Fix a bug in optimising division to shift right

Fix a bug in optimising division to shift right
Division by an integral log2 can't be directly optimised to a shift
right, because shift right behaves like a division that rounds to
negative infinity, whereas we want one that rounds to zero.  Fix this
by adding (divisor-1) to the dividend when it is negative before
shifting.  We do this without jumps, generating very slightly worse
code than gcc, which uses conditional moves on CPUs that support it.