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       case BarrierSet::Epsilon:
 652         break;
 653       default:
 654         ShouldNotReachHere();
 655     }
 656   }
 657 
 658   //  Generate CMS/G1 post-write barrier for array.
 659   //
 660   //  Input:
 661   //     addr     - register containing starting address
 662   //     count    - register containing element count
 663   //     tmp      - scratch register
 664   //
 665   //  The input registers and R0 are overwritten.
 666   //
 667   void gen_write_ref_array_post_barrier(Register addr, Register count, Register tmp, bool branchToEnd) {
 668     BarrierSet* const bs = Universe::heap()->barrier_set();
 669 
 670     switch (bs->kind()) {
 671       case BarrierSet::G1SATBCT:
 672       case BarrierSet::G1SATBCTLogging:
 673         {
 674           if (branchToEnd) {
 675             __ save_LR_CR(R0);
 676             // We need this frame only to spill LR.
 677             __ push_frame_reg_args(0, R0);
 678             __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), addr, count);
 679             __ pop_frame();
 680             __ restore_LR_CR(R0);
 681           } else {
 682             // Tail call: fake call from stub caller by branching without linking.
 683             address entry_point = (address)CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post);
 684             __ mr_if_needed(R3_ARG1, addr);
 685             __ mr_if_needed(R4_ARG2, count);
 686             __ load_const(R11, entry_point, R0);
 687             __ call_c_and_return_to_caller(R11);
 688           }
 689         }
 690         break;
 691       case BarrierSet::CardTableModRef:
 692       case BarrierSet::CardTableExtension:
 693         {
 694           Label Lskip_loop, Lstore_loop;
 695           if (UseConcMarkSweepGC) {
 696             // TODO PPC port: contribute optimization / requires shared changes
 697             __ release();
 698           }
 699 
 700           CardTableModRefBS* const ct = (CardTableModRefBS*)bs;
 701           assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
 702           assert_different_registers(addr, count, tmp);
 703 
 704           __ sldi(count, count, LogBytesPerHeapOop);
 705           __ addi(count, count, -BytesPerHeapOop);
 706           __ add(count, addr, count);
 707           // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.)
 708           __ srdi(addr, addr, CardTableModRefBS::card_shift);
 709           __ srdi(count, count, CardTableModRefBS::card_shift);
 710           __ subf(count, addr, count);
 711           assert_different_registers(R0, addr, count, tmp);
 712           __ load_const(tmp, (address)ct->byte_map_base);
 713           __ addic_(count, count, 1);
 714           __ beq(CCR0, Lskip_loop);
 715           __ li(R0, 0);
 716           __ mtctr(count);
 717           // Byte store loop
 718           __ bind(Lstore_loop);
 719           __ stbx(R0, tmp, addr);
 720           __ addi(addr, addr, 1);
 721           __ bdnz(Lstore_loop);
 722           __ bind(Lskip_loop);
 723 
 724           if (!branchToEnd) __ blr();
 725         }
 726       break;
 727       case BarrierSet::ModRef:
 728       case BarrierSet::Epsilon:
 729         if (!branchToEnd) __ blr();
 730         break;
 731       default:
 732         ShouldNotReachHere();
 733     }
 734   }
 735 
 736   // Support for void zero_words_aligned8(HeapWord* to, size_t count)
 737   //
 738   // Arguments:
 739   //   to:
 740   //   count:
 741   //
 742   // Destroys:
 743   //
 744   address generate_zero_words_aligned8() {
 745     StubCodeMark mark(this, "StubRoutines", "zero_words_aligned8");
 746 
 747     // Implemented as in ClearArray.
 748     address start = __ function_entry();
 749 
 750     Register base_ptr_reg   = R3_ARG1; // tohw (needs to be 8b aligned)
 751     Register cnt_dwords_reg = R4_ARG2; // count (in dwords)
 752     Register tmp1_reg       = R5_ARG3;
 753     Register tmp2_reg       = R6_ARG4;
 754     Register zero_reg       = R7_ARG5;
 755 
 756     // Procedure for large arrays (uses data cache block zero instruction).
 757     Label dwloop, fast, fastloop, restloop, lastdword, done;
 758     int cl_size=VM_Version::get_cache_line_size(), cl_dwords=cl_size>>3, cl_dwordaddr_bits=exact_log2(cl_dwords);
 759     int min_dcbz=2; // Needs to be positive, apply dcbz only to at least min_dcbz cache lines.
 760 
 761     // Clear up to 128byte boundary if long enough, dword_cnt=(16-(base>>3))%16.
 762     __ dcbtst(base_ptr_reg);                    // Indicate write access to first cache line ...
 763     __ andi(tmp2_reg, cnt_dwords_reg, 1);       // to check if number of dwords is even.
 764     __ srdi_(tmp1_reg, cnt_dwords_reg, 1);      // number of double dwords
 765     __ load_const_optimized(zero_reg, 0L);      // Use as zero register.
 766 
 767     __ cmpdi(CCR1, tmp2_reg, 0);                // cnt_dwords even?
 768     __ beq(CCR0, lastdword);                    // size <= 1
 769     __ mtctr(tmp1_reg);                         // Speculatively preload counter for rest loop (>0).
 770     __ cmpdi(CCR0, cnt_dwords_reg, (min_dcbz+1)*cl_dwords-1); // Big enough to ensure >=min_dcbz cache lines are included?
 771     __ neg(tmp1_reg, base_ptr_reg);             // bit 0..58: bogus, bit 57..60: (16-(base>>3))%16, bit 61..63: 000
 772 
 773     __ blt(CCR0, restloop);                     // Too small. (<31=(2*cl_dwords)-1 is sufficient, but bigger performs better.)
 774     __ rldicl_(tmp1_reg, tmp1_reg, 64-3, 64-cl_dwordaddr_bits); // Extract number of dwords to 128byte boundary=(16-(base>>3))%16.
 775 
 776     __ beq(CCR0, fast);                         // already 128byte aligned
 777     __ mtctr(tmp1_reg);                         // Set ctr to hit 128byte boundary (0<ctr<cnt).
 778     __ subf(cnt_dwords_reg, tmp1_reg, cnt_dwords_reg); // rest (>0 since size>=256-8)
 779 
 780     // Clear in first cache line dword-by-dword if not already 128byte aligned.
 781     __ bind(dwloop);
 782       __ std(zero_reg, 0, base_ptr_reg);        // Clear 8byte aligned block.
 783       __ addi(base_ptr_reg, base_ptr_reg, 8);
 784     __ bdnz(dwloop);
 785 
 786     // clear 128byte blocks
 787     __ bind(fast);
 788     __ srdi(tmp1_reg, cnt_dwords_reg, cl_dwordaddr_bits); // loop count for 128byte loop (>0 since size>=256-8)
 789     __ andi(tmp2_reg, cnt_dwords_reg, 1);       // to check if rest even
 790 
 791     __ mtctr(tmp1_reg);                         // load counter
 792     __ cmpdi(CCR1, tmp2_reg, 0);                // rest even?
 793     __ rldicl_(tmp1_reg, cnt_dwords_reg, 63, 65-cl_dwordaddr_bits); // rest in double dwords
 794 
 795     __ bind(fastloop);
 796       __ dcbz(base_ptr_reg);                    // Clear 128byte aligned block.
 797       __ addi(base_ptr_reg, base_ptr_reg, cl_size);
 798     __ bdnz(fastloop);
 799 
 800     //__ dcbtst(base_ptr_reg);                  // Indicate write access to last cache line.
 801     __ beq(CCR0, lastdword);                    // rest<=1
 802     __ mtctr(tmp1_reg);                         // load counter
 803 
 804     // Clear rest.
 805     __ bind(restloop);
 806       __ std(zero_reg, 0, base_ptr_reg);        // Clear 8byte aligned block.
 807       __ std(zero_reg, 8, base_ptr_reg);        // Clear 8byte aligned block.
 808       __ addi(base_ptr_reg, base_ptr_reg, 16);
 809     __ bdnz(restloop);
 810 
 811     __ bind(lastdword);
 812     __ beq(CCR1, done);
 813     __ std(zero_reg, 0, base_ptr_reg);
 814     __ bind(done);
 815     __ blr();                                   // return
 816 
 817     return start;
 818   }
 819 
 820   // The following routine generates a subroutine to throw an asynchronous
 821   // UnknownError when an unsafe access gets a fault that could not be
 822   // reasonably prevented by the programmer.  (Example: SIGBUS/OBJERR.)
 823   //
 824   address generate_handler_for_unsafe_access() {
 825     StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
 826     address start = __ function_entry();
 827     __ unimplemented("StubRoutines::handler_for_unsafe_access", 93);
 828     return start;
 829   }
 830 
 831 #if !defined(PRODUCT)
 832   // Wrapper which calls oopDesc::is_oop_or_null()
 833   // Only called by MacroAssembler::verify_oop
 834   static void verify_oop_helper(const char* message, oop o) {
 835     if (!o->is_oop_or_null()) {
 836       fatal(message);
 837     }
 838     ++ StubRoutines::_verify_oop_count;
 839   }
 840 #endif
 841 
 842   // Return address of code to be called from code generated by
 843   // MacroAssembler::verify_oop.
 844   //
 845   // Don't generate, rather use C++ code.
 846   address generate_verify_oop() {
 847     StubCodeMark mark(this, "StubRoutines", "verify_oop");
 848 
 849     // this is actually a `FunctionDescriptor*'.
 850     address start = 0;
 851 
 852 #if !defined(PRODUCT)
 853     start = CAST_FROM_FN_PTR(address, verify_oop_helper);
 854 #endif
 855 
 856     return start;
 857   }
 858 
 859   // Fairer handling of safepoints for native methods.
 860   //
 861   // Generate code which reads from the polling page. This special handling is needed as the
 862   // linux-ppc64 kernel before 2.6.6 doesn't set si_addr on some segfaults in 64bit mode
 863   // (cf. http://www.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.6), especially when we try
 864   // to read from the safepoint polling page.
 865   address generate_load_from_poll() {
 866     StubCodeMark mark(this, "StubRoutines", "generate_load_from_poll");
 867     address start = __ function_entry();
 868     __ unimplemented("StubRoutines::verify_oop", 95);  // TODO PPC port
 869     return start;
 870   }
 871 
 872   // -XX:+OptimizeFill : convert fill/copy loops into intrinsic
 873   //
 874   // The code is implemented(ported from sparc) as we believe it benefits JVM98, however
 875   // tracing(-XX:+TraceOptimizeFill) shows the intrinsic replacement doesn't happen at all!
