1 /*
   2  * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2012, 2026 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 "asm/macroAssembler.inline.hpp"
  27 #include "code/debugInfoRec.hpp"
  28 #include "code/compiledIC.hpp"
  29 #include "code/vtableStubs.hpp"
  30 #include "frame_ppc.hpp"
  31 #include "compiler/oopMap.hpp"
  32 #include "gc/shared/gcLocker.hpp"
  33 #include "interpreter/interpreter.hpp"
  34 #include "interpreter/interp_masm.hpp"
  35 #include "memory/resourceArea.hpp"
  36 #include "oops/klass.inline.hpp"
  37 #include "prims/methodHandles.hpp"
  38 #include "runtime/continuation.hpp"
  39 #include "runtime/continuationEntry.inline.hpp"
  40 #include "runtime/jniHandles.hpp"
  41 #include "runtime/os.inline.hpp"
  42 #include "runtime/safepointMechanism.hpp"
  43 #include "runtime/sharedRuntime.hpp"
  44 #include "runtime/signature.hpp"
  45 #include "runtime/stubRoutines.hpp"
  46 #include "runtime/timerTrace.hpp"
  47 #include "runtime/vframeArray.hpp"
  48 #include "utilities/align.hpp"
  49 #include "utilities/macros.hpp"
  50 #include "vmreg_ppc.inline.hpp"
  51 #ifdef COMPILER1
  52 #include "c1/c1_Runtime1.hpp"
  53 #endif
  54 #ifdef COMPILER2
  55 #include "opto/ad.hpp"
  56 #include "opto/runtime.hpp"
  57 #endif
  58 
  59 #include <alloca.h>
  60 
  61 #define __ masm->
  62 
  63 #ifdef PRODUCT
  64 #define BLOCK_COMMENT(str) // nothing
  65 #else
  66 #define BLOCK_COMMENT(str) __ block_comment(str)
  67 #endif
  68 
  69 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
  70 
  71 
  72 class RegisterSaver {
  73  // Used for saving volatile registers.
  74  public:
  75 
  76   // Support different return pc locations.
  77   enum ReturnPCLocation {
  78     return_pc_is_lr,
  79     return_pc_is_pre_saved,
  80     return_pc_is_thread_saved_exception_pc
  81   };
  82 
  83   static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
  84                          int* out_frame_size_in_bytes,
  85                          bool generate_oop_map,
  86                          ReturnPCLocation return_pc_location,
  87                          bool save_vectors = false);
  88   static void    restore_live_registers_and_pop_frame(MacroAssembler* masm,
  89                          int frame_size_in_bytes,
  90                          bool restore_ctr,
  91                          bool save_vectors = false);
  92 
  93   static void push_frame_and_save_argument_registers(MacroAssembler* masm,
  94                          Register r_temp,
  95                          int frame_size,
  96                          int total_args,
  97                          const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
  98   static void restore_argument_registers_and_pop_frame(MacroAssembler*masm,
  99                          int frame_size,
 100                          int total_args,
 101                          const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
 102 
 103   // During deoptimization only the result registers need to be restored
 104   // all the other values have already been extracted.
 105   static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes, bool save_vectors);
 106 
 107   // Constants and data structures:
 108 
 109   typedef enum {
 110     int_reg,
 111     float_reg,
 112     special_reg,
 113     vec_reg
 114   } RegisterType;
 115 
 116   typedef enum {
 117     reg_size          = 8,
 118     half_reg_size     = reg_size / 2,
 119     vec_reg_size      = 16
 120   } RegisterConstants;
 121 
 122   typedef struct {
 123     RegisterType        reg_type;
 124     int                 reg_num;
 125     VMReg               vmreg;
 126   } LiveRegType;
 127 };
 128 
 129 
 130 #define RegisterSaver_LiveIntReg(regname) \
 131   { RegisterSaver::int_reg,     regname->encoding(), regname->as_VMReg() }
 132 
 133 #define RegisterSaver_LiveFloatReg(regname) \
 134   { RegisterSaver::float_reg,   regname->encoding(), regname->as_VMReg() }
 135 
 136 #define RegisterSaver_LiveSpecialReg(regname) \
 137   { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() }
 138 
 139 #define RegisterSaver_LiveVecReg(regname) \
 140   { RegisterSaver::vec_reg,      regname->encoding(), regname->as_VMReg() }
 141 
 142 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
 143   // Live registers which get spilled to the stack. Register
 144   // positions in this array correspond directly to the stack layout.
 145 
 146   //
 147   // live special registers:
 148   //
 149   RegisterSaver_LiveSpecialReg(SR_CTR),
 150   //
 151   // live float registers:
 152   //
 153   RegisterSaver_LiveFloatReg( F0  ),
 154   RegisterSaver_LiveFloatReg( F1  ),
 155   RegisterSaver_LiveFloatReg( F2  ),
 156   RegisterSaver_LiveFloatReg( F3  ),
 157   RegisterSaver_LiveFloatReg( F4  ),
 158   RegisterSaver_LiveFloatReg( F5  ),
 159   RegisterSaver_LiveFloatReg( F6  ),
 160   RegisterSaver_LiveFloatReg( F7  ),
 161   RegisterSaver_LiveFloatReg( F8  ),
 162   RegisterSaver_LiveFloatReg( F9  ),
 163   RegisterSaver_LiveFloatReg( F10 ),
 164   RegisterSaver_LiveFloatReg( F11 ),
 165   RegisterSaver_LiveFloatReg( F12 ),
 166   RegisterSaver_LiveFloatReg( F13 ),
 167   RegisterSaver_LiveFloatReg( F14 ),
 168   RegisterSaver_LiveFloatReg( F15 ),
 169   RegisterSaver_LiveFloatReg( F16 ),
 170   RegisterSaver_LiveFloatReg( F17 ),
 171   RegisterSaver_LiveFloatReg( F18 ),
 172   RegisterSaver_LiveFloatReg( F19 ),
 173   RegisterSaver_LiveFloatReg( F20 ),
 174   RegisterSaver_LiveFloatReg( F21 ),
 175   RegisterSaver_LiveFloatReg( F22 ),
 176   RegisterSaver_LiveFloatReg( F23 ),
 177   RegisterSaver_LiveFloatReg( F24 ),
 178   RegisterSaver_LiveFloatReg( F25 ),
 179   RegisterSaver_LiveFloatReg( F26 ),
 180   RegisterSaver_LiveFloatReg( F27 ),
 181   RegisterSaver_LiveFloatReg( F28 ),
 182   RegisterSaver_LiveFloatReg( F29 ),
 183   RegisterSaver_LiveFloatReg( F30 ),
 184   RegisterSaver_LiveFloatReg( F31 ),
 185   //
 186   // live integer registers:
 187   //
 188   RegisterSaver_LiveIntReg(   R0  ),
 189   //RegisterSaver_LiveIntReg( R1  ), // stack pointer
 190   RegisterSaver_LiveIntReg(   R2  ),
 191   RegisterSaver_LiveIntReg(   R3  ),
 192   RegisterSaver_LiveIntReg(   R4  ),
 193   RegisterSaver_LiveIntReg(   R5  ),
 194   RegisterSaver_LiveIntReg(   R6  ),
 195   RegisterSaver_LiveIntReg(   R7  ),
 196   RegisterSaver_LiveIntReg(   R8  ),
 197   RegisterSaver_LiveIntReg(   R9  ),
 198   RegisterSaver_LiveIntReg(   R10 ),
 199   RegisterSaver_LiveIntReg(   R11 ),
 200   RegisterSaver_LiveIntReg(   R12 ),
 201   //RegisterSaver_LiveIntReg( R13 ), // system thread id
 202   RegisterSaver_LiveIntReg(   R14 ),
 203   RegisterSaver_LiveIntReg(   R15 ),
 204   RegisterSaver_LiveIntReg(   R16 ),
 205   RegisterSaver_LiveIntReg(   R17 ),
 206   RegisterSaver_LiveIntReg(   R18 ),
 207   RegisterSaver_LiveIntReg(   R19 ),
 208   RegisterSaver_LiveIntReg(   R20 ),
 209   RegisterSaver_LiveIntReg(   R21 ),
 210   RegisterSaver_LiveIntReg(   R22 ),
 211   RegisterSaver_LiveIntReg(   R23 ),
 212   RegisterSaver_LiveIntReg(   R24 ),
 213   RegisterSaver_LiveIntReg(   R25 ),
 214   RegisterSaver_LiveIntReg(   R26 ),
 215   RegisterSaver_LiveIntReg(   R27 ),
 216   RegisterSaver_LiveIntReg(   R28 ),
 217   RegisterSaver_LiveIntReg(   R29 ),
 218   RegisterSaver_LiveIntReg(   R30 ),
 219   RegisterSaver_LiveIntReg(   R31 )  // must be the last register (see save/restore functions below)
 220 };
 221 
 222 static const RegisterSaver::LiveRegType RegisterSaver_LiveVecRegs[] = {
 223   //
 224   // live vector registers (optional, only these ones are used by C2):
 225   //
 226   RegisterSaver_LiveVecReg( VR0 ),
 227   RegisterSaver_LiveVecReg( VR1 ),
 228   RegisterSaver_LiveVecReg( VR2 ),
 229   RegisterSaver_LiveVecReg( VR3 ),
 230   RegisterSaver_LiveVecReg( VR4 ),
 231   RegisterSaver_LiveVecReg( VR5 ),
 232   RegisterSaver_LiveVecReg( VR6 ),
 233   RegisterSaver_LiveVecReg( VR7 ),
 234   RegisterSaver_LiveVecReg( VR8 ),
 235   RegisterSaver_LiveVecReg( VR9 ),
 236   RegisterSaver_LiveVecReg( VR10 ),
 237   RegisterSaver_LiveVecReg( VR11 ),
 238   RegisterSaver_LiveVecReg( VR12 ),
 239   RegisterSaver_LiveVecReg( VR13 ),
 240   RegisterSaver_LiveVecReg( VR14 ),
 241   RegisterSaver_LiveVecReg( VR15 ),
 242   RegisterSaver_LiveVecReg( VR16 ),
 243   RegisterSaver_LiveVecReg( VR17 ),
 244   RegisterSaver_LiveVecReg( VR18 ),
 245   RegisterSaver_LiveVecReg( VR19 ),
 246   RegisterSaver_LiveVecReg( VR20 ),
 247   RegisterSaver_LiveVecReg( VR21 ),
 248   RegisterSaver_LiveVecReg( VR22 ),
 249   RegisterSaver_LiveVecReg( VR23 ),
 250   RegisterSaver_LiveVecReg( VR24 ),
 251   RegisterSaver_LiveVecReg( VR25 ),
 252   RegisterSaver_LiveVecReg( VR26 ),
 253   RegisterSaver_LiveVecReg( VR27 ),
 254   RegisterSaver_LiveVecReg( VR28 ),
 255   RegisterSaver_LiveVecReg( VR29 ),
 256   RegisterSaver_LiveVecReg( VR30 ),
 257   RegisterSaver_LiveVecReg( VR31 )
 258 };
 259 
 260 
 261 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
 262                          int* out_frame_size_in_bytes,
 263                          bool generate_oop_map,
 264                          ReturnPCLocation return_pc_location,
 265                          bool save_vectors) {
 266   // Push an abi_reg_args-frame and store all registers which may be live.
 267   // If requested, create an OopMap: Record volatile registers as
 268   // callee-save values in an OopMap so their save locations will be
 269   // propagated to the RegisterMap of the caller frame during
 270   // StackFrameStream construction (needed for deoptimization; see
 271   // compiledVFrame::create_stack_value).
 272   // Updated return pc is returned in R31 (if not return_pc_is_pre_saved).
 273 
 274   // calculate frame size
 275   const int regstosave_num       = sizeof(RegisterSaver_LiveRegs) /
 276                                    sizeof(RegisterSaver::LiveRegType);
 277   const int vecregstosave_num    = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
 278                                                    sizeof(RegisterSaver::LiveRegType))
 279                                                 : 0;
 280   const int register_save_size   = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
 281   const int frame_size_in_bytes  = align_up(register_save_size, frame::alignment_in_bytes)
 282                                    + frame::native_abi_reg_args_size;
 283 
 284   *out_frame_size_in_bytes       = frame_size_in_bytes;
 285   const int frame_size_in_slots  = frame_size_in_bytes / sizeof(jint);
 286   const int register_save_offset = frame_size_in_bytes - register_save_size;
 287 
 288   // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
 289   OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : nullptr;
 290 
 291   BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {");
 292 
 293   // push a new frame
 294   __ push_frame(frame_size_in_bytes, noreg);
 295 
 296   // Save some registers in the last (non-vector) slots of the new frame so we
 297   // can use them as scratch regs or to determine the return pc.
 298   __ std(R31, frame_size_in_bytes -   reg_size - vecregstosave_num * vec_reg_size, R1_SP);
 299   __ std(R30, frame_size_in_bytes - 2*reg_size - vecregstosave_num * vec_reg_size, R1_SP);
 300 
 301   // save the flags
 302   // Do the save_LR by hand and adjust the return pc if requested.
 303   switch (return_pc_location) {
 304     case return_pc_is_lr: __ mflr(R31); break;
 305     case return_pc_is_pre_saved: break;
 306     case return_pc_is_thread_saved_exception_pc: __ ld(R31, thread_(saved_exception_pc)); break;
 307     default: ShouldNotReachHere();
 308   }
 309   if (return_pc_location != return_pc_is_pre_saved) {
 310     __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
 311   }
 312 
 313   // save all registers (ints and floats)
 314   int offset = register_save_offset;
 315 
 316   for (int i = 0; i < regstosave_num; i++) {
 317     int reg_num  = RegisterSaver_LiveRegs[i].reg_num;
 318     int reg_type = RegisterSaver_LiveRegs[i].reg_type;
 319 
 320     switch (reg_type) {
 321       case RegisterSaver::int_reg: {
 322         if (reg_num < 30) { // We spilled R30-31 right at the beginning.
 323           __ std(as_Register(reg_num), offset, R1_SP);
 324         }
 325         break;
 326       }
 327       case RegisterSaver::float_reg: {
 328         __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
 329         break;
 330       }
 331       case RegisterSaver::special_reg: {
 332         if (reg_num == SR_CTR.encoding()) {
 333           __ mfctr(R30);
 334           __ std(R30, offset, R1_SP);
 335         } else {
 336           Unimplemented();
 337         }
 338         break;
 339       }
 340       default:
 341         ShouldNotReachHere();
 342     }
 343 
 344     if (generate_oop_map) {
 345       map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2),
 346                             RegisterSaver_LiveRegs[i].vmreg);
 347     }
 348     offset += reg_size;
 349   }
 350 
 351   // Note that generate_oop_map in the following loop is only used for the
 352   // polling_page_vectors_safepoint_handler_blob and the deopt_blob.
 353   // The order in which the vector contents are stored depends on Endianess and
 354   // the utilized instructions (PowerArchitecturePPC64).
 355   assert(is_aligned(offset, StackAlignmentInBytes), "should be");
 356   if (PowerArchitecturePPC64 >= 10) {
 357     assert(is_even(vecregstosave_num), "expectation");
 358     for (int i = 0; i < vecregstosave_num; i += 2) {
 359       int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
 360       assert(RegisterSaver_LiveVecRegs[i + 1].reg_num == reg_num + 1, "or use other instructions!");
 361 
 362       __ stxvp(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
 363       // Note: The contents were read in the same order (see loadV16 node in ppc.ad).
 364       // RegisterMap::pd_location only uses the first VMReg for each VectorRegister.
 365       if (generate_oop_map) {
 366         map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2),
 367                               RegisterSaver_LiveVecRegs[i LITTLE_ENDIAN_ONLY(+1) ].vmreg);
 368         map->set_callee_saved(VMRegImpl::stack2reg((offset + vec_reg_size) >> 2),
 369                               RegisterSaver_LiveVecRegs[i BIG_ENDIAN_ONLY(+1) ].vmreg);
 370       }
 371       offset += (2 * vec_reg_size);
 372     }
 373   } else {
 374     for (int i = 0; i < vecregstosave_num; i++) {
 375       int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
 376 
 377       __ stxv(as_VectorRegister(reg_num)->to_vsr(), offset, R1_SP);
 378       // Note: The contents were read in the same order (see loadV16 node in ppc.ad).
 379       // RegisterMap::pd_location only uses the first VMReg for each VectorRegister.
 380       if (generate_oop_map) {
 381         VMReg vsr = RegisterSaver_LiveVecRegs[i].vmreg;
 382         map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2), vsr);
 383       }
 384       offset += vec_reg_size;
 385     }
 386   }
 387 
 388   assert(offset == frame_size_in_bytes, "consistency check");
 389 
 390   BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers");
 391 
 392   // And we're done.
 393   return map;
 394 }
 395 
 396 
 397 // Pop the current frame and restore all the registers that we
 398 // saved.
 399 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
 400                                                          int frame_size_in_bytes,
 401                                                          bool restore_ctr,
 402                                                          bool save_vectors) {
 403   const int regstosave_num       = sizeof(RegisterSaver_LiveRegs) /
 404                                    sizeof(RegisterSaver::LiveRegType);
 405   const int vecregstosave_num    = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
 406                                                    sizeof(RegisterSaver::LiveRegType))
 407                                                 : 0;
 408   const int register_save_size   = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
 409 
 410   const int register_save_offset = frame_size_in_bytes - register_save_size;
 411 
 412   BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
 413 
 414   // restore all registers (ints and floats)
 415   int offset = register_save_offset;
 416 
 417   for (int i = 0; i < regstosave_num; i++) {
 418     int reg_num  = RegisterSaver_LiveRegs[i].reg_num;
 419     int reg_type = RegisterSaver_LiveRegs[i].reg_type;
 420 
 421     switch (reg_type) {
 422       case RegisterSaver::int_reg: {
 423         if (reg_num != 31) // R31 restored at the end, it's the tmp reg!
