1 /*
   2  * Copyright (c) 2014, 2016, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2015, 2016 SAP SE. 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/bytecodeHistogram.hpp"
  29 #include "interpreter/interpreter.hpp"
  30 #include "interpreter/interpreterRuntime.hpp"
  31 #include "interpreter/interp_masm.hpp"
  32 #include "interpreter/templateInterpreterGenerator.hpp"
  33 #include "interpreter/templateTable.hpp"
  34 #include "oops/arrayOop.hpp"
  35 #include "oops/methodData.hpp"
  36 #include "oops/method.hpp"
  37 #include "oops/oop.inline.hpp"
  38 #include "prims/jvmtiExport.hpp"
  39 #include "prims/jvmtiThreadState.hpp"
  40 #include "runtime/arguments.hpp"
  41 #include "runtime/deoptimization.hpp"
  42 #include "runtime/frame.inline.hpp"
  43 #include "runtime/sharedRuntime.hpp"
  44 #include "runtime/stubRoutines.hpp"
  45 #include "runtime/synchronizer.hpp"
  46 #include "runtime/timer.hpp"
  47 #include "runtime/vframeArray.hpp"
  48 #include "utilities/debug.hpp"
  49 #include "utilities/macros.hpp"
  50 
  51 #undef __
  52 #define __ _masm->
  53 
  54 // Size of interpreter code.  Increase if too small.  Interpreter will
  55 // fail with a guarantee ("not enough space for interpreter generation");
  56 // if too small.
  57 // Run with +PrintInterpreter to get the VM to print out the size.
  58 // Max size with JVMTI
  59 int TemplateInterpreter::InterpreterCodeSize = 230*K;
  60 
  61 #ifdef PRODUCT
  62 #define BLOCK_COMMENT(str) /* nothing */
  63 #else
  64 #define BLOCK_COMMENT(str) __ block_comment(str)
  65 #endif
  66 
  67 #define BIND(label)        __ bind(label); BLOCK_COMMENT(#label ":")
  68 
  69 //-----------------------------------------------------------------------------
  70 
  71 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
  72   // Slow_signature handler that respects the PPC C calling conventions.
  73   //
  74   // We get called by the native entry code with our output register
  75   // area == 8. First we call InterpreterRuntime::get_result_handler
  76   // to copy the pointer to the signature string temporarily to the
  77   // first C-argument and to return the result_handler in
  78   // R3_RET. Since native_entry will copy the jni-pointer to the
  79   // first C-argument slot later on, it is OK to occupy this slot
  80   // temporarilly. Then we copy the argument list on the java
  81   // expression stack into native varargs format on the native stack
  82   // and load arguments into argument registers. Integer arguments in
  83   // the varargs vector will be sign-extended to 8 bytes.
  84   //
  85   // On entry:
  86   //   R3_ARG1        - intptr_t*     Address of java argument list in memory.
  87   //   R15_prev_state - BytecodeInterpreter* Address of interpreter state for
  88   //     this method
  89   //   R19_method
  90   //
  91   // On exit (just before return instruction):
  92   //   R3_RET            - contains the address of the result_handler.
  93   //   R4_ARG2           - is not updated for static methods and contains "this" otherwise.
  94   //   R5_ARG3-R10_ARG8: - When the (i-2)th Java argument is not of type float or double,
  95   //                       ARGi contains this argument. Otherwise, ARGi is not updated.
  96   //   F1_ARG1-F13_ARG13 - contain the first 13 arguments of type float or double.
  97 
  98   const int LogSizeOfTwoInstructions = 3;
  99 
 100   // FIXME: use Argument:: GL: Argument names different numbers!
 101   const int max_fp_register_arguments  = 13;
 102   const int max_int_register_arguments = 6;  // first 2 are reserved
 103 
 104   const Register arg_java       = R21_tmp1;
 105   const Register arg_c          = R22_tmp2;
 106   const Register signature      = R23_tmp3;  // is string
 107   const Register sig_byte       = R24_tmp4;
 108   const Register fpcnt          = R25_tmp5;
 109   const Register argcnt         = R26_tmp6;
 110   const Register intSlot        = R27_tmp7;
 111   const Register target_sp      = R28_tmp8;
 112   const FloatRegister floatSlot = F0;
 113 
 114   address entry = __ function_entry();
 115 
 116   __ save_LR_CR(R0);
 117   __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
 118   // We use target_sp for storing arguments in the C frame.
 119   __ mr(target_sp, R1_SP);
 120   __ push_frame_reg_args_nonvolatiles(0, R11_scratch1);
 121 
 122   __ mr(arg_java, R3_ARG1);
 123 
 124   __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_signature), R16_thread, R19_method);
 125 
 126   // Signature is in R3_RET. Signature is callee saved.
 127   __ mr(signature, R3_RET);
 128 
 129   // Get the result handler.
 130   __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_result_handler), R16_thread, R19_method);
 131 
 132   {
 133     Label L;
 134     // test if static
 135     // _access_flags._flags must be at offset 0.
 136     // TODO PPC port: requires change in shared code.
 137     //assert(in_bytes(AccessFlags::flags_offset()) == 0,
 138     //       "MethodDesc._access_flags == MethodDesc._access_flags._flags");
 139     // _access_flags must be a 32 bit value.
 140     assert(sizeof(AccessFlags) == 4, "wrong size");
 141     __ lwa(R11_scratch1/*access_flags*/, method_(access_flags));
 142     // testbit with condition register.
 143     __ testbitdi(CCR0, R0, R11_scratch1/*access_flags*/, JVM_ACC_STATIC_BIT);
 144     __ btrue(CCR0, L);
 145     // For non-static functions, pass "this" in R4_ARG2 and copy it
 146     // to 2nd C-arg slot.
 147     // We need to box the Java object here, so we use arg_java
 148     // (address of current Java stack slot) as argument and don't
 149     // dereference it as in case of ints, floats, etc.
 150     __ mr(R4_ARG2, arg_java);
 151     __ addi(arg_java, arg_java, -BytesPerWord);
 152     __ std(R4_ARG2, _abi(carg_2), target_sp);
 153     __ bind(L);
 154   }
 155 
 156   // Will be incremented directly after loop_start. argcnt=0
 157   // corresponds to 3rd C argument.
 158   __ li(argcnt, -1);
 159   // arg_c points to 3rd C argument
 160   __ addi(arg_c, target_sp, _abi(carg_3));
 161   // no floating-point args parsed so far
 162   __ li(fpcnt, 0);
 163 
 164   Label move_intSlot_to_ARG, move_floatSlot_to_FARG;
 165   Label loop_start, loop_end;
 166   Label do_int, do_long, do_float, do_double, do_dontreachhere, do_object, do_array, do_boxed;
 167 
 168   // signature points to '(' at entry
 169 #ifdef ASSERT
 170   __ lbz(sig_byte, 0, signature);
 171   __ cmplwi(CCR0, sig_byte, '(');
 172   __ bne(CCR0, do_dontreachhere);
 173 #endif
 174 
 175   __ bind(loop_start);
 176 
 177   __ addi(argcnt, argcnt, 1);
 178   __ lbzu(sig_byte, 1, signature);
 179 
 180   __ cmplwi(CCR0, sig_byte, ')'); // end of signature
 181   __ beq(CCR0, loop_end);
 182 
 183   __ cmplwi(CCR0, sig_byte, 'B'); // byte
 184   __ beq(CCR0, do_int);
 185 
 186   __ cmplwi(CCR0, sig_byte, 'C'); // char
 187   __ beq(CCR0, do_int);
 188 
 189   __ cmplwi(CCR0, sig_byte, 'D'); // double
 190   __ beq(CCR0, do_double);
 191 
 192   __ cmplwi(CCR0, sig_byte, 'F'); // float
 193   __ beq(CCR0, do_float);
 194 
 195   __ cmplwi(CCR0, sig_byte, 'I'); // int
 196   __ beq(CCR0, do_int);
 197 
 198   __ cmplwi(CCR0, sig_byte, 'J'); // long
 199   __ beq(CCR0, do_long);
 200 
 201   __ cmplwi(CCR0, sig_byte, 'S'); // short
 202   __ beq(CCR0, do_int);
 203 
 204   __ cmplwi(CCR0, sig_byte, 'Z'); // boolean
 205   __ beq(CCR0, do_int);
 206 
 207   __ cmplwi(CCR0, sig_byte, 'L'); // object
 208   __ beq(CCR0, do_object);
 209 
 210   __ cmplwi(CCR0, sig_byte, '['); // array
 211   __ beq(CCR0, do_array);
 212 
 213   //  __ cmplwi(CCR0, sig_byte, 'V'); // void cannot appear since we do not parse the return type
 214   //  __ beq(CCR0, do_void);
 215 
 216   __ bind(do_dontreachhere);
 217 
 218   __ unimplemented("ShouldNotReachHere in slow_signature_handler", 120);
 219 
 220   __ bind(do_array);
 221 
 222   {
 223     Label start_skip, end_skip;
 224 
 225     __ bind(start_skip);
 226     __ lbzu(sig_byte, 1, signature);
 227     __ cmplwi(CCR0, sig_byte, '[');
 228     __ beq(CCR0, start_skip); // skip further brackets
 229     __ cmplwi(CCR0, sig_byte, '9');
 230     __ bgt(CCR0, end_skip);   // no optional size
 231     __ cmplwi(CCR0, sig_byte, '0');
 232     __ bge(CCR0, start_skip); // skip optional size
 233     __ bind(end_skip);
 234 
 235     __ cmplwi(CCR0, sig_byte, 'L');
 236     __ beq(CCR0, do_object);  // for arrays of objects, the name of the object must be skipped
 237     __ b(do_boxed);          // otherwise, go directly to do_boxed
 238   }
 239 
 240   __ bind(do_object);
 241   {
 242     Label L;
 243     __ bind(L);
 244     __ lbzu(sig_byte, 1, signature);
 245     __ cmplwi(CCR0, sig_byte, ';');
 246     __ bne(CCR0, L);
 247    }
 248   // Need to box the Java object here, so we use arg_java (address of
 249   // current Java stack slot) as argument and don't dereference it as
 250   // in case of ints, floats, etc.
 251   Label do_null;
 252   __ bind(do_boxed);
 253   __ ld(R0,0, arg_java);
 254   __ cmpdi(CCR0, R0, 0);
 255   __ li(intSlot,0);
 256   __ beq(CCR0, do_null);
 257   __ mr(intSlot, arg_java);
 258   __ bind(do_null);
 259   __ std(intSlot, 0, arg_c);
 260   __ addi(arg_java, arg_java, -BytesPerWord);
 261   __ addi(arg_c, arg_c, BytesPerWord);
 262   __ cmplwi(CCR0, argcnt, max_int_register_arguments);
 263   __ blt(CCR0, move_intSlot_to_ARG);
 264   __ b(loop_start);
 265 
 266   __ bind(do_int);
 267   __ lwa(intSlot, 0, arg_java);
 268   __ std(intSlot, 0, arg_c);
 269   __ addi(arg_java, arg_java, -BytesPerWord);
 270   __ addi(arg_c, arg_c, BytesPerWord);
 271   __ cmplwi(CCR0, argcnt, max_int_register_arguments);
 272   __ blt(CCR0, move_intSlot_to_ARG);
 273   __ b(loop_start);
 274 
 275   __ bind(do_long);
 276   __ ld(intSlot, -BytesPerWord, arg_java);
 277   __ std(intSlot, 0, arg_c);
 278   __ addi(arg_java, arg_java, - 2 * BytesPerWord);
 279   __ addi(arg_c, arg_c, BytesPerWord);
 280   __ cmplwi(CCR0, argcnt, max_int_register_arguments);
 281   __ blt(CCR0, move_intSlot_to_ARG);
 282   __ b(loop_start);
 283 
 284   __ bind(do_float);
 285   __ lfs(floatSlot, 0, arg_java);
 286 #if defined(LINUX)
 287   // Linux uses ELF ABI. Both original ELF and ELFv2 ABIs have float
 288   // in the least significant word of an argument slot.
