1 /*
   2  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright 2012, 2014 SAP AG. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "interpreter/interpreter.hpp"
  29 #include "nativeInst_ppc.hpp"
  30 #include "oops/instanceOop.hpp"
  31 #include "oops/method.hpp"
  32 #include "oops/objArrayKlass.hpp"
  33 #include "oops/oop.inline.hpp"
  34 #include "prims/methodHandles.hpp"
  35 #include "runtime/frame.inline.hpp"
  36 #include "runtime/handles.inline.hpp"
  37 #include "runtime/sharedRuntime.hpp"
  38 #include "runtime/stubCodeGenerator.hpp"
  39 #include "runtime/stubRoutines.hpp"
  40 #include "utilities/top.hpp"
  41 #include "runtime/thread.inline.hpp"
  42 
  43 #define __ _masm->
  44 
  45 #ifdef PRODUCT
  46 #define BLOCK_COMMENT(str) // nothing
  47 #else
  48 #define BLOCK_COMMENT(str) __ block_comment(str)
  49 #endif
  50 
  51 class StubGenerator: public StubCodeGenerator {
  52  private:
  53 
  54   // Call stubs are used to call Java from C
  55   //
  56   // Arguments:
  57   //
  58   //   R3  - call wrapper address     : address
  59   //   R4  - result                   : intptr_t*
  60   //   R5  - result type              : BasicType
  61   //   R6  - method                   : Method
  62   //   R7  - frame mgr entry point    : address
  63   //   R8  - parameter block          : intptr_t*
  64   //   R9  - parameter count in words : int
  65   //   R10 - thread                   : Thread*
  66   //
  67   address generate_call_stub(address& return_address) {
  68     // Setup a new c frame, copy java arguments, call frame manager or
  69     // native_entry, and process result.
  70 
  71     StubCodeMark mark(this, "StubRoutines", "call_stub");
  72 
  73     address start = __ function_entry();
  74 
  75     // some sanity checks
  76     assert((sizeof(frame::abi_minframe) % 16) == 0,           "unaligned");
  77     assert((sizeof(frame::abi_reg_args) % 16) == 0,           "unaligned");
  78     assert((sizeof(frame::spill_nonvolatiles) % 16) == 0,     "unaligned");
  79     assert((sizeof(frame::parent_ijava_frame_abi) % 16) == 0, "unaligned");
  80     assert((sizeof(frame::entry_frame_locals) % 16) == 0,     "unaligned");
  81 
  82     Register r_arg_call_wrapper_addr        = R3;
  83     Register r_arg_result_addr              = R4;
  84     Register r_arg_result_type              = R5;
  85     Register r_arg_method                   = R6;
  86     Register r_arg_entry                    = R7;
  87     Register r_arg_thread                   = R10;
  88 
  89     Register r_temp                         = R24;
  90     Register r_top_of_arguments_addr        = R25;
  91     Register r_entryframe_fp                = R26;
  92 
  93     {
  94       // Stack on entry to call_stub:
  95       //
  96       //      F1      [C_FRAME]
  97       //              ...
  98 
  99       Register r_arg_argument_addr          = R8;
 100       Register r_arg_argument_count         = R9;
 101       Register r_frame_alignment_in_bytes   = R27;
 102       Register r_argument_addr              = R28;
 103       Register r_argumentcopy_addr          = R29;
 104       Register r_argument_size_in_bytes     = R30;
 105       Register r_frame_size                 = R23;
 106 
 107       Label arguments_copied;
 108 
 109       // Save LR/CR to caller's C_FRAME.
 110       __ save_LR_CR(R0);
 111 
 112       // Zero extend arg_argument_count.
 113       __ clrldi(r_arg_argument_count, r_arg_argument_count, 32);
 114 
 115       // Save non-volatiles GPRs to ENTRY_FRAME (not yet pushed, but it's safe).
 116       __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
 117 
 118       // Keep copy of our frame pointer (caller's SP).
 119       __ mr(r_entryframe_fp, R1_SP);
 120 
 121       BLOCK_COMMENT("Push ENTRY_FRAME including arguments");
 122       // Push ENTRY_FRAME including arguments:
 123       //
 124       //      F0      [TOP_IJAVA_FRAME_ABI]
 125       //              alignment (optional)
 126       //              [outgoing Java arguments]
 127       //              [ENTRY_FRAME_LOCALS]
 128       //      F1      [C_FRAME]
 129       //              ...
 130 
 131       // calculate frame size
 132 
 133       // unaligned size of arguments
 134       __ sldi(r_argument_size_in_bytes,
 135                   r_arg_argument_count, Interpreter::logStackElementSize);
 136       // arguments alignment (max 1 slot)
 137       // FIXME: use round_to() here
 138       __ andi_(r_frame_alignment_in_bytes, r_arg_argument_count, 1);
 139       __ sldi(r_frame_alignment_in_bytes,
 140               r_frame_alignment_in_bytes, Interpreter::logStackElementSize);
 141 
 142       // size = unaligned size of arguments + top abi's size
 143       __ addi(r_frame_size, r_argument_size_in_bytes,
 144               frame::top_ijava_frame_abi_size);
 145       // size += arguments alignment
 146       __ add(r_frame_size,
 147              r_frame_size, r_frame_alignment_in_bytes);
 148       // size += size of call_stub locals
 149       __ addi(r_frame_size,
 150               r_frame_size, frame::entry_frame_locals_size);
 151 
 152       // push ENTRY_FRAME
 153       __ push_frame(r_frame_size, r_temp);
 154 
 155       // initialize call_stub locals (step 1)
 156       __ std(r_arg_call_wrapper_addr,
 157              _entry_frame_locals_neg(call_wrapper_address), r_entryframe_fp);
 158       __ std(r_arg_result_addr,
 159              _entry_frame_locals_neg(result_address), r_entryframe_fp);
 160       __ std(r_arg_result_type,
 161              _entry_frame_locals_neg(result_type), r_entryframe_fp);
 162       // we will save arguments_tos_address later
 163 
 164 
 165       BLOCK_COMMENT("Copy Java arguments");
 166       // copy Java arguments
 167 
 168       // Calculate top_of_arguments_addr which will be R17_tos (not prepushed) later.
 169       // FIXME: why not simply use SP+frame::top_ijava_frame_size?
 170       __ addi(r_top_of_arguments_addr,
 171               R1_SP, frame::top_ijava_frame_abi_size);
 172       __ add(r_top_of_arguments_addr,
 173              r_top_of_arguments_addr, r_frame_alignment_in_bytes);
 174 
 175       // any arguments to copy?
 176       __ cmpdi(CCR0, r_arg_argument_count, 0);
 177       __ beq(CCR0, arguments_copied);
 178 
 179       // prepare loop and copy arguments in reverse order
 180       {
 181         // init CTR with arg_argument_count
 182         __ mtctr(r_arg_argument_count);
 183 
 184         // let r_argumentcopy_addr point to last outgoing Java arguments P
 185         __ mr(r_argumentcopy_addr, r_top_of_arguments_addr);
 186 
 187         // let r_argument_addr point to last incoming java argument
 188         __ add(r_argument_addr,
 189                    r_arg_argument_addr, r_argument_size_in_bytes);
 190         __ addi(r_argument_addr, r_argument_addr, -BytesPerWord);
 191 
 192         // now loop while CTR > 0 and copy arguments
 193         {
 194           Label next_argument;
 195           __ bind(next_argument);
 196 
 197           __ ld(r_temp, 0, r_argument_addr);
 198           // argument_addr--;
 199           __ addi(r_argument_addr, r_argument_addr, -BytesPerWord);
 200           __ std(r_temp, 0, r_argumentcopy_addr);
 201           // argumentcopy_addr++;
 202           __ addi(r_argumentcopy_addr, r_argumentcopy_addr, BytesPerWord);
 203 
 204           __ bdnz(next_argument);
 205         }
 206       }
 207 
 208       // Arguments copied, continue.
 209       __ bind(arguments_copied);
 210     }
 211 
 212     {
 213       BLOCK_COMMENT("Call frame manager or native entry.");
 214       // Call frame manager or native entry.
 215       Register r_new_arg_entry = R14;
 216       assert_different_registers(r_new_arg_entry, r_top_of_arguments_addr,
 217                                  r_arg_method, r_arg_thread);
 218 
 219       __ mr(r_new_arg_entry, r_arg_entry);
 220 
 221       // Register state on entry to frame manager / native entry:
 222       //
 223       //   tos         -  intptr_t*    sender tos (prepushed) Lesp = (SP) + copied_arguments_offset - 8
 224       //   R19_method  -  Method
 225       //   R16_thread  -  JavaThread*
 226 
 227       // Tos must point to last argument - element_size.
 228 #ifdef CC_INTERP
 229       const Register tos = R17_tos;
 230 #else
 231       const Register tos = R15_esp;
 232 #endif
 233       __ addi(tos, r_top_of_arguments_addr, -Interpreter::stackElementSize);
 234 
 235       // initialize call_stub locals (step 2)
 236       // now save tos as arguments_tos_address
 237       __ std(tos, _entry_frame_locals_neg(arguments_tos_address), r_entryframe_fp);
 238 
 239       // load argument registers for call
 240       __ mr(R19_method, r_arg_method);
 241       __ mr(R16_thread, r_arg_thread);
 242       assert(tos != r_arg_method, "trashed r_arg_method");
 243       assert(tos != r_arg_thread && R19_method != r_arg_thread, "trashed r_arg_thread");
 244 
 245       // Set R15_prev_state to 0 for simplifying checks in callee.
 246 #ifdef CC_INTERP
 247       __ li(R15_prev_state, 0);
 248 #else
 249       __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
 250 #endif
 251       // Stack on entry to frame manager / native entry:
 252       //
 253       //      F0      [TOP_IJAVA_FRAME_ABI]
 254       //              alignment (optional)
 255       //              [outgoing Java arguments]
 256       //              [ENTRY_FRAME_LOCALS]
 257       //      F1      [C_FRAME]
 258       //              ...
 259       //
 260 
 261       // global toc register
 262       __ load_const(R29, MacroAssembler::global_toc(), R11_scratch1);
 263 
 264       // Load narrow oop base.
 265       __ reinit_heapbase(R30, R11_scratch1);
 266 
 267       // Remember the senderSP so we interpreter can pop c2i arguments off of the stack
 268       // when called via a c2i.
 269 
 270       // Pass initial_caller_sp to framemanager.
 271       __ mr(R21_tmp1, R1_SP);
 272 
 273       // Do a light-weight C-call here, r_new_arg_entry holds the address
 274       // of the interpreter entry point (frame manager or native entry)
 275       // and save runtime-value of LR in return_address.
 276       assert(r_new_arg_entry != tos && r_new_arg_entry != R19_method && r_new_arg_entry != R16_thread,
 277              "trashed r_new_arg_entry");
 278       return_address = __ call_stub(r_new_arg_entry);
 279     }
 280 
 281     {
 282       BLOCK_COMMENT("Returned from frame manager or native entry.");
 283       // Returned from frame manager or native entry.
 284       // Now pop frame, process result, and return to caller.
 285 
 286       // Stack on exit from frame manager / native entry:
 287       //
 288       //      F0      [ABI]
 289       //              ...
 290       //              [ENTRY_FRAME_LOCALS]
 291       //      F1      [C_FRAME]
 292       //              ...
 293       //
 294       // Just pop the topmost frame ...
 295       //
 296 
 297       Label ret_is_object;
 298       Label ret_is_long;
 299       Label ret_is_float;
 300       Label ret_is_double;
 301 
 302       Register r_entryframe_fp = R30;
 303       Register r_lr            = R7_ARG5;
 304       Register r_cr            = R8_ARG6;
 305 
 306       // Reload some volatile registers which we've spilled before the call
 307       // to frame manager / native entry.
 308       // Access all locals via frame pointer, because we know nothing about
 309       // the topmost frame's size.
 310       __ ld(r_entryframe_fp, _abi(callers_sp), R1_SP);
 311       assert_different_registers(r_entryframe_fp, R3_RET, r_arg_result_addr, r_arg_result_type, r_cr, r_lr);
 312       __ ld(r_arg_result_addr,
 313             _entry_frame_locals_neg(result_address), r_entryframe_fp);
 314       __ ld(r_arg_result_type,
 315             _entry_frame_locals_neg(result_type), r_entryframe_fp);
 316       __ ld(r_cr, _abi(cr), r_entryframe_fp);
 317       __ ld(r_lr, _abi(lr), r_entryframe_fp);
 318 
 319       // pop frame and restore non-volatiles, LR and CR
 320       __ mr(R1_SP, r_entryframe_fp);
 321       __ mtcr(r_cr);
 322       __ mtlr(r_lr);
 323 
 324       // Store result depending on type. Everything that is not
 325       // T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE is treated as T_INT.
 326       __ cmpwi(CCR0, r_arg_result_type, T_OBJECT);
 327       __ cmpwi(CCR1, r_arg_result_type, T_LONG);
 328       __ cmpwi(CCR5, r_arg_result_type, T_FLOAT);
 329       __ cmpwi(CCR6, r_arg_result_type, T_DOUBLE);
 330 
 331       // restore non-volatile registers
 332       __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
 333 
 334 
 335       // Stack on exit from call_stub:
 336       //
 337       //      0       [C_FRAME]
 338       //              ...
 339       //
 340       //  no call_stub frames left.
 341 
 342       // All non-volatiles have been restored at this point!!
