The other big difference between my algorithm (and how I tested it) and the once you posted are that mine are not depending on the length of the buffer and are therefor more generic.

I should point out that the hard coded 24s in my code versions are there because I know that is the only length I will be dealing with. Discovering the length from the passed scalar, (and checking it is a multiple of 3), would have little or no effect upon performance.

The point about the mortalisation comes into stark relief when you run your code in a tight loop. Where this version consumes a steady 14.7 MB for as long as you care to run it:

#! perl -slw
use strict;
use Inline C => Config => BUILD_NOISY => 1;
use Inline C => <<'END_C',  NAME => '_24to32', CLEAN_AFTER_BUILD => 0;

void uic( SV *src ) {
    int i = 0;
    STRLEN l;
    unsigned char *s = (unsigned char *)SvPV( src, l );

    inline_stack_vars;
    inline_stack_reset;

    while (i < l) {
        int n = ( s[i] >> 2 ) & 0x3f;
        inline_stack_push( sv_2mortal( newSViv( n ) ) );
        n = ( s[ i++ ] & 0x03 ) << 4;
        n |= ( s[i] >> 4 ) & 0x0f;
        inline_stack_push( sv_2mortal( newSViv( n ) ) );
        n = ( s[ i++ ] & 0x0f ) << 2;
        n |= ( s[i] >> 6 ) & 0x03;
        inline_stack_push( sv_2mortal( newSViv( n ) ) );
        n = s[ i++ ] & 0x3f;
        inline_stack_push( sv_2mortal( newSViv( n ) ) );
     }
     inline_stack_done;
}

END_C

use Math::Random::MT qw[ rand ];

while( 1 ) {
    my $packed = pack 'N6', map rand( 2**32 ), 1 .. 6;
    my @decoded = uic( $packed );
}
[download]

This version:

#! perl -slw
use strict;
use Inline C => Config => BUILD_NOISY => 1;
use Inline C => <<'END_C',  NAME => '_24to32', CLEAN_AFTER_BUILD => 0;

void uic( SV *src ) {
    int i = 0;
    STRLEN l;
    unsigned char *s = (unsigned char *)SvPV( src, l );

    inline_stack_vars;
    inline_stack_reset;

    while (i < l) {
        int n = ( s[i] >> 2 ) & 0x3f;
        inline_stack_push( newSViv( n ) );
        n = ( s[ i++ ] & 0x03 ) << 4;
        n |= ( s[i] >> 4 ) & 0x0f;
        inline_stack_push( newSViv( n ) );
        n = ( s[ i++ ] & 0x0f ) << 2;
        n |= ( s[i] >> 6 ) & 0x03;
        inline_stack_push( newSViv( n ) );
        n = s[ i++ ] & 0x3f;
        inline_stack_push( newSViv( n ) );
     }
     inline_stack_done;
}

END_C

use Math::Random::MT qw[ rand ];

while( 1 ) {
    my $packed = pack 'N6', map rand( 2**32 ), 1 .. 6;
    my @decoded = uic( $packed );
}
[download]

leaks memory at the rate of 3GB per minute, which means it has my machine swapping after less than 2 minutes and dies "Out of memory|", having consume all 16GB of swap space, some time later. Whatever the performance hit, it is unavoidable. It ought (and I believe could), be cheaper though.