1 /*
   2  * Copyright (c) 2014, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2015, 2017, 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 = 256*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     // JNIHandles::resolve result.
 405     __ resolve_jobject(R3_RET, R11_scratch1, R12_scratch2, /* needs_frame */ true); // kills R31
 406     break;
 407   case T_FLOAT:
 408      break;
 409   case T_DOUBLE:
 410      break;
 411   case T_VOID:
 412      break;
 413   default: ShouldNotReachHere();
 414   }
 415 
 416   BIND(done);
 417   __ blr();
 418 
 419   return entry;
 420 }
 421 
 422 // Abstract method entry.
 423 //
 424 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
 425   address entry = __ pc();
 426 
 427   //
 428   // Registers alive
 429   //   R16_thread     - JavaThread*
 430   //   R19_method     - callee's method (method to be invoked)
 431   //   R1_SP          - SP prepared such that caller's outgoing args are near top
 432   //   LR             - return address to caller
 433   //
 434   // Stack layout at this point:
 435   //
 436   //   0       [TOP_IJAVA_FRAME_ABI]         <-- R1_SP
 437   //           alignment (optional)
 438   //           [outgoing Java arguments]
 439   //           ...
 440   //   PARENT  [PARENT_IJAVA_FRAME_ABI]
 441   //            ...
 442   //
 443 
 444   // Can't use call_VM here because we have not set up a new
 445   // interpreter state. Make the call to the vm and make it look like
 446   // our caller set up the JavaFrameAnchor.
 447   __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
 448 
 449   // Push a new C frame and save LR.
 450   __ save_LR_CR(R0);
 451   __ push_frame_reg_args(0, R11_scratch1);
 452 
 453   // This is not a leaf but we have a JavaFrameAnchor now and we will
 454   // check (create) exceptions afterward so this is ok.
 455   __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError),
 456                   R16_thread);
 457 
 458   // Pop the C frame and restore LR.
 459   __ pop_frame();
 460   __ restore_LR_CR(R0);
 461 
 462   // Reset JavaFrameAnchor from call_VM_leaf above.
 463   __ reset_last_Java_frame();
 464 
 465   // We don't know our caller, so jump to the general forward exception stub,
 466   // which will also pop our full frame off. Satisfy the interface of
 467   // SharedRuntime::generate_forward_exception()
 468   __ load_const_optimized(R11_scratch1, StubRoutines::forward_exception_entry(), R0);
 469   __ mtctr(R11_scratch1);
 470   __ bctr();
 471 
 472   return entry;
 473 }
 474 
 475 // Interpreter intrinsic for WeakReference.get().
 476 // 1. Don't push a full blown frame and go on dispatching, but fetch the value
 477 //    into R8 and return quickly
 478 // 2. If G1 is active we *must* execute this intrinsic for corrrectness:
 479 //    It contains a GC barrier which puts the reference into the satb buffer
 480 //    to indicate that someone holds a strong reference to the object the
 481 //    weak ref points to!
 482 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
 483   // Code: _aload_0, _getfield, _areturn
 484   // parameter size = 1
 485   //
 486   // The code that gets generated by this routine is split into 2 parts:
 487   //    1. the "intrinsified" code for G1 (or any SATB based GC),
 488   //    2. the slow path - which is an expansion of the regular method entry.
 489   //
 490   // Notes:
 491   // * In the G1 code we do not check whether we need to block for
 492   //   a safepoint. If G1 is enabled then we must execute the specialized
 493   //   code for Reference.get (except when the Reference object is null)
 494   //   so that we can log the value in the referent field with an SATB
 495   //   update buffer.
 496   //   If the code for the getfield template is modified so that the
 497   //   G1 pre-barrier code is executed when the current method is
 498   //   Reference.get() then going through the normal method entry
 499   //   will be fine.
 500   // * The G1 code can, however, check the receiver object (the instance
 501   //   of java.lang.Reference) and jump to the slow path if null. If the
 502   //   Reference object is null then we obviously cannot fetch the referent
 503   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 504   //   regular method entry code to generate the NPE.
 505   //
 506 
 507   if (UseG1GC) {
 508     address entry = __ pc();
 509 
 510     const int referent_offset = java_lang_ref_Reference::referent_offset;
 511     guarantee(referent_offset > 0, "referent offset not initialized");
 512 
 513     Label slow_path;
 514 
 515     // Debugging not possible, so can't use __ skip_if_jvmti_mode(slow_path, GR31_SCRATCH);
 516 
 517     // In the G1 code we don't check if we need to reach a safepoint. We
 518     // continue and the thread will safepoint at the next bytecode dispatch.
 519 
 520     // If the receiver is null then it is OK to jump to the slow path.
 521     __ ld(R3_RET, Interpreter::stackElementSize, R15_esp); // get receiver
 522 
 523     // Check if receiver == NULL and go the slow path.
 524     __ cmpdi(CCR0, R3_RET, 0);
 525     __ beq(CCR0, slow_path);
 526 
 527     // Load the value of the referent field.
 528     __ load_heap_oop(R3_RET, referent_offset, R3_RET);
 529 
 530     // Generate the G1 pre-barrier code to log the value of
 531     // the referent field in an SATB buffer. Note with
 532     // these parameters the pre-barrier does not generate
 533     // the load of the previous value.
 534 
 535     // Restore caller sp for c2i case.
 536 #ifdef ASSERT
 537       __ ld(R9_ARG7, 0, R1_SP);
 538       __ ld(R10_ARG8, 0, R21_sender_SP);
 539       __ cmpd(CCR0, R9_ARG7, R10_ARG8);
 540       __ asm_assert_eq("backlink", 0x544);
 541 #endif // ASSERT
 542     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
 543 
 544     __ g1_write_barrier_pre(noreg,         // obj
 545                             noreg,         // offset
 546                             R3_RET,        // pre_val
 547                             R11_scratch1,  // tmp
 548                             R12_scratch2,  // tmp
 549                             true);         // needs_frame
 550 
 551     __ blr();
 552 
 553     // Generate regular method entry.
 554     __ bind(slow_path);
 555     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R11_scratch1);
 556     return entry;
 557   }
 558 
 559   return NULL;
 560 }
 561 
 562 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 563   address entry = __ pc();
 564 
 565   // Expression stack must be empty before entering the VM if an
 566   // exception happened.
 567   __ empty_expression_stack();
 568   // Throw exception.
 569   __ call_VM(noreg,
 570              CAST_FROM_FN_PTR(address,
 571                               InterpreterRuntime::throw_StackOverflowError));
 572   return entry;
 573 }
 574 
 575 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 576   address entry = __ pc();
 577   __ empty_expression_stack();
 578   __ load_const_optimized(R4_ARG2, (address) name);
 579   // Index is in R17_tos.
 580   __ mr(R5_ARG3, R17_tos);
 581   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException));
 582   return entry;
 583 }
 584 
 585 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 586   address entry = __ pc();
 587   // Expression stack must be empty before entering the VM if an
 588   // exception happened.
 589   __ empty_expression_stack();
 590 
 591   // Load exception object.
 592   // Thread will be loaded to R3_ARG1.
 593   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException), R17_tos);
 594 #ifdef ASSERT
 595   // Above call must not return here since exception pending.
 596   __ should_not_reach_here();
 597 #endif
 598   return entry;
 599 }
 600 
 601 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
 602   address entry = __ pc();
 603   //__ untested("generate_exception_handler_common");
 604   Register Rexception = R17_tos;
 605 
 606   // Expression stack must be empty before entering the VM if an exception happened.
 607   __ empty_expression_stack();
 608 
 609   __ load_const_optimized(R4_ARG2, (address) name, R11_scratch1);
 610   if (pass_oop) {
 611     __ mr(R5_ARG3, Rexception);
 612     __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), false);
 613   } else {
 614     __ load_const_optimized(R5_ARG3, (address) message, R11_scratch1);
 615     __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), false);
 616   }
 617 
 618   // Throw exception.
 619   __ mr(R3_ARG1, Rexception);
 620   __ load_const_optimized(R11_scratch1, Interpreter::throw_exception_entry(), R12_scratch2);
 621   __ mtctr(R11_scratch1);
 622   __ bctr();
 623 
 624   return entry;
 625 }
 626 
 627 // This entry is returned to when a call returns to the interpreter.
 628 // When we arrive here, we expect that the callee stack frame is already popped.
 629 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
 630   address entry = __ pc();
 631 
 632   // Move the value out of the return register back to the TOS cache of current frame.
 633   switch (state) {
 634     case ltos:
 635     case btos:
 636     case ztos:
 637     case ctos:
 638     case stos:
 639     case atos:
 640     case itos: __ mr(R17_tos, R3_RET); break;   // RET -> TOS cache
 641     case ftos:
 642     case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
 643     case vtos: break;                           // Nothing to do, this was a void return.
 644     default  : ShouldNotReachHere();
 645   }
 646 
 647   __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
 648   __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
 649   __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
 650 
 651   // Compiled code destroys templateTableBase, reload.
 652   __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R12_scratch2);
 653 
 654   if (state == atos) {
 655     __ profile_return_type(R3_RET, R11_scratch1, R12_scratch2);
 656   }
 657 
 658   const Register cache = R11_scratch1;
 659   const Register size  = R12_scratch2;
 660   __ get_cache_and_index_at_bcp(cache, 1, index_size);
 661 
 662   // Get least significant byte of 64 bit value:
 663 #if defined(VM_LITTLE_ENDIAN)
 664   __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()), cache);
 665 #else
 666   __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()) + 7, cache);
 667 #endif
 668   __ sldi(size, size, Interpreter::logStackElementSize);
 669   __ add(R15_esp, R15_esp, size);
 670 
 671  __ check_and_handle_popframe(R11_scratch1);
 672  __ check_and_handle_earlyret(R11_scratch1);
 673 
 674   __ dispatch_next(state, step);
 675   return entry;
 676 }
 677 
 678 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step, address continuation) {
 679   address entry = __ pc();
 680   // If state != vtos, we're returning from a native method, which put it's result
 681   // into the result register. So move the value out of the return register back
 682   // to the TOS cache of current frame.
 683 
 684   switch (state) {
 685     case ltos:
 686     case btos:
 687     case ztos:
 688     case ctos:
 689     case stos:
 690     case atos:
 691     case itos: __ mr(R17_tos, R3_RET); break;   // GR_RET -> TOS cache
 692     case ftos:
 693     case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
 694     case vtos: break;                           // Nothing to do, this was a void return.
 695     default  : ShouldNotReachHere();
 696   }
 697 
 698   // Load LcpoolCache @@@ should be already set!
 699   __ get_constant_pool_cache(R27_constPoolCache);
 700 
 701   // Handle a pending exception, fall through if none.
