test/compiler/6340864/TestLongVect.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/TestLongVect.java	Mon Aug 20 09:07:21 2012 -0700
     1.3 @@ -0,0 +1,917 @@
     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 TestLongVect
    1.34 + */
    1.35 +
    1.36 +public class TestLongVect {
    1.37 +  private static final int ARRLEN = 997;
    1.38 +  private static final int ITERS  = 11000;
    1.39 +  private static final long ADD_INIT = Long.MAX_VALUE-500;
    1.40 +  private static final long BIT_MASK = 0xEC80F731EC80F731L;
    1.41 +  private static final int VALUE = 31;
    1.42 +  private static final int SHIFT = 64;
    1.43 +
    1.44 +  public static void main(String args[]) {
    1.45 +    System.out.println("Testing Long 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 +    long[] a0 = new long[ARRLEN];
    1.56 +    long[] a1 = new long[ARRLEN];
    1.57 +    long[] a2 = new long[ARRLEN];
    1.58 +    long[] a3 = new long[ARRLEN];
    1.59 +    long[] a4 = new long[ARRLEN];
    1.60 +    // Initialize
    1.61 +    long gold_sum = 0;
    1.62 +    for (int i=0; i<ARRLEN; i++) {
    1.63 +      long val = (long)(ADD_INIT+i);
    1.64 +      gold_sum += val;
    1.65 +      a1[i] = val;
    1.66 +      a2[i] = (long)VALUE;
    1.67 +      a3[i] = (long)-VALUE;
    1.68 +      a4[i] = (long)BIT_MASK;
    1.69 +    }
    1.70 +    System.out.println("Warmup");
    1.71 +    for (int i=0; i<ITERS; i++) {
    1.72 +      test_sum(a1);
    1.73 +      test_addc(a0, a1);
    1.74 +      test_addv(a0, a1, (long)VALUE);
    1.75 +      test_adda(a0, a1, a2);
    1.76 +      test_subc(a0, a1);
    1.77 +      test_subv(a0, a1, (long)VALUE);
    1.78 +      test_suba(a0, a1, a2);
    1.79 +      test_mulc(a0, a1);
    1.80 +      test_mulv(a0, a1, (long)VALUE);
    1.81 +      test_mula(a0, a1, a2);
    1.82 +      test_divc(a0, a1);
    1.83 +      test_divv(a0, a1, (long)VALUE);
    1.84 +      test_diva(a0, a1, a2);
    1.85 +      test_mulc_n(a0, a1);
    1.86 +      test_mulv(a0, a1, (long)-VALUE);
    1.87 +      test_mula(a0, a1, a3);
    1.88 +      test_divc_n(a0, a1);
    1.89 +      test_divv(a0, a1, (long)-VALUE);
    1.90 +      test_diva(a0, a1, a3);
    1.91 +      test_andc(a0, a1);
    1.92 +      test_andv(a0, a1, (long)BIT_MASK);
    1.93 +      test_anda(a0, a1, a4);
    1.94 +      test_orc(a0, a1);
    1.95 +      test_orv(a0, a1, (long)BIT_MASK);
    1.96 +      test_ora(a0, a1, a4);
    1.97 +      test_xorc(a0, a1);
    1.98 +      test_xorv(a0, a1, (long)BIT_MASK);
    1.99 +      test_xora(a0, a1, a4);
   1.100 +      test_sllc(a0, a1);
   1.101 +      test_sllv(a0, a1, VALUE);
   1.102 +      test_srlc(a0, a1);
   1.103 +      test_srlv(a0, a1, VALUE);
   1.104 +      test_srac(a0, a1);
   1.105 +      test_srav(a0, a1, VALUE);
   1.106 +      test_sllc_n(a0, a1);
   1.107 +      test_sllv(a0, a1, -VALUE);
   1.108 +      test_srlc_n(a0, a1);
   1.109 +      test_srlv(a0, a1, -VALUE);
   1.110 +      test_srac_n(a0, a1);
   1.111 +      test_srav(a0, a1, -VALUE);
   1.112 +      test_sllc_o(a0, a1);
   1.113 +      test_sllv(a0, a1, SHIFT);
   1.114 +      test_srlc_o(a0, a1);
   1.115 +      test_srlv(a0, a1, SHIFT);
   1.116 +      test_srac_o(a0, a1);
   1.117 +      test_srav(a0, a1, SHIFT);
   1.118 +      test_sllc_on(a0, a1);
   1.119 +      test_sllv(a0, a1, -SHIFT);
   1.120 +      test_srlc_on(a0, a1);
   1.121 +      test_srlv(a0, a1, -SHIFT);
   1.122 +      test_srac_on(a0, a1);
   1.123 +      test_srav(a0, a1, -SHIFT);
   1.124 +    }
   1.125 +    // Test and verify results
   1.126 +    System.out.println("Verification");
   1.127 +    int errn = 0;
   1.128 +    {
   1.129 +      long sum = test_sum(a1);
   1.130 +      if (sum != gold_sum) {
   1.131 +        System.err.println("test_sum:  " + sum + " != " + gold_sum);
   1.132 +        errn++;
   1.133 +      }
   1.134 +
   1.135 +      test_addc(a0, a1);
   1.136 +      for (int i=0; i<ARRLEN; i++) {
   1.137 +        errn += verify("test_addc: ", i, a0[i], (long)((long)(ADD_INIT+i)+VALUE));
   1.138 +      }
   1.139 +      test_addv(a0, a1, (long)VALUE);
   1.140 +      for (int i=0; i<ARRLEN; i++) {
   1.141 +        errn += verify("test_addv: ", i, a0[i], (long)((long)(ADD_INIT+i)+VALUE));
   1.142 +      }
   1.143 +      test_adda(a0, a1, a2);
   1.144 +      for (int i=0; i<ARRLEN; i++) {
   1.145 +        errn += verify("test_adda: ", i, a0[i], (long)((long)(ADD_INIT+i)+VALUE));
   1.146 +      }
   1.147 +
   1.148 +      test_subc(a0, a1);
   1.149 +      for (int i=0; i<ARRLEN; i++) {
