1 /*
   2  * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "interpreter/interpreter.hpp"
  27 #include "interpreter/interpreterRuntime.hpp"
  28 #include "memory/allocation.inline.hpp"
  29 #include "prims/methodHandles.hpp"
  30 
  31 #define __ _masm->
  32 
  33 #ifdef PRODUCT
  34 #define BLOCK_COMMENT(str) /* nothing */
  35 #else
  36 #define BLOCK_COMMENT(str) __ block_comment(str)
  37 #endif
  38 
  39 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
  40 
  41 // Workaround for C++ overloading nastiness on '0' for RegisterOrConstant.
  42 static RegisterOrConstant constant(int value) {
  43   return RegisterOrConstant(value);
  44 }
  45 
  46 address MethodHandleEntry::start_compiled_entry(MacroAssembler* _masm,
  47                                                 address interpreted_entry) {
  48   // Just before the actual machine code entry point, allocate space
  49   // for a MethodHandleEntry::Data record, so that we can manage everything
  50   // from one base pointer.
  51   __ align(wordSize);
  52   address target = __ pc() + sizeof(Data);
  53   while (__ pc() < target) {
  54     __ nop();
  55     __ align(wordSize);
  56   }
  57 
  58   MethodHandleEntry* me = (MethodHandleEntry*) __ pc();
  59   me->set_end_address(__ pc());         // set a temporary end_address
  60   me->set_from_interpreted_entry(interpreted_entry);
  61   me->set_type_checking_entry(NULL);
  62 
  63   return (address) me;
  64 }
  65 
  66 MethodHandleEntry* MethodHandleEntry::finish_compiled_entry(MacroAssembler* _masm,
  67                                                 address start_addr) {
  68   MethodHandleEntry* me = (MethodHandleEntry*) start_addr;
  69   assert(me->end_address() == start_addr, "valid ME");
  70 
  71   // Fill in the real end_address:
  72   __ align(wordSize);
  73   me->set_end_address(__ pc());
  74 
  75   return me;
  76 }
  77 
  78 // stack walking support
  79 
  80 frame MethodHandles::ricochet_frame_sender(const frame& fr, RegisterMap *map) {
  81   RicochetFrame* f = RicochetFrame::from_frame(fr);
  82   if (map->update_map())
  83     frame::update_map_with_saved_link(map, &f->_sender_link);
  84   return frame(f->extended_sender_sp(), f->exact_sender_sp(), f->sender_link(), f->sender_pc());
  85 }
  86 
  87 void MethodHandles::ricochet_frame_oops_do(const frame& fr, OopClosure* blk, const RegisterMap* reg_map) {
  88   RicochetFrame* f = RicochetFrame::from_frame(fr);
  89 
  90   // pick up the argument type descriptor:
  91   Thread* thread = Thread::current();
  92   Handle cookie(thread, f->compute_saved_args_layout(true, true));
  93 
  94   // process fixed part
  95   blk->do_oop((oop*)f->saved_target_addr());
  96   blk->do_oop((oop*)f->saved_args_layout_addr());
  97 
  98   // process variable arguments:
  99   if (cookie.is_null())  return;  // no arguments to describe
 100 
 101   // the cookie is actually the invokeExact method for my target
 102   // his argument signature is what I'm interested in
 103   assert(cookie->is_method(), "");
 104   methodHandle invoker(thread, methodOop(cookie()));
 105   assert(invoker->name() == vmSymbols::invokeExact_name(), "must be this kind of method");
 106   assert(!invoker->is_static(), "must have MH argument");
 107   int slot_count = invoker->size_of_parameters();
 108   assert(slot_count >= 1, "must include 'this'");
 109   intptr_t* base = f->saved_args_base();
 110   intptr_t* retval = NULL;
 111   if (f->has_return_value_slot())
 112     retval = f->return_value_slot_addr();
 113   int slot_num = slot_count;
 114   intptr_t* loc = &base[slot_num -= 1];
 115   //blk->do_oop((oop*) loc);   // original target, which is irrelevant
 116   int arg_num = 0;
 117   for (SignatureStream ss(invoker->signature()); !ss.is_done(); ss.next()) {
 118     if (ss.at_return_type())  continue;
 119     BasicType ptype = ss.type();
 120     if (ptype == T_ARRAY)  ptype = T_OBJECT; // fold all refs to T_OBJECT
 121     assert(ptype >= T_BOOLEAN && ptype <= T_OBJECT, "not array or void");
 122     loc = &base[slot_num -= type2size[ptype]];
 123     bool is_oop = (ptype == T_OBJECT && loc != retval);
 124     if (is_oop)  blk->do_oop((oop*)loc);
 125     arg_num += 1;
 126   }
 127   assert(slot_num == 0, "must have processed all the arguments");
 128 }
 129 
 130 oop MethodHandles::RicochetFrame::compute_saved_args_layout(bool read_cache, bool write_cache) {
 131   oop cookie = NULL;
 132   if (read_cache) {
 133     cookie = saved_args_layout();
 134     if (cookie != NULL)  return cookie;
 135   }
 136   oop target = saved_target();
 137   oop mtype  = java_lang_invoke_MethodHandle::type(target);
 138   oop mtform = java_lang_invoke_MethodType::form(mtype);
 139   cookie = java_lang_invoke_MethodTypeForm::vmlayout(mtform);
 140   if (write_cache)  {
 141     (*saved_args_layout_addr()) = cookie;
 142   }
 143   return cookie;
 144 }
 145 
 146 void MethodHandles::RicochetFrame::generate_ricochet_blob(MacroAssembler* _masm,
 147                                                           // output params:
 148                                                           int* bounce_offset,
 149                                                           int* exception_offset,
 150                                                           int* frame_size_in_words) {
 151   (*frame_size_in_words) = RicochetFrame::frame_size_in_bytes() / wordSize;
 152 
 153   address start = __ pc();
 154 
 155 #ifdef ASSERT
 156   __ hlt(); __ hlt(); __ hlt();
 157   // here's a hint of something special:
 158   __ push(MAGIC_NUMBER_1);
 159   __ push(MAGIC_NUMBER_2);
 160 #endif //ASSERT
 161   __ hlt();  // not reached
 162 
 163   // A return PC has just been popped from the stack.
 164   // Return values are in registers.
 165   // The ebp points into the RicochetFrame, which contains
 166   // a cleanup continuation we must return to.
 167 
 168   (*bounce_offset) = __ pc() - start;
 169   BLOCK_COMMENT("ricochet_blob.bounce");
 170 
 171   if (VerifyMethodHandles)  RicochetFrame::verify_clean(_masm);
 172   trace_method_handle(_masm, "return/ricochet_blob.bounce");
 173 
 174   __ jmp(frame_address(continuation_offset_in_bytes()));
 175   __ hlt();
 176   DEBUG_ONLY(__ push(MAGIC_NUMBER_2));
 177 
 178   (*exception_offset) = __ pc() - start;
 179   BLOCK_COMMENT("ricochet_blob.exception");
 180 
 181   // compare this to Interpreter::rethrow_exception_entry, which is parallel code
 182   // for example, see TemplateInterpreterGenerator::generate_throw_exception
 183   // Live registers in:
 184   //   rax: exception
 185   //   rdx: return address/pc that threw exception (ignored, always equal to bounce addr)
 186   __ verify_oop(rax);
 187 
 188   // no need to empty_FPU_stack or reinit_heapbase, since caller frame will do the same if needed
 189 
 190   // Take down the frame.
 191 
 192   // Cf. InterpreterMacroAssembler::remove_activation.
 193   leave_ricochet_frame(_masm, /*rcx_recv=*/ noreg,
 194                        saved_last_sp_register(),
 195                        /*sender_pc_reg=*/ rdx);
 196 
 197   // In between activations - previous activation type unknown yet
 198   // compute continuation point - the continuation point expects the
 199   // following registers set up:
 200   //
 201   // rax: exception
 202   // rdx: return address/pc that threw exception
 203   // rsp: expression stack of caller
 204   // rbp: ebp of caller
 205   __ push(rax);                                  // save exception
 206   __ push(rdx);                                  // save return address
 207   Register thread_reg = LP64_ONLY(r15_thread) NOT_LP64(rdi);
 208   NOT_LP64(__ get_thread(thread_reg));
 209   __ call_VM_leaf(CAST_FROM_FN_PTR(address,
 210                                    SharedRuntime::exception_handler_for_return_address),
 211                   thread_reg, rdx);
 212   __ mov(rbx, rax);                              // save exception handler
 213   __ pop(rdx);                                   // restore return address
 214   __ pop(rax);                                   // restore exception
 215   __ jmp(rbx);                                   // jump to exception
 216                                                  // handler of caller
 217 }
 218 
 219 void MethodHandles::RicochetFrame::enter_ricochet_frame(MacroAssembler* _masm,
 220                                                         Register rcx_recv,
 221                                                         Register rax_argv,
 222                                                         address return_handler,
 223                                                         Register rbx_temp) {
 224   const Register saved_last_sp = saved_last_sp_register();
 225   Address rcx_mh_vmtarget(    rcx_recv, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes() );
 226   Address rcx_amh_conversion( rcx_recv, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes() );
 227 
 228   // Push the RicochetFrame a word at a time.
 229   // This creates something similar to an interpreter frame.
 230   // Cf. TemplateInterpreterGenerator::generate_fixed_frame.
 231   BLOCK_COMMENT("push RicochetFrame {");
 232   DEBUG_ONLY(int rfo = (int) sizeof(RicochetFrame));
 233   assert((rfo -= wordSize) == RicochetFrame::sender_pc_offset_in_bytes(), "");
 234 #define RF_FIELD(push_value, name)                                      \
 235   { push_value;                                                         \
 236     assert((rfo -= wordSize) == RicochetFrame::name##_offset_in_bytes(), ""); }
 237   RF_FIELD(__ push(rbp),                   sender_link);
 238   RF_FIELD(__ push(saved_last_sp),         exact_sender_sp);  // rsi/r13
 239   RF_FIELD(__ pushptr(rcx_amh_conversion), conversion);
 240   RF_FIELD(__ push(rax_argv),              saved_args_base);   // can be updated if args are shifted
 241   RF_FIELD(__ push((int32_t) NULL_WORD),   saved_args_layout); // cache for GC layout cookie
 242   if (UseCompressedOops) {
 243     __ load_heap_oop(rbx_temp, rcx_mh_vmtarget);
 244     RF_FIELD(__ push(rbx_temp),            saved_target);
 245   } else {
 246     RF_FIELD(__ pushptr(rcx_mh_vmtarget),  saved_target);
 247   }
 248   __ lea(rbx_temp, ExternalAddress(return_handler));
 249   RF_FIELD(__ push(rbx_temp),              continuation);
 250 #undef RF_FIELD
 251   assert(rfo == 0, "fully initialized the RicochetFrame");
 252   // compute new frame pointer:
 253   __ lea(rbp, Address(rsp, RicochetFrame::sender_link_offset_in_bytes()));
 254   // Push guard word #1 in debug mode.
 255   DEBUG_ONLY(__ push((int32_t) RicochetFrame::MAGIC_NUMBER_1));
 256   // For debugging, leave behind an indication of which stub built this frame.
 257   DEBUG_ONLY({ Label L; __ call(L, relocInfo::none); __ bind(L); });
 258   BLOCK_COMMENT("} RicochetFrame");
 259 }
 260 
 261 void MethodHandles::RicochetFrame::leave_ricochet_frame(MacroAssembler* _masm,
 262                                                         Register rcx_recv,
 263                                                         Register new_sp_reg,
 264                                                         Register sender_pc_reg) {
 265   assert_different_registers(rcx_recv, new_sp_reg, sender_pc_reg);
 266   const Register saved_last_sp = saved_last_sp_register();
 267   // Take down the frame.
 268   // Cf. InterpreterMacroAssembler::remove_activation.
 269   BLOCK_COMMENT("end_ricochet_frame {");
 270   // TO DO: If (exact_sender_sp - extended_sender_sp) > THRESH, compact the frame down.
 271   // This will keep stack in bounds even with unlimited tailcalls, each with an adapter.
 272   if (rcx_recv->is_valid())
 273     __ movptr(rcx_recv,    RicochetFrame::frame_address(RicochetFrame::saved_target_offset_in_bytes()));
 274   __ movptr(sender_pc_reg, RicochetFrame::frame_address(RicochetFrame::sender_pc_offset_in_bytes()));
 275   __ movptr(saved_last_sp, RicochetFrame::frame_address(RicochetFrame::exact_sender_sp_offset_in_bytes()));
 276   __ movptr(rbp,           RicochetFrame::frame_address(RicochetFrame::sender_link_offset_in_bytes()));
 277   __ mov(rsp, new_sp_reg);
 278   BLOCK_COMMENT("} end_ricochet_frame");
 279 }
 280 
 281 // Emit code to verify that RBP is pointing at a valid ricochet frame.
 282 #ifdef ASSERT
 283 enum {
 284   ARG_LIMIT = 255, SLOP = 4,
 285   // use this parameter for checking for garbage stack movements:
 286   UNREASONABLE_STACK_MOVE = (ARG_LIMIT + SLOP)
 287   // the slop defends against false alarms due to fencepost errors
 288 };
 289 
 290 void MethodHandles::RicochetFrame::verify_clean(MacroAssembler* _masm) {
 291   // The stack should look like this:
 292   //    ... keep1 | dest=42 | keep2 | RF | magic | handler | magic | recursive args |
 293   // Check various invariants.
