|
|
|
|
|
|
by jerf
4406 days ago
|
|
|
It's easy to lose track of this in the torrent of "PyPy Benchmarks Show It Sped Up 2x!!!!" and "Latest Javascript Engine 50% Faster Than The Last One!!!! OMG!! Node!!!!", but in absolute terms, the dynamic language JITs are still quite slow. It's an urban legend that they are anywhere near compiled performance. Except LuaJIT. "Faster Python" is still slow. Unless you're in a tight loop adding numbers together, but I'm not. Moreover, a plot of their performance over time rather strongly suggests that they've plateaued. Personally, barring any major JIT advances, I'm considering the book closed on the topic of whether JITs can take a dynamic language and make it as fast as C or C++: No. (Recall that a Java JIT is another issue; it starts with the static typing of Java, and then JITs from there. It has a much larger array of much more powerful guarantees to work with, that the dynamic languages simply don't. Go will be running at straight-up C speeds (which it does not currently) long before Python will.) |
|
|
It's possible to close more of the speed gap, but these JITs have to be able to:
* identify and use native primitives (avoid overflow and so forth)
* prefer stack allocation over heap (improved escape analysis)
* inline memory allocation and freeing and remove unused codepaths
* optimize across the boundary between implementation language and hosted language (which I believe requires self-hosting)