   1.150 +        errn += verify("test_subc: ", i, a0[i], (long)((long)(ADD_INIT+i)-VALUE));
   1.151 +      }
   1.152 +      test_subv(a0, a1, (long)VALUE);
   1.153 +      for (int i=0; i<ARRLEN; i++) {
   1.154 +        errn += verify("test_subv: ", i, a0[i], (long)((long)(ADD_INIT+i)-VALUE));
   1.155 +      }
   1.156 +      test_suba(a0, a1, a2);
   1.157 +      for (int i=0; i<ARRLEN; i++) {
   1.158 +        errn += verify("test_suba: ", i, a0[i], (long)((long)(ADD_INIT+i)-VALUE));
   1.159 +      }
   1.160 +
   1.161 +      test_mulc(a0, a1);
   1.162 +      for (int i=0; i<ARRLEN; i++) {
   1.163 +        errn += verify("test_mulc: ", i, a0[i], (long)((long)(ADD_INIT+i)*VALUE));
   1.164 +      }
   1.165 +      test_mulv(a0, a1, (long)VALUE);
   1.166 +      for (int i=0; i<ARRLEN; i++) {
   1.167 +        errn += verify("test_mulv: ", i, a0[i], (long)((long)(ADD_INIT+i)*VALUE));
   1.168 +      }
   1.169 +      test_mula(a0, a1, a2);
   1.170 +      for (int i=0; i<ARRLEN; i++) {
   1.171 +        errn += verify("test_mula: ", i, a0[i], (long)((long)(ADD_INIT+i)*VALUE));
   1.172 +      }
   1.173 +
   1.174 +      test_divc(a0, a1);
   1.175 +      for (int i=0; i<ARRLEN; i++) {
   1.176 +        errn += verify("test_divc: ", i, a0[i], (long)((long)(ADD_INIT+i)/VALUE));
   1.177 +      }
   1.178 +      test_divv(a0, a1, (long)VALUE);
   1.179 +      for (int i=0; i<ARRLEN; i++) {
   1.180 +        errn += verify("test_divv: ", i, a0[i], (long)((long)(ADD_INIT+i)/VALUE));
   1.181 +      }
   1.182 +      test_diva(a0, a1, a2);
   1.183 +      for (int i=0; i<ARRLEN; i++) {
   1.184 +        errn += verify("test_diva: ", i, a0[i], (long)((long)(ADD_INIT+i)/VALUE));
   1.185 +      }
   1.186 +
   1.187 +      test_mulc_n(a0, a1);
   1.188 +      for (int i=0; i<ARRLEN; i++) {
   1.189 +        errn += verify("test_mulc_n: ", i, a0[i], (long)((long)(ADD_INIT+i)*(-VALUE)));
   1.190 +      }
   1.191 +      test_mulv(a0, a1, (long)-VALUE);
   1.192 +      for (int i=0; i<ARRLEN; i++) {
   1.193 +        errn += verify("test_mulv_n: ", i, a0[i], (long)((long)(ADD_INIT+i)*(-VALUE)));
   1.194 +      }
   1.195 +      test_mula(a0, a1, a3);
   1.196 +      for (int i=0; i<ARRLEN; i++) {
   1.197 +        errn += verify("test_mula_n: ", i, a0[i], (long)((long)(ADD_INIT+i)*(-VALUE)));
   1.198 +      }
   1.199 +
   1.200 +      test_divc_n(a0, a1);
   1.201 +      for (int i=0; i<ARRLEN; i++) {
   1.202 +        errn += verify("test_divc_n: ", i, a0[i], (long)((long)(ADD_INIT+i)/(-VALUE)));
   1.203 +      }
   1.204 +      test_divv(a0, a1, (long)-VALUE);
   1.205 +      for (int i=0; i<ARRLEN; i++) {
   1.206 +        errn += verify("test_divv_n: ", i, a0[i], (long)((long)(ADD_INIT+i)/(-VALUE)));
   1.207 +      }
   1.208 +      test_diva(a0, a1, a3);
   1.209 +      for (int i=0; i<ARRLEN; i++) {
   1.210 +        errn += verify("test_diva_n: ", i, a0[i], (long)((long)(ADD_INIT+i)/(-VALUE)));
   1.211 +      }
   1.212 +
   1.213 +      test_andc(a0, a1);
   1.214 +      for (int i=0; i<ARRLEN; i++) {
   1.215 +        errn += verify("test_andc: ", i, a0[i], (long)((long)(ADD_INIT+i)&BIT_MASK));
   1.216 +      }
   1.217 +      test_andv(a0, a1, (long)BIT_MASK);
   1.218 +      for (int i=0; i<ARRLEN; i++) {
   1.219 +        errn += verify("test_andv: ", i, a0[i], (long)((long)(ADD_INIT+i)&BIT_MASK));
   1.220 +      }
   1.221 +      test_anda(a0, a1, a4);
   1.222 +      for (int i=0; i<ARRLEN; i++) {
   1.223 +        errn += verify("test_anda: ", i, a0[i], (long)((long)(ADD_INIT+i)&BIT_MASK));
   1.224 +      }
   1.225 +
   1.226 +      test_orc(a0, a1);
   1.227 +      for (int i=0; i<ARRLEN; i++) {
   1.228 +        errn += verify("test_orc: ", i, a0[i], (long)((long)(ADD_INIT+i)|BIT_MASK));
   1.229 +      }
   1.230 +      test_orv(a0, a1, (long)BIT_MASK);
   1.231 +      for (int i=0; i<ARRLEN; i++) {
   1.232 +        errn += verify("test_orv: ", i, a0[i], (long)((long)(ADD_INIT+i)|BIT_MASK));
   1.233 +      }
   1.234 +      test_ora(a0, a1, a4);
   1.235 +      for (int i=0; i<ARRLEN; i++) {
   1.236 +        errn += verify("test_ora: ", i, a0[i], (long)((long)(ADD_INIT+i)|BIT_MASK));
   1.237 +      }
   1.238 +
   1.239 +      test_xorc(a0, a1);
   1.240 +      for (int i=0; i<ARRLEN; i++) {
   1.241 +        errn += verify("test_xorc: ", i, a0[i], (long)((long)(ADD_INIT+i)^BIT_MASK));
   1.242 +      }
   1.243 +      test_xorv(a0, a1, (long)BIT_MASK);
   1.244 +      for (int i=0; i<ARRLEN; i++) {
