Rscript behaves inconsistently on windows with single and double quotes

Question

If I invoke

Rscript -e "print('hello')"

It correctly prints out the answer

[1] "hello"

However, if I switch single and double quotes, it does not work, and it looks like the double quotes are removed:

Rscript -e 'print("hello")'

gives:

Error in print(hello) : object 'hello' not found
Execution halted

Note that it's not powershell doing the escaping incorrectly. Echoing only gives the expected results:

PS> echo 'print("hello")'
print("hello")
PS> echo "print('hello')"
print('hello')

And the same behavior is not observed on macOs or Linux, where both variants are correctly parsed.

Interestingly enough, it's even crazier for command.com:

C:>Rscript -e "print('hello')"
[1] "hello"

C:>Rscript -e 'print("hello")'
[1] "print(hello)"

I mean... what?!? This has already been mentioned here:

Single line code to run R code from Windows command line

but there's no explanation about it. In my opinion it's a bug of Rscript on windows, but I want to hear other opinions.

Answer 1

Dabombber's helpful answer provides all the pointers, but let me try to boil it down conceptually:

The problem is not specific to RScript.exe and potentially affects calls to any external executable from PowerShell:

Up to at least PowerShell 7.1 (current as of this writing), passing arguments with embedded double quotes (") to external programs is fundamentally broken, as detailed in GitHub issue #1995; in short: behind the scenes, PowerShell constructs a command line for the target program (process) that uses "..."-quoting only, but neglects to escape embedded verbatim " chars. for their syntactically valid inclusion in such double-quoted strings; a fix may be coming in v7.2 - see this answer.

• For now, you have to manually escape embedded " chars. as \".

• However, if and when the bug gets fixed, this workaround will break, because the fix requires that this escaping be applied automatically, which would then escape a verbatim \" as \\\".

# WORKAROUND as of v7.0, which will break if and when the problem gets fixed.
PS> Rscript -e 'print(\"hello\")'

The third-party Native module (install with Install-Module -Scope CurrentUser Native, for instance) offers helper function ie, which compensates for the broken behavior; it is written in a forward-compatible manner so that it will simply defer to the built-in behavior if and when it should get fixed:

# Thanks to `ie`, no workarounds are required.
PS> ie Rscript -e 'print("hello")'

As for ad hoc workarounds - both of them work for Rscript.exe, but can't be expected to be a general solution:

• For target programs that support both '...' and "..." quoting: Swap the quotes to use only embedded ' chars., as shown in your question, but note that '...' and "..." strings have different semantics in PowerShell ("..." strings are expandable (interpolating) strings), and may have different semantics in the target program too (not the case in Rscript):

  • Rscript -e "print('hello')"

• For target programs that accept input via stdin, use the PowerShell pipeline, where the bug doesn't surface (though note that you may have to set the $OutputEncoding preference variable to the character encoding expected by the target program):

  • 'print("hello")' | Rscript -

---

As for your observations and background information, including about cmd.exe and POSIX-compatible shells:

Note that it's not powershell doing the escaping incorrectly.

As Dabombber points out, it is PowerShell that is the problem, but the problem only occurs when calling external programs, whereas echo is a built-in alias for the PowerShell-native
Write-Output cmdlet (verify with Get-Command echo). On Windows, you could see the problem with the flawed parameter passing as follows, by invoking choice.exe (ignore the [Y,N]?N suffix):

PS> 'n' | choice /m 'print("hello")'
print(hello) [Y,N]?N

choice.exe with /m can be used to echo an argument as it would be received by external programs, and as you can see the double quotes were effectively lost, because PowerShell mistakenly placed print("hello") verbatim on the process command line - without escaping the " chars. - which external programs parse as verbatim print(hello), because they allow a single argument to be composed of unquoted and double-quoted parts (print( + hello (stripped of the syntactic double quotes) + )).

• If verbatim print(hello) is interpreted as an R script, it looks for a variable (object) named hello - which in this scenario doesn't exist and triggers the error message you saw.

On Unix-like platforms (macOS, Linux), using the cross-platform PowerShell [Core] edition, /bin/echo 'print("hello")' shows the same problem.

And the same behavior is not observed on macOs or Linux, where both variants are correctly parsed.

