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 
  63 #define __ masm->
  64 
  65 #ifdef PRODUCT
  66 #define BLOCK_COMMENT(str) /* nothing */
  67 #else
  68 #define BLOCK_COMMENT(str) __ block_comment(str)
  69 #endif
  70 
  71 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
  72 
  73 class RegisterSaver {
  74   const bool _save_vectors;
  75  public:
  76   RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
  77   ~RegisterSaver() {}
  78   OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
  79   void restore_live_registers(MacroAssembler* masm);
  80 
  81   // Offsets into the register save area
  82   // Used by deoptimization when it is managing result register
  83   // values on its own
  84   // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
  85   // |---v0---|<---SP
  86   // |---v1---|save vectors only in generate_handler_blob
  87   // |-- .. --|
  88   // |---v31--|-----
  89   // |---f0---|
  90   // |---f1---|
  91   // |   ..   |
  92   // |---f31--|
  93   // |---reserved slot for stack alignment---|
  94   // |---x5---|
  95   // |   x6   |
  96   // |---.. --|
  97   // |---x31--|
  98   // |---fp---|
  99   // |---ra---|
 100   int v0_offset_in_bytes(void) { return 0; }
 101   int f0_offset_in_bytes(void) {
 102     int f0_offset = 0;
 103 #ifdef COMPILER2
 104     if (_save_vectors) {
 105       f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
 106                    BytesPerInt;
 107     }
 108 #endif
 109     return f0_offset;
 110   }
 111   int reserved_slot_offset_in_bytes(void) {
 112     return f0_offset_in_bytes() +
 113            FloatRegister::max_slots_per_register *
 114            FloatRegister::number_of_registers *
 115            BytesPerInt;
 116   }
 117 
 118   int reg_offset_in_bytes(Register r) {
 119     assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
 120     return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
 121   }
 122 
 123   int freg_offset_in_bytes(FloatRegister f) {
 124     return f0_offset_in_bytes() + f->encoding() * wordSize;
 125   }
 126 
 127   int ra_offset_in_bytes(void) {
 128     return reserved_slot_offset_in_bytes() +
 129            (Register::number_of_registers - 3) *
 130            Register::max_slots_per_register *
 131            BytesPerInt;
 132   }
 133 };
 134 
 135 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
 136   int vector_size_in_bytes = 0;
 137   int vector_size_in_slots = 0;
 138 #ifdef COMPILER2
 139   if (_save_vectors) {
 140     vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
 141     vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
 142   }
 143 #endif
 144 
 145   int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
 146   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 147   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 148   // The caller will allocate additional_frame_words
 149   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 150   // CodeBlob frame size is in words.
 151   int frame_size_in_words = frame_size_in_bytes / wordSize;
 152   *total_frame_words = frame_size_in_words;
 153 
 154   // Save Integer, Float and Vector registers.
 155   __ enter();
 156   __ push_CPU_state(_save_vectors, vector_size_in_bytes);
 157 
 158   // Set an oopmap for the call site.  This oopmap will map all
 159   // oop-registers and debug-info registers as callee-saved.  This
 160   // will allow deoptimization at this safepoint to find all possible
 161   // debug-info recordings, as well as let GC find all oops.
 162 
 163   OopMapSet *oop_maps = new OopMapSet();
 164   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 165   assert_cond(oop_maps != nullptr && oop_map != nullptr);
 166 
 167   int sp_offset_in_slots = 0;
 168   int step_in_slots = 0;
 169   if (_save_vectors) {
 170     step_in_slots = vector_size_in_slots;
 171     for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
 172       VectorRegister r = as_VectorRegister(i);
 173       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
 174     }
 175   }
 176 
 177   step_in_slots = FloatRegister::max_slots_per_register;
 178   for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
 179     FloatRegister r = as_FloatRegister(i);
 180     oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
 181   }
 182 
 183   step_in_slots = Register::max_slots_per_register;
 184   // skip the slot reserved for alignment, see MacroAssembler::push_reg;
 185   // also skip x5 ~ x6 on the stack because they are caller-saved registers.
 186   sp_offset_in_slots += Register::max_slots_per_register * 3;
 187   // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
 188   for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
 189     Register r = as_Register(i);
 190     if (r != xthread) {
 191       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
 192     }
 193   }
 194 
 195   return oop_map;
 196 }
 197 
 198 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
 199 #ifdef COMPILER2
 200   __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
 201 #else
 202   assert(!_save_vectors, "vectors are generated only by C2");
 203   __ pop_CPU_state(_save_vectors);
 204 #endif // COMPILER2
 205   __ leave();
 206 }
 207 
 208 // Is vector's size (in bytes) bigger than a size saved by default?
 209 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
 210 bool SharedRuntime::is_wide_vector(int size) {
 211   return UseRVV && size > 0;
 212 }
 213 
 214 // ---------------------------------------------------------------------------
 215 // Read the array of BasicTypes from a signature, and compute where the
 216 // arguments should go.  Values in the VMRegPair regs array refer to 4-byte
 217 // quantities.  Values less than VMRegImpl::stack0 are registers, those above
 218 // refer to 4-byte stack slots.  All stack slots are based off of the stack pointer
 219 // as framesizes are fixed.
 220 // VMRegImpl::stack0 refers to the first slot 0(sp).
 221 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
 222 // Register up to Register::number_of_registers) are the 64-bit
 223 // integer registers.
 224 
 225 // Note: the INPUTS in sig_bt are in units of Java argument words,
 226 // which are 64-bit.  The OUTPUTS are in 32-bit units.
 227 
 228 // The Java calling convention is a "shifted" version of the C ABI.
 229 // By skipping the first C ABI register we can call non-static jni
 230 // methods with small numbers of arguments without having to shuffle
 231 // the arguments at all. Since we control the java ABI we ought to at
 232 // least get some advantage out of it.
 233 
 234 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
 235                                            VMRegPair *regs,
 236                                            int total_args_passed) {
 237   // Create the mapping between argument positions and
 238   // registers.
 239   static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
 240     j_rarg0, j_rarg1, j_rarg2, j_rarg3,
 241     j_rarg4, j_rarg5, j_rarg6, j_rarg7
 242   };
 243   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
 244     j_farg0, j_farg1, j_farg2, j_farg3,
 245     j_farg4, j_farg5, j_farg6, j_farg7
 246   };
 247 
 248   uint int_args = 0;
 249   uint fp_args = 0;
 250   uint stk_args = 0;
 251 
 252   for (int i = 0; i < total_args_passed; i++) {
 253     switch (sig_bt[i]) {
 254       case T_BOOLEAN: // fall through
 255       case T_CHAR:    // fall through
 256       case T_BYTE:    // fall through
 257       case T_SHORT:   // fall through
 258       case T_INT:
 259         if (int_args < Argument::n_int_register_parameters_j) {
 260           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 261         } else {
 262           stk_args = align_up(stk_args, 2);
 263           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 264           stk_args += 1;
 265         }
 266         break;
 267       case T_VOID:
 268         // halves of T_LONG or T_DOUBLE
 269         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 270         regs[i].set_bad();
 271         break;
 272       case T_LONG:      // fall through
 273         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 274       case T_OBJECT:    // fall through
 275       case T_ARRAY:     // fall through
 276       case T_ADDRESS:
 277         if (int_args < Argument::n_int_register_parameters_j) {
 278           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 279         } else {
 280           stk_args = align_up(stk_args, 2);
 281           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 282           stk_args += 2;
 283         }
 284         break;
 285       case T_FLOAT:
 286         if (fp_args < Argument::n_float_register_parameters_j) {
 287           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 288         } else {
 289           stk_args = align_up(stk_args, 2);
 290           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 291           stk_args += 1;
 292         }
 293         break;
 294       case T_DOUBLE:
 295         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 296         if (fp_args < Argument::n_float_register_parameters_j) {
 297           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 298         } else {
 299           stk_args = align_up(stk_args, 2);
 300           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 301           stk_args += 2;
 302         }
 303         break;
 304       default:
 305         ShouldNotReachHere();
 306     }
 307   }
 308 
 309   return stk_args;
 310 }
 311 
 312 // Patch the callers callsite with entry to compiled code if it exists.
 313 static void patch_callers_callsite(MacroAssembler *masm) {
 314   Label L;
 315   __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
 316   __ beqz(t0, L);
 317 
 318   __ enter();
 319   __ push_CPU_state();
 320 
 321   // VM needs caller's callsite
 322   // VM needs target method
 323   // This needs to be a long call since we will relocate this adapter to
 324   // the codeBuffer and it may not reach
 325 
 326 #ifndef PRODUCT
 327   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 328 #endif
 329 
 330   __ mv(c_rarg0, xmethod);
 331   __ mv(c_rarg1, ra);
 332   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
 333 
 334   __ pop_CPU_state();
 335   // restore sp
 336   __ leave();
 337   __ bind(L);
 338 }
 339 
 340 static void gen_c2i_adapter(MacroAssembler *masm,
 341                             int total_args_passed,
 342                             int comp_args_on_stack,
 343                             const BasicType *sig_bt,
 344                             const VMRegPair *regs,
 345                             Label& skip_fixup) {
 346   // Before we get into the guts of the C2I adapter, see if we should be here
 347   // at all.  We've come from compiled code and are attempting to jump to the
 348   // interpreter, which means the caller made a static call to get here
 349   // (vcalls always get a compiled target if there is one).  Check for a
 350   // compiled target.  If there is one, we need to patch the caller's call.
 351   patch_callers_callsite(masm);
 352 
 353   __ bind(skip_fixup);
 354 
 355   int words_pushed = 0;
 356 
 357   // Since all args are passed on the stack, total_args_passed *
 358   // Interpreter::stackElementSize is the space we need.
 359 
 360   int extraspace = total_args_passed * Interpreter::stackElementSize;
 361 
 362   __ mv(x19_sender_sp, sp);
 363 
 364   // stack is aligned, keep it that way
 365   extraspace = align_up(extraspace, 2 * wordSize);
 366 
 367   if (extraspace) {
 368     __ sub(sp, sp, extraspace);
 369   }
 370 
 371   // Now write the args into the outgoing interpreter space
 372   for (int i = 0; i < total_args_passed; i++) {
 373     if (sig_bt[i] == T_VOID) {
 374       assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
 375       continue;
 376     }
 377 
 378     // offset to start parameters
 379     int st_off   = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 380     int next_off = st_off - Interpreter::stackElementSize;
 381 
 382     // Say 4 args:
 383     // i   st_off
 384     // 0   32 T_LONG
 385     // 1   24 T_VOID
 386     // 2   16 T_OBJECT
 387     // 3    8 T_BOOL
 388     // -    0 return address
 389     //
 390     // However to make thing extra confusing. Because we can fit a Java long/double in
 391     // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 392     // leaves one slot empty and only stores to a single slot. In this case the
 393     // slot that is occupied is the T_VOID slot. See I said it was confusing.
