The uzp1/uzp2 implementation has a number of problems  I had to
rewrite it.  It doesn't get the shifting/masking right.  it gets
input1 and input2 wrong.  It checks one bit instead of two for the
size field.  It doesn't fail for the non-full size==3 (1d) case which
should be an unallocated instruction.

The new testcase passes with the patch, and fails without.  The GCC C
testsuite failures go from 2269 to 2227 (-42).

Jim