On 2020/9/23 22:46, Richard Henderson wrote:
On 9/23/20 3:01 AM, LIU Zhiwei wrote:
On 2020/9/19 2:37, Richard Henderson wrote:
For SVE, we potentially have a 4th argument coming from the
movprfx instruction. Currently we do not optimize movprfx,
so the problem is not visible.
Hi Richard,
I am a little confused. If it is not immediately preceded by a MOVPRFX
instruction, the addend will
still be used.
Is it right?
If movprfx is not used, then the addend register will be the same as the
destination register.
Get it.
Could you see again the definition of HELPER(gvec_udot_idx_h) or the
HELPER(gvec_sdot_idx_h)?
I think it is wrong there, it code sequence is like this:
d0 = a[i+0]
//dot calculation
d[i+0] += d0
Because when addend is the destination register, it has no reason to add
destination register twice.
Best Regards,
Zhiwei
r~
