https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66721
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So I have a "solution" that I hope to prettify a bit still. The basic issue
is that SLP is "broken" in that it builds a tree of operations instead of a
graph.
That is, it un-CSEs all n in the testcase:
for (i = 0; i < size; ++i)
{
float n = sqrt (in[i].x * in[i].x + in[i].y * in[i].y + in[i].z *
in[i].z);
out[i].x = in[i].x / n;
out[i].y = in[i].y / n;
out[i].z = in[i].z / n;
}
which results in bloat in the SLP tree and initial vectorized code and also
cost which is good in this case. Usually with nicely vector-size aligned
things later passes can CSE all the mess again and the generated code isn't
so bad (so the costs are overly conservative). But in this case with an
"interleaving size" of 3 we end up with three vectors containing 4 groups
we'd want to CSE.
sqrt { in[0].x * in[0].x + in[0].y * in[0].y + ...,
in[0].x * ....
in[0].x * ...,
in[1].x * in[1].x + in[1].y * ... }
sqrt { ... }
sqrt { ... }
this way we avoid the permutes on the stores of course, but the lack of
CSE here makes the generated code worse (when you look at it with
-fno-vect-cost-model, the cost-model rejects the code even though it
might be profitable at a slightly larger theshold in practice).
.L5:
movups (%r8), %xmm8
addl $1, %r10d
addq $48, %r8
addq $48, %rax
movups -16(%r8), %xmm6
movups -32(%r8), %xmm7
movaps %xmm8, %xmm13
movdqa %xmm6, %xmm11
movdqa %xmm6, %xmm10
movaps %xmm6, %xmm1
movdqa %xmm7, %xmm9
movdqa %xmm7, %xmm0
movaps %xmm7, %xmm12
shufps $252, %xmm6, %xmm1
palignr $8, %xmm7, %xmm11
shufps $175, %xmm7, %xmm13
mulps %xmm1, %xmm1
movaps %xmm7, %xmm2
palignr $12, %xmm7, %xmm10
shufps $252, %xmm11, %xmm11
mulps %xmm11, %xmm11
shufps $240, %xmm7, %xmm12
shufps $252, %xmm10, %xmm10
mulps %xmm10, %xmm10
palignr $4, %xmm8, %xmm9
shufps $5, %xmm6, %xmm2
shufps $192, %xmm9, %xmm9
mulps %xmm9, %xmm9
palignr $8, %xmm8, %xmm0
shufps $192, %xmm0, %xmm0
mulps %xmm0, %xmm0
mulps %xmm13, %xmm13
addps %xmm10, %xmm11
mulps %xmm12, %xmm12
movaps %xmm8, %xmm10
shufps $192, %xmm8, %xmm10
mulps %xmm10, %xmm10
mulps %xmm2, %xmm2
addps %xmm11, %xmm1
addps %xmm13, %xmm12
addps %xmm10, %xmm9
movaps %xmm3, %xmm10
addps %xmm12, %xmm2
addps %xmm9, %xmm0
rsqrtps %xmm0, %xmm9
cmpneqps %xmm0, %xmm10
andps %xmm10, %xmm9
mulps %xmm9, %xmm0
movaps %xmm3, %xmm10
cmpneqps %xmm2, %xmm10
mulps %xmm0, %xmm9
mulps %xmm4, %xmm0
addps %xmm5, %xmm9
mulps %xmm9, %xmm0
rsqrtps %xmm2, %xmm9
andps %xmm10, %xmm9
mulps %xmm9, %xmm2
movaps %xmm3, %xmm10
cmpneqps %xmm1, %xmm10
mulps %xmm2, %xmm9
mulps %xmm4, %xmm2
addps %xmm5, %xmm9
mulps %xmm9, %xmm2
rsqrtps %xmm1, %xmm9
andps %xmm10, %xmm9
mulps %xmm9, %xmm1
mulps %xmm1, %xmm9
mulps %xmm4, %xmm1
addps %xmm5, %xmm9
mulps %xmm9, %xmm1
rcpps %xmm0, %xmm9
mulps %xmm9, %xmm0
mulps %xmm9, %xmm0
addps %xmm9, %xmm9
subps %xmm0, %xmm9
rcpps %xmm2, %xmm0
mulps %xmm9, %xmm8
mulps %xmm0, %xmm2
movups %xmm8, -48(%rax)
mulps %xmm0, %xmm2
addps %xmm0, %xmm0
subps %xmm2, %xmm0
mulps %xmm7, %xmm0
movups %xmm0, -32(%rax)
rcpps %xmm1, %xmm0
mulps %xmm0, %xmm1
mulps %xmm0, %xmm1
addps %xmm0, %xmm0
subps %xmm1, %xmm0
mulps %xmm6, %xmm0
movups %xmm0, -16(%rax)
cmpl %r10d, %r9d
ja .L5
of course it doesn't use any blend instruction. The testcase was to
test interleaving of size 3 support which should better have used
different operations to make SLP impossible (of course the testcase
is also from real-world code which means it is a relevant regression).
