Re: register class constraints question

[Joern Rennecke <joernr at arc dot com>]

I'm not sure if this changed with IRA, but at least with the old reload
code, if the instruction doesn't match, reload will look for
the alternative with the cheapest total cost of alternative and
reloads.  Unfortunately, the fact that some reloads might be impossible
does not figure in the decision.  Where the costs are equal, the order
is the tie-breaker; the first alternative is choosen then.
You must make sure that the HI regs alternative gets precedence over the
HL regs one when the operand getting the HI or HL regs is reloaded.
Uou can do this by tweaking alternative costs, alternative order,
REG_MOVE_COST (for the reload costs), or any combination if these.

Or better, fix the alternative choosing algorithm to avoid alternatives
which are already known to be impossible to fulfill when looking at the
registers that are live during the insn and the constraints of a single
operand in isolation.
I.e. if we require n hard registers of a particular class, check that
there are n hard registers free in that class.
Probably discounting the possibility of having different overlapping
classes between primary output and secondary / tertiary reloads which
can fit individually but not together.  (I.e. leave it to the port
writer to avoid that).

To avoid slowing the compiler down too much, you might want to restrict
this to CLASS_LIKELY_SPILLED and/or delay the test till the alternative
is choosen as the best one, and if the test fails, re-do the alternative
choosing with a bit mask of discarded alternatives.
(In which case a combined register/memory alternative that uses impossible
 address reloads on the first round will be discarded so the register part
 is also disabled - but I suppose that's a limitation we can live with.)
It may be noted that there is nothing really stopping us from considering
all operands of an alternative - for each operand, reload decides on
an actual register class requirement, even if the alternative has multiple
subparts - but I think it would make the code much more complex, and I doubt
it would buy much in making ports easier to write or the generated code
better.