1 /*
   2  * Copyright (c) 2003, 2026, 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 "asm/macroAssembler.hpp"
  28 #include "asm/macroAssembler.inline.hpp"
  29 #include "classfile/symbolTable.hpp"
  30 #include "code/aotCodeCache.hpp"
  31 #include "code/codeCache.hpp"
  32 #include "code/compiledIC.hpp"
  33 #include "code/debugInfoRec.hpp"
  34 #include "code/vtableStubs.hpp"
  35 #include "compiler/oopMap.hpp"
  36 #include "gc/shared/barrierSetAssembler.hpp"
  37 #include "interpreter/interpreter.hpp"
  38 #include "interpreter/interp_masm.hpp"
  39 #include "logging/log.hpp"
  40 #include "memory/resourceArea.hpp"
  41 #include "nativeInst_aarch64.hpp"
  42 #include "oops/klass.inline.hpp"
  43 #include "oops/method.inline.hpp"
  44 #include "prims/methodHandles.hpp"
  45 #include "runtime/continuation.hpp"
  46 #include "runtime/continuationEntry.inline.hpp"
  47 #include "runtime/globals.hpp"
  48 #include "runtime/jniHandles.hpp"
  49 #include "runtime/safepointMechanism.hpp"
  50 #include "runtime/sharedRuntime.hpp"
  51 #include "runtime/signature.hpp"
  52 #include "runtime/stubRoutines.hpp"
  53 #include "runtime/timerTrace.hpp"
  54 #include "runtime/vframeArray.hpp"
  55 #include "utilities/align.hpp"
  56 #include "utilities/formatBuffer.hpp"
  57 #include "vmreg_aarch64.inline.hpp"
  58 #ifdef COMPILER1
  59 #include "c1/c1_Runtime1.hpp"
  60 #endif
  61 #ifdef COMPILER2
  62 #include "adfiles/ad_aarch64.hpp"
  63 #include "opto/runtime.hpp"
  64 #endif
  65 
  66 #define __ masm->
  67 
  68 #ifdef PRODUCT
  69 #define BLOCK_COMMENT(str) /* nothing */
  70 #else
  71 #define BLOCK_COMMENT(str) __ block_comment(str)
  72 #endif
  73 
  74 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
  75 
  76 // FIXME -- this is used by C1
  77 class RegisterSaver {
  78   const bool _save_vectors;
  79  public:
  80   RegisterSaver(bool save_vectors) : _save_vectors(save_vectors) {}
  81 
  82   OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
  83   void restore_live_registers(MacroAssembler* masm);
  84 
  85   // Offsets into the register save area
  86   // Used by deoptimization when it is managing result register
  87   // values on its own
  88 
  89   int reg_offset_in_bytes(Register r);
  90   int r0_offset_in_bytes()    { return reg_offset_in_bytes(r0); }
  91   int rscratch1_offset_in_bytes()    { return reg_offset_in_bytes(rscratch1); }
  92   int v0_offset_in_bytes();
  93 
  94   // Total stack size in bytes for saving sve predicate registers.
  95   int total_sve_predicate_in_bytes();
  96 
  97   // Capture info about frame layout
  98   // Note this is only correct when not saving full vectors.
  99   enum layout {
 100                 fpu_state_off = 0,
 101                 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1,
 102                 // The frame sender code expects that rfp will be in
 103                 // the "natural" place and will override any oopMap
 104                 // setting for it. We must therefore force the layout
 105                 // so that it agrees with the frame sender code.
 106                 r0_off = fpu_state_off + FPUStateSizeInWords,
 107                 rfp_off = r0_off + (Register::number_of_registers - 2) * Register::max_slots_per_register,
 108                 return_off = rfp_off + Register::max_slots_per_register,      // slot for return address
 109                 reg_save_size = return_off + Register::max_slots_per_register};
 110 
 111 };
 112 
 113 int RegisterSaver::reg_offset_in_bytes(Register r) {
 114   // The integer registers are located above the floating point
 115   // registers in the stack frame pushed by save_live_registers() so the
 116   // offset depends on whether we are saving full vectors, and whether
 117   // those vectors are NEON or SVE.
 118 
 119   int slots_per_vect = FloatRegister::save_slots_per_register;
 120 
 121 #ifdef COMPILER2
 122   if (_save_vectors) {
 123     slots_per_vect = FloatRegister::slots_per_neon_register;
 124     if (Matcher::supports_scalable_vector()) {
 125       slots_per_vect = Matcher::scalable_vector_reg_size(T_FLOAT);
 126     }
 127   }
 128 #endif // COMPILER2
 129 
 130   int r0_offset = v0_offset_in_bytes() + (slots_per_vect * FloatRegister::number_of_registers) * BytesPerInt;
 131   return r0_offset + r->encoding() * wordSize;
 132 }
 133 
 134 int RegisterSaver::v0_offset_in_bytes() {
 135   // The floating point registers are located above the predicate registers if
 136   // they are present in the stack frame pushed by save_live_registers(). So the
 137   // offset depends on the saved total predicate vectors in the stack frame.
 138   return (total_sve_predicate_in_bytes() / VMRegImpl::stack_slot_size) * BytesPerInt;
 139 }
 140 
 141 int RegisterSaver::total_sve_predicate_in_bytes() {
 142 #ifdef COMPILER2
 143   if (_save_vectors && Matcher::supports_scalable_vector()) {
 144     return (Matcher::scalable_vector_reg_size(T_BYTE) >> LogBitsPerByte) *
 145            PRegister::number_of_registers;
 146   }
 147 #endif
 148   return 0;
 149 }
 150 
 151 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
 152   bool use_sve = false;
 153   int sve_vector_size_in_bytes = 0;
 154   int sve_vector_size_in_slots = 0;
 155   int sve_predicate_size_in_slots = 0;
 156   int total_predicate_in_bytes = total_sve_predicate_in_bytes();
 157   int total_predicate_in_slots = total_predicate_in_bytes / VMRegImpl::stack_slot_size;
 158 
 159 #ifdef COMPILER2
 160   use_sve = Matcher::supports_scalable_vector();
 161   if (use_sve) {
 162     sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
 163     sve_vector_size_in_slots = Matcher::scalable_vector_reg_size(T_FLOAT);
 164     sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
 165   }
 166 #endif
 167 
 168 #ifdef COMPILER2
 169   if (_save_vectors) {
 170     int extra_save_slots_per_register = 0;
 171     // Save upper half of vector registers
 172     if (use_sve) {
 173       extra_save_slots_per_register = sve_vector_size_in_slots - FloatRegister::save_slots_per_register;
 174     } else {
 175       extra_save_slots_per_register = FloatRegister::extra_save_slots_per_neon_register;
 176     }
 177     int extra_vector_bytes = extra_save_slots_per_register *
 178                              VMRegImpl::stack_slot_size *
 179                              FloatRegister::number_of_registers;
 180     additional_frame_words += ((extra_vector_bytes + total_predicate_in_bytes) / wordSize);
 181   }
 182 #else
 183   assert(!_save_vectors, "vectors are generated only by C2");
 184 #endif // COMPILER2
 185 
 186   int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
 187                                      reg_save_size * BytesPerInt, 16);
 188   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 189   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 190   // The caller will allocate additional_frame_words
 191   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 192   // CodeBlob frame size is in words.
 193   int frame_size_in_words = frame_size_in_bytes / wordSize;
 194   *total_frame_words = frame_size_in_words;
 195 
 196   // Save Integer and Float registers.
 197   __ enter();
 198   __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
 199 
 200   // Set an oopmap for the call site.  This oopmap will map all
 201   // oop-registers and debug-info registers as callee-saved.  This
 202   // will allow deoptimization at this safepoint to find all possible
 203   // debug-info recordings, as well as let GC find all oops.
 204 
 205   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 206 
 207   for (int i = 0; i < Register::number_of_registers; i++) {
 208     Register r = as_Register(i);
 209     if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
 210       // SP offsets are in 4-byte words.
 211       // Register slots are 8 bytes wide, 32 floating-point registers.
 212       int sp_offset = Register::max_slots_per_register * i +
 213                       FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
 214       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
 215     }
 216   }
 217 
 218   for (int i = 0; i < FloatRegister::number_of_registers; i++) {
 219     FloatRegister r = as_FloatRegister(i);
 220     int sp_offset = 0;
 221     if (_save_vectors) {
 222       sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
 223                             (FloatRegister::slots_per_neon_register * i);
 224     } else {
 225       sp_offset = FloatRegister::save_slots_per_register * i;
 226     }
 227     oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), r->as_VMReg());
 228   }
 229 
 230   return oop_map;
 231 }
 232 
 233 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
 234 #ifdef COMPILER2
 235   __ pop_CPU_state(_save_vectors, Matcher::supports_scalable_vector(),
 236                    Matcher::scalable_vector_reg_size(T_BYTE), total_sve_predicate_in_bytes());
 237 #else
 238   assert(!_save_vectors, "vectors are generated only by C2");
 239   __ pop_CPU_state(_save_vectors);
 240 #endif
 241   __ ldp(rfp, lr, Address(__ post(sp, 2 * wordSize)));
 242   __ authenticate_return_address();
 243 }
 244 
 245 // Is vector's size (in bytes) bigger than a size saved by default?
 246 // 8 bytes vector registers are saved by default on AArch64.
 247 // The SVE supported min vector size is 8 bytes and we need to save
 248 // predicate registers when the vector size is 8 bytes as well.
 249 bool SharedRuntime::is_wide_vector(int size) {
 250   return size > 8 || (UseSVE > 0 && size >= 8);
 251 }
 252 
 253 // ---------------------------------------------------------------------------
 254 // Read the array of BasicTypes from a signature, and compute where the
 255 // arguments should go.  Values in the VMRegPair regs array refer to 4-byte
 256 // quantities.  Values less than VMRegImpl::stack0 are registers, those above
 257 // refer to 4-byte stack slots.  All stack slots are based off of the stack pointer
 258 // as framesizes are fixed.
 259 // VMRegImpl::stack0 refers to the first slot 0(sp).
 260 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
 261 // Register up to Register::number_of_registers are the 64-bit
 262 // integer registers.
 263 
 264 // Note: the INPUTS in sig_bt are in units of Java argument words,
 265 // which are 64-bit.  The OUTPUTS are in 32-bit units.
 266 
 267 // The Java calling convention is a "shifted" version of the C ABI.
 268 // By skipping the first C ABI register we can call non-static jni
 269 // methods with small numbers of arguments without having to shuffle
 270 // the arguments at all. Since we control the java ABI we ought to at
 271 // least get some advantage out of it.
 272 
 273 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
 274                                            VMRegPair *regs,
 275                                            int total_args_passed) {
 276 
 277   // Create the mapping between argument positions and
 278   // registers.
 279   static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
 280     j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7
 281   };
 282   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
 283     j_farg0, j_farg1, j_farg2, j_farg3,
 284     j_farg4, j_farg5, j_farg6, j_farg7
 285   };
 286 
 287 
 288   uint int_args = 0;
 289   uint fp_args = 0;
 290   uint stk_args = 0;
 291 
 292   for (int i = 0; i < total_args_passed; i++) {
 293     switch (sig_bt[i]) {
 294     case T_BOOLEAN:
 295     case T_CHAR:
 296     case T_BYTE:
 297     case T_SHORT:
 298     case T_INT:
 299       if (int_args < Argument::n_int_register_parameters_j) {
 300         regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 301       } else {
 302         stk_args = align_up(stk_args, 2);
 303         regs[i].set1(VMRegImpl::stack2reg(stk_args));
 304         stk_args += 1;
 305       }
 306       break;
 307     case T_VOID:
 308       // halves of T_LONG or T_DOUBLE
 309       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 310       regs[i].set_bad();
 311       break;
 312     case T_LONG:
 313       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 314       // fall through
 315     case T_OBJECT:
 316     case T_ARRAY:
 317     case T_ADDRESS:
 318       if (int_args < Argument::n_int_register_parameters_j) {
 319         regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 320       } else {
 321         stk_args = align_up(stk_args, 2);
 322         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 323         stk_args += 2;
 324       }
 325       break;
 326     case T_FLOAT:
 327       if (fp_args < Argument::n_float_register_parameters_j) {
 328         regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 329       } else {
 330         stk_args = align_up(stk_args, 2);
 331         regs[i].set1(VMRegImpl::stack2reg(stk_args));
 332         stk_args += 1;
 333       }
 334       break;
 335     case T_DOUBLE:
 336       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 337       if (fp_args < Argument::n_float_register_parameters_j) {
 338         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 339       } else {
 340         stk_args = align_up(stk_args, 2);
 341         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 342         stk_args += 2;
 343       }
 344       break;
 345     default:
 346       ShouldNotReachHere();
 347       break;
 348     }
 349   }
 350 
 351   return stk_args;
 352 }
 353 
 354 
 355 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
 356 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
 357 
 358 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
 359 
 360   // Create the mapping between argument positions and registers.
 361 
 362   static const Register INT_ArgReg[java_return_convention_max_int] = {
 363     r0 /* j_rarg7 */, j_rarg6, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
 364   };
 365 
 366   static const FloatRegister FP_ArgReg[java_return_convention_max_float] = {
 367     j_farg0, j_farg1, j_farg2, j_farg3, j_farg4, j_farg5, j_farg6, j_farg7
 368   };
 369 
 370   uint int_args = 0;
 371   uint fp_args = 0;
 372 
 373   for (int i = 0; i < total_args_passed; i++) {
 374     switch (sig_bt[i]) {
 375     case T_BOOLEAN:
 376     case T_CHAR:
 377     case T_BYTE:
 378     case T_SHORT:
 379     case T_INT:
 380       if (int_args < SharedRuntime::java_return_convention_max_int) {
 381         regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
 382         int_args ++;
 383       } else {
 384         return -1;
 385       }
 386       break;
 387     case T_VOID:
 388       // halves of T_LONG or T_DOUBLE
 389       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 390       regs[i].set_bad();
 391       break;
 392     case T_LONG:
 393       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 394       // fall through
 395     case T_OBJECT:
 396     case T_ARRAY:
 397     case T_ADDRESS:
 398       // Should T_METADATA be added to java_calling_convention as well ?