 394 
 395     VMReg r_1 = regs[i].first();
 396     VMReg r_2 = regs[i].second();
 397     if (!r_1->is_valid()) {
 398       assert(!r_2->is_valid(), "");
 399       continue;
 400     }
 401     if (r_1->is_stack()) {
 402       // memory to memory use t0
 403       int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
 404                     + extraspace
 405                     + words_pushed * wordSize);
 406       if (!r_2->is_valid()) {
 407         __ lwu(t0, Address(sp, ld_off));
 408         __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 409       } else {
 410         __ ld(t0, Address(sp, ld_off), /*temp register*/esp);
 411 
 412         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 413         // T_DOUBLE and T_LONG use two slots in the interpreter
 414         if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 415           // ld_off == LSW, ld_off+wordSize == MSW
 416           // st_off == MSW, next_off == LSW
 417           __ sd(t0, Address(sp, next_off), /*temp register*/esp);
 418 #ifdef ASSERT
 419           // Overwrite the unused slot with known junk
 420           __ mv(t0, 0xdeadffffdeadaaaaul);
 421           __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 422 #endif /* ASSERT */
 423         } else {
 424           __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 425         }
 426       }
 427     } else if (r_1->is_Register()) {
 428       Register r = r_1->as_Register();
 429       if (!r_2->is_valid()) {
 430         // must be only an int (or less ) so move only 32bits to slot
 431         __ sd(r, Address(sp, st_off));
 432       } else {
 433         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 434         // T_DOUBLE and T_LONG use two slots in the interpreter
 435         if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 436           // long/double in gpr
 437 #ifdef ASSERT
 438           // Overwrite the unused slot with known junk
 439           __ mv(t0, 0xdeadffffdeadaaabul);
 440           __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 441 #endif /* ASSERT */
 442           __ sd(r, Address(sp, next_off));
 443         } else {
 444           __ sd(r, Address(sp, st_off));
 445         }
 446       }
 447     } else {
 448       assert(r_1->is_FloatRegister(), "");
 449       if (!r_2->is_valid()) {
 450         // only a float use just part of the slot
 451         __ fsw(r_1->as_FloatRegister(), Address(sp, st_off));
 452       } else {
 453 #ifdef ASSERT
 454         // Overwrite the unused slot with known junk
 455         __ mv(t0, 0xdeadffffdeadaaacul);
 456         __ sd(t0, Address(sp, st_off), /*temp register*/esp);
 457 #endif /* ASSERT */
 458         __ fsd(r_1->as_FloatRegister(), Address(sp, next_off));
 459       }
 460     }
 461   }
 462 
 463   __ mv(esp, sp); // Interp expects args on caller's expression stack
 464 
 465   __ ld(t1, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
 466   __ jr(t1);
 467 }
 468 
 469 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
 470                                     int total_args_passed,
 471                                     int comp_args_on_stack,
 472                                     const BasicType *sig_bt,
 473                                     const VMRegPair *regs) {
 474   // Note: x19_sender_sp contains the senderSP on entry. We must
 475   // preserve it since we may do a i2c -> c2i transition if we lose a
 476   // race where compiled code goes non-entrant while we get args
 477   // ready.
 478 
 479   // Cut-out for having no stack args.
 480   int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
 481   if (comp_args_on_stack != 0) {
 482     __ sub(t0, sp, comp_words_on_stack * wordSize);
 483     __ andi(sp, t0, -16);
 484   }
 485 
 486   // Will jump to the compiled code just as if compiled code was doing it.
 487   // Pre-load the register-jump target early, to schedule it better.
 488   __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset())));
 489 
 490   // Now generate the shuffle code.
 491   for (int i = 0; i < total_args_passed; i++) {
 492     if (sig_bt[i] == T_VOID) {
 493       assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
 494       continue;
 495     }
 496 
 497     // Pick up 0, 1 or 2 words from SP+offset.
 498 
 499     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
 500            "scrambled load targets?");
 501     // Load in argument order going down.
 502     int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 503     // Point to interpreter value (vs. tag)
 504     int next_off = ld_off - Interpreter::stackElementSize;
 505 
 506     VMReg r_1 = regs[i].first();
 507     VMReg r_2 = regs[i].second();
 508     if (!r_1->is_valid()) {
 509       assert(!r_2->is_valid(), "");
 510       continue;
 511     }
 512     if (r_1->is_stack()) {
 513       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 514       int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
 515       if (!r_2->is_valid()) {
 516         __ lw(t0, Address(esp, ld_off));
 517         __ sd(t0, Address(sp, st_off), /*temp register*/t2);
 518       } else {
 519         //
 520         // We are using two optoregs. This can be either T_OBJECT,
 521         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 522         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 523         // So we must adjust where to pick up the data to match the
 524         // interpreter.
 525         //
 526         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 527         // are accessed as negative so LSW is at LOW address
 528 
 529         // ld_off is MSW so get LSW
 530         const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
 531                            next_off : ld_off;
 532         __ ld(t0, Address(esp, offset));
 533         // st_off is LSW (i.e. reg.first())
 534         __ sd(t0, Address(sp, st_off), /*temp register*/t2);
 535       }
 536     } else if (r_1->is_Register()) {  // Register argument
 537       Register r = r_1->as_Register();
 538       if (r_2->is_valid()) {
 539         //
 540         // We are using two VMRegs. This can be either T_OBJECT,
 541         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 542         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 543         // So we must adjust where to pick up the data to match the
 544         // interpreter.
 545 
 546         const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
 547                            next_off : ld_off;
 548 
 549         // this can be a misaligned move
 550         __ ld(r, Address(esp, offset));
 551       } else {
 552         // sign extend and use a full word?
 553         __ lw(r, Address(esp, ld_off));
 554       }
 555     } else {
 556       if (!r_2->is_valid()) {
 557         __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
 558       } else {
 559         __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
 560       }
 561     }
 562   }
 563 
 564   __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
 565 
 566   // 6243940 We might end up in handle_wrong_method if
 567   // the callee is deoptimized as we race thru here. If that
 568   // happens we don't want to take a safepoint because the
 569   // caller frame will look interpreted and arguments are now
 570   // "compiled" so it is much better to make this transition
 571   // invisible to the stack walking code. Unfortunately if
 572   // we try and find the callee by normal means a safepoint
 573   // is possible. So we stash the desired callee in the thread
 574   // and the vm will find there should this case occur.
 575 
 576   __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
 577 
 578   __ jr(t1);
 579 }
 580 
 581 // ---------------------------------------------------------------
 582 
 583 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
 584                                             int total_args_passed,
 585                                             int comp_args_on_stack,
 586                                             const BasicType *sig_bt,
 587                                             const VMRegPair *regs,
 588                                             address entry_address[AdapterBlob::ENTRY_COUNT]) {
 589   entry_address[AdapterBlob::I2C] = __ pc();
 590   gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
 591 
 592   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
 593   Label skip_fixup;
 594 
 595   const Register receiver = j_rarg0;
 596   const Register data = t0;
 597 
 598   // -------------------------------------------------------------------------
 599   // Generate a C2I adapter.  On entry we know xmethod holds the Method* during calls
 600   // to the interpreter.  The args start out packed in the compiled layout.  They
 601   // need to be unpacked into the interpreter layout.  This will almost always
 602   // require some stack space.  We grow the current (compiled) stack, then repack
 603   // the args.  We  finally end in a jump to the generic interpreter entry point.
 604   // On exit from the interpreter, the interpreter will restore our SP (lest the
 605   // compiled code, which relies solely on SP and not FP, get sick).
 606 
 607   {
 608     __ block_comment("c2i_unverified_entry {");
 609 
 610     __ ic_check();
 611     __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
 612 
 613     __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
 614     __ beqz(t0, skip_fixup);
 615     __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
 616     __ block_comment("} c2i_unverified_entry");
 617   }
 618 
 619   entry_address[AdapterBlob::C2I] = __ pc();
 620 
 621   // Class initialization barrier for static methods
 622   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
 623   assert(VM_Version::supports_fast_class_init_checks(), "sanity");
 624   Label L_skip_barrier;
 625 
 626   // Bypass the barrier for non-static methods
 627   __ load_unsigned_short(t0, Address(xmethod, Method::access_flags_offset()));
 628   __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
 629   __ beqz(t1, L_skip_barrier); // non-static
 630 
 631   __ load_method_holder(t1, xmethod);
 632   __ clinit_barrier(t1, t0, &L_skip_barrier);
 633   __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 634 
 635   __ bind(L_skip_barrier);
 636   entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();
 637 
 638   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 639   bs->c2i_entry_barrier(masm);
 640 
 641   gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
 642   return;
 643 }
 644 
 645 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
 646                                              uint num_bits,
 647                                              uint total_args_passed) {
 648   assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
 649   assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
 650 
 651   // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
 652   static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
 653     v8, v9, v10, v11, v12, v13, v14, v15,
 654     v16, v17, v18, v19, v20, v21, v22, v23
 655   };
 656 
 657   const int next_reg_val = 3;
 658   for (uint i = 0; i < total_args_passed; i++) {
 659     VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
 660     regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
 661   }
 662   return 0;
 663 }
 664 
 665 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
 666                                          VMRegPair *regs,
 667                                          int total_args_passed) {
 668 
 669   // We return the amount of VMRegImpl stack slots we need to reserve for all
 670   // the arguments NOT counting out_preserve_stack_slots.
