nihalpasham · Oct 23, 2026 · Oct 22, 2026
@@ -64,6 +64,7 @@ jobs:
         env:
           CUDA_TOOLKIT_PATH: ${{ steps.cuda-toolkit.outputs.CUDA_PATH }}
           CARGO_TARGET_DIR: ${{ runner.temp }}/clippy-shared-target
+          RUSTC_BOOTSTRAP: "1"
         run: |
           set -e
           example_roots=(

@@ -44,6 +44,8 @@ jobs:
         working-directory: cuda-oxide/crates/cuda-macros/tests/device-only
 
       - name: Check example and model crate formatting
+        env:
+          RUSTC_BOOTSTRAP: "1"
         run: |
           set -euo pipefail
           # `cargo fmt --all` also formats local path dependencies. Running it

@@ -5,6 +5,9 @@ oxide = "run --package cargo-oxide --"
 
 # Environment variables for all builds
 [env]
+# Unlock `-Z` flags and `#![feature(rustc_private)]` on the pinned stable
+# release (`rust-toolchain.toml`).
+RUSTC_BOOTSTRAP = "1"
 # libmathdx installation path (for libmathdx-sys crate)
 # Required only when using --features libmathdx with mathdx_ffi_test.
 # IMPORTANT: Replace the placeholder below with your libmathdx install dir,

@@ -6,7 +6,7 @@ This dev container provides the toolchain expected by cuda-oxide:
 - CUDA Toolkit 13.0
 - LLVM 21 with NVPTX support
 - Clang 21 resource headers for `bindgen`
-- Rust `nightly-2026-10-11` with `rust-src`, `rustc-dev`, `rust-analyzer`,
+- Rust `1.99.1` with `rust-src`, `rustc-dev`, `rust-analyzer`,
   `rustfmt` and `clippy`. `llvm-tools` is not preinstalled in the image;
   rustup adds it on your first cargo command under `cuda-oxide/`, because
   `cuda-oxide/rust-toolchain.toml` names it. The image ships LLVM 21 and

@@ -17,7 +17,7 @@
       "upgradePackages": false
     },
     "ghcr.io/devcontainers/features/rust:1": {
-      "version": "nightly-2026-10-11",
+      "version": "1.99.1",
       "profile": "minimal",
       "components": "rust-src,rustc-dev,rust-analyzer,rustfmt,clippy"
     }

@@ -7,6 +7,10 @@ set windows-shell := ["powershell.exe", "-NoLogo", "-Command"]
 # Use Bash on Unix
 set shell := ["bash", "-c"]
 
+# Unlock unstable Cargo (`cargo-features = ["profile-rustflags"]`) and rustc
+# (`-Z`, `#![feature(rustc_private)]`) gates on the pinned stable toolchain.
+export RUSTC_BOOTSTRAP := "1"
+
 # Format all Rust code (SIMT workspace, codegen, examples)
 fmt:
     cargo oxide fmt

@@ -161,17 +161,17 @@ cargo oxide update
 
 From `cuda-oxide/`, `cargo oxide` works out of the box via a workspace alias.
 
-For use outside the repo (your own projects), install it with the pinned nightly toolchain:
+For use outside the repo (your own projects), install it with the pinned toolchain:
 
 ```bash
-cargo +nightly-2026-10-11 install --git https://github.com/NVIDIA/cuda-rust.git cargo-oxide
+cargo +1.99.1 install --git https://github.com/NVIDIA/cuda-rust.git cargo-oxide
 ```
 
 On first run, `cargo-oxide` will automatically fetch and build the codegen backend.
 
 #### Nix (alternative)
 
-If you have Nix with flakes enabled, `nix develop` from `cuda-oxide/` gives you a reproducible shell with CUDA 13, LLVM 22, Clang, and the pinned Rust nightly — no manual apt installs. The shellHook auto-discovers host NVIDIA drivers on NixOS and non-NixOS systems.
+If you have Nix with flakes enabled, `nix develop` from `cuda-oxide/` gives you a reproducible shell with CUDA 13, LLVM 22, Clang, and the pinned Rust toolchain - no manual apt installs. The shellHook auto-discovers host NVIDIA drivers on NixOS and non-NixOS systems.
 
 ```bash
 # From the repository root:
@@ -187,8 +187,8 @@ nix run 'github:NVIDIA/cuda-rust?dir=cuda-oxide#new' -- my-project
 ```bash
 # Toolchain installed automatically via rust-toolchain.toml
 # Manual install if needed:
-rustup toolchain install nightly-2026-10-11
-rustup component add rust-src rustc-dev rust-analyzer clippy rustfmt llvm-tools --toolchain nightly-2026-10-11
+rustup toolchain install 1.99.1
+rustup component add rust-src rustc-dev rust-analyzer clippy rustfmt llvm-tools --toolchain 1.99.1
 ```
 
 #### CUDA

@@ -10,10 +10,10 @@ Replaces the previous `xtask` pattern with a proper cargo subcommand that works
 
 **External users**:
 
-Install with the project's pinned nightly toolchain:
+Install with the project's pinned toolchain:
 
