HappenLee opened a new pull request, #67896:
URL: https://github.com/apache/doris/pull/67896
### What problem does this PR solve?
Issue Number: N/A
Related PR: N/A
Problem Summary: A non-nullable FLOAT/DOUBLE CASE can return NaN for a
finite selected branch when an unselected branch contains Infinity or NaN. For
example, with ordinary CASE evaluation, selecting `1`, an overflowing
multiplication, and `2` over three rows returns `NaN, Infinity, NaN` instead of
`1, Infinity, 2`. The same result assembly also loses the sign of selected
negative zero.
The result assembly multiplies each branch value by a zero/one mask and adds
it to the result. IEEE-754 arithmetic makes `0 * Infinity` and `0 * NaN` equal
NaN. Replace this floating point path with conditional stores so that selected
values are copied without arithmetic. This form generates AVX2 masked
loads/stores; a ternary assignment can instead become a conditional pointer
load that inhibits vectorization.
Add bitwise unit tests, SQL regression tests and a benchmark that directly
calls the production result assembly function.
### Release note
Fix incorrect FLOAT/DOUBLE CASE results caused by unselected non-finite
branch values, and preserve selected negative zero.
### Check List (For Author)
- Test
- [x] Regression test
- [x] Unit Test
- [x] Manual test (details below)
- [ ] No need to test or manual test.
Validation:
- All 12 new unit tests fail on the original implementation and pass with
ASAN after the fix. Coverage includes FLOAT/DOUBLE, uint8_t/uint16_t indices,
constant columns, non-finite values, signed zero, subnormal values, vector
tails, and 255/257 branches.
- `test_case_float_nonfinite` and `test_short_circuit_evaluation` pass.
The original implementation fails the new SQL regression. Golden output was
generated and independently verified through short-circuit evaluation using the
regression runner.
- ASAN BE build, clang-format 16, header hygiene, and clang-tidy checks
pass.
- RELEASE benchmarks built with Clang 21.1.8 and `-O3 -msse4.2 -mavx2` on
Xeon Platinum 8457C cover 42 scenarios. With identical input, a fixed CPU, five
repetitions per run and before/after/after/before ordering, median CPU time
decreases by approximately 5%–39% (18.5% geometric mean reduction). This
measures the result assembly function on this AVX2 machine, including result
allocation.
- Disassembly of all four floating point/index-width specializations in
the linked binaries confirms `vmaskmovps/pd`, replacing floating point
multiply/add instructions without fast-math.
- Behavior changed:
- [ ] No.
- [x] Yes. Return the selected floating point value without
contamination from other branches or loss of negative zero.
- 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
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
