res_ptr // acc = poly MOVD $-1, acc0 MOVD p256const0<>(SB), acc1 MOVD $0, acc2 MOVD p256const1<>(SB), acc3 // Load the original value LDP 0*16(a_ptr), (t0, t1) LDP 1*16(a_ptr), (t2, t3) // Speculatively subtract SUBS t0, acc0 SBCS t1, acc1 SBCS t2, acc2 SBC t3, acc3 // If condition is 0, keep original value CMP $0, hlp0 CSEL EQ, t0, acc0, acc0 CSEL EQ, t1, acc1, acc1 CSEL EQ, t2, acc2, acc2 CSEL EQ, t3, acc3, acc3 // Store result STP (acc0, acc1), 0*16(res_ptr) STP (acc2, acc3), 1*16(res_ptr) RET /* ----------...