1 /*
   2  * Copyright (c) 2003, 2025, 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 #if INCLUDE_JVMCI
  66 #include "jvmci/jvmciJavaClasses.hpp"
  67 #endif
  68 
  69 #define __ masm->
  70 
  71 #ifdef PRODUCT
  72 #define BLOCK_COMMENT(str) /* nothing */
  73 #else
  74 #define BLOCK_COMMENT(str) __ block_comment(str)
  75 #endif
  76 
  77 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
  78 
  79 // FIXME -- this is used by C1
  80 class RegisterSaver {
  81   const bool _save_vectors;
  82  public:
  83   RegisterSaver(bool save_vectors) : _save_vectors(save_vectors) {}
  84 
  85   OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
  86   void restore_live_registers(MacroAssembler* masm);
  87 
  88   // Offsets into the register save area
  89   // Used by deoptimization when it is managing result register
  90   // values on its own
  91 
  92   int reg_offset_in_bytes(Register r);
  93   int r0_offset_in_bytes()    { return reg_offset_in_bytes(r0); }
  94   int rscratch1_offset_in_bytes()    { return reg_offset_in_bytes(rscratch1); }
  95   int v0_offset_in_bytes();
  96 
  97   // Total stack size in bytes for saving sve predicate registers.
  98   int total_sve_predicate_in_bytes();
  99 
 100   // Capture info about frame layout
 101   // Note this is only correct when not saving full vectors.
 102   enum layout {
 103                 fpu_state_off = 0,
 104                 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1,
 105                 // The frame sender code expects that rfp will be in
 106                 // the "natural" place and will override any oopMap
 107                 // setting for it. We must therefore force the layout
 108                 // so that it agrees with the frame sender code.
 109                 r0_off = fpu_state_off + FPUStateSizeInWords,
 110                 rfp_off = r0_off + (Register::number_of_registers - 2) * Register::max_slots_per_register,
 111                 return_off = rfp_off + Register::max_slots_per_register,      // slot for return address
 112                 reg_save_size = return_off + Register::max_slots_per_register};
 113 
 114 };
 115 
 116 int RegisterSaver::reg_offset_in_bytes(Register r) {
 117   // The integer registers are located above the floating point
 118   // registers in the stack frame pushed by save_live_registers() so the
 119   // offset depends on whether we are saving full vectors, and whether
 120   // those vectors are NEON or SVE.
 121 
 122   int slots_per_vect = FloatRegister::save_slots_per_register;
 123 
 124 #if COMPILER2_OR_JVMCI
 125   if (_save_vectors) {
 126     slots_per_vect = FloatRegister::slots_per_neon_register;
 127 
 128 #ifdef COMPILER2
 129     if (Matcher::supports_scalable_vector()) {
 130       slots_per_vect = Matcher::scalable_vector_reg_size(T_FLOAT);
 131     }
 132 #endif
 133   }
 134 #endif
 135 
 136   int r0_offset = v0_offset_in_bytes() + (slots_per_vect * FloatRegister::number_of_registers) * BytesPerInt;
 137   return r0_offset + r->encoding() * wordSize;
 138 }
 139 
 140 int RegisterSaver::v0_offset_in_bytes() {
 141   // The floating point registers are located above the predicate registers if
 142   // they are present in the stack frame pushed by save_live_registers(). So the
 143   // offset depends on the saved total predicate vectors in the stack frame.
 144   return (total_sve_predicate_in_bytes() / VMRegImpl::stack_slot_size) * BytesPerInt;
 145 }
 146 
 147 int RegisterSaver::total_sve_predicate_in_bytes() {
 148 #ifdef COMPILER2
 149   if (_save_vectors && Matcher::supports_scalable_vector()) {
 150     return (Matcher::scalable_vector_reg_size(T_BYTE) >> LogBitsPerByte) *
 151            PRegister::number_of_registers;
 152   }
 153 #endif
 154   return 0;
 155 }
 156 
 157 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
 158   bool use_sve = false;
 159   int sve_vector_size_in_bytes = 0;
 160   int sve_vector_size_in_slots = 0;
 161   int sve_predicate_size_in_slots = 0;
 162   int total_predicate_in_bytes = total_sve_predicate_in_bytes();
 163   int total_predicate_in_slots = total_predicate_in_bytes / VMRegImpl::stack_slot_size;
 164 
 165 #ifdef COMPILER2
 166   use_sve = Matcher::supports_scalable_vector();
 167   if (use_sve) {
 168     sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
 169     sve_vector_size_in_slots = Matcher::scalable_vector_reg_size(T_FLOAT);
 170     sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
 171   }
 172 #endif
 173 
 174 #if COMPILER2_OR_JVMCI
 175   if (_save_vectors) {
 176     int extra_save_slots_per_register = 0;
 177     // Save upper half of vector registers
 178     if (use_sve) {
 179       extra_save_slots_per_register = sve_vector_size_in_slots - FloatRegister::save_slots_per_register;
 180     } else {
 181       extra_save_slots_per_register = FloatRegister::extra_save_slots_per_neon_register;
 182     }
 183     int extra_vector_bytes = extra_save_slots_per_register *
 184                              VMRegImpl::stack_slot_size *
 185                              FloatRegister::number_of_registers;
 186     additional_frame_words += ((extra_vector_bytes + total_predicate_in_bytes) / wordSize);
 187   }
 188 #else
 189   assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
 190 #endif
 191 
 192   int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
 193                                      reg_save_size * BytesPerInt, 16);
 194   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 195   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 196   // The caller will allocate additional_frame_words
 197   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 198   // CodeBlob frame size is in words.
 199   int frame_size_in_words = frame_size_in_bytes / wordSize;
 200   *total_frame_words = frame_size_in_words;
 201 
 202   // Save Integer and Float registers.
 203   __ enter();
 204   __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
 205 
 206   // Set an oopmap for the call site.  This oopmap will map all
 207   // oop-registers and debug-info registers as callee-saved.  This
 208   // will allow deoptimization at this safepoint to find all possible
 209   // debug-info recordings, as well as let GC find all oops.
 210 
 211   OopMapSet *oop_maps = new OopMapSet();
 212   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 213 
 214   for (int i = 0; i < Register::number_of_registers; i++) {
 215     Register r = as_Register(i);
 216     if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
 217       // SP offsets are in 4-byte words.
 218       // Register slots are 8 bytes wide, 32 floating-point registers.
 219       int sp_offset = Register::max_slots_per_register * i +
 220                       FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
 221       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
 222     }
 223   }
 224 
 225   for (int i = 0; i < FloatRegister::number_of_registers; i++) {
 226     FloatRegister r = as_FloatRegister(i);
 227     int sp_offset = 0;
 228     if (_save_vectors) {
 229       sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
 230                             (FloatRegister::slots_per_neon_register * i);
 231     } else {
 232       sp_offset = FloatRegister::save_slots_per_register * i;
 233     }
 234     oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), r->as_VMReg());
 235   }
 236 
 237   return oop_map;
 238 }
 239 
 240 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
 241 #ifdef COMPILER2
 242   __ pop_CPU_state(_save_vectors, Matcher::supports_scalable_vector(),
 243                    Matcher::scalable_vector_reg_size(T_BYTE), total_sve_predicate_in_bytes());
 244 #else
 245 #if !INCLUDE_JVMCI
 246   assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
 247 #endif
 248   __ pop_CPU_state(_save_vectors);
 249 #endif
 250   __ ldp(rfp, lr, Address(__ post(sp, 2 * wordSize)));
 251   __ authenticate_return_address();
 252 }
 253 
 254 // Is vector's size (in bytes) bigger than a size saved by default?
 255 // 8 bytes vector registers are saved by default on AArch64.
 256 // The SVE supported min vector size is 8 bytes and we need to save
 257 // predicate registers when the vector size is 8 bytes as well.
 258 bool SharedRuntime::is_wide_vector(int size) {
 259   return size > 8 || (UseSVE > 0 && size >= 8);
 260 }
 261 
 262 // ---------------------------------------------------------------------------
 263 // Read the array of BasicTypes from a signature, and compute where the
 264 // arguments should go.  Values in the VMRegPair regs array refer to 4-byte
 265 // quantities.  Values less than VMRegImpl::stack0 are registers, those above
 266 // refer to 4-byte stack slots.  All stack slots are based off of the stack pointer
 267 // as framesizes are fixed.
 268 // VMRegImpl::stack0 refers to the first slot 0(sp).
 269 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
 270 // Register up to Register::number_of_registers are the 64-bit
 271 // integer registers.
 272 
 273 // Note: the INPUTS in sig_bt are in units of Java argument words,
 274 // which are 64-bit.  The OUTPUTS are in 32-bit units.
 275 
 276 // The Java calling convention is a "shifted" version of the C ABI.
 277 // By skipping the first C ABI register we can call non-static jni
 278 // methods with small numbers of arguments without having to shuffle
 279 // the arguments at all. Since we control the java ABI we ought to at
 280 // least get some advantage out of it.
 281 
 282 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
 283                                            VMRegPair *regs,
 284                                            int total_args_passed) {
 285 
 286   // Create the mapping between argument positions and
 287   // registers.
 288   static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
 289     j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7
 290   };
 291   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
 292     j_farg0, j_farg1, j_farg2, j_farg3,
 293     j_farg4, j_farg5, j_farg6, j_farg7
 294   };
 295 
 296 
 297   uint int_args = 0;
 298   uint fp_args = 0;
 299   uint stk_args = 0;
 300 
 301   for (int i = 0; i < total_args_passed; i++) {
 302     switch (sig_bt[i]) {
 303     case T_BOOLEAN:
 304     case T_CHAR:
 305     case T_BYTE:
 306     case T_SHORT:
 307     case T_INT:
 308       if (int_args < Argument::n_int_register_parameters_j) {
 309         regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 310       } else {
 311         stk_args = align_up(stk_args, 2);
 312         regs[i].set1(VMRegImpl::stack2reg(stk_args));
 313         stk_args += 1;
 314       }
 315       break;
 316     case T_VOID:
 317       // halves of T_LONG or T_DOUBLE
 318       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 319       regs[i].set_bad();
 320       break;
 321     case T_LONG:
 322       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 323       // fall through
 324     case T_OBJECT:
 325     case T_ARRAY:
 326     case T_ADDRESS:
 327       if (int_args < Argument::n_int_register_parameters_j) {
 328         regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 329       } else {
 330         stk_args = align_up(stk_args, 2);
 331         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 332         stk_args += 2;
 333       }
 334       break;
 335     case T_FLOAT:
 336       if (fp_args < Argument::n_float_register_parameters_j) {
 337         regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 338       } else {
 339         stk_args = align_up(stk_args, 2);
 340         regs[i].set1(VMRegImpl::stack2reg(stk_args));
 341         stk_args += 1;
 342       }
 343       break;
 344     case T_DOUBLE:
 345       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 346       if (fp_args < Argument::n_float_register_parameters_j) {
 347         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 348       } else {
 349         stk_args = align_up(stk_args, 2);
 350         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 351         stk_args += 2;
 352       }
 353       break;
 354     default:
 355       ShouldNotReachHere();
 356       break;
 357     }
 358   }
 359 
 360   return stk_args;
 361 }
 362 
 363 
 364 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
 365 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
 366 
 367 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
 368 
 369   // Create the mapping between argument positions and registers.
 370 
 371   static const Register INT_ArgReg[java_return_convention_max_int] = {
 372     r0 /* j_rarg7 */, j_rarg6, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
 373   };
 374 
 375   static const FloatRegister FP_ArgReg[java_return_convention_max_float] = {
 376     j_farg0, j_farg1, j_farg2, j_farg3, j_farg4, j_farg5, j_farg6, j_farg7
 377   };
 378 
 379   uint int_args = 0;
 380   uint fp_args = 0;
 381 
 382   for (int i = 0; i < total_args_passed; i++) {
 383     switch (sig_bt[i]) {
 384     case T_BOOLEAN:
 385     case T_CHAR:
 386     case T_BYTE:
 387     case T_SHORT:
 388     case T_INT:
 389       if (int_args < SharedRuntime::java_return_convention_max_int) {
 390         regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
 391         int_args ++;
 392       } else {
 393         return -1;
 394       }
 395       break;
 396     case T_VOID:
 397       // halves of T_LONG or T_DOUBLE
 398       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 399       regs[i].set_bad();
 400       break;
 401     case T_LONG:
 402       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 403       // fall through
 404     case T_OBJECT:
 405     case T_ARRAY:
 406     case T_ADDRESS:
 407       // Should T_METADATA be added to java_calling_convention as well ?
 408     case T_METADATA:
 409       if (int_args < SharedRuntime::java_return_convention_max_int) {
 410         regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
 411         int_args ++;
 412       } else {
 413         return -1;
 414       }
 415       break;
 416     case T_FLOAT:
 417       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 418         regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
 419         fp_args ++;
 420       } else {
 421         return -1;
 422       }
 423       break;
 424     case T_DOUBLE:
 425       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 426       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 427         regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
 428         fp_args ++;
 429       } else {
 430         return -1;
 431       }
 432       break;
 433     default:
 434       ShouldNotReachHere();
 435       break;
 436     }
 437   }
 438 
 439   return int_args + fp_args;
 440 }
 441 
 442 // Patch the callers callsite with entry to compiled code if it exists.
 443 static void patch_callers_callsite(MacroAssembler *masm) {
 444   Label L;
 445   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 446   __ cbz(rscratch1, L);
 447 
 448   __ enter();
 449   __ push_CPU_state();
 450 
 451   // VM needs caller's callsite
 452   // VM needs target method
 453   // This needs to be a long call since we will relocate this adapter to
 454   // the codeBuffer and it may not reach
 455 
 456 #ifndef PRODUCT
 457   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 458 #endif
 459 
 460   __ mov(c_rarg0, rmethod);
 461   __ mov(c_rarg1, lr);
 462   __ authenticate_return_address(c_rarg1);
 463   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 464   __ blr(rscratch1);
 465 
 466   // Explicit isb required because fixup_callers_callsite may change the code
 467   // stream.
 468   __ safepoint_isb();
 469 
 470   __ pop_CPU_state();
 471   // restore sp
 472   __ leave();
 473   __ bind(L);
 474 }
 475 
 476 // For each inline type argument, sig includes the list of fields of
 477 // the inline type. This utility function computes the number of
 478 // arguments for the call if inline types are passed by reference (the
 479 // calling convention the interpreter expects).
 480 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
 481   int total_args_passed = 0;
 482   if (InlineTypePassFieldsAsArgs) {
 483     for (int i = 0; i < sig_extended->length(); i++) {
 484       BasicType bt = sig_extended->at(i)._bt;
 485       if (bt == T_METADATA) {
 486         // In sig_extended, an inline type argument starts with:
 487         // T_METADATA, followed by the types of the fields of the
 488         // inline type and T_VOID to mark the end of the value
 489         // type. Inline types are flattened so, for instance, in the
 490         // case of an inline type with an int field and an inline type
 491         // field that itself has 2 fields, an int and a long:
 492         // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
 493         // slot for the T_LONG) T_VOID (inner inline type) T_VOID
 494         // (outer inline type)
 495         total_args_passed++;
 496         int vt = 1;
 497         do {
 498           i++;
 499           BasicType bt = sig_extended->at(i)._bt;
 500           BasicType prev_bt = sig_extended->at(i-1)._bt;
 501           if (bt == T_METADATA) {
 502             vt++;
 503           } else if (bt == T_VOID &&
 504                      prev_bt != T_LONG &&
 505                      prev_bt != T_DOUBLE) {
 506             vt--;
 507           }
 508         } while (vt != 0);
 509       } else {
 510         total_args_passed++;
 511       }
 512     }
 513   } else {
 514     total_args_passed = sig_extended->length();
 515   }
 516   return total_args_passed;
 517 }
 518 
 519 
 520 static void gen_c2i_adapter_helper(MacroAssembler* masm,
 521                                    BasicType bt,
 522                                    BasicType prev_bt,
 523                                    size_t size_in_bytes,
 524                                    const VMRegPair& reg_pair,
 525                                    const Address& to,
 526                                    Register tmp1,
 527                                    Register tmp2,
 528                                    Register tmp3,
 529                                    int extraspace,
 530                                    bool is_oop) {
 531   if (bt == T_VOID) {
 532     assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
 533     return;
 534   }
 535 
 536   // Say 4 args:
 537   // i   st_off
 538   // 0   32 T_LONG
 539   // 1   24 T_VOID
 540   // 2   16 T_OBJECT
 541   // 3    8 T_BOOL
 542   // -    0 return address
 543   //
 544   // However to make thing extra confusing. Because we can fit a Java long/double in
 545   // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 546   // leaves one slot empty and only stores to a single slot. In this case the
 547   // slot that is occupied is the T_VOID slot. See I said it was confusing.
 548 
 549   bool wide = (size_in_bytes == wordSize);
 550   VMReg r_1 = reg_pair.first();
 551   VMReg r_2 = reg_pair.second();
 552   assert(r_2->is_valid() == wide, "invalid size");
 553   if (!r_1->is_valid()) {
 554     assert(!r_2->is_valid(), "");
 555     return;
 556   }
 557 
 558   if (!r_1->is_FloatRegister()) {
 559     Register val = r25;
 560     if (r_1->is_stack()) {
 561       // memory to memory use r25 (scratch registers is used by store_heap_oop)
 562       int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 563       __ load_sized_value(val, Address(sp, ld_off), size_in_bytes, /* is_signed */ false);
 564     } else {
 565       val = r_1->as_Register();
 566     }
 567     assert_different_registers(to.base(), val, tmp1, tmp2, tmp3);
 568     if (is_oop) {
 569       __ store_heap_oop(to, val, tmp1, tmp2, tmp3, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
 570     } else {
 571       __ store_sized_value(to, val, size_in_bytes);
 572     }
 573   } else {
 574     if (wide) {
 575       __ strd(r_1->as_FloatRegister(), to);
 576     } else {
 577       // only a float use just part of the slot
 578       __ strs(r_1->as_FloatRegister(), to);
 579     }
 580   }
 581 }
 582 
 583 static void gen_c2i_adapter(MacroAssembler *masm,
 584                             const GrowableArray<SigEntry>* sig_extended,


