1 /*
   2  * Copyright (c) 2003, 2023, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2014, 2021, Red Hat Inc. All rights reserved.
   4  * Copyright (c) 2021, Azul Systems, Inc. All rights reserved.
   5  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   6  *
   7  * This code is free software; you can redistribute it and/or modify it
   8  * under the terms of the GNU General Public License version 2 only, as
   9  * published by the Free Software Foundation.
  10  *
  11  * This code is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  14  * version 2 for more details (a copy is included in the LICENSE file that
  15  * accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License version
  18  * 2 along with this work; if not, write to the Free Software Foundation,
  19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  20  *
  21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  22  * or visit www.oracle.com if you need additional information or have any
  23  * questions.
  24  *
  25  */
  26 
  27 #include "precompiled.hpp"
  28 #include "asm/macroAssembler.hpp"
  29 #include "asm/macroAssembler.inline.hpp"
  30 #include "classfile/symbolTable.hpp"
  31 #include "code/codeCache.hpp"
  32 #include "code/compiledIC.hpp"
  33 #include "code/debugInfoRec.hpp"
  34 #include "code/icBuffer.hpp"
  35 #include "code/vtableStubs.hpp"
  36 #include "compiler/oopMap.hpp"
  37 #include "gc/shared/barrierSetAssembler.hpp"
  38 #include "interpreter/interpreter.hpp"
  39 #include "interpreter/interp_masm.hpp"
  40 #include "logging/log.hpp"
  41 #include "memory/resourceArea.hpp"
  42 #include "nativeInst_aarch64.hpp"
  43 #include "oops/compiledICHolder.hpp"
  44 #include "oops/klass.inline.hpp"
  45 #include "oops/method.inline.hpp"
  46 #include "prims/methodHandles.hpp"
  47 #include "runtime/continuation.hpp"
  48 #include "runtime/continuationEntry.inline.hpp"
  49 #include "runtime/globals.hpp"
  50 #include "runtime/jniHandles.hpp"
  51 #include "runtime/safepointMechanism.hpp"
  52 #include "runtime/sharedRuntime.hpp"
  53 #include "runtime/signature.hpp"
  54 #include "runtime/stubRoutines.hpp"
  55 #include "runtime/vframeArray.hpp"
  56 #include "utilities/align.hpp"
  57 #include "utilities/formatBuffer.hpp"
  58 #include "vmreg_aarch64.inline.hpp"
  59 #ifdef COMPILER1
  60 #include "c1/c1_Runtime1.hpp"
  61 #endif
  62 #ifdef COMPILER2
  63 #include "adfiles/ad_aarch64.hpp"
  64 #include "opto/runtime.hpp"
  65 #endif
  66 #if INCLUDE_JVMCI
  67 #include "jvmci/jvmciJavaClasses.hpp"
  68 #endif
  69 
  70 #define __ masm->
  71 
  72 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
  73 
  74 class SimpleRuntimeFrame {
  75 
  76   public:
  77 
  78   // Most of the runtime stubs have this simple frame layout.
  79   // This class exists to make the layout shared in one place.
  80   // Offsets are for compiler stack slots, which are jints.
  81   enum layout {
  82     // The frame sender code expects that rbp will be in the "natural" place and
  83     // will override any oopMap setting for it. We must therefore force the layout
  84     // so that it agrees with the frame sender code.
  85     // we don't expect any arg reg save area so aarch64 asserts that
  86     // frame::arg_reg_save_area_bytes == 0
  87     rfp_off = 0,
  88     rfp_off2,
  89     return_off, return_off2,
  90     framesize
  91   };
  92 };
  93 
  94 // FIXME -- this is used by C1
  95 class RegisterSaver {
  96   const bool _save_vectors;
  97  public:
  98   RegisterSaver(bool save_vectors) : _save_vectors(save_vectors) {}
  99 
 100   OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
 101   void restore_live_registers(MacroAssembler* masm);
 102 
 103   // Offsets into the register save area
 104   // Used by deoptimization when it is managing result register
 105   // values on its own
 106 
 107   int reg_offset_in_bytes(Register r);
 108   int r0_offset_in_bytes()    { return reg_offset_in_bytes(r0); }
 109   int rscratch1_offset_in_bytes()    { return reg_offset_in_bytes(rscratch1); }
 110   int v0_offset_in_bytes();
 111 
 112   // Total stack size in bytes for saving sve predicate registers.
 113   int total_sve_predicate_in_bytes();
 114 
 115   // Capture info about frame layout
 116   // Note this is only correct when not saving full vectors.
 117   enum layout {
 118                 fpu_state_off = 0,
 119                 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1,
 120                 // The frame sender code expects that rfp will be in
 121                 // the "natural" place and will override any oopMap
 122                 // setting for it. We must therefore force the layout
 123                 // so that it agrees with the frame sender code.
 124                 r0_off = fpu_state_off + FPUStateSizeInWords,
 125                 rfp_off = r0_off + (Register::number_of_registers - 2) * Register::max_slots_per_register,
 126                 return_off = rfp_off + Register::max_slots_per_register,      // slot for return address
 127                 reg_save_size = return_off + Register::max_slots_per_register};
 128 
 129 };
 130 
 131 int RegisterSaver::reg_offset_in_bytes(Register r) {
 132   // The integer registers are located above the floating point
 133   // registers in the stack frame pushed by save_live_registers() so the
 134   // offset depends on whether we are saving full vectors, and whether
 135   // those vectors are NEON or SVE.
 136 
 137   int slots_per_vect = FloatRegister::save_slots_per_register;
 138 
 139 #if COMPILER2_OR_JVMCI
 140   if (_save_vectors) {
 141     slots_per_vect = FloatRegister::slots_per_neon_register;
 142 
 143 #ifdef COMPILER2
 144     if (Matcher::supports_scalable_vector()) {
 145       slots_per_vect = Matcher::scalable_vector_reg_size(T_FLOAT);
 146     }
 147 #endif
 148   }
 149 #endif
 150 
 151   int r0_offset = v0_offset_in_bytes() + (slots_per_vect * FloatRegister::number_of_registers) * BytesPerInt;
 152   return r0_offset + r->encoding() * wordSize;
 153 }
 154 
 155 int RegisterSaver::v0_offset_in_bytes() {
 156   // The floating point registers are located above the predicate registers if
 157   // they are present in the stack frame pushed by save_live_registers(). So the
 158   // offset depends on the saved total predicate vectors in the stack frame.
 159   return (total_sve_predicate_in_bytes() / VMRegImpl::stack_slot_size) * BytesPerInt;
 160 }
 161 
 162 int RegisterSaver::total_sve_predicate_in_bytes() {
 163 #ifdef COMPILER2
 164   if (_save_vectors && Matcher::supports_scalable_vector()) {
 165     return (Matcher::scalable_vector_reg_size(T_BYTE) >> LogBitsPerByte) *
 166            PRegister::number_of_registers;
 167   }
 168 #endif
 169   return 0;
 170 }
 171 
 172 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
 173   bool use_sve = false;
 174   int sve_vector_size_in_bytes = 0;
 175   int sve_vector_size_in_slots = 0;
 176   int sve_predicate_size_in_slots = 0;
 177   int total_predicate_in_bytes = total_sve_predicate_in_bytes();
 178   int total_predicate_in_slots = total_predicate_in_bytes / VMRegImpl::stack_slot_size;
 179 
 180 #ifdef COMPILER2
 181   use_sve = Matcher::supports_scalable_vector();
 182   if (use_sve) {
 183     sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
 184     sve_vector_size_in_slots = Matcher::scalable_vector_reg_size(T_FLOAT);
 185     sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
 186   }
 187 #endif
 188 
 189 #if COMPILER2_OR_JVMCI
 190   if (_save_vectors) {
 191     int extra_save_slots_per_register = 0;
 192     // Save upper half of vector registers
 193     if (use_sve) {
 194       extra_save_slots_per_register = sve_vector_size_in_slots - FloatRegister::save_slots_per_register;
 195     } else {
 196       extra_save_slots_per_register = FloatRegister::extra_save_slots_per_neon_register;
 197     }
 198     int extra_vector_bytes = extra_save_slots_per_register *
 199                              VMRegImpl::stack_slot_size *
 200                              FloatRegister::number_of_registers;
 201     additional_frame_words += ((extra_vector_bytes + total_predicate_in_bytes) / wordSize);
 202   }
 203 #else
 204   assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
 205 #endif
 206 
 207   int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
 208                                      reg_save_size * BytesPerInt, 16);
 209   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 210   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 211   // The caller will allocate additional_frame_words
 212   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 213   // CodeBlob frame size is in words.
 214   int frame_size_in_words = frame_size_in_bytes / wordSize;
 215   *total_frame_words = frame_size_in_words;
 216 
 217   // Save Integer and Float registers.
 218   __ enter();
 219   __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
 220 
 221   // Set an oopmap for the call site.  This oopmap will map all
 222   // oop-registers and debug-info registers as callee-saved.  This
 223   // will allow deoptimization at this safepoint to find all possible
 224   // debug-info recordings, as well as let GC find all oops.
 225 
 226   OopMapSet *oop_maps = new OopMapSet();
 227   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 228 
 229   for (int i = 0; i < Register::number_of_registers; i++) {
 230     Register r = as_Register(i);
 231     if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
 232       // SP offsets are in 4-byte words.
 233       // Register slots are 8 bytes wide, 32 floating-point registers.
 234       int sp_offset = Register::max_slots_per_register * i +
 235                       FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
 236       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
 237     }
 238   }
 239 
 240   for (int i = 0; i < FloatRegister::number_of_registers; i++) {
 241     FloatRegister r = as_FloatRegister(i);
 242     int sp_offset = 0;
 243     if (_save_vectors) {
 244       sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
 245                             (FloatRegister::slots_per_neon_register * i);
 246     } else {
 247       sp_offset = FloatRegister::save_slots_per_register * i;
 248     }
 249     oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), r->as_VMReg());
 250   }
 251 
 252   return oop_map;
 253 }
 254 
 255 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
 256 #ifdef COMPILER2
 257   __ pop_CPU_state(_save_vectors, Matcher::supports_scalable_vector(),
 258                    Matcher::scalable_vector_reg_size(T_BYTE), total_sve_predicate_in_bytes());
 259 #else
 260 #if !INCLUDE_JVMCI
 261   assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
 262 #endif
 263   __ pop_CPU_state(_save_vectors);
 264 #endif
 265   __ ldp(rfp, lr, Address(__ post(sp, 2 * wordSize)));
 266   __ authenticate_return_address();
 267 }
 268 
 269 // Is vector's size (in bytes) bigger than a size saved by default?
 270 // 8 bytes vector registers are saved by default on AArch64.
 271 // The SVE supported min vector size is 8 bytes and we need to save
 272 // predicate registers when the vector size is 8 bytes as well.
 273 bool SharedRuntime::is_wide_vector(int size) {
 274   return size > 8 || (UseSVE > 0 && size >= 8);
 275 }
 276 
 277 // ---------------------------------------------------------------------------
 278 // Read the array of BasicTypes from a signature, and compute where the
 279 // arguments should go.  Values in the VMRegPair regs array refer to 4-byte
 280 // quantities.  Values less than VMRegImpl::stack0 are registers, those above
 281 // refer to 4-byte stack slots.  All stack slots are based off of the stack pointer
 282 // as framesizes are fixed.
 283 // VMRegImpl::stack0 refers to the first slot 0(sp).
 284 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
 285 // Register up to Register::number_of_registers are the 64-bit
 286 // integer registers.
 287 
 288 // Note: the INPUTS in sig_bt are in units of Java argument words,
 289 // which are 64-bit.  The OUTPUTS are in 32-bit units.
 290 
 291 // The Java calling convention is a "shifted" version of the C ABI.
 292 // By skipping the first C ABI register we can call non-static jni
 293 // methods with small numbers of arguments without having to shuffle
 294 // the arguments at all. Since we control the java ABI we ought to at
 295 // least get some advantage out of it.
 296 
 297 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
 298                                            VMRegPair *regs,
 299                                            int total_args_passed) {
 300 
 301   // Create the mapping between argument positions and
 302   // registers.
 303   static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
 304     j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7
 305   };
 306   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
 307     j_farg0, j_farg1, j_farg2, j_farg3,
 308     j_farg4, j_farg5, j_farg6, j_farg7
 309   };
 310 
 311 
 312   uint int_args = 0;
 313   uint fp_args = 0;
 314   uint stk_args = 0; // inc by 2 each time
 315 
 316   for (int i = 0; i < total_args_passed; i++) {
 317     switch (sig_bt[i]) {
 318     case T_BOOLEAN:
 319     case T_CHAR:
 320     case T_BYTE:
 321     case T_SHORT:
 322     case T_INT:
 323       if (int_args < Argument::n_int_register_parameters_j) {
 324         regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 325       } else {
 326         regs[i].set1(VMRegImpl::stack2reg(stk_args));
 327         stk_args += 2;
 328       }
 329       break;
 330     case T_VOID:
 331       // halves of T_LONG or T_DOUBLE
 332       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 333       regs[i].set_bad();
 334       break;
 335     case T_LONG:
 336       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 337       // fall through
 338     case T_OBJECT:
 339     case T_ARRAY:
 340     case T_ADDRESS:
 341       if (int_args < Argument::n_int_register_parameters_j) {
 342         regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 343       } else {
 344         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 345         stk_args += 2;
 346       }
 347       break;
 348     case T_FLOAT:
 349       if (fp_args < Argument::n_float_register_parameters_j) {
 350         regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 351       } else {
 352         regs[i].set1(VMRegImpl::stack2reg(stk_args));
 353         stk_args += 2;
 354       }
 355       break;
 356     case T_DOUBLE:
 357       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 358       if (fp_args < Argument::n_float_register_parameters_j) {
 359         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 360       } else {
 361         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 362         stk_args += 2;
 363       }
 364       break;
 365     default:
 366       ShouldNotReachHere();
 367       break;
 368     }
 369   }
 370 
 371   return align_up(stk_args, 2);
 372 }
 373 
 374 
 375 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
 376 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
 377 
 378 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
 379 
 380   // Create the mapping between argument positions and registers.
 381 
 382   static const Register INT_ArgReg[java_return_convention_max_int] = {
 383     r0 /* j_rarg7 */, j_rarg6, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
 384   };
 385 
 386   static const FloatRegister FP_ArgReg[java_return_convention_max_float] = {
 387     j_farg0, j_farg1, j_farg2, j_farg3, j_farg4, j_farg5, j_farg6, j_farg7
 388   };
 389 
 390   uint int_args = 0;
 391   uint fp_args = 0;
 392 
 393   for (int i = 0; i < total_args_passed; i++) {
 394     switch (sig_bt[i]) {
 395     case T_BOOLEAN:
 396     case T_CHAR:
 397     case T_BYTE:
 398     case T_SHORT:
 399     case T_INT:
 400       if (int_args < SharedRuntime::java_return_convention_max_int) {
 401         regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
 402         int_args ++;
 403       } else {
 404         return -1;
 405       }
 406       break;
 407     case T_VOID:
 408       // halves of T_LONG or T_DOUBLE
 409       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 410       regs[i].set_bad();
 411       break;
 412     case T_LONG:
 413       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 414       // fall through
 415     case T_OBJECT:
 416     case T_ARRAY:
 417     case T_ADDRESS:
 418       // Should T_METADATA be added to java_calling_convention as well ?
