[Bug tree-optimization/97832] AoSoA complex caxpy-like loops: AVX2+FMA -Ofast 7 times slower than -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97832

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
There's then also a permute optimization left on the plate:

t.c:16:3: note:   node 0x3a19590 (max_nunits=4, refcnt=2)
t.c:16:3: note:         stmt 0 _153 = f11_im_76 * x1_im_142;
t.c:16:3: note:         stmt 1 _213 = f11_re_72 * x1_re_202;
t.c:16:3: note:         stmt 2 _275 = f11_re_72 * x1_re_264;
t.c:16:3: note:         stmt 3 _337 = f11_re_72 * x1_re_326;
t.c:16:3: note:         stmt 4 _155 = f11_im_76 * x1_re_140;
t.c:16:3: note:         stmt 5 _217 = f11_im_76 * x1_re_202;
t.c:16:3: note:         stmt 6 _279 = f11_im_76 * x1_re_264;
t.c:16:3: note:         stmt 7 _341 = f11_im_76 * x1_re_326;
t.c:16:3: note:         children 0x3a19600 0x3a19670
t.c:16:3: note:   node (external) 0x3a19600 (max_nunits=1, refcnt=1)
t.c:16:3: note:         { f11_im_76, f11_re_72, f11_re_72, f11_re_72,
f11_im_76, f11_im_76, f11_im_76, f11_im_76 }
t.c:16:3: note:   node 0x3a19670 (max_nunits=4, refcnt=1)
t.c:16:3: note:         stmt 0 x1_im_142 = *_141;
t.c:16:3: note:         stmt 1 x1_re_202 = *_201;
t.c:16:3: note:         stmt 2 x1_re_264 = *_263;
t.c:16:3: note:         stmt 3 x1_re_326 = *_325;
t.c:16:3: note:         stmt 4 x1_re_140 = *_139;
t.c:16:3: note:         stmt 5 x1_re_202 = *_201;
t.c:16:3: note:         stmt 6 x1_re_264 = *_263;
t.c:16:3: note:         stmt 7 x1_re_326 = *_325;
t.c:16:3: note:         load permutation { 4 1 2 3 0 1 2 3 }

which we currently do not handle (there's a FIXME as to permute externals,
currently we only handle splats as transparent for permutes).