Re: RFA: reload infrastructure to fix PR target/21623

[Bernd Schmidt <bernds_cb1 at t-online dot de>]

Joern RENNECKE wrote:
> I suppose the best solution for that problem is to make the class 
> describe exclusively a
> temp register.  If no separate temp register is required, reload_class 
> can be returned
> (alternatively, we could say NO_REGS should be returned, and make
> push_secondary_reload test the icode as well as the class when deciding 
> if there is any
> reload at all.)

That would be fine.  So, IIUC the interface would be something like

/* Returns true if a secondary reload is needed for ...input args....
   After return from this function, *PICODE and *PCLASS describe how to
   do the secondary reload.
   If *PICODE is CODE_FOR_nothing, only a simple move instruction with a
   scratch register from *PCLASS is necessary, otherwise we use the
   instruction pattern selected by *PICODE.  CLASS can be NO_REGS in
   that case, or it can describe another scratch register if we need
   a tertiary reload.  In any case, the instruction pattern will be
   examined to determine the necessary class for the secondary reload
   and potential scratch registers needed.  */

bool (*secondary_foo_reload_needed_p)(...input args...,
	enum insn_code *picode, enum reg_class *pclass);

and if an insn code is used, we pick the class of the secondary reload 
and all scratch registers used by that pattern by looking at the operand 
constraints?

Or lose the bool return; I suppose it's redundant with
  icode != CODE_FOR_nothing || class != NO_REGS

Either way would be fine.  If this works for you, go ahead - or do you 
see a need to extend this, now or in the future?  I'd just like to see 
proper typing, no relying on the internal structure of enums, and no 
effect on other parts of the compiler such as gencodes.


Bernd