 671 
 672   static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
 673     c_rarg0, c_rarg1, c_rarg2, c_rarg3,
 674     c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
 675   };
 676   static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
 677     c_farg0, c_farg1, c_farg2, c_farg3,
 678     c_farg4, c_farg5, c_farg6, c_farg7
 679   };
 680 
 681   uint int_args = 0;
 682   uint fp_args = 0;
 683   uint stk_args = 0; // inc by 2 each time
 684 
 685   for (int i = 0; i < total_args_passed; i++) {
 686     switch (sig_bt[i]) {
 687       case T_BOOLEAN:  // fall through
 688       case T_CHAR:     // fall through
 689       case T_BYTE:     // fall through
 690       case T_SHORT:    // fall through
 691       case T_INT:
 692         if (int_args < Argument::n_int_register_parameters_c) {
 693           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 694         } else {
 695           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 696           stk_args += 2;
 697         }
 698         break;
 699       case T_LONG:      // fall through
 700         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 701       case T_OBJECT:    // fall through
 702       case T_ARRAY:     // fall through
 703       case T_ADDRESS:   // fall through
 704       case T_METADATA:
 705         if (int_args < Argument::n_int_register_parameters_c) {
 706           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 707         } else {
 708           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 709           stk_args += 2;
 710         }
 711         break;
 712       case T_FLOAT:
 713         if (fp_args < Argument::n_float_register_parameters_c) {
 714           regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
 715         } else if (int_args < Argument::n_int_register_parameters_c) {
 716           regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
 717         } else {
 718           regs[i].set1(VMRegImpl::stack2reg(stk_args));
 719           stk_args += 2;
 720         }
 721         break;
 722       case T_DOUBLE:
 723         assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 724         if (fp_args < Argument::n_float_register_parameters_c) {
 725           regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 726         } else if (int_args < Argument::n_int_register_parameters_c) {
 727           regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
 728         } else {
 729           regs[i].set2(VMRegImpl::stack2reg(stk_args));
 730           stk_args += 2;
 731         }
 732         break;
 733       case T_VOID: // Halves of longs and doubles
 734         assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 735         regs[i].set_bad();
 736         break;
 737       default:
 738         ShouldNotReachHere();
 739     }
 740   }
 741 
 742   return stk_args;
 743 }
 744 
 745 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
 746   // We always ignore the frame_slots arg and just use the space just below frame pointer
 747   // which by this time is free to use
 748   switch (ret_type) {
 749     case T_FLOAT:
 750       __ fsw(f10, Address(fp, -3 * wordSize));
 751       break;
 752     case T_DOUBLE:
 753       __ fsd(f10, Address(fp, -3 * wordSize));
 754       break;
 755     case T_VOID:  break;
 756     default: {
 757       __ sd(x10, Address(fp, -3 * wordSize));
 758     }
 759   }
 760 }
 761 
 762 void SharedRuntime::restore_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       __ flw(f10, Address(fp, -3 * wordSize));
 768       break;
 769     case T_DOUBLE:
 770       __ fld(f10, Address(fp, -3 * wordSize));
 771       break;
 772     case T_VOID:  break;
 773     default: {
 774       __ ld(x10, Address(fp, -3 * wordSize));
 775     }
 776   }
 777 }
 778 
 779 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
 780   RegSet x;
 781   for ( int i = first_arg ; i < arg_count ; i++ ) {
 782     if (args[i].first()->is_Register()) {
 783       x = x + args[i].first()->as_Register();
 784     } else if (args[i].first()->is_FloatRegister()) {
 785       __ subi(sp, sp, 2 * wordSize);
 786       __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
 787     }
 788   }
 789   __ push_reg(x, sp);
 790 }
 791 
 792 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
 793   RegSet x;
 794   for ( int i = first_arg ; i < arg_count ; i++ ) {
 795     if (args[i].first()->is_Register()) {
 796       x = x + args[i].first()->as_Register();
 797     } else {
 798       ;
 799     }
 800   }
 801   __ pop_reg(x, sp);
 802   for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
 803     if (args[i].first()->is_Register()) {
 804       ;
 805     } else if (args[i].first()->is_FloatRegister()) {
 806       __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
 807       __ addi(sp, sp, 2 * wordSize);
 808     }
 809   }
 810 }
 811 
 812 static void verify_oop_args(MacroAssembler* masm,
 813                             const methodHandle& method,
 814                             const BasicType* sig_bt,
 815                             const VMRegPair* regs) {
 816   const Register temp_reg = x9;  // not part of any compiled calling seq
 817   if (VerifyOops) {
 818     for (int i = 0; i < method->size_of_parameters(); i++) {
 819       if (sig_bt[i] == T_OBJECT ||
 820           sig_bt[i] == T_ARRAY) {
 821         VMReg r = regs[i].first();
 822         assert(r->is_valid(), "bad oop arg");
 823         if (r->is_stack()) {
 824           __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
 825           __ verify_oop(temp_reg);
 826         } else {
 827           __ verify_oop(r->as_Register());
 828         }
 829       }
 830     }
 831   }
 832 }
 833 
 834 // on exit, sp points to the ContinuationEntry
 835 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
 836   assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
 837   assert(in_bytes(ContinuationEntry::cont_offset())  % VMRegImpl::stack_slot_size == 0, "");
 838   assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
 839 
 840   stack_slots += (int)ContinuationEntry::size() / wordSize;
 841   __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
 842 
 843   OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
 844 
 845   __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
 846   __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
 847   __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
 848 
 849   return map;
 850 }
 851 
 852 // on entry c_rarg1 points to the continuation
 853 //          sp points to ContinuationEntry
 854 //          c_rarg3 -- isVirtualThread
 855 static void fill_continuation_entry(MacroAssembler* masm) {
 856 #ifdef ASSERT
 857   __ mv(t0, ContinuationEntry::cookie_value());
 858   __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
 859 #endif
 860 
 861   __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
 862   __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
 863   __ sd(zr,      Address(sp, ContinuationEntry::chunk_offset()));
 864   __ sw(zr,      Address(sp, ContinuationEntry::argsize_offset()));
 865   __ sw(zr,      Address(sp, ContinuationEntry::pin_count_offset()));
 866 
 867   __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
 868   __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
 869 
 870   __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
 871 }
 872 
 873 // on entry, sp points to the ContinuationEntry
 874 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
 875 static void continuation_enter_cleanup(MacroAssembler* masm) {
 876 #ifndef PRODUCT
 877   Label OK;
 878   __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
 879   __ beq(sp, t0, OK);
 880   __ stop("incorrect sp");
 881   __ bind(OK);
 882 #endif
 883 
 884   __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
 885   __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
 886   __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
 887   __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
 888   __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
 889 }
 890 
 891 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
 892 // On entry: c_rarg1 -- the continuation object
 893 //           c_rarg2 -- isContinue
 894 //           c_rarg3 -- isVirtualThread
 895 static void gen_continuation_enter(MacroAssembler* masm,
 896                                    const methodHandle& method,
 897                                    const BasicType* sig_bt,
 898                                    const VMRegPair* regs,
 899                                    int& exception_offset,
 900                                    OopMapSet*oop_maps,
 901                                    int& frame_complete,
 902                                    int& stack_slots,
 903                                    int& interpreted_entry_offset,
 904                                    int& compiled_entry_offset) {
 905   // verify_oop_args(masm, method, sig_bt, regs);
 906   Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
 907 
 908   address start = __ pc();
 909 
 910   Label call_thaw, exit;
 911 
 912   // i2i entry used at interp_only_mode only
 913   interpreted_entry_offset = __ pc() - start;
 914   {
 915 #ifdef ASSERT
 916     Label is_interp_only;
 917     __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
 918     __ bnez(t0, is_interp_only);
 919     __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
 920     __ bind(is_interp_only);
 921 #endif
 922 
 923     // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
 924     __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
 925     __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
 926     __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
 927     __ push_cont_fastpath(xthread);
 928 
 929     __ enter();
 930     stack_slots = 2; // will be adjusted in setup
 931     OopMap* map = continuation_enter_setup(masm, stack_slots);
 932     // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
 933     // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
 934 
 935     fill_continuation_entry(masm);
 936 
 937     __ bnez(c_rarg2, call_thaw);
 938 
 939     address call_pc;
 940     {
 941       Assembler::IncompressibleScope scope(masm);
 942       // Make sure the call is patchable
 943       __ align(NativeInstruction::instruction_size);
 944 
 945       call_pc = __ reloc_call(resolve);
 946       if (call_pc == nullptr) {
 947         fatal("CodeCache is full at gen_continuation_enter");
 948       }
 949 
 950       oop_maps->add_gc_map(__ pc() - start, map);
 951       __ post_call_nop();
 952     }
 953     __ j(exit);
 954 
 955     address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
 956     if (stub == nullptr) {
 957       fatal("CodeCache is full at gen_continuation_enter");
 958     }
 959   }
 960 
 961   // compiled entry
 962   __ align(CodeEntryAlignment);
 963   compiled_entry_offset = __ pc() - start;
 964 
 965   __ enter();
 966   stack_slots = 2; // will be adjusted in setup
 967   OopMap* map = continuation_enter_setup(masm, stack_slots);
 968   frame_complete = __ pc() - start;
 969 
 970   fill_continuation_entry(masm);
 971 
 972   __ bnez(c_rarg2, call_thaw);
 973 
 974   address call_pc;
 975   {
 976     Assembler::IncompressibleScope scope(masm);
 977     // Make sure the call is patchable
 978     __ align(NativeInstruction::instruction_size);
 979 
 980     call_pc = __ reloc_call(resolve);
 981     if (call_pc == nullptr) {
 982       fatal("CodeCache is full at gen_continuation_enter");
 983     }
 984 
 985     oop_maps->add_gc_map(__ pc() - start, map);
 986     __ post_call_nop();
 987   }
 988 
 989   __ j(exit);
 990 
 991   __ bind(call_thaw);
 992 
 993   // Post call nops must be natural aligned due to cmodx rules.
 994   {
 995     Assembler::IncompressibleScope scope(masm);
 996     __ align(NativeInstruction::instruction_size);
 997 
 998     ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
 999     __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1000     oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1001     ContinuationEntry::_return_pc_offset = __ pc() - start;
1002     __ post_call_nop();
1003   }
1004 
1005   __ bind(exit);
1006   ContinuationEntry::_cleanup_offset = __ pc() - start;
1007   continuation_enter_cleanup(masm);
1008   __ leave();
1009   __ ret();
1010 
1011   // exception handling
1012   exception_offset = __ pc() - start;
1013   {
1014     __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1015 
1016     continuation_enter_cleanup(masm);
1017 
1018     __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1019     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1020 
1021     // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1022 
1023     __ mv(x11, x10); // the exception handler
1024     __ mv(x10, x9); // restore return value contaning the exception oop
1025     __ verify_oop(x10);
1026 
1027     __ leave();
1028     __ mv(x13, ra);
1029     __ jr(x11); // the exception handler
1030   }
1031 
1032   address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1033   if (stub == nullptr) {
1034     fatal("CodeCache is full at gen_continuation_enter");
1035   }
1036 }
1037 
1038 static void gen_continuation_yield(MacroAssembler* masm,
1039                                    const methodHandle& method,
1040                                    const BasicType* sig_bt,
1041                                    const VMRegPair* regs,
1042                                    OopMapSet* oop_maps,
1043                                    int& frame_complete,
1044                                    int& stack_slots,
1045                                    int& compiled_entry_offset) {
1046   enum layout {
1047     fp_off,
1048     fp_off2,
1049     return_off,
1050     return_off2,
1051     framesize // inclusive of return address
1052   };
1053   // assert(is_even(framesize/2), "sp not 16-byte aligned");
1054 
1055   stack_slots = framesize / VMRegImpl::slots_per_word;
1056   assert(stack_slots == 2, "recheck layout");
1057 
1058   address start = __ pc();
1059 
1060   compiled_entry_offset = __ pc() - start;
1061   __ enter();
1062 
1063   __ mv(c_rarg1, sp);
1064 
1065   // Post call nops must be natural aligned due to cmodx rules.