 702   __ check_and_forward_exception(R11_scratch1, R12_scratch2);
 703 
 704   // Start executing bytecodes.
 705   if (continuation == NULL) {
 706     __ dispatch_next(state, step);
 707   } else {
 708     __ jump_to_entry(continuation, R11_scratch1);
 709   }
 710 
 711   return entry;
 712 }
 713 
 714 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 715   address entry = __ pc();
 716 
 717   __ push(state);
 718   __ call_VM(noreg, runtime_entry);
 719   __ dispatch_via(vtos, Interpreter::_normal_table.table_for(vtos));
 720 
 721   return entry;
 722 }
 723 
 724 // Helpers for commoning out cases in the various type of method entries.
 725 
 726 // Increment invocation count & check for overflow.
 727 //
 728 // Note: checking for negative value instead of overflow
 729 //       so we have a 'sticky' overflow test.
 730 //
 731 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 732   // Note: In tiered we increment either counters in method or in MDO depending if we're profiling or not.
 733   Register Rscratch1   = R11_scratch1;
 734   Register Rscratch2   = R12_scratch2;
 735   Register R3_counters = R3_ARG1;
 736   Label done;
 737 
 738   if (TieredCompilation) {
 739     const int increment = InvocationCounter::count_increment;
 740     Label no_mdo;
 741     if (ProfileInterpreter) {
 742       const Register Rmdo = R3_counters;
 743       // If no method data exists, go to profile_continue.
 744       __ ld(Rmdo, in_bytes(Method::method_data_offset()), R19_method);
 745       __ cmpdi(CCR0, Rmdo, 0);
 746       __ beq(CCR0, no_mdo);
 747 
 748       // Increment invocation counter in the MDO.
 749       const int mdo_ic_offs = in_bytes(MethodData::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
 750       __ lwz(Rscratch2, mdo_ic_offs, Rmdo);
 751       __ lwz(Rscratch1, in_bytes(MethodData::invoke_mask_offset()), Rmdo);
 752       __ addi(Rscratch2, Rscratch2, increment);
 753       __ stw(Rscratch2, mdo_ic_offs, Rmdo);
 754       __ and_(Rscratch1, Rscratch2, Rscratch1);
 755       __ bne(CCR0, done);
 756       __ b(*overflow);
 757     }
 758 
 759     // Increment counter in MethodCounters*.
 760     const int mo_ic_offs = in_bytes(MethodCounters::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
 761     __ bind(no_mdo);
 762     __ get_method_counters(R19_method, R3_counters, done);
 763     __ lwz(Rscratch2, mo_ic_offs, R3_counters);
 764     __ lwz(Rscratch1, in_bytes(MethodCounters::invoke_mask_offset()), R3_counters);
 765     __ addi(Rscratch2, Rscratch2, increment);
 766     __ stw(Rscratch2, mo_ic_offs, R3_counters);
 767     __ and_(Rscratch1, Rscratch2, Rscratch1);
 768     __ beq(CCR0, *overflow);
 769 
 770     __ bind(done);
 771 
 772   } else {
 773 
 774     // Update standard invocation counters.
 775     Register Rsum_ivc_bec = R4_ARG2;
 776     __ get_method_counters(R19_method, R3_counters, done);
 777     __ increment_invocation_counter(R3_counters, Rsum_ivc_bec, R12_scratch2);
 778     // Increment interpreter invocation counter.
 779     if (ProfileInterpreter) {  // %%% Merge this into methodDataOop.
 780       __ lwz(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
 781       __ addi(R12_scratch2, R12_scratch2, 1);
 782       __ stw(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
 783     }
 784     // Check if we must create a method data obj.
 785     if (ProfileInterpreter && profile_method != NULL) {
 786       const Register profile_limit = Rscratch1;
 787       __ lwz(profile_limit, in_bytes(MethodCounters::interpreter_profile_limit_offset()), R3_counters);
 788       // Test to see if we should create a method data oop.
 789       __ cmpw(CCR0, Rsum_ivc_bec, profile_limit);
 790       __ blt(CCR0, *profile_method_continue);
 791       // If no method data exists, go to profile_method.
 792       __ test_method_data_pointer(*profile_method);
 793     }
 794     // Finally check for counter overflow.
 795     if (overflow) {
 796       const Register invocation_limit = Rscratch1;
 797       __ lwz(invocation_limit, in_bytes(MethodCounters::interpreter_invocation_limit_offset()), R3_counters);
 798       __ cmpw(CCR0, Rsum_ivc_bec, invocation_limit);
 799       __ bge(CCR0, *overflow);
 800     }
 801 
 802     __ bind(done);
 803   }
 804 }
 805 
 806 // Generate code to initiate compilation on invocation counter overflow.
 807 void TemplateInterpreterGenerator::generate_counter_overflow(Label& continue_entry) {
 808   // Generate code to initiate compilation on the counter overflow.
 809 
 810   // InterpreterRuntime::frequency_counter_overflow takes one arguments,
 811   // which indicates if the counter overflow occurs at a backwards branch (NULL bcp)
 812   // We pass zero in.
 813   // The call returns the address of the verified entry point for the method or NULL
 814   // if the compilation did not complete (either went background or bailed out).
 815   //
 816   // Unlike the C++ interpreter above: Check exceptions!
 817   // Assumption: Caller must set the flag "do_not_unlock_if_sychronized" if the monitor of a sync'ed
 818   // method has not yet been created. Thus, no unlocking of a non-existing monitor can occur.
 819 
 820   __ li(R4_ARG2, 0);
 821   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), R4_ARG2, true);
 822 
 823   // Returns verified_entry_point or NULL.
 824   // We ignore it in any case.
 825   __ b(continue_entry);
 826 }
 827 
 828 // See if we've got enough room on the stack for locals plus overhead below
 829 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
 830 // without going through the signal handler, i.e., reserved and yellow zones
 831 // will not be made usable. The shadow zone must suffice to handle the
 832 // overflow.
 833 //
 834 // Kills Rmem_frame_size, Rscratch1.
 835 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rmem_frame_size, Register Rscratch1) {
 836   Label done;
 837   assert_different_registers(Rmem_frame_size, Rscratch1);
 838 
 839   BLOCK_COMMENT("stack_overflow_check_with_compare {");
 840   __ sub(Rmem_frame_size, R1_SP, Rmem_frame_size);
 841   __ ld(Rscratch1, thread_(stack_overflow_limit));
 842   __ cmpld(CCR0/*is_stack_overflow*/, Rmem_frame_size, Rscratch1);
 843   __ bgt(CCR0/*is_stack_overflow*/, done);
 844 
 845   // The stack overflows. Load target address of the runtime stub and call it.
 846   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order");
 847   __ load_const_optimized(Rscratch1, (StubRoutines::throw_StackOverflowError_entry()), R0);
 848   __ mtctr(Rscratch1);
 849   // Restore caller_sp.
 850 #ifdef ASSERT
 851   __ ld(Rscratch1, 0, R1_SP);
 852   __ ld(R0, 0, R21_sender_SP);
 853   __ cmpd(CCR0, R0, Rscratch1);
 854   __ asm_assert_eq("backlink", 0x547);
 855 #endif // ASSERT
 856   __ mr(R1_SP, R21_sender_SP);
 857   __ bctr();
 858 
 859   __ align(32, 12);
 860   __ bind(done);
 861   BLOCK_COMMENT("} stack_overflow_check_with_compare");
 862 }
 863 
 864 // Lock the current method, interpreter register window must be set up!
 865 void TemplateInterpreterGenerator::lock_method(Register Rflags, Register Rscratch1, Register Rscratch2, bool flags_preloaded) {
 866   const Register Robj_to_lock = Rscratch2;
 867 
 868   {
 869     if (!flags_preloaded) {
 870       __ lwz(Rflags, method_(access_flags));
 871     }
 872 
 873 #ifdef ASSERT
 874     // Check if methods needs synchronization.
 875     {
 876       Label Lok;
 877       __ testbitdi(CCR0, R0, Rflags, JVM_ACC_SYNCHRONIZED_BIT);
 878       __ btrue(CCR0,Lok);
 879       __ stop("method doesn't need synchronization");
 880       __ bind(Lok);
 881     }
 882 #endif // ASSERT
 883   }
 884 
 885   // Get synchronization object to Rscratch2.
 886   {
 887     Label Lstatic;
 888     Label Ldone;
 889 
 890     __ testbitdi(CCR0, R0, Rflags, JVM_ACC_STATIC_BIT);
 891     __ btrue(CCR0, Lstatic);
 892 
 893     // Non-static case: load receiver obj from stack and we're done.
 894     __ ld(Robj_to_lock, R18_locals);
 895     __ b(Ldone);
 896 
 897     __ bind(Lstatic); // Static case: Lock the java mirror
 898     // Load mirror from interpreter frame.
 899     __ ld(Robj_to_lock, _abi(callers_sp), R1_SP);
 900     __ ld(Robj_to_lock, _ijava_state_neg(mirror), Robj_to_lock);
 901 
 902     __ bind(Ldone);
 903     __ verify_oop(Robj_to_lock);
 904   }
 905 
 906   // Got the oop to lock => execute!
 907   __ add_monitor_to_stack(true, Rscratch1, R0);
 908 
 909   __ std(Robj_to_lock, BasicObjectLock::obj_offset_in_bytes(), R26_monitor);
 910   __ lock_object(R26_monitor, Robj_to_lock);
 911 }
 912 
 913 // Generate a fixed interpreter frame for pure interpreter
 914 // and I2N native transition frames.
 915 //
 916 // Before (stack grows downwards):
 917 //
 918 //         |  ...         |
 919 //         |------------- |
 920 //         |  java arg0   |
 921 //         |  ...         |
 922 //         |  java argn   |
 923 //         |              |   <-   R15_esp
 924 //         |              |
 925 //         |--------------|
 926 //         | abi_112      |
 927 //         |              |   <-   R1_SP
 928 //         |==============|
 929 //
 930 //
 931 // After:
 932 //
 933 //         |  ...         |
 934 //         |  java arg0   |<-   R18_locals
 935 //         |  ...         |
 936 //         |  java argn   |
 937 //         |--------------|
 938 //         |              |
 939 //         |  java locals |
 940 //         |              |
 941 //         |--------------|
 942 //         |  abi_48      |
 943 //         |==============|
 944 //         |              |
 945 //         |   istate     |
 946 //         |              |
 947 //         |--------------|
 948 //         |   monitor    |<-   R26_monitor
 949 //         |--------------|
 950 //         |              |<-   R15_esp
 951 //         | expression   |
 952 //         | stack        |
 953 //         |              |
 954 //         |--------------|
 955 //         |              |
 956 //         | abi_112      |<-   R1_SP
 957 //         |==============|
 958 //
 959 // The top most frame needs an abi space of 112 bytes. This space is needed,
 960 // since we call to c. The c function may spill their arguments to the caller
 961 // frame. When we call to java, we don't need these spill slots. In order to save
 962 // space on the stack, we resize the caller. However, java locals reside in
 963 // the caller frame and the frame has to be increased. The frame_size for the
 964 // current frame was calculated based on max_stack as size for the expression
 965 // stack. At the call, just a part of the expression stack might be used.
