This is an automated email from the ASF dual-hosted git repository.
yiguolei 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 bcbf74e917b [fix](be) Fix Parquet FLOAT16 compilation with older
libstdc++ (#69666)
bcbf74e917b is described below
commit bcbf74e917b539dbce925a19be8e7e385e9435bf
Author: EscaperLL <[email protected]>
AuthorDate: Tue Oct 27 18:22:38 2026 +0800
[fix](be) Fix Parquet FLOAT16 compilation with older libstdc++ (#69666)
### What problem does this PR solve?
Issue Number: N/A
Related PR: #69647 (closed without merging)
Problem Summary: Parquet FLOAT16 conversion calls `std::powf(2.0F,
-14.0F)`. Older libstdc++ headers do not expose `powf` in namespace
`std`, causing BE compilation to fail even when `<cmath>` is included.
This is reproducible with GCC 12.3.0 and Clang 22.1.8 using their
default libstdc++ configurations, although these compiler versions pass
Doris's compiler version checks.
GCC added the missing namespace declarations, including `powf`, in
[PR79700 /
r14-5341](https://gcc.gnu.org/pipermail/gcc-bugs/2023-November/841727.html).
Include `<cmath>` explicitly and replace the call with `std::pow(2.0F,
-14.0F)`. Both arguments remain `float`, selecting the overload that
returns `float` and preserving the existing conversion semantics. This
fixes standard library API availability without changing the Parquet
format or FLOAT16 encoding.
The PR contains one fix commit changing only
`be/src/format/parquet/parquet_column_convert.h` (+2/-1). It supersedes
#69647 with validation on master baseline `da4f490ae`; no LZ4HC changes
are included.
### Release note
Fix BE compilation for Parquet FLOAT16 conversion with older libstdc++.
### Check List (For Author)
- Test
- [ ] Regression test
- [x] Unit Test
- [x] Manual test (steps and results below)
- [ ] No need to test or manual test. Explain why:
Validation on this branch:
- **Unit tests:** `BUILD_TYPE_UT=ASAN ./run-be-ut.sh --run
--filter='ParquetColumnConvertTest.*' -j 6` passed all 17 tests, with no
failures, errors, disabled tests, or skips. This suite does **not**
directly exercise the FLOAT16 branch.
- **Compilation reproducer:** GCC 12.3.0 and Clang 22.1.8 both reject
the original call and accept the replacement in C++20 mode. A
`static_assert` confirms that the replacement returns `float`;
reproduction steps are below.
- **Numerical check:** The conversion function extracted from the
current header matches a `::powf` baseline bit-for-bit for all 65,536
FLOAT16 patterns with both compilers at `-O0` and `-O2`. All subnormal
values also match an independent exact reference. This standalone check
does not replace a complete FLOAT16 file-reading regression test.
- **Style and static checks:** clang-format 16 formatting/check scripts,
build hygiene, and `git diff --check` passed. The repository clang-tidy
script reported success with a broken-pipe message; a subsequent direct
clang-tidy invocation filtered to the changed lines exited successfully
without diagnostics on those lines.
- **Full BE build limitation:** `BUILD_TYPE=ASAN ./build.sh --be -j 2`
completed source compilation and linked `doris_be`, but the overall
command **failed** the glibc 2.17 compatibility gate. The artifact
requires newer symbols, including `statx@GLIBC_2.28` and
`hypotf@GLIBC_2.35`. The linked `powf` requirement is `GLIBC_2.2.5`,
which does not establish compatibility of the whole artifact. The gate
was not bypassed; dictionary generation and packaging were not verified
complete. A successful standard CI build is still needed before merging.
- **Regression limitation:** `test_hdfs_tvf_float16` was not run because
the external Hive/HDFS test environment is not enabled
(`enableHiveTest=false`).
<details>
<summary>Minimal compilation reproducer</summary>
Run with GCC 12.3.0 or Clang 22.1.8 configured to use the affected
libstdc++ headers:
```bash
cat > /tmp/doris_float16_original.cpp <<'CPP'
#include <cmath>
#include <type_traits>
auto value = std::powf(2.0F, -14.0F);
static_assert(std::is_same_v<decltype(value), float>);
CPP
sed 's/std::powf/std::pow/' /tmp/doris_float16_original.cpp \
> /tmp/doris_float16_fixed.cpp
# Expected: compilation error for the original; success for the replacement.
g++ -std=c++20 -fsyntax-only /tmp/doris_float16_original.cpp
g++ -std=c++20 -fsyntax-only /tmp/doris_float16_fixed.cpp
clang++ -std=c++20 -fsyntax-only /tmp/doris_float16_original.cpp
clang++ -std=c++20 -fsyntax-only /tmp/doris_float16_fixed.cpp
```
</details>
- Behavior changed:
- [x] No. FLOAT16 conversion semantics are preserved.
- [ ] Yes.
- Does this need documentation?
- [x] No.
- [ ] Yes.
### Check List (For Reviewer who merge this PR)
- [ ] Confirm the release note
- [ ] Confirm test cases
- [ ] Confirm document
- [ ] Add branch pick label
---
be/src/format/parquet/parquet_column_convert.h | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/be/src/format/parquet/parquet_column_convert.h
b/be/src/format/parquet/parquet_column_convert.h
index 3374a3e2dc9..98cecad5d13 100644
--- a/be/src/format/parquet/parquet_column_convert.h
+++ b/be/src/format/parquet/parquet_column_convert.h
@@ -22,6 +22,7 @@
#include <libdivide.h>
#include <chrono>
+#include <cmath>
#include <limits>
#include "common/cast_set.h"
@@ -614,7 +615,7 @@ public:
// half subnormal:
// value = (-1)^sign * (mant / 2^10) * 2^(1 - bias)
// half bias = 15 → exponent = 1 - 15 = -14
- float f = (static_cast<float>(mant) / 1024.0F) *
std::powf(2.0F, -14.0F);
+ float f = (static_cast<float>(mant) / 1024.0F) *
std::pow(2.0F, -14.0F);
return sign ? -f : f;
}
} else if (exp == 0x1F) {
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
