test/compiler/6340864/TestShortVect.java

changeset 4001
006050192a5a
child 4204
b2c669fd8114
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/test/compiler/6340864/TestShortVect.java	Mon Aug 20 09:07:21 2012 -0700
     1.3 @@ -0,0 +1,1127 @@
     1.4 +/*
     1.5 + * Copyright (c) 2012, Oracle and/or its affiliates. All rights reserved.
     1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     1.7 + *
     1.8 + * This code is free software; you can redistribute it and/or modify it
     1.9 + * under the terms of the GNU General Public License version 2 only, as
    1.10 + * published by the Free Software Foundation.
    1.11 + *
    1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
    1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    1.14 + * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    1.15 + * version 2 for more details (a copy is included in the LICENSE file that
    1.16 + * accompanied this code).
    1.17 + *
    1.18 + * You should have received a copy of the GNU General Public License version
    1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
    1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    1.21 + *
    1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    1.23 + * or visit www.oracle.com if you need additional information or have any
    1.24 + * questions.
    1.25 + *
    1.26 + */
    1.27 +
    1.28 +/**
    1.29 + * @test
    1.30 + * @bug 6340864
    1.31 + * @summary Implement vectorization optimizations in hotspot-server
    1.32 + *
    1.33 + * @run main/othervm/timeout=400 -Xbatch -Xmx64m TestShortVect
    1.34 + */
    1.35 +
    1.36 +public class TestShortVect {
    1.37 +  private static final int ARRLEN = 997;
    1.38 +  private static final int ITERS  = 11000;
    1.39 +  private static final int ADD_INIT = Short.MAX_VALUE-500;
    1.40 +  private static final int BIT_MASK = 0xB731;
    1.41 +  private static final int VALUE = 7;
    1.42 +  private static final int SHIFT = 16;
    1.43 +
    1.44 +  public static void main(String args[]) {
    1.45 +    System.out.println("Testing Short vectors");
    1.46 +    int errn = test();
    1.47 +    if (errn > 0) {
    1.48 +      System.err.println("FAILED: " + errn + " errors");
    1.49 +      System.exit(97);
    1.50 +    }
    1.51 +    System.out.println("PASSED");
    1.52 +  }
    1.53 +
    1.54 +  static int test() {
    1.55 +    short[] a0 = new short[ARRLEN];
    1.56 +    short[] a1 = new short[ARRLEN];
    1.57 +    short[] a2 = new short[ARRLEN];
    1.58 +    short[] a3 = new short[ARRLEN];
    1.59 +    short[] a4 = new short[ARRLEN];
    1.60 +     int[] p2 = new  int[ARRLEN/2];
    1.61 +    long[] p4 = new long[ARRLEN/4];
    1.62 +    // Initialize
    1.63 +    int gold_sum = 0;
    1.64 +    for (int i=0; i<ARRLEN; i++) {
    1.65 +      short val = (short)(ADD_INIT+i);
    1.66 +      gold_sum += val;
    1.67 +      a1[i] = val;
    1.68 +      a2[i] = (short)VALUE;
    1.69 +      a3[i] = (short)-VALUE;
    1.70 +      a4[i] = (short)BIT_MASK;
    1.71 +    }
    1.72 +    System.out.println("Warmup");
    1.73 +    for (int i=0; i<ITERS; i++) {
    1.74 +      test_sum(a1);
    1.75 +      test_addc(a0, a1);
    1.76 +      test_addv(a0, a1, (short)VALUE);
    1.77 +      test_adda(a0, a1, a2);
    1.78 +      test_subc(a0, a1);
    1.79 +      test_subv(a0, a1, (short)VALUE);
    1.80 +      test_suba(a0, a1, a2);
    1.81 +      test_mulc(a0, a1);
    1.82 +      test_mulv(a0, a1, (short)VALUE);
    1.83 +      test_mula(a0, a1, a2);
    1.84 +      test_divc(a0, a1);
    1.85 +      test_divv(a0, a1, (short)VALUE);
    1.86 +      test_diva(a0, a1, a2);
    1.87 +      test_mulc_n(a0, a1);
    1.88 +      test_mulv(a0, a1, (short)-VALUE);
    1.89 +      test_mula(a0, a1, a3);
    1.90 +      test_divc_n(a0, a1);
    1.91 +      test_divv(a0, a1, (short)-VALUE);
    1.92 +      test_diva(a0, a1, a3);
    1.93 +      test_andc(a0, a1);
    1.94 +      test_andv(a0, a1, (short)BIT_MASK);
    1.95 +      test_anda(a0, a1, a4);
    1.96 +      test_orc(a0, a1);
    1.97 +      test_orv(a0, a1, (short)BIT_MASK);
    1.98 +      test_ora(a0, a1, a4);
    1.99 +      test_xorc(a0, a1);
   1.100 +      test_xorv(a0, a1, (short)BIT_MASK);
   1.101 +      test_xora(a0, a1, a4);
   1.102 +      test_sllc(a0, a1);
   1.103 +      test_sllv(a0, a1, VALUE);
   1.104 +      test_srlc(a0, a1);
   1.105 +      test_srlv(a0, a1, VALUE);
   1.106 +      test_srac(a0, a1);
   1.107 +      test_srav(a0, a1, VALUE);
   1.108 +      test_sllc_n(a0, a1);
   1.109 +      test_sllv(a0, a1, -VALUE);
   1.110 +      test_srlc_n(a0, a1);
   1.111 +      test_srlv(a0, a1, -VALUE);
   1.112 +      test_srac_n(a0, a1);
   1.113 +      test_srav(a0, a1, -VALUE);
   1.114 +      test_sllc_o(a0, a1);
   1.115 +      test_sllv(a0, a1, SHIFT);
   1.116 +      test_srlc_o(a0, a1);
   1.117 +      test_srlv(a0, a1, SHIFT);
   1.118 +      test_srac_o(a0, a1);
   1.119 +      test_srav(a0, a1, SHIFT);
   1.120 +      test_sllc_on(a0, a1);
   1.121 +      test_sllv(a0, a1, -SHIFT);
   1.122 +      test_srlc_on(a0, a1);
   1.123 +      test_srlv(a0, a1, -SHIFT);
   1.124 +      test_srac_on(a0, a1);
   1.125 +      test_srav(a0, a1, -SHIFT);
   1.126 +      test_pack2(p2, a1);
   1.127 +      test_unpack2(a0, p2);
   1.128 +      test_pack2_swap(p2, a1);
   1.129 +      test_unpack2_swap(a0, p2);
   1.130 +      test_pack4(p4, a1);
   1.131 +      test_unpack4(a0, p4);
   1.132 +      test_pack4_swap(p4, a1);
   1.133 +      test_unpack4_swap(a0, p4);
   1.134 +    }
   1.135 +    // Test and verify results
   1.136 +    System.out.println("Verification");
   1.137 +    int errn = 0;
   1.138 +    {
   1.139 +      int sum = test_sum(a1);
   1.140 +      if (sum != gold_sum) {
   1.141 +        System.err.println("test_sum:  " + sum + " != " + gold_sum);
   1.142 +        errn++;
   1.143 +      }
   1.144 +
   1.145 +      test_addc(a0, a1);
