When testing the fix of
MDEV-28371 using its RQG test case on a 32-bit executable, there turned out to be a different bug, reported in CMAKE_BUILD_TYPE=RelWIthDebInfo as follows:
The page number difference is exactly 1<<20, a constant that occurs in the function page_id_t::fold(). This failure is independent of the
There is a debug assertion in mtr_t::page_lock() that also catches this:
The cause is not obvious to me. cmake -DWITH_UBSAN=ON is not reporting anything before the crash. The test does not fail on 64-bit platforms.