 424           __ ld(as_Register(reg_num), offset, R1_SP);
 425         break;
 426       }
 427       case RegisterSaver::float_reg: {
 428         __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
 429         break;
 430       }
 431       case RegisterSaver::special_reg: {
 432         if (reg_num == SR_CTR.encoding()) {
 433           if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
 434             __ ld(R31, offset, R1_SP);
 435             __ mtctr(R31);
 436           }
 437         } else {
 438           Unimplemented();
 439         }
 440         break;
 441       }
 442       default:
 443         ShouldNotReachHere();
 444     }
 445     offset += reg_size;
 446   }
 447 
 448   assert(is_aligned(offset, StackAlignmentInBytes), "should be");
 449   if (PowerArchitecturePPC64 >= 10) {
 450     for (int i = 0; i < vecregstosave_num; i += 2) {
 451       int reg_num  = RegisterSaver_LiveVecRegs[i].reg_num;
 452       assert(RegisterSaver_LiveVecRegs[i + 1].reg_num == reg_num + 1, "or use other instructions!");
 453 
 454       __ lxvp(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
 455 
 456       offset += (2 * vec_reg_size);
 457     }
 458   } else {
 459     for (int i = 0; i < vecregstosave_num; i++) {
 460       int reg_num  = RegisterSaver_LiveVecRegs[i].reg_num;
 461 
 462       __ lxv(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
 463 
 464       offset += vec_reg_size;
 465     }
 466   }
 467 
 468   assert(offset == frame_size_in_bytes, "consistency check");
 469 
 470   // restore link and the flags
 471   __ ld(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
 472   __ mtlr(R31);
 473 
 474   // restore scratch register's value
 475   __ ld(R31, frame_size_in_bytes - reg_size - vecregstosave_num * vec_reg_size, R1_SP);
 476 
 477   // pop the frame
 478   __ addi(R1_SP, R1_SP, frame_size_in_bytes);
 479 
 480   BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
 481 }
 482 
 483 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
 484                                                            int frame_size,int total_args, const VMRegPair *regs,
 485                                                            const VMRegPair *regs2) {
 486   __ push_frame(frame_size, r_temp);
 487   int st_off = frame_size - wordSize;
 488   for (int i = 0; i < total_args; i++) {
 489     VMReg r_1 = regs[i].first();
 490     VMReg r_2 = regs[i].second();
 491     if (!r_1->is_valid()) {
 492       assert(!r_2->is_valid(), "");
 493       continue;
 494     }
 495     if (r_1->is_Register()) {
 496       Register r = r_1->as_Register();
 497       __ std(r, st_off, R1_SP);
 498       st_off -= wordSize;
 499     } else if (r_1->is_FloatRegister()) {
 500       FloatRegister f = r_1->as_FloatRegister();
 501       __ stfd(f, st_off, R1_SP);
 502       st_off -= wordSize;
 503     }
 504   }
 505   if (regs2 != nullptr) {
 506     for (int i = 0; i < total_args; i++) {
 507       VMReg r_1 = regs2[i].first();
 508       VMReg r_2 = regs2[i].second();
 509       if (!r_1->is_valid()) {
 510         assert(!r_2->is_valid(), "");
 511         continue;
 512       }
 513       if (r_1->is_Register()) {
 514         Register r = r_1->as_Register();
 515         __ std(r, st_off, R1_SP);
 516         st_off -= wordSize;
 517       } else if (r_1->is_FloatRegister()) {
 518         FloatRegister f = r_1->as_FloatRegister();
 519         __ stfd(f, st_off, R1_SP);
 520         st_off -= wordSize;
 521       }
 522     }
 523   }
 524 }
 525 
 526 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
 527                                                              int total_args, const VMRegPair *regs,
 528                                                              const VMRegPair *regs2) {
 529   int st_off = frame_size - wordSize;
 530   for (int i = 0; i < total_args; i++) {
 531     VMReg r_1 = regs[i].first();
 532     VMReg r_2 = regs[i].second();
 533     if (r_1->is_Register()) {
 534       Register r = r_1->as_Register();
 535       __ ld(r, st_off, R1_SP);
 536       st_off -= wordSize;
 537     } else if (r_1->is_FloatRegister()) {
 538       FloatRegister f = r_1->as_FloatRegister();
 539       __ lfd(f, st_off, R1_SP);
 540       st_off -= wordSize;
 541     }
 542   }
 543   if (regs2 != nullptr)
 544     for (int i = 0; i < total_args; i++) {
 545       VMReg r_1 = regs2[i].first();
 546       VMReg r_2 = regs2[i].second();
 547       if (r_1->is_Register()) {
 548         Register r = r_1->as_Register();
 549         __ ld(r, st_off, R1_SP);
 550         st_off -= wordSize;
 551       } else if (r_1->is_FloatRegister()) {
 552         FloatRegister f = r_1->as_FloatRegister();
 553         __ lfd(f, st_off, R1_SP);
 554         st_off -= wordSize;
 555       }
 556     }
 557   __ pop_frame();
 558 }
 559 
 560 // Restore the registers that might be holding a result.
 561 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes, bool save_vectors) {
 562   const int regstosave_num       = sizeof(RegisterSaver_LiveRegs) /
 563                                    sizeof(RegisterSaver::LiveRegType);
 564   const int vecregstosave_num    = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
 565                                                    sizeof(RegisterSaver::LiveRegType))
 566                                                 : 0;
 567   const int register_save_size   = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
 568 
 569   const int register_save_offset = frame_size_in_bytes - register_save_size;
 570 
 571   // restore all result registers (ints and floats)
 572   int offset = register_save_offset;
 573   for (int i = 0; i < regstosave_num; i++) {
 574     int reg_num  = RegisterSaver_LiveRegs[i].reg_num;
 575     int reg_type = RegisterSaver_LiveRegs[i].reg_type;
 576     switch (reg_type) {
 577       case RegisterSaver::int_reg: {
 578         if (as_Register(reg_num)==R3_RET) // int result_reg
 579           __ ld(as_Register(reg_num), offset, R1_SP);
 580         break;
 581       }
 582       case RegisterSaver::float_reg: {
 583         if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
 584           __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
 585         break;
 586       }
 587       case RegisterSaver::special_reg: {
 588         // Special registers don't hold a result.
 589         break;
 590       }
 591       default:
 592         ShouldNotReachHere();
 593     }
 594     offset += reg_size;
 595   }
 596 
 597   assert(offset == frame_size_in_bytes - (save_vectors ? vecregstosave_num * vec_reg_size : 0), "consistency check");
 598 }
 599 
 600 // Is vector's size (in bytes) bigger than a size saved by default?
 601 bool SharedRuntime::is_wide_vector(int size) {
 602   // Note, MaxVectorSize == 8/16 on PPC64.
 603   assert(size <= (SuperwordUseVSX ? 16 : 8), "%d bytes vectors are not supported", size);
 604   return size > 8;
 605 }
 606 
 607 static int reg2slot(VMReg r) {
 608   return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
 609 }
 610 
 611 static int reg2offset(VMReg r) {
 612   return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
 613 }
 614 
 615 // ---------------------------------------------------------------------------
 616 // Read the array of BasicTypes from a signature, and compute where the
 617 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
 618 // quantities. Values less than VMRegImpl::stack0 are registers, those above
 619 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
 620 // as framesizes are fixed.
 621 // VMRegImpl::stack0 refers to the first slot 0(sp).
 622 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
 623 // up to Register::number_of_registers) are the 64-bit
 624 // integer registers.
 625 
 626 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
 627 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
 628 // units regardless of build. Of course for i486 there is no 64 bit build
 629 
 630 // The Java calling convention is a "shifted" version of the C ABI.
 631 // By skipping the first C ABI register we can call non-static jni methods
 632 // with small numbers of arguments without having to shuffle the arguments
 633 // at all. Since we control the java ABI we ought to at least get some
 634 // advantage out of it.
 635 
 636 const VMReg java_iarg_reg[8] = {
 637   R3->as_VMReg(),
 638   R4->as_VMReg(),
 639   R5->as_VMReg(),
 640   R6->as_VMReg(),
 641   R7->as_VMReg(),
 642   R8->as_VMReg(),
 643   R9->as_VMReg(),
 644   R10->as_VMReg()
 645 };
 646 
 647 const VMReg java_farg_reg[13] = {
 648   F1->as_VMReg(),
 649   F2->as_VMReg(),
 650   F3->as_VMReg(),
 651   F4->as_VMReg(),
 652   F5->as_VMReg(),
 653   F6->as_VMReg(),
 654   F7->as_VMReg(),
 655   F8->as_VMReg(),
 656   F9->as_VMReg(),
 657   F10->as_VMReg(),
 658   F11->as_VMReg(),
 659   F12->as_VMReg(),
 660   F13->as_VMReg()
 661 };
 662 
 663 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
 664 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
 665 
 666 STATIC_ASSERT(num_java_iarg_registers == Argument::n_int_register_parameters_j);
 667 STATIC_ASSERT(num_java_farg_registers == Argument::n_float_register_parameters_j);
 668 
 669 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
 670                                            VMRegPair *regs,
 671                                            int total_args_passed) {
 672   // C2c calling conventions for compiled-compiled calls.
 673   // Put 8 ints/longs into registers _AND_ 13 float/doubles into
 674   // registers _AND_ put the rest on the stack.
 675 
 676   const int inc_stk_for_intfloat   = 1; // 1 slots for ints and floats
 677   const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
 678 
 679   int i;
 680   VMReg reg;
 681   int stk = 0;
 682   int ireg = 0;
 683   int freg = 0;
 684 
 685   // We put the first 8 arguments into registers and the rest on the
 686   // stack, float arguments are already in their argument registers
 687   // due to c2c calling conventions (see calling_convention).
 688   for (int i = 0; i < total_args_passed; ++i) {
 689     switch(sig_bt[i]) {
 690     case T_BOOLEAN:
 691     case T_CHAR:
 692     case T_BYTE:
 693     case T_SHORT:
 694     case T_INT:
 695       if (ireg < num_java_iarg_registers) {
 696         // Put int/ptr in register
 697         reg = java_iarg_reg[ireg];
 698         ++ireg;
 699       } else {
 700         // Put int/ptr on stack.
 701         reg = VMRegImpl::stack2reg(stk);
 702         stk += inc_stk_for_intfloat;
 703       }
 704       regs[i].set1(reg);
 705       break;
 706     case T_LONG:
 707       assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
 708       if (ireg < num_java_iarg_registers) {
 709         // Put long in register.
 710         reg = java_iarg_reg[ireg];
 711         ++ireg;
 712       } else {
 713         // Put long on stack. They must be aligned to 2 slots.
 714         if (stk & 0x1) ++stk;
 715         reg = VMRegImpl::stack2reg(stk);
 716         stk += inc_stk_for_longdouble;
 717       }
 718       regs[i].set2(reg);
 719       break;
 720     case T_OBJECT:
 721     case T_ARRAY:
 722     case T_ADDRESS:
 723       if (ireg < num_java_iarg_registers) {
 724         // Put ptr in register.
 725         reg = java_iarg_reg[ireg];
 726         ++ireg;
 727       } else {
 728         // Put ptr on stack. Objects must be aligned to 2 slots too,
 729         // because "64-bit pointers record oop-ishness on 2 aligned
 730         // adjacent registers." (see OopFlow::build_oop_map).
 731         if (stk & 0x1) ++stk;
 732         reg = VMRegImpl::stack2reg(stk);
 733         stk += inc_stk_for_longdouble;
 734       }
 735       regs[i].set2(reg);
 736       break;
 737     case T_FLOAT:
 738       if (freg < num_java_farg_registers) {
 739         // Put float in register.
 740         reg = java_farg_reg[freg];
 741         ++freg;
 742       } else {
 743         // Put float on stack.
 744         reg = VMRegImpl::stack2reg(stk);
 745         stk += inc_stk_for_intfloat;
 746       }
 747       regs[i].set1(reg);
 748       break;
 749     case T_DOUBLE:
 750       assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
 751       if (freg < num_java_farg_registers) {
 752         // Put double in register.
 753         reg = java_farg_reg[freg];
 754         ++freg;
 755       } else {
 756         // Put double on stack. They must be aligned to 2 slots.
 757         if (stk & 0x1) ++stk;
 758         reg = VMRegImpl::stack2reg(stk);
 759         stk += inc_stk_for_longdouble;
 760       }
 761       regs[i].set2(reg);
 762       break;
 763     case T_VOID:
 764       // Do not count halves.
 765       regs[i].set_bad();
 766       break;
 767     default:
 768       ShouldNotReachHere();
 769     }
 770   }
 771   return stk;
 772 }
 773 
 774 // Calling convention for calling C code.
 775 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
 776                                         VMRegPair *regs,
 777                                         int total_args_passed) {
 778   // Calling conventions for C runtime calls and calls to JNI native methods.
 779   //
 780   // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
 781   // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
 782   // the first 13 flt/dbl's in the first 13 fp regs but additionally
 783   // copy flt/dbl to the stack if they are beyond the 8th argument.
 784 
 785   const VMReg iarg_reg[8] = {
 786     R3->as_VMReg(),
 787     R4->as_VMReg(),
 788     R5->as_VMReg(),
 789     R6->as_VMReg(),
 790     R7->as_VMReg(),
 791     R8->as_VMReg(),
 792     R9->as_VMReg(),
 793     R10->as_VMReg()
 794   };
 795 
 796   const VMReg farg_reg[13] = {
 797     F1->as_VMReg(),
 798     F2->as_VMReg(),
 799     F3->as_VMReg(),
 800     F4->as_VMReg(),
 801     F5->as_VMReg(),
 802     F6->as_VMReg(),
 803     F7->as_VMReg(),
 804     F8->as_VMReg(),
 805     F9->as_VMReg(),
 806     F10->as_VMReg(),
 807     F11->as_VMReg(),
 808     F12->as_VMReg(),
 809     F13->as_VMReg()
 810   };
 811 
 812   // Check calling conventions consistency.
 813   assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
 814          sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
 815          "consistency");
 816 
 817   const int additional_frame_header_slots = ((frame::native_abi_minframe_size - frame::jit_out_preserve_size)
 818                                             / VMRegImpl::stack_slot_size);
 819   const int float_offset_in_slots = Argument::float_on_stack_offset_in_bytes_c / VMRegImpl::stack_slot_size;
 820 
 821   VMReg reg;
 822   int arg = 0;
 823   int freg = 0;
 824   bool stack_used = false;
 825 
 826   for (int i = 0; i < total_args_passed; ++i, ++arg) {
 827     // Each argument corresponds to a slot in the Parameter Save Area (if not omitted)
 828     int stk = (arg * 2) + additional_frame_header_slots;
 829 
 830     switch(sig_bt[i]) {
 831     //
 832     // If arguments 0-7 are integers, they are passed in integer registers.
 833     // Argument i is placed in iarg_reg[i].
 834     //
 835     case T_BOOLEAN:
 836     case T_CHAR:
 837     case T_BYTE:
 838     case T_SHORT:
 839     case T_INT:
 840       // We must cast ints to longs and use full 64 bit stack slots
 841       // here.  Thus fall through, handle as long.
 842     case T_LONG:
 843     case T_OBJECT:
 844     case T_ARRAY:
 845     case T_ADDRESS:
 846     case T_METADATA:
 847       // Oops are already boxed if required (JNI).
 848       if (arg < Argument::n_int_register_parameters_c) {
 849         reg = iarg_reg[arg];
 850       } else {
 851         reg = VMRegImpl::stack2reg(stk);
 852         stack_used = true;
 853       }
 854       regs[i].set2(reg);
 855       break;
 856 
 857     //
 858     // Floats are treated differently from int regs:  The first 13 float arguments
 859     // are passed in registers (not the float args among the first 13 args).
 860     // Thus argument i is NOT passed in farg_reg[i] if it is float.  It is passed
 861     // in farg_reg[j] if argument i is the j-th float argument of this call.
 862     //
 863     case T_FLOAT:
 864       if (freg < Argument::n_float_register_parameters_c) {
 865         // Put float in register ...
 866         reg = farg_reg[freg];
 867         ++freg;
 868       } else {
 869         // Put float on stack.
 870         reg = VMRegImpl::stack2reg(stk + float_offset_in_slots);
 871         stack_used = true;
 872       }
 873       regs[i].set1(reg);
 874       break;
 875     case T_DOUBLE:
 876       assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
 877       if (freg < Argument::n_float_register_parameters_c) {
 878         // Put double in register ...
 879         reg = farg_reg[freg];
 880         ++freg;
 881       } else {
 882         // Put double on stack.
 883         reg = VMRegImpl::stack2reg(stk);
 884         stack_used = true;
 885       }
 886       regs[i].set2(reg);
 887       break;
 888 
 889     case T_VOID:
 890       // Do not count halves.
 891       regs[i].set_bad();
 892       --arg;
 893       break;
 894     default:
 895       ShouldNotReachHere();
 896     }
 897   }
 898 
 899   // Return size of the stack frame excluding the jit_out_preserve part in single-word slots.
 900 #if defined(ABI_ELFv2)
 901   assert(additional_frame_header_slots == 0, "ABIv2 shouldn't use extra slots");
 902   // ABIv2 allows omitting the Parameter Save Area if the callee's prototype
 903   // indicates that all parameters can be passed in registers.
 904   return stack_used ? (arg * 2) : 0;
 905 #else
 906   // The Parameter Save Area needs to be at least 8 double-word slots for ABIv1.
 907   // We have to add extra slots because ABIv1 uses a larger header.
