https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92499
--- Comment #2 from sandra at gcc dot gnu.org --- As I suspected, MIPS has exactly the same bug. It generates GP-relative addressing in get_flexible and get_flexible_nonzero, and puts flexible_nonzero in .sdata instead of .data. This isn't terribly surprising since they both use int_size_in_bytes (TREE_TYPE (exp)) to get compute the object size in the TARGET_IN_SMALL_DATA_P hook, and SYMBOL_REF_SMALL_P (which uses that hook) to determine whether to use GP-relative addressing. I checked all the targets that define TARGET_IN_SMALL_DATA_P if they use int_size_in_bytes: alpha: Y arc: Y c6x: N (looks like it rejects objects of aggregate type instead) frv: Y ia64: Y lm32: Y m32r: Y microblaze: Y mips: Y nios2: Y riscv: Y rs6000: Y rx: Y Looks like we need a new function in common code that returns 0 for types with flexible array members and int_size_in_bytes otherwise, and use it in all of these backends.
