Auto merge of #101437 - compiler-errors:erase-normalize-ordering, r=tmandry

Normalize before erasing late-bound regions in `equal_up_to_regions`

Normalize erasing regions **first**, before passing the type through a `BottomUpFolder` which erases late-bound regions too.

The root cause of this issue is due to 96d4137deed6c52c6db2dd19568c37d1c160f1e7, which removes a `normalize_erasing_regions` that happens before this call to `equal_up_to_regions`. While reverting that commit might be a fix, I think it was suspicious to be erasing late-bound regions first _then_ normalizing types in the first place in `equal_up_to_regions`.

-----

I am tempted to ask the reviewer to review and `r+` this without a UI test, since the existing issues that I think this fixes are all incredibly difficult to minimize (anything hyper/warp related, given the nature of those libraries :sweat:) or impossible to reproduce locally (the miri test), namely:
* This recently reported issue with tokio + warp: #101430
* This issue from `@RalfJung` about Miri being broken: #101344
* This additional issue reported in a comment by `@tmandry` (issue with fuchsia + hyper): https://github.com/rust-lang/rust/issues/101344#issuecomment-1235974564

I have locally verified that the repro in #101430 is fixed with this PR, but after a couple of hours of attempting to minimize this error and either failing to actually repro the ICE, or being overwhelmed with the number of traits and functions I need to inline into a UI test, I have basically given up. Thoughts are appreciated on how best to handle this.

r? `@oli-obk` who is at the intersection of MIR and types-related stuff who may be able to give advice :sweat_smile:

1 files changed, 15 insertions, 14 deletions

diff --git a/compiler/rustc_const_eval/src/transform/validate.rs b/compiler/rustc_const_eval/src/transform/validate.rs
index 783ab350faa..8576e0f0f7c 100644
--- a/compiler/rustc_const_eval/src/transform/validate.rs
+++ b/compiler/rustc_const_eval/src/transform/validate.rs
@@ -90,20 +90,21 @@ pub fn equal_up_to_regions<'tcx>(
 
     // Normalize lifetimes away on both sides, then compare.
     let normalize = |ty: Ty<'tcx>| {
-        let ty = ty.fold_with(&mut BottomUpFolder {
-            tcx,
-            // FIXME: We erase all late-bound lifetimes, but this is not fully correct.
-            // If you have a type like `<for<'a> fn(&'a u32) as SomeTrait>::Assoc`,
-            // this is not necessarily equivalent to `<fn(&'static u32) as SomeTrait>::Assoc`,
-            // since one may have an `impl SomeTrait for fn(&32)` and
-            // `impl SomeTrait for fn(&'static u32)` at the same time which
-            // specify distinct values for Assoc. (See also #56105)
-            lt_op: |_| tcx.lifetimes.re_erased,
-            // Leave consts and types unchanged.
-            ct_op: |ct| ct,
-            ty_op: |ty| ty,
-        });
-        tcx.try_normalize_erasing_regions(param_env, ty).unwrap_or(ty)
+        tcx.try_normalize_erasing_regions(param_env, ty).unwrap_or(ty).fold_with(
+            &mut BottomUpFolder {
+                tcx,
+                // FIXME: We erase all late-bound lifetimes, but this is not fully correct.
+                // If you have a type like `<for<'a> fn(&'a u32) as SomeTrait>::Assoc`,
+                // this is not necessarily equivalent to `<fn(&'static u32) as SomeTrait>::Assoc`,
+                // since one may have an `impl SomeTrait for fn(&32)` and
+                // `impl SomeTrait for fn(&'static u32)` at the same time which
+                // specify distinct values for Assoc. (See also #56105)
+                lt_op: |_| tcx.lifetimes.re_erased,
+                // Leave consts and types unchanged.
+                ct_op: |ct| ct,
+                ty_op: |ty| ty,
+            },
+        )
     };
     tcx.infer_ctxt().enter(|infcx| infcx.can_eq(param_env, normalize(src), normalize(dest)).is_ok())
 }