perf: Optimize NULL handling in lcm, gcd#21468
Merged
mbutrovich merged 2 commits intoapache:mainfrom Apr 10, 2026
Merged
Conversation
github-actions
Bot
added
sqllogictest
neilconway
changed the title
lcd, gcd NULL handlinglcd, gcd
mbutrovich
changed the title
lcd, gcdlcm, gcd
mbutrovich
approved these changes
Apr 10, 2026
Contributor
mbutrovich
left a comment
mbutrovich
left a comment
There was a problem hiding this comment.
Thanks @neilconway! 🚀
| let prim = arr.as_primitive::<Int64Type>(); | ||
| match scalar { | ||
| Some(scalar_value) if scalar_value != 0 && scalar_value != i64::MIN => { | ||
| // The gcd result divides both inputs' absolute values. When the |
Contributor
There was a problem hiding this comment.
This is a slick optimization.
alamb
added a commit
to apache/arrow-rs
that referenced
this pull request
Apr 14, 2026
…on_many` (#9692) ## Which issue does this PR close? - Closes #8809. ## Rationale for this change Several DataFusion PRs ([#21464](apache/datafusion#21464), [#21468](apache/datafusion#21468), [#21471](apache/datafusion#21471), [#21475](apache/datafusion#21475), [#21477](apache/datafusion#21477), [#21482](apache/datafusion#21482), [#21532](apache/datafusion#21532)) optimize NULL handling in scalar functions by replacing row-by-row null buffer construction with bulk `NullBuffer::union`. When 3+ null buffers need combining, they chain binary `union` calls, each allocating a new `BooleanBuffer`. `NullBuffer::union_many` reduces this to 1 allocation (clone + in-place ANDs). For example, from [#21482](apache/datafusion#21482): Before: ```rust [array.nulls(), from_array.nulls(), to_array.nulls(), stride.and_then(|s| s.nulls())] .into_iter() .fold(None, |acc, nulls| NullBuffer::union(acc.as_ref(), nulls)) ``` After: ```rust NullBuffer::union_many([ array.nulls(), from_array.nulls(), to_array.nulls(), stride.and_then(|s| s.nulls()), ]) ``` Per @alamb's [suggestion](#9692 (comment)), this PR also implements the general-purpose mutable bitwise operations on `BooleanArray` from #8809, following the `PrimitiveArray::unary` / `unary_mut` pattern. This builds on the `BitAndAssign`/`BitOrAssign`/`BitXorAssign` operators added to `BooleanBuffer` in #9567. ## What changes are included in this PR? **`NullBuffer::union_many(impl IntoIterator<Item = Option<&NullBuffer>>)`**: combines multiple null buffers in a single allocation (clone + in-place `&=`). Used by DataFusion for bulk null handling. **`BooleanArray` bitwise operations** (6 new public methods): Unary (`op: FnMut(u64) -> u64`): - `bitwise_unary(&self, op)` — always allocates a new array - `bitwise_unary_mut(self, op) -> Result<Self, Self>` — in-place if uniquely owned, `Err(self)` if shared - `bitwise_unary_mut_or_clone(self, op)` — in-place if uniquely owned, allocates if shared Binary (`op: FnMut(u64, u64) -> u64`): - `bitwise_bin_op(&self, rhs, op)` — always allocates, unions null buffers - `bitwise_bin_op_mut(self, rhs, op) -> Result<Self, Self>` — in-place if uniquely owned, `Err(self)` if shared, unions null buffers - `bitwise_bin_op_mut_or_clone(self, rhs, op)` — in-place if uniquely owned, allocates if shared, unions null buffers Note: #8809 proposed the binary variants take a raw buffer and `right_offset_in_bits`. This PR takes `&BooleanArray` instead, which encapsulates both and matches existing patterns like `BooleanArray::from_binary`. ## Are these changes tested? Yes. 23 tests for the `BooleanArray` bitwise methods and 6 tests for `union_many`, covering: - Basic correctness (AND, OR, NOT) - Null handling (both nullable, one nullable, no nulls, null union) - Buffer ownership (uniquely owned → in-place, shared → `Err` / fallback) - Edge cases (empty arrays, sliced arrays with non-zero offset, misaligned left/right offsets) ## Are there any user-facing changes? Six new public methods on `BooleanArray` and one new public method on `NullBuffer`. --------- Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Rich-T-kid
pushed a commit
to Rich-T-kid/datafusion
that referenced
this pull request
Apr 21, 2026
## Which issue does this PR close? - Closes apache#21467. ## Rationale for this change This PR implements three distinct optimizations: 1. `lcm` was computing the result NULL buffer iteratively. This is relatively slow. Switching to Arrow's `try_binary` kernel makes the implementation more concise and also improves performance by computing the result NULL buffer via the bitwise union of the input NULL buffers. 2. The `gcd` scalar arg path was doing similarly; switching to Arrow's `try_unary` yields a similar speedup. 3. For the `gcd` scalar path, computing the GCD can only fail in a few edge cases (e.g., `gcd(i64::MIN, i64::MIN)`). It is cheap to check for these edge-cases; for most `gcd` inputs, we can use Arrow's `unary` kernel instead of `try_unary`. The former is more efficient because it allows LLVM to vectorize the code more effectively. Benchmarks (ARM64): ``` - gcd array and scalar: 2.9ms → 2.2ms, -25% faster - lcm both array: 2.7ms → 2.0ms, -26% faster ``` ## What changes are included in this PR? * Add benchmark for `lcm` * Improve SLT test coverage * Move Rust unit test for `lcm` to SLT * Optimize `lcm` and `gcm` NULL handling * Optimize `gcm` to avoid overhead for edge cases ## Are these changes tested? Yes. Benchmark results above. I inspected the generated code for the `gcd` case to confirm that LLVM is able to generate better code for the `unary` case than for the `try_unary` case. ## Are there any user-facing changes? No.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Which issue does this PR close?
lcm,gcd#21467.Rationale for this change
This PR implements three distinct optimizations:
lcmwas computing the result NULL buffer iteratively. This is relatively slow. Switching to Arrow'stry_binarykernel makes the implementation more concise and also improves performance by computing the result NULL buffer via the bitwise union of the input NULL buffers.gcdscalar arg path was doing similarly; switching to Arrow'stry_unaryyields a similar speedup.gcdscalar path, computing the GCD can only fail in a few edge cases (e.g.,gcd(i64::MIN, i64::MIN)). It is cheap to check for these edge-cases; for mostgcdinputs, we can use Arrow'sunarykernel instead oftry_unary. The former is more efficient because it allows LLVM to vectorize the code more effectively.Benchmarks (ARM64):
What changes are included in this PR?
lcmlcmto SLTlcmandgcmNULL handlinggcmto avoid overhead for edge casesAre these changes tested?
Yes. Benchmark results above. I inspected the generated code for the
gcdcase to confirm that LLVM is able to generate better code for theunarycase than for thetry_unarycase.Are there any user-facing changes?
No.