py/compile: Remove "return if" optimisation, it's rarely used. #1505
Conversation
This code would optimise "return x if y else z" patterns, to save 2 bytes of RAM in the generated bytecode and eliminate a jump. It costs 56 bytes of code. It was there originally to agree with CPython generated bytecode. This coding pattern is not common and there are better places to spend 56 bytes of code space.
|
Well, my idea would be that compiler grows in complexity, not reduces ;-).
Probably yes, but 56 bytes is not that much.
Yes, replacing this with some other optimization would be more convincing ;-). Overall, if you're working on compiler and plan to add more optimizations then removing this may be grounded, but if it reandomly caught your eye, then why removing something on which previously effort was spent? |
Yes and no. Working on compiler/parser to implement various new ideas lead to finding this (and lead to 64f2b21). I wanted to drastically reduce parse tree size, implement persistent bytecode (#222) and do proper constant tables (#722). But this current PR is neither here nor there for those things. Anyway, reason for PR is to get other's opinions :) |
Well, then persistent bytecode is the most important of these things - if you have that, you can drop entire compiler ;-). |
|
Not yet... :) |
|
No real motivation for this so let's leave it for now. |
|
For the record, this optimisation was made configurable in ae54fbf |
Is it enabled for mpy-cross? |
Yes, all compiler optimisations are enabled for mpy-cross. |
Motivation for this PR is purely to reduce complexity of compiler.
The code that is being removed would optimise "return x if y else z" patterns, to save 2 bytes of RAM in the generated bytecode and eliminate a jump. It costs 56 bytes of code. It was there originally to agree with CPython generated bytecode.
This coding pattern is not common and there are better places to spend 56 bytes of code space.
I'm putting this up as a PR to see what other people's opinions are: is it worth optimising for a rarely used coding pattern? Arguably there are much more common patterns than this that could be optimised instead.