 ```bash
-cargo +nightly-2026-10-11 install --git https://github.com/NVIDIA/cuda-rust.git cargo-oxide
+cargo +1.99.1 install --git https://github.com/NVIDIA/cuda-rust.git cargo-oxide
 ```
 
 On first run, `cargo-oxide` will automatically fetch and build the codegen backend if it's not already available.

@@ -1056,6 +1056,7 @@ fn backend_build_command_in(
     cmd.env("CARGO_ENCODED_RUSTFLAGS", "");
     cmd.env_remove("RUSTFLAGS");
     cmd.env_remove("CARGO_BUILD_RUSTFLAGS");
+    cmd.env("RUSTC_BOOTSTRAP", "1");
 
     if let Some(path) = rustc_lib_path {
         cmd.env("LIBRARY_PATH", build_library_path(path));

@@ -466,6 +466,7 @@ fn apply_codegen_rustflags(
     append_full_debug_rustflags(&mut encoded, cmd, inherited_debug.as_deref());
 
     cmd.env("CARGO_ENCODED_RUSTFLAGS", encoded)
+        .env("RUSTC_BOOTSTRAP", "1")
         .env_remove("RUSTFLAGS");
     apply_profile_pins(cmd, profile);
 }
@@ -553,6 +554,7 @@ pub fn has_configured_arch(ctx: &Context, cli_arch: Option<&str>) -> bool {
 }
 
 pub(super) fn apply_config_env(cmd: &mut Command, ctx: &Context) {
+    cmd.env("RUSTC_BOOTSTRAP", "1");
     for (key, value) in &ctx.config.env {
         if matches!(key.as_str(), "RUSTFLAGS" | "CARGO_ENCODED_RUSTFLAGS") {
             continue;

@@ -144,6 +144,7 @@ fn run_cargo_fmt_manifest(manifest: &Path, check: bool) -> bool {
 }
 
 fn run_fmt_command(mut cmd: Command) -> bool {
+    cmd.env("RUSTC_BOOTSTRAP", "1");
     match cmd.status() {
         Ok(status) => status.success(),
         Err(e) => {

@@ -367,6 +367,7 @@ fn cargo_package_manifest_matches(package: &serde_json::Value, manifest_path: &P
 
 fn cargo_metadata(manifest_dir: &Path) -> Result<serde_json::Value, String> {
     let output = Command::new("cargo")
+        .env("RUSTC_BOOTSTRAP", "1")
         .args(["metadata", "--format-version=1", "--no-deps"])
         .current_dir(manifest_dir)
         .output()

@@ -14,7 +14,7 @@ use super::*;
 const GIT_REPO: &str = "https://github.com/NVIDIA/cuda-rust.git";
 
 const RUST_TOOLCHAIN_TOML: &str = r#"[toolchain]
-channel = "nightly-2026-10-11"
+channel = "1.99.1"
 components = ["rust-src", "rustc-dev", "rust-analyzer", "clippy", "rustfmt", "llvm-tools"]
 "#;
 

@@ -494,7 +494,7 @@ define_float_atomic! {
     ///
     /// `swap` lowers to hardware `atom.exch.b32`; this API does not expose
     /// `compare_exchange` for floats. `fetch_add` uses LLVM `atomicrmw fadd`.
-    /// With the pinned `nightly-2026-10-11` toolchain's LLVM 23 NVPTX defaults,
+    /// With the pinned `1.99.1` toolchain's LLVM 23 NVPTX defaults,
     /// global and generic `f32` adds use a compare-and-swap loop; shared `f32`
     /// and `f64` adds still use native instructions on supported targets.
     ///

@@ -6,8 +6,6 @@ error[E0080]: evaluation panicked: Swizzle<B, M, S> requires M + B <= S, otherwi
   |
  ::: src/swizzle.rs
   |
-  |   impl<const B: usize, const M: usize, const S: usize> Swizzle<B, M, S> {
-  |       /// Compile-time check that the swizzle is an involution.
   |       const INVOLUTION: () = assert!(
   |  ____________________________-
   | |         M + B <= S,

@@ -13,17 +13,3 @@ note: method defined here
    |
    |     pub fn get_mut<'kernel>(&mut self, idx: ThreadIndex<'kernel, IndexSpace>) -> Option<&mut T> {
    |            ^^^^^^^
-note: the `get_mut` call is resolved to the method in `cuda_device::DisjointSlice<'a, T, IndexSpace>`, shadowing the method of the same name on the inherent impl for `cuda_device::DisjointSlice<'a, T, cuda_device::Runtime2DIndex>`
-  --> tests/compile_fail/thread_index_runtime_mismatch.rs:18:17
-   |
-18 |     let _ = out.get_mut(idx);
-   |                 ^^^^^^^ refers to `cuda_device::DisjointSlice::<'a, T, IndexSpace>::get_mut`
-   = note: additionally, there is 1 other available method that isn't in scope
