Re: [PATCH v1] LoongArch: Use current_stack_pointer in current_pt_regs()

Hi, Tiezhu,

On Thu, Aug 13, 2026 at 5:05 PM Tiezhu Yang <[email protected]> wrote:
>
> The current implementation of current_pt_regs() relies on the compiler
> __builtin_frame_address(0). This introduces an unnecessary dependency
> on the frame pointer register, which forces the compiler to generate
> redundant prologue and epilogue code, create a larger stack frame, and
> perform redundant memory operations to preserve the frame pointer even
> in functions where it is otherwise unnecessary.
>
> Optimize this by switching to current_stack_pointer, which explicitly
> maps to the hardware stack pointer register. This allows the compiler
> to compute the stack alignment directly from the natively maintained
> $sp register, completely eliminating the overhead of preserving and
> restoring the frame pointer on the stack memory.
>
> As a prominent example, this optimization significantly improves the
> hot-path function copy_thread(). A localized disassembly comparison
> of copy_thread() illustrates the elimination of the frame pointer,
> the reduction of stack frame size from 48 bytes down to 32 bytes,
> and a more compact epilogue path:
3 instructions and 16 bytes stack space. There is improvement, but not
"significantly".

Huacai

>
> Before:
> 00000000000004f0 <copy_thread>:
>  4f0:   02ff4063        addi.d          $sp, $sp, -48
>  4f4:   29c08076        st.d            $fp, $sp, 32
>  4f8:   29c06077        st.d            $s0, $sp, 24
>  4fc:   29c0a061        st.d            $ra, $sp, 40
>  500:   02c0c076        addi.d          $fp, $sp, 48
>  ...
>  54c:   1400006e        lu12i.w         $t2, 3
>  ...
>  55c:   03bffdce        ori             $t2, $t2, 0xfff
>  560:   00153ace        or              $t2, $fp, $t2
>  564:   02fb05cd        addi.d          $t1, $t2, -319
>  ...
>  628:   28c0a061        ld.d            $ra, $sp, 40
>  62c:   28c08076        ld.d            $fp, $sp, 32
>  630:   28c06077        ld.d            $s0, $sp, 24
>  634:   00150004        move            $a0, $zero
>  638:   02c0c063        addi.d          $sp, $sp, 48
>  63c:   4c000020        ret
>
> After:
> 00000000000004f0 <copy_thread>:
>  4f0:   02ff8063        addi.d          $sp, $sp, -32
>  4f4:   29c04077        st.d            $s0, $sp, 16
>  4f8:   29c06061        st.d            $ra, $sp, 24
>  [ prologue st.d and addi.d for $fp are completely eliminated ]
>  ...
>  544:   1400006e        lu12i.w         $t2, 3
>  ...
>  554:   03bffdce        ori             $t2, $t2, 0xfff
>  558:   0015386e        or              $t2, $sp, $t2
>  55c:   02fb05cd        addi.d          $t1, $t2, -319
>  ...
>  620:   28c06061        ld.d            $ra, $sp, 24
>  624:   28c04077        ld.d            $s0, $sp, 16
>  628:   00150004        move            $a0, $zero
>  [ epilogue ld.d for $fp is eliminated; exit path is shortened ]
>  62c:   02c08063        addi.d          $sp, $sp, 32
>  630:   4c000020        ret
>
> Signed-off-by: Tiezhu Yang <[email protected]>
> ---
>  arch/loongarch/include/asm/ptrace.h | 6 +-----
>  1 file changed, 1 insertion(+), 5 deletions(-)
>
> diff --git a/arch/loongarch/include/asm/ptrace.h b/arch/loongarch/include/asm/ptrace.h
> index e5d21e836d99..2a7ed442d905 100644
> --- a/arch/loongarch/include/asm/ptrace.h
> +++ b/arch/loongarch/include/asm/ptrace.h
> @@ -170,11 +170,7 @@ static inline void die_if_kernel(const char *str, struct pt_regs *regs)
>                 die(str, regs);
>  }
>
> -#define current_pt_regs()                                              \
> -({                                                                     \
> -       unsigned long sp = (unsigned long)__builtin_frame_address(0);   \
> -       (struct pt_regs *)((sp | (THREAD_SIZE - 1)) + 1) - 1;           \
> -})
> +#define current_pt_regs() ((struct pt_regs *)((current_stack_pointer | (THREAD_SIZE - 1)) + 1) - 1)
>
>  /* Helpers for working with the user stack pointer */
>
> --
> 2.42.0
>