Re: [PATCH] loongarch: Only select HAVE_OBJTOOL and allow ORC unwinder if the inline assembler supports R_LARCH_{32,64}_PCREL

[Jinyang He]

On 2024-06-07 13:42, Xi Ruoyao wrote:
> On Thu, 2024-06-06 at 10:10 +0800, Jinyang He wrote:
> > What I just confused is that there is no ".cfi_*"
> > in the eplogue by clang, which may cause wrong backtrace if gdb set
> >
> > breakpoint there and backtrace. (But this is out of this topic.)
> I don't think it'll cause wrong backtrace.  The real assemble code has
> restored the registers and missing .cfi_restore will just make unwinder
> restore them again.  There are redundant works but not breakages.

$ cat hello.c:
extern void __attribute__((noinline)) foo() {}
int main() {
  foo();
  return (int)(long)__builtin_frame_address(0);
}
$ clang hello.c -S -g -o hello.s -O0 -fPIC
$ cat hello.s:
[...]
        addi.d  $sp, $sp, -32
        .cfi_def_cfa_offset 32
        st.d    $ra, $sp, 24                    # 8-byte Folded Spill
        st.d    $fp, $sp, 16                    # 8-byte Folded Spill
        .cfi_offset 1, -8
        .cfi_offset 22, -16
        addi.d  $fp, $sp, 32
        .cfi_def_cfa 22, 0
        move    $a0, $zero
        st.w    $a0, $fp, -20
.Ltmp2:
        .loc    0 3 3 prologue_end              # hello.c:3:3
        bl      %plt(foo)
        .loc    0 4 21                          # hello.c:4:21
        move    $a0, $fp
        .loc    0 4 3 is_stmt 0                 # hello.c:4:3
        addi.w  $a0, $a0, 0
        ld.d    $fp, $sp, 16                    # 8-byte Folded Reload
<use gdb and set break ponint there>
        ld.d    $ra, $sp, 24                    # 8-byte Folded Reload
        .loc    0 4 3 epilogue_begin            # hello.c:4:3
        addi.d  $sp, $sp, 32
        ret
[...]

So how unwinder do if we <use gdb and set break ponint there>? I think
if we not give ".cfi_restore 22" or others info, it will consider
  1) the $ra is in cfa-8
  2) the cfa is $fp
So it will get $ra by $ra = *(long*)($fp-8). So it may unwind failed
because $fp has been restored and not the CFA now.

> For objtool the main difference seems a thing explained in
> https://maskray.me/blog/2020-11-08-stack-unwinding by Fangrui:
>
>     Note: on RISC-V and LoongArch, the stack slot for the previous frame
>     pointer is stored at fp[-2] instead of fp[0]. See [Consider
>     standardising which stack slot fp points
>     to](https://github.com/riscv-non-isa/riscv-elf-psabi-doc/issues/18)
>     for the RISC-V discussion.

In most cases the $fp is saved at cfa-16. But for va args, something
becomes different at LoongArch (I do not know the case of riscv), the
$fp isn't saved at cfa-16. (e.g. printk?)

> So perhaps we just need to code a constant named "PREV_BP_OFFSET"
Can you give some detail?
>   or
> something in arch/ and use it in update_cfi_state() instead of fully re-
> implement the entire function?

I feel that the update_cfi_state should be arch specific. I believe
that some logic can be reused, but each arch may have its own logic.


Thanks,
Jinyang