PR target/21299 (reload accepting invalid asm)

[Richard Sandiford]

Ian Lance Taylor <iant@google.com> writes:
> Richard Sandiford <richard@codesourcery.com> writes:
>> Ian Lance Taylor <iant@google.com> writes:
>> > Richard Sandiford <richard@codesourcery.com> writes:
>> >> I think we need to clarify what it means to have several register
>> >> classes in the same alternative.  Which interpretation is right?
>> >> constrain_operands(), or find_reloads() and preprocess_constraints()?
>> >
>> > Thanks for helping to clarify the issue.  As I noted in my earlier
>> > reply, I believe that both interpretations are right, because MD file
>> > writers are required to follow rules such that both are right.
>> 
>> Wasn't sure whether to reply to that message or this one, but...
>> 
>> I think that message missed the point.  At least it misses the point of
>> what I was trying to say in my reply.  What happens if you have constraints
>> for two register classes C1 and C2, and a union C1+C2?  Must a register
>> belong to either C1 or C2, or may it straddle both?  constrain_operands()
>> does not allow it to straddle both classes, but find_reloads() and
>> preprocess_constraints() do.
>
> I believe that an MD file for which this question arises is buggy.
>
> It can only arise if you use a value which requires multiple
> registers, and you put multiple register class constraints in a single
> alternative, and it is possible, due to register numbering and the
> definition of HARD_REGNO_MODE_OK, for the multiple registers to cross
> from one register class to another.

Right.  I was thinking of cases where we define classes for a subset
of a logical register bank.  Something like Q_REGS for x86.  If you
mix those classes together, HARD_REGNO_MODE_OK may not necessarily
force the register to be in one class or the other.

> I see no reason that we ever need to support this case.  Therefore,
> the MD file is buggy.  You should fix HARD_REGNO_MODE_OK, or you
> should separate the constraints into two separate alternatives, or you
> should use the union class.

I'm happy with that.  But I was trying to pin down what the rules were
so that we can enforce those same rules for user asms.  So if that's
the rule, then I think we should be checking something like:

   "Two register class constraints may not be used in the same
   alternative for a value of mode M if the backend allows values
   of mode M to be stored in a register that straddles both classes."

Hopefully that's a more general fix than the ones suggested so far.

Richard