Hi, On Thu, 2026-04-23 at 09:39 +0200, John Paul Adrian Glaubitz wrote: > Once the alignment issue has been sorted out, the next issue to address is > the lack of the > M68k backend in LLVM to handle 32-bit offsets for variables and jumps [4]. > Currently, the > backend will always generate 16-bit offsets which is not enough to compile > all Rust code, > although some Rust code compiles also with 16-bit offsets. > > I have created a rough patch to add support for 32-bit offsets with the help > of Gemini but > the patch has multiple problems as outlined in the GitHub issue in [4]. One > of the problems > is that it creates the 32-bit offsets in an inefficient way as if the target > was a ColdFire > CPU. However, the original 680x0 CPUs can actually use more efficient code to > address 32-bit > variables and jumps. The inefficient way would still be used when ColdFire > support gets added > to LLVM though [5].
I have used Google Gemini Code Assist now [1] in order to create a patch to add support for 32-bit relocations to the LLVM M68k backend. This time, the code makes use of the full capabilities of 68020+ CPUs and does not the inefficient four-instruction method to implement 32-bit accesses. I'm hesitant to send this in for review as this was basically written by Gemini under my supervision with a lot of testing of build tests. Adrian > [1] https://github.com/llvm/llvm-project/issues/181481#issuecomment-4476933700 -- .''`. John Paul Adrian Glaubitz : :' : Debian Developer `. `' Physicist `- GPG: 62FF 8A75 84E0 2956 9546 0006 7426 3B37 F5B5 F913
