kvn@4001: /*
kvn@4001:  * Copyright (c) 2012, Oracle and/or its affiliates. All rights reserved.
kvn@4001:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
kvn@4001:  *
kvn@4001:  * This code is free software; you can redistribute it and/or modify it
kvn@4001:  * under the terms of the GNU General Public License version 2 only, as
kvn@4001:  * published by the Free Software Foundation.
kvn@4001:  *
kvn@4001:  * This code is distributed in the hope that it will be useful, but WITHOUT
kvn@4001:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
kvn@4001:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
kvn@4001:  * version 2 for more details (a copy is included in the LICENSE file that
kvn@4001:  * accompanied this code).
kvn@4001:  *
kvn@4001:  * You should have received a copy of the GNU General Public License version
kvn@4001:  * 2 along with this work; if not, write to the Free Software Foundation,
kvn@4001:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
kvn@4001:  *
kvn@4001:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
kvn@4001:  * or visit www.oracle.com if you need additional information or have any
kvn@4001:  * questions.
kvn@4001:  *
kvn@4001:  */
kvn@4001: 
kvn@4001: /**
kvn@4001:  * @test
kvn@4001:  * @bug 6340864
kvn@4001:  * @summary Implement vectorization optimizations in hotspot-server
kvn@4001:  *
kvn@4001:  * @run main/othervm/timeout=400 -Xbatch -Xmx64m TestShortVect
kvn@4001:  */
kvn@4001: 
kvn@4001: public class TestShortVect {
kvn@4001:   private static final int ARRLEN = 997;
kvn@4001:   private static final int ITERS  = 11000;
kvn@4001:   private static final int ADD_INIT = Short.MAX_VALUE-500;
kvn@4001:   private static final int BIT_MASK = 0xB731;
kvn@4001:   private static final int VALUE = 7;
kvn@4001:   private static final int SHIFT = 16;
kvn@4001: 
kvn@4001:   public static void main(String args[]) {
kvn@4001:     System.out.println("Testing Short vectors");
kvn@4001:     int errn = test();
kvn@4001:     if (errn > 0) {
kvn@4001:       System.err.println("FAILED: " + errn + " errors");
kvn@4001:       System.exit(97);
kvn@4001:     }
kvn@4001:     System.out.println("PASSED");
kvn@4001:   }
kvn@4001: 
kvn@4001:   static int test() {
kvn@4001:     short[] a0 = new short[ARRLEN];
kvn@4001:     short[] a1 = new short[ARRLEN];
kvn@4001:     short[] a2 = new short[ARRLEN];
kvn@4001:     short[] a3 = new short[ARRLEN];
kvn@4001:     short[] a4 = new short[ARRLEN];
kvn@4001:      int[] p2 = new  int[ARRLEN/2];
kvn@4001:     long[] p4 = new long[ARRLEN/4];
kvn@4001:     // Initialize
kvn@4001:     int gold_sum = 0;
kvn@4001:     for (int i=0; i<ARRLEN; i++) {
kvn@4001:       short val = (short)(ADD_INIT+i);
kvn@4001:       gold_sum += val;
kvn@4001:       a1[i] = val;
kvn@4001:       a2[i] = (short)VALUE;
kvn@4001:       a3[i] = (short)-VALUE;
kvn@4001:       a4[i] = (short)BIT_MASK;
kvn@4001:     }
kvn@4001:     System.out.println("Warmup");
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sum(a1);
kvn@4001:       test_addc(a0, a1);
kvn@4001:       test_addv(a0, a1, (short)VALUE);
kvn@4001:       test_adda(a0, a1, a2);
kvn@4001:       test_subc(a0, a1);
kvn@4001:       test_subv(a0, a1, (short)VALUE);
kvn@4001:       test_suba(a0, a1, a2);
kvn@4204: 
kvn@4001:       test_mulc(a0, a1);
kvn@4001:       test_mulv(a0, a1, (short)VALUE);
kvn@4001:       test_mula(a0, a1, a2);
kvn@4001:       test_divc(a0, a1);
kvn@4001:       test_divv(a0, a1, (short)VALUE);
kvn@4001:       test_diva(a0, a1, a2);
kvn@4001:       test_mulc_n(a0, a1);
kvn@4001:       test_mulv(a0, a1, (short)-VALUE);
kvn@4001:       test_mula(a0, a1, a3);
kvn@4001:       test_divc_n(a0, a1);
kvn@4001:       test_divv(a0, a1, (short)-VALUE);
kvn@4001:       test_diva(a0, a1, a3);
kvn@4204: 
kvn@4001:       test_andc(a0, a1);
kvn@4001:       test_andv(a0, a1, (short)BIT_MASK);
kvn@4001:       test_anda(a0, a1, a4);
kvn@4001:       test_orc(a0, a1);
kvn@4001:       test_orv(a0, a1, (short)BIT_MASK);
kvn@4001:       test_ora(a0, a1, a4);
kvn@4001:       test_xorc(a0, a1);
kvn@4001:       test_xorv(a0, a1, (short)BIT_MASK);
kvn@4001:       test_xora(a0, a1, a4);
kvn@4204: 
kvn@4001:       test_sllc(a0, a1);
kvn@4001:       test_sllv(a0, a1, VALUE);
kvn@4001:       test_srlc(a0, a1);
kvn@4001:       test_srlv(a0, a1, VALUE);
kvn@4001:       test_srac(a0, a1);
kvn@4001:       test_srav(a0, a1, VALUE);
kvn@4204: 
kvn@4001:       test_sllc_n(a0, a1);
kvn@4001:       test_sllv(a0, a1, -VALUE);
kvn@4001:       test_srlc_n(a0, a1);
kvn@4001:       test_srlv(a0, a1, -VALUE);
kvn@4001:       test_srac_n(a0, a1);
kvn@4001:       test_srav(a0, a1, -VALUE);
kvn@4204: 
kvn@4001:       test_sllc_o(a0, a1);
kvn@4001:       test_sllv(a0, a1, SHIFT);
kvn@4001:       test_srlc_o(a0, a1);
kvn@4001:       test_srlv(a0, a1, SHIFT);
kvn@4001:       test_srac_o(a0, a1);
kvn@4001:       test_srav(a0, a1, SHIFT);
kvn@4204: 
kvn@4001:       test_sllc_on(a0, a1);
kvn@4001:       test_sllv(a0, a1, -SHIFT);
kvn@4001:       test_srlc_on(a0, a1);
kvn@4001:       test_srlv(a0, a1, -SHIFT);
kvn@4001:       test_srac_on(a0, a1);
kvn@4001:       test_srav(a0, a1, -SHIFT);
kvn@4204: 
