1  /* This issue can only exist on little-endian P8 targets, since
       2     the built-in functions vec_ld/vec_st can use lxvd2x/stxvd2x
       3     (P8 big-endian) or lxv/stxv (P9 and later) for some cases,
       4     those rldicr instructions fed to them are necessary.  */
       5  /* { dg-do compile { target { powerpc_p8vector_ok && le } } } */
       6  /* { dg-options "-O2 -mdejagnu-cpu=power8" } */
       7  
       8  /* Test there are no useless instructions "rldicr x,y,0,59"
       9     to align the addresses for lvx/stvx.  */
      10  
      11  extern int a, b, c;
      12  extern vector unsigned long long ev5, ev6, ev7, ev8;
      13  extern int dummy (vector unsigned long long);
      14  
      15  int test_vec_ld(unsigned char *pe) {
      16  
      17    vector unsigned long long v1, v2, v3, v4, v9;
      18    vector unsigned long long v5 = ev5;
      19    vector unsigned long long v6 = ev6;
      20    vector unsigned long long v7 = ev7;
      21    vector unsigned long long v8 = ev8;
      22  
      23    unsigned char *e = pe;
      24  
      25    do {
      26      if (a) {
      27        v1 = __builtin_vec_ld(16, (unsigned long long *)e);
      28        v2 = __builtin_vec_ld(32, (unsigned long long *)e);
      29        v3 = __builtin_vec_ld(48, (unsigned long long *)e);
      30        e = e + 8;
      31        for (int i = 0; i < a; i++) {
      32          v4 = v5;
      33          v5 = __builtin_crypto_vpmsumd(v1, v6);
      34          v6 = __builtin_crypto_vpmsumd(v2, v7);
      35          v7 = __builtin_crypto_vpmsumd(v3, v8);
      36          e = e + 8;
      37        }
      38      }
      39      v5 = __builtin_vec_ld(16, (unsigned long long *)e);
      40      v6 = __builtin_vec_ld(32, (unsigned long long *)e);
      41      v7 = __builtin_vec_ld(48, (unsigned long long *)e);
      42      if (c)
      43        b = 1;
      44    } while (b);
      45  
      46    return dummy(v4);
      47  }
      48  
      49  int test_vec_st(unsigned char *pe) {
      50  
      51    vector unsigned long long v1, v2, v3, v4;
      52    vector unsigned long long v5 = ev5;
      53    vector unsigned long long v6 = ev6;
      54    vector unsigned long long v7 = ev7;
      55    vector unsigned long long v8 = ev8;
      56  
      57    unsigned char *e = pe;
      58  
      59    do {
      60      if (a) {
      61        __builtin_vec_st(v1, 16, (unsigned long long *)e);
      62        __builtin_vec_st(v2, 32, (unsigned long long *)e);
      63        __builtin_vec_st(v3, 48, (unsigned long long *)e);
      64        e = e + 8;
      65        for (int i = 0; i < a; i++) {
      66          v4 = v5;
      67          v5 = __builtin_crypto_vpmsumd(v1, v6);
      68          v6 = __builtin_crypto_vpmsumd(v2, v7);
      69          v7 = __builtin_crypto_vpmsumd(v3, v8);
      70          e = e + 8;
      71        }
      72      }
      73      __builtin_vec_st(v5, 16, (unsigned long long *)e);
      74      __builtin_vec_st(v6, 32, (unsigned long long *)e);
      75      __builtin_vec_st(v7, 48, (unsigned long long *)e);
      76      if (c)
      77        b = 1;
      78    } while (b);
      79  
      80    return dummy(v4);
      81  }
      82  
      83  /* { dg-final { scan-assembler-not {(?n)rldicr.*,0,59} } } */