Broken MIR in encoding_rs related to MIR inlining + DestProp

[saethlin]

This was originally #105344 (comment)

---

Trying to compile https://github.com/hsivonen/encoding_rs 5931e21bc20bad0335964f441da0c4309401107a with these flags:

$ RUSTFLAGS="-Zmir-opt-level=3 -Zvalidate-mir -Zinline-mir=yes" cargo +nightly test

Produces this:

error: internal compiler error: broken MIR in Item(WithOptConstParam { did: DefId(0:1315 ~ encoding_rs[5e1f]::utf_16::{impl#0}::decode_to_utf16_raw), const_param_did: None }) (before pass SimplifyCfg-final) at bb218[0]:
                                encountered overlapping memory in `Call` terminator: _454 = bswap::<u16>(move _454) -> bb163
  --> src/utf_16.rs:85:21
   |
85 |                     dest.copy_utf16_from::<LittleEndian>(&mut source)
   |                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   |
   = note: delayed at compiler/rustc_const_eval/src/transform/validate.rs:784:26

Removing -Zinline-mir=yes or -Zmir-opt-level=3 removes the ICE.

---

searched nightlies: from nightly-2022-01-01 to nightly-2022-12-05
regressed nightly: nightly-2022-11-28
searched commit range: 80a9646...1eb62b1
regressed commit: 0e9eee6

bisected with cargo-bisect-rustc v0.6.5

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc --script script --start 2022-01-01 --end 2022-12-05 

[RalfJung]

@tmiasko I am trying to understand what happened here. The test case added in #105420 looks like maybe the analysis done by DestProp is smart enough to ignore dead code and that's why it can introduce UB in dead code?

However the code at the top of this issue looks quite different. Looking at it's source, I can't see how ignoring dead code would lead to introducing overlapping calls?