jmp .L1_test
.p2align 4
-.complex_dotprod_sse64_loop1:
+.Loop1:
pxor %mm0, %mm0
punpcklwd 0(%rdi), %mm0
addps %xmm0, %xmm4
.L1_test:
dec %rax
- jge .complex_dotprod_sse64_loop1
+ jge .Loop1
# set up for primary loop which is unrolled 4 times
# hence enter loop at top
.p2align 4
-.complex_dotprod_sse64_loop2:
+.Loop2:
mulps (%rsi), %xmm0
addps %xmm2, %xmm6
add $0x40, %rsi
add $0x10, %rdi
dec %rdx
- jne .complex_dotprod_sse64_loop2
+ jne .Loop2
# OK, now we've done with all the multiplies, but
# we still need to handle the unaccumulated