On Fri, Feb 18, 2022 at 04:34:28PM -0800, Eric Biggers wrote:
> > +.macro schoolbook1_iteration i xor_sum
> > + .set i, \i
> > + .set xor_sum, \xor_sum
> > + movups (16*i)(OP1), %xmm0
> > + .if(i == 0 && xor_sum == 1)
> > + pxor SUM, %xmm0
> > + .endif
> > + vpclmulqdq $0x01, (16*i)(OP2), %xmm0, %xmm1
> > + vpxor %xmm1, MI, MI
> > + vpclmulqdq $0x00, (16*i)(OP2), %xmm0, %xmm2
> > + vpxor %xmm2, LO, LO
> > + vpclmulqdq $0x11, (16*i)(OP2), %xmm0, %xmm3
> > + vpxor %xmm3, HI, HI
> > + vpclmulqdq $0x10, (16*i)(OP2), %xmm0, %xmm4
> > + vpxor %xmm4, MI, MI
>
> Perhaps the above multiplications and XORs should be reordered slightly so that
> each XOR doesn't depend on the previous instruction? A good ordering might be:
>
> vpclmulqdq $0x01, (16*\i)(OP2), %xmm0, %xmm1
> vpclmulqdq $0x10, (16*\i)(OP2), %xmm0, %xmm2
> vpclmulqdq $0x00, (16*\i)(OP2), %xmm0, %xmm3
> vpclmulqdq $0x11, (16*\i)(OP2), %xmm0, %xmm4
> vpxor %xmm1, MI, MI
> vpxor %xmm3, LO, LO
> vpxor %xmm4, HI, HI
> vpxor %xmm2, MI, MI
>
> With that, no instruction would depend on either of the previous two
> instructions.
>
> This might be more important in the ARM64 version than the x86_64 version, as
> x86_64 CPUs are pretty aggressive about internally reordering instructions. But
> it's something to consider in both versions.
>
> Likewise in schoolbook1_noload.
Or slightly better:
vpclmulqdq $0x01, (16*\i)(OP2), %xmm0, %xmm2
vpclmulqdq $0x00, (16*\i)(OP2), %xmm0, %xmm1
vpclmulqdq $0x10, (16*\i)(OP2), %xmm0, %xmm3
vpclmulqdq $0x11, (16*\i)(OP2), %xmm0, %xmm4
vpxor %xmm2, MI, MI
vpxor %xmm1, LO, LO
vpxor %xmm4, HI, HI
vpxor %xmm3, MI, MI
- Eric