 585                             const VMRegPair *regs,
 586                             bool requires_clinit_barrier,
 587                             address& c2i_no_clinit_check_entry,
 588                             Label& skip_fixup,
 589                             address start,
 590                             OopMapSet* oop_maps,
 591                             int& frame_complete,
 592                             int& frame_size_in_words,
 593                             bool alloc_inline_receiver) {
 594   if (requires_clinit_barrier && VM_Version::supports_fast_class_init_checks()) {
 595     Label L_skip_barrier;
 596 
 597     { // Bypass the barrier for non-static methods
 598       __ ldrh(rscratch1, Address(rmethod, Method::access_flags_offset()));
 599       __ andsw(zr, rscratch1, JVM_ACC_STATIC);
 600       __ br(Assembler::EQ, L_skip_barrier); // non-static
 601     }
 602 
 603     __ load_method_holder(rscratch2, rmethod);
 604     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
 605     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 606 
 607     __ bind(L_skip_barrier);
 608     c2i_no_clinit_check_entry = __ pc();
 609   }
 610 
 611   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 612   bs->c2i_entry_barrier(masm);
 613 
 614   // Before we get into the guts of the C2I adapter, see if we should be here
 615   // at all.  We've come from compiled code and are attempting to jump to the
 616   // interpreter, which means the caller made a static call to get here
 617   // (vcalls always get a compiled target if there is one).  Check for a
 618   // compiled target.  If there is one, we need to patch the caller's call.
 619   patch_callers_callsite(masm);
 620 
 621   __ bind(skip_fixup);
 622 
 623   // TODO 8366717 Is the comment about r13 correct? Isn't that r19_sender_sp?
 624   // Name some registers to be used in the following code. We can use
 625   // anything except r0-r7 which are arguments in the Java calling
 626   // convention, rmethod (r12), and r13 which holds the outgoing sender
 627   // SP for the interpreter.
 628   // TODO 8366717 We need to make sure that buf_array, buf_oop (and potentially other long-life regs) are kept live in slowpath runtime calls in GC barriers
 629   Register buf_array = r10;   // Array of buffered inline types
 630   Register buf_oop = r11;     // Buffered inline type oop
 631   Register tmp1 = r15;
 632   Register tmp2 = r16;
 633   Register tmp3 = r17;
 634 
 635   if (InlineTypePassFieldsAsArgs) {
 636     // Is there an inline type argument?
 637     bool has_inline_argument = false;
 638     for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
 639       has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
 640     }
 641     if (has_inline_argument) {
 642       // There is at least an inline type argument: we're coming from
 643       // compiled code so we have no buffers to back the inline types
 644       // Allocate the buffers here with a runtime call.
 645       // TODO 8366717 Do we need to save vectors here? They could be used as arg registers, right? Same on x64.
 646       RegisterSaver reg_save(true /* save_vectors */);
 647       OopMap* map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
 648 
 649       frame_complete = __ offset();
 650       address the_pc = __ pc();
 651 
 652       Label retaddr;
 653       __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
 654 
 655       __ mov(c_rarg0, rthread);
 656       __ mov(c_rarg1, rmethod);
 657       __ mov(c_rarg2, (int64_t)alloc_inline_receiver);
 658 
 659       __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
 660       __ blr(rscratch1);
 661       __ bind(retaddr);
 662 
 663       oop_maps->add_gc_map(__ pc() - start, map);
 664       __ reset_last_Java_frame(false);
 665 
 666       reg_save.restore_live_registers(masm);





 667 
 668       Label no_exception;
 669       __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
 670       __ cbz(rscratch1, no_exception);













 671 
 672       __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
 673       __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
 674       __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
 675 
 676       __ bind(no_exception);
 677 
 678       // We get an array of objects from the runtime call
 679       __ get_vm_result_oop(buf_array, rthread);
 680       __ get_vm_result_metadata(rmethod, rthread); // TODO: required to keep the callee Method live?
 681     }
 682   }








 683 
 684   // Since all args are passed on the stack, total_args_passed *
 685   // Interpreter::stackElementSize is the space we need.
 686 
 687   int total_args_passed = compute_total_args_passed_int(sig_extended);
 688   int extraspace = total_args_passed * Interpreter::stackElementSize;
 689 
 690   // stack is aligned, keep it that way
 691   extraspace = align_up(extraspace, StackAlignmentInBytes);
 692 
 693   // set senderSP value
 694   __ mov(r19_sender_sp, sp);
 695 
 696   __ sub(sp, sp, extraspace);
 697 
 698   // Now write the args into the outgoing interpreter space
 699 
 700   // next_arg_comp is the next argument from the compiler point of
 701   // view (inline type fields are passed in registers/on the stack). In
 702   // sig_extended, an inline type argument starts with: T_METADATA,
 703   // followed by the types of the fields of the inline type and T_VOID
 704   // to mark the end of the inline type. ignored counts the number of
 705   // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
 706   // used to get the buffer for that argument from the pool of buffers
 707   // we allocated above and want to pass to the
 708   // interpreter. next_arg_int is the next argument from the
 709   // interpreter point of view (inline types are passed by reference).
 710   for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
 711        next_arg_comp < sig_extended->length(); next_arg_comp++) {
 712     assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
 713     assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
 714     BasicType bt = sig_extended->at(next_arg_comp)._bt;
 715     int st_off = (total_args_passed - next_arg_int - 1) * Interpreter::stackElementSize;
 716     if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
 717       int next_off = st_off - Interpreter::stackElementSize;
 718       const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
 719       const VMRegPair reg_pair = regs[next_arg_comp-ignored];
 720       size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
 721       gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 722                              size_in_bytes, reg_pair, Address(sp, offset), tmp1, tmp2, tmp3, extraspace, false);
 723       next_arg_int++;
 724 #ifdef ASSERT
 725       if (bt == T_LONG || bt == T_DOUBLE) {
 726         // Overwrite the unused slot with known junk
 727         __ mov(rscratch1, CONST64(0xdeadffffdeadaaaa));
 728         __ str(rscratch1, Address(sp, st_off));



 729       }















 730 #endif /* ASSERT */
 731     } else {
 732       ignored++;
 733       // get the buffer from the just allocated pool of buffers
 734       int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
 735       __ load_heap_oop(buf_oop, Address(buf_array, index), tmp1, tmp2);
 736       next_vt_arg++; next_arg_int++;
 737       int vt = 1;
 738       // write fields we get from compiled code in registers/stack
 739       // slots to the buffer: we know we are done with that inline type
 740       // argument when we hit the T_VOID that acts as an end of inline
 741       // type delimiter for this inline type. Inline types are flattened
 742       // so we might encounter embedded inline types. Each entry in
 743       // sig_extended contains a field offset in the buffer.
 744       Label L_null;
 745       do {
 746         next_arg_comp++;
 747         BasicType bt = sig_extended->at(next_arg_comp)._bt;
 748         BasicType prev_bt = sig_extended->at(next_arg_comp - 1)._bt;
 749         if (bt == T_METADATA) {
 750           vt++;
 751           ignored++;
 752         } else if (bt == T_VOID && prev_bt != T_LONG && prev_bt != T_DOUBLE) {
 753           vt--;
 754           ignored++;
 755         } else {
 756           int off = sig_extended->at(next_arg_comp)._offset;
 757           if (off == -1) {
 758             // Nullable inline type argument, emit null check
 759             VMReg reg = regs[next_arg_comp-ignored].first();
 760             Label L_notNull;
 761             if (reg->is_stack()) {
 762               int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 763               __ ldrb(tmp1, Address(sp, ld_off));
 764               __ cbnz(tmp1, L_notNull);
 765             } else {
 766               __ cbnz(reg->as_Register(), L_notNull);
 767             }
 768             __ str(zr, Address(sp, st_off));
 769             __ b(L_null);
 770             __ bind(L_notNull);
 771             continue;
 772           }
 773           assert(off > 0, "offset in object should be positive");
 774           size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
 775           bool is_oop = is_reference_type(bt);
 776           gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 777                                  size_in_bytes, regs[next_arg_comp-ignored], Address(buf_oop, off), tmp1, tmp2, tmp3, extraspace, is_oop);
 778         }
 779       } while (vt != 0);
 780       // pass the buffer to the interpreter
 781       __ str(buf_oop, Address(sp, st_off));
 782       __ bind(L_null);










 783     }
 784   }
 785 
 786   __ mov(esp, sp); // Interp expects args on caller's expression stack
 787 
 788   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
 789   __ br(rscratch1);
 790 }
 791 
 792 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, int comp_args_on_stack, const GrowableArray<SigEntry>* sig, const VMRegPair *regs) {
 793 





 794 
 795   // Note: r19_sender_sp contains the senderSP on entry. We must
 796   // preserve it since we may do a i2c -> c2i transition if we lose a
 797   // race where compiled code goes non-entrant while we get args
 798   // ready.
 799 
 800   // Adapters are frameless.
 801 
 802   // An i2c adapter is frameless because the *caller* frame, which is
 803   // interpreted, routinely repairs its own esp (from
 804   // interpreter_frame_last_sp), even if a callee has modified the
 805   // stack pointer.  It also recalculates and aligns sp.
 806 
 807   // A c2i adapter is frameless because the *callee* frame, which is
 808   // interpreted, routinely repairs its caller's sp (from sender_sp,
 809   // which is set up via the senderSP register).
 810 
 811   // In other words, if *either* the caller or callee is interpreted, we can
 812   // get the stack pointer repaired after a call.
 813 
 814   // This is why c2i and i2c adapters cannot be indefinitely composed.
 815   // In particular, if a c2i adapter were to somehow call an i2c adapter,
 816   // both caller and callee would be compiled methods, and neither would
 817   // clean up the stack pointer changes performed by the two adapters.
 818   // If this happens, control eventually transfers back to the compiled
 819   // caller, but with an uncorrected stack, causing delayed havoc.
 820 
 821   // Cut-out for having no stack args.
 822   int comp_words_on_stack = 0;
 823   if (comp_args_on_stack) {
 824      comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
 825      __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
 826      __ andr(sp, rscratch1, -16);
 827   }
 828 
 829   // Will jump to the compiled code just as if compiled code was doing it.
 830   // Pre-load the register-jump target early, to schedule it better.
 831   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_inline_offset())));
 832 
 833 #if INCLUDE_JVMCI
 834   if (EnableJVMCI) {
 835     // check if this call should be routed towards a specific entry point
 836     __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 837     Label no_alternative_target;
 838     __ cbz(rscratch2, no_alternative_target);
 839     __ mov(rscratch1, rscratch2);
 840     __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 841     __ bind(no_alternative_target);
 842   }
 843 #endif // INCLUDE_JVMCI
 844 
 845   int total_args_passed = sig->length();
 846 
 847   // Now generate the shuffle code.
 848   for (int i = 0; i < total_args_passed; i++) {
 849     BasicType bt = sig->at(i)._bt;
 850     if (bt == T_VOID) {
 851       assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
 852       continue;
 853     }
 854 
 855     // Pick up 0, 1 or 2 words from SP+offset.
 856     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), "scrambled load targets?");
 857 