   1.146 +      for (int i=0; i<ARRLEN; i++) {
   1.147 +        errn += verify("test_addc: ", i, a0[i], (short)((short)(ADD_INIT+i)+VALUE));
   1.148 +      }
   1.149 +      test_addv(a0, a1, (short)VALUE);
   1.150 +      for (int i=0; i<ARRLEN; i++) {
   1.151 +        errn += verify("test_addv: ", i, a0[i], (short)((short)(ADD_INIT+i)+VALUE));
   1.152 +      }
   1.153 +      test_adda(a0, a1, a2);
   1.154 +      for (int i=0; i<ARRLEN; i++) {
   1.155 +        errn += verify("test_adda: ", i, a0[i], (short)((short)(ADD_INIT+i)+VALUE));
   1.156 +      }
   1.157 +
   1.158 +      test_subc(a0, a1);
   1.159 +      for (int i=0; i<ARRLEN; i++) {
   1.160 +        errn += verify("test_subc: ", i, a0[i], (short)((short)(ADD_INIT+i)-VALUE));
   1.161 +      }
   1.162 +      test_subv(a0, a1, (short)VALUE);
   1.163 +      for (int i=0; i<ARRLEN; i++) {
   1.164 +        errn += verify("test_subv: ", i, a0[i], (short)((short)(ADD_INIT+i)-VALUE));
   1.165 +      }
   1.166 +      test_suba(a0, a1, a2);
   1.167 +      for (int i=0; i<ARRLEN; i++) {
   1.168 +        errn += verify("test_suba: ", i, a0[i], (short)((short)(ADD_INIT+i)-VALUE));
   1.169 +      }
   1.170 +
   1.171 +      test_mulc(a0, a1);
   1.172 +      for (int i=0; i<ARRLEN; i++) {
   1.173 +        errn += verify("test_mulc: ", i, a0[i], (short)((short)(ADD_INIT+i)*VALUE));
   1.174 +      }
   1.175 +      test_mulv(a0, a1, (short)VALUE);
   1.176 +      for (int i=0; i<ARRLEN; i++) {
   1.177 +        errn += verify("test_mulv: ", i, a0[i], (short)((short)(ADD_INIT+i)*VALUE));
   1.178 +      }
   1.179 +      test_mula(a0, a1, a2);
   1.180 +      for (int i=0; i<ARRLEN; i++) {
   1.181 +        errn += verify("test_mula: ", i, a0[i], (short)((short)(ADD_INIT+i)*VALUE));
   1.182 +      }
   1.183 +
   1.184 +      test_divc(a0, a1);
   1.185 +      for (int i=0; i<ARRLEN; i++) {
   1.186 +        errn += verify("test_divc: ", i, a0[i], (short)((short)(ADD_INIT+i)/VALUE));
   1.187 +      }
   1.188 +      test_divv(a0, a1, (short)VALUE);
   1.189 +      for (int i=0; i<ARRLEN; i++) {
   1.190 +        errn += verify("test_divv: ", i, a0[i], (short)((short)(ADD_INIT+i)/VALUE));
   1.191 +      }
   1.192 +      test_diva(a0, a1, a2);
   1.193 +      for (int i=0; i<ARRLEN; i++) {
   1.194 +        errn += verify("test_diva: ", i, a0[i], (short)((short)(ADD_INIT+i)/VALUE));
   1.195 +      }
   1.196 +
   1.197 +      test_mulc_n(a0, a1);
   1.198 +      for (int i=0; i<ARRLEN; i++) {
   1.199 +        errn += verify("test_mulc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)*(-VALUE)));
   1.200 +      }
   1.201 +      test_mulv(a0, a1, (short)-VALUE);
   1.202 +      for (int i=0; i<ARRLEN; i++) {
   1.203 +        errn += verify("test_mulv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)*(-VALUE)));
   1.204 +      }
   1.205 +      test_mula(a0, a1, a3);
   1.206 +      for (int i=0; i<ARRLEN; i++) {
   1.207 +        errn += verify("test_mula_n: ", i, a0[i], (short)((short)(ADD_INIT+i)*(-VALUE)));
   1.208 +      }
   1.209 +
   1.210 +      test_divc_n(a0, a1);
   1.211 +      for (int i=0; i<ARRLEN; i++) {
   1.212 +        errn += verify("test_divc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)/(-VALUE)));
   1.213 +      }
   1.214 +      test_divv(a0, a1, (short)-VALUE);
   1.215 +      for (int i=0; i<ARRLEN; i++) {
   1.216 +        errn += verify("test_divv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)/(-VALUE)));
   1.217 +      }
   1.218 +      test_diva(a0, a1, a3);
   1.219 +      for (int i=0; i<ARRLEN; i++) {
   1.220 +        errn += verify("test_diva_n: ", i, a0[i], (short)((short)(ADD_INIT+i)/(-VALUE)));
   1.221 +      }
   1.222 +
   1.223 +      test_andc(a0, a1);
   1.224 +      for (int i=0; i<ARRLEN; i++) {
   1.225 +        errn += verify("test_andc: ", i, a0[i], (short)((short)(ADD_INIT+i)&BIT_MASK));
   1.226 +      }
   1.227 +      test_andv(a0, a1, (short)BIT_MASK);
   1.228 +      for (int i=0; i<ARRLEN; i++) {
   1.229 +        errn += verify("test_andv: ", i, a0[i], (short)((short)(ADD_INIT+i)&BIT_MASK));
   1.230 +      }
   1.231 +      test_anda(a0, a1, a4);
   1.232 +      for (int i=0; i<ARRLEN; i++) {
   1.233 +        errn += verify("test_anda: ", i, a0[i], (short)((short)(ADD_INIT+i)&BIT_MASK));
   1.234 +      }
   1.235 +
   1.236 +      test_orc(a0, a1);
   1.237 +      for (int i=0; i<ARRLEN; i++) {
   1.238 +        errn += verify("test_orc: ", i, a0[i], (short)((short)(ADD_INIT+i)|BIT_MASK));
   1.239 +      }
   1.240 +      test_orv(a0, a1, (short)BIT_MASK);
   1.241 +      for (int i=0; i<ARRLEN; i++) {
   1.242 +        errn += verify("test_orv: ", i, a0[i], (short)((short)(ADD_INIT+i)|BIT_MASK));
   1.243 +      }
   1.244 +      test_ora(a0, a1, a4);
   1.245 +      for (int i=0; i<ARRLEN; i++) {
   1.246 +        errn += verify("test_ora: ", i, a0[i], (short)((short)(ADD_INIT+i)|BIT_MASK));
   1.247 +      }
   1.248 +
   1.249 +      test_xorc(a0, a1);
   1.250 +      for (int i=0; i<ARRLEN; i++) {
   1.251 +        errn += verify("test_xorc: ", i, a0[i], (short)((short)(ADD_INIT+i)^BIT_MASK));
   1.252 +      }
   1.253 +      test_xorv(a0, a1, (short)BIT_MASK);
   1.254 +      for (int i=0; i<ARRLEN; i++) {
   1.255 +        errn += verify("test_xorv: ", i, a0[i], (short)((short)(ADD_INIT+i)^BIT_MASK));
   1.256 +      }
   1.257 +      test_xora(a0, a1, a4);
   1.258 +      for (int i=0; i<ARRLEN; i++) {
   1.259 +        errn += verify("test_xora: ", i, a0[i], (short)((short)(ADD_INIT+i)^BIT_MASK));
   1.260 +      }
   1.261 +
   1.262 +      test_sllc(a0, a1);
   1.263 +      for (int i=0; i<ARRLEN; i++) {
   1.264 +        errn += verify("test_sllc: ", i, a0[i], (short)((short)(ADD_INIT+i)<<VALUE));
   1.265 +      }
   1.266 +      test_sllv(a0, a1, VALUE);
   1.267 +      for (int i=0; i<ARRLEN; i++) {
   1.268 +        errn += verify("test_sllv: ", i, a0[i], (short)((short)(ADD_INIT+i)<<VALUE));
