treat .= as syntactic sugar for broadcast!

[stevengj]

As discussed in #16285, this PR treats x .= ... as syntactic sugar for broadcast!(identity, x, ...), and furthermore fuses the broadcast loop with any nested "dot" calls. e.g. x .= sin.(y) becomes broadcast!(sin, x, y).

The tricky parts were all done in #17300, so this PR seems fairly straightforward.

To do:

• Tests
• Documentation
• For vectors, special case broadcast!(identity, x, scalar) to call fill! and broadcast!(identity, x, y) to call copy!.
• Fix expression printing for dotted assignments.

This should be a very non-disruptive change, because the .= assignment syntax is new.

Note that x .+= y and so on is now transformed to x .= x .+ y. However, dot operators like x .+ y are not yet treated as dot calls (+).(x,y) (and hence are not yet fused), which will be a more disruptive change.

[StefanKarpinski]

Good to get these in before 0.5 (might as well take advantage of the otherwise unfortunate delays).

[stevengj]

Should be good to go as soon as I add a NEWS item. (I'd prefer to do that right before merging since NEWS items generate so many conflicts.)

I'm seeing some apparently unrelated Travis and AppVeyor failures on certain platforms, but it appears that master has the same problems.

[tkelman]

does .$= not work?

[stevengj]

@tkelman, .$= is not supported by the parser – it was not added by #17393 because the $ is pretty special (e.g. in macros you can do a.$b) and it seemed like it would require more parser hacking to support .$= than it was worth at this point in the 0.5 cycle.

[stevengj]

@StefanKarpinski, @tkelman, tests are green. Okay to merge once I add a NEWS item?

[StefanKarpinski]

Ok by me. May warrant a @JeffBezanson review due to flisp slinging.

[stevengj]

The flisp changes are pretty trivial (it only looks like a lot of lines because I did a little refactoring of #17300, but most of those lines are just indentation changes and similar).

[stevengj]

Uh oh, I just realized a problem with this PR for backward compatibility.

If you have an expression like:

x[:] .*= 2

then it already updates x in-place. However, that now turns into broadcast!(x -> x .* 2, x[:], x[:]), and because x[:] makes a copy (#13157) this actually does not update x.

I think I need to update the PR so that it detects when the lhs is a getindex expression and calls view to get a view.

[stevengj]

Okay, I have a PR almost ready to fix the x[...] .= ... problems. Fortunately, the fix was easy.