This is a bit more interesting than the last H8 fix.

The H8 has peepholes to optimize masking a value down to a byte.  They
rely on using gen_lowpart to extract the low byte value of the source
operand which works well on the H8 because accessing the low byte of a
reg is particularly cheap.  THey also work on memory operands.

These peepholes don't really work on symbolic operands.  And if one
makes it into the peephole we'll abort in gen_lowpart.

Anyway, the peepholes currently accept "general_operand", which of
course allows those problematical symbolic operands.  Changing the
operand to "nonimmediate_operand" resolves the issue.  Note that
compile-time constants aren't really a concern here as the expression
would have already been simplified.

This fixes a half-dozen failures in the H8 testsuite.  More importantly
it makes it less likely that a test will suddenly fail because of
improvements to constant propagation or forward propagation.

Installing on the trunk.

Jeff