 289 #if defined(VM_LITTLE_ENDIAN)
 290   __ stfs(floatSlot, 0, arg_c);
 291 #else
 292   __ stfs(floatSlot, 4, arg_c);
 293 #endif
 294 #elif defined(AIX)
 295   // Although AIX runs on big endian CPU, float is in most significant
 296   // word of an argument slot.
 297   __ stfs(floatSlot, 0, arg_c);
 298 #else
 299 #error "unknown OS"
 300 #endif
 301   __ addi(arg_java, arg_java, -BytesPerWord);
 302   __ addi(arg_c, arg_c, BytesPerWord);
 303   __ cmplwi(CCR0, fpcnt, max_fp_register_arguments);
 304   __ blt(CCR0, move_floatSlot_to_FARG);
 305   __ b(loop_start);
 306 
 307   __ bind(do_double);
 308   __ lfd(floatSlot, - BytesPerWord, arg_java);
 309   __ stfd(floatSlot, 0, arg_c);
 310   __ addi(arg_java, arg_java, - 2 * BytesPerWord);
 311   __ addi(arg_c, arg_c, BytesPerWord);
 312   __ cmplwi(CCR0, fpcnt, max_fp_register_arguments);
 313   __ blt(CCR0, move_floatSlot_to_FARG);
 314   __ b(loop_start);
 315 
 316   __ bind(loop_end);
 317 
 318   __ pop_frame();
 319   __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
 320   __ restore_LR_CR(R0);
 321 
 322   __ blr();
 323 
 324   Label move_int_arg, move_float_arg;
 325   __ bind(move_int_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions)
 326   __ mr(R5_ARG3, intSlot);  __ b(loop_start);
 327   __ mr(R6_ARG4, intSlot);  __ b(loop_start);
 328   __ mr(R7_ARG5, intSlot);  __ b(loop_start);
 329   __ mr(R8_ARG6, intSlot);  __ b(loop_start);
 330   __ mr(R9_ARG7, intSlot);  __ b(loop_start);
 331   __ mr(R10_ARG8, intSlot); __ b(loop_start);
 332 
 333   __ bind(move_float_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions)
 334   __ fmr(F1_ARG1, floatSlot);   __ b(loop_start);
 335   __ fmr(F2_ARG2, floatSlot);   __ b(loop_start);
 336   __ fmr(F3_ARG3, floatSlot);   __ b(loop_start);
 337   __ fmr(F4_ARG4, floatSlot);   __ b(loop_start);
 338   __ fmr(F5_ARG5, floatSlot);   __ b(loop_start);
 339   __ fmr(F6_ARG6, floatSlot);   __ b(loop_start);
 340   __ fmr(F7_ARG7, floatSlot);   __ b(loop_start);
 341   __ fmr(F8_ARG8, floatSlot);   __ b(loop_start);
 342   __ fmr(F9_ARG9, floatSlot);   __ b(loop_start);
 343   __ fmr(F10_ARG10, floatSlot); __ b(loop_start);
 344   __ fmr(F11_ARG11, floatSlot); __ b(loop_start);
 345   __ fmr(F12_ARG12, floatSlot); __ b(loop_start);
 346   __ fmr(F13_ARG13, floatSlot); __ b(loop_start);
 347 
 348   __ bind(move_intSlot_to_ARG);
 349   __ sldi(R0, argcnt, LogSizeOfTwoInstructions);
 350   __ load_const(R11_scratch1, move_int_arg); // Label must be bound here.
 351   __ add(R11_scratch1, R0, R11_scratch1);
 352   __ mtctr(R11_scratch1/*branch_target*/);
 353   __ bctr();
 354   __ bind(move_floatSlot_to_FARG);
 355   __ sldi(R0, fpcnt, LogSizeOfTwoInstructions);
 356   __ addi(fpcnt, fpcnt, 1);
 357   __ load_const(R11_scratch1, move_float_arg); // Label must be bound here.
 358   __ add(R11_scratch1, R0, R11_scratch1);
 359   __ mtctr(R11_scratch1/*branch_target*/);
 360   __ bctr();
 361 
 362   return entry;
 363 }
 364 
 365 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 366   //
 367   // Registers alive
 368   //   R3_RET
 369   //   LR
 370   //
 371   // Registers updated
 372   //   R3_RET
 373   //
 374 
 375   Label done;
 376   address entry = __ pc();
 377 
 378   switch (type) {
 379   case T_BOOLEAN:
 380     // convert !=0 to 1
 381     __ neg(R0, R3_RET);
 382     __ orr(R0, R3_RET, R0);
 383     __ srwi(R3_RET, R0, 31);
 384     break;
 385   case T_BYTE:
 386      // sign extend 8 bits
 387      __ extsb(R3_RET, R3_RET);
 388      break;
 389   case T_CHAR:
 390      // zero extend 16 bits
 391      __ clrldi(R3_RET, R3_RET, 48);
 392      break;
 393   case T_SHORT:
 394      // sign extend 16 bits
 395      __ extsh(R3_RET, R3_RET);
 396      break;
 397   case T_INT:
 398      // sign extend 32 bits
 399      __ extsw(R3_RET, R3_RET);
 400      break;
 401   case T_LONG:
 402      break;
 403   case T_OBJECT:
 404     // unbox result if not null
 405     __ cmpdi(CCR0, R3_RET, 0);
 406     __ beq(CCR0, done);
 407     __ ld(R3_RET, 0, R3_RET);
 408     __ verify_oop(R3_RET);
 409     break;
 410   case T_FLOAT:
 411      break;
 412   case T_DOUBLE:
 413      break;
 414   case T_VOID:
 415      break;
 416   default: ShouldNotReachHere();
 417   }
 418 
 419   BIND(done);
 420   __ blr();
 421 
 422   return entry;
 423 }
 424 
 425 // Abstract method entry.
 426 //
 427 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
 428   address entry = __ pc();
 429 
 430   //
 431   // Registers alive
 432   //   R16_thread     - JavaThread*
 433   //   R19_method     - callee's method (method to be invoked)
 434   //   R1_SP          - SP prepared such that caller's outgoing args are near top
 435   //   LR             - return address to caller
 436   //
 437   // Stack layout at this point:
 438   //
 439   //   0       [TOP_IJAVA_FRAME_ABI]         <-- R1_SP
 440   //           alignment (optional)
 441   //           [outgoing Java arguments]
 442   //           ...
 443   //   PARENT  [PARENT_IJAVA_FRAME_ABI]
 444   //            ...
 445   //
 446 
 447   // Can't use call_VM here because we have not set up a new
 448   // interpreter state. Make the call to the vm and make it look like
 449   // our caller set up the JavaFrameAnchor.
 450   __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
 451 
 452   // Push a new C frame and save LR.
 453   __ save_LR_CR(R0);
 454   __ push_frame_reg_args(0, R11_scratch1);
 455 
 456   // This is not a leaf but we have a JavaFrameAnchor now and we will
 457   // check (create) exceptions afterward so this is ok.
 458   __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError),
 459                   R16_thread);
 460 
 461   // Pop the C frame and restore LR.
 462   __ pop_frame();
 463   __ restore_LR_CR(R0);
 464 
 465   // Reset JavaFrameAnchor from call_VM_leaf above.
 466   __ reset_last_Java_frame();
 467 
 468   // We don't know our caller, so jump to the general forward exception stub,
 469   // which will also pop our full frame off. Satisfy the interface of
 470   // SharedRuntime::generate_forward_exception()
 471   __ load_const_optimized(R11_scratch1, StubRoutines::forward_exception_entry(), R0);
 472   __ mtctr(R11_scratch1);
 473   __ bctr();
 474 
 475   return entry;
 476 }
 477 
 478 // Interpreter intrinsic for WeakReference.get().
 479 // 1. Don't push a full blown frame and go on dispatching, but fetch the value
 480 //    into R8 and return quickly
 481 // 2. If G1 is active we *must* execute this intrinsic for corrrectness:
 482 //    It contains a GC barrier which puts the reference into the satb buffer
 483 //    to indicate that someone holds a strong reference to the object the
 484 //    weak ref points to!
 485 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
 486   // Code: _aload_0, _getfield, _areturn
 487   // parameter size = 1
 488   //
 489   // The code that gets generated by this routine is split into 2 parts:
 490   //    1. the "intrinsified" code for G1 (or any SATB based GC),
 491   //    2. the slow path - which is an expansion of the regular method entry.
 492   //
 493   // Notes:
 494   // * In the G1 code we do not check whether we need to block for
 495   //   a safepoint. If G1 is enabled then we must execute the specialized
 496   //   code for Reference.get (except when the Reference object is null)
 497   //   so that we can log the value in the referent field with an SATB
 498   //   update buffer.
 499   //   If the code for the getfield template is modified so that the
 500   //   G1 pre-barrier code is executed when the current method is
 501   //   Reference.get() then going through the normal method entry
 502   //   will be fine.
 503   // * The G1 code can, however, check the receiver object (the instance
 504   //   of java.lang.Reference) and jump to the slow path if null. If the
 505   //   Reference object is null then we obviously cannot fetch the referent
 506   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 507   //   regular method entry code to generate the NPE.
 508   //
 509 
 510   if (UseG1GC) {
 511     address entry = __ pc();
 512 
 513     const int referent_offset = java_lang_ref_Reference::referent_offset;
 514     guarantee(referent_offset > 0, "referent offset not initialized");
 515 
 516     Label slow_path;
 517 
 518     // Debugging not possible, so can't use __ skip_if_jvmti_mode(slow_path, GR31_SCRATCH);
 519 
 520     // In the G1 code we don't check if we need to reach a safepoint. We
 521     // continue and the thread will safepoint at the next bytecode dispatch.
 522 
 523     // If the receiver is null then it is OK to jump to the slow path.
 524     __ ld(R3_RET, Interpreter::stackElementSize, R15_esp); // get receiver
 525 
 526     // Check if receiver == NULL and go the slow path.
 527     __ cmpdi(CCR0, R3_RET, 0);
 528     __ beq(CCR0, slow_path);
 529 
 530     // Load the value of the referent field.
 531     __ load_heap_oop(R3_RET, referent_offset, R3_RET);
 532 
 533     // Generate the G1 pre-barrier code to log the value of
 534     // the referent field in an SATB buffer. Note with
 535     // these parameters the pre-barrier does not generate
 536     // the load of the previous value.
 537 
 538     // Restore caller sp for c2i case.
 539 #ifdef ASSERT
 540       __ ld(R9_ARG7, 0, R1_SP);
 541       __ ld(R10_ARG8, 0, R21_sender_SP);
 542       __ cmpd(CCR0, R9_ARG7, R10_ARG8);
 543       __ asm_assert_eq("backlink", 0x544);
 544 #endif // ASSERT
 545     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
 546 
 547     __ g1_write_barrier_pre(noreg,         // obj
 548                             noreg,         // offset
 549                             R3_RET,        // pre_val
 550                             R11_scratch1,  // tmp
 551                             R12_scratch2,  // tmp
 552                             true);         // needs_frame
 553 
 554     __ blr();
 555 
 556     // Generate regular method entry.
 557     __ bind(slow_path);
 558     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R11_scratch1);
 559     return entry;
 560   }
 561 
 562   return NULL;
 563 }
 564 
 565 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 566   address entry = __ pc();
 567 
 568   // Expression stack must be empty before entering the VM if an
 569   // exception happened.
 570   __ empty_expression_stack();
 571   // Throw exception.
 572   __ call_VM(noreg,
 573              CAST_FROM_FN_PTR(address,
 574                               InterpreterRuntime::throw_StackOverflowError));
 575   return entry;
 576 }
 577 
 578 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 579   address entry = __ pc();
 580   __ empty_expression_stack();
 581   __ load_const_optimized(R4_ARG2, (address) name);
 582   // Index is in R17_tos.
 583   __ mr(R5_ARG3, R17_tos);
 584   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException));
 585   return entry;
 586 }
 587 
 588 #if 0
 589 // Call special ClassCastException constructor taking object to cast
 590 // and target class as arguments.
