A generated JIT compiler for Python
| Author: | Samuele Pedroni, Open End AB |
|---|---|
| Date: | December '09 |
Idea
Derive a dynamic compiler from a sufficiently rich intermediate representation of the interpreter while it is turned into a VM
- all features supported by construction
- semantics are more easily encoded as interpreting code
Tracing JIT compiler
Template: use the streamlined approach introduced by tracing JITs
- identify hot loops
- produce linear traces with stay-on-trace checks (guards), follow calls!
- compile the traces to assembler which run the loop
Details
- Identifying loops: hints for the JIT runtime to identify the interpreter loop, and back edges of user loops
- Tracing: build encoding such that interpreter functions can be executed and their operation recorded in the traces
- Exploit when heap allocation can be forgone, mirror Python frames on the CPU stack
Example
Status
- Intel x86-32
- Small and some medium examples run
- Good results for algorithmic code, see benchmarks
- Reasonable heuristic when traces get too long
Benchmarks
Work Left
- Further work on memory footprint
- Correct and improved behavior in the presence of exceptions: work in progress
- Better results for recursion, asm to asm calls
- Bugs, correctness (sys.settrace for example)
Work Left (ii)
- Machine code management, discard no longer valid traces
- Make the interpreter more JIT-friendly
- Interpreter-level algorithmic improvements
- Intel x86-64