   1.245 +        errn += verify("test_xorv: ", i, a0[i], (long)((long)(ADD_INIT+i)^BIT_MASK));
   1.246 +      }
   1.247 +      test_xora(a0, a1, a4);
   1.248 +      for (int i=0; i<ARRLEN; i++) {
   1.249 +        errn += verify("test_xora: ", i, a0[i], (long)((long)(ADD_INIT+i)^BIT_MASK));
   1.250 +      }
   1.251 +
   1.252 +      test_sllc(a0, a1);
   1.253 +      for (int i=0; i<ARRLEN; i++) {
   1.254 +        errn += verify("test_sllc: ", i, a0[i], (long)((long)(ADD_INIT+i)<<VALUE));
   1.255 +      }
   1.256 +      test_sllv(a0, a1, VALUE);
   1.257 +      for (int i=0; i<ARRLEN; i++) {
   1.258 +        errn += verify("test_sllv: ", i, a0[i], (long)((long)(ADD_INIT+i)<<VALUE));
   1.259 +      }
   1.260 +
   1.261 +      test_srlc(a0, a1);
   1.262 +      for (int i=0; i<ARRLEN; i++) {
   1.263 +        errn += verify("test_srlc: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>VALUE));
   1.264 +      }
   1.265 +      test_srlv(a0, a1, VALUE);
   1.266 +      for (int i=0; i<ARRLEN; i++) {
   1.267 +        errn += verify("test_srlv: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>VALUE));
   1.268 +      }
   1.269 +
   1.270 +      test_srac(a0, a1);
   1.271 +      for (int i=0; i<ARRLEN; i++) {
   1.272 +        errn += verify("test_srac: ", i, a0[i], (long)((long)(ADD_INIT+i)>>VALUE));
   1.273 +      }
   1.274 +      test_srav(a0, a1, VALUE);
   1.275 +      for (int i=0; i<ARRLEN; i++) {
   1.276 +        errn += verify("test_srav: ", i, a0[i], (long)((long)(ADD_INIT+i)>>VALUE));
   1.277 +      }
   1.278 +
   1.279 +      test_sllc_n(a0, a1);
   1.280 +      for (int i=0; i<ARRLEN; i++) {
   1.281 +        errn += verify("test_sllc_n: ", i, a0[i], (long)((long)(ADD_INIT+i)<<(-VALUE)));
   1.282 +      }
   1.283 +      test_sllv(a0, a1, -VALUE);
   1.284 +      for (int i=0; i<ARRLEN; i++) {
   1.285 +        errn += verify("test_sllv_n: ", i, a0[i], (long)((long)(ADD_INIT+i)<<(-VALUE)));
   1.286 +      }
   1.287 +
   1.288 +      test_srlc_n(a0, a1);
   1.289 +      for (int i=0; i<ARRLEN; i++) {
   1.290 +        errn += verify("test_srlc_n: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>(-VALUE)));
   1.291 +      }
   1.292 +      test_srlv(a0, a1, -VALUE);
   1.293 +      for (int i=0; i<ARRLEN; i++) {
   1.294 +        errn += verify("test_srlv_n: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>(-VALUE)));
   1.295 +      }
   1.296 +
   1.297 +      test_srac_n(a0, a1);
   1.298 +      for (int i=0; i<ARRLEN; i++) {
   1.299 +        errn += verify("test_srac_n: ", i, a0[i], (long)((long)(ADD_INIT+i)>>(-VALUE)));
   1.300 +      }
   1.301 +      test_srav(a0, a1, -VALUE);
   1.302 +      for (int i=0; i<ARRLEN; i++) {
   1.303 +        errn += verify("test_srav_n: ", i, a0[i], (long)((long)(ADD_INIT+i)>>(-VALUE)));
   1.304 +      }
   1.305 +
   1.306 +      test_sllc_o(a0, a1);
   1.307 +      for (int i=0; i<ARRLEN; i++) {
   1.308 +        errn += verify("test_sllc_o: ", i, a0[i], (long)((long)(ADD_INIT+i)<<SHIFT));
   1.309 +      }
   1.310 +      test_sllv(a0, a1, SHIFT);
   1.311 +      for (int i=0; i<ARRLEN; i++) {
   1.312 +        errn += verify("test_sllv_o: ", i, a0[i], (long)((long)(ADD_INIT+i)<<SHIFT));
   1.313 +      }
   1.314 +
   1.315 +      test_srlc_o(a0, a1);
   1.316 +      for (int i=0; i<ARRLEN; i++) {
   1.317 +        errn += verify("test_srlc_o: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>SHIFT));
   1.318 +      }
   1.319 +      test_srlv(a0, a1, SHIFT);
   1.320 +      for (int i=0; i<ARRLEN; i++) {
   1.321 +        errn += verify("test_srlv_o: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>SHIFT));
   1.322 +      }
   1.323 +
   1.324 +      test_srac_o(a0, a1);
   1.325 +      for (int i=0; i<ARRLEN; i++) {
   1.326 +        errn += verify("test_srac_o: ", i, a0[i], (long)((long)(ADD_INIT+i)>>SHIFT));
   1.327 +      }
   1.328 +      test_srav(a0, a1, SHIFT);
   1.329 +      for (int i=0; i<ARRLEN; i++) {
   1.330 +        errn += verify("test_srav_o: ", i, a0[i], (long)((long)(ADD_INIT+i)>>SHIFT));
   1.331 +      }
   1.332 +
   1.333 +      test_sllc_on(a0, a1);
   1.334 +      for (int i=0; i<ARRLEN; i++) {
   1.335 +        errn += verify("test_sllc_on: ", i, a0[i], (long)((long)(ADD_INIT+i)<<(-SHIFT)));
   1.336 +      }
   1.337 +      test_sllv(a0, a1, -SHIFT);
   1.338 +      for (int i=0; i<ARRLEN; i++) {
   1.339 +        errn += verify("test_sllv_on: ", i, a0[i], (long)((long)(ADD_INIT+i)<<(-SHIFT)));
   1.340 +      }
   1.341 +
   1.342 +      test_srlc_on(a0, a1);
   1.343 +      for (int i=0; i<ARRLEN; i++) {
