https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98542
Bug ID: 98542
Summary: Redundant loads in vectorised loop
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
For the testcase below, based loosely on one from 450.soplex,
the vectoriser loads the v and i fields twice:
struct s { double v; long i; };
double
f (struct s *x, double *y, int n)
{
double res = 0;
for (int i = 0; i < n; ++i)
res += x[i].v * y[x[i].i];
return res;
}
SVE loop:
add x5, x0, 8
...
.L4:
ld2d {z2.d - z3.d}, p0/z, [x5, x4, lsl 3]
ld2d {z4.d - z5.d}, p0/z, [x0, x4, lsl 3]
ld1d z2.d, p0/z, [x1, z2.d, lsl 3]
incd x3
fmla z0.d, p0/m, z4.d, z2.d
incw x4
whilelo p0.d, w3, w2
b.any .L4
where z5 from the second ld2d is the same as z2
from the first ld2d.
In the soplex example, "i" is instead a 32-bit value,
but I guess we need to fix this case first.
