It's a bit target dependent but say for a particular hardware and a shader use case, having a larger shader could mean that it couldn't have as many copies to q

It's a bit target dependent but say for a particular hardware and a shader use case, having a larger shader could mean that it couldn't have as many copies to quickly change context due to an event like a texture read. Context memory is a fixed resource and sometimes smaller shaders means you can fit more. So as the other answer states, profiling. It's also good to have an accurate mental model of the real hardware. Of course this is made difficult unless you are doing specific console development. And it's also usually not as useful to do synthetic profiling but profile in actual use.