-help: you might have meant to call one of the other methods; you can use the fully-qualified path to call one of them explicitly
-   |
-18 -     let _ = out.get_mut(idx);
-18 +     let _ = DisjointSlice::get_mut(&mut out, idx);
-   |
-18 -     let _ = out.get_mut(idx);
-18 +     let _ = <_ as std::slice::SliceIndex<_>>::get_mut(out, idx);
-   |
@@ -13,17 +13,3 @@ note: method defined here
    |
    |     pub fn get_mut<'kernel>(&mut self, idx: ThreadIndex<'kernel, IndexSpace>) -> Option<&mut T> {
    |            ^^^^^^^
-note: the `get_mut` call is resolved to the method in `cuda_device::DisjointSlice<'a, T, IndexSpace>`, shadowing the method of the same name on the inherent impl for `cuda_device::DisjointSlice<'a, T, cuda_device::Runtime2DIndex>`
-  --> tests/compile_fail/thread_index_stride_mismatch.rs:11:17
-   |
-11 |     let _ = out.get_mut(idx);
-   |                 ^^^^^^^ refers to `cuda_device::DisjointSlice::<'a, T, IndexSpace>::get_mut`
-   = note: additionally, there is 1 other available method that isn't in scope
-help: you might have meant to call one of the other methods; you can use the fully-qualified path to call one of them explicitly
-   |
-11 -     let _ = out.get_mut(idx);
-11 +     let _ = DisjointSlice::get_mut(&mut out, idx);
-   |
-11 -     let _ = out.get_mut(idx);
-11 +     let _ = <_ as std::slice::SliceIndex<_>>::get_mut(out, idx);
-   |
@@ -195,7 +195,7 @@ fn type_id_divergence_message(
     format!(
         "generic kernel PTX entry `{requested}` was not found, but the loaded module DOES \
          contain `{base}` under a different type-identity hash: {found}. The host \
-         (core::intrinsics::reflection::type_id) and the device backend (tcx.type_id_hash) disagreed on \
+         (core::intrinsics::type_id) and the device backend (tcx.type_id_hash) disagreed on \
          the kernel's `_TID_` name, which means the TypeId-to-u128 extraction in \
          cuda_host::type_id is no longer value-correct on this toolchain. Host and device \
          code must be built by the same pinned nightly; re-verify the TypeId contract in \
@@ -329,10 +329,7 @@ mod tests {
             "{message}"
         );
         // The message must identify the two hash producers and the remedy.
-        assert!(
-            message.contains("core::intrinsics::reflection::type_id"),
-            "{message}"
-        );
+        assert!(message.contains("core::intrinsics::type_id"), "{message}");
         assert!(message.contains("tcx.type_id_hash"), "{message}");
         assert!(message.contains("same pinned nightly"), "{message}");
         assert!(message.contains("compute_kernel_export_name"), "{message}");

@@ -164,7 +164,7 @@ pub trait CudaKernel {
 /// The bound on `ptx_name` is intentionally just whatever the kernel
 /// itself declared — typically `Copy` on the value-passed generics. No
 /// `'static` is added: the helper `cuda_host::type_id_u128` uses
-/// `core::intrinsics::reflection::type_id`, which is bound `T: ?Sized` rather than
+/// `core::intrinsics::type_id`, which is bound `T: ?Sized` rather than
 /// `T: 'static`, so closures that borrow non-`'static` data still go
 /// through the typed launch path (just like they did under the legacy
 /// `type_name`-based scheme).

@@ -3,7 +3,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-// Required for `type_id_u128`'s use of `core::intrinsics::reflection::type_id`. The
+// Required for `type_id_u128`'s use of `core::intrinsics::type_id`. The
 // intrinsic is the only way to obtain the same 128-bit hash the backend
 // uses while keeping the bound at `T: ?Sized` (stable `TypeId::of` would
 // force `T: 'static` on every kernel marker — see `type_id.rs` for why

@@ -10,7 +10,7 @@
 //! API would force a `T: 'static` bound on the kernel marker, which would in