 399     case T_METADATA:
 400       if (int_args < SharedRuntime::java_return_convention_max_int) {
 401         regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
 402         int_args ++;
 403       } else {
 404         return -1;
 405       }
 406       break;
 407     case T_FLOAT:
 408       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 409         regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
 410         fp_args ++;
 411       } else {
 412         return -1;
 413       }
 414       break;
 415     case T_DOUBLE:
 416       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 417       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 418         regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
 419         fp_args ++;
 420       } else {
 421         return -1;
 422       }
 423       break;
 424     default:
 425       ShouldNotReachHere();
 426       break;
 427     }
 428   }
 429 
 430   return int_args + fp_args;
 431 }
 432 
 433 // Patch the callers callsite with entry to compiled code if it exists.
 434 static void patch_callers_callsite(MacroAssembler *masm) {
 435   Label L;
 436   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 437   __ cbz(rscratch1, L);
 438 
 439   __ enter();
 440   __ push_CPU_state();
 441 
 442   // VM needs caller's callsite
 443   // VM needs target method
 444   // This needs to be a long call since we will relocate this adapter to
 445   // the codeBuffer and it may not reach
 446 
 447 #ifndef PRODUCT
 448   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 449 #endif
 450 
 451   __ mov(c_rarg0, rmethod);
 452   __ mov(c_rarg1, lr);
 453   __ authenticate_return_address(c_rarg1);
 454   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 455   __ blr(rscratch1);
 456 
 457   // Explicit isb required because fixup_callers_callsite may change the code
 458   // stream.
 459   __ safepoint_isb();
 460 
 461   __ pop_CPU_state();
 462   // restore sp
 463   __ leave();
 464   __ bind(L);
 465 }
 466 
 467 // For each inline type argument, sig includes the list of fields of
 468 // the inline type. This utility function computes the number of
 469 // arguments for the call if inline types are passed by reference (the
 470 // calling convention the interpreter expects).
 471 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
 472   int total_args_passed = 0;
 473   if (InlineTypePassFieldsAsArgs) {
 474     for (int i = 0; i < sig_extended->length(); i++) {
 475       BasicType bt = sig_extended->at(i)._bt;
 476       if (bt == T_METADATA) {
 477         // In sig_extended, an inline type argument starts with:
 478         // T_METADATA, followed by the types of the fields of the
 479         // inline type and T_VOID to mark the end of the value
 480         // type. Inline types are flattened so, for instance, in the
 481         // case of an inline type with an int field and an inline type
 482         // field that itself has 2 fields, an int and a long:
 483         // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
 484         // slot for the T_LONG) T_VOID (inner inline type) T_VOID
 485         // (outer inline type)
 486         total_args_passed++;
 487         int vt = 1;
 488         do {
 489           i++;
 490           BasicType bt = sig_extended->at(i)._bt;
 491           BasicType prev_bt = sig_extended->at(i-1)._bt;
 492           if (bt == T_METADATA) {
 493             vt++;
 494           } else if (bt == T_VOID &&
 495                      prev_bt != T_LONG &&
 496                      prev_bt != T_DOUBLE) {
 497             vt--;
 498           }
 499         } while (vt != 0);
 500       } else {
 501         total_args_passed++;
 502       }
 503     }
 504   } else {
 505     total_args_passed = sig_extended->length();
 506   }
 507   return total_args_passed;
 508 }
 509 
 510 
 511 static void gen_c2i_adapter_helper(MacroAssembler* masm,
 512                                    BasicType bt,
 513                                    BasicType prev_bt,
 514                                    size_t size_in_bytes,
 515                                    const VMRegPair& reg_pair,
 516                                    const Address& to,
 517                                    Register tmp1,
 518                                    Register tmp2,
 519                                    Register tmp3,
 520                                    int extraspace,
 521                                    bool is_oop) {
 522   if (bt == T_VOID) {
 523     assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
 524     return;
 525   }
 526 
 527   // Say 4 args:
 528   // i   st_off
 529   // 0   32 T_LONG
 530   // 1   24 T_VOID
 531   // 2   16 T_OBJECT
 532   // 3    8 T_BOOL
 533   // -    0 return address
 534   //
 535   // However to make thing extra confusing. Because we can fit a Java long/double in
 536   // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 537   // leaves one slot empty and only stores to a single slot. In this case the
 538   // slot that is occupied is the T_VOID slot. See I said it was confusing.
 539 
 540   bool wide = (size_in_bytes == wordSize);
 541   VMReg r_1 = reg_pair.first();
 542   VMReg r_2 = reg_pair.second();
 543   assert(r_2->is_valid() == wide, "invalid size");
 544   if (!r_1->is_valid()) {
 545     assert(!r_2->is_valid(), "");
 546     return;
 547   }
 548 
 549   if (!r_1->is_FloatRegister()) {
 550     Register val = r25;
 551     if (r_1->is_stack()) {
 552       // memory to memory use r25 (scratch registers is used by store_heap_oop)
 553       int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 554       __ load_sized_value(val, Address(sp, ld_off), size_in_bytes, /* is_signed */ false);
 555     } else {
 556       val = r_1->as_Register();
 557     }
 558     assert_different_registers(to.base(), val, tmp1, tmp2, tmp3);
 559     if (is_oop) {
 560       // store_heap_oop transitively calls oop_store_at which corrupts to.base(). We need to keep it valid.
 561       __ push(to.base(), sp);
 562       __ store_heap_oop(to, val, tmp1, tmp2, tmp3, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
 563       __ pop(to.base(), sp);
 564     } else {
 565       __ store_sized_value(to, val, size_in_bytes);
 566     }
 567   } else {
 568     if (wide) {
 569       __ strd(r_1->as_FloatRegister(), to);
 570     } else {
 571       // only a float use just part of the slot
 572       __ strs(r_1->as_FloatRegister(), to);
 573     }
 574   }
 575 }
 576 
 577 static void gen_c2i_adapter(MacroAssembler *masm,
 578                             const GrowableArray<SigEntry>* sig_extended,
 579                             const VMRegPair *regs,
 580                             bool requires_clinit_barrier,
 581                             address& c2i_no_clinit_check_entry,
 582                             Label& skip_fixup,
 583                             address start,
 584                             OopMapSet* oop_maps,
 585                             int& frame_complete,
 586                             int& frame_size_in_words,
 587                             bool alloc_inline_receiver) {
 588   if (requires_clinit_barrier) {
 589     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
 590     Label L_skip_barrier;
 591 
 592     { // Bypass the barrier for non-static methods
 593       __ ldrh(rscratch1, Address(rmethod, Method::access_flags_offset()));
 594       __ andsw(zr, rscratch1, JVM_ACC_STATIC);
 595       __ br(Assembler::EQ, L_skip_barrier); // non-static
 596     }
 597 
 598     __ load_method_holder(rscratch2, rmethod);
 599     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
 600     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 601 
 602     __ bind(L_skip_barrier);
 603     c2i_no_clinit_check_entry = __ pc();
 604   }
 605 
 606   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 607   bs->c2i_entry_barrier(masm);
 608 
 609   // Before we get into the guts of the C2I adapter, see if we should be here
 610   // at all.  We've come from compiled code and are attempting to jump to the
 611   // interpreter, which means the caller made a static call to get here
 612   // (vcalls always get a compiled target if there is one).  Check for a
 613   // compiled target.  If there is one, we need to patch the caller's call.
 614   patch_callers_callsite(masm);
 615 
 616   __ bind(skip_fixup);
 617 
 618   // Name some registers to be used in the following code. We can use
 619   // anything except r0-r7 which are arguments in the Java calling
 620   // convention, rmethod (r12), and r19 which holds the outgoing sender
 621   // SP for the interpreter.
 622   Register buf_array = r10;   // Array of buffered inline types
 623   Register buf_oop = r11;     // Buffered inline type oop
 624   Register tmp1 = r15;
 625   Register tmp2 = r16;
 626   Register tmp3 = r17;
 627 
 628 #ifdef ASSERT
 629   RegSet clobbered_gp_regs = MacroAssembler::call_clobbered_gp_registers();
 630   assert(clobbered_gp_regs.contains(buf_array), "buf_array must be saved explicitly if it's not a clobber");
 631   assert(clobbered_gp_regs.contains(buf_oop), "buf_oop must be saved explicitly if it's not a clobber");
 632   assert(clobbered_gp_regs.contains(tmp1), "tmp1 must be saved explicitly if it's not a clobber");
 633   assert(clobbered_gp_regs.contains(tmp2), "tmp2 must be saved explicitly if it's not a clobber");
 634   assert(clobbered_gp_regs.contains(tmp3), "tmp3 must be saved explicitly if it's not a clobber");
 635 #endif
 636 
 637   if (InlineTypePassFieldsAsArgs) {
 638     // Is there an inline type argument?
 639     bool has_inline_argument = false;
 640     for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
 641       has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
 642     }
 643     if (has_inline_argument) {
 644       // There is at least a value type argument: we're coming from
 645       // compiled code so we may not have buffers to back the value
 646       // objects. Allocate the buffers here with a runtime call for
 647       // the value arguments that needs a buffer.
 648       RegisterSaver reg_save(true /* save_vectors */);
 649       OopMap* map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
 650 
 651       frame_complete = __ offset();
 652       address the_pc = __ pc();
 653 
 654       Label retaddr;
 655       __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
 656 
 657       __ mov(c_rarg0, rthread);
 658       __ mov(c_rarg1, rmethod);
 659       __ mov(c_rarg2, (int64_t)alloc_inline_receiver);
 660 
 661       __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
 662       __ blr(rscratch1);
 663       __ bind(retaddr);
 664 
 665       oop_maps->add_gc_map(__ pc() - start, map);
 666       __ reset_last_Java_frame(false);
 667 
 668       reg_save.restore_live_registers(masm);
 669 
 670       Label no_exception;
 671       __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
 672       __ cbz(rscratch1, no_exception);
 673 
 674       __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
 675       __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
 676       __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
 677 
 678       __ bind(no_exception);
 679 
 680       // We get an array of objects from the runtime call
 681       __ get_vm_result_oop(buf_array, rthread);
 682     }
 683   }
 684 
 685   // Since all args are passed on the stack, total_args_passed *
 686   // Interpreter::stackElementSize is the space we need.
 687 
 688   int total_args_passed = compute_total_args_passed_int(sig_extended);
 689   int extraspace = total_args_passed * Interpreter::stackElementSize;
 690 
 691   // stack is aligned, keep it that way
 692   extraspace = align_up(extraspace, StackAlignmentInBytes);
 693 
 694   // set senderSP value
 695   __ mov(r19_sender_sp, sp);
 696 
 697   __ sub(sp, sp, extraspace);
 698 
 699   // Now write the args into the outgoing interpreter space
 700 
 701   // next_arg_comp is the next argument from the compiler point of
 702   // view (inline type fields are passed in registers/on the stack). In
 703   // sig_extended, an inline type argument starts with: T_METADATA,
 704   // followed by the types of the fields of the inline type and T_VOID
 705   // to mark the end of the inline type. ignored counts the number of
 706   // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
 707   // used to get the buffer for that argument from the pool of buffers
 708   // we allocated above and want to pass to the
 709   // interpreter. next_arg_int is the next argument from the
 710   // interpreter point of view (inline types are passed by reference).
 711   for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
 712        next_arg_comp < sig_extended->length(); next_arg_comp++) {
 713     assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
 714     assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
 715     BasicType bt = sig_extended->at(next_arg_comp)._bt;
 716     int st_off = (total_args_passed - next_arg_int - 1) * Interpreter::stackElementSize;
 717     if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
 718       int next_off = st_off - Interpreter::stackElementSize;
 719       const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
 720       const VMRegPair reg_pair = regs[next_arg_comp-ignored];
 721       size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
 722       gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 723                              size_in_bytes, reg_pair, Address(sp, offset), tmp1, tmp2, tmp3, extraspace, false);
 724       next_arg_int++;
 725 #ifdef ASSERT
 726       if (bt == T_LONG || bt == T_DOUBLE) {
 727         // Overwrite the unused slot with known junk
 728         __ mov(rscratch1, CONST64(0xdeadffffdeadaaaa));
 729         __ str(rscratch1, Address(sp, st_off));
 730       }
 731 #endif /* ASSERT */
 732     } else {
 733       ignored++;
 734       next_arg_int++;
 735       int vt = 1;
 736       // write fields we get from compiled code in registers/stack
 737       // slots to the buffer: we know we are done with that inline type
 738       // argument when we hit the T_VOID that acts as an end of inline
 739       // type delimiter for this inline type. Inline types are flattened
 740       // so we might encounter embedded inline types. Each entry in
 741       // sig_extended contains a field offset in the buffer.
 742       Label L_null;
 743       Label not_null_buffer;
 744       do {
 745         next_arg_comp++;
 746         BasicType bt = sig_extended->at(next_arg_comp)._bt;
 747         BasicType prev_bt = sig_extended->at(next_arg_comp - 1)._bt;
 748         if (bt == T_METADATA) {
 749           vt++;
 750           ignored++;
 751         } else if (bt == T_VOID && prev_bt != T_LONG && prev_bt != T_DOUBLE) {
 752           vt--;
 753           ignored++;
 754         } else if (sig_extended->at(next_arg_comp)._vt_oop) {
 755           VMReg buffer = regs[next_arg_comp-ignored].first();
 756           if (buffer->is_stack()) {
 757             int ld_off = buffer->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 758             __ ldr(buf_oop, Address(sp, ld_off));
 759           } else {
 760             __ mov(buf_oop, buffer->as_Register());
 761           }
 762           __ cbnz(buf_oop, not_null_buffer);
 763           // get the buffer from the just allocated pool of buffers
 764           int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
 765           __ load_heap_oop(buf_oop, Address(buf_array, index), rscratch1, tmp2);
 766           next_vt_arg++;
 767         } else {
 768           int off = sig_extended->at(next_arg_comp)._offset;
 769           if (off == -1) {
 770             // Nullable inline type argument, emit null check
 771             VMReg reg = regs[next_arg_comp-ignored].first();
 772             Label L_notNull;
 773             if (reg->is_stack()) {
 774               int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 775               __ ldrb(tmp1, Address(sp, ld_off));
 776               __ cbnz(tmp1, L_notNull);
 777             } else {
 778               __ cbnz(reg->as_Register(), L_notNull);
 779             }
 780             __ str(zr, Address(sp, st_off));
 781             __ b(L_null);
 782             __ bind(L_notNull);
 783             continue;
 784           }
 785           assert(off > 0, "offset in object should be positive");
 786           size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
 787           bool is_oop = is_reference_type(bt);
 788           gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 789                                  size_in_bytes, regs[next_arg_comp-ignored], Address(buf_oop, off), tmp1, tmp2, tmp3, extraspace, is_oop);
 790         }
 791       } while (vt != 0);
 792       // pass the buffer to the interpreter
 793       __ bind(not_null_buffer);
 794       __ str(buf_oop, Address(sp, st_off));
 795       __ bind(L_null);
 796     }
 797   }
 798 
 799   __ mov(esp, sp); // Interp expects args on caller's expression stack
 800 
 801   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
 802   __ br(rscratch1);
 803 }
 804 
 805 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, int comp_args_on_stack, const GrowableArray<SigEntry>* sig, const VMRegPair *regs) {
 806 
 807 
 808   // Note: r19_sender_sp contains the senderSP on entry. We must
 809   // preserve it since we may do a i2c -> c2i transition if we lose a
 810   // race where compiled code goes non-entrant while we get args
 811   // ready.
