Clang/LLVM 15 and above fail to boot on Exynos 990/9830

[ExtremeXT]

Hi, I have tried updating the compiler on my upstreamed Exynos 990/9830 (codename) kernel: https://github.com/ExtremeXT/990_upstream_v2 but have only managed to achieve Clang 14, anything newer (15, 16) fails to boot. Attached is last_kmsg file which is the kernel log. I assume that this is the issue:

<0>[    0.176810]  [0:      swapper/0:    1] undefined instruction: pc=0xffffff8008185a48 (esr=0x02000000)
<6>[    0.176831]  [0:      swapper/0:    1] Code: d4210000 910003e8 91002108 f90007e8 (00000000) 
<0>[    0.176843]  [0:      swapper/0:    1] Internal error: undefined instruction: 2000000 [#1] PREEMPT SMP

But i haven't been able to find what leads to that address, I think there is some kind of relocation going on? It always points to something different on each build.

I hope this can be solved, I have looked over the issues but couldn't find anything similar. Thanks!

[ExtremeXT]

I managed to disable KASLR and fix the address

Now the new one is

<0>[    0.176415]  [0:      swapper/0:    1] undefined instruction: pc=0xffffff80080e5a48 (esr=0x02000000)
<6>[    0.176437]  [0:      swapper/0:    1] Code: d4210000 910003e8 91002108 f90007e8 (00000000) 
<0>[    0.176449]  [0:      swapper/0:    1] Internal error: undefined instruction: 2000000 [#1] PREEMPT SMP

LLVM objdump and addr2line outputs:

❯ llvm-objdump --disassemble --start-address=0xffffff80080e5a30 --stop-address=0xffffff80080e5a60 vmlinux

vmlinux:        file format elf64-littleaarch64

Disassembly of section .text:

ffffff80080e59e4 <$x.125>:
ffffff80080e5a30: 2a b1 1b 39   strb    w10, [x9, #1772]
ffffff80080e5a34: ff 77 ff 97   bl      0xffffff80080c3a30 <__warn_printk>

ffffff80080e5a38 <$x.127>:
ffffff80080e5a38: 00 00 21 d4   brk     #0x800
ffffff80080e5a3c: e8 03 00 91   mov     x8, sp
ffffff80080e5a40: 08 21 00 91   add     x8, x8, #8
ffffff80080e5a44: e8 07 00 f9   str     x8, [sp, #8]
                ...

ffffff80080e5a50 <cancel_work_sync>:
ffffff80080e5a50: fd 7b bf a9   stp     x29, x30, [sp, #-16]!
ffffff80080e5a54: fd 03 00 91   mov     x29, sp
ffffff80080e5a58: e1 03 1f 2a   mov     w1, wzr
ffffff80080e5a5c: 05 00 00 94   bl      0xffffff80080e5a70 <__cancel_work_timer.llvm.17801297389905514748>

❯ llvm-addr2line -e vmlinux 0xffffff80080e5a48
??:0

Attached is last_kmsg and kernel files

[ExtremeXT]

For some reason I can't attach files

https://pastebin.com/snewHGex
https://gofile.io/d/y1hCcU

[nathanchance]

I am not sure I see anything immediately incorrect from the call chain in the logs. It would probably be helpful to see if vanilla LLVM reproduces the same issue, I have optimized prebuilt toolchains on kernel.org. If they do, it should be possible to bisect LLVM to see what commit introduces the boot failure, assuming you have decent hardware to build the toolchain. Otherwise, I would see if it is possible to test newer versions of AOSP LLVM to see if this is something that is resolved closer to tip of tree, like LLVM 19 or 20.