Auto merge of #131650 - saethlin:post-mono-mir-opts, r=<try>

Add post-mono MIR passes to make mono-reachable analysis more accurate

As of #131650 (comment) I believe most of the incr overhead comes from re-computing, re-encoding, and loading a lot more MIR when all we're actually doing is traversing through it. I think that can be addressed by caching a query that looks up the mentioned/used items for an Instance.

I think the full-build regressions are pretty much just the expense of cloning, then monomorphizing, then caching the MIR.

Author: bors

compiler/rustc_codegen_cranelift/src/base.rs

@@ -41,7 +41,7 @@ pub(crate) fn codegen_fn<'tcx>(
     let symbol_name = tcx.symbol_name(instance).name.to_string();
     let _timer = tcx.prof.generic_activity_with_arg("codegen fn", &*symbol_name);
 
-    let mir = tcx.instance_mir(instance.def);
+    let mir = tcx.build_codegen_mir(instance);
     let _mir_guard = crate::PrintOnPanic(|| {
         let mut buf = Vec::new();
         with_no_trimmed_paths!({

compiler/rustc_codegen_ssa/src/base.rs

@@ -416,7 +416,7 @@ pub(crate) fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
     // release builds.
     info!("codegen_instance({})", instance);
 
-    mir::codegen_mir::<Bx>(cx, instance);
+    mir::lower_mir::<Bx>(cx, instance);
 }
 
 /// Creates the `main` function which will initialize the rust runtime and call