https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128005
Bug ID: 128005
Summary: wrong code at -O2 and -O3 for a dot-product loop over
_BitInt(3) arrays
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: kyuwoncho18 at gmail dot com
Target Milestone: ---
Target: x86_64-pc-linux-gnu
_BitInt(3) a[64], b[64];
__attribute__((noipa)) int f(void) {
int acc = 0;
for (int i = 0; i < 64; i++)
acc += (int)a[i] * (int)b[i];
return acc;
}
int main(void) {
for (int i = 0; i < 64; i++) {
_BitInt(3) x = (i % 7) - 3, y = (i % 5) - 2;
__builtin_memcpy(&a[i], &x, sizeof x);
__builtin_memcpy(&b[i], &y, sizeof y);
}
if (f() != -6)
__builtin_abort();
}
godbolt: https://godbolt.org/z/6ehh8doGa
With -O0, -O1 or -O2 -fno-tree-vectorize, f returns -6 and the program exits
normally, but with -O2 or -O3 the loop in f is vectorized and the program
aborts.
Disclaimer: This test case was generated by a fuzzer, and the report was
written with LLM assistance.