enforcing mode matching in SET

[Torbjorn Granlund]

  I've just fixed the violations in the sparc.md file as well.
  I did this only for the mainline for the moment as it seems only
  new patterns I introduced in the sparc backend rewrite have this
  problem, but I'll doublecheck the 1.1 branch later just in case.

I think it would be a mistake to fix these problems in egcs 1.1, since there
is a significant risk that it would break things.  The RTL might be
generated with wrong modes in non-obvious ways, both in the port files and
in the MI compiler.

Mode clashes have been in the md files for a long time and they have proven
rather harmless.  I suggest that we fix them for 2.2.

Torbjorn