Hi!

On Tue, Feb 03, 2026 at 08:30:41AM +0100, Christophe Leroy (CS GROUP) wrote:
> Test robot reports the following error with clang-16.0.6:
> 
>    In file included from kernel/rseq.c:75:
>    include/linux/rseq_entry.h:141:3: error: invalid operand for instruction
>                    unsafe_get_user(offset, &ucs->post_commit_offset, efault);
>                    ^
>    include/linux/uaccess.h:608:2: note: expanded from macro 'unsafe_get_user'
>            arch_unsafe_get_user(x, ptr, local_label);      \
>            ^
>    arch/powerpc/include/asm/uaccess.h:518:2: note: expanded from macro 
> 'arch_unsafe_get_user'
>            __get_user_size_goto(__gu_val, __gu_addr, sizeof(*(p)), e); \
>            ^
>    arch/powerpc/include/asm/uaccess.h:284:2: note: expanded from macro 
> '__get_user_size_goto'
>            __get_user_size_allowed(x, ptr, size, __gus_retval);    \
>            ^
>    arch/powerpc/include/asm/uaccess.h:275:10: note: expanded from macro 
> '__get_user_size_allowed'
>            case 8: __get_user_asm2(x, (u64 __user *)ptr, retval);  break;  \
>                    ^
>    arch/powerpc/include/asm/uaccess.h:258:4: note: expanded from macro 
> '__get_user_asm2'
>                    "       li %1+1,0\n"                    \
>                     ^
>    <inline asm>:7:5: note: instantiated into assembly here
>            li 31+1,0
>               ^
>    1 error generated.
> 
> On PPC32, for 64 bits vars a pair of registers is used. Usually the
> lower register in the pair is the high part and the higher register is
> the low part. GCC uses r3/r4 ... r11/r12 ... r14/r15 ... r30/r31
> 
> In older kernel code inline assembly was using %1 and %1+1 to represent
> 64 bits values. However here it looks like clang uses r31 as high part,
> allthough r32 doesn't exist hence the error.
> 
> Allthoug %1+1 should work, most places now use %L1 instead of %1+1, so
> let's do the same here.
> 
> With that change, the build doesn't fail anymore and a disassembly shows
> clang uses r17/r18 and r31/r14 pair when GCC would have used r16/r17 and
> r30/r31:

This does not fix the problem that somehow LLVM thinks that GPR31/FPR0
is a valid pair for two-register integer things (well, 31+1 in
assembler is not actually valid at all).  Quite worrying.

Maybe you can fix this in a more fundamental way?  In LLVM itself?

(The kernel patch of course is a nice workaround, if it in fact works
reliably, but a big fat comment here would be useful.  Pointing to the
LLVM problem report where this is tracked, etc.)


Segher

Reply via email to