 294   verify_offsets();
 295 
 296   Register rdi_temp = rdi;
 297   Register rcx_temp = rcx;
 298   { __ push(rdi_temp); __ push(rcx_temp); }
 299 #define UNPUSH_TEMPS \
 300   { __ pop(rcx_temp);  __ pop(rdi_temp); }
 301 
 302   Address magic_number_1_addr  = RicochetFrame::frame_address(RicochetFrame::magic_number_1_offset_in_bytes());
 303   Address magic_number_2_addr  = RicochetFrame::frame_address(RicochetFrame::magic_number_2_offset_in_bytes());
 304   Address continuation_addr    = RicochetFrame::frame_address(RicochetFrame::continuation_offset_in_bytes());
 305   Address conversion_addr      = RicochetFrame::frame_address(RicochetFrame::conversion_offset_in_bytes());
 306   Address saved_args_base_addr = RicochetFrame::frame_address(RicochetFrame::saved_args_base_offset_in_bytes());
 307 
 308   Label L_bad, L_ok;
 309   BLOCK_COMMENT("verify_clean {");
 310   // Magic numbers must check out:
 311   __ cmpptr(magic_number_1_addr, (int32_t) MAGIC_NUMBER_1);
 312   __ jcc(Assembler::notEqual, L_bad);
 313   __ cmpptr(magic_number_2_addr, (int32_t) MAGIC_NUMBER_2);
 314   __ jcc(Assembler::notEqual, L_bad);
 315 
 316   // Arguments pointer must look reasonable:
 317   __ movptr(rcx_temp, saved_args_base_addr);
 318   __ cmpptr(rcx_temp, rbp);
 319   __ jcc(Assembler::below, L_bad);
 320   __ subptr(rcx_temp, UNREASONABLE_STACK_MOVE * Interpreter::stackElementSize);
 321   __ cmpptr(rcx_temp, rbp);
 322   __ jcc(Assembler::above, L_bad);
 323 
 324   load_conversion_dest_type(_masm, rdi_temp, conversion_addr);
 325   __ cmpl(rdi_temp, T_VOID);
 326   __ jcc(Assembler::equal, L_ok);
 327   __ movptr(rcx_temp, saved_args_base_addr);
 328   load_conversion_vminfo(_masm, rdi_temp, conversion_addr);
 329   __ cmpptr(Address(rcx_temp, rdi_temp, Interpreter::stackElementScale()),
 330             (int32_t) RETURN_VALUE_PLACEHOLDER);
 331   __ jcc(Assembler::equal, L_ok);
 332   __ BIND(L_bad);
 333   UNPUSH_TEMPS;
 334   __ stop("damaged ricochet frame");
 335   __ BIND(L_ok);
 336   UNPUSH_TEMPS;
 337   BLOCK_COMMENT("} verify_clean");
 338 
 339 #undef UNPUSH_TEMPS
 340 
 341 }
 342 #endif //ASSERT
 343 
 344 void MethodHandles::load_klass_from_Class(MacroAssembler* _masm, Register klass_reg) {
 345   if (VerifyMethodHandles)
 346     verify_klass(_masm, klass_reg, SystemDictionaryHandles::Class_klass(),
 347                  "AMH argument is a Class");
 348   __ load_heap_oop(klass_reg, Address(klass_reg, java_lang_Class::klass_offset_in_bytes()));
 349 }
 350 
 351 void MethodHandles::load_conversion_vminfo(MacroAssembler* _masm, Register reg, Address conversion_field_addr) {
 352   int bits   = BitsPerByte;
 353   int offset = (CONV_VMINFO_SHIFT / bits);
 354   int shift  = (CONV_VMINFO_SHIFT % bits);
 355   __ load_unsigned_byte(reg, conversion_field_addr.plus_disp(offset));
 356   assert(CONV_VMINFO_MASK == right_n_bits(bits - shift), "else change type of previous load");
 357   assert(shift == 0, "no shift needed");
 358 }
 359 
 360 void MethodHandles::load_conversion_dest_type(MacroAssembler* _masm, Register reg, Address conversion_field_addr) {
 361   int bits   = BitsPerByte;
 362   int offset = (CONV_DEST_TYPE_SHIFT / bits);
 363   int shift  = (CONV_DEST_TYPE_SHIFT % bits);
 364   __ load_unsigned_byte(reg, conversion_field_addr.plus_disp(offset));
 365   assert(CONV_TYPE_MASK == right_n_bits(bits - shift), "else change type of previous load");
 366   __ shrl(reg, shift);
 367   DEBUG_ONLY(int conv_type_bits = (int) exact_log2(CONV_TYPE_MASK+1));
 368   assert((shift + conv_type_bits) == bits, "left justified in byte");
 369 }
 370 
 371 void MethodHandles::load_stack_move(MacroAssembler* _masm,
 372                                     Register rdi_stack_move,
 373                                     Register rcx_amh,
 374                                     bool might_be_negative) {
 375   BLOCK_COMMENT("load_stack_move {");
 376   Address rcx_amh_conversion(rcx_amh, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
 377   __ movl(rdi_stack_move, rcx_amh_conversion);
 378   __ sarl(rdi_stack_move, CONV_STACK_MOVE_SHIFT);
 379 #ifdef _LP64
 380   if (might_be_negative) {
 381     // clean high bits of stack motion register (was loaded as an int)
 382     __ movslq(rdi_stack_move, rdi_stack_move);
 383   }
 384 #endif //_LP64
 385 #ifdef ASSERT
 386   if (VerifyMethodHandles) {
 387     Label L_ok, L_bad;
 388     int32_t stack_move_limit = 0x4000;  // extra-large
 389     __ cmpptr(rdi_stack_move, stack_move_limit);
 390     __ jcc(Assembler::greaterEqual, L_bad);
 391     __ cmpptr(rdi_stack_move, -stack_move_limit);
 392     __ jcc(Assembler::greater, L_ok);
 393     __ bind(L_bad);
 394     __ stop("load_stack_move of garbage value");
 395     __ BIND(L_ok);
 396   }
 397 #endif
 398   BLOCK_COMMENT("} load_stack_move");
 399 }
 400 
 401 #ifdef ASSERT
 402 void MethodHandles::RicochetFrame::verify_offsets() {
 403   // Check compatibility of this struct with the more generally used offsets of class frame:
 404   int ebp_off = sender_link_offset_in_bytes();  // offset from struct base to local rbp value
 405   assert(ebp_off + wordSize*frame::interpreter_frame_method_offset      == saved_args_base_offset_in_bytes(), "");
 406   assert(ebp_off + wordSize*frame::interpreter_frame_last_sp_offset     == conversion_offset_in_bytes(), "");
 407   assert(ebp_off + wordSize*frame::interpreter_frame_sender_sp_offset   == exact_sender_sp_offset_in_bytes(), "");
 408   // These last two have to be exact:
 409   assert(ebp_off + wordSize*frame::link_offset                          == sender_link_offset_in_bytes(), "");
 410   assert(ebp_off + wordSize*frame::return_addr_offset                   == sender_pc_offset_in_bytes(), "");
 411 }
 412 
 413 void MethodHandles::RicochetFrame::verify() const {
 414   verify_offsets();
 415   assert(magic_number_1() == MAGIC_NUMBER_1, err_msg(PTR_FORMAT " == " PTR_FORMAT, magic_number_1(), MAGIC_NUMBER_1));
 416   assert(magic_number_2() == MAGIC_NUMBER_2, err_msg(PTR_FORMAT " == " PTR_FORMAT, magic_number_2(), MAGIC_NUMBER_2));
 417   if (!Universe::heap()->is_gc_active()) {
 418     if (saved_args_layout() != NULL) {
 419       assert(saved_args_layout()->is_method(), "must be valid oop");
 420     }
 421     if (saved_target() != NULL) {
 422       assert(java_lang_invoke_MethodHandle::is_instance(saved_target()), "checking frame value");
 423     }
 424   }
 425   int conv_op = adapter_conversion_op(conversion());
 426   assert(conv_op == java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS ||
 427          conv_op == java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS ||
 428          conv_op == java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF,
 429          "must be a sane conversion");
 430   if (has_return_value_slot()) {
 431     assert(*return_value_slot_addr() == RETURN_VALUE_PLACEHOLDER, "");
 432   }
 433 }
 434 #endif //PRODUCT
 435 
 436 #ifdef ASSERT
 437 void MethodHandles::verify_argslot(MacroAssembler* _masm,
 438                                    Register argslot_reg,
 439                                    const char* error_message) {
 440   // Verify that argslot lies within (rsp, rbp].
 441   Label L_ok, L_bad;
 442   BLOCK_COMMENT("verify_argslot {");
 443   __ cmpptr(argslot_reg, rbp);
 444   __ jccb(Assembler::above, L_bad);
 445   __ cmpptr(rsp, argslot_reg);
 446   __ jccb(Assembler::below, L_ok);
 447   __ bind(L_bad);
 448   __ stop(error_message);
 449   __ BIND(L_ok);
 450   BLOCK_COMMENT("} verify_argslot");
 451 }
 452 
 453 void MethodHandles::verify_argslots(MacroAssembler* _masm,
 454                                     RegisterOrConstant arg_slots,
 455                                     Register arg_slot_base_reg,
 456                                     bool negate_argslots,
 457                                     const char* error_message) {
 458   // Verify that [argslot..argslot+size) lies within (rsp, rbp).
 459   Label L_ok, L_bad;
 460   Register rdi_temp = rdi;
 461   BLOCK_COMMENT("verify_argslots {");
 462   __ push(rdi_temp);
 463   if (negate_argslots) {
 464     if (arg_slots.is_constant()) {
 465       arg_slots = -1 * arg_slots.as_constant();
 466     } else {
 467       __ movptr(rdi_temp, arg_slots);
 468       __ negptr(rdi_temp);
 469       arg_slots = rdi_temp;
 470     }
 471   }
 472   __ lea(rdi_temp, Address(arg_slot_base_reg, arg_slots, Interpreter::stackElementScale()));
 473   __ cmpptr(rdi_temp, rbp);
 474   __ pop(rdi_temp);
 475   __ jcc(Assembler::above, L_bad);
 476   __ cmpptr(rsp, arg_slot_base_reg);
 477   __ jcc(Assembler::below, L_ok);
 478   __ bind(L_bad);
 479   __ stop(error_message);
 480   __ BIND(L_ok);
 481   BLOCK_COMMENT("} verify_argslots");
 482 }
 483 
 484 // Make sure that arg_slots has the same sign as the given direction.
 485 // If (and only if) arg_slots is a assembly-time constant, also allow it to be zero.
 486 void MethodHandles::verify_stack_move(MacroAssembler* _masm,
 487                                       RegisterOrConstant arg_slots, int direction) {
 488   bool allow_zero = arg_slots.is_constant();
 489   if (direction == 0) { direction = +1; allow_zero = true; }
 490   assert(stack_move_unit() == -1, "else add extra checks here");
 491   if (arg_slots.is_register()) {
 492     Label L_ok, L_bad;
 493     BLOCK_COMMENT("verify_stack_move {");
 494     // testl(arg_slots.as_register(), -stack_move_unit() - 1);  // no need
 495     // jcc(Assembler::notZero, L_bad);
 496     __ cmpptr(arg_slots.as_register(), (int32_t) NULL_WORD);
 497     if (direction > 0) {
 498       __ jcc(allow_zero ? Assembler::less : Assembler::lessEqual, L_bad);
 499       __ cmpptr(arg_slots.as_register(), (int32_t) UNREASONABLE_STACK_MOVE);
 500       __ jcc(Assembler::less, L_ok);
 501     } else {
 502       __ jcc(allow_zero ? Assembler::greater : Assembler::greaterEqual, L_bad);
 503       __ cmpptr(arg_slots.as_register(), (int32_t) -UNREASONABLE_STACK_MOVE);
 504       __ jcc(Assembler::greater, L_ok);
 505     }
 506     __ bind(L_bad);
 507     if (direction > 0)
 508       __ stop("assert arg_slots > 0");
 509     else
 510       __ stop("assert arg_slots < 0");
 511     __ BIND(L_ok);
 512     BLOCK_COMMENT("} verify_stack_move");
 513   } else {
 514     intptr_t size = arg_slots.as_constant();
 515     if (direction < 0)  size = -size;
 516     assert(size >= 0, "correct direction of constant move");
 517     assert(size < UNREASONABLE_STACK_MOVE, "reasonable size of constant move");
 518   }
 519 }
 520 
 521 void MethodHandles::verify_klass(MacroAssembler* _masm,
 522                                  Register obj, KlassHandle klass,
 523                                  const char* error_message) {
 524   oop* klass_addr = klass.raw_value();
 525   assert(klass_addr >= SystemDictionaryHandles::Object_klass().raw_value() &&
 526          klass_addr <= SystemDictionaryHandles::Long_klass().raw_value(),
 527          "must be one of the SystemDictionaryHandles");
 528   Register temp = rdi;
 529   Label L_ok, L_bad;
 530   BLOCK_COMMENT("verify_klass {");
 531   __ verify_oop(obj);
 532   __ testptr(obj, obj);
 533   __ jcc(Assembler::zero, L_bad);
 534   __ push(temp);
 535   __ load_klass(temp, obj);
 536   __ cmpptr(temp, ExternalAddress((address) klass_addr));
 537   __ jcc(Assembler::equal, L_ok);
 538   intptr_t super_check_offset = klass->super_check_offset();
 539   __ movptr(temp, Address(temp, super_check_offset));
 540   __ cmpptr(temp, ExternalAddress((address) klass_addr));
 541   __ jcc(Assembler::equal, L_ok);
 542   __ pop(temp);
 543   __ bind(L_bad);
 544   __ stop(error_message);
 545   __ BIND(L_ok);
 546   __ pop(temp);
 547   BLOCK_COMMENT("} verify_klass");
 548 }
 549 #endif //ASSERT
 550 
 551 void MethodHandles::jump_from_method_handle(MacroAssembler* _masm, Register method, Register temp) {
 552   if (JvmtiExport::can_post_interpreter_events()) {
 553     Label run_compiled_code;
 554     // JVMTI events, such as single-stepping, are implemented partly by avoiding running
 555     // compiled code in threads for which the event is enabled.  Check here for
 556     // interp_only_mode if these events CAN be enabled.
 557 #ifdef _LP64
 558     Register rthread = r15_thread;
 559 #else
 560     Register rthread = temp;
 561     __ get_thread(rthread);
 562 #endif
 563     // interp_only is an int, on little endian it is sufficient to test the byte only
 564     // Is a cmpl faster?
 565     __ cmpb(Address(rthread, JavaThread::interp_only_mode_offset()), 0);
 566     __ jccb(Assembler::zero, run_compiled_code);
 567     __ jmp(Address(method, methodOopDesc::interpreter_entry_offset()));
 568     __ bind(run_compiled_code);
 569   }
 570   __ jmp(Address(method, methodOopDesc::from_interpreted_offset()));
 571 }
 572 
 573 // Code generation
 574 address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm) {
 575   // rbx: methodOop
 576   // rcx: receiver method handle (must load from sp[MethodTypeForm.vmslots])
 577   // rsi/r13: sender SP (must preserve; see prepare_to_jump_from_interpreted)
 578   // rdx, rdi: garbage temp, blown away
 579 
 580   Register rbx_method = rbx;
 581   Register rcx_recv   = rcx;
 582   Register rax_mtype  = rax;
 583   Register rdx_temp   = rdx;
 584   Register rdi_temp   = rdi;
 585 
 586   // emit WrongMethodType path first, to enable jccb back-branch from main path
 587   Label wrong_method_type;
 588   __ bind(wrong_method_type);
 589   Label invoke_generic_slow_path, invoke_exact_error_path;
 590   assert(methodOopDesc::intrinsic_id_size_in_bytes() == sizeof(u1), "");;
 591   __ cmpb(Address(rbx_method, methodOopDesc::intrinsic_id_offset_in_bytes()), (int) vmIntrinsics::_invokeExact);
 592   __ jcc(Assembler::notEqual, invoke_generic_slow_path);
 593   __ jmp(invoke_exact_error_path);
 594 
 595   // here's where control starts out:
 596   __ align(CodeEntryAlignment);
 597   address entry_point = __ pc();
 598 
 599   // fetch the MethodType from the method handle into rax (the 'check' register)
 600   // FIXME: Interpreter should transmit pre-popped stack pointer, to locate base of arg list.
 601   // This would simplify several touchy bits of code.
