Re: [PATCH] x86: Expand Broadcast to 3 bits

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Thu, Jul 26, 2018 at 8:58 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>> On 26.07.18 at 17:52, <hjl.tools@gmail.com> wrote:
>> On Thu, Jul 26, 2018 at 8:47 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>> On 26.07.18 at 17:02, <hjl.tools@gmail.com> wrote:
>>>> On Thu, Jul 26, 2018 at 7:58 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>>>> On 26.07.18 at 00:05, <hongjiu.lu@intel.com> wrote:
>>>>>> @@ -5008,6 +5010,22 @@ optimize_disp (void)
>>>>>>        }
>>>>>>  }
>>>>>>
>>>>>> +/* Return 1 if there is a match in broadcast bytes between operand
>>>>>> +   GIVEN and instruction template T.   */
>>>>>> +
>>>>>> +static INLINE int
>>>>>> +match_broadcast_size (const insn_template *t, unsigned int given)
>>>>>> +{
>>>>>> +  return ((t->opcode_modifier.broadcast == BYTE_BROADCAST
>>>>>> +        && i.types[given].bitfield.byte)
>>>>>> +       || (t->opcode_modifier.broadcast == WORD_BROADCAST
>>>>>> +           && i.types[given].bitfield.word)
>>>>>> +       || (t->opcode_modifier.broadcast == DWORD_BROADCAST
>>>>>> +           && i.types[given].bitfield.dword)
>>>>>> +       || (t->opcode_modifier.broadcast == QWORD_BROADCAST
>>>>>> +           && i.types[given].bitfield.qword));
>>>>>> +}
>>>>>> +
>>>>>>  /* Check if operands are valid for the instruction.  */
>>>>>>
>>>>>>  static int
>>>>>> @@ -5126,23 +5144,29 @@ check_VecOperands (const insn_template *t)
>>>>>>        i386_operand_type type, overlap;
>>>>>>
>>>>>>        /* Check if specified broadcast is supported in this instruction,
>>>>>> -      and it's applied to memory operand of DWORD or QWORD type.  */
>>>>>> +      and its broadcast bytes match the memory operand.  */
>>>>>>        op = i.broadcast->operand;
>>>>>>        if (!t->opcode_modifier.broadcast
>>>>>>         || !i.types[op].bitfield.mem
>>>>>>         || (!i.types[op].bitfield.unspecified
>>>>>> -           && (t->operand_types[op].bitfield.dword
>>>>>> -               ? !i.types[op].bitfield.dword
>>>>>> -               : !i.types[op].bitfield.qword)))
>>>>>> +           && !match_broadcast_size (t, op)))
>>>>>>       {
>>>>>>       bad_broadcast:
>>>>>>         i.error = unsupported_broadcast;
>>>>>>         return 1;
>>>>>>       }
>>>>>>
>>>>>> +      i.broadcast->bytes = ((1 << (t->opcode_modifier.broadcast - 1))
>>>>>> +                         * i.broadcast->type);
>>>>>
>>>>> So if you moved this up ahead of the earlier if(), and if you used
>>>>> i.broadcast->bytes in place of t->opcode_modifier.broadcast in
>>>>> match_broadcast_size(), I think you could get away without the
>>>>> extension to 3 bits in the templates.
>>>>
>>>> i.broadcast->bytes is set from t->opcode_modifier.broadcast.
>>>> I'd like to avoid check byte, word, dword, qword to compute
>>>> i.broadcast->bytes.
>>>
>>> And this is because of what? This is exactly the kind of redundancy
>>> I'm talking about. Or are there going to be cases where the
>>> broadcast element size is not the smallest among multiple possible
>>> ones for a single template (but then your logic in i386-gen would
>>> be wrong too)?
>>
>> By definition, the broadcast element size is the smalltest non-vector
>> size.
>
> In which case my question stands - what you've said in your earlier
> reply is because of what?

I want to avoid checking byte, word, dword, qword when all I need
is the broadcast element size.


-- 
H.J.