1 /*
   2  * Copyright (c) 2003, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2014, 2020, Red Hat Inc. All rights reserved.
   4  * Copyright (c) 2020, 2023, Huawei Technologies Co., Ltd. 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 "code/compiledIC.hpp"
  30 #include "code/debugInfoRec.hpp"
  31 #include "code/vtableStubs.hpp"
  32 #include "compiler/oopMap.hpp"
  33 #include "gc/shared/barrierSetAssembler.hpp"
  34 #include "interpreter/interp_masm.hpp"
  35 #include "interpreter/interpreter.hpp"
  36 #include "logging/log.hpp"
  37 #include "memory/resourceArea.hpp"
  38 #include "nativeInst_riscv.hpp"
  39 #include "oops/klass.inline.hpp"
  40 #include "oops/method.inline.hpp"
  41 #include "prims/methodHandles.hpp"
  42 #include "runtime/continuation.hpp"
  43 #include "runtime/continuationEntry.inline.hpp"
  44 #include "runtime/globals.hpp"
  45 #include "runtime/jniHandles.hpp"
  46 #include "runtime/safepointMechanism.hpp"
  47 #include "runtime/sharedRuntime.hpp"
  48 #include "runtime/signature.hpp"
  49 #include "runtime/stubRoutines.hpp"
  50 #include "runtime/timerTrace.hpp"
  51 #include "runtime/vframeArray.hpp"
  52 #include "utilities/align.hpp"
  53 #include "utilities/formatBuffer.hpp"
  54 #include "vmreg_riscv.inline.hpp"
  55 #ifdef COMPILER1
  56 #include "c1/c1_Runtime1.hpp"
  57 #endif
  58 #ifdef COMPILER2
  59 #include "adfiles/ad_riscv.hpp"
  60 #include "opto/runtime.hpp"
  61 #endif
  62 #if INCLUDE_JVMCI
  63 #include "jvmci/jvmciJavaClasses.hpp"
  64 #endif
  65 
  66 #define __ masm->
  67 
  68 #ifdef PRODUCT
  69 #define BLOCK_COMMENT(str) /* nothing */
  70 #else
  71 #define BLOCK_COMMENT(str) __ block_comment(str)
  72 #endif
  73 
  74 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
  75 
  76 class RegisterSaver {
  77   const bool _save_vectors;
  78  public:
  79   RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
  80   ~RegisterSaver() {}
  81   OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
  82   void restore_live_registers(MacroAssembler* masm);
  83 
  84   // Offsets into the register save area
  85   // Used by deoptimization when it is managing result register
  86   // values on its own
  87   // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
  88   // |---v0---|<---SP
  89   // |---v1---|save vectors only in generate_handler_blob
  90   // |-- .. --|
  91   // |---v31--|-----
  92   // |---f0---|
  93   // |---f1---|
  94   // |   ..   |
  95   // |---f31--|
  96   // |---reserved slot for stack alignment---|
  97   // |---x5---|
  98   // |   x6   |
  99   // |---.. --|
 100   // |---x31--|
 101   // |---fp---|
 102   // |---ra---|
 103   int v0_offset_in_bytes(void) { return 0; }
 104   int f0_offset_in_bytes(void) {
 105     int f0_offset = 0;
 106 #ifdef COMPILER2
 107     if (_save_vectors) {
 108       f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
 109                    BytesPerInt;
 110     }
 111 #endif
 112     return f0_offset;
 113   }
 114   int reserved_slot_offset_in_bytes(void) {
 115     return f0_offset_in_bytes() +
 116            FloatRegister::max_slots_per_register *
 117            FloatRegister::number_of_registers *
 118            BytesPerInt;
 119   }
 120 
 121   int reg_offset_in_bytes(Register r) {
 122     assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
 123     return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
 124   }
 125 
 126   int freg_offset_in_bytes(FloatRegister f) {
 127     return f0_offset_in_bytes() + f->encoding() * wordSize;
 128   }
 129 
 130   int ra_offset_in_bytes(void) {
 131     return reserved_slot_offset_in_bytes() +
 132            (Register::number_of_registers - 3) *
 133            Register::max_slots_per_register *
 134            BytesPerInt;
 135   }
 136 };
 137 
 138 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
 139   int vector_size_in_bytes = 0;
 140   int vector_size_in_slots = 0;
 141 #ifdef COMPILER2
 142   if (_save_vectors) {
 143     vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
 144     vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
 145   }
 146 #endif
 147 
 148   int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
 149   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 150   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 151   // The caller will allocate additional_frame_words
 152   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 153   // CodeBlob frame size is in words.
 154   int frame_size_in_words = frame_size_in_bytes / wordSize;
 155   *total_frame_words = frame_size_in_words;
 156 
 157   // Save Integer, Float and Vector registers.
 158   __ enter();
 159   __ push_CPU_state(_save_vectors, vector_size_in_bytes);
 160 
 161   // Set an oopmap for the call site.  This oopmap will map all
 162   // oop-registers and debug-info registers as callee-saved.  This
 163   // will allow deoptimization at this safepoint to find all possible
 164   // debug-info recordings, as well as let GC find all oops.
 165 
 166   OopMapSet *oop_maps = new OopMapSet();
 167   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 168   assert_cond(oop_maps != nullptr && oop_map != nullptr);
 169 
 170   int sp_offset_in_slots = 0;
 171   int step_in_slots = 0;
 172   if (_save_vectors) {
 173     step_in_slots = vector_size_in_slots;
 174     for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
 175       VectorRegister r = as_VectorRegister(i);
 176       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
 177     }
 178   }
 179 
 180   step_in_slots = FloatRegister::max_slots_per_register;
 181   for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
 182     FloatRegister r = as_FloatRegister(i);
 183     oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
 184   }
 185 
 186   step_in_slots = Register::max_slots_per_register;
 187   // skip the slot reserved for alignment, see MacroAssembler::push_reg;
 188   // also skip x5 ~ x6 on the stack because they are caller-saved registers.
 189   sp_offset_in_slots += Register::max_slots_per_register * 3;
 190   // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
 191   for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
 192     Register r = as_Register(i);
 193     if (r != xthread) {
 194       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
 195     }
 196   }
 197 
 198   return oop_map;
 199 }
 200 
 201 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
 202 #ifdef COMPILER2
 203   __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
 204 #else
 205 #if !INCLUDE_JVMCI
 206   assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
 207 #endif
 208   __ pop_CPU_state(_save_vectors);
 209 #endif
 210   __ leave();
 211 }
 212 
 213 // Is vector's size (in bytes) bigger than a size saved by default?
 214 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
 215 bool SharedRuntime::is_wide_vector(int size) {
 216   return UseRVV && size > 0;
 217 }
 218 
 219 // ---------------------------------------------------------------------------
 220 // Read the array of BasicTypes from a signature, and compute where the
 221 // arguments should go.  Values in the VMRegPair regs array refer to 4-byte
 222 // quantities.  Values less than VMRegImpl::stack0 are registers, those above
 223 // refer to 4-byte stack slots.  All stack slots are based off of the stack pointer
 224 // as framesizes are fixed.
 225 // VMRegImpl::stack0 refers to the first slot 0(sp).
 226 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
 227 // Register up to Register::number_of_registers) are the 64-bit
 228 // integer registers.
 229 
 230 // Note: the INPUTS in sig_bt are in units of Java argument words,
 231 // which are 64-bit.  The OUTPUTS are in 32-bit units.
 232 
 233 // The Java calling convention is a "shifted" version of the C ABI.
 234 // By skipping the first C ABI register we can call non-static jni
 235 // methods with small numbers of arguments without having to shuffle
 236 // the arguments at all. Since we control the java ABI we ought to at
 237 // least get some advantage out of it.
 238 
 239 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
 240                                            VMRegPair *regs,
 241                                            int total_args_passed) {
 242   // Create the mapping between argument positions and
 243   // registers.
 244   static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
 245     j_rarg0, j_rarg1, j_rarg2, j_rarg3,
 246     j_rarg4, j_rarg5, j_rarg6, j_rarg7
 247   };
 248   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
 249     j_farg0, j_farg1, j_farg2, j_farg3,
 250     j_farg4, j_farg5, j_farg6, j_farg7
 251   };
 252 
 253   uint int_args = 0;
 254   uint fp_args = 0;
 255   uint stk_args = 0;
 256 
 257   for (int i = 0; i < total_args_passed; i++) {
 258     switch (sig_bt[i]) {
 259       case T_BOOLEAN: // fall through
 260       case T_CHAR:    // fall through
 261       case T_BYTE:    // fall through
 262       case T_SHORT:   // fall through
 263       case T_INT:
 264         if (int_args < Argument::n_int_register_parameters_j) {
 265           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 266         } else {
 267           stk_args = align_up(stk_args, 2);
 268           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 269           stk_args += 1;
 270         }
 271         break;
 272       case T_VOID:
 273         // halves of T_LONG or T_DOUBLE
 274         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 275         regs[i].set_bad();
 276         break;
 277       case T_LONG:      // fall through
 278         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 279       case T_OBJECT:    // fall through
 280       case T_ARRAY:     // fall through
 281       case T_ADDRESS:
 282         if (int_args < Argument::n_int_register_parameters_j) {
 283           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 284         } else {
 285           stk_args = align_up(stk_args, 2);
 286           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 287           stk_args += 2;
 288         }
 289         break;
 290       case T_FLOAT:
 291         if (fp_args < Argument::n_float_register_parameters_j) {
 292           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 293         } else {
 294           stk_args = align_up(stk_args, 2);
 295           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 296           stk_args += 1;
 297         }
 298         break;
 299       case T_DOUBLE:
 300         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 301         if (fp_args < Argument::n_float_register_parameters_j) {
 302           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 303         } else {
 304           stk_args = align_up(stk_args, 2);
 305           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 306           stk_args += 2;
 307         }
 308         break;
 309       default:
 310         ShouldNotReachHere();
 311     }
 312   }
 313 
 314   return stk_args;
 315 }
 316 
 317 // Patch the callers callsite with entry to compiled code if it exists.
 318 static void patch_callers_callsite(MacroAssembler *masm) {
 319   Label L;
 320   __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
 321   __ beqz(t0, L);
 322 
 323   __ enter();
 324   __ push_CPU_state();
 325 
 326   // VM needs caller's callsite
 327   // VM needs target method
 328   // This needs to be a long call since we will relocate this adapter to
 329   // the codeBuffer and it may not reach
 330 
 331 #ifndef PRODUCT
 332   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 333 #endif
 334 
 335   __ mv(c_rarg0, xmethod);
 336   __ mv(c_rarg1, ra);
 337   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
 338 
 339   __ pop_CPU_state();
 340   // restore sp
 341   __ leave();
 342   __ bind(L);
 343 }
 344 
 345 static void gen_c2i_adapter(MacroAssembler *masm,
 346                             int total_args_passed,
 347                             int comp_args_on_stack,
 348                             const BasicType *sig_bt,
 349                             const VMRegPair *regs,
 350                             Label& skip_fixup) {
 351   // Before we get into the guts of the C2I adapter, see if we should be here
 352   // at all.  We've come from compiled code and are attempting to jump to the
 353   // interpreter, which means the caller made a static call to get here
 354   // (vcalls always get a compiled target if there is one).  Check for a
 355   // compiled target.  If there is one, we need to patch the caller's call.
 356   patch_callers_callsite(masm);
 357 
 358   __ bind(skip_fixup);
 359 
 360   int words_pushed = 0;
 361 
 362   // Since all args are passed on the stack, total_args_passed *
 363   // Interpreter::stackElementSize is the space we need.
