Bounds check not optimized out after {r}position on a slice

[dtolnay]

All the unrolling makes it hard for llvm to optimize it out. Sample: https://godbolt.org/g/R9nLvC

#45501 has a fix that involves assume(index < len) but we couldn't quite get it to pass tests. Will need to give it another try with some fresh eyes.

[kennytm]

According to #44899 (comment) there was a fix in upstream LLVM. We may reopen the PR if we backport the fix.

[arielb1]

@kennytm

I already backported the fix, it's just that someone needs to make an "up the LLVM" PR that includes it.

[kennytm]

Reopening this because of #48209.

(We can close again if it turns out the assume is not necessary for optimizing out the bounds check on LLVM 6.)

[scottmcm]

It's possible #49551 could help with this, as it should make it easier for LLVM to understand the length of a slice iterator 🤞