Sys::Binmode - A fix for Perl’s system call character encoding

Coverage Status


    use Sys::Binmode;

    my $foo = "\xff";
    $foo .= "\x{100}";
    chop $foo;

    # Prints a single octet (0xFF) and a newline:
    print $foo, $/;

    # In Perl 5.32 this may print the same single octet, or it may
    # print UTF-8-encoded U+00FF. With Sys::Binmode, though, it always
    # gives the single octet, just like print:
    exec 'echo', $foo;


tl;dr: Use this module in all new code.


Ideally, a Perl application doesn’t need to know how the interpreter stores a given string internally. Perl can thus store any Unicode code point while still optimizing for size and speed when storing “bytes-compatible” strings—i.e., strings whose code points all lie below 256. Perl’s “optimized” string storage format is faster and less memory-hungry, but it can only store code points 0-255. The “unoptimized” format, on the other hand, can store any Unicode code point.

Of course, Perl doesn’t always optimize “bytes-compatible” strings; Perl can also, if it wants, store such strings “unoptimized” (i.e., in Perl’s internal “loose UTF-8” format), too. For code points 0-127 (ASCII printables, controls, and DEL) there’s actually no difference between the two forms, but for 128-255 the formats differ. (cf. "The "Unicode Bug"" in perlunicode) This means that anything that reads Perl’s internals MUST differentiate between the two forms in order to use the string correctly.

Alas, that differentiation doesn’t always happen. When it doesn’t, Perl outputs code points 128-255 differently depending on whether the containing string is “optimized” or not.

Remember, though: Perl applications should not care about Perl’s string storage internals like optimized/unoptimized. (This is why, for example, the bytes pragma is discouraged.) The catch, though, is that without that knowledge, the application can’t know what it actually says to the outside world!

Thus, applications must either monitor Perl’s string-storage internals or accept unpredictable behavior, both of which are categorically bad.

(Perl’s documentation calls the “unoptimized” format “upgraded”, while it calls the “optimized” format “downgraded”. The rest of this document will favor Perl’s terms.)


This module provides predictable behavior for Perl’s built-in functions by downgrading all strings before giving them to the operating system. It’s equivalent to—but faster than!—prefixing your system calls with utf8::downgrade() (cf. utf8) on all arguments.

Predictable behavior is always a good thing; ergo, you should use this module in all new code.


If you apply this module injudiciously to existing code you may see exceptions or character corruption where previously things worked fine.

This can happen if you’ve neglected to encode one or more strings before sending them to the OS. Without Sys::Binmode, Perl sends upgraded strings to the OS in UTF-8 encoding. In essence, it’s an implicit UTF-8 auto-encode, which is kind of nice, except that it depends on Perl’s internals, which are unpredictable. Sys::Binmode removes that implicit UTF-8 auto-encode, which of course will break things that need it.

The fix is to apply an explicit UTF-8 encode prior to the system call that throws the error. This is what we should do anyway; Sys::Binmode just enforces that better.

Example: The utf8 Pragma

The widely-used utf8 pragma particularly exemplifies this problem.

If you have code like this:

    use utf8;

    mkdir "épée";

… then adding this module will change your program’s behavior in ways you’ll probably dislike.

Consider the string épée. Without the utf8 pragma (but assuming that the code is actually written in UTF-8) this is 6 characters because the two és are 2 bytes each (so 2 + 1 + 2 + 1), and without the utf8 pragma each byte in a string constant becomes its own character, even if multiple bytes make up a single UTF-8 character. Since nothing probably upgrades that string on its way to mkdir(), the OS will receive the intended 6 bytes and create a directory with a UTF-8-encoded name.

With utf8, though, épée is 4 characters, not 6, because this string is now UTF-8-decoded. Those 4 characters all lie beneath 256, so the string is still bytes-compatible. Thus, if you print() that string you’ll get 4 bytes of Latin-1, which probably isn’t what you want.

mkdir(), though, probably still creates a directory with a 6-byte (UTF-8) name. This happens when Perl itself stores épée in upgraded (i.e., “unoptimized”) form. If that’s the case, that means Perl’s internal buffer of épée is still the 6 bytes of UTF-8, even though to the Perl application it’s a 4-character string. Perl’s mkdir() doesn’t care about characters, though; it just gives Perl’s internal buffer to the OS’s create-directory function. So by violating its own abstraction, Perl happens to achieve something that is sometimes useful.

