1 /*
   2  * Copyright (c) 2008, 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 "memory/allocation.inline.hpp"
  28 #include "prims/methodHandles.hpp"
  29 
  30 #define __ _masm->
  31 
  32 #ifdef PRODUCT
  33 #define BLOCK_COMMENT(str) /* nothing */
  34 #else
  35 #define BLOCK_COMMENT(str) __ block_comment(str)
  36 #endif
  37 
  38 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
  39 
  40 address MethodHandleEntry::start_compiled_entry(MacroAssembler* _masm,
  41                                                 address interpreted_entry) {
  42   // Just before the actual machine code entry point, allocate space
  43   // for a MethodHandleEntry::Data record, so that we can manage everything
  44   // from one base pointer.
  45   __ align(wordSize);
  46   address target = __ pc() + sizeof(Data);
  47   while (__ pc() < target) {
  48     __ nop();
  49     __ align(wordSize);
  50   }
  51 
  52   MethodHandleEntry* me = (MethodHandleEntry*) __ pc();
  53   me->set_end_address(__ pc());         // set a temporary end_address
  54   me->set_from_interpreted_entry(interpreted_entry);
  55   me->set_type_checking_entry(NULL);
  56 
  57   return (address) me;
  58 }
  59 
  60 MethodHandleEntry* MethodHandleEntry::finish_compiled_entry(MacroAssembler* _masm,
  61                                                 address start_addr) {
  62   MethodHandleEntry* me = (MethodHandleEntry*) start_addr;
  63   assert(me->end_address() == start_addr, "valid ME");
  64 
  65   // Fill in the real end_address:
  66   __ align(wordSize);
  67   me->set_end_address(__ pc());
  68 
  69   return me;
  70 }
  71 
  72 // stack walking support
  73 
  74 frame MethodHandles::ricochet_frame_sender(const frame& fr, RegisterMap *map) {
  75   //RicochetFrame* f = RicochetFrame::from_frame(fr);
  76   // Cf. is_interpreted_frame path of frame::sender
  77   intptr_t* younger_sp = fr.sp();
  78   intptr_t* sp         = fr.sender_sp();
  79   map->make_integer_regs_unsaved();
  80   map->shift_window(sp, younger_sp);
  81   bool this_frame_adjusted_stack = true;  // I5_savedSP is live in this RF
  82   return frame(sp, younger_sp, this_frame_adjusted_stack);
  83 }
  84 
  85 void MethodHandles::ricochet_frame_oops_do(const frame& fr, OopClosure* blk, const RegisterMap* reg_map) {
  86   ResourceMark rm;
  87   RicochetFrame* f = RicochetFrame::from_frame(fr);
  88 
  89   // pick up the argument type descriptor:
  90   Thread* thread = Thread::current();
  91   Handle cookie(thread, f->compute_saved_args_layout(true, true));
  92 
  93   // process fixed part
  94   blk->do_oop((oop*)f->saved_target_addr());
  95   blk->do_oop((oop*)f->saved_args_layout_addr());
  96 
  97   // process variable arguments:
  98   if (cookie.is_null())  return;  // no arguments to describe
  99 
 100   // the cookie is actually the invokeExact method for my target
 101   // his argument signature is what I'm interested in
 102   assert(cookie->is_method(), "");
 103   methodHandle invoker(thread, methodOop(cookie()));
 104   assert(invoker->name() == vmSymbols::invokeExact_name(), "must be this kind of method");
 105   assert(!invoker->is_static(), "must have MH argument");
 106   int slot_count = invoker->size_of_parameters();
 107   assert(slot_count >= 1, "must include 'this'");
 108   intptr_t* base = f->saved_args_base();
 109   intptr_t* retval = NULL;
 110   if (f->has_return_value_slot())
 111     retval = f->return_value_slot_addr();
 112   int slot_num = slot_count - 1;
 113   intptr_t* loc = &base[slot_num];
 114   //blk->do_oop((oop*) loc);   // original target, which is irrelevant
 115   int arg_num = 0;
 116   for (SignatureStream ss(invoker->signature()); !ss.is_done(); ss.next()) {
 117     if (ss.at_return_type())  continue;
 118     BasicType ptype = ss.type();
 119     if (ptype == T_ARRAY)  ptype = T_OBJECT; // fold all refs to T_OBJECT
 120     assert(ptype >= T_BOOLEAN && ptype <= T_OBJECT, "not array or void");
 121     slot_num -= type2size[ptype];
 122     loc = &base[slot_num];
 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 // Ricochet Frames
 131 const Register MethodHandles::RicochetFrame::L1_continuation      = L1;
 132 const Register MethodHandles::RicochetFrame::L2_saved_target      = L2;
 133 const Register MethodHandles::RicochetFrame::L3_saved_args_layout = L3;
 134 const Register MethodHandles::RicochetFrame::L4_saved_args_base   = L4; // cf. Gargs = G4
 135 const Register MethodHandles::RicochetFrame::L5_conversion        = L5;
 136 #ifdef ASSERT
 137 const Register MethodHandles::RicochetFrame::L0_magic_number_1    = L0;
 138 #endif //ASSERT
 139 
 140 oop MethodHandles::RicochetFrame::compute_saved_args_layout(bool read_cache, bool write_cache) {
 141   if (read_cache) {
 142     oop cookie = saved_args_layout();
 143     if (cookie != NULL)  return cookie;
 144   }
 145   oop target = saved_target();
 146   oop mtype  = java_lang_invoke_MethodHandle::type(target);
 147   oop mtform = java_lang_invoke_MethodType::form(mtype);
 148   oop cookie = java_lang_invoke_MethodTypeForm::vmlayout(mtform);
 149   if (write_cache)  {
 150     (*saved_args_layout_addr()) = cookie;
 151   }
 152   return cookie;
 153 }
 154 
 155 void MethodHandles::RicochetFrame::generate_ricochet_blob(MacroAssembler* _masm,
 156                                                           // output params:
 157                                                           int* bounce_offset,
 158                                                           int* exception_offset,
 159                                                           int* frame_size_in_words) {
 160   (*frame_size_in_words) = RicochetFrame::frame_size_in_bytes() / wordSize;
 161 
 162   address start = __ pc();
 163 
 164 #ifdef ASSERT
 165   __ illtrap(0); __ illtrap(0); __ illtrap(0);
 166   // here's a hint of something special:
 167   __ set(MAGIC_NUMBER_1, G0);
 168   __ set(MAGIC_NUMBER_2, G0);
 169 #endif //ASSERT
 170   __ illtrap(0);  // not reached
 171 
 172   // Return values are in registers.
 173   // L1_continuation contains a cleanup continuation we must return
 174   // to.
 175 
 176   (*bounce_offset) = __ pc() - start;
 177   BLOCK_COMMENT("ricochet_blob.bounce");
 178 
 179   if (VerifyMethodHandles)  RicochetFrame::verify_clean(_masm);
 180   trace_method_handle(_masm, "ricochet_blob.bounce");
 181 
 182   __ JMP(L1_continuation, 0);
 183   __ delayed()->nop();
 184   __ illtrap(0);
 185 
 186   DEBUG_ONLY(__ set(MAGIC_NUMBER_2, G0));
 187 
 188   (*exception_offset) = __ pc() - start;
 189   BLOCK_COMMENT("ricochet_blob.exception");
 190 
 191   // compare this to Interpreter::rethrow_exception_entry, which is parallel code
 192   // for example, see TemplateInterpreterGenerator::generate_throw_exception
 193   // Live registers in:
 194   //   Oexception  (O0): exception
 195   //   Oissuing_pc (O1): return address/pc that threw exception (ignored, always equal to bounce addr)
 196   __ verify_oop(Oexception);
 197 
 198   // Take down the frame.
 199 
 200   // Cf. InterpreterMacroAssembler::remove_activation.
 201   leave_ricochet_frame(_masm, /*recv_reg=*/ noreg, I5_savedSP, I7);
 202 
 203   // We are done with this activation frame; find out where to go next.
 204   // The continuation point will be an exception handler, which expects
 205   // the following registers set up:
 206   //
 207   // Oexception: exception
 208   // Oissuing_pc: the local call that threw exception
 209   // Other On: garbage
 210   // In/Ln:  the contents of the caller's register window
 211   //
 212   // We do the required restore at the last possible moment, because we
 213   // need to preserve some state across a runtime call.
 214   // (Remember that the caller activation is unknown--it might not be
 215   // interpreted, so things like Lscratch are useless in the caller.)
 216   __ mov(Oexception,  Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
 217   __ add(I7, frame::pc_return_offset, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
 218   __ call_VM_leaf(L7_thread_cache,
 219                   CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
 220                   G2_thread, Oissuing_pc->after_save());
 221 
 222   // The caller's SP was adjusted upon method entry to accomodate
 223   // the callee's non-argument locals. Undo that adjustment.
 224   __ JMP(O0, 0);                         // return exception handler in caller
 225   __ delayed()->restore(I5_savedSP, G0, SP);
 226 
 227   // (same old exception object is already in Oexception; see above)
 228   // Note that an "issuing PC" is actually the next PC after the call
 229 }
 230 
 231 void MethodHandles::RicochetFrame::enter_ricochet_frame(MacroAssembler* _masm,
 232                                                         Register recv_reg,
 233                                                         Register argv_reg,
 234                                                         address return_handler) {
 235   // does not include the __ save()
 236   assert(argv_reg == Gargs, "");
 237   Address G3_mh_vmtarget(   recv_reg, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes());
 238   Address G3_amh_conversion(recv_reg, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
 239 
 240   // Create the RicochetFrame.
 241   // Unlike on x86 we can store all required information in local
 242   // registers.
 243   BLOCK_COMMENT("push RicochetFrame {");
 244   __ set(ExternalAddress(return_handler),          L1_continuation);
 245   __ load_heap_oop(G3_mh_vmtarget,                 L2_saved_target);
 246   __ mov(G0,                                       L3_saved_args_layout);
 247   __ mov(Gargs,                                    L4_saved_args_base);
 248   __ lduw(G3_amh_conversion,                       L5_conversion);  // 32-bit field
 249   // I5, I6, I7 are already set up
 250   DEBUG_ONLY(__ set((int32_t) MAGIC_NUMBER_1,      L0_magic_number_1));
 251   BLOCK_COMMENT("} RicochetFrame");
 252 }
 253 
 254 void MethodHandles::RicochetFrame::leave_ricochet_frame(MacroAssembler* _masm,
 255                                                         Register recv_reg,
 256                                                         Register new_sp_reg,
 257                                                         Register sender_pc_reg) {
 258   assert(new_sp_reg == I5_savedSP, "exact_sender_sp already in place");
 259   assert(sender_pc_reg == I7, "in a fixed place");
 260   // does not include the __ ret() & __ restore()
 261   assert_different_registers(recv_reg, new_sp_reg, sender_pc_reg);
 262   // Take down the frame.
 263   // Cf. InterpreterMacroAssembler::remove_activation.
 264   BLOCK_COMMENT("end_ricochet_frame {");
 265   if (recv_reg->is_valid())
 266     __ mov(L2_saved_target, recv_reg);
 267   BLOCK_COMMENT("} end_ricochet_frame");
 268 }
 269 
 270 // Emit code to verify that FP is pointing at a valid ricochet frame.
 271 #ifdef ASSERT
 272 enum {
 273   ARG_LIMIT = 255, SLOP = 45,
 274   // use this parameter for checking for garbage stack movements:
 275   UNREASONABLE_STACK_MOVE = (ARG_LIMIT + SLOP)
 276   // the slop defends against false alarms due to fencepost errors
 277 };
 278 
 279 void MethodHandles::RicochetFrame::verify_clean(MacroAssembler* _masm) {
 280   // The stack should look like this:
 281   //    ... keep1 | dest=42 | keep2 | magic | handler | magic | recursive args | [RF]
 282   // Check various invariants.
