Make show+mimetype work even if the backend has not decided to set that mime_type

[oxinabox]

If i run show(::IO, "image/png", figure) from the REPL it will do nothing. (and showable will return false) It would be useful if it actually did the thing. we have the technology in GR to save things as PNG In GR.jl: (for a filename) https://github.com/jheinen/GR.jl/blob/39d9cea254d4747ae7db55fdcbca38bc5a92bc17/src/jlgr.jl#L2495-L2501

In Plots.jl (i think easier to apply since it is already a IO) https://github.com/JuliaPlots/Plots.jl/blob/514a7c4fffd05441cf2e3dd1d444f40def7ba7ad/src/backends/gr.jl#L1765-L1788

As I understand it, right now, show(io, mimetype, ::Figure) only works if GR.mime_type is set to the right thing. AFAICT GR.mime_type is a weird global constant the GR sets at start time deoending on if it detect that it is in JuPyTer/Pluto or a bunch of other conditions. I don't get it, and I suspect it is evil, but it is out of scope here.

[oxinabox]

The reason I want this is that then FileIO.jl will work on figures and thus ReferenceTests.jl will work on them

[heliosdrm]

Thanks for the suggestion. To be honest the way in which show works for mime types different from text escapes my knowledge. For GRUtils, I just tried to replicate what GR does. Thus, for producing png files I also wrote a savefig function just like in GR: https://github.com/heliosdrm/GRUtils.jl/blob/f86dfc795eab4af029291b53ab83d02a1305e748/src/frontend.jl#L1437-L1449

Any hint about what should I do to make show(::IO, "image/png", ::Figure) work as wanted will be welcome. :pray:

[oxinabox]

Any hint about what should I do to make show(::IO, "image/png", ::Figure) work as wanted will be welcome. 🙏

I think literally copying what Plots.jl does for GR and updating the types.

Roughly speaking the way the whole display stack works: As I understand it (which may not be perfect).

High level: display(object) calls in and asked all current stack of displays to show an object, it finds one that can and then calls show

Breakdown:

• Julia has a stack of displays (each subtyping AbstractDisplay), they advertise what mimetypes they can display via displayable(this_display, mimetype.
  • Normally in the REPL you have 1 display that can display things like text/plain (you actually have 2 but anyway, both are just text)
  • But in Juno and VSCode you actually have an extra one for InlineDisplay that can do those inline bubbles
  • I think loading up Plots.jl, (possibly depending on the backend) adds an exta one calls a PlotsDisplay that can do various kinds of images.
  • IJulia has just 1 display, but that display can show tonnes of different mimetypes.
• Meanwhile objects advertize which mimetypes they can be shown as via showable(mimetype, object)
• A series of (simple) negotiations happen that eventually you end up with one of the displays saying "I will display this object", then it goes through the mimetypes it likes, and find the best one that is showable by this object and calls show(displays_own_io, best_mimetype, object)

[heliosdrm]

@oxinabox: I'm not sure if #94 fixes the issue in all the relevant cases. Can you try it please? I've made the change only for png and svg, which are the image formats that I have more or less clear how GR deals with.

Probably I could extend this for video formats (MIME"video/mp4" and `MIME"video/webm"). For other formats (pdf, ps...) this straightforward solution does not work.

[heliosdrm]

@oxinabox : is this issue solved by #94?

[oxinabox]

sorry, testing this is in my TODO, but I haven't gotten around to it yet

[oxinabox]

Tested. Works well for me