1066   __ align(NativeInstruction::instruction_size);
1067 
1068   frame_complete = __ pc() - start;
1069   address the_pc = __ pc();
1070 
1071   {
1072     Assembler::IncompressibleScope scope(masm);
1073     __ 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
1074   }
1075 
1076   __ mv(c_rarg0, xthread);
1077   __ set_last_Java_frame(sp, fp, the_pc, t0);
1078   __ call_VM_leaf(Continuation::freeze_entry(), 2);
1079   __ reset_last_Java_frame(true);
1080 
1081   Label pinned;
1082 
1083   __ bnez(x10, pinned);
1084 
1085   // We've succeeded, set sp to the ContinuationEntry
1086   __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1087   continuation_enter_cleanup(masm);
1088 
1089   __ bind(pinned); // pinned -- return to caller
1090 
1091   // handle pending exception thrown by freeze
1092   __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1093   Label ok;
1094   __ beqz(t0, ok);
1095   __ leave();
1096   __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1097   __ bind(ok);
1098 
1099   __ leave();
1100   __ ret();
1101 
1102   OopMap* map = new OopMap(framesize, 1);
1103   oop_maps->add_gc_map(the_pc - start, map);
1104 }
1105 
1106 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1107   ::continuation_enter_cleanup(masm);
1108 }
1109 
1110 static void gen_special_dispatch(MacroAssembler* masm,
1111                                  const methodHandle& method,
1112                                  const BasicType* sig_bt,
1113                                  const VMRegPair* regs) {
1114   verify_oop_args(masm, method, sig_bt, regs);
1115   vmIntrinsics::ID iid = method->intrinsic_id();
1116 
1117   // Now write the args into the outgoing interpreter space
1118   bool     has_receiver   = false;
1119   Register receiver_reg   = noreg;
1120   int      member_arg_pos = -1;
1121   Register member_reg     = noreg;
1122   int      ref_kind       = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1123   if (ref_kind != 0) {
1124     member_arg_pos = method->size_of_parameters() - 1;  // trailing MemberName argument
1125     member_reg = x9;  // known to be free at this point
1126     has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1127   } else if (iid == vmIntrinsics::_invokeBasic) {
1128     has_receiver = true;
1129   } else if (iid == vmIntrinsics::_linkToNative) {
1130     member_arg_pos = method->size_of_parameters() - 1;  // trailing NativeEntryPoint argument
1131     member_reg = x9;  // known to be free at this point
1132   } else {
1133     fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1134   }
1135 
1136   if (member_reg != noreg) {
1137     // Load the member_arg into register, if necessary.
1138     SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1139     VMReg r = regs[member_arg_pos].first();
1140     if (r->is_stack()) {
1141       __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1142     } else {
1143       // no data motion is needed
1144       member_reg = r->as_Register();
1145     }
1146   }
1147 
1148   if (has_receiver) {
1149     // Make sure the receiver is loaded into a register.
1150     assert(method->size_of_parameters() > 0, "oob");
1151     assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1152     VMReg r = regs[0].first();
1153     assert(r->is_valid(), "bad receiver arg");
1154     if (r->is_stack()) {
1155       // Porting note:  This assumes that compiled calling conventions always
1156       // pass the receiver oop in a register.  If this is not true on some
1157       // platform, pick a temp and load the receiver from stack.
1158       fatal("receiver always in a register");
1159       receiver_reg = x12;  // known to be free at this point
1160       __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1161     } else {
1162       // no data motion is needed
1163       receiver_reg = r->as_Register();
1164     }
1165   }
1166 
1167   // Figure out which address we are really jumping to:
1168   MethodHandles::generate_method_handle_dispatch(masm, iid,
1169                                                  receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1170 }
1171 
1172 // ---------------------------------------------------------------------------
1173 // Generate a native wrapper for a given method.  The method takes arguments
1174 // in the Java compiled code convention, marshals them to the native
1175 // convention (handlizes oops, etc), transitions to native, makes the call,
1176 // returns to java state (possibly blocking), unhandlizes any result and
1177 // returns.
1178 //
1179 // Critical native functions are a shorthand for the use of
1180 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1181 // functions.  The wrapper is expected to unpack the arguments before
1182 // passing them to the callee and perform checks before and after the
1183 // native call to ensure that they GCLocker
1184 // lock_critical/unlock_critical semantics are followed.  Some other
1185 // parts of JNI setup are skipped like the tear down of the JNI handle
1186 // block and the check for pending exceptions it's impossible for them
1187 // to be thrown.
1188 //
1189 // They are roughly structured like this:
1190 //    if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1191 //    tranistion to thread_in_native
1192 //    unpack array arguments and call native entry point
1193 //    check for safepoint in progress
1194 //    check if any thread suspend flags are set
1195 //      call into JVM and possible unlock the JNI critical
1196 //      if a GC was suppressed while in the critical native.
1197 //    transition back to thread_in_Java
1198 //    return to caller
1199 //
1200 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1201                                                 const methodHandle& method,
1202                                                 int compile_id,
1203                                                 BasicType* in_sig_bt,
1204                                                 VMRegPair* in_regs,
1205                                                 BasicType ret_type) {
1206   if (method->is_continuation_native_intrinsic()) {
1207     int exception_offset = -1;
1208     OopMapSet* oop_maps = new OopMapSet();
1209     int frame_complete = -1;
1210     int stack_slots = -1;
1211     int interpreted_entry_offset = -1;
1212     int vep_offset = -1;
1213     if (method->is_continuation_enter_intrinsic()) {
1214       gen_continuation_enter(masm,
1215                              method,
1216                              in_sig_bt,
1217                              in_regs,
1218                              exception_offset,
1219                              oop_maps,
1220                              frame_complete,
1221                              stack_slots,
1222                              interpreted_entry_offset,
1223                              vep_offset);
1224     } else if (method->is_continuation_yield_intrinsic()) {
1225       gen_continuation_yield(masm,
1226                              method,
1227                              in_sig_bt,
1228                              in_regs,
1229                              oop_maps,
1230                              frame_complete,
1231                              stack_slots,
1232                              vep_offset);
1233     } else {
1234       guarantee(false, "Unknown Continuation native intrinsic");
1235     }
1236 
1237 #ifdef ASSERT
1238     if (method->is_continuation_enter_intrinsic()) {
1239       assert(interpreted_entry_offset != -1, "Must be set");
1240       assert(exception_offset != -1,         "Must be set");
1241     } else {
1242       assert(interpreted_entry_offset == -1, "Must be unset");
1243       assert(exception_offset == -1,         "Must be unset");
1244     }
1245     assert(frame_complete != -1,    "Must be set");
1246     assert(stack_slots != -1,       "Must be set");
1247     assert(vep_offset != -1,        "Must be set");
1248 #endif
1249 
1250     __ flush();
1251     nmethod* nm = nmethod::new_native_nmethod(method,
1252                                               compile_id,
1253                                               masm->code(),
1254                                               vep_offset,
1255                                               frame_complete,
1256                                               stack_slots,
1257                                               in_ByteSize(-1),
1258                                               in_ByteSize(-1),
1259                                               oop_maps,
1260                                               exception_offset);
1261     if (nm == nullptr) return nm;
1262     if (method->is_continuation_enter_intrinsic()) {
1263       ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1264     } else if (method->is_continuation_yield_intrinsic()) {
1265       _cont_doYield_stub = nm;
1266     } else {
1267       guarantee(false, "Unknown Continuation native intrinsic");
1268     }
1269     return nm;
1270   }
1271 
1272   if (method->is_method_handle_intrinsic()) {
1273     vmIntrinsics::ID iid = method->intrinsic_id();
1274     intptr_t start = (intptr_t)__ pc();
1275     int vep_offset = ((intptr_t)__ pc()) - start;
1276 
1277     // First instruction must be a nop as it may need to be patched on deoptimisation
1278     {
1279       Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1280       MacroAssembler::assert_alignment(__ pc());
1281       __ nop();  // 4 bytes
1282     }
1283     gen_special_dispatch(masm,
1284                          method,
1285                          in_sig_bt,
1286                          in_regs);
1287     int frame_complete = ((intptr_t)__ pc()) - start;  // not complete, period
1288     __ flush();
1289     int stack_slots = SharedRuntime::out_preserve_stack_slots();  // no out slots at all, actually
1290     return nmethod::new_native_nmethod(method,
1291                                        compile_id,
1292                                        masm->code(),
1293                                        vep_offset,
1294                                        frame_complete,
1295                                        stack_slots / VMRegImpl::slots_per_word,
1296                                        in_ByteSize(-1),
1297                                        in_ByteSize(-1),
1298                                        (OopMapSet*)nullptr);
1299   }
1300   address native_func = method->native_function();
1301   assert(native_func != nullptr, "must have function");
1302 
1303   // An OopMap for lock (and class if static)
1304   OopMapSet *oop_maps = new OopMapSet();
1305   assert_cond(oop_maps != nullptr);
1306   intptr_t start = (intptr_t)__ pc();
1307 
1308   // We have received a description of where all the java arg are located
1309   // on entry to the wrapper. We need to convert these args to where
1310   // the jni function will expect them. To figure out where they go
1311   // we convert the java signature to a C signature by inserting
1312   // the hidden arguments as arg[0] and possibly arg[1] (static method)
1313 
1314   const int total_in_args = method->size_of_parameters();
1315   int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1316 
1317   BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1318   VMRegPair* out_regs   = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1319 
1320   int argc = 0;
1321   out_sig_bt[argc++] = T_ADDRESS;
1322   if (method->is_static()) {
1323     out_sig_bt[argc++] = T_OBJECT;
1324   }
1325 
1326   for (int i = 0; i < total_in_args ; i++) {
1327     out_sig_bt[argc++] = in_sig_bt[i];
1328   }
1329 
1330   // Now figure out where the args must be stored and how much stack space
1331   // they require.
1332   int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1333 
1334   // Compute framesize for the wrapper.  We need to handlize all oops in
1335   // incoming registers
1336 
1337   // Calculate the total number of stack slots we will need.
1338 
1339   // First count the abi requirement plus all of the outgoing args
1340   int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1341 
1342   // Now the space for the inbound oop handle area
1343   int total_save_slots = 8 * VMRegImpl::slots_per_word;  // 8 arguments passed in registers
1344 
1345   int oop_handle_offset = stack_slots;
1346   stack_slots += total_save_slots;
1347 
1348   // Now any space we need for handlizing a klass if static method
1349 
1350   int klass_slot_offset = 0;
1351   int klass_offset = -1;
1352   int lock_slot_offset = 0;
1353   bool is_static = false;
1354 
1355   if (method->is_static()) {
1356     klass_slot_offset = stack_slots;
1357     stack_slots += VMRegImpl::slots_per_word;
1358     klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1359     is_static = true;
1360   }
1361 
1362   // Plus a lock if needed
1363 
1364   if (method->is_synchronized()) {
1365     lock_slot_offset = stack_slots;
1366     stack_slots += VMRegImpl::slots_per_word;
1367   }
1368 
1369   // Now a place (+2) to save return values or temp during shuffling
1370   // + 4 for return address (which we own) and saved fp
1371   stack_slots += 6;
1372 
1373   // Ok The space we have allocated will look like:
1374   //
1375   //
1376   // FP-> |                     |
1377   //      | 2 slots (ra)        |
1378   //      | 2 slots (fp)        |
1379   //      |---------------------|
1380   //      | 2 slots for moves   |
1381   //      |---------------------|
1382   //      | lock box (if sync)  |
1383   //      |---------------------| <- lock_slot_offset
1384   //      | klass (if static)   |
1385   //      |---------------------| <- klass_slot_offset
1386   //      | oopHandle area      |
1387   //      |---------------------| <- oop_handle_offset (8 java arg registers)
1388   //      | outbound memory     |
1389   //      | based arguments     |
1390   //      |                     |
1391   //      |---------------------|
1392   //      |                     |
1393   // SP-> | out_preserved_slots |
1394   //
1395   //
1396 
1397 
1398   // Now compute actual number of stack words we need rounding to make
1399   // stack properly aligned.