 966 // We don't want to waste this space and cut the frame back accordingly.
 967 // The resulting amount for resizing is calculated as follows:
 968 // resize =   (number_of_locals - number_of_arguments) * slot_size
 969 //          + (R1_SP - R15_esp) + 48
 970 //
 971 // The size for the callee frame is calculated:
 972 // framesize = 112 + max_stack + monitor + state_size
 973 //
 974 // maxstack:   Max number of slots on the expression stack, loaded from the method.
 975 // monitor:    We statically reserve room for one monitor object.
 976 // state_size: We save the current state of the interpreter to this area.
 977 //
 978 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call, Register Rsize_of_parameters, Register Rsize_of_locals) {
 979   Register parent_frame_resize = R6_ARG4, // Frame will grow by this number of bytes.
 980            top_frame_size      = R7_ARG5,
 981            Rconst_method       = R8_ARG6;
 982 
 983   assert_different_registers(Rsize_of_parameters, Rsize_of_locals, parent_frame_resize, top_frame_size);
 984 
 985   __ ld(Rconst_method, method_(const));
 986   __ lhz(Rsize_of_parameters /* number of params */,
 987          in_bytes(ConstMethod::size_of_parameters_offset()), Rconst_method);
 988   if (native_call) {
 989     // If we're calling a native method, we reserve space for the worst-case signature
 990     // handler varargs vector, which is max(Argument::n_register_parameters, parameter_count+2).
 991     // We add two slots to the parameter_count, one for the jni
 992     // environment and one for a possible native mirror.
 993     Label skip_native_calculate_max_stack;
 994     __ addi(top_frame_size, Rsize_of_parameters, 2);
 995     __ cmpwi(CCR0, top_frame_size, Argument::n_register_parameters);
 996     __ bge(CCR0, skip_native_calculate_max_stack);
 997     __ li(top_frame_size, Argument::n_register_parameters);
 998     __ bind(skip_native_calculate_max_stack);
 999     __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
1000     __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
1001     __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
1002     assert(Rsize_of_locals == noreg, "Rsize_of_locals not initialized"); // Only relevant value is Rsize_of_parameters.
1003   } else {
1004     __ lhz(Rsize_of_locals /* number of params */, in_bytes(ConstMethod::size_of_locals_offset()), Rconst_method);
1005     __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
1006     __ sldi(Rsize_of_locals, Rsize_of_locals, Interpreter::logStackElementSize);
1007     __ lhz(top_frame_size, in_bytes(ConstMethod::max_stack_offset()), Rconst_method);
1008     __ sub(R11_scratch1, Rsize_of_locals, Rsize_of_parameters); // >=0
1009     __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
1010     __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
1011     __ add(parent_frame_resize, parent_frame_resize, R11_scratch1);
1012   }
1013 
1014   // Compute top frame size.
1015   __ addi(top_frame_size, top_frame_size, frame::abi_reg_args_size + frame::ijava_state_size);
1016 
1017   // Cut back area between esp and max_stack.
1018   __ addi(parent_frame_resize, parent_frame_resize, frame::abi_minframe_size - Interpreter::stackElementSize);
1019 
1020   __ round_to(top_frame_size, frame::alignment_in_bytes);
1021   __ round_to(parent_frame_resize, frame::alignment_in_bytes);
1022   // parent_frame_resize = (locals-parameters) - (ESP-SP-ABI48) Rounded to frame alignment size.
1023   // Enlarge by locals-parameters (not in case of native_call), shrink by ESP-SP-ABI48.
1024 
1025   if (!native_call) {
1026     // Stack overflow check.
1027     // Native calls don't need the stack size check since they have no
1028     // expression stack and the arguments are already on the stack and
1029     // we only add a handful of words to the stack.
1030     __ add(R11_scratch1, parent_frame_resize, top_frame_size);
1031     generate_stack_overflow_check(R11_scratch1, R12_scratch2);
1032   }
1033 
1034   // Set up interpreter state registers.
1035 
1036   __ add(R18_locals, R15_esp, Rsize_of_parameters);
1037   __ ld(R27_constPoolCache, in_bytes(ConstMethod::constants_offset()), Rconst_method);
1038   __ ld(R27_constPoolCache, ConstantPool::cache_offset_in_bytes(), R27_constPoolCache);
1039 
1040   // Set method data pointer.
1041   if (ProfileInterpreter) {
1042     Label zero_continue;
1043     __ ld(R28_mdx, method_(method_data));
1044     __ cmpdi(CCR0, R28_mdx, 0);
1045     __ beq(CCR0, zero_continue);
1046     __ addi(R28_mdx, R28_mdx, in_bytes(MethodData::data_offset()));
1047     __ bind(zero_continue);
1048   }
1049 
1050   if (native_call) {
1051     __ li(R14_bcp, 0); // Must initialize.
1052   } else {
1053     __ add(R14_bcp, in_bytes(ConstMethod::codes_offset()), Rconst_method);
1054   }
1055 
1056   // Resize parent frame.
1057   __ mflr(R12_scratch2);
1058   __ neg(parent_frame_resize, parent_frame_resize);
1059   __ resize_frame(parent_frame_resize, R11_scratch1);
1060   __ std(R12_scratch2, _abi(lr), R1_SP);
1061 
1062   // Get mirror and store it in the frame as GC root for this Method*.
1063   __ load_mirror_from_const_method(R12_scratch2, Rconst_method);
1064 
1065   __ addi(R26_monitor, R1_SP, - frame::ijava_state_size);
1066   __ addi(R15_esp, R26_monitor, - Interpreter::stackElementSize);
1067 
1068   // Store values.
1069   // R15_esp, R14_bcp, R26_monitor, R28_mdx are saved at java calls
1070   // in InterpreterMacroAssembler::call_from_interpreter.
1071   __ std(R19_method, _ijava_state_neg(method), R1_SP);
1072   __ std(R12_scratch2, _ijava_state_neg(mirror), R1_SP);
1073   __ std(R21_sender_SP, _ijava_state_neg(sender_sp), R1_SP);
1074   __ std(R27_constPoolCache, _ijava_state_neg(cpoolCache), R1_SP);
1075   __ std(R18_locals, _ijava_state_neg(locals), R1_SP);
1076 
1077   // Note: esp, bcp, monitor, mdx live in registers. Hence, the correct version can only
1078   // be found in the frame after save_interpreter_state is done. This is always true
1079   // for non-top frames. But when a signal occurs, dumping the top frame can go wrong,
1080   // because e.g. frame::interpreter_frame_bcp() will not access the correct value
1081   // (Enhanced Stack Trace).
1082   // The signal handler does not save the interpreter state into the frame.
1083   __ li(R0, 0);
1084 #ifdef ASSERT
1085   // Fill remaining slots with constants.
1086   __ load_const_optimized(R11_scratch1, 0x5afe);
1087   __ load_const_optimized(R12_scratch2, 0xdead);
1088 #endif
1089   // We have to initialize some frame slots for native calls (accessed by GC).
1090   if (native_call) {
1091     __ std(R26_monitor, _ijava_state_neg(monitors), R1_SP);
1092     __ std(R14_bcp, _ijava_state_neg(bcp), R1_SP);
1093     if (ProfileInterpreter) { __ std(R28_mdx, _ijava_state_neg(mdx), R1_SP); }
1094   }
1095 #ifdef ASSERT
1096   else {
1097     __ std(R12_scratch2, _ijava_state_neg(monitors), R1_SP);
1098     __ std(R12_scratch2, _ijava_state_neg(bcp), R1_SP);
1099     __ std(R12_scratch2, _ijava_state_neg(mdx), R1_SP);
1100   }
1101   __ std(R11_scratch1, _ijava_state_neg(ijava_reserved), R1_SP);
1102   __ std(R12_scratch2, _ijava_state_neg(esp), R1_SP);
1103   __ std(R12_scratch2, _ijava_state_neg(lresult), R1_SP);
1104   __ std(R12_scratch2, _ijava_state_neg(fresult), R1_SP);
1105 #endif
1106   __ subf(R12_scratch2, top_frame_size, R1_SP);
1107   __ std(R0, _ijava_state_neg(oop_tmp), R1_SP);
1108   __ std(R12_scratch2, _ijava_state_neg(top_frame_sp), R1_SP);
1109 
1110   // Push top frame.
1111   __ push_frame(top_frame_size, R11_scratch1);
1112 }
1113 
1114 // End of helpers
1115 
1116 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1117 
1118   // Decide what to do: Use same platform specific instructions and runtime calls as compilers.
1119   bool use_instruction = false;
1120   address runtime_entry = NULL;
1121   int num_args = 1;
1122   bool double_precision = true;
1123 
1124   // PPC64 specific:
1125   switch (kind) {
1126     case Interpreter::java_lang_math_sqrt: use_instruction = VM_Version::has_fsqrt(); break;
1127     case Interpreter::java_lang_math_abs:  use_instruction = true; break;
1128     case Interpreter::java_lang_math_fmaF:
1129     case Interpreter::java_lang_math_fmaD: use_instruction = UseFMA; break;
1130     default: break; // Fall back to runtime call.
1131   }
1132 
1133   switch (kind) {
1134     case Interpreter::java_lang_math_sin  : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dsin);   break;
1135     case Interpreter::java_lang_math_cos  : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dcos);   break;
1136     case Interpreter::java_lang_math_tan  : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dtan);   break;
1137     case Interpreter::java_lang_math_abs  : /* run interpreted */ break;
1138     case Interpreter::java_lang_math_sqrt : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dsqrt);  break;
1139     case Interpreter::java_lang_math_log  : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dlog);   break;
1140     case Interpreter::java_lang_math_log10: runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dlog10); break;
1141     case Interpreter::java_lang_math_pow  : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dpow); num_args = 2; break;
1142     case Interpreter::java_lang_math_exp  : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dexp);   break;
1143     case Interpreter::java_lang_math_fmaF : /* run interpreted */ num_args = 3; double_precision = false; break;
1144     case Interpreter::java_lang_math_fmaD : /* run interpreted */ num_args = 3; break;
1145     default: ShouldNotReachHere();
1146   }
1147 
1148   // Use normal entry if neither instruction nor runtime call is used.