 364 
 365   int extraspace = total_args_passed * Interpreter::stackElementSize;
 366 
 367   __ mv(x19_sender_sp, sp);
 368 
 369   // stack is aligned, keep it that way
 370   extraspace = align_up(extraspace, 2 * wordSize);
 371 
 372   if (extraspace) {
 373     __ sub(sp, sp, extraspace);
 374   }
 375 
 376   // Now write the args into the outgoing interpreter space
 377   for (int i = 0; i < total_args_passed; i++) {
 378     if (sig_bt[i] == T_VOID) {
 379       assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
 380       continue;
 381     }
 382 
 383     // offset to start parameters
 384     int st_off   = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 385     int next_off = st_off - Interpreter::stackElementSize;
 386 
 387     // Say 4 args:
 388     // i   st_off
 389     // 0   32 T_LONG
 390     // 1   24 T_VOID
 391     // 2   16 T_OBJECT
 392     // 3    8 T_BOOL
 393     // -    0 return address
 394     //
 395     // However to make thing extra confusing. Because we can fit a Java long/double in
 396     // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 397     // leaves one slot empty and only stores to a single slot. In this case the
 398     // slot that is occupied is the T_VOID slot. See I said it was confusing.
 399 
 400     VMReg r_1 = regs[i].first();
 401     VMReg r_2 = regs[i].second();
 402     if (!r_1->is_valid()) {
 403       assert(!r_2->is_valid(), "");
 404       continue;
 405     }
 406     if (r_1->is_stack()) {
 407       // memory to memory use t0
 408       int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
 409                     + extraspace
 410                     + words_pushed * wordSize);
 411       if (!r_2->is_valid()) {
 412         __ lwu(t0, Address(sp, ld_off));
 413         __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 414       } else {
 415         __ ld(t0, Address(sp, ld_off), /*temp register*/esp);
 416 
 417         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 418         // T_DOUBLE and T_LONG use two slots in the interpreter
 419         if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 420           // ld_off == LSW, ld_off+wordSize == MSW
 421           // st_off == MSW, next_off == LSW
 422           __ sd(t0, Address(sp, next_off), /*temp register*/esp);
 423 #ifdef ASSERT
 424           // Overwrite the unused slot with known junk
 425           __ mv(t0, 0xdeadffffdeadaaaaul);
 426           __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 427 #endif /* ASSERT */
 428         } else {
 429           __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 430         }
 431       }
 432     } else if (r_1->is_Register()) {
 433       Register r = r_1->as_Register();
 434       if (!r_2->is_valid()) {
 435         // must be only an int (or less ) so move only 32bits to slot
 436         __ sd(r, Address(sp, st_off));
 437       } else {
 438         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 439         // T_DOUBLE and T_LONG use two slots in the interpreter
 440         if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 441           // long/double in gpr
 442 #ifdef ASSERT
 443           // Overwrite the unused slot with known junk
 444           __ mv(t0, 0xdeadffffdeadaaabul);
 445           __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 446 #endif /* ASSERT */
 447           __ sd(r, Address(sp, next_off));
 448         } else {
 449           __ sd(r, Address(sp, st_off));
 450         }
 451       }
 452     } else {
 453       assert(r_1->is_FloatRegister(), "");
 454       if (!r_2->is_valid()) {
 455         // only a float use just part of the slot
 456         __ fsw(r_1->as_FloatRegister(), Address(sp, st_off));
 457       } else {
 458 #ifdef ASSERT
 459         // Overwrite the unused slot with known junk
 460         __ mv(t0, 0xdeadffffdeadaaacul);
 461         __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 462 #endif /* ASSERT */
 463         __ fsd(r_1->as_FloatRegister(), Address(sp, next_off));
 464       }
 465     }
 466   }
 467 
 468   __ mv(esp, sp); // Interp expects args on caller's expression stack
 469 
 470   __ ld(t1, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
 471   __ jr(t1);
 472 }
 473 
 474 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
 475                                     int total_args_passed,
 476                                     int comp_args_on_stack,
 477                                     const BasicType *sig_bt,
 478                                     const VMRegPair *regs) {
 479   // Note: x19_sender_sp contains the senderSP on entry. We must
 480   // preserve it since we may do a i2c -> c2i transition if we lose a
 481   // race where compiled code goes non-entrant while we get args
 482   // ready.
 483 
 484   // Cut-out for having no stack args.
 485   int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
 486   if (comp_args_on_stack != 0) {
 487     __ sub(t0, sp, comp_words_on_stack * wordSize);
 488     __ andi(sp, t0, -16);
 489   }
 490 
 491   // Will jump to the compiled code just as if compiled code was doing it.
 492   // Pre-load the register-jump target early, to schedule it better.
 493   __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset())));
 494 
 495 #if INCLUDE_JVMCI
 496   if (EnableJVMCI) {
 497     // check if this call should be routed towards a specific entry point
 498     __ ld(t0, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 499     Label no_alternative_target;
 500     __ beqz(t0, no_alternative_target);
 501     __ mv(t1, t0);
 502     __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 503     __ bind(no_alternative_target);
 504   }
 505 #endif // INCLUDE_JVMCI
 506 
 507   // Now generate the shuffle code.
 508   for (int i = 0; i < total_args_passed; i++) {
 509     if (sig_bt[i] == T_VOID) {
 510       assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
 511       continue;
 512     }
 513 
 514     // Pick up 0, 1 or 2 words from SP+offset.
 515 
 516     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
 517            "scrambled load targets?");
 518     // Load in argument order going down.
 519     int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 520     // Point to interpreter value (vs. tag)
 521     int next_off = ld_off - Interpreter::stackElementSize;
 522 
 523     VMReg r_1 = regs[i].first();
 524     VMReg r_2 = regs[i].second();
 525     if (!r_1->is_valid()) {
 526       assert(!r_2->is_valid(), "");
 527       continue;
 528     }
 529     if (r_1->is_stack()) {
 530       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 531       int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
 532       if (!r_2->is_valid()) {
 533         __ lw(t0, Address(esp, ld_off));
 534         __ sd(t0, Address(sp, st_off), /*temp register*/t2);
 535       } else {
 536         //
 537         // We are using two optoregs. This can be either T_OBJECT,
 538         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 539         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 540         // So we must adjust where to pick up the data to match the
 541         // interpreter.
 542         //
 543         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 544         // are accessed as negative so LSW is at LOW address
 545 
 546         // ld_off is MSW so get LSW
 547         const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
 548                            next_off : ld_off;
 549         __ ld(t0, Address(esp, offset));
 550         // st_off is LSW (i.e. reg.first())
 551         __ sd(t0, Address(sp, st_off), /*temp register*/t2);
 552       }
 553     } else if (r_1->is_Register()) {  // Register argument
 554       Register r = r_1->as_Register();
 555       if (r_2->is_valid()) {
 556         //
 557         // We are using two VMRegs. This can be either T_OBJECT,
 558         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 559         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 560         // So we must adjust where to pick up the data to match the
 561         // interpreter.
 562 
 563         const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
 564                            next_off : ld_off;
 565 
 566         // this can be a misaligned move
 567         __ ld(r, Address(esp, offset));
 568       } else {
 569         // sign extend and use a full word?
 570         __ lw(r, Address(esp, ld_off));
 571       }
 572     } else {
 573       if (!r_2->is_valid()) {
 574         __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
 575       } else {
 576         __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
 577       }
 578     }
 579   }
 580 
 581   __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
 582 
 583   // 6243940 We might end up in handle_wrong_method if
 584   // the callee is deoptimized as we race thru here. If that
 585   // happens we don't want to take a safepoint because the
 586   // caller frame will look interpreted and arguments are now
 587   // "compiled" so it is much better to make this transition
 588   // invisible to the stack walking code. Unfortunately if
 589   // we try and find the callee by normal means a safepoint
 590   // is possible. So we stash the desired callee in the thread
 591   // and the vm will find there should this case occur.
 592 
 593   __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
 594 
 595   __ jr(t1);
 596 }
 597 
 598 // ---------------------------------------------------------------
 599 
 600 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
 601                                             int total_args_passed,
 602                                             int comp_args_on_stack,
 603                                             const BasicType *sig_bt,
 604                                             const VMRegPair *regs,
 605                                             address entry_address[AdapterBlob::ENTRY_COUNT]) {
 606   entry_address[AdapterBlob::I2C] = __ pc();
 607   gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
 608 
 609   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
 610   Label skip_fixup;
 611 
 612   const Register receiver = j_rarg0;
 613   const Register data = t0;
 614 
 615   // -------------------------------------------------------------------------
 616   // Generate a C2I adapter.  On entry we know xmethod holds the Method* during calls
 617   // to the interpreter.  The args start out packed in the compiled layout.  They
 618   // need to be unpacked into the interpreter layout.  This will almost always
 619   // require some stack space.  We grow the current (compiled) stack, then repack
 620   // the args.  We  finally end in a jump to the generic interpreter entry point.
 621   // On exit from the interpreter, the interpreter will restore our SP (lest the
 622   // compiled code, which relies solely on SP and not FP, get sick).
 623 
 624   {
 625     __ block_comment("c2i_unverified_entry {");
 626 
 627     __ ic_check();
 628     __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
 629 
 630     __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
 631     __ beqz(t0, skip_fixup);
 632     __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
 633     __ block_comment("} c2i_unverified_entry");
 634   }
 635 
 636   entry_address[AdapterBlob::C2I] = __ pc();
 637 
 638   // Class initialization barrier for static methods
 639   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
 640   assert(VM_Version::supports_fast_class_init_checks(), "sanity");
 641   Label L_skip_barrier;
 642 
 643   // Bypass the barrier for non-static methods
 644   __ load_unsigned_short(t0, Address(xmethod, Method::access_flags_offset()));
 645   __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
 646   __ beqz(t1, L_skip_barrier); // non-static
 647 
 648   __ load_method_holder(t1, xmethod);
 649   __ clinit_barrier(t1, t0, &L_skip_barrier);
 650   __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 651 
 652   __ bind(L_skip_barrier);
 653   entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();
 654 
 655   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 656   bs->c2i_entry_barrier(masm);
 657 
 658   gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
 659   return;
 660 }
 661 
 662 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
 663                                              uint num_bits,
 664                                              uint total_args_passed) {
 665   assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
 666   assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
 667 
 668   // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
 669   static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
 670     v8, v9, v10, v11, v12, v13, v14, v15,
 671     v16, v17, v18, v19, v20, v21, v22, v23
 672   };
 673 
 674   const int next_reg_val = 3;
 675   for (uint i = 0; i < total_args_passed; i++) {
 676     VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
 677     regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
 678   }
 679   return 0;
 680 }
 681 
 682 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
 683                                          VMRegPair *regs,
 684                                          int total_args_passed) {
 685 
 686   // We return the amount of VMRegImpl stack slots we need to reserve for all
 687   // the arguments NOT counting out_preserve_stack_slots.