   1.269 +      }
   1.270 +
   1.271 +      test_srlc(a0, a1);
   1.272 +      for (int i=0; i<ARRLEN; i++) {
   1.273 +        errn += verify("test_srlc: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>VALUE));
   1.274 +      }
   1.275 +      test_srlv(a0, a1, VALUE);
   1.276 +      for (int i=0; i<ARRLEN; i++) {
   1.277 +        errn += verify("test_srlv: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>VALUE));
   1.278 +      }
   1.279 +
   1.280 +      test_srac(a0, a1);
   1.281 +      for (int i=0; i<ARRLEN; i++) {
   1.282 +        errn += verify("test_srac: ", i, a0[i], (short)((short)(ADD_INIT+i)>>VALUE));
   1.283 +      }
   1.284 +      test_srav(a0, a1, VALUE);
   1.285 +      for (int i=0; i<ARRLEN; i++) {
   1.286 +        errn += verify("test_srav: ", i, a0[i], (short)((short)(ADD_INIT+i)>>VALUE));
   1.287 +      }
   1.288 +
   1.289 +      test_sllc_n(a0, a1);
   1.290 +      for (int i=0; i<ARRLEN; i++) {
   1.291 +        errn += verify("test_sllc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-VALUE)));
   1.292 +      }
   1.293 +      test_sllv(a0, a1, -VALUE);
   1.294 +      for (int i=0; i<ARRLEN; i++) {
   1.295 +        errn += verify("test_sllv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-VALUE)));
   1.296 +      }
   1.297 +
   1.298 +      test_srlc_n(a0, a1);
   1.299 +      for (int i=0; i<ARRLEN; i++) {
   1.300 +        errn += verify("test_srlc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-VALUE)));
   1.301 +      }
   1.302 +      test_srlv(a0, a1, -VALUE);
   1.303 +      for (int i=0; i<ARRLEN; i++) {
   1.304 +        errn += verify("test_srlv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-VALUE)));
   1.305 +      }
   1.306 +
   1.307 +      test_srac_n(a0, a1);
   1.308 +      for (int i=0; i<ARRLEN; i++) {
   1.309 +        errn += verify("test_srac_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-VALUE)));
   1.310 +      }
   1.311 +      test_srav(a0, a1, -VALUE);
   1.312 +      for (int i=0; i<ARRLEN; i++) {
   1.313 +        errn += verify("test_srav_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-VALUE)));
   1.314 +      }
   1.315 +
   1.316 +      test_sllc_o(a0, a1);
   1.317 +      for (int i=0; i<ARRLEN; i++) {
   1.318 +        errn += verify("test_sllc_o: ", i, a0[i], (short)((short)(ADD_INIT+i)<<SHIFT));
   1.319 +      }
   1.320 +      test_sllv(a0, a1, SHIFT);
   1.321 +      for (int i=0; i<ARRLEN; i++) {
   1.322 +        errn += verify("test_sllv_o: ", i, a0[i], (short)((short)(ADD_INIT+i)<<SHIFT));
   1.323 +      }
   1.324 +
   1.325 +      test_srlc_o(a0, a1);
   1.326 +      for (int i=0; i<ARRLEN; i++) {
   1.327 +        errn += verify("test_srlc_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>SHIFT));
   1.328 +      }
   1.329 +      test_srlv(a0, a1, SHIFT);
   1.330 +      for (int i=0; i<ARRLEN; i++) {
   1.331 +        errn += verify("test_srlv_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>SHIFT));
   1.332 +      }
   1.333 +
   1.334 +      test_srac_o(a0, a1);
   1.335 +      for (int i=0; i<ARRLEN; i++) {
   1.336 +        errn += verify("test_srac_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>SHIFT));
   1.337 +      }
   1.338 +      test_srav(a0, a1, SHIFT);
   1.339 +      for (int i=0; i<ARRLEN; i++) {
   1.340 +        errn += verify("test_srav_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>SHIFT));
   1.341 +      }
   1.342 +
   1.343 +      test_sllc_on(a0, a1);
   1.344 +      for (int i=0; i<ARRLEN; i++) {
   1.345 +        errn += verify("test_sllc_on: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-SHIFT)));
   1.346 +      }
   1.347 +      test_sllv(a0, a1, -SHIFT);
   1.348 +      for (int i=0; i<ARRLEN; i++) {
   1.349 +        errn += verify("test_sllv_on: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-SHIFT)));
   1.350 +      }
   1.351 +
   1.352 +      test_srlc_on(a0, a1);
   1.353 +      for (int i=0; i<ARRLEN; i++) {
   1.354 +        errn += verify("test_srlc_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-SHIFT)));
   1.355 +      }
   1.356 +      test_srlv(a0, a1, -SHIFT);
   1.357 +      for (int i=0; i<ARRLEN; i++) {
   1.358 +        errn += verify("test_srlv_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-SHIFT)));
   1.359 +      }
   1.360 +
   1.361 +      test_srac_on(a0, a1);
   1.362 +      for (int i=0; i<ARRLEN; i++) {
   1.363 +        errn += verify("test_srac_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-SHIFT)));
   1.364 +      }
   1.365 +      test_srav(a0, a1, -SHIFT);
   1.366 +      for (int i=0; i<ARRLEN; i++) {
   1.367 +        errn += verify("test_srav_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-SHIFT)));
   1.368 +      }
   1.369 +
   1.370 +      test_pack2(p2, a1);
   1.371 +      for (int i=0; i<ARRLEN/2; i++) {
   1.372 +        errn += verify("test_pack2: ", i, p2[i], ((int)(ADD_INIT+2*i) & 0xFFFF) | ((int)(ADD_INIT+2*i+1) << 16));
   1.373 +      }
   1.374 +      for (int i=0; i<ARRLEN; i++) {
   1.375 +        a0[i] = -1;
   1.376 +      }
   1.377 +      test_unpack2(a0, p2);
   1.378 +      for (int i=0; i<(ARRLEN&(-2)); i++) {
   1.379 +        errn += verify("test_unpack2: ", i, a0[i], (short)(ADD_INIT+i));
   1.380 +      }
   1.381 +
   1.382 +      test_pack2_swap(p2, a1);
   1.383 +      for (int i=0; i<ARRLEN/2; i++) {
   1.384 +        errn += verify("test_pack2_swap: ", i, p2[i], ((int)(ADD_INIT+2*i+1) & 0xFFFF) | ((int)(ADD_INIT+2*i) << 16));
   1.385 +      }
   1.386 +      for (int i=0; i<ARRLEN; i++) {
   1.387 +        a0[i] = -1;
   1.388 +      }
   1.389 +      test_unpack2_swap(a0, p2);
   1.390 +      for (int i=0; i<(ARRLEN&(-2)); i++) {
   1.391 +        errn += verify("test_unpack2_swap: ", i, a0[i], (short)(ADD_INIT+i));
   1.392 +      }
   1.393 +
   1.394 +      test_pack4(p4, a1);
   1.395 +      for (int i=0; i<ARRLEN/4; i++) {
