We use a build of https://github.com/KhronosGroup/Vulkan-Samples from the commit eca122602a5d7a9f63686411dcef1b62feb36a9a for automated testing. When we run these samples on RHEL 10, we observed that the pipeline_cache test crashes, but it works on all the other distros.

On RHEL 10, this is our observation:

• The test serializes some data into a binary file named cache.data under /tmp. As far as we can see, this data is just some GLSL shaders and metadata (size, entry point etc.).
• Later on, the test opens this binary file and reads it back. While reading, it first reads the size, resizes an std::string so that the data would fit in it, and then reads the data into the resized std::string.
• Even though it seems to be reading the file correctly, the test keeps reading the file on RHEL 10 and CentOS 10 where it runs into EOF and breaks the loop on other distros.
• Finally, it reads some garbage metadata for size and tries to create a very large std::string, and it runs out of memory.

On other distros, the test correctly reads a binary file it created, encounters EOF, and exits the reading loop cleanly. However, on RHEL 10, the EOF condition is not reached and the test continues reading garbage data beyond the file boundary and crashes when attempting to process it. Since the behavior is isolated to RHEL 10/CentOS 10 we suspect there may be an issue with the libstdc++ packaged in RHEL 10. It almost looks like a C++ standard library function behaves differently.