 283 
 284   Register O7_temp = O7, O5_temp = O5;
 285 
 286   Label L_ok_1, L_ok_2, L_ok_3, L_ok_4;
 287   BLOCK_COMMENT("verify_clean {");
 288   // Magic numbers must check out:
 289   __ set((int32_t) MAGIC_NUMBER_1, O7_temp);
 290   __ cmp_and_br_short(O7_temp, L0_magic_number_1, Assembler::equal, Assembler::pt, L_ok_1);
 291   __ stop("damaged ricochet frame: MAGIC_NUMBER_1 not found");
 292 
 293   __ BIND(L_ok_1);
 294 
 295   // Arguments pointer must look reasonable:
 296 #ifdef _LP64
 297   Register FP_temp = O5_temp;
 298   __ add(FP, STACK_BIAS, FP_temp);
 299 #else
 300   Register FP_temp = FP;
 301 #endif
 302   __ cmp_and_brx_short(L4_saved_args_base, FP_temp, Assembler::greaterEqualUnsigned, Assembler::pt, L_ok_2);
 303   __ stop("damaged ricochet frame: L4 < FP");
 304 
 305   __ BIND(L_ok_2);
 306   // Disable until we decide on it's fate
 307   // __ sub(L4_saved_args_base, UNREASONABLE_STACK_MOVE * Interpreter::stackElementSize, O7_temp);
 308   // __ cmp(O7_temp, FP_temp);
 309   // __ br(Assembler::lessEqualUnsigned, false, Assembler::pt, L_ok_3);
 310   // __ delayed()->nop();
 311   // __ stop("damaged ricochet frame: (L4 - UNREASONABLE_STACK_MOVE) > FP");
 312 
 313   __ BIND(L_ok_3);
 314   extract_conversion_dest_type(_masm, L5_conversion, O7_temp);
 315   __ cmp_and_br_short(O7_temp, T_VOID, Assembler::equal, Assembler::pt, L_ok_4);
 316   extract_conversion_vminfo(_masm, L5_conversion, O5_temp);
 317   __ ld_ptr(L4_saved_args_base, __ argument_offset(O5_temp, O5_temp), O7_temp);
 318   assert(Assembler::is_simm13(RETURN_VALUE_PLACEHOLDER), "must be simm13");
 319   __ cmp_and_brx_short(O7_temp, (int32_t) RETURN_VALUE_PLACEHOLDER, Assembler::equal, Assembler::pt, L_ok_4);
 320   __ stop("damaged ricochet frame: RETURN_VALUE_PLACEHOLDER not found");
 321   __ BIND(L_ok_4);
 322   BLOCK_COMMENT("} verify_clean");
 323 }
 324 #endif //ASSERT
 325 
 326 void MethodHandles::load_klass_from_Class(MacroAssembler* _masm, Register klass_reg, Register temp_reg, Register temp2_reg) {
 327   if (VerifyMethodHandles)
 328     verify_klass(_masm, klass_reg, SystemDictionaryHandles::Class_klass(), temp_reg, temp2_reg,
 329                  "AMH argument is a Class");
 330   __ load_heap_oop(Address(klass_reg, java_lang_Class::klass_offset_in_bytes()), klass_reg);
 331 }
 332 
 333 void MethodHandles::load_conversion_vminfo(MacroAssembler* _masm, Address conversion_field_addr, Register reg) {
 334   assert(CONV_VMINFO_SHIFT == 0, "preshifted");
 335   assert(CONV_VMINFO_MASK == right_n_bits(BitsPerByte), "else change type of following load");
 336   __ ldub(conversion_field_addr.plus_disp(BytesPerInt - 1), reg);
 337 }
 338 
 339 void MethodHandles::extract_conversion_vminfo(MacroAssembler* _masm, Register conversion_field_reg, Register reg) {
 340   assert(CONV_VMINFO_SHIFT == 0, "preshifted");
 341   __ and3(conversion_field_reg, CONV_VMINFO_MASK, reg);
 342 }
 343 
 344 void MethodHandles::extract_conversion_dest_type(MacroAssembler* _masm, Register conversion_field_reg, Register reg) {
 345   __ srl(conversion_field_reg, CONV_DEST_TYPE_SHIFT, reg);
 346   __ and3(reg, 0x0F, reg);
 347 }
 348 
 349 void MethodHandles::load_stack_move(MacroAssembler* _masm,
 350                                     Address G3_amh_conversion,
 351                                     Register stack_move_reg) {
 352   BLOCK_COMMENT("load_stack_move {");
 353   __ ldsw(G3_amh_conversion, stack_move_reg);
 354   __ sra(stack_move_reg, CONV_STACK_MOVE_SHIFT, stack_move_reg);
 355 #ifdef ASSERT
 356   if (VerifyMethodHandles) {
 357     Label L_ok, L_bad;
 358     int32_t stack_move_limit = 0x0800;  // extra-large
 359     __ cmp_and_br_short(stack_move_reg, stack_move_limit, Assembler::greaterEqual, Assembler::pn, L_bad);
 360     __ cmp(stack_move_reg, -stack_move_limit);
 361     __ br(Assembler::greater, false, Assembler::pt, L_ok);
 362     __ delayed()->nop();
 363     __ BIND(L_bad);
 364     __ stop("load_stack_move of garbage value");
 365     __ BIND(L_ok);
 366   }
 367 #endif
 368   BLOCK_COMMENT("} load_stack_move");
 369 }
 370 
 371 #ifdef ASSERT
 372 void MethodHandles::RicochetFrame::verify() const {
 373   assert(magic_number_1() == MAGIC_NUMBER_1, "");
 374   if (!Universe::heap()->is_gc_active()) {
 375     if (saved_args_layout() != NULL) {
 376       assert(saved_args_layout()->is_method(), "must be valid oop");
 377     }
 378     if (saved_target() != NULL) {
 379       assert(java_lang_invoke_MethodHandle::is_instance(saved_target()), "checking frame value");
 380     }
 381   }
 382   int conv_op = adapter_conversion_op(conversion());
 383   assert(conv_op == java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS ||
 384          conv_op == java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS ||
 385          conv_op == java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF,
 386          "must be a sane conversion");
 387   if (has_return_value_slot()) {
 388     assert(*return_value_slot_addr() == RETURN_VALUE_PLACEHOLDER, "");
 389   }
 390 }
 391 
 392 void MethodHandles::verify_argslot(MacroAssembler* _masm, Register argslot_reg, Register temp_reg, const char* error_message) {
 393   // Verify that argslot lies within (Gargs, FP].
 394   Label L_ok, L_bad;
 395   BLOCK_COMMENT("verify_argslot {");
 396   __ cmp_and_brx_short(Gargs, argslot_reg, Assembler::greaterUnsigned, Assembler::pn, L_bad);
 397   __ add(FP, STACK_BIAS, temp_reg);  // STACK_BIAS is zero on !_LP64
 398   __ cmp_and_brx_short(argslot_reg, temp_reg, Assembler::lessEqualUnsigned, Assembler::pt, L_ok);
 399   __ BIND(L_bad);
 400   __ stop(error_message);
 401   __ BIND(L_ok);
 402   BLOCK_COMMENT("} verify_argslot");
 403 }
 404 
 405 void MethodHandles::verify_argslots(MacroAssembler* _masm,
 406                                     RegisterOrConstant arg_slots,
 407                                     Register arg_slot_base_reg,
 408                                     Register temp_reg,
 409                                     Register temp2_reg,
 410                                     bool negate_argslots,
 411                                     const char* error_message) {
 412   // Verify that [argslot..argslot+size) lies within (Gargs, FP).
 413   Label L_ok, L_bad;
 414   BLOCK_COMMENT("verify_argslots {");
 415   if (negate_argslots) {
 416     if (arg_slots.is_constant()) {
 417       arg_slots = -1 * arg_slots.as_constant();
 418     } else {
 419       __ neg(arg_slots.as_register(), temp_reg);
 420       arg_slots = temp_reg;
 421     }
 422   }
 423   __ add(arg_slot_base_reg, __ argument_offset(arg_slots, temp_reg), temp_reg);
 424   __ add(FP, STACK_BIAS, temp2_reg);  // STACK_BIAS is zero on !_LP64
 425   __ cmp_and_brx_short(temp_reg, temp2_reg, Assembler::greaterUnsigned, Assembler::pn, L_bad);
 426   // Gargs points to the first word so adjust by BytesPerWord
 427   __ add(arg_slot_base_reg, BytesPerWord, temp_reg);
 428   __ cmp_and_brx_short(Gargs, temp_reg, Assembler::lessEqualUnsigned, Assembler::pt, L_ok);
 429   __ BIND(L_bad);
 430   __ stop(error_message);
 431   __ BIND(L_ok);
 432   BLOCK_COMMENT("} verify_argslots");
 433 }
 434 
 435 // Make sure that arg_slots has the same sign as the given direction.
 436 // If (and only if) arg_slots is a assembly-time constant, also allow it to be zero.
 437 void MethodHandles::verify_stack_move(MacroAssembler* _masm,
 438                                       RegisterOrConstant arg_slots, int direction) {
 439   enum { UNREASONABLE_STACK_MOVE = 256 * 4 };  // limit of 255 arguments
 440   bool allow_zero = arg_slots.is_constant();
 441   if (direction == 0) { direction = +1; allow_zero = true; }
 442   assert(stack_move_unit() == -1, "else add extra checks here");
 443   if (arg_slots.is_register()) {
 444     Label L_ok, L_bad;
 445     BLOCK_COMMENT("verify_stack_move {");
 446     // __ btst(-stack_move_unit() - 1, arg_slots.as_register());  // no need
 447     // __ br(Assembler::notZero, false, Assembler::pn, L_bad);
 448     // __ delayed()->nop();
 449     __ cmp(arg_slots.as_register(), (int32_t) NULL_WORD);
 450     if (direction > 0) {
 451       __ br(allow_zero ? Assembler::less : Assembler::lessEqual, false, Assembler::pn, L_bad);
 452       __ delayed()->nop();
 453       __ cmp(arg_slots.as_register(), (int32_t) UNREASONABLE_STACK_MOVE);
 454       __ br(Assembler::less, false, Assembler::pn, L_ok);
 455       __ delayed()->nop();
 456     } else {
 457       __ br(allow_zero ? Assembler::greater : Assembler::greaterEqual, false, Assembler::pn, L_bad);
 458       __ delayed()->nop();
 459       __ cmp(arg_slots.as_register(), (int32_t) -UNREASONABLE_STACK_MOVE);
 460       __ br(Assembler::greater, false, Assembler::pn, L_ok);
 461       __ delayed()->nop();
 462     }
 463     __ BIND(L_bad);
 464     if (direction > 0)
 465       __ stop("assert arg_slots > 0");
 466     else
 467       __ stop("assert arg_slots < 0");
 468     __ BIND(L_ok);
 469     BLOCK_COMMENT("} verify_stack_move");
 470   } else {
 471     intptr_t size = arg_slots.as_constant();
 472     if (direction < 0)  size = -size;
 473     assert(size >= 0, "correct direction of constant move");
 474     assert(size < UNREASONABLE_STACK_MOVE, "reasonable size of constant move");
 475   }
 476 }
 477 
 478 void MethodHandles::verify_klass(MacroAssembler* _masm,
 479                                  Register obj_reg, KlassHandle klass,
 480                                  Register temp_reg, Register temp2_reg,
 481                                  const char* error_message) {
 482   oop* klass_addr = klass.raw_value();
 483   assert(klass_addr >= SystemDictionaryHandles::Object_klass().raw_value() &&
 484          klass_addr <= SystemDictionaryHandles::Long_klass().raw_value(),
 485          "must be one of the SystemDictionaryHandles");
 486   Label L_ok, L_bad;
 487   BLOCK_COMMENT("verify_klass {");
 488   __ verify_oop(obj_reg);
 489   __ br_null_short(obj_reg, Assembler::pn, L_bad);
 490   __ load_klass(obj_reg, temp_reg);
 491   __ set(ExternalAddress(klass_addr), temp2_reg);
 492   __ ld_ptr(Address(temp2_reg, 0), temp2_reg);
 493   __ cmp_and_brx_short(temp_reg, temp2_reg, Assembler::equal, Assembler::pt, L_ok);
 494   intptr_t super_check_offset = klass->super_check_offset();
 495   __ ld_ptr(Address(temp_reg, super_check_offset), temp_reg);
 496   __ set(ExternalAddress(klass_addr), temp2_reg);
 497   __ ld_ptr(Address(temp2_reg, 0), temp2_reg);
 498   __ cmp_and_brx_short(temp_reg, temp2_reg, Assembler::equal, Assembler::pt, L_ok);
 499   __ BIND(L_bad);
 500   __ stop(error_message);
 501   __ BIND(L_ok);
 502   BLOCK_COMMENT("} verify_klass");
 503 }
 504 #endif // ASSERT
 505 
 506 
 507 void MethodHandles::jump_from_method_handle(MacroAssembler* _masm, Register method, Register target, Register temp) {
 508   assert(method == G5_method, "interpreter calling convention");
 509   __ verify_oop(method);
 510   __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_interpreted_offset()), target);
 511   if (JvmtiExport::can_post_interpreter_events()) {
 512     // JVMTI events, such as single-stepping, are implemented partly by avoiding running
 513     // compiled code in threads for which the event is enabled.  Check here for
 514     // interp_only_mode if these events CAN be enabled.
