invalid register name and invalid operand for instruction in arch/powerpc/kernel/vdso{32,64}/gettimeofday.S

[tpimh]

Build of ppc32 and ppc64 kernels with integrated-as enabled failing:

arch/powerpc/kernel/vdso32/gettimeofday.S:34:20: error: invalid register name
  .cfi_register lr,%r12
                   ^
arch/powerpc/kernel/vdso32/gettimeofday.S:71:10: error: invalid operand for instruction
 cmpli 0,%r3,0
         ^
arch/powerpc/kernel/vdso32/gettimeofday.S:72:10: error: invalid operand for instruction
 cmpli 1,%r3,1
         ^
arch/powerpc/kernel/vdso32/gettimeofday.S:77:20: error: invalid register name
  .cfi_register lr,%r12
                   ^
arch/powerpc/kernel/vdso32/gettimeofday.S:107:16: error: invalid operand for instruction
        cmpl 0,%r8,%r0
               ^
arch/powerpc/kernel/vdso32/gettimeofday.S:160:10: error: invalid operand for instruction
 cmpli 0,%r4,0
         ^
arch/powerpc/kernel/vdso32/gettimeofday.S:189:20: error: invalid register name
  .cfi_register lr,%r12
                   ^

Build of ppc64le kernel with integrated-as enabled failing:

arch/powerpc/kernel/vdso64/gettimeofday.S:25:20: error: invalid register name
  .cfi_register lr,%r12
                   ^
arch/powerpc/kernel/vdso64/gettimeofday.S:72:20: error: invalid register name
  .cfi_register lr,%r12
                   ^
arch/powerpc/kernel/vdso64/gettimeofday.S:218:20: error: invalid register name
  .cfi_register lr,%r12
                   ^

[nickdesaulniers]

https://developer.ibm.com/articles/l-ppc/ makes it sound like the %r convention is unconventional.
https://godbolt.org/z/Yme6zt

cc @arndb

for cmpli, looks like Clang wasn't expecting a register operand, I suppose.

[daxtens]

It looks like some of this might be resolved by b2f663073917babb84f5ba9129451e8d564cba1a ("[PowerPC] Allow a '%' prefix for registers in CFI directives"). I'm building llvm from git now to have a look.

[daxtens]

However, it seems a bigger issue might be clang -integrated-as not liking e.g. mulld r3, r3, r4 and only accepting mulld 3, 3, 4; ppc linux asm is exclusively written with the r prefix. Weirdly my reading of llvm/lib/Target/PowerPC/AsmParser/PPCAsmParser.cpp is that it should recognise the r prefix, and should have done since 2013, so maybe I'm misunderstanding how all the pieces fit together...

[daxtens]

@cmriedl points out that this could be due to llvm's assembler not supporting -mregnames: gnu as also doesn't like the r prefix unless you pass that option.

[daxtens]

@mpe pointed out that our register names are actually #defined, so llvm is actually rejecting %r0, %r1 etc, which is even odder. However, I can work around this for now by just changing the defines, for example #define r0 0.

[nickdesaulniers]

Yes, from my godbolt link, it looks like .cfi_directive is fixed, and cmpli is generally broken (regardless of the % or not). So this is two distinct issues.

[nickdesaulniers]

cmpli is #1419 . Let's use this issue to track .cfi_register.

[nemanjai]

The .cfi_register issue was fixed on Nov. 18, 2020 with commit b2f663073917b.

[nickdesaulniers]

Indeed, I can no longer reproduce with:

$ ARCH=powerpc make LLVM=1 -j72 pseries_defconfig arch/powerpc/kernel/vdso32/
# or
$ ARCH=powerpc make LLVM=1 -j72 ppc44x_defconfig arch/powerpc/kernel/vdso32/
# or
$ ARCH=powerpc make LLVM=1 -j72 powernv_defconfig arch/powerpc/kernel/vdso32/

(working around #1484, #1418 and #672).

The .cfi_register issue was fixed on Nov. 18, 2020 with commit b2f663073917b.

lol, I reviewed that patch...forgot about this issue. Looks like this shipped in clang-12. Thanks @gwelymernans !