 //! turn reject perfectly valid non-`'static` borrowing closures (e.g. a kernel
 //! launcher capturing `&[f32]` from a stack frame the caller keeps alive
-//! across the launch). The `core::intrinsics::reflection::type_id` form has bound
+//! across the launch). The `core::intrinsics::type_id` form has bound
 //! `T: ?Sized` — i.e. no `'static` requirement — and produces the exact same
 //! 128-bit value that `tcx.type_id_hash` does for that type, because both go
 //! through the same `erase_and_anonymize_regions` + stable-hash pipeline.
@@ -20,10 +20,10 @@
 //! every monomorphized type and const argument, so one hash covers the complete
 //! specialization without inventing a parallel encoding for const values.
 //!
-//! Framing note for future contributors: `core::intrinsics::reflection::type_id` is an
+//! Framing note for future contributors: `core::intrinsics::type_id` is an
 //! internal API and requires `#![feature(core_intrinsics)]` on the owning
 //! crate. cuda-oxide already ships against `rustc_private` and pins a
-//! nightly toolchain, so this is inside our existing risk surface — but the
+//! toolchain, so this is inside our existing risk surface, but the
 //! helper cannot be lifted into a stable-feeling utility crate without
 //! re-introducing the feature gate there.
 
@@ -54,7 +54,7 @@ static GENERIC_KERNEL_NAMES: OnceLock<Mutex<HashMap<KernelNameKey, &'static str>
 /// caller (the borrow must outlive `stream.synchronize()`).
 #[inline]
 pub fn type_id_u128<T: ?Sized>() -> u128 {
-    let id = const { core::intrinsics::reflection::type_id::<T>() };
+    let id = const { core::intrinsics::type_id::<T>() };
     unsafe { core::mem::transmute::<TypeId, u128>(id) }
 }
 

@@ -1799,6 +1799,22 @@ fn parse_ptxas_version(stdout: &str, stderr: &str) -> Result<String> {
     Ok(version.to_owned())
 }
 
+fn llvm_major_minor_series(version: &str) -> Option<&str> {
+    let rest = version.strip_prefix("LLVM version ")?;
+    let mut parts = rest.split(['.', '-']);
+    let major = parts.next()?;
+    let minor = parts.next()?;
+    if major.is_empty()
+        || !major.bytes().all(|b| b.is_ascii_digit())
+        || minor.is_empty()
+        || !minor.bytes().all(|b| b.is_ascii_digit())
+    {
+        return None;
+    }
+    let prefix_len = "LLVM version ".len() + major.len() + 1 + minor.len();
+    Some(&version[..prefix_len])
+}
+
 fn validate_backend_identity(
     expected_version: &str,
     expected_commit: &RustcCommit,
@@ -1807,8 +1823,12 @@ fn validate_backend_identity(
     let ProbeBackendIdentity::SelectedEvidence(actual) = actual else {
         return Ok(());
     };
+    let same_version_or_series = actual.tool.version == expected_version
+        || (llvm_major_minor_series(&actual.tool.version).is_some()
+            && llvm_major_minor_series(&actual.tool.version)
+                == llvm_major_minor_series(expected_version));
     ensure!(
-        actual.tool.version == expected_version,
+        same_version_or_series,
         "rust-toolchain llc version mismatch: selected evidence records {expected_version:?}, found {:?}; use an explicit `--llc` only for a comparison probe",
         actual.tool.version
     );

@@ -12,4 +12,5 @@ note: required by a bound in `LoadedModule::alias_fakes_row_width`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::alias_fakes_row_width`
 ...
 22 |     pub fn alias_fakes_row_width(mut out: DisjointSlice<f32, Runtime2DIndex>) {
+   |            --------------------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
@@ -12,4 +12,5 @@ note: required by a bound in `LoadedModule::alias_hides_row_width`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::alias_hides_row_width`
 ...
 22 |     pub fn alias_hides_row_width(mut out: DisjointSlice<f32, Rt>) {
