The Universe of Discourse


Wed, 16 Jul 2014

Guess what this does (solution)

A few weeks ago I asked people to predict, without trying it first, what this would print:

 perl -le 'print(two + two == five ? "true" : "false")'

(If you haven't seen this yet, I recommend that you guess, and then test your guess, before reading the rest of this article.)

People familiar with Perl guess that it will print true; that is what I guessed. The reasoning is as follows: Perl is willing to treat the unquoted strings two and five as strings, as if they had been quoted, and is also happy to use the + and == operators on them, converting the strings to numbers in its usual way. If the strings had looked like "2" and "5" Perl would have treated them as 2 and 5, but as they don't look like decimal numerals, Perl interprets them as zeroes. (Perl wants to issue a warning about this, but the warning is not enabled by default. Since the two and five are treated as zeroes, the result of the == comparison are true, and the string "true" should be selected and printed.

So far this is a little bit odd, but not excessively odd; it's the sort of thing you expect from programming languages, all of which more or less suck. For example, Python's behavior, although different, is about equally peculiar. Although Python does require that the strings two and five be quoted, it is happy to do its own peculiar thing with "two" + "two" == "five", which happens to be false: in Python the + operator is overloaded and has completely different behaviors on strings and numbers, so that while in Perl "2" + "2" is the number 4, in Python is it is the string 22, and "two" + "two" yields the string "twotwo". Had the program above actually printed true, as I expected it would, or even false, I would not have found it remarkable.

However, this is not what the program does do. The explanation of two paragraphs earlier is totally wrong. Instead, the program prints nothing, and the reason is incredibly convoluted and bizarre.

First, you must know that print has an optional first argument. (I have plans for an article about how optional first arguments are almost always a bad move, but contrary to my usual practice I will not insert it here.) In Perl, the print function can be invoked in two ways:

   print HANDLE $a, $b, $c, …;
   print $a, $b, $c, …;

The former prints out the list $a, $b, $c, … to the filehandle HANDLE; the latter uses the default handle, which typically points at the terminal. How does Perl decide which of these forms is being used? Specifically, in the second form, how does it know that $a is one of the items to be printed, rather than a variable containing the filehandle to print to?

The answer to this question is further complicated by the fact that the HANDLE in the first form could be either an unquoted string, which is the name of the handle to print to, or it could be a variable containing a filehandle value. Both of these prints should do the same thing:

  my $handle = \*STDERR;
  print STDERR $a, $b, $c;
  print $handle $a, $b, $c;

Perl's method to decide whether a particular print uses an explicit or the default handle is a somewhat complicated heuristic. The basic rule is that the filehandle, if present, can be distinguished because its trailing comma is omitted. But if the filehandle were allowed to be the result of an arbitrary expression, it might be difficult for the parser to decide where there was a a comma; consider the hypothetical expression:

   print $a += EXPRESSION, $b $c, $d, $e;

Here the intention is that the $a += EXPRESSION, $b expression calculates the filehandle value (which is actually retrieved from $b, the $a += … part being executed only for its side effect) and the remaining $c, $d, $e are the values to be printed. To allow this sort of thing would be way too confusing to both Perl and to the programmer. So there is the further rule that the filehandle expression, if present, must be short, either a simple scalar variable such as $fh, or a bare unquoted string that is in the right format for a filehandle name, such as HANDLE. Then the parser need only peek ahead a token or two to see if there is an upcoming comma.

So for example, in

  print STDERR $a, $b, $c;

the print is immediately followed by STDERR, which could be a filehandle name, and STDERR is not followed by a comma, so STDERR is taken to be the name of the output handle. And in

  print $x, $a, $b, $c;

the print is immediately followed by the simple scalar value $x, but this $x is followed by a comma, so is considered one of the things to be printed, and the target of the print is the default output handle.

In

  print STDERR, $a, $b, $c;

Perl has a puzzle: STDERR looks like a filehandle, but it is followed by a comma. This is a compile-time error; Perl complains “No comma allowed after filehandle” and aborts. If you want to print the literal string STDERR, you must quote it, and if you want to print A, B, and C to the standard error handle, you must omit the first comma.

Now we return to the original example.

 perl -le 'print(two + two == five ? "true" : "false")'

Here Perl sees the unquoted string two which could be a filehandle name, and which is not followed by a comma. So it takes the first two to be the output handle name. Then it evaluates the expression

     + two == five ? "true" : "false"

and obtains the value true. (The leading + is a unary plus operator, which is a no-op. The bare two and five are taken to be string constants, which, compared with the numeric == operator, are considered to be numerically zero, eliciting the same warning that I mentioned earlier that I had not enabled. Thus the comparison Perl actually does is is 0 == 0, which is true, and the resulting string is true.)

This value, the string true, is then printed to the filehandle named two. Had we previously opened such a filehandle, say with

open two, ">", "output-file";

then the output would have been sent to the filehandle as usual. Printing to a non-open filehandle elicits an optional warning from Perl, but as I mentioned, I have not enabled warnings, so the print silently fails, yielding a false value.

Had I enabled those optional warnings, we would have seen a plethora of them:

Unquoted string "two" may clash with future reserved word at -e line 1.
Unquoted string "two" may clash with future reserved word at -e line 1.
Unquoted string "five" may clash with future reserved word at -e line 1.
Name "main::two" used only once: possible typo at -e line 1.
Argument "five" isn't numeric in numeric eq (==) at -e line 1.
Argument "two" isn't numeric in numeric eq (==) at -e line 1.
print() on unopened filehandle two at -e line 1.

(The first four are compile-time warnings; the last three are issued at execution time.) The crucial warning is the one at the end, advising us that the output of print was directed to the filehandle two which was never opened for output.

[ Addendum 20140718: I keep thinking of the following remark of Edsger W. Dijkstra:

[This phenomenon] takes one of two different forms: one programmer places a one-line program on the desk of another and … says, "Guess what it does!" From this observation we must conclude that this language as a tool is an open invitation for clever tricks; and while exactly this may be the explanation for some of its appeal, viz., to those who like to show how clever they are, I am sorry, but I must regard this as one of the most damning things that can be said about a programming language.

But my intent is different than what Dijkstra describes. His programmer is proud, but I am disgusted. Incidentally, I believe that Dijkstra was discussing APL here. ]

[ Addendum 20150508: I never have much sympathy for the school of thought that says that you should always always enable warnings in every Perl program; I think Perl produces too many spurious warnings for that. But I also think this example is part of a cogent argument in the other direction. ]


[Other articles in category /prog/perl] permanent link