https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99929
Bug ID: 99929
Summary: SVE: Wrong code at -O2 -ftree-vectorize
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: acoplan at gcc dot gnu.org
Target Milestone: ---
AArch64 GCC miscompiles the following testcase:
#include <arm_sve.h>
static void e(unsigned long long *g, int p2) { *g ^= p2; }
static unsigned long long b;
static int f[1][1][1][1];
static long l[23][2];
static short m[23];
int main() {
for (unsigned i = 0; i < 23; ++i)
for (unsigned j = 0; j < 2; ++j)
l[i][j] = m[i] = 4;
if (svaddv(svptrue_pat_b32(SV_VL1), svdup_u32(1)) != 1)
__builtin_abort();
for (unsigned i = 0; i < 3; ++i)
e(&b, m[i]);
}
with -march=armv8.2-a+sve -O2 -ftree-vectorize. At -O2 (without
-ftree-vectorize), we do the reduction with:
uaddv d0, p0, z0.s
where the predicate is generated by:
ptrue p0.b, vl1
which gives the expected result. With -ftree-vectorize, we do the reduction
with:
uaddv d0, p1, z0.s
where the predicate is generated by:
ptrue p1.h, all
which does not give the expected result.