kvn@4204:       test_sllc_add(a0, a1);
kvn@4204:       test_sllv_add(a0, a1, ADD_INIT);
kvn@4204:       test_srlc_add(a0, a1);
kvn@4204:       test_srlv_add(a0, a1, ADD_INIT);
kvn@4204:       test_srac_add(a0, a1);
kvn@4204:       test_srav_add(a0, a1, ADD_INIT);
kvn@4204: 
kvn@4204:       test_sllc_and(a0, a1);
kvn@4204:       test_sllv_and(a0, a1, BIT_MASK);
kvn@4204:       test_srlc_and(a0, a1);
kvn@4204:       test_srlv_and(a0, a1, BIT_MASK);
kvn@4204:       test_srac_and(a0, a1);
kvn@4204:       test_srav_and(a0, a1, BIT_MASK);
kvn@4204: 
kvn@4001:       test_pack2(p2, a1);
kvn@4001:       test_unpack2(a0, p2);
kvn@4001:       test_pack2_swap(p2, a1);
kvn@4001:       test_unpack2_swap(a0, p2);
kvn@4001:       test_pack4(p4, a1);
kvn@4001:       test_unpack4(a0, p4);
kvn@4001:       test_pack4_swap(p4, a1);
kvn@4001:       test_unpack4_swap(a0, p4);
kvn@4001:     }
kvn@4001:     // Test and verify results
kvn@4001:     System.out.println("Verification");
kvn@4001:     int errn = 0;
kvn@4001:     {
kvn@4001:       int sum = test_sum(a1);
kvn@4001:       if (sum != gold_sum) {
kvn@4001:         System.err.println("test_sum:  " + sum + " != " + gold_sum);
kvn@4001:         errn++;
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_addc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_addc: ", i, a0[i], (short)((short)(ADD_INIT+i)+VALUE));
kvn@4001:       }
kvn@4001:       test_addv(a0, a1, (short)VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_addv: ", i, a0[i], (short)((short)(ADD_INIT+i)+VALUE));
kvn@4001:       }
kvn@4001:       test_adda(a0, a1, a2);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_adda: ", i, a0[i], (short)((short)(ADD_INIT+i)+VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_subc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_subc: ", i, a0[i], (short)((short)(ADD_INIT+i)-VALUE));
kvn@4001:       }
kvn@4001:       test_subv(a0, a1, (short)VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_subv: ", i, a0[i], (short)((short)(ADD_INIT+i)-VALUE));
kvn@4001:       }
kvn@4001:       test_suba(a0, a1, a2);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_suba: ", i, a0[i], (short)((short)(ADD_INIT+i)-VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_mulc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_mulc: ", i, a0[i], (short)((short)(ADD_INIT+i)*VALUE));
kvn@4001:       }
kvn@4001:       test_mulv(a0, a1, (short)VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_mulv: ", i, a0[i], (short)((short)(ADD_INIT+i)*VALUE));
kvn@4001:       }
kvn@4001:       test_mula(a0, a1, a2);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_mula: ", i, a0[i], (short)((short)(ADD_INIT+i)*VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_divc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_divc: ", i, a0[i], (short)((short)(ADD_INIT+i)/VALUE));
kvn@4001:       }
kvn@4001:       test_divv(a0, a1, (short)VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_divv: ", i, a0[i], (short)((short)(ADD_INIT+i)/VALUE));
kvn@4001:       }
kvn@4001:       test_diva(a0, a1, a2);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_diva: ", i, a0[i], (short)((short)(ADD_INIT+i)/VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_mulc_n(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_mulc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)*(-VALUE)));
kvn@4001:       }
kvn@4001:       test_mulv(a0, a1, (short)-VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_mulv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)*(-VALUE)));
kvn@4001:       }
kvn@4001:       test_mula(a0, a1, a3);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_mula_n: ", i, a0[i], (short)((short)(ADD_INIT+i)*(-VALUE)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_divc_n(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_divc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)/(-VALUE)));
kvn@4001:       }
kvn@4001:       test_divv(a0, a1, (short)-VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_divv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)/(-VALUE)));
kvn@4001:       }
kvn@4001:       test_diva(a0, a1, a3);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_diva_n: ", i, a0[i], (short)((short)(ADD_INIT+i)/(-VALUE)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_andc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_andc: ", i, a0[i], (short)((short)(ADD_INIT+i)&BIT_MASK));
kvn@4001:       }
kvn@4001:       test_andv(a0, a1, (short)BIT_MASK);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_andv: ", i, a0[i], (short)((short)(ADD_INIT+i)&BIT_MASK));
kvn@4001:       }
kvn@4001:       test_anda(a0, a1, a4);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_anda: ", i, a0[i], (short)((short)(ADD_INIT+i)&BIT_MASK));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_orc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_orc: ", i, a0[i], (short)((short)(ADD_INIT+i)|BIT_MASK));
kvn@4001:       }
kvn@4001:       test_orv(a0, a1, (short)BIT_MASK);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_orv: ", i, a0[i], (short)((short)(ADD_INIT+i)|BIT_MASK));
kvn@4001:       }
