5.42: Does m// toss a string around?

Anonymous Monk has asked for the wisdom of the Perl Monks concerning the following question:

$s = '0123456789';
printf "%x\n", unpack 'Q', pack 'P', $s
    while $s =~ /\G./g;

# 55fd180fcc60
# 55fd18107f30
# 55fd18107ed0
# 55fd180e6700
# 55fd180fcc60
# 55fd18107f30
# 55fd18107ed0
# 55fd180e6700
# 55fd180fcc60
# 55fd18107f30
[download]

Looks like patterns don't matter, both between slashes (the above is for illusion of being useful, supposedly performed daily by thousands) and alternation of addresses, it just picks what's available.

use strict;
use warnings;
use feature 'say';
use Config;

say "$^V / $Config{ archname }";

for my $len ( 1e4, 1e5, 1e6, 1e7 ) {
    my $s = 'a' x $len;

    my %h;
    my @before = times;
    $s =~ /\G./gs and ++ $h{ pack 'P', $s } 
        for 1 .. $len / 100;                    # do just 1% !!!
    my @after = times;

    printf "Length: %8d, addresses: %6d, user: %6g, system: %6g\n",
        $len, scalar( keys %h ), 
        $after[ 0 ] - $before[ 0 ],
        $after[ 1 ] - $before[ 1 ]
}

__END__

v5.40.3 / x86_64-linux-thread-multi
Length:    10000, addresses:      1, user:      0, system:      0
Length:   100000, addresses:      1, user:      0, system:      0
Length:  1000000, addresses:      1, user:   0.01, system:      0
Length: 10000000, addresses:      1, user:   0.02, system:      0

v5.42.0 / x86_64-linux-thread-multi
Length:    10000, addresses:      2, user:      0, system:      0
Length:   100000, addresses:      3, user:   0.01, system:      0
Length:  1000000, addresses:      3, user:   0.39, system:      0
Length: 10000000, addresses:      3, user: 139.87, system:      0

v5.43.8 / x86_64-linux
Length:    10000, addresses:      2, user:      0, system:      0
Length:   100000, addresses:      3, user:   0.01, system:      0
Length:  1000000, addresses:      3, user:   0.38, system:      0
Length: 10000000, addresses:      3, user: 138.75, system:   0.01

v5.42.0 / MSWin32-x64-multi-thread
Length:    10000, addresses:      2, user:      0, system:      0
Length:   100000, addresses:      2, user:      0, system:      0
Length:  1000000, addresses:      2, user:  0.844, system:  1.063
Length: 10000000, addresses:     70, user: 154.453, system: 253.812
[download]

It's much worse on Windows, with absurdly grotesque "system" time (I understand it's approximation for that OS); I can only assume buffers, when long enough (?), are continuously requested to be de-allocated and allocated, from OS, again and again, which is not the case with Perl on Linux.

If there's a bug it's a shock it goes unnoticed, I'm just a no-one. In fact I'm grateful that yesterday it wasn't possible to post, I now see the above is separate from another similar issue from 5.20 and on; better not to mix them together