 688 
 689   static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
 690     c_rarg0, c_rarg1, c_rarg2, c_rarg3,
 691     c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
 692   };
 693   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
 694     c_farg0, c_farg1, c_farg2, c_farg3,
 695     c_farg4, c_farg5, c_farg6, c_farg7
 696   };
 697 
 698   uint int_args = 0;
 699   uint fp_args = 0;
 700   uint stk_args = 0; // inc by 2 each time
 701 
 702   for (int i = 0; i < total_args_passed; i++) {
 703     switch (sig_bt[i]) {
 704       case T_BOOLEAN:  // fall through
 705       case T_CHAR:     // fall through
 706       case T_BYTE:     // fall through
 707       case T_SHORT:    // fall through
 708       case T_INT:
 709         if (int_args < Argument::n_int_register_parameters_c) {
 710           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 711         } else {
 712           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 713           stk_args += 2;
 714         }
 715         break;
 716       case T_LONG:      // fall through
 717         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 718       case T_OBJECT:    // fall through
 719       case T_ARRAY:     // fall through
 720       case T_ADDRESS:   // fall through
 721       case T_METADATA:
 722         if (int_args < Argument::n_int_register_parameters_c) {
 723           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 724         } else {
 725           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 726           stk_args += 2;
 727         }
 728         break;
 729       case T_FLOAT:
 730         if (fp_args < Argument::n_float_register_parameters_c) {
 731           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 732         } else if (int_args < Argument::n_int_register_parameters_c) {
 733           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 734         } else {
 735           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 736           stk_args += 2;
 737         }
 738         break;
 739       case T_DOUBLE:
 740         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 741         if (fp_args < Argument::n_float_register_parameters_c) {
 742           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 743         } else if (int_args < Argument::n_int_register_parameters_c) {
 744           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 745         } else {
 746           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 747           stk_args += 2;
 748         }
 749         break;
 750       case T_VOID: // Halves of longs and doubles
 751         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 752         regs[i].set_bad();
 753         break;
 754       default:
 755         ShouldNotReachHere();
 756     }
 757   }
 758 
 759   return stk_args;
 760 }
 761 
 762 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
 763   // We always ignore the frame_slots arg and just use the space just below frame pointer
 764   // which by this time is free to use
 765   switch (ret_type) {
 766     case T_FLOAT:
 767       __ fsw(f10, Address(fp, -3 * wordSize));
 768       break;
 769     case T_DOUBLE:
 770       __ fsd(f10, Address(fp, -3 * wordSize));
 771       break;
 772     case T_VOID:  break;
 773     default: {
 774       __ sd(x10, Address(fp, -3 * wordSize));
 775     }
 776   }
 777 }
 778 
 779 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
 780   // We always ignore the frame_slots arg and just use the space just below frame pointer
 781   // which by this time is free to use
 782   switch (ret_type) {
 783     case T_FLOAT:
 784       __ flw(f10, Address(fp, -3 * wordSize));
 785       break;
 786     case T_DOUBLE:
 787       __ fld(f10, Address(fp, -3 * wordSize));
 788       break;
 789     case T_VOID:  break;
 790     default: {
 791       __ ld(x10, Address(fp, -3 * wordSize));
 792     }
 793   }
 794 }
 795 
 796 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
 797   RegSet x;
 798   for ( int i = first_arg ; i < arg_count ; i++ ) {
 799     if (args[i].first()->is_Register()) {
 800       x = x + args[i].first()->as_Register();
 801     } else if (args[i].first()->is_FloatRegister()) {
 802       __ subi(sp, sp, 2 * wordSize);
 803       __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
 804     }
 805   }
 806   __ push_reg(x, sp);
 807 }
 808 
 809 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
 810   RegSet x;
 811   for ( int i = first_arg ; i < arg_count ; i++ ) {
 812     if (args[i].first()->is_Register()) {
 813       x = x + args[i].first()->as_Register();
 814     } else {
 815       ;
 816     }
 817   }
 818   __ pop_reg(x, sp);
 819   for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
 820     if (args[i].first()->is_Register()) {
 821       ;
 822     } else if (args[i].first()->is_FloatRegister()) {
 823       __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
 824       __ addi(sp, sp, 2 * wordSize);
 825     }
 826   }
 827 }
 828 
 829 static void verify_oop_args(MacroAssembler* masm,
 830                             const methodHandle& method,
 831                             const BasicType* sig_bt,
 832                             const VMRegPair* regs) {
 833   const Register temp_reg = x9;  // not part of any compiled calling seq
 834   if (VerifyOops) {
 835     for (int i = 0; i < method->size_of_parameters(); i++) {
 836       if (sig_bt[i] == T_OBJECT ||
 837           sig_bt[i] == T_ARRAY) {
 838         VMReg r = regs[i].first();
 839         assert(r->is_valid(), "bad oop arg");
 840         if (r->is_stack()) {
 841           __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
 842           __ verify_oop(temp_reg);
 843         } else {
 844           __ verify_oop(r->as_Register());
 845         }
 846       }
 847     }
 848   }
 849 }
 850 
 851 // on exit, sp points to the ContinuationEntry
 852 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
 853   assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
 854   assert(in_bytes(ContinuationEntry::cont_offset())  % VMRegImpl::stack_slot_size == 0, "");
 855   assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
 856 
 857   stack_slots += (int)ContinuationEntry::size() / wordSize;
 858   __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
 859 
 860   OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
 861 
 862   __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
 863   __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
 864   __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
 865 
 866   return map;
 867 }
 868 
 869 // on entry c_rarg1 points to the continuation
 870 //          sp points to ContinuationEntry
 871 //          c_rarg3 -- isVirtualThread
 872 static void fill_continuation_entry(MacroAssembler* masm) {
 873 #ifdef ASSERT
 874   __ mv(t0, ContinuationEntry::cookie_value());
 875   __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
 876 #endif
 877 
 878   __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
 879   __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
 880   __ sd(zr,      Address(sp, ContinuationEntry::chunk_offset()));
 881   __ sw(zr,      Address(sp, ContinuationEntry::argsize_offset()));
 882   __ sw(zr,      Address(sp, ContinuationEntry::pin_count_offset()));
 883 
 884   __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
 885   __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
 886 
 887   __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
 888 }
 889 
 890 // on entry, sp points to the ContinuationEntry
 891 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
 892 static void continuation_enter_cleanup(MacroAssembler* masm) {
 893 #ifndef PRODUCT
 894   Label OK;
 895   __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
 896   __ beq(sp, t0, OK);
 897   __ stop("incorrect sp");
 898   __ bind(OK);
 899 #endif
 900 
 901   __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
 902   __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
 903   __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
 904   __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
 905   __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
 906 }
 907 
 908 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
 909 // On entry: c_rarg1 -- the continuation object
 910 //           c_rarg2 -- isContinue
 911 //           c_rarg3 -- isVirtualThread
 912 static void gen_continuation_enter(MacroAssembler* masm,
 913                                    const methodHandle& method,
 914                                    const BasicType* sig_bt,
 915                                    const VMRegPair* regs,
 916                                    int& exception_offset,
 917                                    OopMapSet*oop_maps,
 918                                    int& frame_complete,
 919                                    int& stack_slots,
 920                                    int& interpreted_entry_offset,
 921                                    int& compiled_entry_offset) {
 922   // verify_oop_args(masm, method, sig_bt, regs);
 923   Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
 924 
 925   address start = __ pc();
 926 
 927   Label call_thaw, exit;
 928 
 929   // i2i entry used at interp_only_mode only
 930   interpreted_entry_offset = __ pc() - start;
 931   {
 932 #ifdef ASSERT
 933     Label is_interp_only;
 934     __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
 935     __ bnez(t0, is_interp_only);
 936     __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
 937     __ bind(is_interp_only);
 938 #endif
 939 
 940     // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
 941     __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
 942     __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
 943     __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
 944     __ push_cont_fastpath(xthread);
 945 
 946     __ enter();
 947     stack_slots = 2; // will be adjusted in setup
 948     OopMap* map = continuation_enter_setup(masm, stack_slots);
 949     // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
 950     // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
 951 
 952     fill_continuation_entry(masm);
 953 
 954     __ bnez(c_rarg2, call_thaw);
 955 
 956     address call_pc;
 957     {
 958       Assembler::IncompressibleScope scope(masm);
 959       // Make sure the call is patchable
 960       __ align(NativeInstruction::instruction_size);
 961 
 962       call_pc = __ reloc_call(resolve);
 963       if (call_pc == nullptr) {
 964         fatal("CodeCache is full at gen_continuation_enter");
 965       }
 966 
 967       oop_maps->add_gc_map(__ pc() - start, map);
 968       __ post_call_nop();
 969     }
 970     __ j(exit);
 971 
 972     address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
 973     if (stub == nullptr) {
 974       fatal("CodeCache is full at gen_continuation_enter");
 975     }
 976   }
 977 
 978   // compiled entry
 979   __ align(CodeEntryAlignment);
 980   compiled_entry_offset = __ pc() - start;
 981 
 982   __ enter();
 983   stack_slots = 2; // will be adjusted in setup
 984   OopMap* map = continuation_enter_setup(masm, stack_slots);
 985   frame_complete = __ pc() - start;
 986 
 987   fill_continuation_entry(masm);
 988 
 989   __ bnez(c_rarg2, call_thaw);
 990 
 991   address call_pc;
 992   {
 993     Assembler::IncompressibleScope scope(masm);
 994     // Make sure the call is patchable
 995     __ align(NativeInstruction::instruction_size);
 996 
 997     call_pc = __ reloc_call(resolve);
 998     if (call_pc == nullptr) {
 999       fatal("CodeCache is full at gen_continuation_enter");
1000     }
1001 
1002     oop_maps->add_gc_map(__ pc() - start, map);
1003     __ post_call_nop();
1004   }
1005 
1006   __ j(exit);
1007 
1008   __ bind(call_thaw);
1009 
1010   // Post call nops must be natural aligned due to cmodx rules.
1011   {
1012     Assembler::IncompressibleScope scope(masm);
1013     __ align(NativeInstruction::instruction_size);
1014 
1015     ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1016     __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1017     oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1018     ContinuationEntry::_return_pc_offset = __ pc() - start;
1019     __ post_call_nop();
1020   }
1021 
1022   __ bind(exit);
1023   ContinuationEntry::_cleanup_offset = __ pc() - start;
1024   continuation_enter_cleanup(masm);
1025   __ leave();
1026   __ ret();
1027 
1028   // exception handling
1029   exception_offset = __ pc() - start;
1030   {
1031     __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1032 
1033     continuation_enter_cleanup(masm);
1034 
1035     __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1036     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1037 
1038     // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1039 
1040     __ mv(x11, x10); // the exception handler
1041     __ mv(x10, x9); // restore return value contaning the exception oop
1042     __ verify_oop(x10);
1043 
1044     __ leave();
1045     __ mv(x13, ra);
1046     __ jr(x11); // the exception handler
1047   }
1048 
1049   address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1050   if (stub == nullptr) {
1051     fatal("CodeCache is full at gen_continuation_enter");
1052   }
1053 }
1054 
1055 static void gen_continuation_yield(MacroAssembler* masm,
1056                                    const methodHandle& method,
1057                                    const BasicType* sig_bt,
1058                                    const VMRegPair* regs,
1059                                    OopMapSet* oop_maps,
1060                                    int& frame_complete,
1061                                    int& stack_slots,
1062                                    int& compiled_entry_offset) {
1063   enum layout {
1064     fp_off,
1065     fp_off2,
1066     return_off,
1067     return_off2,
1068     framesize // inclusive of return address
1069   };
1070   // assert(is_even(framesize/2), "sp not 16-byte aligned");
1071 
1072   stack_slots = framesize / VMRegImpl::slots_per_word;
1073   assert(stack_slots == 2, "recheck layout");
1074 
1075   address start = __ pc();
1076 
1077   compiled_entry_offset = __ pc() - start;
1078   __ enter();
1079 
1080   __ mv(c_rarg1, sp);
1081 
1082   // Post call nops must be natural aligned due to cmodx rules.