 812 
 813   // Adapters are frameless.
 814 
 815   // An i2c adapter is frameless because the *caller* frame, which is
 816   // interpreted, routinely repairs its own esp (from
 817   // interpreter_frame_last_sp), even if a callee has modified the
 818   // stack pointer.  It also recalculates and aligns sp.
 819 
 820   // A c2i adapter is frameless because the *callee* frame, which is
 821   // interpreted, routinely repairs its caller's sp (from sender_sp,
 822   // which is set up via the senderSP register).
 823 
 824   // In other words, if *either* the caller or callee is interpreted, we can
 825   // get the stack pointer repaired after a call.
 826 
 827   // This is why c2i and i2c adapters cannot be indefinitely composed.
 828   // In particular, if a c2i adapter were to somehow call an i2c adapter,
 829   // both caller and callee would be compiled methods, and neither would
 830   // clean up the stack pointer changes performed by the two adapters.
 831   // If this happens, control eventually transfers back to the compiled
 832   // caller, but with an uncorrected stack, causing delayed havoc.
 833 
 834   // Cut-out for having no stack args.
 835   int comp_words_on_stack = 0;
 836   if (comp_args_on_stack) {
 837      comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
 838      __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
 839      __ andr(sp, rscratch1, -16);
 840   }
 841 
 842   // Will jump to the compiled code just as if compiled code was doing it.
 843   // Pre-load the register-jump target early, to schedule it better.
 844   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_inline_offset())));
 845 
 846   int total_args_passed = sig->length();
 847 
 848   // Now generate the shuffle code.
 849   for (int i = 0; i < total_args_passed; i++) {
 850     BasicType bt = sig->at(i)._bt;
 851     if (bt == T_VOID) {
 852       assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
 853       continue;
 854     }
 855 
 856     // Pick up 0, 1 or 2 words from SP+offset.
 857     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), "scrambled load targets?");
 858 
 859     // Load in argument order going down.
 860     int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 861     // Point to interpreter value (vs. tag)
 862     int next_off = ld_off - Interpreter::stackElementSize;
 863     //
 864     //
 865     //
 866     VMReg r_1 = regs[i].first();
 867     VMReg r_2 = regs[i].second();
 868     if (!r_1->is_valid()) {
 869       assert(!r_2->is_valid(), "");
 870       continue;
 871     }
 872     if (r_1->is_stack()) {
 873       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 874       int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
 875       if (!r_2->is_valid()) {
 876         // sign extend???
 877         __ ldrsw(rscratch2, Address(esp, ld_off));
 878         __ str(rscratch2, Address(sp, st_off));
 879       } else {
 880         //
 881         // We are using two optoregs. This can be either T_OBJECT,
 882         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 883         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 884         // So we must adjust where to pick up the data to match the
 885         // interpreter.
 886         //
 887         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 888         // are accessed as negative so LSW is at LOW address
 889 
 890         // ld_off is MSW so get LSW
 891         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
 892         __ ldr(rscratch2, Address(esp, offset));
 893         // st_off is LSW (i.e. reg.first())
 894          __ str(rscratch2, Address(sp, st_off));
 895        }
 896      } else if (r_1->is_Register()) {  // Register argument
 897        Register r = r_1->as_Register();
 898        if (r_2->is_valid()) {
 899          //
 900          // We are using two VMRegs. This can be either T_OBJECT,
 901          // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 902          // two slots but only uses one for thr T_LONG or T_DOUBLE case
 903          // So we must adjust where to pick up the data to match the
 904          // interpreter.
 905 
 906         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
 907 
 908          // this can be a misaligned move
 909          __ ldr(r, Address(esp, offset));
 910        } else {
 911          // sign extend and use a full word?
 912          __ ldrw(r, Address(esp, ld_off));
 913        }
 914      } else {
 915        if (!r_2->is_valid()) {
 916          __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
 917        } else {
 918          __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
 919        }
 920      }
 921    }
 922 
 923   __ mov(rscratch2, rscratch1);
 924   __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1
 925   __ mov(rscratch1, rscratch2);
 926 
 927   // 6243940 We might end up in handle_wrong_method if
 928   // the callee is deoptimized as we race thru here. If that
 929   // happens we don't want to take a safepoint because the
 930   // caller frame will look interpreted and arguments are now
 931   // "compiled" so it is much better to make this transition
 932   // invisible to the stack walking code. Unfortunately if
 933   // we try and find the callee by normal means a safepoint
 934   // is possible. So we stash the desired callee in the thread
 935   // and the vm will find there should this case occur.
 936 
 937   __ str(rmethod, Address(rthread, JavaThread::callee_target_offset()));
 938 
 939   __ br(rscratch1);
 940 }
 941 
 942 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
 943   Register data = rscratch2;
 944   __ ic_check(1 /* end_alignment */);
 945   __ ldr(rmethod, Address(data, CompiledICData::speculated_method_offset()));
 946 
 947   // Method might have been compiled since the call site was patched to
 948   // interpreted; if that is the case treat it as a miss so we can get
 949   // the call site corrected.
 950   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 951   __ cbz(rscratch1, skip_fixup);
 952   __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
 953 }
 954 
 955 // ---------------------------------------------------------------
 956 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
 957                                             int comp_args_on_stack,
 958                                             const GrowableArray<SigEntry>* sig,
 959                                             const VMRegPair* regs,
 960                                             const GrowableArray<SigEntry>* sig_cc,
 961                                             const VMRegPair* regs_cc,
 962                                             const GrowableArray<SigEntry>* sig_cc_ro,
 963                                             const VMRegPair* regs_cc_ro,
 964                                             address entry_address[AdapterBlob::ENTRY_COUNT],
 965                                             AdapterBlob*& new_adapter,
 966                                             bool allocate_code_blob) {
 967 
 968   entry_address[AdapterBlob::I2C] = __ pc();
 969   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
 970 
 971   // -------------------------------------------------------------------------
 972   // Generate a C2I adapter.  On entry we know rmethod holds the Method* during calls
 973   // to the interpreter.  The args start out packed in the compiled layout.  They
 974   // need to be unpacked into the interpreter layout.  This will almost always
 975   // require some stack space.  We grow the current (compiled) stack, then repack
 976   // the args.  We  finally end in a jump to the generic interpreter entry point.
 977   // On exit from the interpreter, the interpreter will restore our SP (lest the
 978   // compiled code, which relies solely on SP and not FP, get sick).
 979 
 980   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
 981   entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
 982   Label skip_fixup;
 983 
 984   gen_inline_cache_check(masm, skip_fixup);
 985 
 986   OopMapSet* oop_maps = new OopMapSet();
 987   int frame_complete = CodeOffsets::frame_never_safe;
 988   int frame_size_in_words = 0;
 989 
 990   // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
 991   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
 992   entry_address[AdapterBlob::C2I_Inline_RO] = __ pc();
 993   if (regs_cc != regs_cc_ro) {
 994     // No class init barrier needed because method is guaranteed to be non-static
 995     gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, entry_address[AdapterBlob::C2I_No_Clinit_Check],
 996                     skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
 997     skip_fixup.reset();
 998   }
 999 
1000   // Scalarized c2i adapter
1001   entry_address[AdapterBlob::C2I]        = __ pc();
1002   entry_address[AdapterBlob::C2I_Inline] = __ pc();
1003   gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1004                   skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1005 
1006   // Non-scalarized c2i adapter
1007   if (regs != regs_cc) {
1008     entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
1009     Label inline_entry_skip_fixup;
1010     gen_inline_cache_check(masm, inline_entry_skip_fixup);
1011 
1012     entry_address[AdapterBlob::C2I_Inline] = __ pc();
1013     gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1014                     inline_entry_skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1015   }
1016 
1017   // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
1018   // the GC knows about the location of oop argument locations passed to the c2i adapter.
1019   if (allocate_code_blob) {
1020     bool caller_must_gc_arguments = (regs != regs_cc);
1021     int entry_offset[AdapterHandlerEntry::ENTRIES_COUNT];
1022     assert(AdapterHandlerEntry::ENTRIES_COUNT == 7, "sanity");
1023     AdapterHandlerLibrary::address_to_offset(entry_address, entry_offset);
1024     new_adapter = AdapterBlob::create(masm->code(), entry_offset, frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1025   }
1026 }
1027 
1028 static int c_calling_convention_priv(const BasicType *sig_bt,
1029                                          VMRegPair *regs,
1030                                          int total_args_passed) {
1031 
1032 // We return the amount of VMRegImpl stack slots we need to reserve for all
1033 // the arguments NOT counting out_preserve_stack_slots.
1034 
1035     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1036       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
1037     };
1038     static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1039       c_farg0, c_farg1, c_farg2, c_farg3,
1040       c_farg4, c_farg5, c_farg6, c_farg7
1041     };
1042 
1043     uint int_args = 0;
1044     uint fp_args = 0;
1045     uint stk_args = 0; // inc by 2 each time
1046 
1047     for (int i = 0; i < total_args_passed; i++) {
1048       switch (sig_bt[i]) {
1049       case T_BOOLEAN:
1050       case T_CHAR:
1051       case T_BYTE:
1052       case T_SHORT:
1053       case T_INT:
1054         if (int_args < Argument::n_int_register_parameters_c) {
1055           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
1056         } else {
1057 #ifdef __APPLE__
1058           // Less-than word types are stored one after another.
1059           // The code is unable to handle this so bailout.
1060           return -1;
1061 #endif
1062           regs[i].set1(VMRegImpl::stack2reg(stk_args));
1063           stk_args += 2;
1064         }
1065         break;
1066       case T_LONG:
1067         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1068         // fall through
1069       case T_OBJECT:
1070       case T_ARRAY:
1071       case T_ADDRESS:
1072       case T_METADATA:
1073         if (int_args < Argument::n_int_register_parameters_c) {
1074           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
1075         } else {
1076           regs[i].set2(VMRegImpl::stack2reg(stk_args));
1077           stk_args += 2;
1078         }
1079         break;
1080       case T_FLOAT:
1081         if (fp_args < Argument::n_float_register_parameters_c) {
1082           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
1083         } else {
1084 #ifdef __APPLE__
1085           // Less-than word types are stored one after another.
1086           // The code is unable to handle this so bailout.
1087           return -1;
1088 #endif
1089           regs[i].set1(VMRegImpl::stack2reg(stk_args));
1090           stk_args += 2;
1091         }
1092         break;
1093       case T_DOUBLE:
1094         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1095         if (fp_args < Argument::n_float_register_parameters_c) {
1096           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
1097         } else {
1098           regs[i].set2(VMRegImpl::stack2reg(stk_args));
1099           stk_args += 2;
1100         }
1101         break;
1102       case T_VOID: // Halves of longs and doubles
1103         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
1104         regs[i].set_bad();
1105         break;
1106       default:
1107         ShouldNotReachHere();
1108         break;
1109       }
1110     }
1111 
1112   return stk_args;
1113 }
1114 
1115 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
1116                                              uint num_bits,
1117                                              uint total_args_passed) {
1118   // More than 8 argument inputs are not supported now.
1119   assert(total_args_passed <= Argument::n_float_register_parameters_c, "unsupported");
1120   assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
1121 
1122   static const FloatRegister VEC_ArgReg[Argument::n_float_register_parameters_c] = {
1123     v0, v1, v2, v3, v4, v5, v6, v7
1124   };
1125 
1126   // On SVE, we use the same vector registers with 128-bit vector registers on NEON.
