4k should be plenty to evaluate the performance-critical parts of most applications for a '51 (usually, the CPU spends 90% of its time executing a handful of percent of the code).

And as soon as you look at pipelined architectures (or, even worse, quirks like instructions with delayed execution), writing anything in assembly will become a royal pain in the neck. And after that, you're not faster than the compiler yet, you merely have code that doesn't mess up because of pipeline conflicts - beating the compiler will take another round of work.