 343       assert(R3_RET == R3, "R3_RET should be R3");
 344 
 345       __ beq(CCR0, ret_is_object);
 346       __ beq(CCR1, ret_is_long);
 347       __ beq(CCR5, ret_is_float);
 348       __ beq(CCR6, ret_is_double);
 349 
 350       // default:
 351       __ stw(R3_RET, 0, r_arg_result_addr);
 352       __ blr(); // return to caller
 353 
 354       // case T_OBJECT:
 355       __ bind(ret_is_object);
 356       __ std(R3_RET, 0, r_arg_result_addr);
 357       __ blr(); // return to caller
 358 
 359       // case T_LONG:
 360       __ bind(ret_is_long);
 361       __ std(R3_RET, 0, r_arg_result_addr);
 362       __ blr(); // return to caller
 363 
 364       // case T_FLOAT:
 365       __ bind(ret_is_float);
 366       __ stfs(F1_RET, 0, r_arg_result_addr);
 367       __ blr(); // return to caller
 368 
 369       // case T_DOUBLE:
 370       __ bind(ret_is_double);
 371       __ stfd(F1_RET, 0, r_arg_result_addr);
 372       __ blr(); // return to caller
 373     }
 374 
 375     return start;
 376   }
 377 
 378   // Return point for a Java call if there's an exception thrown in
 379   // Java code.  The exception is caught and transformed into a
 380   // pending exception stored in JavaThread that can be tested from
 381   // within the VM.
 382   //
 383   address generate_catch_exception() {
 384     StubCodeMark mark(this, "StubRoutines", "catch_exception");
 385 
 386     address start = __ pc();
 387 
 388     // Registers alive
 389     //
 390     //  R16_thread
 391     //  R3_ARG1 - address of pending exception
 392     //  R4_ARG2 - return address in call stub
 393 
 394     const Register exception_file = R21_tmp1;
 395     const Register exception_line = R22_tmp2;
 396 
 397     __ load_const(exception_file, (void*)__FILE__);
 398     __ load_const(exception_line, (void*)__LINE__);
 399 
 400     __ std(R3_ARG1, thread_(pending_exception));
 401     // store into `char *'
 402     __ std(exception_file, thread_(exception_file));
 403     // store into `int'
 404     __ stw(exception_line, thread_(exception_line));
 405 
 406     // complete return to VM
 407     assert(StubRoutines::_call_stub_return_address != NULL, "must have been generated before");
 408 
 409     __ mtlr(R4_ARG2);
 410     // continue in call stub
 411     __ blr();
 412 
 413     return start;
 414   }
 415 
 416   // Continuation point for runtime calls returning with a pending
 417   // exception.  The pending exception check happened in the runtime
 418   // or native call stub.  The pending exception in Thread is
 419   // converted into a Java-level exception.
 420   //
 421   address generate_forward_exception() {
 422     StubCodeMark mark(this, "StubRoutines", "forward_exception");
 423     address start = __ pc();
 424 
 425 #if !defined(PRODUCT)
 426     if (VerifyOops) {
 427       // Get pending exception oop.
 428       __ ld(R3_ARG1,
 429                 in_bytes(Thread::pending_exception_offset()),
 430                 R16_thread);
 431       // Make sure that this code is only executed if there is a pending exception.
 432       {
 433         Label L;
 434         __ cmpdi(CCR0, R3_ARG1, 0);
 435         __ bne(CCR0, L);
 436         __ stop("StubRoutines::forward exception: no pending exception (1)");
 437         __ bind(L);
 438       }
 439       __ verify_oop(R3_ARG1, "StubRoutines::forward exception: not an oop");
 440     }
 441 #endif
 442 
 443     // Save LR/CR and copy exception pc (LR) into R4_ARG2.
 444     __ save_LR_CR(R4_ARG2);
 445     __ push_frame_reg_args(0, R0);
 446     // Find exception handler.
 447     __ call_VM_leaf(CAST_FROM_FN_PTR(address,
 448                      SharedRuntime::exception_handler_for_return_address),
 449                     R16_thread,
 450                     R4_ARG2);
 451     // Copy handler's address.
 452     __ mtctr(R3_RET);
 453     __ pop_frame();
 454     __ restore_LR_CR(R0);
 455 
 456     // Set up the arguments for the exception handler:
 457     //  - R3_ARG1: exception oop
 458     //  - R4_ARG2: exception pc.
 459 
 460     // Load pending exception oop.
 461     __ ld(R3_ARG1,
 462               in_bytes(Thread::pending_exception_offset()),
 463               R16_thread);
 464 
 465     // The exception pc is the return address in the caller.
 466     // Must load it into R4_ARG2.
 467     __ mflr(R4_ARG2);
 468 
 469 #ifdef ASSERT
 470     // Make sure exception is set.
 471     {
 472       Label L;
 473       __ cmpdi(CCR0, R3_ARG1, 0);
 474       __ bne(CCR0, L);
 475       __ stop("StubRoutines::forward exception: no pending exception (2)");
 476       __ bind(L);
 477     }
 478 #endif
 479 
 480     // Clear the pending exception.
 481     __ li(R0, 0);
 482     __ std(R0,
 483                in_bytes(Thread::pending_exception_offset()),
 484                R16_thread);
 485     // Jump to exception handler.
 486     __ bctr();
 487 
 488     return start;
 489   }
 490 
 491 #undef __
 492 #define __ masm->
 493   // Continuation point for throwing of implicit exceptions that are
 494   // not handled in the current activation. Fabricates an exception
 495   // oop and initiates normal exception dispatching in this
 496   // frame. Only callee-saved registers are preserved (through the
 497   // normal register window / RegisterMap handling).  If the compiler
 498   // needs all registers to be preserved between the fault point and
 499   // the exception handler then it must assume responsibility for that
 500   // in AbstractCompiler::continuation_for_implicit_null_exception or
 501   // continuation_for_implicit_division_by_zero_exception. All other
 502   // implicit exceptions (e.g., NullPointerException or
 503   // AbstractMethodError on entry) are either at call sites or
 504   // otherwise assume that stack unwinding will be initiated, so
 505   // caller saved registers were assumed volatile in the compiler.
 506   //
 507   // Note that we generate only this stub into a RuntimeStub, because
 508   // it needs to be properly traversed and ignored during GC, so we
 509   // change the meaning of the "__" macro within this method.
 510   //
 511   // Note: the routine set_pc_not_at_call_for_caller in
 512   // SharedRuntime.cpp requires that this code be generated into a
 513   // RuntimeStub.
 514   address generate_throw_exception(const char* name, address runtime_entry, bool restore_saved_exception_pc,
 515                                    Register arg1 = noreg, Register arg2 = noreg) {
 516     CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
 517     MacroAssembler* masm = new MacroAssembler(&code);
 518 
 519     OopMapSet* oop_maps  = new OopMapSet();
 520     int frame_size_in_bytes = frame::abi_reg_args_size;
 521     OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
 522 
 523     StubCodeMark mark(this, "StubRoutines", "throw_exception");
 524 
 525     address start = __ pc();
 526 
 527     __ save_LR_CR(R11_scratch1);
 528 
 529     // Push a frame.
 530     __ push_frame_reg_args(0, R11_scratch1);
 531 
 532     address frame_complete_pc = __ pc();
 533 
 534     if (restore_saved_exception_pc) {
 535       __ unimplemented("StubGenerator::throw_exception with restore_saved_exception_pc", 74);
 536     }
 537 
 538     // Note that we always have a runtime stub frame on the top of
 539     // stack by this point. Remember the offset of the instruction
 540     // whose address will be moved to R11_scratch1.
 541     address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
 542 
 543     __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
 544 
 545     __ mr(R3_ARG1, R16_thread);
 546     if (arg1 != noreg) {
 547       __ mr(R4_ARG2, arg1);
 548     }
 549     if (arg2 != noreg) {
 550       __ mr(R5_ARG3, arg2);
 551     }
 552 #if defined(ABI_ELFv2)
 553     __ call_c(runtime_entry, relocInfo::none);
 554 #else
 555     __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, runtime_entry), relocInfo::none);
 556 #endif
 557 
 558     // Set an oopmap for the call site.
 559     oop_maps->add_gc_map((int)(gc_map_pc - start), map);
 560 
 561     __ reset_last_Java_frame();
 562 
 563 #ifdef ASSERT
 564     // Make sure that this code is only executed if there is a pending
 565     // exception.
 566     {
 567       Label L;
 568       __ ld(R0,
 569                 in_bytes(Thread::pending_exception_offset()),
 570                 R16_thread);
 571       __ cmpdi(CCR0, R0, 0);
 572       __ bne(CCR0, L);
 573       __ stop("StubRoutines::throw_exception: no pending exception");
 574       __ bind(L);
 575     }
 576 #endif
 577 
 578     // Pop frame.
 579     __ pop_frame();
 580 
 581     __ restore_LR_CR(R11_scratch1);
 582 
 583     __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
 584     __ mtctr(R11_scratch1);
 585     __ bctr();
 586 
 587     // Create runtime stub with OopMap.
 588     RuntimeStub* stub =
 589       RuntimeStub::new_runtime_stub(name, &code,
 590                                     /*frame_complete=*/ (int)(frame_complete_pc - start),
 591                                     frame_size_in_bytes/wordSize,
 592                                     oop_maps,
 593                                     false);
 594     return stub->entry_point();
 595   }
 596 #undef __
 597 #define __ _masm->
 598 
 599   //  Generate G1 pre-write barrier for array.
 600   //
 601   //  Input:
 602   //     from     - register containing src address (only needed for spilling)
 603   //     to       - register containing starting address
 604   //     count    - register containing element count
 605   //     tmp      - scratch register
 606   //
 607   //  Kills:
 608   //     nothing
 609   //
 610   void gen_write_ref_array_pre_barrier(Register from, Register to, Register count, bool dest_uninitialized, Register Rtmp1) {
 611     BarrierSet* const bs = Universe::heap()->barrier_set();
 612     switch (bs->kind()) {
 613       case BarrierSet::G1SATBCT:
 614       case BarrierSet::G1SATBCTLogging:
 615         // With G1, don't generate the call if we statically know that the target in uninitialized
 616         if (!dest_uninitialized) {
 617           const int spill_slots = 4 * wordSize;
 618           const int frame_size  = frame::abi_reg_args_size + spill_slots;
 619           Label filtered;
 620 
 621           // Is marking active?
 622           if (in_bytes(PtrQueue::byte_width_of_active()) == 4) {
 623             __ lwz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread);
 624           } else {
 625             guarantee(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption");
 626             __ lbz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread);
 627           }
 628           __ cmpdi(CCR0, Rtmp1, 0);
 629           __ beq(CCR0, filtered);
 630 
 631           __ save_LR_CR(R0);
 632           __ push_frame_reg_args(spill_slots, R0);
 633           __ std(from,  frame_size - 1 * wordSize, R1_SP);
 634           __ std(to,    frame_size - 2 * wordSize, R1_SP);
 635           __ std(count, frame_size - 3 * wordSize, R1_SP);
 636 
 637           __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), to, count);
 638 
 639           __ ld(from,  frame_size - 1 * wordSize, R1_SP);
 640           __ ld(to,    frame_size - 2 * wordSize, R1_SP);
 641           __ ld(count, frame_size - 3 * wordSize, R1_SP);
 642           __ pop_frame();
 643           __ restore_LR_CR(R0);
 644 
 645           __ bind(filtered);
 646         }
 647         break;
 648       case BarrierSet::CardTableModRef:
 649       case BarrierSet::CardTableExtension:
 650       case BarrierSet::ModRef:
 651         break;
 652       default:
 653         ShouldNotReachHere();
 654     }
 655   }
 656 
 657   //  Generate CMS/G1 post-write barrier for array.
 658   //
 659   //  Input:
 660   //     addr     - register containing starting address
 661   //     count    - register containing element count
 662   //     tmp      - scratch register
 663   //
 664   //  The input registers and R0 are overwritten.
 665   //
 666   void gen_write_ref_array_post_barrier(Register addr, Register count, Register tmp, bool branchToEnd) {
 667     BarrierSet* const bs = Universe::heap()->barrier_set();
 668 
 669     switch (bs->kind()) {
 670       case BarrierSet::G1SATBCT:
 671       case BarrierSet::G1SATBCTLogging:
 672         {
 673           if (branchToEnd) {
 674             __ save_LR_CR(R0);
 675             // We need this frame only to spill LR.
 676             __ push_frame_reg_args(0, R0);
 677             __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), addr, count);
 678             __ pop_frame();
 679             __ restore_LR_CR(R0);
 680           } else {
 681             // Tail call: fake call from stub caller by branching without linking.
 682             address entry_point = (address)CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post);
 683             __ mr_if_needed(R3_ARG1, addr);
 684             __ mr_if_needed(R4_ARG2, count);
 685             __ load_const(R11, entry_point, R0);
 686             __ call_c_and_return_to_caller(R11);
 687           }
 688         }
 689         break;
 690       case BarrierSet::CardTableModRef:
 691       case BarrierSet::CardTableExtension:
 692         {
 693           Label Lskip_loop, Lstore_loop;
 694           if (UseConcMarkSweepGC) {
 695             // TODO PPC port: contribute optimization / requires shared changes
 696             __ release();
 697           }
 698 
 699           CardTableModRefBS* const ct = (CardTableModRefBS*)bs;
 700           assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
 701           assert_different_registers(addr, count, tmp);
 702 
 703           __ sldi(count, count, LogBytesPerHeapOop);
 704           __ addi(count, count, -BytesPerHeapOop);
 705           __ add(count, addr, count);
 706           // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.)
