Re: Encoding of V4 32-bit JA

["Jose E. Marchesi" <jose.marchesi@oracle.com>]

>> On Sun, 2023-07-23 at 21:14 +0200, Jose E. Marchesi wrote:
>>> > On Fri, 2023-07-21 at 18:19 +0200, Jose E. Marchesi wrote:
>>> > > Hi Yonghong.
>>> > > 
>>> > > This is from the v4 instructions proposal:
>>> > > 
>>> > >     ========  =====  =========================  ============
>>> > >     code      value  description                notes
>>> > >     ========  =====  =========================  ============
>>> > >     BPF_JA    0x00   PC += imm                  BPF_JMP32 only
>>> > > 
>>> > > Is this instruction using source 1 instead of 0?  Otherwise, it would
>>> > > have exactly the same encoding than the V3< JA instruction.  Is that
>>> > > what is intended?
>>> > > 
>>> > > TIA.
>>> > > 
>>> > 
>>> > Hi Jose,
>>> > 
>>> > I think that assumption is that `BPF_JMP32 | BPF_JA` is currently free:
>>> > - documentation [1] implies that only `BPF_JMP` should be used for `BPF_JA`
>>> >   (see "notes" column for the first line)
>>> > - BPF verifier rejects `BPF_JMP32 | BPF_JA`
>>> > - clang always generates `BPF_JMP | BPF_JA`
>>> 
>>> Makes sense, thanks for the info.
>>> 
>>> Do you know the precise pseudo-c assembly syntax to use for this
>>> instruction?
>>
>> In [1] Yonghong uses the following form:
>>
>>   gotol +0xcd9b
>>
>> But it seems to be not specified in the documentation for the patch-set v3.
>
> I will use that syntax in binutils for now.

https://sourceware.org/pipermail/binutils/2023-July/128543.html

Now we will make GCC to use that instruction whenever needed, but only
with -mcpu=v4 or later:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110781

>> [1] https://reviews.llvm.org/D144829
>>
>>> 
>>> > Thanks,
>>> > Eduard
>>> > 
>>> > [1] https://www.kernel.org/doc/html/latest/bpf/instruction-set.html#jump-instructions