 602   // See 6984712: JSR 292 method handle calls need a clean argument base pointer
 603   {
 604     Register tem = rbx_method;
 605     for (jint* pchase = methodOopDesc::method_type_offsets_chain(); (*pchase) != -1; pchase++) {
 606       __ movptr(rax_mtype, Address(tem, *pchase));
 607       tem = rax_mtype;          // in case there is another indirection
 608     }
 609   }
 610 
 611   // given the MethodType, find out where the MH argument is buried
 612   __ load_heap_oop(rdx_temp, Address(rax_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, rdi_temp)));
 613   Register rdx_vmslots = rdx_temp;
 614   __ movl(rdx_vmslots, Address(rdx_temp, __ delayed_value(java_lang_invoke_MethodTypeForm::vmslots_offset_in_bytes, rdi_temp)));
 615   Address mh_receiver_slot_addr = __ argument_address(rdx_vmslots);
 616   __ movptr(rcx_recv, mh_receiver_slot_addr);
 617 
 618   trace_method_handle(_masm, "invokeExact");
 619 
 620   __ check_method_handle_type(rax_mtype, rcx_recv, rdi_temp, wrong_method_type);
 621 
 622   // Nobody uses the MH receiver slot after this.  Make sure.
 623   DEBUG_ONLY(__ movptr(mh_receiver_slot_addr, (int32_t)0x999999));
 624 
 625   __ jump_to_method_handle_entry(rcx_recv, rdi_temp);
 626 
 627   // error path for invokeExact (only)
 628   __ bind(invoke_exact_error_path);
 629   // ensure that the top of stack is properly aligned.
 630   __ mov(rdi, rsp);
 631   __ andptr(rsp, -StackAlignmentInBytes); // Align the stack for the ABI
 632   __ pushptr(Address(rdi, 0));  // Pick up the return address
 633 
 634   // Stub wants expected type in rax and the actual type in rcx
 635   __ jump(ExternalAddress(StubRoutines::throw_WrongMethodTypeException_entry()));
 636 
 637   // for invokeGeneric (only), apply argument and result conversions on the fly
 638   __ bind(invoke_generic_slow_path);
 639 #ifdef ASSERT
 640   if (VerifyMethodHandles) {
 641     Label L;
 642     __ cmpb(Address(rbx_method, methodOopDesc::intrinsic_id_offset_in_bytes()), (int) vmIntrinsics::_invokeGeneric);
 643     __ jcc(Assembler::equal, L);
 644     __ stop("bad methodOop::intrinsic_id");
 645     __ bind(L);
 646   }
 647 #endif //ASSERT
 648   Register rbx_temp = rbx_method;  // don't need it now
 649 
 650   // make room on the stack for another pointer:
 651   Register rcx_argslot = rcx_recv;
 652   __ lea(rcx_argslot, __ argument_address(rdx_vmslots, 1));
 653   insert_arg_slots(_masm, 2 * stack_move_unit(),
 654                    rcx_argslot, rbx_temp, rdx_temp);
 655 
 656   // load up an adapter from the calling type (Java weaves this)
 657   Register rdx_adapter = rdx_temp;
 658   __ load_heap_oop(rdx_temp,    Address(rax_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes,               rdi_temp)));
 659   __ load_heap_oop(rdx_adapter, Address(rdx_temp,  __ delayed_value(java_lang_invoke_MethodTypeForm::genericInvoker_offset_in_bytes, rdi_temp)));
 660   __ verify_oop(rdx_adapter);
 661   __ movptr(Address(rcx_argslot, 1 * Interpreter::stackElementSize), rdx_adapter);
 662   // As a trusted first argument, pass the type being called, so the adapter knows
 663   // the actual types of the arguments and return values.
 664   // (Generic invokers are shared among form-families of method-type.)
 665   __ movptr(Address(rcx_argslot, 0 * Interpreter::stackElementSize), rax_mtype);
 666   // FIXME: assert that rdx_adapter is of the right method-type.
 667   __ mov(rcx, rdx_adapter);
 668   trace_method_handle(_masm, "invokeGeneric");
 669   __ jump_to_method_handle_entry(rcx, rdi_temp);
 670 
 671   return entry_point;
 672 }
 673 
 674 // Helper to insert argument slots into the stack.
 675 // arg_slots must be a multiple of stack_move_unit() and < 0
 676 // rax_argslot is decremented to point to the new (shifted) location of the argslot
 677 // But, rdx_temp ends up holding the original value of rax_argslot.
 678 void MethodHandles::insert_arg_slots(MacroAssembler* _masm,
 679                                      RegisterOrConstant arg_slots,
 680                                      Register rax_argslot,
 681                                      Register rbx_temp, Register rdx_temp) {
 682   // allow constant zero
 683   if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
 684     return;
 685   assert_different_registers(rax_argslot, rbx_temp, rdx_temp,
 686                              (!arg_slots.is_register() ? rsp : arg_slots.as_register()));
 687   if (VerifyMethodHandles)
 688     verify_argslot(_masm, rax_argslot, "insertion point must fall within current frame");
 689   if (VerifyMethodHandles)
 690     verify_stack_move(_masm, arg_slots, -1);
 691 
 692   // Make space on the stack for the inserted argument(s).
 693   // Then pull down everything shallower than rax_argslot.
 694   // The stacked return address gets pulled down with everything else.
 695   // That is, copy [rsp, argslot) downward by -size words.  In pseudo-code:
 696   //   rsp -= size;
 697   //   for (rdx = rsp + size; rdx < argslot; rdx++)
 698   //     rdx[-size] = rdx[0]
 699   //   argslot -= size;
 700   BLOCK_COMMENT("insert_arg_slots {");
 701   __ mov(rdx_temp, rsp);                        // source pointer for copy
 702   __ lea(rsp, Address(rsp, arg_slots, Interpreter::stackElementScale()));
 703   {
 704     Label loop;
 705     __ BIND(loop);
 706     // pull one word down each time through the loop
 707     __ movptr(rbx_temp, Address(rdx_temp, 0));
 708     __ movptr(Address(rdx_temp, arg_slots, Interpreter::stackElementScale()), rbx_temp);
 709     __ addptr(rdx_temp, wordSize);
 710     __ cmpptr(rdx_temp, rax_argslot);
 711     __ jcc(Assembler::below, loop);
 712   }
 713 
 714   // Now move the argslot down, to point to the opened-up space.
 715   __ lea(rax_argslot, Address(rax_argslot, arg_slots, Interpreter::stackElementScale()));
 716   BLOCK_COMMENT("} insert_arg_slots");
 717 }
 718 
 719 // Helper to remove argument slots from the stack.
 720 // arg_slots must be a multiple of stack_move_unit() and > 0
 721 void MethodHandles::remove_arg_slots(MacroAssembler* _masm,
 722                                      RegisterOrConstant arg_slots,
 723                                      Register rax_argslot,
 724                                      Register rbx_temp, Register rdx_temp) {
 725   // allow constant zero
 726   if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
 727     return;
 728   assert_different_registers(rax_argslot, rbx_temp, rdx_temp,
 729                              (!arg_slots.is_register() ? rsp : arg_slots.as_register()));
 730   if (VerifyMethodHandles)
 731     verify_argslots(_masm, arg_slots, rax_argslot, false,
 732                     "deleted argument(s) must fall within current frame");
 733   if (VerifyMethodHandles)
 734     verify_stack_move(_masm, arg_slots, +1);
 735 
 736   BLOCK_COMMENT("remove_arg_slots {");
 737   // Pull up everything shallower than rax_argslot.
 738   // Then remove the excess space on the stack.
 739   // The stacked return address gets pulled up with everything else.
 740   // That is, copy [rsp, argslot) upward by size words.  In pseudo-code:
 741   //   for (rdx = argslot-1; rdx >= rsp; --rdx)
 742   //     rdx[size] = rdx[0]
 743   //   argslot += size;
 744   //   rsp += size;
 745   __ lea(rdx_temp, Address(rax_argslot, -wordSize)); // source pointer for copy
 746   {
 747     Label loop;
 748     __ BIND(loop);
 749     // pull one word up each time through the loop
 750     __ movptr(rbx_temp, Address(rdx_temp, 0));
 751     __ movptr(Address(rdx_temp, arg_slots, Interpreter::stackElementScale()), rbx_temp);
 752     __ addptr(rdx_temp, -wordSize);
 753     __ cmpptr(rdx_temp, rsp);
 754     __ jcc(Assembler::aboveEqual, loop);
 755   }
 756 
 757   // Now move the argslot up, to point to the just-copied block.
 758   __ lea(rsp, Address(rsp, arg_slots, Interpreter::stackElementScale()));
 759   // And adjust the argslot address to point at the deletion point.
 760   __ lea(rax_argslot, Address(rax_argslot, arg_slots, Interpreter::stackElementScale()));
 761   BLOCK_COMMENT("} remove_arg_slots");
 762 }
 763 
 764 // Helper to copy argument slots to the top of the stack.
 765 // The sequence starts with rax_argslot and is counted by slot_count
 766 // slot_count must be a multiple of stack_move_unit() and >= 0
 767 // This function blows the temps but does not change rax_argslot.
 768 void MethodHandles::push_arg_slots(MacroAssembler* _masm,
 769                                    Register rax_argslot,
 770                                    RegisterOrConstant slot_count,
 771                                    int skip_words_count,
 772                                    Register rbx_temp, Register rdx_temp) {
 773   assert_different_registers(rax_argslot, rbx_temp, rdx_temp,
 774                              (!slot_count.is_register() ? rbp : slot_count.as_register()),
 775                              rsp);
 776   assert(Interpreter::stackElementSize == wordSize, "else change this code");
 777 
 778   if (VerifyMethodHandles)
 779     verify_stack_move(_masm, slot_count, 0);
 780 
 781   // allow constant zero
 782   if (slot_count.is_constant() && slot_count.as_constant() == 0)
 783     return;
 784 
 785   BLOCK_COMMENT("push_arg_slots {");
 786 
 787   Register rbx_top = rbx_temp;
 788 
 789   // There is at most 1 word to carry down with the TOS.
 790   switch (skip_words_count) {
 791   case 1: __ pop(rdx_temp); break;
 792   case 0:                   break;
 793   default: ShouldNotReachHere();
 794   }
 795 
 796   if (slot_count.is_constant()) {
 797     for (int i = slot_count.as_constant() - 1; i >= 0; i--) {
 798       __ pushptr(Address(rax_argslot, i * wordSize));
 799     }
 800   } else {
 801     Label L_plural, L_loop, L_break;
 802     // Emit code to dynamically check for the common cases, zero and one slot.
 803     __ cmpl(slot_count.as_register(), (int32_t) 1);
 804     __ jccb(Assembler::greater, L_plural);
 805     __ jccb(Assembler::less, L_break);
 806     __ pushptr(Address(rax_argslot, 0));
 807     __ jmpb(L_break);
 808     __ BIND(L_plural);
 809 
 810     // Loop for 2 or more:
 811     //   rbx = &rax[slot_count]
 812     //   while (rbx > rax)  *(--rsp) = *(--rbx)
 813     __ lea(rbx_top, Address(rax_argslot, slot_count, Address::times_ptr));
 814     __ BIND(L_loop);
 815     __ subptr(rbx_top, wordSize);
 816     __ pushptr(Address(rbx_top, 0));
 817     __ cmpptr(rbx_top, rax_argslot);
 818     __ jcc(Assembler::above, L_loop);
 819     __ bind(L_break);
 820   }
 821   switch (skip_words_count) {
 822   case 1: __ push(rdx_temp); break;
 823   case 0:                    break;
 824   default: ShouldNotReachHere();
 825   }
 826   BLOCK_COMMENT("} push_arg_slots");
 827 }
 828 
 829 // in-place movement; no change to rsp
 830 // blows rax_temp, rdx_temp
 831 void MethodHandles::move_arg_slots_up(MacroAssembler* _masm,
 832                                       Register rbx_bottom,  // invariant
 833                                       Address  top_addr,     // can use rax_temp
 834                                       RegisterOrConstant positive_distance_in_slots,
 835                                       Register rax_temp, Register rdx_temp) {
 836   BLOCK_COMMENT("move_arg_slots_up {");
 837   assert_different_registers(rbx_bottom,
 838                              rax_temp, rdx_temp,
 839                              positive_distance_in_slots.register_or_noreg());
 840   Label L_loop, L_break;
 841   Register rax_top = rax_temp;
 842   if (!top_addr.is_same_address(Address(rax_top, 0)))
 843     __ lea(rax_top, top_addr);
 844   // Detect empty (or broken) loop:
 845 #ifdef ASSERT
 846   if (VerifyMethodHandles) {
 847     // Verify that &bottom < &top (non-empty interval)
 848     Label L_ok, L_bad;
 849     if (positive_distance_in_slots.is_register()) {
 850       __ cmpptr(positive_distance_in_slots.as_register(), (int32_t) 0);
 851       __ jcc(Assembler::lessEqual, L_bad);
 852     }
 853     __ cmpptr(rbx_bottom, rax_top);
 854     __ jcc(Assembler::below, L_ok);
 855     __ bind(L_bad);
 856     __ stop("valid bounds (copy up)");
 857     __ BIND(L_ok);
 858   }
 859 #endif
 860   __ cmpptr(rbx_bottom, rax_top);
 861   __ jccb(Assembler::aboveEqual, L_break);
 862   // work rax down to rbx, copying contiguous data upwards
 863   // In pseudo-code:
 864   //   [rbx, rax) = &[bottom, top)
 865   //   while (--rax >= rbx) *(rax + distance) = *(rax + 0), rax--;
 866   __ BIND(L_loop);
 867   __ subptr(rax_top, wordSize);
 868   __ movptr(rdx_temp, Address(rax_top, 0));
 869   __ movptr(          Address(rax_top, positive_distance_in_slots, Address::times_ptr), rdx_temp);
 870   __ cmpptr(rax_top, rbx_bottom);
 871   __ jcc(Assembler::above, L_loop);
 872   assert(Interpreter::stackElementSize == wordSize, "else change loop");
 873   __ bind(L_break);
 874   BLOCK_COMMENT("} move_arg_slots_up");
 875 }
 876 
 877 // in-place movement; no change to rsp
 878 // blows rax_temp, rdx_temp
 879 void MethodHandles::move_arg_slots_down(MacroAssembler* _masm,
 880                                         Address  bottom_addr,  // can use rax_temp
 881                                         Register rbx_top,      // invariant
 882                                         RegisterOrConstant negative_distance_in_slots,
 883                                         Register rax_temp, Register rdx_temp) {
 884   BLOCK_COMMENT("move_arg_slots_down {");
 885   assert_different_registers(rbx_top,
 886                              negative_distance_in_slots.register_or_noreg(),
 887                              rax_temp, rdx_temp);
 888   Label L_loop, L_break;
 889   Register rax_bottom = rax_temp;
 890   if (!bottom_addr.is_same_address(Address(rax_bottom, 0)))
 891     __ lea(rax_bottom, bottom_addr);
 892   // Detect empty (or broken) loop:
 893 #ifdef ASSERT
 894   assert(!negative_distance_in_slots.is_constant() || negative_distance_in_slots.as_constant() < 0, "");
 895   if (VerifyMethodHandles) {
 896     // Verify that &bottom < &top (non-empty interval)
 897     Label L_ok, L_bad;
 898     if (negative_distance_in_slots.is_register()) {
 899       __ cmpptr(negative_distance_in_slots.as_register(), (int32_t) 0);
 900       __ jcc(Assembler::greaterEqual, L_bad);
 901     }
 902     __ cmpptr(rax_bottom, rbx_top);
 903     __ jcc(Assembler::below, L_ok);
 904     __ bind(L_bad);
 905     __ stop("valid bounds (copy down)");
 906     __ BIND(L_ok);
 907   }
 908 #endif
 909   __ cmpptr(rax_bottom, rbx_top);
 910   __ jccb(Assembler::aboveEqual, L_break);
 911   // work rax up to rbx, copying contiguous data downwards
 912   // In pseudo-code:
 913   //   [rax, rbx) = &[bottom, top)
 914   //   while (rax < rbx) *(rax - distance) = *(rax + 0), rax++;
 915   __ BIND(L_loop);
 916   __ movptr(rdx_temp, Address(rax_bottom, 0));
 917   __ movptr(          Address(rax_bottom, negative_distance_in_slots, Address::times_ptr), rdx_temp);
 918   __ addptr(rax_bottom, wordSize);
 919   __ cmpptr(rax_bottom, rbx_top);
 920   __ jcc(Assembler::below, L_loop);
 921   assert(Interpreter::stackElementSize == wordSize, "else change loop");
 922   __ bind(L_break);
 923   BLOCK_COMMENT("} move_arg_slots_down");
 924 }
 925 
 926 // Copy from a field or array element to a stacked argument slot.
