It is impossible to write efficient code in any language without a thorough understanding of the underlying architecture.
correct, but, in my opinion, more important is that

It is impossible to write efficient code in any language without a selecting an underlying architecture fit for the purpose.

That said, there will always be compromises. An example would be a control application with some data handling. Since any compiler maker in his right mind would gear the efforts of compiler efficiency towards the most used functions for that particular architecture, you may find that using the "wrong" processor really hurt in such areas. I have an app that run 97% control and 3% heavy data (merging 3 arrays into one). For that particular function I achieved better than 80% throughput improvement using assembler. Now in the "control area" I wrote an ISR in C and took a brief glance and I doubt I could gain 5% by rewriting that in C, so i left it alone.

So, again, when using the '51 for what is it intended for, C is quite efficient, at least Keil C is.

Erik