1400   stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1401 
1402   int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1403 
1404   // First thing make an ic check to see if we should even be here
1405 
1406   // We are free to use all registers as temps without saving them and
1407   // restoring them except fp. fp is the only callee save register
1408   // as far as the interpreter and the compiler(s) are concerned.
1409 
1410   const Register receiver = j_rarg0;
1411 
1412   __ verify_oop(receiver);
1413   assert_different_registers(receiver, t0, t1);
1414 
1415   __ ic_check();
1416 
1417   int vep_offset = ((intptr_t)__ pc()) - start;
1418 
1419   // If we have to make this method not-entrant we'll overwrite its
1420   // first instruction with a jump.
1421   {
1422     Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1423     MacroAssembler::assert_alignment(__ pc());
1424     __ nop();  // 4 bytes
1425   }
1426 
1427   if (method->needs_clinit_barrier()) {
1428     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
1429     Label L_skip_barrier;
1430     __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1431     __ clinit_barrier(t1, t0, &L_skip_barrier);
1432     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1433 
1434     __ bind(L_skip_barrier);
1435   }
1436 
1437   // Generate stack overflow check
1438   __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1439 
1440   // Generate a new frame for the wrapper.
1441   __ enter();
1442   // -2 because return address is already present and so is saved fp
1443   __ sub(sp, sp, stack_size - 2 * wordSize);
1444 
1445   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1446   assert_cond(bs != nullptr);
1447   bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1448 
1449   // Frame is now completed as far as size and linkage.
1450   int frame_complete = ((intptr_t)__ pc()) - start;
1451 
1452   // We use x18 as the oop handle for the receiver/klass
1453   // It is callee save so it survives the call to native
1454 
1455   const Register oop_handle_reg = x18;
1456 
1457   //
1458   // We immediately shuffle the arguments so that any vm call we have to
1459   // make from here on out (sync slow path, jvmti, etc.) we will have
1460   // captured the oops from our caller and have a valid oopMap for
1461   // them.
1462 
1463   // -----------------
1464   // The Grand Shuffle
1465 
1466   // The Java calling convention is either equal (linux) or denser (win64) than the
1467   // c calling convention. However the because of the jni_env argument the c calling
1468   // convention always has at least one more (and two for static) arguments than Java.
1469   // Therefore if we move the args from java -> c backwards then we will never have
1470   // a register->register conflict and we don't have to build a dependency graph
1471   // and figure out how to break any cycles.
1472   //
1473 
1474   // Record esp-based slot for receiver on stack for non-static methods
1475   int receiver_offset = -1;
1476 
1477   // This is a trick. We double the stack slots so we can claim
1478   // the oops in the caller's frame. Since we are sure to have
1479   // more args than the caller doubling is enough to make
1480   // sure we can capture all the incoming oop args from the
1481   // caller.
1482   //
1483   OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1484   assert_cond(map != nullptr);
1485 
1486   int float_args = 0;
1487   int int_args = 0;
1488 
1489 #ifdef ASSERT
1490   bool reg_destroyed[Register::number_of_registers];
1491   bool freg_destroyed[FloatRegister::number_of_registers];
1492   for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1493     reg_destroyed[r] = false;
1494   }
1495   for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1496     freg_destroyed[f] = false;
1497   }
1498 
1499 #endif /* ASSERT */
1500 
1501   // For JNI natives the incoming and outgoing registers are offset upwards.
1502   GrowableArray<int> arg_order(2 * total_in_args);
1503 
1504   for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1505     arg_order.push(i);
1506     arg_order.push(c_arg);
1507   }
1508 
1509   for (int ai = 0; ai < arg_order.length(); ai += 2) {
1510     int i = arg_order.at(ai);
1511     int c_arg = arg_order.at(ai + 1);
1512     __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1513     assert(c_arg != -1 && i != -1, "wrong order");
1514 #ifdef ASSERT
1515     if (in_regs[i].first()->is_Register()) {
1516       assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1517     } else if (in_regs[i].first()->is_FloatRegister()) {
1518       assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1519     }
1520     if (out_regs[c_arg].first()->is_Register()) {
1521       reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1522     } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1523       freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1524     }
1525 #endif /* ASSERT */
1526     switch (in_sig_bt[i]) {
1527       case T_ARRAY:
1528       case T_OBJECT:
1529         __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1530                        ((i == 0) && (!is_static)),
1531                        &receiver_offset);
1532         int_args++;
1533         break;
1534       case T_VOID:
1535         break;
1536 
1537       case T_FLOAT:
1538         __ float_move(in_regs[i], out_regs[c_arg]);
1539         float_args++;
1540         break;
1541 
1542       case T_DOUBLE:
1543         assert( i + 1 < total_in_args &&
1544                 in_sig_bt[i + 1] == T_VOID &&
1545                 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1546         __ double_move(in_regs[i], out_regs[c_arg]);
1547         float_args++;
1548         break;
1549 
1550       case T_LONG :
1551         __ long_move(in_regs[i], out_regs[c_arg]);
1552         int_args++;
1553         break;
1554 
1555       case T_ADDRESS:
1556         assert(false, "found T_ADDRESS in java args");
1557         break;
1558 
1559       default:
1560         __ move32_64(in_regs[i], out_regs[c_arg]);
1561         int_args++;
1562     }
1563   }
1564 
1565   // point c_arg at the first arg that is already loaded in case we
1566   // need to spill before we call out
1567   int c_arg = total_c_args - total_in_args;
1568 
1569   // Pre-load a static method's oop into c_rarg1.
1570   if (method->is_static()) {
1571 
1572     //  load oop into a register
1573     __ movoop(c_rarg1,
1574               JNIHandles::make_local(method->method_holder()->java_mirror()));
1575 
1576     // Now handlize the static class mirror it's known not-null.
1577     __ sd(c_rarg1, Address(sp, klass_offset));
1578     map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1579 
1580     // Now get the handle
1581     __ la(c_rarg1, Address(sp, klass_offset));
1582     // and protect the arg if we must spill
1583     c_arg--;
1584   }
1585 
1586   // Change state to native (we save the return address in the thread, since it might not
1587   // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1588   // points into the right code segment. It does not have to be the correct return pc.
1589   // We use the same pc/oopMap repeatedly when we call out.
1590 
1591   Label native_return;
1592   if (method->is_object_wait0()) {
1593     // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1594     __ set_last_Java_frame(sp, noreg, native_return, t0);
1595   } else {
1596     intptr_t the_pc = (intptr_t) __ pc();
1597     oop_maps->add_gc_map(the_pc - start, map);
1598 
1599     __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1600   }
1601 
1602   Label dtrace_method_entry, dtrace_method_entry_done;
1603   if (DTraceMethodProbes) {
1604     __ j(dtrace_method_entry);
1605     __ bind(dtrace_method_entry_done);
1606   }
1607 
1608   // RedefineClasses() tracing support for obsolete method entry
1609   if (log_is_enabled(Trace, redefine, class, obsolete)) {
1610     // protect the args we've loaded
1611     save_args(masm, total_c_args, c_arg, out_regs);
1612     __ mov_metadata(c_rarg1, method());
1613     __ call_VM_leaf(
1614       CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1615       xthread, c_rarg1);
1616     restore_args(masm, total_c_args, c_arg, out_regs);
1617   }
1618 
1619   // Lock a synchronized method
1620 
1621   // Register definitions used by locking and unlocking
1622 
1623   const Register swap_reg = x10;
1624   const Register obj_reg  = x9;  // Will contain the oop
1625   const Register lock_reg = x30;  // Address of compiler lock object (BasicLock)
1626   const Register old_hdr  = x30;  // value of old header at unlock time
1627   const Register lock_tmp = x31;  // Temporary used by fast_lock/unlock
1628   const Register tmp      = ra;
1629 
1630   Label slow_path_lock;
1631   Label lock_done;
1632 
1633   if (method->is_synchronized()) {
1634     // Get the handle (the 2nd argument)
1635     __ mv(oop_handle_reg, c_rarg1);
1636 
1637     // Get address of the box
1638 
1639     __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1640 
1641     // Load the oop from the handle
1642     __ ld(obj_reg, Address(oop_handle_reg, 0));
1643 
1644     __ fast_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1645 
1646     // Slow path will re-enter here
1647     __ bind(lock_done);
1648   }
1649 
1650 
1651   // Finally just about ready to make the JNI call
1652 
1653   // get JNIEnv* which is first argument to native
1654   __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1655 
1656   // Now set thread in native
1657   __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1658   __ mv(t0, _thread_in_native);
1659   __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1660   __ sw(t0, Address(t1));
1661 
1662   // Clobbers t1
1663   __ rt_call(native_func);
1664 
1665   // Verify or restore cpu control state after JNI call
1666   __ restore_cpu_control_state_after_jni(t0);
1667 
1668   // Unpack native results.
1669   if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1670     __ cast_primitive_type(ret_type, x10);
1671   }
1672 
1673   Label safepoint_in_progress, safepoint_in_progress_done;
1674 
1675   // Switch thread to "native transition" state before reading the synchronization state.
1676   // This additional state is necessary because reading and testing the synchronization
1677   // state is not atomic w.r.t. GC, as this scenario demonstrates:
1678   //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1679   //     VM thread changes sync state to synchronizing and suspends threads for GC.
1680   //     Thread A is resumed to finish this native method, but doesn't block here since it
1681   //     didn't see any synchronization is progress, and escapes.
