refactor(cuda_builder): gate Blackwell default arch on llvm19 cargo feature

Replace the LLVM_CONFIG_19 env-var sniffing in `cuda_builder` with a proper
`llvm19` cargo feature, addressing review feedback on #375.

- `nvvm` gains an `llvm19` feature; `NvvmArch`'s `#[default]` is moved off
  `Compute75` and onto `Compute100` via `cfg_attr` when it's enabled, so
  `NvvmArch::default()` returns the right answer for the active dialect.
- `cuda_builder` gains a matching `llvm19` feature that propagates to
  `nvvm/llvm19` and (when the optional dep is on) `rustc_codegen_nvvm/llvm19`.
  `CudaBuilder::new` goes back to plain `NvvmArch::default()`.
- The build script's nested `cargo build -p rustc_codegen_nvvm` now keys the
  `--features llvm19` flag off `cfg!(feature = "llvm19")` instead of the env
  var, so the toggle lives in one place.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brandonros

crates/cuda_builder/Cargo.toml

@@ -16,6 +16,12 @@ default = []
 # HACK(see rust-gpu/spirv-builder): use `dep:` to avoid Cargo auto-creating a feature
 # with the dependency name. Consumers must explicitly opt-in to compiling the backend.
 rustc_codegen_nvvm = ["dep:rustc_codegen_nvvm"]
+# Build the backend against LLVM 19 instead of LLVM 7. Propagates to `nvvm` (which
+# uses it to flip the default `NvvmArch` to `Compute100`) and, when the optional
+# `rustc_codegen_nvvm` dep is also enabled, to `rustc_codegen_nvvm` itself. Even
+# when the optional dep is disabled, the build script's nested `cargo build -p
+# rustc_codegen_nvvm` reads `cfg(feature = "llvm19")` here and forwards it.
+llvm19 = ["nvvm/llvm19", "rustc_codegen_nvvm?/llvm19"]
 
 [dependencies]
 rustc_codegen_nvvm = { version = "0.3", path = "../rustc_codegen_nvvm", optional = true }

crates/cuda_builder/src/lib.rs

@@ -196,21 +196,6 @@ pub struct CudaBuilder {
     pub final_module_path: Option<PathBuf>,
 }
 
-/// Default arch for new `CudaBuilder`s.
-///
-/// When the backend is being built with LLVM 19 support (detected via the `LLVM_CONFIG_19`
-/// env var — the same signal `rustc_codegen_nvvm`'s build script uses), default to the
-/// lowest Blackwell compute capability (`Compute100`). Pre-Blackwell archs use the legacy
-/// LLVM 7 NVVM dialect, so pairing them with an LLVM 19 backend is never the right choice.
-/// Callers can still override via [`CudaBuilder::arch`].
-fn default_arch() -> NvvmArch {
-    if env::var_os("LLVM_CONFIG_19").is_some() {
-        NvvmArch::Compute100
-    } else {
-        NvvmArch::default()
-    }
-}
-
 impl CudaBuilder {
     pub fn new(path_to_crate_root: impl AsRef<Path>) -> Self {
         Self {
@@ -219,7 +204,7 @@ impl CudaBuilder {
             ptx_file_copy_path: None,
             generate_line_info: true,
             nvvm_opts: true,
-            arch: default_arch(),
+            arch: NvvmArch::default(),
             ftz: false,
             fast_sqrt: false,
             fast_div: false,
@@ -370,7 +355,6 @@ impl CudaBuilder {
     /// ptx file. If [`ptx_file_copy_path`](Self::ptx_file_copy_path) is set, this returns the copied path.
     pub fn build(self) -> Result<PathBuf, CudaBuilderError> {
         println!("cargo:rerun-if-changed={}", self.path_to_crate.display());
-        println!("cargo:rerun-if-env-changed=LLVM_CONFIG_19");
         let path = invoke_rustc(&self)?;
         if let Some(copy_path) = self.ptx_file_copy_path {
             std::fs::copy(path, &copy_path).map_err(CudaBuilderError::FailedToCopyPtxFile)?;
@@ -572,11 +556,10 @@ fn build_backend_and_find(filename: &str) -> Option<PathBuf> {
         .arg(&target_dir)
         .current_dir(&workspace_dir);
 
-    // Propagate the llvm19 cargo feature to the nested build when the surrounding
-    // shell is configured for LLVM 19 (signalled by LLVM_CONFIG_19). Without this
-    // rustc_codegen_nvvm's build.rs defaults to the LLVM 7 path and falls through
-    // to the prebuilt LLVM 7 download, which fails on Linux.
-    if env::var_os("LLVM_CONFIG_19").is_some() {
+    // Propagate the `llvm19` cargo feature to the nested backend build. Without this
+    // `rustc_codegen_nvvm`'s build script falls through to the prebuilt LLVM 7
+    // download, which the LLVM 19 codegen path can't link against.
+    if cfg!(feature = "llvm19") {
         cmd.args(["--features", "llvm19"]);
     }
 

crates/nvvm/Cargo.toml

@@ -8,6 +8,14 @@ description = "High level bindings to libnvvm"
 repository = "https://github.com/Rust-GPU/rust-cuda"
 readme = "../../README.md"
 
+[features]
+default = []
+# Match the `llvm19` feature on `rustc_codegen_nvvm`. Currently only flips the
+# default `NvvmArch` to the lowest Blackwell capability, since the LLVM 7
+# bitcode dialect can't target `compute_100+` and the LLVM 19 dialect can't
+# target pre-Blackwell archs.
+llvm19 = []
+
 [dependencies]
 cust_raw = { version = "0.11.3", path = "../cust_raw", default-features = false, features = ["nvvm"] }
 strum = { version = "0.27", features = ["derive"] }

crates/nvvm/src/lib.rs

@@ -310,12 +310,13 @@ pub enum NvvmArch {
     Compute73,
     /// This default value of 7.5 corresponds to Turing and later devices. We default to this
     /// because it is the minimum supported by CUDA 13.0 while being in the middle of the range
-    /// supported by CUDA 12.x.
+    /// supported by CUDA 12.x. Selected as the default only when the `llvm19` feature is off;
+    /// the LLVM 19 NVVM dialect can't target pre-Blackwell archs.
     // WARNING: If you change the default, consider updating:
     // - The `--target-arch` values used for compiletests in `ci_linux.yml` and
     //   `.github/workflows/ci_{linux,windows}.yml`.
     // - The CUDA versions used in `setup_cuda_environment` in `compiletests`.
-    #[default]
+    #[cfg_attr(not(feature = "llvm19"), default)]
     Compute75,
     Compute80,
     Compute86,
@@ -326,8 +327,10 @@ pub enum NvvmArch {
     Compute90a,
     /// First Blackwell arch and the cutoff for NVVM's modern IR dialect — everything at
     /// or above this capability uses the LLVM 19-flavored bitcode accepted by CUDA 12.9+
-    /// `libnvvm`. See [`NvvmArch::uses_modern_ir_dialect`]. This is also the default arch
-    /// `cuda_builder` picks when the backend is built with `LLVM_CONFIG_19` set.
+    /// `libnvvm`. See [`NvvmArch::uses_modern_ir_dialect`]. Selected as the default when
+    /// the `llvm19` feature is enabled, since the LLVM 7 dialect can't target this and
+    /// the LLVM 19 dialect can't target anything below it.
+    #[cfg_attr(feature = "llvm19", default)]
     Compute100,
     Compute100f,
     Compute100a,