1083   __ align(NativeInstruction::instruction_size);
1084 
1085   frame_complete = __ pc() - start;
1086   address the_pc = __ pc();
1087 
1088   {
1089     Assembler::IncompressibleScope scope(masm);
1090     __ 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
1091   }
1092 
1093   __ mv(c_rarg0, xthread);
1094   __ set_last_Java_frame(sp, fp, the_pc, t0);
1095   __ call_VM_leaf(Continuation::freeze_entry(), 2);
1096   __ reset_last_Java_frame(true);
1097 
1098   Label pinned;
1099 
1100   __ bnez(x10, pinned);
1101 
1102   // We've succeeded, set sp to the ContinuationEntry
1103   __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1104   continuation_enter_cleanup(masm);
1105 
1106   __ bind(pinned); // pinned -- return to caller
1107 
1108   // handle pending exception thrown by freeze
1109   __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1110   Label ok;
1111   __ beqz(t0, ok);
1112   __ leave();
1113   __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1114   __ bind(ok);
1115 
1116   __ leave();
1117   __ ret();
1118 
1119   OopMap* map = new OopMap(framesize, 1);
1120   oop_maps->add_gc_map(the_pc - start, map);
1121 }
1122 
1123 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1124   ::continuation_enter_cleanup(masm);
1125 }
1126 
1127 static void gen_special_dispatch(MacroAssembler* masm,
1128                                  const methodHandle& method,
1129                                  const BasicType* sig_bt,
1130                                  const VMRegPair* regs) {
1131   verify_oop_args(masm, method, sig_bt, regs);
1132   vmIntrinsics::ID iid = method->intrinsic_id();
1133 
1134   // Now write the args into the outgoing interpreter space
1135   bool     has_receiver   = false;
1136   Register receiver_reg   = noreg;
1137   int      member_arg_pos = -1;
1138   Register member_reg     = noreg;
1139   int      ref_kind       = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1140   if (ref_kind != 0) {
1141     member_arg_pos = method->size_of_parameters() - 1;  // trailing MemberName argument
1142     member_reg = x9;  // known to be free at this point
1143     has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1144   } else if (iid == vmIntrinsics::_invokeBasic) {
1145     has_receiver = true;
1146   } else if (iid == vmIntrinsics::_linkToNative) {
1147     member_arg_pos = method->size_of_parameters() - 1;  // trailing NativeEntryPoint argument
1148     member_reg = x9;  // known to be free at this point
1149   } else {
1150     fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1151   }
1152 
1153   if (member_reg != noreg) {
1154     // Load the member_arg into register, if necessary.
1155     SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1156     VMReg r = regs[member_arg_pos].first();
1157     if (r->is_stack()) {
1158       __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1159     } else {
1160       // no data motion is needed
1161       member_reg = r->as_Register();
1162     }
1163   }
1164 
1165   if (has_receiver) {
1166     // Make sure the receiver is loaded into a register.
1167     assert(method->size_of_parameters() > 0, "oob");
1168     assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1169     VMReg r = regs[0].first();
1170     assert(r->is_valid(), "bad receiver arg");
1171     if (r->is_stack()) {
1172       // Porting note:  This assumes that compiled calling conventions always
1173       // pass the receiver oop in a register.  If this is not true on some
1174       // platform, pick a temp and load the receiver from stack.
1175       fatal("receiver always in a register");
1176       receiver_reg = x12;  // known to be free at this point
1177       __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1178     } else {
1179       // no data motion is needed
1180       receiver_reg = r->as_Register();
1181     }
1182   }
1183 
1184   // Figure out which address we are really jumping to:
1185   MethodHandles::generate_method_handle_dispatch(masm, iid,
1186                                                  receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1187 }
1188 
1189 // ---------------------------------------------------------------------------
1190 // Generate a native wrapper for a given method.  The method takes arguments
1191 // in the Java compiled code convention, marshals them to the native
1192 // convention (handlizes oops, etc), transitions to native, makes the call,
1193 // returns to java state (possibly blocking), unhandlizes any result and
1194 // returns.
1195 //
1196 // Critical native functions are a shorthand for the use of
1197 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1198 // functions.  The wrapper is expected to unpack the arguments before
1199 // passing them to the callee and perform checks before and after the
1200 // native call to ensure that they GCLocker
1201 // lock_critical/unlock_critical semantics are followed.  Some other
1202 // parts of JNI setup are skipped like the tear down of the JNI handle
1203 // block and the check for pending exceptions it's impossible for them
1204 // to be thrown.
1205 //
1206 // They are roughly structured like this:
1207 //    if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1208 //    tranistion to thread_in_native
1209 //    unpack array arguments and call native entry point
1210 //    check for safepoint in progress
1211 //    check if any thread suspend flags are set
1212 //      call into JVM and possible unlock the JNI critical
1213 //      if a GC was suppressed while in the critical native.
1214 //    transition back to thread_in_Java
1215 //    return to caller
1216 //
1217 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1218                                                 const methodHandle& method,
1219                                                 int compile_id,
1220                                                 BasicType* in_sig_bt,
1221                                                 VMRegPair* in_regs,
1222                                                 BasicType ret_type) {
1223   if (method->is_continuation_native_intrinsic()) {
1224     int exception_offset = -1;
1225     OopMapSet* oop_maps = new OopMapSet();
1226     int frame_complete = -1;
1227     int stack_slots = -1;
1228     int interpreted_entry_offset = -1;
1229     int vep_offset = -1;
1230     if (method->is_continuation_enter_intrinsic()) {
1231       gen_continuation_enter(masm,
1232                              method,
1233                              in_sig_bt,
1234                              in_regs,
1235                              exception_offset,
1236                              oop_maps,
1237                              frame_complete,
1238                              stack_slots,
1239                              interpreted_entry_offset,
1240                              vep_offset);
1241     } else if (method->is_continuation_yield_intrinsic()) {
1242       gen_continuation_yield(masm,
1243                              method,
1244                              in_sig_bt,
1245                              in_regs,
1246                              oop_maps,
1247                              frame_complete,
1248                              stack_slots,
1249                              vep_offset);
1250     } else {
1251       guarantee(false, "Unknown Continuation native intrinsic");
1252     }
1253 
1254 #ifdef ASSERT
1255     if (method->is_continuation_enter_intrinsic()) {
1256       assert(interpreted_entry_offset != -1, "Must be set");
1257       assert(exception_offset != -1,         "Must be set");
1258     } else {
1259       assert(interpreted_entry_offset == -1, "Must be unset");
1260       assert(exception_offset == -1,         "Must be unset");
1261     }
1262     assert(frame_complete != -1,    "Must be set");
1263     assert(stack_slots != -1,       "Must be set");
1264     assert(vep_offset != -1,        "Must be set");
1265 #endif
1266 
1267     __ flush();
1268     nmethod* nm = nmethod::new_native_nmethod(method,
1269                                               compile_id,
1270                                               masm->code(),
1271                                               vep_offset,
1272                                               frame_complete,
1273                                               stack_slots,
1274                                               in_ByteSize(-1),
1275                                               in_ByteSize(-1),
1276                                               oop_maps,
1277                                               exception_offset);
1278     if (nm == nullptr) return nm;
1279     if (method->is_continuation_enter_intrinsic()) {
1280       ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1281     } else if (method->is_continuation_yield_intrinsic()) {
1282       _cont_doYield_stub = nm;
1283     } else {
1284       guarantee(false, "Unknown Continuation native intrinsic");
1285     }
1286     return nm;
1287   }
1288 
1289   if (method->is_method_handle_intrinsic()) {
1290     vmIntrinsics::ID iid = method->intrinsic_id();
1291     intptr_t start = (intptr_t)__ pc();
1292     int vep_offset = ((intptr_t)__ pc()) - start;
1293 
1294     // First instruction must be a nop as it may need to be patched on deoptimisation
1295     {
1296       Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1297       MacroAssembler::assert_alignment(__ pc());
1298       __ nop();  // 4 bytes
1299     }
1300     gen_special_dispatch(masm,
1301                          method,
1302                          in_sig_bt,
1303                          in_regs);
1304     int frame_complete = ((intptr_t)__ pc()) - start;  // not complete, period
1305     __ flush();
1306     int stack_slots = SharedRuntime::out_preserve_stack_slots();  // no out slots at all, actually
1307     return nmethod::new_native_nmethod(method,
1308                                        compile_id,
1309                                        masm->code(),
1310                                        vep_offset,
1311                                        frame_complete,
1312                                        stack_slots / VMRegImpl::slots_per_word,
1313                                        in_ByteSize(-1),
1314                                        in_ByteSize(-1),
1315                                        (OopMapSet*)nullptr);
1316   }
1317   address native_func = method->native_function();
1318   assert(native_func != nullptr, "must have function");
1319 
1320   // An OopMap for lock (and class if static)
1321   OopMapSet *oop_maps = new OopMapSet();
1322   assert_cond(oop_maps != nullptr);
1323   intptr_t start = (intptr_t)__ pc();
1324 
1325   // We have received a description of where all the java arg are located
1326   // on entry to the wrapper. We need to convert these args to where
1327   // the jni function will expect them. To figure out where they go
1328   // we convert the java signature to a C signature by inserting
1329   // the hidden arguments as arg[0] and possibly arg[1] (static method)
1330 
1331   const int total_in_args = method->size_of_parameters();
1332   int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1333 
1334   BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1335   VMRegPair* out_regs   = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1336 
1337   int argc = 0;
1338   out_sig_bt[argc++] = T_ADDRESS;
1339   if (method->is_static()) {
1340     out_sig_bt[argc++] = T_OBJECT;
1341   }
1342 
1343   for (int i = 0; i < total_in_args ; i++) {
1344     out_sig_bt[argc++] = in_sig_bt[i];
1345   }
1346 
1347   // Now figure out where the args must be stored and how much stack space
1348   // they require.
1349   int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1350 
1351   // Compute framesize for the wrapper.  We need to handlize all oops in
1352   // incoming registers
1353 
1354   // Calculate the total number of stack slots we will need.
1355 
1356   // First count the abi requirement plus all of the outgoing args
1357   int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1358 
1359   // Now the space for the inbound oop handle area
1360   int total_save_slots = 8 * VMRegImpl::slots_per_word;  // 8 arguments passed in registers
1361 
1362   int oop_handle_offset = stack_slots;
1363   stack_slots += total_save_slots;
1364 
1365   // Now any space we need for handlizing a klass if static method
1366 
1367   int klass_slot_offset = 0;
1368   int klass_offset = -1;
1369   int lock_slot_offset = 0;
1370   bool is_static = false;
1371 
1372   if (method->is_static()) {
1373     klass_slot_offset = stack_slots;
1374     stack_slots += VMRegImpl::slots_per_word;
1375     klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1376     is_static = true;
1377   }
1378 
1379   // Plus a lock if needed
1380 
1381   if (method->is_synchronized()) {
1382     lock_slot_offset = stack_slots;
1383     stack_slots += VMRegImpl::slots_per_word;
1384   }
1385 
1386   // Now a place (+2) to save return values or temp during shuffling
1387   // + 4 for return address (which we own) and saved fp
1388   stack_slots += 6;
1389 
1390   // Ok The space we have allocated will look like:
1391   //
1392   //
1393   // FP-> |                     |
1394   //      | 2 slots (ra)        |
1395   //      | 2 slots (fp)        |
1396   //      |---------------------|
1397   //      | 2 slots for moves   |
1398   //      |---------------------|
1399   //      | lock box (if sync)  |
1400   //      |---------------------| <- lock_slot_offset
1401   //      | klass (if static)   |
1402   //      |---------------------| <- klass_slot_offset
1403   //      | oopHandle area      |
1404   //      |---------------------| <- oop_handle_offset (8 java arg registers)
1405   //      | outbound memory     |
1406   //      | based arguments     |
1407   //      |                     |
1408   //      |---------------------|
1409   //      |                     |
1410   // SP-> | out_preserved_slots |
1411   //
1412   //
1413 
1414 
1415   // Now compute actual number of stack words we need rounding to make
1416   // stack properly aligned.