 515     __ verify_thread();
 516     Label skip_compiled_code;
 517 
 518     const Address interp_only(G2_thread, JavaThread::interp_only_mode_offset());
 519     __ ld(interp_only, temp);
 520     __ tst(temp);
 521     __ br(Assembler::notZero, true, Assembler::pn, skip_compiled_code);
 522     __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), target);
 523     __ bind(skip_compiled_code);
 524   }
 525   __ jmp(target, 0);
 526   __ delayed()->nop();
 527 }
 528 
 529 
 530 // Code generation
 531 address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm) {
 532   // I5_savedSP/O5_savedSP: sender SP (must preserve)
 533   // G4 (Gargs): incoming argument list (must preserve)
 534   // G5_method:  invoke methodOop
 535   // G3_method_handle: receiver method handle (must load from sp[MethodTypeForm.vmslots])
 536   // O0, O1, O2, O3, O4: garbage temps, blown away
 537   Register O0_mtype   = O0;
 538   Register O1_scratch = O1;
 539   Register O2_scratch = O2;
 540   Register O3_scratch = O3;
 541   Register O4_argslot = O4;
 542   Register O4_argbase = O4;
 543 
 544   // emit WrongMethodType path first, to enable back-branch from main path
 545   Label wrong_method_type;
 546   __ bind(wrong_method_type);
 547   Label invoke_generic_slow_path;
 548   assert(methodOopDesc::intrinsic_id_size_in_bytes() == sizeof(u1), "");;
 549   __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
 550   __ cmp(O1_scratch, (int) vmIntrinsics::_invokeExact);
 551   __ brx(Assembler::notEqual, false, Assembler::pt, invoke_generic_slow_path);
 552   __ delayed()->nop();
 553   __ mov(O0_mtype, G5_method_type);  // required by throw_WrongMethodType
 554   __ mov(G3_method_handle, G3_method_handle);  // already in this register
 555   // O0 will be filled in with JavaThread in stub
 556   __ jump_to(AddressLiteral(StubRoutines::throw_WrongMethodTypeException_entry()), O3_scratch);
 557   __ delayed()->nop();
 558 
 559   // here's where control starts out:
 560   __ align(CodeEntryAlignment);
 561   address entry_point = __ pc();
 562 
 563   // fetch the MethodType from the method handle
 564   // FIXME: Interpreter should transmit pre-popped stack pointer, to locate base of arg list.
 565   // This would simplify several touchy bits of code.
 566   // See 6984712: JSR 292 method handle calls need a clean argument base pointer
 567   {
 568     Register tem = G5_method;
 569     for (jint* pchase = methodOopDesc::method_type_offsets_chain(); (*pchase) != -1; pchase++) {
 570       __ ld_ptr(Address(tem, *pchase), O0_mtype);
 571       tem = O0_mtype;          // in case there is another indirection
 572     }
 573   }
 574 
 575   // given the MethodType, find out where the MH argument is buried
 576   __ load_heap_oop(Address(O0_mtype,   __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes,        O1_scratch)), O4_argslot);
 577   __ ldsw(         Address(O4_argslot, __ delayed_value(java_lang_invoke_MethodTypeForm::vmslots_offset_in_bytes, O1_scratch)), O4_argslot);
 578   __ add(__ argument_address(O4_argslot, O4_argslot, 1), O4_argbase);
 579   // Note: argument_address uses its input as a scratch register!
 580   Address mh_receiver_slot_addr(O4_argbase, -Interpreter::stackElementSize);
 581   __ ld_ptr(mh_receiver_slot_addr, G3_method_handle);
 582 
 583   trace_method_handle(_masm, "invokeExact");
 584 
 585   __ check_method_handle_type(O0_mtype, G3_method_handle, O1_scratch, wrong_method_type);
 586 
 587   // Nobody uses the MH receiver slot after this.  Make sure.
 588   DEBUG_ONLY(__ set((int32_t) 0x999999, O1_scratch); __ st_ptr(O1_scratch, mh_receiver_slot_addr));
 589 
 590   __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
 591 
 592   // for invokeGeneric (only), apply argument and result conversions on the fly
 593   __ bind(invoke_generic_slow_path);
 594 #ifdef ASSERT
 595   if (VerifyMethodHandles) {
 596     Label L;
 597     __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
 598     __ cmp(O1_scratch, (int) vmIntrinsics::_invokeGeneric);
 599     __ brx(Assembler::equal, false, Assembler::pt, L);
 600     __ delayed()->nop();
 601     __ stop("bad methodOop::intrinsic_id");
 602     __ bind(L);
 603   }
 604 #endif //ASSERT
 605 
 606   // make room on the stack for another pointer:
 607   insert_arg_slots(_masm, 2 * stack_move_unit(), O4_argbase, O1_scratch, O2_scratch, O3_scratch);
 608   // load up an adapter from the calling type (Java weaves this)
 609   Register O2_form    = O2_scratch;
 610   Register O3_adapter = O3_scratch;
 611   __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes,               O1_scratch)), O2_form);
 612   __ load_heap_oop(Address(O2_form,  __ delayed_value(java_lang_invoke_MethodTypeForm::genericInvoker_offset_in_bytes, O1_scratch)), O3_adapter);
 613   __ verify_oop(O3_adapter);
 614   __ st_ptr(O3_adapter, Address(O4_argbase, 1 * Interpreter::stackElementSize));
 615   // As a trusted first argument, pass the type being called, so the adapter knows
 616   // the actual types of the arguments and return values.
 617   // (Generic invokers are shared among form-families of method-type.)
 618   __ st_ptr(O0_mtype,   Address(O4_argbase, 0 * Interpreter::stackElementSize));
 619   // FIXME: assert that O3_adapter is of the right method-type.
 620   __ mov(O3_adapter, G3_method_handle);
 621   trace_method_handle(_masm, "invokeGeneric");
 622   __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
 623 
 624   return entry_point;
 625 }
 626 
 627 // Workaround for C++ overloading nastiness on '0' for RegisterOrConstant.
 628 static RegisterOrConstant constant(int value) {
 629   return RegisterOrConstant(value);
 630 }
 631 
 632 static void load_vmargslot(MacroAssembler* _masm, Address vmargslot_addr, Register result) {
 633   __ ldsw(vmargslot_addr, result);
 634 }
 635 
 636 static RegisterOrConstant adjust_SP_and_Gargs_down_by_slots(MacroAssembler* _masm,
 637                                                             RegisterOrConstant arg_slots,
 638                                                             Register temp_reg, Register temp2_reg) {
 639   // Keep the stack pointer 2*wordSize aligned.
 640   const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
 641   if (arg_slots.is_constant()) {
 642     const int        offset = arg_slots.as_constant() << LogBytesPerWord;
 643     const int masked_offset = round_to(offset, 2 * BytesPerWord);
 644     const int masked_offset2 = (offset + 1*BytesPerWord) & ~TwoWordAlignmentMask;
 645     assert(masked_offset == masked_offset2, "must agree");
 646     __ sub(Gargs,        offset, Gargs);
 647     __ sub(SP,    masked_offset, SP   );
 648     return offset;
 649   } else {
 650 #ifdef ASSERT
 651     {
 652       Label L_ok;
 653       __ cmp_and_br_short(arg_slots.as_register(), 0, Assembler::greaterEqual, Assembler::pt, L_ok);
 654       __ stop("negative arg_slots");
 655       __ bind(L_ok);
 656     }
 657 #endif
 658     __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
 659     __ add( temp_reg,  1*BytesPerWord,       temp2_reg);
 660     __ andn(temp2_reg, TwoWordAlignmentMask, temp2_reg);
 661     __ sub(Gargs, temp_reg,  Gargs);
 662     __ sub(SP,    temp2_reg, SP   );
 663     return temp_reg;
 664   }
 665 }
 666 
 667 static RegisterOrConstant adjust_SP_and_Gargs_up_by_slots(MacroAssembler* _masm,
 668                                                           RegisterOrConstant arg_slots,
 669                                                           Register temp_reg, Register temp2_reg) {
 670   // Keep the stack pointer 2*wordSize aligned.
 671   const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
 672   if (arg_slots.is_constant()) {
 673     const int        offset = arg_slots.as_constant() << LogBytesPerWord;
 674     const int masked_offset = offset & ~TwoWordAlignmentMask;
 675     __ add(Gargs,        offset, Gargs);
 676     __ add(SP,    masked_offset, SP   );
 677     return offset;
 678   } else {
 679     __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
 680     __ andn(temp_reg, TwoWordAlignmentMask, temp2_reg);
 681     __ add(Gargs, temp_reg,  Gargs);
 682     __ add(SP,    temp2_reg, SP   );
 683     return temp_reg;
 684   }
 685 }
 686 
 687 // Helper to insert argument slots into the stack.
 688 // arg_slots must be a multiple of stack_move_unit() and < 0
 689 // argslot_reg is decremented to point to the new (shifted) location of the argslot
 690 // But, temp_reg ends up holding the original value of argslot_reg.
 691 void MethodHandles::insert_arg_slots(MacroAssembler* _masm,
 692                                      RegisterOrConstant arg_slots,
 693                                      Register argslot_reg,
 694                                      Register temp_reg, Register temp2_reg, Register temp3_reg) {
 695   // allow constant zero
 696   if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
 697     return;
 698 
 699   assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
 700                              (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
 701 
 702   BLOCK_COMMENT("insert_arg_slots {");
 703   if (VerifyMethodHandles)
 704     verify_argslot(_masm, argslot_reg, temp_reg, "insertion point must fall within current frame");
 705   if (VerifyMethodHandles)
 706     verify_stack_move(_masm, arg_slots, -1);
 707 
 708   // Make space on the stack for the inserted argument(s).
 709   // Then pull down everything shallower than argslot_reg.
 710   // The stacked return address gets pulled down with everything else.
 711   // That is, copy [sp, argslot) downward by -size words.  In pseudo-code:
 712   //   sp -= size;
 713   //   for (temp = sp + size; temp < argslot; temp++)
 714   //     temp[-size] = temp[0]
 715   //   argslot -= size;
 716 
 717   // offset is temp3_reg in case of arg_slots being a register.
 718   RegisterOrConstant offset = adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
 719   __ sub(Gargs, offset, temp_reg);  // source pointer for copy
 720 
 721   {
 722     Label loop;
 723     __ BIND(loop);
 724     // pull one word down each time through the loop
 725     __ ld_ptr(           Address(temp_reg, 0     ), temp2_reg);
 726     __ st_ptr(temp2_reg, Address(temp_reg, offset)           );
 727     __ add(temp_reg, wordSize, temp_reg);
 728     __ cmp_and_brx_short(temp_reg, argslot_reg, Assembler::lessUnsigned, Assembler::pt, loop);
 729   }
 730 
 731   // Now move the argslot down, to point to the opened-up space.
 732   __ add(argslot_reg, offset, argslot_reg);
 733   BLOCK_COMMENT("} insert_arg_slots");
 734 }
 735 
 736 
 737 // Helper to remove argument slots from the stack.
 738 // arg_slots must be a multiple of stack_move_unit() and > 0
 739 void MethodHandles::remove_arg_slots(MacroAssembler* _masm,
 740                                      RegisterOrConstant arg_slots,
 741                                      Register argslot_reg,
 742                                      Register temp_reg, Register temp2_reg, Register temp3_reg) {
 743   // allow constant zero
 744   if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
 745     return;
 746   assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
 747                              (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
 748 
 749   BLOCK_COMMENT("remove_arg_slots {");
 750   if (VerifyMethodHandles)
 751     verify_argslots(_masm, arg_slots, argslot_reg, temp_reg, temp2_reg, false,
 752                     "deleted argument(s) must fall within current frame");
 753   if (VerifyMethodHandles)
 754     verify_stack_move(_masm, arg_slots, +1);
 755 
 756   // Pull up everything shallower than argslot.
 757   // Then remove the excess space on the stack.
 758   // The stacked return address gets pulled up with everything else.
 759   // That is, copy [sp, argslot) upward by size words.  In pseudo-code:
 760   //   for (temp = argslot-1; temp >= sp; --temp)
 761   //     temp[size] = temp[0]
 762   //   argslot += size;
 763   //   sp += size;
 764 
 765   RegisterOrConstant offset = __ regcon_sll_ptr(arg_slots, LogBytesPerWord, temp3_reg);
 766   __ sub(argslot_reg, wordSize, temp_reg);  // source pointer for copy
 767 
 768   {
 769     Label L_loop;
 770     __ BIND(L_loop);
 771     // pull one word up each time through the loop
 772     __ ld_ptr(           Address(temp_reg, 0     ), temp2_reg);
 773     __ st_ptr(temp2_reg, Address(temp_reg, offset)           );
 774     __ sub(temp_reg, wordSize, temp_reg);
 775     __ cmp_and_brx_short(temp_reg, Gargs, Assembler::greaterEqualUnsigned, Assembler::pt, L_loop);
 776   }
 777 
 778   // And adjust the argslot address to point at the deletion point.