 876   //
 877   // Source code in function is_range_check_if() shows that OptimizeFill relaxed the condition
 878   // for turning on loop predication optimization, and hence the behavior of "array range check"
 879   // and "loop invariant check" could be influenced, which potentially boosted JVM98.
 880   //
 881   // Generate stub for disjoint short fill. If "aligned" is true, the
 882   // "to" address is assumed to be heapword aligned.
 883   //
 884   // Arguments for generated stub:
 885   //   to:    R3_ARG1
 886   //   value: R4_ARG2
 887   //   count: R5_ARG3 treated as signed
 888   //
 889   address generate_fill(BasicType t, bool aligned, const char* name) {
 890     StubCodeMark mark(this, "StubRoutines", name);
 891     address start = __ function_entry();
 892 
 893     const Register to    = R3_ARG1;   // source array address
 894     const Register value = R4_ARG2;   // fill value
 895     const Register count = R5_ARG3;   // elements count
 896     const Register temp  = R6_ARG4;   // temp register
 897 
 898     //assert_clean_int(count, O3);    // Make sure 'count' is clean int.
 899 
 900     Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
 901     Label L_fill_2_bytes, L_fill_4_bytes, L_fill_elements, L_fill_32_bytes;
 902 
 903     int shift = -1;
 904     switch (t) {
 905        case T_BYTE:
 906         shift = 2;
 907         // Clone bytes (zero extend not needed because store instructions below ignore high order bytes).
 908         __ rldimi(value, value, 8, 48);     // 8 bit -> 16 bit
 909         __ cmpdi(CCR0, count, 2<<shift);    // Short arrays (< 8 bytes) fill by element.
 910         __ blt(CCR0, L_fill_elements);
 911         __ rldimi(value, value, 16, 32);    // 16 bit -> 32 bit
 912         break;
 913        case T_SHORT:
 914         shift = 1;
 915         // Clone bytes (zero extend not needed because store instructions below ignore high order bytes).
 916         __ rldimi(value, value, 16, 32);    // 16 bit -> 32 bit
 917         __ cmpdi(CCR0, count, 2<<shift);    // Short arrays (< 8 bytes) fill by element.
 918         __ blt(CCR0, L_fill_elements);
 919         break;
 920       case T_INT:
 921         shift = 0;
 922         __ cmpdi(CCR0, count, 2<<shift);    // Short arrays (< 8 bytes) fill by element.
 923         __ blt(CCR0, L_fill_4_bytes);
 924         break;
 925       default: ShouldNotReachHere();
 926     }
 927 
 928     if (!aligned && (t == T_BYTE || t == T_SHORT)) {
 929       // Align source address at 4 bytes address boundary.
 930       if (t == T_BYTE) {
 931         // One byte misalignment happens only for byte arrays.
 932         __ andi_(temp, to, 1);
 933         __ beq(CCR0, L_skip_align1);
 934         __ stb(value, 0, to);
 935         __ addi(to, to, 1);
 936         __ addi(count, count, -1);
 937         __ bind(L_skip_align1);
 938       }
 939       // Two bytes misalignment happens only for byte and short (char) arrays.
 940       __ andi_(temp, to, 2);
 941       __ beq(CCR0, L_skip_align2);
 942       __ sth(value, 0, to);
 943       __ addi(to, to, 2);
 944       __ addi(count, count, -(1 << (shift - 1)));
 945       __ bind(L_skip_align2);
 946     }
 947 
 948     if (!aligned) {
 949       // Align to 8 bytes, we know we are 4 byte aligned to start.
 950       __ andi_(temp, to, 7);
 951       __ beq(CCR0, L_fill_32_bytes);
 952       __ stw(value, 0, to);
 953       __ addi(to, to, 4);
 954       __ addi(count, count, -(1 << shift));
 955       __ bind(L_fill_32_bytes);
 956     }
 957 
 958     __ li(temp, 8<<shift);                  // Prepare for 32 byte loop.
 959     // Clone bytes int->long as above.
 960     __ rldimi(value, value, 32, 0);         // 32 bit -> 64 bit
 961 
 962     Label L_check_fill_8_bytes;
 963     // Fill 32-byte chunks.
 964     __ subf_(count, temp, count);
 965     __ blt(CCR0, L_check_fill_8_bytes);
 966 
 967     Label L_fill_32_bytes_loop;
 968     __ align(32);
 969     __ bind(L_fill_32_bytes_loop);
 970 
 971     __ std(value, 0, to);
 972     __ std(value, 8, to);
 973     __ subf_(count, temp, count);           // Update count.
 974     __ std(value, 16, to);
 975     __ std(value, 24, to);
 976 
 977     __ addi(to, to, 32);
 978     __ bge(CCR0, L_fill_32_bytes_loop);
 979 
 980     __ bind(L_check_fill_8_bytes);
 981     __ add_(count, temp, count);
 982     __ beq(CCR0, L_exit);
 983     __ addic_(count, count, -(2 << shift));
 984     __ blt(CCR0, L_fill_4_bytes);
 985 
 986     //
 987     // Length is too short, just fill 8 bytes at a time.
 988     //
 989     Label L_fill_8_bytes_loop;
 990     __ bind(L_fill_8_bytes_loop);
 991     __ std(value, 0, to);
 992     __ addic_(count, count, -(2 << shift));
 993     __ addi(to, to, 8);
 994     __ bge(CCR0, L_fill_8_bytes_loop);
 995 
 996     // Fill trailing 4 bytes.
 997     __ bind(L_fill_4_bytes);
 998     __ andi_(temp, count, 1<<shift);
 999     __ beq(CCR0, L_fill_2_bytes);
1000 
1001     __ stw(value, 0, to);
1002     if (t == T_BYTE || t == T_SHORT) {
1003       __ addi(to, to, 4);
1004       // Fill trailing 2 bytes.
1005       __ bind(L_fill_2_bytes);
1006       __ andi_(temp, count, 1<<(shift-1));
1007       __ beq(CCR0, L_fill_byte);
1008       __ sth(value, 0, to);
1009       if (t == T_BYTE) {
1010         __ addi(to, to, 2);
1011         // Fill trailing byte.
1012         __ bind(L_fill_byte);
1013         __ andi_(count, count, 1);
1014         __ beq(CCR0, L_exit);
1015         __ stb(value, 0, to);
1016       } else {
1017         __ bind(L_fill_byte);
1018       }
1019     } else {
1020       __ bind(L_fill_2_bytes);
1021     }
1022     __ bind(L_exit);
1023     __ blr();
1024 
1025     // Handle copies less than 8 bytes. Int is handled elsewhere.
1026     if (t == T_BYTE) {
1027       __ bind(L_fill_elements);
1028       Label L_fill_2, L_fill_4;
1029       __ andi_(temp, count, 1);
1030       __ beq(CCR0, L_fill_2);
1031       __ stb(value, 0, to);
1032       __ addi(to, to, 1);
1033       __ bind(L_fill_2);
1034       __ andi_(temp, count, 2);
1035       __ beq(CCR0, L_fill_4);
1036       __ stb(value, 0, to);
1037       __ stb(value, 0, to);
1038       __ addi(to, to, 2);
1039       __ bind(L_fill_4);
1040       __ andi_(temp, count, 4);
1041       __ beq(CCR0, L_exit);
1042       __ stb(value, 0, to);
1043       __ stb(value, 1, to);
1044       __ stb(value, 2, to);
1045       __ stb(value, 3, to);
1046       __ blr();
1047     }
1048 
1049     if (t == T_SHORT) {
1050       Label L_fill_2;
1051       __ bind(L_fill_elements);
1052       __ andi_(temp, count, 1);
1053       __ beq(CCR0, L_fill_2);
1054       __ sth(value, 0, to);
1055       __ addi(to, to, 2);
1056       __ bind(L_fill_2);
1057       __ andi_(temp, count, 2);
1058       __ beq(CCR0, L_exit);
1059       __ sth(value, 0, to);
1060       __ sth(value, 2, to);
1061       __ blr();
1062     }
1063     return start;
1064   }
1065 
1066 
1067   // Generate overlap test for array copy stubs.
1068   //
1069   // Input:
1070   //   R3_ARG1    -  from
1071   //   R4_ARG2    -  to
1072   //   R5_ARG3    -  element count
1073   //
1074   void array_overlap_test(address no_overlap_target, int log2_elem_size) {
1075     Register tmp1 = R6_ARG4;
1076     Register tmp2 = R7_ARG5;
1077 
1078     Label l_overlap;
1079 #ifdef ASSERT
1080     __ srdi_(tmp2, R5_ARG3, 31);
1081     __ asm_assert_eq("missing zero extend", 0xAFFE);
1082 #endif
1083 
1084     __ subf(tmp1, R3_ARG1, R4_ARG2); // distance in bytes
1085     __ sldi(tmp2, R5_ARG3, log2_elem_size); // size in bytes
1086     __ cmpld(CCR0, R3_ARG1, R4_ARG2); // Use unsigned comparison!
1087     __ cmpld(CCR1, tmp1, tmp2);
1088     __ crand(/*CCR0 lt*/0, /*CCR1 lt*/4+0, /*CCR0 lt*/0);
1089     __ blt(CCR0, l_overlap); // Src before dst and distance smaller than size.
1090 
1091     // need to copy forwards
1092     if (__ is_within_range_of_b(no_overlap_target, __ pc())) {
1093       __ b(no_overlap_target);
1094     } else {
1095       __ load_const(tmp1, no_overlap_target, tmp2);
1096       __ mtctr(tmp1);
1097       __ bctr();
1098     }
1099 
1100     __ bind(l_overlap);
1101     // need to copy backwards
1102   }
1103 
1104   // The guideline in the implementations of generate_disjoint_xxx_copy
1105   // (xxx=byte,short,int,long,oop) is to copy as many elements as possible with
1106   // single instructions, but to avoid alignment interrupts (see subsequent
1107   // comment). Furthermore, we try to minimize misaligned access, even
1108   // though they cause no alignment interrupt.
1109   //
1110   // In Big-Endian mode, the PowerPC architecture requires implementations to
1111   // handle automatically misaligned integer halfword and word accesses,
1112   // word-aligned integer doubleword accesses, and word-aligned floating-point
1113   // accesses. Other accesses may or may not generate an Alignment interrupt
1114   // depending on the implementation.
1115   // Alignment interrupt handling may require on the order of hundreds of cycles,
1116   // so every effort should be made to avoid misaligned memory values.
1117   //
1118   //
1119   // Generate stub for disjoint byte copy.  If "aligned" is true, the
1120   // "from" and "to" addresses are assumed to be heapword aligned.