 707           __ srdi(addr, addr, CardTableModRefBS::card_shift);
 708           __ srdi(count, count, CardTableModRefBS::card_shift);
 709           __ subf(count, addr, count);
 710           assert_different_registers(R0, addr, count, tmp);
 711           __ load_const(tmp, (address)ct->byte_map_base);
 712           __ addic_(count, count, 1);
 713           __ beq(CCR0, Lskip_loop);
 714           __ li(R0, 0);
 715           __ mtctr(count);
 716           // Byte store loop
 717           __ bind(Lstore_loop);
 718           __ stbx(R0, tmp, addr);
 719           __ addi(addr, addr, 1);
 720           __ bdnz(Lstore_loop);
 721           __ bind(Lskip_loop);
 722 
 723           if (!branchToEnd) __ blr();
 724         }
 725       break;
 726       case BarrierSet::ModRef:
 727         if (!branchToEnd) __ blr();
 728         break;
 729       default:
 730         ShouldNotReachHere();
 731     }
 732   }
 733 
 734   // Support for void zero_words_aligned8(HeapWord* to, size_t count)
 735   //
 736   // Arguments:
 737   //   to:
 738   //   count:
 739   //
 740   // Destroys:
 741   //
 742   address generate_zero_words_aligned8() {
 743     StubCodeMark mark(this, "StubRoutines", "zero_words_aligned8");
 744 
 745     // Implemented as in ClearArray.
 746     address start = __ function_entry();
 747 
 748     Register base_ptr_reg   = R3_ARG1; // tohw (needs to be 8b aligned)
 749     Register cnt_dwords_reg = R4_ARG2; // count (in dwords)
 750     Register tmp1_reg       = R5_ARG3;
 751     Register tmp2_reg       = R6_ARG4;
 752     Register zero_reg       = R7_ARG5;
 753 
 754     // Procedure for large arrays (uses data cache block zero instruction).
 755     Label dwloop, fast, fastloop, restloop, lastdword, done;
 756     int cl_size=VM_Version::get_cache_line_size(), cl_dwords=cl_size>>3, cl_dwordaddr_bits=exact_log2(cl_dwords);
 757     int min_dcbz=2; // Needs to be positive, apply dcbz only to at least min_dcbz cache lines.
 758 
 759     // Clear up to 128byte boundary if long enough, dword_cnt=(16-(base>>3))%16.
 760     __ dcbtst(base_ptr_reg);                    // Indicate write access to first cache line ...
 761     __ andi(tmp2_reg, cnt_dwords_reg, 1);       // to check if number of dwords is even.
 762     __ srdi_(tmp1_reg, cnt_dwords_reg, 1);      // number of double dwords
 763     __ load_const_optimized(zero_reg, 0L);      // Use as zero register.
 764 
 765     __ cmpdi(CCR1, tmp2_reg, 0);                // cnt_dwords even?
 766     __ beq(CCR0, lastdword);                    // size <= 1
 767     __ mtctr(tmp1_reg);                         // Speculatively preload counter for rest loop (>0).
 768     __ cmpdi(CCR0, cnt_dwords_reg, (min_dcbz+1)*cl_dwords-1); // Big enough to ensure >=min_dcbz cache lines are included?
 769     __ neg(tmp1_reg, base_ptr_reg);             // bit 0..58: bogus, bit 57..60: (16-(base>>3))%16, bit 61..63: 000
 770 
 771     __ blt(CCR0, restloop);                     // Too small. (<31=(2*cl_dwords)-1 is sufficient, but bigger performs better.)
 772     __ rldicl_(tmp1_reg, tmp1_reg, 64-3, 64-cl_dwordaddr_bits); // Extract number of dwords to 128byte boundary=(16-(base>>3))%16.
 773 
 774     __ beq(CCR0, fast);                         // already 128byte aligned
 775     __ mtctr(tmp1_reg);                         // Set ctr to hit 128byte boundary (0<ctr<cnt).
 776     __ subf(cnt_dwords_reg, tmp1_reg, cnt_dwords_reg); // rest (>0 since size>=256-8)
 777 
 778     // Clear in first cache line dword-by-dword if not already 128byte aligned.
 779     __ bind(dwloop);
 780       __ std(zero_reg, 0, base_ptr_reg);        // Clear 8byte aligned block.
 781       __ addi(base_ptr_reg, base_ptr_reg, 8);
 782     __ bdnz(dwloop);
 783 
 784     // clear 128byte blocks
 785     __ bind(fast);
 786     __ srdi(tmp1_reg, cnt_dwords_reg, cl_dwordaddr_bits); // loop count for 128byte loop (>0 since size>=256-8)
 787     __ andi(tmp2_reg, cnt_dwords_reg, 1);       // to check if rest even
 788 
 789     __ mtctr(tmp1_reg);                         // load counter
 790     __ cmpdi(CCR1, tmp2_reg, 0);                // rest even?
 791     __ rldicl_(tmp1_reg, cnt_dwords_reg, 63, 65-cl_dwordaddr_bits); // rest in double dwords
 792 
 793     __ bind(fastloop);
 794       __ dcbz(base_ptr_reg);                    // Clear 128byte aligned block.
 795       __ addi(base_ptr_reg, base_ptr_reg, cl_size);
 796     __ bdnz(fastloop);
 797 
 798     //__ dcbtst(base_ptr_reg);                  // Indicate write access to last cache line.
 799     __ beq(CCR0, lastdword);                    // rest<=1
 800     __ mtctr(tmp1_reg);                         // load counter
 801 
 802     // Clear rest.
 803     __ bind(restloop);
 804       __ std(zero_reg, 0, base_ptr_reg);        // Clear 8byte aligned block.
 805       __ std(zero_reg, 8, base_ptr_reg);        // Clear 8byte aligned block.
 806       __ addi(base_ptr_reg, base_ptr_reg, 16);
 807     __ bdnz(restloop);
 808 
 809     __ bind(lastdword);
 810     __ beq(CCR1, done);
 811     __ std(zero_reg, 0, base_ptr_reg);
 812     __ bind(done);
 813     __ blr();                                   // return
 814 
 815     return start;
 816   }
 817 
 818   // The following routine generates a subroutine to throw an asynchronous
 819   // UnknownError when an unsafe access gets a fault that could not be
 820   // reasonably prevented by the programmer.  (Example: SIGBUS/OBJERR.)
 821   //
 822   address generate_handler_for_unsafe_access() {
 823     StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
 824     address start = __ function_entry();
 825     __ unimplemented("StubRoutines::handler_for_unsafe_access", 93);
 826     return start;
 827   }
 828 
 829 #if !defined(PRODUCT)
 830   // Wrapper which calls oopDesc::is_oop_or_null()
 831   // Only called by MacroAssembler::verify_oop
 832   static void verify_oop_helper(const char* message, oop o) {
 833     if (!o->is_oop_or_null()) {
 834       fatal(message);
 835     }
 836     ++ StubRoutines::_verify_oop_count;
 837   }
 838 #endif
 839 
 840   // Return address of code to be called from code generated by
 841   // MacroAssembler::verify_oop.
 842   //
 843   // Don't generate, rather use C++ code.
 844   address generate_verify_oop() {
 845     StubCodeMark mark(this, "StubRoutines", "verify_oop");
 846 
 847     // this is actually a `FunctionDescriptor*'.
 848     address start = 0;
 849 
 850 #if !defined(PRODUCT)
 851     start = CAST_FROM_FN_PTR(address, verify_oop_helper);
 852 #endif
 853 
 854     return start;
 855   }
 856 
 857   // Fairer handling of safepoints for native methods.
 858   //
 859   // Generate code which reads from the polling page. This special handling is needed as the
 860   // linux-ppc64 kernel before 2.6.6 doesn't set si_addr on some segfaults in 64bit mode
 861   // (cf. http://www.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.6), especially when we try
 862   // to read from the safepoint polling page.
 863   address generate_load_from_poll() {
 864     StubCodeMark mark(this, "StubRoutines", "generate_load_from_poll");
 865     address start = __ function_entry();
 866     __ unimplemented("StubRoutines::verify_oop", 95);  // TODO PPC port
 867     return start;
 868   }
 869 
 870   // -XX:+OptimizeFill : convert fill/copy loops into intrinsic
 871   //
 872   // The code is implemented(ported from sparc) as we believe it benefits JVM98, however
 873   // tracing(-XX:+TraceOptimizeFill) shows the intrinsic replacement doesn't happen at all!
 874   //
 875   // Source code in function is_range_check_if() shows that OptimizeFill relaxed the condition
 876   // for turning on loop predication optimization, and hence the behavior of "array range check"
 877   // and "loop invariant check" could be influenced, which potentially boosted JVM98.
 878   //
 879   // Generate stub for disjoint short fill. If "aligned" is true, the
 880   // "to" address is assumed to be heapword aligned.
 881   //
 882   // Arguments for generated stub:
 883   //   to:    R3_ARG1
 884   //   value: R4_ARG2
 885   //   count: R5_ARG3 treated as signed
 886   //
 887   address generate_fill(BasicType t, bool aligned, const char* name) {
 888     StubCodeMark mark(this, "StubRoutines", name);
 889     address start = __ function_entry();
 890 
 891     const Register to    = R3_ARG1;   // source array address
 892     const Register value = R4_ARG2;   // fill value
 893     const Register count = R5_ARG3;   // elements count
 894     const Register temp  = R6_ARG4;   // temp register
 895 
 896     //assert_clean_int(count, O3);    // Make sure 'count' is clean int.
 897 
 898     Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
 899     Label L_fill_2_bytes, L_fill_4_bytes, L_fill_elements, L_fill_32_bytes;
 900 
 901     int shift = -1;
 902     switch (t) {
 903        case T_BYTE:
 904         shift = 2;
 905         // Clone bytes (zero extend not needed because store instructions below ignore high order bytes).
 906         __ rldimi(value, value, 8, 48);     // 8 bit -> 16 bit
 907         __ cmpdi(CCR0, count, 2<<shift);    // Short arrays (< 8 bytes) fill by element.
 908         __ blt(CCR0, L_fill_elements);
 909         __ rldimi(value, value, 16, 32);    // 16 bit -> 32 bit
 910         break;
 911        case T_SHORT:
 912         shift = 1;
 913         // Clone bytes (zero extend not needed because store instructions below ignore high order bytes).
 914         __ rldimi(value, value, 16, 32);    // 16 bit -> 32 bit
 915         __ cmpdi(CCR0, count, 2<<shift);    // Short arrays (< 8 bytes) fill by element.
 916         __ blt(CCR0, L_fill_elements);
 917         break;
 918       case T_INT:
 919         shift = 0;
 920         __ cmpdi(CCR0, count, 2<<shift);    // Short arrays (< 8 bytes) fill by element.
 921         __ blt(CCR0, L_fill_4_bytes);
 922         break;
 923       default: ShouldNotReachHere();
 924     }
 925 
 926     if (!aligned && (t == T_BYTE || t == T_SHORT)) {
 927       // Align source address at 4 bytes address boundary.
 928       if (t == T_BYTE) {
 929         // One byte misalignment happens only for byte arrays.
 930         __ andi_(temp, to, 1);
 931         __ beq(CCR0, L_skip_align1);
 932         __ stb(value, 0, to);
 933         __ addi(to, to, 1);
 934         __ addi(count, count, -1);
 935         __ bind(L_skip_align1);
 936       }
 937       // Two bytes misalignment happens only for byte and short (char) arrays.
 938       __ andi_(temp, to, 2);
 939       __ beq(CCR0, L_skip_align2);
 940       __ sth(value, 0, to);
 941       __ addi(to, to, 2);
 942       __ addi(count, count, -(1 << (shift - 1)));
 943       __ bind(L_skip_align2);
 944     }
 945 
 946     if (!aligned) {
 947       // Align to 8 bytes, we know we are 4 byte aligned to start.
 948       __ andi_(temp, to, 7);
 949       __ beq(CCR0, L_fill_32_bytes);
 950       __ stw(value, 0, to);
 951       __ addi(to, to, 4);
 952       __ addi(count, count, -(1 << shift));
 953       __ bind(L_fill_32_bytes);
 954     }
 955 
 956     __ li(temp, 8<<shift);                  // Prepare for 32 byte loop.
 957     // Clone bytes int->long as above.
 958     __ rldimi(value, value, 32, 0);         // 32 bit -> 64 bit
 959 
 960     Label L_check_fill_8_bytes;
 961     // Fill 32-byte chunks.
 962     __ subf_(count, temp, count);
 963     __ blt(CCR0, L_check_fill_8_bytes);
 964 
 965     Label L_fill_32_bytes_loop;
 966     __ align(32);
 967     __ bind(L_fill_32_bytes_loop);
 968 
 969     __ std(value, 0, to);
 970     __ std(value, 8, to);
 971     __ subf_(count, temp, count);           // Update count.
 972     __ std(value, 16, to);
 973     __ std(value, 24, to);
 974 
 975     __ addi(to, to, 32);
 976     __ bge(CCR0, L_fill_32_bytes_loop);
 977 
 978     __ bind(L_check_fill_8_bytes);
 979     __ add_(count, temp, count);
 980     __ beq(CCR0, L_exit);
 981     __ addic_(count, count, -(2 << shift));
 982     __ blt(CCR0, L_fill_4_bytes);
 983 
 984     //
 985     // Length is too short, just fill 8 bytes at a time.
 986     //
 987     Label L_fill_8_bytes_loop;
 988     __ bind(L_fill_8_bytes_loop);
 989     __ std(value, 0, to);
 990     __ addic_(count, count, -(2 << shift));
 991     __ addi(to, to, 8);
 992     __ bge(CCR0, L_fill_8_bytes_loop);
 993 
 994     // Fill trailing 4 bytes.
 995     __ bind(L_fill_4_bytes);
 996     __ andi_(temp, count, 1<<shift);
 997     __ beq(CCR0, L_fill_2_bytes);
 998 
 999     __ stw(value, 0, to);
1000     if (t == T_BYTE || t == T_SHORT) {
1001       __ addi(to, to, 4);
1002       // Fill trailing 2 bytes.
