------- Comment #1 from matz at gcc dot gnu dot org 2009-02-25 13:53 -------
For reference intel fortran (11.0) produces three loops, one where it uses
predictive commoning (that is used when there are only few iterations):
..B1.7: # Preds ..B1.6
movsd 8(%r8), %xmm1 #13.52
movsd (%r8), %xmm0 #13.52
# LOE rax rdx rcx rbx rbp rsi rdi r8 r9 r12 r13
r14 r15 xmm0 xmm1
..B1.8: # Preds ..B1.8 ..B1.7
movaps %xmm1, %xmm2 #13.33
movsd 16(%r8,%rdi,8), %xmm3 #13.52
addsd %xmm3, %xmm2 #13.33
addsd %xmm0, %xmm2 #13.41
movaps %xmm1, %xmm0 #14.7
movaps %xmm3, %xmm1 #14.7
addsd 8(%rdx,%rdi,8), %xmm2 #13.9
movsd %xmm2, 8(%rcx,%rdi,8) #13.9
incq %rdi #14.7
cmpq %rax, %rdi #14.7
jl ..B1.8 # Prob 82% #14.7
And two others which are vectorized (plus four/eight times unrolled), but
do _not_ use something like predictive commoning (i.e. no cross iteration
values). Both loops are just versions of each other, one for aligned
destinations and the other for unaligned. The aligned variant is this:
..B1.15: # Preds ..B1.10 ..B1.15
movsd 8(%rdx,%rax,8), %xmm1 #13.18
movhpd 16(%rdx,%rax,8), %xmm1 #13.18
movsd 8(%r8,%rax,8), %xmm0 #13.34
movhpd 16(%r8,%rax,8), %xmm0 #13.34
movsd 24(%rdx,%rax,8), %xmm4 #13.18
movhpd 32(%rdx,%rax,8), %xmm4 #13.18
movsd 24(%r8,%rax,8), %xmm2 #13.34
movhpd 32(%r8,%rax,8), %xmm2 #13.34
movsd 40(%rdx,%rax,8), %xmm7 #13.18
movhpd 48(%rdx,%rax,8), %xmm7 #13.18
movsd 40(%r8,%rax,8), %xmm5 #13.34
movhpd 48(%r8,%rax,8), %xmm5 #13.34
movsd 56(%rdx,%rax,8), %xmm10 #13.18
movhpd 64(%rdx,%rax,8), %xmm10 #13.18
movsd 56(%r8,%rax,8), %xmm8 #13.34
movhpd 64(%r8,%rax,8), %xmm8 #13.34
addpd %xmm0, %xmm1 #13.33
addpd (%r8,%rax,8), %xmm1 #13.41
addpd %xmm2, %xmm4 #13.33
addpd %xmm5, %xmm7 #13.33
addpd %xmm8, %xmm10 #13.33
movaps 16(%r8,%rax,8), %xmm3 #13.52
addpd %xmm3, %xmm1 #13.9
movaps 32(%r8,%rax,8), %xmm6 #13.52
movaps 48(%r8,%rax,8), %xmm9 #13.52
movaps %xmm1, 8(%rcx,%rax,8) #13.9
addpd %xmm3, %xmm4 #13.41
addpd %xmm6, %xmm4 #13.9
movaps %xmm4, 24(%rcx,%rax,8) #13.9
addpd %xmm6, %xmm7 #13.41
addpd %xmm9, %xmm7 #13.9
movaps %xmm7, 40(%rcx,%rax,8) #13.9
addpd %xmm9, %xmm10 #13.41
addpd 64(%r8,%rax,8), %xmm10 #13.9
movaps %xmm10, 56(%rcx,%rax,8) #13.9
addq $8, %rax #14.7
cmpq %r9, %rax #14.7
jl ..B1.15 # Prob 82% #14.7
Not most optimal, due to not using the cross-iteration values to save
two loads per iteration. But still much better than what GCC uses.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39300