1417   stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1418 
1419   int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1420 
1421   // First thing make an ic check to see if we should even be here
1422 
1423   // We are free to use all registers as temps without saving them and
1424   // restoring them except fp. fp is the only callee save register
1425   // as far as the interpreter and the compiler(s) are concerned.
1426 
1427   const Register receiver = j_rarg0;
1428 
1429   __ verify_oop(receiver);
1430   assert_different_registers(receiver, t0, t1);
1431 
1432   __ ic_check();
1433 
1434   int vep_offset = ((intptr_t)__ pc()) - start;
1435 
1436   // If we have to make this method not-entrant we'll overwrite its
1437   // first instruction with a jump.
1438   {
1439     Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1440     MacroAssembler::assert_alignment(__ pc());
1441     __ nop();  // 4 bytes
1442   }
1443 
1444   if (method->needs_clinit_barrier()) {
1445     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
1446     Label L_skip_barrier;
1447     __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1448     __ clinit_barrier(t1, t0, &L_skip_barrier);
1449     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1450 
1451     __ bind(L_skip_barrier);
1452   }
1453 
1454   // Generate stack overflow check
1455   __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1456 
1457   // Generate a new frame for the wrapper.
1458   __ enter();
1459   // -2 because return address is already present and so is saved fp
1460   __ sub(sp, sp, stack_size - 2 * wordSize);
1461 
1462   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1463   assert_cond(bs != nullptr);
1464   bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1465 
1466   // Frame is now completed as far as size and linkage.
1467   int frame_complete = ((intptr_t)__ pc()) - start;
1468 
1469   // We use x18 as the oop handle for the receiver/klass
1470   // It is callee save so it survives the call to native
1471 
1472   const Register oop_handle_reg = x18;
1473 
1474   //
1475   // We immediately shuffle the arguments so that any vm call we have to
1476   // make from here on out (sync slow path, jvmti, etc.) we will have
1477   // captured the oops from our caller and have a valid oopMap for
1478   // them.
1479 
1480   // -----------------
1481   // The Grand Shuffle
1482 
1483   // The Java calling convention is either equal (linux) or denser (win64) than the
1484   // c calling convention. However the because of the jni_env argument the c calling
1485   // convention always has at least one more (and two for static) arguments than Java.
1486   // Therefore if we move the args from java -> c backwards then we will never have
1487   // a register->register conflict and we don't have to build a dependency graph
1488   // and figure out how to break any cycles.
1489   //
1490 
1491   // Record esp-based slot for receiver on stack for non-static methods
1492   int receiver_offset = -1;
1493 
1494   // This is a trick. We double the stack slots so we can claim
1495   // the oops in the caller's frame. Since we are sure to have
1496   // more args than the caller doubling is enough to make
1497   // sure we can capture all the incoming oop args from the
1498   // caller.
1499   //
1500   OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1501   assert_cond(map != nullptr);
1502 
1503   int float_args = 0;
1504   int int_args = 0;
1505 
1506 #ifdef ASSERT
1507   bool reg_destroyed[Register::number_of_registers];
1508   bool freg_destroyed[FloatRegister::number_of_registers];
1509   for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1510     reg_destroyed[r] = false;
1511   }
1512   for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1513     freg_destroyed[f] = false;
1514   }
1515 
1516 #endif /* ASSERT */
1517 
1518   // For JNI natives the incoming and outgoing registers are offset upwards.
1519   GrowableArray<int> arg_order(2 * total_in_args);
1520 
1521   for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1522     arg_order.push(i);
1523     arg_order.push(c_arg);
1524   }
1525 
1526   for (int ai = 0; ai < arg_order.length(); ai += 2) {
1527     int i = arg_order.at(ai);
1528     int c_arg = arg_order.at(ai + 1);
1529     __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1530     assert(c_arg != -1 && i != -1, "wrong order");
1531 #ifdef ASSERT
1532     if (in_regs[i].first()->is_Register()) {
1533       assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1534     } else if (in_regs[i].first()->is_FloatRegister()) {
1535       assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1536     }
1537     if (out_regs[c_arg].first()->is_Register()) {
1538       reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1539     } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1540       freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1541     }
1542 #endif /* ASSERT */
1543     switch (in_sig_bt[i]) {
1544       case T_ARRAY:
1545       case T_OBJECT:
1546         __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1547                        ((i == 0) && (!is_static)),
1548                        &receiver_offset);
1549         int_args++;
1550         break;
1551       case T_VOID:
1552         break;
1553 
1554       case T_FLOAT:
1555         __ float_move(in_regs[i], out_regs[c_arg]);
1556         float_args++;
1557         break;
1558 
1559       case T_DOUBLE:
1560         assert( i + 1 < total_in_args &&
1561                 in_sig_bt[i + 1] == T_VOID &&
1562                 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1563         __ double_move(in_regs[i], out_regs[c_arg]);
1564         float_args++;
1565         break;
1566 
1567       case T_LONG :
1568         __ long_move(in_regs[i], out_regs[c_arg]);
1569         int_args++;
1570         break;
1571 
1572       case T_ADDRESS:
1573         assert(false, "found T_ADDRESS in java args");
1574         break;
1575 
1576       default:
1577         __ move32_64(in_regs[i], out_regs[c_arg]);
1578         int_args++;
1579     }
1580   }
1581 
1582   // point c_arg at the first arg that is already loaded in case we
1583   // need to spill before we call out
1584   int c_arg = total_c_args - total_in_args;
1585 
1586   // Pre-load a static method's oop into c_rarg1.
1587   if (method->is_static()) {
1588 
1589     //  load oop into a register
1590     __ movoop(c_rarg1,
1591               JNIHandles::make_local(method->method_holder()->java_mirror()));
1592 
1593     // Now handlize the static class mirror it's known not-null.
1594     __ sd(c_rarg1, Address(sp, klass_offset));
1595     map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1596 
1597     // Now get the handle
1598     __ la(c_rarg1, Address(sp, klass_offset));
1599     // and protect the arg if we must spill
1600     c_arg--;
1601   }
1602 
1603   // Change state to native (we save the return address in the thread, since it might not
1604   // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1605   // points into the right code segment. It does not have to be the correct return pc.
1606   // We use the same pc/oopMap repeatedly when we call out.
1607 
1608   Label native_return;
1609   if (method->is_object_wait0()) {
1610     // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1611     __ set_last_Java_frame(sp, noreg, native_return, t0);
1612   } else {
1613     intptr_t the_pc = (intptr_t) __ pc();
1614     oop_maps->add_gc_map(the_pc - start, map);
1615 
1616     __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1617   }
1618 
1619   Label dtrace_method_entry, dtrace_method_entry_done;
1620   if (DTraceMethodProbes) {
1621     __ j(dtrace_method_entry);
1622     __ bind(dtrace_method_entry_done);
1623   }
1624 
1625   // RedefineClasses() tracing support for obsolete method entry
1626   if (log_is_enabled(Trace, redefine, class, obsolete)) {
1627     // protect the args we've loaded
1628     save_args(masm, total_c_args, c_arg, out_regs);
1629     __ mov_metadata(c_rarg1, method());
1630     __ call_VM_leaf(
1631       CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1632       xthread, c_rarg1);
1633     restore_args(masm, total_c_args, c_arg, out_regs);
1634   }
1635 
1636   // Lock a synchronized method
1637 
1638   // Register definitions used by locking and unlocking
1639 
1640   const Register swap_reg = x10;
1641   const Register obj_reg  = x9;  // Will contain the oop
1642   const Register lock_reg = x30;  // Address of compiler lock object (BasicLock)
1643   const Register old_hdr  = x30;  // value of old header at unlock time
1644   const Register lock_tmp = x31;  // Temporary used by fast_lock/unlock
1645   const Register tmp      = ra;
1646 
1647   Label slow_path_lock;
1648   Label lock_done;
1649 
1650   if (method->is_synchronized()) {
1651     // Get the handle (the 2nd argument)
1652     __ mv(oop_handle_reg, c_rarg1);
1653 
1654     // Get address of the box
1655 
1656     __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1657 
1658     // Load the oop from the handle
1659     __ ld(obj_reg, Address(oop_handle_reg, 0));
1660 
1661     __ fast_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1662 
1663     // Slow path will re-enter here
1664     __ bind(lock_done);
1665   }
1666 
1667 
1668   // Finally just about ready to make the JNI call
1669 
1670   // get JNIEnv* which is first argument to native
1671   __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1672 
1673   // Now set thread in native
1674   __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1675   __ mv(t0, _thread_in_native);
1676   __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1677   __ sw(t0, Address(t1));
1678 
1679   // Clobbers t1
1680   __ rt_call(native_func);
1681 
1682   // Verify or restore cpu control state after JNI call
1683   __ restore_cpu_control_state_after_jni(t0);
1684 
1685   // Unpack native results.
1686   if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1687     __ cast_primitive_type(ret_type, x10);
1688   }
1689 
1690   Label safepoint_in_progress, safepoint_in_progress_done;
1691 
1692   // Switch thread to "native transition" state before reading the synchronization state.
1693   // This additional state is necessary because reading and testing the synchronization
1694   // state is not atomic w.r.t. GC, as this scenario demonstrates:
1695   //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1696   //     VM thread changes sync state to synchronizing and suspends threads for GC.
1697   //     Thread A is resumed to finish this native method, but doesn't block here since it
1698   //     didn't see any synchronization is progress, and escapes.
1699   __ mv(t0, _thread_in_native_trans);
1700 
1701   __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1702 
1703   // Force this write out before the read below
1704   if (!UseSystemMemoryBarrier) {
1705     __ membar(MacroAssembler::AnyAny);
1706   }
1707 
1708   // check for safepoint operation in progress and/or pending suspend requests
1709   {
1710     __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
1711     __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1712     __ bnez(t0, safepoint_in_progress);
1713     __ bind(safepoint_in_progress_done);
1714   }
1715 
1716   // change thread state
1717   __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1718   __ mv(t0, _thread_in_Java);
1719   __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1720   __ sw(t0, Address(t1));
1721 
1722   if (method->is_object_wait0()) {
1723     // Check preemption for Object.wait()
1724     __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1725     __ beqz(t1, native_return);
1726     __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1727     __ jr(t1);
1728     __ bind(native_return);
1729 
1730     intptr_t the_pc = (intptr_t) __ pc();
1731     oop_maps->add_gc_map(the_pc - start, map);
1732   }
1733 
1734   Label reguard;
1735   Label reguard_done;
1736   __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1737   __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1738   __ beq(t0, t1, reguard);
1739   __ bind(reguard_done);
1740 
1741   // native result if any is live
1742 
1743   // Unlock
1744   Label unlock_done;
1745   Label slow_path_unlock;
1746   if (method->is_synchronized()) {
1747 
1748     // Get locked oop from the handle we passed to jni
1749     __ ld(obj_reg, Address(oop_handle_reg, 0));
1750 
1751     // Must save x10 if if it is live now because cmpxchg must use it
1752     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1753       save_native_result(masm, ret_type, stack_slots);
1754     }
1755 
1756     __ fast_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1757 
1758     // slow path re-enters here
1759     __ bind(unlock_done);
1760     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1761       restore_native_result(masm, ret_type, stack_slots);
1762     }
1763   }
1764 
1765   Label dtrace_method_exit, dtrace_method_exit_done;
1766   if (DTraceMethodProbes) {
1767     __ j(dtrace_method_exit);
1768     __ bind(dtrace_method_exit_done);
1769   }
1770 
1771   __ reset_last_Java_frame(false);
1772 
1773   // Unbox oop result, e.g. JNIHandles::resolve result.