 858     // Load in argument order going down.
 859     int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 860     // Point to interpreter value (vs. tag)
 861     int next_off = ld_off - Interpreter::stackElementSize;
 862     //
 863     //
 864     //
 865     VMReg r_1 = regs[i].first();
 866     VMReg r_2 = regs[i].second();
 867     if (!r_1->is_valid()) {
 868       assert(!r_2->is_valid(), "");
 869       continue;
 870     }
 871     if (r_1->is_stack()) {
 872       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 873       int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
 874       if (!r_2->is_valid()) {
 875         // sign extend???
 876         __ ldrsw(rscratch2, Address(esp, ld_off));
 877         __ str(rscratch2, Address(sp, st_off));
 878       } else {
 879         //
 880         // We are using two optoregs. This can be either T_OBJECT,
 881         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 882         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 883         // So we must adjust where to pick up the data to match the
 884         // interpreter.
 885         //
 886         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 887         // are accessed as negative so LSW is at LOW address
 888 
 889         // ld_off is MSW so get LSW
 890         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;

 891         __ ldr(rscratch2, Address(esp, offset));
 892         // st_off is LSW (i.e. reg.first())
 893          __ str(rscratch2, Address(sp, st_off));
 894        }
 895      } else if (r_1->is_Register()) {  // Register argument
 896        Register r = r_1->as_Register();
 897        if (r_2->is_valid()) {
 898          //
 899          // We are using two VMRegs. This can be either T_OBJECT,
 900          // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 901          // two slots but only uses one for thr T_LONG or T_DOUBLE case
 902          // So we must adjust where to pick up the data to match the
 903          // interpreter.
 904 
 905         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
 906 
 907          // this can be a misaligned move
 908          __ ldr(r, Address(esp, offset));
 909        } else {
 910          // sign extend and use a full word?
 911          __ ldrw(r, Address(esp, ld_off));
 912        }
 913      } else {
 914        if (!r_2->is_valid()) {
 915          __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
 916        } else {
 917          __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
 918        }
 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   __ br(rscratch1);
 939 }
 940 
 941 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
 942   Register data = rscratch2;
 943   __ ic_check(1 /* end_alignment */);
 944   __ ldr(rmethod, Address(data, CompiledICData::speculated_method_offset()));




 945 
 946   // Method might have been compiled since the call site was patched to
 947   // interpreted; if that is the case treat it as a miss so we can get
 948   // the call site corrected.
 949   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 950   __ cbz(rscratch1, skip_fixup);
 951   __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
 952 }
 953 
 954 // ---------------------------------------------------------------
 955 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
 956                                             int comp_args_on_stack,
 957                                             const GrowableArray<SigEntry>* sig,
 958                                             const VMRegPair* regs,
 959                                             const GrowableArray<SigEntry>* sig_cc,
 960                                             const VMRegPair* regs_cc,
 961                                             const GrowableArray<SigEntry>* sig_cc_ro,
 962                                             const VMRegPair* regs_cc_ro,
 963                                             address entry_address[AdapterBlob::ENTRY_COUNT],
 964                                             AdapterBlob*& new_adapter,
 965                                             bool allocate_code_blob) {
 966 
 967   entry_address[AdapterBlob::I2C] = __ pc();
 968   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);

 969 
 970   // -------------------------------------------------------------------------
 971   // Generate a C2I adapter.  On entry we know rmethod holds the Method* during calls
 972   // to the interpreter.  The args start out packed in the compiled layout.  They
 973   // need to be unpacked into the interpreter layout.  This will almost always
 974   // require some stack space.  We grow the current (compiled) stack, then repack
 975   // the args.  We  finally end in a jump to the generic interpreter entry point.
 976   // On exit from the interpreter, the interpreter will restore our SP (lest the
 977   // compiled code, which relies solely on SP and not FP, get sick).
 978 
 979   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
 980   entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
 981   Label skip_fixup;




 982 
 983   gen_inline_cache_check(masm, skip_fixup);




 984 
 985   OopMapSet* oop_maps = new OopMapSet();
 986   int frame_complete = CodeOffsets::frame_never_safe;
 987   int frame_size_in_words = 0;
 988 
 989   // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
 990   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
 991   entry_address[AdapterBlob::C2I_Inline_RO] = __ pc();
 992   if (regs_cc != regs_cc_ro) {
 993     // No class init barrier needed because method is guaranteed to be non-static
 994     gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, entry_address[AdapterBlob::C2I_No_Clinit_Check],
 995                     skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
 996     skip_fixup.reset();
 997   }
 998 
 999   // Scalarized c2i adapter
1000   entry_address[AdapterBlob::C2I]        = __ pc();
1001   entry_address[AdapterBlob::C2I_Inline] = __ pc();
1002   gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1003                   skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1004 
1005   // Non-scalarized c2i adapter
1006   if (regs != regs_cc) {
1007     entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
1008     Label inline_entry_skip_fixup;
1009     gen_inline_cache_check(masm, inline_entry_skip_fixup);
1010 
1011     entry_address[AdapterBlob::C2I_Inline] = __ pc();
1012     gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1013                     inline_entry_skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1014   }
1015 
1016   // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
1017   // the GC knows about the location of oop argument locations passed to the c2i adapter.
1018   if (allocate_code_blob) {
1019     bool caller_must_gc_arguments = (regs != regs_cc);
1020     int entry_offset[AdapterHandlerEntry::ENTRIES_COUNT];
1021     assert(AdapterHandlerEntry::ENTRIES_COUNT == 7, "sanity");
1022     AdapterHandlerLibrary::address_to_offset(entry_address, entry_offset);
1023     new_adapter = AdapterBlob::create(masm->code(), entry_offset, frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1024   }