   1.396 +        errn += verify("test_pack4: ", i, p4[i],  ((long)(ADD_INIT+4*i+0) & 0xFFFFl) |
   1.397 +                                                 (((long)(ADD_INIT+4*i+1) & 0xFFFFl) << 16)  |
   1.398 +                                                 (((long)(ADD_INIT+4*i+2) & 0xFFFFl) << 32)  |
   1.399 +                                                 (((long)(ADD_INIT+4*i+3) & 0xFFFFl) << 48));
   1.400 +      }
   1.401 +      for (int i=0; i<ARRLEN; i++) {
   1.402 +        a0[i] = -1;
   1.403 +      }
   1.404 +      test_unpack4(a0, p4);
   1.405 +      for (int i=0; i<(ARRLEN&(-4)); i++) {
   1.406 +        errn += verify("test_unpack4: ", i, a0[i], (short)(ADD_INIT+i));
   1.407 +      }
   1.408 +
   1.409 +      test_pack4_swap(p4, a1);
   1.410 +      for (int i=0; i<ARRLEN/4; i++) {
   1.411 +        errn += verify("test_pack4_swap: ", i, p4[i],  ((long)(ADD_INIT+4*i+3) & 0xFFFFl) |
   1.412 +                                                      (((long)(ADD_INIT+4*i+2) & 0xFFFFl) << 16)  |
   1.413 +                                                      (((long)(ADD_INIT+4*i+1) & 0xFFFFl) << 32)  |
   1.414 +                                                      (((long)(ADD_INIT+4*i+0) & 0xFFFFl) << 48));
   1.415 +      }
   1.416 +      for (int i=0; i<ARRLEN; i++) {
   1.417 +        a0[i] = -1;
   1.418 +      }
   1.419 +      test_unpack4_swap(a0, p4);
   1.420 +      for (int i=0; i<(ARRLEN&(-4)); i++) {
   1.421 +        errn += verify("test_unpack4_swap: ", i, a0[i], (short)(ADD_INIT+i));
   1.422 +      }
   1.423 +
   1.424 +    }
   1.425 +
   1.426 +    if (errn > 0)
   1.427 +      return errn;
   1.428 +
   1.429 +    System.out.println("Time");
   1.430 +    long start, end;
   1.431 +
   1.432 +    start = System.currentTimeMillis();
   1.433 +    for (int i=0; i<ITERS; i++) {
   1.434 +      test_sum(a1);
   1.435 +    }
   1.436 +    end = System.currentTimeMillis();
   1.437 +    System.out.println("test_sum: " + (end - start));
   1.438 +
   1.439 +    start = System.currentTimeMillis();
   1.440 +    for (int i=0; i<ITERS; i++) {
   1.441 +      test_addc(a0, a1);
   1.442 +    }
   1.443 +    end = System.currentTimeMillis();
   1.444 +    System.out.println("test_addc: " + (end - start));
   1.445 +    start = System.currentTimeMillis();
   1.446 +    for (int i=0; i<ITERS; i++) {
   1.447 +      test_addv(a0, a1, (short)VALUE);
   1.448 +    }
   1.449 +    end = System.currentTimeMillis();
   1.450 +    System.out.println("test_addv: " + (end - start));
   1.451 +    start = System.currentTimeMillis();
   1.452 +    for (int i=0; i<ITERS; i++) {
   1.453 +      test_adda(a0, a1, a2);
   1.454 +    }
   1.455 +    end = System.currentTimeMillis();
   1.456 +    System.out.println("test_adda: " + (end - start));
   1.457 +
   1.458 +    start = System.currentTimeMillis();
   1.459 +    for (int i=0; i<ITERS; i++) {
   1.460 +      test_subc(a0, a1);
   1.461 +    }
   1.462 +    end = System.currentTimeMillis();
   1.463 +    System.out.println("test_subc: " + (end - start));
   1.464 +    start = System.currentTimeMillis();
   1.465 +    for (int i=0; i<ITERS; i++) {
   1.466 +      test_subv(a0, a1, (short)VALUE);
   1.467 +    }
   1.468 +    end = System.currentTimeMillis();
   1.469 +    System.out.println("test_subv: " + (end - start));
   1.470 +    start = System.currentTimeMillis();
   1.471 +    for (int i=0; i<ITERS; i++) {
   1.472 +      test_suba(a0, a1, a2);
   1.473 +    }
   1.474 +    end = System.currentTimeMillis();
   1.475 +    System.out.println("test_suba: " + (end - start));
   1.476 +
   1.477 +    start = System.currentTimeMillis();
   1.478 +    for (int i=0; i<ITERS; i++) {
   1.479 +      test_mulc(a0, a1);
   1.480 +    }
   1.481 +    end = System.currentTimeMillis();
   1.482 +    System.out.println("test_mulc: " + (end - start));
   1.483 +    start = System.currentTimeMillis();
   1.484 +    for (int i=0; i<ITERS; i++) {
   1.485 +      test_mulv(a0, a1, (short)VALUE);
   1.486 +    }
   1.487 +    end = System.currentTimeMillis();
   1.488 +    System.out.println("test_mulv: " + (end - start));
   1.489 +    start = System.currentTimeMillis();
   1.490 +    for (int i=0; i<ITERS; i++) {
   1.491 +      test_mula(a0, a1, a2);
   1.492 +    }
   1.493 +    end = System.currentTimeMillis();
   1.494 +    System.out.println("test_mula: " + (end - start));
   1.495 +
   1.496 +    start = System.currentTimeMillis();
   1.497 +    for (int i=0; i<ITERS; i++) {
   1.498 +      test_divc(a0, a1);
   1.499 +    }
   1.500 +    end = System.currentTimeMillis();
   1.501 +    System.out.println("test_divc: " + (end - start));
   1.502 +    start = System.currentTimeMillis();
   1.503 +    for (int i=0; i<ITERS; i++) {
   1.504 +      test_divv(a0, a1, (short)VALUE);
   1.505 +    }
   1.506 +    end = System.currentTimeMillis();
   1.507 +    System.out.println("test_divv: " + (end - start));
   1.508 +    start = System.currentTimeMillis();
   1.509 +    for (int i=0; i<ITERS; i++) {
   1.510 +      test_diva(a0, a1, a2);
   1.511 +    }
   1.512 +    end = System.currentTimeMillis();
   1.513 +    System.out.println("test_diva: " + (end - start));
   1.514 +
   1.515 +    start = System.currentTimeMillis();
   1.516 +    for (int i=0; i<ITERS; i++) {
   1.517 +      test_mulc_n(a0, a1);
   1.518 +    }
   1.519 +    end = System.currentTimeMillis();
   1.520 +    System.out.println("test_mulc_n: " + (end - start));
   1.521 +    start = System.currentTimeMillis();
   1.522 +    for (int i=0; i<ITERS; i++) {
   1.523 +      test_mulv(a0, a1, (short)-VALUE);
   1.524 +    }
   1.525 +    end = System.currentTimeMillis();
   1.526 +    System.out.println("test_mulv_n: " + (end - start));
   1.527 +    start = System.currentTimeMillis();
   1.528 +    for (int i=0; i<ITERS; i++) {
   1.529 +      test_mula(a0, a1, a3);
   1.530 +    }
   1.531 +    end = System.currentTimeMillis();
