How to unroll a loop ('D' if you know)


James B. VanBokkelen (JBVB@AI.AI.MIT.EDU)
Wed, 7 Oct 87 01:41:51 EDT


I got asked several times, so I'll reply to everyone. The easy-to-understand
way is:
        ...
        remainder = count % UNROLLING_FACTOR;
        count = (count/UNROLLING_FACTOR) + 1; /* no fenceposts for me */
        switch (remainder) {
                case 0:
                        count--; /* optimize */
                        goto top;
                case 1:
                        goto rem_1;
                ... /* up to UNROLLING_FACTOR - 1 */
        }
top: do_it;
        ... /* UNROLLING_FACTOR times in all */
rem_2: do_it;
rem_1: do_it;
        if (count--)
                goto top;
        ...

The way it is usually done (because speed freaks are rarely willing to put
up with any real or supposed inefficiency in their compilers, and are coding
the critical parts in assembler) is by figuring out how large each occurrence
of "do_it" is in bytes, and multiplying that by the remainder, and either
adding that to the program counter (if you can), or using it as an index
in a JUMP instruction of some sort. If you have a CASE instruction, make sure
it is really more efficient than coding it out. HLL users can take a look
at how their compilers do 'case' or 'computed goto' - maybe it isn't losing
much time.

The smaller the loop test's elapsed time is relative to the time necessary
to "do_it", the less unrolling buys you.

If I coded it wrong, feel free to punish me publicly, but it probably doesn't
belong on tcp-ip. If you are horrified by gotos, please don't tell me...

jbvb



This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:39:34 GMT