Posted on October 24, 2014 at 15:12First, I want to say that the benchmark numbers are really impressive. I wish they had reported a bit more details though.
They compare against a ''Compiler 2'' and Compiler 3''. I wish they had clearly said which version of which compiler with which options was used. My guess is that ''Compiler 2'' is some version of the IAR compiler and ''Compiler 3'' is some version of the Cosmic compiler. They probably consider sdcc too insignificant to include it.
They only report code size, not benchmark results. I suspect that means they only compiled the benchmarks and linked them against the library, but there is no code to set up the device and report the results via RS232. Also it means we have no comparison on the code speed.
They mention the compiler performs link-time optimization. That is kind of an issue with at least Dhrystone: Dhrystone is from a time when there was no link-time optimization. Dhrystone uses multiple source files to simulate data depending on user input, which won't work well with link-time optimization. I suggest to include a more modern benchmark, such as Coremark.
Philipp
P.S.: About half a year ago, Dhrystone code size when compiled with sdcc using strong optimization for code size was at 7948 bytes. Including everything and ready to run on an STM8/128-EVAL and report the results through UART1. Even if Rasonance left out the UART stuff, their result of 2400 bytes is very impressive. It also shows that sdcc still has a long way to go.
On the other hand, with code that is not as much in the spotlight as Dhrystone I usually see sdcc code size only about 20% bigger than Cosmic or IAR (don't have Raisonance to compare).
