In the cases I am talking about, a good C programmer would've created buffer caches which would've pre-allocated memory and reused it. Instead the programmers in question naively malloced when they needed it and freed when they were done. A generational collector does something almost as good as this for you -- they make transient objects very cheap.
Also, the Smalltalk in the block cipher case was written by a very good Smalltalker who knows how to write code that results in fast JITed machine code, who also had the luxury of requesting custom primitive operations like rotations on 32 bit registers. So if the C programs in question had the same level of programmer working on them, they would no doubt have blown away the Smalltalk implementations.
I agree that High Level implementations just get you pretty good for a little effort, and that C will get you darn good for a premium effort.
Also, the Smalltalk in the block cipher case was written by a very good Smalltalker who knows how to write code that results in fast JITed machine code, who also had the luxury of requesting custom primitive operations like rotations on 32 bit registers. So if the C programs in question had the same level of programmer working on them, they would no doubt have blown away the Smalltalk implementations.
I agree that High Level implementations just get you pretty good for a little effort, and that C will get you darn good for a premium effort.