Character matrix upgrade performance is dismal
Extended Description
I'm not sure whether this counts as a Real Bug or not, but I'm going to
stick it in here anyway.
1. upgradeHeap.sac is the integer vector version of APL "upgrade".
Given an integer vector, V, generate a permutation vector PV, such that V[PV]
is in non-decreasing order. Furthermore, PV must be stable, i.e.,
"upgrade 2 2 2 2 2" must produce 0 1 2 3 4. This means that heapsort
(which is what this code implements) can be used, as it is stable, but
quicksort can not be used. This runs about 3x faster than the APL
interpreter's version of upgrade.
2. upgradeChar.sac: If we replace the integer vector argument of upgradeHeap.sac
with a character matrix, and insert upgradeLT and upgradeGT to perform
stupid, incorrect (due to signed char in sac) lessthen/greaterthan on
character vectors, the performance of the sac code is at least twice
as poor as that of the APL interpreter. I suspect the latter does not
do anything special for character upgrade.
3. If I do "sac2c -O3 -dcccall UpgradeChar.sac", then insert "-pg" into
syscall.sh and recompile, the generated code runs a LONG time, then
dies with segmentation fault.
4. If I do "sac2c -O3 -noinl -dcccall upgradeChar.sac", then insert "-pg"
into syscall.sh and recompile, the generated code runs long enough
to drive me to boredom and tears. I'll go back and try again with
a much smaller argument matrix and see if gprof gives any hints as
to where we're thumb-twiddling.
5. Haven't tried upgradeHeap.sac on an integer matrix. That may be
enlightening. Or not.
