array display is a mess

[StefanKarpinski]

Recent changes to array printing have made it impossible to normally provide show methods for custom array types. The writemime function intercepts printing of arrays, so that if you define a new type, e.g. Bytes <: AbstractVector{Uint8} and then define show(io::IO, b::Bytes), that show method never gets called. This whole situation is a pretty huge mess, imo. There's no good reason for show and writemime for arrays to produce such different results and you certainly shouldn't have to know about or override writemime methods to get a custom array type to show up the way you want it to in the repl.

[stevengj]

Didn't a similar problem exist before writemime? Previously, the REPL output called repl_show, which called print_matrix for arrays, as I understand it, rather than show.

Which is not to say that we shouldn't clean up the situation.

One option would be to make writemime call show rather than showarray.

Another would simply be to deprecate show and tell people that they should overload writemime(io, ::MIME"text/plain", x). It seems like, currently, show is for a more "faithful" representation and writemime is for user-readable display, but I'm not sure why this distinction is important since neither text/plain nor show are particularly good formats for data interchange or serialization.

[andreasnoack]

...or UniformScaling<:AbstractMatrix. I couldn't figure out if there was a new right way to define show methods and therefore ended up defining a showarray method. A show(io,Array) method would have felt more julian to me.

[JeffBezanson]

The purpose of showarray is to pass arguments describing how to print the array, because we want several variants. We have to either live with this or switch to a different mechanism for passing these values.

[JeffBezanson]

See also #5709

[JeffBezanson]

Here's a possible way forward that deals with each of the extra parameters in showarray:

1. Leave printing headers to writemime, take it out of show
2. Use a global for limiting output (we have one already)
3. Call tty_rows and tty_cols to get the screen size if you want
4. Make the repr and writemime representations more similar, for example

Float32[
 1.0  1.0  1.0
 1.0  1.0  1.0
 1.0  1.0  1.0]

It is very helpful if the "pretty" output and the "useful" (e.g. parsable) output are the same.

[malmaud]

This also relates to #5564.

[mcg1969]

I just happened upon this issue myself. I created a new abstract array type for which I specifically do not want to print the individual values. I had assumed that overloading show would be sufficient, but of course that failed. I eventually discovered that overloading writemime is necessary to accomplish this.

Given that show(io::IO, X::AbstractArray) immediately calls showarray, it's not clear to me why writemime couldn't just call show instead of showarray. In fact, as the code is currently written, in can be. Here are the two functions:

show(io::IO, X::AbstractArray) = showarray(io, X, header=_limit_output, repr=!_limit_output)
writemime(io::IO, ::MIME"text/plain", v::AbstractArray) =
    with_output_limit(()->showarray(io, v, header=true, repr=false))

Because with_output_limit flips the global _limit_output flag, you should get the same results with this implementation---unless show is overloaded, which of course is what i want :-)

writemime(io::IO, ::MIME"text/plain", v::AbstractArray) =
    with_output_limit(()->show(io, v))

[stevengj]

@mcg1969, that seems reasonable to me.