1149   if (!use_instruction && runtime_entry == NULL) return NULL;
1150 
1151   address entry = __ pc();
1152 
1153   // Load arguments
1154   assert(num_args <= 13, "passed in registers");
1155   if (double_precision) {
1156     int offset = (2 * num_args - 1) * Interpreter::stackElementSize;
1157     for (int i = 0; i < num_args; ++i) {
1158       __ lfd(as_FloatRegister(F1_ARG1->encoding() + i), offset, R15_esp);
1159       offset -= 2 * Interpreter::stackElementSize;
1160     }
1161   } else {
1162     int offset = num_args * Interpreter::stackElementSize;
1163     for (int i = 0; i < num_args; ++i) {
1164       __ lfs(as_FloatRegister(F1_ARG1->encoding() + i), offset, R15_esp);
1165       offset -= Interpreter::stackElementSize;
1166     }
1167   }
1168 
1169   // Pop c2i arguments (if any) off when we return.
1170 #ifdef ASSERT
1171   __ ld(R9_ARG7, 0, R1_SP);
1172   __ ld(R10_ARG8, 0, R21_sender_SP);
1173   __ cmpd(CCR0, R9_ARG7, R10_ARG8);
1174   __ asm_assert_eq("backlink", 0x545);
1175 #endif // ASSERT
1176   __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1177 
1178   if (use_instruction) {
1179     switch (kind) {
1180       case Interpreter::java_lang_math_sqrt: __ fsqrt(F1_RET, F1);          break;
1181       case Interpreter::java_lang_math_abs:  __ fabs(F1_RET, F1);           break;
1182       case Interpreter::java_lang_math_fmaF: __ fmadds(F1_RET, F1, F2, F3); break;
1183       case Interpreter::java_lang_math_fmaD: __ fmadd(F1_RET, F1, F2, F3);  break;
1184       default: ShouldNotReachHere();
1185     }
1186   } else {
1187     // Comment: Can use tail call if the unextended frame is always C ABI compliant:
1188     //__ load_const_optimized(R12_scratch2, runtime_entry, R0);
1189     //__ call_c_and_return_to_caller(R12_scratch2);
1190 
1191     // Push a new C frame and save LR.
1192     __ save_LR_CR(R0);
1193     __ push_frame_reg_args(0, R11_scratch1);
1194 
1195     __ call_VM_leaf(runtime_entry);
1196 
1197     // Pop the C frame and restore LR.
1198     __ pop_frame();
1199     __ restore_LR_CR(R0);
1200   }
1201 
1202   __ blr();
1203 
1204   __ flush();
1205 
1206   return entry;
1207 }
1208 
1209 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1210   // Quick & dirty stack overflow checking: bang the stack & handle trap.
1211   // Note that we do the banging after the frame is setup, since the exception
1212   // handling code expects to find a valid interpreter frame on the stack.
1213   // Doing the banging earlier fails if the caller frame is not an interpreter
1214   // frame.
1215   // (Also, the exception throwing code expects to unlock any synchronized
1216   // method receiever, so do the banging after locking the receiver.)
1217 
1218   // Bang each page in the shadow zone. We can't assume it's been done for
1219   // an interpreter frame with greater than a page of locals, so each page
1220   // needs to be checked.  Only true for non-native.
1221   if (UseStackBanging) {
1222     const int page_size = os::vm_page_size();
1223     const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1224     const int start_page = native_call ? n_shadow_pages : 1;
1225     BLOCK_COMMENT("bang_stack_shadow_pages:");
1226     for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1227       __ bang_stack_with_offset(pages*page_size);
1228     }
1229   }
1230 }
1231 
1232 // Interpreter stub for calling a native method. (asm interpreter)
1233 // This sets up a somewhat different looking stack for calling the
1234 // native method than the typical interpreter frame setup.
1235 //
1236 // On entry:
1237 //   R19_method    - method
1238 //   R16_thread    - JavaThread*
1239 //   R15_esp       - intptr_t* sender tos
1240 //
1241 //   abstract stack (grows up)
1242 //     [  IJava (caller of JNI callee)  ]  <-- ASP
1243 //        ...
1244 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1245 
1246   address entry = __ pc();
1247 
1248   const bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1249 
1250   // -----------------------------------------------------------------------------
1251   // Allocate a new frame that represents the native callee (i2n frame).
1252   // This is not a full-blown interpreter frame, but in particular, the
1253   // following registers are valid after this:
1254   // - R19_method
1255   // - R18_local (points to start of arguments to native function)
1256   //
1257   //   abstract stack (grows up)
1258   //     [  IJava (caller of JNI callee)  ]  <-- ASP
1259   //        ...
1260 
1261   const Register signature_handler_fd = R11_scratch1;
1262   const Register pending_exception    = R0;
1263   const Register result_handler_addr  = R31;
1264   const Register native_method_fd     = R11_scratch1;
1265   const Register access_flags         = R22_tmp2;
1266   const Register active_handles       = R11_scratch1; // R26_monitor saved to state.
1267   const Register sync_state           = R12_scratch2;
1268   const Register sync_state_addr      = sync_state;   // Address is dead after use.
1269   const Register suspend_flags        = R11_scratch1;
1270 
1271   //=============================================================================
1272   // Allocate new frame and initialize interpreter state.
1273 
1274   Label exception_return;
1275   Label exception_return_sync_check;
1276   Label stack_overflow_return;
1277 
1278   // Generate new interpreter state and jump to stack_overflow_return in case of
1279   // a stack overflow.
1280   //generate_compute_interpreter_state(stack_overflow_return);
1281 
1282   Register size_of_parameters = R22_tmp2;
1283 
1284   generate_fixed_frame(true, size_of_parameters, noreg /* unused */);
1285 
1286   //=============================================================================
1287   // Increment invocation counter. On overflow, entry to JNI method
1288   // will be compiled.
1289   Label invocation_counter_overflow, continue_after_compile;
1290   if (inc_counter) {
1291     if (synchronized) {
1292       // Since at this point in the method invocation the exception handler
1293       // would try to exit the monitor of synchronized methods which hasn't
1294       // been entered yet, we set the thread local variable
1295       // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1296       // runtime, exception handling i.e. unlock_if_synchronized_method will
1297       // check this thread local flag.
1298       // This flag has two effects, one is to force an unwind in the topmost
1299       // interpreter frame and not perform an unlock while doing so.
1300       __ li(R0, 1);
1301       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1302     }
1303     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1304 
1305     BIND(continue_after_compile);
1306   }
1307 
1308   bang_stack_shadow_pages(true);
1309 
1310   if (inc_counter) {
1311     // Reset the _do_not_unlock_if_synchronized flag.
1312     if (synchronized) {
1313       __ li(R0, 0);
1314       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1315     }
1316   }
1317 
1318   // access_flags = method->access_flags();
1319   // Load access flags.
1320   assert(access_flags->is_nonvolatile(),
1321          "access_flags must be in a non-volatile register");
1322   // Type check.
1323   assert(4 == sizeof(AccessFlags), "unexpected field size");
1324   __ lwz(access_flags, method_(access_flags));
1325 
1326   // We don't want to reload R19_method and access_flags after calls
1327   // to some helper functions.
1328   assert(R19_method->is_nonvolatile(),
1329          "R19_method must be a non-volatile register");
1330 
1331   // Check for synchronized methods. Must happen AFTER invocation counter
1332   // check, so method is not locked if counter overflows.
1333 
1334   if (synchronized) {
1335     lock_method(access_flags, R11_scratch1, R12_scratch2, true);
1336 
1337     // Update monitor in state.
1338     __ ld(R11_scratch1, 0, R1_SP);
1339     __ std(R26_monitor, _ijava_state_neg(monitors), R11_scratch1);
1340   }
1341 
1342   // jvmti/jvmpi support
1343   __ notify_method_entry();
1344 
1345   //=============================================================================
1346   // Get and call the signature handler.
1347 
1348   __ ld(signature_handler_fd, method_(signature_handler));
1349   Label call_signature_handler;
1350 
1351   __ cmpdi(CCR0, signature_handler_fd, 0);
1352   __ bne(CCR0, call_signature_handler);
1353 
1354   // Method has never been called. Either generate a specialized
1355   // handler or point to the slow one.
1356   //
1357   // Pass parameter 'false' to avoid exception check in call_VM.
1358   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false);
1359 
1360   // Check for an exception while looking up the target method. If we
1361   // incurred one, bail.
1362   __ ld(pending_exception, thread_(pending_exception));
1363   __ cmpdi(CCR0, pending_exception, 0);
1364   __ bne(CCR0, exception_return_sync_check); // Has pending exception.
1365 
1366   // Reload signature handler, it may have been created/assigned in the meanwhile.
1367   __ ld(signature_handler_fd, method_(signature_handler));
1368   __ twi_0(signature_handler_fd); // Order wrt. load of klass mirror and entry point (isync is below).
1369 
1370   BIND(call_signature_handler);
1371 
1372   // Before we call the signature handler we push a new frame to
1373   // protect the interpreter frame volatile registers when we return
1374   // from jni but before we can get back to Java.
1375 
1376   // First set the frame anchor while the SP/FP registers are
1377   // convenient and the slow signature handler can use this same frame
1378   // anchor.
1379 
1380   // We have a TOP_IJAVA_FRAME here, which belongs to us.
1381   __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
1382 
1383   // Now the interpreter frame (and its call chain) have been
1384   // invalidated and flushed. We are now protected against eager
1385   // being enabled in native code. Even if it goes eager the
1386   // registers will be reloaded as clean and we will invalidate after
1387   // the call so no spurious flush should be possible.
1388 
1389   // Call signature handler and pass locals address.
1390   //
1391   // Our signature handlers copy required arguments to the C stack
1392   // (outgoing C args), R3_ARG1 to R10_ARG8, and FARG1 to FARG13.
1393   __ mr(R3_ARG1, R18_locals);
1394 #if !defined(ABI_ELFv2)
1395   __ ld(signature_handler_fd, 0, signature_handler_fd);
1396 #endif
1397 
1398   __ call_stub(signature_handler_fd);
1399 
1400   // Remove the register parameter varargs slots we allocated in
1401   // compute_interpreter_state. SP+16 ends up pointing to the ABI
1402   // outgoing argument area.
1403   //
1404   // Not needed on PPC64.
1405   //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord);
1406 
1407   assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register");
1408   // Save across call to native method.
1409   __ mr(result_handler_addr, R3_RET);
1410 
1411   __ isync(); // Acquire signature handler before trying to fetch the native entry point and klass mirror.
1412 
1413   // Set up fixed parameters and call the native method.
1414   // If the method is static, get mirror into R4_ARG2.
1415   {
1416     Label method_is_not_static;
1417     // Access_flags is non-volatile and still, no need to restore it.
