https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128649
Bug ID: 128649
Summary: [14/15/16/17 Regression] wrong code at -O3: outer-loop
vectorization ignores self output dependence of a[j +
k]
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: needs-bisection, wrong-code
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jeffchen006 at gmail dot com
Target Milestone: ---
int printf(const char *, ...);
int a[32], n = 2;
int main(void) {
for (int j = 0; j < 3; j++)
for (int k = 0; k < n; k++)
a[j + k] = j;
printf("%d\n", a[1]);
}
$ gcc -O0 t.c && ./a.out
1
$ gcc -O3 t.c && ./a.out
0
Expected: 1. a[1] is written by (j=0, k=1) and later by (j=1, k=0), so a[1] =
1. The outer j loop is vectorized ("loop vectorized using 8 byte vectors and
unroll factor 2") and the order changes. A related form (a[i + 2*j]) also fails
on 7.5.0.
Suspected cause: vectorizer data dependence analysis skips the self dependence
(dra == drb) with distance (1,-1) for outer-loop vectorization.
Avoided by: -fno-tree-loop-vectorize.
Known to fail: 10.5.0, 11.4.0, 12.5.0, 13.4.0, 14.3.0, 15.2.0, 16.2.0, 17.0.
Known to work: 7.5.0, 8.5.0.
-O0 -fsanitize=undefined,address is clean.
gcc version 17.0.0 20261111 (experimental) (a2a11abfad3), x86_64-pc-linux-gnu
Configured with: --enable-languages=c,c++ --disable-multilib
--disable-bootstrap --enable-checking=release