Add run-pass regression test for SimplifyComparisonIntegral non-SSA miscompile

[Vastargazing]

Regression test for #150904

#142915 enabled DestinationPropagation by default, which let
SimplifyComparisonIntegral see a switchInt on a comparison local that is
reassigned between the Eq and the switchInt (a non-SSA local). The pass
rewrote switchInt(_cmp) into switchInt(place) without checking that _cmp
is written exactly once, so it used the value from Eq and dropped the later
reassignment, miscompiling at -O.

#150925 fixed this by only relating SSA locals, but it shipped with mir-opt
(MIR-shape) tests only - they assert the pass no longer fires on the bad
shapes, but nothing checks the compiled program actually produces the right
result. This adds that behavioral run-pass test, so the end-to-end miscompile
stays covered even if the MIR shape or surrounding passes change.

Runs locally (./x test tests/ui/mir/simplify-comparison-integral-non-ssa-miscompile.rs → 1 passed)

r? @saethlin

[saethlin]

This adds that behavioral run-pass test, so the end-to-end miscompile
stays covered even if the MIR shape or surrounding passes change.

The bug is that a specific pass miscompiles a specific pattern, so I don't agree with this line of reasoning. Such end-to-end regression tests are vulnerable to spuriously passing because the bug has been reintroduced, but the rest of the compiler has changed.

The way you wrote the PR description makes me think you used an LLM to write it. I generally advise against that. Being able to write a good PR description on your own indicates that you actually understand the code.

[Vastargazing]

This adds that behavioral run-pass test, so the end-to-end miscompile
stays covered even if the MIR shape or surrounding passes change.

The bug is that a specific pass miscompiles a specific pattern, so I don't agree with this line of reasoning. Such end-to-end regression tests are vulnerable to spuriously passing because the bug has been reintroduced, but the rest of the compiler has changed.

The way you wrote the PR description makes me think you used an LLM to write it. I generally advise against that. Being able to write a good PR description on your own indicates that you actually understand the code.

Thanks, that makes total sense. I didn't realize a run-pass test could give a false green like that if the MIR shape changes later. Since #150925 already covers this via mir-opt, this PR is definitely redundant. Also to be honest I leaned on AI for the PR description 'cause digging into rustc solo is intimidating and i'm still finding my bearings here.

[saethlin]

You don't need to go at this alone, if you don't want to. There's a good community Discord server that has a thread about contributing to the compiler called #rustc-dev : https://discord.gg/rust-lang-community and there's an official Zulip but in my opinion it's more of an office and less of a hangout https://forge.rust-lang.org/platforms/zulip.html