update to rustc 1.99.0

[Firestar99]

Requires #612 #640 #642

Warning

This is an in-the-middle nightly

Notes:

• seeing a couple weird compiletest failures, I suspect something changed upstream...

[nazar-pc]

Further bump would be nice, nightly-2026-07-22 introduced macroless_generic_const_args that improved const generics substantially over earlier nightlies.

And for me personally unmerged PRs are not that useful since I can't rely on commits that can disappear any time.

[Firestar99]

afaik github never deletes commits (except when explicitly requested), so any hashes you grab from a repo will always remain valid, even if this PR is rebased.

[Firestar99]

RESOLVED The version I just pushed is completely broken after nightly-2025-07-19 (inclusive)

Upstream commit rust-lang/rust@d93df7f is causing the breakage, even though it looks completely innocent cleanup. Instead of asking the codegen backend for whether some value is an immediate or scalar pair, it removes those APIs to instead just query it directly from BackendRepr. The implementation we and llvm and gcc had is the exact same one they're using, so there's seemingly no semantic change there.

But all compiles are failing with an ICE here:

e(SpirvValue { kind: Def(1137), ty: 1134 }) @ TyAndLayout { ty: math::libm_math::support::big::u256, layout: Layout { size: Size(32 bytes), align: AbiAlign { abi: Align(8 bytes) }, backend_repr: ScalarPair { a: Initialized { value: Int(I128, false), valid_range: 0..=340282366920938463463374607431768211455 }, b: Initialized { value: Int(I128, false), valid_range: 0..=340282366920938463463374607431768211455 }, b_offset: Size(16 bytes) }, fields: Arbitrary { offsets: [Size(0 bytes), Size(16 bytes)], in_memory_order: [0, 1] }, largest_niche: None, uninhabited: false, variants: Single { index: 0 }, max_repr_align: None, unadjusted_abi_align: Align(8 bytes), randomization_seed: 17996504725788008022 } })): not applicable
  --> /home/firestar99/.rustup/toolchains/nightly-2026-07-19-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/compiler-builtins/compiler-builtins/src/math/../../../libm/src/math/support/big.rs:28:5
   |
28 | /     pub fn signed(self) -> i256 {
29 | |         i256 {
30 | |             lo: self.lo,
31 | |             hi: self.hi as i128,
32 | |         }
33 | |     }
   | |_____^