1127   int next_reg_val = num_bits == 64 ? 1 : 3;
1128   for (uint i = 0; i < total_args_passed; i++) {
1129     VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
1130     regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
1131   }
1132   return 0;
1133 }
1134 
1135 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1136                                          VMRegPair *regs,
1137                                          int total_args_passed)
1138 {
1139   int result = c_calling_convention_priv(sig_bt, regs, total_args_passed);
1140   guarantee(result >= 0, "Unsupported arguments configuration");
1141   return result;
1142 }
1143 
1144 
1145 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1146   // We always ignore the frame_slots arg and just use the space just below frame pointer
1147   // which by this time is free to use
1148   switch (ret_type) {
1149   case T_FLOAT:
1150     __ strs(v0, Address(rfp, -wordSize));
1151     break;
1152   case T_DOUBLE:
1153     __ strd(v0, Address(rfp, -wordSize));
1154     break;
1155   case T_VOID:  break;
1156   default: {
1157     __ str(r0, Address(rfp, -wordSize));
1158     }
1159   }
1160 }
1161 
1162 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1163   // We always ignore the frame_slots arg and just use the space just below frame pointer
1164   // which by this time is free to use
1165   switch (ret_type) {
1166   case T_FLOAT:
1167     __ ldrs(v0, Address(rfp, -wordSize));
1168     break;
1169   case T_DOUBLE:
1170     __ ldrd(v0, Address(rfp, -wordSize));
1171     break;
1172   case T_VOID:  break;
1173   default: {
1174     __ ldr(r0, Address(rfp, -wordSize));
1175     }
1176   }
1177 }
1178 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1179   RegSet x;
1180   for ( int i = first_arg ; i < arg_count ; i++ ) {
1181     if (args[i].first()->is_Register()) {
1182       x = x + args[i].first()->as_Register();
1183     } else if (args[i].first()->is_FloatRegister()) {
1184       __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize)));
1185     }
1186   }
1187   __ push(x, sp);
1188 }
1189 
1190 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1191   RegSet x;
1192   for ( int i = first_arg ; i < arg_count ; i++ ) {
1193     if (args[i].first()->is_Register()) {
1194       x = x + args[i].first()->as_Register();
1195     } else {
1196       ;
1197     }
1198   }
1199   __ pop(x, sp);
1200   for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
1201     if (args[i].first()->is_Register()) {
1202       ;
1203     } else if (args[i].first()->is_FloatRegister()) {
1204       __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize)));
1205     }
1206   }
1207 }
1208 
1209 static void verify_oop_args(MacroAssembler* masm,
1210                             const methodHandle& method,
1211                             const BasicType* sig_bt,
1212                             const VMRegPair* regs) {
1213   Register temp_reg = r19;  // not part of any compiled calling seq
1214   if (VerifyOops) {
1215     for (int i = 0; i < method->size_of_parameters(); i++) {
1216       if (sig_bt[i] == T_OBJECT ||
1217           sig_bt[i] == T_ARRAY) {
1218         VMReg r = regs[i].first();
1219         assert(r->is_valid(), "bad oop arg");
1220         if (r->is_stack()) {
1221           __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1222           __ verify_oop(temp_reg);
1223         } else {
1224           __ verify_oop(r->as_Register());
1225         }
1226       }
1227     }
1228   }
1229 }
1230 
1231 // on exit, sp points to the ContinuationEntry
1232 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
1233   assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1234   assert(in_bytes(ContinuationEntry::cont_offset())  % VMRegImpl::stack_slot_size == 0, "");
1235   assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1236 
1237   stack_slots += (int)ContinuationEntry::size()/wordSize;
1238   __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
1239 
1240   OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize)/ VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1241 
1242   __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1243   __ str(rscratch1, Address(sp, ContinuationEntry::parent_offset()));
1244   __ mov(rscratch1, sp); // we can't use sp as the source in str
1245   __ str(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1246 
1247   return map;
1248 }
1249 
1250 // on entry c_rarg1 points to the continuation
1251 //          sp points to ContinuationEntry
1252 //          c_rarg3 -- isVirtualThread
1253 static void fill_continuation_entry(MacroAssembler* masm) {
1254 #ifdef ASSERT
1255   __ movw(rscratch1, ContinuationEntry::cookie_value());
1256   __ strw(rscratch1, Address(sp, ContinuationEntry::cookie_offset()));
1257 #endif
1258 
1259   __ str (c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
1260   __ strw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
1261   __ str (zr,      Address(sp, ContinuationEntry::chunk_offset()));
1262   __ strw(zr,      Address(sp, ContinuationEntry::argsize_offset()));
1263   __ strw(zr,      Address(sp, ContinuationEntry::pin_count_offset()));
1264 
1265   __ ldr(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1266   __ str(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1267 
1268   __ str(zr, Address(rthread, JavaThread::cont_fastpath_offset()));
1269 }
1270 
1271 // on entry, sp points to the ContinuationEntry
1272 // on exit, rfp points to the spilled rfp in the entry frame
1273 static void continuation_enter_cleanup(MacroAssembler* masm) {
1274 #ifndef PRODUCT
1275   Label OK;
1276   __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1277   __ cmp(sp, rscratch1);
1278   __ br(Assembler::EQ, OK);
1279   __ stop("incorrect sp1");
1280   __ bind(OK);
1281 #endif
1282   __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1283   __ str(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1284   __ ldr(rscratch2, Address(sp, ContinuationEntry::parent_offset()));
1285   __ str(rscratch2, Address(rthread, JavaThread::cont_entry_offset()));
1286   __ add(rfp, sp, (int)ContinuationEntry::size());
1287 }
1288 
1289 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1290 // On entry: c_rarg1 -- the continuation object
1291 //           c_rarg2 -- isContinue
1292 //           c_rarg3 -- isVirtualThread
1293 static void gen_continuation_enter(MacroAssembler* masm,
1294                                  const methodHandle& method,
1295                                  const BasicType* sig_bt,
1296                                  const VMRegPair* regs,
1297                                  int& exception_offset,
1298                                  OopMapSet*oop_maps,
1299                                  int& frame_complete,
1300                                  int& stack_slots,
1301                                  int& interpreted_entry_offset,
1302                                  int& compiled_entry_offset) {
1303   //verify_oop_args(masm, method, sig_bt, regs);
1304   Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1305 
1306   address start = __ pc();
1307 
1308   Label call_thaw, exit;
1309 
1310   // i2i entry used at interp_only_mode only
1311   interpreted_entry_offset = __ pc() - start;
1312   {
1313 
1314 #ifdef ASSERT
1315     Label is_interp_only;
1316     __ ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
1317     __ cbnzw(rscratch1, is_interp_only);
1318     __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1319     __ bind(is_interp_only);
1320 #endif
1321 
1322     // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1323     __ ldr(c_rarg1, Address(esp, Interpreter::stackElementSize*2));
1324     __ ldr(c_rarg2, Address(esp, Interpreter::stackElementSize*1));
1325     __ ldr(c_rarg3, Address(esp, Interpreter::stackElementSize*0));
1326     __ push_cont_fastpath(rthread);
1327 
1328     __ enter();
1329     stack_slots = 2; // will be adjusted in setup
1330     OopMap* map = continuation_enter_setup(masm, stack_slots);
1331     // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1332     // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1333 
1334     fill_continuation_entry(masm);
1335 
1336     __ cbnz(c_rarg2, call_thaw);
1337 
1338     const address tr_call = __ trampoline_call(resolve);
1339     if (tr_call == nullptr) {
1340       fatal("CodeCache is full at gen_continuation_enter");
1341     }
1342 
1343     oop_maps->add_gc_map(__ pc() - start, map);
1344     __ post_call_nop();
1345 
1346     __ b(exit);
1347 
1348     address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1349     if (stub == nullptr) {
1350       fatal("CodeCache is full at gen_continuation_enter");
1351     }
1352   }
1353 
1354   // compiled entry
1355   __ align(CodeEntryAlignment);
1356   compiled_entry_offset = __ pc() - start;
1357 
1358   __ enter();
1359   stack_slots = 2; // will be adjusted in setup
1360   OopMap* map = continuation_enter_setup(masm, stack_slots);
1361   frame_complete = __ pc() - start;
1362 
1363   fill_continuation_entry(masm);
1364 
1365   __ cbnz(c_rarg2, call_thaw);
1366 
1367   const address tr_call = __ trampoline_call(resolve);
1368   if (tr_call == nullptr) {
1369     fatal("CodeCache is full at gen_continuation_enter");
1370   }
1371 
1372   oop_maps->add_gc_map(__ pc() - start, map);
1373   __ post_call_nop();
1374 
1375   __ b(exit);
1376 
1377   __ bind(call_thaw);
1378 
1379   ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1380   __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1381   oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1382   ContinuationEntry::_return_pc_offset = __ pc() - start;
1383   __ post_call_nop();
1384 
1385   __ bind(exit);
1386   ContinuationEntry::_cleanup_offset = __ pc() - start;
1387   continuation_enter_cleanup(masm);
1388   __ leave();
1389   __ ret(lr);
1390 
1391   /// exception handling
1392 
1393   exception_offset = __ pc() - start;
1394   {
1395       __ mov(r19, r0); // save return value contaning the exception oop in callee-saved R19
1396 
1397       continuation_enter_cleanup(masm);
1398 
1399       __ ldr(c_rarg1, Address(rfp, wordSize)); // return address
1400       __ authenticate_return_address(c_rarg1);
1401       __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1);
1402 
1403       // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc
1404 
1405       __ mov(r1, r0); // the exception handler
1406       __ mov(r0, r19); // restore return value contaning the exception oop
1407       __ verify_oop(r0);
1408 
1409       __ leave();
1410       __ mov(r3, lr);
1411       __ br(r1); // the exception handler
1412   }
1413 
1414   address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1415   if (stub == nullptr) {
1416     fatal("CodeCache is full at gen_continuation_enter");
1417   }
1418 }
1419 
1420 static void gen_continuation_yield(MacroAssembler* masm,
1421                                    const methodHandle& method,
1422                                    const BasicType* sig_bt,
1423                                    const VMRegPair* regs,
1424                                    OopMapSet* oop_maps,
1425                                    int& frame_complete,
1426                                    int& stack_slots,
1427                                    int& compiled_entry_offset) {
1428     enum layout {
1429       rfp_off1,
1430       rfp_off2,
1431       lr_off,
1432       lr_off2,
1433       framesize // inclusive of return address
1434     };
1435     // assert(is_even(framesize/2), "sp not 16-byte aligned");
1436     stack_slots = framesize /  VMRegImpl::slots_per_word;
1437     assert(stack_slots == 2, "recheck layout");
1438 
1439     address start = __ pc();
1440 
1441     compiled_entry_offset = __ pc() - start;
1442     __ enter();
1443 
1444     __ mov(c_rarg1, sp);
1445 
1446     frame_complete = __ pc() - start;
1447     address the_pc = __ pc();
1448 
1449     __ 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
1450 
1451     __ mov(c_rarg0, rthread);
1452     __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
1453     __ call_VM_leaf(Continuation::freeze_entry(), 2);
1454     __ reset_last_Java_frame(true);
1455 
1456     Label pinned;
1457 
1458     __ cbnz(r0, pinned);
1459 
1460     // We've succeeded, set sp to the ContinuationEntry
1461     __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1462     __ mov(sp, rscratch1);
1463     continuation_enter_cleanup(masm);
1464 
1465     __ bind(pinned); // pinned -- return to caller
1466 
1467     // handle pending exception thrown by freeze
1468     __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
1469     Label ok;
1470     __ cbz(rscratch1, ok);
1471     __ leave();
1472     __ lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
1473     __ br(rscratch1);
1474     __ bind(ok);
1475 
1476     __ leave();
1477     __ ret(lr);
1478 
1479     OopMap* map = new OopMap(framesize, 1);
1480     oop_maps->add_gc_map(the_pc - start, map);
1481 }
1482 
1483 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1484   ::continuation_enter_cleanup(masm);
1485 }
1486 
1487 static void gen_special_dispatch(MacroAssembler* masm,
1488                                  const methodHandle& method,
1489                                  const BasicType* sig_bt,
1490                                  const VMRegPair* regs) {
1491   verify_oop_args(masm, method, sig_bt, regs);
1492   vmIntrinsics::ID iid = method->intrinsic_id();
1493 
1494   // Now write the args into the outgoing interpreter space
1495   bool     has_receiver   = false;
1496   Register receiver_reg   = noreg;
1497   int      member_arg_pos = -1;
1498   Register member_reg     = noreg;
1499   int      ref_kind       = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1500   if (ref_kind != 0) {
1501     member_arg_pos = method->size_of_parameters() - 1;  // trailing MemberName argument
1502     member_reg = r19;  // known to be free at this point
1503     has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1504   } else if (iid == vmIntrinsics::_invokeBasic) {
1505     has_receiver = true;
1506   } else if (iid == vmIntrinsics::_linkToNative) {
1507     member_arg_pos = method->size_of_parameters() - 1;  // trailing NativeEntryPoint argument
1508     member_reg = r19;  // known to be free at this point
1509   } else {
1510     fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1511   }
1512 
1513   if (member_reg != noreg) {
1514     // Load the member_arg into register, if necessary.
1515     SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1516     VMReg r = regs[member_arg_pos].first();
1517     if (r->is_stack()) {
1518       __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1519     } else {
1520       // no data motion is needed
1521       member_reg = r->as_Register();
1522     }
1523   }
1524 
1525   if (has_receiver) {
1526     // Make sure the receiver is loaded into a register.
1527     assert(method->size_of_parameters() > 0, "oob");
1528     assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1529     VMReg r = regs[0].first();
1530     assert(r->is_valid(), "bad receiver arg");
1531     if (r->is_stack()) {
1532       // Porting note:  This assumes that compiled calling conventions always
1533       // pass the receiver oop in a register.  If this is not true on some
1534       // platform, pick a temp and load the receiver from stack.
1535       fatal("receiver always in a register");
1536       receiver_reg = r2;  // known to be free at this point
1537       __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1538     } else {
1539       // no data motion is needed
1540       receiver_reg = r->as_Register();
1541     }
1542   }
1543 
1544   // Figure out which address we are really jumping to:
1545   MethodHandles::generate_method_handle_dispatch(masm, iid,
1546                                                  receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1547 }
1548 
1549 // ---------------------------------------------------------------------------
1550 // Generate a native wrapper for a given method.  The method takes arguments
1551 // in the Java compiled code convention, marshals them to the native
1552 // convention (handlizes oops, etc), transitions to native, makes the call,
1553 // returns to java state (possibly blocking), unhandlizes any result and
1554 // returns.
1555 //
1556 // Critical native functions are a shorthand for the use of
1557 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1558 // functions.  The wrapper is expected to unpack the arguments before
1559 // passing them to the callee. Critical native functions leave the state _in_Java,
1560 // since they block out GC.
1561 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1562 // block and the check for pending exceptions it's impossible for them
1563 // to be thrown.