   1.344 +        errn += verify("test_srlc_on: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>(-SHIFT)));
   1.345 +      }
   1.346 +      test_srlv(a0, a1, -SHIFT);
   1.347 +      for (int i=0; i<ARRLEN; i++) {
   1.348 +        errn += verify("test_srlv_on: ", i, a0[i], (long)((long)(ADD_INIT+i)>>>(-SHIFT)));
   1.349 +      }
   1.350 +
   1.351 +      test_srac_on(a0, a1);
   1.352 +      for (int i=0; i<ARRLEN; i++) {
   1.353 +        errn += verify("test_srac_on: ", i, a0[i], (long)((long)(ADD_INIT+i)>>(-SHIFT)));
   1.354 +      }
   1.355 +      test_srav(a0, a1, -SHIFT);
   1.356 +      for (int i=0; i<ARRLEN; i++) {
   1.357 +        errn += verify("test_srav_on: ", i, a0[i], (long)((long)(ADD_INIT+i)>>(-SHIFT)));
   1.358 +      }
   1.359 +
   1.360 +    }
   1.361 +
   1.362 +    if (errn > 0)
   1.363 +      return errn;
   1.364 +
   1.365 +    System.out.println("Time");
   1.366 +    long start, end;
   1.367 +
   1.368 +    start = System.currentTimeMillis();
   1.369 +    for (int i=0; i<ITERS; i++) {
   1.370 +      test_sum(a1);
   1.371 +    }
   1.372 +    end = System.currentTimeMillis();
   1.373 +    System.out.println("test_sum: " + (end - start));
   1.374 +
   1.375 +    start = System.currentTimeMillis();
   1.376 +    for (int i=0; i<ITERS; i++) {
   1.377 +      test_addc(a0, a1);
   1.378 +    }
   1.379 +    end = System.currentTimeMillis();
   1.380 +    System.out.println("test_addc: " + (end - start));
   1.381 +    start = System.currentTimeMillis();
   1.382 +    for (int i=0; i<ITERS; i++) {
   1.383 +      test_addv(a0, a1, (long)VALUE);
   1.384 +    }
   1.385 +    end = System.currentTimeMillis();
   1.386 +    System.out.println("test_addv: " + (end - start));
   1.387 +    start = System.currentTimeMillis();
   1.388 +    for (int i=0; i<ITERS; i++) {
   1.389 +      test_adda(a0, a1, a2);
   1.390 +    }
   1.391 +    end = System.currentTimeMillis();
   1.392 +    System.out.println("test_adda: " + (end - start));
   1.393 +
   1.394 +    start = System.currentTimeMillis();
   1.395 +    for (int i=0; i<ITERS; i++) {
   1.396 +      test_subc(a0, a1);
   1.397 +    }
   1.398 +    end = System.currentTimeMillis();
   1.399 +    System.out.println("test_subc: " + (end - start));
   1.400 +    start = System.currentTimeMillis();
   1.401 +    for (int i=0; i<ITERS; i++) {
   1.402 +      test_subv(a0, a1, (long)VALUE);
   1.403 +    }
   1.404 +    end = System.currentTimeMillis();
   1.405 +    System.out.println("test_subv: " + (end - start));
   1.406 +    start = System.currentTimeMillis();
   1.407 +    for (int i=0; i<ITERS; i++) {
   1.408 +      test_suba(a0, a1, a2);
   1.409 +    }
   1.410 +    end = System.currentTimeMillis();
   1.411 +    System.out.println("test_suba: " + (end - start));
   1.412 +
   1.413 +    start = System.currentTimeMillis();
   1.414 +    for (int i=0; i<ITERS; i++) {
   1.415 +      test_mulc(a0, a1);
   1.416 +    }
   1.417 +    end = System.currentTimeMillis();
   1.418 +    System.out.println("test_mulc: " + (end - start));
   1.419 +    start = System.currentTimeMillis();
   1.420 +    for (int i=0; i<ITERS; i++) {
   1.421 +      test_mulv(a0, a1, (long)VALUE);
   1.422 +    }
   1.423 +    end = System.currentTimeMillis();
   1.424 +    System.out.println("test_mulv: " + (end - start));
   1.425 +    start = System.currentTimeMillis();
   1.426 +    for (int i=0; i<ITERS; i++) {
   1.427 +      test_mula(a0, a1, a2);
   1.428 +    }
   1.429 +    end = System.currentTimeMillis();
   1.430 +    System.out.println("test_mula: " + (end - start));
   1.431 +
   1.432 +    start = System.currentTimeMillis();
   1.433 +    for (int i=0; i<ITERS; i++) {
   1.434 +      test_divc(a0, a1);
   1.435 +    }
   1.436 +    end = System.currentTimeMillis();
   1.437 +    System.out.println("test_divc: " + (end - start));
   1.438 +    start = System.currentTimeMillis();
   1.439 +    for (int i=0; i<ITERS; i++) {
   1.440 +      test_divv(a0, a1, (long)VALUE);
   1.441 +    }
   1.442 +    end = System.currentTimeMillis();
   1.443 +    System.out.println("test_divv: " + (end - start));
   1.444 +    start = System.currentTimeMillis();
   1.445 +    for (int i=0; i<ITERS; i++) {
   1.446 +      test_diva(a0, a1, a2);
   1.447 +    }
   1.448 +    end = System.currentTimeMillis();
   1.449 +    System.out.println("test_diva: " + (end - start));
   1.450 +
   1.451 +    start = System.currentTimeMillis();
   1.452 +    for (int i=0; i<ITERS; i++) {
   1.453 +      test_mulc_n(a0, a1);
   1.454 +    }
