Re: [PING] Target hook for rewriting inline asm constraints

[Richard Sandiford <rsandifo@nildram.co.uk>]

"Andreas Krebbel" <Andreas.Krebbel@de.ibm.com> writes:
>> Sorry if this has already been suggested, but why not just replace
>> hard-coded uses of 'm' in constraint-handling code with uses of some
>> target macro?  You could then treat your current 'm' as an ordinary
>> define_memory_constraint.
>
> I'm not sure this would work.  I think reload relies on the
> relationship between the 'm' constraint and GO_IF_LEGITIMATE_ADDRESS.
> There is an early loop invoking find_reloads_address.  This reloads
> all addresses which aren't supported by the target at all i.e. all
> addresses for which GO_IF_LEGITIMATE_ADDRESS is false (or does not
> jump to the given label).  If reload later on sees an 'm' constraint
> it considers the address already been fixed by that early loop.  So I
> would expect that in order to make such a hook work I would have to
> mess around with the reload logic which is not preferable I think.

I don't understand.  What I'm saying is that we should look through
gcc/*.c for cases where the C construct "'m'" refers to "the constraint
associated with legitimate addresses", and replace those cases with some
target macro that is 'm' by default.  E.g.:

	      case 'm':
		if (force_reload)
		  break;
		if (MEM_P (operand)
		    || (REG_P (operand)
			&& REGNO (operand) >= FIRST_PSEUDO_REGISTER
			&& reg_renumber[REGNO (operand)] < 0))
		  win = 1;
		if (CONST_POOL_OK_P (operand))
		  badop = 0;
		constmemok = 1;
		break;

would become:

	      case TARGET_MEM_CONSTRAINT:
		if (force_reload)
		  break;
		if (MEM_P (operand)
		    || (REG_P (operand)
			&& REGNO (operand) >= FIRST_PSEUDO_REGISTER
			&& reg_renumber[REGNO (operand)] < 0))
		  win = 1;
		if (CONST_POOL_OK_P (operand))
		  badop = 0;
		constmemok = 1;
		break;

and other places would change similarly.  If we do that consistently,
there would no longer be a connection between 'm' and
GO_IF_LEGITIMATE_ADDRESS.

Replacing hard-coded constants seems like good practice for its own
sake and would avoid the need for the front end to rewrite asms.

Richard