Slow to compile keymaps on armel

Hi,

It takes 25 s to run "ckbcomp -model pc105 us >/dev/null" on a TI EVM board running jaunty/armel; this board is 500 MHz. On a 800 MHz Freescale Babbage board running jaunty/armel but perhaps building a different keymap, it's in the 5 seconds, and on my amd64 desktop it's sub-second.

ckbcomp could benefit from benchmarks and optimizations. We tried to add "use integer;" at the top and it didn't help.

Notes from conversation with Colin below; anybody is welcome to pickup the research.
12:22 < lool> cjwatson: Will it affect first boot?
12:22 < cjwatson> no, it won't
12:23 < cjwatson> it's interesting that it's so slow though
12:23 < lool> cjwatson: Ok, so it's normal that it takes so long on
              console-setup upgrades?
12:23 < lool> Right, I find that really slow
12:23 < cjwatson> ckbcomp only uses maybe 5MB of memory tops
12:23 < cjwatson> $ time ckbcomp -model pc105 us '' '' >/dev/null
12:23 < cjwatson> real 0m0.618s
12:24 < lool> time (ckbcomp -model pc105 us >/dev/null) => 0m26.826s
12:24 < cjwatson> lool: is it possible that 'use integer;' near the top of
                  ckbcomp would speed it up?
12:24 < lool> I don't see any faults in /proc/cpu/alignment
12:25 < cjwatson> I don't think it needs FP math
12:25 < lool> That's likely, trying
12:25 < lool> 0m25.583s
12:25 < lool> no
12:26 < cjwatson> might be worth trying it with -v 10 and seeing what's taking
                  time
12:27 < cjwatson> oh, hmm, -v 10 actually not that informative
12:27 -!- ember [n=pmf@ubuntu/member/ember] has quit [Read error: 104
          (Connection reset by peer)]
12:28 < cjwatson> verbose perl debugging will probably slow the whole thing
                  down to the point of making it hard to tell
12:28 < cjwatson> maybe just print debugging at appropriate points?
12:29 < lool> Ok
12:33 < cjwatson> I don't have any other obvious bright ideas :-(
12:36 < lool> cjwatson: time is spent on iterations of this loop foreach my
              $key (sort {$a <=> $b} (keys %symbols_table)) {
12:44 < cjwatson> lool: would be useful to know how many elements
                  %symbols_table has, and the same recursively for its elements
12:44 < cjwatson> the actual work done in the inner loop is trivial
12:45 < cjwatson> if the hash lookups are taking ages then we could arrange to
                  do that just once
12:46 < lool> cjwatson: flatten() is the one taking most time
12:47 < cjwatson> lool: flatten() isn't in the loop you mentioned ...
12:48 < cjwatson> oh, you're talking about a different loop than I'm looking at
12:48 * cjwatson starts again
12:50 < cjwatson> lool: the Devel::DProf module might help here
12:51 < cjwatson> 'perl -d:DProf /usr/bin/ckbcomp -model pc105 us >/dev/null'
                  will produce tmon.out and dprofpp can analyse that
...
12:48 < cjwatson> oh, you're talking about a different loop than I'm looking at
12:48 * cjwatson starts again
12:50 < cjwatson> lool: the Devel::DProf module might help here
12:51 < cjwatson> 'perl -d:DProf /usr/bin/ckbcomp -model pc105 us >/dev/null'
                  will produce tmon.out and dprofpp can analyse that
12:52 < persia> pitti, When you get a chance, could you look at bug #309396?
                There's some interest in getting it into -proposed before the
                Alpha Freeze.

Bye