1025 }
1026 
1027 static int c_calling_convention_priv(const BasicType *sig_bt,
1028                                          VMRegPair *regs,
1029                                          int total_args_passed) {
1030 
1031 // We return the amount of VMRegImpl stack slots we need to reserve for all
1032 // the arguments NOT counting out_preserve_stack_slots.
1033 
1034     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1035       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
1036     };
1037     static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1038       c_farg0, c_farg1, c_farg2, c_farg3,
1039       c_farg4, c_farg5, c_farg6, c_farg7
1040     };
1041 
1042     uint int_args = 0;
1043     uint fp_args = 0;
1044     uint stk_args = 0; // inc by 2 each time
1045 
1046     for (int i = 0; i < total_args_passed; i++) {
1047       switch (sig_bt[i]) {
1048       case T_BOOLEAN:
1049       case T_CHAR:
1050       case T_BYTE:
1051       case T_SHORT:
1052       case T_INT:
1053         if (int_args < Argument::n_int_register_parameters_c) {
1054           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
1055         } else {
1056 #ifdef __APPLE__
1057           // Less-than word types are stored one after another.
1058           // The code is unable to handle this so bailout.
1059           return -1;
1060 #endif
1061           regs[i].set1(VMRegImpl::stack2reg(stk_args));
1062           stk_args += 2;
1063         }
1064         break;
1065       case T_LONG:
1066         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1067         // fall through
1068       case T_OBJECT:
1069       case T_ARRAY:
1070       case T_ADDRESS:
1071       case T_METADATA:
1072         if (int_args < Argument::n_int_register_parameters_c) {
1073           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
1074         } else {
1075           regs[i].set2(VMRegImpl::stack2reg(stk_args));
1076           stk_args += 2;
1077         }
1078         break;
1079       case T_FLOAT:
1080         if (fp_args < Argument::n_float_register_parameters_c) {
1081           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
1082         } else {
1083 #ifdef __APPLE__
1084           // Less-than word types are stored one after another.
1085           // The code is unable to handle this so bailout.
1086           return -1;
1087 #endif
1088           regs[i].set1(VMRegImpl::stack2reg(stk_args));
1089           stk_args += 2;
1090         }
1091         break;
1092       case T_DOUBLE:
1093         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1094         if (fp_args < Argument::n_float_register_parameters_c) {
1095           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
1096         } else {
1097           regs[i].set2(VMRegImpl::stack2reg(stk_args));
1098           stk_args += 2;
1099         }
1100         break;
1101       case T_VOID: // Halves of longs and doubles
1102         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
1103         regs[i].set_bad();
1104         break;
1105       default:
1106         ShouldNotReachHere();
1107         break;
1108       }
1109     }
1110 
1111   return stk_args;
1112 }
1113 
1114 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
1115                                              uint num_bits,
1116                                              uint total_args_passed) {
1117   // More than 8 argument inputs are not supported now.
1118   assert(total_args_passed <= Argument::n_float_register_parameters_c, "unsupported");
1119   assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
1120 
1121   static const FloatRegister VEC_ArgReg[Argument::n_float_register_parameters_c] = {
1122     v0, v1, v2, v3, v4, v5, v6, v7
1123   };
1124 
1125   // On SVE, we use the same vector registers with 128-bit vector registers on NEON.
1126   int next_reg_val = num_bits == 64 ? 1 : 3;
1127   for (uint i = 0; i < total_args_passed; i++) {
1128     VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
1129     regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
1130   }
1131   return 0;
1132 }
1133 
1134 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1135                                          VMRegPair *regs,
1136                                          int total_args_passed)
1137 {
1138   int result = c_calling_convention_priv(sig_bt, regs, total_args_passed);
1139   guarantee(result >= 0, "Unsupported arguments configuration");
1140   return result;
1141 }
1142 
1143 
1144 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1145   // We always ignore the frame_slots arg and just use the space just below frame pointer
1146   // which by this time is free to use
1147   switch (ret_type) {
1148   case T_FLOAT:
1149     __ strs(v0, Address(rfp, -wordSize));
1150     break;
1151   case T_DOUBLE:
1152     __ strd(v0, Address(rfp, -wordSize));
1153     break;
1154   case T_VOID:  break;
1155   default: {
1156     __ str(r0, Address(rfp, -wordSize));
1157     }
1158   }
1159 }
1160 
1161 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1162   // We always ignore the frame_slots arg and just use the space just below frame pointer
1163   // which by this time is free to use
1164   switch (ret_type) {
1165   case T_FLOAT:
1166     __ ldrs(v0, Address(rfp, -wordSize));
1167     break;
1168   case T_DOUBLE:
1169     __ ldrd(v0, Address(rfp, -wordSize));
1170     break;
1171   case T_VOID:  break;
1172   default: {
1173     __ ldr(r0, Address(rfp, -wordSize));
1174     }
1175   }
1176 }
1177 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1178   RegSet x;
1179   for ( int i = first_arg ; i < arg_count ; i++ ) {
1180     if (args[i].first()->is_Register()) {
1181       x = x + args[i].first()->as_Register();
1182     } else if (args[i].first()->is_FloatRegister()) {
1183       __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize)));
1184     }
1185   }
1186   __ push(x, sp);
1187 }
1188 
1189 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1190   RegSet x;
1191   for ( int i = first_arg ; i < arg_count ; i++ ) {
1192     if (args[i].first()->is_Register()) {
1193       x = x + args[i].first()->as_Register();
1194     } else {
1195       ;
1196     }
1197   }
1198   __ pop(x, sp);
1199   for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
1200     if (args[i].first()->is_Register()) {
1201       ;
1202     } else if (args[i].first()->is_FloatRegister()) {
1203       __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize)));
1204     }
1205   }
1206 }
1207 
1208 static void verify_oop_args(MacroAssembler* masm,
1209                             const methodHandle& method,
1210                             const BasicType* sig_bt,
1211                             const VMRegPair* regs) {
1212   Register temp_reg = r19;  // not part of any compiled calling seq
1213   if (VerifyOops) {
1214     for (int i = 0; i < method->size_of_parameters(); i++) {
1215       if (sig_bt[i] == T_OBJECT ||
1216           sig_bt[i] == T_ARRAY) {
1217         VMReg r = regs[i].first();
1218         assert(r->is_valid(), "bad oop arg");
1219         if (r->is_stack()) {
1220           __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1221           __ verify_oop(temp_reg);
1222         } else {
1223           __ verify_oop(r->as_Register());
1224         }
1225       }
1226     }
1227   }
1228 }
1229 
1230 // on exit, sp points to the ContinuationEntry
1231 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
1232   assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1233   assert(in_bytes(ContinuationEntry::cont_offset())  % VMRegImpl::stack_slot_size == 0, "");
1234   assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1235 
1236   stack_slots += (int)ContinuationEntry::size()/wordSize;
1237   __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
1238 
1239   OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize)/ VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1240 
1241   __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1242   __ str(rscratch1, Address(sp, ContinuationEntry::parent_offset()));
1243   __ mov(rscratch1, sp); // we can't use sp as the source in str
1244   __ str(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1245 
1246   return map;
1247 }
1248 
1249 // on entry c_rarg1 points to the continuation
1250 //          sp points to ContinuationEntry
1251 //          c_rarg3 -- isVirtualThread
1252 static void fill_continuation_entry(MacroAssembler* masm) {
1253 #ifdef ASSERT
1254   __ movw(rscratch1, ContinuationEntry::cookie_value());
1255   __ strw(rscratch1, Address(sp, ContinuationEntry::cookie_offset()));
1256 #endif
1257 
1258   __ str (c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
1259   __ strw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
1260   __ str (zr,      Address(sp, ContinuationEntry::chunk_offset()));
1261   __ strw(zr,      Address(sp, ContinuationEntry::argsize_offset()));
1262   __ strw(zr,      Address(sp, ContinuationEntry::pin_count_offset()));
1263 
1264   __ ldr(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1265   __ str(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1266 
1267   __ str(zr, Address(rthread, JavaThread::cont_fastpath_offset()));
1268 }
1269 
1270 // on entry, sp points to the ContinuationEntry
1271 // on exit, rfp points to the spilled rfp in the entry frame
1272 static void continuation_enter_cleanup(MacroAssembler* masm) {
1273 #ifndef PRODUCT
1274   Label OK;
1275   __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1276   __ cmp(sp, rscratch1);
1277   __ br(Assembler::EQ, OK);
1278   __ stop("incorrect sp1");
1279   __ bind(OK);
1280 #endif
1281   __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1282   __ str(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1283   __ ldr(rscratch2, Address(sp, ContinuationEntry::parent_offset()));
1284   __ str(rscratch2, Address(rthread, JavaThread::cont_entry_offset()));
1285   __ add(rfp, sp, (int)ContinuationEntry::size());
1286 }
1287 
1288 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1289 // On entry: c_rarg1 -- the continuation object
1290 //           c_rarg2 -- isContinue
1291 //           c_rarg3 -- isVirtualThread
1292 static void gen_continuation_enter(MacroAssembler* masm,
1293                                  const methodHandle& method,
1294                                  const BasicType* sig_bt,
1295                                  const VMRegPair* regs,
1296                                  int& exception_offset,
1297                                  OopMapSet*oop_maps,
1298                                  int& frame_complete,
1299                                  int& stack_slots,
1300                                  int& interpreted_entry_offset,
1301                                  int& compiled_entry_offset) {
1302   //verify_oop_args(masm, method, sig_bt, regs);
1303   Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1304 
1305   address start = __ pc();
1306 
1307   Label call_thaw, exit;
1308 
1309   // i2i entry used at interp_only_mode only
1310   interpreted_entry_offset = __ pc() - start;
1311   {
1312 
1313 #ifdef ASSERT
1314     Label is_interp_only;
1315     __ ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
1316     __ cbnzw(rscratch1, is_interp_only);
1317     __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1318     __ bind(is_interp_only);
1319 #endif
1320 
1321     // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1322     __ ldr(c_rarg1, Address(esp, Interpreter::stackElementSize*2));
1323     __ ldr(c_rarg2, Address(esp, Interpreter::stackElementSize*1));
1324     __ ldr(c_rarg3, Address(esp, Interpreter::stackElementSize*0));
1325     __ push_cont_fastpath(rthread);
1326 
1327     __ enter();
1328     stack_slots = 2; // will be adjusted in setup
1329     OopMap* map = continuation_enter_setup(masm, stack_slots);
1330     // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1331     // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1332 
1333     fill_continuation_entry(masm);
1334 
1335     __ cbnz(c_rarg2, call_thaw);
1336 
1337     const address tr_call = __ trampoline_call(resolve);
1338     if (tr_call == nullptr) {
1339       fatal("CodeCache is full at gen_continuation_enter");
1340     }
1341 
1342     oop_maps->add_gc_map(__ pc() - start, map);
1343     __ post_call_nop();
1344 
1345     __ b(exit);
1346 
1347     address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1348     if (stub == nullptr) {
1349       fatal("CodeCache is full at gen_continuation_enter");
1350     }
1351   }
1352 
1353   // compiled entry
1354   __ align(CodeEntryAlignment);
1355   compiled_entry_offset = __ pc() - start;
1356 
1357   __ enter();
1358   stack_slots = 2; // will be adjusted in setup
1359   OopMap* map = continuation_enter_setup(masm, stack_slots);
1360   frame_complete = __ pc() - start;
1361 
1362   fill_continuation_entry(masm);
1363 
1364   __ cbnz(c_rarg2, call_thaw);
1365 
1366   const address tr_call = __ trampoline_call(resolve);
1367   if (tr_call == nullptr) {
1368     fatal("CodeCache is full at gen_continuation_enter");
1369   }
1370 
1371   oop_maps->add_gc_map(__ pc() - start, map);
1372   __ post_call_nop();
1373 
1374   __ b(exit);
1375 
1376   __ bind(call_thaw);
1377 
1378   ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1379   __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1380   oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1381   ContinuationEntry::_return_pc_offset = __ pc() - start;
1382   __ post_call_nop();
1383 
1384   __ bind(exit);
1385   ContinuationEntry::_cleanup_offset = __ pc() - start;
1386   continuation_enter_cleanup(masm);
1387   __ leave();
1388   __ ret(lr);
1389 
1390   /// exception handling
1391 
1392   exception_offset = __ pc() - start;
1393   {
1394       __ mov(r19, r0); // save return value contaning the exception oop in callee-saved R19
1395 
1396       continuation_enter_cleanup(masm);
1397 
1398       __ ldr(c_rarg1, Address(rfp, wordSize)); // return address
1399       __ authenticate_return_address(c_rarg1);
1400       __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1);
1401 
1402       // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc
1403 
1404       __ mov(r1, r0); // the exception handler
1405       __ mov(r0, r19); // restore return value contaning the exception oop
1406       __ verify_oop(r0);
1407 
1408       __ leave();
1409       __ mov(r3, lr);
1410       __ br(r1); // the exception handler
1411   }
1412 
1413   address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1414   if (stub == nullptr) {
1415     fatal("CodeCache is full at gen_continuation_enter");
1416   }
1417 }
1418 
1419 static void gen_continuation_yield(MacroAssembler* masm,
1420                                    const methodHandle& method,
1421                                    const BasicType* sig_bt,
1422                                    const VMRegPair* regs,
1423                                    OopMapSet* oop_maps,
1424                                    int& frame_complete,
1425                                    int& stack_slots,
1426                                    int& compiled_entry_offset) {
1427     enum layout {
1428       rfp_off1,
1429       rfp_off2,
1430       lr_off,
1431       lr_off2,
1432       framesize // inclusive of return address
1433     };
1434     // assert(is_even(framesize/2), "sp not 16-byte aligned");
1435     stack_slots = framesize /  VMRegImpl::slots_per_word;
1436     assert(stack_slots == 2, "recheck layout");
1437 
1438     address start = __ pc();
1439 
1440     compiled_entry_offset = __ pc() - start;
1441     __ enter();
1442 
1443     __ mov(c_rarg1, sp);
1444 
1445     frame_complete = __ pc() - start;
1446     address the_pc = __ pc();
1447 
1448     __ 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
1449 
1450     __ mov(c_rarg0, rthread);
1451     __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
1452     __ call_VM_leaf(Continuation::freeze_entry(), 2);
1453     __ reset_last_Java_frame(true);
1454 
1455     Label pinned;
1456 
1457     __ cbnz(r0, pinned);
1458 
1459     // We've succeeded, set sp to the ContinuationEntry
1460     __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1461     __ mov(sp, rscratch1);
1462     continuation_enter_cleanup(masm);
1463 
1464     __ bind(pinned); // pinned -- return to caller
1465 
1466     // handle pending exception thrown by freeze
1467     __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
1468     Label ok;
1469     __ cbz(rscratch1, ok);
1470     __ leave();
1471     __ lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
1472     __ br(rscratch1);
1473     __ bind(ok);
1474 
1475     __ leave();
1476     __ ret(lr);
1477 
1478     OopMap* map = new OopMap(framesize, 1);
1479     oop_maps->add_gc_map(the_pc - start, map);
1480 }
1481 
1482 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1483   ::continuation_enter_cleanup(masm);
1484 }
1485 
1486 static void gen_special_dispatch(MacroAssembler* masm,
1487                                  const methodHandle& method,
1488                                  const BasicType* sig_bt,
1489                                  const VMRegPair* regs) {
1490   verify_oop_args(masm, method, sig_bt, regs);
1491   vmIntrinsics::ID iid = method->intrinsic_id();
1492 
1493   // Now write the args into the outgoing interpreter space
1494   bool     has_receiver   = false;
1495   Register receiver_reg   = noreg;
1496   int      member_arg_pos = -1;
1497   Register member_reg     = noreg;
1498   int      ref_kind       = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1499   if (ref_kind != 0) {
1500     member_arg_pos = method->size_of_parameters() - 1;  // trailing MemberName argument
