Re^5: Performance penalty of using qr//

You've created a benchmark where the cost of compiling a single regex containing 45000 alternations is overwhelmingly more expensive than running that compiled regex 11 times.

In the qr case, the pattern is compiled once, *before* the benchmark is run. The benchmark cost is 11 matches, plus 11 clonings of the compiled regex's internal struture.

In the str case, the benchmark includes compiling the pattern before matching the first word. For the subsequent 10 matches it attempt to recompile /$str/, but each time it uses an optimisation where it sees if the string has changed since last time and if so skips recompiling it.

So sometimes /$str/ in a loop for unchanging $str can be faster than /$qr/, but that benchmark won't show it.

Or more formally, if C is the time to compile a pattern, M is the time to run (match) against the compiled pattern, and D is the time to duplicate the regex structure, then

$qr = /.../; /$qr/ for 1.N
[download]

takes C + N(M+D), but your benchmark was measuring N(M+D); while

$str = "..."; /$str/ for 1.N
[download]

takes ~~N(C+M)~~ C + NM, which was what your benchmark was measuring. (I just corrected the above - I forgot to include the 'unchanged pattern' optimisation)

Dave.

Comment on Re^5: Performance penalty of using qr// Select or Download Code

Replies are listed 'Best First'.

Re^6: Performance penalty of using qr//
by vr (Curate) on Dec 22, 2017 at 19:26 UTC

C + NM, which was what your benchmark was measuring

Having read your answers and this -- 269035 -- thread, maybe it's better to say that in case of "use_str" benchmark was measuring

C + 2268 * ( 5 - time_for_C ) * 11 * ( M + E )

where E is "light-weight" compilation, consisting only of string equality (eq) check? Next example shows that "proper" compilation happens just once, if I'm reading output correctly:

use strict;
use warnings;
use feature 'say';
use re 'debug';

my $str = 'foobar';

sub foo { map /$str/, 1 .. 3 }

foo for 1 .. 2;
[download]

Compiling REx "foobar"
Final program:
   1: EXACT <foobar> (4)
   4: END (0)
anchored "foobar" at 0 (checking anchored isall) minlen 6
Compiling REx "foobar"
Compiling REx "foobar"
Compiling REx "foobar"
Compiling REx "foobar"
Compiling REx "foobar"
[download]

("use_re" benchmarked concatenation of 45000 words, BTW. Similar, to above, script shows proper compilation happens only once, too.) And "time_for_C" is negligible, we can compile regexp placing dummy /$str/ at the top of the script, and benchmark result won't change. Moreover:

use strict;
use warnings;
use Benchmark qw( cmpthese timethese );

open my $words, "<", "linuxwords.txt" or die "$!";
my @words = <$words>;
chomp @words;

my @search = @words[0..10];

$" = "|";

my $re = qr/^(?:@words)$/;
my $str = "^(?:@words)\$";
my $str1 = $str;

my $r = timethese
(
    -5,
    {
        use_qr  => sub { map /$re/, @search },
        use_str => sub { 
            substr $str1, rand 400_000, 1, '#';     # 'equalize' condi
+tions with 'use_str1'
            map /$str/, @search 
        },
        use_str1 => sub { 
            substr $str1, rand 400_000, 1, '#';     # force proper re-
+compilation below
            map /$str1/, @search 
        },
    }
);
cmpthese $r;
[download]

            Rate use_str1  use_str   use_qr
use_str1  3.65/s       --     -99%    -100%
use_str    436/s   11853%       --     -99%
use_qr   31592/s  865014%    7138%       --
[download]

To summarize, if I may: if pattern is "long enough", then D (time to duplicate the regex structure) is much less than E (time to compare patterns with simple "eq"). For short patterns (as in OP), D is more expensive than E (even multiple E's), so attempt to optimize through use of "qr" failed.

Edit. People "lucky" enough to hit 0..3 with "rand 400_000" should change offset in "substr" to "4 + rand 400_000". :(

Edit 2. I mean, you say: if "N(M+D)" is so much faster than "C+NM", then it's C that is so slow, but I think, no, C is negligible, it's rather "N(M+D)" is faster than "C+N(M+E)" because D is faster than E. Sorry I haven't communicated this thought without edits.

[reply]
[d/l]
[select]


No such thing as a small change
	PerlMonks