1003       __ bind(L_fill_2_bytes);
1004       __ andi_(temp, count, 1<<(shift-1));
1005       __ beq(CCR0, L_fill_byte);
1006       __ sth(value, 0, to);
1007       if (t == T_BYTE) {
1008         __ addi(to, to, 2);
1009         // Fill trailing byte.
1010         __ bind(L_fill_byte);
1011         __ andi_(count, count, 1);
1012         __ beq(CCR0, L_exit);
1013         __ stb(value, 0, to);
1014       } else {
1015         __ bind(L_fill_byte);
1016       }
1017     } else {
1018       __ bind(L_fill_2_bytes);
1019     }
1020     __ bind(L_exit);
1021     __ blr();
1022 
1023     // Handle copies less than 8 bytes. Int is handled elsewhere.
1024     if (t == T_BYTE) {
1025       __ bind(L_fill_elements);
1026       Label L_fill_2, L_fill_4;
1027       __ andi_(temp, count, 1);
1028       __ beq(CCR0, L_fill_2);
1029       __ stb(value, 0, to);
1030       __ addi(to, to, 1);
1031       __ bind(L_fill_2);
1032       __ andi_(temp, count, 2);
1033       __ beq(CCR0, L_fill_4);
1034       __ stb(value, 0, to);
1035       __ stb(value, 0, to);
1036       __ addi(to, to, 2);
1037       __ bind(L_fill_4);
1038       __ andi_(temp, count, 4);
1039       __ beq(CCR0, L_exit);
1040       __ stb(value, 0, to);
1041       __ stb(value, 1, to);
1042       __ stb(value, 2, to);
1043       __ stb(value, 3, to);
1044       __ blr();
1045     }
1046 
1047     if (t == T_SHORT) {
1048       Label L_fill_2;
1049       __ bind(L_fill_elements);
1050       __ andi_(temp, count, 1);
1051       __ beq(CCR0, L_fill_2);
1052       __ sth(value, 0, to);
1053       __ addi(to, to, 2);
1054       __ bind(L_fill_2);
1055       __ andi_(temp, count, 2);
1056       __ beq(CCR0, L_exit);
1057       __ sth(value, 0, to);
1058       __ sth(value, 2, to);
1059       __ blr();
1060     }
1061     return start;
1062   }
1063 
1064 
1065   // Generate overlap test for array copy stubs.
1066   //
1067   // Input:
1068   //   R3_ARG1    -  from
1069   //   R4_ARG2    -  to
1070   //   R5_ARG3    -  element count
1071   //
1072   void array_overlap_test(address no_overlap_target, int log2_elem_size) {
1073     Register tmp1 = R6_ARG4;
1074     Register tmp2 = R7_ARG5;
1075 
1076     Label l_overlap;
1077 #ifdef ASSERT
1078     __ srdi_(tmp2, R5_ARG3, 31);
1079     __ asm_assert_eq("missing zero extend", 0xAFFE);
1080 #endif
1081 
1082     __ subf(tmp1, R3_ARG1, R4_ARG2); // distance in bytes
1083     __ sldi(tmp2, R5_ARG3, log2_elem_size); // size in bytes
1084     __ cmpld(CCR0, R3_ARG1, R4_ARG2); // Use unsigned comparison!
1085     __ cmpld(CCR1, tmp1, tmp2);
1086     __ crand(/*CCR0 lt*/0, /*CCR1 lt*/4+0, /*CCR0 lt*/0);
1087     __ blt(CCR0, l_overlap); // Src before dst and distance smaller than size.
1088 
1089     // need to copy forwards
1090     if (__ is_within_range_of_b(no_overlap_target, __ pc())) {
1091       __ b(no_overlap_target);
1092     } else {
1093       __ load_const(tmp1, no_overlap_target, tmp2);
1094       __ mtctr(tmp1);
1095       __ bctr();
1096     }
1097 
1098     __ bind(l_overlap);
1099     // need to copy backwards
1100   }
1101 
1102   // The guideline in the implementations of generate_disjoint_xxx_copy
1103   // (xxx=byte,short,int,long,oop) is to copy as many elements as possible with
1104   // single instructions, but to avoid alignment interrupts (see subsequent
1105   // comment). Furthermore, we try to minimize misaligned access, even
1106   // though they cause no alignment interrupt.
1107   //
1108   // In Big-Endian mode, the PowerPC architecture requires implementations to
1109   // handle automatically misaligned integer halfword and word accesses,
1110   // word-aligned integer doubleword accesses, and word-aligned floating-point
1111   // accesses. Other accesses may or may not generate an Alignment interrupt
1112   // depending on the implementation.
1113   // Alignment interrupt handling may require on the order of hundreds of cycles,
1114   // so every effort should be made to avoid misaligned memory values.
1115   //
1116   //
1117   // Generate stub for disjoint byte copy.  If "aligned" is true, the
1118   // "from" and "to" addresses are assumed to be heapword aligned.
1119   //
1120   // Arguments for generated stub:
1121   //      from:  R3_ARG1
1122   //      to:    R4_ARG2
1123   //      count: R5_ARG3 treated as signed
1124   //
1125   address generate_disjoint_byte_copy(bool aligned, const char * name) {
1126     StubCodeMark mark(this, "StubRoutines", name);
1127     address start = __ function_entry();
1128 
1129     Register tmp1 = R6_ARG4;
1130     Register tmp2 = R7_ARG5;
1131     Register tmp3 = R8_ARG6;
1132     Register tmp4 = R9_ARG7;
1133 
1134 
1135     Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8, l_9;
1136     // Don't try anything fancy if arrays don't have many elements.
1137     __ li(tmp3, 0);
1138     __ cmpwi(CCR0, R5_ARG3, 17);
1139     __ ble(CCR0, l_6); // copy 4 at a time
1140 
1141     if (!aligned) {
1142       __ xorr(tmp1, R3_ARG1, R4_ARG2);
1143       __ andi_(tmp1, tmp1, 3);
1144       __ bne(CCR0, l_6); // If arrays don't have the same alignment mod 4, do 4 element copy.
1145 
1146       // Copy elements if necessary to align to 4 bytes.
1147       __ neg(tmp1, R3_ARG1); // Compute distance to alignment boundary.
1148       __ andi_(tmp1, tmp1, 3);
1149       __ beq(CCR0, l_2);
1150 
1151       __ subf(R5_ARG3, tmp1, R5_ARG3);
1152       __ bind(l_9);
1153       __ lbz(tmp2, 0, R3_ARG1);
1154       __ addic_(tmp1, tmp1, -1);
1155       __ stb(tmp2, 0, R4_ARG2);
1156       __ addi(R3_ARG1, R3_ARG1, 1);
1157       __ addi(R4_ARG2, R4_ARG2, 1);
1158       __ bne(CCR0, l_9);
1159 
1160       __ bind(l_2);
1161     }
1162 
1163     // copy 8 elements at a time
1164     __ xorr(tmp2, R3_ARG1, R4_ARG2); // skip if src & dest have differing alignment mod 8
1165     __ andi_(tmp1, tmp2, 7);
1166     __ bne(CCR0, l_7); // not same alignment -> to or from is aligned -> copy 8
1167 
1168     // copy a 2-element word if necessary to align to 8 bytes
1169     __ andi_(R0, R3_ARG1, 7);
1170     __ beq(CCR0, l_7);
1171 
1172     __ lwzx(tmp2, R3_ARG1, tmp3);
1173     __ addi(R5_ARG3, R5_ARG3, -4);
1174     __ stwx(tmp2, R4_ARG2, tmp3);
1175     { // FasterArrayCopy
1176       __ addi(R3_ARG1, R3_ARG1, 4);
1177       __ addi(R4_ARG2, R4_ARG2, 4);
1178     }
1179     __ bind(l_7);
1180 
1181     { // FasterArrayCopy
1182       __ cmpwi(CCR0, R5_ARG3, 31);
1183       __ ble(CCR0, l_6); // copy 2 at a time if less than 32 elements remain
1184 
1185       __ srdi(tmp1, R5_ARG3, 5);
1186       __ andi_(R5_ARG3, R5_ARG3, 31);
1187       __ mtctr(tmp1);
1188 
1189       __ bind(l_8);
1190       // Use unrolled version for mass copying (copy 32 elements a time)
1191       // Load feeding store gets zero latency on Power6, however not on Power5.
1192       // Therefore, the following sequence is made for the good of both.
1193       __ ld(tmp1, 0, R3_ARG1);
1194       __ ld(tmp2, 8, R3_ARG1);
1195       __ ld(tmp3, 16, R3_ARG1);
1196       __ ld(tmp4, 24, R3_ARG1);
1197       __ std(tmp1, 0, R4_ARG2);
1198       __ std(tmp2, 8, R4_ARG2);
1199       __ std(tmp3, 16, R4_ARG2);
1200       __ std(tmp4, 24, R4_ARG2);
1201       __ addi(R3_ARG1, R3_ARG1, 32);
1202       __ addi(R4_ARG2, R4_ARG2, 32);
1203       __ bdnz(l_8);
1204     }
1205 
1206     __ bind(l_6);
1207 
1208     // copy 4 elements at a time
1209     __ cmpwi(CCR0, R5_ARG3, 4);
1210     __ blt(CCR0, l_1);
1211     __ srdi(tmp1, R5_ARG3, 2);
1212     __ mtctr(tmp1); // is > 0
1213     __ andi_(R5_ARG3, R5_ARG3, 3);
1214 
1215     { // FasterArrayCopy
1216       __ addi(R3_ARG1, R3_ARG1, -4);
1217       __ addi(R4_ARG2, R4_ARG2, -4);
1218       __ bind(l_3);
1219       __ lwzu(tmp2, 4, R3_ARG1);
1220       __ stwu(tmp2, 4, R4_ARG2);
1221       __ bdnz(l_3);
1222       __ addi(R3_ARG1, R3_ARG1, 4);
1223       __ addi(R4_ARG2, R4_ARG2, 4);
1224     }
1225 
1226     // do single element copy
1227     __ bind(l_1);
1228     __ cmpwi(CCR0, R5_ARG3, 0);
1229     __ beq(CCR0, l_4);
1230 
1231     { // FasterArrayCopy
1232       __ mtctr(R5_ARG3);
1233       __ addi(R3_ARG1, R3_ARG1, -1);
1234       __ addi(R4_ARG2, R4_ARG2, -1);
1235 
1236       __ bind(l_5);
1237       __ lbzu(tmp2, 1, R3_ARG1);
1238       __ stbu(tmp2, 1, R4_ARG2);
1239       __ bdnz(l_5);
1240     }
1241 
1242     __ bind(l_4);
1243     __ blr();
1244 
1245     return start;
1246   }
1247 
1248   // Generate stub for conjoint byte copy.  If "aligned" is true, the
1249   // "from" and "to" addresses are assumed to be heapword aligned.
1250   //
1251   // Arguments for generated stub:
1252   //      from:  R3_ARG1
1253   //      to:    R4_ARG2
1254   //      count: R5_ARG3 treated as signed
1255   //
1256   address generate_conjoint_byte_copy(bool aligned, const char * name) {
1257     StubCodeMark mark(this, "StubRoutines", name);
1258     address start = __ function_entry();
1259 
1260     Register tmp1 = R6_ARG4;
1261     Register tmp2 = R7_ARG5;
1262     Register tmp3 = R8_ARG6;
1263 
1264 #if defined(ABI_ELFv2)
1265      address nooverlap_target = aligned ?
1266        StubRoutines::arrayof_jbyte_disjoint_arraycopy() :
1267        StubRoutines::jbyte_disjoint_arraycopy();
1268 #else
1269     address nooverlap_target = aligned ?
1270       ((FunctionDescriptor*)StubRoutines::arrayof_jbyte_disjoint_arraycopy())->entry() :
1271       ((FunctionDescriptor*)StubRoutines::jbyte_disjoint_arraycopy())->entry();
1272 #endif
1273 
1274     array_overlap_test(nooverlap_target, 0);
1275     // Do reverse copy. We assume the case of actual overlap is rare enough
1276     // that we don't have to optimize it.
1277     Label l_1, l_2;
1278 
1279     __ b(l_2);
1280     __ bind(l_1);
1281     __ stbx(tmp1, R4_ARG2, R5_ARG3);
1282     __ bind(l_2);
1283     __ addic_(R5_ARG3, R5_ARG3, -1);
1284     __ lbzx(tmp1, R3_ARG1, R5_ARG3);
1285     __ bge(CCR0, l_1);
1286 
1287     __ blr();
1288 
1289     return start;
1290   }
1291 
1292   // Generate stub for disjoint short copy.  If "aligned" is true, the
1293   // "from" and "to" addresses are assumed to be heapword aligned.
1294   //
1295   // Arguments for generated stub:
1296   //      from:  R3_ARG1
1297   //      to:    R4_ARG2
1298   //  elm.count: R5_ARG3 treated as signed
1299   //
1300   // Strategy for aligned==true:
1301   //
1302   //  If length <= 9:
1303   //     1. copy 2 elements at a time (l_6)
1304   //     2. copy last element if original element count was odd (l_1)
1305   //
1306   //  If length > 9:
1307   //     1. copy 4 elements at a time until less than 4 elements are left (l_7)
1308   //     2. copy 2 elements at a time until less than 2 elements are left (l_6)
1309   //     3. copy last element if one was left in step 2. (l_1)
1310   //
1311   //
1312   // Strategy for aligned==false:
1313   //
1314   //  If length <= 9: same as aligned==true case, but NOTE: load/stores
1315   //                  can be unaligned (see comment below)
1316   //
1317   //  If length > 9:
1318   //     1. continue with step 6. if the alignment of from and to mod 4
1319   //        is different.
