MaxGekk commented on PR #59069: URL: https://github.com/apache/spark/pull/59069#issuecomment-5969024898
Thank you all for the reviews. Rather than patch this PR further, I redesigned the change around them, and the result is #59225. I am closing this one in its favour. What #59225 does differently, in short: - One input collector (`collectInputs` with an `InputPolicy`) replaces the three walks, including the one this PR added, so the parameters of a split method, a subexpression method and a `With` method come from the same code. - The split is per block: a block that reads an input the operator has not evaluated yet stays inline between the calls, instead of refusing the whole `CASE WHEN`. - One grouping implementation, with at most 8 calls left in the operator's method under whole-stage codegen, since all of a projection's outputs land in one method there. - A stage gate: the stage's code is generated in one piece first and compiled as a trial, and only the `CASE WHEN`s in methods past the JIT limit are split. A stage the JIT compiles whole keeps exactly the code it has today, which answers the 64-branch regression measured here. - The `With` decision is recorded when it is made, the discarded subexpression pass drops its functions, and the public signature of `splitExpressionsWithCurrentInputs` is unchanged. The description there has the runner numbers for JDK 17, 21 and 25 and the differential fuzz. @dongjoon-hyun @cloud-fan @LuciferYang @HyukjinKwon, I would be grateful for your look at #59225. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
