vec_mergel_opt_test.c

来自「Mac OS X 10.4.9 for x86 Source Code gcc」· C语言 代码 · 共 584 行 · 第 1/2 页

  v_vec_maskv = vec_or (tmp0, tmp2);
  v_vec_maskv = vec_or (v_vec_maskv, tmp3);

  intra[0]  = 1; 
  intra[1]  = 1;
  intra[2]  = 2;
  intra[3]  = 3;
  intra[4]  = 2;
  intra[5]  = 2;
  intra[6]  = 2;
  intra[7]  = 1;
  intra[8]  = 1;
  intra[9]  = 5;
  intra[10] = 5;
  intra[11] = 5;
  
  intra[13] = 0;
  intra[14] = 0;
  intra[15] = 0;

  idx0   = vec_ld (0, (signed short *) intra);
  
  v_c1   = (vector unsigned char) 1;

  if (field_MBAFF){
    v8   = (vector signed short) vec_and ((vector unsigned char) idx0, v_c1);
    idx0 = (vector signed short) vec_sra ((vector unsigned char) idx0, v_c1);

    v9   = vec_sld (v8, v8, 15);
    v9   = (vector signed short) vec_pack (v9, v8);
    
    v10  = vec_sld (v9, v9, 2);
    v11  = vec_sld (v9, v9, 10);
    
    v8   = (vector signed short) vec_cmpeq ((vector signed char) v9, (vector signed char) v10);
    v9   = (vector signed short) vec_cmpeq ((vector signed char) v9, (vector signed char) v11);
    v10  = (vector signed short) vec_cmpeq ((vector signed char) v10, (vector signed char) v11);
  }
  else  {
    v8 = v9 = v10 = vec_nor (v_zero, v_zero);
  }

  tmp1   = (vector signed short) vec_sl ((vector unsigned char) idx0, v_c1);
  
  v_c1   = vec_mergel ((vector unsigned char) v_zero, v_c1);
  tmp1   = (vector signed short) vec_add (tmp1, (vector signed short) v_c1); 

if (1){
	unsigned char toto2[16] __attribute__((aligned(16)));
	int m;
	
	vec_st(v_c1, 0, (unsigned char *) toto2);
	for (m=0; m < 16; m +=2)
	   if (toto2[m] != 0)
	     abort();
	for (m=1; m < 16; m +=2)
	   if (toto2[m] != 1)
	     abort();
}

  v_pocl = vec_ld (32, (vector unsigned char *) mv_const);
  v_poch = vec_ld (48, (vector unsigned char *) mv_const);
  tmp2   = (vector signed short) vec_perm (v_pocl, v_poch, (vector unsigned char) tmp1);      

  v_pocl = vec_ld (0,  (vector unsigned char *) mv_const);
  v_poch = vec_ld (16, (vector unsigned char *) mv_const);

  tmp1   = (vector signed short) vec_perm (v_pocl, v_poch, (vector unsigned char) tmp1);
 
  tmp1   = vec_sel (tmp1, tmp2, (vector unsigned short) (0xffff,0xffff,0,0,0,0,0,0));


  tmp3   = (vector signed short) vec_splat ((vector unsigned char) idx0, 12);
  v_c1   = (vector unsigned char) vec_nor (v_zero, v_zero);
  tmp0   = (vector signed short) vec_cmpeq ((vector signed char) idx0, (vector signed char) v_c1);
  tmp1   = vec_sel (tmp1, (vector signed short) tmp3, (vector unsigned short) tmp0);

  tmp2   = vec_sld (tmp1, tmp1, 15);
  tmp1   = (vector signed short) vec_pack (tmp2, tmp1);
  

  tmp2   = vec_sld (tmp1, tmp1, 2);
  tmp3   = vec_sld (tmp1, tmp1, 10);

