https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126159
Bug ID: 126159
Summary: Missed vectorization for load-lanes with 2 groups
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: liuhongt at gcc dot gnu.org
Target Milestone: ---
typedef union { int i; float f; } RtreeCoord;
#define MIN(a,b) ((a)<(b)?(a):(b))
#define MAX(a,b) ((a)>(b)?(a):(b))
void cellUnion_f(RtreeCoord *a1, RtreeCoord *a2, int n) {
int ii = 0;
do {
a1[ii].f = MIN(a1[ii].f, a2[ii].f);
a1[ii+1].f = MAX(a1[ii+1].f, a2[ii+1].f);
ii += 2;
} while (ii < n);
}
void cellUnion_i(RtreeCoord *a1, RtreeCoord *a2, int n) {
int ii = 0;
do {
a1[ii].i = MIN(a1[ii].i, a2[ii].i);
a1[ii+1].i = MAX(a1[ii+1].i, a2[ii+1].i);
ii += 2;
} while (ii < n);
}
GCC recognises the group but asks for an array_mode V8SF[2] (load-lanes with 2
groups); x86 has no such mode, so it gives up instead of doing a shuffle-based
even/odd de-interleave + a vector MIN and vector MAX + re-interleave (all cheap
AVX-512 ops). A genuine missed opportunity: a manual split into even/odd
MIN/MAX vectorizes fine.
