https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96208
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Smaller testcase, avoiding reductions and negative step:
void
test(double * __restrict a, double *b, double *k, int n)
{
for (int i = 0; i < n; ++i)
{
a[2*i] = b[2*i] * k[ONE*i];
a[2*i + 1] = b[2*i + 1] * k[ONE*i];
}
}
this is vectorized with interleaving with -DONE=1 as SLP discovery fails:
t.c:4:21: missed: Build SLP failed: not grouped load _9 = *_8;
.L4:
movupd (%r8,%rax,2), %xmm1
movupd (%rdx,%rax), %xmm2
movupd 16(%r8,%rax,2), %xmm0
movlpd 8(%r8,%rax,2), %xmm0
movhpd 16(%r8,%rax,2), %xmm1
mulpd %xmm2, %xmm1
mulpd %xmm2, %xmm0
movapd %xmm1, %xmm2
unpcklpd %xmm0, %xmm2
unpckhpd %xmm0, %xmm1
movups %xmm2, (%rdi,%rax,2)
movups %xmm1, 16(%rdi,%rax,2)
addq $16, %rax
cmpq %rax, %rsi
jne .L4
but when we make the access to 'k' non-unit-stride with -DONE=2
SLP works fine
t.c:4:21: note: Vectorizing SLP tree:
t.c:4:21: note: node 0x48580d0 (max_nunits=2, refcnt=1) vector(2) double
t.c:4:21: note: op template: *_8 = _9;
t.c:4:21: note: stmt 0 *_8 = _9;
t.c:4:21: note: stmt 1 *_14 = _15;
t.c:4:21: note: children 0x4858158
t.c:4:21: note: node 0x4858158 (max_nunits=2, refcnt=1) vector(2) double
t.c:4:21: note: op template: _9 = _5 * _7;
t.c:4:21: note: stmt 0 _9 = _5 * _7;
t.c:4:21: note: stmt 1 _15 = _7 * _13;
t.c:4:21: note: children 0x4858268 0x48582f0
t.c:4:21: note: node 0x4858268 (max_nunits=2, refcnt=1) vector(2) double
t.c:4:21: note: op template: _5 = *_4;
t.c:4:21: note: stmt 0 _5 = *_4;
t.c:4:21: note: stmt 1 _13 = *_12;
t.c:4:21: note: node 0x48582f0 (max_nunits=2, refcnt=1) vector(2) double
t.c:4:21: note: op template: _7 = *_6;
t.c:4:21: note: stmt 0 _7 = *_6;
t.c:4:21: note: stmt 1 _7 = *_6;
t.c:4:21: note: load permutation { 0 0 }
and we get
.L4:
movupd (%rdx,%rax), %xmm0
movupd (%rsi,%rax,2), %xmm2
unpcklpd %xmm0, %xmm0
mulpd %xmm2, %xmm0
movups %xmm0, (%rdi,%rax,2)
addq $8, %rax
cmpq %rax, %r8
jne .L4
