[PATCH net-next 2/3] arm64: bpf: optimize JMP_CALL

On Mon, Jun 06, 2016 at 09:36:03PM -0700, Z Lim wrote:

On Mon, Jun 6, 2016 at 10:05 AM, Will Deacon [off-list ref] wrote:

quoted

On Sat, Jun 04, 2016 at 03:00:29PM -0700, Zi Shen Lim wrote:

quoted

Remove superfluous stack frame, saving us 3 instructions for
every JMP_CALL.

Signed-off-by: Zi Shen Lim <redacted>
---
 arch/arm64/net/bpf_jit_comp.c | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/arch/arm64/net/bpf_jit_comp.c b/arch/arm64/net/bpf_jit_comp.c
index 51abc97..7ae304e 100644
--- a/arch/arm64/net/bpf_jit_comp.c
+++ b/arch/arm64/net/bpf_jit_comp.c

@@ -578,11 +578,8 @@ emit_cond_jmp:
              const u64 func = (u64)__bpf_call_base + imm;

              emit_a64_mov_i64(tmp, func, ctx);
-             emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
-             emit(A64_MOV(1, A64_FP, A64_SP), ctx);
              emit(A64_BLR(tmp), ctx);
              emit(A64_MOV(1, r0, A64_R(0)), ctx);
-             emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
              break;
      }

Is the jitted code intended to be unwindable by standard tools?

Before this patch:
    bpf_prologue => push stack frame
    ...
    jmp_call => push stack frame, call bpf_helper*, pop stack frame
    ...
    bpf_epilogue => pop stack frame, ret

Now:
    bpf_prologue => push stack frame
    ...
    jmp_call => call bpf_helper*
    ...
    bpf_epilogue => pop stack frame, ret

*Note: bpf_helpers in kernel/bpf/helper.c

So yes, it's still unwindable.

Sure, I'm not disputing that. I just wondered whether or not it needs to
be unwindable at all...

Will