`ggplot_add.list()` uses the `%+%` function instead of the `+` generic.

[eliocamp]

ggplot_add.list() adds each element of the list using %+% instead of + here:

ggplot2/R/plot-construction.R

Lines 156 to 161 in ca47270

	ggplot_add.list <- function(object, plot, object_name) {
	for (o in object) {
	plot <- plot %+% o
	}
	plot
	}

The idea is that ggplot() + list(e1, e2, e3) should be the same as ggplot() + e1 + e2 + e3, but this is not true if someone (😇) wanted to create a gg subclass with a custom + method.

I'm not sure why that code uses %+% instead of +. I modified the code and run tests both ways and everything passes. This logic has been since 3c16f9c from 2017

Would it be possible to change ggplot_add.list() to use the + generic? Barring that, would it be possible to make %+% an exported generic (same as +)?

[teunbrand]

I think that %+% here is used to allow member of the list to be a <data.frame>, which cannot use + due to S3 method precedence. It was discussed in #3818 that double dispatch would be the preferred way of handling addition of objects to subclassed plots. At the time S7 didn't come out, but now that it is on CRAN, we might consider adding the double dispatch.

[eliocamp]

Ah, that's why! I might need to read up on S7, apparently.

Then, in the meantime, how about exporting a %+% generic? That would also work.

[teunbrand]

Can you give an (dummy) example of what you'd like to work that currently doesn't?
There might be no need to do double dispatch or reconfiguring %+% as a generic if there is an easier way out.
With 'an easier way out', I might mean something like the following:

#' @export
ggplot_add.list <- function(object, plot, object_name) {
  for (o in object) {
    plot <- ggplot_add(o, plot, object_name)
  }
  plot
}

[eliocamp]

I'm trying to change ggnewscale so that thenew_scale() call modifies the next layers instead of the previous ones. This solves a hard-to-debug problem and enables renaming aes to something meaningful. So I'm doing this:

"+.gg_subclass" <- function(e1, e2) {

  e2 <- rename_aes(layer = e2)

  NextMethod()
}

So layers get modified before being added to the plot.

The problem is that something like this:

ggplot(mapping = aes(x, y)) +
  list(
    geom_contour(....),
    scale_color_viridis_c(...),

    new_scale_color(),

    geom_point(...),
    scale_color_viridis_c(option = "A")
  )

doesn't work, because the first + calls ggplot_add.list(), which then uses %+% to add the elements, bypassing my custom method.

[eliocamp]

Another possibility would be to have somethig like

 ggplot_add.list <- function(object, plot, object_name) { 
   for (o in object) { 
    if (is.data.frame(object)) {
       plot <- plot %+% o 
    } else {
      plot <- plot + o 
   } 
   plot 
 } 

[teunbrand]

Thanks, yes that makes sense to me. I still think that overriding +.class is an undesired approach because of S3 method precedence hairiness, and that instead this problem would be better solved by double dispatch. I'll experiment around a bit to see how feasible it is to get double dispatch working (with or without S7).

[billdenney]

For a method that would work starting at R version 4.3, we could add a chooseOpsMethod.gg function.

I assume that ggplot2 itself would not want to require version 4.3, but it would allow other packages to take advantage of not needing %+%, if they were comfortable requiring 4.3.

[teunbrand]

Interesting, I was only peripherally aware of chooseOpsMethod, but that looks promising!

[billdenney]

It's what makes the ggtibble package possible! 😄