Re: Register constraint in NEG instructions

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

On 7/26/23 2:16 AM, Jose E. Marchesi wrote:
> 
> I see this in the verifier (bpf-next):
> 
>     if (opcode == BPF_NEG) {
> 	if (BPF_SRC(insn->code) != BPF_K ||
> 	    insn->src_reg != BPF_REG_0 ||
> 	    insn->off != 0 || insn->imm != 0) {
> 		verbose(env, "BPF_NEG uses reserved fields\n");
> 	return -EINVAL;
>     }
> 
> And along this llvm assembler test:
> 
>                 |
>                 v
>    // CHECK: 84 01 00 00 00 00 00 00	w1 = -w1
>    w1 = -w1
> 
> Is enough evidence that NEG is supposed to use only dst and not src.  I

Yes, in llvm (BPFInstrInfo.td),

class NEG_RR<BPFOpClass Class, BPFArithOp Opc,
              dag outs, dag ins, string asmstr, list<dag> pattern>
     : TYPE_ALU_JMP<Opc.Value, 0, outs, ins, asmstr, pattern> {
   bits<4> dst;

   let Inst{51-48} = dst;
   let BPFClass = Class;
}

You can see only dst register is used. The further evidence
is from the above kernel check.


> am sending a fix for standarization/instruction-set.rst.
> 
>> Hello.
>>
>> The neg (and neg32) instructions are documented to use (and encode) both
>> src and dst register operands in standarization/instruction-set.rst:
>>
>>    BPF_NEG   0x80   dst = -src
>>
>> However, in llvm's BPFAsmParser::PreMatchCheck, it is checked that both
>> source and destination registers refer to the same register.  If they
>> are not, an error is raised.
>>
>> Is this to speed up JIT to different architectures, some like x86
>> featuring `NEG reg' and others like aarch64 featuring `NEG reg1,reg2'?
>>
>> Should I send a patch for instruction-set.rst documenting the
>> requirement?
>>
>> Thanks.