  tmp0   = (vector signed short) vec_cmpeq ((vector signed char) tmp1, (vector signed char) tmp2);
  tmp4   = (vector signed short) vec_cmpeq ((vector signed char) tmp1, (vector signed char) tmp3);
  tmp1   = (vector signed short) vec_cmpeq ((vector signed char) tmp2, (vector signed char) tmp3);
  tmp0   = vec_and (tmp0, v8);
  tmp4   = vec_and (tmp4, v9);
  tmp1   = vec_and (tmp1, v10);
  tmp2   = vec_sld ((vector signed short) tmp0, (vector signed short) tmp0, 8);
  tmp3   = vec_sld ((vector signed short) tmp4, (vector signed short) tmp4, 8);
  tmp5   = vec_sld ((vector signed short) tmp1, (vector signed short) tmp1, 8);
  tmp0   = vec_and (tmp0, tmp2);
  tmp4   = vec_and (tmp4, tmp3);
  tmp1   = vec_and (tmp1, tmp5);
  v_ref_mask00 = vec_mergel ((vector signed short) tmp0, (vector signed short) v_c1);
  v_ref_mask01 = vec_mergel ((vector signed short) tmp4, (vector signed short) tmp1);
  v_ref_mask00 = (vector signed short) vec_mergel ((vector unsigned char) v_ref_mask00, (vector unsigned char) v_ref_mask00);
  v_ref_mask01 = (vector signed short) vec_mergel ((vector unsigned char) v_ref_mask01, (vector unsigned char) v_ref_mask01);
  

  v_permv= vec_ld (0, (vector unsigned char *) mv_const);
  v8     = vec_ld (0,  (vector signed short *) mv_const);
  v9     = vec_ld (16, (vector signed short *) mv_const);
  tmp2   = vec_perm (v0, v0, v_permv);
  tmp3   = vec_sub (vec_max (v8, v0), vec_min (v8, v0));
  tmp4   = vec_sub (vec_max (v8, tmp2), vec_min (v8, tmp2));
  tmp3   = (vector signed short) vec_cmpgt (tmp3, v_three); 
  tmp4   = (vector signed short) vec_cmpgt (tmp4, v_three);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  
  tmp2   = vec_perm (v2, v2, v_permv);
  tmp5   = vec_sub (vec_max (v9, v2), vec_min (v9, v2));
  tmp6   = vec_sub (vec_max (v9, tmp2), vec_min (v9, tmp2));
  tmp5   = (vector signed short) vec_cmpgt (tmp5, v_three); 
  tmp6   = (vector signed short) vec_cmpgt (tmp6, v_three);
  tmp0   = vec_sld (tmp5, tmp5, 14);
  tmp1   = vec_sld (tmp6, tmp6, 14);
  tmp5   = vec_or (tmp0, tmp5);
  tmp6   = vec_or (tmp1, tmp6);
  
  tmp3   = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp5);
  tmp4   = (vector signed short) vec_pack ((vector unsigned int) tmp4, (vector unsigned int) tmp6);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  tmp00  = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp4);

  tmp2   = vec_perm (v1, v1, v_permv);
  tmp3   = vec_sub (vec_max (v0, v1), vec_min (v0, v1));
  tmp4   = vec_sub (vec_max (v0, tmp2), vec_min (v0, tmp2));
  tmp3   = (vector signed short) vec_cmpgt (tmp3, v_three); 
  tmp4   = (vector signed short) vec_cmpgt (tmp4, v_three);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  
  tmp2   = vec_perm (v3, v3, v_permv);
  tmp5   = vec_sub (vec_max (v2, v3), vec_min (v2, v3));
  tmp6   = vec_sub (vec_max (v2, tmp2), vec_min (v2, tmp2));
  tmp5   = (vector signed short) vec_cmpgt (tmp5, v_three); 
  tmp6   = (vector signed short) vec_cmpgt (tmp6, v_three);
  tmp0   = vec_sld (tmp5, tmp5, 14);
  tmp1   = vec_sld (tmp6, tmp6, 14);
  tmp5   = vec_or (tmp0, tmp5);
  tmp6   = vec_or (tmp1, tmp6);
  
  tmp3   = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp5);
  tmp4   = (vector signed short) vec_pack ((vector unsigned int) tmp4, (vector unsigned int) tmp6);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  tmp01  = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp4);

