On 2026/2/20 01:50, Alexei Starovoitov wrote:

On Thu, Feb 19, 2026 at 6:30 AM Leon Hwang [off-list ref] wrote:

quoted

+static bool bpf_kfunc_is_fastcall(struct bpf_verifier_env *env, u32 func_id, u32 flags)
+{
+       if (!(flags & KF_FASTCALL))
+               return false;
+
+       if (!env->prog->jit_requested)
+               return true;
+
+       if (func_id == special_kfunc_list[KF_bpf_clz64])
+               return bpf_jit_inlines_kfunc_call(bpf_clz64);
+       if (func_id == special_kfunc_list[KF_bpf_ctz64])
+               return bpf_jit_inlines_kfunc_call(bpf_ctz64);
+       if (func_id == special_kfunc_list[KF_bpf_ffs64])
+               return bpf_jit_inlines_kfunc_call(bpf_ffs64);
+       if (func_id == special_kfunc_list[KF_bpf_fls64])
+               return bpf_jit_inlines_kfunc_call(bpf_fls64);
+       if (func_id == special_kfunc_list[KF_bpf_bitrev64])
+               return bpf_jit_inlines_kfunc_call(bpf_bitrev64);
+       if (func_id == special_kfunc_list[KF_bpf_popcnt64])
+               return bpf_jit_inlines_kfunc_call(bpf_popcnt64);
+       if (func_id == special_kfunc_list[KF_bpf_rol64])
+               return bpf_jit_inlines_kfunc_call(bpf_rol64);
+       if (func_id == special_kfunc_list[KF_bpf_ror64])
+               return bpf_jit_inlines_kfunc_call(bpf_ror64);

This is too ugly. Find a way to do it differently.

Agreed.

I'd like to introduce a new flag KF_JIT_MAY_INLINE to indicate the kfunc
will be inlined by JIT backends if possible. As for those kfuncs w/
KF_FASTCALL w/o KF_JIT_MAY_INLINE, they are fastcall always.

Thanks,
Leon