To quote @lqd in #156141 (comment):

While trying to expand new solver testing in CI to keep it working, I've noticed that not only had we regressed building the stdlib (we knew that and now have CI to prevent that from happening before stabilization), we also regressed bootstrapping with the new solver enabled. That issue is still present on the main branch.

Building a stage 2 compiler with a stage 1 w/ new solver compiler, or RUSTFLAGS, will ICE very strangely while building regex_syntax (and I can't easily reproduce building that crate by itself...):

error: internal compiler error: compiler/rustc_ty_utils/src/sig_types.rs:114:13: Ctor(Variant, Fn) has not seen any uses of `walk_types` yet, ping oli-obk if you'd like any help
    --> ~/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/regex-syntax-0.8.8/src/ast/mod.rs:1134:5
     |
1134 |     Item(ClassSetItem),
     |     ^^^^


thread 'rustc' (1986053) panicked at compiler/rustc_ty_utils/src/sig_types.rs:114:13:

(e.g. with RUSTFLAGS_NOT_BOOTSTRAP="-Znext-solver=globally" ./x build library --stage 2. Fun: I can reproduce the ICE locally on different machines, but not on CI. what.)

I don't really know what's going on, but it looks like weirdness w/ opaque types, during codegen. I've bisected that to 68f9143 from this PR so I wanted to let you and @lcnr know in case you hadn't noticed.

@rustbot claim