grep(pattern, ...) et al.: Optionally escalate warning on "argument 'pattern' has length > 1 ..." to an error

[HenrikBengtsson]

Issue

The grep() family of functions produces a warning if pattern specified more than one regular expression, e.g.

> pattern <- c("Volvo", "Fiat")
> grep(pattern, rownames(mtcars))
[1] 32
Warning message:
In grep(pattern, rownames(mtcars)) :
  argument 'pattern' has length > 1 and only the first element will be used

I cannot come up with a valid use case where using length(pattern) != 1 is not a mistake. Because of this, I argue it's a bug. Because warnings might be ignored, or drown in the noise of other warnings, this is almost as problematic as a silent bug. There is a risk that there is code out there that produces incorrect results, without being noticed.

I've checked R 4.3.1 and found the following functions with this lax behavior:

In the base package:

• agrep()
• agrepl()
• grep()
• grepl()
• grepRaw()
• regexpr()
• gregexpr()
• regexec()
• gsub()
• sub()
• ...?

There might be more.

Suggestion

For the sprint

As a starter, introduce an "internal" environment variable, e.g. _R_CHECK_LENGTH_1_PATTERN_, that can be used to escalate the warning to an error:

> Sys.setenv("_R_CHECK_LENGTH_1_PATTERN_" = "true")

> pattern <- c("Volvo", "Fiat")
> grep(pattern, rownames(mtcars))
[1] 32
Error in grep(pattern, rownames(mtcars)) : 
  argument 'pattern' has length != 1

Long-term

I'd argue that using length(pattern) > 1 is a bug, similar to how length(cond) > 1 is a bug in if (cond) { ... }. For the latter, we started out with _R_CHECK_LENGTH_1_CONDITION_=true as above in R (>= 3.4.0), but since that was too harsh to enable on CRAN, support for per-package checking was added as:

_R_CHECK_LENGTH_1_CONDITION_=package:_R_CHECK_PACKAGE_NAME_,verbose

in R (>= 4.0.0). That would only escalate to an error, if the buggy code was part of the package tested, otherwise it remained a warning. This allowed CRAN to have packages fixed one by one without breaking the package ecosystem. Eventually, enough packages were fixed, and they made _R_CHECK_LENGTH_1_CONDITION_=true the default. And in R (>= 4.2.0), erroring was the only outcome and _R_CHECK_LENGTH_1_CONDITION_ could be dropped.

A similar path forward can be made to fix length(pattern) > 1 bugs, which means the next step after the above would be to add support for:

_R_CHECK_LENGTH_1_PATTERN_=package:_R_CHECK_PACKAGE_NAME_,verbose

There is no need to implement the latter during the sprint. Having _R_CHECK_LENGTH_1_PATTERN_=true will be a big step forward.

Comment: I think it was @kalibera who implemented _R_CHECK_LENGTH_1_CONDITION_=package:_R_CHECK_PACKAGE_NAME_,verbose.

PS. This proposal was taken from HenrikBengtsson/Wishlist-for-R#122.

[HenrikBengtsson]

I've submitted this to BugZilla, cf. PR18577.