Avoid boundchecking iteration

[carnaval]

Add an unsafe_next method which contract is that the iteration protocol was followed correctly. It has a generic fallback to next.
Lowering uses this instead of next. Wrapper iterators can also forward it.
This is demonstrated for Enumerate in this commit.

With this we can proudly say we compile a dead loop to nothing, yay...

[carnaval]

This will allow us to avoid sprinkinling @inbounds at least for the trivial cases. At some point we'll probably need a redundant @inbounds checker to be able to remove the useless ones.

[mbauman]

Not sure if this is still the case, but @inbounds return was difficult for inference to reason about at one point in recent history. You're probably more aware of the current design than I am, but just FYI: #13461.

[vtjnash]

would it be possible to use the new @boundscheck mechanism?

[carnaval]

I think boundscheck is too coarse since if lowering insert a @inbounds next() and someone has a custom iterator that happens to index into an array in next() then we would be incorrectly removing the bound check.

unsafe_next has the advantage that the contract is only about the iterator passed as an argument, which is the only thing for-loop lowering guarantees

[JeffBezanson]

I wonder if this plus #15259 will fix the performance gap in #15123 (which implements map using collect and zip).

[carnaval]

I'm open to suggestions for design, maybe unsafe_next is too painful and it apparently could use the new boundscheck thing according to Jameson. In any case it is kind of silly that we boundcheck one of the only guaranteed safe indexing construct in the language :-)

[blakejohnson]

I guess an alternative approach would be to directly inject (inbounds true) ... (inbounds false) around the call to next when lowering the expression.

[blakejohnson]

Or, write next as something like:

next(a::Array, i) = (@boundscheck checkbounds(a, i); @inbounds r=a[i]; (r, i+1))

Though, to ellide the bounds check you still need to decorate the next call site with @inbounds.

[vtjnash]

Closing, as I believe we do this now via @boundscheck

[yuyichao]

Do we? I think this is about removing bounds check automatically for for ele in a::Array without @inbounds The @boundscheck change was actually before this PR.

[stevengj]

Any chance of resurrecting this PR in some form?