1121   //
1122   // Arguments for generated stub:
1123   //      from:  R3_ARG1
1124   //      to:    R4_ARG2
1125   //      count: R5_ARG3 treated as signed
1126   //
1127   address generate_disjoint_byte_copy(bool aligned, const char * name) {
1128     StubCodeMark mark(this, "StubRoutines", name);
1129     address start = __ function_entry();
1130 
1131     Register tmp1 = R6_ARG4;
1132     Register tmp2 = R7_ARG5;
1133     Register tmp3 = R8_ARG6;
1134     Register tmp4 = R9_ARG7;
1135 
1136 
1137     Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8, l_9;
1138     // Don't try anything fancy if arrays don't have many elements.
1139     __ li(tmp3, 0);
1140     __ cmpwi(CCR0, R5_ARG3, 17);
1141     __ ble(CCR0, l_6); // copy 4 at a time
1142 
1143     if (!aligned) {
1144       __ xorr(tmp1, R3_ARG1, R4_ARG2);
1145       __ andi_(tmp1, tmp1, 3);
1146       __ bne(CCR0, l_6); // If arrays don't have the same alignment mod 4, do 4 element copy.
1147 
1148       // Copy elements if necessary to align to 4 bytes.
1149       __ neg(tmp1, R3_ARG1); // Compute distance to alignment boundary.
1150       __ andi_(tmp1, tmp1, 3);
1151       __ beq(CCR0, l_2);
1152 
1153       __ subf(R5_ARG3, tmp1, R5_ARG3);
1154       __ bind(l_9);
1155       __ lbz(tmp2, 0, R3_ARG1);
1156       __ addic_(tmp1, tmp1, -1);
1157       __ stb(tmp2, 0, R4_ARG2);
1158       __ addi(R3_ARG1, R3_ARG1, 1);
1159       __ addi(R4_ARG2, R4_ARG2, 1);
1160       __ bne(CCR0, l_9);
1161 
1162       __ bind(l_2);
1163     }
1164 
1165     // copy 8 elements at a time
1166     __ xorr(tmp2, R3_ARG1, R4_ARG2); // skip if src & dest have differing alignment mod 8
1167     __ andi_(tmp1, tmp2, 7);
1168     __ bne(CCR0, l_7); // not same alignment -> to or from is aligned -> copy 8
1169 
1170     // copy a 2-element word if necessary to align to 8 bytes
1171     __ andi_(R0, R3_ARG1, 7);
1172     __ beq(CCR0, l_7);
1173 
1174     __ lwzx(tmp2, R3_ARG1, tmp3);
1175     __ addi(R5_ARG3, R5_ARG3, -4);
1176     __ stwx(tmp2, R4_ARG2, tmp3);
1177     { // FasterArrayCopy
1178       __ addi(R3_ARG1, R3_ARG1, 4);
1179       __ addi(R4_ARG2, R4_ARG2, 4);
1180     }
1181     __ bind(l_7);
1182 
1183     { // FasterArrayCopy
1184       __ cmpwi(CCR0, R5_ARG3, 31);
1185       __ ble(CCR0, l_6); // copy 2 at a time if less than 32 elements remain
1186 
1187       __ srdi(tmp1, R5_ARG3, 5);
1188       __ andi_(R5_ARG3, R5_ARG3, 31);
1189       __ mtctr(tmp1);
1190 
1191       __ bind(l_8);
1192       // Use unrolled version for mass copying (copy 32 elements a time)
1193       // Load feeding store gets zero latency on Power6, however not on Power5.
1194       // Therefore, the following sequence is made for the good of both.
1195       __ ld(tmp1, 0, R3_ARG1);
1196       __ ld(tmp2, 8, R3_ARG1);
1197       __ ld(tmp3, 16, R3_ARG1);
1198       __ ld(tmp4, 24, R3_ARG1);
1199       __ std(tmp1, 0, R4_ARG2);
1200       __ std(tmp2, 8, R4_ARG2);
1201       __ std(tmp3, 16, R4_ARG2);
1202       __ std(tmp4, 24, R4_ARG2);
1203       __ addi(R3_ARG1, R3_ARG1, 32);
1204       __ addi(R4_ARG2, R4_ARG2, 32);
1205       __ bdnz(l_8);
1206     }
1207 
1208     __ bind(l_6);
1209 
1210     // copy 4 elements at a time
1211     __ cmpwi(CCR0, R5_ARG3, 4);
1212     __ blt(CCR0, l_1);
1213     __ srdi(tmp1, R5_ARG3, 2);
1214     __ mtctr(tmp1); // is > 0
1215     __ andi_(R5_ARG3, R5_ARG3, 3);
1216 
1217     { // FasterArrayCopy
1218       __ addi(R3_ARG1, R3_ARG1, -4);
1219       __ addi(R4_ARG2, R4_ARG2, -4);
1220       __ bind(l_3);
1221       __ lwzu(tmp2, 4, R3_ARG1);
1222       __ stwu(tmp2, 4, R4_ARG2);
1223       __ bdnz(l_3);
1224       __ addi(R3_ARG1, R3_ARG1, 4);
1225       __ addi(R4_ARG2, R4_ARG2, 4);
1226     }
1227 
1228     // do single element copy
1229     __ bind(l_1);
1230     __ cmpwi(CCR0, R5_ARG3, 0);
1231     __ beq(CCR0, l_4);
1232 
1233     { // FasterArrayCopy
1234       __ mtctr(R5_ARG3);
1235       __ addi(R3_ARG1, R3_ARG1, -1);
1236       __ addi(R4_ARG2, R4_ARG2, -1);
1237 
1238       __ bind(l_5);
1239       __ lbzu(tmp2, 1, R3_ARG1);
1240       __ stbu(tmp2, 1, R4_ARG2);
1241       __ bdnz(l_5);
1242     }
1243 
1244     __ bind(l_4);
1245     __ blr();
1246 
1247     return start;
1248   }
1249 
1250   // Generate stub for conjoint byte copy.  If "aligned" is true, the
1251   // "from" and "to" addresses are assumed to be heapword aligned.
1252   //
1253   // Arguments for generated stub:
1254   //      from:  R3_ARG1
1255   //      to:    R4_ARG2
1256   //      count: R5_ARG3 treated as signed
1257   //
1258   address generate_conjoint_byte_copy(bool aligned, const char * name) {
1259     StubCodeMark mark(this, "StubRoutines", name);
1260     address start = __ function_entry();
1261 
1262     Register tmp1 = R6_ARG4;
1263     Register tmp2 = R7_ARG5;
1264     Register tmp3 = R8_ARG6;
1265 
1266 #if defined(ABI_ELFv2)
1267      address nooverlap_target = aligned ?
1268        StubRoutines::arrayof_jbyte_disjoint_arraycopy() :
1269        StubRoutines::jbyte_disjoint_arraycopy();
1270 #else
1271     address nooverlap_target = aligned ?
1272       ((FunctionDescriptor*)StubRoutines::arrayof_jbyte_disjoint_arraycopy())->entry() :
1273       ((FunctionDescriptor*)StubRoutines::jbyte_disjoint_arraycopy())->entry();
1274 #endif
1275 
1276     array_overlap_test(nooverlap_target, 0);
1277     // Do reverse copy. We assume the case of actual overlap is rare enough
1278     // that we don't have to optimize it.
1279     Label l_1, l_2;
1280 
1281     __ b(l_2);
1282     __ bind(l_1);
1283     __ stbx(tmp1, R4_ARG2, R5_ARG3);
1284     __ bind(l_2);
1285     __ addic_(R5_ARG3, R5_ARG3, -1);
1286     __ lbzx(tmp1, R3_ARG1, R5_ARG3);
1287     __ bge(CCR0, l_1);
1288 
1289     __ blr();
1290 
1291     return start;
1292   }
1293 
1294   // Generate stub for disjoint short copy.  If "aligned" is true, the
1295   // "from" and "to" addresses are assumed to be heapword aligned.
1296   //
1297   // Arguments for generated stub:
1298   //      from:  R3_ARG1
1299   //      to:    R4_ARG2
1300   //  elm.count: R5_ARG3 treated as signed
1301   //
1302   // Strategy for aligned==true:
1303   //
1304   //  If length <= 9:
1305   //     1. copy 2 elements at a time (l_6)
1306   //     2. copy last element if original element count was odd (l_1)
1307   //
1308   //  If length > 9:
1309   //     1. copy 4 elements at a time until less than 4 elements are left (l_7)
1310   //     2. copy 2 elements at a time until less than 2 elements are left (l_6)
1311   //     3. copy last element if one was left in step 2. (l_1)
1312   //
1313   //
1314   // Strategy for aligned==false:
1315   //
1316   //  If length <= 9: same as aligned==true case, but NOTE: load/stores
1317   //                  can be unaligned (see comment below)
1318   //
1319   //  If length > 9:
1320   //     1. continue with step 6. if the alignment of from and to mod 4
1321   //        is different.
1322   //     2. align from and to to 4 bytes by copying 1 element if necessary
1323   //     3. at l_2 from and to are 4 byte aligned; continue with
1324   //        5. if they cannot be aligned to 8 bytes because they have
1325   //        got different alignment mod 8.
1326   //     4. at this point we know that both, from and to, have the same
1327   //        alignment mod 8, now copy one element if necessary to get
1328   //        8 byte alignment of from and to.
1329   //     5. copy 4 elements at a time until less than 4 elements are
1330   //        left; depending on step 3. all load/stores are aligned or
1331   //        either all loads or all stores are unaligned.
1332   //     6. copy 2 elements at a time until less than 2 elements are
1333   //        left (l_6); arriving here from step 1., there is a chance
1334   //        that all accesses are unaligned.
1335   //     7. copy last element if one was left in step 6. (l_1)
1336   //
1337   //  There are unaligned data accesses using integer load/store
1338   //  instructions in this stub. POWER allows such accesses.
1339   //
1340   //  According to the manuals (PowerISA_V2.06_PUBLIC, Book II,
1341   //  Chapter 2: Effect of Operand Placement on Performance) unaligned
1342   //  integer load/stores have good performance. Only unaligned
1343   //  floating point load/stores can have poor performance.
1344   //
1345   //  TODO:
1346   //
1347   //  1. check if aligning the backbranch target of loops is beneficial
1348   //
1349   address generate_disjoint_short_copy(bool aligned, const char * name) {
1350     StubCodeMark mark(this, "StubRoutines", name);
1351 
1352     Register tmp1 = R6_ARG4;
1353     Register tmp2 = R7_ARG5;
1354     Register tmp3 = R8_ARG6;
1355     Register tmp4 = R9_ARG7;
1356 
1357     address start = __ function_entry();
1358 
1359       Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8;
1360     // don't try anything fancy if arrays don't have many elements
1361     __ li(tmp3, 0);
1362     __ cmpwi(CCR0, R5_ARG3, 9);
1363     __ ble(CCR0, l_6); // copy 2 at a time
1364 
1365     if (!aligned) {
1366       __ xorr(tmp1, R3_ARG1, R4_ARG2);
1367       __ andi_(tmp1, tmp1, 3);
1368       __ bne(CCR0, l_6); // if arrays don't have the same alignment mod 4, do 2 element copy
1369 
1370       // At this point it is guaranteed that both, from and to have the same alignment mod 4.