Comment on 5.42: Does m// toss a string around? Select or Download Code

Replies are listed 'Best First'.
Re: 5.42: Does m// toss a string around? by ikegami (Patriarch) on Jan 21, 2026 at 20:54 UTC
It's `pack 'P'` that's making a copy of the string buffer. While matching creates a copy of the string matched, matching uses the COW mechanism to avoid making a copy of the string buffer when possible.^[1] The COW mechanism is being used here, so no copy is being made by matching. However, `pack 'P'` forces the scalar to have a string buffer and for it to be modifiable.^[2] Since you are starting with a buffer that's shared with the literal constant and the scalar associated with `$&` and other regex vars, this means `pack 'P'` forces a copy to be made. `use Config qw( %Config ); use Devel::Peek qw( Dump ); my $ptr_size = $Config{ ptrsize }; my $ptr_format = $ptr_size == 4 ? "L" : $ptr_size == 8 ? "Q" : die( "Unsupported pointer size $ptr_size\n" ); # https://perldoc.perl.org/perlapi#SvPV_force sub SvPV_force { unpack $ptr_format, pack "P", $_[0] } my $s = "abc"; Dump $s; printf "%x\n", SvPV_force( $s ); Dump $s;` [download] `SV = PV(0x57565c72fee0) at 0x57565c75e0d0 REFCNT = 1 FLAGS = (POK,IsCOW,pPOK) PV = 0x57565c761a90 "abc"\0 CUR = 3 LEN = 16 COW_REFCNT = 1 00.00.57.56.5c.78.5b.c0 SV = PV(0x57565c72fee0) at 0x57565c75e0d0 REFCNT = 1 FLAGS = (POK,pPOK) PV = 0x57565c785bc0 "abc"\0 CUR = 3 LEN = 16` [download] Notice how `Dump` showed a shared buffer (`IsCOW`) before `pack 'P'`, but one that isn't shared afterwards. You can use the following to get the address of the string buffer without forcing it to exist and to be modifiable: `use B qw( svref_2object ); # https://perldoc.perl.org/perlapi#SvPVX sub SvPVX { svref_2object( \$_[0] )->PV } my $s = "0123456789"; printf "%x\n", SvPVX( $s ) while $s =~ /\G./g;` [download] Matching will make a copy of the string buffer if it can't use COW. Example. From perl5420delta, `pack("p", ...)` and `pack("P", ...)` now SvPV_force() the supplied SV unless it is read only. This will remove CoW from the SV and prevents code that writes through the generated pointer from modifying the value of other SVs that happen the share the same CoWed string buffer. Note: this does not make `pack("p",... )` safe, if the SV is magical then any writes to the buffer will likely be discarded on the next read. [GH #22380] Update: Renamed `SvPV` to `SvPVX` for accuracy.	[reply] [d/l] [select]
Re^2: 5.42: Does m// toss a string around? by Anonymous Monk on Jan 22, 2026 at 10:50 UTC
Ah, thanks. If I understand correctly, it was 5.42 which fixed the long-standing issue with "`pack 'P',`". (I think the PV method constructs new string rather than returns numeric value. I'll use "`pack 'P',`" for the next example, for brevity of output over `Devel::Peek::Dump`; there's no 5.42 below) Wild goose chase, then, being sidetracked with imagined issue with 5.42, but initially I investigated `s///`, not `m//`. I now suspect similar reasons, related to COW, can you explain please? The 5.18 re-uses string buffer when new length is either the same or less than original. The 5.20 and then 5.26 loose both abilities. `print "$^V\n"; my $x = '0123456789'; { my $s = "$x$x"; $s =~ s/(.)$/a/; printf "%x\n", unpack "Q", pack "P", $s; $s =~ s/(.)$/b/; printf "%x\n", unpack "Q", pack "P", $s; } { my $s = "$x$x"; $s =~ s/.$//; printf "%x\n", unpack "Q", pack "P", $s; $s =~ s/.$//; printf "%x\n", unpack "Q", pack "P", $s; } __END__ v5.18.4 29cdda8 29cdda8 29ce1e8 29ce1e8 v5.20.3 2b774f8 2b774f8 2b76b78 2b76f38 v5.26.3 2c9e118 2c9d7b8 2c9da88 2b4ae88` [download]	[reply] [d/l] [select]
Re^3: 5.42: Does m// toss a string around? by ikegami (Patriarch) on Jan 22, 2026 at 13:50 UTC
I think the PV method constructs new string rather than returns numeric value No. `SvPVX` creates an object that provide information about a scalar (`$_[0]` aka `$s`), then uses that object's `PV` method to obtain the address of the string buffer of `$s`. I now suspect similar reasons, related to COW, can you explain please? COW was introduced in 5.20. Before 5.20, hacks which malfunctioned and/or a more expensive alternative had to be used.	[reply] [d/l] [select]
Re^4: 5.42: Does m// toss a string around? by Anonymous Monk on Jan 29, 2026 at 23:30 UTC
Re^5: 5.42: Does m// toss a string around? by ikegami (Patriarch) on Jan 30, 2026 at 18:59 UTC
Some notes below your chosen depth have not been shown here
Re: 5.42: Does m// toss a string around? by Anonymous Monk on Jan 21, 2026 at 22:41 UTC
M3 Macbook Air `v5.38.2 / darwin-2level Length: 10000, addresses: 1, user: 0, system: 0 Length: 100000, addresses: 1, user: 0, system: 0 Length: 1000000, addresses: 1, user: 0, system: 0 Length: 10000000, addresses: 1, user: 0.02, system: 0` [download] Run 4 times under 5.42.0: v5.42.0 / darwin-2level Length: 10000, addresses: 2, user: 0, system: 0 Length: 100000, addresses: 4, user: 0, system: 0 Length: 1000000, addresses: 11, user: 0.2, system: 0 Length: 10000000, addresses: 2, user: 23.86, system: 0.03 v5.42.0 / darwin-2level Length: 10000, addresses: 2, user: 0, system: 0 Length: 100000, addresses: 2, user: 0.01, system: 0 Length: 1000000, addresses: 7, user: 0.18, system: 0 Length: 10000000, addresses: 2, user: 23.15, system: 0.03 v5.42.0 / darwin-2level Length: 10000, addresses: 2, user: 0, system: 0 Length: 100000, addresses: 2, user: 0.01, system: 0 Length: 1000000, addresses: 8, user: 0.21, system: 0 Length: 10000000, addresses: 2, user: 23.23, system: 0.02 v5.42.0 / darwin-2level Length: 10000, addresses: 2, user: 0, system: 0 Length: 100000, addresses: 2, user: 0.01, system: 0 Length: 1000000, addresses: 9, user: 0.21, system: 0 Length: 10000000, addresses: 2, user: 23.26, system: 0.03 [download]	[reply] [d/l] [select]