A brief note and question about how JRuby optimizes case/when
statements in some cases.
When a case/when is homogeneous, JRuby may emit an optimized version
that is faster than O(n) performance, as in other implementations. For
example, if the elements are all Fixnums, JRuby will emit a JVM
bytecode "lookupswitch" like this:
case a
when 1
when 2
when 3
when 4
end
...yields...
LOOKUPSWITCH
1: L2
2: L3
3: L4
4: L5
default: L6
This optimization is done for Fixnum, Symbol, and String. In the case
of Symbol and String, there are two modes: single-character (byte),
which is compiled similar to the Fixnum version; and hash-based. The
hash-based version uses the hash of the String or Symbol as the key
for the lookupswitch, as in this case:
case a
when 'foo'
when 'bar'
when 'baz'
end
...yields...
LOOKUPSWITCH
841490416: L2
841490424: L3
876516207: L4
default: L5
Now there's actually a subtle bug here I need to fix: it doesn't
compensate for hash collisions. Nobody's ever reported it, so it's
probably not common, but I'll fix it anyway.
So, my question is this: how common is it for folks to use homogeneous
case/when constructs with literal Fixnums, Symbols, or Strings? Is
this optimization worth keeping (and fixing)?
- Charlie
---------------------------------------------------------------------
To unsubscribe from this list, please visit:
http://xircles.codehaus.org/manage_email
