This is an automated email from the ASF dual-hosted git repository.
hello-stephen 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 0127314e7a5 [opt](exec) optimize fixed key packing (#65572)
0127314e7a5 is described below
commit 0127314e7a50fd43fa9a8b959bb73ba600fb304b
Author: Mryange <[email protected]>
AuthorDate: Tue Jul 28 10:39:16 2026 +0800
[opt](exec) optimize fixed key packing (#65572)
### What problem does this PR solve?
Fixed-key packing could lose `restrict` alias information when the
source data pointer was captured by reference in an outlined lambda,
producing a less efficient copy loop. This change obtains the restricted
pointer inside the lambda after nullable key replacement, preserving
alias information without changing the key encoding or query behavior.
---
be/src/exec/common/hash_table/hash_map_context.h | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/be/src/exec/common/hash_table/hash_map_context.h
b/be/src/exec/common/hash_table/hash_map_context.h
index 1dc33319dfe..9c5bbab2101 100644
--- a/be/src/exec/common/hash_table/hash_map_context.h
+++ b/be/src/exec/common/hash_table/hash_map_context.h
@@ -947,17 +947,21 @@ struct MethodKeysFixed : public MethodBase<TData> {
}
for (size_t j = 0; j < key_columns.size(); ++j) {
- const char* __restrict data = key_columns[j]->get_raw_data().data;
-
auto goo = [&]<typename Fixed, bool aligned>(Fixed zero) {
CHECK_EQ(sizeof(Fixed), key_sizes[j]);
if (has_null_column.size() && has_null_column[j]) {
const auto* nullmap =
assert_cast<const
ColumnUInt8&>(*nullmap_columns[j]).get_data().data();
// make sure null cell is filled by 0x0
+ // This mutates the same underlying buffer returned by
get_raw_data(), so get
+ // the restricted data pointer only after the replacement
finishes.
const_cast<IColumn*>(key_columns[j])->replace_column_null_data(nullmap);
}
auto* __restrict current = result_data + offset;
+ // Do not hoist data out of goo. A reference capture is
lowered to an ordinary
+ // closure field, so restrict/noalias information is lost when
goo is not inlined,
+ // which prevents the copy loop from being optimized.
+ const char* __restrict data =
key_columns[j]->get_raw_data().data;
for (size_t i = 0; i < row_numbers; ++i) {
memcpy_fixed<Fixed, aligned>(current, data);
current += sizeof(T);
@@ -969,6 +973,7 @@ struct MethodKeysFixed : public MethodBase<TData> {
// Also verify that the stride sizeof(T) is a multiple of
alignof(Fixed),
// otherwise alignment will be lost on subsequent loop
iterations
// (e.g. UInt96 has sizeof=12, stride 12 is not a multiple of
alignof(uint64_t)=8).
+ const char* data = key_columns[j]->get_raw_data().data;
if (sizeof(T) % alignof(Fixed) == 0 &&
reinterpret_cast<uintptr_t>(result_data + offset) %
alignof(Fixed) == 0 &&
reinterpret_cast<uintptr_t>(data) % alignof(Fixed) == 0) {
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