1418 
1419     // Restore access flags.
1420     __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT);
1421     __ bfalse(CCR0, method_is_not_static);
1422 
1423     __ ld(R11_scratch1, _abi(callers_sp), R1_SP);
1424     // Load mirror from interpreter frame.
1425     __ ld(R12_scratch2, _ijava_state_neg(mirror), R11_scratch1);
1426     // R4_ARG2 = &state->_oop_temp;
1427     __ addi(R4_ARG2, R11_scratch1, _ijava_state_neg(oop_tmp));
1428     __ std(R12_scratch2/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1);
1429     BIND(method_is_not_static);
1430   }
1431 
1432   // At this point, arguments have been copied off the stack into
1433   // their JNI positions. Oops are boxed in-place on the stack, with
1434   // handles copied to arguments. The result handler address is in a
1435   // register.
1436 
1437   // Pass JNIEnv address as first parameter.
1438   __ addir(R3_ARG1, thread_(jni_environment));
1439 
1440   // Load the native_method entry before we change the thread state.
1441   __ ld(native_method_fd, method_(native_function));
1442 
1443   //=============================================================================
1444   // Transition from _thread_in_Java to _thread_in_native. As soon as
1445   // we make this change the safepoint code needs to be certain that
1446   // the last Java frame we established is good. The pc in that frame
1447   // just needs to be near here not an actual return address.
1448 
1449   // We use release_store_fence to update values like the thread state, where
1450   // we don't want the current thread to continue until all our prior memory
1451   // accesses (including the new thread state) are visible to other threads.
1452   __ li(R0, _thread_in_native);
1453   __ release();
1454 
1455   // TODO PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
1456   __ stw(R0, thread_(thread_state));
1457 
1458   //=============================================================================
1459   // Call the native method. Argument registers must not have been
1460   // overwritten since "__ call_stub(signature_handler);" (except for
1461   // ARG1 and ARG2 for static methods).
1462   __ call_c(native_method_fd);
1463 
1464   __ li(R0, 0);
1465   __ ld(R11_scratch1, 0, R1_SP);
1466   __ std(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1467   __ stfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1468   __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1); // reset
1469 
1470   // Note: C++ interpreter needs the following here:
1471   // The frame_manager_lr field, which we use for setting the last
1472   // java frame, gets overwritten by the signature handler. Restore
1473   // it now.
1474   //__ get_PC_trash_LR(R11_scratch1);
1475   //__ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1476 
1477   // Because of GC R19_method may no longer be valid.
1478 
1479   // Block, if necessary, before resuming in _thread_in_Java state.
1480   // In order for GC to work, don't clear the last_Java_sp until after
1481   // blocking.
1482 
1483   //=============================================================================
1484   // Switch thread to "native transition" state before reading the
1485   // synchronization state. This additional state is necessary
1486   // because reading and testing the synchronization state is not
1487   // atomic w.r.t. GC, as this scenario demonstrates: Java thread A,
1488   // in _thread_in_native state, loads _not_synchronized and is
1489   // preempted. VM thread changes sync state to synchronizing and
1490   // suspends threads for GC. Thread A is resumed to finish this
1491   // native method, but doesn't block here since it didn't see any
1492   // synchronization in progress, and escapes.
1493 
1494   // We use release_store_fence to update values like the thread state, where
1495   // we don't want the current thread to continue until all our prior memory
1496   // accesses (including the new thread state) are visible to other threads.
1497   __ li(R0/*thread_state*/, _thread_in_native_trans);
1498   __ release();
1499   __ stw(R0/*thread_state*/, thread_(thread_state));
1500   if (UseMembar) {
1501     __ fence();
1502   }
1503   // Write serialization page so that the VM thread can do a pseudo remote
1504   // membar. We use the current thread pointer to calculate a thread
1505   // specific offset to write to within the page. This minimizes bus
1506   // traffic due to cache line collision.
1507   else {
1508     __ serialize_memory(R16_thread, R11_scratch1, R12_scratch2);
1509   }
1510 
1511   // Now before we return to java we must look for a current safepoint
1512   // (a new safepoint can not start since we entered native_trans).
1513   // We must check here because a current safepoint could be modifying
1514   // the callers registers right this moment.
1515 
1516   // Acquire isn't strictly necessary here because of the fence, but
1517   // sync_state is declared to be volatile, so we do it anyway
1518   // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
1519 
1520   Label do_safepoint, sync_check_done;
1521   // No synchronization in progress nor yet synchronized.
1522   __ safepoint_poll(do_safepoint, sync_state);
1523 
1524   // Not suspended.
1525   // TODO PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
1526   __ lwz(suspend_flags, thread_(suspend_flags));
1527   __ cmpwi(CCR1, suspend_flags, 0);
1528   __ beq(CCR1, sync_check_done);
1529 
1530   __ bind(do_safepoint);
1531   __ isync();
1532   // Block. We do the call directly and leave the current
1533   // last_Java_frame setup undisturbed. We must save any possible
1534   // native result across the call. No oop is present.
1535 
1536   __ mr(R3_ARG1, R16_thread);
1537 #if defined(ABI_ELFv2)
1538   __ call_c(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1539             relocInfo::none);
1540 #else
1541   __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans),
1542             relocInfo::none);
1543 #endif
1544 
1545   __ bind(sync_check_done);
1546 
1547   //=============================================================================
1548   // <<<<<< Back in Interpreter Frame >>>>>
1549 
1550   // We are in thread_in_native_trans here and back in the normal
1551   // interpreter frame. We don't have to do anything special about
1552   // safepoints and we can switch to Java mode anytime we are ready.
1553 
1554   // Note: frame::interpreter_frame_result has a dependency on how the
1555   // method result is saved across the call to post_method_exit. For
1556   // native methods it assumes that the non-FPU/non-void result is
1557   // saved in _native_lresult and a FPU result in _native_fresult. If
1558   // this changes then the interpreter_frame_result implementation
1559   // will need to be updated too.
1560 
1561   // On PPC64, we have stored the result directly after the native call.
1562 
1563   //=============================================================================
1564   // Back in Java
1565 
1566   // We use release_store_fence to update values like the thread state, where
1567   // we don't want the current thread to continue until all our prior memory
1568   // accesses (including the new thread state) are visible to other threads.
1569   __ li(R0/*thread_state*/, _thread_in_Java);
1570   __ lwsync(); // Acquire safepoint and suspend state, release thread state.
1571   __ stw(R0/*thread_state*/, thread_(thread_state));
1572 
1573   if (CheckJNICalls) {
1574     // clear_pending_jni_exception_check
1575     __ load_const_optimized(R0, 0L);
1576     __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
1577   }
1578 
1579   __ reset_last_Java_frame();
1580 
1581   // Jvmdi/jvmpi support. Whether we've got an exception pending or
1582   // not, and whether unlocking throws an exception or not, we notify
1583   // on native method exit. If we do have an exception, we'll end up
1584   // in the caller's context to handle it, so if we don't do the
1585   // notify here, we'll drop it on the floor.
1586   __ notify_method_exit(true/*native method*/,
1587                         ilgl /*illegal state (not used for native methods)*/,
1588                         InterpreterMacroAssembler::NotifyJVMTI,
1589                         false /*check_exceptions*/);
1590 
1591   //=============================================================================
1592   // Handle exceptions
1593 
1594   if (synchronized) {
1595     // Don't check for exceptions since we're still in the i2n frame. Do that
1596     // manually afterwards.
1597     __ unlock_object(R26_monitor, false); // Can also unlock methods.
1598   }
1599 
1600   // Reset active handles after returning from native.
1601   // thread->active_handles()->clear();
1602   __ ld(active_handles, thread_(active_handles));
1603   // TODO PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
1604   __ li(R0, 0);
1605   __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles);
1606 
1607   Label exception_return_sync_check_already_unlocked;
1608   __ ld(R0/*pending_exception*/, thread_(pending_exception));
1609   __ cmpdi(CCR0, R0/*pending_exception*/, 0);
1610   __ bne(CCR0, exception_return_sync_check_already_unlocked);
1611 
1612   //-----------------------------------------------------------------------------
1613   // No exception pending.
1614 
1615   // Move native method result back into proper registers and return.
1616   // Invoke result handler (may unbox/promote).
1617   __ ld(R11_scratch1, 0, R1_SP);
1618   __ ld(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1619   __ lfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1620   __ call_stub(result_handler_addr);
1621 
1622   __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ R0, R11_scratch1, R12_scratch2);
1623 
1624   // Must use the return pc which was loaded from the caller's frame
1625   // as the VM uses return-pc-patching for deoptimization.
1626   __ mtlr(R0);
1627   __ blr();
1628 
1629   //-----------------------------------------------------------------------------
1630   // An exception is pending. We call into the runtime only if the
1631   // caller was not interpreted. If it was interpreted the
1632   // interpreter will do the correct thing. If it isn't interpreted
1633   // (call stub/compiled code) we will change our return and continue.
1634 
1635   BIND(exception_return_sync_check);
1636 
1637   if (synchronized) {
1638     // Don't check for exceptions since we're still in the i2n frame. Do that
1639     // manually afterwards.
1640     __ unlock_object(R26_monitor, false); // Can also unlock methods.
1641   }
1642   BIND(exception_return_sync_check_already_unlocked);
1643 
1644   const Register return_pc = R31;
1645 
1646   __ ld(return_pc, 0, R1_SP);
1647   __ ld(return_pc, _abi(lr), return_pc);
1648 
1649   // Get the address of the exception handler.
1650   __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1651                   R16_thread,
1652                   return_pc /* return pc */);
1653   __ merge_frames(/*top_frame_sp*/ R21_sender_SP, noreg, R11_scratch1, R12_scratch2);
1654 
1655   // Load the PC of the the exception handler into LR.
1656   __ mtlr(R3_RET);
1657 
1658   // Load exception into R3_ARG1 and clear pending exception in thread.
1659   __ ld(R3_ARG1/*exception*/, thread_(pending_exception));
1660   __ li(R4_ARG2, 0);
1661   __ std(R4_ARG2, thread_(pending_exception));
1662 
1663   // Load the original return pc into R4_ARG2.
1664   __ mr(R4_ARG2/*issuing_pc*/, return_pc);
1665 
1666   // Return to exception handler.
1667   __ blr();
1668 
1669   //=============================================================================
1670   // Counter overflow.
1671 
1672   if (inc_counter) {
1673     // Handle invocation counter overflow.
1674     __ bind(invocation_counter_overflow);
1675 
1676     generate_counter_overflow(continue_after_compile);
1677   }
1678 
1679   return entry;
1680 }
1681 
1682 // Generic interpreted method entry to (asm) interpreter.
1683 //
1684 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1685   bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1686   address entry = __ pc();
1687   // Generate the code to allocate the interpreter stack frame.