Yes, if you use a native, POSIX-compatible shell there, such as bash, you'll get the correct behavior (see below).

it's even crazier for command.com:

As an aside: You probably meant cmd.exe, the legacy command processor (Command Prompt) on NT-based Windows platforms up to the current Windows 10. (command.com was the command processor on the extinct DOS-based Windows versions that ended with Windows ME).

cmd.exe only recognizes double-quoting ("...") to demarcate argument boundaries for itself, not also single-quoting ('...'); irrespective of that, it essentially passes the original quoting through to the target executable (after performing its own interpretation of the command line, such as environment-variable expansion).

This differs fundamentally from what PowerShell and POSIX-compatible shells do:

• On Unix-like platforms - where POSIX-compatible shells recognize '...'-quoted arguments - the concept of a process command line doesn't exist, and whatever arguments a POSIX-like shell has itself parsed out of its command line are passed as-is - as an array of verbatim arguments - to the target executable; thus shell string literals "print('hello')" and 'print("hello")' are passed as verbatim print('hello') and print("hello"), respectively, which works as expected, given that R too recognizes both '...' and "..." string literals.

• PowerShell too has '...' strings (it treats their content verbatim), but on Windows it translates them to "..." strings behind the scenes (if quoting is needed), which is where the aforementioned bug can surface as of v7.0. The bug aside, this translation makes sense, because only "..." quoting can be assumed to have syntactic meaning on the command line for other programs (see bottom section). Unfortunately, PowerShell does the same thing on Unix-like platforms, even though it shouldn't (it constructs a pseudo command line that the .NET API then parses into an array of verbatim arguments passed to the target process), so the bug surfaces there as well.

Because cmd.exe preserves the original quoting, RScript interprets 'print("hello")' in command line Rscript -e 'print("hello")' as a string literal rather than as a command, because it removes any " chars. with syntactic function on the command line first (whereas ' (single quotes) by convention do not have syntactic meaning on the command line), before the result is interpreted as an R script:

• 'print("hello")' is therefore parsed as 'print( + hello (the command-line " are stripped) + ), resulting in verbatim 'print(hello)' getting interpreted as R code, which is an R string literal that therefore prints as-is (the output uses "..." quoting, but that's just an artifact of output formatting; note that an explicit call to print() isn't necessary, the result of an expression - such as string literal 'print(hello)' in this case - is automatically printed).

• By contrast, "print('hello')" is parsed as verbatim print('hello') (the command-line " are stripped), which - due to the absence of enclosing quoting - is then interpreted as a command, namely a print() function call, as intended.

---

Ultimately, there are no hard and fast rules in the anarchic world of process command-line parsing on Windows: it is ultimately up to each program to interpret its command line - this answer contains excellent background information.

Fortunately, however, there are widely adhered-to conventions, as implemented in the MS C/C++/.NET compilers and documented here.

Unfortunately, as of PowerShell 7.0, PowerShell doesn't adhere to these conventions, due to the aforementioned bug. Since the bug has been around since v1, users have learned to work around it, such as with manual \"-escaping, as shown above. The problem is that fixing the bug would break all workarounds. Implementing a fix as an experimental feature is now being considered, for v7.1 at the earliest - see this PR on GitHub and the associated discussion here, which suggests that, in addition to implementing the widely established conventions, accommodations be made for calls to batch files and msiexec.exe-style programs, which have non-conventional quoting requirements.

Answer 2

It might be worth taking a look through this PowerShell issue: Arguments for external executables aren't correctly escaped. The Native module by Michael Klement provides a workaround until the problem is fixed (and shouldn't be broken post-fix like many current workarounds).

Note that it's not powershell doing the escaping incorrectly. Echoing only gives the expected results

echo is a PowerShell function rather than an external program so you won't notice the broken behaviour when using it.

PS> Get-Command echo

CommandType     Name                                               Version    Source
-----------     ----                                               -------    ------
Alias           echo -> Write-Output

A better test would be to use the EchoArgs.exe command line tool from PowerShell Community Extensions (downloadable here).

PS> echoargs.exe 'print("hello")'
Arg 0 is <print(hello)>

Command line:
"E:\echoargs.exe"  print("hello")

PS> echoargs.exe "print('hello')"
Arg 0 is <print('hello')>

Command line:
"E:\echoargs.exe"  print('hello')