On Thu, Aug 11, 2022 at 11:10:34AM +0700, John Naylor wrote:
>> I wonder if reusing a zero vector (instead of creating a new one every
>> time) has any noticeable effect on performance.
>
> Creating a zeroed register is just FOO PXOR FOO, which should get
> hoisted out of the (unrolled in this case) loop, and which a recent
> CPU will just map to a hard-coded zero in the register file, in which
> case the execution latency is 0 cycles. :-)
Ah, indeed. At -O2, my compiler seems to zero out two registers before the
loop with either approach:
pxor %xmm0, %xmm0 ; accumulator
pxor %xmm2, %xmm2 ; always zeros
And within the loop, I see the following:
movdqu (%rdi), %xmm1
movdqu (%rdi), %xmm3
addq $16, %rdi
pcmpeqb %xmm2, %xmm1 ; check for zeros
por %xmm3, %xmm0 ; OR data into accumulator
por %xmm1, %xmm0 ; OR zero check results into accumulator
cmpq %rdi, %rsi
So the call to _mm_setzero_si128() within the loop is fine. Apologies for
the noise.
--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com