 927 // is_element (ignored) says whether caller is loading an array element instead of an instance field.
 928 void MethodHandles::move_typed_arg(MacroAssembler* _masm,
 929                                    BasicType type, bool is_element,
 930                                    Address slot_dest, Address value_src,
 931                                    Register rbx_temp, Register rdx_temp) {
 932   BLOCK_COMMENT(!is_element ? "move_typed_arg {" : "move_typed_arg { (array element)");
 933   if (type == T_OBJECT || type == T_ARRAY) {
 934     __ load_heap_oop(rbx_temp, value_src);
 935     __ movptr(slot_dest, rbx_temp);
 936   } else if (type != T_VOID) {
 937     int  arg_size      = type2aelembytes(type);
 938     bool arg_is_signed = is_signed_subword_type(type);
 939     int  slot_size     = (arg_size > wordSize) ? arg_size : wordSize;
 940     __ load_sized_value(  rdx_temp,  value_src, arg_size, arg_is_signed, rbx_temp);
 941     __ store_sized_value( slot_dest, rdx_temp,  slot_size,               rbx_temp);
 942   }
 943   BLOCK_COMMENT("} move_typed_arg");
 944 }
 945 
 946 void MethodHandles::move_return_value(MacroAssembler* _masm, BasicType type,
 947                                       Address return_slot) {
 948   BLOCK_COMMENT("move_return_value {");
 949   // Old versions of the JVM must clean the FPU stack after every return.
 950 #ifndef _LP64
 951 #ifdef COMPILER2
 952   // The FPU stack is clean if UseSSE >= 2 but must be cleaned in other cases
 953   if ((type == T_FLOAT && UseSSE < 1) || (type == T_DOUBLE && UseSSE < 2)) {
 954     for (int i = 1; i < 8; i++) {
 955         __ ffree(i);
 956     }
 957   } else if (UseSSE < 2) {
 958     __ empty_FPU_stack();
 959   }
 960 #endif //COMPILER2
 961 #endif //!_LP64
 962 
 963   // Look at the type and pull the value out of the corresponding register.
 964   if (type == T_VOID) {
 965     // nothing to do
 966   } else if (type == T_OBJECT) {
 967     __ movptr(return_slot, rax);
 968   } else if (type == T_INT || is_subword_type(type)) {
 969     // write the whole word, even if only 32 bits is significant
 970     __ movptr(return_slot, rax);
 971   } else if (type == T_LONG) {
 972     // store the value by parts
 973     // Note: We assume longs are continguous (if misaligned) on the interpreter stack.
 974     __ store_sized_value(return_slot, rax, BytesPerLong, rdx);
 975   } else if (NOT_LP64((type == T_FLOAT  && UseSSE < 1) ||
 976                       (type == T_DOUBLE && UseSSE < 2) ||)
 977              false) {
 978     // Use old x86 FPU registers:
 979     if (type == T_FLOAT)
 980       __ fstp_s(return_slot);
 981     else
 982       __ fstp_d(return_slot);
 983   } else if (type == T_FLOAT) {
 984     __ movflt(return_slot, xmm0);
 985   } else if (type == T_DOUBLE) {
 986     __ movdbl(return_slot, xmm0);
 987   } else {
 988     ShouldNotReachHere();
 989   }
 990   BLOCK_COMMENT("} move_return_value");
 991 }
 992 
 993 
 994 #ifndef PRODUCT
 995 extern "C" void print_method_handle(oop mh);
 996 void trace_method_handle_stub(const char* adaptername,
 997                               oop mh,
 998                               intptr_t* saved_regs,
 999                               intptr_t* entry_sp,
1000                               intptr_t* saved_sp,
1001                               intptr_t* saved_bp) {
1002   // called as a leaf from native code: do not block the JVM!
1003   bool has_mh = (strstr(adaptername, "return/") == NULL);  // return adapters don't have rcx_mh
1004   intptr_t* last_sp = (intptr_t*) saved_bp[frame::interpreter_frame_last_sp_offset];
1005   intptr_t* base_sp = last_sp;
1006   typedef MethodHandles::RicochetFrame RicochetFrame;
1007   RicochetFrame* rfp = (RicochetFrame*)((address)saved_bp - RicochetFrame::sender_link_offset_in_bytes());
1008   if (!UseRicochetFrames || Universe::heap()->is_in((address) rfp->saved_args_base())) {
1009     // Probably an interpreter frame.
1010     base_sp = (intptr_t*) saved_bp[frame::interpreter_frame_monitor_block_top_offset];
1011   }
1012   intptr_t    mh_reg = (intptr_t)mh;
1013   const char* mh_reg_name = "rcx_mh";
1014   if (!has_mh)  mh_reg_name = "rcx";
1015   tty->print_cr("MH %s %s="PTR_FORMAT" sp=("PTR_FORMAT"+"INTX_FORMAT") stack_size="INTX_FORMAT" bp="PTR_FORMAT,
1016                 adaptername, mh_reg_name, mh_reg,
1017                 (intptr_t)entry_sp, (intptr_t)(saved_sp - entry_sp), (intptr_t)(base_sp - last_sp), (intptr_t)saved_bp);
1018   if (Verbose) {
1019     tty->print(" reg dump: ");
1020     int saved_regs_count = (entry_sp-1) - saved_regs;
1021     // 32 bit: rdi rsi rbp rsp; rbx rdx rcx (*) rax
1022     int i;
1023     for (i = 0; i <= saved_regs_count; i++) {
1024       if (i > 0 && i % 4 == 0 && i != saved_regs_count) {
1025         tty->cr();
1026         tty->print("   + dump: ");
1027       }
1028       tty->print(" %d: "PTR_FORMAT, i, saved_regs[i]);
1029     }
1030     tty->cr();
1031     if (last_sp != saved_sp && last_sp != NULL)
1032       tty->print_cr("*** last_sp="PTR_FORMAT, (intptr_t)last_sp);
1033     int stack_dump_count = 16;
1034     if (stack_dump_count < (int)(saved_bp + 2 - saved_sp))
1035       stack_dump_count = (int)(saved_bp + 2 - saved_sp);
1036     if (stack_dump_count > 64)  stack_dump_count = 48;
1037     for (i = 0; i < stack_dump_count; i += 4) {
1038       tty->print_cr(" dump at SP[%d] "PTR_FORMAT": "PTR_FORMAT" "PTR_FORMAT" "PTR_FORMAT" "PTR_FORMAT,
1039                     i, (intptr_t) &entry_sp[i+0], entry_sp[i+0], entry_sp[i+1], entry_sp[i+2], entry_sp[i+3]);
1040     }
1041     if (has_mh)
1042       print_method_handle(mh);
1043   }
1044 }
1045 
1046 // The stub wraps the arguments in a struct on the stack to avoid
1047 // dealing with the different calling conventions for passing 6
1048 // arguments.
1049 struct MethodHandleStubArguments {
1050   const char* adaptername;
1051   oopDesc* mh;
1052   intptr_t* saved_regs;
1053   intptr_t* entry_sp;
1054   intptr_t* saved_sp;
1055   intptr_t* saved_bp;
1056 };
1057 void trace_method_handle_stub_wrapper(MethodHandleStubArguments* args) {
1058   trace_method_handle_stub(args->adaptername,
1059                            args->mh,
1060                            args->saved_regs,
1061                            args->entry_sp,
1062                            args->saved_sp,
1063                            args->saved_bp);
1064 }
1065 
1066 void MethodHandles::trace_method_handle(MacroAssembler* _masm, const char* adaptername) {
1067   if (!TraceMethodHandles)  return;
1068   BLOCK_COMMENT("trace_method_handle {");
1069   __ push(rax);
1070   __ lea(rax, Address(rsp, wordSize * NOT_LP64(6) LP64_ONLY(14))); // entry_sp  __ pusha();
1071   __ pusha();
1072   __ mov(rbx, rsp);
1073   __ enter();
1074   // incoming state:
1075   // rcx: method handle
1076   // r13 or rsi: saved sp
1077   // To avoid calling convention issues, build a record on the stack and pass the pointer to that instead.
1078   __ push(rbp);               // saved_bp
1079   __ push(rsi);               // saved_sp
1080   __ push(rax);               // entry_sp
1081   __ push(rbx);               // pusha saved_regs
1082   __ push(rcx);               // mh
1083   __ push(rcx);               // adaptername
1084   __ movptr(Address(rsp, 0), (intptr_t) adaptername);
1085   __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, trace_method_handle_stub_wrapper), rsp);
1086   __ leave();
1087   __ popa();
1088   __ pop(rax);
1089   BLOCK_COMMENT("} trace_method_handle");
1090 }
1091 #endif //PRODUCT
1092 
1093 // which conversion op types are implemented here?
1094 int MethodHandles::adapter_conversion_ops_supported_mask() {
1095   return ((1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_ONLY)
1096          |(1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_RAW)
1097          |(1<<java_lang_invoke_AdapterMethodHandle::OP_CHECK_CAST)
1098          |(1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_PRIM)
1099          |(1<<java_lang_invoke_AdapterMethodHandle::OP_REF_TO_PRIM)
1100           //OP_PRIM_TO_REF is below...
1101          |(1<<java_lang_invoke_AdapterMethodHandle::OP_SWAP_ARGS)
1102          |(1<<java_lang_invoke_AdapterMethodHandle::OP_ROT_ARGS)
1103          |(1<<java_lang_invoke_AdapterMethodHandle::OP_DUP_ARGS)
1104          |(1<<java_lang_invoke_AdapterMethodHandle::OP_DROP_ARGS)
1105           //OP_COLLECT_ARGS is below...
1106          |(1<<java_lang_invoke_AdapterMethodHandle::OP_SPREAD_ARGS)
1107          |(!UseRicochetFrames ? 0 :
1108            java_lang_invoke_MethodTypeForm::vmlayout_offset_in_bytes() <= 0 ? 0 :
1109            ((1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF)
1110            |(1<<java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS)
1111            |(1<<java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS)
1112             ))
1113          );
1114 }
1115 
1116 //------------------------------------------------------------------------------
1117 // MethodHandles::generate_method_handle_stub
1118 //
1119 // Generate an "entry" field for a method handle.
1120 // This determines how the method handle will respond to calls.
1121 void MethodHandles::generate_method_handle_stub(MacroAssembler* _masm, MethodHandles::EntryKind ek) {
1122   MethodHandles::EntryKind ek_orig = ek_original_kind(ek);
1123 
1124   // Here is the register state during an interpreted call,
1125   // as set up by generate_method_handle_interpreter_entry():
1126   // - rbx: garbage temp (was MethodHandle.invoke methodOop, unused)
1127   // - rcx: receiver method handle
1128   // - rax: method handle type (only used by the check_mtype entry point)
1129   // - rsi/r13: sender SP (must preserve; see prepare_to_jump_from_interpreted)
1130   // - rdx: garbage temp, can blow away
1131 
1132   const Register rcx_recv    = rcx;
1133   const Register rax_argslot = rax;
1134   const Register rbx_temp    = rbx;
1135   const Register rdx_temp    = rdx;
1136   const Register rdi_temp    = rdi;
1137 
1138   // This guy is set up by prepare_to_jump_from_interpreted (from interpreted calls)
1139   // and gen_c2i_adapter (from compiled calls):
1140   const Register saved_last_sp = saved_last_sp_register();
1141 
1142   // Argument registers for _raise_exception.
1143   // 32-bit: Pass first two oop/int args in registers ECX and EDX.
1144   const Register rarg0_code     = LP64_ONLY(j_rarg0) NOT_LP64(rcx);
1145   const Register rarg1_actual   = LP64_ONLY(j_rarg1) NOT_LP64(rdx);
1146   const Register rarg2_required = LP64_ONLY(j_rarg2) NOT_LP64(rdi);
1147   assert_different_registers(rarg0_code, rarg1_actual, rarg2_required, saved_last_sp);
1148 
1149   guarantee(java_lang_invoke_MethodHandle::vmentry_offset_in_bytes() != 0, "must have offsets");
1150 
1151   // some handy addresses
1152   Address rcx_mh_vmtarget(    rcx_recv, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes() );
1153   Address rcx_dmh_vmindex(    rcx_recv, java_lang_invoke_DirectMethodHandle::vmindex_offset_in_bytes() );
1154 
1155   Address rcx_bmh_vmargslot(  rcx_recv, java_lang_invoke_BoundMethodHandle::vmargslot_offset_in_bytes() );
1156   Address rcx_bmh_argument(   rcx_recv, java_lang_invoke_BoundMethodHandle::argument_offset_in_bytes() );
1157 
1158   Address rcx_amh_vmargslot(  rcx_recv, java_lang_invoke_AdapterMethodHandle::vmargslot_offset_in_bytes() );
1159   Address rcx_amh_argument(   rcx_recv, java_lang_invoke_AdapterMethodHandle::argument_offset_in_bytes() );
1160   Address rcx_amh_conversion( rcx_recv, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes() );
1161   Address vmarg;                // __ argument_address(vmargslot)
1162 
1163   const int java_mirror_offset = in_bytes(Klass::java_mirror_offset());
1164 
1165   if (have_entry(ek)) {
1166     __ nop();                   // empty stubs make SG sick
1167     return;
1168   }
1169 
1170 #ifdef ASSERT
1171   __ push((int32_t) 0xEEEEEEEE);
1172   __ push((int32_t) (intptr_t) entry_name(ek));
1173   LP64_ONLY(__ push((int32_t) high((intptr_t) entry_name(ek))));
1174   __ push((int32_t) 0x33333333);
1175 #endif //ASSERT
1176 
1177   address interp_entry = __ pc();
1178 
1179   trace_method_handle(_masm, entry_name(ek));
1180 
1181   BLOCK_COMMENT(err_msg("Entry %s {", entry_name(ek)));
1182 
1183   switch ((int) ek) {
1184   case _raise_exception:
1185     {
1186       // Not a real MH entry, but rather shared code for raising an
1187       // exception.  Since we use the compiled entry, arguments are
1188       // expected in compiler argument registers.