 591 address TemplateInterpreterGenerator::generate_ClassCastException_verbose_handler() {
 592   address entry = __ pc();
 593 
 594   // Expression stack must be empty before entering the VM if an
 595   // exception happened.
 596   __ empty_expression_stack();
 597 
 598   // Thread will be loaded to R3_ARG1.
 599   // Target class oop is in register R5_ARG3 by convention!
 600   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException_verbose), R17_tos, R5_ARG3);
 601   // Above call must not return here since exception pending.
 602   DEBUG_ONLY(__ should_not_reach_here();)
 603   return entry;
 604 }
 605 #endif
 606 
 607 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 608   address entry = __ pc();
 609   // Expression stack must be empty before entering the VM if an
 610   // exception happened.
 611   __ empty_expression_stack();
 612 
 613   // Load exception object.
 614   // Thread will be loaded to R3_ARG1.
 615   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException), R17_tos);
 616 #ifdef ASSERT
 617   // Above call must not return here since exception pending.
 618   __ should_not_reach_here();
 619 #endif
 620   return entry;
 621 }
 622 
 623 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
 624   address entry = __ pc();
 625   //__ untested("generate_exception_handler_common");
 626   Register Rexception = R17_tos;
 627 
 628   // Expression stack must be empty before entering the VM if an exception happened.
 629   __ empty_expression_stack();
 630 
 631   __ load_const_optimized(R4_ARG2, (address) name, R11_scratch1);
 632   if (pass_oop) {
 633     __ mr(R5_ARG3, Rexception);
 634     __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), false);
 635   } else {
 636     __ load_const_optimized(R5_ARG3, (address) message, R11_scratch1);
 637     __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), false);
 638   }
 639 
 640   // Throw exception.
 641   __ mr(R3_ARG1, Rexception);
 642   __ load_const_optimized(R11_scratch1, Interpreter::throw_exception_entry(), R12_scratch2);
 643   __ mtctr(R11_scratch1);
 644   __ bctr();
 645 
 646   return entry;
 647 }
 648 
 649 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
 650   address entry = __ pc();
 651   __ unimplemented("generate_continuation_for");
 652   return entry;
 653 }
 654 
 655 // This entry is returned to when a call returns to the interpreter.
 656 // When we arrive here, we expect that the callee stack frame is already popped.
 657 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
 658   address entry = __ pc();
 659 
 660   // Move the value out of the return register back to the TOS cache of current frame.
 661   switch (state) {
 662     case ltos:
 663     case btos:
 664     case ztos:
 665     case ctos:
 666     case stos:
 667     case atos:
 668     case itos: __ mr(R17_tos, R3_RET); break;   // RET -> TOS cache
 669     case ftos:
 670     case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
 671     case vtos: break;                           // Nothing to do, this was a void return.
 672     default  : ShouldNotReachHere();
 673   }
 674 
 675   __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
 676   __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
 677   __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
 678 
 679   // Compiled code destroys templateTableBase, reload.
 680   __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R12_scratch2);
 681 
 682   if (state == atos) {
 683     __ profile_return_type(R3_RET, R11_scratch1, R12_scratch2);
 684   }
 685 
 686   const Register cache = R11_scratch1;
 687   const Register size  = R12_scratch2;
 688   __ get_cache_and_index_at_bcp(cache, 1, index_size);
 689 
 690   // Get least significant byte of 64 bit value:
 691 #if defined(VM_LITTLE_ENDIAN)
 692   __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()), cache);
 693 #else
 694   __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()) + 7, cache);
 695 #endif
 696   __ sldi(size, size, Interpreter::logStackElementSize);
 697   __ add(R15_esp, R15_esp, size);
 698   __ dispatch_next(state, step);
 699   return entry;
 700 }
 701 
 702 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
 703   address entry = __ pc();
 704   // If state != vtos, we're returning from a native method, which put it's result
 705   // into the result register. So move the value out of the return register back
 706   // to the TOS cache of current frame.
 707 
 708   switch (state) {
 709     case ltos:
 710     case btos:
 711     case ztos:
 712     case ctos:
 713     case stos:
 714     case atos:
 715     case itos: __ mr(R17_tos, R3_RET); break;   // GR_RET -> TOS cache
 716     case ftos:
 717     case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
 718     case vtos: break;                           // Nothing to do, this was a void return.
 719     default  : ShouldNotReachHere();
 720   }
 721 
 722   // Load LcpoolCache @@@ should be already set!
 723   __ get_constant_pool_cache(R27_constPoolCache);
 724 
 725   // Handle a pending exception, fall through if none.
 726   __ check_and_forward_exception(R11_scratch1, R12_scratch2);
 727 
 728   // Start executing bytecodes.
 729   __ dispatch_next(state, step);
 730 
 731   return entry;
 732 }
 733 
 734 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 735   address entry = __ pc();
 736 
 737   __ push(state);
 738   __ call_VM(noreg, runtime_entry);
 739   __ dispatch_via(vtos, Interpreter::_normal_table.table_for(vtos));
 740 
 741   return entry;
 742 }
 743 
 744 // Helpers for commoning out cases in the various type of method entries.
 745 
 746 // Increment invocation count & check for overflow.
 747 //
 748 // Note: checking for negative value instead of overflow
 749 //       so we have a 'sticky' overflow test.
 750 //
 751 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 752   // Note: In tiered we increment either counters in method or in MDO depending if we're profiling or not.
 753   Register Rscratch1   = R11_scratch1;
 754   Register Rscratch2   = R12_scratch2;
 755   Register R3_counters = R3_ARG1;
 756   Label done;
 757 
 758   if (TieredCompilation) {
 759     const int increment = InvocationCounter::count_increment;
 760     Label no_mdo;
 761     if (ProfileInterpreter) {
 762       const Register Rmdo = R3_counters;
 763       // If no method data exists, go to profile_continue.
 764       __ ld(Rmdo, in_bytes(Method::method_data_offset()), R19_method);
 765       __ cmpdi(CCR0, Rmdo, 0);
 766       __ beq(CCR0, no_mdo);
 767 
 768       // Increment invocation counter in the MDO.
 769       const int mdo_ic_offs = in_bytes(MethodData::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
 770       __ lwz(Rscratch2, mdo_ic_offs, Rmdo);
 771       __ lwz(Rscratch1, in_bytes(MethodData::invoke_mask_offset()), Rmdo);
 772       __ addi(Rscratch2, Rscratch2, increment);
 773       __ stw(Rscratch2, mdo_ic_offs, Rmdo);
 774       __ and_(Rscratch1, Rscratch2, Rscratch1);
 775       __ bne(CCR0, done);
 776       __ b(*overflow);
 777     }
 778 
 779     // Increment counter in MethodCounters*.
 780     const int mo_ic_offs = in_bytes(MethodCounters::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
 781     __ bind(no_mdo);
 782     __ get_method_counters(R19_method, R3_counters, done);
 783     __ lwz(Rscratch2, mo_ic_offs, R3_counters);
 784     __ lwz(Rscratch1, in_bytes(MethodCounters::invoke_mask_offset()), R3_counters);
 785     __ addi(Rscratch2, Rscratch2, increment);
 786     __ stw(Rscratch2, mo_ic_offs, R3_counters);
 787     __ and_(Rscratch1, Rscratch2, Rscratch1);
 788     __ beq(CCR0, *overflow);
 789 
 790     __ bind(done);
 791 
 792   } else {
 793 
 794     // Update standard invocation counters.
 795     Register Rsum_ivc_bec = R4_ARG2;
 796     __ get_method_counters(R19_method, R3_counters, done);
 797     __ increment_invocation_counter(R3_counters, Rsum_ivc_bec, R12_scratch2);
 798     // Increment interpreter invocation counter.
 799     if (ProfileInterpreter) {  // %%% Merge this into methodDataOop.
 800       __ lwz(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
 801       __ addi(R12_scratch2, R12_scratch2, 1);
 802       __ stw(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
 803     }
 804     // Check if we must create a method data obj.
 805     if (ProfileInterpreter && profile_method != NULL) {
 806       const Register profile_limit = Rscratch1;
 807       __ lwz(profile_limit, in_bytes(MethodCounters::interpreter_profile_limit_offset()), R3_counters);
 808       // Test to see if we should create a method data oop.
 809       __ cmpw(CCR0, Rsum_ivc_bec, profile_limit);
 810       __ blt(CCR0, *profile_method_continue);
 811       // If no method data exists, go to profile_method.
 812       __ test_method_data_pointer(*profile_method);
 813     }
 814     // Finally check for counter overflow.
 815     if (overflow) {
 816       const Register invocation_limit = Rscratch1;
 817       __ lwz(invocation_limit, in_bytes(MethodCounters::interpreter_invocation_limit_offset()), R3_counters);
 818       __ cmpw(CCR0, Rsum_ivc_bec, invocation_limit);
 819       __ bge(CCR0, *overflow);
 820     }
 821 
 822     __ bind(done);
 823   }
 824 }
 825 
 826 // Generate code to initiate compilation on invocation counter overflow.
 827 void TemplateInterpreterGenerator::generate_counter_overflow(Label& continue_entry) {
 828   // Generate code to initiate compilation on the counter overflow.
 829 
 830   // InterpreterRuntime::frequency_counter_overflow takes one arguments,
 831   // which indicates if the counter overflow occurs at a backwards branch (NULL bcp)
 832   // We pass zero in.
 833   // The call returns the address of the verified entry point for the method or NULL
 834   // if the compilation did not complete (either went background or bailed out).
 835   //
 836   // Unlike the C++ interpreter above: Check exceptions!
 837   // Assumption: Caller must set the flag "do_not_unlock_if_sychronized" if the monitor of a sync'ed
 838   // method has not yet been created. Thus, no unlocking of a non-existing monitor can occur.
 839 
 840   __ li(R4_ARG2, 0);
 841   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), R4_ARG2, true);
 842 
 843   // Returns verified_entry_point or NULL.
 844   // We ignore it in any case.
 845   __ b(continue_entry);
 846 }
 847 
 848 // See if we've got enough room on the stack for locals plus overhead below
 849 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
 850 // without going through the signal handler, i.e., reserved and yellow zones
 851 // will not be made usable. The shadow zone must suffice to handle the
 852 // overflow.
 853 //
 854 // Kills Rmem_frame_size, Rscratch1.
 855 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rmem_frame_size, Register Rscratch1) {
 856   Label done;
 857   assert_different_registers(Rmem_frame_size, Rscratch1);
 858 
 859   BLOCK_COMMENT("stack_overflow_check_with_compare {");
 860   __ sub(Rmem_frame_size, R1_SP, Rmem_frame_size);
 861   __ ld(Rscratch1, thread_(stack_overflow_limit));
 862   __ cmpld(CCR0/*is_stack_overflow*/, Rmem_frame_size, Rscratch1);
 863   __ bgt(CCR0/*is_stack_overflow*/, done);
 864 
 865   // The stack overflows. Load target address of the runtime stub and call it.
 866   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order");
 867   __ load_const_optimized(Rscratch1, (StubRoutines::throw_StackOverflowError_entry()), R0);
 868   __ mtctr(Rscratch1);
 869   // Restore caller_sp.
 870 #ifdef ASSERT
 871   __ ld(Rscratch1, 0, R1_SP);
 872   __ ld(R0, 0, R21_sender_SP);
 873   __ cmpd(CCR0, R0, Rscratch1);
 874   __ asm_assert_eq("backlink", 0x547);
 875 #endif // ASSERT
 876   __ mr(R1_SP, R21_sender_SP);
 877   __ bctr();
 878 
 879   __ align(32, 12);
 880   __ bind(done);
 881   BLOCK_COMMENT("} stack_overflow_check_with_compare");
 882 }
 883 
 884 void TemplateInterpreterGenerator::unlock_method(bool check_exceptions) {
 885   __ unlock_object(R26_monitor, check_exceptions);
 886 }
 887 
 888 // Lock the current method, interpreter register window must be set up!