kvn@4001:       test_ora(a0, a1, a4);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_ora: ", i, a0[i], (short)((short)(ADD_INIT+i)|BIT_MASK));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_xorc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_xorc: ", i, a0[i], (short)((short)(ADD_INIT+i)^BIT_MASK));
kvn@4001:       }
kvn@4001:       test_xorv(a0, a1, (short)BIT_MASK);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_xorv: ", i, a0[i], (short)((short)(ADD_INIT+i)^BIT_MASK));
kvn@4001:       }
kvn@4001:       test_xora(a0, a1, a4);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_xora: ", i, a0[i], (short)((short)(ADD_INIT+i)^BIT_MASK));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_sllc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllc: ", i, a0[i], (short)((short)(ADD_INIT+i)<<VALUE));
kvn@4001:       }
kvn@4001:       test_sllv(a0, a1, VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllv: ", i, a0[i], (short)((short)(ADD_INIT+i)<<VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srlc(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlc: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>VALUE));
kvn@4001:       }
kvn@4001:       test_srlv(a0, a1, VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlv: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srac(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srac: ", i, a0[i], (short)((short)(ADD_INIT+i)>>VALUE));
kvn@4001:       }
kvn@4001:       test_srav(a0, a1, VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srav: ", i, a0[i], (short)((short)(ADD_INIT+i)>>VALUE));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_sllc_n(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-VALUE)));
kvn@4001:       }
kvn@4001:       test_sllv(a0, a1, -VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-VALUE)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srlc_n(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlc_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-VALUE)));
kvn@4001:       }
kvn@4001:       test_srlv(a0, a1, -VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlv_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-VALUE)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srac_n(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srac_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-VALUE)));
kvn@4001:       }
kvn@4001:       test_srav(a0, a1, -VALUE);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srav_n: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-VALUE)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_sllc_o(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllc_o: ", i, a0[i], (short)((short)(ADD_INIT+i)<<SHIFT));
kvn@4001:       }
kvn@4001:       test_sllv(a0, a1, SHIFT);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllv_o: ", i, a0[i], (short)((short)(ADD_INIT+i)<<SHIFT));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srlc_o(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlc_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>SHIFT));
kvn@4001:       }
kvn@4001:       test_srlv(a0, a1, SHIFT);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlv_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>SHIFT));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srac_o(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srac_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>SHIFT));
kvn@4001:       }
kvn@4001:       test_srav(a0, a1, SHIFT);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srav_o: ", i, a0[i], (short)((short)(ADD_INIT+i)>>SHIFT));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_sllc_on(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllc_on: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-SHIFT)));
kvn@4001:       }
kvn@4001:       test_sllv(a0, a1, -SHIFT);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_sllv_on: ", i, a0[i], (short)((short)(ADD_INIT+i)<<(-SHIFT)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srlc_on(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlc_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-SHIFT)));
kvn@4001:       }
kvn@4001:       test_srlv(a0, a1, -SHIFT);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srlv_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>>(-SHIFT)));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_srac_on(a0, a1);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srac_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-SHIFT)));
kvn@4001:       }
kvn@4001:       test_srav(a0, a1, -SHIFT);
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         errn += verify("test_srav_on: ", i, a0[i], (short)((short)(ADD_INIT+i)>>(-SHIFT)));
kvn@4001:       }
kvn@4001: 
kvn@4204:       test_sllc_add(a0, a1);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_sllc_add: ", i, a0[i], (short)(((short)(ADD_INIT+i) + ADD_INIT)<<VALUE));
kvn@4204:       }
kvn@4204:       test_sllv_add(a0, a1, ADD_INIT);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_sllv_add: ", i, a0[i], (short)(((short)(ADD_INIT+i) + ADD_INIT)<<VALUE));
kvn@4204:       }
kvn@4204: 