 419     case T_METADATA:
 420       if (int_args < SharedRuntime::java_return_convention_max_int) {
 421         regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
 422         int_args ++;
 423       } else {
 424         return -1;
 425       }
 426       break;
 427     case T_FLOAT:
 428       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 429         regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
 430         fp_args ++;
 431       } else {
 432         return -1;
 433       }
 434       break;
 435     case T_DOUBLE:
 436       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 437       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 438         regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
 439         fp_args ++;
 440       } else {
 441         return -1;
 442       }
 443       break;
 444     default:
 445       ShouldNotReachHere();
 446       break;
 447     }
 448   }
 449 
 450   return int_args + fp_args;
 451 }
 452 
 453 // Patch the callers callsite with entry to compiled code if it exists.
 454 static void patch_callers_callsite(MacroAssembler *masm) {
 455   Label L;
 456   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 457   __ cbz(rscratch1, L);
 458 
 459   __ enter();
 460   __ push_CPU_state();
 461 
 462   // VM needs caller's callsite
 463   // VM needs target method
 464   // This needs to be a long call since we will relocate this adapter to
 465   // the codeBuffer and it may not reach
 466 
 467 #ifndef PRODUCT
 468   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 469 #endif
 470 
 471   __ mov(c_rarg0, rmethod);
 472   __ mov(c_rarg1, lr);
 473   __ authenticate_return_address(c_rarg1);
 474   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 475   __ blr(rscratch1);
 476 
 477   // Explicit isb required because fixup_callers_callsite may change the code
 478   // stream.
 479   __ safepoint_isb();
 480 
 481   __ pop_CPU_state();
 482   // restore sp
 483   __ leave();
 484   __ bind(L);
 485 }
 486 
 487 // For each inline type argument, sig includes the list of fields of
 488 // the inline type. This utility function computes the number of
 489 // arguments for the call if inline types are passed by reference (the
 490 // calling convention the interpreter expects).
 491 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
 492   int total_args_passed = 0;
 493   if (InlineTypePassFieldsAsArgs) {
 494      for (int i = 0; i < sig_extended->length(); i++) {
 495        BasicType bt = sig_extended->at(i)._bt;
 496        if (bt == T_METADATA) {
 497          // In sig_extended, an inline type argument starts with:
 498          // T_METADATA, followed by the types of the fields of the
 499          // inline type and T_VOID to mark the end of the value
 500          // type. Inline types are flattened so, for instance, in the
 501          // case of an inline type with an int field and an inline type
 502          // field that itself has 2 fields, an int and a long:
 503          // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
 504          // slot for the T_LONG) T_VOID (inner inline type) T_VOID
 505          // (outer inline type)
 506          total_args_passed++;
 507          int vt = 1;
 508          do {
 509            i++;
 510            BasicType bt = sig_extended->at(i)._bt;
 511            BasicType prev_bt = sig_extended->at(i-1)._bt;
 512            if (bt == T_METADATA) {
 513              vt++;
 514            } else if (bt == T_VOID &&
 515                       prev_bt != T_LONG &&
 516                       prev_bt != T_DOUBLE) {
 517              vt--;
 518            }
 519          } while (vt != 0);
 520        } else {
 521          total_args_passed++;
 522        }
 523      }
 524   } else {
 525     total_args_passed = sig_extended->length();
 526   }
 527 
 528   return total_args_passed;
 529 }
 530 
 531 
 532 static void gen_c2i_adapter_helper(MacroAssembler* masm,
 533                                    BasicType bt,
 534                                    BasicType prev_bt,
 535                                    size_t size_in_bytes,
 536                                    const VMRegPair& reg_pair,
 537                                    const Address& to,
 538                                    Register tmp1,
 539                                    Register tmp2,
 540                                    Register tmp3,
 541                                    int extraspace,
 542                                    bool is_oop) {
 543   if (bt == T_VOID) {
 544     assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
 545     return;
 546   }
 547 
 548   // Say 4 args:
 549   // i   st_off
 550   // 0   32 T_LONG
 551   // 1   24 T_VOID
 552   // 2   16 T_OBJECT
 553   // 3    8 T_BOOL
 554   // -    0 return address
 555   //
 556   // However to make thing extra confusing. Because we can fit a Java long/double in
 557   // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 558   // leaves one slot empty and only stores to a single slot. In this case the
 559   // slot that is occupied is the T_VOID slot. See I said it was confusing.
 560 
 561   bool wide = (size_in_bytes == wordSize);
 562   VMReg r_1 = reg_pair.first();
 563   VMReg r_2 = reg_pair.second();
 564   assert(r_2->is_valid() == wide, "invalid size");
 565   if (!r_1->is_valid()) {
 566     assert(!r_2->is_valid(), "");
 567     return;
 568   }
 569 
 570   if (!r_1->is_FloatRegister()) {
 571     Register val = r25;
 572     if (r_1->is_stack()) {
 573       // memory to memory use r25 (scratch registers is used by store_heap_oop)
 574       int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 575       __ load_sized_value(val, Address(sp, ld_off), size_in_bytes, /* is_signed */ false);
 576     } else {
 577       val = r_1->as_Register();
 578     }
 579     assert_different_registers(to.base(), val, tmp1, tmp2, tmp3);
 580     if (is_oop) {
 581       __ store_heap_oop(to, val, tmp1, tmp2, tmp3, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
 582     } else {
 583       __ store_sized_value(to, val, size_in_bytes);
 584     }
 585   } else {
 586     if (wide) {
 587       __ strd(r_1->as_FloatRegister(), to);
 588     } else {
 589       // only a float use just part of the slot
 590       __ strs(r_1->as_FloatRegister(), to);
 591     }
 592   }
 593 }
 594 
 595 static void gen_c2i_adapter(MacroAssembler *masm,
 596                             const GrowableArray<SigEntry>* sig_extended,
 597                             const VMRegPair *regs,
 598                             bool requires_clinit_barrier,
 599                             address& c2i_no_clinit_check_entry,
 600                             Label& skip_fixup,
 601                             address start,
 602                             OopMapSet* oop_maps,
 603                             int& frame_complete,
 604                             int& frame_size_in_words,
 605                             bool alloc_inline_receiver) {
 606   if (requires_clinit_barrier && VM_Version::supports_fast_class_init_checks()) {
 607     Label L_skip_barrier;
 608 
 609     { // Bypass the barrier for non-static methods
 610       __ ldrw(rscratch1, Address(rmethod, Method::access_flags_offset()));
 611       __ andsw(zr, rscratch1, JVM_ACC_STATIC);
 612       __ br(Assembler::EQ, L_skip_barrier); // non-static
 613     }
 614 
 615     __ load_method_holder(rscratch2, rmethod);
 616     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
 617     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 618 
 619     __ bind(L_skip_barrier);
 620     c2i_no_clinit_check_entry = __ pc();
 621   }
 622 
 623   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 624   bs->c2i_entry_barrier(masm);
 625 
 626   // Before we get into the guts of the C2I adapter, see if we should be here
 627   // at all.  We've come from compiled code and are attempting to jump to the
 628   // interpreter, which means the caller made a static call to get here
 629   // (vcalls always get a compiled target if there is one).  Check for a
 630   // compiled target.  If there is one, we need to patch the caller's call.
 631   patch_callers_callsite(masm);
 632 
 633   __ bind(skip_fixup);
 634 
 635   // Name some registers to be used in the following code. We can use
 636   // anything except r0-r7 which are arguments in the Java calling
 637   // convention, rmethod (r12), and r13 which holds the outgoing sender
 638   // SP for the interpreter.
 639   Register buf_array = r10;   // Array of buffered inline types
 640   Register buf_oop = r11;     // Buffered inline type oop
 641   Register tmp1 = r15;
 642   Register tmp2 = r16;
 643   Register tmp3 = r17;
 644 
 645   if (InlineTypePassFieldsAsArgs) {
 646     // Is there an inline type argument?
 647     bool has_inline_argument = false;
 648     for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
 649       has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
 650     }
 651     if (has_inline_argument) {
 652       // There is at least an inline type argument: we're coming from
 653       // compiled code so we have no buffers to back the inline types
 654       // Allocate the buffers here with a runtime call.
 655       RegisterSaver reg_save(false /* save_vectors */);
 656       OopMap* map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
 657 
 658       frame_complete = __ offset();
 659       address the_pc = __ pc();
 660 
 661       Label retaddr;
 662       __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
 663 
 664       __ mov(c_rarg0, rthread);
 665       __ mov(c_rarg1, rmethod);
 666       __ mov(c_rarg2, (int64_t)alloc_inline_receiver);
 667 
 668       __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
 669       __ blr(rscratch1);
 670       __ bind(retaddr);
 671 
 672       oop_maps->add_gc_map(__ pc() - start, map);
 673       __ reset_last_Java_frame(false);
 674 
 675       reg_save.restore_live_registers(masm);
 676 
 677       Label no_exception;
 678       __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
 679       __ cbz(rscratch1, no_exception);
 680 
 681       __ str(zr, Address(rthread, JavaThread::vm_result_offset()));
 682       __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
 683       __ b(RuntimeAddress(StubRoutines::forward_exception_entry()));
 684 
 685       __ bind(no_exception);
 686 
 687       // We get an array of objects from the runtime call
 688       __ get_vm_result(buf_array, rthread);
 689       __ get_vm_result_2(rmethod, rthread); // TODO: required to keep the callee Method live?
 690     }
 691   }
 692 
 693   // Since all args are passed on the stack, total_args_passed *
 694   // Interpreter::stackElementSize is the space we need.
 695 
 696   int total_args_passed = compute_total_args_passed_int(sig_extended);
 697   int extraspace = total_args_passed * Interpreter::stackElementSize;
 698 
 699   // stack is aligned, keep it that way
 700   extraspace = align_up(extraspace, StackAlignmentInBytes);
 701 
 702   // set senderSP value
 703   __ mov(r19_sender_sp, sp);
 704 
 705   __ sub(sp, sp, extraspace);
 706 
 707   // Now write the args into the outgoing interpreter space
 708 
 709   // next_arg_comp is the next argument from the compiler point of
 710   // view (inline type fields are passed in registers/on the stack). In
 711   // sig_extended, an inline type argument starts with: T_METADATA,
 712   // followed by the types of the fields of the inline type and T_VOID
 713   // to mark the end of the inline type. ignored counts the number of
 714   // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
 715   // used to get the buffer for that argument from the pool of buffers
 716   // we allocated above and want to pass to the
 717   // interpreter. next_arg_int is the next argument from the
 718   // interpreter point of view (inline types are passed by reference).
 719   for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
 720        next_arg_comp < sig_extended->length(); next_arg_comp++) {
 721     assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
 722     assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
 723     BasicType bt = sig_extended->at(next_arg_comp)._bt;
 724     int st_off = (total_args_passed - next_arg_int - 1) * Interpreter::stackElementSize;
 725     if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
 726       int next_off = st_off - Interpreter::stackElementSize;
 727       const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
 728       const VMRegPair reg_pair = regs[next_arg_comp-ignored];
 729       size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
 730       gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 731                              size_in_bytes, reg_pair, Address(sp, offset), tmp1, tmp2, tmp3, extraspace, false);
 732       next_arg_int++;
 733 #ifdef ASSERT
 734       if (bt == T_LONG || bt == T_DOUBLE) {
 735         // Overwrite the unused slot with known junk
 736         __ mov(rscratch1, CONST64(0xdeadffffdeadaaaa));
 737         __ str(rscratch1, Address(sp, st_off));
 738       }
 739 #endif /* ASSERT */
 740     } else {
 741       ignored++;
 742       // get the buffer from the just allocated pool of buffers
 743       int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
 744       __ load_heap_oop(buf_oop, Address(buf_array, index), tmp1, tmp2);
 745       next_vt_arg++; next_arg_int++;
 746       int vt = 1;
 747       // write fields we get from compiled code in registers/stack
 748       // slots to the buffer: we know we are done with that inline type
 749       // argument when we hit the T_VOID that acts as an end of inline
 750       // type delimiter for this inline type. Inline types are flattened
 751       // so we might encounter embedded inline types. Each entry in
 752       // sig_extended contains a field offset in the buffer.
 753       Label L_null;
 754       do {
 755         next_arg_comp++;
 756         BasicType bt = sig_extended->at(next_arg_comp)._bt;
 757         BasicType prev_bt = sig_extended->at(next_arg_comp - 1)._bt;
 758         if (bt == T_METADATA) {
 759           vt++;
 760           ignored++;
 761         } else if (bt == T_VOID && prev_bt != T_LONG && prev_bt != T_DOUBLE) {
 762           vt--;
 763           ignored++;
 764         } else {
 765           int off = sig_extended->at(next_arg_comp)._offset;
 766           if (off == -1) {
 767             // Nullable inline type argument, emit null check
 768             VMReg reg = regs[next_arg_comp-ignored].first();
 769             Label L_notNull;
 770             if (reg->is_stack()) {
 771               int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 772               __ ldrb(tmp1, Address(sp, ld_off));
 773               __ cbnz(tmp1, L_notNull);
 774             } else {
 775               __ cbnz(reg->as_Register(), L_notNull);
 776             }
 777             __ str(zr, Address(sp, st_off));
 778             __ b(L_null);
 779             __ bind(L_notNull);
 780             continue;
 781           }
 782           assert(off > 0, "offset in object should be positive");
 783           size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
 784           bool is_oop = is_reference_type(bt);
 785           gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 786                                  size_in_bytes, regs[next_arg_comp-ignored], Address(buf_oop, off), tmp1, tmp2, tmp3, extraspace, is_oop);
 787         }
 788       } while (vt != 0);
 789       // pass the buffer to the interpreter
 790       __ str(buf_oop, Address(sp, st_off));
 791       __ bind(L_null);
 792     }
 793   }
 794 
 795   __ mov(esp, sp); // Interp expects args on caller's expression stack
 796 
 797   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
 798   __ br(rscratch1);
 799 }
 800 
 801 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, int comp_args_on_stack, const GrowableArray<SigEntry>* sig, const VMRegPair *regs) {
 802 
 803 
 804   // Note: r19_sender_sp contains the senderSP on entry. We must
 805   // preserve it since we may do a i2c -> c2i transition if we lose a
 806   // race where compiled code goes non-entrant while we get args
 807   // ready.
 808 
 809   // Adapters are frameless.
 810 
 811   // An i2c adapter is frameless because the *caller* frame, which is
 812   // interpreted, routinely repairs its own esp (from
 813   // interpreter_frame_last_sp), even if a callee has modified the
 814   // stack pointer.  It also recalculates and aligns sp.
 815 
 816   // A c2i adapter is frameless because the *callee* frame, which is
 817   // interpreted, routinely repairs its caller's sp (from sender_sp,
 818   // which is set up via the senderSP register).
 819 
 820   // In other words, if *either* the caller or callee is interpreted, we can
 821   // get the stack pointer repaired after a call.
 822 
 823   // This is why c2i and i2c adapters cannot be indefinitely composed.
 824   // In particular, if a c2i adapter were to somehow call an i2c adapter,
 825   // both caller and callee would be compiled methods, and neither would
 826   // clean up the stack pointer changes performed by the two adapters.
 827   // If this happens, control eventually transfers back to the compiled
 828   // caller, but with an uncorrected stack, causing delayed havoc.
 829 
 830   if (VerifyAdapterCalls &&
 831       (Interpreter::code() != nullptr || StubRoutines::final_stubs_code() != nullptr)) {
 832 #if 0
 833     // So, let's test for cascading c2i/i2c adapters right now.
