Re: clang/objtool failures on linux-next

[Nathan Chancellor]

On Fri, Mar 06, 2026 at 05:35:20PM -0800, Josh Poimboeuf wrote:
> On Fri, Mar 06, 2026 at 11:04:50PM +0100, Arnd Bergmann wrote:
> > ==> build/x86/0x8FD83285_defconfig/log <==
> > drivers/gpu/drm/amd/amdgpu/amdgpu_cs.o: warning: objtool: amdgpu_cs_parser_bos+0x723: unreachable instruction
> > drivers/gpu/drm/xe/xe_bo.o: warning: objtool: xe_bo_create_pin_map_at_novm+0x11d: unreachable instruction
> > drivers/gpu/drm/panthor/panthor_drv.o: warning: objtool: panthor_ioctl_vm_bind+0x2df: unreachable instruction
> 
> These three warnings are accurate, for each of these, Clang is leaving
> an unreachable JMP at the end of the function:
> 
>     166e:	4c 89 ea             	mov    %r13,%rdx
>     1671:	e8 00 00 00 00       	call   1676 <amdgpu_cs_parser_bos+0x716>	1672: R_X86_64_PLT32	__ubsan_handle_shift_out_of_bounds-0x4
>     1676:	49 8b 94 24 08 1d 01 00 	mov    0x11d08(%r12),%rdx
>     167e:	e9 c6 fd ff ff       	jmp    1449 <amdgpu_cs_parser_bos+0x4e9>
>     1683:	e9 8c f9 ff ff       	jmp    1014 <amdgpu_cs_parser_bos+0xb4>
>     1688:	0f 1f 84 00 00 00 00 00 	nopl   0x0(%rax,%rax,1)
> 
> Nathan, any idea?  I haven't seen this pattern before.

This seems pretty similar to

  https://github.com/ClangBuiltLinux/linux/issues/2130
  https://github.com/llvm/llvm-project/issues/167774

as they all use drm_exec_until_all_locked(). Justin fixed this in main
(23.0.0) and it was backported to 22.1.0:

  https://github.com/llvm/llvm-project/commit/404920838836a022b0ae4b9a462befba89ee0cdc
  https://github.com/llvm/llvm-project/commit/2fb8921a7c70b264f670569ac319d1c96e9f5acc

I don't think Arnd has that fix based on the .comment section:

  Debian clang version 22.0.0 (++20251025104945+e219cf60598c-1~exp1~20251025225011.1789)

I don't see a configuration file for these so I cannot verify that they
are actually fixed with a newer version of clang.

Unfortunately, the problematic change landed in the 21.1.0 release so
that version will still see these warnings. I am guessing it is probably
not too easy to workaround this in objtool though, so maybe we will just
have to live with them?

Cheers,
Nathan