kvn@4204:       test_srlc_add(a0, a1);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srlc_add: ", i, a0[i], (short)(((short)(ADD_INIT+i) + ADD_INIT)>>>VALUE));
kvn@4204:       }
kvn@4204:       test_srlv_add(a0, a1, ADD_INIT);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srlv_add: ", i, a0[i], (short)(((short)(ADD_INIT+i) + ADD_INIT)>>>VALUE));
kvn@4204:       }
kvn@4204: 
kvn@4204:       test_srac_add(a0, a1);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srac_add: ", i, a0[i], (short)(((short)(ADD_INIT+i) + ADD_INIT)>>VALUE));
kvn@4204:       }
kvn@4204:       test_srav_add(a0, a1, ADD_INIT);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srav_add: ", i, a0[i], (short)(((short)(ADD_INIT+i) + ADD_INIT)>>VALUE));
kvn@4204:       }
kvn@4204: 
kvn@4204:       test_sllc_and(a0, a1);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_sllc_and: ", i, a0[i], (short)(((short)(ADD_INIT+i) & BIT_MASK)<<VALUE));
kvn@4204:       }
kvn@4204:       test_sllv_and(a0, a1, BIT_MASK);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_sllv_and: ", i, a0[i], (short)(((short)(ADD_INIT+i) & BIT_MASK)<<VALUE));
kvn@4204:       }
kvn@4204: 
kvn@4204:       test_srlc_and(a0, a1);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srlc_and: ", i, a0[i], (short)(((short)(ADD_INIT+i) & BIT_MASK)>>>VALUE));
kvn@4204:       }
kvn@4204:       test_srlv_and(a0, a1, BIT_MASK);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srlv_and: ", i, a0[i], (short)(((short)(ADD_INIT+i) & BIT_MASK)>>>VALUE));
kvn@4204:       }
kvn@4204: 
kvn@4204:       test_srac_and(a0, a1);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srac_and: ", i, a0[i], (short)(((short)(ADD_INIT+i) & BIT_MASK)>>VALUE));
kvn@4204:       }
kvn@4204:       test_srav_and(a0, a1, BIT_MASK);
kvn@4204:       for (int i=0; i<ARRLEN; i++) {
kvn@4204:         errn += verify("test_srav_and: ", i, a0[i], (short)(((short)(ADD_INIT+i) & BIT_MASK)>>VALUE));
kvn@4204:       }
kvn@4204: 
kvn@4001:       test_pack2(p2, a1);
kvn@4001:       for (int i=0; i<ARRLEN/2; i++) {
kvn@4001:         errn += verify("test_pack2: ", i, p2[i], ((int)(ADD_INIT+2*i) & 0xFFFF) | ((int)(ADD_INIT+2*i+1) << 16));
kvn@4001:       }
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         a0[i] = -1;
kvn@4001:       }
kvn@4001:       test_unpack2(a0, p2);
kvn@4001:       for (int i=0; i<(ARRLEN&(-2)); i++) {
kvn@4001:         errn += verify("test_unpack2: ", i, a0[i], (short)(ADD_INIT+i));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_pack2_swap(p2, a1);
kvn@4001:       for (int i=0; i<ARRLEN/2; i++) {
kvn@4001:         errn += verify("test_pack2_swap: ", i, p2[i], ((int)(ADD_INIT+2*i+1) & 0xFFFF) | ((int)(ADD_INIT+2*i) << 16));
kvn@4001:       }
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         a0[i] = -1;
kvn@4001:       }
kvn@4001:       test_unpack2_swap(a0, p2);
kvn@4001:       for (int i=0; i<(ARRLEN&(-2)); i++) {
kvn@4001:         errn += verify("test_unpack2_swap: ", i, a0[i], (short)(ADD_INIT+i));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_pack4(p4, a1);
kvn@4001:       for (int i=0; i<ARRLEN/4; i++) {
kvn@4001:         errn += verify("test_pack4: ", i, p4[i],  ((long)(ADD_INIT+4*i+0) & 0xFFFFl) |
kvn@4001:                                                  (((long)(ADD_INIT+4*i+1) & 0xFFFFl) << 16)  |
kvn@4001:                                                  (((long)(ADD_INIT+4*i+2) & 0xFFFFl) << 32)  |
kvn@4001:                                                  (((long)(ADD_INIT+4*i+3) & 0xFFFFl) << 48));
kvn@4001:       }
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         a0[i] = -1;
kvn@4001:       }
kvn@4001:       test_unpack4(a0, p4);
kvn@4001:       for (int i=0; i<(ARRLEN&(-4)); i++) {
kvn@4001:         errn += verify("test_unpack4: ", i, a0[i], (short)(ADD_INIT+i));
kvn@4001:       }
kvn@4001: 
kvn@4001:       test_pack4_swap(p4, a1);
kvn@4001:       for (int i=0; i<ARRLEN/4; i++) {
kvn@4001:         errn += verify("test_pack4_swap: ", i, p4[i],  ((long)(ADD_INIT+4*i+3) & 0xFFFFl) |
kvn@4001:                                                       (((long)(ADD_INIT+4*i+2) & 0xFFFFl) << 16)  |
kvn@4001:                                                       (((long)(ADD_INIT+4*i+1) & 0xFFFFl) << 32)  |
kvn@4001:                                                       (((long)(ADD_INIT+4*i+0) & 0xFFFFl) << 48));
kvn@4001:       }
kvn@4001:       for (int i=0; i<ARRLEN; i++) {
kvn@4001:         a0[i] = -1;
kvn@4001:       }
kvn@4001:       test_unpack4_swap(a0, p4);
kvn@4001:       for (int i=0; i<(ARRLEN&(-4)); i++) {
kvn@4001:         errn += verify("test_unpack4_swap: ", i, a0[i], (short)(ADD_INIT+i));
kvn@4001:       }
kvn@4001: 
kvn@4001:     }
kvn@4001: 
kvn@4001:     if (errn > 0)
kvn@4001:       return errn;
kvn@4001: 
kvn@4001:     System.out.println("Time");
kvn@4001:     long start, end;
