https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106217
Bug ID: 106217
Summary: [11/12/13 Regression] sinking of loads prevents
vectorization
Product: gcc
Version: 11.3.1
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: tnfchris at gcc dot gnu.org
Target Milestone: ---
The following example
void f(int n, float * __restrict x, float *mass, float max, float &ax)
{
float lax = 0.0f;
for (int i = 0; i < n; ++i) {
float f = x[i] * mass[i];
lax += x[i] >= max ? 0 : f * x[i];
}
ax += lax;
return;
}
vectorizes with GCC 10 with -Ofast but fails starting with GCC 11.
The difference is that in the sink1 pass we now sink the unconditional load of
mass[i] into the conditional.
Sinking # VUSE <.MEM_14(D)>
_7 = *_6;
from bb 3 to bb 4
Sinking _6 = mass_18(D) + _2;
from bb 3 to bb 4
consequently this causes if-convert to no longer be able to if-convert the loop
because the load is now conditional, and we only accept unconditional loads in
ifcvt_memrefs_wont_trap.
-------------------------
_6 = mass_18(D) + _2;
-------------------------
_7 = *_6;
tree could trap...
perhaps during the sinking we should mark the load as non-trapping? as it was
originally unconditional anyway.