1501     member_reg = r19;  // known to be free at this point
1502     has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1503   } else if (iid == vmIntrinsics::_invokeBasic) {
1504     has_receiver = true;
1505   } else if (iid == vmIntrinsics::_linkToNative) {
1506     member_arg_pos = method->size_of_parameters() - 1;  // trailing NativeEntryPoint argument
1507     member_reg = r19;  // known to be free at this point
1508   } else {
1509     fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1510   }
1511 
1512   if (member_reg != noreg) {
1513     // Load the member_arg into register, if necessary.
1514     SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1515     VMReg r = regs[member_arg_pos].first();
1516     if (r->is_stack()) {
1517       __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1518     } else {
1519       // no data motion is needed
1520       member_reg = r->as_Register();
1521     }
1522   }
1523 
1524   if (has_receiver) {
1525     // Make sure the receiver is loaded into a register.
1526     assert(method->size_of_parameters() > 0, "oob");
1527     assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1528     VMReg r = regs[0].first();
1529     assert(r->is_valid(), "bad receiver arg");
1530     if (r->is_stack()) {
1531       // Porting note:  This assumes that compiled calling conventions always
1532       // pass the receiver oop in a register.  If this is not true on some
1533       // platform, pick a temp and load the receiver from stack.
1534       fatal("receiver always in a register");
1535       receiver_reg = r2;  // known to be free at this point
1536       __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1537     } else {
1538       // no data motion is needed
1539       receiver_reg = r->as_Register();
1540     }
1541   }
1542 
1543   // Figure out which address we are really jumping to:
1544   MethodHandles::generate_method_handle_dispatch(masm, iid,
1545                                                  receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1546 }
1547 
1548 // ---------------------------------------------------------------------------
1549 // Generate a native wrapper for a given method.  The method takes arguments
1550 // in the Java compiled code convention, marshals them to the native
1551 // convention (handlizes oops, etc), transitions to native, makes the call,
1552 // returns to java state (possibly blocking), unhandlizes any result and
1553 // returns.
1554 //
1555 // Critical native functions are a shorthand for the use of
1556 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1557 // functions.  The wrapper is expected to unpack the arguments before
1558 // passing them to the callee. Critical native functions leave the state _in_Java,
1559 // since they block out GC.
1560 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1561 // block and the check for pending exceptions it's impossible for them
1562 // to be thrown.
1563 //
1564 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1565                                                 const methodHandle& method,
1566                                                 int compile_id,
1567                                                 BasicType* in_sig_bt,
1568                                                 VMRegPair* in_regs,
1569                                                 BasicType ret_type) {
1570   if (method->is_continuation_native_intrinsic()) {
1571     int exception_offset = -1;
1572     OopMapSet* oop_maps = new OopMapSet();
1573     int frame_complete = -1;
1574     int stack_slots = -1;
1575     int interpreted_entry_offset = -1;
1576     int vep_offset = -1;
1577     if (method->is_continuation_enter_intrinsic()) {
1578       gen_continuation_enter(masm,
1579                              method,
1580                              in_sig_bt,
1581                              in_regs,
1582                              exception_offset,
1583                              oop_maps,
1584                              frame_complete,
1585                              stack_slots,
1586                              interpreted_entry_offset,
1587                              vep_offset);
1588     } else if (method->is_continuation_yield_intrinsic()) {
1589       gen_continuation_yield(masm,
1590                              method,
1591                              in_sig_bt,
1592                              in_regs,
1593                              oop_maps,
1594                              frame_complete,
1595                              stack_slots,
1596                              vep_offset);
1597     } else {
1598       guarantee(false, "Unknown Continuation native intrinsic");
1599     }
1600 
1601 #ifdef ASSERT
1602     if (method->is_continuation_enter_intrinsic()) {
1603       assert(interpreted_entry_offset != -1, "Must be set");
1604       assert(exception_offset != -1,         "Must be set");
1605     } else {
1606       assert(interpreted_entry_offset == -1, "Must be unset");
1607       assert(exception_offset == -1,         "Must be unset");
1608     }
1609     assert(frame_complete != -1,    "Must be set");
1610     assert(stack_slots != -1,       "Must be set");
1611     assert(vep_offset != -1,        "Must be set");
1612 #endif
1613 
1614     __ flush();
1615     nmethod* nm = nmethod::new_native_nmethod(method,
1616                                               compile_id,
1617                                               masm->code(),
1618                                               vep_offset,
1619                                               frame_complete,
1620                                               stack_slots,
1621                                               in_ByteSize(-1),
1622                                               in_ByteSize(-1),
1623                                               oop_maps,
1624                                               exception_offset);
1625     if (nm == nullptr) return nm;
1626     if (method->is_continuation_enter_intrinsic()) {
1627       ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1628     } else if (method->is_continuation_yield_intrinsic()) {
1629       _cont_doYield_stub = nm;
1630     } else {
1631       guarantee(false, "Unknown Continuation native intrinsic");
1632     }
1633     return nm;
1634   }
1635 
1636   if (method->is_method_handle_intrinsic()) {
1637     vmIntrinsics::ID iid = method->intrinsic_id();
1638     intptr_t start = (intptr_t)__ pc();
1639     int vep_offset = ((intptr_t)__ pc()) - start;
1640 
1641     // First instruction must be a nop as it may need to be patched on deoptimisation
1642     __ nop();
1643     gen_special_dispatch(masm,
1644                          method,
1645                          in_sig_bt,
1646                          in_regs);
1647     int frame_complete = ((intptr_t)__ pc()) - start;  // not complete, period
1648     __ flush();
1649     int stack_slots = SharedRuntime::out_preserve_stack_slots();  // no out slots at all, actually
1650     return nmethod::new_native_nmethod(method,
1651                                        compile_id,
1652                                        masm->code(),
1653                                        vep_offset,
1654                                        frame_complete,
1655                                        stack_slots / VMRegImpl::slots_per_word,
1656                                        in_ByteSize(-1),
1657                                        in_ByteSize(-1),
1658                                        nullptr);
1659   }
1660   address native_func = method->native_function();
1661   assert(native_func != nullptr, "must have function");
1662 
1663   // An OopMap for lock (and class if static)
1664   OopMapSet *oop_maps = new OopMapSet();
1665   intptr_t start = (intptr_t)__ pc();
1666 
1667   // We have received a description of where all the java arg are located
1668   // on entry to the wrapper. We need to convert these args to where
1669   // the jni function will expect them. To figure out where they go
1670   // we convert the java signature to a C signature by inserting
1671   // the hidden arguments as arg[0] and possibly arg[1] (static method)
1672 
1673   const int total_in_args = method->size_of_parameters();
1674   int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1675 
1676   BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1677   VMRegPair* out_regs   = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1678 
1679   int argc = 0;
1680   out_sig_bt[argc++] = T_ADDRESS;
1681   if (method->is_static()) {
1682     out_sig_bt[argc++] = T_OBJECT;
1683   }
1684 
1685   for (int i = 0; i < total_in_args ; i++ ) {
1686     out_sig_bt[argc++] = in_sig_bt[i];
1687   }
1688 
1689   // Now figure out where the args must be stored and how much stack space
1690   // they require.
1691   int out_arg_slots;
1692   out_arg_slots = c_calling_convention_priv(out_sig_bt, out_regs, total_c_args);
1693 
1694   if (out_arg_slots < 0) {
1695     return nullptr;
1696   }
1697 
1698   // Compute framesize for the wrapper.  We need to handlize all oops in
1699   // incoming registers
1700 
1701   // Calculate the total number of stack slots we will need.
1702 
1703   // First count the abi requirement plus all of the outgoing args
1704   int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1705 
1706   // Now the space for the inbound oop handle area
1707   int total_save_slots = 8 * VMRegImpl::slots_per_word;  // 8 arguments passed in registers
1708 
1709   int oop_handle_offset = stack_slots;
1710   stack_slots += total_save_slots;
1711 
1712   // Now any space we need for handlizing a klass if static method
1713 
1714   int klass_slot_offset = 0;
1715   int klass_offset = -1;
1716   int lock_slot_offset = 0;
1717   bool is_static = false;
1718 
1719   if (method->is_static()) {
1720     klass_slot_offset = stack_slots;
1721     stack_slots += VMRegImpl::slots_per_word;
1722     klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1723     is_static = true;
1724   }
1725 
1726   // Plus a lock if needed
1727 
1728   if (method->is_synchronized()) {
1729     lock_slot_offset = stack_slots;
1730     stack_slots += VMRegImpl::slots_per_word;
1731   }
1732 
1733   // Now a place (+2) to save return values or temp during shuffling
1734   // + 4 for return address (which we own) and saved rfp
1735   stack_slots += 6;
1736 
1737   // Ok The space we have allocated will look like:
1738   //
1739   //
1740   // FP-> |                     |
1741   //      |---------------------|
1742   //      | 2 slots for moves   |
1743   //      |---------------------|
1744   //      | lock box (if sync)  |
1745   //      |---------------------| <- lock_slot_offset
1746   //      | klass (if static)   |
1747   //      |---------------------| <- klass_slot_offset
1748   //      | oopHandle area      |
1749   //      |---------------------| <- oop_handle_offset (8 java arg registers)
1750   //      | outbound memory     |
1751   //      | based arguments     |
1752   //      |                     |
1753   //      |---------------------|
1754   //      |                     |
1755   // SP-> | out_preserved_slots |
1756   //
1757   //
1758 
1759 
1760   // Now compute actual number of stack words we need rounding to make
1761   // stack properly aligned.
1762   stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1763 
1764   int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1765 
1766   // First thing make an ic check to see if we should even be here
1767 
1768   // We are free to use all registers as temps without saving them and
1769   // restoring them except rfp. rfp is the only callee save register
1770   // as far as the interpreter and the compiler(s) are concerned.
1771 
1772   const Register receiver = j_rarg0;
1773 
1774   Label exception_pending;
1775 
1776   assert_different_registers(receiver, rscratch1);
1777   __ verify_oop(receiver);
1778   __ ic_check(8 /* end_alignment */);
1779 
1780   // Verified entry point must be aligned
1781   int vep_offset = ((intptr_t)__ pc()) - start;
1782 
1783   // If we have to make this method not-entrant we'll overwrite its
1784   // first instruction with a jump.  For this action to be legal we
1785   // must ensure that this first instruction is a B, BL, NOP, BKPT,
1786   // SVC, HVC, or SMC.  Make it a NOP.
1787   __ nop();
1788 
1789   if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1790     Label L_skip_barrier;
1791     __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass*
1792     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
1793     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1794 
1795     __ bind(L_skip_barrier);
1796   }
1797 
1798   // Generate stack overflow check
1799   __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1800 
1801   // Generate a new frame for the wrapper.
1802   __ enter();
1803   // -2 because return address is already present and so is saved rfp
1804   __ sub(sp, sp, stack_size - 2*wordSize);
1805 
1806   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1807   bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1808 
1809   // Frame is now completed as far as size and linkage.
1810   int frame_complete = ((intptr_t)__ pc()) - start;
1811 
1812   // We use r20 as the oop handle for the receiver/klass
1813   // It is callee save so it survives the call to native
1814 
1815   const Register oop_handle_reg = r20;
1816 
1817   //
1818   // We immediately shuffle the arguments so that any vm call we have to
1819   // make from here on out (sync slow path, jvmti, etc.) we will have
1820   // captured the oops from our caller and have a valid oopMap for
1821   // them.
1822 
1823   // -----------------
1824   // The Grand Shuffle
1825 
1826   // The Java calling convention is either equal (linux) or denser (win64) than the
1827   // c calling convention. However the because of the jni_env argument the c calling
1828   // convention always has at least one more (and two for static) arguments than Java.
1829   // Therefore if we move the args from java -> c backwards then we will never have
1830   // a register->register conflict and we don't have to build a dependency graph
1831   // and figure out how to break any cycles.
1832   //
1833 
1834   // Record esp-based slot for receiver on stack for non-static methods
1835   int receiver_offset = -1;
1836 
1837   // This is a trick. We double the stack slots so we can claim
1838   // the oops in the caller's frame. Since we are sure to have
1839   // more args than the caller doubling is enough to make
1840   // sure we can capture all the incoming oop args from the
1841   // caller.
1842   //
1843   OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1844 
1845   // Mark location of rfp (someday)
1846   // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp));
1847 
1848 
1849   int float_args = 0;
1850   int int_args = 0;
1851 
1852 #ifdef ASSERT
1853   bool reg_destroyed[Register::number_of_registers];
1854   bool freg_destroyed[FloatRegister::number_of_registers];
1855   for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1856     reg_destroyed[r] = false;
1857   }
1858   for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1859     freg_destroyed[f] = false;
1860   }
1861 
1862 #endif /* ASSERT */
1863 
1864   // For JNI natives the incoming and outgoing registers are offset upwards.
1865   GrowableArray<int> arg_order(2 * total_in_args);
1866 
1867   for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1868     arg_order.push(i);
1869     arg_order.push(c_arg);
1870   }
1871 
1872   for (int ai = 0; ai < arg_order.length(); ai += 2) {
1873     int i = arg_order.at(ai);
1874     int c_arg = arg_order.at(ai + 1);
1875     __ block_comment(err_msg("move %d -> %d", i, c_arg));
1876     assert(c_arg != -1 && i != -1, "wrong order");
1877 #ifdef ASSERT
1878     if (in_regs[i].first()->is_Register()) {
1879       assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1880     } else if (in_regs[i].first()->is_FloatRegister()) {
1881       assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1882     }
1883     if (out_regs[c_arg].first()->is_Register()) {
1884       reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1885     } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1886       freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1887     }
1888 #endif /* ASSERT */
1889     switch (in_sig_bt[i]) {
1890       case T_ARRAY:
1891       case T_OBJECT:
1892         __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1893                        ((i == 0) && (!is_static)),
1894                        &receiver_offset);
1895         int_args++;
1896         break;
1897       case T_VOID:
1898         break;
1899 
1900       case T_FLOAT:
1901         __ float_move(in_regs[i], out_regs[c_arg]);
1902         float_args++;
1903         break;
1904 
1905       case T_DOUBLE:
1906         assert( i + 1 < total_in_args &&
1907                 in_sig_bt[i + 1] == T_VOID &&
1908                 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1909         __ double_move(in_regs[i], out_regs[c_arg]);
1910         float_args++;
1911         break;
1912 
1913       case T_LONG :
1914         __ long_move(in_regs[i], out_regs[c_arg]);
1915         int_args++;
1916         break;
1917 
1918       case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1919 
1920       default:
1921         __ move32_64(in_regs[i], out_regs[c_arg]);
1922         int_args++;
1923     }
1924   }
1925 
1926   // point c_arg at the first arg that is already loaded in case we
1927   // need to spill before we call out
1928   int c_arg = total_c_args - total_in_args;