 779   __ add(argslot_reg, offset, argslot_reg);
 780 
 781   // We don't need the offset at this point anymore, just adjust SP and Gargs.
 782   (void) adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
 783 
 784   BLOCK_COMMENT("} remove_arg_slots");
 785 }
 786 
 787 // Helper to copy argument slots to the top of the stack.
 788 // The sequence starts with argslot_reg and is counted by slot_count
 789 // slot_count must be a multiple of stack_move_unit() and >= 0
 790 // This function blows the temps but does not change argslot_reg.
 791 void MethodHandles::push_arg_slots(MacroAssembler* _masm,
 792                                    Register argslot_reg,
 793                                    RegisterOrConstant slot_count,
 794                                    Register temp_reg, Register temp2_reg) {
 795   // allow constant zero
 796   if (slot_count.is_constant() && slot_count.as_constant() == 0)
 797     return;
 798   assert_different_registers(argslot_reg, temp_reg, temp2_reg,
 799                              (!slot_count.is_register() ? Gargs : slot_count.as_register()),
 800                              SP);
 801   assert(Interpreter::stackElementSize == wordSize, "else change this code");
 802 
 803   BLOCK_COMMENT("push_arg_slots {");
 804   if (VerifyMethodHandles)
 805     verify_stack_move(_masm, slot_count, 0);
 806 
 807   RegisterOrConstant offset = adjust_SP_and_Gargs_down_by_slots(_masm, slot_count, temp2_reg, temp_reg);
 808 
 809   if (slot_count.is_constant()) {
 810     for (int i = slot_count.as_constant() - 1; i >= 0; i--) {
 811       __ ld_ptr(          Address(argslot_reg, i * wordSize), temp_reg);
 812       __ st_ptr(temp_reg, Address(Gargs,       i * wordSize));
 813     }
 814   } else {
 815     Label L_plural, L_loop, L_break;
 816     // Emit code to dynamically check for the common cases, zero and one slot.
 817     __ cmp(slot_count.as_register(), (int32_t) 1);
 818     __ br(Assembler::greater, false, Assembler::pn, L_plural);
 819     __ delayed()->nop();
 820     __ br(Assembler::less, false, Assembler::pn, L_break);
 821     __ delayed()->nop();
 822     __ ld_ptr(          Address(argslot_reg, 0), temp_reg);
 823     __ st_ptr(temp_reg, Address(Gargs,       0));
 824     __ ba_short(L_break);
 825     __ BIND(L_plural);
 826 
 827     // Loop for 2 or more:
 828     //   top = &argslot[slot_count]
 829     //   while (top > argslot)  *(--Gargs) = *(--top)
 830     Register top_reg = temp_reg;
 831     __ add(argslot_reg, offset, top_reg);
 832     __ add(Gargs,       offset, Gargs  );  // move back up again so we can go down
 833     __ BIND(L_loop);
 834     __ sub(top_reg, wordSize, top_reg);
 835     __ sub(Gargs,   wordSize, Gargs  );
 836     __ ld_ptr(           Address(top_reg, 0), temp2_reg);
 837     __ st_ptr(temp2_reg, Address(Gargs,   0));
 838     __ cmp_and_brx_short(top_reg, argslot_reg, Assembler::greaterUnsigned, Assembler::pt, L_loop);
 839     __ BIND(L_break);
 840   }
 841   BLOCK_COMMENT("} push_arg_slots");
 842 }
 843 
 844 // in-place movement; no change to Gargs
 845 // blows temp_reg, temp2_reg
 846 void MethodHandles::move_arg_slots_up(MacroAssembler* _masm,
 847                                       Register bottom_reg,  // invariant
 848                                       Address  top_addr,    // can use temp_reg
 849                                       RegisterOrConstant positive_distance_in_slots,  // destroyed if register
 850                                       Register temp_reg, Register temp2_reg) {
 851   assert_different_registers(bottom_reg,
 852                              temp_reg, temp2_reg,
 853                              positive_distance_in_slots.register_or_noreg());
 854   BLOCK_COMMENT("move_arg_slots_up {");
 855   Label L_loop, L_break;
 856   Register top_reg = temp_reg;
 857   if (!top_addr.is_same_address(Address(top_reg, 0))) {
 858     __ add(top_addr, top_reg);
 859   }
 860   // Detect empty (or broken) loop:
 861 #ifdef ASSERT
 862   if (VerifyMethodHandles) {
 863     // Verify that &bottom < &top (non-empty interval)
 864     Label L_ok, L_bad;
 865     if (positive_distance_in_slots.is_register()) {
 866       __ cmp(positive_distance_in_slots.as_register(), (int32_t) 0);
 867       __ br(Assembler::lessEqual, false, Assembler::pn, L_bad);
 868       __ delayed()->nop();
 869     }
 870     __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::lessUnsigned, Assembler::pt, L_ok);
 871     __ BIND(L_bad);
 872     __ stop("valid bounds (copy up)");
 873     __ BIND(L_ok);
 874   }
 875 #endif
 876   __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::greaterEqualUnsigned, Assembler::pn, L_break);
 877   // work top down to bottom, copying contiguous data upwards
 878   // In pseudo-code:
 879   //   while (--top >= bottom) *(top + distance) = *(top + 0);
 880   RegisterOrConstant offset = __ argument_offset(positive_distance_in_slots, positive_distance_in_slots.register_or_noreg());
 881   __ BIND(L_loop);
 882   __ sub(top_reg, wordSize, top_reg);
 883   __ ld_ptr(           Address(top_reg, 0     ), temp2_reg);
 884   __ st_ptr(temp2_reg, Address(top_reg, offset)           );
 885   __ cmp_and_brx_short(top_reg, bottom_reg, Assembler::greaterUnsigned, Assembler::pt, L_loop);
 886   assert(Interpreter::stackElementSize == wordSize, "else change loop");
 887   __ BIND(L_break);
 888   BLOCK_COMMENT("} move_arg_slots_up");
 889 }
 890 
 891 // in-place movement; no change to rsp
 892 // blows temp_reg, temp2_reg
 893 void MethodHandles::move_arg_slots_down(MacroAssembler* _masm,
 894                                         Address  bottom_addr,  // can use temp_reg
 895                                         Register top_reg,      // invariant
 896                                         RegisterOrConstant negative_distance_in_slots,  // destroyed if register
 897                                         Register temp_reg, Register temp2_reg) {
 898   assert_different_registers(top_reg,
 899                              negative_distance_in_slots.register_or_noreg(),
 900                              temp_reg, temp2_reg);
 901   BLOCK_COMMENT("move_arg_slots_down {");
 902   Label L_loop, L_break;
 903   Register bottom_reg = temp_reg;
 904   if (!bottom_addr.is_same_address(Address(bottom_reg, 0))) {
 905     __ add(bottom_addr, bottom_reg);
 906   }
 907   // Detect empty (or broken) loop:
 908 #ifdef ASSERT
 909   assert(!negative_distance_in_slots.is_constant() || negative_distance_in_slots.as_constant() < 0, "");
 910   if (VerifyMethodHandles) {
 911     // Verify that &bottom < &top (non-empty interval)
 912     Label L_ok, L_bad;
 913     if (negative_distance_in_slots.is_register()) {
 914       __ cmp(negative_distance_in_slots.as_register(), (int32_t) 0);
 915       __ br(Assembler::greaterEqual, false, Assembler::pn, L_bad);
 916       __ delayed()->nop();
 917     }
 918     __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::lessUnsigned, Assembler::pt, L_ok);
 919     __ BIND(L_bad);
 920     __ stop("valid bounds (copy down)");
 921     __ BIND(L_ok);
 922   }
 923 #endif
 924   __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::greaterEqualUnsigned, Assembler::pn, L_break);
 925   // work bottom up to top, copying contiguous data downwards
 926   // In pseudo-code:
 927   //   while (bottom < top) *(bottom - distance) = *(bottom + 0), bottom++;
 928   RegisterOrConstant offset = __ argument_offset(negative_distance_in_slots, negative_distance_in_slots.register_or_noreg());
 929   __ BIND(L_loop);
 930   __ ld_ptr(           Address(bottom_reg, 0     ), temp2_reg);
 931   __ st_ptr(temp2_reg, Address(bottom_reg, offset)           );
 932   __ add(bottom_reg, wordSize, bottom_reg);
 933   __ cmp_and_brx_short(bottom_reg, top_reg, Assembler::lessUnsigned, Assembler::pt, L_loop);
 934   assert(Interpreter::stackElementSize == wordSize, "else change loop");
 935   __ BIND(L_break);
 936   BLOCK_COMMENT("} move_arg_slots_down");
 937 }
 938 
 939 // Copy from a field or array element to a stacked argument slot.
 940 // is_element (ignored) says whether caller is loading an array element instead of an instance field.
 941 void MethodHandles::move_typed_arg(MacroAssembler* _masm,
 942                                    BasicType type, bool is_element,
 943                                    Address value_src, Address slot_dest,
 944                                    Register temp_reg) {
 945   assert(!slot_dest.uses(temp_reg), "must be different register");
 946   BLOCK_COMMENT(!is_element ? "move_typed_arg {" : "move_typed_arg { (array element)");
 947   if (type == T_OBJECT || type == T_ARRAY) {
 948     __ load_heap_oop(value_src, temp_reg);
 949     __ verify_oop(temp_reg);
 950     __ st_ptr(temp_reg, slot_dest);
 951   } else if (type != T_VOID) {
 952     int  arg_size      = type2aelembytes(type);
 953     bool arg_is_signed = is_signed_subword_type(type);
 954     int  slot_size     = is_subword_type(type) ? type2aelembytes(T_INT) : arg_size;  // store int sub-words as int
 955     __ load_sized_value( value_src, temp_reg, arg_size, arg_is_signed);
 956     __ store_sized_value(temp_reg, slot_dest, slot_size              );
 957   }
 958   BLOCK_COMMENT("} move_typed_arg");
 959 }
 960 
 961 // Cf. TemplateInterpreterGenerator::generate_return_entry_for and
 962 // InterpreterMacroAssembler::save_return_value
 963 void MethodHandles::move_return_value(MacroAssembler* _masm, BasicType type,
 964                                       Address return_slot) {
 965   BLOCK_COMMENT("move_return_value {");
 966   // Look at the type and pull the value out of the corresponding register.
 967   if (type == T_VOID) {
 968     // nothing to do
 969   } else if (type == T_OBJECT) {
 970     __ verify_oop(O0);
 971     __ st_ptr(O0, return_slot);
 972   } else if (type == T_INT || is_subword_type(type)) {
 973     int type_size = type2aelembytes(T_INT);
 974     __ store_sized_value(O0, return_slot, type_size);
 975   } else if (type == T_LONG) {
 976     // store the value by parts
 977     // Note: We assume longs are continguous (if misaligned) on the interpreter stack.
 978 #if !defined(_LP64) && defined(COMPILER2)
 979     __ stx(G1, return_slot);
 980 #else
 981   #ifdef _LP64
 982     __ stx(O0, return_slot);
 983   #else
 984     if (return_slot.has_disp()) {
 985       // The displacement is a constant
 986       __ st(O0, return_slot);
 987       __ st(O1, return_slot.plus_disp(Interpreter::stackElementSize));
 988     } else {
 989       __ std(O0, return_slot);
 990     }
 991   #endif
 992 #endif
 993   } else if (type == T_FLOAT) {
 994     __ stf(FloatRegisterImpl::S, Ftos_f, return_slot);
 995   } else if (type == T_DOUBLE) {
 996     __ stf(FloatRegisterImpl::D, Ftos_f, return_slot);
 997   } else {
 998     ShouldNotReachHere();
 999   }
1000   BLOCK_COMMENT("} move_return_value");
1001 }
1002 
1003 #ifndef PRODUCT
1004 extern "C" void print_method_handle(oop mh);
1005 void trace_method_handle_stub(const char* adaptername,
1006                               oopDesc* mh,
1007                               intptr_t* saved_sp) {
1008   bool has_mh = (strstr(adaptername, "return/") == NULL);  // return adapters don't have mh
1009   tty->print_cr("MH %s mh="INTPTR_FORMAT " saved_sp=" INTPTR_FORMAT, adaptername, (intptr_t) mh, saved_sp);
1010   if (has_mh)
1011     print_method_handle(mh);
1012 }
1013 void MethodHandles::trace_method_handle(MacroAssembler* _masm, const char* adaptername) {
1014   if (!TraceMethodHandles)  return;
1015   BLOCK_COMMENT("trace_method_handle {");
1016   // save: Gargs, O5_savedSP
1017   __ save_frame(16);
1018   __ set((intptr_t) adaptername, O0);
1019   __ mov(G3_method_handle, O1);
1020   __ mov(I5_savedSP, O2);
1021   __ mov(G3_method_handle, L3);
1022   __ mov(Gargs, L4);
1023   __ mov(G5_method_type, L5);
1024   __ call_VM_leaf(L7, CAST_FROM_FN_PTR(address, trace_method_handle_stub));
1025 
1026   __ mov(L3, G3_method_handle);
1027   __ mov(L4, Gargs);
1028   __ mov(L5, G5_method_type);
1029   __ restore();
1030   BLOCK_COMMENT("} trace_method_handle");
1031 }
1032 #endif // PRODUCT
1033 
1034 // which conversion op types are implemented here?
