https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112327
Bug ID: 112327
Summary: RVV: Redundant vmv1r for widen reduction
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: ---
#include "riscv_vector.h"
void rvv_dot_prod(int16_t *pSrcA, int16_t *pSrcB, uint32_t n, int64_t *result)
{
size_t vl;
vint16m4_t vSrcA, vSrcB;
vint64m1_t vSum = __riscv_vmv_s_x_i64m1(0, 1);
while (n > 0) {
vl = __riscv_vsetvl_e16m4(n);
vSrcA = __riscv_vle16_v_i16m4(pSrcA, vl);
vSrcB = __riscv_vle16_v_i16m4(pSrcB, vl);
vSum = __riscv_vwredsum_vs_i32m8_i64m1(__riscv_vwmul_vv_i32m8(vSrcA,
vSrcB, vl), vSum, vl);
pSrcA += vl;
pSrcB += vl;
n -= vl;
}
*result = __riscv_vmv_x_s_i64m1_i64(vSum);
}
https://godbolt.org/z/sb8G7ExKP
GCC:
...
vmv1r.v v2,v1
...
vwredsum.vs v1,v8,v2
Clang:
vwredsum.vs v8, v24, v8
The root cause is that we don't allow vwredsum.vs vd,vs2,vs1, vs1 overlaps vd.