Comparison with byteplay and codetransformer

History of the bytecode API design

The design of the bytecode module started with a single use case: reimplement the CPython peephole optimizer (implemented in C) in pure Python. The design of the API required many iterations to get the current API.

bytecode now has a clear separation between concrete instructions using integer arguments and abstract instructions which use Python objects for arguments. Jump targets are labels or basic blocks. And the control flow graph abstraction is now an API well separated from the regular abstract bytecode which is a simple list of instructions.

byteplay and codetransformer

The byteplay and codetransformer are clear inspiration for the design of the bytecode API. Sadly, byteplay and codetransformer API have design issues (at least for my specific use cases).

Free and cell variables

Converting a code object to bytecode and then back to code must no modify the code object. It is an important requirement.

The LOAD_DEREF instruction supports free variables and cell variables. byteplay and codetransformer use a simple string for the variable name. When the bytecode is converted to a code object, they check if the variable is a free variable, or fallback to a cell variable.

The CPython code base contains a corner case: code having a free variable and a cell variable with the same name. The heuristic produces invalid code which can lead to a crash.

bytecode uses FreeVar and CellVar classes to tag the type of the variable. Trying to use a simple string raise a TypeError in the Instr constructor.

Note

It’s possible to fix this issue in byteplay and codetransformer, maybe even with keeping support for simple string for free/cell variables for backward compatibility.

Line numbers

codetransformer uses internally a dictionary mapping offsets to line numbers. It is updated when the .steal() method is used.

byteplay uses a pseudo-instruction SetLineno to set the current line number of the following instructions. It requires to handle these pseudo-instructions when you modify the bytecode, especially when instructions are moved.

In FAT Python, some optimizations move instructions but their line numbers must be kept. That’s also why Python 3.6 was modified to support negative line number delta in code.co_lntotab.

bytecode has a different design: line numbers are stored directly inside instructions (Instr.lineno attribute). Moving an instruction keeps the line number information by design.

bytecode also supports the pseudo-instruction SetLineno. It was added to simplify functions emitting bytecode. It’s not used when an existing code object is converted to bytecode.

Jump targets

In codetransformer, a jump target is an instruction. Jump targets are computed when the bytecode is converted to a code object.

byteplay and bytecode use labels. Jump targets are computed when the abstract bytecode is converted to a code object.

Note

A loop is need in the conversion from bytecode to code: if the jump target is larger than 2**16, the size of the jump instruction changes (from 3 to 6 bytes). So other jump targets must be recomputed.

bytecode handles this corner case. byteplay and codetransformer don’t, but it should be easy to fix them.

Control flow graph

The peephole optimizer has strong requirements on the control flow: an optimization must not modify two instructions which are part of two different basic blocks. Otherwise, the optimizer produces invalid code.

bytecode provides a control flow graph API for this use case.

byteplay and codetransformer don’t.

Functions or methods

This point is a matter of taste.

In bytecode, instructions are objects with methods like is_final(), has_cond_jump(), etc.

The byteplay project uses functions taking an instruction as parameter.