 889 void TemplateInterpreterGenerator::lock_method(Register Rflags, Register Rscratch1, Register Rscratch2, bool flags_preloaded) {
 890   const Register Robj_to_lock = Rscratch2;
 891 
 892   {
 893     if (!flags_preloaded) {
 894       __ lwz(Rflags, method_(access_flags));
 895     }
 896 
 897 #ifdef ASSERT
 898     // Check if methods needs synchronization.
 899     {
 900       Label Lok;
 901       __ testbitdi(CCR0, R0, Rflags, JVM_ACC_SYNCHRONIZED_BIT);
 902       __ btrue(CCR0,Lok);
 903       __ stop("method doesn't need synchronization");
 904       __ bind(Lok);
 905     }
 906 #endif // ASSERT
 907   }
 908 
 909   // Get synchronization object to Rscratch2.
 910   {
 911     Label Lstatic;
 912     Label Ldone;
 913 
 914     __ testbitdi(CCR0, R0, Rflags, JVM_ACC_STATIC_BIT);
 915     __ btrue(CCR0, Lstatic);
 916 
 917     // Non-static case: load receiver obj from stack and we're done.
 918     __ ld(Robj_to_lock, R18_locals);
 919     __ b(Ldone);
 920 
 921     __ bind(Lstatic); // Static case: Lock the java mirror
 922     __ load_mirror(Robj_to_lock, R19_method);
 923 
 924     __ bind(Ldone);
 925     __ verify_oop(Robj_to_lock);
 926   }
 927 
 928   // Got the oop to lock => execute!
 929   __ add_monitor_to_stack(true, Rscratch1, R0);
 930 
 931   __ std(Robj_to_lock, BasicObjectLock::obj_offset_in_bytes(), R26_monitor);
 932   __ lock_object(R26_monitor, Robj_to_lock);
 933 }
 934 
 935 // Generate a fixed interpreter frame for pure interpreter
 936 // and I2N native transition frames.
 937 //
 938 // Before (stack grows downwards):
 939 //
 940 //         |  ...         |
 941 //         |------------- |
 942 //         |  java arg0   |
 943 //         |  ...         |
 944 //         |  java argn   |
 945 //         |              |   <-   R15_esp
 946 //         |              |
 947 //         |--------------|
 948 //         | abi_112      |
 949 //         |              |   <-   R1_SP
 950 //         |==============|
 951 //
 952 //
 953 // After:
 954 //
 955 //         |  ...         |
 956 //         |  java arg0   |<-   R18_locals
 957 //         |  ...         |
 958 //         |  java argn   |
 959 //         |--------------|
 960 //         |              |
 961 //         |  java locals |
 962 //         |              |
 963 //         |--------------|
 964 //         |  abi_48      |
 965 //         |==============|
 966 //         |              |
 967 //         |   istate     |
 968 //         |              |
 969 //         |--------------|
 970 //         |   monitor    |<-   R26_monitor
 971 //         |--------------|
 972 //         |              |<-   R15_esp
 973 //         | expression   |
 974 //         | stack        |
 975 //         |              |
 976 //         |--------------|
 977 //         |              |
 978 //         | abi_112      |<-   R1_SP
 979 //         |==============|
 980 //
 981 // The top most frame needs an abi space of 112 bytes. This space is needed,
 982 // since we call to c. The c function may spill their arguments to the caller
 983 // frame. When we call to java, we don't need these spill slots. In order to save
 984 // space on the stack, we resize the caller. However, java locals reside in
 985 // the caller frame and the frame has to be increased. The frame_size for the
 986 // current frame was calculated based on max_stack as size for the expression
 987 // stack. At the call, just a part of the expression stack might be used.
 988 // We don't want to waste this space and cut the frame back accordingly.
 989 // The resulting amount for resizing is calculated as follows:
 990 // resize =   (number_of_locals - number_of_arguments) * slot_size
 991 //          + (R1_SP - R15_esp) + 48
 992 //
 993 // The size for the callee frame is calculated:
 994 // framesize = 112 + max_stack + monitor + state_size
 995 //
 996 // maxstack:   Max number of slots on the expression stack, loaded from the method.
 997 // monitor:    We statically reserve room for one monitor object.
 998 // state_size: We save the current state of the interpreter to this area.
 999 //
1000 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call, Register Rsize_of_parameters, Register Rsize_of_locals) {
1001   Register parent_frame_resize = R6_ARG4, // Frame will grow by this number of bytes.
1002            top_frame_size      = R7_ARG5,
1003            Rconst_method       = R8_ARG6;
1004 
1005   assert_different_registers(Rsize_of_parameters, Rsize_of_locals, parent_frame_resize, top_frame_size);
1006 
1007   __ ld(Rconst_method, method_(const));
1008   __ lhz(Rsize_of_parameters /* number of params */,
1009          in_bytes(ConstMethod::size_of_parameters_offset()), Rconst_method);
1010   if (native_call) {
1011     // If we're calling a native method, we reserve space for the worst-case signature
1012     // handler varargs vector, which is max(Argument::n_register_parameters, parameter_count+2).
1013     // We add two slots to the parameter_count, one for the jni
1014     // environment and one for a possible native mirror.
1015     Label skip_native_calculate_max_stack;
1016     __ addi(top_frame_size, Rsize_of_parameters, 2);
1017     __ cmpwi(CCR0, top_frame_size, Argument::n_register_parameters);
1018     __ bge(CCR0, skip_native_calculate_max_stack);
1019     __ li(top_frame_size, Argument::n_register_parameters);
1020     __ bind(skip_native_calculate_max_stack);
1021     __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
1022     __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
1023     __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
1024     assert(Rsize_of_locals == noreg, "Rsize_of_locals not initialized"); // Only relevant value is Rsize_of_parameters.
1025   } else {
1026     __ lhz(Rsize_of_locals /* number of params */, in_bytes(ConstMethod::size_of_locals_offset()), Rconst_method);
1027     __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
1028     __ sldi(Rsize_of_locals, Rsize_of_locals, Interpreter::logStackElementSize);
1029     __ lhz(top_frame_size, in_bytes(ConstMethod::max_stack_offset()), Rconst_method);
1030     __ sub(R11_scratch1, Rsize_of_locals, Rsize_of_parameters); // >=0
1031     __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
1032     __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
1033     __ add(parent_frame_resize, parent_frame_resize, R11_scratch1);
1034   }
1035 
1036   // Compute top frame size.
1037   __ addi(top_frame_size, top_frame_size, frame::abi_reg_args_size + frame::ijava_state_size);
1038 
1039   // Cut back area between esp and max_stack.
1040   __ addi(parent_frame_resize, parent_frame_resize, frame::abi_minframe_size - Interpreter::stackElementSize);
1041 
1042   __ round_to(top_frame_size, frame::alignment_in_bytes);
1043   __ round_to(parent_frame_resize, frame::alignment_in_bytes);
1044   // parent_frame_resize = (locals-parameters) - (ESP-SP-ABI48) Rounded to frame alignment size.
1045   // Enlarge by locals-parameters (not in case of native_call), shrink by ESP-SP-ABI48.
1046 
1047   if (!native_call) {
1048     // Stack overflow check.
1049     // Native calls don't need the stack size check since they have no
1050     // expression stack and the arguments are already on the stack and
1051     // we only add a handful of words to the stack.
1052     __ add(R11_scratch1, parent_frame_resize, top_frame_size);
1053     generate_stack_overflow_check(R11_scratch1, R12_scratch2);
1054   }
1055 
1056   // Set up interpreter state registers.
1057 
1058   __ add(R18_locals, R15_esp, Rsize_of_parameters);
1059   __ ld(R27_constPoolCache, in_bytes(ConstMethod::constants_offset()), Rconst_method);
1060   __ ld(R27_constPoolCache, ConstantPool::cache_offset_in_bytes(), R27_constPoolCache);
1061 
1062   // Set method data pointer.
1063   if (ProfileInterpreter) {
1064     Label zero_continue;
1065     __ ld(R28_mdx, method_(method_data));
1066     __ cmpdi(CCR0, R28_mdx, 0);
1067     __ beq(CCR0, zero_continue);
1068     __ addi(R28_mdx, R28_mdx, in_bytes(MethodData::data_offset()));
1069     __ bind(zero_continue);
1070   }
1071 
1072   if (native_call) {
1073     __ li(R14_bcp, 0); // Must initialize.
1074   } else {
1075     __ add(R14_bcp, in_bytes(ConstMethod::codes_offset()), Rconst_method);
1076   }
1077 
1078   // Resize parent frame.
1079   __ mflr(R12_scratch2);
1080   __ neg(parent_frame_resize, parent_frame_resize);
1081   __ resize_frame(parent_frame_resize, R11_scratch1);
1082   __ std(R12_scratch2, _abi(lr), R1_SP);
1083 
1084   __ addi(R26_monitor, R1_SP, - frame::ijava_state_size);
1085   __ addi(R15_esp, R26_monitor, - Interpreter::stackElementSize);
1086 
1087   // Get mirror and store it in the frame as GC root for this Method*.
1088   __ load_mirror(R12_scratch2, R19_method);
1089 
1090   // Store values.
1091   // R15_esp, R14_bcp, R26_monitor, R28_mdx are saved at java calls
1092   // in InterpreterMacroAssembler::call_from_interpreter.
1093   __ std(R19_method, _ijava_state_neg(method), R1_SP);
1094   __ std(R12_scratch2, _ijava_state_neg(mirror), R1_SP);
1095   __ std(R21_sender_SP, _ijava_state_neg(sender_sp), R1_SP);
1096   __ std(R27_constPoolCache, _ijava_state_neg(cpoolCache), R1_SP);
1097   __ std(R18_locals, _ijava_state_neg(locals), R1_SP);
1098 
1099   // Note: esp, bcp, monitor, mdx live in registers. Hence, the correct version can only
1100   // be found in the frame after save_interpreter_state is done. This is always true
1101   // for non-top frames. But when a signal occurs, dumping the top frame can go wrong,
1102   // because e.g. frame::interpreter_frame_bcp() will not access the correct value
1103   // (Enhanced Stack Trace).
1104   // The signal handler does not save the interpreter state into the frame.
1105   __ li(R0, 0);
1106 #ifdef ASSERT
1107   // Fill remaining slots with constants.
1108   __ load_const_optimized(R11_scratch1, 0x5afe);
1109   __ load_const_optimized(R12_scratch2, 0xdead);
1110 #endif
1111   // We have to initialize some frame slots for native calls (accessed by GC).
1112   if (native_call) {
1113     __ std(R26_monitor, _ijava_state_neg(monitors), R1_SP);
1114     __ std(R14_bcp, _ijava_state_neg(bcp), R1_SP);
1115     if (ProfileInterpreter) { __ std(R28_mdx, _ijava_state_neg(mdx), R1_SP); }
1116   }
1117 #ifdef ASSERT
1118   else {
1119     __ std(R12_scratch2, _ijava_state_neg(monitors), R1_SP);
1120     __ std(R12_scratch2, _ijava_state_neg(bcp), R1_SP);
1121     __ std(R12_scratch2, _ijava_state_neg(mdx), R1_SP);
1122   }
1123   __ std(R11_scratch1, _ijava_state_neg(ijava_reserved), R1_SP);
1124   __ std(R12_scratch2, _ijava_state_neg(esp), R1_SP);
1125   __ std(R12_scratch2, _ijava_state_neg(lresult), R1_SP);
1126   __ std(R12_scratch2, _ijava_state_neg(fresult), R1_SP);
1127 #endif
1128   __ subf(R12_scratch2, top_frame_size, R1_SP);
1129   __ std(R0, _ijava_state_neg(oop_tmp), R1_SP);
1130   __ std(R12_scratch2, _ijava_state_neg(top_frame_sp), R1_SP);
1131 
1132   // Push top frame.
1133   __ push_frame(top_frame_size, R11_scratch1);
1134 }
1135 
1136 // End of helpers
1137 
1138 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1139   if (!Interpreter::math_entry_available(kind)) {
1140     NOT_PRODUCT(__ should_not_reach_here();)
1141     return NULL;
1142   }
1143 
1144   address entry = __ pc();
1145 
1146   __ lfd(F1_RET, Interpreter::stackElementSize, R15_esp);
1147 
1148   // Pop c2i arguments (if any) off when we return.