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sum(a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sum: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_addc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_addc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_addv(a0, a1, (short)VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_addv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_adda(a0, a1, a2);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_adda: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_subc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_subc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_subv(a0, a1, (short)VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_subv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_suba(a0, a1, a2);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_suba: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_mulc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_mulc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_mulv(a0, a1, (short)VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_mulv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_mula(a0, a1, a2);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_mula: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_divc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_divc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_divv(a0, a1, (short)VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_divv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_diva(a0, a1, a2);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_diva: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_mulc_n(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_mulc_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_mulv(a0, a1, (short)-VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_mulv_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_mula(a0, a1, a3);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_mula_n: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_divc_n(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_divc_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_divv(a0, a1, (short)-VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_divv_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_diva(a0, a1, a3);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_diva_n: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_andc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_andc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_andv(a0, a1, (short)BIT_MASK);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_andv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_anda(a0, a1, a4);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_anda: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_orc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_orc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_orv(a0, a1, (short)BIT_MASK);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_orv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_ora(a0, a1, a4);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_ora: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_xorc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_xorc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_xorv(a0, a1, (short)BIT_MASK);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_xorv: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_xora(a0, a1, a4);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_xora: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllv(a0, a1, VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllv: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlc(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlc: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlv(a0, a1, VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlv: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srac(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srac: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srav(a0, a1, VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srav: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllc_n(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllc_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllv(a0, a1, -VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllv_n: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlc_n(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlc_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlv(a0, a1, -VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlv_n: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srac_n(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srac_n: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srav(a0, a1, -VALUE);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srav_n: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllc_o(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllc_o: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllv(a0, a1, SHIFT);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllv_o: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlc_o(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlc_o: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlv(a0, a1, SHIFT);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlv_o: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srac_o(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