   1.532 +    System.out.println("test_mula_n: " + (end - start));
   1.533 +
   1.534 +    start = System.currentTimeMillis();
   1.535 +    for (int i=0; i<ITERS; i++) {
   1.536 +      test_divc_n(a0, a1);
   1.537 +    }
   1.538 +    end = System.currentTimeMillis();
   1.539 +    System.out.println("test_divc_n: " + (end - start));
   1.540 +    start = System.currentTimeMillis();
   1.541 +    for (int i=0; i<ITERS; i++) {
   1.542 +      test_divv(a0, a1, (short)-VALUE);
   1.543 +    }
   1.544 +    end = System.currentTimeMillis();
   1.545 +    System.out.println("test_divv_n: " + (end - start));
   1.546 +    start = System.currentTimeMillis();
   1.547 +    for (int i=0; i<ITERS; i++) {
   1.548 +      test_diva(a0, a1, a3);
   1.549 +    }
   1.550 +    end = System.currentTimeMillis();
   1.551 +    System.out.println("test_diva_n: " + (end - start));
   1.552 +
   1.553 +    start = System.currentTimeMillis();
   1.554 +    for (int i=0; i<ITERS; i++) {
   1.555 +      test_andc(a0, a1);
   1.556 +    }
   1.557 +    end = System.currentTimeMillis();
   1.558 +    System.out.println("test_andc: " + (end - start));
   1.559 +    start = System.currentTimeMillis();
   1.560 +    for (int i=0; i<ITERS; i++) {
   1.561 +      test_andv(a0, a1, (short)BIT_MASK);
   1.562 +    }
   1.563 +    end = System.currentTimeMillis();
   1.564 +    System.out.println("test_andv: " + (end - start));
   1.565 +    start = System.currentTimeMillis();
   1.566 +    for (int i=0; i<ITERS; i++) {
   1.567 +      test_anda(a0, a1, a4);
   1.568 +    }
   1.569 +    end = System.currentTimeMillis();
   1.570 +    System.out.println("test_anda: " + (end - start));
   1.571 +
   1.572 +    start = System.currentTimeMillis();
   1.573 +    for (int i=0; i<ITERS; i++) {
   1.574 +      test_orc(a0, a1);
   1.575 +    }
   1.576 +    end = System.currentTimeMillis();
   1.577 +    System.out.println("test_orc: " + (end - start));
   1.578 +    start = System.currentTimeMillis();
   1.579 +    for (int i=0; i<ITERS; i++) {
   1.580 +      test_orv(a0, a1, (short)BIT_MASK);
   1.581 +    }
   1.582 +    end = System.currentTimeMillis();
   1.583 +    System.out.println("test_orv: " + (end - start));
   1.584 +    start = System.currentTimeMillis();
   1.585 +    for (int i=0; i<ITERS; i++) {
   1.586 +      test_ora(a0, a1, a4);
   1.587 +    }
   1.588 +    end = System.currentTimeMillis();
   1.589 +    System.out.println("test_ora: " + (end - start));
   1.590 +
   1.591 +    start = System.currentTimeMillis();
   1.592 +    for (int i=0; i<ITERS; i++) {
   1.593 +      test_xorc(a0, a1);
   1.594 +    }
   1.595 +    end = System.currentTimeMillis();
   1.596 +    System.out.println("test_xorc: " + (end - start));
   1.597 +    start = System.currentTimeMillis();
   1.598 +    for (int i=0; i<ITERS; i++) {
   1.599 +      test_xorv(a0, a1, (short)BIT_MASK);
   1.600 +    }
   1.601 +    end = System.currentTimeMillis();
   1.602 +    System.out.println("test_xorv: " + (end - start));
   1.603 +    start = System.currentTimeMillis();
   1.604 +    for (int i=0; i<ITERS; i++) {
   1.605 +      test_xora(a0, a1, a4);
   1.606 +    }
   1.607 +    end = System.currentTimeMillis();
   1.608 +    System.out.println("test_xora: " + (end - start));
   1.609 +
   1.610 +    start = System.currentTimeMillis();
   1.611 +    for (int i=0; i<ITERS; i++) {
   1.612 +      test_sllc(a0, a1);
   1.613 +    }
   1.614 +    end = System.currentTimeMillis();
   1.615 +    System.out.println("test_sllc: " + (end - start));
   1.616 +    start = System.currentTimeMillis();
   1.617 +    for (int i=0; i<ITERS; i++) {
   1.618 +      test_sllv(a0, a1, VALUE);
   1.619 +    }
   1.620 +    end = System.currentTimeMillis();
   1.621 +    System.out.println("test_sllv: " + (end - start));
   1.622 +
   1.623 +    start = System.currentTimeMillis();
   1.624 +    for (int i=0; i<ITERS; i++) {
   1.625 +      test_srlc(a0, a1);
   1.626 +    }
   1.627 +    end = System.currentTimeMillis();
   1.628 +    System.out.println("test_srlc: " + (end - start));
   1.629 +    start = System.currentTimeMillis();
   1.630 +    for (int i=0; i<ITERS; i++) {
   1.631 +      test_srlv(a0, a1, VALUE);
   1.632 +    }
   1.633 +    end = System.currentTimeMillis();
   1.634 +    System.out.println("test_srlv: " + (end - start));
   1.635 +
   1.636 +    start = System.currentTimeMillis();
   1.637 +    for (int i=0; i<ITERS; i++) {
   1.638 +      test_srac(a0, a1);
   1.639 +    }
   1.640 +    end = System.currentTimeMillis();
   1.641 +    System.out.println("test_srac: " + (end - start));
   1.642 +    start = System.currentTimeMillis();
   1.643 +    for (int i=0; i<ITERS; i++) {
   1.644 +      test_srav(a0, a1, VALUE);
   1.645 +    }
   1.646 +    end = System.currentTimeMillis();
   1.647 +    System.out.println("test_srav: " + (end - start));
   1.648 +
   1.649 +    start = System.currentTimeMillis();
   1.650 +    for (int i=0; i<ITERS; i++) {
   1.651 +      test_sllc_n(a0, a1);
   1.652 +    }
   1.653 +    end = System.currentTimeMillis();
   1.654 +    System.out.println("test_sllc_n: " + (end - start));
   1.655 +    start = System.currentTimeMillis();
   1.656 +    for (int i=0; i<ITERS; i++) {
   1.657 +      test_sllv(a0, a1, -VALUE);
   1.658 +    }
   1.659 +    end = System.currentTimeMillis();
   1.660 +    System.out.println("test_sllv_n: " + (end - start));
   1.661 +
   1.662 +    start = System.currentTimeMillis();
   1.663 +    for (int i=0; i<ITERS; i++) {
   1.664 +      test_srlc_n(a0, a1);
   1.665 +    }
   1.666 +    end = System.currentTimeMillis();
   1.667 +    System.out.println("test_srlc_n: " + (end - start));
   1.668 +    start = System.currentTimeMillis();
   1.669 +    for (int i=0; i<ITERS; i++) {
   1.670 +      test_srlv(a0, a1, -VALUE);
   1.671 +    }
   1.672 +    end = System.currentTimeMillis();
   1.673 +    System.out.println("test_srlv_n: " + (end - start));
   1.674 +
