[llvm] [RISCV] Enable shouldMaximizeVectorBandwidth (PR #216005)

[Pengcheng Wang via llvm-commits]

wangpc-pp wrote:

This improves some internal micro benchmarks as this avoids some fractional LMULs when mixed precision operations are common. @lukel97 Can measure this on LNT? 

https://github.com/llvm/llvm-project/pull/216005