 834     //  assert(Interpreter::contains($return_addr) ||
 835     //         StubRoutines::contains($return_addr),
 836     //         "i2c adapter must return to an interpreter frame");
 837     __ block_comment("verify_i2c { ");
 838     Label L_ok;
 839     if (Interpreter::code() != nullptr) {
 840       range_check(masm, rax, r11,
 841                   Interpreter::code()->code_start(), Interpreter::code()->code_end(),
 842                   L_ok);
 843     }
 844     if (StubRoutines::initial_stubs_code() != nullptr) {
 845       range_check(masm, rax, r11,
 846                   StubRoutines::initial_stubs_code()->code_begin(),
 847                   StubRoutines::initial_stubs_code()->code_end(),
 848                   L_ok);
 849     }
 850     if (StubRoutines::final_stubs_code() != nullptr) {
 851       range_check(masm, rax, r11,
 852                   StubRoutines::final_stubs_code()->code_begin(),
 853                   StubRoutines::final_stubs_code()->code_end(),
 854                   L_ok);
 855     }
 856     const char* msg = "i2c adapter must return to an interpreter frame";
 857     __ block_comment(msg);
 858     __ stop(msg);
 859     __ bind(L_ok);
 860     __ block_comment("} verify_i2ce ");
 861 #endif
 862   }
 863 
 864   // Cut-out for having no stack args.
 865   int comp_words_on_stack = 0;
 866   if (comp_args_on_stack) {
 867      comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
 868      __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
 869      __ andr(sp, rscratch1, -16);
 870   }
 871 
 872   // Will jump to the compiled code just as if compiled code was doing it.
 873   // Pre-load the register-jump target early, to schedule it better.
 874   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_inline_offset())));
 875 
 876 #if INCLUDE_JVMCI
 877   if (EnableJVMCI) {
 878     // check if this call should be routed towards a specific entry point
 879     __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 880     Label no_alternative_target;
 881     __ cbz(rscratch2, no_alternative_target);
 882     __ mov(rscratch1, rscratch2);
 883     __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 884     __ bind(no_alternative_target);
 885   }
 886 #endif // INCLUDE_JVMCI
 887 
 888   int total_args_passed = sig->length();
 889 
 890   // Now generate the shuffle code.
 891   for (int i = 0; i < total_args_passed; i++) {
 892     BasicType bt = sig->at(i)._bt;
 893     if (bt == T_VOID) {
 894       assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
 895       continue;
 896     }
 897 
 898     // Pick up 0, 1 or 2 words from SP+offset.
 899     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), "scrambled load targets?");
 900 
 901     // Load in argument order going down.
 902     int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 903     // Point to interpreter value (vs. tag)
 904     int next_off = ld_off - Interpreter::stackElementSize;
 905     //
 906     //
 907     //
 908     VMReg r_1 = regs[i].first();
 909     VMReg r_2 = regs[i].second();
 910     if (!r_1->is_valid()) {
 911       assert(!r_2->is_valid(), "");
 912       continue;
 913     }
 914     if (r_1->is_stack()) {
 915       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 916       int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
 917       if (!r_2->is_valid()) {
 918         // sign extend???
 919         __ ldrsw(rscratch2, Address(esp, ld_off));
 920         __ str(rscratch2, Address(sp, st_off));
 921       } else {
 922         //
 923         // We are using two optoregs. This can be either T_OBJECT,
 924         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 925         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 926         // So we must adjust where to pick up the data to match the
 927         // interpreter.
 928         //
 929         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 930         // are accessed as negative so LSW is at LOW address
 931 
 932         // ld_off is MSW so get LSW
 933         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
 934         __ ldr(rscratch2, Address(esp, offset));
 935         // st_off is LSW (i.e. reg.first())
 936          __ str(rscratch2, Address(sp, st_off));
 937        }
 938      } else if (r_1->is_Register()) {  // Register argument
 939        Register r = r_1->as_Register();
 940        if (r_2->is_valid()) {
 941          //
 942          // We are using two VMRegs. This can be either T_OBJECT,
 943          // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 944          // two slots but only uses one for thr T_LONG or T_DOUBLE case
 945          // So we must adjust where to pick up the data to match the
 946          // interpreter.
 947 
 948         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
 949 
 950          // this can be a misaligned move
 951          __ ldr(r, Address(esp, offset));
 952        } else {
 953          // sign extend and use a full word?
 954          __ ldrw(r, Address(esp, ld_off));
 955        }
 956      } else {
 957        if (!r_2->is_valid()) {
 958          __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
 959        } else {
 960          __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
 961        }
 962      }
 963    }
 964 
 965 
 966   __ mov(rscratch2, rscratch1);
 967   __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1
 968   __ mov(rscratch1, rscratch2);
 969 
 970   // 6243940 We might end up in handle_wrong_method if
 971   // the callee is deoptimized as we race thru here. If that
 972   // happens we don't want to take a safepoint because the
 973   // caller frame will look interpreted and arguments are now
 974   // "compiled" so it is much better to make this transition
 975   // invisible to the stack walking code. Unfortunately if
 976   // we try and find the callee by normal means a safepoint
 977   // is possible. So we stash the desired callee in the thread
 978   // and the vm will find there should this case occur.
 979 
 980   __ str(rmethod, Address(rthread, JavaThread::callee_target_offset()));
 981   __ br(rscratch1);
 982 }
 983 
 984 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
 985 
 986   Label ok;
 987 
 988   Register holder = rscratch2;
 989   Register receiver = j_rarg0;
 990   Register tmp = r10;  // A call-clobbered register not used for arg passing
 991 
 992   // -------------------------------------------------------------------------
 993   // Generate a C2I adapter.  On entry we know rmethod holds the Method* during calls
 994   // to the interpreter.  The args start out packed in the compiled layout.  They
 995   // need to be unpacked into the interpreter layout.  This will almost always
 996   // require some stack space.  We grow the current (compiled) stack, then repack
 997   // the args.  We  finally end in a jump to the generic interpreter entry point.
 998   // On exit from the interpreter, the interpreter will restore our SP (lest the
 999   // compiled code, which relies solely on SP and not FP, get sick).
1000 
1001   {
1002     __ block_comment("c2i_unverified_entry {");
1003     __ load_klass(rscratch1, receiver);
1004     __ ldr(tmp, Address(holder, CompiledICHolder::holder_klass_offset()));
1005     __ cmp(rscratch1, tmp);
1006     __ ldr(rmethod, Address(holder, CompiledICHolder::holder_metadata_offset()));
1007     __ br(Assembler::EQ, ok);
1008     __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1009 
1010     __ bind(ok);
1011     // Method might have been compiled since the call site was patched to
1012     // interpreted; if that is the case treat it as a miss so we can get
1013     // the call site corrected.
1014     __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
1015     __ cbz(rscratch1, skip_fixup);
1016     __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1017     __ block_comment("} c2i_unverified_entry");
1018   }
1019 }
1020 
1021 
1022 // ---------------------------------------------------------------
1023 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
1024                                                             int comp_args_on_stack,
1025                                                             const GrowableArray<SigEntry>* sig,
1026                                                             const VMRegPair* regs,
1027                                                             const GrowableArray<SigEntry>* sig_cc,
1028                                                             const VMRegPair* regs_cc,
1029                                                             const GrowableArray<SigEntry>* sig_cc_ro,
1030                                                             const VMRegPair* regs_cc_ro,
1031                                                             AdapterFingerPrint* fingerprint,
1032                                                             AdapterBlob*& new_adapter,
1033                                                             bool allocate_code_blob) {
1034 
1035   address i2c_entry = __ pc();
1036   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
1037 
1038   address c2i_unverified_entry        = __ pc();
1039   address c2i_unverified_inline_entry = __ pc();
1040   Label skip_fixup;
1041 
1042   gen_inline_cache_check(masm, skip_fixup);
1043 
1044   OopMapSet* oop_maps = new OopMapSet();
1045   int frame_complete = CodeOffsets::frame_never_safe;
1046   int frame_size_in_words = 0;
1047 
1048   // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
1049   address c2i_no_clinit_check_entry = nullptr;
1050   address c2i_inline_ro_entry = __ pc();
1051   if (regs_cc != regs_cc_ro) {
1052     // No class init barrier needed because method is guaranteed to be non-static
1053     gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, c2i_no_clinit_check_entry,
1054                     skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1055     skip_fixup.reset();
1056   }
1057 
1058   // Scalarized c2i adapter
1059   address c2i_entry        = __ pc();
1060   address c2i_inline_entry = __ pc();
1061   gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1062                   skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1063 
1064   // Non-scalarized c2i adapter
1065   if (regs != regs_cc) {
1066     c2i_unverified_inline_entry = __ pc();
1067     Label inline_entry_skip_fixup;
1068     gen_inline_cache_check(masm, inline_entry_skip_fixup);
1069 
1070     c2i_inline_entry = __ pc();
1071     gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1072                     inline_entry_skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1073   }
1074 
1075 
1076   // The c2i adapter might safepoint and trigger a GC. The caller must make sure that
1077   // the GC knows about the location of oop argument locations passed to the c2i adapter.
1078   if (allocate_code_blob) {
1079     bool caller_must_gc_arguments = (regs != regs_cc);
1080     new_adapter = AdapterBlob::create(masm->code(), frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1081   }
1082 
1083   return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_inline_entry, c2i_inline_ro_entry, c2i_unverified_entry, c2i_unverified_inline_entry, c2i_no_clinit_check_entry);
1084 }
1085 
1086 static int c_calling_convention_priv(const BasicType *sig_bt,
1087                                          VMRegPair *regs,
1088                                          int total_args_passed) {
1089 
1090 // We return the amount of VMRegImpl stack slots we need to reserve for all
1091 // the arguments NOT counting out_preserve_stack_slots.
1092 
1093     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1094       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
1095     };
1096     static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1097       c_farg0, c_farg1, c_farg2, c_farg3,
1098       c_farg4, c_farg5, c_farg6, c_farg7
1099     };
1100 
1101     uint int_args = 0;
1102     uint fp_args = 0;
1103     uint stk_args = 0; // inc by 2 each time
1104 
1105     for (int i = 0; i < total_args_passed; i++) {
1106       switch (sig_bt[i]) {
1107       case T_BOOLEAN:
1108       case T_CHAR:
1109       case T_BYTE:
1110       case T_SHORT:
1111       case T_INT:
1112         if (int_args < Argument::n_int_register_parameters_c) {
1113           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
1114         } else {
1115 #ifdef __APPLE__
1116           // Less-than word types are stored one after another.
1117           // The code is unable to handle this so bailout.
1118           return -1;
1119 #endif
1120           regs[i].set1(VMRegImpl::stack2reg(stk_args));
1121           stk_args += 2;
1122         }
1123         break;
1124       case T_LONG:
1125         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1126         // fall through
1127       case T_OBJECT:
1128       case T_ARRAY:
1129       case T_ADDRESS:
1130       case T_METADATA:
1131         if (int_args < Argument::n_int_register_parameters_c) {
1132           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
1133         } else {
1134           regs[i].set2(VMRegImpl::stack2reg(stk_args));
1135           stk_args += 2;
1136         }
1137         break;
1138       case T_FLOAT:
1139         if (fp_args < Argument::n_float_register_parameters_c) {
1140           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
1141         } else {
1142 #ifdef __APPLE__
1143           // Less-than word types are stored one after another.
1144           // The code is unable to handle this so bailout.
