addss   xmm1, xmm3
addps   xmm0, xmm2
movss   dword ptr [rsp - 8], xmm1
movsd   xmm1, qword ptr [rsp - 8]   ; upper 4 bytes are uninitialized padding
ret
```

`probe_vec3a_add` after (`b465490b`):

```asm
addss   xmm3, xmm1
addps   xmm0, xmm2
movss   dword ptr [rsp - 8], xmm3
movss   dword ptr [rsp - 4], xmm3   ; <--... (continued)

Coverage Stats

4015 of 4025 new or added lines in 37 files covered. (99.75%)

71 existing lines in 19 files now uncovered.

91349 of 92161 relevant lines covered (99.12%)

126567.04 hits per line