1682   __ mv(t0, _thread_in_native_trans);
1683 
1684   __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1685 
1686   // Force this write out before the read below
1687   if (!UseSystemMemoryBarrier) {
1688     __ membar(MacroAssembler::AnyAny);
1689   }
1690 
1691   // check for safepoint operation in progress and/or pending suspend requests
1692   {
1693     __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
1694     __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1695     __ bnez(t0, safepoint_in_progress);
1696     __ bind(safepoint_in_progress_done);
1697   }
1698 
1699   // change thread state
1700   __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1701   __ mv(t0, _thread_in_Java);
1702   __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1703   __ sw(t0, Address(t1));
1704 
1705   if (method->is_object_wait0()) {
1706     // Check preemption for Object.wait()
1707     __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1708     __ beqz(t1, native_return);
1709     __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1710     __ jr(t1);
1711     __ bind(native_return);
1712 
1713     intptr_t the_pc = (intptr_t) __ pc();
1714     oop_maps->add_gc_map(the_pc - start, map);
1715   }
1716 
1717   Label reguard;
1718   Label reguard_done;
1719   __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1720   __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1721   __ beq(t0, t1, reguard);
1722   __ bind(reguard_done);
1723 
1724   // native result if any is live
1725 
1726   // Unlock
1727   Label unlock_done;
1728   Label slow_path_unlock;
1729   if (method->is_synchronized()) {
1730 
1731     // Get locked oop from the handle we passed to jni
1732     __ ld(obj_reg, Address(oop_handle_reg, 0));
1733 
1734     // Must save x10 if if it is live now because cmpxchg must use it
1735     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1736       save_native_result(masm, ret_type, stack_slots);
1737     }
1738 
1739     __ fast_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1740 
1741     // slow path re-enters here
1742     __ bind(unlock_done);
1743     if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1744       restore_native_result(masm, ret_type, stack_slots);
1745     }
1746   }
1747 
1748   Label dtrace_method_exit, dtrace_method_exit_done;
1749   if (DTraceMethodProbes) {
1750     __ j(dtrace_method_exit);
1751     __ bind(dtrace_method_exit_done);
1752   }
1753 
1754   __ reset_last_Java_frame(false);
1755 
1756   // Unbox oop result, e.g. JNIHandles::resolve result.
1757   if (is_reference_type(ret_type)) {
1758     __ resolve_jobject(x10, x11, x12);
1759   }
1760 
1761   if (CheckJNICalls) {
1762     // clear_pending_jni_exception_check
1763     __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1764   }
1765 
1766   // reset handle block
1767   __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1768   __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1769 
1770   __ leave();
1771 
1772   #if INCLUDE_JFR
1773   // We need to do a poll test after unwind in case the sampler
1774   // managed to sample the native frame after returning to Java.
1775   Label L_return;
1776   __ ld(t0, Address(xthread, JavaThread::polling_word_offset()));
1777   address poll_test_pc = __ pc();
1778   __ relocate(relocInfo::poll_return_type);
1779   __ test_bit(t0, t0, log2i_exact(SafepointMechanism::poll_bit()));
1780   __ beqz(t0, L_return);
1781   assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
1782          "polling page return stub not created yet");
1783   address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
1784   __ la(t0, InternalAddress(poll_test_pc));
1785   __ sd(t0, Address(xthread, JavaThread::saved_exception_pc_offset()));
1786   __ far_jump(RuntimeAddress(stub));
1787   __ bind(L_return);
1788 #endif // INCLUDE_JFR
1789 
1790   // Any exception pending?
1791   Label exception_pending;
1792   __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1793   __ bnez(t0, exception_pending);
1794 
1795   // We're done
1796   __ ret();
1797 
1798   // Unexpected paths are out of line and go here
1799 
1800   // forward the exception
1801   __ bind(exception_pending);
1802 
1803   // and forward the exception
1804   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1805 
1806   // Slow path locking & unlocking
1807   if (method->is_synchronized()) {
1808 
1809     __ block_comment("Slow path lock {");
1810     __ bind(slow_path_lock);
1811 
1812     // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1813     // args are (oop obj, BasicLock* lock, JavaThread* thread)
1814 
1815     // protect the args we've loaded
1816     save_args(masm, total_c_args, c_arg, out_regs);
1817 
1818     __ mv(c_rarg0, obj_reg);
1819     __ mv(c_rarg1, lock_reg);
1820     __ mv(c_rarg2, xthread);
1821 
1822     // Not a leaf but we have last_Java_frame setup as we want.
1823     // We don't want to unmount in case of contention since that would complicate preserving
1824     // the arguments that had already been marshalled into the native convention. So we force
1825     // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1826     // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1827     __ push_cont_fastpath();
1828     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1829     __ pop_cont_fastpath();
1830     restore_args(masm, total_c_args, c_arg, out_regs);
1831 
1832 #ifdef ASSERT
1833     { Label L;
1834       __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1835       __ beqz(t0, L);
1836       __ stop("no pending exception allowed on exit from monitorenter");
1837       __ bind(L);
1838     }
1839 #endif
1840     __ j(lock_done);
1841 
1842     __ block_comment("} Slow path lock");
1843 
1844     __ block_comment("Slow path unlock {");
1845     __ bind(slow_path_unlock);
1846 
1847     if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1848       save_native_result(masm, ret_type, stack_slots);
1849     }
1850 
1851     __ mv(c_rarg2, xthread);
1852     __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1853     __ mv(c_rarg0, obj_reg);
1854 
1855     // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1856     // NOTE that obj_reg == x9 currently
1857     __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1858     __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1859 
1860     __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1861 
1862 #ifdef ASSERT
1863     {
1864       Label L;
1865       __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1866       __ beqz(t0, L);
1867       __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1868       __ bind(L);
1869     }
1870 #endif /* ASSERT */
1871 
1872     __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1873 
1874     if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1875       restore_native_result(masm, ret_type, stack_slots);
1876     }
1877     __ j(unlock_done);
1878 
1879     __ block_comment("} Slow path unlock");
1880 
1881   } // synchronized
1882 
1883   // SLOW PATH Reguard the stack if needed
1884 
1885   __ bind(reguard);
1886   save_native_result(masm, ret_type, stack_slots);
1887   __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
1888   restore_native_result(masm, ret_type, stack_slots);
1889   // and continue
1890   __ j(reguard_done);
1891 
1892   // SLOW PATH safepoint
1893   {
1894     __ block_comment("safepoint {");
1895     __ bind(safepoint_in_progress);
1896 
1897     // Don't use call_VM as it will see a possible pending exception and forward it
1898     // and never return here preventing us from clearing _last_native_pc down below.
1899     //
1900     save_native_result(masm, ret_type, stack_slots);
1901     __ mv(c_rarg0, xthread);
1902 #ifndef PRODUCT
1903     assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
1904 #endif
1905     __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
1906 
1907     // Restore any method result value
1908     restore_native_result(masm, ret_type, stack_slots);
1909 
1910     __ j(safepoint_in_progress_done);
1911     __ block_comment("} safepoint");
1912   }
1913 
1914   // SLOW PATH dtrace support
1915   if (DTraceMethodProbes) {
1916     {
1917       __ block_comment("dtrace entry {");
1918       __ bind(dtrace_method_entry);
1919 
1920       // We have all of the arguments setup at this point. We must not touch any register
1921       // argument registers at this point (what if we save/restore them there are no oop?
1922 
1923       save_args(masm, total_c_args, c_arg, out_regs);
1924       __ mov_metadata(c_rarg1, method());
1925       __ call_VM_leaf(
1926         CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1927         xthread, c_rarg1);
1928       restore_args(masm, total_c_args, c_arg, out_regs);
1929       __ j(dtrace_method_entry_done);
1930       __ block_comment("} dtrace entry");
1931     }
1932 
1933     {
1934       __ block_comment("dtrace exit {");
1935       __ bind(dtrace_method_exit);
1936       save_native_result(masm, ret_type, stack_slots);
1937       __ mov_metadata(c_rarg1, method());
1938       __ call_VM_leaf(
1939            CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1940            xthread, c_rarg1);
1941       restore_native_result(masm, ret_type, stack_slots);
1942       __ j(dtrace_method_exit_done);
1943       __ block_comment("} dtrace exit");
1944     }
1945   }
1946 
1947   __ flush();
1948 
1949   nmethod *nm = nmethod::new_native_nmethod(method,
1950                                             compile_id,
1951                                             masm->code(),
1952                                             vep_offset,
1953                                             frame_complete,
1954                                             stack_slots / VMRegImpl::slots_per_word,
1955                                             (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
1956                                             in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
1957                                             oop_maps);
1958   assert(nm != nullptr, "create native nmethod fail!");
1959   return nm;
1960 }
1961 
1962 // this function returns the adjust size (in number of words) to a c2i adapter
1963 // activation for use during deoptimization
1964 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
1965   assert(callee_locals >= callee_parameters,
1966          "test and remove; got more parms than locals");
1967   if (callee_locals < callee_parameters) {
1968     return 0;                   // No adjustment for negative locals
1969   }
1970   int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
1971   // diff is counted in stack words
1972   return align_up(diff, 2);
1973 }
1974 
1975 //------------------------------generate_deopt_blob----------------------------
1976 void SharedRuntime::generate_deopt_blob() {
1977   // Allocate space for the code
1978   ResourceMark rm;
1979   // Setup code generation tools
1980   int pad = 0;
1981   const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
1982   CodeBuffer buffer(name, 2048 + pad, 1024);
1983   MacroAssembler* masm = new MacroAssembler(&buffer);
1984   int frame_size_in_words = -1;
1985   OopMap* map = nullptr;
1986   OopMapSet *oop_maps = new OopMapSet();
1987   assert_cond(masm != nullptr && oop_maps != nullptr);
1988   RegisterSaver reg_saver(COMPILER2_PRESENT(true) NOT_COMPILER2(false));
1989 
1990   // -------------
1991   // This code enters when returning to a de-optimized nmethod.  A return
1992   // address has been pushed on the stack, and return values are in
1993   // registers.
1994   // If we are doing a normal deopt then we were called from the patched
1995   // nmethod from the point we returned to the nmethod. So the return
1996   // address on the stack is wrong by NativeCall::instruction_size
1997   // We will adjust the value so it looks like we have the original return
1998   // address on the stack (like when we eagerly deoptimized).
1999   // In the case of an exception pending when deoptimizing, we enter
2000   // with a return address on the stack that points after the call we patched
2001   // into the exception handler. We have the following register state from,
2002   // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2003   //    x10: exception oop
2004   //    x9: exception handler
2005   //    x13: throwing pc
2006   // So in this case we simply jam x13 into the useless return address and
2007   // the stack looks just like we want.
2008   //
2009   // At this point we need to de-opt.  We save the argument return
2010   // registers.  We call the first C routine, fetch_unroll_info().  This
2011   // routine captures the return values and returns a structure which
2012   // describes the current frame size and the sizes of all replacement frames.
2013   // The current frame is compiled code and may contain many inlined
2014   // functions, each with their own JVM state.  We pop the current frame, then
2015   // push all the new frames.  Then we call the C routine unpack_frames() to
2016   // populate these frames.  Finally unpack_frames() returns us the new target
2017   // address.  Notice that callee-save registers are BLOWN here; they have
2018   // already been captured in the vframeArray at the time the return PC was
2019   // patched.
2020   address start = __ pc();
2021   Label cont;
2022 
2023   // Prolog for non exception case!
2024 
2025   // Save everything in sight.
2026   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2027 
2028   // Normal deoptimization.  Save exec mode for unpack_frames.
2029   __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2030   __ j(cont);
2031 
2032   int reexecute_offset = __ pc() - start;
2033   // Reexecute case
2034   // return address is the pc describes what bci to do re-execute at
2035 
2036   // No need to update map as each call to save_live_registers will produce identical oopmap
2037   (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2038 
2039   __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2040   __ j(cont);
2041 
2042   int exception_offset = __ pc() - start;
2043 
2044   // Prolog for exception case
2045 
2046   // all registers are dead at this entry point, except for x10, and
2047   // x13 which contain the exception oop and exception pc
2048   // respectively.  Set them in TLS and fall thru to the
2049   // unpack_with_exception_in_tls entry point.
