[PATCH] Fix PR middle-end/28690, indexed load/store performance + reload bug

[Paolo Bonzini]

> You mean something like below (which passed bootstrap on powerpc64-linux
> and is currently regtesting) or something different?  If something
> different, I'd really need to know why x86 prefers set X X Y over
> set X Y X to be able to do anything about it.

Yes, mostly (see comments down in the patch).  Unless somebody knows why 
x86 prefers set X X Y, however, I can only point you to the testcase. 
:-(  I'm CCing Jan.

> I will note that I actually tried combining swap_commutative_operands_p
> and simplify_plus_minus_op_data_cmp, but I ended up hitting an ICE.
> It ended up optabs.c:emit_cmp_and_jump_insns() doesn't like it when we
> include the sort the REG's heuristic that simplify_plus_minus_op_data_cmp
> has to swap_commutative_operands_p.

Interesting.

> +      if (REG_P (op))
> +	return (REG_POINTER (op)) ? -1 : -3;
> +      else
> +	return -2;

Actually, I meant

   if (REG_P (op))
     return REG_POINTER (op) ? -1 : -2;
   else
     return -3;

And then swap_commutative_operands_with_target could die entirely.

Paolo