There are still two problems, though:

  • 1. Inconsistency: print() sends 4 bytes to the OS while mkdir() (again, probably) outputs 6.

  • 2. Uncertainty: épée could be stored downgraded rather than upgraded, which would cause mkdir() to send 4 bytes instead.

print()’s outputting of 4 bytes here is actually the correct behavior because it doesn’t depend on whether Perl stores the string upgraded or downgraded. Sys::Binmode extends that correct behavior to mkdir() and other such Perl commands.

Of course, in the end, we want mkdir() to receive 6 bytes of UTF-8, not 4 bytes of Latin-1. To achieve that, just do as you normally do with print(): encode your string before you give it to the OS.

    use utf8;
    use Encode;

    mkdir encode("UTF-8", "épée");

This is what your code should look like, regardless of Sys::Binmode; the omitted encoding step was a bug that Perl’s own abstraction-violation bug might have obscured for you. Sys::Binmode fixes Perl’s bug, which makes you fix your own bug, too.

Non-POSIX Operating Systems (e.g., Windows)

In a POSIX operating system, an application’s communication with the OS happens entirely through byte strings. Thus, treating all OS-destined strings as byte strings is good and natural.

In Windows, though, things are weirder. For example, Windows exposes multiple APIs for creating a directory, and the one Perl uses (as of 5.32, anyway) only accepts code points 0-255. In this context Sys::Binmode doesn’t break anything, but it does reinforce one of Perl’s unfortunate limitations on Windows.

Sys::Binmode is a good idea anywhere that Perl sends byte strings to the OS. For now, as far as I know, that’s everywhere that Perl runs. If that’s not true, please file a bug.


The unpredictable-behavior problem that this module fixes in core Perl is also common in CPAN’s XS modules due to rampant use of the SvPV macro and variants. SvPV is basically Perl’s bytes pragma in C: it gives you the string’s internal bytes with no regard for what those bytes represent. This, of course, is problematic for the same reason why the bytes pragma is. XS authors generally should prefer SvPVbyte or SvPVutf8 in lieu of SvPV unless the C code in question handles Perl’s encoding abstraction.

Note in particular that, as of Perl 5.32, the default XS typemap converts scalars to C char * and const char * via an SvPV variant. This means that any module that uses that conversion logic also has this problem. So XS authors should also avoid the default typemap for such conversions. (Again, though, use of the default typemap in this context is regrettably commonplace.)

Before Perl 5.18 this problem also affected %ENV. 5.18 introduced an auto-downgrade when setting %ENV similar to what this module does.


If, for some reason, you want Perl’s unpredictable default behavior, you can disable this module for a given block via no Sys::Binmode, thus:

    use Sys::Binmode;

    system 'echo', $foo;        # predictable/sane/happy


        # You should probably explain here why you’re doing this.
        no Sys::Binmode;

        system 'echo', $foo;    # nasal demons


  • exec, system, and readpipe

  • do and require

  • File tests (e.g., -e) and the following: chdir, chmod, chown, chroot, ioctl, link, lstat, mkdir, open, opendir, readlink, rename, rmdir, stat, symlink, sysopen, truncate, unlink, utime

  • bind, connect, setsockopt, and send (last argument)

  • syscall


  • crypt already does as Sys::Binmode would make it do.

  • select (the 4-argument one) has the bug that Sys::Binmode fixes, but since it’s a performance-sensitive call where upgraded strings are unlikely, this library doesn’t wrap it.


autodie creates functions named, e.g., chmod in the namespace of the module that import()s it. Those functions lack the compiler “hint” that tells Sys::Binmode to do its work; thus, autodie “clobbers” Sys::Binmode. CORE::* functions will still have Sys::Binmode, but of course they won’t throw exceptions.


  • dbmopen and the System V IPC functions aren’t covered here. If you’d like them, ask.

  • There’s room for optimization, if that’s gainful.

  • Ideally this behavior should be in Perl’s core distribution.

  • Even more ideally, Perl should adopt this behavior as default. Maybe someday!


Thanks to Leon Timmermans (LEONT) and Paul Evans (PEVANS) for some debugging and design help.


Copyright 2021 Gasper Software Consulting. All rights reserved.

This library is licensed under the same license as Perl.