Two commits: first one separates the latch flip flops for the read and write data directions in the vme core. Not sharing flip-flops means that the router has more freedom in placement. (And might rid it of LUTs in the path to them). Please check the changes, especially in the commented out sections.
The second commit tried to only get data for the low 16 bits if we are not in 32/64 bit mode. Please check if I got the polarity of lword* right. Also not sure if the test for MBLT is needed there.