 908   return MAX2(arg, 8) * 2 + additional_frame_header_slots;
 909 #endif
 910 }
 911 
 912 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
 913                                              uint num_bits,
 914                                              uint total_args_passed) {
 915   Unimplemented();
 916   return 0;
 917 }
 918 
 919 static address gen_c2i_adapter(MacroAssembler *masm,
 920                                int comp_args_on_stack,
 921                                const GrowableArray<SigEntry>* sig,
 922                                const VMRegPair *regs,
 923                                Label& call_interpreter,
 924                                const Register& ientry) {
 925 
 926   address c2i_entrypoint;
 927 
 928   const Register sender_SP = R21_sender_SP; // == R21_tmp1
 929   const Register code      = R22_tmp2;
 930   //const Register ientry  = R23_tmp3;
 931   const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
 932   const int num_value_regs = sizeof(value_regs) / sizeof(Register);
 933   int value_regs_index = 0;
 934   int total_args_passed = sig->length();
 935 
 936   const Register return_pc = R27_tmp7;
 937   const Register tmp       = R28_tmp8;
 938 
 939   assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
 940 
 941   // Adapter needs TOP_IJAVA_FRAME_ABI.
 942   const int adapter_size = frame::top_ijava_frame_abi_size +
 943                            align_up(total_args_passed * wordSize, frame::alignment_in_bytes);
 944 
 945   // regular (verified) c2i entry point
 946   c2i_entrypoint = __ pc();
 947 
 948   // Does compiled code exists? If yes, patch the caller's callsite.
 949   __ ld(code, method_(code));
 950   __ cmpdi(CR0, code, 0);
 951   __ ld(ientry, method_(interpreter_entry)); // preloaded
 952   __ beq(CR0, call_interpreter);
 953 
 954 
 955   // Patch caller's callsite, method_(code) was not null which means that
 956   // compiled code exists.
 957   __ mflr(return_pc);
 958   __ std(return_pc, _abi0(lr), R1_SP);
 959   RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
 960 
 961   __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
 962 
 963   RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
 964   __ ld(return_pc, _abi0(lr), R1_SP);
 965   __ ld(ientry, method_(interpreter_entry)); // preloaded
 966   __ mtlr(return_pc);
 967 
 968 
 969   // Call the interpreter.
 970   __ BIND(call_interpreter);
 971   __ mtctr(ientry);
 972 
 973   // Get a copy of the current SP for loading caller's arguments.
 974   __ mr(sender_SP, R1_SP);
 975 
 976   // Add space for the adapter.
 977   __ resize_frame(-adapter_size, R12_scratch2);
 978 
 979   int st_off = adapter_size - wordSize;
 980 
 981   // Write the args into the outgoing interpreter space.
 982   for (int i = 0; i < total_args_passed; i++) {
 983     BasicType bt = sig->at(i)._bt;
 984 
 985     VMReg r_1 = regs[i].first();
 986     VMReg r_2 = regs[i].second();
 987     if (!r_1->is_valid()) {
 988       assert(!r_2->is_valid(), "");
 989       continue;
 990     }
 991     if (r_1->is_stack()) {
 992       Register tmp_reg = value_regs[value_regs_index];
 993       value_regs_index = (value_regs_index + 1) % num_value_regs;
 994       // The calling convention produces OptoRegs that ignore the out
 995       // preserve area (JIT's ABI). We must account for it here.
 996       int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
 997       if (!r_2->is_valid()) {
 998         __ lwz(tmp_reg, ld_off, sender_SP);
 999       } else {
1000         __ ld(tmp_reg, ld_off, sender_SP);
1001       }
1002       // Pretend stack targets were loaded into tmp_reg.
1003       r_1 = tmp_reg->as_VMReg();
1004     }
1005 
1006     if (r_1->is_Register()) {
1007       Register r = r_1->as_Register();
1008       if (!r_2->is_valid()) {
1009         __ stw(r, st_off, R1_SP);
1010         st_off-=wordSize;
1011       } else {
1012         // Longs are given 2 64-bit slots in the interpreter, but the
1013         // data is passed in only 1 slot.
1014         if (bt == T_LONG || bt == T_DOUBLE) {
1015           DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
1016           st_off-=wordSize;
1017         }
1018         __ std(r, st_off, R1_SP);
1019         st_off-=wordSize;
1020       }
1021     } else {
1022       assert(r_1->is_FloatRegister(), "");
1023       FloatRegister f = r_1->as_FloatRegister();
1024       if (!r_2->is_valid()) {
1025         __ stfs(f, st_off, R1_SP);
1026         st_off-=wordSize;
1027       } else {
1028         // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
1029         // data is passed in only 1 slot.
1030         // One of these should get known junk...
1031         DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
1032         st_off-=wordSize;
1033         __ stfd(f, st_off, R1_SP);
1034         st_off-=wordSize;
1035       }
1036     }
1037   }
1038 
1039   // Jump to the interpreter just as if interpreter was doing it.
1040 
1041   __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1042 
1043   // load TOS
1044   __ addi(R15_esp, R1_SP, st_off);
1045 
1046   // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
1047   assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
1048   __ bctr();
1049 
1050   return c2i_entrypoint;
1051 }
1052 
1053 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, int comp_args_on_stack, const GrowableArray<SigEntry>* sig, const VMRegPair *regs) {
1054 
1055   // Load method's entry-point from method.
1056   __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
1057   __ mtctr(R12_scratch2);
1058 
1059   // We will only enter here from an interpreted frame and never from after
1060   // passing thru a c2i. Azul allowed this but we do not. If we lose the
1061   // race and use a c2i we will remain interpreted for the race loser(s).
1062   // This removes all sorts of headaches on the x86 side and also eliminates
1063   // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
1064 
1065   // Note: r13 contains the senderSP on entry. We must preserve it since
1066   // we may do a i2c -> c2i transition if we lose a race where compiled
1067   // code goes non-entrant while we get args ready.
1068   // In addition we use r13 to locate all the interpreter args as
1069   // we must align the stack to 16 bytes on an i2c entry else we
1070   // lose alignment we expect in all compiled code and register
1071   // save code can segv when fxsave instructions find improperly
1072   // aligned stack pointer.
1073 
1074   const Register ld_ptr = R15_esp;
1075   const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
1076   const int num_value_regs = sizeof(value_regs) / sizeof(Register);
1077   int value_regs_index = 0;
1078 
1079   int total_args_passed = sig->length();
1080   int ld_offset = total_args_passed*wordSize;
1081 
1082   // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1083   // in registers, we will occasionally have no stack args.
1084   int comp_words_on_stack = 0;
1085   if (comp_args_on_stack) {
1086     // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1087     // registers are below. By subtracting stack0, we either get a negative
1088     // number (all values in registers) or the maximum stack slot accessed.
1089 
1090     // Convert 4-byte c2 stack slots to words.
1091     comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1092     // Round up to miminum stack alignment, in wordSize.
1093     comp_words_on_stack = align_up(comp_words_on_stack, 2);
1094     __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1095   }
1096 
1097   // Now generate the shuffle code.  Pick up all register args and move the
1098   // rest through register value=Z_R12.
1099   BLOCK_COMMENT("Shuffle arguments");
1100 
1101   for (int i = 0; i < total_args_passed; i++) {
1102     BasicType bt = sig->at(i)._bt;
1103     if (bt == T_VOID) {
1104       assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
1105       continue;
1106     }
1107 
1108     // Pick up 0, 1 or 2 words from ld_ptr.
1109     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1110             "scrambled load targets?");
1111     VMReg r_1 = regs[i].first();
1112     VMReg r_2 = regs[i].second();
1113     if (!r_1->is_valid()) {
1114       assert(!r_2->is_valid(), "");
1115       continue;
1116     }
1117     if (r_1->is_FloatRegister()) {
1118       if (!r_2->is_valid()) {
1119         __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1120         ld_offset-=wordSize;
1121       } else {
1122         // Skip the unused interpreter slot.
1123         __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1124         ld_offset-=2*wordSize;
1125       }
1126     } else {
1127       Register r;
1128       if (r_1->is_stack()) {
1129         // Must do a memory to memory move thru "value".
1130         r = value_regs[value_regs_index];
1131         value_regs_index = (value_regs_index + 1) % num_value_regs;
1132       } else {
1133         r = r_1->as_Register();
1134       }
1135       if (!r_2->is_valid()) {
1136         // Not sure we need to do this but it shouldn't hurt.
1137         if (is_reference_type(bt) || bt == T_ADDRESS) {
1138           __ ld(r, ld_offset, ld_ptr);
1139           ld_offset-=wordSize;
1140         } else {
1141           __ lwz(r, ld_offset, ld_ptr);
1142           ld_offset-=wordSize;
1143         }
1144       } else {
1145         // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1146         // data is passed in only 1 slot.
1147         if (bt == T_LONG || bt == T_DOUBLE) {
1148           ld_offset-=wordSize;
1149         }
1150         __ ld(r, ld_offset, ld_ptr);
1151         ld_offset-=wordSize;
1152       }
1153 
1154       if (r_1->is_stack()) {
1155         // Now store value where the compiler expects it
1156         int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1157 
1158         if (bt == T_INT   || bt == T_FLOAT || bt == T_BOOLEAN ||
1159             bt == T_SHORT || bt == T_CHAR  || bt == T_BYTE) {
1160           __ stw(r, st_off, R1_SP);
1161         } else {
1162           __ std(r, st_off, R1_SP);
1163         }
1164       }
1165     }
1166   }
1167 
1168   __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1169 
1170   BLOCK_COMMENT("Store method");
1171   // Store method into thread->callee_target.
1172   // We might end up in handle_wrong_method if the callee is
1173   // deoptimized as we race thru here. If that happens we don't want
1174   // to take a safepoint because the caller frame will look
1175   // interpreted and arguments are now "compiled" so it is much better
1176   // to make this transition invisible to the stack walking
1177   // code. Unfortunately if we try and find the callee by normal means
1178   // a safepoint is possible. So we stash the desired callee in the
1179   // thread and the vm will find there should this case occur.
1180   __ std(R19_method, thread_(callee_target));
1181 
1182   // Jump to the compiled code just as if compiled code was doing it.
1183   __ bctr();
1184 }
1185 
1186 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
1187                                             int comp_args_on_stack,
1188                                             const GrowableArray<SigEntry>* sig,
1189                                             const VMRegPair* regs,
1190                                             const GrowableArray<SigEntry>* sig_cc,
1191                                             const VMRegPair* regs_cc,
1192                                             const GrowableArray<SigEntry>* sig_cc_ro,
1193                                             const VMRegPair* regs_cc_ro,
1194                                             address entry_address[AdapterBlob::ENTRY_COUNT],
1195                                             AdapterBlob*& new_adapter,
1196                                             bool allocate_code_blob) {
1197   // entry: i2c
1198 
1199   __ align(CodeEntryAlignment);
1200   entry_address[AdapterBlob::I2C] = __ pc();
1201   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
1202 
1203 
1204   // entry: c2i unverified
1205 
1206   __ align(CodeEntryAlignment);
1207   BLOCK_COMMENT("c2i unverified entry");
1208   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
1209 
1210   // inline_cache contains a CompiledICData
1211   const Register ic             = R19_inline_cache_reg;
1212   const Register ic_klass       = R11_scratch1;
1213   const Register receiver_klass = R12_scratch2;
1214   const Register code           = R21_tmp1;
1215   const Register ientry         = R23_tmp3;
1216 
1217   assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1218   assert(R11_scratch1 == R11, "need prologue scratch register");
1219 
1220   Label call_interpreter;
1221 
1222   __ ic_check(4 /* end_alignment */);
1223   __ ld(R19_method, CompiledICData::speculated_method_offset(), ic);
1224   // Argument is valid and klass is as expected, continue.
1225 
1226   __ ld(code, method_(code));
1227   __ cmpdi(CR0, code, 0);
1228   __ ld(ientry, method_(interpreter_entry)); // preloaded
1229   __ beq_predict_taken(CR0, call_interpreter);
1230 
1231   // Branch to ic_miss_stub.
1232   __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1233 
1234   // entry: c2i
1235 
1236   entry_address[AdapterBlob::C2I] = __ pc();
1237 
1238   // Class initialization barrier for static methods
1239   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
1240   assert(VM_Version::supports_fast_class_init_checks(), "sanity");
1241   Label L_skip_barrier;
1242 
1243   // Bypass the barrier for non-static methods
1244   __ lhz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1245   __ andi_(R0, R0, JVM_ACC_STATIC);
1246   __ beq(CR0, L_skip_barrier); // non-static
1247 
1248   Register klass = R11_scratch1;
1249   __ load_method_holder(klass, R19_method);
1250   __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
1251 
1252   __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
1253   __ mtctr(klass);
1254   __ bctr();
1255 
1256   __ bind(L_skip_barrier);
1257   entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();
1258 
1259   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1260   bs->c2i_entry_barrier(masm, /* tmp register*/ ic_klass, /* tmp register*/ receiver_klass, /* tmp register*/ code);
1261 
1262   gen_c2i_adapter(masm, comp_args_on_stack, sig, regs, call_interpreter, ientry);
1263   return;
1264 }
1265 
1266 // An oop arg. Must pass a handle not the oop itself.
1267 static void object_move(MacroAssembler* masm,
1268                         int frame_size_in_slots,
1269                         OopMap* oop_map, int oop_handle_offset,
1270                         bool is_receiver, int* receiver_offset,
1271                         VMRegPair src, VMRegPair dst,
1272                         Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1273   assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1274          "receiver has already been moved");
1275 
1276   // We must pass a handle. First figure out the location we use as a handle.
1277 
1278   if (src.first()->is_stack()) {
1279     // stack to stack or reg
1280 
1281     const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1282     Label skip;
1283     const int oop_slot_in_callers_frame = reg2slot(src.first());
1284 
1285     guarantee(!is_receiver, "expecting receiver in register");
1286     oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1287 
1288     __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1289     __ ld(  r_temp_2, reg2offset(src.first()), r_caller_sp);
1290     __ cmpdi(CR0, r_temp_2, 0);
1291     __ bne(CR0, skip);
1292     // Use a null handle if oop is null.
1293     __ li(r_handle, 0);
1294     __ bind(skip);
1295 
1296     if (dst.first()->is_stack()) {
1297       // stack to stack
1298       __ std(r_handle, reg2offset(dst.first()), R1_SP);
1299     } else {
1300       // stack to reg
1301       // Nothing to do, r_handle is already the dst register.
1302     }
1303   } else {
1304     // reg to stack or reg
1305     const Register r_oop      = src.first()->as_Register();
1306     const Register r_handle   = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1307     const int oop_slot        = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1308                                 + oop_handle_offset; // in slots
1309     const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1310     Label skip;
1311 
1312     if (is_receiver) {
1313       *receiver_offset = oop_offset;
1314     }
1315     oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1316 
1317     __ std( r_oop,    oop_offset, R1_SP);
1318     __ addi(r_handle, R1_SP, oop_offset);
1319 
1320     __ cmpdi(CR0, r_oop, 0);
1321     __ bne(CR0, skip);
1322     // Use a null handle if oop is null.
1323     __ li(r_handle, 0);
1324     __ bind(skip);
1325 
1326     if (dst.first()->is_stack()) {
1327       // reg to stack
1328       __ std(r_handle, reg2offset(dst.first()), R1_SP);
1329     } else {
1330       // reg to reg
1331       // Nothing to do, r_handle is already the dst register.