1929 
1930   // Pre-load a static method's oop into c_rarg1.
1931   if (method->is_static()) {
1932 
1933     //  load oop into a register
1934     __ movoop(c_rarg1,
1935               JNIHandles::make_local(method->method_holder()->java_mirror()));
1936 
1937     // Now handlize the static class mirror it's known not-null.
1938     __ str(c_rarg1, Address(sp, klass_offset));
1939     map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1940 
1941     // Now get the handle
1942     __ lea(c_rarg1, Address(sp, klass_offset));
1943     // and protect the arg if we must spill
1944     c_arg--;
1945   }
1946 
1947   // Change state to native (we save the return address in the thread, since it might not
1948   // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1949   // points into the right code segment. It does not have to be the correct return pc.
1950   // We use the same pc/oopMap repeatedly when we call out.
1951 
1952   Label native_return;
1953   if (method->is_object_wait0()) {
1954     // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1955     __ set_last_Java_frame(sp, noreg, native_return, rscratch1);
1956   } else {
1957     intptr_t the_pc = (intptr_t) __ pc();
1958     oop_maps->add_gc_map(the_pc - start, map);
1959 
1960     __ set_last_Java_frame(sp, noreg, __ pc(), rscratch1);
1961   }
1962 
1963   Label dtrace_method_entry, dtrace_method_entry_done;
1964   if (DTraceMethodProbes) {
1965     __ b(dtrace_method_entry);
1966     __ bind(dtrace_method_entry_done);
1967   }
1968 
1969   // RedefineClasses() tracing support for obsolete method entry
1970   if (log_is_enabled(Trace, redefine, class, obsolete)) {
1971     // protect the args we've loaded
1972     save_args(masm, total_c_args, c_arg, out_regs);
1973     __ mov_metadata(c_rarg1, method());
1974     __ call_VM_leaf(
1975       CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1976       rthread, c_rarg1);
1977     restore_args(masm, total_c_args, c_arg, out_regs);
1978   }
1979 
1980   // Lock a synchronized method
1981 
1982   // Register definitions used by locking and unlocking
1983 
1984   const Register swap_reg = r0;
1985   const Register obj_reg  = r19;  // Will contain the oop
1986   const Register lock_reg = r13;  // Address of compiler lock object (BasicLock)
1987   const Register old_hdr  = r13;  // value of old header at unlock time
1988   const Register lock_tmp = r14;  // Temporary used by lightweight_lock/unlock
1989   const Register tmp = lr;
1990 
1991   Label slow_path_lock;
1992   Label lock_done;
1993 
1994   if (method->is_synchronized()) {
1995     // Get the handle (the 2nd argument)
1996     __ mov(oop_handle_reg, c_rarg1);
1997 
1998     // Get address of the box
1999 
2000     __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2001 
2002     // Load the oop from the handle
2003     __ ldr(obj_reg, Address(oop_handle_reg, 0));
2004 
2005     __ lightweight_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
2006 
2007     // Slow path will re-enter here
2008     __ bind(lock_done);
2009   }
2010 
2011 
2012   // Finally just about ready to make the JNI call
2013 
2014   // get JNIEnv* which is first argument to native
2015   __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset())));
2016 
2017   // Now set thread in native
2018   __ mov(rscratch1, _thread_in_native);
2019   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2020   __ stlrw(rscratch1, rscratch2);
2021 
2022   __ rt_call(native_func);
2023 
2024   // Verify or restore cpu control state after JNI call
2025   __ restore_cpu_control_state_after_jni(rscratch1, rscratch2);
2026 
2027   // Unpack native results.
2028   switch (ret_type) {
2029   case T_BOOLEAN: __ c2bool(r0);                     break;
2030   case T_CHAR   : __ ubfx(r0, r0, 0, 16);            break;
2031   case T_BYTE   : __ sbfx(r0, r0, 0, 8);             break;
2032   case T_SHORT  : __ sbfx(r0, r0, 0, 16);            break;
2033   case T_INT    : __ sbfx(r0, r0, 0, 32);            break;
2034   case T_DOUBLE :
2035   case T_FLOAT  :
2036     // Result is in v0 we'll save as needed
2037     break;
2038   case T_ARRAY:                 // Really a handle
2039   case T_OBJECT:                // Really a handle
2040       break; // can't de-handlize until after safepoint check
2041   case T_VOID: break;
2042   case T_LONG: break;
2043   default       : ShouldNotReachHere();
2044   }
2045 
2046   Label safepoint_in_progress, safepoint_in_progress_done;
2047 
2048   // Switch thread to "native transition" state before reading the synchronization state.
2049   // This additional state is necessary because reading and testing the synchronization
2050   // state is not atomic w.r.t. GC, as this scenario demonstrates:
2051   //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2052   //     VM thread changes sync state to synchronizing and suspends threads for GC.
2053   //     Thread A is resumed to finish this native method, but doesn't block here since it
2054   //     didn't see any synchronization is progress, and escapes.
2055   __ mov(rscratch1, _thread_in_native_trans);
2056 
2057   __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset()));
2058 
2059   // Force this write out before the read below
2060   if (!UseSystemMemoryBarrier) {
2061     __ dmb(Assembler::ISH);
2062   }
2063 
2064   __ verify_sve_vector_length();
2065 
2066   // Check for safepoint operation in progress and/or pending suspend requests.
2067   {
2068     // No need for acquire as Java threads always disarm themselves.
2069     __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
2070     __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset()));
2071     __ cbnzw(rscratch1, safepoint_in_progress);
2072     __ bind(safepoint_in_progress_done);
2073   }
2074 
2075   // change thread state
2076   __ mov(rscratch1, _thread_in_Java);
2077   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2078   __ stlrw(rscratch1, rscratch2);
2079 
2080   if (method->is_object_wait0()) {
2081     // Check preemption for Object.wait()
2082     __ ldr(rscratch1, Address(rthread, JavaThread::preempt_alternate_return_offset()));
2083     __ cbz(rscratch1, native_return);
2084     __ str(zr, Address(rthread, JavaThread::preempt_alternate_return_offset()));
2085     __ br(rscratch1);
2086     __ bind(native_return);
2087 
2088     intptr_t the_pc = (intptr_t) __ pc();
2089     oop_maps->add_gc_map(the_pc - start, map);
2090   }
2091 
2092   Label reguard;
2093   Label reguard_done;
2094   __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset()));
2095   __ cmpw(rscratch1, StackOverflow::stack_guard_yellow_reserved_disabled);
2096   __ br(Assembler::EQ, reguard);
2097   __ bind(reguard_done);
2098 
2099   // native result if any is live
2100 
2101   // Unlock
2102   Label unlock_done;
2103   Label slow_path_unlock;
2104   if (method->is_synchronized()) {
2105 
2106     // Get locked oop from the handle we passed to jni
2107     __ ldr(obj_reg, Address(oop_handle_reg, 0));
2108 
2109     // Must save r0 if if it is live now because cmpxchg must use it
2110     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2111       save_native_result(masm, ret_type, stack_slots);
2112     }
2113 
2114     __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
2115 
2116     // slow path re-enters here
2117     __ bind(unlock_done);
2118     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2119       restore_native_result(masm, ret_type, stack_slots);
2120     }
2121   }
2122 
2123   Label dtrace_method_exit, dtrace_method_exit_done;
2124   if (DTraceMethodProbes) {
2125     __ b(dtrace_method_exit);
2126     __ bind(dtrace_method_exit_done);
2127   }
2128 
2129   __ reset_last_Java_frame(false);
2130 
2131   // Unbox oop result, e.g. JNIHandles::resolve result.
2132   if (is_reference_type(ret_type)) {
2133     __ resolve_jobject(r0, r1, r2);
2134   }
2135 
2136   if (CheckJNICalls) {
2137     // clear_pending_jni_exception_check
2138     __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset()));
2139   }
2140 
2141   // reset handle block
2142   __ ldr(r2, Address(rthread, JavaThread::active_handles_offset()));
2143   __ str(zr, Address(r2, JNIHandleBlock::top_offset()));
2144 
2145   __ leave();
2146 
2147   #if INCLUDE_JFR
2148   // We need to do a poll test after unwind in case the sampler
2149   // managed to sample the native frame after returning to Java.
2150   Label L_return;
2151   __ ldr(rscratch1, Address(rthread, JavaThread::polling_word_offset()));
2152   address poll_test_pc = __ pc();
2153   __ relocate(relocInfo::poll_return_type);
2154   __ tbz(rscratch1, log2i_exact(SafepointMechanism::poll_bit()), L_return);
2155   assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
2156     "polling page return stub not created yet");
2157   address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
2158   __ adr(rscratch1, InternalAddress(poll_test_pc));
2159   __ str(rscratch1, Address(rthread, JavaThread::saved_exception_pc_offset()));
2160   __ far_jump(RuntimeAddress(stub));
2161   __ bind(L_return);
2162 #endif // INCLUDE_JFR
2163 
2164   // Any exception pending?
2165   __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2166   __ cbnz(rscratch1, exception_pending);
2167 
2168   // We're done
2169   __ ret(lr);
2170 
2171   // Unexpected paths are out of line and go here
2172 
2173   // forward the exception
2174   __ bind(exception_pending);
2175 
2176   // and forward the exception
2177   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2178 
2179   // Slow path locking & unlocking
2180   if (method->is_synchronized()) {
2181 
2182     __ block_comment("Slow path lock {");
2183     __ bind(slow_path_lock);
2184 
2185     // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2186     // args are (oop obj, BasicLock* lock, JavaThread* thread)
2187 
2188     // protect the args we've loaded
2189     save_args(masm, total_c_args, c_arg, out_regs);
2190 
2191     __ mov(c_rarg0, obj_reg);
2192     __ mov(c_rarg1, lock_reg);
2193     __ mov(c_rarg2, rthread);
2194 
2195     // Not a leaf but we have last_Java_frame setup as we want.
2196     // We don't want to unmount in case of contention since that would complicate preserving
2197     // the arguments that had already been marshalled into the native convention. So we force
2198     // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
2199     // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
2200     __ push_cont_fastpath();
2201     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
2202     __ pop_cont_fastpath();
2203     restore_args(masm, total_c_args, c_arg, out_regs);
2204 
2205 #ifdef ASSERT
2206     { Label L;
2207       __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2208       __ cbz(rscratch1, L);
2209       __ stop("no pending exception allowed on exit from monitorenter");
2210       __ bind(L);
2211     }
2212 #endif
2213     __ b(lock_done);
2214 
2215     __ block_comment("} Slow path lock");
2216 
2217     __ block_comment("Slow path unlock {");
2218     __ bind(slow_path_unlock);
2219 
2220     // If we haven't already saved the native result we must save it now as xmm registers
2221     // are still exposed.
2222 
2223     if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2224       save_native_result(masm, ret_type, stack_slots);
2225     }
2226 
2227     __ mov(c_rarg2, rthread);
2228     __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2229     __ mov(c_rarg0, obj_reg);
2230 
2231     // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2232     // NOTE that obj_reg == r19 currently
2233     __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2234     __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2235 
2236     __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2237 
2238 #ifdef ASSERT
2239     {
2240       Label L;
2241       __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2242       __ cbz(rscratch1, L);
2243       __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2244       __ bind(L);
2245     }
2246 #endif /* ASSERT */
2247 
2248     __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2249 
2250     if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2251       restore_native_result(masm, ret_type, stack_slots);
2252     }
2253     __ b(unlock_done);
2254 
2255     __ block_comment("} Slow path unlock");
2256 
2257   } // synchronized
2258 
2259   // SLOW PATH Reguard the stack if needed
2260 
2261   __ bind(reguard);
2262   save_native_result(masm, ret_type, stack_slots);
2263   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2264   restore_native_result(masm, ret_type, stack_slots);
2265   // and continue
2266   __ b(reguard_done);
2267 
2268   // SLOW PATH safepoint
2269   {
2270     __ block_comment("safepoint {");
2271     __ bind(safepoint_in_progress);
2272 
2273     // Don't use call_VM as it will see a possible pending exception and forward it
2274     // and never return here preventing us from clearing _last_native_pc down below.
2275     //
2276     save_native_result(masm, ret_type, stack_slots);
2277     __ mov(c_rarg0, rthread);
2278 #ifndef PRODUCT
2279   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2280 #endif
2281     __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)));
2282     __ blr(rscratch1);
2283 
2284     // Restore any method result value
2285     restore_native_result(masm, ret_type, stack_slots);
2286 
2287     __ b(safepoint_in_progress_done);
2288     __ block_comment("} safepoint");
2289   }
2290 
2291   // SLOW PATH dtrace support
2292   if (DTraceMethodProbes) {
2293     {
2294       __ block_comment("dtrace entry {");
2295       __ bind(dtrace_method_entry);
2296 
2297       // We have all of the arguments setup at this point. We must not touch any register
2298       // argument registers at this point (what if we save/restore them there are no oop?
2299 
2300       save_args(masm, total_c_args, c_arg, out_regs);
2301       __ mov_metadata(c_rarg1, method());
2302       __ call_VM_leaf(
2303         CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2304         rthread, c_rarg1);
2305       restore_args(masm, total_c_args, c_arg, out_regs);
2306       __ b(dtrace_method_entry_done);
2307       __ block_comment("} dtrace entry");
2308     }
2309 
2310     {
2311       __ block_comment("dtrace exit {");
2312       __ bind(dtrace_method_exit);
2313       save_native_result(masm, ret_type, stack_slots);
2314       __ mov_metadata(c_rarg1, method());
2315       __ call_VM_leaf(
2316         CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2317         rthread, c_rarg1);
2318       restore_native_result(masm, ret_type, stack_slots);
2319       __ b(dtrace_method_exit_done);
2320       __ block_comment("} dtrace exit");
2321     }
2322   }
2323 
2324   __ flush();
2325 
2326   nmethod *nm = nmethod::new_native_nmethod(method,
2327                                             compile_id,
2328                                             masm->code(),
2329                                             vep_offset,
2330                                             frame_complete,
2331                                             stack_slots / VMRegImpl::slots_per_word,
2332                                             (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2333                                             in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2334                                             oop_maps);
2335 
2336   return nm;
2337 }
2338 
2339 // this function returns the adjust size (in number of words) to a c2i adapter
2340 // activation for use during deoptimization
2341 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2342   assert(callee_locals >= callee_parameters,
2343           "test and remove; got more parms than locals");
2344   if (callee_locals < callee_parameters)
2345     return 0;                   // No adjustment for negative locals
2346   int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2347   // diff is counted in stack words
2348   return align_up(diff, 2);
2349 }
2350 
2351 
2352 //------------------------------generate_deopt_blob----------------------------
2353 void SharedRuntime::generate_deopt_blob() {
2354   // Allocate space for the code
2355   ResourceMark rm;
2356   // Setup code generation tools
2357   int pad = 0;
2358 #if INCLUDE_JVMCI
2359   if (EnableJVMCI) {
2360     pad += 512; // Increase the buffer size when compiling for JVMCI
2361   }
2362 #endif
2363   const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2364   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, BlobId::shared_deopt_id);
2365   if (blob != nullptr) {
2366     _deopt_blob = blob->as_deoptimization_blob();
2367     return;
2368   }
2369 
2370   CodeBuffer buffer(name, 2048+pad, 1024);
2371   MacroAssembler* masm = new MacroAssembler(&buffer);
2372   int frame_size_in_words;
2373   OopMap* map = nullptr;
2374   OopMapSet *oop_maps = new OopMapSet();
2375   RegisterSaver reg_save(COMPILER2_OR_JVMCI != 0);
2376 
2377   // -------------
2378   // This code enters when returning to a de-optimized nmethod.  A return
2379   // address has been pushed on the stack, and return values are in
2380   // registers.
2381   // If we are doing a normal deopt then we were called from the patched
2382   // nmethod from the point we returned to the nmethod. So the return
2383   // address on the stack is wrong by NativeCall::instruction_size
2384   // We will adjust the value so it looks like we have the original return
2385   // address on the stack (like when we eagerly deoptimized).
2386   // In the case of an exception pending when deoptimizing, we enter
