https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127075
Bug ID: 127075
Summary: riscv: -mrvv-max-lmul=conv-dynamic causes regression
on cpu2017 x264
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: bergner at gcc dot gnu.org
Target Milestone: ---
The -mrvv-max-lmul=conv-dynamic option causes a perfromance regression on
cpu2017 x264 by choosing an overly large LMUL value leading to spilling when
compared to a smaller LMUL value. The following test case reduced from x264
shows the problem:
bergner@rvsw-c-02-bergner-18288:X264$ cat bug.c
#include <stdint.h>
void
wide (uint8_t *__restrict a, uint8_t *__restrict b,
uint8_t *__restrict c, uint8_t *__restrict d,
int64_t *__restrict o0, int64_t *__restrict o1,
int64_t *__restrict o2, int64_t *__restrict o3,
int64_t *__restrict o4, int64_t *__restrict o5, int n)
{
for (int i = 0; i < n; i++)
{
int64_t va = a[i], vb = b[i], vc = c[i], vd = d[i];
int64_t t0 = va + vb, t1 = vb + vc, t2 = vc + vd, t3 = vd + va;
int64_t t4 = va * vb, t5 = vb * vc, t6 = vc * vd, t7 = vd * va;
o0[i] = t0 + t4;
o1[i] = t1 + t5;
o2[i] = t2 + t6;
o3[i] = t3 + t7;
o4[i] = t0 * t1 + t2 * t3;
o5[i] = t4 * t5 + t6 * t7;
}
}
bergner@rvsw-c-02-bergner-18288:X264$ gcc -S -march=rv64gcv -mabi=lp64d -O3
-ftree-vectorize -mno-vector-strict-align -mrvv-max-lmul=conv-dynamic bug.c
