This is an automated email from the ASF dual-hosted git repository.
Mryange pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/doris.git
The following commit(s) were added to refs/heads/master by this push:
new 7596090adee [opt](column) Inline column creation to enable cast
vectorization (#69840)
7596090adee is described below
commit 7596090adee512f24cf7fe66dbc806468ce89901
Author: Mryange <[email protected]>
AuthorDate: Fri Oct 30 09:55:09 2026 +0800
[opt](column) Inline column creation to enable cast vectorization (#69840)
### What problem does this PR solve?
Issue Number: N/A
Problem Summary:
Some simple CAST loops remain scalar because
`COWHelper::create(Args&&...)` is not inlined. Root cause: forwarding a
row-count lvalue passes its address to the factory. When output stores
may alias that escaped variable, the compiler cannot prove the loop
bound is invariant and reports `UnsupportedUncountableLoop`.
On x86 with Clang 20.1.8, `-O3` and AVX2, the factory exceeds the inline
threshold for INT-to-BIGINT (cost 260 vs. 250) and
numeric/Decimal-to-BOOLEAN (255 vs. 250). This is not ARM-specific. Mark
the small factory wrapper `ALWAYS_INLINE` so the compiler can see
through column construction and optimize affected callers.
### Release note
None
### Check List (For Author)
- Test
- [x] Manual test: compiled `function_cast_int.cpp` before/after with
Clang optimization remarks and inspected assembly. The non-null
INT-to-BIGINT loop changes from scalar code to AVX2 `vpmovsxdq`, with
vectorization width 4 and interleave count 4. The null-map path still
has a separate vectorization blocker.
- [x] Manual test: diagnostic compilation of `function_cast_bool.cpp`
confirms the same escaped-bound blocker in 26 numeric/Decimal-to-BOOLEAN
instances; an isolated temporary-row-count experiment removes that
blocker. This BOOLEAN experiment validates the cause, not the final
patch across every instance.
- [x] Formatting, `git diff --check`, and
`build-support/check-build-hygiene.sh` passed.
- Full BE build, runtime benchmarks, SQL regression tests and ARM
recompilation were not run. In the INT translation-unit experiment,
object text grows from 448,711 to 478,128 bytes (about 6.6%); runtime
speedup has not been measured.
- Behavior changed:
- [x] No. SQL conversion semantics are unchanged.
- Does this need documentation?
- [x] No.
### Check List (For Reviewer who merge this PR)
- [ ] Confirm the release note
- [ ] Confirm test cases
- [ ] Confirm document
- [ ] Add branch pick label
---
be/src/core/cow.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/be/src/core/cow.h b/be/src/core/cow.h
index 7fb4cb4f754..89c6228223d 100644
--- a/be/src/core/cow.h
+++ b/be/src/core/cow.h
@@ -429,7 +429,7 @@ public:
// | ~~~~~~~ ^~~~~~~~~~~~~~~~~~~~~~~~
// ColumnPtr res_data_column = ColumnUInt8::create(1, 1);
template <typename... Args>
- static MutablePtr create(Args&&... args) {
+ ALWAYS_INLINE static MutablePtr create(Args&&... args) {
return MutablePtr(new Derived(std::forward<Args>(args)...));
}
#include "common/compile_check_avoid_end.h"
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
