[PATCH 2/2] gcc/genrecog: Don't warn for missing mode on special predicates

[Richard Sandiford]

Andrew Burgess <andrew.burgess@embecosm.com> writes:
> In md.texi it says:
>
>   Predicates written with @code{define_special_predicate} do not get any
>   automatic mode checks, and are treated as having special mode handling
>   by @command{genrecog}.
>
> However, in genrecog, when validating a SET pattern, if either the
> source or destination is missing a mode then a warning is given, even if
> there's a predicate defined with define_special_predicate.
>
> This commit silences the warning for special predicates.
>
> gcc/ChangeLog:
>
> 	* genrecog.c (validate_pattern): Don't warn about missing mode for
> 	define_special_predicate predicates.
> Acked-by: Andrew Burgess <andrew.burgess@embecosm.com>
> ---
>  gcc/ChangeLog  |  5 +++++
>  gcc/genrecog.c | 22 +++++++++++++++++++---
>  2 files changed, 24 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/genrecog.c b/gcc/genrecog.c
> index a9f5a4a..7596552 100644
> --- a/gcc/genrecog.c
> +++ b/gcc/genrecog.c
> @@ -674,9 +674,25 @@ validate_pattern (rtx pattern, md_rtx_info *info, rtx set, int set_code)
>  		 && !CONST_WIDE_INT_P (src)
>  		 && GET_CODE (src) != CALL)
>  	  {
> -	    const char *which;
> -	    which = (dmode == VOIDmode ? "destination" : "source");
> -	    message_at (info->loc, "warning: %s missing a mode?", which);
> +	    const char *which_msg;
> +	    rtx which;
> +	    const char *pred_name;
> +	    const struct pred_data *pred;
> +
> +	    which_msg = (dmode == VOIDmode ? "destination" : "source");
> +	    which = (dmode == VOIDmode ? dest : src);
> +	    pred_name = XSTR (which, 1);
> +	    if (pred_name[0] != 0)
> +	      {
> +		pred = lookup_predicate (pred_name);
> +		if (!pred)
> +		  error_at (info->loc, "unknown predicate '%s'", pred_name);
> +	      }
> +	    else
> +	      pred = 0;
> +	    if (!pred || !pred->special)
> +	      message_at (info->loc, "warning: %s missing a mode?",
> +			  which_msg);

There's no guarantee at this point that "which" is a match_operand.
Also, I think the earlier:

        /* The operands of a SET must have the same mode unless one
	   is VOIDmode.  */
        else if (dmode != VOIDmode && smode != VOIDmode && dmode != smode)
	  error_at (info->loc, "mode mismatch in set: %smode vs %smode",
		    GET_MODE_NAME (dmode), GET_MODE_NAME (smode));

should be skipped for special predicates too.

How about generalising:

	/* The mode of an ADDRESS_OPERAND is the mode of the memory
	   reference, not the mode of the address.  */
	if (GET_CODE (src) == MATCH_OPERAND
	    && ! strcmp (XSTR (src, 1), "address_operand"))
	  ;

to:

	if (special_predicate_operand_p (src)
	    || special_predicate_operand_p (dest))
	  ;

with a new special_predicate_operand_p helper?  I don't think we should
duplicate the "unknown predicate" error here; the helper can just return
false for unknown predicates.

Thanks,
Richard