When is it valid to use RTX_FRAME_RELATED_P ?

[Richard Earnshaw]

> I'm trying to find out why gcc-3.2 cvs builds (with checking enabled) fail on arm-linux.
> 
> The first problem that I had to resolve was with 'CONSTANT_POOL_ADDRESS_P'
> where Krister came up with a solution in
> <http://gcc.gnu.org/ml/gcc-patches/2002-06/msg00365.html>
> 
> (although I would rather rewrite it as :
> 
> #define LEGITIMATE_PIC_OPERAND_P(X)					\
> 	(   ! symbol_mentioned_p (X)					\
> 	 && ! label_mentioned_p (X)					\
> 	 && (! ((GET_CODE(X) == SYMBOL_REF) && CONSTANT_POOL_ADDRESS_P (X))				\
> 	     || (   ! symbol_mentioned_p (get_pool_constant (X))  	\
> 	         && ! label_mentioned_p (get_pool_constant (X)))))
> 
> )
> 
> )
> 

This statement needs rearranging to remove all those inversions, it can be 
written much more cleanly as:

	(!(symbol_mentioned_p (X)
	   || label_mentioned_p (X)
	   || (GET_CODE (X) == SYMBOL_REF
	       && CONSTANT_POOL_ADDRESS_P (X)
	       && (symbol_mentioned_p (get_pool_constant (X))
		   || label_mentioned_p (get_pool_constant (X))))))


> static rtx
> emit_sfm (base_reg, count)
>       int base_reg;
>       int count;
> {
>    rtx par;
>    rtx dwarf;
>    rtx tmp, reg;
>    int i;
> 
>    par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (count));
>    dwarf = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (count));
>    RTX_FRAME_RELATED_P (dwarf) = 1; /* <==== here it fails*/
> 
>    reg = gen_rtx_REG (XFmode, base_reg++);
> [...]
> 
> 
> So what should I do now ? RTX_FRAME_RELATED_P doesn't expect 'parallel', but
> in arm.c we explicitely create a parallel and set the RTX_FRAME_RELATED_P for it ??
> 

Simply delete the line where we set RTX_FRAME_RELATED_P for 'dwarf'.  The 
argument goes: this expression is only used inside a 
REG_FRAME_RELATED_EXPR note, so we don't need to annotate it further.

R.