1149 #ifdef ASSERT
1150   __ ld(R9_ARG7, 0, R1_SP);
1151   __ ld(R10_ARG8, 0, R21_sender_SP);
1152   __ cmpd(CCR0, R9_ARG7, R10_ARG8);
1153   __ asm_assert_eq("backlink", 0x545);
1154 #endif // ASSERT
1155   __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1156 
1157   if (kind == Interpreter::java_lang_math_sqrt) {
1158     __ fsqrt(F1_RET, F1_RET);
1159   } else if (kind == Interpreter::java_lang_math_abs) {
1160     __ fabs(F1_RET, F1_RET);
1161   } else {
1162     ShouldNotReachHere();
1163   }
1164 
1165   // And we're done.
1166   __ blr();
1167 
1168   __ flush();
1169 
1170   return entry;
1171 }
1172 
1173 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1174   // Quick & dirty stack overflow checking: bang the stack & handle trap.
1175   // Note that we do the banging after the frame is setup, since the exception
1176   // handling code expects to find a valid interpreter frame on the stack.
1177   // Doing the banging earlier fails if the caller frame is not an interpreter
1178   // frame.
1179   // (Also, the exception throwing code expects to unlock any synchronized
1180   // method receiever, so do the banging after locking the receiver.)
1181 
1182   // Bang each page in the shadow zone. We can't assume it's been done for
1183   // an interpreter frame with greater than a page of locals, so each page
1184   // needs to be checked.  Only true for non-native.
1185   if (UseStackBanging) {
1186     const int page_size = os::vm_page_size();
1187     const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1188     const int start_page = native_call ? n_shadow_pages : 1;
1189     BLOCK_COMMENT("bang_stack_shadow_pages:");
1190     for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1191       __ bang_stack_with_offset(pages*page_size);
1192     }
1193   }
1194 }
1195 
1196 // Interpreter stub for calling a native method. (asm interpreter)
1197 // This sets up a somewhat different looking stack for calling the
1198 // native method than the typical interpreter frame setup.
1199 //
1200 // On entry:
1201 //   R19_method    - method
1202 //   R16_thread    - JavaThread*
1203 //   R15_esp       - intptr_t* sender tos
1204 //
1205 //   abstract stack (grows up)
1206 //     [  IJava (caller of JNI callee)  ]  <-- ASP
1207 //        ...
1208 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1209 
1210   address entry = __ pc();
1211 
1212   const bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1213 
1214   // -----------------------------------------------------------------------------
1215   // Allocate a new frame that represents the native callee (i2n frame).
1216   // This is not a full-blown interpreter frame, but in particular, the
1217   // following registers are valid after this:
1218   // - R19_method
1219   // - R18_local (points to start of arguments to native function)
1220   //
1221   //   abstract stack (grows up)
1222   //     [  IJava (caller of JNI callee)  ]  <-- ASP
1223   //        ...
1224 
1225   const Register signature_handler_fd = R11_scratch1;
1226   const Register pending_exception    = R0;
1227   const Register result_handler_addr  = R31;
1228   const Register native_method_fd     = R11_scratch1;
1229   const Register access_flags         = R22_tmp2;
1230   const Register active_handles       = R11_scratch1; // R26_monitor saved to state.
1231   const Register sync_state           = R12_scratch2;
1232   const Register sync_state_addr      = sync_state;   // Address is dead after use.
1233   const Register suspend_flags        = R11_scratch1;
1234 
1235   //=============================================================================
1236   // Allocate new frame and initialize interpreter state.
1237 
1238   Label exception_return;
1239   Label exception_return_sync_check;
1240   Label stack_overflow_return;
1241 
1242   // Generate new interpreter state and jump to stack_overflow_return in case of
1243   // a stack overflow.
1244   //generate_compute_interpreter_state(stack_overflow_return);
1245 
1246   Register size_of_parameters = R22_tmp2;
1247 
1248   generate_fixed_frame(true, size_of_parameters, noreg /* unused */);
1249 
1250   //=============================================================================
1251   // Increment invocation counter. On overflow, entry to JNI method
1252   // will be compiled.
1253   Label invocation_counter_overflow, continue_after_compile;
1254   if (inc_counter) {
1255     if (synchronized) {
1256       // Since at this point in the method invocation the exception handler
1257       // would try to exit the monitor of synchronized methods which hasn't
1258       // been entered yet, we set the thread local variable
1259       // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1260       // runtime, exception handling i.e. unlock_if_synchronized_method will
1261       // check this thread local flag.
1262       // This flag has two effects, one is to force an unwind in the topmost
1263       // interpreter frame and not perform an unlock while doing so.
1264       __ li(R0, 1);
1265       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1266     }
1267     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1268 
1269     BIND(continue_after_compile);
1270   }
1271 
1272   bang_stack_shadow_pages(true);
1273 
1274   if (inc_counter) {
1275     // Reset the _do_not_unlock_if_synchronized flag.
1276     if (synchronized) {
1277       __ li(R0, 0);
1278       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1279     }
1280   }
1281 
1282   // access_flags = method->access_flags();
1283   // Load access flags.
1284   assert(access_flags->is_nonvolatile(),
1285          "access_flags must be in a non-volatile register");
1286   // Type check.
1287   assert(4 == sizeof(AccessFlags), "unexpected field size");
1288   __ lwz(access_flags, method_(access_flags));
1289 
1290   // We don't want to reload R19_method and access_flags after calls
1291   // to some helper functions.
1292   assert(R19_method->is_nonvolatile(),
1293          "R19_method must be a non-volatile register");
1294 
1295   // Check for synchronized methods. Must happen AFTER invocation counter
1296   // check, so method is not locked if counter overflows.
1297 
1298   if (synchronized) {
1299     lock_method(access_flags, R11_scratch1, R12_scratch2, true);
1300 
1301     // Update monitor in state.
1302     __ ld(R11_scratch1, 0, R1_SP);
1303     __ std(R26_monitor, _ijava_state_neg(monitors), R11_scratch1);
1304   }
1305 
1306   // jvmti/jvmpi support
1307   __ notify_method_entry();
1308 
1309   //=============================================================================
1310   // Get and call the signature handler.
1311 
1312   __ ld(signature_handler_fd, method_(signature_handler));
1313   Label call_signature_handler;
1314 
1315   __ cmpdi(CCR0, signature_handler_fd, 0);
1316   __ bne(CCR0, call_signature_handler);
1317 
1318   // Method has never been called. Either generate a specialized
1319   // handler or point to the slow one.
1320   //
1321   // Pass parameter 'false' to avoid exception check in call_VM.
1322   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false);
1323 
1324   // Check for an exception while looking up the target method. If we
1325   // incurred one, bail.
1326   __ ld(pending_exception, thread_(pending_exception));
1327   __ cmpdi(CCR0, pending_exception, 0);
1328   __ bne(CCR0, exception_return_sync_check); // Has pending exception.
1329 
1330   // Reload signature handler, it may have been created/assigned in the meanwhile.
1331   __ ld(signature_handler_fd, method_(signature_handler));
1332   __ twi_0(signature_handler_fd); // Order wrt. load of klass mirror and entry point (isync is below).
1333 
1334   BIND(call_signature_handler);
1335 
1336   // Before we call the signature handler we push a new frame to
1337   // protect the interpreter frame volatile registers when we return
1338   // from jni but before we can get back to Java.
1339 
1340   // First set the frame anchor while the SP/FP registers are
1341   // convenient and the slow signature handler can use this same frame
1342   // anchor.
1343 
1344   // We have a TOP_IJAVA_FRAME here, which belongs to us.
1345   __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
1346 
1347   // Now the interpreter frame (and its call chain) have been
1348   // invalidated and flushed. We are now protected against eager
1349   // being enabled in native code. Even if it goes eager the
1350   // registers will be reloaded as clean and we will invalidate after
1351   // the call so no spurious flush should be possible.
1352 
1353   // Call signature handler and pass locals address.
1354   //
1355   // Our signature handlers copy required arguments to the C stack
1356   // (outgoing C args), R3_ARG1 to R10_ARG8, and FARG1 to FARG13.
1357   __ mr(R3_ARG1, R18_locals);
1358 #if !defined(ABI_ELFv2)
1359   __ ld(signature_handler_fd, 0, signature_handler_fd);
1360 #endif
1361 
1362   __ call_stub(signature_handler_fd);
1363 
1364   // Remove the register parameter varargs slots we allocated in
1365   // compute_interpreter_state. SP+16 ends up pointing to the ABI
1366   // outgoing argument area.
1367   //
1368   // Not needed on PPC64.
1369   //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord);
1370 
1371   assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register");
1372   // Save across call to native method.
1373   __ mr(result_handler_addr, R3_RET);
1374 
1375   __ isync(); // Acquire signature handler before trying to fetch the native entry point and klass mirror.
1376 
1377   // Set up fixed parameters and call the native method.
1378   // If the method is static, get mirror into R4_ARG2.
1379   {
1380     Label method_is_not_static;
1381     // Access_flags is non-volatile and still, no need to restore it.
1382 
1383     // Restore access flags.
1384     __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT);
1385     __ bfalse(CCR0, method_is_not_static);
1386 
1387     __ load_mirror(R12_scratch2, R19_method);
1388     // state->_native_mirror = mirror;
1389 
1390     __ ld(R11_scratch1, 0, R1_SP);
1391     __ std(R12_scratch2/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1);
1392     // R4_ARG2 = &state->_oop_temp;
1393     __ addi(R4_ARG2, R11_scratch1, _ijava_state_neg(oop_tmp));
1394     BIND(method_is_not_static);
1395   }
1396 
1397   // At this point, arguments have been copied off the stack into
1398   // their JNI positions. Oops are boxed in-place on the stack, with
1399   // handles copied to arguments. The result handler address is in a
1400   // register.
1401 
1402   // Pass JNIEnv address as first parameter.
1403   __ addir(R3_ARG1, thread_(jni_environment));
1404 
1405   // Load the native_method entry before we change the thread state.
1406   __ ld(native_method_fd, method_(native_function));
1407 
1408   //=============================================================================
1409   // Transition from _thread_in_Java to _thread_in_native. As soon as
1410   // we make this change the safepoint code needs to be certain that
1411   // the last Java frame we established is good. The pc in that frame
1412   // just needs to be near here not an actual return address.
1413 
1414   // We use release_store_fence to update values like the thread state, where
1415   // we don't want the current thread to continue until all our prior memory
1416   // accesses (including the new thread state) are visible to other threads.
1417   __ li(R0, _thread_in_native);
1418   __ release();
1419 
1420   // TODO PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
1421   __ stw(R0, thread_(thread_state));
1422 
1423   if (UseMembar) {
1424     __ fence();
1425   }
1426 
1427   //=============================================================================
1428   // Call the native method. Argument registers must not have been
1429   // overwritten since "__ call_stub(signature_handler);" (except for
1430   // ARG1 and ARG2 for static methods).
1431   __ call_c(native_method_fd);
1432 
1433   __ li(R0, 0);
1434   __ ld(R11_scratch1, 0, R1_SP);
1435   __ std(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1436   __ stfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1437   __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1); // reset
1438 
1439   // Note: C++ interpreter needs the following here:
1440   // The frame_manager_lr field, which we use for setting the last
1441   // java frame, gets overwritten by the signature handler. Restore
1442   // it now.
1443   //__ get_PC_trash_LR(R11_scratch1);
1444   //__ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1445 
1446   // Because of GC R19_method may no longer be valid.
1447 
1448   // Block, if necessary, before resuming in _thread_in_Java state.
1449   // In order for GC to work, don't clear the last_Java_sp until after
1450   // blocking.
1451 
1452   //=============================================================================
1453   // Switch thread to "native transition" state before reading the
1454   // synchronization state. This additional state is necessary
1455   // because reading and testing the synchronization state is not
1456   // atomic w.r.t. GC, as this scenario demonstrates: Java thread A,
1457   // in _thread_in_native state, loads _not_synchronized and is
1458   // preempted. VM thread changes sync state to synchronizing and
1459   // suspends threads for GC. Thread A is resumed to finish this
1460   // native method, but doesn't block here since it didn't see any
1461   // synchronization in progress, and escapes.