1371 
1372       // Copy 1 element if necessary to align to 4 bytes.
1373       __ andi_(tmp1, R3_ARG1, 3);
1374       __ beq(CCR0, l_2);
1375 
1376       __ lhz(tmp2, 0, R3_ARG1);
1377       __ addi(R3_ARG1, R3_ARG1, 2);
1378       __ sth(tmp2, 0, R4_ARG2);
1379       __ addi(R4_ARG2, R4_ARG2, 2);
1380       __ addi(R5_ARG3, R5_ARG3, -1);
1381       __ bind(l_2);
1382 
1383       // At this point the positions of both, from and to, are at least 4 byte aligned.
1384 
1385       // Copy 4 elements at a time.
1386       // Align to 8 bytes, but only if both, from and to, have same alignment mod 8.
1387       __ xorr(tmp2, R3_ARG1, R4_ARG2);
1388       __ andi_(tmp1, tmp2, 7);
1389       __ bne(CCR0, l_7); // not same alignment mod 8 -> copy 4, either from or to will be unaligned
1390 
1391       // Copy a 2-element word if necessary to align to 8 bytes.
1392       __ andi_(R0, R3_ARG1, 7);
1393       __ beq(CCR0, l_7);
1394 
1395       __ lwzx(tmp2, R3_ARG1, tmp3);
1396       __ addi(R5_ARG3, R5_ARG3, -2);
1397       __ stwx(tmp2, R4_ARG2, tmp3);
1398       { // FasterArrayCopy
1399         __ addi(R3_ARG1, R3_ARG1, 4);
1400         __ addi(R4_ARG2, R4_ARG2, 4);
1401       }
1402     }
1403 
1404     __ bind(l_7);
1405 
1406     // Copy 4 elements at a time; either the loads or the stores can
1407     // be unaligned if aligned == false.
1408 
1409     { // FasterArrayCopy
1410       __ cmpwi(CCR0, R5_ARG3, 15);
1411       __ ble(CCR0, l_6); // copy 2 at a time if less than 16 elements remain
1412 
1413       __ srdi(tmp1, R5_ARG3, 4);
1414       __ andi_(R5_ARG3, R5_ARG3, 15);
1415       __ mtctr(tmp1);
1416 
1417       __ bind(l_8);
1418       // Use unrolled version for mass copying (copy 16 elements a time).
1419       // Load feeding store gets zero latency on Power6, however not on Power5.
1420       // Therefore, the following sequence is made for the good of both.
1421       __ ld(tmp1, 0, R3_ARG1);
1422       __ ld(tmp2, 8, R3_ARG1);
1423       __ ld(tmp3, 16, R3_ARG1);
1424       __ ld(tmp4, 24, R3_ARG1);
1425       __ std(tmp1, 0, R4_ARG2);
1426       __ std(tmp2, 8, R4_ARG2);
1427       __ std(tmp3, 16, R4_ARG2);
1428       __ std(tmp4, 24, R4_ARG2);
1429       __ addi(R3_ARG1, R3_ARG1, 32);
1430       __ addi(R4_ARG2, R4_ARG2, 32);
1431       __ bdnz(l_8);
1432     }
1433     __ bind(l_6);
1434 
1435     // copy 2 elements at a time
1436     { // FasterArrayCopy
1437       __ cmpwi(CCR0, R5_ARG3, 2);
1438       __ blt(CCR0, l_1);
1439       __ srdi(tmp1, R5_ARG3, 1);
1440       __ andi_(R5_ARG3, R5_ARG3, 1);
1441 
1442       __ addi(R3_ARG1, R3_ARG1, -4);
1443       __ addi(R4_ARG2, R4_ARG2, -4);
1444       __ mtctr(tmp1);
1445 
1446       __ bind(l_3);
1447       __ lwzu(tmp2, 4, R3_ARG1);
1448       __ stwu(tmp2, 4, R4_ARG2);
1449       __ bdnz(l_3);
1450 
1451       __ addi(R3_ARG1, R3_ARG1, 4);
1452       __ addi(R4_ARG2, R4_ARG2, 4);
1453     }
1454 
1455     // do single element copy
1456     __ bind(l_1);
1457     __ cmpwi(CCR0, R5_ARG3, 0);
1458     __ beq(CCR0, l_4);
1459 
1460     { // FasterArrayCopy
1461       __ mtctr(R5_ARG3);
1462       __ addi(R3_ARG1, R3_ARG1, -2);
1463       __ addi(R4_ARG2, R4_ARG2, -2);
1464 
1465       __ bind(l_5);
1466       __ lhzu(tmp2, 2, R3_ARG1);
1467       __ sthu(tmp2, 2, R4_ARG2);
1468       __ bdnz(l_5);
1469     }
1470     __ bind(l_4);
1471     __ blr();
1472 
1473     return start;
1474   }
1475 
1476   // Generate stub for conjoint short copy.  If "aligned" is true, the
1477   // "from" and "to" addresses are assumed to be heapword aligned.
1478   //
1479   // Arguments for generated stub:
1480   //      from:  R3_ARG1
1481   //      to:    R4_ARG2
1482   //      count: R5_ARG3 treated as signed
1483   //
1484   address generate_conjoint_short_copy(bool aligned, const char * name) {
1485     StubCodeMark mark(this, "StubRoutines", name);
1486     address start = __ function_entry();
1487 
1488     Register tmp1 = R6_ARG4;
1489     Register tmp2 = R7_ARG5;
1490     Register tmp3 = R8_ARG6;
1491 
1492 #if defined(ABI_ELFv2)
1493     address nooverlap_target = aligned ?
1494         StubRoutines::arrayof_jshort_disjoint_arraycopy() :
1495         StubRoutines::jshort_disjoint_arraycopy();
1496 #else
1497     address nooverlap_target = aligned ?
1498         ((FunctionDescriptor*)StubRoutines::arrayof_jshort_disjoint_arraycopy())->entry() :
1499         ((FunctionDescriptor*)StubRoutines::jshort_disjoint_arraycopy())->entry();
1500 #endif
1501 
1502     array_overlap_test(nooverlap_target, 1);
1503 
1504     Label l_1, l_2;
1505     __ sldi(tmp1, R5_ARG3, 1);
1506     __ b(l_2);
1507     __ bind(l_1);
1508     __ sthx(tmp2, R4_ARG2, tmp1);
1509     __ bind(l_2);
1510     __ addic_(tmp1, tmp1, -2);
1511     __ lhzx(tmp2, R3_ARG1, tmp1);
1512     __ bge(CCR0, l_1);
1513 
1514     __ blr();
1515 
1516     return start;
1517   }
1518 
1519   // Generate core code for disjoint int copy (and oop copy on 32-bit).  If "aligned"
1520   // is true, the "from" and "to" addresses are assumed to be heapword aligned.
1521   //
1522   // Arguments:
1523   //      from:  R3_ARG1
1524   //      to:    R4_ARG2
1525   //      count: R5_ARG3 treated as signed
1526   //
1527   void generate_disjoint_int_copy_core(bool aligned) {
1528     Register tmp1 = R6_ARG4;
1529     Register tmp2 = R7_ARG5;
1530     Register tmp3 = R8_ARG6;
1531     Register tmp4 = R0;
1532 
1533     Label l_1, l_2, l_3, l_4, l_5, l_6;
1534     // for short arrays, just do single element copy
1535     __ li(tmp3, 0);
1536     __ cmpwi(CCR0, R5_ARG3, 5);
1537     __ ble(CCR0, l_2);
1538 
1539     if (!aligned) {
1540         // check if arrays have same alignment mod 8.
1541         __ xorr(tmp1, R3_ARG1, R4_ARG2);
1542         __ andi_(R0, tmp1, 7);
1543         // Not the same alignment, but ld and std just need to be 4 byte aligned.
1544         __ bne(CCR0, l_4); // to OR from is 8 byte aligned -> copy 2 at a time
1545 
1546         // copy 1 element to align to and from on an 8 byte boundary
1547         __ andi_(R0, R3_ARG1, 7);
1548         __ beq(CCR0, l_4);
1549 
1550         __ lwzx(tmp2, R3_ARG1, tmp3);
1551         __ addi(R5_ARG3, R5_ARG3, -1);
1552         __ stwx(tmp2, R4_ARG2, tmp3);
1553         { // FasterArrayCopy
1554           __ addi(R3_ARG1, R3_ARG1, 4);
1555           __ addi(R4_ARG2, R4_ARG2, 4);
1556         }
1557         __ bind(l_4);
1558       }
1559 
1560     { // FasterArrayCopy
1561       __ cmpwi(CCR0, R5_ARG3, 7);
1562       __ ble(CCR0, l_2); // copy 1 at a time if less than 8 elements remain
1563 
1564       __ srdi(tmp1, R5_ARG3, 3);
1565       __ andi_(R5_ARG3, R5_ARG3, 7);
1566       __ mtctr(tmp1);
1567 
1568       __ bind(l_6);
1569       // Use unrolled version for mass copying (copy 8 elements a time).
1570       // Load feeding store gets zero latency on power6, however not on power 5.
1571       // Therefore, the following sequence is made for the good of both.
1572       __ ld(tmp1, 0, R3_ARG1);
1573       __ ld(tmp2, 8, R3_ARG1);
1574       __ ld(tmp3, 16, R3_ARG1);
1575       __ ld(tmp4, 24, R3_ARG1);
1576       __ std(tmp1, 0, R4_ARG2);
1577       __ std(tmp2, 8, R4_ARG2);
1578       __ std(tmp3, 16, R4_ARG2);
1579       __ std(tmp4, 24, R4_ARG2);
1580       __ addi(R3_ARG1, R3_ARG1, 32);
1581       __ addi(R4_ARG2, R4_ARG2, 32);
1582       __ bdnz(l_6);
1583     }
1584 
1585     // copy 1 element at a time
1586     __ bind(l_2);
1587     __ cmpwi(CCR0, R5_ARG3, 0);
1588     __ beq(CCR0, l_1);
1589 
1590     { // FasterArrayCopy
1591       __ mtctr(R5_ARG3);
1592       __ addi(R3_ARG1, R3_ARG1, -4);
1593       __ addi(R4_ARG2, R4_ARG2, -4);
1594 
1595       __ bind(l_3);
1596       __ lwzu(tmp2, 4, R3_ARG1);
1597       __ stwu(tmp2, 4, R4_ARG2);
1598       __ bdnz(l_3);
1599     }
1600 
1601     __ bind(l_1);
1602     return;
1603   }
1604 
1605   // Generate stub for disjoint int copy.  If "aligned" is true, the
1606   // "from" and "to" addresses are assumed to be heapword aligned.
