Speed of development is key, and Python was a faster way to code (and re-code) the programs that control this production pipeline.

The mission of the Python Software Foundation is to promote, protect, and advance the Python programming language, and to support and facilitate the growth of a diverse and international community of Python programmers.

ILM runs a batch processing environment capable of modeling, rendering and compositing tens of thousands of motion picture frames per day.

Thousands of machines running Linux, IRIX, Compaq Tru64, OS X, Solaris, and Windows join together to provide a production pipeline used by ~800 users daily.

We could say "never trace nested loops" and not give up too much. Maybe we can figure all that out statically, but we might be able to tell better when we're ready to try an outer loop by how the inner loop is tracing.dmandelin, great analysis. It's most distinctive feature is that the main loop is enormous -- almost 20,000 LIR instructions get put into the writer pipeline.

(In comparison, most traced fragments have 10s or 100s of LIR instructions.) In one respect it's an excellent loop to trace, because it's nothing but arithmetic.

) runs 3967 times, which I would expect would give a nice speedup and be well worth tracing.

Is the trace compile time linear in the length of the loop body, or can it grow faster (either asymptotically, or at thresholds or something)? Otherwise it suggests some kind of "don't trace 'giant loops'" element to the heuristics. It looks like the other loops in md5 are don't cares.But the compile-time cost is really high, and the loop isn't run enough times to make up that cost.(I'd love to know how we compare against V8/Nitro if this benchmark checksummed 10x or 100x more text.) Not tracing this loop seems the right decision.It then runs the new benchmark suite with -m, -j, and then -m -j, and outputs a chart describing how long each loop took, and which mode is best for that loop. Attached is the initial spew for Sun Spider.bz: That's a good question.I am trying to find a good way to automatically correlate results with information like that, right now it's by hand. MCTS results Results for the four remaining interesting benchmarks, using a random search.For bugs involving browser objects such as "window" and "document", use the "DOM" component.