[Python-Dev] Do more at compile time; less at runtime

[Mark Shannon]

Hi all,

The current CPython bytecode interpreter is rather more complex than it 
needs to be. A number of bytecodes could be eliminated and a few more 
simplified by moving the work involved in handling compound statements 
(loops, try-blocks, etc) from the interpreter to the compiler.

This simplest example of this is the while loop...
while cond:
    body

This currently compiled as

start:
    if not cond goto end
    body
    goto start
end:

but it could be compiled as

goto test:
start:
     body
     if cond goto start

which eliminates one instruction per iteration.

A more complex example is a return in a try-finally block.

try:
     part1
     if cond:
         return X
     part2
finally:
     part3

Currently, handling the return is complex and involves "pseudo 
exceptions", but if part3 were duplicated by the compiler, then the 
RETURN bytecode could just perform a simple return.
The code above would be compiled thus...

     PUSH_BLOCK try
     part1
     if not X goto endif
     push X
     POP_BLOCK
     part3           <<< duplicated
     RETURN_VALUE
endif:
     part2
     POP_BLOCK
     part3           <<< duplicated

The changes I am proposing are:

Allow negative line deltas in the lnotab array (bytecode deltas would 
remain non-negative)
Remove the SETUP_LOOP, BREAK and CONTINUE bytecodes
Simplify the RETURN bytecode
Eliminate "pseudo exceptions" from the interpreter
Simplify (or perhaps eliminate) SETUP_TRY, END_FINALLY, END_WITH.
Reverse the sense of the FOR_ITER bytecode (ie. jump on not-exhausted)


The net effect of these changes would be:
Reduced code size and reduced code complexity.
A small (1-5%)? increase in speed, due the simplification of the
bytecodes and a very small change in the number of bytecodes executed.
A small change in the static size of the bytecodes (-2% to +2%)?

Although this is a quite intrusive change, I think it is worthwhile as 
it simplifies ceval.c considerably.
The interpreter has become rather convoluted and any simplification has 
to be a good thing.

I've already implemented negative line deltas and the transformed while 
loop: https://bitbucket.org/markshannon/cpython-lnotab-signed
I'm currently working on the block unwinding.

So,
Good idea? Bad idea?
Should I write a PEP or is the bug tracker sufficient?

Cheers,
Mark.