Auto merge of rust-lang#137586 - nnethercote:SetImpliedBits, r=bjorn3

Speed up target feature computation

The LLVM backend calls `LLVMRustHasFeature` twice for every feature. In short-running rustc invocations, this accounts for a surprising amount of work.

r? `@bjorn3`

Author: bors

Diff for: src/lib.rs

@@ -176,13 +176,9 @@ impl CodegenBackend for CraneliftCodegenBackend {
         }
     }
 
-    fn target_features_cfg(
-        &self,
-        sess: &Session,
-        _allow_unstable: bool,
-    ) -> Vec<rustc_span::Symbol> {
+    fn target_features_cfg(&self, sess: &Session) -> (Vec<Symbol>, Vec<Symbol>) {
         // FIXME return the actually used target features. this is necessary for #[cfg(target_feature)]
-        if sess.target.arch == "x86_64" && sess.target.os != "none" {
+        let target_features = if sess.target.arch == "x86_64" && sess.target.os != "none" {
             // x86_64 mandates SSE2 support and rustc requires the x87 feature to be enabled
             vec![sym::fsxr, sym::sse, sym::sse2, Symbol::intern("x87")]
         } else if sess.target.arch == "aarch64" {
@@ -196,7 +192,10 @@ impl CodegenBackend for CraneliftCodegenBackend {
             }
         } else {
             vec![]
-        }
+        };
+        // FIXME do `unstable_target_features` properly
+        let unstable_target_features = target_features.clone();
+        (target_features, unstable_target_features)
     }
 
     fn print_version(&self) {