srac_o: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srav(a0, a1, SHIFT);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srav_o: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllc_on(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllc_on: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_sllv(a0, a1, -SHIFT);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_sllv_on: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlc_on(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlc_on: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srlv(a0, a1, -SHIFT);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srlv_on: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srac_on(a0, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srac_on: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_srav(a0, a1, -SHIFT);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_srav_on: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_sllc_add(a0, a1);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_sllc_add: " + (end - start));
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_sllv_add(a0, a1, ADD_INIT);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_sllv_add: " + (end - start));
kvn@4204: 
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srlc_add(a0, a1);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srlc_add: " + (end - start));
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srlv_add(a0, a1, ADD_INIT);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srlv_add: " + (end - start));
kvn@4204: 
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srac_add(a0, a1);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srac_add: " + (end - start));
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srav_add(a0, a1, ADD_INIT);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srav_add: " + (end - start));
kvn@4204: 
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_sllc_and(a0, a1);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_sllc_and: " + (end - start));
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_sllv_and(a0, a1, BIT_MASK);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_sllv_and: " + (end - start));
kvn@4204: 
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srlc_and(a0, a1);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srlc_and: " + (end - start));
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srlv_and(a0, a1, BIT_MASK);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srlv_and: " + (end - start));
kvn@4204: 
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srac_and(a0, a1);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srac_and: " + (end - start));
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4204:       test_srav_and(a0, a1, BIT_MASK);
kvn@4204:     }
kvn@4204:     end = System.currentTimeMillis();
kvn@4204:     System.out.println("test_srav_and: " + (end - start));
kvn@4204: 
kvn@4204:     start = System.currentTimeMillis();
kvn@4204:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_pack2(p2, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_pack2: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_unpack2(a0, p2);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_unpack2: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_pack2_swap(p2, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_pack2_swap: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_unpack2_swap(a0, p2);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_unpack2_swap: " + (end - start));
kvn@4001: 
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_pack4(p4, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_pack4: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_unpack4(a0, p4);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_unpack4: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_pack4_swap(p4, a1);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_pack4_swap: " + (end - start));
kvn@4001:     start = System.currentTimeMillis();
kvn@4001:     for (int i=0; i<ITERS; i++) {
kvn@4001:       test_unpack4_swap(a0, p4);
kvn@4001:     }
kvn@4001:     end = System.currentTimeMillis();
kvn@4001:     System.out.println("test_unpack4_swap: " + (end - start));
kvn@4001: 
kvn@4001:     return errn;
kvn@4001:   }
kvn@4001: 
kvn@4001:   static int test_sum(short[] a1) {
kvn@4001:     int sum = 0;
kvn@4001:     for (int i = 0; i < a1.length; i+=1) {
kvn@4001:       sum += a1[i];
kvn@4001:     }
kvn@4001:     return sum;