thread 'rustc' (819536) panicked at /home/firestar99/workspace/rust-gpu/target/release/build/rustc_codegen_spirv-d246d51479a59b87/out/pqp_cg_ssa/src/mir/operand.rs:423:21:
Box<dyn Any>
stack backtrace:
[...]
  16:     0x75e402e148e8 - rustc_middle[96c442875f2cb62c]::util::bug::span_bug_fmt::<rustc_span[157f53c51de8f998]::span_encoding::Span>
  17:     0x75e3f2d0019b - <rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::mir::operand::OperandRef<rustc_codegen_spirv[36c5a3c0a874027]::builder_spirv::SpirvValue>>::extract_field::<rustc_codegen_spirv[36c5a3c0a874027]::builder::Builder>
  18:     0x75e3f2a8f808 - <rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::mir::FunctionCx<rustc_codegen_spirv[36c5a3c0a874027]::builder::Builder>>::codegen_consume
  19:     0x75e3f2a9011c - <rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::mir::FunctionCx<rustc_codegen_spirv[36c5a3c0a874027]::builder::Builder>>::codegen_operand
  20:     0x75e3f2a79262 - rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::mir::codegen_mir::<rustc_codegen_spirv[36c5a3c0a874027]::builder::Builder>
  21:     0x75e3f2af90f1 - rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::base::codegen_instance::<rustc_codegen_spirv[36c5a3c0a874027]::builder::Builder>
  22:     0x75e3f29b2827 - <rustc_middle[96c442875f2cb62c]::mono::MonoItem as rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::mono_item::MonoItemExt>::define::<rustc_codegen_spirv[36c5a3c0a874027]::builder::Builder>
  23:     0x75e3f2b8aea5 - <rustc_codegen_spirv[36c5a3c0a874027]::SpirvCodegenBackend as rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::traits::backend::ExtraBackendMethods>::compile_codegen_unit::{closure#0}
  24:     0x75e3f2b8cd1d - <rustc_codegen_spirv[36c5a3c0a874027]::SpirvCodegenBackend as rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::traits::backend::ExtraBackendMethods>::compile_codegen_unit
  25:     0x75e3f2af895b - rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::base::codegen_crate::<rustc_codegen_spirv[36c5a3c0a874027]::SpirvCodegenBackend, rspirv[3ee83ac1199cf24a]::dr::constructs::Module>
  26:     0x75e3f2b8b4cd - <rustc_codegen_spirv[36c5a3c0a874027]::SpirvCodegenBackend as rustc_codegen_spirv[36c5a3c0a874027]::maybe_pqp_cg_ssa::traits::backend::CodegenBackend>::codegen_crate
  27:     0x75e406819ca4 - <rustc_interface[bb04609e4c960450]::queries::Linker>::codegen_and_build_linker
  28:     0x75e40690470e - rustc_interface[bb04609e4c960450]::interface::run_compiler::<(), rustc_driver_impl[c2a64f8504684822]::run_compiler::{closure#0}>::{closure#2}
  29:     0x75e406972ae4 - std[2a8c3cd46992084f]::sys::backtrace::__rust_begin_short_backtrace::<rustc_interface[bb04609e4c960450]::util::run_in_thread_with_globals<rustc_interface[bb04609e4c960450]::util::run_in_thread_pool_with_globals<rustc_interface[bb04609e4c960450]::interface::run_compiler<(), rustc_driver_impl[c2a64f8504684822]::run_compiler::{closure#0}>::{closure#2}, ()>::{closure#0}, ()>::{closure#0}::{closure#0}, ()>
  30:     0x75e4069728bb - <std[2a8c3cd46992084f]::thread::lifecycle::spawn_unchecked<rustc_interface[bb04609e4c960450]::util::run_in_thread_with_globals<rustc_interface[bb04609e4c960450]::util::run_in_thread_pool_with_globals<rustc_interface[bb04609e4c960450]::interface::run_compiler<(), rustc_driver_impl[c2a64f8504684822]::run_compiler::{closure#0}>::{closure#2}, ()>::{closure#0}, ()>::{closure#0}::{closure#0}, ()>::{closure#1} as core[f0447a45d83adb47]::ops::function::FnOnce<()>>::call_once::{shim:vtable#0}
  31:     0x75e40697cec8 - <std[2a8c3cd46992084f]::sys::thread::unix::Thread>::new::thread_start
  32:     0x75e40029dd53 - start_thread
  33:     0x75e40032563c - __clone3
  34:                0x0 - <unknown>

This function wasn't changed in the mentioned commit. But if you look at the match statement:

let (in_scalar, imm) = match (self.val, self.layout.backend_repr) {
    // Extract a scalar component from a pair.
    (
        OperandValue::Pair(a_llval, b_llval),
        BackendRepr::ScalarPair { a, b, b_offset },
    ) => {

The issue is that we have an OperandValue::Pair with a BackendRepr::Immediate due to us patching scalar pairs to immediates:

// FIXME(eddyb) detect `#[rust_gpu::vector::v1]` more specifically,
// to avoid affecting anything should actually be passed as a pair.
if let PassMode::Pair(..) = arg.mode {
    // HACK(eddyb) this avoids breaking e.g. `&[T]` pairs.
    if let TyKind::Adt(..) = arg.layout.ty.kind() {
        arg.mode = PassMode::Direct(ArgAttributes::new());
    }
}

EDIT: was added in PR #437

Removing that makes it "work" again, with around 50 compiletest failing on function calls, so removing this isn't trivial. But may very well be necessary going forward, as we're hit problems due to this patching many times.

So going forward we could:

• try to patch all the cases where cg_ssa is failing due to our patch, who knows how many more patches it'll be in the future?
• I have yet to pin point the exact reason it's starting to fail to patch just that
• remove the abi patching and make our codegen backend without it

[Firestar99]

@nazar-pc fixed the issues, feel free to give it a try!

I know compiletests are failing, with the new build dir structure our compiletest infra can't find the dependent libs anymore and needs to be adjusted. I've been able to test everything apart from the last commit in this PR, there may be dragons!

[Firestar99]

nightly-2026-07-30 is the last version using the old build dir, CI should pass for it

[nazar-pc]

There is (for now) __CARGO_TEMPORARY_BUILD_DIR_NEW_LAYOUT_OPT_OUT=1 if you want to opt out in newer nightlies. I had to use it due to some incompatibilities with Miri.

[Firestar99]

I've seen that flag in rust-lang/rust's history, not sure how long they'll support that flag. Updating to the new layout shouldn't be too hard anyway.

[nazar-pc]

Just tried e720714, it worked for me