https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112331
Bug ID: 112331
Summary: middle-end: Fail vectorization
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: juzhe.zhong at rivai dot ai
Target Milestone: ---
https://gcc.godbolt.org/z/x7GGzezGh
#include <stdio.h>
#define LEN 32000
#define ntimes 200000
#define TYPE float
#define lll LEN
#define LEN2 256
#define ALIGNMENT 16
__attribute__ ((aligned(ALIGNMENT))) TYPE X[lll],Y[lll],Z[lll],U[lll],V[lll];
struct GlobalData {
__attribute__((aligned(ALIGNMENT))) TYPE a[LEN];
int pad1[3];
__attribute__((aligned(ALIGNMENT))) TYPE b[LEN];
int pad2[5];
__attribute__((aligned(ALIGNMENT))) TYPE c[LEN];
int pad3[7];
__attribute__((aligned(ALIGNMENT))) TYPE d[LEN];
int pad4[11];
__attribute__((aligned(ALIGNMENT))) TYPE e[LEN];
int pad5[13];
__attribute__((aligned(ALIGNMENT))) TYPE aa[LEN2][LEN2];
int pad6[17];
__attribute__((aligned(ALIGNMENT))) TYPE bb[LEN2][LEN2];
int pad7[19];
__attribute__((aligned(ALIGNMENT))) TYPE cc[LEN2][LEN2];
int pad8[23];
__attribute__((aligned(ALIGNMENT))) TYPE tt[LEN2][LEN2];
} global_data;
__attribute__((aligned(ALIGNMENT))) TYPE * const a = global_data.a;
__attribute__((aligned(ALIGNMENT))) TYPE * const b = global_data.b;
__attribute__((aligned(ALIGNMENT))) TYPE * const c = global_data.c;
__attribute__((aligned(ALIGNMENT))) TYPE * const d = global_data.d;
__attribute__((aligned(ALIGNMENT))) TYPE * const e = global_data.e;
__attribute__((aligned(ALIGNMENT))) TYPE (* const aa)[LEN2] = global_data.aa;
__attribute__((aligned(ALIGNMENT))) TYPE (* const bb)[LEN2] = global_data.bb;
__attribute__((aligned(ALIGNMENT))) TYPE (* const cc)[LEN2] = global_data.cc;
__attribute__((aligned(ALIGNMENT))) TYPE (* const tt)[LEN2] = global_data.tt;
int foo()
{
// linear dependence testing
// no dependence - vectorizable
for (int nl = 0; nl < 2*ntimes; nl++) {
// #pragma vector always
for (int i = 1; i < LEN; i += 2) {
a[i] = a[i - 1] + b[i];
}
}
return 0;
}
Both RVV and ARM SVE faild to vectorize it wheras Clang can vectorize it.