2050 
2051   __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2052   __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2053 
2054   int exception_in_tls_offset = __ pc() - start;
2055 
2056   // new implementation because exception oop is now passed in JavaThread
2057 
2058   // Prolog for exception case
2059   // All registers must be preserved because they might be used by LinearScan
2060   // Exceptiop oop and throwing PC are passed in JavaThread
2061   // tos: stack at point of call to method that threw the exception (i.e. only
2062   // args are on the stack, no return address)
2063 
2064   // The return address pushed by save_live_registers will be patched
2065   // later with the throwing pc. The correct value is not available
2066   // now because loading it from memory would destroy registers.
2067 
2068   // NB: The SP at this point must be the SP of the method that is
2069   // being deoptimized.  Deoptimization assumes that the frame created
2070   // here by save_live_registers is immediately below the method's SP.
2071   // This is a somewhat fragile mechanism.
2072 
2073   // Save everything in sight.
2074   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2075 
2076   // Now it is safe to overwrite any register
2077 
2078   // Deopt during an exception.  Save exec mode for unpack_frames.
2079   __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2080 
2081   // load throwing pc from JavaThread and patch it as the return address
2082   // of the current frame. Then clear the field in JavaThread
2083 
2084   __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2085   __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2086   __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2087 
2088 #ifdef ASSERT
2089   // verify that there is really an exception oop in JavaThread
2090   __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2091   __ verify_oop(x10);
2092 
2093   // verify that there is no pending exception
2094   Label no_pending_exception;
2095   __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2096   __ beqz(t0, no_pending_exception);
2097   __ stop("must not have pending exception here");
2098   __ bind(no_pending_exception);
2099 #endif
2100 
2101   __ bind(cont);
2102 
2103   // Call C code.  Need thread and this frame, but NOT official VM entry
2104   // crud.  We cannot block on this call, no GC can happen.
2105   //
2106   // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2107 
2108   // fetch_unroll_info needs to call last_java_frame().
2109 
2110   Label retaddr;
2111   __ set_last_Java_frame(sp, noreg, retaddr, t0);
2112 #ifdef ASSERT
2113   {
2114     Label L;
2115     __ ld(t0, Address(xthread,
2116                               JavaThread::last_Java_fp_offset()));
2117     __ beqz(t0, L);
2118     __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2119     __ bind(L);
2120   }
2121 #endif // ASSERT
2122   __ mv(c_rarg0, xthread);
2123   __ mv(c_rarg1, xcpool);
2124   __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2125   __ bind(retaddr);
2126 
2127   // Need to have an oopmap that tells fetch_unroll_info where to
2128   // find any register it might need.
2129   oop_maps->add_gc_map(__ pc() - start, map);
2130 
2131   __ reset_last_Java_frame(false);
2132 
2133   // Load UnrollBlock* into x15
2134   __ mv(x15, x10);
2135 
2136   __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2137   Label noException;
2138   __ mv(t0, Deoptimization::Unpack_exception);
2139   __ bne(xcpool, t0, noException); // Was exception pending?
2140   __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2141   __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2142   __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2143   __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2144 
2145   __ verify_oop(x10);
2146 
2147   // Overwrite the result registers with the exception results.
2148   __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2149 
2150   __ bind(noException);
2151 
2152   // Only register save data is on the stack.
2153   // Now restore the result registers.  Everything else is either dead
2154   // or captured in the vframeArray.
2155 
2156   // Restore fp result register
2157   __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2158   // Restore integer result register
2159   __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2160 
2161   // Pop all of the register save area off the stack
2162   __ add(sp, sp, frame_size_in_words * wordSize);
2163 
2164   // All of the register save area has been popped of the stack. Only the
2165   // return address remains.
2166 
2167   // Pop all the frames we must move/replace.
2168   //
2169   // Frame picture (youngest to oldest)
2170   // 1: self-frame (no frame link)
2171   // 2: deopting frame  (no frame link)
2172   // 3: caller of deopting frame (could be compiled/interpreted).
2173   //
2174   // Note: by leaving the return address of self-frame on the stack
2175   // and using the size of frame 2 to adjust the stack
2176   // when we are done the return to frame 3 will still be on the stack.
2177 
2178   // Pop deoptimized frame
2179   __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2180   __ subi(x12, x12, 2 * wordSize);
2181   __ add(sp, sp, x12);
2182   __ ld(fp, Address(sp, 0));
2183   __ ld(ra, Address(sp, wordSize));
2184   __ addi(sp, sp, 2 * wordSize);
2185   // RA should now be the return address to the caller (3)
2186 
2187 #ifdef ASSERT
2188   // Compilers generate code that bang the stack by as much as the
2189   // interpreter would need. So this stack banging should never
2190   // trigger a fault. Verify that it does not on non product builds.
2191   __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2192   __ bang_stack_size(x9, x12);
2193 #endif
2194   // Load address of array of frame pcs into x12
2195   __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2196 
2197   // Load address of array of frame sizes into x14
2198   __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2199 
2200   // Load counter into x13
2201   __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2202 
2203   // Now adjust the caller's stack to make up for the extra locals
2204   // but record the original sp so that we can save it in the skeletal interpreter
2205   // frame and the stack walking of interpreter_sender will get the unextended sp
2206   // value and not the "real" sp value.
2207 
2208   const Register sender_sp = x16;
2209 
2210   __ mv(sender_sp, sp);
2211   __ lwu(x9, Address(x15,
2212                      Deoptimization::UnrollBlock::
2213                      caller_adjustment_offset()));
2214   __ sub(sp, sp, x9);
2215 
2216   // Push interpreter frames in a loop
2217   __ mv(t0, 0xDEADDEAD);               // Make a recognizable pattern
2218   __ mv(t1, t0);
2219   Label loop;
2220   __ bind(loop);
2221   __ ld(x9, Address(x14, 0));          // Load frame size
2222   __ addi(x14, x14, wordSize);
2223   __ subi(x9, x9, 2 * wordSize);       // We'll push pc and fp by hand
2224   __ ld(ra, Address(x12, 0));          // Load pc
2225   __ addi(x12, x12, wordSize);
2226   __ enter();                          // Save old & set new fp
2227   __ sub(sp, sp, x9);                  // Prolog
2228   // This value is corrected by layout_activation_impl
2229   __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2230   __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2231   __ mv(sender_sp, sp);                // Pass sender_sp to next frame
2232   __ subi(x13, x13, 1);                // Decrement counter
2233   __ bnez(x13, loop);
2234 
2235     // Re-push self-frame
2236   __ ld(ra, Address(x12));
2237   __ enter();
2238 
2239   // Allocate a full sized register save area.  We subtract 2 because
2240   // enter() just pushed 2 words
2241   __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2242 
2243   // Restore frame locals after moving the frame
2244   __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2245   __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2246 
2247   // Call C code.  Need thread but NOT official VM entry
2248   // crud.  We cannot block on this call, no GC can happen.  Call should
2249   // restore return values to their stack-slots with the new SP.
2250   //
2251   // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2252 
2253   // Use fp because the frames look interpreted now
2254   // Don't need the precise return PC here, just precise enough to point into this code blob.
2255   address the_pc = __ pc();
2256   __ set_last_Java_frame(sp, fp, the_pc, t0);
2257 
2258   __ mv(c_rarg0, xthread);
2259   __ mv(c_rarg1, xcpool); // second arg: exec_mode
2260   __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2261 
2262   // Set an oopmap for the call site
2263   // Use the same PC we used for the last java frame
2264   oop_maps->add_gc_map(the_pc - start,
2265                        new OopMap(frame_size_in_words, 0));
2266 
2267   // Clear fp AND pc
2268   __ reset_last_Java_frame(true);
2269 
2270   // Collect return values
2271   __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2272   __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2273 
2274   // Pop self-frame.
2275   __ leave();                           // Epilog
2276 
2277   // Jump to interpreter
2278   __ ret();
2279 
2280   // Make sure all code is generated
2281   masm->flush();
2282 
2283   _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2284   assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2285   _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2286 }
2287 
2288 // Number of stack slots between incoming argument block and the start of
2289 // a new frame. The PROLOG must add this many slots to the stack. The
2290 // EPILOG must remove this many slots.
2291 // RISCV needs two words for RA (return address) and FP (frame pointer).
2292 uint SharedRuntime::in_preserve_stack_slots() {
2293   return 2 * VMRegImpl::slots_per_word + (VerifyStackAtCalls ? 0 : 2) ;
2294 }
2295 
2296 uint SharedRuntime::out_preserve_stack_slots() {
2297   return 0;
2298 }
2299 
2300 VMReg SharedRuntime::thread_register() {
2301   return xthread->as_VMReg();
2302 }
2303 
2304 //------------------------------generate_handler_blob------
2305 //
2306 // Generate a special Compile2Runtime blob that saves all registers,
2307 // and setup oopmap.
2308 //
2309 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2310   assert(is_polling_page_id(id), "expected a polling page stub id");
2311 
2312   ResourceMark rm;
2313   OopMapSet *oop_maps = new OopMapSet();
2314   assert_cond(oop_maps != nullptr);
2315   OopMap* map = nullptr;
2316 
2317   // Allocate space for the code.  Setup code generation tools.
2318   const char* name = SharedRuntime::stub_name(id);
2319   CodeBuffer buffer(name, 2048, 1024);
2320   MacroAssembler* masm = new MacroAssembler(&buffer);
2321   assert_cond(masm != nullptr);
2322 
2323   address start   = __ pc();
2324   address call_pc = nullptr;
2325   int frame_size_in_words = -1;
2326   bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2327   RegisterSaver reg_saver(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2328 
2329   // Save Integer and Float registers.
2330   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2331 
2332   // The following is basically a call_VM.  However, we need the precise
2333   // address of the call in order to generate an oopmap. Hence, we do all the
2334   // work ourselves.
2335 
2336   Label retaddr;
2337   __ set_last_Java_frame(sp, noreg, retaddr, t0);
2338 
2339   // The return address must always be correct so that frame constructor never
2340   // sees an invalid pc.
2341 
2342   if (!cause_return) {
2343     // overwrite the return address pushed by save_live_registers
2344     // Additionally, x18 is a callee-saved register so we can look at
2345     // it later to determine if someone changed the return address for
2346     // us!
2347     __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2348     __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2349   }
2350 
2351   // Do the call
2352   __ mv(c_rarg0, xthread);
2353   __ rt_call(call_ptr);
2354   __ bind(retaddr);
2355 
2356   // Set an oopmap for the call site.  This oopmap will map all
2357   // oop-registers and debug-info registers as callee-saved.  This
2358   // will allow deoptimization at this safepoint to find all possible
2359   // debug-info recordings, as well as let GC find all oops.