1145           return -1;
1146 #endif
1147           regs[i].set1(VMRegImpl::stack2reg(stk_args));
1148           stk_args += 2;
1149         }
1150         break;
1151       case T_DOUBLE:
1152         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1153         if (fp_args < Argument::n_float_register_parameters_c) {
1154           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
1155         } else {
1156           regs[i].set2(VMRegImpl::stack2reg(stk_args));
1157           stk_args += 2;
1158         }
1159         break;
1160       case T_VOID: // Halves of longs and doubles
1161         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
1162         regs[i].set_bad();
1163         break;
1164       default:
1165         ShouldNotReachHere();
1166         break;
1167       }
1168     }
1169 
1170   return stk_args;
1171 }
1172 
1173 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
1174                                              uint num_bits,
1175                                              uint total_args_passed) {
1176   Unimplemented();
1177   return 0;
1178 }
1179 
1180 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1181                                          VMRegPair *regs,
1182                                          int total_args_passed)
1183 {
1184   int result = c_calling_convention_priv(sig_bt, regs, total_args_passed);
1185   guarantee(result >= 0, "Unsupported arguments configuration");
1186   return result;
1187 }
1188 
1189 
1190 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1191   // We always ignore the frame_slots arg and just use the space just below frame pointer
1192   // which by this time is free to use
1193   switch (ret_type) {
1194   case T_FLOAT:
1195     __ strs(v0, Address(rfp, -wordSize));
1196     break;
1197   case T_DOUBLE:
1198     __ strd(v0, Address(rfp, -wordSize));
1199     break;
1200   case T_VOID:  break;
1201   default: {
1202     __ str(r0, Address(rfp, -wordSize));
1203     }
1204   }
1205 }
1206 
1207 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1208   // We always ignore the frame_slots arg and just use the space just below frame pointer
1209   // which by this time is free to use
1210   switch (ret_type) {
1211   case T_FLOAT:
1212     __ ldrs(v0, Address(rfp, -wordSize));
1213     break;
1214   case T_DOUBLE:
1215     __ ldrd(v0, Address(rfp, -wordSize));
1216     break;
1217   case T_VOID:  break;
1218   default: {
1219     __ ldr(r0, Address(rfp, -wordSize));
1220     }
1221   }
1222 }
1223 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1224   RegSet x;
1225   for ( int i = first_arg ; i < arg_count ; i++ ) {
1226     if (args[i].first()->is_Register()) {
1227       x = x + args[i].first()->as_Register();
1228     } else if (args[i].first()->is_FloatRegister()) {
1229       __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize)));
1230     }
1231   }
1232   __ push(x, sp);
1233 }
1234 
1235 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1236   RegSet x;
1237   for ( int i = first_arg ; i < arg_count ; i++ ) {
1238     if (args[i].first()->is_Register()) {
1239       x = x + args[i].first()->as_Register();
1240     } else {
1241       ;
1242     }
1243   }
1244   __ pop(x, sp);
1245   for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
1246     if (args[i].first()->is_Register()) {
1247       ;
1248     } else if (args[i].first()->is_FloatRegister()) {
1249       __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize)));
1250     }
1251   }
1252 }
1253 
1254 static void verify_oop_args(MacroAssembler* masm,
1255                             const methodHandle& method,
1256                             const BasicType* sig_bt,
1257                             const VMRegPair* regs) {
1258   Register temp_reg = r19;  // not part of any compiled calling seq
1259   if (VerifyOops) {
1260     for (int i = 0; i < method->size_of_parameters(); i++) {
1261       if (sig_bt[i] == T_OBJECT ||
1262           sig_bt[i] == T_ARRAY) {
1263         VMReg r = regs[i].first();
1264         assert(r->is_valid(), "bad oop arg");
1265         if (r->is_stack()) {
1266           __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1267           __ verify_oop(temp_reg);
1268         } else {
1269           __ verify_oop(r->as_Register());
1270         }
1271       }
1272     }
1273   }
1274 }
1275 
1276 // on exit, sp points to the ContinuationEntry
1277 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
1278   assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1279   assert(in_bytes(ContinuationEntry::cont_offset())  % VMRegImpl::stack_slot_size == 0, "");
1280   assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1281 
1282   stack_slots += (int)ContinuationEntry::size()/wordSize;
1283   __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
1284 
1285   OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize)/ VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1286 
1287   __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1288   __ str(rscratch1, Address(sp, ContinuationEntry::parent_offset()));
1289   __ mov(rscratch1, sp); // we can't use sp as the source in str
1290   __ str(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1291 
1292   return map;
1293 }
1294 
1295 // on entry c_rarg1 points to the continuation
1296 //          sp points to ContinuationEntry
1297 //          c_rarg3 -- isVirtualThread
1298 static void fill_continuation_entry(MacroAssembler* masm) {
1299 #ifdef ASSERT
1300   __ movw(rscratch1, ContinuationEntry::cookie_value());
1301   __ strw(rscratch1, Address(sp, ContinuationEntry::cookie_offset()));
1302 #endif
1303 
1304   __ str (c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
1305   __ strw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
1306   __ str (zr,      Address(sp, ContinuationEntry::chunk_offset()));
1307   __ strw(zr,      Address(sp, ContinuationEntry::argsize_offset()));
1308   __ strw(zr,      Address(sp, ContinuationEntry::pin_count_offset()));
1309 
1310   __ ldr(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1311   __ str(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1312   __ ldr(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset()));
1313   __ str(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
1314 
1315   __ str(zr, Address(rthread, JavaThread::cont_fastpath_offset()));
1316   __ str(zr, Address(rthread, JavaThread::held_monitor_count_offset()));
1317 }
1318 
1319 // on entry, sp points to the ContinuationEntry
1320 // on exit, rfp points to the spilled rfp in the entry frame
1321 static void continuation_enter_cleanup(MacroAssembler* masm) {
1322 #ifndef PRODUCT
1323   Label OK;
1324   __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1325   __ cmp(sp, rscratch1);
1326   __ br(Assembler::EQ, OK);
1327   __ stop("incorrect sp1");
1328   __ bind(OK);
1329 #endif
1330 
1331   __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1332   __ str(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1333   __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
1334   __ str(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset()));
1335 
1336   __ ldr(rscratch2, Address(sp, ContinuationEntry::parent_offset()));
1337   __ str(rscratch2, Address(rthread, JavaThread::cont_entry_offset()));
1338   __ add(rfp, sp, (int)ContinuationEntry::size());
1339 }
1340 
1341 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1342 // On entry: c_rarg1 -- the continuation object
1343 //           c_rarg2 -- isContinue
1344 //           c_rarg3 -- isVirtualThread
1345 static void gen_continuation_enter(MacroAssembler* masm,
1346                                  const methodHandle& method,
1347                                  const BasicType* sig_bt,
1348                                  const VMRegPair* regs,
1349                                  int& exception_offset,
1350                                  OopMapSet*oop_maps,
1351                                  int& frame_complete,
1352                                  int& stack_slots,
1353                                  int& interpreted_entry_offset,
1354                                  int& compiled_entry_offset) {
1355   //verify_oop_args(masm, method, sig_bt, regs);
1356   Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1357 
1358   address start = __ pc();
1359 
1360   Label call_thaw, exit;
1361 
1362   // i2i entry used at interp_only_mode only
1363   interpreted_entry_offset = __ pc() - start;
1364   {
1365 
1366 #ifdef ASSERT
1367     Label is_interp_only;
1368     __ ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
1369     __ cbnzw(rscratch1, is_interp_only);
1370     __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1371     __ bind(is_interp_only);
1372 #endif
1373 
1374     // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1375     __ ldr(c_rarg1, Address(esp, Interpreter::stackElementSize*2));
1376     __ ldr(c_rarg2, Address(esp, Interpreter::stackElementSize*1));
1377     __ ldr(c_rarg3, Address(esp, Interpreter::stackElementSize*0));
1378     __ push_cont_fastpath(rthread);
1379 
1380     __ enter();
1381     stack_slots = 2; // will be adjusted in setup
1382     OopMap* map = continuation_enter_setup(masm, stack_slots);
1383     // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1384     // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1385 
1386     fill_continuation_entry(masm);
1387 
1388     __ cbnz(c_rarg2, call_thaw);
1389 
1390     const address tr_call = __ trampoline_call(resolve);
1391     if (tr_call == nullptr) {
1392       fatal("CodeCache is full at gen_continuation_enter");
1393     }
1394 
1395     oop_maps->add_gc_map(__ pc() - start, map);
1396     __ post_call_nop();
1397 
1398     __ b(exit);
1399 
1400     CodeBuffer* cbuf = masm->code_section()->outer();
1401     address stub = CompiledStaticCall::emit_to_interp_stub(*cbuf, tr_call);
1402     if (stub == nullptr) {
1403       fatal("CodeCache is full at gen_continuation_enter");
1404     }
1405   }
1406 
1407   // compiled entry
1408   __ align(CodeEntryAlignment);
1409   compiled_entry_offset = __ pc() - start;
1410 
1411   __ enter();
1412   stack_slots = 2; // will be adjusted in setup
1413   OopMap* map = continuation_enter_setup(masm, stack_slots);
1414   frame_complete = __ pc() - start;
1415 
1416   fill_continuation_entry(masm);
1417 
1418   __ cbnz(c_rarg2, call_thaw);
1419 
1420   const address tr_call = __ trampoline_call(resolve);
1421   if (tr_call == nullptr) {
1422     fatal("CodeCache is full at gen_continuation_enter");
1423   }
1424 
1425   oop_maps->add_gc_map(__ pc() - start, map);
1426   __ post_call_nop();
1427 
1428   __ b(exit);
1429 
1430   __ bind(call_thaw);
1431 
1432   __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1433   oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1434   ContinuationEntry::_return_pc_offset = __ pc() - start;
1435   __ post_call_nop();
1436 
1437   __ bind(exit);
1438   continuation_enter_cleanup(masm);
1439   __ leave();
1440   __ ret(lr);
1441 
1442   /// exception handling
1443 
1444   exception_offset = __ pc() - start;
1445   {
1446       __ mov(r19, r0); // save return value contaning the exception oop in callee-saved R19
1447 
1448       continuation_enter_cleanup(masm);
1449 
1450       __ ldr(c_rarg1, Address(rfp, wordSize)); // return address
1451       __ authenticate_return_address(c_rarg1);
1452       __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1);
1453 
1454       // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc
1455 
1456       __ mov(r1, r0); // the exception handler
1457       __ mov(r0, r19); // restore return value contaning the exception oop
1458       __ verify_oop(r0);
1459 
1460       __ leave();
1461       __ mov(r3, lr);
1462       __ br(r1); // the exception handler
1463   }
1464 
1465   CodeBuffer* cbuf = masm->code_section()->outer();
1466   address stub = CompiledStaticCall::emit_to_interp_stub(*cbuf, tr_call);
1467   if (stub == nullptr) {
1468     fatal("CodeCache is full at gen_continuation_enter");
1469   }
1470 }
1471 
1472 static void gen_continuation_yield(MacroAssembler* masm,
1473                                    const methodHandle& method,
1474                                    const BasicType* sig_bt,
1475                                    const VMRegPair* regs,
1476                                    OopMapSet* oop_maps,
1477                                    int& frame_complete,
1478                                    int& stack_slots,
1479                                    int& compiled_entry_offset) {
1480     enum layout {
1481       rfp_off1,
1482       rfp_off2,
1483       lr_off,
1484       lr_off2,
1485       framesize // inclusive of return address
1486     };
1487     // assert(is_even(framesize/2), "sp not 16-byte aligned");
1488     stack_slots = framesize /  VMRegImpl::slots_per_word;
1489     assert(stack_slots == 2, "recheck layout");
1490 
1491     address start = __ pc();
1492 
1493     compiled_entry_offset = __ pc() - start;
1494     __ enter();
1495 
1496     __ mov(c_rarg1, sp);
1497 
1498     frame_complete = __ pc() - start;
1499     address the_pc = __ pc();
1500 
1501     __ post_call_nop(); // this must be exactly after the pc value that is pushed into the frame info, we use this nop for fast CodeBlob lookup
1502 
1503     __ mov(c_rarg0, rthread);
1504     __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
1505     __ call_VM_leaf(Continuation::freeze_entry(), 2);
1506     __ reset_last_Java_frame(true);
1507 
1508     Label pinned;
1509 
1510     __ cbnz(r0, pinned);
1511 
1512     // We've succeeded, set sp to the ContinuationEntry
1513     __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1514     __ mov(sp, rscratch1);
1515     continuation_enter_cleanup(masm);
1516 
1517     __ bind(pinned); // pinned -- return to caller
1518 
1519     // handle pending exception thrown by freeze
1520     __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
1521     Label ok;
1522     __ cbz(rscratch1, ok);
1523     __ leave();
1524     __ lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
1525     __ br(rscratch1);
1526     __ bind(ok);
1527 
1528     __ leave();
1529     __ ret(lr);
1530 
1531     OopMap* map = new OopMap(framesize, 1);
1532     oop_maps->add_gc_map(the_pc - start, map);
1533 }
1534 
1535 static void gen_special_dispatch(MacroAssembler* masm,
1536                                  const methodHandle& method,
1537                                  const BasicType* sig_bt,
1538                                  const VMRegPair* regs) {
1539   verify_oop_args(masm, method, sig_bt, regs);
1540   vmIntrinsics::ID iid = method->intrinsic_id();
1541 
1542   // Now write the args into the outgoing interpreter space
1543   bool     has_receiver   = false;
1544   Register receiver_reg   = noreg;
1545   int      member_arg_pos = -1;
1546   Register member_reg     = noreg;
1547   int      ref_kind       = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1548   if (ref_kind != 0) {
1549     member_arg_pos = method->size_of_parameters() - 1;  // trailing MemberName argument
1550     member_reg = r19;  // known to be free at this point
1551     has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1552   } else if (iid == vmIntrinsics::_invokeBasic) {
1553     has_receiver = true;
1554   } else if (iid == vmIntrinsics::_linkToNative) {
1555     member_arg_pos = method->size_of_parameters() - 1;  // trailing NativeEntryPoint argument
1556     member_reg = r19;  // known to be free at this point
1557   } else {
1558     fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1559   }
1560 
1561   if (member_reg != noreg) {
1562     // Load the member_arg into register, if necessary.
1563     SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1564     VMReg r = regs[member_arg_pos].first();
1565     if (r->is_stack()) {
1566       __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1567     } else {
1568       // no data motion is needed
1569       member_reg = r->as_Register();
1570     }
1571   }
1572 
1573   if (has_receiver) {
1574     // Make sure the receiver is loaded into a register.
1575     assert(method->size_of_parameters() > 0, "oob");
1576     assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1577     VMReg r = regs[0].first();
1578     assert(r->is_valid(), "bad receiver arg");
1579     if (r->is_stack()) {
1580       // Porting note:  This assumes that compiled calling conventions always
1581       // pass the receiver oop in a register.  If this is not true on some
1582       // platform, pick a temp and load the receiver from stack.
1583       fatal("receiver always in a register");
1584       receiver_reg = r2;  // known to be free at this point
1585       __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1586     } else {
1587       // no data motion is needed
1588       receiver_reg = r->as_Register();
1589     }
1590   }
1591 
1592   // Figure out which address we are really jumping to:
1593   MethodHandles::generate_method_handle_dispatch(masm, iid,
1594                                                  receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1595 }
1596 
1597 // ---------------------------------------------------------------------------
1598 // Generate a native wrapper for a given method.  The method takes arguments
1599 // in the Java compiled code convention, marshals them to the native
1600 // convention (handlizes oops, etc), transitions to native, makes the call,
1601 // returns to java state (possibly blocking), unhandlizes any result and
1602 // returns.
1603 //
1604 // Critical native functions are a shorthand for the use of
1605 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1606 // functions.  The wrapper is expected to unpack the arguments before
1607 // passing them to the callee. Critical native functions leave the state _in_Java,
1608 // since they block out GC.
1609 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1610 // block and the check for pending exceptions it's impossible for them
1611 // to be thrown.
1612 //
1613 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1614                                                 const methodHandle& method,
1615                                                 int compile_id,
1616                                                 BasicType* in_sig_bt,
1617                                                 VMRegPair* in_regs,
1618                                                 BasicType ret_type) {
1619   if (method->is_continuation_native_intrinsic()) {
1620     int exception_offset = -1;
1621     OopMapSet* oop_maps = new OopMapSet();
1622     int frame_complete = -1;
1623     int stack_slots = -1;
1624     int interpreted_entry_offset = -1;
1625     int vep_offset = -1;
1626     if (method->is_continuation_enter_intrinsic()) {
1627       gen_continuation_enter(masm,
1628                              method,
1629                              in_sig_bt,
1630                              in_regs,
1631                              exception_offset,
1632                              oop_maps,
1633                              frame_complete,
1634                              stack_slots,
1635                              interpreted_entry_offset,
1636                              vep_offset);
1637     } else if (method->is_continuation_yield_intrinsic()) {
1638       gen_continuation_yield(masm,
1639                              method,
1640                              in_sig_bt,
1641                              in_regs,
1642                              oop_maps,
1643                              frame_complete,
1644                              stack_slots,
1645                              vep_offset);
1646     } else {
1647       guarantee(false, "Unknown Continuation native intrinsic");
1648     }
1649 
1650 #ifdef ASSERT
1651     if (method->is_continuation_enter_intrinsic()) {
1652       assert(interpreted_entry_offset != -1, "Must be set");
1653       assert(exception_offset != -1,         "Must be set");
1654     } else {
1655       assert(interpreted_entry_offset == -1, "Must be unset");
1656       assert(exception_offset == -1,         "Must be unset");
1657     }
1658     assert(frame_complete != -1,    "Must be set");
1659     assert(stack_slots != -1,       "Must be set");
1660     assert(vep_offset != -1,        "Must be set");
1661 #endif
1662 
1663     __ flush();
1664     nmethod* nm = nmethod::new_native_nmethod(method,
1665                                               compile_id,
1666                                               masm->code(),
1667                                               vep_offset,
1668                                               frame_complete,
1669                                               stack_slots,
1670                                               in_ByteSize(-1),
1671                                               in_ByteSize(-1),
1672                                               oop_maps,
1673                                               exception_offset);
1674     if (method->is_continuation_enter_intrinsic()) {
1675       ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1676     } else if (method->is_continuation_yield_intrinsic()) {
1677       _cont_doYield_stub = nm;
1678     } else {
1679       guarantee(false, "Unknown Continuation native intrinsic");
1680     }
1681     return nm;
1682   }
1683 
1684   if (method->is_method_handle_intrinsic()) {
1685     vmIntrinsics::ID iid = method->intrinsic_id();
1686     intptr_t start = (intptr_t)__ pc();
1687     int vep_offset = ((intptr_t)__ pc()) - start;
1688 
1689     // First instruction must be a nop as it may need to be patched on deoptimisation
1690     __ nop();
1691     gen_special_dispatch(masm,
1692                          method,
1693                          in_sig_bt,
1694                          in_regs);
1695     int frame_complete = ((intptr_t)__ pc()) - start;  // not complete, period
1696     __ flush();
1697     int stack_slots = SharedRuntime::out_preserve_stack_slots();  // no out slots at all, actually
1698     return nmethod::new_native_nmethod(method,
1699                                        compile_id,
1700                                        masm->code(),
1701                                        vep_offset,
1702                                        frame_complete,
1703                                        stack_slots / VMRegImpl::slots_per_word,
1704                                        in_ByteSize(-1),
1705                                        in_ByteSize(-1),
1706                                        nullptr);
1707   }
1708   address native_func = method->native_function();
1709   assert(native_func != nullptr, "must have function");
1710 
1711   // An OopMap for lock (and class if static)
1712   OopMapSet *oop_maps = new OopMapSet();
1713   intptr_t start = (intptr_t)__ pc();
1714 
1715   // We have received a description of where all the java arg are located
1716   // on entry to the wrapper. We need to convert these args to where
1717   // the jni function will expect them. To figure out where they go
1718   // we convert the java signature to a C signature by inserting
1719   // the hidden arguments as arg[0] and possibly arg[1] (static method)
1720 
1721   const int total_in_args = method->size_of_parameters();
1722   int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1723 
1724   BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1725   VMRegPair* out_regs   = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1726   BasicType* in_elem_bt = nullptr;
1727 
1728   int argc = 0;
1729   out_sig_bt[argc++] = T_ADDRESS;
1730   if (method->is_static()) {
1731     out_sig_bt[argc++] = T_OBJECT;
1732   }
1733 
1734   for (int i = 0; i < total_in_args ; i++ ) {
1735     out_sig_bt[argc++] = in_sig_bt[i];
1736   }
1737 
1738   // Now figure out where the args must be stored and how much stack space
1739   // they require.