1607   //
1608   // Arguments for generated stub:
1609   //      from:  R3_ARG1
1610   //      to:    R4_ARG2
1611   //      count: R5_ARG3 treated as signed
1612   //
1613   address generate_disjoint_int_copy(bool aligned, const char * name) {
1614     StubCodeMark mark(this, "StubRoutines", name);
1615     address start = __ function_entry();
1616     generate_disjoint_int_copy_core(aligned);
1617     __ blr();
1618     return start;
1619   }
1620 
1621   // Generate core code for conjoint int copy (and oop copy on
1622   // 32-bit).  If "aligned" is true, the "from" and "to" addresses
1623   // are assumed to be heapword aligned.
1624   //
1625   // Arguments:
1626   //      from:  R3_ARG1
1627   //      to:    R4_ARG2
1628   //      count: R5_ARG3 treated as signed
1629   //
1630   void generate_conjoint_int_copy_core(bool aligned) {
1631     // Do reverse copy.  We assume the case of actual overlap is rare enough
1632     // that we don't have to optimize it.
1633 
1634     Label l_1, l_2, l_3, l_4, l_5, l_6;
1635 
1636     Register tmp1 = R6_ARG4;
1637     Register tmp2 = R7_ARG5;
1638     Register tmp3 = R8_ARG6;
1639     Register tmp4 = R0;
1640 
1641     { // FasterArrayCopy
1642       __ cmpwi(CCR0, R5_ARG3, 0);
1643       __ beq(CCR0, l_6);
1644 
1645       __ sldi(R5_ARG3, R5_ARG3, 2);
1646       __ add(R3_ARG1, R3_ARG1, R5_ARG3);
1647       __ add(R4_ARG2, R4_ARG2, R5_ARG3);
1648       __ srdi(R5_ARG3, R5_ARG3, 2);
1649 
1650       __ cmpwi(CCR0, R5_ARG3, 7);
1651       __ ble(CCR0, l_5); // copy 1 at a time if less than 8 elements remain
1652 
1653       __ srdi(tmp1, R5_ARG3, 3);
1654       __ andi(R5_ARG3, R5_ARG3, 7);
1655       __ mtctr(tmp1);
1656 
1657       __ bind(l_4);
1658       // Use unrolled version for mass copying (copy 4 elements a time).
1659       // Load feeding store gets zero latency on Power6, however not on Power5.
1660       // Therefore, the following sequence is made for the good of both.
1661       __ addi(R3_ARG1, R3_ARG1, -32);
1662       __ addi(R4_ARG2, R4_ARG2, -32);
1663       __ ld(tmp4, 24, R3_ARG1);
1664       __ ld(tmp3, 16, R3_ARG1);
1665       __ ld(tmp2, 8, R3_ARG1);
1666       __ ld(tmp1, 0, R3_ARG1);
1667       __ std(tmp4, 24, R4_ARG2);
1668       __ std(tmp3, 16, R4_ARG2);
1669       __ std(tmp2, 8, R4_ARG2);
1670       __ std(tmp1, 0, R4_ARG2);
1671       __ bdnz(l_4);
1672 
1673       __ cmpwi(CCR0, R5_ARG3, 0);
1674       __ beq(CCR0, l_6);
1675 
1676       __ bind(l_5);
1677       __ mtctr(R5_ARG3);
1678       __ bind(l_3);
1679       __ lwz(R0, -4, R3_ARG1);
1680       __ stw(R0, -4, R4_ARG2);
1681       __ addi(R3_ARG1, R3_ARG1, -4);
1682       __ addi(R4_ARG2, R4_ARG2, -4);
1683       __ bdnz(l_3);
1684 
1685       __ bind(l_6);
1686     }
1687   }
1688 
1689   // Generate stub for conjoint int copy.  If "aligned" is true, the
1690   // "from" and "to" addresses are assumed to be heapword aligned.
1691   //
1692   // Arguments for generated stub:
1693   //      from:  R3_ARG1
1694   //      to:    R4_ARG2
1695   //      count: R5_ARG3 treated as signed
1696   //
1697   address generate_conjoint_int_copy(bool aligned, const char * name) {
1698     StubCodeMark mark(this, "StubRoutines", name);
1699     address start = __ function_entry();
1700 
1701 #if defined(ABI_ELFv2)
1702     address nooverlap_target = aligned ?
1703       StubRoutines::arrayof_jint_disjoint_arraycopy() :
1704       StubRoutines::jint_disjoint_arraycopy();
1705 #else
1706     address nooverlap_target = aligned ?
1707       ((FunctionDescriptor*)StubRoutines::arrayof_jint_disjoint_arraycopy())->entry() :
1708       ((FunctionDescriptor*)StubRoutines::jint_disjoint_arraycopy())->entry();
1709 #endif
1710 
1711     array_overlap_test(nooverlap_target, 2);
1712 
1713     generate_conjoint_int_copy_core(aligned);
1714 
1715     __ blr();
1716 
1717     return start;
1718   }
1719 
1720   // Generate core code for disjoint long copy (and oop copy on
1721   // 64-bit).  If "aligned" is true, the "from" and "to" addresses
1722   // are assumed to be heapword aligned.
1723   //
1724   // Arguments:
1725   //      from:  R3_ARG1
1726   //      to:    R4_ARG2
1727   //      count: R5_ARG3 treated as signed
1728   //
1729   void generate_disjoint_long_copy_core(bool aligned) {
1730     Register tmp1 = R6_ARG4;
1731     Register tmp2 = R7_ARG5;
1732     Register tmp3 = R8_ARG6;
1733     Register tmp4 = R0;
1734 
1735     Label l_1, l_2, l_3, l_4;
1736 
1737     { // FasterArrayCopy
1738       __ cmpwi(CCR0, R5_ARG3, 3);
1739       __ ble(CCR0, l_3); // copy 1 at a time if less than 4 elements remain
1740 
1741       __ srdi(tmp1, R5_ARG3, 2);
1742       __ andi_(R5_ARG3, R5_ARG3, 3);
1743       __ mtctr(tmp1);
1744 
1745       __ bind(l_4);
1746       // Use unrolled version for mass copying (copy 4 elements a time).
1747       // Load feeding store gets zero latency on Power6, however not on Power5.
1748       // Therefore, the following sequence is made for the good of both.
1749       __ ld(tmp1, 0, R3_ARG1);
1750       __ ld(tmp2, 8, R3_ARG1);
1751       __ ld(tmp3, 16, R3_ARG1);
1752       __ ld(tmp4, 24, R3_ARG1);
1753       __ std(tmp1, 0, R4_ARG2);
1754       __ std(tmp2, 8, R4_ARG2);
1755       __ std(tmp3, 16, R4_ARG2);
1756       __ std(tmp4, 24, R4_ARG2);
1757       __ addi(R3_ARG1, R3_ARG1, 32);
1758       __ addi(R4_ARG2, R4_ARG2, 32);
1759       __ bdnz(l_4);
1760     }
1761 
1762     // copy 1 element at a time
1763     __ bind(l_3);
1764     __ cmpwi(CCR0, R5_ARG3, 0);
1765     __ beq(CCR0, l_1);
1766 
1767     { // FasterArrayCopy
1768       __ mtctr(R5_ARG3);
1769       __ addi(R3_ARG1, R3_ARG1, -8);
1770       __ addi(R4_ARG2, R4_ARG2, -8);
1771 
1772       __ bind(l_2);
1773       __ ldu(R0, 8, R3_ARG1);
1774       __ stdu(R0, 8, R4_ARG2);
1775       __ bdnz(l_2);
1776 
1777     }
1778     __ bind(l_1);
1779   }
1780 
1781   // Generate stub for disjoint long copy.  If "aligned" is true, the
1782   // "from" and "to" addresses are assumed to be heapword aligned.
1783   //
1784   // Arguments for generated stub:
1785   //      from:  R3_ARG1
1786   //      to:    R4_ARG2
1787   //      count: R5_ARG3 treated as signed
1788   //
1789   address generate_disjoint_long_copy(bool aligned, const char * name) {
1790     StubCodeMark mark(this, "StubRoutines", name);
1791     address start = __ function_entry();
1792     generate_disjoint_long_copy_core(aligned);
1793     __ blr();
1794 
1795     return start;
1796   }
1797 
1798   // Generate core code for conjoint long copy (and oop copy on
1799   // 64-bit).  If "aligned" is true, the "from" and "to" addresses
1800   // are assumed to be heapword aligned.
1801   //
1802   // Arguments:
1803   //      from:  R3_ARG1
1804   //      to:    R4_ARG2
1805   //      count: R5_ARG3 treated as signed
1806   //
1807   void generate_conjoint_long_copy_core(bool aligned) {
1808     Register tmp1 = R6_ARG4;
1809     Register tmp2 = R7_ARG5;
1810     Register tmp3 = R8_ARG6;
1811     Register tmp4 = R0;
1812 
1813     Label l_1, l_2, l_3, l_4, l_5;
1814 
1815     __ cmpwi(CCR0, R5_ARG3, 0);
1816     __ beq(CCR0, l_1);
1817 
1818     { // FasterArrayCopy
1819       __ sldi(R5_ARG3, R5_ARG3, 3);
1820       __ add(R3_ARG1, R3_ARG1, R5_ARG3);
1821       __ add(R4_ARG2, R4_ARG2, R5_ARG3);
1822       __ srdi(R5_ARG3, R5_ARG3, 3);
1823 
1824       __ cmpwi(CCR0, R5_ARG3, 3);
1825       __ ble(CCR0, l_5); // copy 1 at a time if less than 4 elements remain
1826 
1827       __ srdi(tmp1, R5_ARG3, 2);
1828       __ andi(R5_ARG3, R5_ARG3, 3);
1829       __ mtctr(tmp1);
1830 
1831       __ bind(l_4);
1832       // Use unrolled version for mass copying (copy 4 elements a time).
1833       // Load feeding store gets zero latency on Power6, however not on Power5.
1834       // Therefore, the following sequence is made for the good of both.
