lucasfang opened a new issue, #529:
URL: https://github.com/apache/paimon-cpp/issues/529
### Search before asking
- [x] I searched the repository issues and pull requests for
`fields_comparator`, GCC ICE, and `internal compiler error`, and found no
matching report.
### Paimon-cpp version
`main` at `caba0ff8210939d71b78ad385a1d864968ae5d1a`.
Environment: Ubuntu 24.04 x86_64, GCC 13.3.0 (Ubuntu
13.3.0-6ubuntu2~24.04.1), CMake 3.28.3, Debug / C++17. Clang 18.1.3
successfully compiles the same source.
### Minimal reproduce step
Configure the default feature set with tests and build the common test
executable:
```sh
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Debug -DPAIMON_BUILD_TESTS=ON
cmake --build build --target paimon-common-test -j 96
```
GCC crashes while instantiating
`FieldsComparatorTest::CheckFloatingPointOrder<float>`:
```text
fields_comparator_test.cpp:122:5: internal compiler error:
in gimplify_var_or_parm_decl, at gimplify.cc:3059
```
The triggering expression is the direct conversion of `{-nan}` to the
variant-backed `BinaryRowGenerator::ValueType`:
```cpp
const T nan = std::numeric_limits<T>::quiet_NaN();
// ...
BinaryRow negative_nan_row =
BinaryRowGenerator::GenerateRow({-nan}, pool.get());
```
The failure can also be reproduced without Paimon, Arrow, or GoogleTest:
```cpp
#include <cstdint>
#include <limits>
#include <string>
#include <variant>
#include <vector>
using Values = std::vector<std::variant<
bool, int8_t, int16_t, int32_t, int64_t, float, double, std::string>>;
void consume(const Values&);
template <typename T>
void check() {
const T nan = std::numeric_limits<T>::quiet_NaN();
consume({-nan});
}
template void check<float>();
template void check<double>();
```
Compile with `g++ -std=c++17 -O0 -g -c repro.cpp`. GCC 13.3.0 produces the
same ICE; Clang 18.1.3 succeeds. The reduced experiment also reproduces with
infinity and finite constants, so this is not specific to the NaN value. A
float/double-only variant does not reproduce in this experiment.
### What doesn't meet your expectations?
The default development configuration should compile the floating-point
comparator regression tests with GCC 13.3.0. Instead, the compiler crashes
before the tests can run.
### Anything else?
Materializing the negated value as `T` before passing it to the row
generator avoids the GCC failure:
```cpp
const T negative_nan = -nan;
BinaryRow negative_nan_row =
BinaryRowGenerator::GenerateRow({negative_nan}, pool.get());
```
This keeps the float/double special-value ordering assertions, including
sign-flipped NaN equality, intact.
Validation of this workaround:
- Full default Debug builds completed with GCC 13.3.0 and Clang 18.1.3.
- All 6 `FieldsComparatorTest.*` tests passed with each compiler.
- `ctest --test-dir build -L unittest --output-on-failure -j 96`: 32/32 test
executables passed (201.32 seconds).
- All configured pre-commit hooks passed against an isolated checkout
containing the patch.
- A scan of 2,195 repository-owned C++ files and manual review of the other
negated floating/template-constant candidates found no additional confirmed
instance of this conversion pattern.
### Are you willing to submit a PR?
- [x] I'm willing to submit a PR!
Generated-by: OpenAI Codex 26.1029.70936
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]