Improving GPU Performance through Instruction Redistribution and Diversification