1564 //
1565 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1566                                                 const methodHandle& method,
1567                                                 int compile_id,
1568                                                 BasicType* in_sig_bt,
1569                                                 VMRegPair* in_regs,
1570                                                 BasicType ret_type) {
1571   if (method->is_continuation_native_intrinsic()) {
1572     int exception_offset = -1;
1573     OopMapSet* oop_maps = new OopMapSet();
1574     int frame_complete = -1;
1575     int stack_slots = -1;
1576     int interpreted_entry_offset = -1;
1577     int vep_offset = -1;
1578     if (method->is_continuation_enter_intrinsic()) {
1579       gen_continuation_enter(masm,
1580                              method,
1581                              in_sig_bt,
1582                              in_regs,
1583                              exception_offset,
1584                              oop_maps,
1585                              frame_complete,
1586                              stack_slots,
1587                              interpreted_entry_offset,
1588                              vep_offset);
1589     } else if (method->is_continuation_yield_intrinsic()) {
1590       gen_continuation_yield(masm,
1591                              method,
1592                              in_sig_bt,
1593                              in_regs,
1594                              oop_maps,
1595                              frame_complete,
1596                              stack_slots,
1597                              vep_offset);
1598     } else {
1599       guarantee(false, "Unknown Continuation native intrinsic");
1600     }
1601 
1602 #ifdef ASSERT
1603     if (method->is_continuation_enter_intrinsic()) {
1604       assert(interpreted_entry_offset != -1, "Must be set");
1605       assert(exception_offset != -1,         "Must be set");
1606     } else {
1607       assert(interpreted_entry_offset == -1, "Must be unset");
1608       assert(exception_offset == -1,         "Must be unset");
1609     }
1610     assert(frame_complete != -1,    "Must be set");
1611     assert(stack_slots != -1,       "Must be set");
1612     assert(vep_offset != -1,        "Must be set");
1613 #endif
1614 
1615     __ flush();
1616     nmethod* nm = nmethod::new_native_nmethod(method,
1617                                               compile_id,
1618                                               masm->code(),
1619                                               vep_offset,
1620                                               frame_complete,
1621                                               stack_slots,
1622                                               in_ByteSize(-1),
1623                                               in_ByteSize(-1),
1624                                               oop_maps,
1625                                               exception_offset);
1626     if (nm == nullptr) return nm;
1627     if (method->is_continuation_enter_intrinsic()) {
1628       ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1629     } else if (method->is_continuation_yield_intrinsic()) {
1630       _cont_doYield_stub = nm;
1631     } else {
1632       guarantee(false, "Unknown Continuation native intrinsic");
1633     }
1634     return nm;
1635   }
1636 
1637   if (method->is_method_handle_intrinsic()) {
1638     vmIntrinsics::ID iid = method->intrinsic_id();
1639     intptr_t start = (intptr_t)__ pc();
1640     int vep_offset = ((intptr_t)__ pc()) - start;
1641 
1642     // First instruction must be a nop as it may need to be patched on deoptimisation
1643     __ nop();
1644     gen_special_dispatch(masm,
1645                          method,
1646                          in_sig_bt,
1647                          in_regs);
1648     int frame_complete = ((intptr_t)__ pc()) - start;  // not complete, period
1649     __ flush();
1650     int stack_slots = SharedRuntime::out_preserve_stack_slots();  // no out slots at all, actually
1651     return nmethod::new_native_nmethod(method,
1652                                        compile_id,
1653                                        masm->code(),
1654                                        vep_offset,
1655                                        frame_complete,
1656                                        stack_slots / VMRegImpl::slots_per_word,
1657                                        in_ByteSize(-1),
1658                                        in_ByteSize(-1),
1659                                        nullptr);
1660   }
1661   address native_func = method->native_function();
1662   assert(native_func != nullptr, "must have function");
1663 
1664   // An OopMap for lock (and class if static)
1665   OopMapSet *oop_maps = new OopMapSet();
1666   intptr_t start = (intptr_t)__ pc();
1667 
1668   // We have received a description of where all the java arg are located
1669   // on entry to the wrapper. We need to convert these args to where
1670   // the jni function will expect them. To figure out where they go
1671   // we convert the java signature to a C signature by inserting
1672   // the hidden arguments as arg[0] and possibly arg[1] (static method)
1673 
1674   const int total_in_args = method->size_of_parameters();
1675   int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1676 
1677   BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1678   VMRegPair* out_regs   = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1679 
1680   int argc = 0;
1681   out_sig_bt[argc++] = T_ADDRESS;
1682   if (method->is_static()) {
1683     out_sig_bt[argc++] = T_OBJECT;
1684   }
1685 
1686   for (int i = 0; i < total_in_args ; i++ ) {
1687     out_sig_bt[argc++] = in_sig_bt[i];
1688   }
1689 
1690   // Now figure out where the args must be stored and how much stack space
1691   // they require.
1692   int out_arg_slots;
1693   out_arg_slots = c_calling_convention_priv(out_sig_bt, out_regs, total_c_args);
1694 
1695   if (out_arg_slots < 0) {
1696     return nullptr;
1697   }
1698 
1699   // Compute framesize for the wrapper.  We need to handlize all oops in
1700   // incoming registers
1701 
1702   // Calculate the total number of stack slots we will need.
1703 
1704   // First count the abi requirement plus all of the outgoing args
1705   int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1706 
1707   // Now the space for the inbound oop handle area
1708   int total_save_slots = 8 * VMRegImpl::slots_per_word;  // 8 arguments passed in registers
1709 
1710   int oop_handle_offset = stack_slots;
1711   stack_slots += total_save_slots;
1712 
1713   // Now any space we need for handlizing a klass if static method
1714 
1715   int klass_slot_offset = 0;
1716   int klass_offset = -1;
1717   int lock_slot_offset = 0;
1718   bool is_static = false;
1719 
1720   if (method->is_static()) {
1721     klass_slot_offset = stack_slots;
1722     stack_slots += VMRegImpl::slots_per_word;
1723     klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1724     is_static = true;
1725   }
1726 
1727   // Plus a lock if needed
1728 
1729   if (method->is_synchronized()) {
1730     lock_slot_offset = stack_slots;
1731     stack_slots += VMRegImpl::slots_per_word;
1732   }
1733 
1734   // Now a place (+2) to save return values or temp during shuffling
1735   // + 4 for return address (which we own) and saved rfp
1736   stack_slots += 6;
1737 
1738   // Ok The space we have allocated will look like:
1739   //
1740   //
1741   // FP-> |                     |
1742   //      |---------------------|
1743   //      | 2 slots for moves   |
1744   //      |---------------------|
1745   //      | lock box (if sync)  |
1746   //      |---------------------| <- lock_slot_offset
1747   //      | klass (if static)   |
1748   //      |---------------------| <- klass_slot_offset
1749   //      | oopHandle area      |
1750   //      |---------------------| <- oop_handle_offset (8 java arg registers)
1751   //      | outbound memory     |
1752   //      | based arguments     |
1753   //      |                     |
1754   //      |---------------------|
1755   //      |                     |
1756   // SP-> | out_preserved_slots |
1757   //
1758   //
1759 
1760 
1761   // Now compute actual number of stack words we need rounding to make
1762   // stack properly aligned.
1763   stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1764 
1765   int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1766 
1767   // First thing make an ic check to see if we should even be here
1768 
1769   // We are free to use all registers as temps without saving them and
1770   // restoring them except rfp. rfp is the only callee save register
1771   // as far as the interpreter and the compiler(s) are concerned.
1772 
1773   const Register receiver = j_rarg0;
1774 
1775   Label exception_pending;
1776 
1777   assert_different_registers(receiver, rscratch1);
1778   __ verify_oop(receiver);
1779   __ ic_check(8 /* end_alignment */);
1780 
1781   // Verified entry point must be aligned
1782   int vep_offset = ((intptr_t)__ pc()) - start;
1783 
1784   // If we have to make this method not-entrant we'll overwrite its
1785   // first instruction with a jump.  For this action to be legal we
1786   // must ensure that this first instruction is a B, BL, NOP, BKPT,
1787   // SVC, HVC, or SMC.  Make it a NOP.
1788   __ nop();
1789 
1790   if (method->needs_clinit_barrier()) {
1791     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
1792     Label L_skip_barrier;
1793     __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass*
1794     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
1795     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1796 
1797     __ bind(L_skip_barrier);
1798   }
1799 
1800   // Generate stack overflow check
1801   __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1802 
1803   // Generate a new frame for the wrapper.
1804   __ enter();
1805   // -2 because return address is already present and so is saved rfp
1806   __ sub(sp, sp, stack_size - 2*wordSize);
1807 
1808   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1809   bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1810 
1811   // Frame is now completed as far as size and linkage.
1812   int frame_complete = ((intptr_t)__ pc()) - start;
1813 
1814   // We use r20 as the oop handle for the receiver/klass
1815   // It is callee save so it survives the call to native
1816 
1817   const Register oop_handle_reg = r20;
1818 
1819   //
1820   // We immediately shuffle the arguments so that any vm call we have to
1821   // make from here on out (sync slow path, jvmti, etc.) we will have
1822   // captured the oops from our caller and have a valid oopMap for
1823   // them.
1824 
1825   // -----------------
1826   // The Grand Shuffle
1827 
1828   // The Java calling convention is either equal (linux) or denser (win64) than the
1829   // c calling convention. However the because of the jni_env argument the c calling
1830   // convention always has at least one more (and two for static) arguments than Java.
1831   // Therefore if we move the args from java -> c backwards then we will never have
1832   // a register->register conflict and we don't have to build a dependency graph
1833   // and figure out how to break any cycles.
1834   //
1835 
1836   // Record esp-based slot for receiver on stack for non-static methods
1837   int receiver_offset = -1;
1838 
1839   // This is a trick. We double the stack slots so we can claim
1840   // the oops in the caller's frame. Since we are sure to have
1841   // more args than the caller doubling is enough to make
1842   // sure we can capture all the incoming oop args from the
1843   // caller.
1844   //
1845   OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1846 
1847   // Mark location of rfp (someday)
1848   // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp));
1849 
1850 
1851   int float_args = 0;
1852   int int_args = 0;
1853 
1854 #ifdef ASSERT
1855   bool reg_destroyed[Register::number_of_registers];
1856   bool freg_destroyed[FloatRegister::number_of_registers];
1857   for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1858     reg_destroyed[r] = false;
1859   }
1860   for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1861     freg_destroyed[f] = false;
1862   }
1863 
1864 #endif /* ASSERT */
1865 
1866   // For JNI natives the incoming and outgoing registers are offset upwards.
1867   GrowableArray<int> arg_order(2 * total_in_args);
1868 
1869   for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1870     arg_order.push(i);
1871     arg_order.push(c_arg);
1872   }
1873 
1874   for (int ai = 0; ai < arg_order.length(); ai += 2) {
1875     int i = arg_order.at(ai);
1876     int c_arg = arg_order.at(ai + 1);
1877     __ block_comment(err_msg("move %d -> %d", i, c_arg));
1878     assert(c_arg != -1 && i != -1, "wrong order");
1879 #ifdef ASSERT
1880     if (in_regs[i].first()->is_Register()) {
1881       assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1882     } else if (in_regs[i].first()->is_FloatRegister()) {
1883       assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1884     }
1885     if (out_regs[c_arg].first()->is_Register()) {
1886       reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1887     } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1888       freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1889     }
1890 #endif /* ASSERT */
1891     switch (in_sig_bt[i]) {
1892       case T_ARRAY:
1893       case T_OBJECT:
1894         __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1895                        ((i == 0) && (!is_static)),
1896                        &receiver_offset);
1897         int_args++;
1898         break;
1899       case T_VOID:
1900         break;
1901 
1902       case T_FLOAT:
1903         __ float_move(in_regs[i], out_regs[c_arg]);
1904         float_args++;
1905         break;
1906 
1907       case T_DOUBLE:
1908         assert( i + 1 < total_in_args &&
1909                 in_sig_bt[i + 1] == T_VOID &&
1910                 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1911         __ double_move(in_regs[i], out_regs[c_arg]);
1912         float_args++;
1913         break;
1914 
1915       case T_LONG :
1916         __ long_move(in_regs[i], out_regs[c_arg]);
1917         int_args++;
1918         break;
1919 
1920       case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1921 
1922       default:
1923         __ move32_64(in_regs[i], out_regs[c_arg]);
1924         int_args++;
1925     }
1926   }
1927 
1928   // point c_arg at the first arg that is already loaded in case we
1929   // need to spill before we call out
1930   int c_arg = total_c_args - total_in_args;
1931 
1932   // Pre-load a static method's oop into c_rarg1.
1933   if (method->is_static()) {
1934 
1935     //  load oop into a register
1936     __ movoop(c_rarg1,
1937               JNIHandles::make_local(method->method_holder()->java_mirror()));
1938 
1939     // Now handlize the static class mirror it's known not-null.
1940     __ str(c_rarg1, Address(sp, klass_offset));
1941     map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1942 
1943     // Now get the handle
1944     __ lea(c_rarg1, Address(sp, klass_offset));
1945     // and protect the arg if we must spill
1946     c_arg--;
1947   }
1948 
1949   // Change state to native (we save the return address in the thread, since it might not
1950   // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1951   // points into the right code segment. It does not have to be the correct return pc.
1952   // We use the same pc/oopMap repeatedly when we call out.
1953 
1954   Label native_return;
1955   if (method->is_object_wait0()) {
1956     // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1957     __ set_last_Java_frame(sp, noreg, native_return, rscratch1);
1958   } else {
1959     intptr_t the_pc = (intptr_t) __ pc();
1960     oop_maps->add_gc_map(the_pc - start, map);
1961 
1962     __ set_last_Java_frame(sp, noreg, __ pc(), rscratch1);
1963   }
1964 
1965   Label dtrace_method_entry, dtrace_method_entry_done;
1966   if (DTraceMethodProbes) {
1967     __ b(dtrace_method_entry);
1968     __ bind(dtrace_method_entry_done);
1969   }
1970 
1971   // RedefineClasses() tracing support for obsolete method entry
1972   if (log_is_enabled(Trace, redefine, class, obsolete)) {
1973     // protect the args we've loaded
1974     save_args(masm, total_c_args, c_arg, out_regs);
1975     __ mov_metadata(c_rarg1, method());
1976     __ call_VM_leaf(
1977       CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1978       rthread, c_rarg1);
1979     restore_args(masm, total_c_args, c_arg, out_regs);
1980   }
1981 
1982   // Lock a synchronized method
1983 
1984   // Register definitions used by locking and unlocking
1985 
1986   const Register swap_reg = r0;
1987   const Register obj_reg  = r19;  // Will contain the oop
1988   const Register lock_reg = r13;  // Address of compiler lock object (BasicLock)
1989   const Register old_hdr  = r13;  // value of old header at unlock time
1990   const Register lock_tmp = r14;  // Temporary used by fast_lock/unlock
1991   const Register tmp = lr;
1992 
1993   Label slow_path_lock;
1994   Label lock_done;
1995 
1996   if (method->is_synchronized()) {
1997     // Get the handle (the 2nd argument)
1998     __ mov(oop_handle_reg, c_rarg1);
1999 
2000     // Get address of the box
2001 
2002     __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2003 
2004     // Load the oop from the handle
2005     __ ldr(obj_reg, Address(oop_handle_reg, 0));
2006 
2007     __ fast_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
2008 
2009     // Slow path will re-enter here
2010     __ bind(lock_done);
2011   }
2012 
2013 
2014   // Finally just about ready to make the JNI call
2015 
2016   // get JNIEnv* which is first argument to native
2017   __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset())));
2018 
2019   // Now set thread in native
2020   __ mov(rscratch1, _thread_in_native);
2021   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2022   __ stlrw(rscratch1, rscratch2);
2023 
2024   __ rt_call(native_func);
2025 
2026   // Verify or restore cpu control state after JNI call
2027   __ restore_cpu_control_state_after_jni(rscratch1, rscratch2);
2028 
2029   // Unpack native results.