1774   if (is_reference_type(ret_type)) {
1775     __ resolve_jobject(x10, x11, x12);
1776   }
1777 
1778   if (CheckJNICalls) {
1779     // clear_pending_jni_exception_check
1780     __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1781   }
1782 
1783   // reset handle block
1784   __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1785   __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1786 
1787   __ leave();
1788 
1789   #if INCLUDE_JFR
1790   // We need to do a poll test after unwind in case the sampler
1791   // managed to sample the native frame after returning to Java.
1792   Label L_return;
1793   __ ld(t0, Address(xthread, JavaThread::polling_word_offset()));
1794   address poll_test_pc = __ pc();
1795   __ relocate(relocInfo::poll_return_type);
1796   __ test_bit(t0, t0, log2i_exact(SafepointMechanism::poll_bit()));
1797   __ beqz(t0, L_return);
1798   assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
1799          "polling page return stub not created yet");
1800   address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
1801   __ la(t0, InternalAddress(poll_test_pc));
1802   __ sd(t0, Address(xthread, JavaThread::saved_exception_pc_offset()));
1803   __ far_jump(RuntimeAddress(stub));
1804   __ bind(L_return);
1805 #endif // INCLUDE_JFR
1806 
1807   // Any exception pending?
1808   Label exception_pending;
1809   __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1810   __ bnez(t0, exception_pending);
1811 
1812   // We're done
1813   __ ret();
1814 
1815   // Unexpected paths are out of line and go here
1816 
1817   // forward the exception
1818   __ bind(exception_pending);
1819 
1820   // and forward the exception
1821   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1822 
1823   // Slow path locking & unlocking
1824   if (method->is_synchronized()) {
1825 
1826     __ block_comment("Slow path lock {");
1827     __ bind(slow_path_lock);
1828 
1829     // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1830     // args are (oop obj, BasicLock* lock, JavaThread* thread)
1831 
1832     // protect the args we've loaded
1833     save_args(masm, total_c_args, c_arg, out_regs);
1834 
1835     __ mv(c_rarg0, obj_reg);
1836     __ mv(c_rarg1, lock_reg);
1837     __ mv(c_rarg2, xthread);
1838 
1839     // Not a leaf but we have last_Java_frame setup as we want.
1840     // We don't want to unmount in case of contention since that would complicate preserving
1841     // the arguments that had already been marshalled into the native convention. So we force
1842     // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1843     // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1844     __ push_cont_fastpath();
1845     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1846     __ pop_cont_fastpath();
1847     restore_args(masm, total_c_args, c_arg, out_regs);
1848 
1849 #ifdef ASSERT
1850     { Label L;
1851       __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1852       __ beqz(t0, L);
1853       __ stop("no pending exception allowed on exit from monitorenter");
1854       __ bind(L);
1855     }
1856 #endif
1857     __ j(lock_done);
1858 
1859     __ block_comment("} Slow path lock");
1860 
1861     __ block_comment("Slow path unlock {");
1862     __ bind(slow_path_unlock);
1863 
1864     if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1865       save_native_result(masm, ret_type, stack_slots);
1866     }
1867 
1868     __ mv(c_rarg2, xthread);
1869     __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1870     __ mv(c_rarg0, obj_reg);
1871 
1872     // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1873     // NOTE that obj_reg == x9 currently
1874     __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1875     __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1876 
1877     __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1878 
1879 #ifdef ASSERT
1880     {
1881       Label L;
1882       __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1883       __ beqz(t0, L);
1884       __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1885       __ bind(L);
1886     }
1887 #endif /* ASSERT */
1888 
1889     __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1890 
1891     if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1892       restore_native_result(masm, ret_type, stack_slots);
1893     }
1894     __ j(unlock_done);
1895 
1896     __ block_comment("} Slow path unlock");
1897 
1898   } // synchronized
1899 
1900   // SLOW PATH Reguard the stack if needed
1901 
1902   __ bind(reguard);
1903   save_native_result(masm, ret_type, stack_slots);
1904   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
1905   restore_native_result(masm, ret_type, stack_slots);
1906   // and continue
1907   __ j(reguard_done);
1908 
1909   // SLOW PATH safepoint
1910   {
1911     __ block_comment("safepoint {");
1912     __ bind(safepoint_in_progress);
1913 
1914     // Don't use call_VM as it will see a possible pending exception and forward it
1915     // and never return here preventing us from clearing _last_native_pc down below.
1916     //
1917     save_native_result(masm, ret_type, stack_slots);
1918     __ mv(c_rarg0, xthread);
1919 #ifndef PRODUCT
1920     assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
1921 #endif
1922     __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
1923 
1924     // Restore any method result value
1925     restore_native_result(masm, ret_type, stack_slots);
1926 
1927     __ j(safepoint_in_progress_done);
1928     __ block_comment("} safepoint");
1929   }
1930 
1931   // SLOW PATH dtrace support
1932   if (DTraceMethodProbes) {
1933     {
1934       __ block_comment("dtrace entry {");
1935       __ bind(dtrace_method_entry);
1936 
1937       // We have all of the arguments setup at this point. We must not touch any register
1938       // argument registers at this point (what if we save/restore them there are no oop?
1939 
1940       save_args(masm, total_c_args, c_arg, out_regs);
1941       __ mov_metadata(c_rarg1, method());
1942       __ call_VM_leaf(
1943         CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1944         xthread, c_rarg1);
1945       restore_args(masm, total_c_args, c_arg, out_regs);
1946       __ j(dtrace_method_entry_done);
1947       __ block_comment("} dtrace entry");
1948     }
1949 
1950     {
1951       __ block_comment("dtrace exit {");
1952       __ bind(dtrace_method_exit);
1953       save_native_result(masm, ret_type, stack_slots);
1954       __ mov_metadata(c_rarg1, method());
1955       __ call_VM_leaf(
1956            CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1957            xthread, c_rarg1);
1958       restore_native_result(masm, ret_type, stack_slots);
1959       __ j(dtrace_method_exit_done);
1960       __ block_comment("} dtrace exit");
1961     }
1962   }
1963 
1964   __ flush();
1965 
1966   nmethod *nm = nmethod::new_native_nmethod(method,
1967                                             compile_id,
1968                                             masm->code(),
1969                                             vep_offset,
1970                                             frame_complete,
1971                                             stack_slots / VMRegImpl::slots_per_word,
1972                                             (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
1973                                             in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
1974                                             oop_maps);
1975   assert(nm != nullptr, "create native nmethod fail!");
1976   return nm;
1977 }
1978 
1979 // this function returns the adjust size (in number of words) to a c2i adapter
1980 // activation for use during deoptimization
1981 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
1982   assert(callee_locals >= callee_parameters,
1983          "test and remove; got more parms than locals");
1984   if (callee_locals < callee_parameters) {
1985     return 0;                   // No adjustment for negative locals
1986   }
1987   int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
1988   // diff is counted in stack words
1989   return align_up(diff, 2);
1990 }
1991 
1992 //------------------------------generate_deopt_blob----------------------------
1993 void SharedRuntime::generate_deopt_blob() {
1994   // Allocate space for the code
1995   ResourceMark rm;
1996   // Setup code generation tools
1997   int pad = 0;
1998 #if INCLUDE_JVMCI
1999   if (EnableJVMCI) {
2000     pad += 512; // Increase the buffer size when compiling for JVMCI
2001   }
2002 #endif
2003   const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2004   CodeBuffer buffer(name, 2048 + pad, 1024);
2005   MacroAssembler* masm = new MacroAssembler(&buffer);
2006   int frame_size_in_words = -1;
2007   OopMap* map = nullptr;
2008   OopMapSet *oop_maps = new OopMapSet();
2009   assert_cond(masm != nullptr && oop_maps != nullptr);
2010   RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0);
2011 
2012   // -------------
2013   // This code enters when returning to a de-optimized nmethod.  A return
2014   // address has been pushed on the stack, and return values are in
2015   // registers.
2016   // If we are doing a normal deopt then we were called from the patched
2017   // nmethod from the point we returned to the nmethod. So the return
2018   // address on the stack is wrong by NativeCall::instruction_size
2019   // We will adjust the value so it looks like we have the original return
2020   // address on the stack (like when we eagerly deoptimized).
2021   // In the case of an exception pending when deoptimizing, we enter
2022   // with a return address on the stack that points after the call we patched
2023   // into the exception handler. We have the following register state from,
2024   // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2025   //    x10: exception oop
2026   //    x9: exception handler
2027   //    x13: throwing pc
2028   // So in this case we simply jam x13 into the useless return address and
2029   // the stack looks just like we want.
2030   //
2031   // At this point we need to de-opt.  We save the argument return
2032   // registers.  We call the first C routine, fetch_unroll_info().  This
2033   // routine captures the return values and returns a structure which
2034   // describes the current frame size and the sizes of all replacement frames.
2035   // The current frame is compiled code and may contain many inlined
2036   // functions, each with their own JVM state.  We pop the current frame, then
2037   // push all the new frames.  Then we call the C routine unpack_frames() to
2038   // populate these frames.  Finally unpack_frames() returns us the new target
2039   // address.  Notice that callee-save registers are BLOWN here; they have
2040   // already been captured in the vframeArray at the time the return PC was
2041   // patched.
2042   address start = __ pc();
2043   Label cont;
2044 
2045   // Prolog for non exception case!
2046 
2047   // Save everything in sight.
2048   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2049 
2050   // Normal deoptimization.  Save exec mode for unpack_frames.
2051   __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2052   __ j(cont);
2053 
2054   int reexecute_offset = __ pc() - start;
2055 #if INCLUDE_JVMCI && !defined(COMPILER1)
2056   if (UseJVMCICompiler) {
2057     // JVMCI does not use this kind of deoptimization
2058     __ should_not_reach_here();
2059   }
2060 #endif
2061 
2062   // Reexecute case
2063   // return address is the pc describes what bci to do re-execute at
2064 
2065   // No need to update map as each call to save_live_registers will produce identical oopmap
2066   (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2067 
2068   __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2069   __ j(cont);
2070 
2071 #if INCLUDE_JVMCI
2072   Label after_fetch_unroll_info_call;
2073   int implicit_exception_uncommon_trap_offset = 0;
2074   int uncommon_trap_offset = 0;
2075 
2076   if (EnableJVMCI) {
2077     implicit_exception_uncommon_trap_offset = __ pc() - start;
2078 
2079     __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2080     __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2081 
2082     uncommon_trap_offset = __ pc() - start;
2083 
2084     // Save everything in sight.
