18

From this question about whether printf is a built-in for yash, comes this answer that quotes the POSIX standard.

The answer points out that the POSIX search sequence is to find an external implementation of the desired command, and then, if the shell has implemented it as a built-in, run the built-in. (For built-ins that aren't special built-ins.)

Why does POSIX have this requirement for an external implementation to exist before allowing an internal implementation to be run?

It seems... arbitrary, so I am curious.

  • I believe that is a way to enable/disable builtins if desired/required. – Isaac Jan 23 at 22:43
  • 2
    Disabling the built-in by removing the external implementation? Now there are no commands of name printf available. – studog Jan 23 at 22:49
  • @studog, so create an empty file with the same name as the built-in, turn on the execute bit, and put it in a directory in your PATH. :P – Wildcard Jan 23 at 23:13
  • @Wildcard A strictly compliant shell would then see the name while searching the PATH and then call the built-in utility, not the external script. What if you'd want to call the external script in your path? Hmm... This seems to call for a table describing the different possibilities. There is one here, but it doesn't make sense to me. – Kusalananda Jan 23 at 23:23
  • @Kusalananda, re your first sentence, that was my point. Hence why I said to create an empty file. – Wildcard Jan 23 at 23:28
15

This is an "as if" rule.

Simply put: The behaviour of the shell as users see it should not change if an implementation decides to make a standard external command also available as shell built-in.

The contrast that I showed at https://unix.stackexchange.com/a/496291/5132 between the behaviours of (on the one hand) the PD Korn, MirBSD Korn, and Heirloom Bourne shells; (on the other hand) the Z, 93 Korn, Bourne Again, and Debian Almquist shells; and (on the gripping hand) the Watanabe shell highlights this.

For the shells that do not have printf as a built-in, removing /usr/bin from PATH makes an invocation of printf stop working. The POSIX conformant behaviour, exhibited by the Watanabe shell in its conformant mode, causes the same result. The behaviour of the shell that has a printf built-in is as if it were invoking an external command.

Whereas the behaviour of all of the non-conformant shells does not alter if /usr/bin is removed from PATH, and they do not behave as if they were invoking an external command.

What the standard is trying to guarantee to you is that shells can build-in all sorts of normally external commands (or implement them as its own shell functions), and you'll still get the same behaviour from the built-ins as you did with the external commands if you adjust PATH to stop the commands from being found. PATH remains your tool for selecting and controlling what commands you can invoke.

(As explained at https://unix.stackexchange.com/a/448799/5132, years ago people chose the personality of their Unix by changing what was on PATH.)

One might opine that making the command always work irrespective of whether it can be found on PATH is in fact the point of making normally external commands built-in. (It's why my nosh toolset just gained a built-in printenv command in version 1.38, in fact. Although this is not a shell.)

But the standard is giving you the guarantee that you'll see the same behaviour for regular external commands that are not on PATH from the shell as you will see from other non-shell programs invoking the execvpe() function, and the shell will not magically be able to run (apparently) ordinary external commands that other programs cannot find with the same PATH. Everything works self-consistently from the user's perspective, and PATH is the tool for controlling how it works.

Further reading

13

That's quite absurd and that's why no shell is implementing it in its default mode.

The standard's rationale and its illustrating example suggest that this was a botched attempt to have a regular built-in associated with a path, and let the user override it by having their own binary appear before it in PATH (eg. a printf built-in associated with /usr/bin/printf could be overridden by the /foo/bin/printf external command by setting PATH=/foo/bin:$PATH).

However, the standard did not end up requiring that, but something completely different (and also useless and unexpected).

You can read more about it in this bug report. Quoting from from the final accepted text:

Many existing implementations execute a regular built-in without performing a PATH search. This behavior does not match the normative text, and it does not allow script authors to override regular built-in utilities via a specially crafted PATH. In addition, the rationale explains that the intention is to allow authors to override built-ins by modifying PATH, but this is not what the normative text says.

FWIW, I don't think there's any shell implementing the revised requirements from the accepted text, either.

  • See also the discussion at article.gmane.org/gmane.comp.standards.posix.austin.general/… (and there have been several others). – Stéphane Chazelas Jan 24 at 17:29
  • Also github.com/att/ast/issues/370 (long) – Stéphane Chazelas Jan 24 at 17:41
  • No, (eg. a printf built-in associated with /usr/bin/printf could be overridden by the /foo/bin/printf external command by setting PATH=/foo/bin:$PATH)., that is incorrect. The existence of either/both/any of /usr/bin/printf or /foo/bin/printf in the PATH will activate the builtin printf. The only thing that a missing (in the PATH) external printf will do is to disable the builtin. (By the letter of the spec). – Isaac Feb 1 at 5:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.