https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97428
--- Comment #6 from Michael_S <already5chosen at yahoo dot com> ---
(In reply to Richard Biener from comment #4)
>
> 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
>
I don't follow. AVX2 indeed can't transpose with 4 shuffles per loop iteration,
but it is fully capable to do it with 8 shuffles per iteration, as you
demonstrated few posts above:
.L8:
vmovupd (%rdi), %ymm1
vmovupd 32(%rdi), %ymm4
vmovupd (%rax), %ymm0
vmovupd 32(%rax), %ymm3
vunpcklpd %ymm4, %ymm1, %ymm2
vunpckhpd %ymm4, %ymm1, %ymm1
vpermpd $216, %ymm1, %ymm1
vmovupd %ymm1, 32(%r8)
vunpcklpd %ymm3, %ymm0, %ymm1
vunpckhpd %ymm3, %ymm0, %ymm0
vpermpd $216, %ymm2, %ymm2
vpermpd $216, %ymm1, %ymm1
vpermpd $216, %ymm0, %ymm0
addq $64, %rdi
vmovupd %ymm2, (%r8)
vmovupd %ymm1, 64(%r8)
vmovupd %ymm0, 96(%r8)
addq $64, %rax
subq $-128, %r8
cmpq %rdi, %rdx
jne .L8