2085     reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2086     // fetch_unroll_info needs to call last_java_frame()
2087     Label retaddr;
2088     __ set_last_Java_frame(sp, noreg, retaddr, t0);
2089 
2090     __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2091     __ mv(t0, -1);
2092     __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2093 
2094     __ mv(xcpool, Deoptimization::Unpack_reexecute);
2095     __ mv(c_rarg0, xthread);
2096     __ orrw(c_rarg2, zr, xcpool); // exec mode
2097     __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2098     __ bind(retaddr);
2099     oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2100 
2101     __ reset_last_Java_frame(false);
2102 
2103     __ j(after_fetch_unroll_info_call);
2104   } // EnableJVMCI
2105 #endif // INCLUDE_JVMCI
2106 
2107   int exception_offset = __ pc() - start;
2108 
2109   // Prolog for exception case
2110 
2111   // all registers are dead at this entry point, except for x10, and
2112   // x13 which contain the exception oop and exception pc
2113   // respectively.  Set them in TLS and fall thru to the
2114   // unpack_with_exception_in_tls entry point.
2115 
2116   __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2117   __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2118 
2119   int exception_in_tls_offset = __ pc() - start;
2120 
2121   // new implementation because exception oop is now passed in JavaThread
2122 
2123   // Prolog for exception case
2124   // All registers must be preserved because they might be used by LinearScan
2125   // Exceptiop oop and throwing PC are passed in JavaThread
2126   // tos: stack at point of call to method that threw the exception (i.e. only
2127   // args are on the stack, no return address)
2128 
2129   // The return address pushed by save_live_registers will be patched
2130   // later with the throwing pc. The correct value is not available
2131   // now because loading it from memory would destroy registers.
2132 
2133   // NB: The SP at this point must be the SP of the method that is
2134   // being deoptimized.  Deoptimization assumes that the frame created
2135   // here by save_live_registers is immediately below the method's SP.
2136   // This is a somewhat fragile mechanism.
2137 
2138   // Save everything in sight.
2139   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2140 
2141   // Now it is safe to overwrite any register
2142 
2143   // Deopt during an exception.  Save exec mode for unpack_frames.
2144   __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2145 
2146   // load throwing pc from JavaThread and patch it as the return address
2147   // of the current frame. Then clear the field in JavaThread
2148 
2149   __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2150   __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2151   __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2152 
2153 #ifdef ASSERT
2154   // verify that there is really an exception oop in JavaThread
2155   __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2156   __ verify_oop(x10);
2157 
2158   // verify that there is no pending exception
2159   Label no_pending_exception;
2160   __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2161   __ beqz(t0, no_pending_exception);
2162   __ stop("must not have pending exception here");
2163   __ bind(no_pending_exception);
2164 #endif
2165 
2166   __ bind(cont);
2167 
2168   // Call C code.  Need thread and this frame, but NOT official VM entry
2169   // crud.  We cannot block on this call, no GC can happen.
2170   //
2171   // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2172 
2173   // fetch_unroll_info needs to call last_java_frame().
2174 
2175   Label retaddr;
2176   __ set_last_Java_frame(sp, noreg, retaddr, t0);
2177 #ifdef ASSERT
2178   {
2179     Label L;
2180     __ ld(t0, Address(xthread,
2181                               JavaThread::last_Java_fp_offset()));
2182     __ beqz(t0, L);
2183     __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2184     __ bind(L);
2185   }
2186 #endif // ASSERT
2187   __ mv(c_rarg0, xthread);
2188   __ mv(c_rarg1, xcpool);
2189   __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2190   __ bind(retaddr);
2191 
2192   // Need to have an oopmap that tells fetch_unroll_info where to
2193   // find any register it might need.
2194   oop_maps->add_gc_map(__ pc() - start, map);
2195 
2196   __ reset_last_Java_frame(false);
2197 
2198 #if INCLUDE_JVMCI
2199   if (EnableJVMCI) {
2200     __ bind(after_fetch_unroll_info_call);
2201   }
2202 #endif
2203 
2204   // Load UnrollBlock* into x15
2205   __ mv(x15, x10);
2206 
2207   __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2208   Label noException;
2209   __ mv(t0, Deoptimization::Unpack_exception);
2210   __ bne(xcpool, t0, noException); // Was exception pending?
2211   __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2212   __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2213   __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2214   __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2215 
2216   __ verify_oop(x10);
2217 
2218   // Overwrite the result registers with the exception results.
2219   __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2220 
2221   __ bind(noException);
2222 
2223   // Only register save data is on the stack.
2224   // Now restore the result registers.  Everything else is either dead
2225   // or captured in the vframeArray.
2226 
2227   // Restore fp result register
2228   __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2229   // Restore integer result register
2230   __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2231 
2232   // Pop all of the register save area off the stack
2233   __ add(sp, sp, frame_size_in_words * wordSize);
2234 
2235   // All of the register save area has been popped of the stack. Only the
2236   // return address remains.
2237 
2238   // Pop all the frames we must move/replace.
2239   //
2240   // Frame picture (youngest to oldest)
2241   // 1: self-frame (no frame link)
2242   // 2: deopting frame  (no frame link)
2243   // 3: caller of deopting frame (could be compiled/interpreted).
2244   //
2245   // Note: by leaving the return address of self-frame on the stack
2246   // and using the size of frame 2 to adjust the stack
2247   // when we are done the return to frame 3 will still be on the stack.
2248 
2249   // Pop deoptimized frame
2250   __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2251   __ subi(x12, x12, 2 * wordSize);
2252   __ add(sp, sp, x12);
2253   __ ld(fp, Address(sp, 0));
2254   __ ld(ra, Address(sp, wordSize));
2255   __ addi(sp, sp, 2 * wordSize);
2256   // RA should now be the return address to the caller (3)
2257 
2258 #ifdef ASSERT
2259   // Compilers generate code that bang the stack by as much as the
2260   // interpreter would need. So this stack banging should never
2261   // trigger a fault. Verify that it does not on non product builds.
2262   __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2263   __ bang_stack_size(x9, x12);
2264 #endif
2265   // Load address of array of frame pcs into x12
2266   __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2267 
2268   // Load address of array of frame sizes into x14
2269   __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2270 
2271   // Load counter into x13
2272   __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2273 
2274   // Now adjust the caller's stack to make up for the extra locals
2275   // but record the original sp so that we can save it in the skeletal interpreter
2276   // frame and the stack walking of interpreter_sender will get the unextended sp
2277   // value and not the "real" sp value.
2278 
2279   const Register sender_sp = x16;
2280 
2281   __ mv(sender_sp, sp);
2282   __ lwu(x9, Address(x15,
2283                      Deoptimization::UnrollBlock::
2284                      caller_adjustment_offset()));
2285   __ sub(sp, sp, x9);
2286 
2287   // Push interpreter frames in a loop
2288   __ mv(t0, 0xDEADDEAD);               // Make a recognizable pattern
2289   __ mv(t1, t0);
2290   Label loop;
2291   __ bind(loop);
2292   __ ld(x9, Address(x14, 0));          // Load frame size
2293   __ addi(x14, x14, wordSize);
2294   __ subi(x9, x9, 2 * wordSize);       // We'll push pc and fp by hand
2295   __ ld(ra, Address(x12, 0));          // Load pc
2296   __ addi(x12, x12, wordSize);
2297   __ enter();                          // Save old & set new fp
2298   __ sub(sp, sp, x9);                  // Prolog
2299   // This value is corrected by layout_activation_impl
2300   __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2301   __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2302   __ mv(sender_sp, sp);                // Pass sender_sp to next frame
2303   __ subi(x13, x13, 1);                // Decrement counter
2304   __ bnez(x13, loop);
2305 
2306     // Re-push self-frame
2307   __ ld(ra, Address(x12));
2308   __ enter();
2309 
2310   // Allocate a full sized register save area.  We subtract 2 because
2311   // enter() just pushed 2 words
2312   __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2313 
2314   // Restore frame locals after moving the frame
2315   __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2316   __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2317 
2318   // Call C code.  Need thread but NOT official VM entry
2319   // crud.  We cannot block on this call, no GC can happen.  Call should
2320   // restore return values to their stack-slots with the new SP.
2321   //
2322   // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2323 
2324   // Use fp because the frames look interpreted now
2325   // Don't need the precise return PC here, just precise enough to point into this code blob.
2326   address the_pc = __ pc();
2327   __ set_last_Java_frame(sp, fp, the_pc, t0);
2328 
2329   __ mv(c_rarg0, xthread);
2330   __ mv(c_rarg1, xcpool); // second arg: exec_mode
2331   __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2332 
2333   // Set an oopmap for the call site
2334   // Use the same PC we used for the last java frame
2335   oop_maps->add_gc_map(the_pc - start,
2336                        new OopMap(frame_size_in_words, 0));
2337 
2338   // Clear fp AND pc
2339   __ reset_last_Java_frame(true);
2340 
2341   // Collect return values
2342   __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2343   __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2344 
2345   // Pop self-frame.
2346   __ leave();                           // Epilog
2347 
2348   // Jump to interpreter
2349   __ ret();
2350 
2351   // Make sure all code is generated
2352   masm->flush();
2353 
2354   _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2355   assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2356   _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2357 #if INCLUDE_JVMCI
2358   if (EnableJVMCI) {
2359     _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2360     _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2361   }
2362 #endif
2363 }
2364 
2365 // Number of stack slots between incoming argument block and the start of
2366 // a new frame. The PROLOG must add this many slots to the stack. The
2367 // EPILOG must remove this many slots.
2368 // RISCV needs two words for RA (return address) and FP (frame pointer).
2369 uint SharedRuntime::in_preserve_stack_slots() {
2370   return 2 * VMRegImpl::slots_per_word + (VerifyStackAtCalls ? 0 : 2) ;
2371 }
2372 
2373 uint SharedRuntime::out_preserve_stack_slots() {
2374   return 0;
2375 }
2376 
2377 VMReg SharedRuntime::thread_register() {
2378   return xthread->as_VMReg();
2379 }
2380 
2381 //------------------------------generate_handler_blob------
2382 //
2383 // Generate a special Compile2Runtime blob that saves all registers,
2384 // and setup oopmap.
2385 //
2386 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2387   assert(is_polling_page_id(id), "expected a polling page stub id");
2388 
2389   ResourceMark rm;
2390   OopMapSet *oop_maps = new OopMapSet();
2391   assert_cond(oop_maps != nullptr);
2392   OopMap* map = nullptr;
2393 
2394   // Allocate space for the code.  Setup code generation tools.
2395   const char* name = SharedRuntime::stub_name(id);
2396   CodeBuffer buffer(name, 2048, 1024);
2397   MacroAssembler* masm = new MacroAssembler(&buffer);
2398   assert_cond(masm != nullptr);
2399 
2400   address start   = __ pc();
2401   address call_pc = nullptr;
2402   int frame_size_in_words = -1;
2403   bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2404   RegisterSaver reg_saver(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2405 
2406   // Save Integer and Float registers.
2407   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2408 
2409   // The following is basically a call_VM.  However, we need the precise
2410   // address of the call in order to generate an oopmap. Hence, we do all the
2411   // work ourselves.
2412 
2413   Label retaddr;
2414   __ set_last_Java_frame(sp, noreg, retaddr, t0);
2415 
2416   // The return address must always be correct so that frame constructor never
2417   // sees an invalid pc.
2418 
2419   if (!cause_return) {
2420     // overwrite the return address pushed by save_live_registers
2421     // Additionally, x18 is a callee-saved register so we can look at
2422     // it later to determine if someone changed the return address for
2423     // us!