1835       __ addi(R3_ARG1, R3_ARG1, -32);
1836       __ addi(R4_ARG2, R4_ARG2, -32);
1837       __ ld(tmp4, 24, R3_ARG1);
1838       __ ld(tmp3, 16, R3_ARG1);
1839       __ ld(tmp2, 8, R3_ARG1);
1840       __ ld(tmp1, 0, R3_ARG1);
1841       __ std(tmp4, 24, R4_ARG2);
1842       __ std(tmp3, 16, R4_ARG2);
1843       __ std(tmp2, 8, R4_ARG2);
1844       __ std(tmp1, 0, R4_ARG2);
1845       __ bdnz(l_4);
1846 
1847       __ cmpwi(CCR0, R5_ARG3, 0);
1848       __ beq(CCR0, l_1);
1849 
1850       __ bind(l_5);
1851       __ mtctr(R5_ARG3);
1852       __ bind(l_3);
1853       __ ld(R0, -8, R3_ARG1);
1854       __ std(R0, -8, R4_ARG2);
1855       __ addi(R3_ARG1, R3_ARG1, -8);
1856       __ addi(R4_ARG2, R4_ARG2, -8);
1857       __ bdnz(l_3);
1858 
1859     }
1860     __ bind(l_1);
1861   }
1862 
1863   // Generate stub for conjoint long copy.  If "aligned" is true, the
1864   // "from" and "to" addresses are assumed to be heapword aligned.
1865   //
1866   // Arguments for generated stub:
1867   //      from:  R3_ARG1
1868   //      to:    R4_ARG2
1869   //      count: R5_ARG3 treated as signed
1870   //
1871   address generate_conjoint_long_copy(bool aligned, const char * name) {
1872     StubCodeMark mark(this, "StubRoutines", name);
1873     address start = __ function_entry();
1874 
1875 #if defined(ABI_ELFv2)
1876     address nooverlap_target = aligned ?
1877       StubRoutines::arrayof_jlong_disjoint_arraycopy() :
1878       StubRoutines::jlong_disjoint_arraycopy();
1879 #else
1880     address nooverlap_target = aligned ?
1881       ((FunctionDescriptor*)StubRoutines::arrayof_jlong_disjoint_arraycopy())->entry() :
1882       ((FunctionDescriptor*)StubRoutines::jlong_disjoint_arraycopy())->entry();
1883 #endif
1884 
1885     array_overlap_test(nooverlap_target, 3);
1886     generate_conjoint_long_copy_core(aligned);
1887 
1888     __ blr();
1889 
1890     return start;
1891   }
1892 
1893   // Generate stub for conjoint oop copy.  If "aligned" is true, the
1894   // "from" and "to" addresses are assumed to be heapword aligned.
1895   //
1896   // Arguments for generated stub:
1897   //      from:  R3_ARG1
1898   //      to:    R4_ARG2
1899   //      count: R5_ARG3 treated as signed
1900   //      dest_uninitialized: G1 support
1901   //
1902   address generate_conjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) {
1903     StubCodeMark mark(this, "StubRoutines", name);
1904 
1905     address start = __ function_entry();
1906 
1907 #if defined(ABI_ELFv2)
1908     address nooverlap_target = aligned ?
1909       StubRoutines::arrayof_oop_disjoint_arraycopy() :
1910       StubRoutines::oop_disjoint_arraycopy();
1911 #else
1912     address nooverlap_target = aligned ?
1913       ((FunctionDescriptor*)StubRoutines::arrayof_oop_disjoint_arraycopy())->entry() :
1914       ((FunctionDescriptor*)StubRoutines::oop_disjoint_arraycopy())->entry();
1915 #endif
1916 
1917     gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7);
1918 
1919     // Save arguments.
1920     __ mr(R9_ARG7, R4_ARG2);
1921     __ mr(R10_ARG8, R5_ARG3);
1922 
1923     if (UseCompressedOops) {
1924       array_overlap_test(nooverlap_target, 2);
1925       generate_conjoint_int_copy_core(aligned);
1926     } else {
1927       array_overlap_test(nooverlap_target, 3);
1928       generate_conjoint_long_copy_core(aligned);
1929     }
1930 
1931     gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1, /*branchToEnd*/ false);
1932     return start;
1933   }
1934 
1935   // Generate stub for disjoint oop copy.  If "aligned" is true, the
1936   // "from" and "to" addresses are assumed to be heapword aligned.
1937   //
1938   // Arguments for generated stub:
1939   //      from:  R3_ARG1
1940   //      to:    R4_ARG2
1941   //      count: R5_ARG3 treated as signed
1942   //      dest_uninitialized: G1 support
1943   //
1944   address generate_disjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) {
1945     StubCodeMark mark(this, "StubRoutines", name);
1946     address start = __ function_entry();
1947 
1948     gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7);
1949 
1950     // save some arguments, disjoint_long_copy_core destroys them.
1951     // needed for post barrier
1952     __ mr(R9_ARG7, R4_ARG2);
1953     __ mr(R10_ARG8, R5_ARG3);
1954 
1955     if (UseCompressedOops) {
1956       generate_disjoint_int_copy_core(aligned);
1957     } else {
1958       generate_disjoint_long_copy_core(aligned);
1959     }
1960 
1961     gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1, /*branchToEnd*/ false);
1962 
1963     return start;
1964   }
1965 
1966   // Arguments for generated stub (little endian only):
1967   //   R3_ARG1   - source byte array address
1968   //   R4_ARG2   - destination byte array address
1969   //   R5_ARG3   - round key array
1970   address generate_aescrypt_encryptBlock() {
1971     assert(UseAES, "need AES instructions and misaligned SSE support");
1972     StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
1973 
1974     address start = __ function_entry();
1975 
1976     Label L_doLast;
1977 
1978     Register from           = R3_ARG1;  // source array address
1979     Register to             = R4_ARG2;  // destination array address
1980     Register key            = R5_ARG3;  // round key array
1981 
1982     Register keylen         = R8;
1983     Register temp           = R9;
1984     Register keypos         = R10;
1985     Register hex            = R11;
1986     Register fifteen        = R12;
1987 
1988     VectorRegister vRet     = VR0;
1989 
1990     VectorRegister vKey1    = VR1;
1991     VectorRegister vKey2    = VR2;
1992     VectorRegister vKey3    = VR3;
1993     VectorRegister vKey4    = VR4;
1994 
1995     VectorRegister fromPerm = VR5;
1996     VectorRegister keyPerm  = VR6;
1997     VectorRegister toPerm   = VR7;
1998     VectorRegister fSplt    = VR8;
1999 
2000     VectorRegister vTmp1    = VR9;
2001     VectorRegister vTmp2    = VR10;
2002     VectorRegister vTmp3    = VR11;
2003     VectorRegister vTmp4    = VR12;
2004 
2005     VectorRegister vLow     = VR13;
2006     VectorRegister vHigh    = VR14;
2007 
2008     __ li              (hex, 16);
2009     __ li              (fifteen, 15);
2010     __ vspltisb        (fSplt, 0x0f);
2011 
2012     // load unaligned from[0-15] to vsRet
2013     __ lvx             (vRet, from);
2014     __ lvx             (vTmp1, fifteen, from);
2015     __ lvsl            (fromPerm, from);
2016     __ vxor            (fromPerm, fromPerm, fSplt);
2017     __ vperm           (vRet, vRet, vTmp1, fromPerm);
2018 
2019     // load keylen (44 or 52 or 60)
2020     __ lwz             (keylen, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT), key);
2021 
2022     // to load keys
2023     __ lvsr            (keyPerm, key);
2024     __ vxor            (vTmp2, vTmp2, vTmp2);
2025     __ vspltisb        (vTmp2, -16);
2026     __ vrld            (keyPerm, keyPerm, vTmp2);
2027     __ vrld            (keyPerm, keyPerm, vTmp2);
2028     __ vsldoi          (keyPerm, keyPerm, keyPerm, -8);
2029 
2030     // load the 1st round key to vKey1
2031     __ li              (keypos, 0);
2032     __ lvx             (vKey1, keypos, key);
2033     __ addi            (keypos, keypos, 16);
2034     __ lvx             (vTmp1, keypos, key);
2035     __ vperm           (vKey1, vTmp1, vKey1, keyPerm);
2036 
2037     // 1st round
2038     __ vxor (vRet, vRet, vKey1);
2039 
2040     // load the 2nd round key to vKey1
2041     __ addi            (keypos, keypos, 16);
2042     __ lvx             (vTmp2, keypos, key);
2043     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2044 
2045     // load the 3rd round key to vKey2
2046     __ addi            (keypos, keypos, 16);
2047     __ lvx             (vTmp1, keypos, key);
2048     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2049 
2050     // load the 4th round key to vKey3
2051     __ addi            (keypos, keypos, 16);
2052     __ lvx             (vTmp2, keypos, key);
2053     __ vperm           (vKey3, vTmp2, vTmp1, keyPerm);
2054 
2055     // load the 5th round key to vKey4
2056     __ addi            (keypos, keypos, 16);
2057     __ lvx             (vTmp1, keypos, key);
2058     __ vperm           (vKey4, vTmp1, vTmp2, keyPerm);
2059 
2060     // 2nd - 5th rounds
2061     __ vcipher (vRet, vRet, vKey1);
2062     __ vcipher (vRet, vRet, vKey2);
2063     __ vcipher (vRet, vRet, vKey3);
2064     __ vcipher (vRet, vRet, vKey4);
2065 
2066     // load the 6th round key to vKey1
2067     __ addi            (keypos, keypos, 16);
2068     __ lvx             (vTmp2, keypos, key);
2069     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2070 
2071     // load the 7th round key to vKey2
2072     __ addi            (keypos, keypos, 16);
2073     __ lvx             (vTmp1, keypos, key);
2074     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2075 
2076     // load the 8th round key to vKey3
2077     __ addi            (keypos, keypos, 16);
2078     __ lvx             (vTmp2, keypos, key);
2079     __ vperm           (vKey3, vTmp2, vTmp1, keyPerm);
2080 
2081     // load the 9th round key to vKey4
2082     __ addi            (keypos, keypos, 16);
2083     __ lvx             (vTmp1, keypos, key);
2084     __ vperm           (vKey4, vTmp1, vTmp2, keyPerm);
2085 
2086     // 6th - 9th rounds
2087     __ vcipher (vRet, vRet, vKey1);
2088     __ vcipher (vRet, vRet, vKey2);
2089     __ vcipher (vRet, vRet, vKey3);
2090     __ vcipher (vRet, vRet, vKey4);
2091 
2092     // load the 10th round key to vKey1
2093     __ addi            (keypos, keypos, 16);
2094     __ lvx             (vTmp2, keypos, key);
2095     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2096 
2097     // load the 11th round key to vKey2
2098     __ addi            (keypos, keypos, 16);
2099     __ lvx             (vTmp1, keypos, key);
2100     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2101 
2102     // if all round keys are loaded, skip next 4 rounds
2103     __ cmpwi           (CCR0, keylen, 44);
2104     __ beq             (CCR0, L_doLast);
2105 
2106     // 10th - 11th rounds
2107     __ vcipher (vRet, vRet, vKey1);
2108     __ vcipher (vRet, vRet, vKey2);
2109 
2110     // load the 12th round key to vKey1
2111     __ addi            (keypos, keypos, 16);
2112     __ lvx             (vTmp2, keypos, key);
2113     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2114 
2115     // load the 13th round key to vKey2
2116     __ addi            (keypos, keypos, 16);
2117     __ lvx             (vTmp1, keypos, key);
2118     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2119 
2120     // if all round keys are loaded, skip next 2 rounds
2121     __ cmpwi           (CCR0, keylen, 52);
