Update to nightly-2025-07-07.

[nnethercote]

• We now get warnings about unnecessary parentheses around dyn types.
• Spelling of sf.name.prefer_remapped_unconditionaly was fixed (now ends in lly).
• Pointer::into_parts has been replaced with Pointer::prov_and_relative_offset.

[nnethercote]

Currently getting failures like this:

 thread 'rustc' panicked at /home/runner/work/rust-gpu/rust-gpu/target/release/build/rustc_codegen_spirv-7aed03f54af259ed/out/pqp_cg_ssa/src/mir/operand.rs:355:13:
assertion `left == right` failed
  left: Size(4 bytes)
 right: Size(16 bytes)

I will bisect locally to work out what's happening.

EDIT: I am no longer getting this error since #380 merged. It's now a different error, see below.

[nnethercote]

Nightly 2025-07-05 is the last good one; Nightly 2025-07-06 is the first bad one.

The rust-lang/rust/ commit range between those two is e3843659e9f65f589d184d1221ac6149d5fa07b5..5adb489a8034f7b56b29f3b28af0813c866f679c.

I don't know if it's possible to bisect more finely than that because I don't know how to build rust-gpu with a locally built compiler. Having said that, the likely PR causing the problem is one of:

• Block SIMD in transmute_immediate; delete OperandValueKind rust-lang/rust#143410
• Allow enum and union literals to also create SSA values rust-lang/rust#138759

Because they are the only two PRs in the range that modify rustc_codegen_ssa, which is likely to affect rust-gpu.

I suspect it's the former (143410) because that's about SIMD and the failing assertion has this comment:

            // codegen_transmute_operand doesn't support SIMD, but since the previous
            // check handled ZSTs, the only possible field access into something SIMD
            // is to the `non_1zst_field` that's the same SIMD. (Other things, even
            // just padding, would change the wrapper's representation type.)
            assert_eq!(field.size, self.layout.size);

i.e. it's some kind of SIMD padding issue.

I'm out of my depth here, particularly on the rust-gpu side but even on the rustc side. @eddyb, do you know how to fix this?

[nazar-pc]

#380 replaces #[repr(simd)] with #[rust_gpu::vector::v1]. Does it help with this PR?

[nnethercote]

I did a whole lot of follow-up on this today. The problem identified above turned out to be a rustc bug and the change was later reverted, but then there were other problems with later nightlys as well and it got very complicated. But in short, yes, landing #380 will probably help.

[nazar-pc]

#380 is in 🎉

[nnethercote]

I have updated, now trying to update to 2025-08-04, but I have cargo compiletest failures:

failures:
    [ui] ui/arch/atomic_i_increment.rs
    [ui] ui/arch/index_unchecked.rs
    [ui] ui/lang/core/unwrap_or.rs
    [ui] ui/storage_class/runtime_descriptor_array.rs
    [ui] ui/storage_class/typed-buffer-simple.rs
    [ui] ui/storage_class/typed-buffer-unbound-struct.rs
    [ui] ui/storage_class/typed-buffer-unbound.rs
    [ui] ui/storage_class/typed_buffer.rs
    [ui] ui/storage_class/typed_buffer_descriptor_array.rs
    [ui] ui/storage_class/typed_buffer_descriptor_array_slice.rs
    [ui] ui/storage_class/typed_buffer_slice.rs

Lots of errors like this:

error: cannot cast between pointer types
       from `*void`
         to `*[u32]`
   --> /home/njn/.rustup/toolchains/nightly-2025-08-04-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/mem/maybe_uninit.rs:332:6

[nnethercote]

@eddyb: to summarize here:

• I could update to Nightly 2025-07-05 successful, but Nightly 2025-07-06 caused problems.
• I think Block SIMD in transmute_immediate; delete OperandValueKind rust-lang/rust#143410 is the problematic rustc PR.
• Some of the code from that PR has since been changed, possibly due to bugs. E.g. Let codegen_transmute_operand just handle everything rust-lang/rust#143860. So I tried updating to as late as 2025-08-04, but I still get the same failures.
• tests/compiletests/ui/lang/core/unwrap_or.rs might be a good test failure to start with. The stderr diff shows that a OpCompositeInsert and OpCompositeExtract op are no longer in the output. This might also relate to the "cannot cast between pointer types" failures.