1740   int out_arg_slots;
1741   out_arg_slots = c_calling_convention_priv(out_sig_bt, out_regs, total_c_args);
1742 
1743   if (out_arg_slots < 0) {
1744     return nullptr;
1745   }
1746 
1747   // Compute framesize for the wrapper.  We need to handlize all oops in
1748   // incoming registers
1749 
1750   // Calculate the total number of stack slots we will need.
1751 
1752   // First count the abi requirement plus all of the outgoing args
1753   int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1754 
1755   // Now the space for the inbound oop handle area
1756   int total_save_slots = 8 * VMRegImpl::slots_per_word;  // 8 arguments passed in registers
1757 
1758   int oop_handle_offset = stack_slots;
1759   stack_slots += total_save_slots;
1760 
1761   // Now any space we need for handlizing a klass if static method
1762 
1763   int klass_slot_offset = 0;
1764   int klass_offset = -1;
1765   int lock_slot_offset = 0;
1766   bool is_static = false;
1767 
1768   if (method->is_static()) {
1769     klass_slot_offset = stack_slots;
1770     stack_slots += VMRegImpl::slots_per_word;
1771     klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1772     is_static = true;
1773   }
1774 
1775   // Plus a lock if needed
1776 
1777   if (method->is_synchronized()) {
1778     lock_slot_offset = stack_slots;
1779     stack_slots += VMRegImpl::slots_per_word;
1780   }
1781 
1782   // Now a place (+2) to save return values or temp during shuffling
1783   // + 4 for return address (which we own) and saved rfp
1784   stack_slots += 6;
1785 
1786   // Ok The space we have allocated will look like:
1787   //
1788   //
1789   // FP-> |                     |
1790   //      |---------------------|
1791   //      | 2 slots for moves   |
1792   //      |---------------------|
1793   //      | lock box (if sync)  |
1794   //      |---------------------| <- lock_slot_offset
1795   //      | klass (if static)   |
1796   //      |---------------------| <- klass_slot_offset
1797   //      | oopHandle area      |
1798   //      |---------------------| <- oop_handle_offset (8 java arg registers)
1799   //      | outbound memory     |
1800   //      | based arguments     |
1801   //      |                     |
1802   //      |---------------------|
1803   //      |                     |
1804   // SP-> | out_preserved_slots |
1805   //
1806   //
1807 
1808 
1809   // Now compute actual number of stack words we need rounding to make
1810   // stack properly aligned.
1811   stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1812 
1813   int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1814 
1815   // First thing make an ic check to see if we should even be here
1816 
1817   // We are free to use all registers as temps without saving them and
1818   // restoring them except rfp. rfp is the only callee save register
1819   // as far as the interpreter and the compiler(s) are concerned.
1820 
1821 
1822   const Register ic_reg = rscratch2;
1823   const Register receiver = j_rarg0;
1824 
1825   Label hit;
1826   Label exception_pending;
1827 
1828   assert_different_registers(ic_reg, receiver, rscratch1);
1829   __ verify_oop(receiver);
1830   __ cmp_klass(receiver, ic_reg, rscratch1);
1831   __ br(Assembler::EQ, hit);
1832 
1833   __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1834 
1835   // Verified entry point must be aligned
1836   __ align(8);
1837 
1838   __ bind(hit);
1839 
1840   int vep_offset = ((intptr_t)__ pc()) - start;
1841 
1842   // If we have to make this method not-entrant we'll overwrite its
1843   // first instruction with a jump.  For this action to be legal we
1844   // must ensure that this first instruction is a B, BL, NOP, BKPT,
1845   // SVC, HVC, or SMC.  Make it a NOP.
1846   __ nop();
1847 
1848   if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1849     Label L_skip_barrier;
1850     __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass*
1851     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
1852     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1853 
1854     __ bind(L_skip_barrier);
1855   }
1856 
1857   // Generate stack overflow check
1858   __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1859 
1860   // Generate a new frame for the wrapper.
1861   __ enter();
1862   // -2 because return address is already present and so is saved rfp
1863   __ sub(sp, sp, stack_size - 2*wordSize);
1864 
1865   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1866   bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1867 
1868   // Frame is now completed as far as size and linkage.
1869   int frame_complete = ((intptr_t)__ pc()) - start;
1870 
1871   // We use r20 as the oop handle for the receiver/klass
1872   // It is callee save so it survives the call to native
1873 
1874   const Register oop_handle_reg = r20;
1875 
1876   //
1877   // We immediately shuffle the arguments so that any vm call we have to
1878   // make from here on out (sync slow path, jvmti, etc.) we will have
1879   // captured the oops from our caller and have a valid oopMap for
1880   // them.
1881 
1882   // -----------------
1883   // The Grand Shuffle
1884 
1885   // The Java calling convention is either equal (linux) or denser (win64) than the
1886   // c calling convention. However the because of the jni_env argument the c calling
1887   // convention always has at least one more (and two for static) arguments than Java.
1888   // Therefore if we move the args from java -> c backwards then we will never have
1889   // a register->register conflict and we don't have to build a dependency graph
1890   // and figure out how to break any cycles.
1891   //
1892 
1893   // Record esp-based slot for receiver on stack for non-static methods
1894   int receiver_offset = -1;
1895 
1896   // This is a trick. We double the stack slots so we can claim
1897   // the oops in the caller's frame. Since we are sure to have
1898   // more args than the caller doubling is enough to make
1899   // sure we can capture all the incoming oop args from the
1900   // caller.
1901   //
1902   OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1903 
1904   // Mark location of rfp (someday)
1905   // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp));
1906 
1907 
1908   int float_args = 0;
1909   int int_args = 0;
1910 
1911 #ifdef ASSERT
1912   bool reg_destroyed[Register::number_of_registers];
1913   bool freg_destroyed[FloatRegister::number_of_registers];
1914   for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1915     reg_destroyed[r] = false;
1916   }
1917   for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1918     freg_destroyed[f] = false;
1919   }
1920 
1921 #endif /* ASSERT */
1922 
1923   // For JNI natives the incoming and outgoing registers are offset upwards.
1924   GrowableArray<int> arg_order(2 * total_in_args);
1925   VMRegPair tmp_vmreg;
1926   tmp_vmreg.set2(r19->as_VMReg());
1927 
1928   for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1929     arg_order.push(i);
1930     arg_order.push(c_arg);
1931   }
1932 
1933   int temploc = -1;
1934   for (int ai = 0; ai < arg_order.length(); ai += 2) {
1935     int i = arg_order.at(ai);
1936     int c_arg = arg_order.at(ai + 1);
1937     __ block_comment(err_msg("move %d -> %d", i, c_arg));
1938     assert(c_arg != -1 && i != -1, "wrong order");
1939 #ifdef ASSERT
1940     if (in_regs[i].first()->is_Register()) {
1941       assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1942     } else if (in_regs[i].first()->is_FloatRegister()) {
1943       assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1944     }
1945     if (out_regs[c_arg].first()->is_Register()) {
1946       reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1947     } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1948       freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1949     }
1950 #endif /* ASSERT */
1951     switch (in_sig_bt[i]) {
1952       case T_ARRAY:
1953       case T_OBJECT:
1954         __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1955                        ((i == 0) && (!is_static)),
1956                        &receiver_offset);
1957         int_args++;
1958         break;
1959       case T_VOID:
1960         break;
1961 
1962       case T_FLOAT:
1963         __ float_move(in_regs[i], out_regs[c_arg]);
1964         float_args++;
1965         break;
1966 
1967       case T_DOUBLE:
1968         assert( i + 1 < total_in_args &&
1969                 in_sig_bt[i + 1] == T_VOID &&
1970                 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1971         __ double_move(in_regs[i], out_regs[c_arg]);
1972         float_args++;
1973         break;
1974 
1975       case T_LONG :
1976         __ long_move(in_regs[i], out_regs[c_arg]);
1977         int_args++;
1978         break;
1979 
1980       case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1981 
1982       default:
1983         __ move32_64(in_regs[i], out_regs[c_arg]);
1984         int_args++;
1985     }
1986   }
1987 
1988   // point c_arg at the first arg that is already loaded in case we
1989   // need to spill before we call out
1990   int c_arg = total_c_args - total_in_args;
1991 
1992   // Pre-load a static method's oop into c_rarg1.
1993   if (method->is_static()) {
1994 
1995     //  load oop into a register
1996     __ movoop(c_rarg1,
1997               JNIHandles::make_local(method->method_holder()->java_mirror()));
1998 
1999     // Now handlize the static class mirror it's known not-null.
2000     __ str(c_rarg1, Address(sp, klass_offset));
2001     map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2002 
2003     // Now get the handle
2004     __ lea(c_rarg1, Address(sp, klass_offset));
2005     // and protect the arg if we must spill
2006     c_arg--;
2007   }
2008 
2009   // Change state to native (we save the return address in the thread, since it might not
2010   // be pushed on the stack when we do a stack traversal).
2011   // We use the same pc/oopMap repeatedly when we call out
2012 
2013   Label native_return;
2014   __ set_last_Java_frame(sp, noreg, native_return, rscratch1);
2015 
2016   Label dtrace_method_entry, dtrace_method_entry_done;
2017   {
2018     uint64_t offset;
2019     __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset);
2020     __ ldrb(rscratch1, Address(rscratch1, offset));
2021     __ cbnzw(rscratch1, dtrace_method_entry);
2022     __ bind(dtrace_method_entry_done);
2023   }
2024 
2025   // RedefineClasses() tracing support for obsolete method entry
2026   if (log_is_enabled(Trace, redefine, class, obsolete)) {
2027     // protect the args we've loaded
2028     save_args(masm, total_c_args, c_arg, out_regs);
2029     __ mov_metadata(c_rarg1, method());
2030     __ call_VM_leaf(
2031       CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
2032       rthread, c_rarg1);
2033     restore_args(masm, total_c_args, c_arg, out_regs);
2034   }
2035 
2036   // Lock a synchronized method
2037 
2038   // Register definitions used by locking and unlocking
2039 
2040   const Register swap_reg = r0;
2041   const Register obj_reg  = r19;  // Will contain the oop
2042   const Register lock_reg = r13;  // Address of compiler lock object (BasicLock)
2043   const Register old_hdr  = r13;  // value of old header at unlock time
2044   const Register lock_tmp = r14;  // Temporary used by lightweight_lock/unlock
2045   const Register tmp = lr;
2046 
2047   Label slow_path_lock;
2048   Label lock_done;
2049 
2050   if (method->is_synchronized()) {
2051     Label count;
2052     const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
2053 
2054     // Get the handle (the 2nd argument)
2055     __ mov(oop_handle_reg, c_rarg1);
2056 
2057     // Get address of the box
2058 
2059     __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2060 
2061     // Load the oop from the handle
2062     __ ldr(obj_reg, Address(oop_handle_reg, 0));
2063 
2064     if (LockingMode == LM_MONITOR) {
2065       __ b(slow_path_lock);
2066     } else if (LockingMode == LM_LEGACY) {
2067       // Load (object->mark() | 1) into swap_reg %r0
2068       __ ldr(rscratch1, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2069       __ orr(swap_reg, rscratch1, 1);
2070       if (EnableValhalla) {
2071         // Mask inline_type bit such that we go to the slow path if object is an inline type
2072         __ andr(swap_reg, swap_reg, ~((int) markWord::inline_type_bit_in_place));
2073       }
2074 
2075       // Save (object->mark() | 1) into BasicLock's displaced header
2076       __ str(swap_reg, Address(lock_reg, mark_word_offset));
2077 
2078       // src -> dest iff dest == r0 else r0 <- dest
2079       __ cmpxchg_obj_header(r0, lock_reg, obj_reg, rscratch1, count, /*fallthrough*/nullptr);
2080 
2081       // Hmm should this move to the slow path code area???
2082 
2083       // Test if the oopMark is an obvious stack pointer, i.e.,
2084       //  1) (mark & 3) == 0, and
2085       //  2) sp <= mark < mark + os::pagesize()
2086       // These 3 tests can be done by evaluating the following
2087       // expression: ((mark - sp) & (3 - os::vm_page_size())),
2088       // assuming both stack pointer and pagesize have their
2089       // least significant 2 bits clear.
2090       // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg
2091 
2092       __ sub(swap_reg, sp, swap_reg);
2093       __ neg(swap_reg, swap_reg);
2094       __ ands(swap_reg, swap_reg, 3 - (int)os::vm_page_size());
2095 
2096       // Save the test result, for recursive case, the result is zero
2097       __ str(swap_reg, Address(lock_reg, mark_word_offset));
2098       __ br(Assembler::NE, slow_path_lock);
2099     } else {
2100       assert(LockingMode == LM_LIGHTWEIGHT, "must be");
2101       __ ldr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2102       __ lightweight_lock(obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
2103     }
2104     __ bind(count);
2105     __ increment(Address(rthread, JavaThread::held_monitor_count_offset()));
2106 
2107     // Slow path will re-enter here
2108     __ bind(lock_done);
2109   }
2110 
2111 
2112   // Finally just about ready to make the JNI call
2113 
2114   // get JNIEnv* which is first argument to native
2115   __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset())));
2116 
2117   // Now set thread in native
2118   __ mov(rscratch1, _thread_in_native);
2119   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2120   __ stlrw(rscratch1, rscratch2);
2121 
2122   __ rt_call(native_func);
2123 
2124   __ bind(native_return);
2125 
2126   intptr_t return_pc = (intptr_t) __ pc();
2127   oop_maps->add_gc_map(return_pc - start, map);
2128 
2129   // Unpack native results.
2130   switch (ret_type) {
2131   case T_BOOLEAN: __ c2bool(r0);                     break;
2132   case T_CHAR   : __ ubfx(r0, r0, 0, 16);            break;
2133   case T_BYTE   : __ sbfx(r0, r0, 0, 8);             break;
2134   case T_SHORT  : __ sbfx(r0, r0, 0, 16);            break;
2135   case T_INT    : __ sbfx(r0, r0, 0, 32);            break;
2136   case T_DOUBLE :
2137   case T_FLOAT  :
2138     // Result is in v0 we'll save as needed
2139     break;
2140   case T_ARRAY:                 // Really a handle
2141   case T_OBJECT:                // Really a handle
2142       break; // can't de-handlize until after safepoint check
2143   case T_VOID: break;
2144   case T_LONG: break;
2145   default       : ShouldNotReachHere();
2146   }
2147 
2148   Label safepoint_in_progress, safepoint_in_progress_done;
2149   Label after_transition;
2150 
2151   // Switch thread to "native transition" state before reading the synchronization state.
2152   // This additional state is necessary because reading and testing the synchronization
2153   // state is not atomic w.r.t. GC, as this scenario demonstrates:
2154   //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2155   //     VM thread changes sync state to synchronizing and suspends threads for GC.
2156   //     Thread A is resumed to finish this native method, but doesn't block here since it
2157   //     didn't see any synchronization is progress, and escapes.
2158   __ mov(rscratch1, _thread_in_native_trans);
2159 
2160   __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset()));
2161 
2162   // Force this write out before the read below
2163   if (!UseSystemMemoryBarrier) {
2164     __ dmb(Assembler::ISH);
2165   }
2166 
2167   __ verify_sve_vector_length();
2168 
2169   // Check for safepoint operation in progress and/or pending suspend requests.
