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