2387   // with a return address on the stack that points after the call we patched
2388   // into the exception handler. We have the following register state from,
2389   // e.g., the forward exception stub (see stubGenerator_x86_64.cpp).
2390   //    r0: exception oop
2391   //    r19: exception handler
2392   //    r3: throwing pc
2393   // So in this case we simply jam r3 into the useless return address and
2394   // the stack looks just like we want.
2395   //
2396   // At this point we need to de-opt.  We save the argument return
2397   // registers.  We call the first C routine, fetch_unroll_info().  This
2398   // routine captures the return values and returns a structure which
2399   // describes the current frame size and the sizes of all replacement frames.
2400   // The current frame is compiled code and may contain many inlined
2401   // functions, each with their own JVM state.  We pop the current frame, then
2402   // push all the new frames.  Then we call the C routine unpack_frames() to
2403   // populate these frames.  Finally unpack_frames() returns us the new target
2404   // address.  Notice that callee-save registers are BLOWN here; they have
2405   // already been captured in the vframeArray at the time the return PC was
2406   // patched.
2407   address start = __ pc();
2408   Label cont;
2409 
2410   // Prolog for non exception case!
2411 
2412   // Save everything in sight.
2413   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2414 
2415   // Normal deoptimization.  Save exec mode for unpack_frames.
2416   __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved
2417   __ b(cont);
2418 
2419   int reexecute_offset = __ pc() - start;
2420 #if INCLUDE_JVMCI && !defined(COMPILER1)
2421   if (UseJVMCICompiler) {
2422     // JVMCI does not use this kind of deoptimization
2423     __ should_not_reach_here();
2424   }
2425 #endif
2426 
2427   // Reexecute case
2428   // return address is the pc describes what bci to do re-execute at
2429 
2430   // No need to update map as each call to save_live_registers will produce identical oopmap
2431   (void) reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2432 
2433   __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved
2434   __ b(cont);
2435 
2436 #if INCLUDE_JVMCI
2437   Label after_fetch_unroll_info_call;
2438   int implicit_exception_uncommon_trap_offset = 0;
2439   int uncommon_trap_offset = 0;
2440 
2441   if (EnableJVMCI) {
2442     implicit_exception_uncommon_trap_offset = __ pc() - start;
2443 
2444     __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2445     __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2446 
2447     uncommon_trap_offset = __ pc() - start;
2448 
2449     // Save everything in sight.
2450     reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2451     // fetch_unroll_info needs to call last_java_frame()
2452     Label retaddr;
2453     __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2454 
2455     __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2456     __ movw(rscratch1, -1);
2457     __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2458 
2459     __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute);
2460     __ mov(c_rarg0, rthread);
2461     __ movw(c_rarg2, rcpool); // exec mode
2462     __ lea(rscratch1,
2463            RuntimeAddress(CAST_FROM_FN_PTR(address,
2464                                            Deoptimization::uncommon_trap)));
2465     __ blr(rscratch1);
2466     __ bind(retaddr);
2467     oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2468 
2469     __ reset_last_Java_frame(false);
2470 
2471     __ b(after_fetch_unroll_info_call);
2472   } // EnableJVMCI
2473 #endif // INCLUDE_JVMCI
2474 
2475   int exception_offset = __ pc() - start;
2476 
2477   // Prolog for exception case
2478 
2479   // all registers are dead at this entry point, except for r0, and
2480   // r3 which contain the exception oop and exception pc
2481   // respectively.  Set them in TLS and fall thru to the
2482   // unpack_with_exception_in_tls entry point.
2483 
2484   __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
2485   __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
2486 
2487   int exception_in_tls_offset = __ pc() - start;
2488 
2489   // new implementation because exception oop is now passed in JavaThread
2490 
2491   // Prolog for exception case
2492   // All registers must be preserved because they might be used by LinearScan
2493   // Exceptiop oop and throwing PC are passed in JavaThread
2494   // tos: stack at point of call to method that threw the exception (i.e. only
2495   // args are on the stack, no return address)
2496 
2497   // The return address pushed by save_live_registers will be patched
2498   // later with the throwing pc. The correct value is not available
2499   // now because loading it from memory would destroy registers.
2500 
2501   // NB: The SP at this point must be the SP of the method that is
2502   // being deoptimized.  Deoptimization assumes that the frame created
2503   // here by save_live_registers is immediately below the method's SP.
2504   // This is a somewhat fragile mechanism.
2505 
2506   // Save everything in sight.
2507   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2508 
2509   // Now it is safe to overwrite any register
2510 
2511   // Deopt during an exception.  Save exec mode for unpack_frames.
2512   __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved
2513 
2514   // load throwing pc from JavaThread and patch it as the return address
2515   // of the current frame. Then clear the field in JavaThread
2516   __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2517   __ protect_return_address(r3);
2518   __ str(r3, Address(rfp, wordSize));
2519   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2520 
2521 #ifdef ASSERT
2522   // verify that there is really an exception oop in JavaThread
2523   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2524   __ verify_oop(r0);
2525 
2526   // verify that there is no pending exception
2527   Label no_pending_exception;
2528   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2529   __ cbz(rscratch1, no_pending_exception);
2530   __ stop("must not have pending exception here");
2531   __ bind(no_pending_exception);
2532 #endif
2533 
2534   __ bind(cont);
2535 
2536   // Call C code.  Need thread and this frame, but NOT official VM entry
2537   // crud.  We cannot block on this call, no GC can happen.
2538   //
2539   // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2540 
2541   // fetch_unroll_info needs to call last_java_frame().
2542 
2543   Label retaddr;
2544   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2545 #ifdef ASSERT
2546   { Label L;
2547     __ ldr(rscratch1, Address(rthread, JavaThread::last_Java_fp_offset()));
2548     __ cbz(rscratch1, L);
2549     __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2550     __ bind(L);
2551   }
2552 #endif // ASSERT
2553   __ mov(c_rarg0, rthread);
2554   __ mov(c_rarg1, rcpool);
2555   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2556   __ blr(rscratch1);
2557   __ bind(retaddr);
2558 
2559   // Need to have an oopmap that tells fetch_unroll_info where to
2560   // find any register it might need.
2561   oop_maps->add_gc_map(__ pc() - start, map);
2562 
2563   __ reset_last_Java_frame(false);
2564 
2565 #if INCLUDE_JVMCI
2566   if (EnableJVMCI) {
2567     __ bind(after_fetch_unroll_info_call);
2568   }
2569 #endif
2570 
2571   // Load UnrollBlock* into r5
2572   __ mov(r5, r0);
2573 
2574   __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset()));
2575    Label noException;
2576   __ cmpw(rcpool, Deoptimization::Unpack_exception);   // Was exception pending?
2577   __ br(Assembler::NE, noException);
2578   __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2579   // QQQ this is useless it was null above
2580   __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2581   __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
2582   __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2583 
2584   __ verify_oop(r0);
2585 
2586   // Overwrite the result registers with the exception results.
2587   __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2588   // I think this is useless
2589   // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2590 
2591   __ bind(noException);
2592 
2593   // Only register save data is on the stack.
2594   // Now restore the result registers.  Everything else is either dead
2595   // or captured in the vframeArray.
2596 
2597   // Restore fp result register
2598   __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2599   // Restore integer result register
2600   __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2601 
2602   // Pop all of the register save area off the stack
2603   __ add(sp, sp, frame_size_in_words * wordSize);
2604 
2605   // All of the register save area has been popped of the stack. Only the
2606   // return address remains.
2607 
2608   // Pop all the frames we must move/replace.
2609   //
2610   // Frame picture (youngest to oldest)
2611   // 1: self-frame (no frame link)
2612   // 2: deopting frame  (no frame link)
2613   // 3: caller of deopting frame (could be compiled/interpreted).
2614   //
2615   // Note: by leaving the return address of self-frame on the stack
2616   // and using the size of frame 2 to adjust the stack
2617   // when we are done the return to frame 3 will still be on the stack.
2618 
2619   // Pop deoptimized frame
2620   __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2621   __ sub(r2, r2, 2 * wordSize);
2622   __ add(sp, sp, r2);
2623   __ ldp(rfp, zr, __ post(sp, 2 * wordSize));
2624 
2625 #ifdef ASSERT
2626   // Compilers generate code that bang the stack by as much as the
2627   // interpreter would need. So this stack banging should never
2628   // trigger a fault. Verify that it does not on non product builds.
2629   __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2630   __ bang_stack_size(r19, r2);
2631 #endif
2632   // Load address of array of frame pcs into r2
2633   __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset()));
2634 
2635   // Trash the old pc
2636   // __ addptr(sp, wordSize);  FIXME ????
2637 
2638   // Load address of array of frame sizes into r4
2639   __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset()));
2640 
2641   // Load counter into r3
2642   __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset()));
2643 
2644   // Now adjust the caller's stack to make up for the extra locals
2645   // but record the original sp so that we can save it in the skeletal interpreter
2646   // frame and the stack walking of interpreter_sender will get the unextended sp
2647   // value and not the "real" sp value.
2648 
2649   const Register sender_sp = r6;
2650 
2651   __ mov(sender_sp, sp);
2652   __ ldrw(r19, Address(r5,
2653                        Deoptimization::UnrollBlock::
2654                        caller_adjustment_offset()));
2655   __ sub(sp, sp, r19);
2656 
2657   // Push interpreter frames in a loop
2658   __ mov(rscratch1, (uint64_t)0xDEADDEAD);        // Make a recognizable pattern
2659   __ mov(rscratch2, rscratch1);
2660   Label loop;
2661   __ bind(loop);
2662   __ ldr(r19, Address(__ post(r4, wordSize)));          // Load frame size
2663   __ sub(r19, r19, 2*wordSize);           // We'll push pc and fp by hand
2664   __ ldr(lr, Address(__ post(r2, wordSize)));  // Load pc
2665   __ enter();                           // Save old & set new fp
2666   __ sub(sp, sp, r19);                  // Prolog
2667   // This value is corrected by layout_activation_impl
2668   __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2669   __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2670   __ mov(sender_sp, sp);               // Pass sender_sp to next frame
2671   __ sub(r3, r3, 1);                   // Decrement counter
2672   __ cbnz(r3, loop);
2673 
2674     // Re-push self-frame
2675   __ ldr(lr, Address(r2));
2676   __ enter();
2677 
2678   // Allocate a full sized register save area.  We subtract 2 because
2679   // enter() just pushed 2 words
2680   __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2681 
2682   // Restore frame locals after moving the frame
2683   __ strd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2684   __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2685 
2686   // Call C code.  Need thread but NOT official VM entry
2687   // crud.  We cannot block on this call, no GC can happen.  Call should
2688   // restore return values to their stack-slots with the new SP.
2689   //
2690   // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2691 
2692   // Use rfp because the frames look interpreted now
2693   // Don't need the precise return PC here, just precise enough to point into this code blob.
2694   address the_pc = __ pc();
2695   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
2696 
2697   __ mov(c_rarg0, rthread);
2698   __ movw(c_rarg1, rcpool); // second arg: exec_mode
2699   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2700   __ blr(rscratch1);
2701 
2702   // Set an oopmap for the call site
2703   // Use the same PC we used for the last java frame
2704   oop_maps->add_gc_map(the_pc - start,
2705                        new OopMap( frame_size_in_words, 0 ));
2706 
2707   // Clear fp AND pc
2708   __ reset_last_Java_frame(true);
2709 
2710   // Collect return values
2711   __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2712   __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2713   // I think this is useless (throwing pc?)
2714   // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2715 
2716   // Pop self-frame.
2717   __ leave();                           // Epilog
2718 
2719   // Jump to interpreter
2720   __ ret(lr);
2721 
2722   // Make sure all code is generated
2723   masm->flush();
2724 
2725   _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2726   _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2727 #if INCLUDE_JVMCI
2728   if (EnableJVMCI) {
2729     _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2730     _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2731   }
2732 #endif
2733 
2734   AOTCodeCache::store_code_blob(*_deopt_blob, AOTCodeEntry::SharedBlob, BlobId::shared_deopt_id);
2735 }
2736 
2737 // Number of stack slots between incoming argument block and the start of
2738 // a new frame.  The PROLOG must add this many slots to the stack.  The
2739 // EPILOG must remove this many slots. aarch64 needs two slots for
2740 // return address and fp.
2741 // TODO think this is correct but check
2742 uint SharedRuntime::in_preserve_stack_slots() {
2743   return 4;
2744 }
2745 
2746 uint SharedRuntime::out_preserve_stack_slots() {
2747   return 0;
2748 }
2749 
2750 
2751 VMReg SharedRuntime::thread_register() {
2752   return rthread->as_VMReg();
2753 }
2754 
2755 //------------------------------generate_handler_blob------
2756 //
2757 // Generate a special Compile2Runtime blob that saves all registers,
2758 // and setup oopmap.
2759 //
2760 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2761   assert(is_polling_page_id(id), "expected a polling page stub id");
2762 
2763   // Allocate space for the code.  Setup code generation tools.
2764   const char* name = SharedRuntime::stub_name(id);
2765   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2766   if (blob != nullptr) {
2767     return blob->as_safepoint_blob();
2768   }
2769 
2770   ResourceMark rm;
2771   OopMapSet *oop_maps = new OopMapSet();
2772   OopMap* map;
2773   CodeBuffer buffer(name, 2048, 1024);
2774   MacroAssembler* masm = new MacroAssembler(&buffer);
2775 
2776   address start   = __ pc();
2777   address call_pc = nullptr;
2778   int frame_size_in_words;
2779   bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2780   RegisterSaver reg_save(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2781 
2782   // When the signal occurred, the LR was either signed and stored on the stack (in which
2783   // case it will be restored from the stack before being used) or unsigned and not stored
2784   // on the stack. Stipping ensures we get the right value.
2785   __ strip_return_address();
2786 
2787   // Save Integer and Float registers.
2788   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2789 
2790   // The following is basically a call_VM.  However, we need the precise
2791   // address of the call in order to generate an oopmap. Hence, we do all the
2792   // work ourselves.
2793 
2794   Label retaddr;
2795   __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2796 
2797   // The return address must always be correct so that frame constructor never
2798   // sees an invalid pc.
2799 
2800   if (!cause_return) {
2801     // overwrite the return address pushed by save_live_registers
2802     // Additionally, r20 is a callee-saved register so we can look at
2803     // it later to determine if someone changed the return address for
2804     // us!
2805     __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset()));
2806     __ protect_return_address(r20);
2807     __ str(r20, Address(rfp, wordSize));
2808   }
2809 
2810   // Do the call
2811   __ mov(c_rarg0, rthread);
2812   __ lea(rscratch1, RuntimeAddress(call_ptr));
2813   __ blr(rscratch1);
2814   __ bind(retaddr);
2815 
2816   // Set an oopmap for the call site.  This oopmap will map all
2817   // oop-registers and debug-info registers as callee-saved.  This
2818   // will allow deoptimization at this safepoint to find all possible
2819   // debug-info recordings, as well as let GC find all oops.
2820 
2821   oop_maps->add_gc_map( __ pc() - start, map);
2822 
2823   Label noException;
2824 
2825   __ reset_last_Java_frame(false);
2826 