2170   {
2171     // We need an acquire here to ensure that any subsequent load of the
2172     // global SafepointSynchronize::_state flag is ordered after this load
2173     // of the thread-local polling word.  We don't want this poll to
2174     // return false (i.e. not safepointing) and a later poll of the global
2175     // SafepointSynchronize::_state spuriously to return true.
2176     //
2177     // This is to avoid a race when we're in a native->Java transition
2178     // racing the code which wakes up from a safepoint.
2179 
2180     __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */);
2181     __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset()));
2182     __ cbnzw(rscratch1, safepoint_in_progress);
2183     __ bind(safepoint_in_progress_done);
2184   }
2185 
2186   // change thread state
2187   __ mov(rscratch1, _thread_in_Java);
2188   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2189   __ stlrw(rscratch1, rscratch2);
2190   __ bind(after_transition);
2191 
2192   Label reguard;
2193   Label reguard_done;
2194   __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset()));
2195   __ cmpw(rscratch1, StackOverflow::stack_guard_yellow_reserved_disabled);
2196   __ br(Assembler::EQ, reguard);
2197   __ bind(reguard_done);
2198 
2199   // native result if any is live
2200 
2201   // Unlock
2202   Label unlock_done;
2203   Label slow_path_unlock;
2204   if (method->is_synchronized()) {
2205 
2206     // Get locked oop from the handle we passed to jni
2207     __ ldr(obj_reg, Address(oop_handle_reg, 0));
2208 
2209     Label done, not_recursive;
2210 
2211     if (LockingMode == LM_LEGACY) {
2212       // Simple recursive lock?
2213       __ ldr(rscratch1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2214       __ cbnz(rscratch1, not_recursive);
2215       __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
2216       __ b(done);
2217     }
2218 
2219     __ bind(not_recursive);
2220 
2221     // Must save r0 if if it is live now because cmpxchg must use it
2222     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2223       save_native_result(masm, ret_type, stack_slots);
2224     }
2225 
2226     if (LockingMode == LM_MONITOR) {
2227       __ b(slow_path_unlock);
2228     } else if (LockingMode == LM_LEGACY) {
2229       // get address of the stack lock
2230       __ lea(r0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2231       //  get old displaced header
2232       __ ldr(old_hdr, Address(r0, 0));
2233 
2234       // Atomic swap old header if oop still contains the stack lock
2235       Label count;
2236       __ cmpxchg_obj_header(r0, old_hdr, obj_reg, rscratch1, count, &slow_path_unlock);
2237       __ bind(count);
2238       __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
2239     } else {
2240       assert(LockingMode == LM_LIGHTWEIGHT, "");
2241       __ ldr(old_hdr, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2242       __ tbnz(old_hdr, exact_log2(markWord::monitor_value), slow_path_unlock);
2243       __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
2244       __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
2245     }
2246 
2247     // slow path re-enters here
2248     __ bind(unlock_done);
2249     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2250       restore_native_result(masm, ret_type, stack_slots);
2251     }
2252 
2253     __ bind(done);
2254   }
2255 
2256   Label dtrace_method_exit, dtrace_method_exit_done;
2257   {
2258     uint64_t offset;
2259     __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset);
2260     __ ldrb(rscratch1, Address(rscratch1, offset));
2261     __ cbnzw(rscratch1, dtrace_method_exit);
2262     __ bind(dtrace_method_exit_done);
2263   }
2264 
2265   __ reset_last_Java_frame(false);
2266 
2267   // Unbox oop result, e.g. JNIHandles::resolve result.
2268   if (is_reference_type(ret_type)) {
2269     __ resolve_jobject(r0, r1, r2);
2270   }
2271 
2272   if (CheckJNICalls) {
2273     // clear_pending_jni_exception_check
2274     __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset()));
2275   }
2276 
2277   // reset handle block
2278   __ ldr(r2, Address(rthread, JavaThread::active_handles_offset()));
2279   __ str(zr, Address(r2, JNIHandleBlock::top_offset()));
2280 
2281   __ leave();
2282 
2283   // Any exception pending?
2284   __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2285   __ cbnz(rscratch1, exception_pending);
2286 
2287   // We're done
2288   __ ret(lr);
2289 
2290   // Unexpected paths are out of line and go here
2291 
2292   // forward the exception
2293   __ bind(exception_pending);
2294 
2295   // and forward the exception
2296   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2297 
2298   // Slow path locking & unlocking
2299   if (method->is_synchronized()) {
2300 
2301     __ block_comment("Slow path lock {");
2302     __ bind(slow_path_lock);
2303 
2304     // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2305     // args are (oop obj, BasicLock* lock, JavaThread* thread)
2306 
2307     // protect the args we've loaded
2308     save_args(masm, total_c_args, c_arg, out_regs);
2309 
2310     __ mov(c_rarg0, obj_reg);
2311     __ mov(c_rarg1, lock_reg);
2312     __ mov(c_rarg2, rthread);
2313 
2314     // Not a leaf but we have last_Java_frame setup as we want
2315     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
2316     restore_args(masm, total_c_args, c_arg, out_regs);
2317 
2318 #ifdef ASSERT
2319     { Label L;
2320       __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2321       __ cbz(rscratch1, L);
2322       __ stop("no pending exception allowed on exit from monitorenter");
2323       __ bind(L);
2324     }
2325 #endif
2326     __ b(lock_done);
2327 
2328     __ block_comment("} Slow path lock");
2329 
2330     __ block_comment("Slow path unlock {");
2331     __ bind(slow_path_unlock);
2332 
2333     // If we haven't already saved the native result we must save it now as xmm registers
2334     // are still exposed.
2335 
2336     if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2337       save_native_result(masm, ret_type, stack_slots);
2338     }
2339 
2340     __ mov(c_rarg2, rthread);
2341     __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2342     __ mov(c_rarg0, obj_reg);
2343 
2344     // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2345     // NOTE that obj_reg == r19 currently
2346     __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2347     __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2348 
2349     __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2350 
2351 #ifdef ASSERT
2352     {
2353       Label L;
2354       __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2355       __ cbz(rscratch1, L);
2356       __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2357       __ bind(L);
2358     }
2359 #endif /* ASSERT */
2360 
2361     __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2362 
2363     if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2364       restore_native_result(masm, ret_type, stack_slots);
2365     }
2366     __ b(unlock_done);
2367 
2368     __ block_comment("} Slow path unlock");
2369 
2370   } // synchronized
2371 
2372   // SLOW PATH Reguard the stack if needed
2373 
2374   __ bind(reguard);
2375   save_native_result(masm, ret_type, stack_slots);
2376   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2377   restore_native_result(masm, ret_type, stack_slots);
2378   // and continue
2379   __ b(reguard_done);
2380 
2381   // SLOW PATH safepoint
2382   {
2383     __ block_comment("safepoint {");
2384     __ bind(safepoint_in_progress);
2385 
2386     // Don't use call_VM as it will see a possible pending exception and forward it
2387     // and never return here preventing us from clearing _last_native_pc down below.
2388     //
2389     save_native_result(masm, ret_type, stack_slots);
2390     __ mov(c_rarg0, rthread);
2391 #ifndef PRODUCT
2392   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2393 #endif
2394     __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)));
2395     __ blr(rscratch1);
2396 
2397     // Restore any method result value
2398     restore_native_result(masm, ret_type, stack_slots);
2399 
2400     __ b(safepoint_in_progress_done);
2401     __ block_comment("} safepoint");
2402   }
2403 
2404   // SLOW PATH dtrace support
2405   {
2406     __ block_comment("dtrace entry {");
2407     __ bind(dtrace_method_entry);
2408 
2409     // We have all of the arguments setup at this point. We must not touch any register
2410     // argument registers at this point (what if we save/restore them there are no oop?
2411 
2412     save_args(masm, total_c_args, c_arg, out_regs);
2413     __ mov_metadata(c_rarg1, method());
2414     __ call_VM_leaf(
2415       CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2416       rthread, c_rarg1);
2417     restore_args(masm, total_c_args, c_arg, out_regs);
2418     __ b(dtrace_method_entry_done);
2419     __ block_comment("} dtrace entry");
2420   }
2421 
2422   {
2423     __ block_comment("dtrace exit {");
2424     __ bind(dtrace_method_exit);
2425     save_native_result(masm, ret_type, stack_slots);
2426     __ mov_metadata(c_rarg1, method());
2427     __ call_VM_leaf(
2428          CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2429          rthread, c_rarg1);
2430     restore_native_result(masm, ret_type, stack_slots);
2431     __ b(dtrace_method_exit_done);
2432     __ block_comment("} dtrace exit");
2433   }
2434 
2435 
2436   __ flush();
2437 
2438   nmethod *nm = nmethod::new_native_nmethod(method,
2439                                             compile_id,
2440                                             masm->code(),
2441                                             vep_offset,
2442                                             frame_complete,
2443                                             stack_slots / VMRegImpl::slots_per_word,
2444                                             (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2445                                             in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2446                                             oop_maps);
2447 
2448   return nm;
2449 }
2450 
2451 // this function returns the adjust size (in number of words) to a c2i adapter
2452 // activation for use during deoptimization
2453 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2454   assert(callee_locals >= callee_parameters,
2455           "test and remove; got more parms than locals");
2456   if (callee_locals < callee_parameters)
2457     return 0;                   // No adjustment for negative locals
2458   int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2459   // diff is counted in stack words
2460   return align_up(diff, 2);
2461 }
2462 
2463 
2464 //------------------------------generate_deopt_blob----------------------------
2465 void SharedRuntime::generate_deopt_blob() {
2466   // Allocate space for the code
2467   ResourceMark rm;
2468   // Setup code generation tools
2469   int pad = 0;
2470 #if INCLUDE_JVMCI
2471   if (EnableJVMCI) {
2472     pad += 512; // Increase the buffer size when compiling for JVMCI
2473   }
2474 #endif
2475   CodeBuffer buffer("deopt_blob", 2048+pad, 1024);
2476   MacroAssembler* masm = new MacroAssembler(&buffer);
2477   int frame_size_in_words;
2478   OopMap* map = nullptr;
2479   OopMapSet *oop_maps = new OopMapSet();
2480   RegisterSaver reg_save(COMPILER2_OR_JVMCI != 0);
2481 
2482   // -------------
2483   // This code enters when returning to a de-optimized nmethod.  A return
2484   // address has been pushed on the stack, and return values are in
2485   // registers.
2486   // If we are doing a normal deopt then we were called from the patched
2487   // nmethod from the point we returned to the nmethod. So the return
2488   // address on the stack is wrong by NativeCall::instruction_size
2489   // We will adjust the value so it looks like we have the original return
2490   // address on the stack (like when we eagerly deoptimized).
2491   // In the case of an exception pending when deoptimizing, we enter
2492   // with a return address on the stack that points after the call we patched
2493   // into the exception handler. We have the following register state from,
2494   // e.g., the forward exception stub (see stubGenerator_x86_64.cpp).
2495   //    r0: exception oop
2496   //    r19: exception handler
2497   //    r3: throwing pc
2498   // So in this case we simply jam r3 into the useless return address and
2499   // the stack looks just like we want.
2500   //
2501   // At this point we need to de-opt.  We save the argument return
2502   // registers.  We call the first C routine, fetch_unroll_info().  This
2503   // routine captures the return values and returns a structure which
2504   // describes the current frame size and the sizes of all replacement frames.
2505   // The current frame is compiled code and may contain many inlined
2506   // functions, each with their own JVM state.  We pop the current frame, then
2507   // push all the new frames.  Then we call the C routine unpack_frames() to
2508   // populate these frames.  Finally unpack_frames() returns us the new target
2509   // address.  Notice that callee-save registers are BLOWN here; they have
2510   // already been captured in the vframeArray at the time the return PC was
2511   // patched.
2512   address start = __ pc();
2513   Label cont;
2514 
2515   // Prolog for non exception case!
2516 
2517   // Save everything in sight.
2518   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2519 
2520   // Normal deoptimization.  Save exec mode for unpack_frames.
2521   __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved
2522   __ b(cont);
2523 
2524   int reexecute_offset = __ pc() - start;
2525 #if INCLUDE_JVMCI && !defined(COMPILER1)
2526   if (EnableJVMCI && UseJVMCICompiler) {
2527     // JVMCI does not use this kind of deoptimization
2528     __ should_not_reach_here();
2529   }
2530 #endif
2531 
2532   // Reexecute case
2533   // return address is the pc describes what bci to do re-execute at
2534 
2535   // No need to update map as each call to save_live_registers will produce identical oopmap
2536   (void) reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2537 
2538   __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved
2539   __ b(cont);
2540 
2541 #if INCLUDE_JVMCI
2542   Label after_fetch_unroll_info_call;
2543   int implicit_exception_uncommon_trap_offset = 0;
2544   int uncommon_trap_offset = 0;
2545 
2546   if (EnableJVMCI) {
2547     implicit_exception_uncommon_trap_offset = __ pc() - start;
2548 
2549     __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2550     __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2551 
2552     uncommon_trap_offset = __ pc() - start;
2553 
2554     // Save everything in sight.
2555     reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2556     // fetch_unroll_info needs to call last_java_frame()
2557     Label retaddr;
2558     __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2559 
2560     __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2561     __ movw(rscratch1, -1);
2562     __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2563 
2564     __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute);
2565     __ mov(c_rarg0, rthread);
2566     __ movw(c_rarg2, rcpool); // exec mode
2567     __ lea(rscratch1,
2568            RuntimeAddress(CAST_FROM_FN_PTR(address,
2569                                            Deoptimization::uncommon_trap)));
2570     __ blr(rscratch1);
2571     __ bind(retaddr);
2572     oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2573 
2574     __ reset_last_Java_frame(false);
2575 
2576     __ b(after_fetch_unroll_info_call);
2577   } // EnableJVMCI
2578 #endif // INCLUDE_JVMCI
2579 
2580   int exception_offset = __ pc() - start;
2581 
2582   // Prolog for exception case
2583 
2584   // all registers are dead at this entry point, except for r0, and
2585   // r3 which contain the exception oop and exception pc
2586   // respectively.  Set them in TLS and fall thru to the
2587   // unpack_with_exception_in_tls entry point.
2588 
2589   __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
2590   __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
2591 
2592   int exception_in_tls_offset = __ pc() - start;
2593 
2594   // new implementation because exception oop is now passed in JavaThread
2595 
2596   // Prolog for exception case
2597   // All registers must be preserved because they might be used by LinearScan
2598   // Exceptiop oop and throwing PC are passed in JavaThread
2599   // tos: stack at point of call to method that threw the exception (i.e. only
2600   // args are on the stack, no return address)
2601 
2602   // The return address pushed by save_live_registers will be patched
2603   // later with the throwing pc. The correct value is not available
2604   // now because loading it from memory would destroy registers.
2605 
2606   // NB: The SP at this point must be the SP of the method that is
2607   // being deoptimized.  Deoptimization assumes that the frame created
2608   // here by save_live_registers is immediately below the method's SP.
2609   // This is a somewhat fragile mechanism.
2610 
2611   // Save everything in sight.
2612   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2613 
2614   // Now it is safe to overwrite any register
2615 
2616   // Deopt during an exception.  Save exec mode for unpack_frames.
