Racket 7.6

The idea is that someday the CS version will be faster than the traditional version. Now the CS is faster in some benchmarks, and slightly slower (205? 50%?) in others. IIRC the worst cases are in code that uses too many mutable pairs, and the futures block more often.

IIRC Many of the changes of the last year were to improve the transformation from Racket to Chez Scheme. For example, some structs can be used as function in Racket, so in (f 1 2 3) the transformation must add a check to see if f is a real function or a struct that contains a function. If the transformation know which variables contain real functions and which struct, then the checks can be avoided, and hopefully the function can used in the constant folding and other transformation.

(Some of the checks can be eliminated by the optimizations in Chez Scheme, but some are just too difficult, so it's important to improve the transformation. For example, I think that the unnecessary continuation marks are very difficult to remove using only the information available in Chez Scheme, and the continuation marks are used heavily when the debug mode is enabled.)

EDIT: I have found disk IO to be perhaps slow, or at least in an example I wrote that loads one of my trained Keras/TensorFlow models (saved as a CSV file) into memory and uses the native matrix library to use the model for predictions. The whole thing takes about two seconds to load which seems like a long time.