Python 3.11 made .pyc files almost twice as large. There are two main reasons for this:
- PEP 659 made the bytecode stream ~3x as large as 3.10.
- PEP 657 made the location tables ~9x as large as 3.10.
(Note that these effects compound each other, since longer bytecode means more location entries.)
However, there is low-hanging fruit for improving this situation in 3.12:
- Bytecode can be compressed using a fairly simple scheme (one byte for instructions without an oparg, two bytes for instructions with an oparg, and zero bytes for CACHE entries). This results in serialized bytecode that is ~66% smaller than 3.11.
- The location table format already has a mechanism for compressing multiple code units into a single entry. Currently it's only used for EXTENDED_ARGs and CACHEs corresponding to a single instruction, but with slight changes the compiler can use the same mechanism to share location table entries between adjacent instructions. This is a double-win, since it not only makes .pyc files smaller, but also shrinks the memory footprint of all code objects in the process. Experiments show that this makes location tables ~33% smaller than 3.11.
When both of these optimizations are applied, .pyc files become ~33% smaller than 3.11. This is only ~33% larger than 3.10, despite all of the rich new debugging information present.
Linked PRs
Reactions are currently unavailable
Python 3.11 made .pyc files almost twice as large. There are two main reasons for this:
(Note that these effects compound each other, since longer bytecode means more location entries.)
However, there is low-hanging fruit for improving this situation in 3.12:
When both of these optimizations are applied, .pyc files become ~33% smaller than 3.11. This is only ~33% larger than 3.10, despite all of the rich new debugging information present.
Linked PRs