1320   //     2. align from and to to 4 bytes by copying 1 element if necessary
1321   //     3. at l_2 from and to are 4 byte aligned; continue with
1322   //        5. if they cannot be aligned to 8 bytes because they have
1323   //        got different alignment mod 8.
1324   //     4. at this point we know that both, from and to, have the same
1325   //        alignment mod 8, now copy one element if necessary to get
1326   //        8 byte alignment of from and to.
1327   //     5. copy 4 elements at a time until less than 4 elements are
1328   //        left; depending on step 3. all load/stores are aligned or
1329   //        either all loads or all stores are unaligned.
1330   //     6. copy 2 elements at a time until less than 2 elements are
1331   //        left (l_6); arriving here from step 1., there is a chance
1332   //        that all accesses are unaligned.
1333   //     7. copy last element if one was left in step 6. (l_1)
1334   //
1335   //  There are unaligned data accesses using integer load/store
1336   //  instructions in this stub. POWER allows such accesses.
1337   //
1338   //  According to the manuals (PowerISA_V2.06_PUBLIC, Book II,
1339   //  Chapter 2: Effect of Operand Placement on Performance) unaligned
1340   //  integer load/stores have good performance. Only unaligned
1341   //  floating point load/stores can have poor performance.
1342   //
1343   //  TODO:
1344   //
1345   //  1. check if aligning the backbranch target of loops is beneficial
1346   //
1347   address generate_disjoint_short_copy(bool aligned, const char * name) {
1348     StubCodeMark mark(this, "StubRoutines", name);
1349 
1350     Register tmp1 = R6_ARG4;
1351     Register tmp2 = R7_ARG5;
1352     Register tmp3 = R8_ARG6;
1353     Register tmp4 = R9_ARG7;
1354 
1355     address start = __ function_entry();
1356 
1357       Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8;
1358     // don't try anything fancy if arrays don't have many elements
1359     __ li(tmp3, 0);
1360     __ cmpwi(CCR0, R5_ARG3, 9);
1361     __ ble(CCR0, l_6); // copy 2 at a time
1362 
1363     if (!aligned) {
1364       __ xorr(tmp1, R3_ARG1, R4_ARG2);
1365       __ andi_(tmp1, tmp1, 3);
1366       __ bne(CCR0, l_6); // if arrays don't have the same alignment mod 4, do 2 element copy
1367 
1368       // At this point it is guaranteed that both, from and to have the same alignment mod 4.
1369 
1370       // Copy 1 element if necessary to align to 4 bytes.
1371       __ andi_(tmp1, R3_ARG1, 3);
1372       __ beq(CCR0, l_2);
1373 
1374       __ lhz(tmp2, 0, R3_ARG1);
1375       __ addi(R3_ARG1, R3_ARG1, 2);
1376       __ sth(tmp2, 0, R4_ARG2);
1377       __ addi(R4_ARG2, R4_ARG2, 2);
1378       __ addi(R5_ARG3, R5_ARG3, -1);
1379       __ bind(l_2);
1380 
1381       // At this point the positions of both, from and to, are at least 4 byte aligned.
1382 
1383       // Copy 4 elements at a time.
1384       // Align to 8 bytes, but only if both, from and to, have same alignment mod 8.
1385       __ xorr(tmp2, R3_ARG1, R4_ARG2);
1386       __ andi_(tmp1, tmp2, 7);
1387       __ bne(CCR0, l_7); // not same alignment mod 8 -> copy 4, either from or to will be unaligned
1388 
1389       // Copy a 2-element word if necessary to align to 8 bytes.
1390       __ andi_(R0, R3_ARG1, 7);
1391       __ beq(CCR0, l_7);
1392 
1393       __ lwzx(tmp2, R3_ARG1, tmp3);
1394       __ addi(R5_ARG3, R5_ARG3, -2);
1395       __ stwx(tmp2, R4_ARG2, tmp3);
1396       { // FasterArrayCopy
1397         __ addi(R3_ARG1, R3_ARG1, 4);
1398         __ addi(R4_ARG2, R4_ARG2, 4);
1399       }
1400     }
1401 
1402     __ bind(l_7);
1403 
1404     // Copy 4 elements at a time; either the loads or the stores can
1405     // be unaligned if aligned == false.
1406 
1407     { // FasterArrayCopy
1408       __ cmpwi(CCR0, R5_ARG3, 15);
1409       __ ble(CCR0, l_6); // copy 2 at a time if less than 16 elements remain
1410 
1411       __ srdi(tmp1, R5_ARG3, 4);
1412       __ andi_(R5_ARG3, R5_ARG3, 15);
1413       __ mtctr(tmp1);
1414 
1415       __ bind(l_8);
1416       // Use unrolled version for mass copying (copy 16 elements a time).
1417       // Load feeding store gets zero latency on Power6, however not on Power5.
1418       // Therefore, the following sequence is made for the good of both.
1419       __ ld(tmp1, 0, R3_ARG1);
1420       __ ld(tmp2, 8, R3_ARG1);
1421       __ ld(tmp3, 16, R3_ARG1);
1422       __ ld(tmp4, 24, R3_ARG1);
1423       __ std(tmp1, 0, R4_ARG2);
1424       __ std(tmp2, 8, R4_ARG2);
1425       __ std(tmp3, 16, R4_ARG2);
1426       __ std(tmp4, 24, R4_ARG2);
1427       __ addi(R3_ARG1, R3_ARG1, 32);
1428       __ addi(R4_ARG2, R4_ARG2, 32);
1429       __ bdnz(l_8);
1430     }
1431     __ bind(l_6);
1432 
1433     // copy 2 elements at a time
1434     { // FasterArrayCopy
1435       __ cmpwi(CCR0, R5_ARG3, 2);
1436       __ blt(CCR0, l_1);
1437       __ srdi(tmp1, R5_ARG3, 1);
1438       __ andi_(R5_ARG3, R5_ARG3, 1);
1439 
1440       __ addi(R3_ARG1, R3_ARG1, -4);
1441       __ addi(R4_ARG2, R4_ARG2, -4);
1442       __ mtctr(tmp1);
1443 
1444       __ bind(l_3);
1445       __ lwzu(tmp2, 4, R3_ARG1);
1446       __ stwu(tmp2, 4, R4_ARG2);
1447       __ bdnz(l_3);
1448 
1449       __ addi(R3_ARG1, R3_ARG1, 4);
1450       __ addi(R4_ARG2, R4_ARG2, 4);
1451     }
1452 
1453     // do single element copy
1454     __ bind(l_1);
1455     __ cmpwi(CCR0, R5_ARG3, 0);
1456     __ beq(CCR0, l_4);
1457 
1458     { // FasterArrayCopy
1459       __ mtctr(R5_ARG3);
1460       __ addi(R3_ARG1, R3_ARG1, -2);
1461       __ addi(R4_ARG2, R4_ARG2, -2);
1462 
1463       __ bind(l_5);
1464       __ lhzu(tmp2, 2, R3_ARG1);
1465       __ sthu(tmp2, 2, R4_ARG2);
1466       __ bdnz(l_5);
1467     }
1468     __ bind(l_4);
1469     __ blr();
1470 
1471     return start;
1472   }
1473 
1474   // Generate stub for conjoint short copy.  If "aligned" is true, the
1475   // "from" and "to" addresses are assumed to be heapword aligned.
1476   //
1477   // Arguments for generated stub:
1478   //      from:  R3_ARG1
1479   //      to:    R4_ARG2
1480   //      count: R5_ARG3 treated as signed
1481   //
1482   address generate_conjoint_short_copy(bool aligned, const char * name) {
1483     StubCodeMark mark(this, "StubRoutines", name);
1484     address start = __ function_entry();
1485 
1486     Register tmp1 = R6_ARG4;
1487     Register tmp2 = R7_ARG5;
1488     Register tmp3 = R8_ARG6;
1489 
1490 #if defined(ABI_ELFv2)
1491     address nooverlap_target = aligned ?
1492         StubRoutines::arrayof_jshort_disjoint_arraycopy() :
1493         StubRoutines::jshort_disjoint_arraycopy();
1494 #else
1495     address nooverlap_target = aligned ?
1496         ((FunctionDescriptor*)StubRoutines::arrayof_jshort_disjoint_arraycopy())->entry() :
1497         ((FunctionDescriptor*)StubRoutines::jshort_disjoint_arraycopy())->entry();
1498 #endif
1499 
1500     array_overlap_test(nooverlap_target, 1);
1501 
1502     Label l_1, l_2;
1503     __ sldi(tmp1, R5_ARG3, 1);
1504     __ b(l_2);
1505     __ bind(l_1);
1506     __ sthx(tmp2, R4_ARG2, tmp1);
1507     __ bind(l_2);
1508     __ addic_(tmp1, tmp1, -2);
1509     __ lhzx(tmp2, R3_ARG1, tmp1);
1510     __ bge(CCR0, l_1);
1511 
1512     __ blr();
1513 
1514     return start;
1515   }
1516 
1517   // Generate core code for disjoint int copy (and oop copy on 32-bit).  If "aligned"
1518   // is true, the "from" and "to" addresses are assumed to be heapword aligned.
1519   //
1520   // Arguments:
1521   //      from:  R3_ARG1
1522   //      to:    R4_ARG2
1523   //      count: R5_ARG3 treated as signed
1524   //
1525   void generate_disjoint_int_copy_core(bool aligned) {
1526     Register tmp1 = R6_ARG4;
1527     Register tmp2 = R7_ARG5;
1528     Register tmp3 = R8_ARG6;
1529     Register tmp4 = R0;
1530 
1531     Label l_1, l_2, l_3, l_4, l_5, l_6;
1532     // for short arrays, just do single element copy
1533     __ li(tmp3, 0);
1534     __ cmpwi(CCR0, R5_ARG3, 5);
1535     __ ble(CCR0, l_2);
1536 
1537     if (!aligned) {
1538         // check if arrays have same alignment mod 8.
1539         __ xorr(tmp1, R3_ARG1, R4_ARG2);
1540         __ andi_(R0, tmp1, 7);
1541         // Not the same alignment, but ld and std just need to be 4 byte aligned.
1542         __ bne(CCR0, l_4); // to OR from is 8 byte aligned -> copy 2 at a time
1543 
1544         // copy 1 element to align to and from on an 8 byte boundary
1545         __ andi_(R0, R3_ARG1, 7);
1546         __ beq(CCR0, l_4);
1547 
1548         __ lwzx(tmp2, R3_ARG1, tmp3);
1549         __ addi(R5_ARG3, R5_ARG3, -1);
1550         __ stwx(tmp2, R4_ARG2, tmp3);
1551         { // FasterArrayCopy
1552           __ addi(R3_ARG1, R3_ARG1, 4);
1553           __ addi(R4_ARG2, R4_ARG2, 4);
1554         }
1555         __ bind(l_4);
1556       }
1557 
1558     { // FasterArrayCopy
1559       __ cmpwi(CCR0, R5_ARG3, 7);
1560       __ ble(CCR0, l_2); // copy 1 at a time if less than 8 elements remain
1561 
1562       __ srdi(tmp1, R5_ARG3, 3);
1563       __ andi_(R5_ARG3, R5_ARG3, 7);
1564       __ mtctr(tmp1);
1565 
1566       __ bind(l_6);
1567       // Use unrolled version for mass copying (copy 8 elements a time).
1568       // Load feeding store gets zero latency on power6, however not on power 5.
1569       // Therefore, the following sequence is made for the good of both.
1570       __ ld(tmp1, 0, R3_ARG1);
1571       __ ld(tmp2, 8, R3_ARG1);
1572       __ ld(tmp3, 16, R3_ARG1);
1573       __ ld(tmp4, 24, R3_ARG1);
1574       __ std(tmp1, 0, R4_ARG2);
1575       __ std(tmp2, 8, R4_ARG2);
1576       __ std(tmp3, 16, R4_ARG2);
1577       __ std(tmp4, 24, R4_ARG2);
1578       __ addi(R3_ARG1, R3_ARG1, 32);
1579       __ addi(R4_ARG2, R4_ARG2, 32);
1580       __ bdnz(l_6);
1581     }
1582 
1583     // copy 1 element at a time
1584     __ bind(l_2);
1585     __ cmpwi(CCR0, R5_ARG3, 0);
1586     __ beq(CCR0, l_1);
1587 
1588     { // FasterArrayCopy
1589       __ mtctr(R5_ARG3);
1590       __ addi(R3_ARG1, R3_ARG1, -4);
1591       __ addi(R4_ARG2, R4_ARG2, -4);
1592 
1593       __ bind(l_3);
1594       __ lwzu(tmp2, 4, R3_ARG1);
1595       __ stwu(tmp2, 4, R4_ARG2);
1596       __ bdnz(l_3);
1597     }
1598 
1599     __ bind(l_1);
1600     return;
1601   }
1602 
1603   // Generate stub for disjoint int copy.  If "aligned" is true, the
1604   // "from" and "to" addresses are assumed to be heapword aligned.
1605   //
1606   // Arguments for generated stub:
1607   //      from:  R3_ARG1
1608   //      to:    R4_ARG2
1609   //      count: R5_ARG3 treated as signed
1610   //
1611   address generate_disjoint_int_copy(bool aligned, const char * name) {
1612     StubCodeMark mark(this, "StubRoutines", name);
1613     address start = __ function_entry();
1614     generate_disjoint_int_copy_core(aligned);
1615     __ blr();
1616     return start;
1617   }
1618 
1619   // Generate core code for conjoint int copy (and oop copy on
1620   // 32-bit).  If "aligned" is true, the "from" and "to" addresses
1621   // are assumed to be heapword aligned.
