When switch detection looks at a disjunction, it loops over all the nonlocal
variables that are already-bound on entry to the disjunction, to see whether
the disjunction is a switch on that variable. We used to stop this search
as soon as we found what looked like an good candidate, but in some
rare cases, (a) committing to this candidate leads to a determinism that is
not as tight as possible, and (b), if we chose to transform the disjunction
into another candidate, the resulting determinism *would* be as tight
as possible.
The fix is to always gather all the candidate switches, rank them in order of
how likely they are to yield a tight determinism, and choose the candidate
with the highest rank. Doing this requires computing whether *each candidate
switch* is a can_fail switch, whereas the old approach required computing
only whether the *chosen switch* is a can_fail switch.
While avoiding this small amount of extra computation was probably a good
tradeoff in the mid 1990s, it isn't anymore. The new code is also simpler
and clearer, for two distinct reasons. One is that committing to a good-enough
candidate as soon as we find it required mixing the code that *finds* candidate
switches with the code that *acts* on the selected candidate. The second is
that the old code could choose the best possible switch (none of whose arms is
reachable due the switched-on variable's initial inst, and which can therefore
be translated to nothing more than just `fail') only by accident, never by
design, since it tested which arms are reachable only *after* it has already
selected the variable to switch on.
This diff fixes Mantis bug #414.
compiler/switch_detection.m:
As above.
tests/valid/bug414.m:
The test case for the bug.
tests/valid/Mmakefile:
Enable the test case.
