https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69873
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
ICC decides to have a cut-off at 8 iterations, having an unrolled vectorized
iteration
..B1.4: # Preds ..B1.4 ..B1.3
addl $8, %eax #4.3
movaps %xmm0, a(,%rdx,8) #5.5
cmpl %ecx, %eax #4.3
movaps %xmm0, 16+a(,%rdx,8) #5.5
movaps %xmm0, 32+a(,%rdx,8) #5.5
movaps %xmm0, 48+a(,%rdx,8) #5.5
movl %eax, %edx #4.3
jb ..B1.4
with a shared scalar epilogue (shared with the non-vectorized path)
..B1.8: # Preds ..B1.8 ..B1.7
incl %edx #4.3
movq %rax, a(,%rcx,8) #5.5
incl %ecx #4.3
cmpl %edi, %edx #4.3
jb ..B1.8
where at least that scalar loop with two IVs looks sub-optimal.
This also shows that unrolling from within the vectorizer might be more
profitable than doing it afterwards (generating an additional vectorized
epilogue for the remaining vectorized iterations). Code-size wise, of course.
