Logged In: YES 
user_id=4771

Optimizing BINARY_ADD was what I did in the first patch too, but it has the drawback of being harder to explain.  The main trouble I see is that a=a+b+c will NOT be accelerated, while a=a+b or a+=b+c will.

The rationale for only tampering with INPLACE_ADD was to make the result more predictable.  Moreover the behavior would be superficially similar to what occurs with lists with a=a+b vs a+=b.

In-place operators in Python are a dark corner of the language.  I think it is OK for people to think about it as a possibly optimized (in-place-modifying) version of the standard operators.  Thus the idea of only tampering with INPLACE_ADD for the optimization here.

Well, maybe you can find a convincingly clear explanation of why a=a+b works but a=a+b+c doesn't.