Repository navigation
Performance degradation 28.9.0+ #4804
Description
Activity
Iterator_INT_T::Get()runs once per accessed row.SetCurBlock()runs only when moving to another 64K block, butFORCE_INLINEforces its large reader and header-decoding path intoGet(). So, a large, rarely executed path was forced into a per-row function.After the reader templating change, introduced in manticoresoftware/columnar@dbc40aa this caused:
- A much larger hot function.
- A larger stack frame and stack-canary handling.
- Approximately 468 million additional retired branches per query.
- Additional instructions and register pressure on every row.
Removing single
FORCE_INLINElets compiler (both GCC and Clang) correctly keepSetCurBlock()out of line.- added 2 commits that reference this issue
on Oct 2, 2026
Metadata
Metadata
Assignees
Labels
No labels
Here we see a slight degradation after 28.9.0, followed by a more noticeable degradation after 28.9.1.
https://nightly.db-benchmarks.com/?cache=fast_avg&engines=manticoresearch_columnar_tuned_28.6.6%2Cmanticoresearch_columnar_tuned_28.9.0%2Cmanticoresearch_columnar_tuned_28.9.1%2Cmanticoresearch_columnar_tuned_28.9.3%2Cmanticoresearch_columnar_tuned_29.0.1&tests=hn&memory=110000&queries=0%2C1%2C2%2C3%2C4%2C5%2C6%2C7%2C8%2C9%2C10%2C11%2C12%2C13%2C14%2C15%2C16%2C17%2C18%2C19%2C20%2C21%2C22%2C23%2C24%2C25%2C26%2C27