2030   switch (ret_type) {
2031   case T_BOOLEAN: __ c2bool(r0);                     break;
2032   case T_CHAR   : __ ubfx(r0, r0, 0, 16);            break;
2033   case T_BYTE   : __ sbfx(r0, r0, 0, 8);             break;
2034   case T_SHORT  : __ sbfx(r0, r0, 0, 16);            break;
2035   case T_INT    : __ sbfx(r0, r0, 0, 32);            break;
2036   case T_DOUBLE :
2037   case T_FLOAT  :
2038     // Result is in v0 we'll save as needed
2039     break;
2040   case T_ARRAY:                 // Really a handle
2041   case T_OBJECT:                // Really a handle
2042       break; // can't de-handlize until after safepoint check
2043   case T_VOID: break;
2044   case T_LONG: break;
2045   default       : ShouldNotReachHere();
2046   }
2047 
2048   Label safepoint_in_progress, safepoint_in_progress_done;
2049 
2050   // Switch thread to "native transition" state before reading the synchronization state.
2051   // This additional state is necessary because reading and testing the synchronization
2052   // state is not atomic w.r.t. GC, as this scenario demonstrates:
2053   //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2054   //     VM thread changes sync state to synchronizing and suspends threads for GC.
2055   //     Thread A is resumed to finish this native method, but doesn't block here since it
2056   //     didn't see any synchronization is progress, and escapes.
2057   __ mov(rscratch1, _thread_in_native_trans);
2058 
2059   __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset()));
2060 
2061   // Force this write out before the read below
2062   if (!UseSystemMemoryBarrier) {
2063     __ dmb(Assembler::ISH);
2064   }
2065 
2066   __ verify_sve_vector_length();
2067 
2068   // Check for safepoint operation in progress and/or pending suspend requests.
2069   {
2070     // No need for acquire as Java threads always disarm themselves.
2071     __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
2072     __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset()));
2073     __ cbnzw(rscratch1, safepoint_in_progress);
2074     __ bind(safepoint_in_progress_done);
2075   }
2076 
2077   // change thread state
2078   __ mov(rscratch1, _thread_in_Java);
2079   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2080   __ stlrw(rscratch1, rscratch2);
2081 
2082   if (method->is_object_wait0()) {
2083     // Check preemption for Object.wait()
2084     __ ldr(rscratch1, Address(rthread, JavaThread::preempt_alternate_return_offset()));
2085     __ cbz(rscratch1, native_return);
2086     __ str(zr, Address(rthread, JavaThread::preempt_alternate_return_offset()));
2087     __ br(rscratch1);
2088     __ bind(native_return);
2089 
2090     intptr_t the_pc = (intptr_t) __ pc();
2091     oop_maps->add_gc_map(the_pc - start, map);
2092   }
2093 
2094   Label reguard;
2095   Label reguard_done;
2096   __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset()));
2097   __ cmpw(rscratch1, StackOverflow::stack_guard_yellow_reserved_disabled);
2098   __ br(Assembler::EQ, reguard);
2099   __ bind(reguard_done);
2100 
2101   // native result if any is live
2102 
2103   // Unlock
2104   Label unlock_done;
2105   Label slow_path_unlock;
2106   if (method->is_synchronized()) {
2107 
2108     // Get locked oop from the handle we passed to jni
2109     __ ldr(obj_reg, Address(oop_handle_reg, 0));
2110 
2111     // Must save r0 if if it is live now because cmpxchg must use it
2112     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2113       save_native_result(masm, ret_type, stack_slots);
2114     }
2115 
2116     __ fast_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
2117 
2118     // slow path re-enters here
2119     __ bind(unlock_done);
2120     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2121       restore_native_result(masm, ret_type, stack_slots);
2122     }
2123   }
2124 
2125   Label dtrace_method_exit, dtrace_method_exit_done;
2126   if (DTraceMethodProbes) {
2127     __ b(dtrace_method_exit);
2128     __ bind(dtrace_method_exit_done);
2129   }
2130 
2131   __ reset_last_Java_frame(false);
2132 
2133   // Unbox oop result, e.g. JNIHandles::resolve result.
2134   if (is_reference_type(ret_type)) {
2135     __ resolve_jobject(r0, r1, r2);
2136   }
2137 
2138   if (CheckJNICalls) {
2139     // clear_pending_jni_exception_check
2140     __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset()));
2141   }
2142 
2143   // reset handle block
2144   __ ldr(r2, Address(rthread, JavaThread::active_handles_offset()));
2145   __ str(zr, Address(r2, JNIHandleBlock::top_offset()));
2146 
2147   __ leave();
2148 
2149   #if INCLUDE_JFR
2150   // We need to do a poll test after unwind in case the sampler
2151   // managed to sample the native frame after returning to Java.
2152   Label L_return;
2153   __ ldr(rscratch1, Address(rthread, JavaThread::polling_word_offset()));
2154   address poll_test_pc = __ pc();
2155   __ relocate(relocInfo::poll_return_type);
2156   __ tbz(rscratch1, log2i_exact(SafepointMechanism::poll_bit()), L_return);
2157   assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
2158     "polling page return stub not created yet");
2159   address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
2160   __ adr(rscratch1, InternalAddress(poll_test_pc));
2161   __ str(rscratch1, Address(rthread, JavaThread::saved_exception_pc_offset()));
2162   __ far_jump(RuntimeAddress(stub));
2163   __ bind(L_return);
2164 #endif // INCLUDE_JFR
2165 
2166   // Any exception pending?
2167   __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2168   __ cbnz(rscratch1, exception_pending);
2169 
2170   // We're done
2171   __ ret(lr);
2172 
2173   // Unexpected paths are out of line and go here
2174 
2175   // forward the exception
2176   __ bind(exception_pending);
2177 
2178   // and forward the exception
2179   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2180 
2181   // Slow path locking & unlocking
2182   if (method->is_synchronized()) {
2183 
2184     __ block_comment("Slow path lock {");
2185     __ bind(slow_path_lock);
2186 
2187     // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2188     // args are (oop obj, BasicLock* lock, JavaThread* thread)
2189 
2190     // protect the args we've loaded
2191     save_args(masm, total_c_args, c_arg, out_regs);
2192 
2193     __ mov(c_rarg0, obj_reg);
2194     __ mov(c_rarg1, lock_reg);
2195     __ mov(c_rarg2, rthread);
2196 
2197     // Not a leaf but we have last_Java_frame setup as we want.
2198     // We don't want to unmount in case of contention since that would complicate preserving
2199     // the arguments that had already been marshalled into the native convention. So we force
2200     // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
2201     // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
2202     __ push_cont_fastpath();
2203     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
2204     __ pop_cont_fastpath();
2205     restore_args(masm, total_c_args, c_arg, out_regs);
2206 
2207 #ifdef ASSERT
2208     { Label L;
2209       __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2210       __ cbz(rscratch1, L);
2211       __ stop("no pending exception allowed on exit from monitorenter");
2212       __ bind(L);
2213     }
2214 #endif
2215     __ b(lock_done);
2216 
2217     __ block_comment("} Slow path lock");
2218 
2219     __ block_comment("Slow path unlock {");
2220     __ bind(slow_path_unlock);
2221 
2222     // If we haven't already saved the native result we must save it now as xmm registers
2223     // are still exposed.
2224 
2225     if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2226       save_native_result(masm, ret_type, stack_slots);
2227     }
2228 
2229     __ mov(c_rarg2, rthread);
2230     __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2231     __ mov(c_rarg0, obj_reg);
2232 
2233     // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2234     // NOTE that obj_reg == r19 currently
2235     __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2236     __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2237 
2238     __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2239 
2240 #ifdef ASSERT
2241     {
2242       Label L;
2243       __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2244       __ cbz(rscratch1, L);
2245       __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2246       __ bind(L);
2247     }
2248 #endif /* ASSERT */
2249 
2250     __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2251 
2252     if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2253       restore_native_result(masm, ret_type, stack_slots);
2254     }
2255     __ b(unlock_done);
2256 
2257     __ block_comment("} Slow path unlock");
2258 
2259   } // synchronized
2260 
2261   // SLOW PATH Reguard the stack if needed
2262 
2263   __ bind(reguard);
2264   save_native_result(masm, ret_type, stack_slots);
2265   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2266   restore_native_result(masm, ret_type, stack_slots);
2267   // and continue
2268   __ b(reguard_done);
2269 
2270   // SLOW PATH safepoint
2271   {
2272     __ block_comment("safepoint {");
2273     __ bind(safepoint_in_progress);
2274 
2275     // Don't use call_VM as it will see a possible pending exception and forward it
2276     // and never return here preventing us from clearing _last_native_pc down below.
2277     //
2278     save_native_result(masm, ret_type, stack_slots);
2279     __ mov(c_rarg0, rthread);
2280 #ifndef PRODUCT
2281   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2282 #endif
2283     __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)));
2284     __ blr(rscratch1);
2285 
2286     // Restore any method result value
2287     restore_native_result(masm, ret_type, stack_slots);
2288 
2289     __ b(safepoint_in_progress_done);
2290     __ block_comment("} safepoint");
2291   }
2292 
2293   // SLOW PATH dtrace support
2294   if (DTraceMethodProbes) {
2295     {
2296       __ block_comment("dtrace entry {");
2297       __ bind(dtrace_method_entry);
2298 
2299       // We have all of the arguments setup at this point. We must not touch any register
2300       // argument registers at this point (what if we save/restore them there are no oop?
2301 
2302       save_args(masm, total_c_args, c_arg, out_regs);
2303       __ mov_metadata(c_rarg1, method());
2304       __ call_VM_leaf(
2305         CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2306         rthread, c_rarg1);
2307       restore_args(masm, total_c_args, c_arg, out_regs);
2308       __ b(dtrace_method_entry_done);
2309       __ block_comment("} dtrace entry");
2310     }
2311 
2312     {
2313       __ block_comment("dtrace exit {");
2314       __ bind(dtrace_method_exit);
2315       save_native_result(masm, ret_type, stack_slots);
2316       __ mov_metadata(c_rarg1, method());
2317       __ call_VM_leaf(
2318         CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2319         rthread, c_rarg1);
2320       restore_native_result(masm, ret_type, stack_slots);
2321       __ b(dtrace_method_exit_done);
2322       __ block_comment("} dtrace exit");
2323     }
2324   }
2325 
2326   __ flush();
2327 
2328   nmethod *nm = nmethod::new_native_nmethod(method,
2329                                             compile_id,
2330                                             masm->code(),
2331                                             vep_offset,
2332                                             frame_complete,
2333                                             stack_slots / VMRegImpl::slots_per_word,
2334                                             (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2335                                             in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2336                                             oop_maps);
2337 
2338   return nm;
2339 }
2340 
2341 // this function returns the adjust size (in number of words) to a c2i adapter
2342 // activation for use during deoptimization
2343 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2344   assert(callee_locals >= callee_parameters,
2345           "test and remove; got more parms than locals");
2346   if (callee_locals < callee_parameters)
2347     return 0;                   // No adjustment for negative locals
2348   int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2349   // diff is counted in stack words
2350   return align_up(diff, 2);
2351 }
2352 
2353 
2354 //------------------------------generate_deopt_blob----------------------------
2355 void SharedRuntime::generate_deopt_blob() {
2356   // Allocate space for the code
2357   ResourceMark rm;
2358   // Setup code generation tools
2359   int pad = 0;
2360   const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2361   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, BlobId::shared_deopt_id);
2362   if (blob != nullptr) {
2363     _deopt_blob = blob->as_deoptimization_blob();
2364     return;
2365   }
2366 
2367   CodeBuffer buffer(name, 2048+pad, 1024);
2368   MacroAssembler* masm = new MacroAssembler(&buffer);
2369   int frame_size_in_words;
2370   OopMap* map = nullptr;
2371   OopMapSet *oop_maps = new OopMapSet();
2372   RegisterSaver reg_save(COMPILER2_PRESENT(true) NOT_COMPILER2(false));
2373 
2374   // -------------
2375   // This code enters when returning to a de-optimized nmethod.  A return
2376   // address has been pushed on the stack, and return values are in
2377   // registers.
2378   // If we are doing a normal deopt then we were called from the patched
2379   // nmethod from the point we returned to the nmethod. So the return
2380   // address on the stack is wrong by NativeCall::instruction_size
2381   // We will adjust the value so it looks like we have the original return
2382   // address on the stack (like when we eagerly deoptimized).
2383   // In the case of an exception pending when deoptimizing, we enter
2384   // with a return address on the stack that points after the call we patched
2385   // into the exception handler. We have the following register state from,
2386   // e.g., the forward exception stub (see stubGenerator_x86_64.cpp).
2387   //    r0: exception oop
2388   //    r19: exception handler
2389   //    r3: throwing pc
2390   // So in this case we simply jam r3 into the useless return address and
2391   // the stack looks just like we want.
2392   //
2393   // At this point we need to de-opt.  We save the argument return
2394   // registers.  We call the first C routine, fetch_unroll_info().  This
2395   // routine captures the return values and returns a structure which
2396   // describes the current frame size and the sizes of all replacement frames.
2397   // The current frame is compiled code and may contain many inlined
2398   // functions, each with their own JVM state.  We pop the current frame, then
2399   // push all the new frames.  Then we call the C routine unpack_frames() to
2400   // populate these frames.  Finally unpack_frames() returns us the new target
2401   // address.  Notice that callee-save registers are BLOWN here; they have
2402   // already been captured in the vframeArray at the time the return PC was
2403   // patched.
2404   address start = __ pc();
2405   Label cont;
2406 
2407   // Prolog for non exception case!
2408 
2409   // Save everything in sight.
2410   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2411 
2412   // Normal deoptimization.  Save exec mode for unpack_frames.
2413   __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved
2414   __ b(cont);
2415 
2416   int reexecute_offset = __ pc() - start;
2417   // Reexecute case
2418   // return address is the pc describes what bci to do re-execute at
2419 
2420   // No need to update map as each call to save_live_registers will produce identical oopmap
2421   (void) reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2422 
2423   __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved
2424   __ b(cont);
2425 
2426   int exception_offset = __ pc() - start;
2427 
2428   // Prolog for exception case
2429 
2430   // all registers are dead at this entry point, except for r0, and
2431   // r3 which contain the exception oop and exception pc
2432   // respectively.  Set them in TLS and fall thru to the
2433   // unpack_with_exception_in_tls entry point.
