RFA: Fix rtl-optimization/22258

[Bernd Schmidt]

Joern RENNECKE wrote:
> The problem in rtl-optimization/22258 is that we combine an instruction 
> that
> requires a spill into a return value copy to a likely spilled register.
> If any part of the return value is live during i1, i2 and/or i3, it will be
> live in i3.  It is therefore sufficient to check liveness for i3.
> For a value that can be held in a single machine mode, a HARD_REG_SET
> would be overkill, so I use a bitmask in a single int, which the GNU coding
> standard says can be assumed to contain at least 32 bits on any of our host
> platforms.

Example?  Why isn't this caught by cant_combine_insn_p?


Bernd