1035 int MethodHandles::adapter_conversion_ops_supported_mask() {
1036   return ((1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_ONLY)
1037          |(1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_RAW)
1038          |(1<<java_lang_invoke_AdapterMethodHandle::OP_CHECK_CAST)
1039          |(1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_PRIM)
1040          |(1<<java_lang_invoke_AdapterMethodHandle::OP_REF_TO_PRIM)
1041           // OP_PRIM_TO_REF is below...
1042          |(1<<java_lang_invoke_AdapterMethodHandle::OP_SWAP_ARGS)
1043          |(1<<java_lang_invoke_AdapterMethodHandle::OP_ROT_ARGS)
1044          |(1<<java_lang_invoke_AdapterMethodHandle::OP_DUP_ARGS)
1045          |(1<<java_lang_invoke_AdapterMethodHandle::OP_DROP_ARGS)
1046           // OP_COLLECT_ARGS is below...
1047          |(1<<java_lang_invoke_AdapterMethodHandle::OP_SPREAD_ARGS)
1048          |(!UseRicochetFrames ? 0 :
1049            java_lang_invoke_MethodTypeForm::vmlayout_offset_in_bytes() <= 0 ? 0 :
1050            ((1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF)
1051            |(1<<java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS)
1052            |(1<<java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS)
1053            )
1054           )
1055          );
1056 }
1057 
1058 //------------------------------------------------------------------------------
1059 // MethodHandles::generate_method_handle_stub
1060 //
1061 // Generate an "entry" field for a method handle.
1062 // This determines how the method handle will respond to calls.
1063 void MethodHandles::generate_method_handle_stub(MacroAssembler* _masm, MethodHandles::EntryKind ek) {
1064   MethodHandles::EntryKind ek_orig = ek_original_kind(ek);
1065 
1066   // Here is the register state during an interpreted call,
1067   // as set up by generate_method_handle_interpreter_entry():
1068   // - G5: garbage temp (was MethodHandle.invoke methodOop, unused)
1069   // - G3: receiver method handle
1070   // - O5_savedSP: sender SP (must preserve)
1071 
1072   const Register O0_scratch = O0;
1073   const Register O1_scratch = O1;
1074   const Register O2_scratch = O2;
1075   const Register O3_scratch = O3;
1076   const Register O4_scratch = O4;
1077   const Register G5_scratch = G5;
1078 
1079   // Often used names:
1080   const Register O0_argslot = O0;
1081 
1082   // Argument registers for _raise_exception:
1083   const Register O0_code     = O0;
1084   const Register O1_actual   = O1;
1085   const Register O2_required = O2;
1086 
1087   guarantee(java_lang_invoke_MethodHandle::vmentry_offset_in_bytes() != 0, "must have offsets");
1088 
1089   // Some handy addresses:
1090   Address G3_mh_vmtarget(   G3_method_handle, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes());
1091 
1092   Address G3_dmh_vmindex(   G3_method_handle, java_lang_invoke_DirectMethodHandle::vmindex_offset_in_bytes());
1093 
1094   Address G3_bmh_vmargslot( G3_method_handle, java_lang_invoke_BoundMethodHandle::vmargslot_offset_in_bytes());
1095   Address G3_bmh_argument(  G3_method_handle, java_lang_invoke_BoundMethodHandle::argument_offset_in_bytes());
1096 
1097   Address G3_amh_vmargslot( G3_method_handle, java_lang_invoke_AdapterMethodHandle::vmargslot_offset_in_bytes());
1098   Address G3_amh_argument ( G3_method_handle, java_lang_invoke_AdapterMethodHandle::argument_offset_in_bytes());
1099   Address G3_amh_conversion(G3_method_handle, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
1100 
1101   const int java_mirror_offset = in_bytes(Klass::java_mirror_offset());
1102 
1103   if (have_entry(ek)) {
1104     __ nop();  // empty stubs make SG sick
1105     return;
1106   }
1107 
1108   address interp_entry = __ pc();
1109 
1110   trace_method_handle(_masm, entry_name(ek));
1111 
1112   BLOCK_COMMENT(err_msg("Entry %s {", entry_name(ek)));
1113 
1114   switch ((int) ek) {
1115   case _raise_exception:
1116     {
1117       // Not a real MH entry, but rather shared code for raising an
1118       // exception.  For sharing purposes the arguments are passed into registers
1119       // and then placed in the intepreter calling convention here.
1120       assert(raise_exception_method(), "must be set");
1121       assert(raise_exception_method()->from_compiled_entry(), "method must be linked");
1122 
1123       __ set(AddressLiteral((address) &_raise_exception_method), G5_method);
1124       __ ld_ptr(Address(G5_method, 0), G5_method);
1125 
1126       const int jobject_oop_offset = 0;
1127       __ ld_ptr(Address(G5_method, jobject_oop_offset), G5_method);
1128 
1129       adjust_SP_and_Gargs_down_by_slots(_masm, 3, noreg, noreg);
1130 
1131       __ st    (O0_code,     __ argument_address(constant(2), noreg, 0));
1132       __ st_ptr(O1_actual,   __ argument_address(constant(1), noreg, 0));
1133       __ st_ptr(O2_required, __ argument_address(constant(0), noreg, 0));
1134       jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1135     }
1136     break;
1137 
1138   case _invokestatic_mh:
1139   case _invokespecial_mh:
1140     {
1141       __ load_heap_oop(G3_mh_vmtarget, G5_method);  // target is a methodOop
1142       // Same as TemplateTable::invokestatic or invokespecial,
1143       // minus the CP setup and profiling:
1144       if (ek == _invokespecial_mh) {
1145         // Must load & check the first argument before entering the target method.
1146         __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1147         __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1148         __ null_check(G3_method_handle);
1149         __ verify_oop(G3_method_handle);
1150       }
1151       jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1152     }
1153     break;
1154 
1155   case _invokevirtual_mh:
1156     {
1157       // Same as TemplateTable::invokevirtual,
1158       // minus the CP setup and profiling:
1159 
1160       // Pick out the vtable index and receiver offset from the MH,
1161       // and then we can discard it:
1162       Register O2_index = O2_scratch;
1163       __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1164       __ ldsw(G3_dmh_vmindex, O2_index);
1165       // Note:  The verifier allows us to ignore G3_mh_vmtarget.
1166       __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1167       __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1168 
1169       // Get receiver klass:
1170       Register O0_klass = O0_argslot;
1171       __ load_klass(G3_method_handle, O0_klass);
1172       __ verify_oop(O0_klass);
1173 
1174       // Get target methodOop & entry point:
1175       const int base = instanceKlass::vtable_start_offset() * wordSize;
1176       assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1177 
1178       __ sll_ptr(O2_index, LogBytesPerWord, O2_index);
1179       __ add(O0_klass, O2_index, O0_klass);
1180       Address vtable_entry_addr(O0_klass, base + vtableEntry::method_offset_in_bytes());
1181       __ ld_ptr(vtable_entry_addr, G5_method);
1182 
1183       jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1184     }
1185     break;
1186 
1187   case _invokeinterface_mh:
1188     {
1189       // Same as TemplateTable::invokeinterface,
1190       // minus the CP setup and profiling:
1191       __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1192       Register O1_intf  = O1_scratch;
1193       Register G5_index = G5_scratch;
1194       __ load_heap_oop(G3_mh_vmtarget, O1_intf);
1195       __ ldsw(G3_dmh_vmindex, G5_index);
1196       __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1197       __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1198 
1199       // Get receiver klass:
1200       Register O0_klass = O0_argslot;
1201       __ load_klass(G3_method_handle, O0_klass);
1202       __ verify_oop(O0_klass);
1203 
1204       // Get interface:
1205       Label no_such_interface;
1206       __ verify_oop(O1_intf);
1207       __ lookup_interface_method(O0_klass, O1_intf,
1208                                  // Note: next two args must be the same:
1209                                  G5_index, G5_method,
1210                                  O2_scratch,
1211                                  O3_scratch,
1212                                  no_such_interface);
1213 
1214       jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1215 
1216       __ bind(no_such_interface);
1217       // Throw an exception.
1218       // For historical reasons, it will be IncompatibleClassChangeError.
1219       __ unimplemented("not tested yet");
1220       __ ld_ptr(Address(O1_intf, java_mirror_offset), O2_required);  // required interface
1221       __ mov(   O0_klass,                             O1_actual);    // bad receiver
1222       __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1223       __ delayed()->mov(Bytecodes::_invokeinterface,  O0_code);      // who is complaining?
1224     }
1225     break;
1226 
1227   case _bound_ref_mh:
1228   case _bound_int_mh:
1229   case _bound_long_mh:
1230   case _bound_ref_direct_mh:
1231   case _bound_int_direct_mh:
1232   case _bound_long_direct_mh:
1233     {
1234       const bool direct_to_method = (ek >= _bound_ref_direct_mh);
1235       BasicType arg_type  = ek_bound_mh_arg_type(ek);
1236       int       arg_slots = type2size[arg_type];
1237 
1238       // Make room for the new argument:
1239       load_vmargslot(_masm, G3_bmh_vmargslot, O0_argslot);
1240       __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1241 
1242       insert_arg_slots(_masm, arg_slots * stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1243 
1244       // Store bound argument into the new stack slot:
1245       __ load_heap_oop(G3_bmh_argument, O1_scratch);
1246       if (arg_type == T_OBJECT) {
1247         __ st_ptr(O1_scratch, Address(O0_argslot, 0));
1248       } else {
1249         Address prim_value_addr(O1_scratch, java_lang_boxing_object::value_offset_in_bytes(arg_type));
1250         move_typed_arg(_masm, arg_type, false,
1251                        prim_value_addr,
1252                        Address(O0_argslot, 0),
1253                        O2_scratch);  // must be an even register for !_LP64 long moves (uses O2/O3)
1254       }
1255 
1256       if (direct_to_method) {
1257         __ load_heap_oop(G3_mh_vmtarget, G5_method);  // target is a methodOop
1258         jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1259       } else {
1260         __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);  // target is a methodOop
1261         __ verify_oop(G3_method_handle);
1262         __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1263       }
1264     }
1265     break;
1266 
1267   case _adapter_opt_profiling:
1268     if (java_lang_invoke_CountingMethodHandle::vmcount_offset_in_bytes() != 0) {
1269       Address G3_mh_vmcount(G3_method_handle, java_lang_invoke_CountingMethodHandle::vmcount_offset_in_bytes());
1270       __ ld(G3_mh_vmcount, O1_scratch);
1271       __ add(O1_scratch, 1, O1_scratch);
1272       __ st(O1_scratch, G3_mh_vmcount);
1273     }
1274     // fall through
1275 
1276   case _adapter_retype_only:
1277   case _adapter_retype_raw:
1278     // Immediately jump to the next MH layer:
1279     __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1280     __ verify_oop(G3_method_handle);
1281     __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1282     // This is OK when all parameter types widen.
1283     // It is also OK when a return type narrows.
1284     break;
1285 
1286   case _adapter_check_cast:
1287     {
1288       // Check a reference argument before jumping to the next layer of MH:
1289       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1290       Address vmarg = __ argument_address(O0_argslot, O0_argslot);
1291 
1292       // What class are we casting to?
1293       Register O1_klass = O1_scratch;  // Interesting AMH data.
1294       __ load_heap_oop(G3_amh_argument, O1_klass);  // This is a Class object!
1295       load_klass_from_Class(_masm, O1_klass, O2_scratch, O3_scratch);
1296 
1297       Label L_done;
1298       __ ld_ptr(vmarg, O2_scratch);
1299       __ br_null_short(O2_scratch, Assembler::pn, L_done);  // No cast if null.
1300       __ load_klass(O2_scratch, O2_scratch);
1301 
1302       // Live at this point:
1303       // - O0_argslot      :  argslot index in vmarg; may be required in the failing path
1304       // - O1_klass        :  klass required by the target method
1305       // - O2_scratch      :  argument klass to test
1306       // - G3_method_handle:  adapter method handle
1307       __ check_klass_subtype(O2_scratch, O1_klass, O3_scratch, O4_scratch, L_done);
1308 
1309       // If we get here, the type check failed!