2122     __ beq             (CCR0, L_doLast);
2123 
2124     // 12th - 13th rounds
2125     __ vcipher (vRet, vRet, vKey1);
2126     __ vcipher (vRet, vRet, vKey2);
2127 
2128     // load the 14th round key to vKey1
2129     __ addi            (keypos, keypos, 16);
2130     __ lvx             (vTmp2, keypos, key);
2131     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2132 
2133     // load the 15th round key to vKey2
2134     __ addi            (keypos, keypos, 16);
2135     __ lvx             (vTmp1, keypos, key);
2136     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2137 
2138     __ bind(L_doLast);
2139 
2140     // last two rounds
2141     __ vcipher (vRet, vRet, vKey1);
2142     __ vcipherlast (vRet, vRet, vKey2);
2143 
2144     __ neg             (temp, to);
2145     __ lvsr            (toPerm, temp);
2146     __ vspltisb        (vTmp2, -1);
2147     __ vxor            (vTmp1, vTmp1, vTmp1);
2148     __ vperm           (vTmp2, vTmp2, vTmp1, toPerm);
2149     __ vxor            (toPerm, toPerm, fSplt);
2150     __ lvx             (vTmp1, to);
2151     __ vperm           (vRet, vRet, vRet, toPerm);
2152     __ vsel            (vTmp1, vTmp1, vRet, vTmp2);
2153     __ lvx             (vTmp4, fifteen, to);
2154     __ stvx            (vTmp1, to);
2155     __ vsel            (vRet, vRet, vTmp4, vTmp2);
2156     __ stvx            (vRet, fifteen, to);
2157 
2158     __ blr();
2159      return start;
2160   }
2161 
2162   // Arguments for generated stub (little endian only):
2163   //   R3_ARG1   - source byte array address
2164   //   R4_ARG2   - destination byte array address
2165   //   R5_ARG3   - K (key) in little endian int array
2166   address generate_aescrypt_decryptBlock() {
2167     assert(UseAES, "need AES instructions and misaligned SSE support");
2168     StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
2169 
2170     address start = __ function_entry();
2171 
2172     Label L_doLast;
2173     Label L_do44;
2174     Label L_do52;
2175     Label L_do60;
2176 
2177     Register from           = R3_ARG1;  // source array address
2178     Register to             = R4_ARG2;  // destination array address
2179     Register key            = R5_ARG3;  // round key array
2180 
2181     Register keylen         = R8;
2182     Register temp           = R9;
2183     Register keypos         = R10;
2184     Register hex            = R11;
2185     Register fifteen        = R12;
2186 
2187     VectorRegister vRet     = VR0;
2188 
2189     VectorRegister vKey1    = VR1;
2190     VectorRegister vKey2    = VR2;
2191     VectorRegister vKey3    = VR3;
2192     VectorRegister vKey4    = VR4;
2193     VectorRegister vKey5    = VR5;
2194 
2195     VectorRegister fromPerm = VR6;
2196     VectorRegister keyPerm  = VR7;
2197     VectorRegister toPerm   = VR8;
2198     VectorRegister fSplt    = VR9;
2199 
2200     VectorRegister vTmp1    = VR10;
2201     VectorRegister vTmp2    = VR11;
2202     VectorRegister vTmp3    = VR12;
2203     VectorRegister vTmp4    = VR13;
2204 
2205     VectorRegister vLow     = VR14;
2206     VectorRegister vHigh    = VR15;
2207 
2208     __ li              (hex, 16);
2209     __ li              (fifteen, 15);
2210     __ vspltisb        (fSplt, 0x0f);
2211 
2212     // load unaligned from[0-15] to vsRet
2213     __ lvx             (vRet, from);
2214     __ lvx             (vTmp1, fifteen, from);
2215     __ lvsl            (fromPerm, from);
2216     __ vxor            (fromPerm, fromPerm, fSplt);
2217     __ vperm           (vRet, vRet, vTmp1, fromPerm); // align [and byte swap in LE]
2218 
2219     // load keylen (44 or 52 or 60)
2220     __ lwz             (keylen, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT), key);
2221 
2222     // to load keys
2223     __ lvsr            (keyPerm, key);
2224     __ vxor            (vTmp2, vTmp2, vTmp2);
2225     __ vspltisb        (vTmp2, -16);
2226     __ vrld            (keyPerm, keyPerm, vTmp2);
2227     __ vrld            (keyPerm, keyPerm, vTmp2);
2228     __ vsldoi          (keyPerm, keyPerm, keyPerm, -8);
2229 
2230     __ cmpwi           (CCR0, keylen, 44);
2231     __ beq             (CCR0, L_do44);
2232 
2233     __ cmpwi           (CCR0, keylen, 52);
2234     __ beq             (CCR0, L_do52);
2235 
2236     // load the 15th round key to vKey11
2237     __ li              (keypos, 240);
2238     __ lvx             (vTmp1, keypos, key);
2239     __ addi            (keypos, keypos, -16);
2240     __ lvx             (vTmp2, keypos, key);
2241     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2242 
2243     // load the 14th round key to vKey10
2244     __ addi            (keypos, keypos, -16);
2245     __ lvx             (vTmp1, keypos, key);
2246     __ vperm           (vKey2, vTmp2, vTmp1, keyPerm);
2247 
2248     // load the 13th round key to vKey10
2249     __ addi            (keypos, keypos, -16);
2250     __ lvx             (vTmp2, keypos, key);
2251     __ vperm           (vKey3, vTmp1, vTmp2, keyPerm);
2252 
2253     // load the 12th round key to vKey10
2254     __ addi            (keypos, keypos, -16);
2255     __ lvx             (vTmp1, keypos, key);
2256     __ vperm           (vKey4, vTmp2, vTmp1, keyPerm);
2257 
2258     // load the 11th round key to vKey10
2259     __ addi            (keypos, keypos, -16);
2260     __ lvx             (vTmp2, keypos, key);
2261     __ vperm           (vKey5, vTmp1, vTmp2, keyPerm);
2262 
2263     // 1st - 5th rounds
2264     __ vxor            (vRet, vRet, vKey1);
2265     __ vncipher        (vRet, vRet, vKey2);
2266     __ vncipher        (vRet, vRet, vKey3);
2267     __ vncipher        (vRet, vRet, vKey4);
2268     __ vncipher        (vRet, vRet, vKey5);
2269 
2270     __ b               (L_doLast);
2271 
2272     __ bind            (L_do52);
2273 
2274     // load the 13th round key to vKey11
2275     __ li              (keypos, 208);
2276     __ lvx             (vTmp1, keypos, key);
2277     __ addi            (keypos, keypos, -16);
2278     __ lvx             (vTmp2, keypos, key);
2279     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2280 
2281     // load the 12th round key to vKey10
2282     __ addi            (keypos, keypos, -16);
2283     __ lvx             (vTmp1, keypos, key);
2284     __ vperm           (vKey2, vTmp2, vTmp1, keyPerm);
2285 
2286     // load the 11th round key to vKey10
2287     __ addi            (keypos, keypos, -16);
2288     __ lvx             (vTmp2, keypos, key);
2289     __ vperm           (vKey3, vTmp1, vTmp2, keyPerm);
2290 
2291     // 1st - 3rd rounds
2292     __ vxor            (vRet, vRet, vKey1);
2293     __ vncipher        (vRet, vRet, vKey2);
2294     __ vncipher        (vRet, vRet, vKey3);
2295 
2296     __ b               (L_doLast);
2297 
2298     __ bind            (L_do44);
2299 
2300     // load the 11th round key to vKey11
2301     __ li              (keypos, 176);
2302     __ lvx             (vTmp1, keypos, key);
2303     __ addi            (keypos, keypos, -16);
2304     __ lvx             (vTmp2, keypos, key);
2305     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2306 
2307     // 1st round
2308     __ vxor            (vRet, vRet, vKey1);
2309 
2310     __ bind            (L_doLast);
2311 
2312     // load the 10th round key to vKey10
2313     __ addi            (keypos, keypos, -16);
2314     __ lvx             (vTmp1, keypos, key);
2315     __ vperm           (vKey1, vTmp2, vTmp1, keyPerm);
2316 
2317     // load the 9th round key to vKey10
2318     __ addi            (keypos, keypos, -16);
2319     __ lvx             (vTmp2, keypos, key);
2320     __ vperm           (vKey2, vTmp1, vTmp2, keyPerm);
2321 
2322     // load the 8th round key to vKey10
2323     __ addi            (keypos, keypos, -16);
2324     __ lvx             (vTmp1, keypos, key);
2325     __ vperm           (vKey3, vTmp2, vTmp1, keyPerm);
2326 
2327     // load the 7th round key to vKey10
2328     __ addi            (keypos, keypos, -16);
2329     __ lvx             (vTmp2, keypos, key);
2330     __ vperm           (vKey4, vTmp1, vTmp2, keyPerm);
2331 
2332     // load the 6th round key to vKey10
2333     __ addi            (keypos, keypos, -16);
2334     __ lvx             (vTmp1, keypos, key);
2335     __ vperm           (vKey5, vTmp2, vTmp1, keyPerm);
2336 
2337     // last 10th - 6th rounds
2338     __ vncipher        (vRet, vRet, vKey1);
2339     __ vncipher        (vRet, vRet, vKey2);
2340     __ vncipher        (vRet, vRet, vKey3);
2341     __ vncipher        (vRet, vRet, vKey4);
2342     __ vncipher        (vRet, vRet, vKey5);
2343 
2344     // load the 5th round key to vKey10
2345     __ addi            (keypos, keypos, -16);
2346     __ lvx             (vTmp2, keypos, key);
2347     __ vperm           (vKey1, vTmp1, vTmp2, keyPerm);
2348 
2349     // load the 4th round key to vKey10
2350     __ addi            (keypos, keypos, -16);
2351     __ lvx             (vTmp1, keypos, key);
2352     __ vperm           (vKey2, vTmp2, vTmp1, keyPerm);
2353 
2354     // load the 3rd round key to vKey10
2355     __ addi            (keypos, keypos, -16);
2356     __ lvx             (vTmp2, keypos, key);
2357     __ vperm           (vKey3, vTmp1, vTmp2, keyPerm);
2358 
2359     // load the 2nd round key to vKey10
2360     __ addi            (keypos, keypos, -16);
2361     __ lvx             (vTmp1, keypos, key);
2362     __ vperm           (vKey4, vTmp2, vTmp1, keyPerm);
2363 
2364     // load the 1st round key to vKey10
2365     __ addi            (keypos, keypos, -16);
2366     __ lvx             (vTmp2, keypos, key);
2367     __ vperm           (vKey5, vTmp1, vTmp2, keyPerm);
2368 
2369     // last 5th - 1th rounds
2370     __ vncipher        (vRet, vRet, vKey1);
2371     __ vncipher        (vRet, vRet, vKey2);
2372     __ vncipher        (vRet, vRet, vKey3);
2373     __ vncipher        (vRet, vRet, vKey4);
2374     __ vncipherlast    (vRet, vRet, vKey5);
2375 
2376     __ neg             (temp, to);
2377     __ lvsr            (toPerm, temp);
2378     __ vspltisb        (vTmp2, -1);
2379     __ vxor            (vTmp1, vTmp1, vTmp1);
2380     __ vperm           (vTmp2, vTmp2, vTmp1, toPerm);
2381     __ vxor            (toPerm, toPerm, fSplt);
2382     __ lvx             (vTmp1, to);
2383     __ vperm           (vRet, vRet, vRet, toPerm);
2384     __ vsel            (vTmp1, vTmp1, vRet, vTmp2);
2385     __ lvx             (vTmp4, fifteen, to);
2386     __ stvx            (vTmp1, to);
2387     __ vsel            (vRet, vRet, vTmp4, vTmp2);
2388     __ stvx            (vRet, fifteen, to);
2389 
2390     __ blr();
2391      return start;
2392   }
2393 
2394   void generate_arraycopy_stubs() {
2395     // Note: the disjoint stubs must be generated first, some of
2396     // the conjoint stubs use them.