2617   __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved
2618 
2619   // load throwing pc from JavaThread and patch it as the return address
2620   // of the current frame. Then clear the field in JavaThread
2621   __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2622   __ protect_return_address(r3);
2623   __ str(r3, Address(rfp, wordSize));
2624   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2625 
2626 #ifdef ASSERT
2627   // verify that there is really an exception oop in JavaThread
2628   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2629   __ verify_oop(r0);
2630 
2631   // verify that there is no pending exception
2632   Label no_pending_exception;
2633   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2634   __ cbz(rscratch1, no_pending_exception);
2635   __ stop("must not have pending exception here");
2636   __ bind(no_pending_exception);
2637 #endif
2638 
2639   __ bind(cont);
2640 
2641   // Call C code.  Need thread and this frame, but NOT official VM entry
2642   // crud.  We cannot block on this call, no GC can happen.
2643   //
2644   // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2645 
2646   // fetch_unroll_info needs to call last_java_frame().
2647 
2648   Label retaddr;
2649   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2650 #ifdef ASSERT
2651   { Label L;
2652     __ ldr(rscratch1, Address(rthread, JavaThread::last_Java_fp_offset()));
2653     __ cbz(rscratch1, L);
2654     __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2655     __ bind(L);
2656   }
2657 #endif // ASSERT
2658   __ mov(c_rarg0, rthread);
2659   __ mov(c_rarg1, rcpool);
2660   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2661   __ blr(rscratch1);
2662   __ bind(retaddr);
2663 
2664   // Need to have an oopmap that tells fetch_unroll_info where to
2665   // find any register it might need.
2666   oop_maps->add_gc_map(__ pc() - start, map);
2667 
2668   __ reset_last_Java_frame(false);
2669 
2670 #if INCLUDE_JVMCI
2671   if (EnableJVMCI) {
2672     __ bind(after_fetch_unroll_info_call);
2673   }
2674 #endif
2675 
2676   // Load UnrollBlock* into r5
2677   __ mov(r5, r0);
2678 
2679   __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset()));
2680    Label noException;
2681   __ cmpw(rcpool, Deoptimization::Unpack_exception);   // Was exception pending?
2682   __ br(Assembler::NE, noException);
2683   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2684   // QQQ this is useless it was null above
2685   __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2686   __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
2687   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2688 
2689   __ verify_oop(r0);
2690 
2691   // Overwrite the result registers with the exception results.
2692   __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2693   // I think this is useless
2694   // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2695 
2696   __ bind(noException);
2697 
2698   // Only register save data is on the stack.
2699   // Now restore the result registers.  Everything else is either dead
2700   // or captured in the vframeArray.
2701 
2702   // Restore fp result register
2703   __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2704   // Restore integer result register
2705   __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2706 
2707   // Pop all of the register save area off the stack
2708   __ add(sp, sp, frame_size_in_words * wordSize);
2709 
2710   // All of the register save area has been popped of the stack. Only the
2711   // return address remains.
2712 
2713   // Pop all the frames we must move/replace.
2714   //
2715   // Frame picture (youngest to oldest)
2716   // 1: self-frame (no frame link)
2717   // 2: deopting frame  (no frame link)
2718   // 3: caller of deopting frame (could be compiled/interpreted).
2719   //
2720   // Note: by leaving the return address of self-frame on the stack
2721   // and using the size of frame 2 to adjust the stack
2722   // when we are done the return to frame 3 will still be on the stack.
2723 
2724   // Pop deoptimized frame
2725   __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2726   __ sub(r2, r2, 2 * wordSize);
2727   __ add(sp, sp, r2);
2728   __ ldp(rfp, zr, __ post(sp, 2 * wordSize));
2729 
2730 #ifdef ASSERT
2731   // Compilers generate code that bang the stack by as much as the
2732   // interpreter would need. So this stack banging should never
2733   // trigger a fault. Verify that it does not on non product builds.
2734   __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2735   __ bang_stack_size(r19, r2);
2736 #endif
2737   // Load address of array of frame pcs into r2
2738   __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset()));
2739 
2740   // Trash the old pc
2741   // __ addptr(sp, wordSize);  FIXME ????
2742 
2743   // Load address of array of frame sizes into r4
2744   __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset()));
2745 
2746   // Load counter into r3
2747   __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset()));
2748 
2749   // Now adjust the caller's stack to make up for the extra locals
2750   // but record the original sp so that we can save it in the skeletal interpreter
2751   // frame and the stack walking of interpreter_sender will get the unextended sp
2752   // value and not the "real" sp value.
2753 
2754   const Register sender_sp = r6;
2755 
2756   __ mov(sender_sp, sp);
2757   __ ldrw(r19, Address(r5,
2758                        Deoptimization::UnrollBlock::
2759                        caller_adjustment_offset()));
2760   __ sub(sp, sp, r19);
2761 
2762   // Push interpreter frames in a loop
2763   __ mov(rscratch1, (uint64_t)0xDEADDEAD);        // Make a recognizable pattern
2764   __ mov(rscratch2, rscratch1);
2765   Label loop;
2766   __ bind(loop);
2767   __ ldr(r19, Address(__ post(r4, wordSize)));          // Load frame size
2768   __ sub(r19, r19, 2*wordSize);           // We'll push pc and fp by hand
2769   __ ldr(lr, Address(__ post(r2, wordSize)));  // Load pc
2770   __ enter();                           // Save old & set new fp
2771   __ sub(sp, sp, r19);                  // Prolog
2772   // This value is corrected by layout_activation_impl
2773   __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2774   __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2775   __ mov(sender_sp, sp);               // Pass sender_sp to next frame
2776   __ sub(r3, r3, 1);                   // Decrement counter
2777   __ cbnz(r3, loop);
2778 
2779     // Re-push self-frame
2780   __ ldr(lr, Address(r2));
2781   __ enter();
2782 
2783   // Allocate a full sized register save area.  We subtract 2 because
2784   // enter() just pushed 2 words
2785   __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2786 
2787   // Restore frame locals after moving the frame
2788   __ strd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2789   __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2790 
2791   // Call C code.  Need thread but NOT official VM entry
2792   // crud.  We cannot block on this call, no GC can happen.  Call should
2793   // restore return values to their stack-slots with the new SP.
2794   //
2795   // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2796 
2797   // Use rfp because the frames look interpreted now
2798   // Don't need the precise return PC here, just precise enough to point into this code blob.
2799   address the_pc = __ pc();
2800   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
2801 
2802   __ mov(c_rarg0, rthread);
2803   __ movw(c_rarg1, rcpool); // second arg: exec_mode
2804   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2805   __ blr(rscratch1);
2806 
2807   // Set an oopmap for the call site
2808   // Use the same PC we used for the last java frame
2809   oop_maps->add_gc_map(the_pc - start,
2810                        new OopMap( frame_size_in_words, 0 ));
2811 
2812   // Clear fp AND pc
2813   __ reset_last_Java_frame(true);
2814 
2815   // Collect return values
2816   __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2817   __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2818   // I think this is useless (throwing pc?)
2819   // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2820 
2821   // Pop self-frame.
2822   __ leave();                           // Epilog
2823 
2824   // Jump to interpreter
2825   __ ret(lr);
2826 
2827   // Make sure all code is generated
2828   masm->flush();
2829 
2830   _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2831   _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2832 #if INCLUDE_JVMCI
2833   if (EnableJVMCI) {
2834     _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2835     _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2836   }
2837 #endif
2838 }
2839 
2840 // Number of stack slots between incoming argument block and the start of
2841 // a new frame.  The PROLOG must add this many slots to the stack.  The
2842 // EPILOG must remove this many slots. aarch64 needs two slots for
2843 // return address and fp.
2844 // TODO think this is correct but check
2845 uint SharedRuntime::in_preserve_stack_slots() {
2846   return 4;
2847 }
2848 
2849 uint SharedRuntime::out_preserve_stack_slots() {
2850   return 0;
2851 }
2852 
2853 #ifdef COMPILER2
2854 //------------------------------generate_uncommon_trap_blob--------------------
2855 void SharedRuntime::generate_uncommon_trap_blob() {
2856   // Allocate space for the code
2857   ResourceMark rm;
2858   // Setup code generation tools
2859   CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
2860   MacroAssembler* masm = new MacroAssembler(&buffer);
2861 
2862   assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned");
2863 
2864   address start = __ pc();
2865 
2866   // Push self-frame.  We get here with a return address in LR
2867   // and sp should be 16 byte aligned
2868   // push rfp and retaddr by hand
2869   __ protect_return_address();
2870   __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize)));
2871   // we don't expect an arg reg save area
2872 #ifndef PRODUCT
2873   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2874 #endif
2875   // compiler left unloaded_class_index in j_rarg0 move to where the
2876   // runtime expects it.
2877   if (c_rarg1 != j_rarg0) {
2878     __ movw(c_rarg1, j_rarg0);
2879   }
2880 
2881   // we need to set the past SP to the stack pointer of the stub frame
2882   // and the pc to the address where this runtime call will return
2883   // although actually any pc in this code blob will do).
2884   Label retaddr;
2885   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2886 
2887   // Call C code.  Need thread but NOT official VM entry
2888   // crud.  We cannot block on this call, no GC can happen.  Call should
2889   // capture callee-saved registers as well as return values.
2890   // Thread is in rdi already.
2891   //
2892   // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index);
2893   //
2894   // n.b. 2 gp args, 0 fp args, integral return type
2895 
2896   __ mov(c_rarg0, rthread);
2897   __ movw(c_rarg2, (unsigned)Deoptimization::Unpack_uncommon_trap);
2898   __ lea(rscratch1,
2899          RuntimeAddress(CAST_FROM_FN_PTR(address,
2900                                          Deoptimization::uncommon_trap)));
2901   __ blr(rscratch1);
2902   __ bind(retaddr);
2903 
2904   // Set an oopmap for the call site
2905   OopMapSet* oop_maps = new OopMapSet();
2906   OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0);
2907 
2908   // location of rfp is known implicitly by the frame sender code
2909 
2910   oop_maps->add_gc_map(__ pc() - start, map);
2911 
2912   __ reset_last_Java_frame(false);
2913 
2914   // move UnrollBlock* into r4
2915   __ mov(r4, r0);
2916 
2917 #ifdef ASSERT
2918   { Label L;
2919     __ ldrw(rscratch1, Address(r4, Deoptimization::UnrollBlock::unpack_kind_offset()));
2920     __ cmpw(rscratch1, (unsigned)Deoptimization::Unpack_uncommon_trap);
2921     __ br(Assembler::EQ, L);
2922     __ stop("SharedRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
2923     __ bind(L);
2924   }
2925 #endif
2926 
2927   // Pop all the frames we must move/replace.
2928   //
2929   // Frame picture (youngest to oldest)
2930   // 1: self-frame (no frame link)
2931   // 2: deopting frame  (no frame link)
2932   // 3: caller of deopting frame (could be compiled/interpreted).
2933 
2934   // Pop self-frame.  We have no frame, and must rely only on r0 and sp.
2935   __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog!
2936 
2937   // Pop deoptimized frame (int)
2938   __ ldrw(r2, Address(r4,
2939                       Deoptimization::UnrollBlock::
2940                       size_of_deoptimized_frame_offset()));
2941   __ sub(r2, r2, 2 * wordSize);
2942   __ add(sp, sp, r2);
2943   __ ldp(rfp, zr, __ post(sp, 2 * wordSize));
2944 
2945 #ifdef ASSERT
2946   // Compilers generate code that bang the stack by as much as the
2947   // interpreter would need. So this stack banging should never
2948   // trigger a fault. Verify that it does not on non product builds.
2949   __ ldrw(r1, Address(r4,
2950                       Deoptimization::UnrollBlock::
2951                       total_frame_sizes_offset()));
2952   __ bang_stack_size(r1, r2);
2953 #endif
2954 
2955   // Load address of array of frame pcs into r2 (address*)
2956   __ ldr(r2, Address(r4,
2957                      Deoptimization::UnrollBlock::frame_pcs_offset()));
2958 
2959   // Load address of array of frame sizes into r5 (intptr_t*)
2960   __ ldr(r5, Address(r4,
2961                      Deoptimization::UnrollBlock::
2962                      frame_sizes_offset()));
2963 
2964   // Counter
2965   __ ldrw(r3, Address(r4,
2966                       Deoptimization::UnrollBlock::
2967                       number_of_frames_offset())); // (int)
2968 
2969   // Now adjust the caller's stack to make up for the extra locals but
2970   // record the original sp so that we can save it in the skeletal
2971   // interpreter frame and the stack walking of interpreter_sender
2972   // will get the unextended sp value and not the "real" sp value.
2973 
2974   const Register sender_sp = r8;
2975 
2976   __ mov(sender_sp, sp);
2977   __ ldrw(r1, Address(r4,
2978                       Deoptimization::UnrollBlock::
2979                       caller_adjustment_offset())); // (int)
2980   __ sub(sp, sp, r1);
2981 
2982   // Push interpreter frames in a loop
2983   Label loop;
2984   __ bind(loop);
2985   __ ldr(r1, Address(r5, 0));       // Load frame size
2986   __ sub(r1, r1, 2 * wordSize);     // We'll push pc and rfp by hand
2987   __ ldr(lr, Address(r2, 0));       // Save return address
2988   __ enter();                       // and old rfp & set new rfp
2989   __ sub(sp, sp, r1);               // Prolog
2990   __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2991   // This value is corrected by layout_activation_impl
2992   __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2993   __ mov(sender_sp, sp);          // Pass sender_sp to next frame
2994   __ add(r5, r5, wordSize);       // Bump array pointer (sizes)
2995   __ add(r2, r2, wordSize);       // Bump array pointer (pcs)
2996   __ subsw(r3, r3, 1);            // Decrement counter
2997   __ br(Assembler::GT, loop);
2998   __ ldr(lr, Address(r2, 0));     // save final return address
2999   // Re-push self-frame
3000   __ enter();                     // & old rfp & set new rfp
3001 
3002   // Use rfp because the frames look interpreted now
3003   // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP.
3004   // Don't need the precise return PC here, just precise enough to point into this code blob.
3005   address the_pc = __ pc();
3006   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3007 
3008   // Call C code.  Need thread but NOT official VM entry
3009   // crud.  We cannot block on this call, no GC can happen.  Call should
3010   // restore return values to their stack-slots with the new SP.
3011   // Thread is in rdi already.
3012   //
3013   // BasicType unpack_frames(JavaThread* thread, int exec_mode);
3014   //
3015   // n.b. 2 gp args, 0 fp args, integral return type
3016 
3017   // sp should already be aligned
3018   __ mov(c_rarg0, rthread);
3019   __ movw(c_rarg1, (unsigned)Deoptimization::Unpack_uncommon_trap);
3020   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
3021   __ blr(rscratch1);
3022 
3023   // Set an oopmap for the call site
3024   // Use the same PC we used for the last java frame
3025   oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0));
3026 
3027   // Clear fp AND pc
3028   __ reset_last_Java_frame(true);
3029 
3030   // Pop self-frame.
3031   __ leave();                 // Epilog
3032 
3033   // Jump to interpreter
3034   __ ret(lr);
3035 
3036   // Make sure all code is generated
3037   masm->flush();
3038 
3039   _uncommon_trap_blob =  UncommonTrapBlob::create(&buffer, oop_maps,
3040                                                  SimpleRuntimeFrame::framesize >> 1);
3041 }
3042 #endif // COMPILER2
3043 
3044 
3045 //------------------------------generate_handler_blob------
3046 //
3047 // Generate a special Compile2Runtime blob that saves all registers,
3048 // and setup oopmap.
3049 //
3050 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3051   ResourceMark rm;
3052   OopMapSet *oop_maps = new OopMapSet();
3053   OopMap* map;
3054 
3055   // Allocate space for the code.  Setup code generation tools.
