2342 
2343     generate_conjoint_int_copy_core(aligned);
2344 
2345     // O3, O4 are used as temp registers
2346     inc_counter_np(SharedRuntime::_jint_array_copy_ctr, O3, O4);
2347     __ retl();
2348     __ delayed()->mov(G0, O0); // return 0
2349     return start;
2350   }
2351 
2352   //
2353   // Helper methods for generate_disjoint_long_copy_core()
2354   //
2355   void copy_64_bytes_loop(Register from, Register to, Register count, int count_dec,
2356                           Label& L_loop, bool use_prefetch, bool use_bis) {
2357     __ align(OptoLoopAlignment);
2358     __ BIND(L_loop);
2359     for (int off = 0; off < 64; off += 16) {
2360       if (use_prefetch && (off & 31) == 0) {
2361         if (ArraycopySrcPrefetchDistance > 0) {
2362           __ prefetch(from, ArraycopySrcPrefetchDistance, Assembler::severalReads);
2363         }
2364         if (ArraycopyDstPrefetchDistance > 0) {
2365           __ prefetch(to, ArraycopyDstPrefetchDistance, Assembler::severalWritesAndPossiblyReads);
2366         }
2367       }
2368       __ ldx(from,  off+0, O4);
2369       __ ldx(from,  off+8, O5);
2370       if (use_bis) {
2371         __ stxa(O4, to,  off+0);
2372         __ stxa(O5, to,  off+8);
2373       } else {
2374         __ stx(O4, to,  off+0);
2375         __ stx(O5, to,  off+8);
2376       }
2377     }
2378     __ deccc(count, 8);
2379     __ inc(from, 64);
2380     __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
2381     __ delayed()->inc(to, 64);
2382   }
2383 
2384   //
2385   //  Generate core code for disjoint long copy (and oop copy on 64-bit).

2342 
2343     generate_conjoint_int_copy_core(aligned);
2344 
2345     // O3, O4 are used as temp registers
2346     inc_counter_np(SharedRuntime::_jint_array_copy_ctr, O3, O4);
2347     __ retl();
2348     __ delayed()->mov(G0, O0); // return 0
2349     return start;
2350   }
2351 
2352   //
2353   // Helper methods for generate_disjoint_long_copy_core()
2354   //
2355   void copy_64_bytes_loop(Register from, Register to, Register count, int count_dec,
2356                           Label& L_loop, bool use_prefetch, bool use_bis) {
2357     __ align(OptoLoopAlignment);
2358     __ BIND(L_loop);
2359     for (int off = 0; off < 64; off += 16) {
2360       if (use_prefetch && (off & 31) == 0) {
2361         if (ArraycopySrcPrefetchDistance > 0) {
2362           __ prefetch(from, ArraycopySrcPrefetchDistance+off, Assembler::severalReads);
2363         }
2364         if (ArraycopyDstPrefetchDistance > 0) {
2365           __ prefetch(to, ArraycopyDstPrefetchDistance+off, Assembler::severalWritesAndPossiblyReads);
2366         }
2367       }
2368       __ ldx(from,  off+0, O4);
2369       __ ldx(from,  off+8, O5);
2370       if (use_bis) {
2371         __ stxa(O4, to,  off+0);
2372         __ stxa(O5, to,  off+8);
2373       } else {
2374         __ stx(O4, to,  off+0);
2375         __ stx(O5, to,  off+8);
2376       }
2377     }
2378     __ deccc(count, 8);
2379     __ inc(from, 64);
2380     __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
2381     __ delayed()->inc(to, 64);
2382   }
2383 
2384   //
2385   //  Generate core code for disjoint long copy (and oop copy on 64-bit).