1310       __ load_heap_oop(G3_amh_argument,        O2_required);  // required class
1311       __ ld_ptr(       vmarg,                  O1_actual);    // bad object
1312       __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1313       __ delayed()->mov(Bytecodes::_checkcast, O0_code);      // who is complaining?
1314 
1315       __ BIND(L_done);
1316       // Get the new MH:
1317       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1318       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1319     }
1320     break;
1321 
1322   case _adapter_prim_to_prim:
1323   case _adapter_ref_to_prim:
1324     // Handled completely by optimized cases.
1325     __ stop("init_AdapterMethodHandle should not issue this");
1326     break;
1327 
1328   case _adapter_opt_i2i:        // optimized subcase of adapt_prim_to_prim
1329 //case _adapter_opt_f2i:        // optimized subcase of adapt_prim_to_prim
1330   case _adapter_opt_l2i:        // optimized subcase of adapt_prim_to_prim
1331   case _adapter_opt_unboxi:     // optimized subcase of adapt_ref_to_prim
1332     {
1333       // Perform an in-place conversion to int or an int subword.
1334       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1335       Address value;
1336       Address vmarg;
1337       bool value_left_justified = false;
1338 
1339       switch (ek) {
1340       case _adapter_opt_i2i:
1341         value = vmarg = __ argument_address(O0_argslot, O0_argslot);
1342         break;
1343       case _adapter_opt_l2i:
1344         {
1345           // just delete the extra slot
1346 #ifdef _LP64
1347           // In V9, longs are given 2 64-bit slots in the interpreter, but the
1348           // data is passed in only 1 slot.
1349           // Keep the second slot.
1350           __ add(__ argument_address(O0_argslot, O0_argslot, -1), O0_argslot);
1351           remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1352           value = Address(O0_argslot, 4);  // Get least-significant 32-bit of 64-bit value.
1353           vmarg = Address(O0_argslot, Interpreter::stackElementSize);
1354 #else
1355           // Keep the first slot.
1356           __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1357           remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1358           value = Address(O0_argslot, 0);
1359           vmarg = value;
1360 #endif
1361         }
1362         break;
1363       case _adapter_opt_unboxi:
1364         {
1365           vmarg = __ argument_address(O0_argslot, O0_argslot);
1366           // Load the value up from the heap.
1367           __ ld_ptr(vmarg, O1_scratch);
1368           int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_INT);
1369 #ifdef ASSERT
1370           for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1371             if (is_subword_type(BasicType(bt)))
1372               assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(BasicType(bt)), "");
1373           }
1374 #endif
1375           __ null_check(O1_scratch, value_offset);
1376           value = Address(O1_scratch, value_offset);
1377 #ifdef _BIG_ENDIAN
1378           // Values stored in objects are packed.
1379           value_left_justified = true;
1380 #endif
1381         }
1382         break;
1383       default:
1384         ShouldNotReachHere();
1385       }
1386 
1387       // This check is required on _BIG_ENDIAN
1388       Register G5_vminfo = G5_scratch;
1389       __ ldsw(G3_amh_conversion, G5_vminfo);
1390       assert(CONV_VMINFO_SHIFT == 0, "preshifted");
1391 
1392       // Original 32-bit vmdata word must be of this form:
1393       // | MBZ:6 | signBitCount:8 | srcDstTypes:8 | conversionOp:8 |
1394       __ lduw(value, O1_scratch);
1395       if (!value_left_justified)
1396         __ sll(O1_scratch, G5_vminfo, O1_scratch);
1397       Label zero_extend, done;
1398       __ btst(CONV_VMINFO_SIGN_FLAG, G5_vminfo);
1399       __ br(Assembler::zero, false, Assembler::pn, zero_extend);
1400       __ delayed()->nop();
1401 
1402       // this path is taken for int->byte, int->short
1403       __ sra(O1_scratch, G5_vminfo, O1_scratch);
1404       __ ba_short(done);
1405 
1406       __ bind(zero_extend);
1407       // this is taken for int->char
1408       __ srl(O1_scratch, G5_vminfo, O1_scratch);
1409 
1410       __ bind(done);
1411       __ st(O1_scratch, vmarg);
1412 
1413       // Get the new MH:
1414       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1415       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1416     }
1417     break;
1418 
1419   case _adapter_opt_i2l:        // optimized subcase of adapt_prim_to_prim
1420   case _adapter_opt_unboxl:     // optimized subcase of adapt_ref_to_prim
1421     {
1422       // Perform an in-place int-to-long or ref-to-long conversion.
1423       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1424 
1425       // On big-endian machine we duplicate the slot and store the MSW
1426       // in the first slot.
1427       __ add(__ argument_address(O0_argslot, O0_argslot, 1), O0_argslot);
1428 
1429       insert_arg_slots(_masm, stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1430 
1431       Address arg_lsw(O0_argslot, 0);
1432       Address arg_msw(O0_argslot, -Interpreter::stackElementSize);
1433 
1434       switch (ek) {
1435       case _adapter_opt_i2l:
1436         {
1437 #ifdef _LP64
1438           __ ldsw(arg_lsw, O2_scratch);                 // Load LSW sign-extended
1439 #else
1440           __ ldsw(arg_lsw, O3_scratch);                 // Load LSW sign-extended
1441           __ srlx(O3_scratch, BitsPerInt, O2_scratch);  // Move MSW value to lower 32-bits for std
1442 #endif
1443           __ st_long(O2_scratch, arg_msw);              // Uses O2/O3 on !_LP64
1444         }
1445         break;
1446       case _adapter_opt_unboxl:
1447         {
1448           // Load the value up from the heap.
1449           __ ld_ptr(arg_lsw, O1_scratch);
1450           int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_LONG);
1451           assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(T_DOUBLE), "");
1452           __ null_check(O1_scratch, value_offset);
1453           __ ld_long(Address(O1_scratch, value_offset), O2_scratch);  // Uses O2/O3 on !_LP64
1454           __ st_long(O2_scratch, arg_msw);
1455         }
1456         break;
1457       default:
1458         ShouldNotReachHere();
1459       }
1460 
1461       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1462       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1463     }
1464     break;
1465 
1466   case _adapter_opt_f2d:        // optimized subcase of adapt_prim_to_prim
1467   case _adapter_opt_d2f:        // optimized subcase of adapt_prim_to_prim
1468     {
1469       // perform an in-place floating primitive conversion
1470       __ unimplemented(entry_name(ek));
1471     }
1472     break;
1473 
1474   case _adapter_prim_to_ref:
1475     __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
1476     break;
1477 
1478   case _adapter_swap_args:
1479   case _adapter_rot_args:
1480     // handled completely by optimized cases
1481     __ stop("init_AdapterMethodHandle should not issue this");
1482     break;
1483 
1484   case _adapter_opt_swap_1:
1485   case _adapter_opt_swap_2:
1486   case _adapter_opt_rot_1_up:
1487   case _adapter_opt_rot_1_down:
1488   case _adapter_opt_rot_2_up:
1489   case _adapter_opt_rot_2_down:
1490     {
1491       int swap_slots = ek_adapter_opt_swap_slots(ek);
1492       int rotate     = ek_adapter_opt_swap_mode(ek);
1493 
1494       // 'argslot' is the position of the first argument to swap.
1495       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1496       __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1497       if (VerifyMethodHandles)
1498         verify_argslot(_masm, O0_argslot, O2_scratch, "swap point must fall within current frame");
1499 
1500       // 'vminfo' is the second.
1501       Register O1_destslot = O1_scratch;
1502       load_conversion_vminfo(_masm, G3_amh_conversion, O1_destslot);
1503       __ add(__ argument_address(O1_destslot, O1_destslot), O1_destslot);
1504       if (VerifyMethodHandles)
1505         verify_argslot(_masm, O1_destslot, O2_scratch, "swap point must fall within current frame");
1506 
1507       assert(Interpreter::stackElementSize == wordSize, "else rethink use of wordSize here");
1508       if (!rotate) {
1509         // simple swap
1510         for (int i = 0; i < swap_slots; i++) {
1511           __ ld_ptr(            Address(O0_argslot,  i * wordSize), O2_scratch);
1512           __ ld_ptr(            Address(O1_destslot, i * wordSize), O3_scratch);
1513           __ st_ptr(O3_scratch, Address(O0_argslot,  i * wordSize));
1514           __ st_ptr(O2_scratch, Address(O1_destslot, i * wordSize));
1515         }
1516       } else {
1517         // A rotate is actually pair of moves, with an "odd slot" (or pair)
1518         // changing place with a series of other slots.
1519         // First, push the "odd slot", which is going to get overwritten
1520         switch (swap_slots) {
1521         case 2 :  __ ld_ptr(Address(O0_argslot, 1 * wordSize), O4_scratch); // fall-thru
1522         case 1 :  __ ld_ptr(Address(O0_argslot, 0 * wordSize), O3_scratch); break;
1523         default:  ShouldNotReachHere();
1524         }
1525         if (rotate > 0) {
1526           // Here is rotate > 0:
1527           // (low mem)                                          (high mem)
1528           //     | dest:     more_slots...     | arg: odd_slot :arg+1 |
1529           // =>
1530           //     | dest: odd_slot | dest+1: more_slots...      :arg+1 |
1531           // work argslot down to destslot, copying contiguous data upwards
1532           // pseudo-code:
1533           //   argslot  = src_addr - swap_bytes
1534           //   destslot = dest_addr
1535           //   while (argslot >= destslot) *(argslot + swap_bytes) = *(argslot + 0), argslot--;
1536           move_arg_slots_up(_masm,
1537                             O1_destslot,
1538                             Address(O0_argslot, 0),
1539                             swap_slots,
1540                             O0_argslot, O2_scratch);
1541         } else {
1542           // Here is the other direction, rotate < 0:
1543           // (low mem)                                          (high mem)
1544           //     | arg: odd_slot | arg+1: more_slots...       :dest+1 |
1545           // =>
1546           //     | arg:    more_slots...     | dest: odd_slot :dest+1 |
1547           // work argslot up to destslot, copying contiguous data downwards
1548           // pseudo-code:
1549           //   argslot  = src_addr + swap_bytes
1550           //   destslot = dest_addr
1551           //   while (argslot <= destslot) *(argslot - swap_bytes) = *(argslot + 0), argslot++;
1552           // dest_slot denotes an exclusive upper limit
1553           int limit_bias = OP_ROT_ARGS_DOWN_LIMIT_BIAS;
1554           if (limit_bias != 0)
1555             __ add(O1_destslot, - limit_bias * wordSize, O1_destslot);
1556           move_arg_slots_down(_masm,
1557                               Address(O0_argslot, swap_slots * wordSize),
1558                               O1_destslot,
1559                               -swap_slots,
1560                               O0_argslot, O2_scratch);
1561 
1562           __ sub(O1_destslot, swap_slots * wordSize, O1_destslot);
1563         }
1564         // pop the original first chunk into the destination slot, now free
1565         switch (swap_slots) {
1566         case 2 :  __ st_ptr(O4_scratch, Address(O1_destslot, 1 * wordSize)); // fall-thru
1567         case 1 :  __ st_ptr(O3_scratch, Address(O1_destslot, 0 * wordSize)); break;
1568         default:  ShouldNotReachHere();
1569         }
1570       }
1571 
1572       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1573       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1574     }
1575     break;
1576 
1577   case _adapter_dup_args:
1578     {
1579       // 'argslot' is the position of the first argument to duplicate.
1580       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1581       __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1582 
1583       // 'stack_move' is negative number of words to duplicate.
1584       Register O1_stack_move = O1_scratch;
1585       load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1586 
1587       if (VerifyMethodHandles) {
1588         verify_argslots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, true,
1589                         "copied argument(s) must fall within current frame");
1590       }
1591 
1592       // insert location is always the bottom of the argument list:
1593       __ neg(O1_stack_move);
1594       push_arg_slots(_masm, O0_argslot, O1_stack_move, O2_scratch, O3_scratch);
1595 
1596       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1597       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1598     }
1599     break;
1600 
1601   case _adapter_drop_args:
1602     {
1603       // 'argslot' is the position of the first argument to nuke.
1604       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1605       __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1606 
1607       // 'stack_move' is number of words to drop.
1608       Register O1_stack_move = O1_scratch;
1609       load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1610 
1611       remove_arg_slots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, O4_scratch);
1612 
1613       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1614       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1615     }
1616     break;
1617 
1618   case _adapter_collect_args:
1619   case _adapter_fold_args:
1620   case _adapter_spread_args:
1621     // Handled completely by optimized cases.