   1.455 +    end = System.currentTimeMillis();
   1.456 +    System.out.println("test_mulc_n: " + (end - start));
   1.457 +    start = System.currentTimeMillis();
   1.458 +    for (int i=0; i<ITERS; i++) {
   1.459 +      test_mulv(a0, a1, (long)-VALUE);
   1.460 +    }
   1.461 +    end = System.currentTimeMillis();
   1.462 +    System.out.println("test_mulv_n: " + (end - start));
   1.463 +    start = System.currentTimeMillis();
   1.464 +    for (int i=0; i<ITERS; i++) {
   1.465 +      test_mula(a0, a1, a3);
   1.466 +    }
   1.467 +    end = System.currentTimeMillis();
   1.468 +    System.out.println("test_mula_n: " + (end - start));
   1.469 +
   1.470 +    start = System.currentTimeMillis();
   1.471 +    for (int i=0; i<ITERS; i++) {
   1.472 +      test_divc_n(a0, a1);
   1.473 +    }
   1.474 +    end = System.currentTimeMillis();
   1.475 +    System.out.println("test_divc_n: " + (end - start));
   1.476 +    start = System.currentTimeMillis();
   1.477 +    for (int i=0; i<ITERS; i++) {
   1.478 +      test_divv(a0, a1, (long)-VALUE);
   1.479 +    }
   1.480 +    end = System.currentTimeMillis();
   1.481 +    System.out.println("test_divv_n: " + (end - start));
   1.482 +    start = System.currentTimeMillis();
   1.483 +    for (int i=0; i<ITERS; i++) {
   1.484 +      test_diva(a0, a1, a3);
   1.485 +    }
   1.486 +    end = System.currentTimeMillis();
   1.487 +    System.out.println("test_diva_n: " + (end - start));
   1.488 +
   1.489 +    start = System.currentTimeMillis();
   1.490 +    for (int i=0; i<ITERS; i++) {
   1.491 +      test_andc(a0, a1);
   1.492 +    }
   1.493 +    end = System.currentTimeMillis();
   1.494 +    System.out.println("test_andc: " + (end - start));
   1.495 +    start = System.currentTimeMillis();
   1.496 +    for (int i=0; i<ITERS; i++) {
   1.497 +      test_andv(a0, a1, (long)BIT_MASK);
   1.498 +    }
   1.499 +    end = System.currentTimeMillis();
   1.500 +    System.out.println("test_andv: " + (end - start));
   1.501 +    start = System.currentTimeMillis();
   1.502 +    for (int i=0; i<ITERS; i++) {
   1.503 +      test_anda(a0, a1, a4);
   1.504 +    }
   1.505 +    end = System.currentTimeMillis();
   1.506 +    System.out.println("test_anda: " + (end - start));
   1.507 +
   1.508 +    start = System.currentTimeMillis();
   1.509 +    for (int i=0; i<ITERS; i++) {
   1.510 +      test_orc(a0, a1);
   1.511 +    }
   1.512 +    end = System.currentTimeMillis();
   1.513 +    System.out.println("test_orc: " + (end - start));
   1.514 +    start = System.currentTimeMillis();
   1.515 +    for (int i=0; i<ITERS; i++) {
   1.516 +      test_orv(a0, a1, (long)BIT_MASK);
   1.517 +    }
   1.518 +    end = System.currentTimeMillis();
   1.519 +    System.out.println("test_orv: " + (end - start));
   1.520 +    start = System.currentTimeMillis();
   1.521 +    for (int i=0; i<ITERS; i++) {
   1.522 +      test_ora(a0, a1, a4);
   1.523 +    }
   1.524 +    end = System.currentTimeMillis();
   1.525 +    System.out.println("test_ora: " + (end - start));
   1.526 +
   1.527 +    start = System.currentTimeMillis();
   1.528 +    for (int i=0; i<ITERS; i++) {
   1.529 +      test_xorc(a0, a1);
   1.530 +    }
   1.531 +    end = System.currentTimeMillis();
   1.532 +    System.out.println("test_xorc: " + (end - start));
   1.533 +    start = System.currentTimeMillis();
   1.534 +    for (int i=0; i<ITERS; i++) {
   1.535 +      test_xorv(a0, a1, (long)BIT_MASK);
   1.536 +    }
   1.537 +    end = System.currentTimeMillis();
   1.538 +    System.out.println("test_xorv: " + (end - start));
   1.539 +    start = System.currentTimeMillis();
   1.540 +    for (int i=0; i<ITERS; i++) {
   1.541 +      test_xora(a0, a1, a4);
   1.542 +    }
   1.543 +    end = System.currentTimeMillis();
   1.544 +    System.out.println("test_xora: " + (end - start));
   1.545 +
   1.546 +    start = System.currentTimeMillis();
   1.547 +    for (int i=0; i<ITERS; i++) {
   1.548 +      test_sllc(a0, a1);
   1.549 +    }
   1.550 +    end = System.currentTimeMillis();
   1.551 +    System.out.println("test_sllc: " + (end - start));
   1.552 +    start = System.currentTimeMillis();
   1.553 +    for (int i=0; i<ITERS; i++) {
   1.554 +      test_sllv(a0, a1, VALUE);
   1.555 +    }
   1.556 +    end = System.currentTimeMillis();
   1.557 +    System.out.println("test_sllv: " + (end - start));
   1.558 +
   1.559 +    start = System.currentTimeMillis();
   1.560 +    for (int i=0; i<ITERS; i++) {
   1.561 +      test_srlc(a0, a1);
   1.562 +    }