2360 
2361   oop_maps->add_gc_map( __ pc() - start, map);
2362 
2363   Label noException;
2364 
2365   __ reset_last_Java_frame(false);
2366 
2367   __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2368 
2369   __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2370   __ beqz(t0, noException);
2371 
2372   // Exception pending
2373 
2374   reg_saver.restore_live_registers(masm);
2375 
2376   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2377 
2378   // No exception case
2379   __ bind(noException);
2380 
2381   Label no_adjust, bail;
2382   if (!cause_return) {
2383     // If our stashed return pc was modified by the runtime we avoid touching it
2384     __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2385     __ bne(x18, t0, no_adjust);
2386 
2387 #ifdef ASSERT
2388     // Verify the correct encoding of the poll we're about to skip.
2389     // See NativeInstruction::is_lwu_to_zr()
2390     __ lwu(t0, Address(x18));
2391     __ andi(t1, t0, 0b1111111);
2392     __ mv(t2, 0b0000011);
2393     __ bne(t1, t2, bail); // 0-6:0b0000011
2394     __ srli(t1, t0, 7);
2395     __ andi(t1, t1, 0b11111);
2396     __ bnez(t1, bail);    // 7-11:0b00000
2397     __ srli(t1, t0, 12);
2398     __ andi(t1, t1, 0b111);
2399     __ mv(t2, 0b110);
2400     __ bne(t1, t2, bail); // 12-14:0b110
2401 #endif
2402 
2403     // Adjust return pc forward to step over the safepoint poll instruction
2404     __ addi(x18, x18, NativeInstruction::instruction_size);
2405     __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2406   }
2407 
2408   __ bind(no_adjust);
2409   // Normal exit, restore registers and exit.
2410 
2411   reg_saver.restore_live_registers(masm);
2412   __ ret();
2413 
2414 #ifdef ASSERT
2415   __ bind(bail);
2416   __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2417 #endif
2418 
2419   // Make sure all code is generated
2420   masm->flush();
2421 
2422   // Fill-out other meta info
2423   return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2424 }
2425 
2426 //
2427 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2428 //
2429 // Generate a stub that calls into vm to find out the proper destination
2430 // of a java call. All the argument registers are live at this point
2431 // but since this is generic code we don't know what they are and the caller
2432 // must do any gc of the args.
2433 //
2434 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2435   assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2436   assert(is_resolve_id(id), "expected a resolve stub id");
2437 
2438   // allocate space for the code
2439   ResourceMark rm;
2440 
2441   const char* name = SharedRuntime::stub_name(id);
2442   CodeBuffer buffer(name, 1000, 512);
2443   MacroAssembler* masm = new MacroAssembler(&buffer);
2444   assert_cond(masm != nullptr);
2445 
2446   int frame_size_in_words = -1;
2447   RegisterSaver reg_saver(false /* save_vectors */);
2448 
2449   OopMapSet *oop_maps = new OopMapSet();
2450   assert_cond(oop_maps != nullptr);
2451   OopMap* map = nullptr;
2452 
2453   int start = __ offset();
2454 
2455   map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2456 
2457   int frame_complete = __ offset();
2458 
2459   {
2460     Label retaddr;
2461     __ set_last_Java_frame(sp, noreg, retaddr, t0);
2462 
2463     __ mv(c_rarg0, xthread);
2464     __ rt_call(destination);
2465     __ bind(retaddr);
2466   }
2467 
2468   // Set an oopmap for the call site.
2469   // We need this not only for callee-saved registers, but also for volatile
2470   // registers that the compiler might be keeping live across a safepoint.
2471 
2472   oop_maps->add_gc_map( __ offset() - start, map);
2473 
2474   // x10 contains the address we are going to jump to assuming no exception got installed
2475 
2476   // clear last_Java_sp
2477   __ reset_last_Java_frame(false);
2478   // check for pending exceptions
2479   Label pending;
2480   __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2481   __ bnez(t1, pending);
2482 
2483   // get the returned Method*
2484   __ get_vm_result_metadata(xmethod, xthread);
2485   __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2486 
2487   // x10 is where we want to jump, overwrite t1 which is saved and temporary
2488   __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2489   reg_saver.restore_live_registers(masm);
2490 
2491   // We are back to the original state on entry and ready to go.
2492   __ jr(t1);
2493 
2494   // Pending exception after the safepoint
2495 
2496   __ bind(pending);
2497 
2498   reg_saver.restore_live_registers(masm);
2499 
2500   // exception pending => remove activation and forward to exception handler
2501 
2502   __ sd(zr, Address(xthread, JavaThread::vm_result_oop_offset()));
2503 
2504   __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2505   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2506 
2507   // -------------
2508   // make sure all code is generated
2509   masm->flush();
2510 
2511   // return the  blob
2512   return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2513 }
2514 
2515 // Continuation point for throwing of implicit exceptions that are
2516 // not handled in the current activation. Fabricates an exception
2517 // oop and initiates normal exception dispatching in this
2518 // frame. Since we need to preserve callee-saved values (currently
2519 // only for C2, but done for C1 as well) we need a callee-saved oop
2520 // map and therefore have to make these stubs into RuntimeStubs
2521 // rather than BufferBlobs.  If the compiler needs all registers to
2522 // be preserved between the fault point and the exception handler
2523 // then it must assume responsibility for that in
2524 // AbstractCompiler::continuation_for_implicit_null_exception or
2525 // continuation_for_implicit_division_by_zero_exception. All other
2526 // implicit exceptions (e.g., NullPointerException or
2527 // AbstractMethodError on entry) are either at call sites or
2528 // otherwise assume that stack unwinding will be initiated, so
2529 // caller saved registers were assumed volatile in the compiler.
2530 
2531 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2532   assert(is_throw_id(id), "expected a throw stub id");
2533 
2534   const char* name = SharedRuntime::stub_name(id);
2535 
2536   // Information about frame layout at time of blocking runtime call.
2537   // Note that we only have to preserve callee-saved registers since
2538   // the compilers are responsible for supplying a continuation point
2539   // if they expect all registers to be preserved.
2540   // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2541   assert_cond(runtime_entry != nullptr);
2542   enum layout {
2543     fp_off = 0,
2544     fp_off2,
2545     return_off,
2546     return_off2,
2547     framesize // inclusive of return address
2548   };
2549 
2550   const int insts_size = 1024;
2551   const int locs_size  = 64;
2552 
2553   ResourceMark rm;
2554   const char* timer_msg = "SharedRuntime generate_throw_exception";
2555   TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2556 
2557   CodeBuffer code(name, insts_size, locs_size);
2558   OopMapSet* oop_maps  = new OopMapSet();
2559   MacroAssembler* masm = new MacroAssembler(&code);
2560   assert_cond(oop_maps != nullptr && masm != nullptr);
2561 
2562   address start = __ pc();
2563 
2564   // This is an inlined and slightly modified version of call_VM
2565   // which has the ability to fetch the return PC out of
2566   // thread-local storage and also sets up last_Java_sp slightly
2567   // differently than the real call_VM
2568 
2569   __ enter(); // Save FP and RA before call
2570 
2571   assert(is_even(framesize / 2), "sp not 16-byte aligned");
2572 
2573   // ra and fp are already in place
2574   __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2575 
2576   int frame_complete = __ pc() - start;
2577 
2578   // Set up last_Java_sp and last_Java_fp
2579   address the_pc = __ pc();
2580   __ set_last_Java_frame(sp, fp, the_pc, t0);
2581 
2582   // Call runtime
2583   __ mv(c_rarg0, xthread);
2584   BLOCK_COMMENT("call runtime_entry");
2585   __ rt_call(runtime_entry);
2586 
2587   // Generate oop map
2588   OopMap* map = new OopMap(framesize, 0);
2589   assert_cond(map != nullptr);
2590 
2591   oop_maps->add_gc_map(the_pc - start, map);
2592 
2593   __ reset_last_Java_frame(true);
2594 
2595   __ leave();
2596 
2597   // check for pending exceptions
2598 #ifdef ASSERT
2599   Label L;
2600   __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2601   __ bnez(t0, L);
2602   __ should_not_reach_here();
2603   __ bind(L);
2604 #endif // ASSERT
2605   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2606 
2607   // codeBlob framesize is in words (not VMRegImpl::slot_size)
2608   RuntimeStub* stub =
2609     RuntimeStub::new_runtime_stub(name,
2610                                   &code,
2611                                   frame_complete,
2612                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2613                                   oop_maps, false);
2614   assert(stub != nullptr, "create runtime stub fail!");
2615   return stub;
2616 }
2617 
2618 #if INCLUDE_JFR
2619 
2620 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2621   __ set_last_Java_frame(sp, fp, the_pc, t0);
2622   __ mv(c_rarg0, thread);
2623 }
2624 
2625 static void jfr_epilogue(MacroAssembler* masm) {
2626   __ reset_last_Java_frame(true);
2627 }
2628 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2629 // It returns a jobject handle to the event writer.
2630 // The handle is dereferenced and the return value is the event writer oop.
2631 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2632   enum layout {
2633     fp_off,
2634     fp_off2,
2635     return_off,
2636     return_off2,
2637     framesize // inclusive of return address
2638   };
2639 
2640   int insts_size = 1024;
2641   int locs_size = 64;
2642   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
2643   CodeBuffer code(name, insts_size, locs_size);
2644   OopMapSet* oop_maps = new OopMapSet();
2645   MacroAssembler* masm = new MacroAssembler(&code);
2646 
2647   address start = __ pc();
2648   __ enter();
2649   int frame_complete = __ pc() - start;
2650   address the_pc = __ pc();
2651   jfr_prologue(the_pc, masm, xthread);
2652   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2653 
2654   jfr_epilogue(masm);
2655   __ resolve_global_jobject(x10, t0, t1);
2656   __ leave();
2657   __ ret();
2658 
2659   OopMap* map = new OopMap(framesize, 1);
2660   oop_maps->add_gc_map(the_pc - start, map);
2661 
2662   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2663     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2664                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2665                                   oop_maps, false);
2666   return stub;
2667 }
2668 
2669 // For c2: call to return a leased buffer.
2670 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2671   enum layout {
2672     fp_off,
2673     fp_off2,
2674     return_off,
2675     return_off2,
2676     framesize // inclusive of return address
2677   };
2678 
2679   int insts_size = 1024;
2680   int locs_size = 64;
2681   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
2682   CodeBuffer code(name, insts_size, locs_size);
2683   OopMapSet* oop_maps = new OopMapSet();
2684   MacroAssembler* masm = new MacroAssembler(&code);
2685 
2686   address start = __ pc();
2687   __ enter();
2688   int frame_complete = __ pc() - start;
2689   address the_pc = __ pc();
2690   jfr_prologue(the_pc, masm, xthread);
2691   __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2692 
2693   jfr_epilogue(masm);
2694   __ leave();
2695   __ ret();
2696 
2697   OopMap* map = new OopMap(framesize, 1);
2698   oop_maps->add_gc_map(the_pc - start, map);
2699 
2700   RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2701     RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2702                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2703                                   oop_maps, false);
2704   return stub;
2705 }
2706 
2707 #endif // INCLUDE_JFR