benchmarks avoid dead code elimination

[bitfaster]

This appears to have little effect on the result, but is best practice as mentioned here.

[coveralls]

coverage: 99.217%. remained the same
when pulling 8693304 on users/alexpeck/benchfix
into 1e7011e on main.