1332     }
1333   }
1334 }
1335 
1336 static void int_move(MacroAssembler*masm,
1337                      VMRegPair src, VMRegPair dst,
1338                      Register r_caller_sp, Register r_temp) {
1339   assert(src.first()->is_valid(), "incoming must be int");
1340   assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1341 
1342   if (src.first()->is_stack()) {
1343     if (dst.first()->is_stack()) {
1344       // stack to stack
1345       __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1346       __ std(r_temp, reg2offset(dst.first()), R1_SP);
1347     } else {
1348       // stack to reg
1349       __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1350     }
1351   } else if (dst.first()->is_stack()) {
1352     // reg to stack
1353     __ extsw(r_temp, src.first()->as_Register());
1354     __ std(r_temp, reg2offset(dst.first()), R1_SP);
1355   } else {
1356     // reg to reg
1357     __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1358   }
1359 }
1360 
1361 static void long_move(MacroAssembler*masm,
1362                       VMRegPair src, VMRegPair dst,
1363                       Register r_caller_sp, Register r_temp) {
1364   assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1365   assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1366 
1367   if (src.first()->is_stack()) {
1368     if (dst.first()->is_stack()) {
1369       // stack to stack
1370       __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1371       __ std(r_temp, reg2offset(dst.first()), R1_SP);
1372     } else {
1373       // stack to reg
1374       __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1375     }
1376   } else if (dst.first()->is_stack()) {
1377     // reg to stack
1378     __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1379   } else {
1380     // reg to reg
1381     if (dst.first()->as_Register() != src.first()->as_Register())
1382       __ mr(dst.first()->as_Register(), src.first()->as_Register());
1383   }
1384 }
1385 
1386 static void float_move(MacroAssembler*masm,
1387                        VMRegPair src, VMRegPair dst,
1388                        Register r_caller_sp, Register r_temp) {
1389   assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1390   assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1391 
1392   if (src.first()->is_stack()) {
1393     if (dst.first()->is_stack()) {
1394       // stack to stack
1395       __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1396       __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1397     } else {
1398       // stack to reg
1399       __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1400     }
1401   } else if (dst.first()->is_stack()) {
1402     // reg to stack
1403     __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1404   } else {
1405     // reg to reg
1406     if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1407       __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1408   }
1409 }
1410 
1411 static void double_move(MacroAssembler*masm,
1412                         VMRegPair src, VMRegPair dst,
1413                         Register r_caller_sp, Register r_temp) {
1414   assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1415   assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1416 
1417   if (src.first()->is_stack()) {
1418     if (dst.first()->is_stack()) {
1419       // stack to stack
1420       __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1421       __ std(r_temp, reg2offset(dst.first()), R1_SP);
1422     } else {
1423       // stack to reg
1424       __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1425     }
1426   } else if (dst.first()->is_stack()) {
1427     // reg to stack
1428     __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1429   } else {
1430     // reg to reg
1431     if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1432       __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1433   }
1434 }
1435 
1436 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1437   switch (ret_type) {
1438     case T_BOOLEAN:
1439     case T_CHAR:
1440     case T_BYTE:
1441     case T_SHORT:
1442     case T_INT:
1443       __ stw (R3_RET,  frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1444       break;
1445     case T_ARRAY:
1446     case T_OBJECT:
1447     case T_LONG:
1448       __ std (R3_RET,  frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1449       break;
1450     case T_FLOAT:
1451       __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1452       break;
1453     case T_DOUBLE:
1454       __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1455       break;
1456     case T_VOID:
1457       break;
1458     default:
1459       ShouldNotReachHere();
1460       break;
1461   }
1462 }
1463 
1464 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1465   switch (ret_type) {
1466     case T_BOOLEAN:
1467     case T_CHAR:
1468     case T_BYTE:
1469     case T_SHORT:
1470     case T_INT:
1471       __ lwz(R3_RET,  frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1472       break;
1473     case T_ARRAY:
1474     case T_OBJECT:
1475     case T_LONG:
1476       __ ld (R3_RET,  frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1477       break;
1478     case T_FLOAT:
1479       __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1480       break;
1481     case T_DOUBLE:
1482       __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1483       break;
1484     case T_VOID:
1485       break;
1486     default:
1487       ShouldNotReachHere();
1488       break;
1489   }
1490 }
1491 
1492 static void verify_oop_args(MacroAssembler* masm,
1493                             const methodHandle& method,
1494                             const BasicType* sig_bt,
1495                             const VMRegPair* regs) {
1496   Register temp_reg = R19_method;  // not part of any compiled calling seq
1497   if (VerifyOops) {
1498     for (int i = 0; i < method->size_of_parameters(); i++) {
1499       if (is_reference_type(sig_bt[i])) {
1500         VMReg r = regs[i].first();
1501         assert(r->is_valid(), "bad oop arg");
1502         if (r->is_stack()) {
1503           __ ld(temp_reg, reg2offset(r), R1_SP);
1504           __ verify_oop(temp_reg, FILE_AND_LINE);
1505         } else {
1506           __ verify_oop(r->as_Register(), FILE_AND_LINE);
1507         }
1508       }
1509     }
1510   }
1511 }
1512 
1513 static void gen_special_dispatch(MacroAssembler* masm,
1514                                  const methodHandle& method,
1515                                  const BasicType* sig_bt,
1516                                  const VMRegPair* regs) {
1517   verify_oop_args(masm, method, sig_bt, regs);
1518   vmIntrinsics::ID iid = method->intrinsic_id();
1519 
1520   // Now write the args into the outgoing interpreter space
1521   bool     has_receiver   = false;
1522   Register receiver_reg   = noreg;
1523   int      member_arg_pos = -1;
1524   Register member_reg     = noreg;
1525   int      ref_kind       = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1526   if (ref_kind != 0) {
1527     member_arg_pos = method->size_of_parameters() - 1;  // trailing MemberName argument
1528     member_reg = R19_method;  // known to be free at this point
1529     has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1530   } else if (iid == vmIntrinsics::_invokeBasic) {
1531     has_receiver = true;
1532   } else if (iid == vmIntrinsics::_linkToNative) {
1533     member_arg_pos = method->size_of_parameters() - 1;  // trailing NativeEntryPoint argument
1534     member_reg = R19_method;  // known to be free at this point
1535   } else {
1536     fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1537   }
1538 
1539   if (member_reg != noreg) {
1540     // Load the member_arg into register, if necessary.
1541     SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1542     VMReg r = regs[member_arg_pos].first();
1543     if (r->is_stack()) {
1544       __ ld(member_reg, reg2offset(r), R1_SP);
1545     } else {
1546       // no data motion is needed
1547       member_reg = r->as_Register();
1548     }
1549   }
1550 
1551   if (has_receiver) {
1552     // Make sure the receiver is loaded into a register.
1553     assert(method->size_of_parameters() > 0, "oob");
1554     assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1555     VMReg r = regs[0].first();
1556     assert(r->is_valid(), "bad receiver arg");
1557     if (r->is_stack()) {
1558       // Porting note:  This assumes that compiled calling conventions always
1559       // pass the receiver oop in a register.  If this is not true on some
1560       // platform, pick a temp and load the receiver from stack.
1561       fatal("receiver always in a register");
1562       receiver_reg = R11_scratch1;  // TODO (hs24): is R11_scratch1 really free at this point?
1563       __ ld(receiver_reg, reg2offset(r), R1_SP);
1564     } else {
1565       // no data motion is needed
1566       receiver_reg = r->as_Register();
1567     }
1568   }
1569 
1570   // Figure out which address we are really jumping to:
1571   MethodHandles::generate_method_handle_dispatch(masm, iid,
1572                                                  receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1573 }
1574 
1575 //---------------------------- continuation_enter_setup ---------------------------
1576 //
1577 // Frame setup.
1578 //
1579 // Arguments:
1580 //   None.
1581 //
1582 // Results:
1583 //   R1_SP: pointer to blank ContinuationEntry in the pushed frame.
1584 //
1585 // Kills:
1586 //   R0, R20
1587 //
1588 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& framesize_words) {
1589   assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1590   assert(in_bytes(ContinuationEntry::cont_offset())  % VMRegImpl::stack_slot_size == 0, "");
1591   assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1592 
1593   const int frame_size_in_bytes = (int)ContinuationEntry::size();
1594   assert(is_aligned(frame_size_in_bytes, frame::alignment_in_bytes), "alignment error");
1595 
1596   framesize_words = frame_size_in_bytes / wordSize;
1597 
1598   DEBUG_ONLY(__ block_comment("setup {"));
1599   // Save return pc and push entry frame
1600   const Register return_pc = R20;
1601   __ mflr(return_pc);
1602   __ std(return_pc, _abi0(lr), R1_SP);     // SP->lr = return_pc
1603   __ push_frame(frame_size_in_bytes , R0); // SP -= frame_size_in_bytes
1604 
1605   OopMap* map = new OopMap((int)frame_size_in_bytes / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1606 
1607   __ ld_ptr(R0, JavaThread::cont_entry_offset(), R16_thread);
1608   __ st_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1609   __ st_ptr(R0, ContinuationEntry::parent_offset(), R1_SP);
1610   DEBUG_ONLY(__ block_comment("} setup"));
1611 
1612   return map;
1613 }
1614 
1615 //---------------------------- fill_continuation_entry ---------------------------
1616 //
1617 // Initialize the new ContinuationEntry.
1618 //
1619 // Arguments:
1620 //   R1_SP: pointer to blank Continuation entry
1621 //   reg_cont_obj: pointer to the continuation
1622 //   reg_flags: flags
1623 //
1624 // Results:
1625 //   R1_SP: pointer to filled out ContinuationEntry
1626 //
1627 // Kills:
1628 //   R8_ARG6, R9_ARG7, R10_ARG8
1629 //
1630 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1631   assert_different_registers(reg_cont_obj, reg_flags);
1632   Register zero = R8_ARG6;
1633   Register tmp2 = R9_ARG7;
1634 
1635   DEBUG_ONLY(__ block_comment("fill {"));
1636 #ifdef ASSERT
1637   __ load_const_optimized(tmp2, ContinuationEntry::cookie_value());
1638   __ stw(tmp2, in_bytes(ContinuationEntry::cookie_offset()), R1_SP);
1639 #endif //ASSERT
1640 
1641   __ li(zero, 0);
1642   __ st_ptr(reg_cont_obj, ContinuationEntry::cont_offset(), R1_SP);
1643   __ stw(reg_flags, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1644   __ st_ptr(zero, ContinuationEntry::chunk_offset(), R1_SP);
1645   __ stw(zero, in_bytes(ContinuationEntry::argsize_offset()), R1_SP);
1646   __ stw(zero, in_bytes(ContinuationEntry::pin_count_offset()), R1_SP);
1647 
1648   __ ld_ptr(tmp2, JavaThread::cont_fastpath_offset(), R16_thread);
1649   __ st_ptr(tmp2, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1650 
1651   __ st_ptr(zero, JavaThread::cont_fastpath_offset(), R16_thread);
1652   DEBUG_ONLY(__ block_comment("} fill"));
1653 }
1654 
1655 //---------------------------- continuation_enter_cleanup ---------------------------
1656 //
1657 // Copy corresponding attributes from the top ContinuationEntry to the JavaThread
1658 // before deleting it.
1659 //
1660 // Arguments:
1661 //   R1_SP: pointer to the ContinuationEntry
1662 //
1663 // Results:
1664 //   None.
1665 //
1666 // Kills:
1667 //   R8_ARG6, R9_ARG7, R10_ARG8, R15_esp
1668 //
1669 static void continuation_enter_cleanup(MacroAssembler* masm) {
1670   Register tmp1 = R8_ARG6;
1671   Register tmp2 = R9_ARG7;
1672 
1673 #ifdef ASSERT
1674   __ block_comment("clean {");
1675   __ ld_ptr(tmp1, JavaThread::cont_entry_offset(), R16_thread);
1676   __ cmpd(CR0, R1_SP, tmp1);
1677   __ asm_assert_eq(FILE_AND_LINE ": incorrect R1_SP");
1678 #endif
1679 
1680   __ ld_ptr(tmp1, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1681   __ st_ptr(tmp1, JavaThread::cont_fastpath_offset(), R16_thread);
1682   __ ld_ptr(tmp2, ContinuationEntry::parent_offset(), R1_SP);
1683   __ st_ptr(tmp2, JavaThread::cont_entry_offset(), R16_thread);
1684   DEBUG_ONLY(__ block_comment("} clean"));
1685 }
1686 
1687 static void check_continuation_enter_argument(VMReg actual_vmreg,
1688                                               Register expected_reg,
1689                                               const char* name) {
1690   assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1691   assert(actual_vmreg->as_Register() == expected_reg,
1692          "%s is in unexpected register: %s instead of %s",
1693          name, actual_vmreg->as_Register()->name(), expected_reg->name());
1694 }
1695 
1696 static void gen_continuation_enter(MacroAssembler* masm,
1697                                    const VMRegPair* regs,
1698                                    int& exception_offset,
1699                                    OopMapSet* oop_maps,
1700                                    int& frame_complete,
1701                                    int& framesize_words,
1702                                    int& interpreted_entry_offset,
1703                                    int& compiled_entry_offset) {
1704 
1705   // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1706   int pos_cont_obj   = 0;
1707   int pos_is_cont    = 1;
1708   int pos_is_virtual = 2;
1709 
1710   // The platform-specific calling convention may present the arguments in various registers.
1711   // To simplify the rest of the code, we expect the arguments to reside at these known
1712   // registers, and we additionally check the placement here in case calling convention ever
1713   // changes.
1714   Register reg_cont_obj   = R3_ARG1;
1715   Register reg_is_cont    = R4_ARG2;
1716   Register reg_is_virtual = R5_ARG3;
1717 
1718   check_continuation_enter_argument(regs[pos_cont_obj].first(),   reg_cont_obj,   "Continuation object");
1719   check_continuation_enter_argument(regs[pos_is_cont].first(),    reg_is_cont,    "isContinue");
1720   check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1721 
1722   AddressLiteral resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1723   address start = __ pc();
1724   Label L_thaw, L_exit;
1725 
1726   // i2i entry used at interp_only_mode only
1727   interpreted_entry_offset = __ pc() - start;
1728   {
1729 #ifdef ASSERT
1730     Label is_interp_only;
1731     __ lwz(R0, in_bytes(JavaThread::interp_only_mode_offset()), R16_thread);
1732     __ cmpwi(CR0, R0, 0);
1733     __ bne(CR0, is_interp_only);
1734     __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1735     __ bind(is_interp_only);
1736 #endif
1737 
1738     // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1739     __ ld(reg_cont_obj,    Interpreter::stackElementSize*3, R15_esp);
1740     __ lwz(reg_is_cont,    Interpreter::stackElementSize*2, R15_esp);
1741     __ lwz(reg_is_virtual, Interpreter::stackElementSize*1, R15_esp);
1742 
1743     __ push_cont_fastpath();
1744 
1745     OopMap* map = continuation_enter_setup(masm, framesize_words);
1746 
1747     // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1748     // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1749 
1750     fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1751 
1752     // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1753     __ cmpwi(CR0, reg_is_cont, 0);
1754     __ bne(CR0, L_thaw);
1755 
1756     // --- call Continuation.enter(Continuation c, boolean isContinue)
1757 
1758     // Emit compiled static call. The call will be always resolved to the c2i
1759     // entry of Continuation.enter(Continuation c, boolean isContinue).
1760     address c2i_call_pc = __ trampoline_call(resolve);
1761     guarantee(c2i_call_pc != nullptr, "CodeCache is full at gen_continuation_enter");
1762 
1763     // Emit stub for static call
1764     address stub = CompiledDirectCall::emit_to_interp_stub(masm, c2i_call_pc);
1765     guarantee(stub != nullptr, "CodeCache is full at gen_continuation_enter");
1766 
1767     oop_maps->add_gc_map(__ pc() - start, map);
1768     __ post_call_nop();
1769 
1770     __ b(L_exit);
1771   }
1772 
1773   // compiled entry
1774   __ align(CodeEntryAlignment);
1775   compiled_entry_offset = __ pc() - start;
1776 
1777   OopMap* map = continuation_enter_setup(masm, framesize_words);
1778 
1779   // Frame is now completed as far as size and linkage.
1780   frame_complete =__ pc() - start;
1781 
1782   fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1783 
1784   // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1785   __ cmpwi(CR0, reg_is_cont, 0);
1786   __ bne(CR0, L_thaw);
1787 
1788   // --- call Continuation.enter(Continuation c, boolean isContinue)
1789 
1790   // Emit compiled static call
1791   // The call needs to be resolved. There's a special case for this in
1792   // SharedRuntime::find_callee_info_helper() which calls
1793   // LinkResolver::resolve_continuation_enter() which resolves the call to
1794   // Continuation.enter(Continuation c, boolean isContinue).
1795   address call_pc = __ trampoline_call(resolve);
1796   guarantee(call_pc != nullptr, "CodeCache is full at gen_continuation_enter");
1797 
1798   oop_maps->add_gc_map(__ pc() - start, map);
1799   __ post_call_nop();
1800 
1801   __ b(L_exit);
1802 
1803   // --- Thawing path
1804 
1805   __ bind(L_thaw);
1806   ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1807   __ add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(StubRoutines::cont_thaw()));
1808   __ mtctr(R0);
1809   __ bctrl();
1810   oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1811   ContinuationEntry::_return_pc_offset = __ pc() - start;
1812   __ post_call_nop();
1813 
1814   // --- Normal exit (resolve/thawing)
1815 
1816   __ bind(L_exit);
1817   ContinuationEntry::_cleanup_offset = __ pc() - start;
1818   continuation_enter_cleanup(masm);
1819 
1820   // Pop frame and return
1821   DEBUG_ONLY(__ ld_ptr(R0, 0, R1_SP));
1822   __ addi(R1_SP, R1_SP, framesize_words*wordSize);
1823   DEBUG_ONLY(__ cmpd(CR0, R0, R1_SP));
1824   __ asm_assert_eq(FILE_AND_LINE ": inconsistent frame size");
1825   __ ld(R0, _abi0(lr), R1_SP); // Return pc
1826   __ mtlr(R0);
1827   __ blr();
1828 
1829   // --- Exception handling path
1830 
1831   exception_offset = __ pc() - start;
1832 
1833   continuation_enter_cleanup(masm);
1834   Register ex_pc  = R17_tos;   // nonvolatile register
1835   Register ex_oop = R15_esp;   // nonvolatile register
1836   __ ld(ex_pc, _abi0(callers_sp), R1_SP); // Load caller's return pc
1837   __ ld(ex_pc, _abi0(lr), ex_pc);
1838   __ mr(ex_oop, R3_RET);                  // save return value containing the exception oop
1839   __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, ex_pc);
1840   __ mtlr(R3_RET);                        // the exception handler
1841   __ ld(R1_SP, _abi0(callers_sp), R1_SP); // remove enterSpecial frame
1842 
1843   // Continue at exception handler
1844   // See OptoRuntime::generate_exception_blob for register arguments
1845   __ mr(R3_ARG1, ex_oop); // pass exception oop
1846   __ mr(R4_ARG2, ex_pc);  // pass exception pc
1847   __ blr();
1848 
1849   // static stub for the call above
1850   address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1851   guarantee(stub != nullptr, "CodeCache is full at gen_continuation_enter");
1852 }
1853 
1854 static void gen_continuation_yield(MacroAssembler* masm,
1855                                    const VMRegPair* regs,
1856                                    OopMapSet* oop_maps,
1857                                    int& frame_complete,
1858                                    int& framesize_words,
1859                                    int& compiled_entry_offset) {
1860   Register tmp = R10_ARG8;
1861 
1862   const int framesize_bytes = (int)align_up((int)frame::native_abi_reg_args_size, frame::alignment_in_bytes);
1863   framesize_words = framesize_bytes / wordSize;
1864 
1865   address start = __ pc();
1866   compiled_entry_offset = __ pc() - start;
1867 
1868   // Save return pc and push entry frame
1869   __ mflr(tmp);
1870   __ std(tmp, _abi0(lr), R1_SP);       // SP->lr = return_pc
1871   __ push_frame(framesize_bytes , R0); // SP -= frame_size_in_bytes
1872 
1873   DEBUG_ONLY(__ block_comment("Frame Complete"));
1874   frame_complete = __ pc() - start;
1875   address last_java_pc = __ pc();
1876 
1877   // This nop must be exactly at the PC we push into the frame info.
1878   // We use this nop for fast CodeBlob lookup, associate the OopMap
1879   // with it right away.
