On 11/24/23 3:28 AM, Kewen.Lin wrote:
>> + int regoff = INTVAL (operands[2]) * GET_MODE_SIZE (V16QImode);
>
> Is it intentional to keep GET_MODE_SIZE (V16QImode) instead of 16?
> I think if one day NUM_POLY_INT_COEFFS isn't 1 on rs6000 any more,
> we have to add one explicit .to_constant () here. So I prefer this
> to use 16 directly, maybe one comment above to indicate what's for
> the value 16.
I normally don't like hard coding constants in the code, so used
GET_MODE_SIZE (V16QImode) as the number of bytes of a vector register,
but if that's going to cause an issue in the future, I'm fine using 16.
Changed.
>> + int regoff = INTVAL (operands[2]) * GET_MODE_SIZE (V16QImode);
>
> Likewise.
Changed.
>> diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
>> index 5f56c3ed85b..f2efa46c147 100644
>> --- a/gcc/config/rs6000/rs6000.cc
>> +++ b/gcc/config/rs6000/rs6000.cc
>> @@ -1964,9 +1964,12 @@ rs6000_hard_regno_mode_ok (unsigned int regno,
>> machine_mode mode)
>> static bool
>> rs6000_modes_tieable_p (machine_mode mode1, machine_mode mode2)
>> {
>> - if (mode1 == PTImode || mode1 == OOmode || mode1 == XOmode
>> - || mode2 == PTImode || mode2 == OOmode || mode2 == XOmode)
>> - return mode1 == mode2;
>> + if (mode1 == PTImode || mode1 == OOmode || mode1 == XOmode
>> + || mode2 == PTImode || mode2 == XOmode)
>> + return mode1 == mode2;
>> +
>> + if (mode2 == OOmode)
>> + return ALTIVEC_OR_VSX_VECTOR_MODE (mode1);
>
> I vaguely remembered that Segher mentioned it's unexpected for opaque
> modes to have tieable modes excepting for themselves, but if this is the
> only way to get rid of those extra moves, I guess we can special-case
> them here. Looking forward to Segher's comments on this part.
To be honest, my original patch didn't have this. I think it was Mike who
said we want or need this. Mike, why do we want/need this again?
That said, the documentation for TARGET_MODES_TIEABLE_P says:
This hook returns true if a value of mode mode1 is accessible in
mode mode2 without copying.
Given OOmode (ie, __vector_pair) under the covers is two consecutive
vector registers, and we use them/initialize them with two vectors,
then mode1 being a vector mode could be accesible from an OOmode mode2
without copying, meaning we could access it directly from the registers
holding mode2.
Segher, your input to the above an the subreg portion of the patch in general?
Peter
