github-actions[bot] commented on code in PR #68113: URL: https://github.com/apache/doris/pull/68113#discussion_r4033283514
########## regression-test/data/query_p0/sql_functions/window_functions/test_window_float_sliding_frame.out: ########## @@ -0,0 +1,66 @@ +-- This file is automatically generated. You should know what you did if you want to edit this +-- !avg_double_preceding_1 -- +1 1.8014398509481984e+16 1.8014398509481984e+16 +2 1 9007199254740992 +3 1 1 +4 1 1 + +-- !sum_double_preceding_1 -- +1 1.8014398509481984e+16 1.8014398509481984e+16 +2 1 1.8014398509481984e+16 +3 1 2 +4 1 2 + +-- !avg_sum_float_preceding_1 -- +1 1.6777216E7 16777216 16777216 +2 1.0 8388608.5 16777217 +3 1.0 1 2 +4 1.0 1 2 + +-- !avg_sum_double_sign_flip -- +5 1.8014398509481984e+16 1.8014398509481984e+16 1.8014398509481984e+16 +6 1 9007199254740992 1.8014398509481984e+16 +7 -1.8014398509481984e+16 0 0 Review Comment: [P2] Avoid pinning a reassociation-dependent result This frame is `[2^54, 1, -2^54]`. The recorded `0` assumes left-to-right addition, but the equally legal grouping `(2^54 + -2^54) + 1` yields `1` (and AVG yields `1/3`). Both SUM and AVG mark their additions with `ALLOW_FP_REASSOCIATION`; [Clang's contract](https://clang.llvm.org/docs/LanguageExtensions.html#extensions-to-specify-floating-point-flags) permits reassociation across statements. `order_qt` fixes row order, not the compiler's grouping, so this exact output can fail after a supported compiler/target/codegen change even when the implementation is correct. Please remove this sign-cancellation assertion or use association-invariant values; the direct outgoing-row cases already cover the reported bug. -- 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]
