Rollup merge of #122287 - RalfJung:simd-static-assert, r=pnkfelix

add test ensuring simd codegen checks don't run when a static assertion failed

stdarch relies on this to ensure that SIMD indices are in bounds.

I would love to know why this works, but I can't figure out where codegen decides to not codegen a function if a required-const does not evaluate. `@oli-obk` `@bjorn3` do you have any idea?

compiler/rustc_codegen_ssa/src/mir/constant.rs

@@ -21,11 +21,11 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
     }
 
     pub fn eval_mir_constant(&self, constant: &mir::ConstOperand<'tcx>) -> mir::ConstValue<'tcx> {
-        // `MirUsedCollector` visited all constants before codegen began, so if we got here there
-        // can be no more constants that fail to evaluate.
+        // `MirUsedCollector` visited all required_consts before codegen began, so if we got here
+        // there can be no more constants that fail to evaluate.
         self.monomorphize(constant.const_)
             .eval(self.cx.tcx(), ty::ParamEnv::reveal_all(), Some(constant.span))
-            .expect("erroneous constant not captured by required_consts")
+            .expect("erroneous constant missed by mono item collection")
     }
 
     /// This is a convenience helper for `simd_shuffle_indices`. It has the precondition

compiler/rustc_codegen_ssa/src/mir/mod.rs

@@ -211,7 +211,8 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 
     // It may seem like we should iterate over `required_consts` to ensure they all successfully
     // evaluate; however, the `MirUsedCollector` already did that during the collection phase of
-    // monomorphization so we don't have to do it again.
+    // monomorphization, and if there is an error during collection then codegen never starts -- so
+    // we don't have to do it again.
 
     fx.per_local_var_debug_info = fx.compute_per_local_var_debug_info(&mut start_bx);