RFC: subreg fixing (extract_bit_field or simplify_subreg)

[Aldy Hernandez]

Yay.  More subreg stupidity.

On the following testcase, we call extract_bit_field trying to extract
a DF out of a memory block:

	struct __attribute__ ((packed)) W { double l; } sW;
	double dd;
	testW () { dd = sW.l; }

We first call operand_subword with wordnum=1, and somewhere down the call
chain, simplify_gen_subreg calls simplify_subreg, which returns 0.
This causes simplify_gen_subreg to force creation of the subreg
with gen_rtx_SUBREG.

...this creates invalid rtl: (subreg:SI (reg:DF 118) 4).

What is advisable?

Would it be ok to modify simplify_gen_subreg to return NULL_RTX for
class incompatabilities, or is the problem further up?  (Obviously,
I'd have to change operand_subword to operand_subword_force).

On a similar note, invalid rtl keeps biting my butt, and has for about
5 years.   Would be ok to add checks in gen_rtx_SUBREG to abort on 
invalid RTL (hard coding the common cases)...or am I back to putting
yellow stickies on my monitor with the stupid cases, and matching RTL
dumps with stickies for the rest of my life?

Aldy