Fix MIPS ICE on HImode subregs of frame pointer

[Richard Sandiford]

"Joseph S. Myers" <joseph@codesourcery.com> writes:
> 2008-09-04  Joseph Myers  <joseph@codesourcery.com>
>
> 	* config/mips/mips.h (enum reg_class): Add FRAME_REGS.
> 	(REG_CLASS_NAMES): Update.
> 	(REG_CLASS_CONTENTS): Update.
> 	* config/mips/mips.c (mips_regno_to_class): Use FRAME_REGS instead
> 	of ALL_REGS for regs 77 and 78.
> 	* function.c (instantiate_virtual_regs_in_insn): Assert that
> 	return value of simplify_gen_subreg is not NULL.
>
> testsuite:
> 2008-09-04  Joseph Myers  <joseph@codesourcery.com>
>
> 	* gcc.c-torture/compile/20080903-1.c: New test.

You're right that the existing code mapped 77 and 78 to ALL_REGS, but
there was another bug; the existing code didn't include 77 and 78 in...

> +  { 0x00000000, 0x00000000, 0x00006000, 0x00000000, 0x00000000, 0x00000000 },	/* frame registers */	\
>    { 0xffffffff, 0xffffffff, 0xffff07ff, 0xffffffff, 0xffffffff, 0x0fffffff }	/* all registers */	\

...this definition of ALL_REGS.  ALL_REGS really does have to be union
of other classes though, so please add 77 and 78 to ALL_REGS as part
of the same patch.

The MIPS parts are OK with that change, thanks.  I can't approve the
assert, but I think it counts as obvious.

(I'm not very keen on the idea of creating subregs of frame_pointer_rtx
and arg_pointer_rtx, but it's clear from this code that we do that on
other ports.)

Richard