https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66051
Bug ID: 66051
Summary: can't vectorize reductions inside an SLP group
Product: gcc
Version: 6.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Blocks: 53947
Target Milestone: ---
void foo (int *p, short *q, int n)
{
int i;
for (i = 0; i < n; ++i)
{
p[i*4+0] = q[i*8+0] + q[i*8+4];
p[i*4+1] = q[i*8+1] + q[i*8+5];
p[i*4+2] = q[i*8+2] + q[i*8+6];
p[i*4+3] = q[i*8+3] + q[i*8+7];
}
}
is vectorized by unrolling the loop 4 times instead of using SLP because
t3.c:4:3: note: Build SLP for _15 = *_14;
t3.c:4:3: note: Build SLP failed: grouped loads have gaps _15 = *_14;
which isn't the whole story (I don't think we support this kind of
"reductions"). The SLP build runs into two load children, one loading
[0, 3] and one loading [4, 7] of a single group (and thus producing gaps).
Ideally we vectorize this with a single load of [0, 7], shuffle, add
and only unpack the low part.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