1622     __ stop("init_AdapterMethodHandle should not issue this");
1623     break;
1624 
1625   case _adapter_opt_collect_ref:
1626   case _adapter_opt_collect_int:
1627   case _adapter_opt_collect_long:
1628   case _adapter_opt_collect_float:
1629   case _adapter_opt_collect_double:
1630   case _adapter_opt_collect_void:
1631   case _adapter_opt_collect_0_ref:
1632   case _adapter_opt_collect_1_ref:
1633   case _adapter_opt_collect_2_ref:
1634   case _adapter_opt_collect_3_ref:
1635   case _adapter_opt_collect_4_ref:
1636   case _adapter_opt_collect_5_ref:
1637   case _adapter_opt_filter_S0_ref:
1638   case _adapter_opt_filter_S1_ref:
1639   case _adapter_opt_filter_S2_ref:
1640   case _adapter_opt_filter_S3_ref:
1641   case _adapter_opt_filter_S4_ref:
1642   case _adapter_opt_filter_S5_ref:
1643   case _adapter_opt_collect_2_S0_ref:
1644   case _adapter_opt_collect_2_S1_ref:
1645   case _adapter_opt_collect_2_S2_ref:
1646   case _adapter_opt_collect_2_S3_ref:
1647   case _adapter_opt_collect_2_S4_ref:
1648   case _adapter_opt_collect_2_S5_ref:
1649   case _adapter_opt_fold_ref:
1650   case _adapter_opt_fold_int:
1651   case _adapter_opt_fold_long:
1652   case _adapter_opt_fold_float:
1653   case _adapter_opt_fold_double:
1654   case _adapter_opt_fold_void:
1655   case _adapter_opt_fold_1_ref:
1656   case _adapter_opt_fold_2_ref:
1657   case _adapter_opt_fold_3_ref:
1658   case _adapter_opt_fold_4_ref:
1659   case _adapter_opt_fold_5_ref:
1660     {
1661       // Given a fresh incoming stack frame, build a new ricochet frame.
1662       // On entry, TOS points at a return PC, and FP is the callers frame ptr.
1663       // RSI/R13 has the caller's exact stack pointer, which we must also preserve.
1664       // RCX contains an AdapterMethodHandle of the indicated kind.
1665 
1666       // Relevant AMH fields:
1667       // amh.vmargslot:
1668       //   points to the trailing edge of the arguments
1669       //   to filter, collect, or fold.  For a boxing operation,
1670       //   it points just after the single primitive value.
1671       // amh.argument:
1672       //   recursively called MH, on |collect| arguments
1673       // amh.vmtarget:
1674       //   final destination MH, on return value, etc.
1675       // amh.conversion.dest:
1676       //   tells what is the type of the return value
1677       //   (not needed here, since dest is also derived from ek)
1678       // amh.conversion.vminfo:
1679       //   points to the trailing edge of the return value
1680       //   when the vmtarget is to be called; this is
1681       //   equal to vmargslot + (retained ? |collect| : 0)
1682 
1683       // Pass 0 or more argument slots to the recursive target.
1684       int collect_count_constant = ek_adapter_opt_collect_count(ek);
1685 
1686       // The collected arguments are copied from the saved argument list:
1687       int collect_slot_constant = ek_adapter_opt_collect_slot(ek);
1688 
1689       assert(ek_orig == _adapter_collect_args ||
1690              ek_orig == _adapter_fold_args, "");
1691       bool retain_original_args = (ek_orig == _adapter_fold_args);
1692 
1693       // The return value is replaced (or inserted) at the 'vminfo' argslot.
1694       // Sometimes we can compute this statically.
1695       int dest_slot_constant = -1;
1696       if (!retain_original_args)
1697         dest_slot_constant = collect_slot_constant;
1698       else if (collect_slot_constant >= 0 && collect_count_constant >= 0)
1699         // We are preserving all the arguments, and the return value is prepended,
1700         // so the return slot is to the left (above) the |collect| sequence.
1701         dest_slot_constant = collect_slot_constant + collect_count_constant;
1702 
1703       // Replace all those slots by the result of the recursive call.
1704       // The result type can be one of ref, int, long, float, double, void.
1705       // In the case of void, nothing is pushed on the stack after return.
1706       BasicType dest = ek_adapter_opt_collect_type(ek);
1707       assert(dest == type2wfield[dest], "dest is a stack slot type");
1708       int dest_count = type2size[dest];
1709       assert(dest_count == 1 || dest_count == 2 || (dest_count == 0 && dest == T_VOID), "dest has a size");
1710 
1711       // Choose a return continuation.
1712       EntryKind ek_ret = _adapter_opt_return_any;
1713       if (dest != T_CONFLICT && OptimizeMethodHandles) {
1714         switch (dest) {
1715         case T_INT    : ek_ret = _adapter_opt_return_int;     break;
1716         case T_LONG   : ek_ret = _adapter_opt_return_long;    break;
1717         case T_FLOAT  : ek_ret = _adapter_opt_return_float;   break;
1718         case T_DOUBLE : ek_ret = _adapter_opt_return_double;  break;
1719         case T_OBJECT : ek_ret = _adapter_opt_return_ref;     break;
1720         case T_VOID   : ek_ret = _adapter_opt_return_void;    break;
1721         default       : ShouldNotReachHere();
1722         }
1723         if (dest == T_OBJECT && dest_slot_constant >= 0) {
1724           EntryKind ek_try = EntryKind(_adapter_opt_return_S0_ref + dest_slot_constant);
1725           if (ek_try <= _adapter_opt_return_LAST &&
1726               ek_adapter_opt_return_slot(ek_try) == dest_slot_constant) {
1727             ek_ret = ek_try;
1728           }
1729         }
1730         assert(ek_adapter_opt_return_type(ek_ret) == dest, "");
1731       }
1732 
1733       // Already pushed:  ... keep1 | collect | keep2 |
1734 
1735       // Push a few extra argument words, if we need them to store the return value.
1736       {
1737         int extra_slots = 0;
1738         if (retain_original_args) {
1739           extra_slots = dest_count;
1740         } else if (collect_count_constant == -1) {
1741           extra_slots = dest_count;  // collect_count might be zero; be generous
1742         } else if (dest_count > collect_count_constant) {
1743           extra_slots = (dest_count - collect_count_constant);
1744         } else {
1745           // else we know we have enough dead space in |collect| to repurpose for return values
1746         }
1747         if (extra_slots != 0) {
1748           __ sub(SP, round_to(extra_slots, 2) * Interpreter::stackElementSize, SP);
1749         }
1750       }
1751 
1752       // Set up Ricochet Frame.
1753       __ mov(SP, O5_savedSP);  // record SP for the callee
1754 
1755       // One extra (empty) slot for outgoing target MH (see Gargs computation below).
1756       __ save_frame(2);  // Note: we need to add 2 slots since frame::memory_parameter_word_sp_offset is 23.
1757 
1758       // Note: Gargs is live throughout the following, until we make our recursive call.
1759       // And the RF saves a copy in L4_saved_args_base.
1760 
1761       RicochetFrame::enter_ricochet_frame(_masm, G3_method_handle, Gargs,
1762                                           entry(ek_ret)->from_interpreted_entry());
1763 
1764       // Compute argument base:
1765       // Set up Gargs for current frame, extra (empty) slot is for outgoing target MH (space reserved by save_frame above).
1766       __ add(FP, STACK_BIAS - (1 * Interpreter::stackElementSize), Gargs);
1767 
1768       // Now pushed:  ... keep1 | collect | keep2 | extra | [RF]
1769 
1770 #ifdef ASSERT
1771       if (VerifyMethodHandles && dest != T_CONFLICT) {
1772         BLOCK_COMMENT("verify AMH.conv.dest {");
1773         extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O1_scratch);
1774         Label L_dest_ok;
1775         __ cmp(O1_scratch, (int) dest);
1776         __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1777         __ delayed()->nop();
1778         if (dest == T_INT) {
1779           for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1780             if (is_subword_type(BasicType(bt))) {
1781               __ cmp(O1_scratch, (int) bt);
1782               __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1783               __ delayed()->nop();
1784             }
1785           }
1786         }
1787         __ stop("bad dest in AMH.conv");
1788         __ BIND(L_dest_ok);
1789         BLOCK_COMMENT("} verify AMH.conv.dest");
1790       }
1791 #endif //ASSERT
1792 
1793       // Find out where the original copy of the recursive argument sequence begins.
1794       Register O0_coll = O0_scratch;
1795       {
1796         RegisterOrConstant collect_slot = collect_slot_constant;
1797         if (collect_slot_constant == -1) {
1798           load_vmargslot(_masm, G3_amh_vmargslot, O1_scratch);
1799           collect_slot = O1_scratch;
1800         }
1801         // collect_slot might be 0, but we need the move anyway.
1802         __ add(RicochetFrame::L4_saved_args_base, __ argument_offset(collect_slot, collect_slot.register_or_noreg()), O0_coll);
1803         // O0_coll now points at the trailing edge of |collect| and leading edge of |keep2|
1804       }
1805 
1806       // Replace the old AMH with the recursive MH.  (No going back now.)
1807       // In the case of a boxing call, the recursive call is to a 'boxer' method,
1808       // such as Integer.valueOf or Long.valueOf.  In the case of a filter
1809       // or collect call, it will take one or more arguments, transform them,
1810       // and return some result, to store back into argument_base[vminfo].
1811       __ load_heap_oop(G3_amh_argument, G3_method_handle);
1812       if (VerifyMethodHandles)  verify_method_handle(_masm, G3_method_handle, O1_scratch, O2_scratch);
1813 
1814       // Calculate |collect|, the number of arguments we are collecting.
1815       Register O1_collect_count = O1_scratch;
1816       RegisterOrConstant collect_count;
1817       if (collect_count_constant < 0) {
1818         __ load_method_handle_vmslots(O1_collect_count, G3_method_handle, O2_scratch);
1819         collect_count = O1_collect_count;
1820       } else {
1821         collect_count = collect_count_constant;
1822 #ifdef ASSERT
1823         if (VerifyMethodHandles) {
1824           BLOCK_COMMENT("verify collect_count_constant {");
1825           __ load_method_handle_vmslots(O3_scratch, G3_method_handle, O2_scratch);
1826           Label L_count_ok;
1827           __ cmp_and_br_short(O3_scratch, collect_count_constant, Assembler::equal, Assembler::pt, L_count_ok);
1828           __ stop("bad vminfo in AMH.conv");
1829           __ BIND(L_count_ok);
1830           BLOCK_COMMENT("} verify collect_count_constant");
1831         }
1832 #endif //ASSERT
1833       }
1834 
1835       // copy |collect| slots directly to TOS:
1836       push_arg_slots(_masm, O0_coll, collect_count, O2_scratch, O3_scratch);
1837       // Now pushed:  ... keep1 | collect | keep2 | RF... | collect |
1838       // O0_coll still points at the trailing edge of |collect| and leading edge of |keep2|
1839 
1840       // If necessary, adjust the saved arguments to make room for the eventual return value.
1841       // Normal adjustment:  ... keep1 | +dest+ | -collect- | keep2 | RF... | collect |
1842       // If retaining args:  ... keep1 | +dest+ |  collect  | keep2 | RF... | collect |
1843       // In the non-retaining case, this might move keep2 either up or down.
1844       // We don't have to copy the whole | RF... collect | complex,
1845       // but we must adjust RF.saved_args_base.
1846       // Also, from now on, we will forget about the original copy of |collect|.
1847       // If we are retaining it, we will treat it as part of |keep2|.
1848       // For clarity we will define |keep3| = |collect|keep2| or |keep2|.
1849 
1850       BLOCK_COMMENT("adjust trailing arguments {");
1851       // Compare the sizes of |+dest+| and |-collect-|, which are opposed opening and closing movements.
1852       int                open_count  = dest_count;
1853       RegisterOrConstant close_count = collect_count_constant;
1854       Register O1_close_count = O1_collect_count;
1855       if (retain_original_args) {
1856         close_count = constant(0);
1857       } else if (collect_count_constant == -1) {
1858         close_count = O1_collect_count;
1859       }
1860 
1861       // How many slots need moving?  This is simply dest_slot (0 => no |keep3|).