3056   CodeBuffer buffer("handler_blob", 2048, 1024);
3057   MacroAssembler* masm = new MacroAssembler(&buffer);
3058 
3059   address start   = __ pc();
3060   address call_pc = nullptr;
3061   int frame_size_in_words;
3062   bool cause_return = (poll_type == POLL_AT_RETURN);
3063   RegisterSaver reg_save(poll_type == POLL_AT_VECTOR_LOOP /* save_vectors */);
3064 
3065   // When the signal occurred, the LR was either signed and stored on the stack (in which
3066   // case it will be restored from the stack before being used) or unsigned and not stored
3067   // on the stack. Stipping ensures we get the right value.
3068   __ strip_return_address();
3069 
3070   // Save Integer and Float registers.
3071   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
3072 
3073   // The following is basically a call_VM.  However, we need the precise
3074   // address of the call in order to generate an oopmap. Hence, we do all the
3075   // work ourselves.
3076 
3077   Label retaddr;
3078   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
3079 
3080   // The return address must always be correct so that frame constructor never
3081   // sees an invalid pc.
3082 
3083   if (!cause_return) {
3084     // overwrite the return address pushed by save_live_registers
3085     // Additionally, r20 is a callee-saved register so we can look at
3086     // it later to determine if someone changed the return address for
3087     // us!
3088     __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset()));
3089     __ protect_return_address(r20);
3090     __ str(r20, Address(rfp, wordSize));
3091   }
3092 
3093   // Do the call
3094   __ mov(c_rarg0, rthread);
3095   __ lea(rscratch1, RuntimeAddress(call_ptr));
3096   __ blr(rscratch1);
3097   __ bind(retaddr);
3098 
3099   // Set an oopmap for the call site.  This oopmap will map all
3100   // oop-registers and debug-info registers as callee-saved.  This
3101   // will allow deoptimization at this safepoint to find all possible
3102   // debug-info recordings, as well as let GC find all oops.
3103 
3104   oop_maps->add_gc_map( __ pc() - start, map);
3105 
3106   Label noException;
3107 
3108   __ reset_last_Java_frame(false);
3109 
3110   __ membar(Assembler::LoadLoad | Assembler::LoadStore);
3111 
3112   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
3113   __ cbz(rscratch1, noException);
3114 
3115   // Exception pending
3116 
3117   reg_save.restore_live_registers(masm);
3118 
3119   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3120 
3121   // No exception case
3122   __ bind(noException);
3123 
3124   Label no_adjust, bail;
3125   if (!cause_return) {
3126     // If our stashed return pc was modified by the runtime we avoid touching it
3127     __ ldr(rscratch1, Address(rfp, wordSize));
3128     __ cmp(r20, rscratch1);
3129     __ br(Assembler::NE, no_adjust);
3130     __ authenticate_return_address(r20);
3131 
3132 #ifdef ASSERT
3133     // Verify the correct encoding of the poll we're about to skip.
3134     // See NativeInstruction::is_ldrw_to_zr()
3135     __ ldrw(rscratch1, Address(r20));
3136     __ ubfx(rscratch2, rscratch1, 22, 10);
3137     __ cmpw(rscratch2, 0b1011100101);
3138     __ br(Assembler::NE, bail);
3139     __ ubfx(rscratch2, rscratch1, 0, 5);
3140     __ cmpw(rscratch2, 0b11111);
3141     __ br(Assembler::NE, bail);
3142 #endif
3143     // Adjust return pc forward to step over the safepoint poll instruction
3144     __ add(r20, r20, NativeInstruction::instruction_size);
3145     __ protect_return_address(r20);
3146     __ str(r20, Address(rfp, wordSize));
3147   }
3148 
3149   __ bind(no_adjust);
3150   // Normal exit, restore registers and exit.
3151   reg_save.restore_live_registers(masm);
3152 
3153   __ ret(lr);
3154 
3155 #ifdef ASSERT
3156   __ bind(bail);
3157   __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
3158 #endif
3159 
3160   // Make sure all code is generated
3161   masm->flush();
3162 
3163   // Fill-out other meta info
3164   return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
3165 }
3166 
3167 //
3168 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3169 //
3170 // Generate a stub that calls into vm to find out the proper destination
3171 // of a java call. All the argument registers are live at this point
3172 // but since this is generic code we don't know what they are and the caller
3173 // must do any gc of the args.
3174 //
3175 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3176   assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
3177 
3178   // allocate space for the code
3179   ResourceMark rm;
3180 
3181   CodeBuffer buffer(name, 1000, 512);
3182   MacroAssembler* masm                = new MacroAssembler(&buffer);
3183 
3184   int frame_size_in_words;
3185   RegisterSaver reg_save(false /* save_vectors */);
3186 
3187   OopMapSet *oop_maps = new OopMapSet();
3188   OopMap* map = nullptr;
3189 
3190   int start = __ offset();
3191 
3192   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
3193 
3194   int frame_complete = __ offset();
3195 
3196   {
3197     Label retaddr;
3198     __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
3199 
3200     __ mov(c_rarg0, rthread);
3201     __ lea(rscratch1, RuntimeAddress(destination));
3202 
3203     __ blr(rscratch1);
3204     __ bind(retaddr);
3205   }
3206 
3207   // Set an oopmap for the call site.
3208   // We need this not only for callee-saved registers, but also for volatile
3209   // registers that the compiler might be keeping live across a safepoint.
3210 
3211   oop_maps->add_gc_map( __ offset() - start, map);
3212 
3213   // r0 contains the address we are going to jump to assuming no exception got installed
3214 
3215   // clear last_Java_sp
3216   __ reset_last_Java_frame(false);
3217   // check for pending exceptions
3218   Label pending;
3219   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
3220   __ cbnz(rscratch1, pending);
3221 
3222   // get the returned Method*
3223   __ get_vm_result_2(rmethod, rthread);
3224   __ str(rmethod, Address(sp, reg_save.reg_offset_in_bytes(rmethod)));
3225 
3226   // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch
3227   __ str(r0, Address(sp, reg_save.rscratch1_offset_in_bytes()));
3228   reg_save.restore_live_registers(masm);
3229 
3230   // We are back to the original state on entry and ready to go.
3231 
3232   __ br(rscratch1);
3233 
3234   // Pending exception after the safepoint
3235 
3236   __ bind(pending);
3237 
3238   reg_save.restore_live_registers(masm);
3239 
3240   // exception pending => remove activation and forward to exception handler
3241 
3242   __ str(zr, Address(rthread, JavaThread::vm_result_offset()));
3243 
3244   __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
3245   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3246 
3247   // -------------
3248   // make sure all code is generated
3249   masm->flush();
3250 
3251   // return the  blob
3252   // frame_size_words or bytes??
3253   return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
3254 }
3255 
3256 #ifdef COMPILER2
3257 // This is here instead of runtime_aarch64_64.cpp because it uses SimpleRuntimeFrame
3258 //
3259 //------------------------------generate_exception_blob---------------------------
3260 // creates exception blob at the end
3261 // Using exception blob, this code is jumped from a compiled method.
3262 // (see emit_exception_handler in x86_64.ad file)
3263 //
3264 // Given an exception pc at a call we call into the runtime for the
3265 // handler in this method. This handler might merely restore state
3266 // (i.e. callee save registers) unwind the frame and jump to the
3267 // exception handler for the nmethod if there is no Java level handler
3268 // for the nmethod.
3269 //
3270 // This code is entered with a jmp.
3271 //
3272 // Arguments:
3273 //   r0: exception oop
3274 //   r3: exception pc
3275 //
3276 // Results:
3277 //   r0: exception oop
3278 //   r3: exception pc in caller or ???
3279 //   destination: exception handler of caller
3280 //
3281 // Note: the exception pc MUST be at a call (precise debug information)
3282 //       Registers r0, r3, r2, r4, r5, r8-r11 are not callee saved.
3283 //
3284 
3285 void OptoRuntime::generate_exception_blob() {
3286   assert(!OptoRuntime::is_callee_saved_register(R3_num), "");
3287   assert(!OptoRuntime::is_callee_saved_register(R0_num), "");
3288   assert(!OptoRuntime::is_callee_saved_register(R2_num), "");
3289 
3290   assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned");
3291 
3292   // Allocate space for the code
3293   ResourceMark rm;
3294   // Setup code generation tools
3295   CodeBuffer buffer("exception_blob", 2048, 1024);
3296   MacroAssembler* masm = new MacroAssembler(&buffer);
3297 
3298   // TODO check various assumptions made here
3299   //
3300   // make sure we do so before running this
3301 
3302   address start = __ pc();
3303 
3304   // push rfp and retaddr by hand
3305   // Exception pc is 'return address' for stack walker
3306   __ protect_return_address();
3307   __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize)));
3308   // there are no callee save registers and we don't expect an
3309   // arg reg save area
3310 #ifndef PRODUCT
3311   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
3312 #endif
3313   // Store exception in Thread object. We cannot pass any arguments to the
3314   // handle_exception call, since we do not want to make any assumption
3315   // about the size of the frame where the exception happened in.
3316   __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
3317   __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
3318 
3319   // This call does all the hard work.  It checks if an exception handler
3320   // exists in the method.
3321   // If so, it returns the handler address.
3322   // If not, it prepares for stack-unwinding, restoring the callee-save
3323   // registers of the frame being removed.
3324   //
3325   // address OptoRuntime::handle_exception_C(JavaThread* thread)
3326   //
3327   // n.b. 1 gp arg, 0 fp args, integral return type
3328 
3329   // the stack should always be aligned
3330   address the_pc = __ pc();
3331   __ set_last_Java_frame(sp, noreg, the_pc, rscratch1);
3332   __ mov(c_rarg0, rthread);
3333   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C)));
3334   __ blr(rscratch1);
3335   // handle_exception_C is a special VM call which does not require an explicit
3336   // instruction sync afterwards.
3337 
3338   // May jump to SVE compiled code
3339   __ reinitialize_ptrue();
3340 
3341   // Set an oopmap for the call site.  This oopmap will only be used if we
3342   // are unwinding the stack.  Hence, all locations will be dead.
3343   // Callee-saved registers will be the same as the frame above (i.e.,
3344   // handle_exception_stub), since they were restored when we got the
3345   // exception.
3346 
3347   OopMapSet* oop_maps = new OopMapSet();
3348 
3349   oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0));
3350 
3351   __ reset_last_Java_frame(false);
3352 
3353   // Restore callee-saved registers
3354 
3355   // rfp is an implicitly saved callee saved register (i.e. the calling
3356   // convention will save restore it in prolog/epilog) Other than that
3357   // there are no callee save registers now that adapter frames are gone.
3358   // and we dont' expect an arg reg save area
3359   __ ldp(rfp, r3, Address(__ post(sp, 2 * wordSize)));
3360   __ authenticate_return_address(r3);
3361 
3362   // r0: exception handler
3363 
3364   // We have a handler in r0 (could be deopt blob).
3365   __ mov(r8, r0);
3366 
3367   // Get the exception oop
3368   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
3369   // Get the exception pc in case we are deoptimized
3370   __ ldr(r4, Address(rthread, JavaThread::exception_pc_offset()));
3371 #ifdef ASSERT
3372   __ str(zr, Address(rthread, JavaThread::exception_handler_pc_offset()));
3373   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
3374 #endif
3375   // Clear the exception oop so GC no longer processes it as a root.
3376   __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
3377 
3378   // r0: exception oop
3379   // r8:  exception handler
3380   // r4: exception pc
3381   // Jump to handler
3382 
3383   __ br(r8);
3384 
3385   // Make sure all code is generated
3386   masm->flush();
3387 
3388   // Set exception blob
3389   _exception_blob =  ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1);
3390 }
3391 
3392 #endif // COMPILER2
3393 
3394 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
3395   BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
3396   CodeBuffer buffer(buf);
3397   short buffer_locs[20];
3398   buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
3399                                          sizeof(buffer_locs)/sizeof(relocInfo));
3400 
3401   MacroAssembler _masm(&buffer);
3402   MacroAssembler* masm = &_masm;
3403 
3404   const Array<SigEntry>* sig_vk = vk->extended_sig();
3405   const Array<VMRegPair>* regs = vk->return_regs();
3406 
3407   int pack_fields_jobject_off = __ offset();
3408   // Resolve pre-allocated buffer from JNI handle.
3409   // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
3410   Register Rresult = r14;  // See StubGenerator::generate_call_stub().
3411   __ ldr(r0, Address(Rresult));
3412   __ resolve_jobject(r0 /* value */,
3413                      rthread /* thread */,
3414                      r12 /* tmp */);
3415   __ str(r0, Address(Rresult));
3416 
3417   int pack_fields_off = __ offset();
3418 
3419   int j = 1;
3420   for (int i = 0; i < sig_vk->length(); i++) {
3421     BasicType bt = sig_vk->at(i)._bt;
3422     if (bt == T_METADATA) {
3423       continue;
3424     }
3425     if (bt == T_VOID) {
3426       if (sig_vk->at(i-1)._bt == T_LONG ||
3427           sig_vk->at(i-1)._bt == T_DOUBLE) {
3428         j++;
3429       }
3430       continue;
3431     }
3432     int off = sig_vk->at(i)._offset;
3433     VMRegPair pair = regs->at(j);
3434     VMReg r_1 = pair.first();
3435     VMReg r_2 = pair.second();
3436     Address to(r0, off);
3437     if (bt == T_FLOAT) {
3438       __ strs(r_1->as_FloatRegister(), to);
3439     } else if (bt == T_DOUBLE) {
3440       __ strd(r_1->as_FloatRegister(), to);
3441     } else {
3442       Register val = r_1->as_Register();
3443       assert_different_registers(to.base(), val, r15, r16, r17);
3444       if (is_reference_type(bt)) {
3445         __ store_heap_oop(to, val, r15, r16, r17, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
3446       } else {
3447         __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
3448       }
3449     }
3450     j++;
3451   }
3452   assert(j == regs->length(), "missed a field?");
3453 
3454   __ ret(lr);
3455 
3456   int unpack_fields_off = __ offset();
3457 
3458   Label skip;
3459   __ cbz(r0, skip);
3460 
3461   j = 1;
3462   for (int i = 0; i < sig_vk->length(); i++) {
3463     BasicType bt = sig_vk->at(i)._bt;
3464     if (bt == T_METADATA) {
3465       continue;
3466     }
3467     if (bt == T_VOID) {
3468       if (sig_vk->at(i-1)._bt == T_LONG ||
3469           sig_vk->at(i-1)._bt == T_DOUBLE) {
3470         j++;
3471       }
3472       continue;
3473     }
3474     int off = sig_vk->at(i)._offset;
3475     assert(off > 0, "offset in object should be positive");
3476     VMRegPair pair = regs->at(j);
3477     VMReg r_1 = pair.first();
3478     VMReg r_2 = pair.second();
3479     Address from(r0, off);
3480     if (bt == T_FLOAT) {
3481       __ ldrs(r_1->as_FloatRegister(), from);
3482     } else if (bt == T_DOUBLE) {
3483       __ ldrd(r_1->as_FloatRegister(), from);
3484     } else if (bt == T_OBJECT || bt == T_ARRAY) {
3485       assert_different_registers(r0, r_1->as_Register());
3486       __ load_heap_oop(r_1->as_Register(), from, rscratch1, rscratch2);
3487     } else {
3488       assert(is_java_primitive(bt), "unexpected basic type");
3489       assert_different_registers(r0, r_1->as_Register());
3490 
3491       size_t size_in_bytes = type2aelembytes(bt);
3492       __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3493     }
3494     j++;
3495   }
3496   assert(j == regs->length(), "missed a field?");
3497 
3498   __ bind(skip);
3499 
3500   __ ret(lr);
3501 
3502   __ flush();
3503 
3504   return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3505 }