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_addc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]+VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_addv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]+b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_adda(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]+a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_subc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]-VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_subv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]-b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_suba(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]-a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_mulc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]*VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_mulc_n(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]*(-VALUE));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_mulv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]*b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_mula(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]*a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_divc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]/VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_divc_n(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]/(-VALUE));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_divv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]/b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_diva(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]/a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_andc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]&BIT_MASK);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_andv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]&b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_anda(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]&a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_orc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]|BIT_MASK);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_orv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]|b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_ora(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]|a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_xorc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]^BIT_MASK);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_xorv(short[] a0, short[] a1, short b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]^b);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_xora(short[] a0, short[] a1, short[] a2) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]^a2[i]);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_sllc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]<<VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_sllc_n(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]<<(-VALUE));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_sllc_o(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]<<SHIFT);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_sllc_on(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]<<(-SHIFT));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_sllv(short[] a0, short[] a1, int b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]<<b);
kvn@4001:     }
kvn@4001:   }
kvn@4204:   static void test_sllc_add(short[] a0, short[] a1) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] + ADD_INIT)<<VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_sllv_add(short[] a0, short[] a1, int b) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] + b)<<VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_sllc_and(short[] a0, short[] a1) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] & BIT_MASK)<<VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_sllv_and(short[] a0, short[] a1, int b) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] & b)<<VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4001: 
kvn@4001:   static void test_srlc(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>>VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srlc_n(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>>(-VALUE));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srlc_o(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>>SHIFT);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srlc_on(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>>(-SHIFT));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srlv(short[] a0, short[] a1, int b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>>b);
kvn@4001:     }
kvn@4001:   }
kvn@4204:   static void test_srlc_add(short[] a0, short[] a1) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] + ADD_INIT)>>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_srlv_add(short[] a0, short[] a1, int b) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] + b)>>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_srlc_and(short[] a0, short[] a1) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] & BIT_MASK)>>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_srlv_and(short[] a0, short[] a1, int b) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] & b)>>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4001: 
kvn@4001:   static void test_srac(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>VALUE);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srac_n(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>(-VALUE));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srac_o(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>SHIFT);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srac_on(short[] a0, short[] a1) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>(-SHIFT));
