[jira] [Created] (AVRO-4360) [C++] Float and double binary codecs use native byte order on big-endian hosts

["Manu Zhang (Jira)" <ji...@apache.org>]

Manu Zhang created AVRO-4360:
--------------------------------

             Summary: [C++] Float and double binary codecs use native byte order on big-endian hosts
                 Key: AVRO-4360
                 URL: https://issues.apache.org/jira/browse/AVRO-4360
             Project: Apache Avro
          Issue Type: Bug
          Components: c++
            Reporter: Manu Zhang


Avro C++ binary float/double encoding and decoding copy native memory bytes directly, while the Avro binary format requires little-endian IEEE 754 bytes.

On a big-endian host such as s390x:
* Encoding float 1.0 produces 3f 80 00 00 instead of 00 00 80 3f.
* Encoding double 1.0 produces 3f f0 00 00 00 00 00 00 instead of 00 00 00 00 00 00 f0 3f.
* Decoding valid little-endian Avro bytes produces incorrect values.

Affected methods:
* lang/c++/impl/BinaryEncoder.cc: BinaryEncoder::encodeFloat and encodeDouble.
* lang/c++/impl/BinaryDecoder.cc: BinaryDecoder::decodeFloat and decodeDouble.

The defect is present on current main, commit 28cb08c15543c8c2dd8d186db2029bfe760940a5. An encode/decode round trip on a single big-endian host can conceal it because the two incorrect operations cancel each other out.

Reproduction and validation:
* Independent fixed wire-byte tests for float and double, covering positive and negative zero, 1.0, -2.5, subnormal and normal limits, largest finite values, infinities, and quiet NaN payloads.
* Stream chunk sizes of 1, 3 and 16 bytes.
* GCC 14 s390x cross-build under QEMU: the original codec fails 120 of 132 assertions; the corrected codec passes all 132. This is emulation rather than physical s390x hardware.
* Complete CodecTests passes on native macOS and Linux ARM64: 852 test cases and 103,659 assertions on each.

Expected behavior: encode and decode the required little-endian wire bytes while preserving raw IEEE 754 bits.

Proposed fix uses C++20 bit_cast and endian-aware byte reversal in all four methods. Regression tests check each direction independently.

Pull request: https://github.com/apache/avro/pull/4043
Specification: https://avro.apache.org/docs/1.12.0/specification/




--
This message was sent by Atlassian Jira
(v8.20.10#820010)