Use &'static str instead of String for target features.

It's a more precise type -- these are static values. Also it avoids
unnecessary allocations.

Author: nnethercote

crates/nvvm/src/lib.rs

@@ -393,40 +393,40 @@ impl NvvmArch {
     }
 
     /// Get the target feature string (e.g., "compute_35" for Compute35, "compute_90a" for Compute90a)
-    pub fn target_feature(&self) -> String {
+    pub fn target_feature(&self) -> &'static str {
         match self {
-            Self::Compute35 => "compute_35".to_string(),
-            Self::Compute37 => "compute_37".to_string(),
-            Self::Compute50 => "compute_50".to_string(),
-            Self::Compute52 => "compute_52".to_string(),
-            Self::Compute53 => "compute_53".to_string(),
-            Self::Compute60 => "compute_60".to_string(),
-            Self::Compute61 => "compute_61".to_string(),
-            Self::Compute62 => "compute_62".to_string(),
-            Self::Compute70 => "compute_70".to_string(),
-            Self::Compute72 => "compute_72".to_string(),
-            Self::Compute75 => "compute_75".to_string(),
-            Self::Compute80 => "compute_80".to_string(),
-            Self::Compute86 => "compute_86".to_string(),
-            Self::Compute87 => "compute_87".to_string(),
-            Self::Compute89 => "compute_89".to_string(),
-            Self::Compute90 => "compute_90".to_string(),
-            Self::Compute90a => "compute_90a".to_string(),
-            Self::Compute100 => "compute_100".to_string(),
-            Self::Compute100f => "compute_100f".to_string(),
-            Self::Compute100a => "compute_100a".to_string(),
-            Self::Compute101 => "compute_101".to_string(),
-            Self::Compute101f => "compute_101f".to_string(),
-            Self::Compute101a => "compute_101a".to_string(),
-            Self::Compute103 => "compute_103".to_string(),
-            Self::Compute103f => "compute_103f".to_string(),
-            Self::Compute103a => "compute_103a".to_string(),
-            Self::Compute120 => "compute_120".to_string(),
-            Self::Compute120f => "compute_120f".to_string(),
-            Self::Compute120a => "compute_120a".to_string(),
-            Self::Compute121 => "compute_121".to_string(),
-            Self::Compute121f => "compute_121f".to_string(),
-            Self::Compute121a => "compute_121a".to_string(),
+            Self::Compute35 => "compute_35",
+            Self::Compute37 => "compute_37",
+            Self::Compute50 => "compute_50",
+            Self::Compute52 => "compute_52",
+            Self::Compute53 => "compute_53",
+            Self::Compute60 => "compute_60",
+            Self::Compute61 => "compute_61",
+            Self::Compute62 => "compute_62",
+            Self::Compute70 => "compute_70",
+            Self::Compute72 => "compute_72",
+            Self::Compute75 => "compute_75",
+            Self::Compute80 => "compute_80",
+            Self::Compute86 => "compute_86",
+            Self::Compute87 => "compute_87",
+            Self::Compute89 => "compute_89",
+            Self::Compute90 => "compute_90",
+            Self::Compute90a => "compute_90a",
+            Self::Compute100 => "compute_100",
+            Self::Compute100f => "compute_100f",
+            Self::Compute100a => "compute_100a",
+            Self::Compute101 => "compute_101",
+            Self::Compute101f => "compute_101f",
+            Self::Compute101a => "compute_101a",
+            Self::Compute103 => "compute_103",
+            Self::Compute103f => "compute_103f",
+            Self::Compute103a => "compute_103a",
+            Self::Compute120 => "compute_120",
+            Self::Compute120f => "compute_120f",
+            Self::Compute120a => "compute_120a",
+            Self::Compute121 => "compute_121",
+            Self::Compute121f => "compute_121f",
+            Self::Compute121a => "compute_121a",
         }
     }
 
@@ -448,18 +448,18 @@ impl NvvmArch {
     ///
     /// For more details on family and architecture-specific features, see:
     /// <https://developer.nvidia.com/blog/nvidia-blackwell-and-nvidia-cuda-12-9-introduce-family-specific-architecture-features/>
-    pub fn all_target_features(&self) -> Vec<String> {
-        let mut features: Vec<String> = if self.is_architecture_variant() {
+    pub fn all_target_features(&self) -> Vec<&'static str> {
+        let mut features: Vec<_> = if self.is_architecture_variant() {
             // 'a' variants: include all available instructions for the architecture
             // This means: all base variants up to same version, all 'f' variants with same major and <= minor, plus itself
-            let base_features: Vec<String> = NvvmArch::iter()
+            let base_features: Vec<_> = NvvmArch::iter()
                 .filter(|arch| {
                     arch.is_base_variant() && arch.capability_value() <= self.capability_value()
                 })
                 .map(|arch| arch.target_feature())
                 .collect();
 
-            let family_features: Vec<String> = NvvmArch::iter()
+            let family_features: Vec<_> = NvvmArch::iter()
                 .filter(|arch| {
                     arch.is_family_variant()
                         && arch.major_version() == self.major_version()

crates/rustc_codegen_nvvm/src/lib.rs

@@ -146,7 +146,11 @@ impl CodegenBackend for NvvmCodegenBackend {
             for opt in &args.nvvm_options {
                 if let ::nvvm::NvvmOption::Arch(arch) = opt {
                     // Add all features up to and including the current architecture
-                    features.extend(arch.all_target_features());
+                    features.extend(
+                        arch.all_target_features()
+                            .into_iter()
+                            .map(|s| s.to_string()),
+                    );
                     break;
                 }
             }