Repository navigation
Description
Since #672 (v0.11.0) a shader does not compile if it uses an enum without repr whose payload is 16-byte aligned. Option<Vec4>, Mat4::try_inverse and Vec4::try_normalize fail with:
error: unsupported constant of type `u128`
With i128:128 in the data layout rustc gives such an enum a u128 tag, and the backend does not support u128 values.
Which direction do you prefer?
- Go back to the 8-byte alignment of
i128/u128. It undoes Canu128/i128alignment be made16? #665. - Keep the 16-byte alignment and lower the
u128tag to supported integer types.rustc_codegen_ssastores, loads and converts the tag as a plainu128, so I think it's theu128part of Integer polyfills #307.
I'll be happy to send a PR for 1.
Reproducer, affected cases, bisect
tests/compiletests/ui/option_vec4.rs:
// build-pass
#![no_std]
use spirv_std::glam::Vec4;
use spirv_std::spirv;
#[spirv(fragment)]
pub fn main(input: Vec4, output: &mut Vec4) {
let o = if input.x > 0.0 { Some(input) } else { None };
*output = o.unwrap_or(Vec4::ZERO);
}To reproduce:
cargo run -p compiletests --release --no-default-features --features use-installed-tools -- --target-env vulkan1.2 option_vec4
Error:
error: unsupported constant of type `u128`
|
note: used from within `option_vec4::main`
--> <ui/option_vec4.rs>:9:53
|
LL | let o = if input.x > 0.0 { Some(input) } else { None };
| ^^^^
note: called by `main`
--> <ui/option_vec4.rs>:8:8
|
LL | pub fn main(input: Vec4, output: &mut Vec4) {
| ^^^^
The same error comes from:
Mat4::try_inverse,Vec4::try_normalizeandVec3A::try_normalize(at theNoneinsideglam);Option<Quat>;- a user enum
enum Hit { Miss, At(Vec4) }; Option<A>with#[repr(align(16))] struct A(f32);, so it does not needglam.
Option<Vec3>, Mat3::try_inverse, Vec3::try_normalize and the Hit enum with #[repr(u32)] build.
All the cases above build at 02770fb and fail at its child c4c5e15 (#672), with the same toolchain.
The tag width comes from Integer::for_align in rustc_abi: an enum without repr(C) or repr(<int>) gets a tag as wide as the smallest alignment among the first fields of its variants, when an integer type with that size and alignment exists. The error comes from def_constant_cx at
rust-gpu/crates/rustc_codegen_spirv/src/builder_spirv.rs
Lines 598 to 610 in 60d1e25
| let const_op = |builder: &mut Builder, op, lhs, maybe_rhs: Option<_>| { | |
| // HACK(eddyb) remove after `OpSpecConstantOp` support gets added to SPIR-T. | |
| let spirt_has_const_op = false; | |
| if !spirt_has_const_op { | |
| let zombie = builder.undef(ty, None); | |
| cx.zombie_with_span( | |
| zombie, | |
| DUMMY_SP, | |
| &format!("unsupported constant of type `{}`", cx.debug_type(ty)), | |
| ); | |
| return zombie; | |
| } |
#684 removes the spirt_has_const_op hack that prints this error, but the SPIR-T validator still rejects OpTypeInt 128. So I expect the error to change to "u128 type unsupported in SPIR-V" there (read from the code, I did not build it).
- rust-gpu:
mainat 60d1e25 - Rust:
nightly-2026-08-15(rustc 1.99.0-nightly (d453bdd8f 2026-08-14)) - Target:
spirv-unknown-vulkan1.2 - OS: Debian (Linux, aarch64)
- SPIR-V: SPIRV-Tools v2026.4 unknown hash, 2026-10-02T11:54:22+00:00
I found it by running glam's own tests in a shader, with the help of an AI coding agent.
Activity
Firestar99 commentedon Nov 16, 2026
rust-gpu is quite messy at it's core and doesn't properly support unions, thus no rust-style enums, thus no
Option. It just happens that sometimes rustc optimizes away the enum before we see it, and a few months ago we've had a PR that as a side-effect made some parts of enums work.Sadly, we're still at the stance that rust-like enums are unsupported, and to properly support them, we'd likely have to rewrite significant parts of rust-gpu.
greenhat commentedon Nov 16, 2026
@Firestar99 Got it. Thank you for the explanation!