Re: [PATCH v3 6/9] Support APX NDD

[Jan Beulich <jbeulich@suse.com>]

On 26.03.2024 09:22, Cui, Lili wrote:
>> On 22.03.2024 11:02, Jan Beulich wrote:
>>> On 08.12.2023 15:12, Jan Beulich wrote:
>>>> On 24.11.2023 08:02, Cui, Lili wrote:
>>>>> +rol, 0xd0/0, APX_F,
>>>>>
>> +W|Modrm|No_sSuf|CheckOperandSize|DstVVVV|EVex128|EVexMap4|NF, {
>>>>> +Imm1,
>>>>>
>> +Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseInd
>> ex,
>>>>> +Reg8|Reg16|Reg32|Reg64 }
>>>>>  rol, 0xd0/0, 0, W|Modrm|No_sSuf, { Imm1,
>>>>>
>> Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseIndex
>>  }
>>>>> +rol, 0xc0/0, APX_F,
>>>>>
>> +W|Modrm|No_sSuf|CheckOperandSize|DstVVVV|EVex128|EVexMap4|NF, {
>>>>> +Imm8|Imm8S,
>>>>>
>> +Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseInd
>> ex,
>>>>> +Reg8|Reg16|Reg32|Reg64 }
>>>>>  rol, 0xc0/0, i186, W|Modrm|No_sSuf, { Imm8|Imm8S,
>>>>>
>> Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseIndex
>>  }
>>>>> +rol, 0xd2/0, APX_F,
>>>>>
>> +W|Modrm|No_sSuf|CheckOperandSize|DstVVVV|EVex128|EVexMap4|NF, {
>>>>> +ShiftCount,
>>>>>
>> +Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseInd
>> ex,
>>>>> +Reg8|Reg16|Reg32|Reg64 }
>>>>>  rol, 0xd2/0, 0, W|Modrm|No_sSuf, { ShiftCount,
>>>>>
>> Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseIndex
>>  }
>>>>> +rol, 0xd0/0, APX_F,
>>>>>
>> +W|Modrm|No_sSuf|CheckOperandSize|DstVVVV|EVex128|EVexMap4|NF, {
>>>>>
>> +Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseInd
>> ex,
>>>>> +Reg8|Reg16|Reg32|Reg64 }
>>>>
>>>> Didn't we agree to avoid adding this (and its sibling) template, for
>>>> the omitted shift count being ambiguous? Consider
>>>>
>>>>     rol %cl, %al
>>>>
>>>> Is this a rotate by %cl, or a 1-bit NDD rotate?
>>>
>>> Btw, while this comment was taken into account for the "normal" shifts
>>> and rotates, SHLD / SHRD still have this odd extra form.
>>
>> I have to correct myself here: RCL and RCR had such an odd form retained, too
>> (as, perhaps, a side effect of prematurely adding the non-NDD forms there).
>>
> 
> For RCL/RCR, we dropped the format of omitting $1.
> rcl, 0xd0/2, APX_F, W|Modrm|No_sSuf|CheckOperandSize|DstVVVV|EVexMap4, { Reg8|Reg16|Reg32|Reg64|Unspecified|BaseIndex, Reg8|Reg16|Reg32|Reg64 }

That's my point: This was supposed to be dropped, but is still there. Only
rol/ror and the four shifts are where it was properly dropped. The rcl/rcr
ones disappear in "x86: templatize shift/rotate insns" now.

> For example, when the register is %rcx, it will conflict with the following template.
> 
> rcl, 0xd2/2, 0, W|Modrm|No_sSuf, { ShiftCount, Reg8|Reg16|Reg32|Reg64|Unspecified|BaseIndex }
> 
> I'm confused if we really want to omit the %cl case. I'll confirm it later.

I'm confused by this. What are you talking about? Hmm, perhaps there was
some confusion from me originally saying "had such an odd form retained".
That wasn't specifically about it being %cl or $1 omitted, but more
generally about insn forms with no explicit shift count (of whatever
shape).

Jan