  tmp2   = vec_perm (v4, v4, v_permv);
  tmp3   = vec_sub (vec_max (v1, v4), vec_min (v1, v4));
  tmp4   = vec_sub (vec_max (v1, tmp2), vec_min (v1, tmp2));
  tmp3   = (vector signed short) vec_cmpgt (tmp3, v_three); 
  tmp4   = (vector signed short) vec_cmpgt (tmp4, v_three);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  
  tmp2   = vec_perm (v6, v6, v_permv);
  tmp5   = vec_sub (vec_max (v3, v6), vec_min (v3, v6));
  tmp6   = vec_sub (vec_max (v3, tmp2), vec_min (v3, tmp2));
  tmp5   = (vector signed short) vec_cmpgt (tmp5, v_three); 
  tmp6   = (vector signed short) vec_cmpgt (tmp6, v_three);
  tmp0   = vec_sld (tmp5, tmp5, 14);
  tmp1   = vec_sld (tmp6, tmp6, 14);
  tmp5   = vec_or (tmp0, tmp5);
  tmp6   = vec_or (tmp1, tmp6);
  
  tmp3   = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp5);
  tmp4   = (vector signed short) vec_pack ((vector unsigned int) tmp4, (vector unsigned int) tmp6);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  tmp02  = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp4);


  tmp2   = vec_perm (v5, v5, v_permv);
  tmp3   = vec_sub (vec_max (v4, v5), vec_min (v4, v5));
  tmp4   = vec_sub (vec_max (v4, tmp2), vec_min (v4, tmp2));
  tmp3   = (vector signed short) vec_cmpgt (tmp3, v_three); 
  tmp4   = (vector signed short) vec_cmpgt (tmp4, v_three);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  
  tmp2   = vec_perm (v7, v7, v_permv);
  tmp5   = vec_sub (vec_max (v6, v7), vec_min (v6, v7));
  tmp6   = vec_sub (vec_max (v6, tmp2), vec_min (v6, tmp2));
  tmp5   = (vector signed short) vec_cmpgt (tmp5, v_three); 
  tmp6   = (vector signed short) vec_cmpgt (tmp6, v_three);
  tmp0   = vec_sld (tmp5, tmp5, 14);
  tmp1   = vec_sld (tmp6, tmp6, 14);
  tmp5   = vec_or (tmp0, tmp5);
  tmp6   = vec_or (tmp1, tmp6);
  
  tmp3   = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp5);
  tmp4   = (vector signed short) vec_pack ((vector unsigned int) tmp4, (vector unsigned int) tmp6);
  tmp5   = vec_sld (tmp3, tmp3, 14);
  tmp6   = vec_sld (tmp4, tmp4, 14);
  tmp3   = vec_or (tmp3, tmp5);
  tmp4   = vec_or (tmp4, tmp6);
  tmp03  = (vector signed short) vec_pack ((vector unsigned int) tmp3, (vector unsigned int) tmp4);

  tmp0   = (vector signed short) vec_pack ((vector unsigned short) tmp00, (vector unsigned short) tmp01);
  tmp1   = (vector signed short) vec_pack ((vector unsigned short) tmp02, (vector unsigned short) tmp03);
  tmp2   = (vector signed short) vec_mergel ((vector signed int) tmp0, (vector signed int) tmp1);
  tmp3   = (vector signed short) vec_mergel ((vector signed int) tmp0, (vector signed int) tmp1);
  tmp4   = (vector signed short) vec_mergel ((vector signed int) tmp2, (vector signed int) tmp3);
  tmp5   = (vector signed short) vec_mergel ((vector signed int) tmp2, (vector signed int) tmp3);
  tmp4   = vec_and (v_ref_mask00, tmp4);
  tmp5   = vec_and (v_ref_mask01, tmp5);

