On Sat, Aug 21, 2021 at 1:56 PM Aileen Honess <aileen@xxxxxxxxxxxxxxxxx> wrote: > When I configure and create a gcc cross-compiler for aarch64-unknown-elf, > I get a valid and working libatomic. Critically, this contains > implementations of the methods that underlie the gcc atomic builtins > (__atomic_load(), etc.). I mean things like '__atomic_fetch_and_1()'. > When I do similar for riscv64-unknown-elf, I do not get libatomic. The > “configure.tgt” for libatomic says that the platform is UNSUPPORTED. The > build continues, but if I use the atomic builtins, I’ll get undefined > references to those underlying functions. > libatomic fails to build for all *-elf targets. So it is failing to build for both aarch64-elf and riscv-elf. The difference here is that rv32 only has atomic support for 32-bit words, and rv64 only has atomic support for 32-bit words and 64-bit double words. Whereas aarch64 v8.1-a has atomic support for all integer types, so can emit code directly for operations like __atomic_fetch_and_1. If compiling for RISC-V or armv8.0 then you will get calls to functions for unsupported types. aarch64 apparently has a library somewhere to implement these functions for *-elf toolchains. RISC-V does not yet. We have long term plans to open code the char and short operations using lr.[wd] and sc.[wd] but we don't know when this work will be done. Meanwhile, you should avoid char/short atomic operations. Or you could write your own library routines. You could take the linux riscv libatomic and disassemble it to get the code you need. Jim