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_srav(short[] a0, short[] a1, int b) {
kvn@4001:     for (int i = 0; i < a0.length; i+=1) {
kvn@4001:       a0[i] = (short)(a1[i]>>b);
kvn@4001:     }
kvn@4001:   }
kvn@4204:   static void test_srac_add(short[] a0, short[] a1) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] + ADD_INIT)>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_srav_add(short[] a0, short[] a1, int b) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] + b)>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_srac_and(short[] a0, short[] a1) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] & BIT_MASK)>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4204:   static void test_srav_and(short[] a0, short[] a1, int b) {
kvn@4204:     for (int i = 0; i < a0.length; i+=1) {
kvn@4204:       a0[i] = (short)((a1[i] & b)>>VALUE);
kvn@4204:     }
kvn@4204:   }
kvn@4001: 
kvn@4001:   static void test_pack2(int[] p2, short[] a1) {
kvn@4001:     if (p2.length*2 > a1.length) return;
kvn@4001:     for (int i = 0; i < p2.length; i+=1) {
kvn@4001:       int l0 = (int)a1[i*2+0];
kvn@4001:       int l1 = (int)a1[i*2+1];
kvn@4001:       p2[i] = (l1 << 16) | (l0 & 0xFFFF);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_unpack2(short[] a0, int[] p2) {
kvn@4001:     if (p2.length*2 > a0.length) return;
kvn@4001:     for (int i = 0; i < p2.length; i+=1) {
kvn@4001:       int l = p2[i];
kvn@4001:       a0[i*2+0] = (short)(l & 0xFFFF);
kvn@4001:       a0[i*2+1] = (short)(l >> 16);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_pack2_swap(int[] p2, short[] a1) {
kvn@4001:     if (p2.length*2 > a1.length) return;
kvn@4001:     for (int i = 0; i < p2.length; i+=1) {
kvn@4001:       int l0 = (int)a1[i*2+0];
kvn@4001:       int l1 = (int)a1[i*2+1];
kvn@4001:       p2[i] = (l0 << 16) | (l1 & 0xFFFF);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_unpack2_swap(short[] a0, int[] p2) {
kvn@4001:     if (p2.length*2 > a0.length) return;
kvn@4001:     for (int i = 0; i < p2.length; i+=1) {
kvn@4001:       int l = p2[i];
kvn@4001:       a0[i*2+0] = (short)(l >> 16);
kvn@4001:       a0[i*2+1] = (short)(l & 0xFFFF);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static void test_pack4(long[] p4, short[] a1) {
kvn@4001:     if (p4.length*4 > a1.length) return;
kvn@4001:     for (int i = 0; i < p4.length; i+=1) {
kvn@4001:       long l0 = (long)a1[i*4+0];
kvn@4001:       long l1 = (long)a1[i*4+1];
kvn@4001:       long l2 = (long)a1[i*4+2];
kvn@4001:       long l3 = (long)a1[i*4+3];
kvn@4001:       p4[i] = (l0 & 0xFFFFl) |
kvn@4001:              ((l1 & 0xFFFFl) << 16) |
kvn@4001:              ((l2 & 0xFFFFl) << 32) |
kvn@4001:              ((l3 & 0xFFFFl) << 48);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_unpack4(short[] a0, long[] p4) {
kvn@4001:     if (p4.length*4 > a0.length) return;
kvn@4001:     for (int i = 0; i < p4.length; i+=1) {
kvn@4001:       long l = p4[i];
kvn@4001:       a0[i*4+0] = (short)(l & 0xFFFFl);
kvn@4001:       a0[i*4+1] = (short)(l >> 16);
kvn@4001:       a0[i*4+2] = (short)(l >> 32);
kvn@4001:       a0[i*4+3] = (short)(l >> 48);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_pack4_swap(long[] p4, short[] a1) {
kvn@4001:     if (p4.length*4 > a1.length) return;
kvn@4001:     for (int i = 0; i < p4.length; i+=1) {
kvn@4001:       long l0 = (long)a1[i*4+0];
kvn@4001:       long l1 = (long)a1[i*4+1];
kvn@4001:       long l2 = (long)a1[i*4+2];
kvn@4001:       long l3 = (long)a1[i*4+3];
kvn@4001:       p4[i] = (l3 & 0xFFFFl) |
kvn@4001:              ((l2 & 0xFFFFl) << 16) |
kvn@4001:              ((l1 & 0xFFFFl) << 32) |
kvn@4001:              ((l0 & 0xFFFFl) << 48);
kvn@4001:     }
kvn@4001:   }
kvn@4001:   static void test_unpack4_swap(short[] a0, long[] p4) {
kvn@4001:     if (p4.length*4 > a0.length) return;
kvn@4001:     for (int i = 0; i < p4.length; i+=1) {
kvn@4001:       long l = p4[i];
kvn@4001:       a0[i*4+0] = (short)(l >> 48);
kvn@4001:       a0[i*4+1] = (short)(l >> 32);
kvn@4001:       a0[i*4+2] = (short)(l >> 16);
kvn@4001:       a0[i*4+3] = (short)(l & 0xFFFFl);
kvn@4001:     }
kvn@4001:   }
kvn@4001: 
kvn@4001:   static int verify(String text, int i, short elem, short val) {
kvn@4001:     if (elem != val) {
kvn@4001:       System.err.println(text + "[" + i + "] = " + elem + " != " + val);
kvn@4001:       return 1;
kvn@4001:     }
kvn@4001:     return 0;
kvn@4001:   }
kvn@4001: 
kvn@4001:   static int verify(String text, int i, int elem, int val) {
kvn@4001:     if (elem != val) {
kvn@4001:       System.err.println(text + "[" + i + "] = " + Integer.toHexString(elem) + " != " + Integer.toHexString(val));
kvn@4001:       return 1;
kvn@4001:     }
kvn@4001:     return 0;
kvn@4001:   }
kvn@4001: 
kvn@4001:   static int verify(String text, int i, long elem, long val) {
kvn@4001:     if (elem != val) {
kvn@4001:       System.err.println(text + "[" + i + "] = " + Long.toHexString(elem) + " != " + Long.toHexString(val));
kvn@4001:       return 1;
kvn@4001:     }
kvn@4001:     return 0;
kvn@4001:   }
kvn@4001: }