1622   //
1623   // Arguments:
1624   //      from:  R3_ARG1
1625   //      to:    R4_ARG2
1626   //      count: R5_ARG3 treated as signed
1627   //
1628   void generate_conjoint_int_copy_core(bool aligned) {
1629     // Do reverse copy.  We assume the case of actual overlap is rare enough
1630     // that we don't have to optimize it.
1631 
1632     Label l_1, l_2, l_3, l_4, l_5, l_6;
1633 
1634     Register tmp1 = R6_ARG4;
1635     Register tmp2 = R7_ARG5;
1636     Register tmp3 = R8_ARG6;
1637     Register tmp4 = R0;
1638 
1639     { // FasterArrayCopy
1640       __ cmpwi(CCR0, R5_ARG3, 0);
1641       __ beq(CCR0, l_6);
1642 
1643       __ sldi(R5_ARG3, R5_ARG3, 2);
1644       __ add(R3_ARG1, R3_ARG1, R5_ARG3);
1645       __ add(R4_ARG2, R4_ARG2, R5_ARG3);
1646       __ srdi(R5_ARG3, R5_ARG3, 2);
1647 
1648       __ cmpwi(CCR0, R5_ARG3, 7);
1649       __ ble(CCR0, l_5); // copy 1 at a time if less than 8 elements remain
1650 
1651       __ srdi(tmp1, R5_ARG3, 3);
1652       __ andi(R5_ARG3, R5_ARG3, 7);
1653       __ mtctr(tmp1);
1654 
1655       __ bind(l_4);
1656       // Use unrolled version for mass copying (copy 4 elements a time).
1657       // Load feeding store gets zero latency on Power6, however not on Power5.
1658       // Therefore, the following sequence is made for the good of both.
1659       __ addi(R3_ARG1, R3_ARG1, -32);
1660       __ addi(R4_ARG2, R4_ARG2, -32);
1661       __ ld(tmp4, 24, R3_ARG1);
1662       __ ld(tmp3, 16, R3_ARG1);
1663       __ ld(tmp2, 8, R3_ARG1);
1664       __ ld(tmp1, 0, R3_ARG1);
1665       __ std(tmp4, 24, R4_ARG2);
1666       __ std(tmp3, 16, R4_ARG2);
1667       __ std(tmp2, 8, R4_ARG2);
1668       __ std(tmp1, 0, R4_ARG2);
1669       __ bdnz(l_4);
1670 
1671       __ cmpwi(CCR0, R5_ARG3, 0);
1672       __ beq(CCR0, l_6);
1673 
1674       __ bind(l_5);
1675       __ mtctr(R5_ARG3);
1676       __ bind(l_3);
1677       __ lwz(R0, -4, R3_ARG1);
1678       __ stw(R0, -4, R4_ARG2);
1679       __ addi(R3_ARG1, R3_ARG1, -4);
1680       __ addi(R4_ARG2, R4_ARG2, -4);
1681       __ bdnz(l_3);
1682 
1683       __ bind(l_6);
1684     }
1685   }
1686 
1687   // Generate stub for conjoint int copy.  If "aligned" is true, the
1688   // "from" and "to" addresses are assumed to be heapword aligned.
1689   //
1690   // Arguments for generated stub:
1691   //      from:  R3_ARG1
1692   //      to:    R4_ARG2
1693   //      count: R5_ARG3 treated as signed
1694   //
1695   address generate_conjoint_int_copy(bool aligned, const char * name) {
1696     StubCodeMark mark(this, "StubRoutines", name);
1697     address start = __ function_entry();
1698 
1699 #if defined(ABI_ELFv2)
1700     address nooverlap_target = aligned ?
1701       StubRoutines::arrayof_jint_disjoint_arraycopy() :
1702       StubRoutines::jint_disjoint_arraycopy();
1703 #else
1704     address nooverlap_target = aligned ?
1705       ((FunctionDescriptor*)StubRoutines::arrayof_jint_disjoint_arraycopy())->entry() :
1706       ((FunctionDescriptor*)StubRoutines::jint_disjoint_arraycopy())->entry();
1707 #endif
1708 
1709     array_overlap_test(nooverlap_target, 2);
1710 
1711     generate_conjoint_int_copy_core(aligned);
1712 
1713     __ blr();
1714 
1715     return start;
1716   }
1717 
1718   // Generate core code for disjoint long copy (and oop copy on
1719   // 64-bit).  If "aligned" is true, the "from" and "to" addresses
1720   // are assumed to be heapword aligned.
1721   //
1722   // Arguments:
1723   //      from:  R3_ARG1
1724   //      to:    R4_ARG2
1725   //      count: R5_ARG3 treated as signed
1726   //
1727   void generate_disjoint_long_copy_core(bool aligned) {
1728     Register tmp1 = R6_ARG4;
1729     Register tmp2 = R7_ARG5;
1730     Register tmp3 = R8_ARG6;
1731     Register tmp4 = R0;
1732 
1733     Label l_1, l_2, l_3, l_4;
1734 
1735     { // FasterArrayCopy
1736       __ cmpwi(CCR0, R5_ARG3, 3);
1737       __ ble(CCR0, l_3); // copy 1 at a time if less than 4 elements remain
1738 
1739       __ srdi(tmp1, R5_ARG3, 2);
1740       __ andi_(R5_ARG3, R5_ARG3, 3);
1741       __ mtctr(tmp1);
1742 
1743       __ bind(l_4);
1744       // Use unrolled version for mass copying (copy 4 elements a time).
1745       // Load feeding store gets zero latency on Power6, however not on Power5.
1746       // Therefore, the following sequence is made for the good of both.
1747       __ ld(tmp1, 0, R3_ARG1);
1748       __ ld(tmp2, 8, R3_ARG1);
1749       __ ld(tmp3, 16, R3_ARG1);
1750       __ ld(tmp4, 24, R3_ARG1);
1751       __ std(tmp1, 0, R4_ARG2);
1752       __ std(tmp2, 8, R4_ARG2);
1753       __ std(tmp3, 16, R4_ARG2);
1754       __ std(tmp4, 24, R4_ARG2);
1755       __ addi(R3_ARG1, R3_ARG1, 32);
1756       __ addi(R4_ARG2, R4_ARG2, 32);
1757       __ bdnz(l_4);
1758     }
1759 
1760     // copy 1 element at a time
1761     __ bind(l_3);
1762     __ cmpwi(CCR0, R5_ARG3, 0);
1763     __ beq(CCR0, l_1);
1764 
1765     { // FasterArrayCopy
1766       __ mtctr(R5_ARG3);
1767       __ addi(R3_ARG1, R3_ARG1, -8);
1768       __ addi(R4_ARG2, R4_ARG2, -8);
1769 
1770       __ bind(l_2);
1771       __ ldu(R0, 8, R3_ARG1);
1772       __ stdu(R0, 8, R4_ARG2);
1773       __ bdnz(l_2);
1774 
1775     }
1776     __ bind(l_1);
1777   }
1778 
1779   // Generate stub for disjoint long copy.  If "aligned" is true, the
1780   // "from" and "to" addresses are assumed to be heapword aligned.
1781   //
1782   // Arguments for generated stub:
1783   //      from:  R3_ARG1
1784   //      to:    R4_ARG2
1785   //      count: R5_ARG3 treated as signed
1786   //
1787   address generate_disjoint_long_copy(bool aligned, const char * name) {
1788     StubCodeMark mark(this, "StubRoutines", name);
1789     address start = __ function_entry();
1790     generate_disjoint_long_copy_core(aligned);
1791     __ blr();
1792 
1793     return start;
1794   }
1795 
1796   // Generate core code for conjoint long copy (and oop copy on
1797   // 64-bit).  If "aligned" is true, the "from" and "to" addresses
1798   // are assumed to be heapword aligned.
1799   //
1800   // Arguments:
1801   //      from:  R3_ARG1
1802   //      to:    R4_ARG2
1803   //      count: R5_ARG3 treated as signed
1804   //
1805   void generate_conjoint_long_copy_core(bool aligned) {
1806     Register tmp1 = R6_ARG4;
1807     Register tmp2 = R7_ARG5;
1808     Register tmp3 = R8_ARG6;
1809     Register tmp4 = R0;
1810 
1811     Label l_1, l_2, l_3, l_4, l_5;
1812 
1813     __ cmpwi(CCR0, R5_ARG3, 0);
1814     __ beq(CCR0, l_1);
1815 
1816     { // FasterArrayCopy
1817       __ sldi(R5_ARG3, R5_ARG3, 3);
1818       __ add(R3_ARG1, R3_ARG1, R5_ARG3);
1819       __ add(R4_ARG2, R4_ARG2, R5_ARG3);
1820       __ srdi(R5_ARG3, R5_ARG3, 3);
1821 
1822       __ cmpwi(CCR0, R5_ARG3, 3);
1823       __ ble(CCR0, l_5); // copy 1 at a time if less than 4 elements remain
1824 
1825       __ srdi(tmp1, R5_ARG3, 2);
1826       __ andi(R5_ARG3, R5_ARG3, 3);
1827       __ mtctr(tmp1);
1828 
1829       __ bind(l_4);
1830       // Use unrolled version for mass copying (copy 4 elements a time).
1831       // Load feeding store gets zero latency on Power6, however not on Power5.
1832       // Therefore, the following sequence is made for the good of both.
1833       __ addi(R3_ARG1, R3_ARG1, -32);
1834       __ addi(R4_ARG2, R4_ARG2, -32);
1835       __ ld(tmp4, 24, R3_ARG1);
1836       __ ld(tmp3, 16, R3_ARG1);
1837       __ ld(tmp2, 8, R3_ARG1);
1838       __ ld(tmp1, 0, R3_ARG1);
1839       __ std(tmp4, 24, R4_ARG2);
1840       __ std(tmp3, 16, R4_ARG2);
1841       __ std(tmp2, 8, R4_ARG2);
1842       __ std(tmp1, 0, R4_ARG2);
1843       __ bdnz(l_4);
1844 
1845       __ cmpwi(CCR0, R5_ARG3, 0);
1846       __ beq(CCR0, l_1);
1847 
1848       __ bind(l_5);
1849       __ mtctr(R5_ARG3);
1850       __ bind(l_3);
1851       __ ld(R0, -8, R3_ARG1);
1852       __ std(R0, -8, R4_ARG2);
1853       __ addi(R3_ARG1, R3_ARG1, -8);
1854       __ addi(R4_ARG2, R4_ARG2, -8);
1855       __ bdnz(l_3);
1856 
1857     }
1858     __ bind(l_1);
1859   }
1860 
1861   // Generate stub for conjoint long copy.  If "aligned" is true, the
1862   // "from" and "to" addresses are assumed to be heapword aligned.
1863   //
1864   // Arguments for generated stub:
1865   //      from:  R3_ARG1
1866   //      to:    R4_ARG2
1867   //      count: R5_ARG3 treated as signed
1868   //
1869   address generate_conjoint_long_copy(bool aligned, const char * name) {
1870     StubCodeMark mark(this, "StubRoutines", name);
1871     address start = __ function_entry();
1872 
1873 #if defined(ABI_ELFv2)
1874     address nooverlap_target = aligned ?
1875       StubRoutines::arrayof_jlong_disjoint_arraycopy() :
1876       StubRoutines::jlong_disjoint_arraycopy();
1877 #else
1878     address nooverlap_target = aligned ?
1879       ((FunctionDescriptor*)StubRoutines::arrayof_jlong_disjoint_arraycopy())->entry() :
1880       ((FunctionDescriptor*)StubRoutines::jlong_disjoint_arraycopy())->entry();
1881 #endif
1882 
1883     array_overlap_test(nooverlap_target, 3);
1884     generate_conjoint_long_copy_core(aligned);
1885 
1886     __ blr();
1887 
1888     return start;
1889   }
1890 
1891   // Generate stub for conjoint oop copy.  If "aligned" is true, the
1892   // "from" and "to" addresses are assumed to be heapword aligned.
1893   //
1894   // Arguments for generated stub:
1895   //      from:  R3_ARG1
1896   //      to:    R4_ARG2
1897   //      count: R5_ARG3 treated as signed
1898   //      dest_uninitialized: G1 support
1899   //
1900   address generate_conjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) {
1901     StubCodeMark mark(this, "StubRoutines", name);
1902 
1903     address start = __ function_entry();
1904 
1905 #if defined(ABI_ELFv2)
1906     address nooverlap_target = aligned ?
1907       StubRoutines::arrayof_oop_disjoint_arraycopy() :
1908       StubRoutines::oop_disjoint_arraycopy();
1909 #else
1910     address nooverlap_target = aligned ?
1911       ((FunctionDescriptor*)StubRoutines::arrayof_oop_disjoint_arraycopy())->entry() :
1912       ((FunctionDescriptor*)StubRoutines::oop_disjoint_arraycopy())->entry();
1913 #endif
1914 
1915     gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7);
1916 
1917     // Save arguments.
1918     __ mr(R9_ARG7, R4_ARG2);
1919     __ mr(R10_ARG8, R5_ARG3);
1920 
1921     if (UseCompressedOops) {
1922       array_overlap_test(nooverlap_target, 2);
1923       generate_conjoint_int_copy_core(aligned);
1924     } else {
1925       array_overlap_test(nooverlap_target, 3);
1926       generate_conjoint_long_copy_core(aligned);
1927     }
1928 
1929     gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1, /*branchToEnd*/ false);
1930     return start;
1931   }
1932 
1933   // Generate stub for disjoint oop copy.  If "aligned" is true, the
1934   // "from" and "to" addresses are assumed to be heapword aligned.
