I guess our approaches are different. I am a student and currently my ultimate goal here is to learn. I think this is achieved better going from simpler things to more complex ones. According to your approach, if I were given a black box that solves my problem in a split of a second, I should have used it and be over with it. This will solve the problem obviously, but will teach me nothing. Anyway, I'm sorry to hear that you think I'm "bordering on irrational obsession". This was actually quite offending.

UPDATE: re. your update, I did. Quite surprisingly, it now takes significantly longer to run. I guess I have some error, but I can't see it.

Inline version:

use strict;
use warnings;
use List::Util qw(max min);
use Time::HiRes qw(time);

# this builds a structure that is usually retrieved from disk.
# in this example we will use this structure again and again,
# but in the real program we obviously retrieve a fresh structure
# at each iteration

my $simulation_h = {};
for ( 1 .. 70000 ) {
    my $random_start  = int( rand(5235641) );
    my $random_length = int( rand(40000) );
    push @{ $simulation_h->{$random_start} }, $random_length;
}

my $zone_o = {
    _chromosome_length => 5235641,
    _legal_range       => [ { FROM => 100000, TO => 200000 } ]
};

my $start_time = time;
scenario();
print "total loop time: " . ( time - $start_time ) . " seconds\n";
my $temp_gene_to_legal_range;
my $gene_to;

sub scenario {
    for ( my $i = 0 ; $i < 50 ; $i++ ) {
        print "i=$i time=" . ( time - $start_time ) . " seconds\n";

        # originally there was a retreive of $simulation_h from disk h
+ere

        # iterate genes

        foreach my $gene_from ( keys %{$simulation_h} ) {
            foreach my $gene_length ( @{ $simulation_h->{$gene_from} }
+ ) {

### inlining gene_to_legal_range
                $gene_to =
                  ( ( $gene_from + $gene_length - 1 )
                    % ( $zone_o->{_chromosome_length} ) ) + 1;

                if ( $gene_to < $gene_from ) {

                    # split
                    # low range first
                    $temp_gene_to_legal_range = [
                        { FROM => 0, TO => $gene_to },
                        {
                            FROM => $gene_from,
                            TO   => $zone_o->{_chromosome_length}
                        }
                    ];
                }
                else {

                    # single
                    $temp_gene_to_legal_range =
                      [ { FROM => $gene_from, TO => $gene_to } ];
                }

            }
        }
    }
}
[download]

Previous version (with subroutine call):

use strict;
use warnings;
use List::Util qw(max min);
use Time::HiRes qw(time);

# this builds a structure that is usually retrieved from disk.
# in this example we will use this structure again and again,
# but in the real program we obviously retrieve a fresh structure
# at each iteration

my $simulation_h = {};
for ( 1 .. 70000 ) {
    my $random_start  = int( rand(5235641) );
    my $random_length = int( rand(40000) );
    push @{ $simulation_h->{$random_start} }, $random_length;
}

my $zone_o = {
    _chromosome_length => 5235641,
    _legal_range       => [ { FROM => 100000, TO => 200000 } ]
};

my $start_time = time;
scenario();
print "total loop time: " . ( time - $start_time ) . " seconds\n";
my $temp_gene_to_legal_range;
my $gene_to;

sub scenario {
    for ( my $i = 0 ; $i < 50 ; $i++ ) {
        print "i=$i time=" . ( time - $start_time ) . " seconds\n";

        # originally there was a retreive of $simulation_h from disk h
+ere

        # iterate genes

        foreach my $gene_from ( keys %{$simulation_h} ) {
            foreach my $gene_length ( @{ $simulation_h->{$gene_from} }
+ ) {

#### inlining gene_to_legal_range
#                $gene_to =
#                  ( ( $gene_from + $gene_length - 1 )
#                    % ( $zone_o->{_chromosome_length} ) ) + 1;
#
#                if ( $gene_to < $gene_from ) {
#
#                    # split
#                    # low range first
#                    $temp_gene_to_legal_range = [
#                        { FROM => 0, TO => $gene_to },
#                        {
#                            FROM => $gene_from,
#                            TO   => $zone_o->{_chromosome_length}
#                        }
#                    ];
#                }
#                else {
#
#                    # single
#                    $temp_gene_to_legal_range =
#                      [ { FROM => $gene_from, TO => $gene_to } ];
#                }
#
#                next; ####
                
                $temp_gene_to_legal_range =
                  gene_to_legal_range( $gene_from, $gene_length,
                    $zone_o->{_chromosome_length} );

            }
        }
    }
}

sub gene_to_legal_range($$$) {
    return;
    my ( $gene_from, $gene_length, $legal_length ) = @_;

    my $ret;
    my $gene_to = ( ( $gene_from + $gene_length - 1 ) % ($legal_length
+) ) + 1;

    if ( $gene_to < $gene_from ) {

        # split
        # low range first
        $ret = [
            { FROM => 0,          TO => $gene_to },
            { FROM => $gene_from, TO => $legal_length }
        ];
    }
    else {

        # single
        $ret = [ { FROM => $gene_from, TO => $gene_to } ];
    }

    return $ret;
}
[download]

You are comparing "running the logic of the function in-line" vs. "calling a function that does not execute the logic". That's not the comparison that matters. Comment out the "return" at the top of the function in the latter version, and the timing shows the function call to be more expensive than the inline code.

Another minor detail: in your inline version, you do two hash lookups for the value of  $zone_o->{_chromosome_length} whereas the function call version does only one lookup. If you change the inline version to assign that value to a "my" variable, and use that variable twice (just like it was used twice in the function call), you'll see a reduction of a few sec -- i.e. the improvement over the function call will be more evident.

Now, get back to the part that really matters.

Thanks. Indeed I forgot to omit the return. I will now go on to follow the reco's regarding the second part and look forward to coming back here updating you.

I must comment that I am still surprised that such a simple function/piece of inline code takes so much time to execute. I never looked at this level of benchmarking, so I guess I had mistaken expectations.

Sorry, that wasn't meant as a personal insult -- it was an exaggerated comment about the thought pattern (which I recognize, because I fall prey to it myself, even after 25 years of professional programming in numerous languages).

It was just a bit jarring for me to see the contradiction laid out so plainly: you are worried (appropriately) about how long it'll take for your script to run on real data and you want to learn how to fix it, but you'll put off the work of addressing the most costly part of the algorithm, because you want to do the simpler part first?

I guess that's okay if you want to remain a student indefinitely. But in terms of holding a job, that approach is not just "different"... Anyway, good luck.