1880   __ post_call_nop();
1881   OopMap* map = new OopMap(framesize_bytes / VMRegImpl::stack_slot_size, 1);
1882   oop_maps->add_gc_map(last_java_pc - start, map);
1883 
1884   __ calculate_address_from_global_toc(tmp, last_java_pc); // will be relocated
1885   __ set_last_Java_frame(R1_SP, tmp);
1886   __ call_VM_leaf(Continuation::freeze_entry(), R16_thread, R1_SP);
1887   __ reset_last_Java_frame();
1888 
1889   Label L_pinned;
1890 
1891   __ cmpwi(CR0, R3_RET, 0);
1892   __ bne(CR0, L_pinned);
1893 
1894   // yield succeeded
1895 
1896   // Pop frames of continuation including this stub's frame
1897   __ ld_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1898   // The frame pushed by gen_continuation_enter is on top now again
1899   continuation_enter_cleanup(masm);
1900 
1901   // Pop frame and return
1902   Label L_return;
1903   __ bind(L_return);
1904   __ pop_frame();
1905   __ ld(R0, _abi0(lr), R1_SP); // Return pc
1906   __ mtlr(R0);
1907   __ blr();
1908 
1909   // yield failed - continuation is pinned
1910 
1911   __ bind(L_pinned);
1912 
1913   // handle pending exception thrown by freeze
1914   __ ld(tmp, in_bytes(JavaThread::pending_exception_offset()), R16_thread);
1915   __ cmpdi(CR0, tmp, 0);
1916   __ beq(CR0, L_return); // return if no exception is pending
1917   __ pop_frame();
1918   __ ld(R0, _abi0(lr), R1_SP); // Return pc
1919   __ mtlr(R0);
1920   __ load_const_optimized(tmp, StubRoutines::forward_exception_entry(), R0);
1921   __ mtctr(tmp);
1922   __ bctr();
1923 }
1924 
1925 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1926   ::continuation_enter_cleanup(masm);
1927 }
1928 
1929 // ---------------------------------------------------------------------------
1930 // Generate a native wrapper for a given method. The method takes arguments
1931 // in the Java compiled code convention, marshals them to the native
1932 // convention (handlizes oops, etc), transitions to native, makes the call,
1933 // returns to java state (possibly blocking), unhandlizes any result and
1934 // returns.
1935 //
1936 // Critical native functions are a shorthand for the use of
1937 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1938 // functions.  The wrapper is expected to unpack the arguments before
1939 // passing them to the callee. Critical native functions leave the state _in_Java,
1940 // since they cannot stop for GC.
1941 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1942 // block and the check for pending exceptions it's impossible for them
1943 // to be thrown.
1944 //
1945 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
1946                                                 const methodHandle& method,
1947                                                 int compile_id,
1948                                                 BasicType *in_sig_bt,
1949                                                 VMRegPair *in_regs,
1950                                                 BasicType ret_type) {
1951   if (method->is_continuation_native_intrinsic()) {
1952     int exception_offset = -1;
1953     OopMapSet* oop_maps = new OopMapSet();
1954     int frame_complete = -1;
1955     int stack_slots = -1;
1956     int interpreted_entry_offset = -1;
1957     int vep_offset = -1;
1958     if (method->is_continuation_enter_intrinsic()) {
1959       gen_continuation_enter(masm,
1960                              in_regs,
1961                              exception_offset,
1962                              oop_maps,
1963                              frame_complete,
1964                              stack_slots,
1965                              interpreted_entry_offset,
1966                              vep_offset);
1967     } else if (method->is_continuation_yield_intrinsic()) {
1968       gen_continuation_yield(masm,
1969                              in_regs,
1970                              oop_maps,
1971                              frame_complete,
1972                              stack_slots,
1973                              vep_offset);
1974     } else {
1975       guarantee(false, "Unknown Continuation native intrinsic");
1976     }
1977 
1978 #ifdef ASSERT
1979     if (method->is_continuation_enter_intrinsic()) {
1980       assert(interpreted_entry_offset != -1, "Must be set");
1981       assert(exception_offset != -1,         "Must be set");
1982     } else {
1983       assert(interpreted_entry_offset == -1, "Must be unset");
1984       assert(exception_offset == -1,         "Must be unset");
1985     }
1986     assert(frame_complete != -1,    "Must be set");
1987     assert(stack_slots != -1,       "Must be set");
1988     assert(vep_offset != -1,        "Must be set");
1989 #endif
1990 
1991     __ flush();
1992     nmethod* nm = nmethod::new_native_nmethod(method,
1993                                               compile_id,
1994                                               masm->code(),
1995                                               vep_offset,
1996                                               frame_complete,
1997                                               stack_slots,
1998                                               in_ByteSize(-1),
1999                                               in_ByteSize(-1),
2000                                               oop_maps,
2001                                               exception_offset);
2002     if (nm == nullptr) return nm;
2003     if (method->is_continuation_enter_intrinsic()) {
2004       ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2005     } else if (method->is_continuation_yield_intrinsic()) {
2006       _cont_doYield_stub = nm;
2007     }
2008     return nm;
2009   }
2010 
2011   if (method->is_method_handle_intrinsic()) {
2012     vmIntrinsics::ID iid = method->intrinsic_id();
2013     intptr_t start = (intptr_t)__ pc();
2014     int vep_offset = ((intptr_t)__ pc()) - start;
2015     gen_special_dispatch(masm,
2016                          method,
2017                          in_sig_bt,
2018                          in_regs);
2019     int frame_complete = ((intptr_t)__ pc()) - start;  // not complete, period
2020     __ flush();
2021     int stack_slots = SharedRuntime::out_preserve_stack_slots();  // no out slots at all, actually
2022     return nmethod::new_native_nmethod(method,
2023                                        compile_id,
2024                                        masm->code(),
2025                                        vep_offset,
2026                                        frame_complete,
2027                                        stack_slots / VMRegImpl::slots_per_word,
2028                                        in_ByteSize(-1),
2029                                        in_ByteSize(-1),
2030                                        (OopMapSet*)nullptr);
2031   }
2032 
2033   address native_func = method->native_function();
2034   assert(native_func != nullptr, "must have function");
2035 
2036   // First, create signature for outgoing C call
2037   // --------------------------------------------------------------------------
2038 
2039   int total_in_args = method->size_of_parameters();
2040   // We have received a description of where all the java args are located
2041   // on entry to the wrapper. We need to convert these args to where
2042   // the jni function will expect them. To figure out where they go
2043   // we convert the java signature to a C signature by inserting
2044   // the hidden arguments as arg[0] and possibly arg[1] (static method)
2045 
2046   // Calculate the total number of C arguments and create arrays for the
2047   // signature and the outgoing registers.
2048   // On ppc64, we have two arrays for the outgoing registers, because
2049   // some floating-point arguments must be passed in registers _and_
2050   // in stack locations.
2051   bool method_is_static = method->is_static();
2052   int  total_c_args     = total_in_args + (method_is_static ? 2 : 1);
2053 
2054   BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2055   VMRegPair *out_regs   = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2056 
2057   // Create the signature for the C call:
2058   //   1) add the JNIEnv*
2059   //   2) add the class if the method is static
2060   //   3) copy the rest of the incoming signature (shifted by the number of
2061   //      hidden arguments).
2062 
2063   int argc = 0;
2064   out_sig_bt[argc++] = T_ADDRESS;
2065   if (method->is_static()) {
2066     out_sig_bt[argc++] = T_OBJECT;
2067   }
2068 
2069   for (int i = 0; i < total_in_args ; i++ ) {
2070     out_sig_bt[argc++] = in_sig_bt[i];
2071   }
2072 
2073 
2074   // Compute the wrapper's frame size.
2075   // --------------------------------------------------------------------------
2076 
2077   // Now figure out where the args must be stored and how much stack space
2078   // they require.
2079   //
2080   // Compute framesize for the wrapper. We need to handlize all oops in
2081   // incoming registers.
2082   //
2083   // Calculate the total number of stack slots we will need:
2084   //   1) abi requirements
2085   //   2) outgoing arguments
2086   //   3) space for inbound oop handle area
2087   //   4) space for handlizing a klass if static method
2088   //   5) space for a lock if synchronized method
2089   //   6) workspace for saving return values, int <-> float reg moves, etc.
2090   //   7) alignment
2091   //
2092   // Layout of the native wrapper frame:
2093   // (stack grows upwards, memory grows downwards)
2094   //
2095   // NW     [ABI_REG_ARGS]             <-- 1) R1_SP
2096   //        [outgoing arguments]       <-- 2) R1_SP + out_arg_slot_offset
2097   //        [oopHandle area]           <-- 3) R1_SP + oop_handle_offset
2098   //        klass                      <-- 4) R1_SP + klass_offset
2099   //        lock                       <-- 5) R1_SP + lock_offset
2100   //        [workspace]                <-- 6) R1_SP + workspace_offset
2101   //        [alignment] (optional)     <-- 7)
2102   // caller [JIT_TOP_ABI_48]           <-- r_callers_sp
2103   //
2104   // - *_slot_offset Indicates offset from SP in number of stack slots.
2105   // - *_offset      Indicates offset from SP in bytes.
2106 
2107   int stack_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args) + // 1+2)
2108                     SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
2109 
2110   // Now the space for the inbound oop handle area.
2111   int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
2112 
2113   int oop_handle_slot_offset = stack_slots;
2114   stack_slots += total_save_slots;                                                // 3)
2115 
2116   int klass_slot_offset = 0;
2117   int klass_offset      = -1;
2118   if (method_is_static) {                                                         // 4)
2119     klass_slot_offset  = stack_slots;
2120     klass_offset       = klass_slot_offset * VMRegImpl::stack_slot_size;
2121     stack_slots       += VMRegImpl::slots_per_word;
2122   }
2123 
2124   int lock_slot_offset = 0;
2125   int lock_offset      = -1;
2126   if (method->is_synchronized()) {                                                // 5)
2127     lock_slot_offset   = stack_slots;
2128     lock_offset        = lock_slot_offset * VMRegImpl::stack_slot_size;
2129     stack_slots       += VMRegImpl::slots_per_word;
2130   }
2131 
2132   int workspace_slot_offset = stack_slots;                                        // 6)
2133   stack_slots         += 2;
2134 
2135   // Now compute actual number of stack words we need.
2136   // Rounding to make stack properly aligned.
2137   stack_slots = align_up(stack_slots,                                             // 7)
2138                          frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
2139   int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
2140 
2141 
2142   // Now we can start generating code.
2143   // --------------------------------------------------------------------------
2144 
2145   intptr_t start_pc = (intptr_t)__ pc();
2146   intptr_t vep_start_pc;
2147   intptr_t frame_done_pc;
2148 
2149   Label    handle_pending_exception;
2150   Label    last_java_pc;
2151 
2152   Register r_callers_sp = R21;
2153   Register r_temp_1     = R22;
2154   Register r_temp_2     = R23;
2155   Register r_temp_3     = R24;
2156   Register r_temp_4     = R25;
2157   Register r_temp_5     = R26;
2158   Register r_temp_6     = R27;
2159   Register r_last_java_pc = R28;
2160 
2161   Register r_carg1_jnienv        = noreg;
2162   Register r_carg2_classorobject = noreg;
2163   r_carg1_jnienv        = out_regs[0].first()->as_Register();
2164   r_carg2_classorobject = out_regs[1].first()->as_Register();
2165 
2166 
2167   // Generate the Unverified Entry Point (UEP).
2168   // --------------------------------------------------------------------------
2169   assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
2170 
2171   // Check ic: object class == cached class?
2172   if (!method_is_static) {
2173     __ ic_check(4 /* end_alignment */);
2174   }
2175 
2176   // Generate the Verified Entry Point (VEP).
2177   // --------------------------------------------------------------------------
2178   vep_start_pc = (intptr_t)__ pc();
2179 
2180   if (method->needs_clinit_barrier()) {
2181     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
2182     Label L_skip_barrier;
2183     Register klass = r_temp_1;
2184     // Notify OOP recorder (don't need the relocation)
2185     AddressLiteral md = __ constant_metadata_address(method->method_holder());
2186     __ load_const_optimized(klass, md.value(), R0);
2187     __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
2188 
2189     __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
2190     __ mtctr(klass);
2191     __ bctr();
2192 
2193     __ bind(L_skip_barrier);
2194   }
2195 
2196   __ save_LR(r_temp_1);
2197   __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2198   __ mr(r_callers_sp, R1_SP);                            // Remember frame pointer.
2199   __ push_frame(frame_size_in_bytes, r_temp_1);          // Push the c2n adapter's frame.
2200 
2201   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2202   bs->nmethod_entry_barrier(masm, r_temp_1);
2203 
2204   frame_done_pc = (intptr_t)__ pc();
2205 
2206   // Native nmethod wrappers never take possession of the oop arguments.
2207   // So the caller will gc the arguments.
2208   // The only thing we need an oopMap for is if the call is static.
2209   //
2210   // An OopMap for lock (and class if static), and one for the VM call itself.
2211   OopMapSet *oop_maps = new OopMapSet();
2212   OopMap    *oop_map  = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2213 
2214   // Move arguments from register/stack to register/stack.
2215   // --------------------------------------------------------------------------
2216   //
2217   // We immediately shuffle the arguments so that for any vm call we have
2218   // to make from here on out (sync slow path, jvmti, etc.) we will have
2219   // captured the oops from our caller and have a valid oopMap for them.
2220   //
2221   // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2222   // (derived from JavaThread* which is in R16_thread) and, if static,
2223   // the class mirror instead of a receiver. This pretty much guarantees that
2224   // register layout will not match. We ignore these extra arguments during
2225   // the shuffle. The shuffle is described by the two calling convention
2226   // vectors we have in our possession. We simply walk the java vector to
2227   // get the source locations and the c vector to get the destinations.
2228 
2229   // Record sp-based slot for receiver on stack for non-static methods.
2230   int receiver_offset = -1;
2231 
2232   // We move the arguments backward because the floating point registers
2233   // destination will always be to a register with a greater or equal
2234   // register number or the stack.
2235   //   in  is the index of the incoming Java arguments
2236   //   out is the index of the outgoing C arguments
2237 
2238 #ifdef ASSERT
2239   bool reg_destroyed[Register::number_of_registers];
2240   bool freg_destroyed[FloatRegister::number_of_registers];
2241   for (int r = 0 ; r < Register::number_of_registers ; r++) {
2242     reg_destroyed[r] = false;
2243   }
2244   for (int f = 0 ; f < FloatRegister::number_of_registers ; f++) {
2245     freg_destroyed[f] = false;
2246   }
2247 #endif // ASSERT
2248 
2249   for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2250 
2251 #ifdef ASSERT
2252     if (in_regs[in].first()->is_Register()) {
2253       assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2254     } else if (in_regs[in].first()->is_FloatRegister()) {
2255       assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2256     }
2257     if (out_regs[out].first()->is_Register()) {
2258       reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2259     } else if (out_regs[out].first()->is_FloatRegister()) {
2260       freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2261     }
2262 #endif // ASSERT
2263 
2264     switch (in_sig_bt[in]) {
2265       case T_BOOLEAN:
2266       case T_CHAR:
2267       case T_BYTE:
2268       case T_SHORT:
2269       case T_INT:
2270         // Move int and do sign extension.
2271         int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2272         break;
2273       case T_LONG:
2274         long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2275         break;
2276       case T_ARRAY:
2277       case T_OBJECT:
2278         object_move(masm, stack_slots,
2279                     oop_map, oop_handle_slot_offset,
2280                     ((in == 0) && (!method_is_static)), &receiver_offset,
2281                     in_regs[in], out_regs[out],
2282                     r_callers_sp, r_temp_1, r_temp_2);
2283         break;
2284       case T_VOID:
2285         break;
2286       case T_FLOAT:
2287         float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2288         break;
2289       case T_DOUBLE:
2290         double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2291         break;
2292       case T_ADDRESS:
2293         fatal("found type (T_ADDRESS) in java args");
2294         break;
2295       default:
2296         ShouldNotReachHere();
2297         break;
2298     }
2299   }
2300 
2301   // Pre-load a static method's oop into ARG2.
2302   // Used both by locking code and the normal JNI call code.
2303   if (method_is_static) {
2304     __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2305                         r_carg2_classorobject);
2306 
2307     // Now handlize the static class mirror in carg2. It's known not-null.
2308     __ std(r_carg2_classorobject, klass_offset, R1_SP);
2309     oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2310     __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2311   }
2312 
2313   // Get JNIEnv* which is first argument to native.
2314   __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2315 
2316   // NOTE:
2317   //
2318   // We have all of the arguments setup at this point.
2319   // We MUST NOT touch any outgoing regs from this point on.
2320   // So if we must call out we must push a new frame.
2321 
2322   // The last java pc will also be used as resume pc if this is the wrapper for wait0.
2323   // For this purpose the precise location matters but not for oopmap lookup.
2324   __ calculate_address_from_global_toc(r_last_java_pc, last_java_pc, true, true, true, true);
2325 
2326   // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2327   assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2328 
2329 # if 0
2330   // DTrace method entry
2331 # endif
2332 
2333   // Lock a synchronized method.
2334   // --------------------------------------------------------------------------
2335 
2336   if (method->is_synchronized()) {
2337     Register          r_oop  = r_temp_4;
2338     const Register    r_box  = r_temp_5;
2339     Label             done, locked;
2340 
2341     // Load the oop for the object or class. r_carg2_classorobject contains
2342     // either the handlized oop from the incoming arguments or the handlized
2343     // class mirror (if the method is static).
2344     __ ld(r_oop, 0, r_carg2_classorobject);
2345 
2346     // Get the lock box slot's address.
2347     __ addi(r_box, R1_SP, lock_offset);
2348 
2349     // Try fastpath for locking.
2350     // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2351     Register r_temp_3_or_noreg = UseObjectMonitorTable ? r_temp_3 : noreg;
2352     __ compiler_fast_lock_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3_or_noreg);
2353     __ beq(CR0, locked);
2354 
2355     // None of the above fast optimizations worked so we have to get into the
2356     // slow case of monitor enter. Inline a special case of call_VM that
2357     // disallows any pending_exception.