1935   //
1936   // Arguments for generated stub:
1937   //      from:  R3_ARG1
1938   //      to:    R4_ARG2
1939   //      count: R5_ARG3 treated as signed
1940   //      dest_uninitialized: G1 support
1941   //
1942   address generate_disjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) {
1943     StubCodeMark mark(this, "StubRoutines", name);
1944     address start = __ function_entry();
1945 
1946     gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7);
1947 
1948     // save some arguments, disjoint_long_copy_core destroys them.
1949     // needed for post barrier
1950     __ mr(R9_ARG7, R4_ARG2);
1951     __ mr(R10_ARG8, R5_ARG3);
1952 
1953     if (UseCompressedOops) {
1954       generate_disjoint_int_copy_core(aligned);
1955     } else {
1956       generate_disjoint_long_copy_core(aligned);
1957     }
1958 
1959     gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1, /*branchToEnd*/ false);
1960 
1961     return start;
1962   }
1963 
1964   // Arguments for generated stub (little endian only):
1965   //   R3_ARG1   - source byte array address
1966   //   R4_ARG2   - destination byte array address
1967   //   R5_ARG3   - round key array
1968   address generate_aescrypt_encryptBlock() {
1969     assert(UseAES, "need AES instructions and misaligned SSE support");
1970     StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
1971 
1972     address start = __ function_entry();
1973 
1974     Label L_doLast;
1975 
1976     Register from           = R3_ARG1;  // source array address
1977     Register to             = R4_ARG2;  // destination array address
1978     Register key            = R5_ARG3;  // round key array
1979 
1980     Register keylen         = R8;
1981     Register temp           = R9;
1982     Register keypos         = R10;
1983     Register hex            = R11;
1984     Register fifteen        = R12;
1985 
1986     VectorRegister vRet     = VR0;
1987 
1988     VectorRegister vKey1    = VR1;
1989     VectorRegister vKey2    = VR2;
1990     VectorRegister vKey3    = VR3;
1991     VectorRegister vKey4    = VR4;
1992 
1993     VectorRegister fromPerm = VR5;
1994     VectorRegister keyPerm  = VR6;
1995     VectorRegister toPerm   = VR7;
1996     VectorRegister fSplt    = VR8;
1997 
1998     VectorRegister vTmp1    = VR9;
1999     VectorRegister vTmp2    = VR10;
2000     VectorRegister vTmp3    = VR11;
2001     VectorRegister vTmp4    = VR12;
2002 
2003     VectorRegister vLow     = VR13;
2004     VectorRegister vHigh    = VR14;
2005 
2006     __ li              (hex, 16);
2007     __ li              (fifteen, 15);
2008     __ vspltisb        (fSplt, 0x0f);
2009 
2010     // load unaligned from[0-15] to vsRet
2011     __ lvx             (vRet, from);
2012     __ lvx             (vTmp1, fifteen, from);
2013     __ lvsl            (fromPerm, from);
2014     __ vxor            (fromPerm, fromPerm, fSplt);
2015     __ vperm           (vRet, vRet, vTmp1, fromPerm);
2016 
2017     // load keylen (44 or 52 or 60)
2018     __ lwz             (keylen, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT), key);
2019 
2020     // to load keys
2021     __ lvsr            (keyPerm, key);
2022     __ vxor            (vTmp2, vTmp2, vTmp2);
2023     __ vspltisb        (vTmp2, -16);
2024     __ vrld            (keyPerm, keyPerm, vTmp2);
2025     __ vrld            (keyPerm, keyPerm, vTmp2);
2026     __ vsldoi          (keyPerm, keyPerm, keyPerm, -8);
2027 
2028     // load the 1st round key to vKey1
2029     __ li              (keypos, 0);
2030     __ lvx             (vKey1, keypos, key);
2031     __ addi            (keypos, keypos, 16);
2032     __ lvx             (vTmp1, keypos, key);
2033     __ vperm           (vKey1, vTmp1, vKey1, keyPerm);
2034 
2035     // 1st round
2036     __ vxor (vRet, vRet, vKey1);
2037 
2038     // load the 2nd round key to vKey1
2039     __ addi            (keypos, keypos, 16);
2040     __ lvx             (vTmp2, keypos, key);
2041     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2042 
2043     // load the 3rd round key to vKey2
2044     __ addi            (keypos, keypos, 16);
2045     __ lvx             (vTmp1, keypos, key);
2046     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2047 
2048     // load the 4th round key to vKey3
2049     __ addi            (keypos, keypos, 16);
2050     __ lvx             (vTmp2, keypos, key);
2051     __ vperm           (vKey3, vTmp2, vTmp1, keyPerm);
2052 
2053     // load the 5th round key to vKey4
2054     __ addi            (keypos, keypos, 16);
2055     __ lvx             (vTmp1, keypos, key);
2056     __ vperm           (vKey4, vTmp1, vTmp2, keyPerm);
2057 
2058     // 2nd - 5th rounds
2059     __ vcipher (vRet, vRet, vKey1);
2060     __ vcipher (vRet, vRet, vKey2);
2061     __ vcipher (vRet, vRet, vKey3);
2062     __ vcipher (vRet, vRet, vKey4);
2063 
2064     // load the 6th round key to vKey1
2065     __ addi            (keypos, keypos, 16);
2066     __ lvx             (vTmp2, keypos, key);
2067     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2068 
2069     // load the 7th round key to vKey2
2070     __ addi            (keypos, keypos, 16);
2071     __ lvx             (vTmp1, keypos, key);
2072     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2073 
2074     // load the 8th round key to vKey3
2075     __ addi            (keypos, keypos, 16);
2076     __ lvx             (vTmp2, keypos, key);
2077     __ vperm           (vKey3, vTmp2, vTmp1, keyPerm);
2078 
2079     // load the 9th round key to vKey4
2080     __ addi            (keypos, keypos, 16);
2081     __ lvx             (vTmp1, keypos, key);
2082     __ vperm           (vKey4, vTmp1, vTmp2, keyPerm);
2083 
2084     // 6th - 9th rounds
2085     __ vcipher (vRet, vRet, vKey1);
2086     __ vcipher (vRet, vRet, vKey2);
2087     __ vcipher (vRet, vRet, vKey3);
2088     __ vcipher (vRet, vRet, vKey4);
2089 
2090     // load the 10th round key to vKey1
2091     __ addi            (keypos, keypos, 16);
2092     __ lvx             (vTmp2, keypos, key);
2093     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2094 
2095     // load the 11th round key to vKey2
2096     __ addi            (keypos, keypos, 16);
2097     __ lvx             (vTmp1, keypos, key);
2098     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2099 
2100     // if all round keys are loaded, skip next 4 rounds
2101     __ cmpwi           (CCR0, keylen, 44);
2102     __ beq             (CCR0, L_doLast);
2103 
2104     // 10th - 11th rounds
2105     __ vcipher (vRet, vRet, vKey1);
2106     __ vcipher (vRet, vRet, vKey2);
2107 
2108     // load the 12th round key to vKey1
2109     __ addi            (keypos, keypos, 16);
2110     __ lvx             (vTmp2, keypos, key);
2111     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2112 
2113     // load the 13th round key to vKey2
2114     __ addi            (keypos, keypos, 16);
2115     __ lvx             (vTmp1, keypos, key);
2116     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2117 
2118     // if all round keys are loaded, skip next 2 rounds
2119     __ cmpwi           (CCR0, keylen, 52);
2120     __ beq             (CCR0, L_doLast);
2121 
2122     // 12th - 13th rounds
2123     __ vcipher (vRet, vRet, vKey1);
2124     __ vcipher (vRet, vRet, vKey2);
2125 
2126     // load the 14th round key to vKey1
2127     __ addi            (keypos, keypos, 16);
2128     __ lvx             (vTmp2, keypos, key);
2129     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2130 
2131     // load the 15th round key to vKey2
2132     __ addi            (keypos, keypos, 16);
2133     __ lvx             (vTmp1, keypos, key);
2134     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2135 
2136     __ bind(L_doLast);
2137 
2138     // last two rounds
2139     __ vcipher (vRet, vRet, vKey1);
2140     __ vcipherlast (vRet, vRet, vKey2);
2141 
2142     __ neg             (temp, to);
2143     __ lvsr            (toPerm, temp);
2144     __ vspltisb        (vTmp2, -1);
2145     __ vxor            (vTmp1, vTmp1, vTmp1);
2146     __ vperm           (vTmp2, vTmp2, vTmp1, toPerm);
2147     __ vxor            (toPerm, toPerm, fSplt);
2148     __ lvx             (vTmp1, to);
2149     __ vperm           (vRet, vRet, vRet, toPerm);
2150     __ vsel            (vTmp1, vTmp1, vRet, vTmp2);
2151     __ lvx             (vTmp4, fifteen, to);
2152     __ stvx            (vTmp1, to);
2153     __ vsel            (vRet, vRet, vTmp4, vTmp2);
2154     __ stvx            (vRet, fifteen, to);
2155 
2156     __ blr();
2157      return start;
2158   }
2159 
2160   // Arguments for generated stub (little endian only):
2161   //   R3_ARG1   - source byte array address
2162   //   R4_ARG2   - destination byte array address
2163   //   R5_ARG3   - K (key) in little endian int array
2164   address generate_aescrypt_decryptBlock() {
2165     assert(UseAES, "need AES instructions and misaligned SSE support");
2166     StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
2167 
2168     address start = __ function_entry();
2169 
2170     Label L_doLast;
2171     Label L_do44;
2172     Label L_do52;
2173     Label L_do60;
2174 
2175     Register from           = R3_ARG1;  // source array address
2176     Register to             = R4_ARG2;  // destination array address
2177     Register key            = R5_ARG3;  // round key array
2178 
2179     Register keylen         = R8;
2180     Register temp           = R9;
2181     Register keypos         = R10;
2182     Register hex            = R11;
2183     Register fifteen        = R12;
2184 
2185     VectorRegister vRet     = VR0;
2186 
2187     VectorRegister vKey1    = VR1;
2188     VectorRegister vKey2    = VR2;
2189     VectorRegister vKey3    = VR3;
2190     VectorRegister vKey4    = VR4;
2191     VectorRegister vKey5    = VR5;
2192 
2193     VectorRegister fromPerm = VR6;
2194     VectorRegister keyPerm  = VR7;
2195     VectorRegister toPerm   = VR8;
2196     VectorRegister fSplt    = VR9;
2197 
2198     VectorRegister vTmp1    = VR10;
2199     VectorRegister vTmp2    = VR11;
2200     VectorRegister vTmp3    = VR12;
2201     VectorRegister vTmp4    = VR13;
2202 
2203     VectorRegister vLow     = VR14;
2204     VectorRegister vHigh    = VR15;
2205 
2206     __ li              (hex, 16);
2207     __ li              (fifteen, 15);
2208     __ vspltisb        (fSplt, 0x0f);
2209 
2210     // load unaligned from[0-15] to vsRet
2211     __ lvx             (vRet, from);
2212     __ lvx             (vTmp1, fifteen, from);
2213     __ lvsl            (fromPerm, from);
2214     __ vxor            (fromPerm, fromPerm, fSplt);
2215     __ vperm           (vRet, vRet, vTmp1, fromPerm); // align [and byte swap in LE]
2216 
2217     // load keylen (44 or 52 or 60)
2218     __ lwz             (keylen, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT), key);
2219 
2220     // to load keys
2221     __ lvsr            (keyPerm, key);
2222     __ vxor            (vTmp2, vTmp2, vTmp2);
2223     __ vspltisb        (vTmp2, -16);
2224     __ vrld            (keyPerm, keyPerm, vTmp2);
2225     __ vrld            (keyPerm, keyPerm, vTmp2);
2226     __ vsldoi          (keyPerm, keyPerm, keyPerm, -8);
2227 
2228     __ cmpwi           (CCR0, keylen, 44);
2229     __ beq             (CCR0, L_do44);
2230 
2231     __ cmpwi           (CCR0, keylen, 52);
2232     __ beq             (CCR0, L_do52);
2233 
2234     // load the 15th round key to vKey11
2235     __ li              (keypos, 240);
2236     __ lvx             (vTmp1, keypos, key);
2237     __ addi            (keypos, keypos, -16);
2238     __ lvx             (vTmp2, keypos, key);
2239     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2240 
2241     // load the 14th round key to vKey10
2242     __ addi            (keypos, keypos, -16);
2243     __ lvx             (vTmp1, keypos, key);
2244     __ vperm           (vKey2, vTmp2, vTmp1, keyPerm);
2245 
2246     // load the 13th round key to vKey10
2247     __ addi            (keypos, keypos, -16);
2248     __ lvx             (vTmp2, keypos, key);
2249     __ vperm           (vKey3, vTmp1, vTmp2, keyPerm);
2250 
2251     // load the 12th round key to vKey10
2252     __ addi            (keypos, keypos, -16);
2253     __ lvx             (vTmp1, keypos, key);
2254     __ vperm           (vKey4, vTmp2, vTmp1, keyPerm);
2255 
2256     // load the 11th round key to vKey10
2257     __ addi            (keypos, keypos, -16);
2258     __ lvx             (vTmp2, keypos, key);
2259     __ vperm           (vKey5, vTmp1, vTmp2, keyPerm);
2260 
2261     // 1st - 5th rounds
2262     __ vxor            (vRet, vRet, vKey1);
2263     __ vncipher        (vRet, vRet, vKey2);
2264     __ vncipher        (vRet, vRet, vKey3);
2265     __ vncipher        (vRet, vRet, vKey4);
2266     __ vncipher        (vRet, vRet, vKey5);
2267 
2268     __ b               (L_doLast);
2269 
2270     __ bind            (L_do52);
2271 
2272     // load the 13th round key to vKey11
2273     __ li              (keypos, 208);
2274     __ lvx             (vTmp1, keypos, key);
2275     __ addi            (keypos, keypos, -16);
2276     __ lvx             (vTmp2, keypos, key);
2277     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2278 
