[Python-Dev] Re: Annotating pure functions to improve inline caching/optimization
> I wonder how this caching works, given that the dynamic nature means > that virtually every operation could have side effects, causing wrong > behaviour when cached. The only mitigation for this that I can imagine > is that caching just occurs for basic operations defined in the standard > library, where it is known that they are free of side effects or "pure". I've frequently explored the new adaptive, inline caching code generated by 3.11. "inline caching" does not mean result caching (like what C/C++ might do) here, but rather it should mean the caching of info used for the adaptive instructions. That means the bytecode stays the same except for the adaptive instructions and the changed `CACHE`s below each adaptive instruction, which should *always* be skipped due to the possibility of misinterpretation as other instructions. ___ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-le...@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/2EYWR5OLDNCFRP5QEG7K5RV4XNPQX6WS/ Code of Conduct: http://python.org/psf/codeofconduct/
[Python-Dev] Re: Annotating pure functions to improve inline caching/optimization
You should bring this up on discuss.python.org. It's not going to see much if any discussion here. Eric On 9/14/2022 10:05 AM, Philipp Burch wrote: Hello everyone, the docs on the upcoming 3.11 release state > This [specializing adaptive interpreter] also brings in another concept called inline caching, where Python caches the results of expensive operations directly in the bytecode. I wonder how this caching works, given that the dynamic nature means that virtually every operation could have side effects, causing wrong behaviour when cached. The only mitigation for this that I can imagine is that caching just occurs for basic operations defined in the standard library, where it is known that they are free of side effects or "pure". A web search did reveal some discussions[1,2] and a module related to dealing with pure functions, but, as far as I see, not related to optimization. As an example, consider a code like this: ```py @pure def rot_factor(angle_deg: float) -> complex: # This could also be a much more expensive calculation. return cmath.exp(angle_deg / 180 * cmath.pi * 1j) # ... res: List[Tuple(complex, complex, complex, float)] = [] for x in many: res.append(( x * rot_factor(90), x * rot_factor(45), x * rot_factor(-45), x * math.sin(math.pi/8), )) ``` The problem with this code is obvious, every loop iteration calls `rot_factor()` with 90, 45 and -45 and will get exactly the same set of results. The last factor might already be inline cached by the interpreter, since it probably knows that `math.pi` is a constant and `math.sin()` is a pure function. Optimizing this by hand (not considering a list comprehension or other more sophisticated improvements) is easy, but not very pretty: ```py f_p90 = rot_factor(90) f_p45 = rot_factor(45) f_m45 = rot_factor(-45) f_sin = math.sin(math.pi / 8) res: List[Tuple(complex, complex, complex, float)] = [] for x in many: res.append(( x * f_p90, x * f_p45, x * f_m45, x * f_sin, )) ``` I actually find myself often factoring such data out of loops in Python, whereas in C I would just leave that to the optimizer/compiler. An available option would be to use `@lru_cache` for `rot_factor()`, but this will still cause the same dictionary lookups in every iteration and it may not work at all in case the function argument(s) is/are not hashable. Now, if the interpreter understood the `@pure` decorator for `rot_factor()` indicated above would give it the same opportunity to cache the three results throughout the loop, basically creating the manually-optimized code above. For these completely static values, it could even precompute the results and integrate them into the bytecode. Has anything like this been considered already, or is the interpreter itself capable to perform such optimizations? Thanks and best regards, Philipp [1] 'pure' type annotation for mypy: https://github.com/python/mypy/issues/4468 [2] pure-func module: https://pypi.org/project/pure-func/ ___ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-le...@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/XDYZRC6L7LBPE3T6RO6S5IVY3J6IMRSJ/ Code of Conduct: http://python.org/psf/codeofconduct/ ___ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-le...@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/A5CATJEKKNZN4XOWXVJ2ICK2Z4A3WFYU/ Code of Conduct: http://python.org/psf/codeofconduct/
[Python-Dev] Annotating pure functions to improve inline caching/optimization
Hello everyone, the docs on the upcoming 3.11 release state > This [specializing adaptive interpreter] also brings in another concept called inline caching, where Python caches the results of expensive operations directly in the bytecode. I wonder how this caching works, given that the dynamic nature means that virtually every operation could have side effects, causing wrong behaviour when cached. The only mitigation for this that I can imagine is that caching just occurs for basic operations defined in the standard library, where it is known that they are free of side effects or "pure". A web search did reveal some discussions[1,2] and a module related to dealing with pure functions, but, as far as I see, not related to optimization. As an example, consider a code like this: ```py @pure def rot_factor(angle_deg: float) -> complex: # This could also be a much more expensive calculation. return cmath.exp(angle_deg / 180 * cmath.pi * 1j) # ... res: List[Tuple(complex, complex, complex, float)] = [] for x in many: res.append(( x * rot_factor(90), x * rot_factor(45), x * rot_factor(-45), x * math.sin(math.pi/8), )) ``` The problem with this code is obvious, every loop iteration calls `rot_factor()` with 90, 45 and -45 and will get exactly the same set of results. The last factor might already be inline cached by the interpreter, since it probably knows that `math.pi` is a constant and `math.sin()` is a pure function. Optimizing this by hand (not considering a list comprehension or other more sophisticated improvements) is easy, but not very pretty: ```py f_p90 = rot_factor(90) f_p45 = rot_factor(45) f_m45 = rot_factor(-45) f_sin = math.sin(math.pi / 8) res: List[Tuple(complex, complex, complex, float)] = [] for x in many: res.append(( x * f_p90, x * f_p45, x * f_m45, x * f_sin, )) ``` I actually find myself often factoring such data out of loops in Python, whereas in C I would just leave that to the optimizer/compiler. An available option would be to use `@lru_cache` for `rot_factor()`, but this will still cause the same dictionary lookups in every iteration and it may not work at all in case the function argument(s) is/are not hashable. Now, if the interpreter understood the `@pure` decorator for `rot_factor()` indicated above would give it the same opportunity to cache the three results throughout the loop, basically creating the manually-optimized code above. For these completely static values, it could even precompute the results and integrate them into the bytecode. Has anything like this been considered already, or is the interpreter itself capable to perform such optimizations? Thanks and best regards, Philipp [1] 'pure' type annotation for mypy: https://github.com/python/mypy/issues/4468 [2] pure-func module: https://pypi.org/project/pure-func/ ___ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-le...@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/XDYZRC6L7LBPE3T6RO6S5IVY3J6IMRSJ/ Code of Conduct: http://python.org/psf/codeofconduct/