1462 
1463   // We use release_store_fence to update values like the thread state, where
1464   // we don't want the current thread to continue until all our prior memory
1465   // accesses (including the new thread state) are visible to other threads.
1466   __ li(R0/*thread_state*/, _thread_in_native_trans);
1467   __ release();
1468   __ stw(R0/*thread_state*/, thread_(thread_state));
1469   if (UseMembar) {
1470     __ fence();
1471   }
1472   // Write serialization page so that the VM thread can do a pseudo remote
1473   // membar. We use the current thread pointer to calculate a thread
1474   // specific offset to write to within the page. This minimizes bus
1475   // traffic due to cache line collision.
1476   else {
1477     __ serialize_memory(R16_thread, R11_scratch1, R12_scratch2);
1478   }
1479 
1480   // Now before we return to java we must look for a current safepoint
1481   // (a new safepoint can not start since we entered native_trans).
1482   // We must check here because a current safepoint could be modifying
1483   // the callers registers right this moment.
1484 
1485   // Acquire isn't strictly necessary here because of the fence, but
1486   // sync_state is declared to be volatile, so we do it anyway
1487   // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
1488   int sync_state_offs = __ load_const_optimized(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/R0, true);
1489 
1490   // TODO PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size");
1491   __ lwz(sync_state, sync_state_offs, sync_state_addr);
1492 
1493   // TODO PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
1494   __ lwz(suspend_flags, thread_(suspend_flags));
1495 
1496   Label sync_check_done;
1497   Label do_safepoint;
1498   // No synchronization in progress nor yet synchronized.
1499   __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized);
1500   // Not suspended.
1501   __ cmpwi(CCR1, suspend_flags, 0);
1502 
1503   __ bne(CCR0, do_safepoint);
1504   __ beq(CCR1, sync_check_done);
1505   __ bind(do_safepoint);
1506   __ isync();
1507   // Block. We do the call directly and leave the current
1508   // last_Java_frame setup undisturbed. We must save any possible
1509   // native result across the call. No oop is present.
1510 
1511   __ mr(R3_ARG1, R16_thread);
1512 #if defined(ABI_ELFv2)
1513   __ call_c(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1514             relocInfo::none);
1515 #else
1516   __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans),
1517             relocInfo::none);
1518 #endif
1519 
1520   __ bind(sync_check_done);
1521 
1522   //=============================================================================
1523   // <<<<<< Back in Interpreter Frame >>>>>
1524 
1525   // We are in thread_in_native_trans here and back in the normal
1526   // interpreter frame. We don't have to do anything special about
1527   // safepoints and we can switch to Java mode anytime we are ready.
1528 
1529   // Note: frame::interpreter_frame_result has a dependency on how the
1530   // method result is saved across the call to post_method_exit. For
1531   // native methods it assumes that the non-FPU/non-void result is
1532   // saved in _native_lresult and a FPU result in _native_fresult. If
1533   // this changes then the interpreter_frame_result implementation
1534   // will need to be updated too.
1535 
1536   // On PPC64, we have stored the result directly after the native call.
1537 
1538   //=============================================================================
1539   // Back in Java
1540 
1541   // We use release_store_fence to update values like the thread state, where
1542   // we don't want the current thread to continue until all our prior memory
1543   // accesses (including the new thread state) are visible to other threads.
1544   __ li(R0/*thread_state*/, _thread_in_Java);
1545   __ release();
1546   __ stw(R0/*thread_state*/, thread_(thread_state));
1547   if (UseMembar) {
1548     __ fence();
1549   }
1550 
1551   __ reset_last_Java_frame();
1552 
1553   // Jvmdi/jvmpi support. Whether we've got an exception pending or
1554   // not, and whether unlocking throws an exception or not, we notify
1555   // on native method exit. If we do have an exception, we'll end up
1556   // in the caller's context to handle it, so if we don't do the
1557   // notify here, we'll drop it on the floor.
1558   __ notify_method_exit(true/*native method*/,
1559                         ilgl /*illegal state (not used for native methods)*/,
1560                         InterpreterMacroAssembler::NotifyJVMTI,
1561                         false /*check_exceptions*/);
1562 
1563   //=============================================================================
1564   // Handle exceptions
1565 
1566   if (synchronized) {
1567     // Don't check for exceptions since we're still in the i2n frame. Do that
1568     // manually afterwards.
1569     unlock_method(false);
1570   }
1571 
1572   // Reset active handles after returning from native.
1573   // thread->active_handles()->clear();
1574   __ ld(active_handles, thread_(active_handles));
1575   // TODO PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
1576   __ li(R0, 0);
1577   __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles);
1578 
1579   Label exception_return_sync_check_already_unlocked;
1580   __ ld(R0/*pending_exception*/, thread_(pending_exception));
1581   __ cmpdi(CCR0, R0/*pending_exception*/, 0);
1582   __ bne(CCR0, exception_return_sync_check_already_unlocked);
1583 
1584   //-----------------------------------------------------------------------------
1585   // No exception pending.
1586 
1587   // Move native method result back into proper registers and return.
1588   // Invoke result handler (may unbox/promote).
1589   __ ld(R11_scratch1, 0, R1_SP);
1590   __ ld(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1591   __ lfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1592   __ call_stub(result_handler_addr);
1593 
1594   __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ R0, R11_scratch1, R12_scratch2);
1595 
1596   // Must use the return pc which was loaded from the caller's frame
1597   // as the VM uses return-pc-patching for deoptimization.
1598   __ mtlr(R0);
1599   __ blr();
1600 
1601   //-----------------------------------------------------------------------------
1602   // An exception is pending. We call into the runtime only if the
1603   // caller was not interpreted. If it was interpreted the
1604   // interpreter will do the correct thing. If it isn't interpreted
1605   // (call stub/compiled code) we will change our return and continue.
1606 
1607   BIND(exception_return_sync_check);
1608 
1609   if (synchronized) {
1610     // Don't check for exceptions since we're still in the i2n frame. Do that
1611     // manually afterwards.
1612     unlock_method(false);
1613   }
1614   BIND(exception_return_sync_check_already_unlocked);
1615 
1616   const Register return_pc = R31;
1617 
1618   __ ld(return_pc, 0, R1_SP);
1619   __ ld(return_pc, _abi(lr), return_pc);
1620 
1621   // Get the address of the exception handler.
1622   __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1623                   R16_thread,
1624                   return_pc /* return pc */);
1625   __ merge_frames(/*top_frame_sp*/ R21_sender_SP, noreg, R11_scratch1, R12_scratch2);
1626 
1627   // Load the PC of the the exception handler into LR.
1628   __ mtlr(R3_RET);
1629 
1630   // Load exception into R3_ARG1 and clear pending exception in thread.
1631   __ ld(R3_ARG1/*exception*/, thread_(pending_exception));
1632   __ li(R4_ARG2, 0);
1633   __ std(R4_ARG2, thread_(pending_exception));
1634 
1635   // Load the original return pc into R4_ARG2.
1636   __ mr(R4_ARG2/*issuing_pc*/, return_pc);
1637 
1638   // Return to exception handler.
1639   __ blr();
1640 
1641   //=============================================================================
1642   // Counter overflow.
1643 
1644   if (inc_counter) {
1645     // Handle invocation counter overflow.
1646     __ bind(invocation_counter_overflow);
1647 
1648     generate_counter_overflow(continue_after_compile);
1649   }
1650 
1651   return entry;
1652 }
1653 
1654 // Generic interpreted method entry to (asm) interpreter.
1655 //
1656 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1657   bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1658   address entry = __ pc();
1659   // Generate the code to allocate the interpreter stack frame.
1660   Register Rsize_of_parameters = R4_ARG2, // Written by generate_fixed_frame.
1661            Rsize_of_locals     = R5_ARG3; // Written by generate_fixed_frame.
1662 
1663   // Does also a stack check to assure this frame fits on the stack.
1664   generate_fixed_frame(false, Rsize_of_parameters, Rsize_of_locals);
1665 
1666   // --------------------------------------------------------------------------
1667   // Zero out non-parameter locals.
1668   // Note: *Always* zero out non-parameter locals as Sparc does. It's not
1669   // worth to ask the flag, just do it.
1670   Register Rslot_addr = R6_ARG4,
1671            Rnum       = R7_ARG5;
1672   Label Lno_locals, Lzero_loop;
1673 
1674   // Set up the zeroing loop.
1675   __ subf(Rnum, Rsize_of_parameters, Rsize_of_locals);
1676   __ subf(Rslot_addr, Rsize_of_parameters, R18_locals);
1677   __ srdi_(Rnum, Rnum, Interpreter::logStackElementSize);
1678   __ beq(CCR0, Lno_locals);
1679   __ li(R0, 0);
1680   __ mtctr(Rnum);
1681 
1682   // The zero locals loop.
1683   __ bind(Lzero_loop);
1684   __ std(R0, 0, Rslot_addr);
1685   __ addi(Rslot_addr, Rslot_addr, -Interpreter::stackElementSize);
1686   __ bdnz(Lzero_loop);
1687 
1688   __ bind(Lno_locals);
1689 
1690   // --------------------------------------------------------------------------
1691   // Counter increment and overflow check.
1692   Label invocation_counter_overflow,
1693         profile_method,
1694         profile_method_continue;
1695   if (inc_counter || ProfileInterpreter) {
1696 
1697     Register Rdo_not_unlock_if_synchronized_addr = R11_scratch1;
1698     if (synchronized) {
1699       // Since at this point in the method invocation the exception handler
1700       // would try to exit the monitor of synchronized methods which hasn't
1701       // been entered yet, we set the thread local variable
1702       // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1703       // runtime, exception handling i.e. unlock_if_synchronized_method will
1704       // check this thread local flag.
1705       // This flag has two effects, one is to force an unwind in the topmost
1706       // interpreter frame and not perform an unlock while doing so.
1707       __ li(R0, 1);
1708       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1709     }
1710 
1711     // Argument and return type profiling.
1712     __ profile_parameters_type(R3_ARG1, R4_ARG2, R5_ARG3, R6_ARG4);
1713 
1714     // Increment invocation counter and check for overflow.
1715     if (inc_counter) {
1716       generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1717     }
1718 
1719     __ bind(profile_method_continue);
1720   }
1721 
1722   bang_stack_shadow_pages(false);
1723 
1724   if (inc_counter || ProfileInterpreter) {
1725     // Reset the _do_not_unlock_if_synchronized flag.
1726     if (synchronized) {
1727       __ li(R0, 0);
1728       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1729     }
1730   }
1731 
1732   // --------------------------------------------------------------------------
1733   // Locking of synchronized methods. Must happen AFTER invocation_counter
1734   // check and stack overflow check, so method is not locked if overflows.
1735   if (synchronized) {
1736     lock_method(R3_ARG1, R4_ARG2, R5_ARG3);
1737   }
1738 #ifdef ASSERT
1739   else {
1740     Label Lok;
1741     __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1742     __ andi_(R0, R0, JVM_ACC_SYNCHRONIZED);
1743     __ asm_assert_eq("method needs synchronization", 0x8521);
1744     __ bind(Lok);
1745   }
1746 #endif // ASSERT
1747 
1748   __ verify_thread();
1749 
1750   // --------------------------------------------------------------------------
1751   // JVMTI support
1752   __ notify_method_entry();
1753 
1754   // --------------------------------------------------------------------------
1755   // Start executing instructions.
1756   __ dispatch_next(vtos);
1757 
1758   // --------------------------------------------------------------------------
1759   // Out of line counter overflow and MDO creation code.
1760   if (ProfileInterpreter) {
1761     // We have decided to profile this method in the interpreter.
1762     __ bind(profile_method);
1763     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1764     __ set_method_data_pointer_for_bcp();
1765     __ b(profile_method_continue);
1766   }
1767 
1768   if (inc_counter) {
1769     // Handle invocation counter overflow.
