Re: [Bpf] [PATCH bpf-next v2] bpf, docs: Add callx instructions in new conformance group

[Yonghong Song <yonghong.song@linux.dev>]

On 2/12/24 1:52 PM, dthaler1968@googlemail.com wrote:
>> -----Original Message-----
>> From: Yonghong Song <yonghong.song@linux.dev>
>> Sent: Monday, February 12, 2024 1:49 PM
>> To: Jose E. Marchesi <jose.marchesi@oracle.com>; Dave Thaler
>> <dthaler1968=40googlemail.com@dmarc.ietf.org>
>> Cc: bpf@vger.kernel.org; bpf@ietf.org; Dave Thaler
>> <dthaler1968@gmail.com>
>> Subject: Re: [Bpf] [PATCH bpf-next v2] bpf, docs: Add callx instructions in new
>> conformance group
>>
>>
>> On 2/12/24 1:28 PM, Jose E. Marchesi wrote:
>>>> +BPF_CALL  0x8    0x1  call PC += reg_val(imm)          BPF_JMP | BPF_X
>> only, see `Program-local functions`_
>>> If the instruction requires a register operand, why not using one of
>>> the register fields?  Is there any reason for not doing that?
>> Talked to Alexei and we think using dst_reg for the register for callx insn is
>> better. I will craft a llvm patch for this today. Thanks!
> Why dst_reg instead of src_reg?
> BPF_X is supposed to mean use src_reg.

Let us use dst_reg. Currently, for BPF_K, we have src_reg for a bunch
of flags (pseudo call, kfunc call, etc.). So for BPF_X, let us preserve this
property as well in case in the future we will introduce variants for
callx. The following is the llvm diff:

https://github.com/llvm/llvm-project/pull/81546

>
> But this thread is about reserving/documenting the existing practice,
> since anyone trying to use it would run into interop issues because
> of existing clang.   Should we document both and list one as deprecated?

I think just documenting the new encoding is good enough. But other
people can chime in just in case that I missed something.

>
> Dave
>