https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115385
Bug ID: 115385
Summary: Peeling for gaps can be optimized more or needs to
peel more than one iteration
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Consider
void __attribute__((noipa)) foo(unsigned char * __restrict x,
unsigned char *y, int n)
{
for (int i = 0; i < n; ++i)
{
x[16*i+0] = y[3*i+0];
x[16*i+1] = y[3*i+1];
x[16*i+2] = y[3*i+2];
x[16*i+3] = y[3*i+0];
x[16*i+4] = y[3*i+1];
x[16*i+5] = y[3*i+2];
x[16*i+6] = y[3*i+0];
x[16*i+7] = y[3*i+1];
x[16*i+8] = y[3*i+2];
x[16*i+9] = y[3*i+0];
x[16*i+10] = y[3*i+1];
x[16*i+11] = y[3*i+2];
x[16*i+12] = y[3*i+0];
x[16*i+13] = y[3*i+1];
x[16*i+14] = y[3*i+2];
x[16*i+15] = y[3*i+0];
}
}
and
void __attribute__((noipa)) bar(unsigned char * __restrict x,
unsigned char *y, int n)
{
for (int i = 0; i < n; ++i)
{
x[16*i+0] = y[5*i+0];
x[16*i+1] = y[5*i+1];
x[16*i+2] = y[5*i+2];
x[16*i+3] = y[5*i+3];
x[16*i+4] = y[5*i+4];
x[16*i+5] = y[5*i+0];
x[16*i+6] = y[5*i+1];
x[16*i+7] = y[5*i+2];
x[16*i+8] = y[5*i+3];
x[16*i+9] = y[5*i+4];
x[16*i+10] = y[5*i+0];
x[16*i+11] = y[5*i+1];
x[16*i+12] = y[5*i+2];
x[16*i+13] = y[5*i+3];
x[16*i+14] = y[5*i+4];
x[16*i+15] = y[5*i+0];
}
}
for both loops we currently cannot reduce the access for the load from 'y' to
not touch extra elements so we force peeling for gaps. But in both cases
peeling a single scalar iteration is not sufficient and we get
t.c:5:21: note: ==> examining statement: _3 = y[_1];
t.c:5:21: missed: peeling for gaps insufficient for access
t.c:7:20: missed: not vectorized: relevant stmt not supported: _3 = y[_1];
we can avoid this excessive peeling for gaps if we narrow the load from 'y'
to the next power-of-two size where then it's always sufficient to just
peel a single scalar iteration. When the target cannot construct a vector
with those elements we'd have to peel more than one iteration.