  tmp0   = vec_nor (v_ref_mask00, v_ref_mask01);
  tmp1   = vec_and (v_ref_mask00, v_ref_mask01);
  tmp2   = vec_and (tmp4, tmp5);
  tmp2   = vec_and (tmp2, tmp1);
  tmp3   = vec_nor (tmp4, tmp5);
  tmp3   = vec_nor (tmp3, tmp1);
  v_vec_maskh = vec_or (tmp0, tmp2);
  v_vec_maskh = vec_or (v_vec_maskh, tmp3);


  v_intra_maskvn = vec_nor (v_intra_maskv, v_intra_maskv);
  v_intra_maskhn = vec_nor (v_intra_maskh, v_intra_maskh);
  v_cbp_maskvn = (vector unsigned char) vec_cmpeq ((vector unsigned char) v_cbp_maskv, (vector unsigned char) v_zero);
  v_cbp_maskhn = (vector unsigned char) vec_cmpeq ((vector unsigned char) v_cbp_maskh, (vector unsigned char) v_zero);

  v_cbp_maskv  = vec_and (v_cbp_maskv, v_intra_maskvn);
  v_cbp_maskh  = vec_and (v_cbp_maskh, v_intra_maskhn);
  v_vec_maskv  = vec_and (v_vec_maskv, (vector signed short) v_intra_maskvn);
  v_vec_maskv  = vec_and (v_vec_maskv, (vector signed short) v_cbp_maskvn);
  v_vec_maskh  = vec_and (v_vec_maskh, (vector signed short) v_intra_maskhn);
  v_vec_maskh  = vec_and (v_vec_maskh, (vector signed short) v_cbp_maskhn);

  tmp9        = vec_splat_u8(2);
  tmp8        = vec_splat_u8(1);
  v_bS        = vec_ld (0, (vector unsigned char *) mv_const);
  
  v_bSv       = vec_and ((vector unsigned char) v_bS, (vector unsigned char)v_intra_maskv);
  tmp7        = vec_and ((vector unsigned char)tmp9, (vector unsigned char)v_cbp_maskv);
  tmp6        = (vector signed short) vec_and ((vector unsigned char)tmp8, (vector unsigned char)v_vec_maskv);
  tmp7        = vec_or  ((vector unsigned char)tmp7, (vector unsigned char)tmp6);
  v_bSv       = vec_or  ((vector unsigned char)tmp7, (vector unsigned char)v_bSv);

  v_bS        = vec_ld (0, (vector unsigned char *) mv_const);
  v_bSh       = vec_and ((vector unsigned char) v_bS, (vector unsigned char)v_intra_maskh);
  tmp7        = vec_and ((vector unsigned char)tmp9, (vector unsigned char)v_cbp_maskh);
  tmp6        = (vector signed short) vec_and ((vector unsigned char)tmp8, (vector unsigned char)v_vec_maskh);
  tmp7        = vec_or  ((vector unsigned char)tmp7, (vector unsigned char)tmp6);
  v_bSh       = vec_or  ((vector unsigned char)tmp7, (vector unsigned char)v_bSh);

  v_permh     = (vector unsigned char) vec_ld (0 , (vector unsigned char *) mv_const);
  v_permv     = (vector unsigned char) vec_ld (0, (vector unsigned char *) mv_const);
  v_bSv       = vec_and (v_bSv, v_permv);
  v_bSh       = vec_and (v_bSh, v_permh);

  vec_st (v_bSv, 0, (unsigned char *) mv_const);
  vec_st (v_bSh, 0, (unsigned char *) mv_const);

  v_bSv = vec_mergel (v_bSv, v_bSv);
  v_bSv = vec_mergel (v_bSv, v_bSv);
  v_bSh = vec_mergel (v_bSh, v_bSh);
  v_bSh = vec_mergel (v_bSh, v_bSh);
    
  vec_st (v_bSv, 0, (vector unsigned char *) mv_const);
  vec_st (v_bSh, 0,(vector unsigned char *) mv_const);
}


int main(int argc, char **argv)
{
    char toto[32] __attribute__((aligned(16)));
    foo(toto, toto, 0, 0);
    return 0;
}
/* APPLE LOCAL end radar 4110116 */