[PATCH] rs6000: Use subreg for QI/HI vector init

[Kewen.Lin]

Hi Segher,

on 2020/12/15 下午10:40, Segher Boessenkool wrote:
> Hi Ke Wen,
> 
> On Tue, Dec 15, 2020 at 03:53:29PM +0800, Kewen.Lin wrote:
>> on 2020/12/15 上午2:51, Segher Boessenkool wrote:
>>> On Wed, Dec 02, 2020 at 05:44:24PM +0800, Kewen.Lin wrote:
>>>> --- a/gcc/config/rs6000/rs6000.c
>>>> +++ b/gcc/config/rs6000/rs6000.c
>>>> @@ -6793,17 +6793,8 @@ rs6000_expand_vector_init (rtx target, rtx vals)
>>>>        /* Force the values into word_mode registers.  */
>>>>        for (i = 0; i < n_elts; i++)
>>>>  	{
>>>> -	  rtx tmp = force_reg (GET_MODE_INNER (mode), XVECEXP (vals, 0, i));
>>>> -	  if (TARGET_POWERPC64)
>>>> -	    {
>>>> -	      op[i] = gen_reg_rtx (DImode);
>>>> -	      emit_insn (gen_zero_extendqidi2 (op[i], tmp));
>>>> -	    }
>>>> -	  else
>>>> -	    {
>>>> -	      op[i] = gen_reg_rtx (SImode);
>>>> -	      emit_insn (gen_zero_extendqisi2 (op[i], tmp));
>>>> -	    }
>>>> +	  rtx tmp = force_reg (inner_mode, XVECEXP (vals, 0, i));
>>>> +	  op[i] = simplify_gen_subreg (Pmode, tmp, inner_mode, 0);
>>>>  	}
>>>
>>> Pmode is defined based on TARGET_64BIT, not TARGET_POWERPC64.
>>
>> Good point, you are right, is it ok to change this part with one
>> explicit mode based on TARGET_POWERPC64?
>>
>>           rtx tmp = force_reg (inner_mode, XVECEXP (vals, 0, i));
>>           machine_mode tmode = TARGET_POWERPC64 ? DImode : SImode;
>>           op[i] = simplify_gen_subreg (tmode, tmp, inner_mode, 0);
> 
> That looks fine, yes.
> 
>>> But, can you not always use SImode here?
>>
>> Sorry that I didn't quite follow here.
> 
> I mean do
> 
>           rtx tmp = force_reg (inner_mode, XVECEXP (vals, 0, i));
>           op[i] = simplify_gen_subreg (SImode, tmp, inner_mode, 0);
> 
> If that works (also in 64-bit mode), that is preferred.  It might need
> some more adjustment elsewhere, not sure if that is worth it.
> 

Thanks for the clarification!  Yes, as you said it doesn't work without
any adjustments, the gen_vsx_concat_v2di requires DImode operands, I
think it's not worthy to expand it just for this case.

> It is okay for trunk with either of those changes.  Thanks!
> 

Thanks!  Committed with the former way via r11-6109.

BR,
Kewen