2397 
2398     // non-aligned disjoint versions
2399     StubRoutines::_jbyte_disjoint_arraycopy       = generate_disjoint_byte_copy(false, "jbyte_disjoint_arraycopy");
2400     StubRoutines::_jshort_disjoint_arraycopy      = generate_disjoint_short_copy(false, "jshort_disjoint_arraycopy");
2401     StubRoutines::_jint_disjoint_arraycopy        = generate_disjoint_int_copy(false, "jint_disjoint_arraycopy");
2402     StubRoutines::_jlong_disjoint_arraycopy       = generate_disjoint_long_copy(false, "jlong_disjoint_arraycopy");
2403     StubRoutines::_oop_disjoint_arraycopy         = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy", false);
2404     StubRoutines::_oop_disjoint_arraycopy_uninit  = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy_uninit", true);
2405 
2406     // aligned disjoint versions
2407     StubRoutines::_arrayof_jbyte_disjoint_arraycopy      = generate_disjoint_byte_copy(true, "arrayof_jbyte_disjoint_arraycopy");
2408     StubRoutines::_arrayof_jshort_disjoint_arraycopy     = generate_disjoint_short_copy(true, "arrayof_jshort_disjoint_arraycopy");
2409     StubRoutines::_arrayof_jint_disjoint_arraycopy       = generate_disjoint_int_copy(true, "arrayof_jint_disjoint_arraycopy");
2410     StubRoutines::_arrayof_jlong_disjoint_arraycopy      = generate_disjoint_long_copy(true, "arrayof_jlong_disjoint_arraycopy");
2411     StubRoutines::_arrayof_oop_disjoint_arraycopy        = generate_disjoint_oop_copy(true, "arrayof_oop_disjoint_arraycopy", false);
2412     StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(true, "oop_disjoint_arraycopy_uninit", true);
2413 
2414     // non-aligned conjoint versions
2415     StubRoutines::_jbyte_arraycopy      = generate_conjoint_byte_copy(false, "jbyte_arraycopy");
2416     StubRoutines::_jshort_arraycopy     = generate_conjoint_short_copy(false, "jshort_arraycopy");
2417     StubRoutines::_jint_arraycopy       = generate_conjoint_int_copy(false, "jint_arraycopy");
2418     StubRoutines::_jlong_arraycopy      = generate_conjoint_long_copy(false, "jlong_arraycopy");
2419     StubRoutines::_oop_arraycopy        = generate_conjoint_oop_copy(false, "oop_arraycopy", false);
2420     StubRoutines::_oop_arraycopy_uninit = generate_conjoint_oop_copy(false, "oop_arraycopy_uninit", true);
2421 
2422     // aligned conjoint versions
2423     StubRoutines::_arrayof_jbyte_arraycopy      = generate_conjoint_byte_copy(true, "arrayof_jbyte_arraycopy");
2424     StubRoutines::_arrayof_jshort_arraycopy     = generate_conjoint_short_copy(true, "arrayof_jshort_arraycopy");
2425     StubRoutines::_arrayof_jint_arraycopy       = generate_conjoint_int_copy(true, "arrayof_jint_arraycopy");
2426     StubRoutines::_arrayof_jlong_arraycopy      = generate_conjoint_long_copy(true, "arrayof_jlong_arraycopy");
2427     StubRoutines::_arrayof_oop_arraycopy        = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", false);
2428     StubRoutines::_arrayof_oop_arraycopy_uninit = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", true);
2429 
2430     // fill routines
2431     StubRoutines::_jbyte_fill          = generate_fill(T_BYTE,  false, "jbyte_fill");
2432     StubRoutines::_jshort_fill         = generate_fill(T_SHORT, false, "jshort_fill");
2433     StubRoutines::_jint_fill           = generate_fill(T_INT,   false, "jint_fill");
2434     StubRoutines::_arrayof_jbyte_fill  = generate_fill(T_BYTE,  true, "arrayof_jbyte_fill");
2435     StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
2436     StubRoutines::_arrayof_jint_fill   = generate_fill(T_INT,   true, "arrayof_jint_fill");
2437   }
2438 
2439   // Safefetch stubs.
2440   void generate_safefetch(const char* name, int size, address* entry, address* fault_pc, address* continuation_pc) {
2441     // safefetch signatures:
2442     //   int      SafeFetch32(int*      adr, int      errValue);
2443     //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
2444     //
2445     // arguments:
2446     //   R3_ARG1 = adr
2447     //   R4_ARG2 = errValue
2448     //
2449     // result:
2450     //   R3_RET  = *adr or errValue
2451 
2452     StubCodeMark mark(this, "StubRoutines", name);
2453 
2454     // Entry point, pc or function descriptor.
2455     *entry = __ function_entry();
2456 
2457     // Load *adr into R4_ARG2, may fault.
2458     *fault_pc = __ pc();
2459     switch (size) {
2460       case 4:
2461         // int32_t, signed extended
2462         __ lwa(R4_ARG2, 0, R3_ARG1);
2463         break;
2464       case 8:
2465         // int64_t
2466         __ ld(R4_ARG2, 0, R3_ARG1);
2467         break;
2468       default:
2469         ShouldNotReachHere();
2470     }
2471 
2472     // return errValue or *adr
2473     *continuation_pc = __ pc();
2474     __ mr(R3_RET, R4_ARG2);
2475     __ blr();
2476   }
2477 
2478   // Initialization
2479   void generate_initial() {
2480     // Generates all stubs and initializes the entry points
2481 
2482     // Entry points that exist in all platforms.
2483     // Note: This is code that could be shared among different platforms - however the
2484     // benefit seems to be smaller than the disadvantage of having a
2485     // much more complicated generator structure. See also comment in
2486     // stubRoutines.hpp.
2487 
2488     StubRoutines::_forward_exception_entry          = generate_forward_exception();
2489     StubRoutines::_call_stub_entry                  = generate_call_stub(StubRoutines::_call_stub_return_address);
2490     StubRoutines::_catch_exception_entry            = generate_catch_exception();
2491 
2492     // Build this early so it's available for the interpreter.
2493     StubRoutines::_throw_StackOverflowError_entry   =
2494       generate_throw_exception("StackOverflowError throw_exception",
2495                                CAST_FROM_FN_PTR(address, SharedRuntime::throw_StackOverflowError), false);
2496   }
2497 
2498   void generate_all() {
2499     // Generates all stubs and initializes the entry points
2500 
2501     // These entry points require SharedInfo::stack0 to be set up in
2502     // non-core builds
2503     StubRoutines::_throw_AbstractMethodError_entry         = generate_throw_exception("AbstractMethodError throw_exception",          CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError),  false);
2504     // Handle IncompatibleClassChangeError in itable stubs.
2505     StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError),  false);
2506     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);
2507 
2508     StubRoutines::_handler_for_unsafe_access_entry         = generate_handler_for_unsafe_access();
2509 
2510     // support for verify_oop (must happen after universe_init)
2511     StubRoutines::_verify_oop_subroutine_entry             = generate_verify_oop();
2512 
2513     // arraycopy stubs used by compilers
2514     generate_arraycopy_stubs();
2515 
2516     // Safefetch stubs.
2517     generate_safefetch("SafeFetch32", sizeof(int),     &StubRoutines::_safefetch32_entry,
2518                                                        &StubRoutines::_safefetch32_fault_pc,
2519                                                        &StubRoutines::_safefetch32_continuation_pc);
2520     generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
2521                                                        &StubRoutines::_safefetchN_fault_pc,
2522                                                        &StubRoutines::_safefetchN_continuation_pc);
2523 
2524     if (UseAESIntrinsics) {
2525       StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
2526       StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
2527     }
2528 
2529     if (UseMontgomeryMultiplyIntrinsic) {
2530       StubRoutines::_montgomeryMultiply
2531         = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
2532     }
2533     if (UseMontgomerySquareIntrinsic) {
2534       StubRoutines::_montgomerySquare
2535         = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
2536     }
2537   }
2538 
2539  public:
2540   StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
2541     // replace the standard masm with a special one:
2542     _masm = new MacroAssembler(code);
2543     if (all) {
2544       generate_all();
2545     } else {
2546       generate_initial();
2547     }
2548   }
2549 };
2550 
2551 void StubGenerator_generate(CodeBuffer* code, bool all) {
2552   StubGenerator g(code, all);
2553 }