2358 
2359     // Save argument registers and leave room for C-compatible ABI_REG_ARGS.
2360     int frame_size = frame::native_abi_reg_args_size + align_up(total_c_args * wordSize, frame::alignment_in_bytes);
2361     __ mr(R11_scratch1, R1_SP);
2362     RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs);
2363 
2364     // Do the call.
2365     __ set_last_Java_frame(R11_scratch1, r_last_java_pc);
2366     assert(r_last_java_pc->is_nonvolatile(), "r_last_java_pc needs to be preserved accross complete_monitor_locking_C call");
2367     // The following call will not be preempted.
2368     // push_cont_fastpath forces freeze slow path in case we try to preempt where we will pin the
2369     // vthread to the carrier (see FreezeBase::recurse_freeze_native_frame()).
2370     __ push_cont_fastpath();
2371     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2372     __ pop_cont_fastpath();
2373     __ reset_last_Java_frame();
2374 
2375     RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs);
2376 
2377     __ asm_assert_mem8_is_zero(thread_(pending_exception),
2378        "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C");
2379 
2380     __ bind(locked);
2381   }
2382 
2383   __ set_last_Java_frame(R1_SP, r_last_java_pc);
2384 
2385   // Publish thread state
2386   // --------------------------------------------------------------------------
2387 
2388   // Transition from _thread_in_Java to _thread_in_native.
2389   __ li(R0, _thread_in_native);
2390   __ release();
2391   // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2392   __ stw(R0, thread_(thread_state));
2393 
2394 
2395   // The JNI call
2396   // --------------------------------------------------------------------------
2397   __ call_c(native_func, relocInfo::runtime_call_type);
2398 
2399 
2400   // Now, we are back from the native code.
2401 
2402 
2403   // Unpack the native result.
2404   // --------------------------------------------------------------------------
2405 
2406   // For int-types, we do any needed sign-extension required.
2407   // Care must be taken that the return values (R3_RET and F1_RET)
2408   // will survive any VM calls for blocking or unlocking.
2409   // An OOP result (handle) is done specially in the slow-path code.
2410 
2411   switch (ret_type) {
2412     case T_VOID:    break;        // Nothing to do!
2413     case T_FLOAT:   break;        // Got it where we want it (unless slow-path).
2414     case T_DOUBLE:  break;        // Got it where we want it (unless slow-path).
2415     case T_LONG:    break;        // Got it where we want it (unless slow-path).
2416     case T_OBJECT:  break;        // Really a handle.
2417                                   // Cannot de-handlize until after reclaiming jvm_lock.
2418     case T_ARRAY:   break;
2419 
2420     case T_BOOLEAN: {             // 0 -> false(0); !0 -> true(1)
2421       __ normalize_bool(R3_RET);
2422       break;
2423       }
2424     case T_BYTE: {                // sign extension
2425       __ extsb(R3_RET, R3_RET);
2426       break;
2427       }
2428     case T_CHAR: {                // unsigned result
2429       __ andi(R3_RET, R3_RET, 0xffff);
2430       break;
2431       }
2432     case T_SHORT: {               // sign extension
2433       __ extsh(R3_RET, R3_RET);
2434       break;
2435       }
2436     case T_INT:                   // nothing to do
2437       break;
2438     default:
2439       ShouldNotReachHere();
2440       break;
2441   }
2442 
2443   // Publish thread state
2444   // --------------------------------------------------------------------------
2445 
2446   // Switch thread to "native transition" state before reading the
2447   // synchronization state. This additional state is necessary because reading
2448   // and testing the synchronization state is not atomic w.r.t. GC, as this
2449   // scenario demonstrates:
2450   //   - Java thread A, in _thread_in_native state, loads _not_synchronized
2451   //     and is preempted.
2452   //   - VM thread changes sync state to synchronizing and suspends threads
2453   //     for GC.
2454   //   - Thread A is resumed to finish this native method, but doesn't block
2455   //     here since it didn't see any synchronization in progress, and escapes.
2456 
2457   // Transition from _thread_in_native to _thread_in_native_trans.
2458   __ li(R0, _thread_in_native_trans);
2459   __ release();
2460   // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2461   __ stw(R0, thread_(thread_state));
2462 
2463 
2464   // Must we block?
2465   // --------------------------------------------------------------------------
2466 
2467   // Block, if necessary, before resuming in _thread_in_Java state.
2468   // In order for GC to work, don't clear the last_Java_sp until after blocking.
2469   {
2470     Label no_block, sync;
2471 
2472     // Force this write out before the read below.
2473     if (!UseSystemMemoryBarrier) {
2474       __ fence();
2475     }
2476 
2477     Register sync_state_addr = r_temp_4;
2478     Register sync_state      = r_temp_5;
2479     Register suspend_flags   = r_temp_6;
2480 
2481     // No synchronization in progress nor yet synchronized
2482     // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
2483     __ safepoint_poll(sync, sync_state, true /* at_return */, false /* in_nmethod */);
2484 
2485     // Not suspended.
2486     // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2487     __ lwz(suspend_flags, thread_(suspend_flags));
2488     __ cmpwi(CR1, suspend_flags, 0);
2489     __ beq(CR1, no_block);
2490 
2491     // Block. Save any potential method result value before the operation and
2492     // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2493     // lets us share the oopMap we used when we went native rather than create
2494     // a distinct one for this pc.
2495     __ bind(sync);
2496     __ isync();
2497 
2498     address entry_point =
2499       CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2500     save_native_result(masm, ret_type, workspace_slot_offset);
2501     __ call_VM_leaf(entry_point, R16_thread);
2502     restore_native_result(masm, ret_type, workspace_slot_offset);
2503 
2504     __ bind(no_block);
2505 
2506     // Publish thread state.
2507     // --------------------------------------------------------------------------
2508 
2509     // Thread state is thread_in_native_trans. Any safepoint blocking has
2510     // already happened so we can now change state to _thread_in_Java.
2511 
2512     // Transition from _thread_in_native_trans to _thread_in_Java.
2513     __ li(R0, _thread_in_Java);
2514     __ lwsync(); // Acquire safepoint and suspend state, release thread state.
2515     // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2516     __ stw(R0, thread_(thread_state));
2517 
2518     // Check preemption for Object.wait()
2519     if (method->is_object_wait0()) {
2520       Label not_preempted;
2521       __ ld(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2522       __ cmpdi(CR0, R0, 0);
2523       __ beq(CR0, not_preempted);
2524       __ mtlr(R0);
2525       __ li(R0, 0);
2526       __ std(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2527       __ blr();
2528       __ bind(not_preempted);
2529     }
2530     __ bind(last_java_pc);
2531     // We use the same pc/oopMap repeatedly when we call out above.
2532     intptr_t oopmap_pc = (intptr_t) __ pc();
2533     oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2534   }
2535 
2536   // Reguard any pages if necessary.
2537   // --------------------------------------------------------------------------
2538 
2539   Label no_reguard;
2540   __ lwz(r_temp_1, thread_(stack_guard_state));
2541   __ cmpwi(CR0, r_temp_1, StackOverflow::stack_guard_yellow_reserved_disabled);
2542   __ bne(CR0, no_reguard);
2543 
2544   save_native_result(masm, ret_type, workspace_slot_offset);
2545   __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2546   restore_native_result(masm, ret_type, workspace_slot_offset);
2547 
2548   __ bind(no_reguard);
2549 
2550 
2551   // Unlock
2552   // --------------------------------------------------------------------------
2553 
2554   if (method->is_synchronized()) {
2555     const Register r_oop       = r_temp_4;
2556     const Register r_box       = r_temp_5;
2557     const Register r_exception = r_temp_6;
2558     Label done;
2559 
2560     // Get oop and address of lock object box.
2561     if (method_is_static) {
2562       assert(klass_offset != -1, "");
2563       __ ld(r_oop, klass_offset, R1_SP);
2564     } else {
2565       assert(receiver_offset != -1, "");
2566       __ ld(r_oop, receiver_offset, R1_SP);
2567     }
2568     __ addi(r_box, R1_SP, lock_offset);
2569 
2570     // Try fastpath for unlocking.
2571     __ compiler_fast_unlock_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2572     __ beq(CR0, done);
2573 
2574     // Save and restore any potential method result value around the unlocking operation.
2575     save_native_result(masm, ret_type, workspace_slot_offset);
2576 
2577     // Must save pending exception around the slow-path VM call. Since it's a
2578     // leaf call, the pending exception (if any) can be kept in a register.
2579     __ ld(r_exception, thread_(pending_exception));
2580     assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2581     __ li(R0, 0);
2582     __ std(R0, thread_(pending_exception));
2583 
2584     // Slow case of monitor enter.
2585     // Inline a special case of call_VM that disallows any pending_exception.
2586     // Arguments are (oop obj, BasicLock* lock, JavaThread* thread).
2587     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread);
2588 
2589     __ asm_assert_mem8_is_zero(thread_(pending_exception),
2590        "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C");
2591 
2592     restore_native_result(masm, ret_type, workspace_slot_offset);
2593 
2594     // Check_forward_pending_exception jump to forward_exception if any pending
2595     // exception is set. The forward_exception routine expects to see the
2596     // exception in pending_exception and not in a register. Kind of clumsy,
2597     // since all folks who branch to forward_exception must have tested
2598     // pending_exception first and hence have it in a register already.
2599     __ std(r_exception, thread_(pending_exception));
2600 
2601     __ bind(done);
2602   }
2603 
2604 # if 0
2605   // DTrace method exit
2606 # endif
2607 
2608   // Clear "last Java frame" SP and PC.
2609   // --------------------------------------------------------------------------
2610 
2611   // Last java frame won't be set if we're resuming after preemption
2612   bool maybe_preempted = method->is_object_wait0();
2613   __ reset_last_Java_frame(!maybe_preempted /* check_last_java_sp */);
2614 
2615   // Unbox oop result, e.g. JNIHandles::resolve value.
2616   // --------------------------------------------------------------------------
2617 
2618   if (is_reference_type(ret_type)) {
2619     __ resolve_jobject(R3_RET, r_temp_1, r_temp_2, MacroAssembler::PRESERVATION_NONE);
2620   }
2621 
2622   if (CheckJNICalls) {
2623     // clear_pending_jni_exception_check
2624     __ load_const_optimized(R0, 0L);
2625     __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
2626   }
2627 
2628   // Reset handle block.
2629   // --------------------------------------------------------------------------
2630   __ ld(r_temp_1, thread_(active_handles));
2631   // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2632   __ li(r_temp_2, 0);
2633   __ stw(r_temp_2, in_bytes(JNIHandleBlock::top_offset()), r_temp_1);
2634 
2635   // Prepare for return
2636   // --------------------------------------------------------------------------
2637   __ pop_frame();
2638   __ restore_LR(R11);
2639 
2640 #if INCLUDE_JFR
2641   // We need to do a poll test after unwind in case the sampler
2642   // managed to sample the native frame after returning to Java.
2643   Label L_stub;
2644   int safepoint_offset = __ offset();
2645   if (!UseSIGTRAP) {
2646     __ relocate(relocInfo::poll_return_type);
2647   }
2648   __ safepoint_poll(L_stub, r_temp_2, true /* at_return */, true /* in_nmethod: frame already popped */);
2649 #endif // INCLUDE_JFR
2650 
2651   // Check for pending exceptions.
2652   // --------------------------------------------------------------------------
2653   __ ld(r_temp_2, thread_(pending_exception));
2654   __ cmpdi(CR0, r_temp_2, 0);
2655   __ bne(CR0, handle_pending_exception);
2656 
2657   // Return.
2658   __ blr();
2659 
2660   // Handler for return safepoint (out-of-line).
2661 #if INCLUDE_JFR
2662   if (!UseSIGTRAP) {
2663     __ bind(L_stub);
2664     __ jump_to_polling_page_return_handler_blob(safepoint_offset);
2665   }
2666 #endif // INCLUDE_JFR
2667 
2668   // Handler for pending exceptions (out-of-line).
2669   // --------------------------------------------------------------------------
2670   // Since this is a native call, we know the proper exception handler
2671   // is the empty function. We just pop this frame and then jump to
2672   // forward_exception_entry.
2673   __ bind(handle_pending_exception);
2674   __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2675                        relocInfo::runtime_call_type);
2676 
2677   // Done.
2678   // --------------------------------------------------------------------------
2679 
2680   __ flush();
2681 
2682   nmethod *nm = nmethod::new_native_nmethod(method,
2683                                             compile_id,
2684                                             masm->code(),
2685                                             vep_start_pc-start_pc,
2686                                             frame_done_pc-start_pc,
2687                                             stack_slots / VMRegImpl::slots_per_word,
2688                                             (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2689                                             in_ByteSize(lock_offset),
2690                                             oop_maps);
2691 
2692   return nm;
2693 }
2694 
2695 // This function returns the adjust size (in number of words) to a c2i adapter
2696 // activation for use during deoptimization.
2697 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2698   return align_up((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::frame_alignment_in_words);
2699 }
2700 
2701 uint SharedRuntime::in_preserve_stack_slots() {
2702   return frame::jit_in_preserve_size / VMRegImpl::stack_slot_size;
2703 }
2704 
2705 uint SharedRuntime::out_preserve_stack_slots() {
2706 #if defined(COMPILER1) || defined(COMPILER2)
2707   return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2708 #else
2709   return 0;
2710 #endif
2711 }
2712 
2713 VMReg SharedRuntime::thread_register() {
2714   // On PPC virtual threads don't save the JavaThread* in their context (e.g. C1 stub frames).
2715   ShouldNotCallThis();
2716   return nullptr;
2717 }
2718 
2719 #if defined(COMPILER1) || defined(COMPILER2)
2720 // Frame generation for deopt and uncommon trap blobs.
2721 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2722                                 /* Read */
2723                                 Register unroll_block_reg,
2724                                 /* Update */
2725                                 Register frame_sizes_reg,
2726                                 Register number_of_frames_reg,
2727                                 Register pcs_reg,
2728                                 /* Invalidate */
2729                                 Register frame_size_reg,
2730                                 Register pc_reg) {
2731 
2732   __ ld(pc_reg, 0, pcs_reg);
2733   __ ld(frame_size_reg, 0, frame_sizes_reg);
2734   __ std(pc_reg, _abi0(lr), R1_SP);
2735   __ push_frame(frame_size_reg, R0/*tmp*/);
2736   __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP);
2737   __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2738   __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2739   __ addi(pcs_reg, pcs_reg, wordSize);
2740 }
2741 
2742 // Loop through the UnrollBlock info and create new frames.
2743 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2744                                  /* read */
2745                                  Register unroll_block_reg,
2746                                  /* invalidate */
2747                                  Register frame_sizes_reg,
2748                                  Register number_of_frames_reg,
2749                                  Register pcs_reg,
2750                                  Register frame_size_reg,
2751                                  Register pc_reg) {
2752   Label loop;
2753 
2754  // _number_of_frames is of type int (deoptimization.hpp)
2755   __ lwa(number_of_frames_reg,
2756              in_bytes(Deoptimization::UnrollBlock::number_of_frames_offset()),
2757              unroll_block_reg);
2758   __ ld(pcs_reg,
2759             in_bytes(Deoptimization::UnrollBlock::frame_pcs_offset()),
2760             unroll_block_reg);
2761   __ ld(frame_sizes_reg,
2762             in_bytes(Deoptimization::UnrollBlock::frame_sizes_offset()),
2763             unroll_block_reg);
2764 
2765   // stack: (caller_of_deoptee, ...).
2766 
2767   // At this point we either have an interpreter frame or a compiled
2768   // frame on top of stack. If it is a compiled frame we push a new c2i
2769   // adapter here
2770 
2771   // Memorize top-frame stack-pointer.
2772   __ mr(frame_size_reg/*old_sp*/, R1_SP);
2773 
2774   // Resize interpreter top frame OR C2I adapter.
2775 
2776   // At this moment, the top frame (which is the caller of the deoptee) is
2777   // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2778   // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2779   // outgoing arguments.
2780   //
2781   // In order to push the interpreter frame for the deoptee, we need to
2782   // resize the top frame such that we are able to place the deoptee's
2783   // locals in the frame.
2784   // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2785   // into a valid PARENT_IJAVA_FRAME_ABI.
2786 
2787   __ lwa(R11_scratch1,
2788              in_bytes(Deoptimization::UnrollBlock::caller_adjustment_offset()),
2789              unroll_block_reg);
2790   __ neg(R11_scratch1, R11_scratch1);
2791 
2792   // R11_scratch1 contains size of locals for frame resizing.
2793   // R12_scratch2 contains top frame's lr.
2794 
2795   // Resize frame by complete frame size prevents TOC from being
2796   // overwritten by locals. A more stack space saving way would be
2797   // to copy the TOC to its location in the new abi.
2798   __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2799 
2800   // now, resize the frame
2801   __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2802 
2803   // In the case where we have resized a c2i frame above, the optional
2804   // alignment below the locals has size 32 (why?).
2805   __ std(R12_scratch2, _abi0(lr), R1_SP);
2806 
2807   // Initialize initial_caller_sp.
2808  __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP);
2809 
2810 #ifdef ASSERT
2811   // Make sure that there is at least one entry in the array.
2812   __ cmpdi(CR0, number_of_frames_reg, 0);
2813   __ asm_assert_ne("array_size must be > 0");
2814 #endif
2815 
2816   // Now push the new interpreter frames.
2817   //
2818   __ bind(loop);
2819   // Allocate a new frame, fill in the pc.
2820   push_skeleton_frame(masm, deopt,
2821                       unroll_block_reg,
2822                       frame_sizes_reg,
2823                       number_of_frames_reg,
2824                       pcs_reg,
2825                       frame_size_reg,
2826                       pc_reg);
2827   __ cmpdi(CR0, number_of_frames_reg, 0);
2828   __ bne(CR0, loop);
2829 
2830   // Get the return address pointing into the template interpreter.
