[Python-3000] Wither PEP 335 (Overloadable Boolean Operators)? (original) (raw)

Guido van Rossum guido at python.org
Fri May 25 04:53:40 CEST 2007

On 5/24/07, Greg Ewing <greg.ewing at canterbury.ac.nz> wrote:

Guido van Rossum wrote:

> Last call for discussion! I'm tempted to reject this -- the ability to > generate optimized code based on the shortcut semantics of and/or is > pretty important to me. Please don't be hasty. I've had to think about this issue a bit. The conclusion I've come to is that there may be a small loss in the theoretical amount of optimization opportunity available, but not much. Furthermore, if you take into account some other improvements that can be made (which I'll explain below) the result is actually better than what 2.5 currently generates. For example, Python 2.5 currently compiles if a and b: into JUMPIFFALSE L1 POPTOP JUMPIFFALSE L1 POPTOP JUMPFORWARD L2 L1: 15 POPTOP L2: Under my PEP, without any other changes, this would become LOGICALAND1 L1 LOGICALAND2 L1: JUMPIFFALSE L2 POPTOP JUMPFORWARD L3 L2: 15 POPTOP L3: The fastest path through this involves executing one extra bytecode. However, since we're not using JUMPIFFALSE to do the short-circuiting any more, there's no need for it to leave its operand on the stack. So let's redefine it and change its name to POPJUMPIFFALSE. This allows us to get rid of all the POPTOPs, plus the jump at the end of the statement body. Now we have LOGICALAND1 L1 LOGICALAND2 L1: POPJUMPIFFALSE L2 L2: The fastest path through this executes one less bytecode than in the current 2.5-generated code. Also, any path that ends up executing the body benefits from the lack of a jump at the end. The same benefits also result when the boolean expression is more complex, e.g. if a or b and c: becomes LOGICALOR1 L1 LOGICALAND1 L2 LOGICALAND2 L2: LOGICALOR2 L1: POPJUMPIFFALSE L3 L3: which contains 3 fewer instructions overall than the corresponding 2.5-generated code. So I contend that optimization is not an argument for rejecting this PEP, and may even be one for accepting it.

Do you have an implementation available to measure this? In most cases the cost is not in the number of bytecode instructions executed but in the total amount of work. Two cheap bytecodes might well be cheaper than one expensive one.

However, I'm happy to keep your PEP open until you have code that we can measure. (However, adding additional optimizations elsewhere to make up for the loss wouldn't be fair -- we would have to compare with a 2.5 or trunk (2.6) interpreter with the same additional optimizations added.)

-- --Guido van Rossum (home page: http://www.python.org/~guido/)

More information about the Python-3000 mailing list