1189       assert(raise_exception_method(), "must be set");
1190       assert(raise_exception_method()->from_compiled_entry(), "method must be linked");
1191 
1192       const Register rax_pc = rax;
1193       __ pop(rax_pc);  // caller PC
1194       __ mov(rsp, saved_last_sp);  // cut the stack back to where the caller started
1195 
1196       Register rbx_method = rbx_temp;
1197       __ movptr(rbx_method, ExternalAddress((address) &_raise_exception_method));
1198 
1199       const int jobject_oop_offset = 0;
1200       __ movptr(rbx_method, Address(rbx_method, jobject_oop_offset));  // dereference the jobject
1201 
1202       __ movptr(saved_last_sp, rsp);
1203       __ subptr(rsp, 3 * wordSize);
1204       __ push(rax_pc);         // restore caller PC
1205 
1206       __ movl  (__ argument_address(constant(2)), rarg0_code);
1207       __ movptr(__ argument_address(constant(1)), rarg1_actual);
1208       __ movptr(__ argument_address(constant(0)), rarg2_required);
1209       jump_from_method_handle(_masm, rbx_method, rax);
1210     }
1211     break;
1212 
1213   case _invokestatic_mh:
1214   case _invokespecial_mh:
1215     {
1216       Register rbx_method = rbx_temp;
1217       __ load_heap_oop(rbx_method, rcx_mh_vmtarget); // target is a methodOop
1218       __ verify_oop(rbx_method);
1219       // same as TemplateTable::invokestatic or invokespecial,
1220       // minus the CP setup and profiling:
1221       if (ek == _invokespecial_mh) {
1222         // Must load & check the first argument before entering the target method.
1223         __ load_method_handle_vmslots(rax_argslot, rcx_recv, rdx_temp);
1224         __ movptr(rcx_recv, __ argument_address(rax_argslot, -1));
1225         __ null_check(rcx_recv);
1226         __ verify_oop(rcx_recv);
1227       }
1228       jump_from_method_handle(_masm, rbx_method, rax);
1229     }
1230     break;
1231 
1232   case _invokevirtual_mh:
1233     {
1234       // same as TemplateTable::invokevirtual,
1235       // minus the CP setup and profiling:
1236 
1237       // pick out the vtable index and receiver offset from the MH,
1238       // and then we can discard it:
1239       __ load_method_handle_vmslots(rax_argslot, rcx_recv, rdx_temp);
1240       Register rbx_index = rbx_temp;
1241       __ movl(rbx_index, rcx_dmh_vmindex);
1242       // Note:  The verifier allows us to ignore rcx_mh_vmtarget.
1243       __ movptr(rcx_recv, __ argument_address(rax_argslot, -1));
1244       __ null_check(rcx_recv, oopDesc::klass_offset_in_bytes());
1245 
1246       // get receiver klass
1247       Register rax_klass = rax_argslot;
1248       __ load_klass(rax_klass, rcx_recv);
1249       __ verify_oop(rax_klass);
1250 
1251       // get target methodOop & entry point
1252       const int base = instanceKlass::vtable_start_offset() * wordSize;
1253       assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1254       Address vtable_entry_addr(rax_klass,
1255                                 rbx_index, Address::times_ptr,
1256                                 base + vtableEntry::method_offset_in_bytes());
1257       Register rbx_method = rbx_temp;
1258       __ movptr(rbx_method, vtable_entry_addr);
1259 
1260       __ verify_oop(rbx_method);
1261       jump_from_method_handle(_masm, rbx_method, rax);
1262     }
1263     break;
1264 
1265   case _invokeinterface_mh:
1266     {
1267       // same as TemplateTable::invokeinterface,
1268       // minus the CP setup and profiling:
1269 
1270       // pick out the interface and itable index from the MH.
1271       __ load_method_handle_vmslots(rax_argslot, rcx_recv, rdx_temp);
1272       Register rdx_intf  = rdx_temp;
1273       Register rbx_index = rbx_temp;
1274       __ load_heap_oop(rdx_intf, rcx_mh_vmtarget);
1275       __ movl(rbx_index, rcx_dmh_vmindex);
1276       __ movptr(rcx_recv, __ argument_address(rax_argslot, -1));
1277       __ null_check(rcx_recv, oopDesc::klass_offset_in_bytes());
1278 
1279       // get receiver klass
1280       Register rax_klass = rax_argslot;
1281       __ load_klass(rax_klass, rcx_recv);
1282       __ verify_oop(rax_klass);
1283 
1284       Register rbx_method = rbx_index;
1285 
1286       // get interface klass
1287       Label no_such_interface;
1288       __ verify_oop(rdx_intf);
1289       __ lookup_interface_method(rax_klass, rdx_intf,
1290                                  // note: next two args must be the same:
1291                                  rbx_index, rbx_method,
1292                                  rdi_temp,
1293                                  no_such_interface);
1294 
1295       __ verify_oop(rbx_method);
1296       jump_from_method_handle(_masm, rbx_method, rax);
1297       __ hlt();
1298 
1299       __ bind(no_such_interface);
1300       // Throw an exception.
1301       // For historical reasons, it will be IncompatibleClassChangeError.
1302       __ mov(rbx_temp, rcx_recv);  // rarg2_required might be RCX
1303       assert_different_registers(rarg2_required, rbx_temp);
1304       __ movptr(rarg2_required, Address(rdx_intf, java_mirror_offset));  // required interface
1305       __ mov(   rarg1_actual,   rbx_temp);                               // bad receiver
1306       __ movl(  rarg0_code,     (int) Bytecodes::_invokeinterface);      // who is complaining?
1307       __ jump(ExternalAddress(from_interpreted_entry(_raise_exception)));
1308     }
1309     break;
1310 
1311   case _bound_ref_mh:
1312   case _bound_int_mh:
1313   case _bound_long_mh:
1314   case _bound_ref_direct_mh:
1315   case _bound_int_direct_mh:
1316   case _bound_long_direct_mh:
1317     {
1318       const bool direct_to_method = (ek >= _bound_ref_direct_mh);
1319       BasicType arg_type  = ek_bound_mh_arg_type(ek);
1320       int       arg_slots = type2size[arg_type];
1321 
1322       // make room for the new argument:
1323       __ movl(rax_argslot, rcx_bmh_vmargslot);
1324       __ lea(rax_argslot, __ argument_address(rax_argslot));
1325 
1326       insert_arg_slots(_masm, arg_slots * stack_move_unit(), rax_argslot, rbx_temp, rdx_temp);
1327 
1328       // store bound argument into the new stack slot:
1329       __ load_heap_oop(rbx_temp, rcx_bmh_argument);
1330       if (arg_type == T_OBJECT) {
1331         __ movptr(Address(rax_argslot, 0), rbx_temp);
1332       } else {
1333         Address prim_value_addr(rbx_temp, java_lang_boxing_object::value_offset_in_bytes(arg_type));
1334         move_typed_arg(_masm, arg_type, false,
1335                        Address(rax_argslot, 0),
1336                        prim_value_addr,
1337                        rbx_temp, rdx_temp);
1338       }
1339 
1340       if (direct_to_method) {
1341         Register rbx_method = rbx_temp;
1342         __ load_heap_oop(rbx_method, rcx_mh_vmtarget);
1343         __ verify_oop(rbx_method);
1344         jump_from_method_handle(_masm, rbx_method, rax);
1345       } else {
1346         __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1347         __ verify_oop(rcx_recv);
1348         __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1349       }
1350     }
1351     break;
1352 
1353   case _adapter_opt_profiling:
1354     if (java_lang_invoke_CountingMethodHandle::vmcount_offset_in_bytes() != 0) {
1355       Address rcx_mh_vmcount(rcx_recv, java_lang_invoke_CountingMethodHandle::vmcount_offset_in_bytes());
1356       __ incrementl(rcx_mh_vmcount);
1357     }
1358     // fall through
1359 
1360   case _adapter_retype_only:
1361   case _adapter_retype_raw:
1362     // immediately jump to the next MH layer:
1363     __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1364     __ verify_oop(rcx_recv);
1365     __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1366     // This is OK when all parameter types widen.
1367     // It is also OK when a return type narrows.
1368     break;
1369 
1370   case _adapter_check_cast:
1371     {
1372       // temps:
1373       Register rbx_klass = rbx_temp; // interesting AMH data
1374 
1375       // check a reference argument before jumping to the next layer of MH:
1376       __ movl(rax_argslot, rcx_amh_vmargslot);
1377       vmarg = __ argument_address(rax_argslot);
1378 
1379       // What class are we casting to?
1380       __ load_heap_oop(rbx_klass, rcx_amh_argument); // this is a Class object!
1381       load_klass_from_Class(_masm, rbx_klass);
1382 
1383       Label done;
1384       __ movptr(rdx_temp, vmarg);
1385       __ testptr(rdx_temp, rdx_temp);
1386       __ jcc(Assembler::zero, done);         // no cast if null
1387       __ load_klass(rdx_temp, rdx_temp);
1388 
1389       // live at this point:
1390       // - rbx_klass:  klass required by the target method
1391       // - rdx_temp:   argument klass to test
1392       // - rcx_recv:   adapter method handle
1393       __ check_klass_subtype(rdx_temp, rbx_klass, rax_argslot, done);
1394 
1395       // If we get here, the type check failed!
1396       // Call the wrong_method_type stub, passing the failing argument type in rax.
1397       Register rax_mtype = rax_argslot;
1398       __ movl(rax_argslot, rcx_amh_vmargslot);  // reload argslot field
1399       __ movptr(rdx_temp, vmarg);
1400 
1401       assert_different_registers(rarg2_required, rdx_temp);
1402       __ load_heap_oop(rarg2_required, rcx_amh_argument);             // required class
1403       __ mov(          rarg1_actual,   rdx_temp);                     // bad object
1404       __ movl(         rarg0_code,     (int) Bytecodes::_checkcast);  // who is complaining?
1405       __ jump(ExternalAddress(from_interpreted_entry(_raise_exception)));
1406 
1407       __ bind(done);
1408       // get the new MH:
1409       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1410       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1411     }
1412     break;
1413 
1414   case _adapter_prim_to_prim:
1415   case _adapter_ref_to_prim:
1416   case _adapter_prim_to_ref:
1417     // handled completely by optimized cases
1418     __ stop("init_AdapterMethodHandle should not issue this");
1419     break;
1420 
1421   case _adapter_opt_i2i:        // optimized subcase of adapt_prim_to_prim
1422 //case _adapter_opt_f2i:        // optimized subcase of adapt_prim_to_prim
1423   case _adapter_opt_l2i:        // optimized subcase of adapt_prim_to_prim
1424   case _adapter_opt_unboxi:     // optimized subcase of adapt_ref_to_prim
1425     {
1426       // perform an in-place conversion to int or an int subword
1427       __ movl(rax_argslot, rcx_amh_vmargslot);
1428       vmarg = __ argument_address(rax_argslot);
1429 
1430       switch (ek) {
1431       case _adapter_opt_i2i:
1432         __ movl(rdx_temp, vmarg);
1433         break;
1434       case _adapter_opt_l2i:
1435         {
1436           // just delete the extra slot; on a little-endian machine we keep the first
1437           __ lea(rax_argslot, __ argument_address(rax_argslot, 1));
1438           remove_arg_slots(_masm, -stack_move_unit(),
1439                            rax_argslot, rbx_temp, rdx_temp);
1440           vmarg = Address(rax_argslot, -Interpreter::stackElementSize);
1441           __ movl(rdx_temp, vmarg);
1442         }
1443         break;
1444       case _adapter_opt_unboxi:
1445         {
1446           // Load the value up from the heap.
1447           __ movptr(rdx_temp, vmarg);
1448           int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_INT);
1449 #ifdef ASSERT
1450           for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1451             if (is_subword_type(BasicType(bt)))
1452               assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(BasicType(bt)), "");
1453           }
1454 #endif
1455           __ null_check(rdx_temp, value_offset);
1456           __ movl(rdx_temp, Address(rdx_temp, value_offset));
1457           // We load this as a word.  Because we are little-endian,
1458           // the low bits will be correct, but the high bits may need cleaning.
1459           // The vminfo will guide us to clean those bits.
1460         }
1461         break;
1462       default:
1463         ShouldNotReachHere();
1464       }
1465 
1466       // Do the requested conversion and store the value.
1467       Register rbx_vminfo = rbx_temp;
1468       load_conversion_vminfo(_masm, rbx_vminfo, rcx_amh_conversion);
1469 
1470       // get the new MH:
1471       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1472       // (now we are done with the old MH)
1473 
1474       // original 32-bit vmdata word must be of this form:
1475       //    | MBZ:6 | signBitCount:8 | srcDstTypes:8 | conversionOp:8 |
1476       __ xchgptr(rcx, rbx_vminfo);                // free rcx for shifts
1477       __ shll(rdx_temp /*, rcx*/);
1478       Label zero_extend, done;
1479       __ testl(rcx, CONV_VMINFO_SIGN_FLAG);
1480       __ jccb(Assembler::zero, zero_extend);
1481 
1482       // this path is taken for int->byte, int->short
1483       __ sarl(rdx_temp /*, rcx*/);
1484       __ jmpb(done);
1485 
1486       __ bind(zero_extend);
1487       // this is taken for int->char
1488       __ shrl(rdx_temp /*, rcx*/);
1489 
1490       __ bind(done);
1491       __ movl(vmarg, rdx_temp);  // Store the value.