2831   __ ld(R0, 0, pcs_reg);
2832   // Store it in the top interpreter frame.
2833   __ std(R0, _abi0(lr), R1_SP);
2834   // Initialize frame_manager_lr of interpreter top frame.
2835 }
2836 #endif
2837 
2838 void SharedRuntime::generate_deopt_blob() {
2839   // Allocate space for the code
2840   ResourceMark rm;
2841   // Setup code generation tools
2842   const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2843   CodeBuffer buffer(name, 2048, 1024);
2844   InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2845   Label exec_mode_initialized;
2846   OopMap* map = nullptr;
2847   OopMapSet *oop_maps = new OopMapSet();
2848 
2849   // size of ABI112 plus spill slots for R3_RET and F1_RET.
2850   const int frame_size_in_bytes = frame::native_abi_reg_args_spill_size;
2851   const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2852   int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2853 
2854   const Register exec_mode_reg = R21_tmp1;
2855 
2856   const address start = __ pc();
2857   int exception_offset = 0;
2858   int exception_in_tls_offset = 0;
2859   int reexecute_offset = 0;
2860 
2861 #if defined(COMPILER1) || defined(COMPILER2)
2862   // --------------------------------------------------------------------------
2863   // Prolog for non exception case!
2864 
2865   // We have been called from the deopt handler of the deoptee.
2866   //
2867   // deoptee:
2868   //                      ...
2869   //                      call X
2870   //                      ...
2871   //  deopt_handler:      call_deopt_stub
2872   //  cur. return pc  --> ...
2873   //
2874   // The return_pc has been stored in the frame of the deoptee and
2875   // will replace the address of the deopt_handler in the call
2876   // to Deoptimization::fetch_unroll_info below.
2877 
2878   // Push the "unpack frame"
2879   // Save everything in sight.
2880   map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2881                                                                    &first_frame_size_in_bytes,
2882                                                                    /*generate_oop_map=*/ true,
2883                                                                    RegisterSaver::return_pc_is_lr,
2884                                                                    /*save_vectors*/ SuperwordUseVSX);
2885   assert(map != nullptr, "OopMap must have been created");
2886 
2887   __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
2888   // Save exec mode for unpack_frames.
2889   __ b(exec_mode_initialized);
2890 
2891   // --------------------------------------------------------------------------
2892   // Prolog for exception case
2893 
2894   // An exception is pending.
2895   // We have been called with a return (interpreter) or a jump (exception blob).
2896   //
2897   // - R3_ARG1: exception oop
2898   // - R4_ARG2: exception pc
2899 
2900   exception_offset = __ pc() - start;
2901 
2902   BLOCK_COMMENT("Prolog for exception case");
2903 
2904   // Store exception oop and pc in thread (location known to GC).
2905   // This is needed since the call to "fetch_unroll_info()" may safepoint.
2906   __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2907   __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()),  R16_thread);
2908   __ std(R4_ARG2, _abi0(lr), R1_SP);
2909 
2910   // Vanilla deoptimization with an exception pending in exception_oop.
2911   exception_in_tls_offset = __ pc() - start;
2912 
2913   // Push the "unpack frame".
2914   // Save everything in sight.
2915   RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2916                                                              &first_frame_size_in_bytes,
2917                                                              /*generate_oop_map=*/ false,
2918                                                              RegisterSaver::return_pc_is_pre_saved,
2919                                                              /*save_vectors*/ SuperwordUseVSX);
2920 
2921   // Deopt during an exception. Save exec mode for unpack_frames.
2922   __ li(exec_mode_reg, Deoptimization::Unpack_exception);
2923 
2924   // fall through
2925 #ifdef COMPILER1
2926   __ b(exec_mode_initialized);
2927 
2928   // Reexecute entry, similar to c2 uncommon trap
2929   reexecute_offset = __ pc() - start;
2930 
2931   RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2932                                                              &first_frame_size_in_bytes,
2933                                                              /*generate_oop_map=*/ false,
2934                                                              RegisterSaver::return_pc_is_pre_saved,
2935                                                              /*save_vectors*/ SuperwordUseVSX);
2936   __ li(exec_mode_reg, Deoptimization::Unpack_reexecute);
2937 #endif
2938 
2939   // --------------------------------------------------------------------------
2940   __ BIND(exec_mode_initialized);
2941 
2942   const Register unroll_block_reg = R22_tmp2;
2943 
2944   // We need to set `last_Java_frame' because `fetch_unroll_info' will
2945   // call `last_Java_frame()'. The value of the pc in the frame is not
2946   // particularly important. It just needs to identify this blob.
2947   __ set_last_Java_frame(R1_SP, noreg);
2948 
2949   // With EscapeAnalysis turned on, this call may safepoint!
2950   __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg);
2951   address calls_return_pc = __ last_calls_return_pc();
2952   // Set an oopmap for the call site that describes all our saved registers.
2953   oop_maps->add_gc_map(calls_return_pc - start, map);
2954 
2955   __ reset_last_Java_frame();
2956   // Save the return value.
2957   __ mr(unroll_block_reg, R3_RET);
2958 
2959   // Restore only the result registers that have been saved
2960   // by save_volatile_registers(...).
2961   RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes, /*save_vectors*/ SuperwordUseVSX);
2962 
2963   // reload the exec mode from the UnrollBlock (it might have changed)
2964   __ lwz(exec_mode_reg, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
2965   // In excp_deopt_mode, restore and clear exception oop which we
2966   // stored in the thread during exception entry above. The exception
2967   // oop will be the return value of this stub.
2968   Label skip_restore_excp;
2969   __ cmpdi(CR0, exec_mode_reg, Deoptimization::Unpack_exception);
2970   __ bne(CR0, skip_restore_excp);
2971   __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2972   __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2973   __ li(R0, 0);
2974   __ std(R0, in_bytes(JavaThread::exception_pc_offset()),  R16_thread);
2975   __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2976   __ BIND(skip_restore_excp);
2977 
2978   __ pop_frame();
2979 
2980   // stack: (deoptee, optional i2c, caller of deoptee, ...).
2981 
2982   // pop the deoptee's frame
2983   __ pop_frame();
2984 
2985   // stack: (caller_of_deoptee, ...).
2986 
2987   // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
2988   // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
2989   // and the frame is effectively not resized.
2990   Register caller_sp = R23_tmp3;
2991   __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
2992   __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
2993 
2994   // Loop through the `UnrollBlock' info and create interpreter frames.
2995   push_skeleton_frames(masm, true/*deopt*/,
2996                        unroll_block_reg,
2997                        R23_tmp3,
2998                        R24_tmp4,
2999                        R25_tmp5,
3000                        R26_tmp6,
3001                        R27_tmp7);
3002 
3003   // stack: (skeletal interpreter frame, ..., optional skeletal
3004   // interpreter frame, optional c2i, caller of deoptee, ...).
3005 
3006   // push an `unpack_frame' taking care of float / int return values.
3007   __ push_frame(frame_size_in_bytes, R0/*tmp*/);
3008 
3009   // stack: (unpack frame, skeletal interpreter frame, ..., optional
3010   // skeletal interpreter frame, optional c2i, caller of deoptee,
3011   // ...).
3012 
3013   // Spill live volatile registers since we'll do a call.
3014   __ std( R3_RET, _native_abi_reg_args_spill(spill_ret),  R1_SP);
3015   __ stfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3016 
3017   // Let the unpacker layout information in the skeletal frames just
3018   // allocated.
3019   __ calculate_address_from_global_toc(R3_RET, calls_return_pc, true, true, true, true);
3020   __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
3021   // This is a call to a LEAF method, so no oop map is required.
3022   __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3023                   R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3024   __ reset_last_Java_frame();
3025 
3026   // Restore the volatiles saved above.
3027   __ ld( R3_RET, _native_abi_reg_args_spill(spill_ret),  R1_SP);
3028   __ lfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3029 
3030   // Pop the unpack frame.
3031   __ pop_frame();
3032   __ restore_LR(R0);
3033 
3034   // stack: (top interpreter frame, ..., optional interpreter frame,
3035   // optional c2i, caller of deoptee, ...).
3036 
3037   // Initialize R14_state.
3038   __ restore_interpreter_state(R11_scratch1);
3039   __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3040 
3041   // Return to the interpreter entry point.
3042   __ blr();
3043 #else // !defined(COMPILER1) && !defined(COMPILER2)
3044   __ unimplemented("deopt blob needed only with compiler");
3045 #endif
3046 
3047   // Make sure all code is generated
3048   __ flush();
3049 
3050   _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset,
3051                                            reexecute_offset, first_frame_size_in_bytes / wordSize);
3052   _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3053 }
3054 
3055 #ifdef COMPILER2
3056 UncommonTrapBlob* OptoRuntime::generate_uncommon_trap_blob() {
3057   // Allocate space for the code.
3058   ResourceMark rm;
3059   // Setup code generation tools.
3060   const char* name = OptoRuntime::stub_name(StubId::c2_uncommon_trap_id);
3061   CodeBuffer buffer(name, 2048, 1024);
3062   if (buffer.blob() == nullptr) {
3063     return nullptr;
3064   }
3065   InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3066   address start = __ pc();
3067 
3068   Register unroll_block_reg = R21_tmp1;
3069   Register klass_index_reg  = R22_tmp2;
3070   Register unc_trap_reg     = R23_tmp3;
3071   Register r_return_pc      = R27_tmp7;
3072 
3073   OopMapSet* oop_maps = new OopMapSet();
3074   int frame_size_in_bytes = frame::native_abi_reg_args_size;
3075   OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3076 
3077   // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3078 
3079   // Push a dummy `unpack_frame' and call
3080   // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3081   // vframe array and return the `UnrollBlock' information.
3082 
3083   // Save LR to compiled frame.
3084   __ save_LR(R11_scratch1);
3085 
3086   // Push an "uncommon_trap" frame.
3087   __ push_frame_reg_args(0, R11_scratch1);
3088 
3089   // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
3090 
3091   // Set the `unpack_frame' as last_Java_frame.
3092   // `Deoptimization::uncommon_trap' expects it and considers its
3093   // sender frame as the deoptee frame.
3094   // Remember the offset of the instruction whose address will be
3095   // moved to R11_scratch1.
3096   address gc_map_pc = __ pc();
3097   __ calculate_address_from_global_toc(r_return_pc, gc_map_pc, true, true, true, true);
3098   __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3099 
3100   __ mr(klass_index_reg, R3);
3101   __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap);
3102   __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
3103                   R16_thread, klass_index_reg, R5_ARG3);
3104 
3105   // Set an oopmap for the call site.
3106   oop_maps->add_gc_map(gc_map_pc - start, map);
3107 
3108   __ reset_last_Java_frame();
3109 
3110   // Pop the `unpack frame'.
3111   __ pop_frame();
3112 
3113   // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3114 
3115   // Save the return value.
3116   __ mr(unroll_block_reg, R3_RET);
3117 
3118   // Pop the uncommon_trap frame.
3119   __ pop_frame();
3120 
3121   // stack: (caller_of_deoptee, ...).
3122 
3123 #ifdef ASSERT
3124   __ lwz(R22_tmp2, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3125   __ cmpdi(CR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap);
3126   __ asm_assert_eq("OptoRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
3127 #endif
3128 
3129   // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3130   // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3131   // and the frame is effectively not resized.
3132   Register caller_sp = R23_tmp3;
3133   __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3134   __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3135 
3136   // Allocate new interpreter frame(s) and possibly a c2i adapter
3137   // frame.
3138   push_skeleton_frames(masm, false/*deopt*/,
3139                        unroll_block_reg,
3140                        R22_tmp2,
3141                        R23_tmp3,
3142                        R24_tmp4,
3143                        R25_tmp5,
3144                        R26_tmp6);
3145 
3146   // stack: (skeletal interpreter frame, ..., optional skeletal
3147   // interpreter frame, optional c2i, caller of deoptee, ...).
3148 
3149   // Push a dummy `unpack_frame' taking care of float return values.
3150   // Call `Deoptimization::unpack_frames' to layout information in the
3151   // interpreter frames just created.
3152 
3153   // Push a simple "unpack frame" here.
3154   __ push_frame_reg_args(0, R11_scratch1);
3155 
3156   // stack: (unpack frame, skeletal interpreter frame, ..., optional
3157   // skeletal interpreter frame, optional c2i, caller of deoptee,
3158   // ...).
3159 
3160   // Set the "unpack_frame" as last_Java_frame.
3161   __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3162 
3163   // Indicate it is the uncommon trap case.
3164   __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3165   // Let the unpacker layout information in the skeletal frames just
3166   // allocated.
3167   __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3168                   R16_thread, unc_trap_reg);
3169 
3170   __ reset_last_Java_frame();
3171   // Pop the `unpack frame'.
3172   __ pop_frame();
3173   // Restore LR from top interpreter frame.
3174   __ restore_LR(R11_scratch1);
3175 
3176   // stack: (top interpreter frame, ..., optional interpreter frame,
3177   // optional c2i, caller of deoptee, ...).
3178 
3179   __ restore_interpreter_state(R11_scratch1);
3180   __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3181 
3182   // Return to the interpreter entry point.
3183   __ blr();
3184 
3185   masm->flush();
3186 
3187   return UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3188 }
3189 #endif // COMPILER2
3190 
3191 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3192 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
3193   assert(StubRoutines::forward_exception_entry() != nullptr,
3194          "must be generated before");
3195   assert(is_polling_page_id(id), "expected a polling page stub id");
3196 
3197   ResourceMark rm;
3198   OopMapSet *oop_maps = new OopMapSet();
3199   OopMap* map;
3200 
3201   // Allocate space for the code. Setup code generation tools.
3202   const char* name = SharedRuntime::stub_name(id);
3203   CodeBuffer buffer(name, 2048, 1024);
3204   MacroAssembler* masm = new MacroAssembler(&buffer);
3205 
3206   address start = __ pc();
3207   int frame_size_in_bytes = 0;
3208 
3209   RegisterSaver::ReturnPCLocation return_pc_location;
3210   bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
3211   if (cause_return) {
3212     // Nothing to do here. The frame has already been popped in MachEpilogNode.
3213     // Register LR already contains the return pc.
3214     return_pc_location = RegisterSaver::return_pc_is_pre_saved;
3215   } else {
3216     // Use thread()->saved_exception_pc() as return pc.
3217     return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3218   }
3219 
3220   bool save_vectors = (id == StubId::shared_polling_page_vectors_safepoint_handler_id);
3221 
3222   // Save registers, fpu state, and flags. Set R31 = return pc.
3223   map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3224                                                                    &frame_size_in_bytes,
3225                                                                    /*generate_oop_map=*/ true,
3226                                                                    return_pc_location, save_vectors);
3227 
3228   // The following is basically a call_VM. However, we need the precise
3229   // address of the call in order to generate an oopmap. Hence, we do all the
3230   // work ourselves.
3231   __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3232 
3233   // The return address must always be correct so that the frame constructor
3234   // never sees an invalid pc.
3235 
3236   // Do the call
3237   __ call_VM_leaf(call_ptr, R16_thread);
3238   address calls_return_pc = __ last_calls_return_pc();
3239 
3240   // Set an oopmap for the call site. This oopmap will map all
3241   // oop-registers and debug-info registers as callee-saved. This
3242   // will allow deoptimization at this safepoint to find all possible
3243   // debug-info recordings, as well as let GC find all oops.
3244   oop_maps->add_gc_map(calls_return_pc - start, map);
3245 
3246   Label noException;
3247 
3248   // Clear the last Java frame.
3249   __ reset_last_Java_frame();
3250 
3251   BLOCK_COMMENT("  Check pending exception.");
3252   const Register pending_exception = R0;
3253   __ ld(pending_exception, thread_(pending_exception));
3254   __ cmpdi(CR0, pending_exception, 0);
3255   __ beq(CR0, noException);
3256 
3257   // Exception pending
3258   RegisterSaver::restore_live_registers_and_pop_frame(masm,
3259                                                       frame_size_in_bytes,
3260                                                       /*restore_ctr=*/true, save_vectors);
3261 
3262   BLOCK_COMMENT("  Jump to forward_exception_entry.");
3263   // Jump to forward_exception_entry, with the issuing PC in LR
3264   // so it looks like the original nmethod called forward_exception_entry.
3265   __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3266 
3267   // No exception case.
3268   __ BIND(noException);
3269 
3270   if (!cause_return) {
3271     Label no_adjust;
3272     // If our stashed return pc was modified by the runtime we avoid touching it
3273     __ ld(R0, frame_size_in_bytes + _abi0(lr), R1_SP);
3274     __ cmpd(CR0, R0, R31);
3275     __ bne(CR0, no_adjust);
3276 
3277     // Adjust return pc forward to step over the safepoint poll instruction
3278     __ addi(R31, R31, 4);
3279     __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
3280 
3281     __ bind(no_adjust);
3282   }
3283 
3284   // Normal exit, restore registers and exit.
3285   RegisterSaver::restore_live_registers_and_pop_frame(masm,
3286                                                       frame_size_in_bytes,
3287                                                       /*restore_ctr=*/true, save_vectors);
3288 
3289   __ blr();
3290 
3291   // Make sure all code is generated
3292   masm->flush();
3293 
3294   // Fill-out other meta info
3295   // CodeBlob frame size is in words.
3296   return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3297 }
3298 
3299 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3300 //
3301 // Generate a stub that calls into the vm to find out the proper destination
3302 // of a java call. All the argument registers are live at this point
3303 // but since this is generic code we don't know what they are and the caller
3304 // must do any gc of the args.