   1.563 +    end = System.currentTimeMillis();
   1.564 +    System.out.println("test_srlc: " + (end - start));
   1.565 +    start = System.currentTimeMillis();
   1.566 +    for (int i=0; i<ITERS; i++) {
   1.567 +      test_srlv(a0, a1, VALUE);
   1.568 +    }
   1.569 +    end = System.currentTimeMillis();
   1.570 +    System.out.println("test_srlv: " + (end - start));
   1.571 +
   1.572 +    start = System.currentTimeMillis();
   1.573 +    for (int i=0; i<ITERS; i++) {
   1.574 +      test_srac(a0, a1);
   1.575 +    }
   1.576 +    end = System.currentTimeMillis();
   1.577 +    System.out.println("test_srac: " + (end - start));
   1.578 +    start = System.currentTimeMillis();
   1.579 +    for (int i=0; i<ITERS; i++) {
   1.580 +      test_srav(a0, a1, VALUE);
   1.581 +    }
   1.582 +    end = System.currentTimeMillis();
   1.583 +    System.out.println("test_srav: " + (end - start));
   1.584 +
   1.585 +    start = System.currentTimeMillis();
   1.586 +    for (int i=0; i<ITERS; i++) {
   1.587 +      test_sllc_n(a0, a1);
   1.588 +    }
   1.589 +    end = System.currentTimeMillis();
   1.590 +    System.out.println("test_sllc_n: " + (end - start));
   1.591 +    start = System.currentTimeMillis();
   1.592 +    for (int i=0; i<ITERS; i++) {
   1.593 +      test_sllv(a0, a1, -VALUE);
   1.594 +    }
   1.595 +    end = System.currentTimeMillis();
   1.596 +    System.out.println("test_sllv_n: " + (end - start));
   1.597 +
   1.598 +    start = System.currentTimeMillis();
   1.599 +    for (int i=0; i<ITERS; i++) {
   1.600 +      test_srlc_n(a0, a1);
   1.601 +    }
   1.602 +    end = System.currentTimeMillis();
   1.603 +    System.out.println("test_srlc_n: " + (end - start));
   1.604 +    start = System.currentTimeMillis();
   1.605 +    for (int i=0; i<ITERS; i++) {
   1.606 +      test_srlv(a0, a1, -VALUE);
   1.607 +    }
   1.608 +    end = System.currentTimeMillis();
   1.609 +    System.out.println("test_srlv_n: " + (end - start));
   1.610 +
   1.611 +    start = System.currentTimeMillis();
   1.612 +    for (int i=0; i<ITERS; i++) {
   1.613 +      test_srac_n(a0, a1);
   1.614 +    }
   1.615 +    end = System.currentTimeMillis();
   1.616 +    System.out.println("test_srac_n: " + (end - start));
   1.617 +    start = System.currentTimeMillis();
   1.618 +    for (int i=0; i<ITERS; i++) {
   1.619 +      test_srav(a0, a1, -VALUE);
   1.620 +    }
   1.621 +    end = System.currentTimeMillis();
   1.622 +    System.out.println("test_srav_n: " + (end - start));
   1.623 +
   1.624 +    start = System.currentTimeMillis();
   1.625 +    for (int i=0; i<ITERS; i++) {
   1.626 +      test_sllc_o(a0, a1);
   1.627 +    }
   1.628 +    end = System.currentTimeMillis();
   1.629 +    System.out.println("test_sllc_o: " + (end - start));
   1.630 +    start = System.currentTimeMillis();
   1.631 +    for (int i=0; i<ITERS; i++) {
   1.632 +      test_sllv(a0, a1, SHIFT);
   1.633 +    }
   1.634 +    end = System.currentTimeMillis();
   1.635 +    System.out.println("test_sllv_o: " + (end - start));
   1.636 +
   1.637 +    start = System.currentTimeMillis();
   1.638 +    for (int i=0; i<ITERS; i++) {
   1.639 +      test_srlc_o(a0, a1);
   1.640 +    }
   1.641 +    end = System.currentTimeMillis();
   1.642 +    System.out.println("test_srlc_o: " + (end - start));
   1.643 +    start = System.currentTimeMillis();
   1.644 +    for (int i=0; i<ITERS; i++) {
   1.645 +      test_srlv(a0, a1, SHIFT);
   1.646 +    }
   1.647 +    end = System.currentTimeMillis();
   1.648 +    System.out.println("test_srlv_o: " + (end - start));
   1.649 +
   1.650 +    start = System.currentTimeMillis();
   1.651 +    for (int i=0; i<ITERS; i++) {
   1.652 +      test_srac_o(a0, a1);
   1.653 +    }
   1.654 +    end = System.currentTimeMillis();
   1.655 +    System.out.println("test_srac_o: " + (end - start));
   1.656 +    start = System.currentTimeMillis();
   1.657 +    for (int i=0; i<ITERS; i++) {
   1.658 +      test_srav(a0, a1, SHIFT);
   1.659 +    }
   1.660 +    end = System.currentTimeMillis();
   1.661 +    System.out.println("test_srav_o: " + (end - start));
   1.662 +
   1.663 +    start = System.currentTimeMillis();
   1.664 +    for (int i=0; i<ITERS; i++) {
   1.665 +      test_sllc_on(a0, a1);
   1.666 +    }
   1.667 +    end = System.currentTimeMillis();
   1.668 +    System.out.println("test_sllc_on: " + (end - start));
   1.669 +    start = System.currentTimeMillis();
   1.670 +    for (int i=0; i<ITERS; i++) {
   1.671 +      test_sllv(a0, a1, -SHIFT);
   1.672 +    }