   1.675 +    start = System.currentTimeMillis();
   1.676 +    for (int i=0; i<ITERS; i++) {
   1.677 +      test_srac_n(a0, a1);
   1.678 +    }
   1.679 +    end = System.currentTimeMillis();
   1.680 +    System.out.println("test_srac_n: " + (end - start));
   1.681 +    start = System.currentTimeMillis();
   1.682 +    for (int i=0; i<ITERS; i++) {
   1.683 +      test_srav(a0, a1, -VALUE);
   1.684 +    }
   1.685 +    end = System.currentTimeMillis();
   1.686 +    System.out.println("test_srav_n: " + (end - start));
   1.687 +
   1.688 +    start = System.currentTimeMillis();
   1.689 +    for (int i=0; i<ITERS; i++) {
   1.690 +      test_sllc_o(a0, a1);
   1.691 +    }
   1.692 +    end = System.currentTimeMillis();
   1.693 +    System.out.println("test_sllc_o: " + (end - start));
   1.694 +    start = System.currentTimeMillis();
   1.695 +    for (int i=0; i<ITERS; i++) {
   1.696 +      test_sllv(a0, a1, SHIFT);
   1.697 +    }
   1.698 +    end = System.currentTimeMillis();
   1.699 +    System.out.println("test_sllv_o: " + (end - start));
   1.700 +
   1.701 +    start = System.currentTimeMillis();
   1.702 +    for (int i=0; i<ITERS; i++) {
   1.703 +      test_srlc_o(a0, a1);
   1.704 +    }
   1.705 +    end = System.currentTimeMillis();
   1.706 +    System.out.println("test_srlc_o: " + (end - start));
   1.707 +    start = System.currentTimeMillis();
   1.708 +    for (int i=0; i<ITERS; i++) {
   1.709 +      test_srlv(a0, a1, SHIFT);
   1.710 +    }
   1.711 +    end = System.currentTimeMillis();
   1.712 +    System.out.println("test_srlv_o: " + (end - start));
   1.713 +
   1.714 +    start = System.currentTimeMillis();
   1.715 +    for (int i=0; i<ITERS; i++) {
   1.716 +      test_srac_o(a0, a1);
   1.717 +    }
   1.718 +    end = System.currentTimeMillis();
   1.719 +    System.out.println("test_srac_o: " + (end - start));
   1.720 +    start = System.currentTimeMillis();
   1.721 +    for (int i=0; i<ITERS; i++) {
   1.722 +      test_srav(a0, a1, SHIFT);
   1.723 +    }
   1.724 +    end = System.currentTimeMillis();
   1.725 +    System.out.println("test_srav_o: " + (end - start));
   1.726 +
   1.727 +    start = System.currentTimeMillis();
   1.728 +    for (int i=0; i<ITERS; i++) {
   1.729 +      test_sllc_on(a0, a1);
   1.730 +    }
   1.731 +    end = System.currentTimeMillis();
   1.732 +    System.out.println("test_sllc_on: " + (end - start));
   1.733 +    start = System.currentTimeMillis();
   1.734 +    for (int i=0; i<ITERS; i++) {
   1.735 +      test_sllv(a0, a1, -SHIFT);
   1.736 +    }
   1.737 +    end = System.currentTimeMillis();
   1.738 +    System.out.println("test_sllv_on: " + (end - start));
   1.739 +
   1.740 +    start = System.currentTimeMillis();
   1.741 +    for (int i=0; i<ITERS; i++) {
   1.742 +      test_srlc_on(a0, a1);
   1.743 +    }
   1.744 +    end = System.currentTimeMillis();
   1.745 +    System.out.println("test_srlc_on: " + (end - start));
   1.746 +    start = System.currentTimeMillis();
   1.747 +    for (int i=0; i<ITERS; i++) {
   1.748 +      test_srlv(a0, a1, -SHIFT);
   1.749 +    }
   1.750 +    end = System.currentTimeMillis();
   1.751 +    System.out.println("test_srlv_on: " + (end - start));
   1.752 +
   1.753 +    start = System.currentTimeMillis();
   1.754 +    for (int i=0; i<ITERS; i++) {
   1.755 +      test_srac_on(a0, a1);
   1.756 +    }
   1.757 +    end = System.currentTimeMillis();
   1.758 +    System.out.println("test_srac_on: " + (end - start));
   1.759 +    start = System.currentTimeMillis();
   1.760 +    for (int i=0; i<ITERS; i++) {
   1.761 +      test_srav(a0, a1, -SHIFT);
   1.762 +    }
   1.763 +    end = System.currentTimeMillis();
   1.764 +    System.out.println("test_srav_on: " + (end - start));
   1.765 +
   1.766 +    start = System.currentTimeMillis();
   1.767 +    for (int i=0; i<ITERS; i++) {
   1.768 +      test_pack2(p2, a1);
   1.769 +    }
   1.770 +    end = System.currentTimeMillis();
   1.771 +    System.out.println("test_pack2: " + (end - start));
   1.772 +    start = System.currentTimeMillis();
   1.773 +    for (int i=0; i<ITERS; i++) {
   1.774 +      test_unpack2(a0, p2);
   1.775 +    }
   1.776 +    end = System.currentTimeMillis();
   1.777 +    System.out.println("test_unpack2: " + (end - start));
   1.778 +    start = System.currentTimeMillis();
   1.779 +    for (int i=0; i<ITERS; i++) {
   1.780 +      test_pack2_swap(p2, a1);
   1.781 +    }
   1.782 +    end = System.currentTimeMillis();
   1.783 +    System.out.println("test_pack2_swap: " + (end - start));
   1.784 +    start = System.currentTimeMillis();
   1.785 +    for (int i=0; i<ITERS; i++) {
   1.786 +      test_unpack2_swap(a0, p2);
   1.787 +    }
   1.788 +    end = System.currentTimeMillis();
   1.789 +    System.out.println("test_unpack2_swap: " + (end - start));
   1.790 +
   1.791 +    start = System.currentTimeMillis();
   1.792 +    for (int i=0; i<ITERS; i++) {
   1.793 +      test_pack4(p4, a1);
   1.794 +    }
   1.795 +    end = System.currentTimeMillis();
   1.796 +    System.out.println("test_pack4: " + (end - start));
   1.797 +    start = System.currentTimeMillis();
   1.798 +    for (int i=0; i<ITERS; i++) {
   1.799 +      test_unpack4(a0, p4);
   1.800 +    }
   1.801 +    end = System.currentTimeMillis();
   1.802 +    System.out.println("test_unpack4: " + (end - start));
   1.803 +    start = System.currentTimeMillis();
   1.804 +    for (int i=0; i<ITERS; i++) {
   1.805 +      test_pack4_swap(p4, a1);
   1.806 +    }
   1.807 +    end = System.currentTimeMillis();
   1.808 +    System.out.println("test_pack4_swap: " + (end - start));
   1.809 +    start = System.currentTimeMillis();
   1.810 +    for (int i=0; i<ITERS; i++) {
   1.811 +      test_unpack4_swap(a0, p4);
   1.812 +    }
   1.813 +    end = System.currentTimeMillis();
   1.814 +    System.out.println("test_unpack4_swap: " + (end - start));
   1.815 +
   1.816 +    return errn;