3305 //
3306 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
3307   assert(is_resolve_id(id), "expected a resolve stub id");
3308 
3309   // allocate space for the code
3310   ResourceMark rm;
3311 
3312   const char* name = SharedRuntime::stub_name(id);
3313   CodeBuffer buffer(name, 1000, 512);
3314   MacroAssembler* masm = new MacroAssembler(&buffer);
3315 
3316   int frame_size_in_bytes;
3317 
3318   OopMapSet *oop_maps = new OopMapSet();
3319   OopMap* map = nullptr;
3320 
3321   address start = __ pc();
3322 
3323   map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3324                                                                    &frame_size_in_bytes,
3325                                                                    /*generate_oop_map*/ true,
3326                                                                    RegisterSaver::return_pc_is_lr);
3327 
3328   // Use noreg as last_Java_pc, the return pc will be reconstructed
3329   // from the physical frame.
3330   __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3331 
3332   int frame_complete = __ offset();
3333 
3334   // Pass R19_method as 2nd (optional) argument, used by
3335   // counter_overflow_stub.
3336   __ call_VM_leaf(destination, R16_thread, R19_method);
3337   address calls_return_pc = __ last_calls_return_pc();
3338   // Set an oopmap for the call site.
3339   // We need this not only for callee-saved registers, but also for volatile
3340   // registers that the compiler might be keeping live across a safepoint.
3341   // Create the oopmap for the call's return pc.
3342   oop_maps->add_gc_map(calls_return_pc - start, map);
3343 
3344   // R3_RET contains the address we are going to jump to assuming no exception got installed.
3345 
3346   // clear last_Java_sp
3347   __ reset_last_Java_frame();
3348 
3349   // Check for pending exceptions.
3350   BLOCK_COMMENT("Check for pending exceptions.");
3351   Label pending;
3352   __ ld(R11_scratch1, thread_(pending_exception));
3353   __ cmpdi(CR0, R11_scratch1, 0);
3354   __ bne(CR0, pending);
3355 
3356   __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3357 
3358   RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3359 
3360   // Get the returned method.
3361   __ get_vm_result_metadata(R19_method);
3362 
3363   __ bctr();
3364 
3365 
3366   // Pending exception after the safepoint.
3367   __ BIND(pending);
3368 
3369   RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3370 
3371   // exception pending => remove activation and forward to exception handler
3372 
3373   __ li(R11_scratch1, 0);
3374   __ ld(R3_ARG1, thread_(pending_exception));
3375   __ std(R11_scratch1, in_bytes(JavaThread::vm_result_oop_offset()), R16_thread);
3376   __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3377 
3378   // -------------
3379   // Make sure all code is generated.
3380   masm->flush();
3381 
3382   // return the blob
3383   // frame_size_words or bytes??
3384   return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3385                                        oop_maps, true);
3386 }
3387 
3388 // Continuation point for throwing of implicit exceptions that are
3389 // not handled in the current activation. Fabricates an exception
3390 // oop and initiates normal exception dispatching in this
3391 // frame. Only callee-saved registers are preserved (through the
3392 // normal register window / RegisterMap handling).  If the compiler
3393 // needs all registers to be preserved between the fault point and
3394 // the exception handler then it must assume responsibility for that
3395 // in AbstractCompiler::continuation_for_implicit_null_exception or
3396 // continuation_for_implicit_division_by_zero_exception. All other
3397 // implicit exceptions (e.g., NullPointerException or
3398 // AbstractMethodError on entry) are either at call sites or
3399 // otherwise assume that stack unwinding will be initiated, so
3400 // caller saved registers were assumed volatile in the compiler.
3401 //
3402 // Note that we generate only this stub into a RuntimeStub, because
3403 // it needs to be properly traversed and ignored during GC, so we
3404 // change the meaning of the "__" macro within this method.
3405 //
3406 // Note: the routine set_pc_not_at_call_for_caller in
3407 // SharedRuntime.cpp requires that this code be generated into a
3408 // RuntimeStub.
3409 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
3410   assert(is_throw_id(id), "expected a throw stub id");
3411 
3412   const char* name = SharedRuntime::stub_name(id);
3413 
3414   ResourceMark rm;
3415   const char* timer_msg = "SharedRuntime generate_throw_exception";
3416   TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3417 
3418   CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
3419   MacroAssembler* masm = new MacroAssembler(&code);
3420 
3421   OopMapSet* oop_maps  = new OopMapSet();
3422   int frame_size_in_bytes = frame::native_abi_reg_args_size;
3423   OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3424 
3425   address start = __ pc();
3426 
3427   __ save_LR(R11_scratch1);
3428 
3429   // Push a frame.
3430   __ push_frame_reg_args(0, R11_scratch1);
3431 
3432   address frame_complete_pc = __ pc();
3433 
3434   // Note that we always have a runtime stub frame on the top of
3435   // stack by this point. Remember the offset of the instruction
3436   // whose address will be moved to R11_scratch1.
3437   address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
3438 
3439   __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
3440 
3441   __ mr(R3_ARG1, R16_thread);
3442   __ call_c(runtime_entry);
3443 
3444   // Set an oopmap for the call site.
3445   oop_maps->add_gc_map((int)(gc_map_pc - start), map);
3446 
3447   __ reset_last_Java_frame();
3448 
3449 #ifdef ASSERT
3450   // Make sure that this code is only executed if there is a pending
3451   // exception.
3452   {
3453     Label L;
3454     __ ld(R0,
3455           in_bytes(Thread::pending_exception_offset()),
3456           R16_thread);
3457     __ cmpdi(CR0, R0, 0);
3458     __ bne(CR0, L);
3459     __ stop("SharedRuntime::throw_exception: no pending exception");
3460     __ bind(L);
3461   }
3462 #endif
3463 
3464   // Pop frame.
3465   __ pop_frame();
3466 
3467   __ restore_LR(R11_scratch1);
3468 
3469   __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
3470   __ mtctr(R11_scratch1);
3471   __ bctr();
3472 
3473   // Create runtime stub with OopMap.
3474   RuntimeStub* stub =
3475     RuntimeStub::new_runtime_stub(name, &code,
3476                                   /*frame_complete=*/ (int)(frame_complete_pc - start),
3477                                   frame_size_in_bytes/wordSize,
3478                                   oop_maps,
3479                                   false);
3480   return stub;
3481 }
3482 
3483 //------------------------------Montgomery multiplication------------------------
3484 //
3485 
3486 // Subtract 0:b from carry:a. Return carry.
3487 static unsigned long
3488 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3489   long i = 0;
3490   unsigned long tmp, tmp2;
3491   __asm__ __volatile__ (
3492     "subfc  %[tmp], %[tmp], %[tmp]   \n" // pre-set CA
3493     "mtctr  %[len]                   \n"
3494     "0:                              \n"
3495     "ldx    %[tmp], %[i], %[a]       \n"
3496     "ldx    %[tmp2], %[i], %[b]      \n"
3497     "subfe  %[tmp], %[tmp2], %[tmp]  \n" // subtract extended
3498     "stdx   %[tmp], %[i], %[a]       \n"
3499     "addi   %[i], %[i], 8            \n"
3500     "bdnz   0b                       \n"
3501     "addme  %[tmp], %[carry]         \n" // carry + CA - 1
3502     : [i]"+b"(i), [tmp]"=&r"(tmp), [tmp2]"=&r"(tmp2)
3503     : [a]"r"(a), [b]"r"(b), [carry]"r"(carry), [len]"r"(len)
3504     : "ctr", "xer", "memory"
3505   );
3506   return tmp;
3507 }
3508 
3509 // Multiply (unsigned) Long A by Long B, accumulating the double-
3510 // length result into the accumulator formed of T0, T1, and T2.
3511 inline void MACC(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3512   unsigned long hi, lo;
3513   __asm__ __volatile__ (
3514     "mulld  %[lo], %[A], %[B]    \n"
3515     "mulhdu %[hi], %[A], %[B]    \n"
3516     "addc   %[T0], %[T0], %[lo]  \n"
3517     "adde   %[T1], %[T1], %[hi]  \n"
3518     "addze  %[T2], %[T2]         \n"
3519     : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3520     : [A]"r"(A), [B]"r"(B)
3521     : "xer"
3522   );
3523 }
3524 
3525 // As above, but add twice the double-length result into the
3526 // accumulator.
3527 inline void MACC2(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3528   unsigned long hi, lo;
3529   __asm__ __volatile__ (
3530     "mulld  %[lo], %[A], %[B]    \n"
3531     "mulhdu %[hi], %[A], %[B]    \n"
3532     "addc   %[T0], %[T0], %[lo]  \n"
3533     "adde   %[T1], %[T1], %[hi]  \n"
3534     "addze  %[T2], %[T2]         \n"
3535     "addc   %[T0], %[T0], %[lo]  \n"
3536     "adde   %[T1], %[T1], %[hi]  \n"
3537     "addze  %[T2], %[T2]         \n"
3538     : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3539     : [A]"r"(A), [B]"r"(B)
3540     : "xer"
3541   );
3542 }
3543 
3544 // Fast Montgomery multiplication. The derivation of the algorithm is
3545 // in "A Cryptographic Library for the Motorola DSP56000,
3546 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3547 static void
3548 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3549                     unsigned long m[], unsigned long inv, int len) {
3550   unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3551   int i;
3552 
3553   assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3554 
3555   for (i = 0; i < len; i++) {
3556     int j;
3557     for (j = 0; j < i; j++) {
3558       MACC(a[j], b[i-j], t0, t1, t2);
3559       MACC(m[j], n[i-j], t0, t1, t2);
3560     }
3561     MACC(a[i], b[0], t0, t1, t2);
3562     m[i] = t0 * inv;
3563     MACC(m[i], n[0], t0, t1, t2);
3564 
3565     assert(t0 == 0, "broken Montgomery multiply");
3566 
3567     t0 = t1; t1 = t2; t2 = 0;
3568   }
3569 
3570   for (i = len; i < 2*len; i++) {
3571     int j;
3572     for (j = i-len+1; j < len; j++) {
3573       MACC(a[j], b[i-j], t0, t1, t2);
3574       MACC(m[j], n[i-j], t0, t1, t2);
3575     }
3576     m[i-len] = t0;
3577     t0 = t1; t1 = t2; t2 = 0;
3578   }
3579 
3580   while (t0) {
3581     t0 = sub(m, n, t0, len);
3582   }
3583 }
3584 
3585 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3586 // multiplies so it should be up to 25% faster than Montgomery
3587 // multiplication. However, its loop control is more complex and it
3588 // may actually run slower on some machines.
3589 static void
3590 montgomery_square(unsigned long a[], unsigned long n[],
3591                   unsigned long m[], unsigned long inv, int len) {
3592   unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3593   int i;
3594 
3595   assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3596 
3597   for (i = 0; i < len; i++) {
3598     int j;
3599     int end = (i+1)/2;
3600     for (j = 0; j < end; j++) {
3601       MACC2(a[j], a[i-j], t0, t1, t2);
3602       MACC(m[j], n[i-j], t0, t1, t2);
3603     }
3604     if ((i & 1) == 0) {
3605       MACC(a[j], a[j], t0, t1, t2);
3606     }
3607     for (; j < i; j++) {
3608       MACC(m[j], n[i-j], t0, t1, t2);
3609     }
3610     m[i] = t0 * inv;
3611     MACC(m[i], n[0], t0, t1, t2);
3612 
3613     assert(t0 == 0, "broken Montgomery square");
3614 
3615     t0 = t1; t1 = t2; t2 = 0;
3616   }
3617 
3618   for (i = len; i < 2*len; i++) {
3619     int start = i-len+1;
3620     int end = start + (len - start)/2;
3621     int j;
3622     for (j = start; j < end; j++) {
3623       MACC2(a[j], a[i-j], t0, t1, t2);
3624       MACC(m[j], n[i-j], t0, t1, t2);
3625     }
3626     if ((i & 1) == 0) {
3627       MACC(a[j], a[j], t0, t1, t2);
3628     }
3629     for (; j < len; j++) {
3630       MACC(m[j], n[i-j], t0, t1, t2);
3631     }
3632     m[i-len] = t0;
3633     t0 = t1; t1 = t2; t2 = 0;
3634   }
3635 
3636   while (t0) {
3637     t0 = sub(m, n, t0, len);
3638   }
3639 }
3640 
3641 // The threshold at which squaring is advantageous was determined
3642 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3643 // Doesn't seem to be relevant for Power8 so we use the same value.
3644 #define MONTGOMERY_SQUARING_THRESHOLD 64
3645 
3646 // Copy len longwords from s to d, word-swapping as we go. The
3647 // destination array is reversed.
3648 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3649   d += len;
3650   while(len-- > 0) {
3651     d--;
3652     unsigned long s_val = *s;
3653     // Swap words in a longword on little endian machines.
3654 #ifdef VM_LITTLE_ENDIAN
3655      s_val = (s_val << 32) | (s_val >> 32);
3656 #endif
3657     *d = s_val;
3658     s++;
3659   }
3660 }
3661 
3662 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3663                                         jint len, jlong inv,
3664                                         jint *m_ints) {
3665   len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3666   assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3667   int longwords = len/2;
3668 
3669   // Make very sure we don't use so much space that the stack might
3670   // overflow. 512 jints corresponds to an 16384-bit integer and
3671   // will use here a total of 8k bytes of stack space.
3672   int divisor = sizeof(unsigned long) * 4;
3673   guarantee(longwords <= 8192 / divisor, "must be");
3674   int total_allocation = longwords * sizeof (unsigned long) * 4;
3675   unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3676 
3677   // Local scratch arrays
3678   unsigned long
3679     *a = scratch + 0 * longwords,
3680     *b = scratch + 1 * longwords,
3681     *n = scratch + 2 * longwords,
3682     *m = scratch + 3 * longwords;
3683 
3684   reverse_words((unsigned long *)a_ints, a, longwords);
3685   reverse_words((unsigned long *)b_ints, b, longwords);
3686   reverse_words((unsigned long *)n_ints, n, longwords);
3687 
3688   ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3689 
3690   reverse_words(m, (unsigned long *)m_ints, longwords);
3691 }
3692 
3693 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3694                                       jint len, jlong inv,
3695                                       jint *m_ints) {
3696   len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3697   assert(len % 2 == 0, "array length in montgomery_square must be even");
3698   int longwords = len/2;
3699 
3700   // Make very sure we don't use so much space that the stack might
3701   // overflow. 512 jints corresponds to an 16384-bit integer and
3702   // will use here a total of 6k bytes of stack space.
3703   int divisor = sizeof(unsigned long) * 3;
3704   guarantee(longwords <= (8192 / divisor), "must be");
3705   int total_allocation = longwords * sizeof (unsigned long) * 3;
3706   unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3707 
3708   // Local scratch arrays
3709   unsigned long
3710     *a = scratch + 0 * longwords,
3711     *n = scratch + 1 * longwords,
3712     *m = scratch + 2 * longwords;
3713 
3714   reverse_words((unsigned long *)a_ints, a, longwords);
3715   reverse_words((unsigned long *)n_ints, n, longwords);
3716 
3717   if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3718     ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3719   } else {
3720     ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3721   }
3722 
3723   reverse_words(m, (unsigned long *)m_ints, longwords);
3724 }
3725 
3726 #if INCLUDE_JFR
3727 
3728 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3729 // It returns a jobject handle to the event writer.
3730 // The handle is dereferenced and the return value is the event writer oop.
3731 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3732   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
3733   CodeBuffer code(name, 512, 64);
3734   MacroAssembler* masm = new MacroAssembler(&code);
3735 
3736   Register tmp1 = R10_ARG8;
3737   Register tmp2 = R9_ARG7;
3738 
3739   int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3740   address start = __ pc();
3741   __ mflr(tmp1);
3742   __ std(tmp1, _abi0(lr), R1_SP);  // save return pc
3743   __ push_frame_reg_args(0, tmp1);
3744   int frame_complete = __ pc() - start;
3745   __ set_last_Java_frame(R1_SP, noreg);
3746   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), R16_thread);
3747   address calls_return_pc = __ last_calls_return_pc();
3748   __ reset_last_Java_frame();
3749   // The handle is dereferenced through a load barrier.
3750   __ resolve_global_jobject(R3_RET, tmp1, tmp2, MacroAssembler::PRESERVATION_NONE);
3751   __ pop_frame();
3752   __ ld(tmp1, _abi0(lr), R1_SP);
3753   __ mtlr(tmp1);
3754   __ blr();
3755 
3756   OopMapSet* oop_maps = new OopMapSet();
3757   OopMap* map = new OopMap(framesize, 0);
3758   oop_maps->add_gc_map(calls_return_pc - start, map);
3759 
3760   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3761     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3762                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3763                                   oop_maps, false);
3764   return stub;
3765 }
3766 
3767 // For c2: call to return a leased buffer.
3768 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3769   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
3770   CodeBuffer code(name, 512, 64);
3771   MacroAssembler* masm = new MacroAssembler(&code);
3772 
3773   Register tmp1 = R10_ARG8;
3774   Register tmp2 = R9_ARG7;
3775 
3776   int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3777   address start = __ pc();
3778   __ mflr(tmp1);
3779   __ std(tmp1, _abi0(lr), R1_SP);  // save return pc
3780   __ push_frame_reg_args(0, tmp1);
3781   int frame_complete = __ pc() - start;
3782   __ set_last_Java_frame(R1_SP, noreg);
3783   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), R16_thread);
3784   address calls_return_pc = __ last_calls_return_pc();
3785   __ reset_last_Java_frame();
3786   __ pop_frame();
3787   __ ld(tmp1, _abi0(lr), R1_SP);
3788   __ mtlr(tmp1);
3789   __ blr();
3790 
3791   OopMapSet* oop_maps = new OopMapSet();
3792   OopMap* map = new OopMap(framesize, 0);
3793   oop_maps->add_gc_map(calls_return_pc - start, map);
3794 
3795   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3796     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3797                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3798                                   oop_maps, false);
3799   return stub;
3800 }
3801 #endif // INCLUDE_JFR
3802 
3803 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
3804 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
3805 
3806 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
3807   Unimplemented();
3808   return 0;
3809 }
3810 
3811 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
3812   Unimplemented();
3813   return nullptr;
3814 }
3815