1492       __ xchgptr(rcx, rbx_vminfo);                // restore rcx_recv
1493 
1494       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1495     }
1496     break;
1497 
1498   case _adapter_opt_i2l:        // optimized subcase of adapt_prim_to_prim
1499   case _adapter_opt_unboxl:     // optimized subcase of adapt_ref_to_prim
1500     {
1501       // perform an in-place int-to-long or ref-to-long conversion
1502       __ movl(rax_argslot, rcx_amh_vmargslot);
1503 
1504       // on a little-endian machine we keep the first slot and add another after
1505       __ lea(rax_argslot, __ argument_address(rax_argslot, 1));
1506       insert_arg_slots(_masm, stack_move_unit(),
1507                        rax_argslot, rbx_temp, rdx_temp);
1508       Address vmarg1(rax_argslot, -Interpreter::stackElementSize);
1509       Address vmarg2 = vmarg1.plus_disp(Interpreter::stackElementSize);
1510 
1511       switch (ek) {
1512       case _adapter_opt_i2l:
1513         {
1514 #ifdef _LP64
1515           __ movslq(rdx_temp, vmarg1);  // Load sign-extended
1516           __ movq(vmarg1, rdx_temp);    // Store into first slot
1517 #else
1518           __ movl(rdx_temp, vmarg1);
1519           __ sarl(rdx_temp, BitsPerInt - 1);  // __ extend_sign()
1520           __ movl(vmarg2, rdx_temp); // store second word
1521 #endif
1522         }
1523         break;
1524       case _adapter_opt_unboxl:
1525         {
1526           // Load the value up from the heap.
1527           __ movptr(rdx_temp, vmarg1);
1528           int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_LONG);
1529           assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(T_DOUBLE), "");
1530           __ null_check(rdx_temp, value_offset);
1531 #ifdef _LP64
1532           __ movq(rbx_temp, Address(rdx_temp, value_offset));
1533           __ movq(vmarg1, rbx_temp);
1534 #else
1535           __ movl(rbx_temp, Address(rdx_temp, value_offset + 0*BytesPerInt));
1536           __ movl(rdx_temp, Address(rdx_temp, value_offset + 1*BytesPerInt));
1537           __ movl(vmarg1, rbx_temp);
1538           __ movl(vmarg2, rdx_temp);
1539 #endif
1540         }
1541         break;
1542       default:
1543         ShouldNotReachHere();
1544       }
1545 
1546       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1547       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1548     }
1549     break;
1550 
1551   case _adapter_opt_f2d:        // optimized subcase of adapt_prim_to_prim
1552   case _adapter_opt_d2f:        // optimized subcase of adapt_prim_to_prim
1553     {
1554       // perform an in-place floating primitive conversion
1555       __ movl(rax_argslot, rcx_amh_vmargslot);
1556       __ lea(rax_argslot, __ argument_address(rax_argslot, 1));
1557       if (ek == _adapter_opt_f2d) {
1558         insert_arg_slots(_masm, stack_move_unit(),
1559                          rax_argslot, rbx_temp, rdx_temp);
1560       }
1561       Address vmarg(rax_argslot, -Interpreter::stackElementSize);
1562 
1563 #ifdef _LP64
1564       if (ek == _adapter_opt_f2d) {
1565         __ movflt(xmm0, vmarg);
1566         __ cvtss2sd(xmm0, xmm0);
1567         __ movdbl(vmarg, xmm0);
1568       } else {
1569         __ movdbl(xmm0, vmarg);
1570         __ cvtsd2ss(xmm0, xmm0);
1571         __ movflt(vmarg, xmm0);
1572       }
1573 #else //_LP64
1574       if (ek == _adapter_opt_f2d) {
1575         __ fld_s(vmarg);        // load float to ST0
1576         __ fstp_d(vmarg);       // store double
1577       } else {
1578         __ fld_d(vmarg);        // load double to ST0
1579         __ fstp_s(vmarg);       // store single
1580       }
1581 #endif //_LP64
1582 
1583       if (ek == _adapter_opt_d2f) {
1584         remove_arg_slots(_masm, -stack_move_unit(),
1585                          rax_argslot, rbx_temp, rdx_temp);
1586       }
1587 
1588       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1589       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1590     }
1591     break;
1592 
1593   case _adapter_swap_args:
1594   case _adapter_rot_args:
1595     // handled completely by optimized cases
1596     __ stop("init_AdapterMethodHandle should not issue this");
1597     break;
1598 
1599   case _adapter_opt_swap_1:
1600   case _adapter_opt_swap_2:
1601   case _adapter_opt_rot_1_up:
1602   case _adapter_opt_rot_1_down:
1603   case _adapter_opt_rot_2_up:
1604   case _adapter_opt_rot_2_down:
1605     {
1606       int swap_slots = ek_adapter_opt_swap_slots(ek);
1607       int rotate     = ek_adapter_opt_swap_mode(ek);
1608 
1609       // 'argslot' is the position of the first argument to swap
1610       __ movl(rax_argslot, rcx_amh_vmargslot);
1611       __ lea(rax_argslot, __ argument_address(rax_argslot));
1612 
1613       // 'vminfo' is the second
1614       Register rbx_destslot = rbx_temp;
1615       load_conversion_vminfo(_masm, rbx_destslot, rcx_amh_conversion);
1616       __ lea(rbx_destslot, __ argument_address(rbx_destslot));
1617       if (VerifyMethodHandles)
1618         verify_argslot(_masm, rbx_destslot, "swap point must fall within current frame");
1619 
1620       assert(Interpreter::stackElementSize == wordSize, "else rethink use of wordSize here");
1621       if (!rotate) {
1622         // simple swap
1623         for (int i = 0; i < swap_slots; i++) {
1624           __ movptr(rdi_temp, Address(rax_argslot,  i * wordSize));
1625           __ movptr(rdx_temp, Address(rbx_destslot, i * wordSize));
1626           __ movptr(Address(rax_argslot,  i * wordSize), rdx_temp);
1627           __ movptr(Address(rbx_destslot, i * wordSize), rdi_temp);
1628         }
1629       } else {
1630         // A rotate is actually pair of moves, with an "odd slot" (or pair)
1631         // changing place with a series of other slots.
1632         // First, push the "odd slot", which is going to get overwritten
1633         for (int i = swap_slots - 1; i >= 0; i--) {
1634           // handle one with rdi_temp instead of a push:
1635           if (i == 0)  __ movptr(rdi_temp, Address(rax_argslot, i * wordSize));
1636           else         __ pushptr(         Address(rax_argslot, i * wordSize));
1637         }
1638         if (rotate > 0) {
1639           // Here is rotate > 0:
1640           // (low mem)                                          (high mem)
1641           //     | dest:     more_slots...     | arg: odd_slot :arg+1 |
1642           // =>
1643           //     | dest: odd_slot | dest+1: more_slots...      :arg+1 |
1644           // work argslot down to destslot, copying contiguous data upwards
1645           // pseudo-code:
1646           //   rax = src_addr - swap_bytes
1647           //   rbx = dest_addr
1648           //   while (rax >= rbx) *(rax + swap_bytes) = *(rax + 0), rax--;
1649           move_arg_slots_up(_masm,
1650                             rbx_destslot,
1651                             Address(rax_argslot, 0),
1652                             swap_slots,
1653                             rax_argslot, rdx_temp);
1654         } else {
1655           // Here is the other direction, rotate < 0:
1656           // (low mem)                                          (high mem)
1657           //     | arg: odd_slot | arg+1: more_slots...       :dest+1 |
1658           // =>
1659           //     | arg:    more_slots...     | dest: odd_slot :dest+1 |
1660           // work argslot up to destslot, copying contiguous data downwards
1661           // pseudo-code:
1662           //   rax = src_addr + swap_bytes
1663           //   rbx = dest_addr
1664           //   while (rax <= rbx) *(rax - swap_bytes) = *(rax + 0), rax++;
1665           // dest_slot denotes an exclusive upper limit
1666           int limit_bias = OP_ROT_ARGS_DOWN_LIMIT_BIAS;
1667           if (limit_bias != 0)
1668             __ addptr(rbx_destslot, - limit_bias * wordSize);
1669           move_arg_slots_down(_masm,
1670                               Address(rax_argslot, swap_slots * wordSize),
1671                               rbx_destslot,
1672                               -swap_slots,
1673                               rax_argslot, rdx_temp);
1674           __ subptr(rbx_destslot, swap_slots * wordSize);
1675         }
1676         // pop the original first chunk into the destination slot, now free
1677         for (int i = 0; i < swap_slots; i++) {
1678           if (i == 0)  __ movptr(Address(rbx_destslot, i * wordSize), rdi_temp);
1679           else         __ popptr(Address(rbx_destslot, i * wordSize));
1680         }
1681       }
1682 
1683       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1684       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1685     }
1686     break;
1687 
1688   case _adapter_dup_args:
1689     {
1690       // 'argslot' is the position of the first argument to duplicate
1691       __ movl(rax_argslot, rcx_amh_vmargslot);
1692       __ lea(rax_argslot, __ argument_address(rax_argslot));
1693 
1694       // 'stack_move' is negative number of words to duplicate
1695       Register rdi_stack_move = rdi_temp;
1696       load_stack_move(_masm, rdi_stack_move, rcx_recv, true);
1697 
1698       if (VerifyMethodHandles) {
1699         verify_argslots(_masm, rdi_stack_move, rax_argslot, true,
1700                         "copied argument(s) must fall within current frame");
1701       }
1702 
1703       // insert location is always the bottom of the argument list:
1704       Address insert_location = __ argument_address(constant(0));
1705       int pre_arg_words = insert_location.disp() / wordSize;   // return PC is pushed
1706       assert(insert_location.base() == rsp, "");
1707 
1708       __ negl(rdi_stack_move);
1709       push_arg_slots(_masm, rax_argslot, rdi_stack_move,
1710                      pre_arg_words, rbx_temp, rdx_temp);
1711 
1712       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1713       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1714     }
1715     break;
1716 
1717   case _adapter_drop_args:
1718     {
1719       // 'argslot' is the position of the first argument to nuke
1720       __ movl(rax_argslot, rcx_amh_vmargslot);
1721       __ lea(rax_argslot, __ argument_address(rax_argslot));
1722 
1723       // (must do previous push after argslot address is taken)
1724 
1725       // 'stack_move' is number of words to drop
1726       Register rdi_stack_move = rdi_temp;
1727       load_stack_move(_masm, rdi_stack_move, rcx_recv, false);
1728       remove_arg_slots(_masm, rdi_stack_move,
1729                        rax_argslot, rbx_temp, rdx_temp);
1730 
1731       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
1732       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
1733     }
1734     break;
1735 
1736   case _adapter_collect_args:
1737   case _adapter_fold_args:
1738   case _adapter_spread_args:
1739     // handled completely by optimized cases
1740     __ stop("init_AdapterMethodHandle should not issue this");
1741     break;
1742 
1743   case _adapter_opt_collect_ref:
1744   case _adapter_opt_collect_int:
1745   case _adapter_opt_collect_long:
1746   case _adapter_opt_collect_float:
1747   case _adapter_opt_collect_double:
1748   case _adapter_opt_collect_void:
1749   case _adapter_opt_collect_0_ref:
1750   case _adapter_opt_collect_1_ref:
1751   case _adapter_opt_collect_2_ref:
1752   case _adapter_opt_collect_3_ref:
1753   case _adapter_opt_collect_4_ref:
1754   case _adapter_opt_collect_5_ref:
1755   case _adapter_opt_filter_S0_ref:
1756   case _adapter_opt_filter_S1_ref:
1757   case _adapter_opt_filter_S2_ref:
1758   case _adapter_opt_filter_S3_ref:
1759   case _adapter_opt_filter_S4_ref:
1760   case _adapter_opt_filter_S5_ref:
1761   case _adapter_opt_collect_2_S0_ref:
1762   case _adapter_opt_collect_2_S1_ref:
1763   case _adapter_opt_collect_2_S2_ref:
1764   case _adapter_opt_collect_2_S3_ref:
1765   case _adapter_opt_collect_2_S4_ref:
1766   case _adapter_opt_collect_2_S5_ref:
1767   case _adapter_opt_fold_ref:
1768   case _adapter_opt_fold_int:
1769   case _adapter_opt_fold_long:
1770   case _adapter_opt_fold_float:
1771   case _adapter_opt_fold_double:
1772   case _adapter_opt_fold_void:
1773   case _adapter_opt_fold_1_ref:
1774   case _adapter_opt_fold_2_ref:
1775   case _adapter_opt_fold_3_ref:
1776   case _adapter_opt_fold_4_ref:
1777   case _adapter_opt_fold_5_ref:
1778     {
1779       // Given a fresh incoming stack frame, build a new ricochet frame.
1780       // On entry, TOS points at a return PC, and RBP is the callers frame ptr.
1781       // RSI/R13 has the caller's exact stack pointer, which we must also preserve.
1782       // RCX contains an AdapterMethodHandle of the indicated kind.
1783 
1784       // Relevant AMH fields:
1785       // amh.vmargslot:
1786       //   points to the trailing edge of the arguments
1787       //   to filter, collect, or fold.  For a boxing operation,
1788       //   it points just after the single primitive value.
1789       // amh.argument:
1790       //   recursively called MH, on |collect| arguments
1791       // amh.vmtarget:
1792       //   final destination MH, on return value, etc.
1793       // amh.conversion.dest:
1794       //   tells what is the type of the return value
1795       //   (not needed here, since dest is also derived from ek)
1796       // amh.conversion.vminfo:
1797       //   points to the trailing edge of the return value
1798       //   when the vmtarget is to be called; this is
1799       //   equal to vmargslot + (retained ? |collect| : 0)
1800 
1801       // Pass 0 or more argument slots to the recursive target.
1802       int collect_count_constant = ek_adapter_opt_collect_count(ek);
1803 
1804       // The collected arguments are copied from the saved argument list:
1805       int collect_slot_constant = ek_adapter_opt_collect_slot(ek);
1806 
1807       assert(ek_orig == _adapter_collect_args ||
1808              ek_orig == _adapter_fold_args, "");
1809       bool retain_original_args = (ek_orig == _adapter_fold_args);
1810 
1811       // The return value is replaced (or inserted) at the 'vminfo' argslot.
1812       // Sometimes we can compute this statically.
1813       int dest_slot_constant = -1;
1814       if (!retain_original_args)
1815         dest_slot_constant = collect_slot_constant;
1816       else if (collect_slot_constant >= 0 && collect_count_constant >= 0)
1817         // We are preserving all the arguments, and the return value is prepended,
1818         // so the return slot is to the left (above) the |collect| sequence.
1819         dest_slot_constant = collect_slot_constant + collect_count_constant;
1820 
1821       // Replace all those slots by the result of the recursive call.
1822       // The result type can be one of ref, int, long, float, double, void.
1823       // In the case of void, nothing is pushed on the stack after return.
1824       BasicType dest = ek_adapter_opt_collect_type(ek);
1825       assert(dest == type2wfield[dest], "dest is a stack slot type");
1826       int dest_count = type2size[dest];
1827       assert(dest_count == 1 || dest_count == 2 || (dest_count == 0 && dest == T_VOID), "dest has a size");
1828 
1829       // Choose a return continuation.
