EncodeUtf8 iterator implementation results in bad code

[nagisa]

Provided #35098 (more precisely f729e17 commit) lands, iterating x.encode_utf8().as_slice() results in considerably better code compared to x.encode_utf8().

One major problem with the EncodeUtf8 iterator is the fact that it uses checked indexing, despite it being obvious this iterator cannot index out-of-bounds. However, even if that problem is fixed, noticeably worse code is still generated, but I couldn’t exactly pinpoint why.

cc @alexcrichton #32204

[eefriedman]

Maybe it's a problem that EncodeUtf8::next() isn't marked #[inline]? Or are you not running into that issue because you're writing code inside libstd?

[nagisa]

Yes, I was writing code inside libstd (libcore, rather), so the method got inlined regardless of attribute (or lack of it thereof).

[eefriedman]

I looked into this a bit; there are a few related problems, but I think the biggest problem is that LLVM can't really figure out that accesses to EncodeUtf8.buf don't alias accesses to EncodeUtf8.pos, which prevents a lot of simplifications. There's been some discussion of this on the LLVM mailing lists; see http://lists.llvm.org/pipermail/llvm-dev/2016-July/102867.html and http://lists.llvm.org/pipermail/llvm-dev/2016-July/102472.html .

[nagisa]

Not relevant anymore as a different approach is now used, that generates pretty nice code.