Skip to end of metadata
Go to start of metadata

Problem

Bit operations don't have the same semantics that the rest of Clojure's numeric functions do. Primitive semantics should be the default, there should be no conditionals, and bit ops should map directly to primitive ops.

Steps

  1. Bit ops have primitive semantics by default (CLJ-772 bit ops to have primitive semantics by default, no conditionals, direct mapping to JVM primitive ops)
  2. Bit ops with boxed semantics are in another ns (CLJ-767 Remove support for non-primitive bit-shift operations)
  3. n-ary bit ops are inlined (CLJ-184 n-ary bit functions, also inlining of n-ary bit and math operations)

Questions

  1. Does "boxed semantics" mean negative shift distance flips the direction?  Does it mean anything else, e.g., returning BigInt on overflow?  Can this just be left to the API on BigInteger? SDH: negative shifts should be eliminated everywhere. There is (should be?) no such thing as overflow in bit ops, and there should be no promotion to BigInt.
  2. Changing bit ops Numbers methods to map directly to primitive ops means a breaking change with respect to negative bit-shift distances.  Is this intended/acceptable? SDH: No negative shifts. Patch that simply removes the negative shift stuff would be welcome.
  3. Assuming no overflow support, is there any reason bit ops should not all return long, as opposed to sometimes Object? SDH: Always long is good.
    1. If no, do we still need the BitOps interface? SDH: Good point, BitOps no longer needed.
    2. Can the bit ops fns be written using prim interfaces? SDH: Should be able to work with no interfaces, merely four overloads: long/long, long/Object, Object/long, and Object/Object.
  4. Support of n-ary ops via reduce causes a multiple orders of magnitude performance hit. Is unbounded expansion an acceptable solution for inline n-ary support?  Is there work underway for better support for reduce when working with primitive-taking/returning fns? SDH: Macro expansion is good, won't be unbounded unless your code is infinite. There is a patch that needs review/refreshing.

Related Tickets

Labels:
  1. Apr 25, 2011

    seems like for #4, when inlining n-ary ops, the inline function should generate a form via reduce at compile time

    1. Apr 25, 2011

      If you mean generate something like (foo x (foo y (foo z ...))), then I agree.  That's what I meant by "unbounded expansion".