1688   Register Rsize_of_parameters = R4_ARG2, // Written by generate_fixed_frame.
1689            Rsize_of_locals     = R5_ARG3; // Written by generate_fixed_frame.
1690 
1691   // Does also a stack check to assure this frame fits on the stack.
1692   generate_fixed_frame(false, Rsize_of_parameters, Rsize_of_locals);
1693 
1694   // --------------------------------------------------------------------------
1695   // Zero out non-parameter locals.
1696   // Note: *Always* zero out non-parameter locals as Sparc does. It's not
1697   // worth to ask the flag, just do it.
1698   Register Rslot_addr = R6_ARG4,
1699            Rnum       = R7_ARG5;
1700   Label Lno_locals, Lzero_loop;
1701 
1702   // Set up the zeroing loop.
1703   __ subf(Rnum, Rsize_of_parameters, Rsize_of_locals);
1704   __ subf(Rslot_addr, Rsize_of_parameters, R18_locals);
1705   __ srdi_(Rnum, Rnum, Interpreter::logStackElementSize);
1706   __ beq(CCR0, Lno_locals);
1707   __ li(R0, 0);
1708   __ mtctr(Rnum);
1709 
1710   // The zero locals loop.
1711   __ bind(Lzero_loop);
1712   __ std(R0, 0, Rslot_addr);
1713   __ addi(Rslot_addr, Rslot_addr, -Interpreter::stackElementSize);
1714   __ bdnz(Lzero_loop);
1715 
1716   __ bind(Lno_locals);
1717 
1718   // --------------------------------------------------------------------------
1719   // Counter increment and overflow check.
1720   Label invocation_counter_overflow,
1721         profile_method,
1722         profile_method_continue;
1723   if (inc_counter || ProfileInterpreter) {
1724 
1725     Register Rdo_not_unlock_if_synchronized_addr = R11_scratch1;
1726     if (synchronized) {
1727       // Since at this point in the method invocation the exception handler
1728       // would try to exit the monitor of synchronized methods which hasn't
1729       // been entered yet, we set the thread local variable
1730       // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1731       // runtime, exception handling i.e. unlock_if_synchronized_method will
1732       // check this thread local flag.
1733       // This flag has two effects, one is to force an unwind in the topmost
1734       // interpreter frame and not perform an unlock while doing so.
1735       __ li(R0, 1);
1736       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1737     }
1738 
1739     // Argument and return type profiling.
1740     __ profile_parameters_type(R3_ARG1, R4_ARG2, R5_ARG3, R6_ARG4);
1741 
1742     // Increment invocation counter and check for overflow.
1743     if (inc_counter) {
1744       generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1745     }
1746 
1747     __ bind(profile_method_continue);
1748   }
1749 
1750   bang_stack_shadow_pages(false);
1751 
1752   if (inc_counter || ProfileInterpreter) {
1753     // Reset the _do_not_unlock_if_synchronized flag.
1754     if (synchronized) {
1755       __ li(R0, 0);
1756       __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1757     }
1758   }
1759 
1760   // --------------------------------------------------------------------------
1761   // Locking of synchronized methods. Must happen AFTER invocation_counter
1762   // check and stack overflow check, so method is not locked if overflows.
1763   if (synchronized) {
1764     lock_method(R3_ARG1, R4_ARG2, R5_ARG3);
1765   }
1766 #ifdef ASSERT
1767   else {
1768     Label Lok;
1769     __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1770     __ andi_(R0, R0, JVM_ACC_SYNCHRONIZED);
1771     __ asm_assert_eq("method needs synchronization", 0x8521);
1772     __ bind(Lok);
1773   }
1774 #endif // ASSERT
1775 
1776   __ verify_thread();
1777 
1778   // --------------------------------------------------------------------------
1779   // JVMTI support
1780   __ notify_method_entry();
1781 
1782   // --------------------------------------------------------------------------
1783   // Start executing instructions.
1784   __ dispatch_next(vtos);
1785 
1786   // --------------------------------------------------------------------------
1787   // Out of line counter overflow and MDO creation code.
1788   if (ProfileInterpreter) {
1789     // We have decided to profile this method in the interpreter.
1790     __ bind(profile_method);
1791     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1792     __ set_method_data_pointer_for_bcp();
1793     __ b(profile_method_continue);
1794   }
1795 
1796   if (inc_counter) {
1797     // Handle invocation counter overflow.
1798     __ bind(invocation_counter_overflow);
1799     generate_counter_overflow(profile_method_continue);
1800   }
1801   return entry;
1802 }
1803 
1804 // CRC32 Intrinsics.
1805 //
1806 // Contract on scratch and work registers.
1807 // =======================================
1808 //
1809 // On ppc, the register set {R2..R12} is available in the interpreter as scratch/work registers.
1810 // You should, however, keep in mind that {R3_ARG1..R10_ARG8} is the C-ABI argument register set.
1811 // You can't rely on these registers across calls.
1812 //
1813 // The generators for CRC32_update and for CRC32_updateBytes use the
1814 // scratch/work register set internally, passing the work registers
1815 // as arguments to the MacroAssembler emitters as required.
1816 //
1817 // R3_ARG1..R6_ARG4 are preset to hold the incoming java arguments.
1818 // Their contents is not constant but may change according to the requirements
1819 // of the emitted code.
1820 //
1821 // All other registers from the scratch/work register set are used "internally"
1822 // and contain garbage (i.e. unpredictable values) once blr() is reached.
1823 // Basically, only R3_RET contains a defined value which is the function result.
1824 //
1825 /**
1826  * Method entry for static native methods:
1827  *   int java.util.zip.CRC32.update(int crc, int b)
1828  */
1829 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
1830   if (UseCRC32Intrinsics) {
1831     address start = __ pc();  // Remember stub start address (is rtn value).
1832     Label slow_path;
1833 
1834     // Safepoint check
1835     const Register sync_state = R11_scratch1;
1836     __ safepoint_poll(slow_path, sync_state);
1837 
1838     // We don't generate local frame and don't align stack because
1839     // we not even call stub code (we generate the code inline)
1840     // and there is no safepoint on this path.
1841 
1842     // Load java parameters.
1843     // R15_esp is callers operand stack pointer, i.e. it points to the parameters.
1844     const Register argP    = R15_esp;
1845     const Register crc     = R3_ARG1;  // crc value
1846     const Register data    = R4_ARG2;  // address of java byte value (kernel_crc32 needs address)
1847     const Register dataLen = R5_ARG3;  // source data len (1 byte). Not used because calling the single-byte emitter.
1848     const Register table   = R6_ARG4;  // address of crc32 table
1849     const Register tmp     = dataLen;  // Reuse unused len register to show we don't actually need a separate tmp here.
1850 
1851     BLOCK_COMMENT("CRC32_update {");
1852 
1853     // Arguments are reversed on java expression stack
1854 #ifdef VM_LITTLE_ENDIAN
1855     __ addi(data, argP, 0+1*wordSize); // (stack) address of byte value. Emitter expects address, not value.
1856                                        // Being passed as an int, the single byte is at offset +0.
1857 #else
1858     __ addi(data, argP, 3+1*wordSize); // (stack) address of byte value. Emitter expects address, not value.
1859                                        // Being passed from java as an int, the single byte is at offset +3.
1860 #endif
1861     __ lwz(crc,  2*wordSize, argP);    // Current crc state, zero extend to 64 bit to have a clean register.
1862 
1863     StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table);
1864     __ kernel_crc32_singleByte(crc, data, dataLen, table, tmp, true);
1865 
1866     // Restore caller sp for c2i case and return.
1867     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1868     __ blr();
1869 
1870     // Generate a vanilla native entry as the slow path.
1871     BLOCK_COMMENT("} CRC32_update");
1872     BIND(slow_path);
1873     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native), R11_scratch1);
1874     return start;
1875   }
1876 
1877   return NULL;
1878 }
1879 
1880 /**
1881  * Method entry for static native methods:
1882  *   int java.util.zip.CRC32.updateBytes(     int crc, byte[] b,  int off, int len)
1883  *   int java.util.zip.CRC32.updateByteBuffer(int crc, long* buf, int off, int len)
1884  */
1885 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1886   if (UseCRC32Intrinsics) {
1887     address start = __ pc();  // Remember stub start address (is rtn value).
1888     Label slow_path;
1889 
1890     // Safepoint check
1891     const Register sync_state = R11_scratch1;
1892     __ safepoint_poll(slow_path, sync_state);
1893 
1894     // We don't generate local frame and don't align stack because
1895     // we not even call stub code (we generate the code inline)
1896     // and there is no safepoint on this path.
1897 
1898     // Load parameters.
1899     // Z_esp is callers operand stack pointer, i.e. it points to the parameters.
1900     const Register argP    = R15_esp;
1901     const Register crc     = R3_ARG1;  // crc value
1902     const Register data    = R4_ARG2;  // address of java byte array
1903     const Register dataLen = R5_ARG3;  // source data len
1904     const Register table   = R6_ARG4;  // address of crc32 table
1905 
1906     const Register t0      = R9;       // scratch registers for crc calculation
1907     const Register t1      = R10;
1908     const Register t2      = R11;
1909     const Register t3      = R12;
1910 
1911     const Register tc0     = R2;       // registers to hold pre-calculated column addresses
1912     const Register tc1     = R7;
1913     const Register tc2     = R8;
1914     const Register tc3     = table;    // table address is reconstructed at the end of kernel_crc32_* emitters
1915 
1916     const Register tmp     = t0;       // Only used very locally to calculate byte buffer address.
1917 
1918     // Arguments are reversed on java expression stack.
1919     // Calculate address of start element.
1920     if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) { // Used for "updateByteBuffer direct".
1921       BLOCK_COMMENT("CRC32_updateByteBuffer {");
1922       // crc     @ (SP + 5W) (32bit)
1923       // buf     @ (SP + 3W) (64bit ptr to long array)
1924       // off     @ (SP + 2W) (32bit)
1925       // dataLen @ (SP + 1W) (32bit)
1926       // data = buf + off
1927       __ ld(  data,    3*wordSize, argP);  // start of byte buffer
1928       __ lwa( tmp,     2*wordSize, argP);  // byte buffer offset
1929       __ lwa( dataLen, 1*wordSize, argP);  // #bytes to process
1930       __ lwz( crc,     5*wordSize, argP);  // current crc state
1931       __ add( data, data, tmp);            // Add byte buffer offset.
1932     } else {                                                         // Used for "updateBytes update".
