https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97428
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
I have a fix that, with -mavx512f generates just
.L3:
vmovupd (%rcx,%rax), %zmm0
vpermpd (%rsi,%rax), %zmm1, %zmm2
vpermpd %zmm0, %zmm1, %zmm0
vmovupd %zmm2, (%rdi,%rax,2)
vmovupd %zmm0, 64(%rdi,%rax,2)
addq $64, %rax
cmpq %rax, %rdx
jne .L3
while the lack of cross-lane shuffles in AVX2 requires a
.L3:
vmovupd (%rsi,%rax), %xmm5
vmovupd 32(%rsi,%rax), %xmm6
vinsertf128 $0x1, 16(%rsi,%rax), %ymm5, %ymm1
vinsertf128 $0x1, 48(%rsi,%rax), %ymm6, %ymm3
vmovupd (%rcx,%rax), %xmm7
vmovupd 32(%rcx,%rax), %xmm5
vinsertf128 $0x1, 16(%rcx,%rax), %ymm7, %ymm0
vinsertf128 $0x1, 48(%rcx,%rax), %ymm5, %ymm2
vunpcklpd %ymm3, %ymm1, %ymm4
vunpckhpd %ymm3, %ymm1, %ymm1
vpermpd $216, %ymm4, %ymm4
vpermpd $216, %ymm1, %ymm1
vmovupd %xmm4, (%rdi,%rax,2)
vextractf128 $0x1, %ymm4, 16(%rdi,%rax,2)
vmovupd %xmm1, 32(%rdi,%rax,2)
vextractf128 $0x1, %ymm1, 48(%rdi,%rax,2)
vunpcklpd %ymm2, %ymm0, %ymm1
vunpckhpd %ymm2, %ymm0, %ymm0
vpermpd $216, %ymm1, %ymm1
vpermpd $216, %ymm0, %ymm0
vmovupd %xmm1, 64(%rdi,%rax,2)
vextractf128 $0x1, %ymm1, 80(%rdi,%rax,2)
vextractf128 $0x1, %ymm0, 112(%rdi,%rax,2)
vmovupd %xmm0, 96(%rdi,%rax,2)
addq $64, %rax
cmpq %rax, %rdx
jne .L3
not much better than what SSE provides:
.L3:
movupd (%rsi), %xmm3
movupd 16(%rsi), %xmm7
movq %rax, %rdx
subq $-128, %rdi
movupd 32(%rsi), %xmm2
movupd 48(%rsi), %xmm6
addq $64, %rax
addq $64, %rsi
movapd %xmm3, %xmm8
movupd -64(%rax), %xmm1
movupd -48(%rax), %xmm5
unpckhpd %xmm7, %xmm3
unpcklpd %xmm7, %xmm8
movupd -32(%rax), %xmm0
movupd -16(%rax), %xmm4
movups %xmm3, -96(%rdi)
movups %xmm8, -128(%rdi)
movapd %xmm2, %xmm8
unpckhpd %xmm6, %xmm2
movups %xmm2, -80(%rdi)
movapd %xmm1, %xmm2
unpcklpd %xmm6, %xmm8
unpckhpd %xmm5, %xmm1
unpcklpd %xmm5, %xmm2
movups %xmm8, -112(%rdi)
movups %xmm2, -64(%rdi)
movapd %xmm0, %xmm2
unpckhpd %xmm4, %xmm0
unpcklpd %xmm4, %xmm2
movups %xmm1, -32(%rdi)
movups %xmm2, -48(%rdi)
movups %xmm0, -16(%rdi)
cmpq %rdx, %rcx
jne .L3