   1.673 +    end = System.currentTimeMillis();
   1.674 +    System.out.println("test_sllv_on: " + (end - start));
   1.675 +
   1.676 +    start = System.currentTimeMillis();
   1.677 +    for (int i=0; i<ITERS; i++) {
   1.678 +      test_srlc_on(a0, a1);
   1.679 +    }
   1.680 +    end = System.currentTimeMillis();
   1.681 +    System.out.println("test_srlc_on: " + (end - start));
   1.682 +    start = System.currentTimeMillis();
   1.683 +    for (int i=0; i<ITERS; i++) {
   1.684 +      test_srlv(a0, a1, -SHIFT);
   1.685 +    }
   1.686 +    end = System.currentTimeMillis();
   1.687 +    System.out.println("test_srlv_on: " + (end - start));
   1.688 +
   1.689 +    start = System.currentTimeMillis();
   1.690 +    for (int i=0; i<ITERS; i++) {
   1.691 +      test_srac_on(a0, a1);
   1.692 +    }
   1.693 +    end = System.currentTimeMillis();
   1.694 +    System.out.println("test_srac_on: " + (end - start));
   1.695 +    start = System.currentTimeMillis();
   1.696 +    for (int i=0; i<ITERS; i++) {
   1.697 +      test_srav(a0, a1, -SHIFT);
   1.698 +    }
   1.699 +    end = System.currentTimeMillis();
   1.700 +    System.out.println("test_srav_on: " + (end - start));
   1.701 +
   1.702 +    return errn;
   1.703 +  }
   1.704 +
   1.705 +  static long test_sum(long[] a1) {
   1.706 +    long sum = 0;
   1.707 +    for (int i = 0; i < a1.length; i+=1) {
   1.708 +      sum += a1[i];
   1.709 +    }
   1.710 +    return sum;
   1.711 +  }
   1.712 +
   1.713 +  static void test_addc(long[] a0, long[] a1) {
   1.714 +    for (int i = 0; i < a0.length; i+=1) {
   1.715 +      a0[i] = (long)(a1[i]+VALUE);
   1.716 +    }
   1.717 +  }
   1.718 +  static void test_addv(long[] a0, long[] a1, long b) {
   1.719 +    for (int i = 0; i < a0.length; i+=1) {
   1.720 +      a0[i] = (long)(a1[i]+b);
   1.721 +    }
   1.722 +  }
   1.723 +  static void test_adda(long[] a0, long[] a1, long[] a2) {
   1.724 +    for (int i = 0; i < a0.length; i+=1) {
   1.725 +      a0[i] = (long)(a1[i]+a2[i]);
   1.726 +    }
   1.727 +  }
   1.728 +
   1.729 +  static void test_subc(long[] a0, long[] a1) {
   1.730 +    for (int i = 0; i < a0.length; i+=1) {
   1.731 +      a0[i] = (long)(a1[i]-VALUE);
   1.732 +    }
   1.733 +  }
   1.734 +  static void test_subv(long[] a0, long[] a1, long b) {
   1.735 +    for (int i = 0; i < a0.length; i+=1) {
   1.736 +      a0[i] = (long)(a1[i]-b);
   1.737 +    }
   1.738 +  }
   1.739 +  static void test_suba(long[] a0, long[] a1, long[] a2) {
   1.740 +    for (int i = 0; i < a0.length; i+=1) {
   1.741 +      a0[i] = (long)(a1[i]-a2[i]);
   1.742 +    }
   1.743 +  }
   1.744 +
   1.745 +  static void test_mulc(long[] a0, long[] a1) {
   1.746 +    for (int i = 0; i < a0.length; i+=1) {
   1.747 +      a0[i] = (long)(a1[i]*VALUE);
   1.748 +    }
   1.749 +  }
   1.750 +  static void test_mulc_n(long[] a0, long[] a1) {
   1.751 +    for (int i = 0; i < a0.length; i+=1) {
   1.752 +      a0[i] = (long)(a1[i]*(-VALUE));
   1.753 +    }
   1.754 +  }
   1.755 +  static void test_mulv(long[] a0, long[] a1, long b) {
   1.756 +    for (int i = 0; i < a0.length; i+=1) {
   1.757 +      a0[i] = (long)(a1[i]*b);
   1.758 +    }
   1.759 +  }
   1.760 +  static void test_mula(long[] a0, long[] a1, long[] a2) {
   1.761 +    for (int i = 0; i < a0.length; i+=1) {
   1.762 +      a0[i] = (long)(a1[i]*a2[i]);
   1.763 +    }
   1.764 +  }
   1.765 +
   1.766 +  static void test_divc(long[] a0, long[] a1) {
   1.767 +    for (int i = 0; i < a0.length; i+=1) {
   1.768 +      a0[i] = (long)(a1[i]/VALUE);
   1.769 +    }
   1.770 +  }
   1.771 +  static void test_divc_n(long[] a0, long[] a1) {
   1.772 +    for (int i = 0; i < a0.length; i+=1) {
   1.773 +      a0[i] = (long)(a1[i]/(-VALUE));
   1.774 +    }
   1.775 +  }
   1.776 +  static void test_divv(long[] a0, long[] a1, long b) {
   1.777 +    for (int i = 0; i < a0.length; i+=1) {
   1.778 +      a0[i] = (long)(a1[i]/b);
   1.779 +    }
   1.780 +  }
   1.781 +  static void test_diva(long[] a0, long[] a1, long[] a2) {
   1.782 +    for (int i = 0; i < a0.length; i+=1) {
   1.783 +      a0[i] = (long)(a1[i]/a2[i]);
   1.784 +    }
   1.785 +  }
   1.786 +
   1.787 +  static void test_andc(long[] a0, long[] a1) {
   1.788 +    for (int i = 0; i < a0.length; i+=1) {
   1.789 +      a0[i] = (long)(a1[i]&BIT_MASK);
   1.790 +    }
   1.791 +  }
   1.792 +  static void test_andv(long[] a0, long[] a1, long b) {
   1.793 +    for (int i = 0; i < a0.length; i+=1) {
   1.794 +      a0[i] = (long)(a1[i]&b);
   1.795 +    }
