[PATCH] riscv: eliminate unreliable __builtin_frame_address(1)

[Jessica Clarke]

On 17 Jan 2022, at 15:44, Changbin Du <changbin.du at gmail.com> wrote:
> 
> I tried different pieces of code which uses __builtin_frame_address(1)
> (with both gcc version 7.5.0 and 10.3.0) to verify whether it works as
> expected on riscv64. The result is negative.
> 
> What the compiler had generated is as below:
> 31                      fp = (unsigned long)__builtin_frame_address(1);
>   0xffffffff80006024 <+200>:   ld      s1,0(s0)
> 
> It takes '0(s0)' as the address of frame 1 (caller), but the actual address
> should be '-16(s0)'.
> 
>          |       ...       | <-+
>          +-----------------+   |
>          | return address  |   |
>          | previous fp     |   |
>          | saved registers |   |
>          | local variables |   |
>  $fp --> |       ...       |   |
>          +-----------------+   |
>          | return address  |   |
>          | previous fp --------+
>          | saved registers |
>  $sp --> | local variables |
>          +-----------------+
> 
> This leads the kernel can not dump the full stack trace on riscv.
> 
> [    7.222126][    T1] Call Trace:
> [    7.222804][    T1] [<ffffffff80006058>] dump_backtrace+0x2c/0x3a
> 
> This problem is not exposed on most riscv builds just because the '0(s0)'
> occasionally is the address frame 2 (caller's caller), if only ra and fp
> are stored in frame 1 (caller).
> 
>          |       ...       | <-+
>          +-----------------+   |
>          | return address  |   |
>  $fp --> | previous fp     |   |
>          +-----------------+   |
>          | return address  |   |
>          | previous fp --------+
>          | saved registers |
>  $sp --> | local variables |
>          +-----------------+
> 
> This could be a *bug* of gcc that should be fixed. But as noted in gcc
> manual "Calling this function with a nonzero argument can have
> unpredictable effects, including crashing the calling program.", let's
> remove the '__builtin_frame_address(1)' in backtrace code.

Yes, this is a bug, that is always wrong. LLVM gets this right.

https://godbolt.org/z/MrhsoPPM6

Jess