2827   __ membar(Assembler::LoadLoad | Assembler::LoadStore);
2828 
2829   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2830   __ cbz(rscratch1, noException);
2831 
2832   // Exception pending
2833 
2834   reg_save.restore_live_registers(masm);
2835 
2836   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2837 
2838   // No exception case
2839   __ bind(noException);
2840 
2841   Label no_adjust, bail;
2842   if (!cause_return) {
2843     // If our stashed return pc was modified by the runtime we avoid touching it
2844     __ ldr(rscratch1, Address(rfp, wordSize));
2845     __ cmp(r20, rscratch1);
2846     __ br(Assembler::NE, no_adjust);
2847     __ authenticate_return_address(r20);
2848 
2849 #ifdef ASSERT
2850     // Verify the correct encoding of the poll we're about to skip.
2851     // See NativeInstruction::is_ldrw_to_zr()
2852     __ ldrw(rscratch1, Address(r20));
2853     __ ubfx(rscratch2, rscratch1, 22, 10);
2854     __ cmpw(rscratch2, 0b1011100101);
2855     __ br(Assembler::NE, bail);
2856     __ ubfx(rscratch2, rscratch1, 0, 5);
2857     __ cmpw(rscratch2, 0b11111);
2858     __ br(Assembler::NE, bail);
2859 #endif
2860     // Adjust return pc forward to step over the safepoint poll instruction
2861     __ add(r20, r20, NativeInstruction::instruction_size);
2862     __ protect_return_address(r20);
2863     __ str(r20, Address(rfp, wordSize));
2864   }
2865 
2866   __ bind(no_adjust);
2867   // Normal exit, restore registers and exit.
2868   reg_save.restore_live_registers(masm);
2869 
2870   __ ret(lr);
2871 
2872 #ifdef ASSERT
2873   __ bind(bail);
2874   __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2875 #endif
2876 
2877   // Make sure all code is generated
2878   masm->flush();
2879 
2880   // Fill-out other meta info
2881   SafepointBlob* sp_blob = SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2882 
2883   AOTCodeCache::store_code_blob(*sp_blob, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2884   return sp_blob;
2885 }
2886 
2887 //
2888 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2889 //
2890 // Generate a stub that calls into vm to find out the proper destination
2891 // of a java call. All the argument registers are live at this point
2892 // but since this is generic code we don't know what they are and the caller
2893 // must do any gc of the args.
2894 //
2895 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2896   assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2897   assert(is_resolve_id(id), "expected a resolve stub id");
2898 
2899   const char* name = SharedRuntime::stub_name(id);
2900   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2901   if (blob != nullptr) {
2902     return blob->as_runtime_stub();
2903   }
2904 
2905   // allocate space for the code
2906   ResourceMark rm;
2907   CodeBuffer buffer(name, 1000, 512);
2908   MacroAssembler* masm                = new MacroAssembler(&buffer);
2909 
2910   int frame_size_in_words;
2911   RegisterSaver reg_save(false /* save_vectors */);
2912 
2913   OopMapSet *oop_maps = new OopMapSet();
2914   OopMap* map = nullptr;
2915 
2916   int start = __ offset();
2917 
2918   map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2919 
2920   int frame_complete = __ offset();
2921 
2922   {
2923     Label retaddr;
2924     __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2925 
2926     __ mov(c_rarg0, rthread);
2927     __ lea(rscratch1, RuntimeAddress(destination));
2928 
2929     __ blr(rscratch1);
2930     __ bind(retaddr);
2931   }
2932 
2933   // Set an oopmap for the call site.
2934   // We need this not only for callee-saved registers, but also for volatile
2935   // registers that the compiler might be keeping live across a safepoint.
2936 
2937   oop_maps->add_gc_map( __ offset() - start, map);
2938 
2939   // r0 contains the address we are going to jump to assuming no exception got installed
2940 
2941   // clear last_Java_sp
2942   __ reset_last_Java_frame(false);
2943   // check for pending exceptions
2944   Label pending;
2945   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2946   __ cbnz(rscratch1, pending);
2947 
2948   // get the returned Method*
2949   __ get_vm_result_metadata(rmethod, rthread);
2950   __ str(rmethod, Address(sp, reg_save.reg_offset_in_bytes(rmethod)));
2951 
2952   // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch
2953   __ str(r0, Address(sp, reg_save.rscratch1_offset_in_bytes()));
2954   reg_save.restore_live_registers(masm);
2955 
2956   // We are back to the original state on entry and ready to go.
2957 
2958   __ br(rscratch1);
2959 
2960   // Pending exception after the safepoint
2961 
2962   __ bind(pending);
2963 
2964   reg_save.restore_live_registers(masm);
2965 
2966   // exception pending => remove activation and forward to exception handler
2967 
2968   __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
2969 
2970   __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
2971   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2972 
2973   // -------------
2974   // make sure all code is generated
2975   masm->flush();
2976 
2977   // return the  blob
2978   // frame_size_words or bytes??
2979   RuntimeStub* rs_blob = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2980 
2981   AOTCodeCache::store_code_blob(*rs_blob, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2982   return rs_blob;
2983 }
2984 
2985 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
2986   BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
2987   if (buf == nullptr) {
2988     return nullptr;
2989   }
2990   CodeBuffer buffer(buf);
2991   short buffer_locs[20];
2992   buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
2993                                          sizeof(buffer_locs)/sizeof(relocInfo));
2994 
2995   MacroAssembler _masm(&buffer);
2996   MacroAssembler* masm = &_masm;
2997 
2998   const Array<SigEntry>* sig_vk = vk->extended_sig();
2999   const Array<VMRegPair>* regs = vk->return_regs();
3000 
3001   int pack_fields_jobject_off = __ offset();
3002   // Resolve pre-allocated buffer from JNI handle.
3003   // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
3004   Register Rresult = r14;  // See StubGenerator::generate_call_stub().
3005   __ ldr(r0, Address(Rresult));
3006   __ resolve_jobject(r0 /* value */,
3007                      rthread /* thread */,
3008                      r12 /* tmp */);
3009   __ str(r0, Address(Rresult));
3010 
3011   int pack_fields_off = __ offset();
3012 
3013   int j = 1;
3014   for (int i = 0; i < sig_vk->length(); i++) {
3015     BasicType bt = sig_vk->at(i)._bt;
3016     if (bt == T_METADATA) {
3017       continue;
3018     }
3019     if (bt == T_VOID) {
3020       if (sig_vk->at(i-1)._bt == T_LONG ||
3021           sig_vk->at(i-1)._bt == T_DOUBLE) {
3022         j++;
3023       }
3024       continue;
3025     }
3026     int off = sig_vk->at(i)._offset;
3027     VMRegPair pair = regs->at(j);
3028     VMReg r_1 = pair.first();
3029     VMReg r_2 = pair.second();
3030     Address to(r0, off);
3031     if (bt == T_FLOAT) {
3032       __ strs(r_1->as_FloatRegister(), to);
3033     } else if (bt == T_DOUBLE) {
3034       __ strd(r_1->as_FloatRegister(), to);
3035     } else {
3036       Register val = r_1->as_Register();
3037       assert_different_registers(to.base(), val, r15, r16, r17);
3038       if (is_reference_type(bt)) {
3039         __ store_heap_oop(to, val, r15, r16, r17, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
3040       } else {
3041         __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
3042       }
3043     }
3044     j++;
3045   }
3046   assert(j == regs->length(), "missed a field?");
3047   if (vk->has_nullable_atomic_layout()) {
3048     // Zero the null marker (setting it to 1 would be better but would require an additional register)
3049     __ strb(zr, Address(r0, vk->null_marker_offset()));
3050   }
3051   __ ret(lr);
3052 
3053   int unpack_fields_off = __ offset();
3054 
3055   Label skip;
3056   Label not_null;
3057   __ cbnz(r0, not_null);
3058 
3059   // Return value is null. Zero oop registers to make the GC happy.
3060   j = 1;
3061   for (int i = 0; i < sig_vk->length(); i++) {
3062     BasicType bt = sig_vk->at(i)._bt;
3063     if (bt == T_METADATA) {
3064       continue;
3065     }
3066     if (bt == T_VOID) {
3067       if (sig_vk->at(i-1)._bt == T_LONG ||
3068           sig_vk->at(i-1)._bt == T_DOUBLE) {
3069         j++;
3070       }
3071       continue;
3072     }
3073     if (bt == T_OBJECT || bt == T_ARRAY) {
3074       VMRegPair pair = regs->at(j);
3075       VMReg r_1 = pair.first();
3076       __ mov(r_1->as_Register(), zr);
3077     }
3078     j++;
3079   }
3080   __ b(skip);
3081   __ bind(not_null);
3082 
3083   j = 1;
3084   for (int i = 0; i < sig_vk->length(); i++) {
3085     BasicType bt = sig_vk->at(i)._bt;
3086     if (bt == T_METADATA) {
3087       continue;
3088     }
3089     if (bt == T_VOID) {
3090       if (sig_vk->at(i-1)._bt == T_LONG ||
3091           sig_vk->at(i-1)._bt == T_DOUBLE) {
3092         j++;
3093       }
3094       continue;
3095     }
3096     int off = sig_vk->at(i)._offset;
3097     assert(off > 0, "offset in object should be positive");
3098     VMRegPair pair = regs->at(j);
3099     VMReg r_1 = pair.first();
3100     VMReg r_2 = pair.second();
3101     Address from(r0, off);
3102     if (bt == T_FLOAT) {
3103       __ ldrs(r_1->as_FloatRegister(), from);
3104     } else if (bt == T_DOUBLE) {
3105       __ ldrd(r_1->as_FloatRegister(), from);
3106     } else if (bt == T_OBJECT || bt == T_ARRAY) {
3107       assert_different_registers(r0, r_1->as_Register());
3108       __ load_heap_oop(r_1->as_Register(), from, rscratch1, rscratch2);
3109     } else {
3110       assert(is_java_primitive(bt), "unexpected basic type");
3111       assert_different_registers(r0, r_1->as_Register());
3112       size_t size_in_bytes = type2aelembytes(bt);
3113       __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3114     }
3115     j++;
3116   }
3117   assert(j == regs->length(), "missed a field?");
3118 
3119   __ bind(skip);
3120 
3121   __ ret(lr);
3122 
3123   __ flush();
3124 
3125   return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3126 }
3127 
3128 // Continuation point for throwing of implicit exceptions that are
3129 // not handled in the current activation. Fabricates an exception
3130 // oop and initiates normal exception dispatching in this
3131 // frame. Since we need to preserve callee-saved values (currently
3132 // only for C2, but done for C1 as well) we need a callee-saved oop
3133 // map and therefore have to make these stubs into RuntimeStubs
3134 // rather than BufferBlobs.  If the compiler needs all registers to
3135 // be preserved between the fault point and the exception handler
3136 // then it must assume responsibility for that in
3137 // AbstractCompiler::continuation_for_implicit_null_exception or
3138 // continuation_for_implicit_division_by_zero_exception. All other
3139 // implicit exceptions (e.g., NullPointerException or
3140 // AbstractMethodError on entry) are either at call sites or
3141 // otherwise assume that stack unwinding will be initiated, so
3142 // caller saved registers were assumed volatile in the compiler.
3143 
3144 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
3145   assert(is_throw_id(id), "expected a throw stub id");
3146 
3147   const char* name = SharedRuntime::stub_name(id);
3148 
3149   // Information about frame layout at time of blocking runtime call.
3150   // Note that we only have to preserve callee-saved registers since
3151   // the compilers are responsible for supplying a continuation point
3152   // if they expect all registers to be preserved.
3153   // n.b. aarch64 asserts that frame::arg_reg_save_area_bytes == 0
3154   enum layout {
3155     rfp_off = 0,
3156     rfp_off2,
3157     return_off,
3158     return_off2,
3159     framesize // inclusive of return address
3160   };
3161 
3162   int insts_size = 512;
3163   int locs_size  = 64;
3164 
3165   const char* timer_msg = "SharedRuntime generate_throw_exception";
3166   TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3167 
3168   CodeBlob* blob = AOTCodeCache::load_code_blob(AOTCodeEntry::SharedBlob, StubInfo::blob(id));
3169   if (blob != nullptr) {
3170     return blob->as_runtime_stub();
3171   }
3172 
3173   ResourceMark rm;
3174   CodeBuffer code(name, insts_size, locs_size);
3175   OopMapSet* oop_maps  = new OopMapSet();
3176   MacroAssembler* masm = new MacroAssembler(&code);
3177 
3178   address start = __ pc();
3179 
3180   // This is an inlined and slightly modified version of call_VM
3181   // which has the ability to fetch the return PC out of
3182   // thread-local storage and also sets up last_Java_sp slightly
3183   // differently than the real call_VM
3184 
3185   __ enter(); // Save FP and LR before call
3186 
3187   assert(is_even(framesize/2), "sp not 16-byte aligned");
3188 
3189   // lr and fp are already in place
3190   __ sub(sp, rfp, ((uint64_t)framesize-4) << LogBytesPerInt); // prolog
3191 
3192   int frame_complete = __ pc() - start;
3193 
3194   // Set up last_Java_sp and last_Java_fp
3195   address the_pc = __ pc();
3196   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3197 
3198   __ mov(c_rarg0, rthread);
3199   BLOCK_COMMENT("call runtime_entry");
3200   __ lea(rscratch1, RuntimeAddress(runtime_entry));
3201   __ blr(rscratch1);
3202 
3203   // Generate oop map
3204   OopMap* map = new OopMap(framesize, 0);
3205 
3206   oop_maps->add_gc_map(the_pc - start, map);
3207 
3208   __ reset_last_Java_frame(true);
3209 
3210   // Reinitialize the ptrue predicate register, in case the external runtime
3211   // call clobbers ptrue reg, as we may return to SVE compiled code.
3212   __ reinitialize_ptrue();
3213 
3214   __ leave();
3215 
3216   // check for pending exceptions
3217 #ifdef ASSERT
3218   Label L;
3219   __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
3220   __ cbnz(rscratch1, L);
3221   __ should_not_reach_here();
3222   __ bind(L);
3223 #endif // ASSERT
3224   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3225 
3226   // codeBlob framesize is in words (not VMRegImpl::slot_size)
3227   RuntimeStub* stub =
3228     RuntimeStub::new_runtime_stub(name,
3229                                   &code,
3230                                   frame_complete,
3231                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3232                                   oop_maps, false);
3233   AOTCodeCache::store_code_blob(*stub, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
3234 
3235   return stub;
3236 }
3237 
3238 #if INCLUDE_JFR
3239 
3240 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
3241   __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3242   __ mov(c_rarg0, thread);
3243 }
3244 
3245 // The handle is dereferenced through a load barrier.
3246 static void jfr_epilogue(MacroAssembler* masm) {
3247   __ reset_last_Java_frame(true);
3248 }
3249 
3250 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3251 // It returns a jobject handle to the event writer.
3252 // The handle is dereferenced and the return value is the event writer oop.
3253 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3254   enum layout {
3255     rbp_off,
3256     rbpH_off,
3257     return_off,
3258     return_off2,
3259     framesize // inclusive of return address
3260   };
3261 
3262   int insts_size = 1024;
3263   int locs_size = 64;
3264   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
3265   CodeBuffer code(name, insts_size, locs_size);
3266   OopMapSet* oop_maps = new OopMapSet();
3267   MacroAssembler* masm = new MacroAssembler(&code);
3268 
3269   address start = __ pc();
3270   __ enter();
3271   int frame_complete = __ pc() - start;
3272   address the_pc = __ pc();
3273   jfr_prologue(the_pc, masm, rthread);
3274   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
3275   jfr_epilogue(masm);
3276   __ resolve_global_jobject(r0, rscratch1, rscratch2);
3277   __ leave();
3278   __ ret(lr);
3279 
3280   OopMap* map = new OopMap(framesize, 1); // rfp
3281   oop_maps->add_gc_map(the_pc - start, map);
3282 
3283   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3284     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3285                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3286                                   oop_maps, false);
3287   return stub;
3288 }
3289 
3290 // For c2: call to return a leased buffer.
3291 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3292   enum layout {
3293     rbp_off,
3294     rbpH_off,
3295     return_off,
3296     return_off2,
3297     framesize // inclusive of return address
3298   };
3299 
3300   int insts_size = 1024;
3301   int locs_size = 64;
3302 
3303   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
3304   CodeBuffer code(name, insts_size, locs_size);
3305   OopMapSet* oop_maps = new OopMapSet();
3306   MacroAssembler* masm = new MacroAssembler(&code);
3307 
3308   address start = __ pc();
3309   __ enter();
3310   int frame_complete = __ pc() - start;
3311   address the_pc = __ pc();
3312   jfr_prologue(the_pc, masm, rthread);
3313   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
3314   jfr_epilogue(masm);
3315 
3316   __ leave();
3317   __ ret(lr);
3318 
3319   OopMap* map = new OopMap(framesize, 1); // rfp
3320   oop_maps->add_gc_map(the_pc - start, map);
3321 
3322   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3323     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3324                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3325                                   oop_maps, false);
3326   return stub;
3327 }
3328 
3329 #endif // INCLUDE_JFR
--- EOF ---