2434 
2435   __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
2436   __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
2437 
2438   int exception_in_tls_offset = __ pc() - start;
2439 
2440   // new implementation because exception oop is now passed in JavaThread
2441 
2442   // Prolog for exception case
2443   // All registers must be preserved because they might be used by LinearScan
2444   // Exceptiop oop and throwing PC are passed in JavaThread
2445   // tos: stack at point of call to method that threw the exception (i.e. only
2446   // args are on the stack, no return address)
2447 
2448   // The return address pushed by save_live_registers will be patched
2449   // later with the throwing pc. The correct value is not available
2450   // now because loading it from memory would destroy registers.
2451 
2452   // NB: The SP at this point must be the SP of the method that is
2453   // being deoptimized.  Deoptimization assumes that the frame created
2454   // here by save_live_registers is immediately below the method's SP.
2455   // This is a somewhat fragile mechanism.
2456 
2457   // Save everything in sight.
2458   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2459 
2460   // Now it is safe to overwrite any register
2461 
2462   // Deopt during an exception.  Save exec mode for unpack_frames.
2463   __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved
2464 
2465   // load throwing pc from JavaThread and patch it as the return address
2466   // of the current frame. Then clear the field in JavaThread
2467   __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2468   __ protect_return_address(r3);
2469   __ str(r3, Address(rfp, wordSize));
2470   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2471 
2472 #ifdef ASSERT
2473   // verify that there is really an exception oop in JavaThread
2474   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2475   __ verify_oop(r0);
2476 
2477   // verify that there is no pending exception
2478   Label no_pending_exception;
2479   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2480   __ cbz(rscratch1, no_pending_exception);
2481   __ stop("must not have pending exception here");
2482   __ bind(no_pending_exception);
2483 #endif
2484 
2485   __ bind(cont);
2486 
2487   // Call C code.  Need thread and this frame, but NOT official VM entry
2488   // crud.  We cannot block on this call, no GC can happen.
2489   //
2490   // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2491 
2492   // fetch_unroll_info needs to call last_java_frame().
2493 
2494   Label retaddr;
2495   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2496 #ifdef ASSERT
2497   { Label L;
2498     __ ldr(rscratch1, Address(rthread, JavaThread::last_Java_fp_offset()));
2499     __ cbz(rscratch1, L);
2500     __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2501     __ bind(L);
2502   }
2503 #endif // ASSERT
2504   __ mov(c_rarg0, rthread);
2505   __ mov(c_rarg1, rcpool);
2506   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2507   __ blr(rscratch1);
2508   __ bind(retaddr);
2509 
2510   // Need to have an oopmap that tells fetch_unroll_info where to
2511   // find any register it might need.
2512   oop_maps->add_gc_map(__ pc() - start, map);
2513 
2514   __ reset_last_Java_frame(false);
2515 
2516   // Load UnrollBlock* into r5
2517   __ mov(r5, r0);
2518 
2519   __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset()));
2520    Label noException;
2521   __ cmpw(rcpool, Deoptimization::Unpack_exception);   // Was exception pending?
2522   __ br(Assembler::NE, noException);
2523   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2524   // QQQ this is useless it was null above
2525   __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2526   __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
2527   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2528 
2529   __ verify_oop(r0);
2530 
2531   // Overwrite the result registers with the exception results.
2532   __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2533   // I think this is useless
2534   // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2535 
2536   __ bind(noException);
2537 
2538   // Only register save data is on the stack.
2539   // Now restore the result registers.  Everything else is either dead
2540   // or captured in the vframeArray.
2541 
2542   // Restore fp result register
2543   __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2544   // Restore integer result register
2545   __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2546 
2547   // Pop all of the register save area off the stack
2548   __ add(sp, sp, frame_size_in_words * wordSize);
2549 
2550   // All of the register save area has been popped of the stack. Only the
2551   // return address remains.
2552 
2553   // Pop all the frames we must move/replace.
2554   //
2555   // Frame picture (youngest to oldest)
2556   // 1: self-frame (no frame link)
2557   // 2: deopting frame  (no frame link)
2558   // 3: caller of deopting frame (could be compiled/interpreted).
2559   //
2560   // Note: by leaving the return address of self-frame on the stack
2561   // and using the size of frame 2 to adjust the stack
2562   // when we are done the return to frame 3 will still be on the stack.
2563 
2564   // Pop deoptimized frame
2565   __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2566   __ sub(r2, r2, 2 * wordSize);
2567   __ add(sp, sp, r2);
2568   __ ldp(rfp, zr, __ post(sp, 2 * wordSize));
2569 
2570 #ifdef ASSERT
2571   // Compilers generate code that bang the stack by as much as the
2572   // interpreter would need. So this stack banging should never
2573   // trigger a fault. Verify that it does not on non product builds.
2574   __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2575   __ bang_stack_size(r19, r2);
2576 #endif
2577   // Load address of array of frame pcs into r2
2578   __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset()));
2579 
2580   // Trash the old pc
2581   // __ addptr(sp, wordSize);  FIXME ????
2582 
2583   // Load address of array of frame sizes into r4
2584   __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset()));
2585 
2586   // Load counter into r3
2587   __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset()));
2588 
2589   // Now adjust the caller's stack to make up for the extra locals
2590   // but record the original sp so that we can save it in the skeletal interpreter
2591   // frame and the stack walking of interpreter_sender will get the unextended sp
2592   // value and not the "real" sp value.
2593 
2594   const Register sender_sp = r6;
2595 
2596   __ mov(sender_sp, sp);
2597   __ ldrw(r19, Address(r5,
2598                        Deoptimization::UnrollBlock::
2599                        caller_adjustment_offset()));
2600   __ sub(sp, sp, r19);
2601 
2602   // Push interpreter frames in a loop
2603   __ mov(rscratch1, (uint64_t)0xDEADDEAD);        // Make a recognizable pattern
2604   __ mov(rscratch2, rscratch1);
2605   Label loop;
2606   __ bind(loop);
2607   __ ldr(r19, Address(__ post(r4, wordSize)));          // Load frame size
2608   __ sub(r19, r19, 2*wordSize);           // We'll push pc and fp by hand
2609   __ ldr(lr, Address(__ post(r2, wordSize)));  // Load pc
2610   __ enter();                           // Save old & set new fp
2611   __ sub(sp, sp, r19);                  // Prolog
2612   // This value is corrected by layout_activation_impl
2613   __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2614   __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2615   __ mov(sender_sp, sp);               // Pass sender_sp to next frame
2616   __ sub(r3, r3, 1);                   // Decrement counter
2617   __ cbnz(r3, loop);
2618 
2619     // Re-push self-frame
2620   __ ldr(lr, Address(r2));
2621   __ enter();
2622 
2623   // Allocate a full sized register save area.  We subtract 2 because
2624   // enter() just pushed 2 words
2625   __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2626 
2627   // Restore frame locals after moving the frame
2628   __ strd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2629   __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2630 
2631   // Call C code.  Need thread but NOT official VM entry
2632   // crud.  We cannot block on this call, no GC can happen.  Call should
2633   // restore return values to their stack-slots with the new SP.
2634   //
2635   // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2636 
2637   // Use rfp because the frames look interpreted now
2638   // Don't need the precise return PC here, just precise enough to point into this code blob.
2639   address the_pc = __ pc();
2640   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
2641 
2642   __ mov(c_rarg0, rthread);
2643   __ movw(c_rarg1, rcpool); // second arg: exec_mode
2644   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2645   __ blr(rscratch1);
2646 
2647   // Set an oopmap for the call site
2648   // Use the same PC we used for the last java frame
2649   oop_maps->add_gc_map(the_pc - start,
2650                        new OopMap( frame_size_in_words, 0 ));
2651 
2652   // Clear fp AND pc
2653   __ reset_last_Java_frame(true);
2654 
2655   // Collect return values
2656   __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2657   __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2658   // I think this is useless (throwing pc?)
2659   // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2660 
2661   // Pop self-frame.
2662   __ leave();                           // Epilog
2663 
2664   // Jump to interpreter
2665   __ ret(lr);
2666 
2667   // Make sure all code is generated
2668   masm->flush();
2669 
2670   _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2671   _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2672 
2673   AOTCodeCache::store_code_blob(*_deopt_blob, AOTCodeEntry::SharedBlob, BlobId::shared_deopt_id);
2674 }
2675 
2676 // Number of stack slots between incoming argument block and the start of
2677 // a new frame.  The PROLOG must add this many slots to the stack.  The
2678 // EPILOG must remove this many slots. aarch64 needs two slots for
2679 // return address and fp.
2680 // TODO think this is correct but check
2681 uint SharedRuntime::in_preserve_stack_slots() {
2682   return 4;
2683 }
2684 
2685 uint SharedRuntime::out_preserve_stack_slots() {
2686   return 0;
2687 }
2688 
2689 
2690 VMReg SharedRuntime::thread_register() {
2691   return rthread->as_VMReg();
2692 }
2693 
2694 //------------------------------generate_handler_blob------
2695 //
2696 // Generate a special Compile2Runtime blob that saves all registers,
2697 // and setup oopmap.
2698 //
2699 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2700   assert(is_polling_page_id(id), "expected a polling page stub id");
2701 
2702   // Allocate space for the code.  Setup code generation tools.
2703   const char* name = SharedRuntime::stub_name(id);
2704   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2705   if (blob != nullptr) {
2706     return blob->as_safepoint_blob();
2707   }
2708 
2709   ResourceMark rm;
2710   OopMapSet *oop_maps = new OopMapSet();
2711   OopMap* map;
2712   CodeBuffer buffer(name, 2048, 1024);
2713   MacroAssembler* masm = new MacroAssembler(&buffer);
2714 
2715   address start   = __ pc();
2716   address call_pc = nullptr;
2717   int frame_size_in_words;
2718   bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2719   RegisterSaver reg_save(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2720 
2721   // When the signal occurred, the LR was either signed and stored on the stack (in which
2722   // case it will be restored from the stack before being used) or unsigned and not stored
2723   // on the stack. Stipping ensures we get the right value.
2724   __ strip_return_address();
2725 
2726   // Save Integer and Float registers.
2727   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2728 
2729   // The following is basically a call_VM.  However, we need the precise
2730   // address of the call in order to generate an oopmap. Hence, we do all the
2731   // work ourselves.
2732 
2733   Label retaddr;
2734   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2735 
2736   // The return address must always be correct so that frame constructor never
2737   // sees an invalid pc.
2738 
2739   if (!cause_return) {
2740     // overwrite the return address pushed by save_live_registers
2741     // Additionally, r20 is a callee-saved register so we can look at
2742     // it later to determine if someone changed the return address for
2743     // us!
2744     __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset()));
2745     __ protect_return_address(r20);
2746     __ str(r20, Address(rfp, wordSize));
2747   }
2748 
2749   // Do the call
2750   __ mov(c_rarg0, rthread);
2751   __ lea(rscratch1, RuntimeAddress(call_ptr));
2752   __ blr(rscratch1);
2753   __ bind(retaddr);
2754 
2755   // Set an oopmap for the call site.  This oopmap will map all
2756   // oop-registers and debug-info registers as callee-saved.  This
2757   // will allow deoptimization at this safepoint to find all possible
2758   // debug-info recordings, as well as let GC find all oops.
2759 
2760   oop_maps->add_gc_map( __ pc() - start, map);
2761 
2762   Label noException;
2763 
2764   __ reset_last_Java_frame(false);
2765 
2766   __ membar(Assembler::LoadLoad | Assembler::LoadStore);
2767 
2768   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2769   __ cbz(rscratch1, noException);
2770 
2771   // Exception pending
2772 
2773   reg_save.restore_live_registers(masm);
2774 
2775   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2776 
2777   // No exception case
2778   __ bind(noException);
2779 
2780   Label no_adjust, bail;
2781   if (!cause_return) {
2782     // If our stashed return pc was modified by the runtime we avoid touching it
2783     __ ldr(rscratch1, Address(rfp, wordSize));
2784     __ cmp(r20, rscratch1);
2785     __ br(Assembler::NE, no_adjust);
2786     __ authenticate_return_address(r20);
2787 
2788 #ifdef ASSERT
2789     // Verify the correct encoding of the poll we're about to skip.
2790     // See NativeInstruction::is_ldrw_to_zr()
2791     __ ldrw(rscratch1, Address(r20));
2792     __ ubfx(rscratch2, rscratch1, 22, 10);
2793     __ cmpw(rscratch2, 0b1011100101);
2794     __ br(Assembler::NE, bail);
2795     __ ubfx(rscratch2, rscratch1, 0, 5);
2796     __ cmpw(rscratch2, 0b11111);
2797     __ br(Assembler::NE, bail);
2798 #endif
2799     // Adjust return pc forward to step over the safepoint poll instruction
2800     __ add(r20, r20, NativeInstruction::instruction_size);
2801     __ protect_return_address(r20);
2802     __ str(r20, Address(rfp, wordSize));
2803   }
2804 
2805   __ bind(no_adjust);
2806   // Normal exit, restore registers and exit.
2807   reg_save.restore_live_registers(masm);
2808 
2809   __ ret(lr);
2810 
2811 #ifdef ASSERT
2812   __ bind(bail);
2813   __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2814 #endif
2815 
2816   // Make sure all code is generated
2817   masm->flush();
2818 
2819   // Fill-out other meta info
2820   SafepointBlob* sp_blob = SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2821 
2822   AOTCodeCache::store_code_blob(*sp_blob, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2823   return sp_blob;
2824 }
2825 
2826 //
2827 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2828 //
2829 // Generate a stub that calls into vm to find out the proper destination
2830 // of a java call. All the argument registers are live at this point
2831 // but since this is generic code we don't know what they are and the caller
2832 // must do any gc of the args.
2833 //
2834 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2835   assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2836   assert(is_resolve_id(id), "expected a resolve stub id");
2837 
2838   const char* name = SharedRuntime::stub_name(id);
2839   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2840   if (blob != nullptr) {
2841     return blob->as_runtime_stub();
2842   }
2843 
2844   // allocate space for the code
2845   ResourceMark rm;
2846   CodeBuffer buffer(name, 1000, 512);
2847   MacroAssembler* masm                = new MacroAssembler(&buffer);
2848 
2849   int frame_size_in_words;
2850   RegisterSaver reg_save(false /* save_vectors */);
2851 
2852   OopMapSet *oop_maps = new OopMapSet();
2853   OopMap* map = nullptr;
2854 
2855   int start = __ offset();
2856 
2857   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2858 
2859   int frame_complete = __ offset();
2860 
2861   {
2862     Label retaddr;
2863     __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2864 
2865     __ mov(c_rarg0, rthread);
2866     __ lea(rscratch1, RuntimeAddress(destination));
2867 
2868     __ blr(rscratch1);
2869     __ bind(retaddr);
2870   }
2871 
2872   // Set an oopmap for the call site.