1830       EntryKind ek_ret = _adapter_opt_return_any;
1831       if (dest != T_CONFLICT && OptimizeMethodHandles) {
1832         switch (dest) {
1833         case T_INT    : ek_ret = _adapter_opt_return_int;     break;
1834         case T_LONG   : ek_ret = _adapter_opt_return_long;    break;
1835         case T_FLOAT  : ek_ret = _adapter_opt_return_float;   break;
1836         case T_DOUBLE : ek_ret = _adapter_opt_return_double;  break;
1837         case T_OBJECT : ek_ret = _adapter_opt_return_ref;     break;
1838         case T_VOID   : ek_ret = _adapter_opt_return_void;    break;
1839         default       : ShouldNotReachHere();
1840         }
1841         if (dest == T_OBJECT && dest_slot_constant >= 0) {
1842           EntryKind ek_try = EntryKind(_adapter_opt_return_S0_ref + dest_slot_constant);
1843           if (ek_try <= _adapter_opt_return_LAST &&
1844               ek_adapter_opt_return_slot(ek_try) == dest_slot_constant) {
1845             ek_ret = ek_try;
1846           }
1847         }
1848         assert(ek_adapter_opt_return_type(ek_ret) == dest, "");
1849       }
1850 
1851       // Already pushed:  ... keep1 | collect | keep2 | sender_pc |
1852       // push(sender_pc);
1853 
1854       // Compute argument base:
1855       Register rax_argv = rax_argslot;
1856       __ lea(rax_argv, __ argument_address(constant(0)));
1857 
1858       // Push a few extra argument words, if we need them to store the return value.
1859       {
1860         int extra_slots = 0;
1861         if (retain_original_args) {
1862           extra_slots = dest_count;
1863         } else if (collect_count_constant == -1) {
1864           extra_slots = dest_count;  // collect_count might be zero; be generous
1865         } else if (dest_count > collect_count_constant) {
1866           extra_slots = (dest_count - collect_count_constant);
1867         } else {
1868           // else we know we have enough dead space in |collect| to repurpose for return values
1869         }
1870         DEBUG_ONLY(extra_slots += 1);
1871         if (extra_slots > 0) {
1872           __ pop(rbx_temp);   // return value
1873           __ subptr(rsp, (extra_slots * Interpreter::stackElementSize));
1874           // Push guard word #2 in debug mode.
1875           DEBUG_ONLY(__ movptr(Address(rsp, 0), (int32_t) RicochetFrame::MAGIC_NUMBER_2));
1876           __ push(rbx_temp);
1877         }
1878       }
1879 
1880       RicochetFrame::enter_ricochet_frame(_masm, rcx_recv, rax_argv,
1881                                           entry(ek_ret)->from_interpreted_entry(), rbx_temp);
1882 
1883       // Now pushed:  ... keep1 | collect | keep2 | RF |
1884       // some handy frame slots:
1885       Address exact_sender_sp_addr = RicochetFrame::frame_address(RicochetFrame::exact_sender_sp_offset_in_bytes());
1886       Address conversion_addr      = RicochetFrame::frame_address(RicochetFrame::conversion_offset_in_bytes());
1887       Address saved_args_base_addr = RicochetFrame::frame_address(RicochetFrame::saved_args_base_offset_in_bytes());
1888 
1889 #ifdef ASSERT
1890       if (VerifyMethodHandles && dest != T_CONFLICT) {
1891         BLOCK_COMMENT("verify AMH.conv.dest");
1892         load_conversion_dest_type(_masm, rbx_temp, conversion_addr);
1893         Label L_dest_ok;
1894         __ cmpl(rbx_temp, (int) dest);
1895         __ jcc(Assembler::equal, L_dest_ok);
1896         if (dest == T_INT) {
1897           for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1898             if (is_subword_type(BasicType(bt))) {
1899               __ cmpl(rbx_temp, (int) bt);
1900               __ jcc(Assembler::equal, L_dest_ok);
1901             }
1902           }
1903         }
1904         __ stop("bad dest in AMH.conv");
1905         __ BIND(L_dest_ok);
1906       }
1907 #endif //ASSERT
1908 
1909       // Find out where the original copy of the recursive argument sequence begins.
1910       Register rax_coll = rax_argv;
1911       {
1912         RegisterOrConstant collect_slot = collect_slot_constant;
1913         if (collect_slot_constant == -1) {
1914           __ movl(rdi_temp, rcx_amh_vmargslot);
1915           collect_slot = rdi_temp;
1916         }
1917         if (collect_slot_constant != 0)
1918           __ lea(rax_coll, Address(rax_argv, collect_slot, Interpreter::stackElementScale()));
1919         // rax_coll now points at the trailing edge of |collect| and leading edge of |keep2|
1920       }
1921 
1922       // Replace the old AMH with the recursive MH.  (No going back now.)
1923       // In the case of a boxing call, the recursive call is to a 'boxer' method,
1924       // such as Integer.valueOf or Long.valueOf.  In the case of a filter
1925       // or collect call, it will take one or more arguments, transform them,
1926       // and return some result, to store back into argument_base[vminfo].
1927       __ load_heap_oop(rcx_recv, rcx_amh_argument);
1928       if (VerifyMethodHandles)  verify_method_handle(_masm, rcx_recv);
1929 
1930       // Push a space for the recursively called MH first:
1931       __ push((int32_t)NULL_WORD);
1932 
1933       // Calculate |collect|, the number of arguments we are collecting.
1934       Register rdi_collect_count = rdi_temp;
1935       RegisterOrConstant collect_count;
1936       if (collect_count_constant >= 0) {
1937         collect_count = collect_count_constant;
1938       } else {
1939         __ load_method_handle_vmslots(rdi_collect_count, rcx_recv, rdx_temp);
1940         collect_count = rdi_collect_count;
1941       }
1942 #ifdef ASSERT
1943       if (VerifyMethodHandles && collect_count_constant >= 0) {
1944         __ load_method_handle_vmslots(rbx_temp, rcx_recv, rdx_temp);
1945         Label L_count_ok;
1946         __ cmpl(rbx_temp, collect_count_constant);
1947         __ jcc(Assembler::equal, L_count_ok);
1948         __ stop("bad vminfo in AMH.conv");
1949         __ BIND(L_count_ok);
1950       }
1951 #endif //ASSERT
1952 
1953       // copy |collect| slots directly to TOS:
1954       push_arg_slots(_masm, rax_coll, collect_count, 0, rbx_temp, rdx_temp);
1955       // Now pushed:  ... keep1 | collect | keep2 | RF... | collect |
1956       // rax_coll still points at the trailing edge of |collect| and leading edge of |keep2|
1957 
1958       // If necessary, adjust the saved arguments to make room for the eventual return value.
1959       // Normal adjustment:  ... keep1 | +dest+ | -collect- | keep2 | RF... | collect |
1960       // If retaining args:  ... keep1 | +dest+ |  collect  | keep2 | RF... | collect |
1961       // In the non-retaining case, this might move keep2 either up or down.
1962       // We don't have to copy the whole | RF... collect | complex,
1963       // but we must adjust RF.saved_args_base.
1964       // Also, from now on, we will forget about the original copy of |collect|.
1965       // If we are retaining it, we will treat it as part of |keep2|.
1966       // For clarity we will define |keep3| = |collect|keep2| or |keep2|.
1967 
1968       BLOCK_COMMENT("adjust trailing arguments {");
1969       // Compare the sizes of |+dest+| and |-collect-|, which are opposed opening and closing movements.
1970       int                open_count  = dest_count;
1971       RegisterOrConstant close_count = collect_count_constant;
1972       Register rdi_close_count = rdi_collect_count;
1973       if (retain_original_args) {
1974         close_count = constant(0);
1975       } else if (collect_count_constant == -1) {
1976         close_count = rdi_collect_count;
1977       }
1978 
1979       // How many slots need moving?  This is simply dest_slot (0 => no |keep3|).
1980       RegisterOrConstant keep3_count;
1981       Register rsi_keep3_count = rsi;  // can repair from RF.exact_sender_sp
1982       if (dest_slot_constant >= 0) {
1983         keep3_count = dest_slot_constant;
1984       } else  {
1985         load_conversion_vminfo(_masm, rsi_keep3_count, conversion_addr);
1986         keep3_count = rsi_keep3_count;
1987       }
1988 #ifdef ASSERT
1989       if (VerifyMethodHandles && dest_slot_constant >= 0) {
1990         load_conversion_vminfo(_masm, rbx_temp, conversion_addr);
1991         Label L_vminfo_ok;
1992         __ cmpl(rbx_temp, dest_slot_constant);
1993         __ jcc(Assembler::equal, L_vminfo_ok);
1994         __ stop("bad vminfo in AMH.conv");
1995         __ BIND(L_vminfo_ok);
1996       }
1997 #endif //ASSERT
1998 
1999       // tasks remaining:
2000       bool move_keep3 = (!keep3_count.is_constant() || keep3_count.as_constant() != 0);
2001       bool stomp_dest = (NOT_DEBUG(dest == T_OBJECT) DEBUG_ONLY(dest_count != 0));
2002       bool fix_arg_base = (!close_count.is_constant() || open_count != close_count.as_constant());
2003 
2004       if (stomp_dest | fix_arg_base) {
2005         // we will probably need an updated rax_argv value
2006         if (collect_slot_constant >= 0) {
2007           // rax_coll already holds the leading edge of |keep2|, so tweak it
2008           assert(rax_coll == rax_argv, "elided a move");
2009           if (collect_slot_constant != 0)
2010             __ subptr(rax_argv, collect_slot_constant * Interpreter::stackElementSize);
2011         } else {
2012           // Just reload from RF.saved_args_base.
2013           __ movptr(rax_argv, saved_args_base_addr);
2014         }
2015       }
2016 
2017       // Old and new argument locations (based at slot 0).
2018       // Net shift (&new_argv - &old_argv) is (close_count - open_count).
2019       bool zero_open_count = (open_count == 0);  // remember this bit of info
2020       if (move_keep3 && fix_arg_base) {
2021         // It will be easier to have everything in one register:
2022         if (close_count.is_register()) {
2023           // Deduct open_count from close_count register to get a clean +/- value.
2024           __ subptr(close_count.as_register(), open_count);
2025         } else {
2026           close_count = close_count.as_constant() - open_count;
2027         }
2028         open_count = 0;
2029       }
2030       Address old_argv(rax_argv, 0);
2031       Address new_argv(rax_argv, close_count,  Interpreter::stackElementScale(),
2032                                 - open_count * Interpreter::stackElementSize);
2033 
2034       // First decide if any actual data are to be moved.
2035       // We can skip if (a) |keep3| is empty, or (b) the argument list size didn't change.
2036       // (As it happens, all movements involve an argument list size change.)
2037 
2038       // If there are variable parameters, use dynamic checks to skip around the whole mess.
2039       Label L_done;
2040       if (!keep3_count.is_constant()) {
2041         __ testl(keep3_count.as_register(), keep3_count.as_register());
2042         __ jcc(Assembler::zero, L_done);
2043       }
2044       if (!close_count.is_constant()) {
2045         __ cmpl(close_count.as_register(), open_count);
2046         __ jcc(Assembler::equal, L_done);
2047       }
2048 
2049       if (move_keep3 && fix_arg_base) {
2050         bool emit_move_down = false, emit_move_up = false, emit_guard = false;
2051         if (!close_count.is_constant()) {
2052           emit_move_down = emit_guard = !zero_open_count;
2053           emit_move_up   = true;
2054         } else if (open_count != close_count.as_constant()) {
2055           emit_move_down = (open_count > close_count.as_constant());
2056           emit_move_up   = !emit_move_down;
2057         }
2058         Label L_move_up;
2059         if (emit_guard) {
2060           __ cmpl(close_count.as_register(), open_count);
2061           __ jcc(Assembler::greater, L_move_up);
2062         }
2063 
2064         if (emit_move_down) {
2065           // Move arguments down if |+dest+| > |-collect-|
2066           // (This is rare, except when arguments are retained.)
2067           // This opens space for the return value.
2068           if (keep3_count.is_constant()) {
2069             for (int i = 0; i < keep3_count.as_constant(); i++) {
2070               __ movptr(rdx_temp, old_argv.plus_disp(i * Interpreter::stackElementSize));
2071               __ movptr(          new_argv.plus_disp(i * Interpreter::stackElementSize), rdx_temp);
2072             }
2073           } else {
2074             Register rbx_argv_top = rbx_temp;
2075             __ lea(rbx_argv_top, old_argv.plus_disp(keep3_count, Interpreter::stackElementScale()));
2076             move_arg_slots_down(_masm,
2077                                 old_argv,     // beginning of old argv
2078                                 rbx_argv_top, // end of old argv
2079                                 close_count,  // distance to move down (must be negative)
2080                                 rax_argv, rdx_temp);
2081             // Used argv as an iteration variable; reload from RF.saved_args_base.
2082             __ movptr(rax_argv, saved_args_base_addr);
2083           }
2084         }
2085 
2086         if (emit_guard) {
2087           __ jmp(L_done);  // assumes emit_move_up is true also
2088           __ BIND(L_move_up);
2089         }
2090 
2091         if (emit_move_up) {
2092 
2093           // Move arguments up if |+dest+| < |-collect-|
2094           // (This is usual, except when |keep3| is empty.)
2095           // This closes up the space occupied by the now-deleted collect values.
2096           if (keep3_count.is_constant()) {
2097             for (int i = keep3_count.as_constant() - 1; i >= 0; i--) {
2098               __ movptr(rdx_temp, old_argv.plus_disp(i * Interpreter::stackElementSize));
2099               __ movptr(          new_argv.plus_disp(i * Interpreter::stackElementSize), rdx_temp);
2100             }
2101           } else {
2102             Address argv_top = old_argv.plus_disp(keep3_count, Interpreter::stackElementScale());
2103             move_arg_slots_up(_masm,
2104                               rax_argv,     // beginning of old argv
2105                               argv_top,     // end of old argv
2106                               close_count,  // distance to move up (must be positive)
2107                               rbx_temp, rdx_temp);
2108           }
2109         }
2110       }
2111       __ BIND(L_done);
2112 
2113       if (fix_arg_base) {
2114         // adjust RF.saved_args_base by adding (close_count - open_count)
2115         if (!new_argv.is_same_address(Address(rax_argv, 0)))
2116           __ lea(rax_argv, new_argv);
2117         __ movptr(saved_args_base_addr, rax_argv);
2118       }
2119 
2120       if (stomp_dest) {
2121         // Stomp the return slot, so it doesn't hold garbage.
2122         // This isn't strictly necessary, but it may help detect bugs.
2123         int forty_two = RicochetFrame::RETURN_VALUE_PLACEHOLDER;
2124         __ movptr(Address(rax_argv, keep3_count, Address::times_ptr),
2125                   (int32_t) forty_two);
2126         // uses rsi_keep3_count
2127       }
2128       BLOCK_COMMENT("} adjust trailing arguments");
2129 
2130       BLOCK_COMMENT("do_recursive_call");
2131       __ mov(saved_last_sp, rsp);    // set rsi/r13 for callee
2132       __ pushptr(ExternalAddress(SharedRuntime::ricochet_blob()->bounce_addr()).addr());
2133       // The globally unique bounce address has two purposes:
2134       // 1. It helps the JVM recognize this frame (frame::is_ricochet_frame).