   1.817 +  }
   1.818 +
   1.819 +  static int test_sum(short[] a1) {
   1.820 +    int sum = 0;
   1.821 +    for (int i = 0; i < a1.length; i+=1) {
   1.822 +      sum += a1[i];
   1.823 +    }
   1.824 +    return sum;
   1.825 +  }
   1.826 +
   1.827 +  static void test_addc(short[] a0, short[] a1) {
   1.828 +    for (int i = 0; i < a0.length; i+=1) {
   1.829 +      a0[i] = (short)(a1[i]+VALUE);
   1.830 +    }
   1.831 +  }
   1.832 +  static void test_addv(short[] a0, short[] a1, short b) {
   1.833 +    for (int i = 0; i < a0.length; i+=1) {
   1.834 +      a0[i] = (short)(a1[i]+b);
   1.835 +    }
   1.836 +  }
   1.837 +  static void test_adda(short[] a0, short[] a1, short[] a2) {
   1.838 +    for (int i = 0; i < a0.length; i+=1) {
   1.839 +      a0[i] = (short)(a1[i]+a2[i]);
   1.840 +    }
   1.841 +  }
   1.842 +
   1.843 +  static void test_subc(short[] a0, short[] a1) {
   1.844 +    for (int i = 0; i < a0.length; i+=1) {
   1.845 +      a0[i] = (short)(a1[i]-VALUE);
   1.846 +    }
   1.847 +  }
   1.848 +  static void test_subv(short[] a0, short[] a1, short b) {
   1.849 +    for (int i = 0; i < a0.length; i+=1) {
   1.850 +      a0[i] = (short)(a1[i]-b);
   1.851 +    }
   1.852 +  }
   1.853 +  static void test_suba(short[] a0, short[] a1, short[] a2) {
   1.854 +    for (int i = 0; i < a0.length; i+=1) {
   1.855 +      a0[i] = (short)(a1[i]-a2[i]);
   1.856 +    }
   1.857 +  }
   1.858 +
   1.859 +  static void test_mulc(short[] a0, short[] a1) {
   1.860 +    for (int i = 0; i < a0.length; i+=1) {
   1.861 +      a0[i] = (short)(a1[i]*VALUE);
   1.862 +    }
   1.863 +  }
   1.864 +  static void test_mulc_n(short[] a0, short[] a1) {
   1.865 +    for (int i = 0; i < a0.length; i+=1) {
   1.866 +      a0[i] = (short)(a1[i]*(-VALUE));
   1.867 +    }
   1.868 +  }
   1.869 +  static void test_mulv(short[] a0, short[] a1, short b) {
   1.870 +    for (int i = 0; i < a0.length; i+=1) {
   1.871 +      a0[i] = (short)(a1[i]*b);
   1.872 +    }
   1.873 +  }
   1.874 +  static void test_mula(short[] a0, short[] a1, short[] a2) {
   1.875 +    for (int i = 0; i < a0.length; i+=1) {
   1.876 +      a0[i] = (short)(a1[i]*a2[i]);
   1.877 +    }
   1.878 +  }
   1.879 +
   1.880 +  static void test_divc(short[] a0, short[] a1) {
   1.881 +    for (int i = 0; i < a0.length; i+=1) {
   1.882 +      a0[i] = (short)(a1[i]/VALUE);
   1.883 +    }
   1.884 +  }
   1.885 +  static void test_divc_n(short[] a0, short[] a1) {
   1.886 +    for (int i = 0; i < a0.length; i+=1) {
   1.887 +      a0[i] = (short)(a1[i]/(-VALUE));
   1.888 +    }
   1.889 +  }
   1.890 +  static void test_divv(short[] a0, short[] a1, short b) {
   1.891 +    for (int i = 0; i < a0.length; i+=1) {
   1.892 +      a0[i] = (short)(a1[i]/b);
   1.893 +    }
   1.894 +  }
   1.895 +  static void test_diva(short[] a0, short[] a1, short[] a2) {
   1.896 +    for (int i = 0; i < a0.length; i+=1) {
   1.897 +      a0[i] = (short)(a1[i]/a2[i]);
   1.898 +    }
   1.899 +  }
   1.900 +
   1.901 +  static void test_andc(short[] a0, short[] a1) {
   1.902 +    for (int i = 0; i < a0.length; i+=1) {
   1.903 +      a0[i] = (short)(a1[i]&BIT_MASK);
   1.904 +    }
   1.905 +  }
   1.906 +  static void test_andv(short[] a0, short[] a1, short b) {
   1.907 +    for (int i = 0; i < a0.length; i+=1) {
   1.908 +      a0[i] = (short)(a1[i]&b);
   1.909 +    }
   1.910 +  }
   1.911 +  static void test_anda(short[] a0, short[] a1, short[] a2) {
   1.912 +    for (int i = 0; i < a0.length; i+=1) {
   1.913 +      a0[i] = (short)(a1[i]&a2[i]);
   1.914 +    }
   1.915 +  }
   1.916 +
   1.917 +  static void test_orc(short[] a0, short[] a1) {
   1.918 +    for (int i = 0; i < a0.length; i+=1) {
   1.919 +      a0[i] = (short)(a1[i]|BIT_MASK);
   1.920 +    }
   1.921 +  }
   1.922 +  static void test_orv(short[] a0, short[] a1, short b) {
   1.923 +    for (int i = 0; i < a0.length; i+=1) {
   1.924 +      a0[i] = (short)(a1[i]|b);
   1.925 +    }
   1.926 +  }
   1.927 +  static void test_ora(short[] a0, short[] a1, short[] a2) {
   1.928 +    for (int i = 0; i < a0.length; i+=1) {
   1.929 +      a0[i] = (short)(a1[i]|a2[i]);
   1.930 +    }
   1.931 +  }
   1.932 +
   1.933 +  static void test_xorc(short[] a0, short[] a1) {
   1.934 +    for (int i = 0; i < a0.length; i+=1) {
   1.935 +      a0[i] = (short)(a1[i]^BIT_MASK);
   1.936 +    }
   1.937 +  }
   1.938 +  static void test_xorv(short[] a0, short[] a1, short b) {
   1.939 +    for (int i = 0; i < a0.length; i+=1) {
   1.940 +      a0[i] = (short)(a1[i]^b);
   1.941 +    }
   1.942 +  }
   1.943 +  static void test_xora(short[] a0, short[] a1, short[] a2) {
   1.944 +    for (int i = 0; i < a0.length; i+=1) {
   1.945 +      a0[i] = (short)(a1[i]^a2[i]);
   1.946 +    }
   1.947 +  }
   1.948 +
   1.949 +  static void test_sllc(short[] a0, short[] a1) {
   1.950 +    for (int i = 0; i < a0.length; i+=1) {
   1.951 +      a0[i] = (short)(a1[i]<<VALUE);
   1.952 +    }
   1.953 +  }
   1.954 +  static void test_sllc_n(short[] a0, short[] a1) {
   1.955 +    for (int i = 0; i < a0.length; i+=1) {
   1.956 +      a0[i] = (short)(a1[i]<<(-VALUE));
   1.957 +    }
   1.958 +  }
   1.959 +  static void test_sllc_o(short[] a0, short[] a1) {
   1.960 +    for (int i = 0; i < a0.length; i+=1) {
   1.961 +      a0[i] = (short)(a1[i]<<SHIFT);
   1.962 +    }
   1.963 +  }
   1.964 +  static void test_sllc_on(short[] a0, short[] a1) {
   1.965 +    for (int i = 0; i < a0.length; i+=1) {
   1.966 +      a0[i] = (short)(a1[i]<<(-SHIFT));
   1.967 +    }
   1.968 +  }
   1.969 +  static void test_sllv(short[] a0, short[] a1, int b) {
   1.970 +    for (int i = 0; i < a0.length; i+=1) {
   1.971 +      a0[i] = (short)(a1[i]<<b);
   1.972 +    }
   1.973 +  }
   1.974 +
   1.975 +  static void test_srlc(short[] a0, short[] a1) {
   1.976 +    for (int i = 0; i < a0.length; i+=1) {
   1.977 +      a0[i] = (short)(a1[i]>>>VALUE);
   1.978 +    }
   1.979 +  }
