Bound checks not eliminated in simple cases

[ghost]

I've found several simple cases where bound checks do not get eliminated: https://is.gd/OMzVUa

There are four variants of the same piece of code. It baffles me why bound checks don't get eliminated in the first three functions, and they finally do in the fourth (the function compiles to noop).

Any ideas?

[arthurprs]

Probably related to under/overflow confusion by the optimizer. Noticed the same when I was trying to optimize bounds check in binary search.

[nikic]

The bounds checks are optimized away since rust 1.25, see https://rust.godbolt.org/z/xTXz0n. However, there are two peculiarities:

a) In 1.25 the whole function is optimized away, but since 1.28 this is no longer the case. This looks like a SROA failure.

b) In nightly first and second are merged, but fourth isn't, even though the IR is strictly the same. This looks like a MergeFuncs bug.

[KamilaBorowska]

This appears to be fixed. first, second and fourth are merged. third is correctly optimized to always panic as a - 1 >= 0 isn't a correct check as a is usize.