2873   // We need this not only for callee-saved registers, but also for volatile
2874   // registers that the compiler might be keeping live across a safepoint.
2875 
2876   oop_maps->add_gc_map( __ offset() - start, map);
2877 
2878   // r0 contains the address we are going to jump to assuming no exception got installed
2879 
2880   // clear last_Java_sp
2881   __ reset_last_Java_frame(false);
2882   // check for pending exceptions
2883   Label pending;
2884   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2885   __ cbnz(rscratch1, pending);
2886 
2887   // get the returned Method*
2888   __ get_vm_result_metadata(rmethod, rthread);
2889   __ str(rmethod, Address(sp, reg_save.reg_offset_in_bytes(rmethod)));
2890 
2891   // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch
2892   __ str(r0, Address(sp, reg_save.rscratch1_offset_in_bytes()));
2893   reg_save.restore_live_registers(masm);
2894 
2895   // We are back to the original state on entry and ready to go.
2896 
2897   __ br(rscratch1);
2898 
2899   // Pending exception after the safepoint
2900 
2901   __ bind(pending);
2902 
2903   reg_save.restore_live_registers(masm);
2904 
2905   // exception pending => remove activation and forward to exception handler
2906 
2907   __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
2908 
2909   __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
2910   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2911 
2912   // -------------
2913   // make sure all code is generated
2914   masm->flush();
2915 
2916   // return the  blob
2917   // frame_size_words or bytes??
2918   RuntimeStub* rs_blob = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2919 
2920   AOTCodeCache::store_code_blob(*rs_blob, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2921   return rs_blob;
2922 }
2923 
2924 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
2925   BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
2926   if (buf == nullptr) {
2927     return nullptr;
2928   }
2929   CodeBuffer buffer(buf);
2930   short buffer_locs[20];
2931   buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
2932                                          sizeof(buffer_locs)/sizeof(relocInfo));
2933 
2934   MacroAssembler _masm(&buffer);
2935   MacroAssembler* masm = &_masm;
2936 
2937   const Array<SigEntry>* sig_vk = vk->extended_sig();
2938   const Array<VMRegPair>* regs = vk->return_regs();
2939 
2940   int pack_fields_jobject_off = __ offset();
2941   // Resolve pre-allocated buffer from JNI handle.
2942   // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
2943   Register Rresult = r14;  // See StubGenerator::generate_call_stub().
2944   __ ldr(r0, Address(Rresult));
2945   __ resolve_jobject(r0 /* value */,
2946                      rthread /* thread */,
2947                      r12 /* tmp */);
2948   __ str(r0, Address(Rresult));
2949 
2950   int pack_fields_off = __ offset();
2951 
2952   int j = 1;
2953   for (int i = 0; i < sig_vk->length(); i++) {
2954     BasicType bt = sig_vk->at(i)._bt;
2955     if (bt == T_METADATA) {
2956       continue;
2957     }
2958     if (bt == T_VOID) {
2959       if (sig_vk->at(i-1)._bt == T_LONG ||
2960           sig_vk->at(i-1)._bt == T_DOUBLE) {
2961         j++;
2962       }
2963       continue;
2964     }
2965     int off = sig_vk->at(i)._offset;
2966     VMRegPair pair = regs->at(j);
2967     VMReg r_1 = pair.first();
2968     VMReg r_2 = pair.second();
2969     Address to(r0, off);
2970     if (bt == T_FLOAT) {
2971       __ strs(r_1->as_FloatRegister(), to);
2972     } else if (bt == T_DOUBLE) {
2973       __ strd(r_1->as_FloatRegister(), to);
2974     } else {
2975       Register val = r_1->as_Register();
2976       assert_different_registers(to.base(), val, r15, r16, r17);
2977       if (is_reference_type(bt)) {
2978         // store_heap_oop transitively calls oop_store_at which corrupts to.base(). We need to keep r0 valid.
2979         __ mov(r17, r0);
2980         Address to_with_r17(r17, off);
2981         __ store_heap_oop(to_with_r17, val, r15, r16, r17, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
2982       } else {
2983         __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
2984       }
2985     }
2986     j++;
2987   }
2988   assert(j == regs->length(), "missed a field?");
2989   if (vk->supports_nullable_layouts()) {
2990     // Zero the null marker (setting it to 1 would be better but would require an additional register)
2991     __ strb(zr, Address(r0, vk->null_marker_offset()));
2992   }
2993   __ ret(lr);
2994 
2995   int unpack_fields_off = __ offset();
2996 
2997   Label skip;
2998   Label not_null;
2999   __ cbnz(r0, not_null);
3000 
3001   // Return value is null. Zero all registers because the runtime requires a canonical
3002   // representation of a flat null.
3003   j = 1;
3004   for (int i = 0; i < sig_vk->length(); i++) {
3005     BasicType bt = sig_vk->at(i)._bt;
3006     if (bt == T_METADATA) {
3007       continue;
3008     }
3009     if (bt == T_VOID) {
3010       if (sig_vk->at(i-1)._bt == T_LONG ||
3011           sig_vk->at(i-1)._bt == T_DOUBLE) {
3012         j++;
3013       }
3014       continue;
3015     }
3016 
3017     VMRegPair pair = regs->at(j);
3018     VMReg r_1 = pair.first();
3019     if (r_1->is_FloatRegister()) {
3020       __ mov(r_1->as_FloatRegister(), Assembler::T2S, 0);
3021     } else {
3022       __ mov(r_1->as_Register(), zr);
3023     }
3024     j++;
3025   }
3026   __ b(skip);
3027   __ bind(not_null);
3028 
3029   j = 1;
3030   for (int i = 0; i < sig_vk->length(); i++) {
3031     BasicType bt = sig_vk->at(i)._bt;
3032     if (bt == T_METADATA) {
3033       continue;
3034     }
3035     if (bt == T_VOID) {
3036       if (sig_vk->at(i-1)._bt == T_LONG ||
3037           sig_vk->at(i-1)._bt == T_DOUBLE) {
3038         j++;
3039       }
3040       continue;
3041     }
3042     int off = sig_vk->at(i)._offset;
3043     assert(off > 0, "offset in object should be positive");
3044     VMRegPair pair = regs->at(j);
3045     VMReg r_1 = pair.first();
3046     VMReg r_2 = pair.second();
3047     Address from(r0, off);
3048     if (bt == T_FLOAT) {
3049       __ ldrs(r_1->as_FloatRegister(), from);
3050     } else if (bt == T_DOUBLE) {
3051       __ ldrd(r_1->as_FloatRegister(), from);
3052     } else if (bt == T_OBJECT || bt == T_ARRAY) {
3053       assert_different_registers(r0, r_1->as_Register());
3054       __ load_heap_oop(r_1->as_Register(), from, rscratch1, rscratch2);
3055     } else {
3056       assert(is_java_primitive(bt), "unexpected basic type");
3057       assert_different_registers(r0, r_1->as_Register());
3058       size_t size_in_bytes = type2aelembytes(bt);
3059       __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3060     }
3061     j++;
3062   }
3063   assert(j == regs->length(), "missed a field?");
3064 
3065   __ bind(skip);
3066 
3067   __ ret(lr);
3068 
3069   __ flush();
3070 
3071   return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3072 }
3073 
3074 // Continuation point for throwing of implicit exceptions that are
3075 // not handled in the current activation. Fabricates an exception
3076 // oop and initiates normal exception dispatching in this
3077 // frame. Since we need to preserve callee-saved values (currently
3078 // only for C2, but done for C1 as well) we need a callee-saved oop
3079 // map and therefore have to make these stubs into RuntimeStubs
3080 // rather than BufferBlobs.  If the compiler needs all registers to
3081 // be preserved between the fault point and the exception handler
3082 // then it must assume responsibility for that in
3083 // AbstractCompiler::continuation_for_implicit_null_exception or
3084 // continuation_for_implicit_division_by_zero_exception. All other
3085 // implicit exceptions (e.g., NullPointerException or
3086 // AbstractMethodError on entry) are either at call sites or
3087 // otherwise assume that stack unwinding will be initiated, so
3088 // caller saved registers were assumed volatile in the compiler.
3089 
3090 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
3091   assert(is_throw_id(id), "expected a throw stub id");
3092 
3093   const char* name = SharedRuntime::stub_name(id);
3094 
3095   // Information about frame layout at time of blocking runtime call.
3096   // Note that we only have to preserve callee-saved registers since
3097   // the compilers are responsible for supplying a continuation point
3098   // if they expect all registers to be preserved.
3099   // n.b. aarch64 asserts that frame::arg_reg_save_area_bytes == 0
3100   enum layout {
3101     rfp_off = 0,
3102     rfp_off2,
3103     return_off,
3104     return_off2,
3105     framesize // inclusive of return address
3106   };
3107 
3108   int insts_size = 512;
3109   int locs_size  = 64;
3110 
3111   const char* timer_msg = "SharedRuntime generate_throw_exception";
3112   TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3113 
3114   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, StubInfo::blob(id));
3115   if (blob != nullptr) {
3116     return blob->as_runtime_stub();
3117   }
3118 
3119   ResourceMark rm;
3120   CodeBuffer code(name, insts_size, locs_size);
3121   OopMapSet* oop_maps  = new OopMapSet();
3122   MacroAssembler* masm = new MacroAssembler(&code);
3123 
3124   address start = __ pc();
3125 
3126   // This is an inlined and slightly modified version of call_VM
3127   // which has the ability to fetch the return PC out of
3128   // thread-local storage and also sets up last_Java_sp slightly
3129   // differently than the real call_VM
3130 
3131   __ enter(); // Save FP and LR before call
3132 
3133   assert(is_even(framesize/2), "sp not 16-byte aligned");
3134 
3135   // lr and fp are already in place
3136   __ sub(sp, rfp, ((uint64_t)framesize-4) << LogBytesPerInt); // prolog
3137 
3138   int frame_complete = __ pc() - start;
3139 
3140   // Set up last_Java_sp and last_Java_fp
3141   address the_pc = __ pc();
3142   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3143 
3144   __ mov(c_rarg0, rthread);
3145   BLOCK_COMMENT("call runtime_entry");
3146   __ lea(rscratch1, RuntimeAddress(runtime_entry));
3147   __ blr(rscratch1);
3148 
3149   // Generate oop map
3150   OopMap* map = new OopMap(framesize, 0);
3151 
3152   oop_maps->add_gc_map(the_pc - start, map);
3153 
3154   __ reset_last_Java_frame(true);
3155 
3156   // Reinitialize the ptrue predicate register, in case the external runtime
3157   // call clobbers ptrue reg, as we may return to SVE compiled code.
3158   __ reinitialize_ptrue();
3159 
3160   __ leave();
3161 
3162   // check for pending exceptions
3163 #ifdef ASSERT
3164   Label L;
3165   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
3166   __ cbnz(rscratch1, L);
3167   __ should_not_reach_here();
3168   __ bind(L);
3169 #endif // ASSERT
3170   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3171 
3172   // codeBlob framesize is in words (not VMRegImpl::slot_size)
3173   RuntimeStub* stub =
3174     RuntimeStub::new_runtime_stub(name,
3175                                   &code,
3176                                   frame_complete,
3177                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3178                                   oop_maps, false);
3179   AOTCodeCache::store_code_blob(*stub, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
3180 
3181   return stub;
3182 }
3183 
3184 #if INCLUDE_JFR
3185 
3186 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
3187   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3188   __ mov(c_rarg0, thread);
3189 }
3190 
3191 // The handle is dereferenced through a load barrier.
3192 static void jfr_epilogue(MacroAssembler* masm) {
3193   __ reset_last_Java_frame(true);
3194 }
3195 
3196 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3197 // It returns a jobject handle to the event writer.
3198 // The handle is dereferenced and the return value is the event writer oop.
3199 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3200   enum layout {
3201     rbp_off,
3202     rbpH_off,
3203     return_off,
3204     return_off2,
3205     framesize // inclusive of return address
3206   };
3207 
3208   int insts_size = 1024;
3209   int locs_size = 64;
3210   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
3211   CodeBuffer code(name, insts_size, locs_size);
3212   OopMapSet* oop_maps = new OopMapSet();
3213   MacroAssembler* masm = new MacroAssembler(&code);
3214 
3215   address start = __ pc();
3216   __ enter();
3217   int frame_complete = __ pc() - start;
3218   address the_pc = __ pc();
3219   jfr_prologue(the_pc, masm, rthread);
3220   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
3221   jfr_epilogue(masm);
3222   __ resolve_global_jobject(r0, rscratch1, rscratch2);
3223   __ leave();
3224   __ ret(lr);
3225 
3226   OopMap* map = new OopMap(framesize, 1); // rfp
3227   oop_maps->add_gc_map(the_pc - start, map);
3228 
3229   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3230     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3231                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3232                                   oop_maps, false);
3233   return stub;
3234 }
3235 
3236 // For c2: call to return a leased buffer.
3237 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3238   enum layout {
3239     rbp_off,
3240     rbpH_off,
3241     return_off,
3242     return_off2,
3243     framesize // inclusive of return address
3244   };
3245 
3246   int insts_size = 1024;
3247   int locs_size = 64;
3248 
3249   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
3250   CodeBuffer code(name, insts_size, locs_size);
3251   OopMapSet* oop_maps = new OopMapSet();
3252   MacroAssembler* masm = new MacroAssembler(&code);
3253 
3254   address start = __ pc();
3255   __ enter();
3256   int frame_complete = __ pc() - start;
3257   address the_pc = __ pc();
3258   jfr_prologue(the_pc, masm, rthread);
3259   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
3260   jfr_epilogue(masm);
3261 
3262   __ leave();
3263   __ ret(lr);
3264 
3265   OopMap* map = new OopMap(framesize, 1); // rfp
3266   oop_maps->add_gc_map(the_pc - start, map);
3267 
3268   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3269     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3270                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3271                                   oop_maps, false);
3272   return stub;
3273 }
3274 
3275 #endif // INCLUDE_JFR