https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125260
Bug ID: 125260
Summary: New signed overflow introduced by vectorization
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: kristerw at gcc dot gnu.org
Blocks: 118443
Target Milestone: ---
The vectorizer introduces new signed overflows when the function below (taken
from testsuite/gcc.dg/vect/vect-uncounted-run_2.c) is compiled for x86_64 with
-O2 -march=x86-64-v3
#include <assert.h>
#define N 9
__attribute__((noipa, noinline))
void
test01 ()
{
{
int x[N] = {2, 4, 6, 8, 10, 12, 14, 16, 18};
const int y[N] = {3, 5, 7, 9, 11, 13, 15, 17, 19};
int z[N] = {5, 9, 13, 17, 21, 25, 29, 33, 37};
int *x0 = x;
int *xN = x+N;
const int *y0 = y;
const int *yN = y+N;
int *res = x;
for (; x0 != xN && y0 != yN; ++x0, (void)++y0, ++res)
*res = *x0 + *y0;
assert (x0 == x+N && y0 == y+N && res == x+N);
assert (x[0] == z[0] && x[1] == z[1]);
}
}
The vectorized loop performs the addition as
vect__1.10_76 = MEM <vector(8) int> [(int *)vectp_x.8_74];
vect__2.13_79 = MEM <const vector(8) int> [(const int *)vectp_y.11_77];
vect__3.14_80 = vect__1.10_76 + vect__2.13_79;
The loads reads out of bounds in the second iteration, which is OK because of
the alignment, and later comparisons ensure the extra elements are not stored.
But adding the out of bounds elements may overflow.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118443
[Bug 118443] [Meta bug] Bugs triggered by and blocking more smtgcc testing
