feat(llvm): add llvm19 support for compute_100+

.gitignore

@@ -2,3 +2,7 @@ book
 /target
 **/.vscode
 .devcontainer
+.codex
+rustc-ice-*.txt
+.nix-driver-libs
+.claude

crates/cuda_builder/src/lib.rs

@@ -196,6 +196,21 @@ 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 {
@@ -204,7 +219,7 @@ impl CudaBuilder {
             ptx_file_copy_path: None,
             generate_line_info: true,
             nvvm_opts: true,
-            arch: NvvmArch::default(),
+            arch: default_arch(),
             ftz: false,
             fast_sqrt: false,
             fast_div: false,
@@ -355,6 +370,7 @@ 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)?;
@@ -550,13 +566,21 @@ fn build_backend_and_find(filename: &str) -> Option<PathBuf> {
 
     let target_dir = workspace_dir.join("target").join("cuda-builder-codegen");
 
-    let status = Command::new("cargo")
-        .args(["build", "-p", "rustc_codegen_nvvm"])
+    let mut cmd = Command::new("cargo");
+    cmd.args(["build", "-p", "rustc_codegen_nvvm"])
         .arg("--target-dir")
         .arg(&target_dir)
-        .current_dir(&workspace_dir)
-        .status()
-        .ok()?;
+        .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() {
+        cmd.args(["--features", "llvm19"]);
+    }
+
+    let status = cmd.status().ok()?;
 
     if !status.success() {
         return None;

crates/cust/build.rs

@@ -40,4 +40,12 @@ fn main() {
         println!("cargo::rustc-cfg=cuGraphGetEdges_v2");
         println!("cargo::rustc-cfg=cuCtxCreate_v4");
     }
+
+    // In CUDA 13.2 the `id` field in `CUmemLocation_st` was placed inside an anonymous union.
+    // Bindgen renders this as `__bindgen_anon_1: CUmemLocation_st__bindgen_ty_1` instead of a
+    // direct `id` field. This cfg gates the struct initialization syntax accordingly.
+    println!("cargo::rustc-check-cfg=cfg(cuMemLocation_anon_id)");
+    if driver_version >= 13020 {
+        println!("cargo::rustc-cfg=cuMemLocation_anon_id");
+    }
 }

crates/cust/src/memory/unified.rs

@@ -647,6 +647,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemPrefetchAsync_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemPrefetchAsync_v2))]
@@ -693,6 +696,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemPrefetchAsync_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemPrefetchAsync_v2))]
@@ -735,6 +741,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemAdvise_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemAdvise_v2))]
@@ -777,6 +786,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemAdvise_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemAdvise_v2))]
@@ -801,6 +813,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemAdvise_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemAdvise_v2))]

crates/cust_raw/build/cuda_sdk.rs

@@ -215,6 +215,7 @@ impl CudaSdk {
             vec![
                 cuda_root.join("nvvm").join("bin"),
                 cuda_root.join("nvvm").join("lib64"),
+                cuda_root.join("nvvm").join("lib"),
             ]
         };
         let library_dirs = Self::normalize_dirpaths(search_dirs);

crates/cust_raw/build/main.rs

@@ -99,8 +99,12 @@ fn main() {
             println!("cargo::rustc-link-search=native={}", libdir.display());
         }
         println!("cargo::rustc-link-lib=dylib=nvvm");
-        // Handle libdevice support.
-        fs::copy(sdk.libdevice_bitcode_path(), outdir.join("libdevice.bc"))
+        // `fs::copy` preserves source mode. When libdevice.10.bc comes from
+        // the Nix store (0444), re-running this build can't overwrite the
+        // previous copy in OUT_DIR. Drop it first.
+        let dest = outdir.join("libdevice.bc");
+        let _ = fs::remove_file(&dest);
+        fs::copy(sdk.libdevice_bitcode_path(), &dest)
             .expect("Cannot copy libdevice bitcode file.");
     }
 }

crates/nvvm/src/lib.rs

@@ -4,7 +4,6 @@ use std::{
     ffi::{CStr, CString},
     fmt::Display,
     mem::MaybeUninit,
-    ptr::null_mut,
     str::FromStr,
 };
 
@@ -325,6 +324,10 @@ pub enum NvvmArch {
     Compute89,
     Compute90,
     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.
     Compute100,
     Compute100f,
     Compute100a,
@@ -448,6 +451,14 @@ impl NvvmArch {
         self.capability_value() % 10
     }
 
+    /// Whether this target uses NVVM's modern IR dialect rather than the legacy LLVM 7 dialect.
+    ///
+    /// CUDA 13.2 documents the modern dialect as Blackwell-and-later only, which begins at
+    /// `compute_100`.
+    pub fn uses_modern_ir_dialect(&self) -> bool {
+        self.capability_value() >= 100
+    }
+
     /// Get the target feature string (e.g., "compute_50" for `Compute50`, "compute_90a" for
     /// `Compute90a`).
     pub fn target_feature(&self) -> &'static str {
@@ -739,7 +750,24 @@ impl NvvmProgram {
     /// Verify the program without actually compiling it. In the case of invalid IR, you can find
     /// more detailed error info by calling [`compiler_log`](Self::compiler_log).
     pub fn verify(&self) -> Result<(), NvvmError> {
-        unsafe { nvvm_sys::nvvmVerifyProgram(self.raw, 0, null_mut()).to_result() }
+        self.verify_with_options(&[])
+    }
+
+    /// Like [`verify`](Self::verify), but runs the verifier with the same `NvvmOption`s that will
+    /// be passed to [`compile`](Self::compile). Passing the user-selected `-arch=compute_XXX` in
+    /// particular matters for CUDA 12.9+ / LLVM 19 bitcode: without it the verifier can fall back
+    /// to the legacy LLVM 7 parser and reject modern-dialect bitcode that would otherwise compile
+    /// fine.
+    pub fn verify_with_options(&self, options: &[NvvmOption]) -> Result<(), NvvmError> {
+        unsafe {
+            let options = options.iter().map(|x| format!("{x}\0")).collect::<Vec<_>>();
+            let mut options_ptr = options
+                .iter()
+                .map(|x| x.as_ptr().cast())
+                .collect::<Vec<_>>();
+            nvvm_sys::nvvmVerifyProgram(self.raw, options.len() as i32, options_ptr.as_mut_ptr())
+                .to_result()
+        }
     }
 }
 

crates/rustc_codegen_nvvm/Cargo.toml

@@ -14,6 +14,10 @@ readme = "../../README.md"
 [lib]
 crate-type = ["dylib"]
 
+[features]
+default = []
+llvm19 = []
+
 [dependencies]
 nvvm = { version = "0.1", path = "../nvvm" }
 rustc-demangle = "0.1.24"