Help optimize out bounds checks in median_of_medians

[kornelski]

LLVM can't see that assert!((v / 200) <= (v / 2)) is always true, so it couldn't remove bounds in median_of_ninthers.

.unwrap() outside of min_index() was losing bounds information. I've even tried assume(index < slice.len()); return Some(index), but it didn't help. Only moving the unwrap() to inside the function seems to keep the range information (and doesn't need assume).

Part of #144326

Currently blocked by #144522

[rustbot]

r? @tgross35

rustbot has assigned @tgross35.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[tgross35]

Is this something that could get a codegen test to make sure we don't have a jump to a panic? Seems easy to accidentally regress.

[kornelski]

A codegen test would be valuable, but these functions are inlined into other functions that still need bounds checks fixed, so I can't realistically codegen tests until more of this work is done.

[Dylan-DPC]

@kornelski any updates on this? thanks

[kornelski]

I've implemented further refactorings that should eliminate further bounds checks in theory, but tracking of already-checked ranges of variables turned out to be too fickle in LLVM.

I got stuck at ninther that has 9 variables to check, but LLVM gets amnesia after every else, so this function would have to be full of unsafe. This is blocked by llvm/llvm-project#151078 #144522

I can't add codegen tests, because I can't make LLVM generate good code.