print.data.table: check for knitr at sys.calls()-6

[aitap]

This is a band-aid for #6509. Previous check for mimicsAutoPrint (i.e. knit_print.default) at sys.calls()[[length(sys.calls())-3L]] no longer works after a knitr upgrade. Currently, the call to the autoprint-like function resides at position -6.

Edit: unfortunately there doesn't seem to be a right way to solve this problem. Returning invisible(anything) from [.data.table sets it back to visible because PRIMPRINT(.Primitive("[")) says so, and knitr respects visibility of the evaluation results.

[github-actions]

Generated via commit ffeb415

Download link for the artifact containing the test results: ↓ atime-results.zip

Time taken to finish the standard R installation steps: 11 minutes and 25 seconds

Time taken to run atime::atime_pkg on the tests: 1 minutes and 50 seconds

[tdhock]

I don't really understand the issue/fix, so comments would be super helpful here.
Could you please add somethign like # suppress printing in knitr, #6509
and maybe some explanation about where the magic number 6 comes from?

[aitap]

Does ffeb415 help?

The original problem is that [.data.table would like to return normally for DT[<any read-only operation>] but invisibly for DT[,<var> := <assignment by reference>].

Unfortunately, when evaluating calls to primitives, R checks a special flag for whether to restore the visibility, and the primitive `[` has this flag set to 0 (force R_Visible = TRUE), so there doesn't seem to be a way to return invisible(anything) from [.data.table.

To achieve something like that, data.table remembers the address of the last data.table on which a by-reference operation has been computed. print.data.table checks if the address of the argument matches the remembered one and, if it suspects it's being called for auto-printing (call stack very shallow or called by a specific knitr function), it refuses to print anything. After something in the knitr stack got upgraded, the check for "am I being called for auto-printing by knitr" no longer works. We could also change the check to "is knit_print.default anywhere in the call stack, not just at positions -3 or -6" and risk false positives.

We could try asking R core to change the flag on .Primitive("["), but almost always forcing visibility on for results of `[` is the right thing and it'll take years before we can enjoy one less workaround in the codebase.

[MichaelChirico]

Thanks for the really good summary, really this should be in the codebase somewhere.

[tdhock]

great comments thanks!

for a medium-term solution I wonder if rendering vignettes using litedown would fix this, or cause other problems? https://yihui.org/litedown/#sec-package-vignettes

haha

[MichaelChirico]

for a medium-term solution I wonder if rendering vignettes using litedown would fix this, or cause other problems?

Not really. The test is here to ensure that anyone using {knitr}+{data.table} gets the expected behavior when knitting their vignettes, regardless of what engine we use.