2424     __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2425     __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2426   }
2427 
2428   // Do the call
2429   __ mv(c_rarg0, xthread);
2430   __ rt_call(call_ptr);
2431   __ bind(retaddr);
2432 
2433   // Set an oopmap for the call site.  This oopmap will map all
2434   // oop-registers and debug-info registers as callee-saved.  This
2435   // will allow deoptimization at this safepoint to find all possible
2436   // debug-info recordings, as well as let GC find all oops.
2437 
2438   oop_maps->add_gc_map( __ pc() - start, map);
2439 
2440   Label noException;
2441 
2442   __ reset_last_Java_frame(false);
2443 
2444   __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2445 
2446   __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2447   __ beqz(t0, noException);
2448 
2449   // Exception pending
2450 
2451   reg_saver.restore_live_registers(masm);
2452 
2453   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2454 
2455   // No exception case
2456   __ bind(noException);
2457 
2458   Label no_adjust, bail;
2459   if (!cause_return) {
2460     // If our stashed return pc was modified by the runtime we avoid touching it
2461     __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2462     __ bne(x18, t0, no_adjust);
2463 
2464 #ifdef ASSERT
2465     // Verify the correct encoding of the poll we're about to skip.
2466     // See NativeInstruction::is_lwu_to_zr()
2467     __ lwu(t0, Address(x18));
2468     __ andi(t1, t0, 0b1111111);
2469     __ mv(t2, 0b0000011);
2470     __ bne(t1, t2, bail); // 0-6:0b0000011
2471     __ srli(t1, t0, 7);
2472     __ andi(t1, t1, 0b11111);
2473     __ bnez(t1, bail);    // 7-11:0b00000
2474     __ srli(t1, t0, 12);
2475     __ andi(t1, t1, 0b111);
2476     __ mv(t2, 0b110);
2477     __ bne(t1, t2, bail); // 12-14:0b110
2478 #endif
2479 
2480     // Adjust return pc forward to step over the safepoint poll instruction
2481     __ addi(x18, x18, NativeInstruction::instruction_size);
2482     __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2483   }
2484 
2485   __ bind(no_adjust);
2486   // Normal exit, restore registers and exit.
2487 
2488   reg_saver.restore_live_registers(masm);
2489   __ ret();
2490 
2491 #ifdef ASSERT
2492   __ bind(bail);
2493   __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2494 #endif
2495 
2496   // Make sure all code is generated
2497   masm->flush();
2498 
2499   // Fill-out other meta info
2500   return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2501 }
2502 
2503 //
2504 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2505 //
2506 // Generate a stub that calls into vm to find out the proper destination
2507 // of a java call. All the argument registers are live at this point
2508 // but since this is generic code we don't know what they are and the caller
2509 // must do any gc of the args.
2510 //
2511 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2512   assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2513   assert(is_resolve_id(id), "expected a resolve stub id");
2514 
2515   // allocate space for the code
2516   ResourceMark rm;
2517 
2518   const char* name = SharedRuntime::stub_name(id);
2519   CodeBuffer buffer(name, 1000, 512);
2520   MacroAssembler* masm = new MacroAssembler(&buffer);
2521   assert_cond(masm != nullptr);
2522 
2523   int frame_size_in_words = -1;
2524   RegisterSaver reg_saver(false /* save_vectors */);
2525 
2526   OopMapSet *oop_maps = new OopMapSet();
2527   assert_cond(oop_maps != nullptr);
2528   OopMap* map = nullptr;
2529 
2530   int start = __ offset();
2531 
2532   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2533 
2534   int frame_complete = __ offset();
2535 
2536   {
2537     Label retaddr;
2538     __ set_last_Java_frame(sp, noreg, retaddr, t0);
2539 
2540     __ mv(c_rarg0, xthread);
2541     __ rt_call(destination);
2542     __ bind(retaddr);
2543   }
2544 
2545   // Set an oopmap for the call site.
2546   // We need this not only for callee-saved registers, but also for volatile
2547   // registers that the compiler might be keeping live across a safepoint.
2548 
2549   oop_maps->add_gc_map( __ offset() - start, map);
2550 
2551   // x10 contains the address we are going to jump to assuming no exception got installed
2552 
2553   // clear last_Java_sp
2554   __ reset_last_Java_frame(false);
2555   // check for pending exceptions
2556   Label pending;
2557   __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2558   __ bnez(t1, pending);
2559 
2560   // get the returned Method*
2561   __ get_vm_result_metadata(xmethod, xthread);
2562   __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2563 
2564   // x10 is where we want to jump, overwrite t1 which is saved and temporary
2565   __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2566   reg_saver.restore_live_registers(masm);
2567 
2568   // We are back to the original state on entry and ready to go.
2569   __ jr(t1);
2570 
2571   // Pending exception after the safepoint
2572 
2573   __ bind(pending);
2574 
2575   reg_saver.restore_live_registers(masm);
2576 
2577   // exception pending => remove activation and forward to exception handler
2578 
2579   __ sd(zr, Address(xthread, JavaThread::vm_result_oop_offset()));
2580 
2581   __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2582   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2583 
2584   // -------------
2585   // make sure all code is generated
2586   masm->flush();
2587 
2588   // return the  blob
2589   return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2590 }
2591 
2592 // Continuation point for throwing of implicit exceptions that are
2593 // not handled in the current activation. Fabricates an exception
2594 // oop and initiates normal exception dispatching in this
2595 // frame. Since we need to preserve callee-saved values (currently
2596 // only for C2, but done for C1 as well) we need a callee-saved oop
2597 // map and therefore have to make these stubs into RuntimeStubs
2598 // rather than BufferBlobs.  If the compiler needs all registers to
2599 // be preserved between the fault point and the exception handler
2600 // then it must assume responsibility for that in
2601 // AbstractCompiler::continuation_for_implicit_null_exception or
2602 // continuation_for_implicit_division_by_zero_exception. All other
2603 // implicit exceptions (e.g., NullPointerException or
2604 // AbstractMethodError on entry) are either at call sites or
2605 // otherwise assume that stack unwinding will be initiated, so
2606 // caller saved registers were assumed volatile in the compiler.
2607 
2608 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2609   assert(is_throw_id(id), "expected a throw stub id");
2610 
2611   const char* name = SharedRuntime::stub_name(id);
2612 
2613   // Information about frame layout at time of blocking runtime call.
2614   // Note that we only have to preserve callee-saved registers since
2615   // the compilers are responsible for supplying a continuation point
2616   // if they expect all registers to be preserved.
2617   // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2618   assert_cond(runtime_entry != nullptr);
2619   enum layout {
2620     fp_off = 0,
2621     fp_off2,
2622     return_off,
2623     return_off2,
2624     framesize // inclusive of return address
2625   };
2626 
2627   const int insts_size = 1024;
2628   const int locs_size  = 64;
2629 
2630   ResourceMark rm;
2631   const char* timer_msg = "SharedRuntime generate_throw_exception";
2632   TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2633 
2634   CodeBuffer code(name, insts_size, locs_size);
2635   OopMapSet* oop_maps  = new OopMapSet();
2636   MacroAssembler* masm = new MacroAssembler(&code);
2637   assert_cond(oop_maps != nullptr && masm != nullptr);
2638 
2639   address start = __ pc();
2640 
2641   // This is an inlined and slightly modified version of call_VM
2642   // which has the ability to fetch the return PC out of
2643   // thread-local storage and also sets up last_Java_sp slightly
2644   // differently than the real call_VM
2645 
2646   __ enter(); // Save FP and RA before call
2647 
2648   assert(is_even(framesize / 2), "sp not 16-byte aligned");
2649 
2650   // ra and fp are already in place
2651   __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2652 
2653   int frame_complete = __ pc() - start;
2654 
2655   // Set up last_Java_sp and last_Java_fp
2656   address the_pc = __ pc();
2657   __ set_last_Java_frame(sp, fp, the_pc, t0);
2658 
2659   // Call runtime
2660   __ mv(c_rarg0, xthread);
2661   BLOCK_COMMENT("call runtime_entry");
2662   __ rt_call(runtime_entry);
2663 
2664   // Generate oop map
2665   OopMap* map = new OopMap(framesize, 0);
2666   assert_cond(map != nullptr);
2667 
2668   oop_maps->add_gc_map(the_pc - start, map);
2669 
2670   __ reset_last_Java_frame(true);
2671 
2672   __ leave();
2673 
2674   // check for pending exceptions
2675 #ifdef ASSERT
2676   Label L;
2677   __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2678   __ bnez(t0, L);
2679   __ should_not_reach_here();
2680   __ bind(L);
2681 #endif // ASSERT
2682   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2683 
2684   // codeBlob framesize is in words (not VMRegImpl::slot_size)
2685   RuntimeStub* stub =
2686     RuntimeStub::new_runtime_stub(name,
2687                                   &code,
2688                                   frame_complete,
2689                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2690                                   oop_maps, false);
2691   assert(stub != nullptr, "create runtime stub fail!");
2692   return stub;
2693 }
2694 
2695 #if INCLUDE_JFR
2696 
2697 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2698   __ set_last_Java_frame(sp, fp, the_pc, t0);
2699   __ mv(c_rarg0, thread);
2700 }
2701 
2702 static void jfr_epilogue(MacroAssembler* masm) {
2703   __ reset_last_Java_frame(true);
2704 }
2705 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2706 // It returns a jobject handle to the event writer.
2707 // The handle is dereferenced and the return value is the event writer oop.
2708 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2709   enum layout {
2710     fp_off,
2711     fp_off2,
2712     return_off,
2713     return_off2,
2714     framesize // inclusive of return address
2715   };
2716 
2717   int insts_size = 1024;
2718   int locs_size = 64;
2719   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
2720   CodeBuffer code(name, insts_size, locs_size);
2721   OopMapSet* oop_maps = new OopMapSet();
2722   MacroAssembler* masm = new MacroAssembler(&code);
2723 
2724   address start = __ pc();
2725   __ enter();
2726   int frame_complete = __ pc() - start;
2727   address the_pc = __ pc();
2728   jfr_prologue(the_pc, masm, xthread);
2729   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2730 
2731   jfr_epilogue(masm);
2732   __ resolve_global_jobject(x10, t0, t1);
2733   __ leave();
2734   __ ret();
2735 
2736   OopMap* map = new OopMap(framesize, 1);
2737   oop_maps->add_gc_map(the_pc - start, map);
2738 
2739   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2740     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2741                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2742                                   oop_maps, false);
2743   return stub;
2744 }
2745 
2746 // For c2: call to return a leased buffer.
2747 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2748   enum layout {
2749     fp_off,
2750     fp_off2,
2751     return_off,
2752     return_off2,
2753     framesize // inclusive of return address
2754   };
2755 
2756   int insts_size = 1024;
2757   int locs_size = 64;
2758   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
2759   CodeBuffer code(name, insts_size, locs_size);
2760   OopMapSet* oop_maps = new OopMapSet();
2761   MacroAssembler* masm = new MacroAssembler(&code);
2762 
2763   address start = __ pc();
2764   __ enter();
2765   int frame_complete = __ pc() - start;
2766   address the_pc = __ pc();
2767   jfr_prologue(the_pc, masm, xthread);
2768   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2769 
2770   jfr_epilogue(masm);
2771   __ leave();
2772   __ ret();
2773 
2774   OopMap* map = new OopMap(framesize, 1);
2775   oop_maps->add_gc_map(the_pc - start, map);
2776 
2777   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2778     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2779                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2780                                   oop_maps, false);
2781   return stub;
2782 }
2783 
2784 #endif // INCLUDE_JFR