1933       BLOCK_COMMENT("CRC32_updateBytes {");
1934       // crc     @ (SP + 4W) (32bit)
1935       // buf     @ (SP + 3W) (64bit ptr to byte array)
1936       // off     @ (SP + 2W) (32bit)
1937       // dataLen @ (SP + 1W) (32bit)
1938       // data = buf + off + base_offset
1939       __ ld(  data,    3*wordSize, argP);  // start of byte buffer
1940       __ lwa( tmp,     2*wordSize, argP);  // byte buffer offset
1941       __ lwa( dataLen, 1*wordSize, argP);  // #bytes to process
1942       __ add( data, data, tmp);            // add byte buffer offset
1943       __ lwz( crc,     4*wordSize, argP);  // current crc state
1944       __ addi(data, data, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1945     }
1946 
1947     StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table);
1948 
1949     // Performance measurements show the 1word and 2word variants to be almost equivalent,
1950     // with very light advantages for the 1word variant. We chose the 1word variant for
1951     // code compactness.
1952     __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, tc3, true);
1953 
1954     // Restore caller sp for c2i case and return.
1955     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1956     __ blr();
1957 
1958     // Generate a vanilla native entry as the slow path.
1959     BLOCK_COMMENT("} CRC32_updateBytes(Buffer)");
1960     BIND(slow_path);
1961     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native), R11_scratch1);
1962     return start;
1963   }
1964 
1965   return NULL;
1966 }
1967 
1968 
1969 /**
1970  * Method entry for intrinsic-candidate (non-native) methods:
1971  *   int java.util.zip.CRC32C.updateBytes(           int crc, byte[] b,  int off, int end)
1972  *   int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long* buf, int off, int end)
1973  * Unlike CRC32, CRC32C does not have any methods marked as native
1974  * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1975  **/
1976 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1977   if (UseCRC32CIntrinsics) {
1978     address start = __ pc();  // Remember stub start address (is rtn value).
1979 
1980     // We don't generate local frame and don't align stack because
1981     // we not even call stub code (we generate the code inline)
1982     // and there is no safepoint on this path.
1983 
1984     // Load parameters.
1985     // Z_esp is callers operand stack pointer, i.e. it points to the parameters.
1986     const Register argP    = R15_esp;
1987     const Register crc     = R3_ARG1;  // crc value
1988     const Register data    = R4_ARG2;  // address of java byte array
1989     const Register dataLen = R5_ARG3;  // source data len
1990     const Register table   = R6_ARG4;  // address of crc32c table
1991 
1992     const Register t0      = R9;       // scratch registers for crc calculation
1993     const Register t1      = R10;
1994     const Register t2      = R11;
1995     const Register t3      = R12;
1996 
1997     const Register tc0     = R2;       // registers to hold pre-calculated column addresses
1998     const Register tc1     = R7;
1999     const Register tc2     = R8;
2000     const Register tc3     = table;    // table address is reconstructed at the end of kernel_crc32_* emitters
2001 
2002     const Register tmp     = t0;       // Only used very locally to calculate byte buffer address.
2003 
2004     // Arguments are reversed on java expression stack.
2005     // Calculate address of start element.
2006     if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) { // Used for "updateDirectByteBuffer".
2007       BLOCK_COMMENT("CRC32C_updateDirectByteBuffer {");
2008       // crc     @ (SP + 5W) (32bit)
2009       // buf     @ (SP + 3W) (64bit ptr to long array)
2010       // off     @ (SP + 2W) (32bit)
2011       // dataLen @ (SP + 1W) (32bit)
2012       // data = buf + off
2013       __ ld(  data,    3*wordSize, argP);  // start of byte buffer
2014       __ lwa( tmp,     2*wordSize, argP);  // byte buffer offset
2015       __ lwa( dataLen, 1*wordSize, argP);  // #bytes to process
2016       __ lwz( crc,     5*wordSize, argP);  // current crc state
2017       __ add( data, data, tmp);            // Add byte buffer offset.
2018       __ sub( dataLen, dataLen, tmp);      // (end_index - offset)
2019     } else {                                                         // Used for "updateBytes update".
2020       BLOCK_COMMENT("CRC32C_updateBytes {");
2021       // crc     @ (SP + 4W) (32bit)
2022       // buf     @ (SP + 3W) (64bit ptr to byte array)
2023       // off     @ (SP + 2W) (32bit)
2024       // dataLen @ (SP + 1W) (32bit)
2025       // data = buf + off + base_offset
2026       __ ld(  data,    3*wordSize, argP);  // start of byte buffer
2027       __ lwa( tmp,     2*wordSize, argP);  // byte buffer offset
2028       __ lwa( dataLen, 1*wordSize, argP);  // #bytes to process
2029       __ add( data, data, tmp);            // add byte buffer offset
2030       __ sub( dataLen, dataLen, tmp);      // (end_index - offset)
2031       __ lwz( crc,     4*wordSize, argP);  // current crc state
2032       __ addi(data, data, arrayOopDesc::base_offset_in_bytes(T_BYTE));
2033     }
2034 
2035     StubRoutines::ppc64::generate_load_crc32c_table_addr(_masm, table);
2036 
2037     // Performance measurements show the 1word and 2word variants to be almost equivalent,
2038     // with very light advantages for the 1word variant. We chose the 1word variant for
2039     // code compactness.
2040     __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, tc3, false);
2041 
2042     // Restore caller sp for c2i case and return.
2043     __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
2044     __ blr();
2045 
2046     BLOCK_COMMENT("} CRC32C_update{Bytes|DirectByteBuffer}");
2047     return start;
2048   }
2049 
2050   return NULL;
2051 }
2052 
2053 // =============================================================================
2054 // Exceptions
2055 
2056 void TemplateInterpreterGenerator::generate_throw_exception() {
2057   Register Rexception    = R17_tos,
2058            Rcontinuation = R3_RET;
2059 
2060   // --------------------------------------------------------------------------
2061   // Entry point if an method returns with a pending exception (rethrow).
2062   Interpreter::_rethrow_exception_entry = __ pc();
2063   {
2064     __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
2065     __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
2066     __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
2067 
2068     // Compiled code destroys templateTableBase, reload.
2069     __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
2070   }
2071 
2072   // Entry point if a interpreted method throws an exception (throw).
2073   Interpreter::_throw_exception_entry = __ pc();
2074   {
2075     __ mr(Rexception, R3_RET);
2076 
2077     __ verify_thread();
2078     __ verify_oop(Rexception);
2079 
2080     // Expression stack must be empty before entering the VM in case of an exception.
2081     __ empty_expression_stack();
2082     // Find exception handler address and preserve exception oop.
2083     // Call C routine to find handler and jump to it.
2084     __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Rexception);
2085     __ mtctr(Rcontinuation);
2086     // Push exception for exception handler bytecodes.
2087     __ push_ptr(Rexception);
2088 
2089     // Jump to exception handler (may be remove activation entry!).
2090     __ bctr();
2091   }
2092 
2093   // If the exception is not handled in the current frame the frame is
2094   // removed and the exception is rethrown (i.e. exception
2095   // continuation is _rethrow_exception).
2096   //
2097   // Note: At this point the bci is still the bxi for the instruction
2098   // which caused the exception and the expression stack is
2099   // empty. Thus, for any VM calls at this point, GC will find a legal
2100   // oop map (with empty expression stack).
2101 
2102   // In current activation
2103   // tos: exception
2104   // bcp: exception bcp
2105 
2106   // --------------------------------------------------------------------------
2107   // JVMTI PopFrame support
2108 
2109   Interpreter::_remove_activation_preserving_args_entry = __ pc();
2110   {
2111     // Set the popframe_processing bit in popframe_condition indicating that we are
2112     // currently handling popframe, so that call_VMs that may happen later do not
2113     // trigger new popframe handling cycles.
2114     __ lwz(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2115     __ ori(R11_scratch1, R11_scratch1, JavaThread::popframe_processing_bit);
2116     __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2117 
2118     // Empty the expression stack, as in normal exception handling.
2119     __ empty_expression_stack();
2120     __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
2121 
2122     // Check to see whether we are returning to a deoptimized frame.
2123     // (The PopFrame call ensures that the caller of the popped frame is
2124     // either interpreted or compiled and deoptimizes it if compiled.)
2125     // Note that we don't compare the return PC against the
2126     // deoptimization blob's unpack entry because of the presence of
2127     // adapter frames in C2.
2128     Label Lcaller_not_deoptimized;
2129     Register return_pc = R3_ARG1;
2130     __ ld(return_pc, 0, R1_SP);
2131     __ ld(return_pc, _abi(lr), return_pc);
2132     __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), return_pc);
2133     __ cmpdi(CCR0, R3_RET, 0);
2134     __ bne(CCR0, Lcaller_not_deoptimized);
2135 
2136     // The deoptimized case.
2137     // In this case, we can't call dispatch_next() after the frame is
2138     // popped, but instead must save the incoming arguments and restore
2139     // them after deoptimization has occurred.
2140     __ ld(R4_ARG2, in_bytes(Method::const_offset()), R19_method);
2141     __ lhz(R4_ARG2 /* number of params */, in_bytes(ConstMethod::size_of_parameters_offset()), R4_ARG2);
2142     __ slwi(R4_ARG2, R4_ARG2, Interpreter::logStackElementSize);
2143     __ addi(R5_ARG3, R18_locals, Interpreter::stackElementSize);
2144     __ subf(R5_ARG3, R4_ARG2, R5_ARG3);
2145     // Save these arguments.
2146     __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), R16_thread, R4_ARG2, R5_ARG3);
2147 
2148     // Inform deoptimization that it is responsible for restoring these arguments.
2149     __ load_const_optimized(R11_scratch1, JavaThread::popframe_force_deopt_reexecution_bit);
2150     __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2151 
2152     // Return from the current method into the deoptimization blob. Will eventually
2153     // end up in the deopt interpeter entry, deoptimization prepared everything that
2154     // we will reexecute the call that called us.
2155     __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*reload return_pc*/ return_pc, R11_scratch1, R12_scratch2);
2156     __ mtlr(return_pc);
2157     __ blr();
2158 
2159     // The non-deoptimized case.
2160     __ bind(Lcaller_not_deoptimized);
2161 
2162     // Clear the popframe condition flag.
2163     __ li(R0, 0);
2164     __ stw(R0, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2165 
2166     // Get out of the current method and re-execute the call that called us.