1770     __ bind(invocation_counter_overflow);
1771     generate_counter_overflow(profile_method_continue);
1772   }
1773   return entry;
1774 }
1775 
1776 // CRC32 Intrinsics.
1777 //
1778 // Contract on scratch and work registers.
1779 // =======================================
1780 //
1781 // On ppc, the register set {R2..R12} is available in the interpreter as scratch/work registers.
1782 // You should, however, keep in mind that {R3_ARG1..R10_ARG8} is the C-ABI argument register set.
1783 // You can't rely on these registers across calls.
1784 //
1785 // The generators for CRC32_update and for CRC32_updateBytes use the
1786 // scratch/work register set internally, passing the work registers
1787 // as arguments to the MacroAssembler emitters as required.
1788 //
1789 // R3_ARG1..R6_ARG4 are preset to hold the incoming java arguments.
1790 // Their contents is not constant but may change according to the requirements
1791 // of the emitted code.
1792 //
1793 // All other registers from the scratch/work register set are used "internally"
1794 // and contain garbage (i.e. unpredictable values) once blr() is reached.
1795 // Basically, only R3_RET contains a defined value which is the function result.
1796 //
1797 /**
1798  * Method entry for static native methods:
1799  *   int java.util.zip.CRC32.update(int crc, int b)
1800  */
1801 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
1802   if (UseCRC32Intrinsics) {
1803     address start = __ pc();  // Remember stub start address (is rtn value).
1804     Label slow_path;
1805 
1806     // Safepoint check
1807     const Register sync_state = R11_scratch1;
1808     int sync_state_offs = __ load_const_optimized(sync_state, SafepointSynchronize::address_of_state(), /*temp*/R0, true);
1809     __ lwz(sync_state, sync_state_offs, sync_state);
1810     __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized);
1811     __ bne(CCR0, slow_path);
1812 
1813     // We don't generate local frame and don't align stack because
1814     // we not even call stub code (we generate the code inline)
1815     // and there is no safepoint on this path.
1816 
1817     // Load java parameters.
1818     // R15_esp is callers operand stack pointer, i.e. it points to the parameters.
1819     const Register argP    = R15_esp;
1820     const Register crc     = R3_ARG1;  // crc value
1821     const Register data    = R4_ARG2;  // address of java byte value (kernel_crc32 needs address)
1822     const Register dataLen = R5_ARG3;  // source data len (1 byte). Not used because calling the single-byte emitter.
1823     const Register table   = R6_ARG4;  // address of crc32 table
1824     const Register tmp     = dataLen;  // Reuse unused len register to show we don't actually need a separate tmp here.
1825 
1826     BLOCK_COMMENT("CRC32_update {");
1827 
1828     // Arguments are reversed on java expression stack
1829 #ifdef VM_LITTLE_ENDIAN
1830     __ addi(data, argP, 0+1*wordSize); // (stack) address of byte value. Emitter expects address, not value.
1831                                        // Being passed as an int, the single byte is at offset +0.
1832 #else
1833     __ addi(data, argP, 3+1*wordSize); // (stack) address of byte value. Emitter expects address, not value.
1834                                        // Being passed from java as an int, the single byte is at offset +3.
1835 #endif
1836     __ lwz(crc,  2*wordSize, argP);    // Current crc state, zero extend to 64 bit to have a clean register.
1837 
1838     StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table);
1839     __ kernel_crc32_singleByte(crc, data, dataLen, table, tmp);
1840 
1841     // Restore caller sp for c2i case and return.
1842     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1843     __ blr();
1844 
1845     // Generate a vanilla native entry as the slow path.
1846     BLOCK_COMMENT("} CRC32_update");
1847     BIND(slow_path);
1848     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native), R11_scratch1);
1849     return start;
1850   }
1851 
1852   return NULL;
1853 }
1854 
1855 // CRC32 Intrinsics.
1856 /**
1857  * Method entry for static native methods:
1858  *   int java.util.zip.CRC32.updateBytes(     int crc, byte[] b,  int off, int len)
1859  *   int java.util.zip.CRC32.updateByteBuffer(int crc, long* buf, int off, int len)
1860  */
1861 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1862   if (UseCRC32Intrinsics) {
1863     address start = __ pc();  // Remember stub start address (is rtn value).
1864     Label slow_path;
1865 
1866     // Safepoint check
1867     const Register sync_state = R11_scratch1;
1868     int sync_state_offs = __ load_const_optimized(sync_state, SafepointSynchronize::address_of_state(), /*temp*/R0, true);
1869     __ lwz(sync_state, sync_state_offs, sync_state);
1870     __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized);
1871     __ bne(CCR0, slow_path);
1872 
1873     // We don't generate local frame and don't align stack because
1874     // we not even call stub code (we generate the code inline)
1875     // and there is no safepoint on this path.
1876 
1877     // Load parameters.
1878     // Z_esp is callers operand stack pointer, i.e. it points to the parameters.
1879     const Register argP    = R15_esp;
1880     const Register crc     = R3_ARG1;  // crc value
1881     const Register data    = R4_ARG2;  // address of java byte array
1882     const Register dataLen = R5_ARG3;  // source data len
1883     const Register table   = R6_ARG4;  // address of crc32 table
1884 
1885     const Register t0      = R9;       // scratch registers for crc calculation
1886     const Register t1      = R10;
1887     const Register t2      = R11;
1888     const Register t3      = R12;
1889 
1890     const Register tc0     = R2;       // registers to hold pre-calculated column addresses
1891     const Register tc1     = R7;
1892     const Register tc2     = R8;
1893     const Register tc3     = table;    // table address is reconstructed at the end of kernel_crc32_* emitters
1894 
1895     const Register tmp     = t0;       // Only used very locally to calculate byte buffer address.
1896 
1897     // Arguments are reversed on java expression stack.
1898     // Calculate address of start element.
1899     if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) { // Used for "updateByteBuffer direct".
1900       BLOCK_COMMENT("CRC32_updateByteBuffer {");
1901       // crc     @ (SP + 5W) (32bit)
1902       // buf     @ (SP + 3W) (64bit ptr to long array)
1903       // off     @ (SP + 2W) (32bit)
1904       // dataLen @ (SP + 1W) (32bit)
1905       // data = buf + off
1906       __ ld(  data,    3*wordSize, argP);  // start of byte buffer
1907       __ lwa( tmp,     2*wordSize, argP);  // byte buffer offset
1908       __ lwa( dataLen, 1*wordSize, argP);  // #bytes to process
1909       __ lwz( crc,     5*wordSize, argP);  // current crc state
1910       __ add( data, data, tmp);            // Add byte buffer offset.
1911     } else {                                                         // Used for "updateBytes update".
1912       BLOCK_COMMENT("CRC32_updateBytes {");
1913       // crc     @ (SP + 4W) (32bit)
1914       // buf     @ (SP + 3W) (64bit ptr to byte array)
1915       // off     @ (SP + 2W) (32bit)
1916       // dataLen @ (SP + 1W) (32bit)
1917       // data = buf + off + base_offset
1918       __ ld(  data,    3*wordSize, argP);  // start of byte buffer
1919       __ lwa( tmp,     2*wordSize, argP);  // byte buffer offset
1920       __ lwa( dataLen, 1*wordSize, argP);  // #bytes to process
1921       __ add( data, data, tmp);            // add byte buffer offset
1922       __ lwz( crc,     4*wordSize, argP);  // current crc state
1923       __ addi(data, data, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1924     }
1925 
1926     StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table);
1927 
1928     // Performance measurements show the 1word and 2word variants to be almost equivalent,
1929     // with very light advantages for the 1word variant. We chose the 1word variant for
1930     // code compactness.
1931     __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, tc3);
1932 
1933     // Restore caller sp for c2i case and return.
1934     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1935     __ blr();
1936 
1937     // Generate a vanilla native entry as the slow path.
1938     BLOCK_COMMENT("} CRC32_updateBytes(Buffer)");
1939     BIND(slow_path);
1940     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native), R11_scratch1);
1941     return start;
1942   }
1943 
1944   return NULL;
1945 }
1946 
1947 // Not supported
1948 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1949   return NULL;
1950 }
1951 
1952 // =============================================================================
1953 // Exceptions
1954 
1955 void TemplateInterpreterGenerator::generate_throw_exception() {
1956   Register Rexception    = R17_tos,
1957            Rcontinuation = R3_RET;
1958 
1959   // --------------------------------------------------------------------------
1960   // Entry point if an method returns with a pending exception (rethrow).
1961   Interpreter::_rethrow_exception_entry = __ pc();
1962   {
1963     __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
1964     __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
1965     __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
1966 
1967     // Compiled code destroys templateTableBase, reload.
1968     __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1969   }
1970 
1971   // Entry point if a interpreted method throws an exception (throw).
1972   Interpreter::_throw_exception_entry = __ pc();
1973   {
1974     __ mr(Rexception, R3_RET);
1975 
1976     __ verify_thread();
1977     __ verify_oop(Rexception);
1978 
1979     // Expression stack must be empty before entering the VM in case of an exception.
1980     __ empty_expression_stack();
1981     // Find exception handler address and preserve exception oop.
1982     // Call C routine to find handler and jump to it.
1983     __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Rexception);
1984     __ mtctr(Rcontinuation);
1985     // Push exception for exception handler bytecodes.
1986     __ push_ptr(Rexception);
1987 
1988     // Jump to exception handler (may be remove activation entry!).
1989     __ bctr();
1990   }
1991 
1992   // If the exception is not handled in the current frame the frame is
1993   // removed and the exception is rethrown (i.e. exception
1994   // continuation is _rethrow_exception).
1995   //
1996   // Note: At this point the bci is still the bxi for the instruction
1997   // which caused the exception and the expression stack is
1998   // empty. Thus, for any VM calls at this point, GC will find a legal
1999   // oop map (with empty expression stack).
2000 
2001   // In current activation
2002   // tos: exception
2003   // bcp: exception bcp
2004 
2005   // --------------------------------------------------------------------------
2006   // JVMTI PopFrame support
2007 
2008   Interpreter::_remove_activation_preserving_args_entry = __ pc();
2009   {
2010     // Set the popframe_processing bit in popframe_condition indicating that we are
2011     // currently handling popframe, so that call_VMs that may happen later do not
2012     // trigger new popframe handling cycles.
2013     __ lwz(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2014     __ ori(R11_scratch1, R11_scratch1, JavaThread::popframe_processing_bit);
2015     __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2016 
2017     // Empty the expression stack, as in normal exception handling.
2018     __ empty_expression_stack();
2019     __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
2020 
2021     // Check to see whether we are returning to a deoptimized frame.
2022     // (The PopFrame call ensures that the caller of the popped frame is
2023     // either interpreted or compiled and deoptimizes it if compiled.)
2024     // Note that we don't compare the return PC against the
2025     // deoptimization blob's unpack entry because of the presence of
2026     // adapter frames in C2.
2027     Label Lcaller_not_deoptimized;
2028     Register return_pc = R3_ARG1;
2029     __ ld(return_pc, 0, R1_SP);
2030     __ ld(return_pc, _abi(lr), return_pc);
2031     __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), return_pc);
2032     __ cmpdi(CCR0, R3_RET, 0);
2033     __ bne(CCR0, Lcaller_not_deoptimized);
2034 
2035     // The deoptimized case.
2036     // In this case, we can't call dispatch_next() after the frame is
2037     // popped, but instead must save the incoming arguments and restore
2038     // them after deoptimization has occurred.
2039     __ ld(R4_ARG2, in_bytes(Method::const_offset()), R19_method);
2040     __ lhz(R4_ARG2 /* number of params */, in_bytes(ConstMethod::size_of_parameters_offset()), R4_ARG2);
2041     __ slwi(R4_ARG2, R4_ARG2, Interpreter::logStackElementSize);
2042     __ addi(R5_ARG3, R18_locals, Interpreter::stackElementSize);
2043     __ subf(R5_ARG3, R4_ARG2, R5_ARG3);
2044     // Save these arguments.