   1.796 +  }
   1.797 +  static void test_anda(long[] a0, long[] a1, long[] a2) {
   1.798 +    for (int i = 0; i < a0.length; i+=1) {
   1.799 +      a0[i] = (long)(a1[i]&a2[i]);
   1.800 +    }
   1.801 +  }
   1.802 +
   1.803 +  static void test_orc(long[] a0, long[] a1) {
   1.804 +    for (int i = 0; i < a0.length; i+=1) {
   1.805 +      a0[i] = (long)(a1[i]|BIT_MASK);
   1.806 +    }
   1.807 +  }
   1.808 +  static void test_orv(long[] a0, long[] a1, long b) {
   1.809 +    for (int i = 0; i < a0.length; i+=1) {
   1.810 +      a0[i] = (long)(a1[i]|b);
   1.811 +    }
   1.812 +  }
   1.813 +  static void test_ora(long[] a0, long[] a1, long[] a2) {
   1.814 +    for (int i = 0; i < a0.length; i+=1) {
   1.815 +      a0[i] = (long)(a1[i]|a2[i]);
   1.816 +    }
   1.817 +  }
   1.818 +
   1.819 +  static void test_xorc(long[] a0, long[] a1) {
   1.820 +    for (int i = 0; i < a0.length; i+=1) {
   1.821 +      a0[i] = (long)(a1[i]^BIT_MASK);
   1.822 +    }
   1.823 +  }
   1.824 +  static void test_xorv(long[] a0, long[] a1, long b) {
   1.825 +    for (int i = 0; i < a0.length; i+=1) {
   1.826 +      a0[i] = (long)(a1[i]^b);
   1.827 +    }
   1.828 +  }
   1.829 +  static void test_xora(long[] a0, long[] a1, long[] a2) {
   1.830 +    for (int i = 0; i < a0.length; i+=1) {
   1.831 +      a0[i] = (long)(a1[i]^a2[i]);
   1.832 +    }
   1.833 +  }
   1.834 +
   1.835 +  static void test_sllc(long[] a0, long[] a1) {
   1.836 +    for (int i = 0; i < a0.length; i+=1) {
   1.837 +      a0[i] = (long)(a1[i]<<VALUE);
   1.838 +    }
   1.839 +  }
   1.840 +  static void test_sllc_n(long[] a0, long[] a1) {
   1.841 +    for (int i = 0; i < a0.length; i+=1) {
   1.842 +      a0[i] = (long)(a1[i]<<(-VALUE));
   1.843 +    }
   1.844 +  }
   1.845 +  static void test_sllc_o(long[] a0, long[] a1) {
   1.846 +    for (int i = 0; i < a0.length; i+=1) {
   1.847 +      a0[i] = (long)(a1[i]<<SHIFT);
   1.848 +    }
   1.849 +  }
   1.850 +  static void test_sllc_on(long[] a0, long[] a1) {
   1.851 +    for (int i = 0; i < a0.length; i+=1) {
   1.852 +      a0[i] = (long)(a1[i]<<(-SHIFT));
   1.853 +    }
   1.854 +  }
   1.855 +  static void test_sllv(long[] a0, long[] a1, int b) {
   1.856 +    for (int i = 0; i < a0.length; i+=1) {
   1.857 +      a0[i] = (long)(a1[i]<<b);
   1.858 +    }
   1.859 +  }
   1.860 +
   1.861 +  static void test_srlc(long[] a0, long[] a1) {
   1.862 +    for (int i = 0; i < a0.length; i+=1) {
   1.863 +      a0[i] = (long)(a1[i]>>>VALUE);
   1.864 +    }
   1.865 +  }
   1.866 +  static void test_srlc_n(long[] a0, long[] a1) {
   1.867 +    for (int i = 0; i < a0.length; i+=1) {
   1.868 +      a0[i] = (long)(a1[i]>>>(-VALUE));
   1.869 +    }
   1.870 +  }
   1.871 +  static void test_srlc_o(long[] a0, long[] a1) {
   1.872 +    for (int i = 0; i < a0.length; i+=1) {
   1.873 +      a0[i] = (long)(a1[i]>>>SHIFT);
   1.874 +    }
   1.875 +  }
   1.876 +  static void test_srlc_on(long[] a0, long[] a1) {
   1.877 +    for (int i = 0; i < a0.length; i+=1) {
   1.878 +      a0[i] = (long)(a1[i]>>>(-SHIFT));
   1.879 +    }
   1.880 +  }
   1.881 +  static void test_srlv(long[] a0, long[] a1, int b) {
   1.882 +    for (int i = 0; i < a0.length; i+=1) {
   1.883 +      a0[i] = (long)(a1[i]>>>b);
   1.884 +    }
   1.885 +  }
   1.886 +
   1.887 +  static void test_srac(long[] a0, long[] a1) {
   1.888 +    for (int i = 0; i < a0.length; i+=1) {
   1.889 +      a0[i] = (long)(a1[i]>>VALUE);
   1.890 +    }
   1.891 +  }
   1.892 +  static void test_srac_n(long[] a0, long[] a1) {
   1.893 +    for (int i = 0; i < a0.length; i+=1) {
   1.894 +      a0[i] = (long)(a1[i]>>(-VALUE));
   1.895 +    }
   1.896 +  }
   1.897 +  static void test_srac_o(long[] a0, long[] a1) {
   1.898 +    for (int i = 0; i < a0.length; i+=1) {
   1.899 +      a0[i] = (long)(a1[i]>>SHIFT);
   1.900 +    }
   1.901 +  }
   1.902 +  static void test_srac_on(long[] a0, long[] a1) {
   1.903 +    for (int i = 0; i < a0.length; i+=1) {
   1.904 +      a0[i] = (long)(a1[i]>>(-SHIFT));
   1.905 +    }
   1.906 +  }
   1.907 +  static void test_srav(long[] a0, long[] a1, int b) {
   1.908 +    for (int i = 0; i < a0.length; i+=1) {
   1.909 +      a0[i] = (long)(a1[i]>>b);
   1.910 +    }
   1.911 +  }
   1.912 +
   1.913 +  static int verify(String text, int i, long elem, long val) {
   1.914 +    if (elem != val) {
   1.915 +      System.err.println(text + "[" + i + "] = " + elem + " != " + val);
   1.916 +      return 1;
   1.917 +    }
   1.918 +    return 0;
   1.919 +  }
   1.920 +}

mercurial