   1.980 +  static void test_srlc_n(short[] a0, short[] a1) {
   1.981 +    for (int i = 0; i < a0.length; i+=1) {
   1.982 +      a0[i] = (short)(a1[i]>>>(-VALUE));
   1.983 +    }
   1.984 +  }
   1.985 +  static void test_srlc_o(short[] a0, short[] a1) {
   1.986 +    for (int i = 0; i < a0.length; i+=1) {
   1.987 +      a0[i] = (short)(a1[i]>>>SHIFT);
   1.988 +    }
   1.989 +  }
   1.990 +  static void test_srlc_on(short[] a0, short[] a1) {
   1.991 +    for (int i = 0; i < a0.length; i+=1) {
   1.992 +      a0[i] = (short)(a1[i]>>>(-SHIFT));
   1.993 +    }
   1.994 +  }
   1.995 +  static void test_srlv(short[] a0, short[] a1, int b) {
   1.996 +    for (int i = 0; i < a0.length; i+=1) {
   1.997 +      a0[i] = (short)(a1[i]>>>b);
   1.998 +    }
   1.999 +  }
  1.1000 +
  1.1001 +  static void test_srac(short[] a0, short[] a1) {
  1.1002 +    for (int i = 0; i < a0.length; i+=1) {
  1.1003 +      a0[i] = (short)(a1[i]>>VALUE);
  1.1004 +    }
  1.1005 +  }
  1.1006 +  static void test_srac_n(short[] a0, short[] a1) {
  1.1007 +    for (int i = 0; i < a0.length; i+=1) {
  1.1008 +      a0[i] = (short)(a1[i]>>(-VALUE));
  1.1009 +    }
  1.1010 +  }
  1.1011 +  static void test_srac_o(short[] a0, short[] a1) {
  1.1012 +    for (int i = 0; i < a0.length; i+=1) {
  1.1013 +      a0[i] = (short)(a1[i]>>SHIFT);
  1.1014 +    }
  1.1015 +  }
  1.1016 +  static void test_srac_on(short[] a0, short[] a1) {
  1.1017 +    for (int i = 0; i < a0.length; i+=1) {
  1.1018 +      a0[i] = (short)(a1[i]>>(-SHIFT));
  1.1019 +    }
  1.1020 +  }
  1.1021 +  static void test_srav(short[] a0, short[] a1, int b) {
  1.1022 +    for (int i = 0; i < a0.length; i+=1) {
  1.1023 +      a0[i] = (short)(a1[i]>>b);
  1.1024 +    }
  1.1025 +  }
  1.1026 +
  1.1027 +  static void test_pack2(int[] p2, short[] a1) {
  1.1028 +    if (p2.length*2 > a1.length) return;
  1.1029 +    for (int i = 0; i < p2.length; i+=1) {
  1.1030 +      int l0 = (int)a1[i*2+0];
  1.1031 +      int l1 = (int)a1[i*2+1];
  1.1032 +      p2[i] = (l1 << 16) | (l0 & 0xFFFF);
  1.1033 +    }
  1.1034 +  }
  1.1035 +  static void test_unpack2(short[] a0, int[] p2) {
  1.1036 +    if (p2.length*2 > a0.length) return;
  1.1037 +    for (int i = 0; i < p2.length; i+=1) {
  1.1038 +      int l = p2[i];
  1.1039 +      a0[i*2+0] = (short)(l & 0xFFFF);
  1.1040 +      a0[i*2+1] = (short)(l >> 16);
  1.1041 +    }
  1.1042 +  }
  1.1043 +  static void test_pack2_swap(int[] p2, short[] a1) {
  1.1044 +    if (p2.length*2 > a1.length) return;
  1.1045 +    for (int i = 0; i < p2.length; i+=1) {
  1.1046 +      int l0 = (int)a1[i*2+0];
  1.1047 +      int l1 = (int)a1[i*2+1];
  1.1048 +      p2[i] = (l0 << 16) | (l1 & 0xFFFF);
  1.1049 +    }
  1.1050 +  }
  1.1051 +  static void test_unpack2_swap(short[] a0, int[] p2) {
  1.1052 +    if (p2.length*2 > a0.length) return;
  1.1053 +    for (int i = 0; i < p2.length; i+=1) {
  1.1054 +      int l = p2[i];
  1.1055 +      a0[i*2+0] = (short)(l >> 16);
  1.1056 +      a0[i*2+1] = (short)(l & 0xFFFF);
  1.1057 +    }
  1.1058 +  }
  1.1059 +
  1.1060 +  static void test_pack4(long[] p4, short[] a1) {
  1.1061 +    if (p4.length*4 > a1.length) return;
  1.1062 +    for (int i = 0; i < p4.length; i+=1) {
  1.1063 +      long l0 = (long)a1[i*4+0];
  1.1064 +      long l1 = (long)a1[i*4+1];
  1.1065 +      long l2 = (long)a1[i*4+2];
  1.1066 +      long l3 = (long)a1[i*4+3];
  1.1067 +      p4[i] = (l0 & 0xFFFFl) |
  1.1068 +             ((l1 & 0xFFFFl) << 16) |
  1.1069 +             ((l2 & 0xFFFFl) << 32) |
  1.1070 +             ((l3 & 0xFFFFl) << 48);
  1.1071 +    }
  1.1072 +  }
  1.1073 +  static void test_unpack4(short[] a0, long[] p4) {
  1.1074 +    if (p4.length*4 > a0.length) return;
  1.1075 +    for (int i = 0; i < p4.length; i+=1) {
  1.1076 +      long l = p4[i];
  1.1077 +      a0[i*4+0] = (short)(l & 0xFFFFl);
  1.1078 +      a0[i*4+1] = (short)(l >> 16);
  1.1079 +      a0[i*4+2] = (short)(l >> 32);
  1.1080 +      a0[i*4+3] = (short)(l >> 48);
  1.1081 +    }
  1.1082 +  }
  1.1083 +  static void test_pack4_swap(long[] p4, short[] a1) {
  1.1084 +    if (p4.length*4 > a1.length) return;
  1.1085 +    for (int i = 0; i < p4.length; i+=1) {
  1.1086 +      long l0 = (long)a1[i*4+0];
  1.1087 +      long l1 = (long)a1[i*4+1];
  1.1088 +      long l2 = (long)a1[i*4+2];
  1.1089 +      long l3 = (long)a1[i*4+3];
  1.1090 +      p4[i] = (l3 & 0xFFFFl) |
  1.1091 +             ((l2 & 0xFFFFl) << 16) |
  1.1092 +             ((l1 & 0xFFFFl) << 32) |
  1.1093 +             ((l0 & 0xFFFFl) << 48);
  1.1094 +    }
  1.1095 +  }
  1.1096 +  static void test_unpack4_swap(short[] a0, long[] p4) {
  1.1097 +    if (p4.length*4 > a0.length) return;
  1.1098 +    for (int i = 0; i < p4.length; i+=1) {
  1.1099 +      long l = p4[i];
  1.1100 +      a0[i*4+0] = (short)(l >> 48);
  1.1101 +      a0[i*4+1] = (short)(l >> 32);
  1.1102 +      a0[i*4+2] = (short)(l >> 16);
  1.1103 +      a0[i*4+3] = (short)(l & 0xFFFFl);
  1.1104 +    }
  1.1105 +  }
  1.1106 +
  1.1107 +  static int verify(String text, int i, short elem, short val) {
  1.1108 +    if (elem != val) {
  1.1109 +      System.err.println(text + "[" + i + "] = " + elem + " != " + val);
  1.1110 +      return 1;
  1.1111 +    }
  1.1112 +    return 0;
  1.1113 +  }
  1.1114 +
  1.1115 +  static int verify(String text, int i, int elem, int val) {
  1.1116 +    if (elem != val) {
  1.1117 +      System.err.println(text + "[" + i + "] = " + Integer.toHexString(elem) + " != " + Integer.toHexString(val));
  1.1118 +      return 1;
  1.1119 +    }
  1.1120 +    return 0;
  1.1121 +  }
  1.1122 +
  1.1123 +  static int verify(String text, int i, long elem, long val) {
  1.1124 +    if (elem != val) {
  1.1125 +      System.err.println(text + "[" + i + "] = " + Long.toHexString(elem) + " != " + Long.toHexString(val));
  1.1126 +      return 1;
  1.1127 +    }
  1.1128 +    return 0;
  1.1129 +  }
  1.1130 +}

mercurial