2045     __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), R16_thread, R4_ARG2, R5_ARG3);
2046 
2047     // Inform deoptimization that it is responsible for restoring these arguments.
2048     __ load_const_optimized(R11_scratch1, JavaThread::popframe_force_deopt_reexecution_bit);
2049     __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2050 
2051     // Return from the current method into the deoptimization blob. Will eventually
2052     // end up in the deopt interpeter entry, deoptimization prepared everything that
2053     // we will reexecute the call that called us.
2054     __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*reload return_pc*/ return_pc, R11_scratch1, R12_scratch2);
2055     __ mtlr(return_pc);
2056     __ blr();
2057 
2058     // The non-deoptimized case.
2059     __ bind(Lcaller_not_deoptimized);
2060 
2061     // Clear the popframe condition flag.
2062     __ li(R0, 0);
2063     __ stw(R0, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2064 
2065     // Get out of the current method and re-execute the call that called us.
2066     __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
2067     __ restore_interpreter_state(R11_scratch1);
2068     __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
2069     __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
2070     if (ProfileInterpreter) {
2071       __ set_method_data_pointer_for_bcp();
2072       __ ld(R11_scratch1, 0, R1_SP);
2073       __ std(R28_mdx, _ijava_state_neg(mdx), R11_scratch1);
2074     }
2075 #if INCLUDE_JVMTI
2076     Label L_done;
2077 
2078     __ lbz(R11_scratch1, 0, R14_bcp);
2079     __ cmpwi(CCR0, R11_scratch1, Bytecodes::_invokestatic);
2080     __ bne(CCR0, L_done);
2081 
2082     // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
2083     // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
2084     __ ld(R4_ARG2, 0, R18_locals);
2085     __ MacroAssembler::call_VM(R4_ARG2, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), R4_ARG2, R19_method, R14_bcp, false);
2086     __ restore_interpreter_state(R11_scratch1, /*bcp_and_mdx_only*/ true);
2087     __ cmpdi(CCR0, R4_ARG2, 0);
2088     __ beq(CCR0, L_done);
2089     __ std(R4_ARG2, wordSize, R15_esp);
2090     __ bind(L_done);
2091 #endif // INCLUDE_JVMTI
2092     __ dispatch_next(vtos);
2093   }
2094   // end of JVMTI PopFrame support
2095 
2096   // --------------------------------------------------------------------------
2097   // Remove activation exception entry.
2098   // This is jumped to if an interpreted method can't handle an exception itself
2099   // (we come from the throw/rethrow exception entry above). We're going to call
2100   // into the VM to find the exception handler in the caller, pop the current
2101   // frame and return the handler we calculated.
2102   Interpreter::_remove_activation_entry = __ pc();
2103   {
2104     __ pop_ptr(Rexception);
2105     __ verify_thread();
2106     __ verify_oop(Rexception);
2107     __ std(Rexception, in_bytes(JavaThread::vm_result_offset()), R16_thread);
2108 
2109     __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, true);
2110     __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI, false);
2111 
2112     __ get_vm_result(Rexception);
2113 
2114     // We are done with this activation frame; find out where to go next.
2115     // The continuation point will be an exception handler, which expects
2116     // the following registers set up:
2117     //
2118     // RET:  exception oop
2119     // ARG2: Issuing PC (see generate_exception_blob()), only used if the caller is compiled.
2120 
2121     Register return_pc = R31; // Needs to survive the runtime call.
2122     __ ld(return_pc, 0, R1_SP);
2123     __ ld(return_pc, _abi(lr), return_pc);
2124     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, return_pc);
2125 
2126     // Remove the current activation.
2127     __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
2128 
2129     __ mr(R4_ARG2, return_pc);
2130     __ mtlr(R3_RET);
2131     __ mr(R3_RET, Rexception);
2132     __ blr();
2133   }
2134 }
2135 
2136 // JVMTI ForceEarlyReturn support.
2137 // Returns "in the middle" of a method with a "fake" return value.
2138 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
2139 
2140   Register Rscratch1 = R11_scratch1,
2141            Rscratch2 = R12_scratch2;
2142 
2143   address entry = __ pc();
2144   __ empty_expression_stack();
2145 
2146   __ load_earlyret_value(state, Rscratch1);
2147 
2148   __ ld(Rscratch1, in_bytes(JavaThread::jvmti_thread_state_offset()), R16_thread);
2149   // Clear the earlyret state.
2150   __ li(R0, 0);
2151   __ stw(R0, in_bytes(JvmtiThreadState::earlyret_state_offset()), Rscratch1);
2152 
2153   __ remove_activation(state, false, false);
2154   // Copied from TemplateTable::_return.
2155   // Restoration of lr done by remove_activation.
2156   switch (state) {
2157     // Narrow result if state is itos but result type is smaller.
2158     case itos: __ narrow(R17_tos); /* fall through */
2159     case ltos:
2160     case btos:
2161     case ztos:
2162     case ctos:
2163     case stos:
2164     case atos: __ mr(R3_RET, R17_tos); break;
2165     case ftos:
2166     case dtos: __ fmr(F1_RET, F15_ftos); break;
2167     case vtos: // This might be a constructor. Final fields (and volatile fields on PPC64) need
2168                // to get visible before the reference to the object gets stored anywhere.
2169                __ membar(Assembler::StoreStore); break;
2170     default  : ShouldNotReachHere();
2171   }
2172   __ blr();
2173 
2174   return entry;
2175 } // end of ForceEarlyReturn support
2176 
2177 //-----------------------------------------------------------------------------
2178 // Helper for vtos entry point generation
2179 
2180 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t,
2181                                                          address& bep,
2182                                                          address& cep,
2183                                                          address& sep,
2184                                                          address& aep,
2185                                                          address& iep,
2186                                                          address& lep,
2187                                                          address& fep,
2188                                                          address& dep,
2189                                                          address& vep) {
2190   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2191   Label L;
2192 
2193   aep = __ pc();  __ push_ptr();  __ b(L);
2194   fep = __ pc();  __ push_f();    __ b(L);
2195   dep = __ pc();  __ push_d();    __ b(L);
2196   lep = __ pc();  __ push_l();    __ b(L);
2197   __ align(32, 12, 24); // align L
2198   bep = cep = sep =
2199   iep = __ pc();  __ push_i();
2200   vep = __ pc();
2201   __ bind(L);
2202   generate_and_dispatch(t);
2203 }
2204 
2205 //-----------------------------------------------------------------------------
2206 
2207 // Non-product code
2208 #ifndef PRODUCT
2209 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2210   //__ flush_bundle();
2211   address entry = __ pc();
2212 
2213   const char *bname = NULL;
2214   uint tsize = 0;
2215   switch(state) {
2216   case ftos:
2217     bname = "trace_code_ftos {";
2218     tsize = 2;
2219     break;
2220   case btos:
2221     bname = "trace_code_btos {";
2222     tsize = 2;
2223     break;
2224   case ztos:
2225     bname = "trace_code_ztos {";
2226     tsize = 2;
2227     break;
2228   case ctos:
2229     bname = "trace_code_ctos {";
2230     tsize = 2;
2231     break;
2232   case stos:
2233     bname = "trace_code_stos {";
2234     tsize = 2;
2235     break;
2236   case itos:
2237     bname = "trace_code_itos {";
2238     tsize = 2;
2239     break;
2240   case ltos:
2241     bname = "trace_code_ltos {";
2242     tsize = 3;
2243     break;
2244   case atos:
2245     bname = "trace_code_atos {";
2246     tsize = 2;
2247     break;
2248   case vtos:
2249     // Note: In case of vtos, the topmost of stack value could be a int or doubl
2250     // In case of a double (2 slots) we won't see the 2nd stack value.
2251     // Maybe we simply should print the topmost 3 stack slots to cope with the problem.
2252     bname = "trace_code_vtos {";
2253     tsize = 2;
2254 
2255     break;
2256   case dtos:
2257     bname = "trace_code_dtos {";
2258     tsize = 3;
2259     break;
2260   default:
2261     ShouldNotReachHere();
2262   }
2263   BLOCK_COMMENT(bname);
2264 
2265   // Support short-cut for TraceBytecodesAt.
2266   // Don't call into the VM if we don't want to trace to speed up things.
2267   Label Lskip_vm_call;
2268   if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
2269     int offs1 = __ load_const_optimized(R11_scratch1, (address) &TraceBytecodesAt, R0, true);
2270     int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
2271     __ ld(R11_scratch1, offs1, R11_scratch1);
2272     __ lwa(R12_scratch2, offs2, R12_scratch2);
2273     __ cmpd(CCR0, R12_scratch2, R11_scratch1);
2274     __ blt(CCR0, Lskip_vm_call);
2275   }
2276 
2277   __ push(state);
2278   // Load 2 topmost expression stack values.
2279   __ ld(R6_ARG4, tsize*Interpreter::stackElementSize, R15_esp);
2280   __ ld(R5_ARG3, Interpreter::stackElementSize, R15_esp);
2281   __ mflr(R31);
2282   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), /* unused */ R4_ARG2, R5_ARG3, R6_ARG4, false);
2283   __ mtlr(R31);
2284   __ pop(state);
2285 
2286   if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
2287     __ bind(Lskip_vm_call);
2288   }
2289   __ blr();
2290   BLOCK_COMMENT("} trace_code");
2291   return entry;
2292 }
2293 
2294 void TemplateInterpreterGenerator::count_bytecode() {
2295   int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeCounter::_counter_value, R12_scratch2, true);
2296   __ lwz(R12_scratch2, offs, R11_scratch1);
2297   __ addi(R12_scratch2, R12_scratch2, 1);
2298   __ stw(R12_scratch2, offs, R11_scratch1);
2299 }
2300 
2301 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2302   int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeHistogram::_counters[t->bytecode()], R12_scratch2, true);
2303   __ lwz(R12_scratch2, offs, R11_scratch1);
2304   __ addi(R12_scratch2, R12_scratch2, 1);
2305   __ stw(R12_scratch2, offs, R11_scratch1);
2306 }
2307 
2308 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2309   const Register addr = R11_scratch1,
2310                  tmp  = R12_scratch2;
2311   // Get index, shift out old bytecode, bring in new bytecode, and store it.
2312   // _index = (_index >> log2_number_of_codes) |
2313   //          (bytecode << log2_number_of_codes);
2314   int offs1 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_index, tmp, true);
2315   __ lwz(tmp, offs1, addr);
2316   __ srwi(tmp, tmp, BytecodePairHistogram::log2_number_of_codes);
2317   __ ori(tmp, tmp, ((int) t->bytecode()) << BytecodePairHistogram::log2_number_of_codes);
2318   __ stw(tmp, offs1, addr);
2319 
2320   // Bump bucket contents.
2321   // _counters[_index] ++;
2322   int offs2 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_counters, R0, true);
2323   __ sldi(tmp, tmp, LogBytesPerInt);
2324   __ add(addr, tmp, addr);
2325   __ lwz(tmp, offs2, addr);
2326   __ addi(tmp, tmp, 1);
2327   __ stw(tmp, offs2, addr);
2328 }
2329 
2330 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2331   // Call a little run-time stub to avoid blow-up for each bytecode.
2332   // The run-time runtime saves the right registers, depending on
2333   // the tosca in-state for the given template.
2334 
2335   assert(Interpreter::trace_code(t->tos_in()) != NULL,
2336          "entry must have been generated");
2337 
2338   // Note: we destroy LR here.
2339   __ bl(Interpreter::trace_code(t->tos_in()));
2340 }
2341 
2342 void TemplateInterpreterGenerator::stop_interpreter_at() {
2343   Label L;
2344   int offs1 = __ load_const_optimized(R11_scratch1, (address) &StopInterpreterAt, R0, true);
2345   int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
2346   __ ld(R11_scratch1, offs1, R11_scratch1);
2347   __ lwa(R12_scratch2, offs2, R12_scratch2);
2348   __ cmpd(CCR0, R12_scratch2, R11_scratch1);
2349   __ bne(CCR0, L);
2350   __ illtrap();
2351   __ bind(L);
2352 }
2353 
2354 #endif // !PRODUCT