Some performance oddities in numerical comparisons on merge tables