2167     __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
2168     __ restore_interpreter_state(R11_scratch1);
2169     __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
2170     __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
2171     if (ProfileInterpreter) {
2172       __ set_method_data_pointer_for_bcp();
2173       __ ld(R11_scratch1, 0, R1_SP);
2174       __ std(R28_mdx, _ijava_state_neg(mdx), R11_scratch1);
2175     }
2176 #if INCLUDE_JVMTI
2177     Label L_done;
2178 
2179     __ lbz(R11_scratch1, 0, R14_bcp);
2180     __ cmpwi(CCR0, R11_scratch1, Bytecodes::_invokestatic);
2181     __ bne(CCR0, L_done);
2182 
2183     // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
2184     // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
2185     __ ld(R4_ARG2, 0, R18_locals);
2186     __ MacroAssembler::call_VM(R4_ARG2, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), R4_ARG2, R19_method, R14_bcp, false);
2187     __ restore_interpreter_state(R11_scratch1, /*bcp_and_mdx_only*/ true);
2188     __ cmpdi(CCR0, R4_ARG2, 0);
2189     __ beq(CCR0, L_done);
2190     __ std(R4_ARG2, wordSize, R15_esp);
2191     __ bind(L_done);
2192 #endif // INCLUDE_JVMTI
2193     __ dispatch_next(vtos);
2194   }
2195   // end of JVMTI PopFrame support
2196 
2197   // --------------------------------------------------------------------------
2198   // Remove activation exception entry.
2199   // This is jumped to if an interpreted method can't handle an exception itself
2200   // (we come from the throw/rethrow exception entry above). We're going to call
2201   // into the VM to find the exception handler in the caller, pop the current
2202   // frame and return the handler we calculated.
2203   Interpreter::_remove_activation_entry = __ pc();
2204   {
2205     __ pop_ptr(Rexception);
2206     __ verify_thread();
2207     __ verify_oop(Rexception);
2208     __ std(Rexception, in_bytes(JavaThread::vm_result_offset()), R16_thread);
2209 
2210     __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, true);
2211     __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI, false);
2212 
2213     __ get_vm_result(Rexception);
2214 
2215     // We are done with this activation frame; find out where to go next.
2216     // The continuation point will be an exception handler, which expects
2217     // the following registers set up:
2218     //
2219     // RET:  exception oop
2220     // ARG2: Issuing PC (see generate_exception_blob()), only used if the caller is compiled.
2221 
2222     Register return_pc = R31; // Needs to survive the runtime call.
2223     __ ld(return_pc, 0, R1_SP);
2224     __ ld(return_pc, _abi(lr), return_pc);
2225     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, return_pc);
2226 
2227     // Remove the current activation.
2228     __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
2229 
2230     __ mr(R4_ARG2, return_pc);
2231     __ mtlr(R3_RET);
2232     __ mr(R3_RET, Rexception);
2233     __ blr();
2234   }
2235 }
2236 
2237 // JVMTI ForceEarlyReturn support.
2238 // Returns "in the middle" of a method with a "fake" return value.
2239 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
2240 
2241   Register Rscratch1 = R11_scratch1,
2242            Rscratch2 = R12_scratch2;
2243 
2244   address entry = __ pc();
2245   __ empty_expression_stack();
2246 
2247   __ load_earlyret_value(state, Rscratch1);
2248 
2249   __ ld(Rscratch1, in_bytes(JavaThread::jvmti_thread_state_offset()), R16_thread);
2250   // Clear the earlyret state.
2251   __ li(R0, 0);
2252   __ stw(R0, in_bytes(JvmtiThreadState::earlyret_state_offset()), Rscratch1);
2253 
2254   __ remove_activation(state, false, false);
2255   // Copied from TemplateTable::_return.
2256   // Restoration of lr done by remove_activation.
2257   switch (state) {
2258     // Narrow result if state is itos but result type is smaller.
2259     case btos:
2260     case ztos:
2261     case ctos:
2262     case stos:
2263     case itos: __ narrow(R17_tos); /* fall through */
2264     case ltos:
2265     case atos: __ mr(R3_RET, R17_tos); break;
2266     case ftos:
2267     case dtos: __ fmr(F1_RET, F15_ftos); break;
2268     case vtos: // This might be a constructor. Final fields (and volatile fields on PPC64) need
2269                // to get visible before the reference to the object gets stored anywhere.
2270                __ membar(Assembler::StoreStore); break;
2271     default  : ShouldNotReachHere();
2272   }
2273   __ blr();
2274 
2275   return entry;
2276 } // end of ForceEarlyReturn support
2277 
2278 //-----------------------------------------------------------------------------
2279 // Helper for vtos entry point generation
2280 
2281 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t,
2282                                                          address& bep,
2283                                                          address& cep,
2284                                                          address& sep,
2285                                                          address& aep,
2286                                                          address& iep,
2287                                                          address& lep,
2288                                                          address& fep,
2289                                                          address& dep,
2290                                                          address& vep) {
2291   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2292   Label L;
2293 
2294   aep = __ pc();  __ push_ptr();  __ b(L);
2295   fep = __ pc();  __ push_f();    __ b(L);
2296   dep = __ pc();  __ push_d();    __ b(L);
2297   lep = __ pc();  __ push_l();    __ b(L);
2298   __ align(32, 12, 24); // align L
2299   bep = cep = sep =
2300   iep = __ pc();  __ push_i();
2301   vep = __ pc();
2302   __ bind(L);
2303   generate_and_dispatch(t);
2304 }
2305 
2306 //-----------------------------------------------------------------------------
2307 
2308 // Non-product code
2309 #ifndef PRODUCT
2310 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2311   //__ flush_bundle();
2312   address entry = __ pc();
2313 
2314   const char *bname = NULL;
2315   uint tsize = 0;
2316   switch(state) {
2317   case ftos:
2318     bname = "trace_code_ftos {";
2319     tsize = 2;
2320     break;
2321   case btos:
2322     bname = "trace_code_btos {";
2323     tsize = 2;
2324     break;
2325   case ztos:
2326     bname = "trace_code_ztos {";
2327     tsize = 2;
2328     break;
2329   case ctos:
2330     bname = "trace_code_ctos {";
2331     tsize = 2;
2332     break;
2333   case stos:
2334     bname = "trace_code_stos {";
2335     tsize = 2;
2336     break;
2337   case itos:
2338     bname = "trace_code_itos {";
2339     tsize = 2;
2340     break;
2341   case ltos:
2342     bname = "trace_code_ltos {";
2343     tsize = 3;
2344     break;
2345   case atos:
2346     bname = "trace_code_atos {";
2347     tsize = 2;
2348     break;
2349   case vtos:
2350     // Note: In case of vtos, the topmost of stack value could be a int or doubl
2351     // In case of a double (2 slots) we won't see the 2nd stack value.
2352     // Maybe we simply should print the topmost 3 stack slots to cope with the problem.
2353     bname = "trace_code_vtos {";
2354     tsize = 2;
2355 
2356     break;
2357   case dtos:
2358     bname = "trace_code_dtos {";
2359     tsize = 3;
2360     break;
2361   default:
2362     ShouldNotReachHere();
2363   }
2364   BLOCK_COMMENT(bname);
2365 
2366   // Support short-cut for TraceBytecodesAt.
2367   // Don't call into the VM if we don't want to trace to speed up things.
2368   Label Lskip_vm_call;
2369   if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
2370     int offs1 = __ load_const_optimized(R11_scratch1, (address) &TraceBytecodesAt, R0, true);
2371     int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
2372     __ ld(R11_scratch1, offs1, R11_scratch1);
2373     __ lwa(R12_scratch2, offs2, R12_scratch2);
2374     __ cmpd(CCR0, R12_scratch2, R11_scratch1);
2375     __ blt(CCR0, Lskip_vm_call);
2376   }
2377 
2378   __ push(state);
2379   // Load 2 topmost expression stack values.
2380   __ ld(R6_ARG4, tsize*Interpreter::stackElementSize, R15_esp);
2381   __ ld(R5_ARG3, Interpreter::stackElementSize, R15_esp);
2382   __ mflr(R31);
2383   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), /* unused */ R4_ARG2, R5_ARG3, R6_ARG4, false);
2384   __ mtlr(R31);
2385   __ pop(state);
2386 
2387   if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
2388     __ bind(Lskip_vm_call);
2389   }
2390   __ blr();
2391   BLOCK_COMMENT("} trace_code");
2392   return entry;
2393 }
2394 
2395 void TemplateInterpreterGenerator::count_bytecode() {
2396   int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeCounter::_counter_value, R12_scratch2, true);
2397   __ lwz(R12_scratch2, offs, R11_scratch1);
2398   __ addi(R12_scratch2, R12_scratch2, 1);
2399   __ stw(R12_scratch2, offs, R11_scratch1);
2400 }
2401 
2402 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2403   int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeHistogram::_counters[t->bytecode()], R12_scratch2, true);
2404   __ lwz(R12_scratch2, offs, R11_scratch1);
2405   __ addi(R12_scratch2, R12_scratch2, 1);
2406   __ stw(R12_scratch2, offs, R11_scratch1);
2407 }
2408 
2409 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2410   const Register addr = R11_scratch1,
2411                  tmp  = R12_scratch2;
2412   // Get index, shift out old bytecode, bring in new bytecode, and store it.
2413   // _index = (_index >> log2_number_of_codes) |
2414   //          (bytecode << log2_number_of_codes);
2415   int offs1 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_index, tmp, true);
2416   __ lwz(tmp, offs1, addr);
2417   __ srwi(tmp, tmp, BytecodePairHistogram::log2_number_of_codes);
2418   __ ori(tmp, tmp, ((int) t->bytecode()) << BytecodePairHistogram::log2_number_of_codes);
2419   __ stw(tmp, offs1, addr);
2420 
2421   // Bump bucket contents.
2422   // _counters[_index] ++;
2423   int offs2 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_counters, R0, true);
2424   __ sldi(tmp, tmp, LogBytesPerInt);
2425   __ add(addr, tmp, addr);
2426   __ lwz(tmp, offs2, addr);
2427   __ addi(tmp, tmp, 1);
2428   __ stw(tmp, offs2, addr);
2429 }
2430 
2431 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2432   // Call a little run-time stub to avoid blow-up for each bytecode.
2433   // The run-time runtime saves the right registers, depending on
2434   // the tosca in-state for the given template.
2435 
2436   assert(Interpreter::trace_code(t->tos_in()) != NULL,
2437          "entry must have been generated");
2438 
2439   // Note: we destroy LR here.
2440   __ bl(Interpreter::trace_code(t->tos_in()));
2441 }
2442 
2443 void TemplateInterpreterGenerator::stop_interpreter_at() {
2444   Label L;
2445   int offs1 = __ load_const_optimized(R11_scratch1, (address) &StopInterpreterAt, R0, true);
2446   int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
2447   __ ld(R11_scratch1, offs1, R11_scratch1);
2448   __ lwa(R12_scratch2, offs2, R12_scratch2);
2449   __ cmpd(CCR0, R12_scratch2, R11_scratch1);
2450   __ bne(CCR0, L);
2451   __ illtrap();
2452   __ bind(L);
2453 }
2454 
2455 #endif // !PRODUCT