2135       // 2. When returned to, it cuts back the stack and redirects control flow
2136       //    to the return handler.
2137       // The return handler will further cut back the stack when it takes
2138       // down the RF.  Perhaps there is a way to streamline this further.
2139 
2140       // State during recursive call:
2141       // ... keep1 | dest | dest=42 | keep3 | RF... | collect | bounce_pc |
2142       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
2143 
2144       break;
2145     }
2146 
2147   case _adapter_opt_return_ref:
2148   case _adapter_opt_return_int:
2149   case _adapter_opt_return_long:
2150   case _adapter_opt_return_float:
2151   case _adapter_opt_return_double:
2152   case _adapter_opt_return_void:
2153   case _adapter_opt_return_S0_ref:
2154   case _adapter_opt_return_S1_ref:
2155   case _adapter_opt_return_S2_ref:
2156   case _adapter_opt_return_S3_ref:
2157   case _adapter_opt_return_S4_ref:
2158   case _adapter_opt_return_S5_ref:
2159     {
2160       BasicType dest_type_constant = ek_adapter_opt_return_type(ek);
2161       int       dest_slot_constant = ek_adapter_opt_return_slot(ek);
2162 
2163       if (VerifyMethodHandles)  RicochetFrame::verify_clean(_masm);
2164 
2165       if (dest_slot_constant == -1) {
2166         // The current stub is a general handler for this dest_type.
2167         // It can be called from _adapter_opt_return_any below.
2168         // Stash the address in a little table.
2169         assert((dest_type_constant & CONV_TYPE_MASK) == dest_type_constant, "oob");
2170         address return_handler = __ pc();
2171         _adapter_return_handlers[dest_type_constant] = return_handler;
2172         if (dest_type_constant == T_INT) {
2173           // do the subword types too
2174           for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
2175             if (is_subword_type(BasicType(bt)) &&
2176                 _adapter_return_handlers[bt] == NULL) {
2177               _adapter_return_handlers[bt] = return_handler;
2178             }
2179           }
2180         }
2181       }
2182 
2183       Register rbx_arg_base = rbx_temp;
2184       assert_different_registers(rax, rdx,  // possibly live return value registers
2185                                  rdi_temp, rbx_arg_base);
2186 
2187       Address conversion_addr      = RicochetFrame::frame_address(RicochetFrame::conversion_offset_in_bytes());
2188       Address saved_args_base_addr = RicochetFrame::frame_address(RicochetFrame::saved_args_base_offset_in_bytes());
2189 
2190       __ movptr(rbx_arg_base, saved_args_base_addr);
2191       RegisterOrConstant dest_slot = dest_slot_constant;
2192       if (dest_slot_constant == -1) {
2193         load_conversion_vminfo(_masm, rdi_temp, conversion_addr);
2194         dest_slot = rdi_temp;
2195       }
2196       // Store the result back into the argslot.
2197       // This code uses the interpreter calling sequence, in which the return value
2198       // is usually left in the TOS register, as defined by InterpreterMacroAssembler::pop.
2199       // There are certain irregularities with floating point values, which can be seen
2200       // in TemplateInterpreterGenerator::generate_return_entry_for.
2201       move_return_value(_masm, dest_type_constant, Address(rbx_arg_base, dest_slot, Interpreter::stackElementScale()));
2202 
2203       RicochetFrame::leave_ricochet_frame(_masm, rcx_recv, rbx_arg_base, rdx_temp);
2204       __ push(rdx_temp);  // repush the return PC
2205 
2206       // Load the final target and go.
2207       if (VerifyMethodHandles)  verify_method_handle(_masm, rcx_recv);
2208       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
2209       __ hlt(); // --------------------
2210       break;
2211     }
2212 
2213   case _adapter_opt_return_any:
2214     {
2215       if (VerifyMethodHandles)  RicochetFrame::verify_clean(_masm);
2216       Register rdi_conv = rdi_temp;
2217       assert_different_registers(rax, rdx,  // possibly live return value registers
2218                                  rdi_conv, rbx_temp);
2219 
2220       Address conversion_addr = RicochetFrame::frame_address(RicochetFrame::conversion_offset_in_bytes());
2221       load_conversion_dest_type(_masm, rdi_conv, conversion_addr);
2222       __ lea(rbx_temp, ExternalAddress((address) &_adapter_return_handlers[0]));
2223       __ movptr(rbx_temp, Address(rbx_temp, rdi_conv, Address::times_ptr));
2224 
2225 #ifdef ASSERT
2226       { Label L_badconv;
2227         __ testptr(rbx_temp, rbx_temp);
2228         __ jccb(Assembler::zero, L_badconv);
2229         __ jmp(rbx_temp);
2230         __ bind(L_badconv);
2231         __ stop("bad method handle return");
2232       }
2233 #else //ASSERT
2234       __ jmp(rbx_temp);
2235 #endif //ASSERT
2236       break;
2237     }
2238 
2239   case _adapter_opt_spread_0:
2240   case _adapter_opt_spread_1_ref:
2241   case _adapter_opt_spread_2_ref:
2242   case _adapter_opt_spread_3_ref:
2243   case _adapter_opt_spread_4_ref:
2244   case _adapter_opt_spread_5_ref:
2245   case _adapter_opt_spread_ref:
2246   case _adapter_opt_spread_byte:
2247   case _adapter_opt_spread_char:
2248   case _adapter_opt_spread_short:
2249   case _adapter_opt_spread_int:
2250   case _adapter_opt_spread_long:
2251   case _adapter_opt_spread_float:
2252   case _adapter_opt_spread_double:
2253     {
2254       // spread an array out into a group of arguments
2255       int length_constant = ek_adapter_opt_spread_count(ek);
2256       bool length_can_be_zero = (length_constant == 0);
2257       if (length_constant < 0) {
2258         // some adapters with variable length must handle the zero case
2259         if (!OptimizeMethodHandles ||
2260             ek_adapter_opt_spread_type(ek) != T_OBJECT)
2261           length_can_be_zero = true;
2262       }
2263 
2264       // find the address of the array argument
2265       __ movl(rax_argslot, rcx_amh_vmargslot);
2266       __ lea(rax_argslot, __ argument_address(rax_argslot));
2267 
2268       // grab another temp
2269       Register rsi_temp = rsi;
2270       { if (rsi_temp == saved_last_sp)  __ push(saved_last_sp); }
2271       // (preceding push must be done after argslot address is taken!)
2272 #define UNPUSH_RSI \
2273       { if (rsi_temp == saved_last_sp)  __ pop(saved_last_sp); }
2274 
2275       // arx_argslot points both to the array and to the first output arg
2276       vmarg = Address(rax_argslot, 0);
2277 
2278       // Get the array value.
2279       Register  rsi_array       = rsi_temp;
2280       Register  rdx_array_klass = rdx_temp;
2281       BasicType elem_type = ek_adapter_opt_spread_type(ek);
2282       int       elem_slots = type2size[elem_type];  // 1 or 2
2283       int       array_slots = 1;  // array is always a T_OBJECT
2284       int       length_offset   = arrayOopDesc::length_offset_in_bytes();
2285       int       elem0_offset    = arrayOopDesc::base_offset_in_bytes(elem_type);
2286       __ movptr(rsi_array, vmarg);
2287 
2288       Label L_array_is_empty, L_insert_arg_space, L_copy_args, L_args_done;
2289       if (length_can_be_zero) {
2290         // handle the null pointer case, if zero is allowed
2291         Label L_skip;
2292         if (length_constant < 0) {
2293           load_conversion_vminfo(_masm, rbx_temp, rcx_amh_conversion);
2294           __ testl(rbx_temp, rbx_temp);
2295           __ jcc(Assembler::notZero, L_skip);
2296         }
2297         __ testptr(rsi_array, rsi_array);
2298         __ jcc(Assembler::zero, L_array_is_empty);
2299         __ bind(L_skip);
2300       }
2301       __ null_check(rsi_array, oopDesc::klass_offset_in_bytes());
2302       __ load_klass(rdx_array_klass, rsi_array);
2303 
2304       // Check the array type.
2305       Register rbx_klass = rbx_temp;
2306       __ load_heap_oop(rbx_klass, rcx_amh_argument); // this is a Class object!
2307       load_klass_from_Class(_masm, rbx_klass);
2308 
2309       Label ok_array_klass, bad_array_klass, bad_array_length;
2310       __ check_klass_subtype(rdx_array_klass, rbx_klass, rdi_temp, ok_array_klass);
2311       // If we get here, the type check failed!
2312       __ jmp(bad_array_klass);
2313       __ BIND(ok_array_klass);
2314 
2315       // Check length.
2316       if (length_constant >= 0) {
2317         __ cmpl(Address(rsi_array, length_offset), length_constant);
2318       } else {
2319         Register rbx_vminfo = rbx_temp;
2320         load_conversion_vminfo(_masm, rbx_vminfo, rcx_amh_conversion);
2321         __ cmpl(rbx_vminfo, Address(rsi_array, length_offset));
2322       }
2323       __ jcc(Assembler::notEqual, bad_array_length);
2324 
2325       Register rdx_argslot_limit = rdx_temp;
2326 
2327       // Array length checks out.  Now insert any required stack slots.
2328       if (length_constant == -1) {
2329         // Form a pointer to the end of the affected region.
2330         __ lea(rdx_argslot_limit, Address(rax_argslot, Interpreter::stackElementSize));
2331         // 'stack_move' is negative number of words to insert
2332         // This number already accounts for elem_slots.
2333         Register rdi_stack_move = rdi_temp;
2334         load_stack_move(_masm, rdi_stack_move, rcx_recv, true);
2335         __ cmpptr(rdi_stack_move, 0);
2336         assert(stack_move_unit() < 0, "else change this comparison");
2337         __ jcc(Assembler::less, L_insert_arg_space);
2338         __ jcc(Assembler::equal, L_copy_args);
2339         // single argument case, with no array movement
2340         __ BIND(L_array_is_empty);
2341         remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2342                          rax_argslot, rbx_temp, rdx_temp);
2343         __ jmp(L_args_done);  // no spreading to do
2344         __ BIND(L_insert_arg_space);
2345         // come here in the usual case, stack_move < 0 (2 or more spread arguments)
2346         Register rsi_temp = rsi_array;  // spill this
2347         insert_arg_slots(_masm, rdi_stack_move,
2348                          rax_argslot, rbx_temp, rsi_temp);
2349         // reload the array since rsi was killed
2350         // reload from rdx_argslot_limit since rax_argslot is now decremented
2351         __ movptr(rsi_array, Address(rdx_argslot_limit, -Interpreter::stackElementSize));
2352       } else if (length_constant >= 1) {
2353         int new_slots = (length_constant * elem_slots) - array_slots;
2354         insert_arg_slots(_masm, new_slots * stack_move_unit(),
2355                          rax_argslot, rbx_temp, rdx_temp);
2356       } else if (length_constant == 0) {
2357         __ BIND(L_array_is_empty);
2358         remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2359                          rax_argslot, rbx_temp, rdx_temp);
2360       } else {
2361         ShouldNotReachHere();
2362       }
2363 
2364       // Copy from the array to the new slots.
2365       // Note: Stack change code preserves integrity of rax_argslot pointer.
2366       // So even after slot insertions, rax_argslot still points to first argument.
2367       // Beware:  Arguments that are shallow on the stack are deep in the array,
2368       // and vice versa.  So a downward-growing stack (the usual) has to be copied
2369       // elementwise in reverse order from the source array.
2370       __ BIND(L_copy_args);
2371       if (length_constant == -1) {
2372         // [rax_argslot, rdx_argslot_limit) is the area we are inserting into.
2373         // Array element [0] goes at rdx_argslot_limit[-wordSize].
2374         Register rsi_source = rsi_array;
2375         __ lea(rsi_source, Address(rsi_array, elem0_offset));
2376         Register rdx_fill_ptr = rdx_argslot_limit;
2377         Label loop;
2378         __ BIND(loop);
2379         __ addptr(rdx_fill_ptr, -Interpreter::stackElementSize * elem_slots);
2380         move_typed_arg(_masm, elem_type, true,
2381                        Address(rdx_fill_ptr, 0), Address(rsi_source, 0),
2382                        rbx_temp, rdi_temp);
2383         __ addptr(rsi_source, type2aelembytes(elem_type));
2384         __ cmpptr(rdx_fill_ptr, rax_argslot);
2385         __ jcc(Assembler::above, loop);
2386       } else if (length_constant == 0) {
2387         // nothing to copy
2388       } else {
2389         int elem_offset = elem0_offset;
2390         int slot_offset = length_constant * Interpreter::stackElementSize;
2391         for (int index = 0; index < length_constant; index++) {
2392           slot_offset -= Interpreter::stackElementSize * elem_slots;  // fill backward
2393           move_typed_arg(_masm, elem_type, true,
2394                          Address(rax_argslot, slot_offset), Address(rsi_array, elem_offset),
2395                          rbx_temp, rdi_temp);
2396           elem_offset += type2aelembytes(elem_type);
2397         }
2398       }
2399       __ BIND(L_args_done);
2400 
2401       // Arguments are spread.  Move to next method handle.
2402       UNPUSH_RSI;
2403       __ load_heap_oop(rcx_recv, rcx_mh_vmtarget);
2404       __ jump_to_method_handle_entry(rcx_recv, rdx_temp);
2405 
2406       __ bind(bad_array_klass);
2407       UNPUSH_RSI;
2408       assert(!vmarg.uses(rarg2_required), "must be different registers");
2409       __ load_heap_oop( rarg2_required, Address(rdx_array_klass, java_mirror_offset));  // required type
2410       __ movptr(        rarg1_actual,   vmarg);                                         // bad array
2411       __ movl(          rarg0_code,     (int) Bytecodes::_aaload);                      // who is complaining?
2412       __ jump(ExternalAddress(from_interpreted_entry(_raise_exception)));
2413 
2414       __ bind(bad_array_length);
2415       UNPUSH_RSI;
2416       assert(!vmarg.uses(rarg2_required), "must be different registers");
2417       __ mov(    rarg2_required, rcx_recv);                       // AMH requiring a certain length
2418       __ movptr( rarg1_actual,   vmarg);                          // bad array
2419       __ movl(   rarg0_code,     (int) Bytecodes::_arraylength);  // who is complaining?
2420       __ jump(ExternalAddress(from_interpreted_entry(_raise_exception)));
2421 #undef UNPUSH_RSI
2422 
2423       break;
2424     }
2425 
2426   default:
2427     // do not require all platforms to recognize all adapter types
2428     __ nop();
2429     return;
2430   }
2431   BLOCK_COMMENT(err_msg("} Entry %s", entry_name(ek)));
2432   __ hlt();
2433 
2434   address me_cookie = MethodHandleEntry::start_compiled_entry(_masm, interp_entry);
2435   __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
2436 
2437   init_entry(ek, MethodHandleEntry::finish_compiled_entry(_masm, me_cookie));
2438 }