+   |            --------------------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
@@ -26,6 +26,7 @@ note: required by a bound in `LoadedModule::prepare_lookalike`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::prepare_lookalike`
 ...
 21 |     pub fn lookalike(mut out: DisjointSlice<u32>) {
+   |            --------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
 
 error[E0277]: the trait bound `for<'__cuda_oxide_disjoint> DisjointSlice<'__cuda_oxide_disjoint, u32>: cuda_device::disjoint::__LaunchContractDisjointSliceAbi<u32, false>` is not satisfied
@@ -56,4 +57,5 @@ note: required by a bound in `LoadedModule::prepare_lookalike`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::prepare_lookalike`
 ...
 21 |     pub fn lookalike(mut out: DisjointSlice<u32>) {
+   |            --------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
@@ -22,4 +22,5 @@ note: required by a bound in `LoadedModule::prepare_misleading_name`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::prepare_misleading_name`
 ...
 13 |     pub fn misleading_name(mut out: DisjointSlice<u32, Index2D>) {
+   |            --------------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
@@ -22,6 +22,7 @@ note: required by a bound in `LoadedModule::prepare_reordered`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::prepare_reordered`
 ...
 17 |     pub fn reordered(mut out: DisjointSlice<Index1D, u64>) {
+   |            --------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
 
 error[E0277]: the trait bound `for<'__cuda_oxide_disjoint> cuda_device::DisjointSlice<'__cuda_oxide_disjoint, u64>: cuda_device::disjoint::__LaunchContractDisjointSliceAbi<cuda_device::Index1D, false>` is not satisfied
@@ -48,4 +49,5 @@ note: required by a bound in `LoadedModule::prepare_reordered`
    | ^^^^^^^^^^^^^^ required by this bound in `LoadedModule::prepare_reordered`
 ...
 17 |     pub fn reordered(mut out: DisjointSlice<Index1D, u64>) {
+   |            --------- required by a bound in this associated function
    = note: this error originates in the attribute macro `cuda_module` (in Nightly builds, run with -Z macro-backtrace for more info)
@@ -1,7 +1,6 @@
 error[E0080]: evaluation panicked: partial unroll factor must be in 2..=1024, or 0 for full unrolling
  --> $WORKSPACE/crates/cuda-device/src/thread.rs
   |
-  | pub fn __unroll_config<const FACTOR: u32>() {
   |     const { validate_unroll_factor(FACTOR) }
   |             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ evaluation of `cuda_device::__unroll_config::<1>::{constant#0}` failed inside this call
   |

@@ -1,14 +1,3 @@
-warning: `feature(generic_const_exprs)` is not supported with the next-generation trait solver
- --> tests/compile_fail/policy_config_unresolved.rs:4:12
-  |
-4 | #![feature(generic_const_exprs)]
-  |            ^^^^^^^^^^^^^^^^^^^
-  |
-  = note: `-Znext-solver=globally` is currently enabled by default for testing
-  = note: reverted the setting to `-Znext-solver=coherence` for this crate
-  = note: the currently stable trait solver will be used for this crate
-  = note: see issues #160895 <https://github.com/rust-lang/rust/issues/160895> for more information
-
 error[E0599]: no associated function or constant named `MAX_THREADS` found for type parameter `P` in the current scope
   --> tests/compile_fail/policy_config_unresolved.rs:10:20
    |

@@ -1,7 +1,6 @@
 error[E0080]: evaluation panicked: launch_bounds maximum threads must be greater than zero
  --> $WORKSPACE/crates/cuda-device/src/thread.rs
   |
-  | pub fn __launch_bounds_config<const MAX_THREADS: u32, const MIN_BLOCKS: u32>() {
   |     const { validate_launch_bounds(MAX_THREADS) }
   |             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ evaluation of `cuda_device::__launch_bounds_config::<0, 0>::{constant#0}` failed inside this call
   |