2279     // load the 12th round key to vKey10
2280     __ addi            (keypos, keypos, -16);
2281     __ lvx             (vTmp1, keypos, key);
2282     __ vperm           (vKey2, vTmp2, vTmp1, keyPerm);
2283 
2284     // load the 11th round key to vKey10
2285     __ addi            (keypos, keypos, -16);
2286     __ lvx             (vTmp2, keypos, key);
2287     __ vperm           (vKey3, vTmp1, vTmp2, keyPerm);
2288 
2289     // 1st - 3rd rounds
2290     __ vxor            (vRet, vRet, vKey1);
2291     __ vncipher        (vRet, vRet, vKey2);
2292     __ vncipher        (vRet, vRet, vKey3);
2293 
2294     __ b               (L_doLast);
2295 
2296     __ bind            (L_do44);
2297 
2298     // load the 11th round key to vKey11
2299     __ li              (keypos, 176);
2300     __ lvx             (vTmp1, keypos, key);
2301     __ addi            (keypos, keypos, -16);
2302     __ lvx             (vTmp2, keypos, key);
2303     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2304 
2305     // 1st round
2306     __ vxor            (vRet, vRet, vKey1);
2307 
2308     __ bind            (L_doLast);
2309 
2310     // load the 10th round key to vKey10
2311     __ addi            (keypos, keypos, -16);
2312     __ lvx             (vTmp1, keypos, key);
2313     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2314 
2315     // load the 9th round key to vKey10
2316     __ addi            (keypos, keypos, -16);
2317     __ lvx             (vTmp2, keypos, key);
2318     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2319 
2320     // load the 8th round key to vKey10
2321     __ addi            (keypos, keypos, -16);
2322     __ lvx             (vTmp1, keypos, key);
2323     __ vperm           (vKey3, vTmp2, vTmp1, keyPerm);
2324 
2325     // load the 7th round key to vKey10
2326     __ addi            (keypos, keypos, -16);
2327     __ lvx             (vTmp2, keypos, key);
2328     __ vperm           (vKey4, vTmp1, vTmp2, keyPerm);
2329 
2330     // load the 6th round key to vKey10
2331     __ addi            (keypos, keypos, -16);
2332     __ lvx             (vTmp1, keypos, key);
2333     __ vperm           (vKey5, vTmp2, vTmp1, keyPerm);
2334 
2335     // last 10th - 6th rounds
2336     __ vncipher        (vRet, vRet, vKey1);
2337     __ vncipher        (vRet, vRet, vKey2);
2338     __ vncipher        (vRet, vRet, vKey3);
2339     __ vncipher        (vRet, vRet, vKey4);
2340     __ vncipher        (vRet, vRet, vKey5);
2341 
2342     // load the 5th round key to vKey10
2343     __ addi            (keypos, keypos, -16);
2344     __ lvx             (vTmp2, keypos, key);
2345     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2346 
2347     // load the 4th round key to vKey10
2348     __ addi            (keypos, keypos, -16);
2349     __ lvx             (vTmp1, keypos, key);
2350     __ vperm           (vKey2, vTmp2, vTmp1, keyPerm);
2351 
2352     // load the 3rd round key to vKey10
2353     __ addi            (keypos, keypos, -16);
2354     __ lvx             (vTmp2, keypos, key);
2355     __ vperm           (vKey3, vTmp1, vTmp2, keyPerm);
2356 
2357     // load the 2nd round key to vKey10
2358     __ addi            (keypos, keypos, -16);
2359     __ lvx             (vTmp1, keypos, key);
2360     __ vperm           (vKey4, vTmp2, vTmp1, keyPerm);
2361 
2362     // load the 1st round key to vKey10
2363     __ addi            (keypos, keypos, -16);
2364     __ lvx             (vTmp2, keypos, key);
2365     __ vperm           (vKey5, vTmp1, vTmp2, keyPerm);
2366 
2367     // last 5th - 1th rounds
2368     __ vncipher        (vRet, vRet, vKey1);
2369     __ vncipher        (vRet, vRet, vKey2);
2370     __ vncipher        (vRet, vRet, vKey3);
2371     __ vncipher        (vRet, vRet, vKey4);
2372     __ vncipherlast    (vRet, vRet, vKey5);
2373 
2374     __ neg             (temp, to);
2375     __ lvsr            (toPerm, temp);
2376     __ vspltisb        (vTmp2, -1);
2377     __ vxor            (vTmp1, vTmp1, vTmp1);
2378     __ vperm           (vTmp2, vTmp2, vTmp1, toPerm);
2379     __ vxor            (toPerm, toPerm, fSplt);
2380     __ lvx             (vTmp1, to);
2381     __ vperm           (vRet, vRet, vRet, toPerm);
2382     __ vsel            (vTmp1, vTmp1, vRet, vTmp2);
2383     __ lvx             (vTmp4, fifteen, to);
2384     __ stvx            (vTmp1, to);
2385     __ vsel            (vRet, vRet, vTmp4, vTmp2);
2386     __ stvx            (vRet, fifteen, to);
2387 
2388     __ blr();
2389      return start;
2390   }
2391 
2392   void generate_arraycopy_stubs() {
2393     // Note: the disjoint stubs must be generated first, some of
2394     // the conjoint stubs use them.
2395 
2396     // non-aligned disjoint versions
2397     StubRoutines::_jbyte_disjoint_arraycopy       = generate_disjoint_byte_copy(false, "jbyte_disjoint_arraycopy");
2398     StubRoutines::_jshort_disjoint_arraycopy      = generate_disjoint_short_copy(false, "jshort_disjoint_arraycopy");
2399     StubRoutines::_jint_disjoint_arraycopy        = generate_disjoint_int_copy(false, "jint_disjoint_arraycopy");
2400     StubRoutines::_jlong_disjoint_arraycopy       = generate_disjoint_long_copy(false, "jlong_disjoint_arraycopy");
2401     StubRoutines::_oop_disjoint_arraycopy         = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy", false);
2402     StubRoutines::_oop_disjoint_arraycopy_uninit  = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy_uninit", true);
2403 
2404     // aligned disjoint versions
2405     StubRoutines::_arrayof_jbyte_disjoint_arraycopy      = generate_disjoint_byte_copy(true, "arrayof_jbyte_disjoint_arraycopy");
2406     StubRoutines::_arrayof_jshort_disjoint_arraycopy     = generate_disjoint_short_copy(true, "arrayof_jshort_disjoint_arraycopy");
2407     StubRoutines::_arrayof_jint_disjoint_arraycopy       = generate_disjoint_int_copy(true, "arrayof_jint_disjoint_arraycopy");
2408     StubRoutines::_arrayof_jlong_disjoint_arraycopy      = generate_disjoint_long_copy(true, "arrayof_jlong_disjoint_arraycopy");
2409     StubRoutines::_arrayof_oop_disjoint_arraycopy        = generate_disjoint_oop_copy(true, "arrayof_oop_disjoint_arraycopy", false);
2410     StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(true, "oop_disjoint_arraycopy_uninit", true);
2411 
2412     // non-aligned conjoint versions
2413     StubRoutines::_jbyte_arraycopy      = generate_conjoint_byte_copy(false, "jbyte_arraycopy");
2414     StubRoutines::_jshort_arraycopy     = generate_conjoint_short_copy(false, "jshort_arraycopy");
2415     StubRoutines::_jint_arraycopy       = generate_conjoint_int_copy(false, "jint_arraycopy");
2416     StubRoutines::_jlong_arraycopy      = generate_conjoint_long_copy(false, "jlong_arraycopy");
2417     StubRoutines::_oop_arraycopy        = generate_conjoint_oop_copy(false, "oop_arraycopy", false);
2418     StubRoutines::_oop_arraycopy_uninit = generate_conjoint_oop_copy(false, "oop_arraycopy_uninit", true);
2419 
2420     // aligned conjoint versions
2421     StubRoutines::_arrayof_jbyte_arraycopy      = generate_conjoint_byte_copy(true, "arrayof_jbyte_arraycopy");
2422     StubRoutines::_arrayof_jshort_arraycopy     = generate_conjoint_short_copy(true, "arrayof_jshort_arraycopy");
2423     StubRoutines::_arrayof_jint_arraycopy       = generate_conjoint_int_copy(true, "arrayof_jint_arraycopy");
2424     StubRoutines::_arrayof_jlong_arraycopy      = generate_conjoint_long_copy(true, "arrayof_jlong_arraycopy");
2425     StubRoutines::_arrayof_oop_arraycopy        = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", false);
2426     StubRoutines::_arrayof_oop_arraycopy_uninit = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", true);
2427 
2428     // fill routines
2429     StubRoutines::_jbyte_fill          = generate_fill(T_BYTE,  false, "jbyte_fill");
2430     StubRoutines::_jshort_fill         = generate_fill(T_SHORT, false, "jshort_fill");
2431     StubRoutines::_jint_fill           = generate_fill(T_INT,   false, "jint_fill");
2432     StubRoutines::_arrayof_jbyte_fill  = generate_fill(T_BYTE,  true, "arrayof_jbyte_fill");
2433     StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
2434     StubRoutines::_arrayof_jint_fill   = generate_fill(T_INT,   true, "arrayof_jint_fill");
2435   }
2436 
2437   // Safefetch stubs.
2438   void generate_safefetch(const char* name, int size, address* entry, address* fault_pc, address* continuation_pc) {
2439     // safefetch signatures:
2440     //   int      SafeFetch32(int*      adr, int      errValue);
2441     //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
2442     //
2443     // arguments:
2444     //   R3_ARG1 = adr
2445     //   R4_ARG2 = errValue
2446     //
2447     // result:
2448     //   R3_RET  = *adr or errValue
2449 
2450     StubCodeMark mark(this, "StubRoutines", name);
2451 
2452     // Entry point, pc or function descriptor.
2453     *entry = __ function_entry();
2454 
2455     // Load *adr into R4_ARG2, may fault.
2456     *fault_pc = __ pc();
2457     switch (size) {
2458       case 4:
2459         // int32_t, signed extended
2460         __ lwa(R4_ARG2, 0, R3_ARG1);
2461         break;
2462       case 8:
2463         // int64_t
2464         __ ld(R4_ARG2, 0, R3_ARG1);
2465         break;
2466       default:
2467         ShouldNotReachHere();
2468     }
2469 
2470     // return errValue or *adr
2471     *continuation_pc = __ pc();
2472     __ mr(R3_RET, R4_ARG2);
2473     __ blr();
2474   }
2475 
2476   // Initialization
2477   void generate_initial() {
2478     // Generates all stubs and initializes the entry points
2479 
2480     // Entry points that exist in all platforms.
2481     // Note: This is code that could be shared among different platforms - however the
2482     // benefit seems to be smaller than the disadvantage of having a
2483     // much more complicated generator structure. See also comment in
2484     // stubRoutines.hpp.
2485 
2486     StubRoutines::_forward_exception_entry          = generate_forward_exception();
2487     StubRoutines::_call_stub_entry                  = generate_call_stub(StubRoutines::_call_stub_return_address);
2488     StubRoutines::_catch_exception_entry            = generate_catch_exception();
2489 
2490     // Build this early so it's available for the interpreter.
2491     StubRoutines::_throw_StackOverflowError_entry   =
2492       generate_throw_exception("StackOverflowError throw_exception",
2493                                CAST_FROM_FN_PTR(address, SharedRuntime::throw_StackOverflowError), false);
2494   }
2495 
2496   void generate_all() {
2497     // Generates all stubs and initializes the entry points
2498 
2499     // These entry points require SharedInfo::stack0 to be set up in
2500     // non-core builds
2501     StubRoutines::_throw_AbstractMethodError_entry         = generate_throw_exception("AbstractMethodError throw_exception",          CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError),  false);
2502     // Handle IncompatibleClassChangeError in itable stubs.
2503     StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError),  false);
2504     StubRoutines::_throw_NullPointerException_at_call_entry= generate_throw_exception("NullPointerException at call throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_NullPointerException_at_call), false);
2505 
2506     StubRoutines::_handler_for_unsafe_access_entry         = generate_handler_for_unsafe_access();
2507 
2508     // support for verify_oop (must happen after universe_init)
2509     StubRoutines::_verify_oop_subroutine_entry             = generate_verify_oop();
2510 
2511     // arraycopy stubs used by compilers
2512     generate_arraycopy_stubs();
2513 
2514     // Safefetch stubs.
2515     generate_safefetch("SafeFetch32", sizeof(int),     &StubRoutines::_safefetch32_entry,
2516                                                        &StubRoutines::_safefetch32_fault_pc,
2517                                                        &StubRoutines::_safefetch32_continuation_pc);
2518     generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
2519                                                        &StubRoutines::_safefetchN_fault_pc,
2520                                                        &StubRoutines::_safefetchN_continuation_pc);
2521 
2522     if (UseAESIntrinsics) {
2523       StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
2524       StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
2525     }
2526 
2527     if (UseMontgomeryMultiplyIntrinsic) {
2528       StubRoutines::_montgomeryMultiply
2529         = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
2530     }
2531     if (UseMontgomerySquareIntrinsic) {
2532       StubRoutines::_montgomerySquare
2533         = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
2534     }
2535   }
2536 
2537  public:
2538   StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
2539     // replace the standard masm with a special one:
2540     _masm = new MacroAssembler(code);
2541     if (all) {
2542       generate_all();
2543     } else {
2544       generate_initial();
2545     }
2546   }
2547 };
2548 
2549 void StubGenerator_generate(CodeBuffer* code, bool all) {
2550   StubGenerator g(code, all);
2551 }