Some opt-level 2 optimizations make rustc a *lot* slower

[RalfJung]

This Miri testcase, when built via rustc, usually only takes a few seconds. With -Zmir-opt-level=2, however, it takes at least two minutes (then I aborted compilation).

To reproduce, I did this in a Miri checkout (no need to build Miri, we are just using the example):

$ time rustc tests/run-pass/float.rs

real    0m7,099s
user    0m6,219s
sys     0m0,964s
$ time rustc tests/run-pass/float.rs -Zmir-opt-level=2
^C

real    2m12,074s
user    2m10,970s
sys     0m0,985s

Cc @rust-lang/wg-mir-opt

[tmiasko]

Copy propagation takes most of time here; tracked in #36673. Is this the same source as in #73717?

[RalfJung]

oops it is... I thought I had reported this before but couldn't find it.^^ Sorry.