1862       RegisterOrConstant keep3_count;
1863       Register O2_keep3_count = O2_scratch;
1864       if (dest_slot_constant < 0) {
1865         extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O2_keep3_count);
1866         keep3_count = O2_keep3_count;
1867       } else  {
1868         keep3_count = dest_slot_constant;
1869 #ifdef ASSERT
1870         if (VerifyMethodHandles && dest_slot_constant < 0) {
1871           BLOCK_COMMENT("verify dest_slot_constant {");
1872           extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O3_scratch);
1873           Label L_vminfo_ok;
1874           __ cmp_and_br_short(O3_scratch, dest_slot_constant, Assembler::equal, Assembler::pt, L_vminfo_ok);
1875           __ stop("bad vminfo in AMH.conv");
1876           __ BIND(L_vminfo_ok);
1877           BLOCK_COMMENT("} verify dest_slot_constant");
1878         }
1879 #endif //ASSERT
1880       }
1881 
1882       // tasks remaining:
1883       bool move_keep3 = (!keep3_count.is_constant() || keep3_count.as_constant() != 0);
1884       bool stomp_dest = (NOT_DEBUG(dest == T_OBJECT) DEBUG_ONLY(dest_count != 0));
1885       bool fix_arg_base = (!close_count.is_constant() || open_count != close_count.as_constant());
1886 
1887       // Old and new argument locations (based at slot 0).
1888       // Net shift (&new_argv - &old_argv) is (close_count - open_count).
1889       bool zero_open_count = (open_count == 0);  // remember this bit of info
1890       if (move_keep3 && fix_arg_base) {
1891         // It will be easier to have everything in one register:
1892         if (close_count.is_register()) {
1893           // Deduct open_count from close_count register to get a clean +/- value.
1894           __ sub(close_count.as_register(), open_count, close_count.as_register());
1895         } else {
1896           close_count = close_count.as_constant() - open_count;
1897         }
1898         open_count = 0;
1899       }
1900       Register L4_old_argv = RicochetFrame::L4_saved_args_base;
1901       Register O3_new_argv = O3_scratch;
1902       if (fix_arg_base) {
1903         __ add(L4_old_argv, __ argument_offset(close_count, O4_scratch), O3_new_argv,
1904                -(open_count * Interpreter::stackElementSize));
1905       }
1906 
1907       // First decide if any actual data are to be moved.
1908       // We can skip if (a) |keep3| is empty, or (b) the argument list size didn't change.
1909       // (As it happens, all movements involve an argument list size change.)
1910 
1911       // If there are variable parameters, use dynamic checks to skip around the whole mess.
1912       Label L_done;
1913       if (keep3_count.is_register()) {
1914         __ cmp_and_br_short(keep3_count.as_register(), 0, Assembler::equal, Assembler::pn, L_done);
1915       }
1916       if (close_count.is_register()) {
1917         __ cmp_and_br_short(close_count.as_register(), open_count, Assembler::equal, Assembler::pn, L_done);
1918       }
1919 
1920       if (move_keep3 && fix_arg_base) {
1921         bool emit_move_down = false, emit_move_up = false, emit_guard = false;
1922         if (!close_count.is_constant()) {
1923           emit_move_down = emit_guard = !zero_open_count;
1924           emit_move_up   = true;
1925         } else if (open_count != close_count.as_constant()) {
1926           emit_move_down = (open_count > close_count.as_constant());
1927           emit_move_up   = !emit_move_down;
1928         }
1929         Label L_move_up;
1930         if (emit_guard) {
1931           __ cmp(close_count.as_register(), open_count);
1932           __ br(Assembler::greater, false, Assembler::pn, L_move_up);
1933           __ delayed()->nop();
1934         }
1935 
1936         if (emit_move_down) {
1937           // Move arguments down if |+dest+| > |-collect-|
1938           // (This is rare, except when arguments are retained.)
1939           // This opens space for the return value.
1940           if (keep3_count.is_constant()) {
1941             for (int i = 0; i < keep3_count.as_constant(); i++) {
1942               __ ld_ptr(            Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
1943               __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize)            );
1944             }
1945           } else {
1946             // Live: O1_close_count, O2_keep3_count, O3_new_argv
1947             Register argv_top = O0_scratch;
1948             __ add(L4_old_argv, __ argument_offset(keep3_count, O4_scratch), argv_top);
1949             move_arg_slots_down(_masm,
1950                                 Address(L4_old_argv, 0),  // beginning of old argv
1951                                 argv_top,                 // end of old argv
1952                                 close_count,              // distance to move down (must be negative)
1953                                 O4_scratch, G5_scratch);
1954           }
1955         }
1956 
1957         if (emit_guard) {
1958           __ ba_short(L_done);  // assumes emit_move_up is true also
1959           __ BIND(L_move_up);
1960         }
1961 
1962         if (emit_move_up) {
1963           // Move arguments up if |+dest+| < |-collect-|
1964           // (This is usual, except when |keep3| is empty.)
1965           // This closes up the space occupied by the now-deleted collect values.
1966           if (keep3_count.is_constant()) {
1967             for (int i = keep3_count.as_constant() - 1; i >= 0; i--) {
1968               __ ld_ptr(            Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
1969               __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize)            );
1970             }
1971           } else {
1972             Address argv_top(L4_old_argv, __ argument_offset(keep3_count, O4_scratch));
1973             // Live: O1_close_count, O2_keep3_count, O3_new_argv
1974             move_arg_slots_up(_masm,
1975                               L4_old_argv,  // beginning of old argv
1976                               argv_top,     // end of old argv
1977                               close_count,  // distance to move up (must be positive)
1978                               O4_scratch, G5_scratch);
1979           }
1980         }
1981       }
1982       __ BIND(L_done);
1983 
1984       if (fix_arg_base) {
1985         // adjust RF.saved_args_base
1986         __ mov(O3_new_argv, RicochetFrame::L4_saved_args_base);
1987       }
1988 
1989       if (stomp_dest) {
1990         // Stomp the return slot, so it doesn't hold garbage.
1991         // This isn't strictly necessary, but it may help detect bugs.
1992         __ set(RicochetFrame::RETURN_VALUE_PLACEHOLDER, O4_scratch);
1993         __ st_ptr(O4_scratch, Address(RicochetFrame::L4_saved_args_base,
1994                                       __ argument_offset(keep3_count, keep3_count.register_or_noreg())));  // uses O2_keep3_count
1995       }
1996       BLOCK_COMMENT("} adjust trailing arguments");
1997 
1998       BLOCK_COMMENT("do_recursive_call");
1999       __ mov(SP, O5_savedSP);  // record SP for the callee
2000       __ set(ExternalAddress(SharedRuntime::ricochet_blob()->bounce_addr() - frame::pc_return_offset), O7);
2001       // The globally unique bounce address has two purposes:
2002       // 1. It helps the JVM recognize this frame (frame::is_ricochet_frame).
2003       // 2. When returned to, it cuts back the stack and redirects control flow
2004       //    to the return handler.
2005       // The return handler will further cut back the stack when it takes
2006       // down the RF.  Perhaps there is a way to streamline this further.
2007 
2008       // State during recursive call:
2009       // ... keep1 | dest | dest=42 | keep3 | RF... | collect | bounce_pc |
2010       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2011     }
2012     break;
2013 
2014   case _adapter_opt_return_ref:
2015   case _adapter_opt_return_int:
2016   case _adapter_opt_return_long:
2017   case _adapter_opt_return_float:
2018   case _adapter_opt_return_double:
2019   case _adapter_opt_return_void:
2020   case _adapter_opt_return_S0_ref:
2021   case _adapter_opt_return_S1_ref:
2022   case _adapter_opt_return_S2_ref:
2023   case _adapter_opt_return_S3_ref:
2024   case _adapter_opt_return_S4_ref:
2025   case _adapter_opt_return_S5_ref:
2026     {
2027       BasicType dest_type_constant = ek_adapter_opt_return_type(ek);
2028       int       dest_slot_constant = ek_adapter_opt_return_slot(ek);
2029 
2030       if (VerifyMethodHandles)  RicochetFrame::verify_clean(_masm);
2031 
2032       if (dest_slot_constant == -1) {
2033         // The current stub is a general handler for this dest_type.
2034         // It can be called from _adapter_opt_return_any below.
2035         // Stash the address in a little table.
2036         assert((dest_type_constant & CONV_TYPE_MASK) == dest_type_constant, "oob");
2037         address return_handler = __ pc();
2038         _adapter_return_handlers[dest_type_constant] = return_handler;
2039         if (dest_type_constant == T_INT) {
2040           // do the subword types too
2041           for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
2042             if (is_subword_type(BasicType(bt)) &&
2043                 _adapter_return_handlers[bt] == NULL) {
2044               _adapter_return_handlers[bt] = return_handler;
2045             }
2046           }
2047         }
2048       }
2049 
2050       // On entry to this continuation handler, make Gargs live again.
2051       __ mov(RicochetFrame::L4_saved_args_base, Gargs);
2052 
2053       Register O7_temp   = O7;
2054       Register O5_vminfo = O5;
2055 
2056       RegisterOrConstant dest_slot = dest_slot_constant;
2057       if (dest_slot_constant == -1) {
2058         extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O5_vminfo);
2059         dest_slot = O5_vminfo;
2060       }
2061       // Store the result back into the argslot.
2062       // This code uses the interpreter calling sequence, in which the return value
2063       // is usually left in the TOS register, as defined by InterpreterMacroAssembler::pop.
2064       // There are certain irregularities with floating point values, which can be seen
2065       // in TemplateInterpreterGenerator::generate_return_entry_for.
2066       move_return_value(_masm, dest_type_constant, __ argument_address(dest_slot, O7_temp));
2067 
2068       RicochetFrame::leave_ricochet_frame(_masm, G3_method_handle, I5_savedSP, I7);
2069 
2070       // Load the final target and go.
2071       if (VerifyMethodHandles)  verify_method_handle(_masm, G3_method_handle, O0_scratch, O1_scratch);
2072       __ restore(I5_savedSP, G0, SP);
2073       __ jump_to_method_handle_entry(G3_method_handle, O0_scratch);
2074       __ illtrap(0);
2075     }
2076     break;
2077 
2078   case _adapter_opt_return_any:
2079     {
2080       Register O7_temp      = O7;
2081       Register O5_dest_type = O5;
2082 
2083       if (VerifyMethodHandles)  RicochetFrame::verify_clean(_masm);
2084       extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O5_dest_type);
2085       __ set(ExternalAddress((address) &_adapter_return_handlers[0]), O7_temp);
2086       __ sll_ptr(O5_dest_type, LogBytesPerWord, O5_dest_type);
2087       __ ld_ptr(O7_temp, O5_dest_type, O7_temp);
2088 
2089 #ifdef ASSERT
2090       { Label L_ok;
2091         __ br_notnull_short(O7_temp, Assembler::pt, L_ok);
2092         __ stop("bad method handle return");
2093         __ BIND(L_ok);
2094       }
2095 #endif //ASSERT
2096       __ JMP(O7_temp, 0);
2097       __ delayed()->nop();
2098     }
2099     break;
2100 
2101   case _adapter_opt_spread_0:
2102   case _adapter_opt_spread_1_ref:
2103   case _adapter_opt_spread_2_ref:
2104   case _adapter_opt_spread_3_ref:
2105   case _adapter_opt_spread_4_ref:
2106   case _adapter_opt_spread_5_ref:
2107   case _adapter_opt_spread_ref:
2108   case _adapter_opt_spread_byte:
2109   case _adapter_opt_spread_char:
2110   case _adapter_opt_spread_short:
2111   case _adapter_opt_spread_int:
2112   case _adapter_opt_spread_long:
2113   case _adapter_opt_spread_float:
2114   case _adapter_opt_spread_double:
2115     {
2116       // spread an array out into a group of arguments
2117       int  length_constant    = ek_adapter_opt_spread_count(ek);
2118       bool length_can_be_zero = (length_constant == 0);
2119       if (length_constant < 0) {
2120         // some adapters with variable length must handle the zero case
2121         if (!OptimizeMethodHandles ||
2122             ek_adapter_opt_spread_type(ek) != T_OBJECT)
2123           length_can_be_zero = true;
2124       }
2125 
2126       // find the address of the array argument
2127       load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
2128       __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
2129 
2130       // O0_argslot points both to the array and to the first output arg
2131       Address vmarg = Address(O0_argslot, 0);
2132 
2133       // Get the array value.
2134       Register  O1_array       = O1_scratch;
2135       Register  O2_array_klass = O2_scratch;
2136       BasicType elem_type      = ek_adapter_opt_spread_type(ek);
2137       int       elem_slots     = type2size[elem_type];  // 1 or 2
2138       int       array_slots    = 1;  // array is always a T_OBJECT
2139       int       length_offset  = arrayOopDesc::length_offset_in_bytes();
2140       int       elem0_offset   = arrayOopDesc::base_offset_in_bytes(elem_type);
2141       __ ld_ptr(vmarg, O1_array);
2142 
2143       Label L_array_is_empty, L_insert_arg_space, L_copy_args, L_args_done;
2144       if (length_can_be_zero) {
2145         // handle the null pointer case, if zero is allowed
2146         Label L_skip;
2147         if (length_constant < 0) {
2148           load_conversion_vminfo(_masm, G3_amh_conversion, O3_scratch);
2149           __ cmp_zero_and_br(Assembler::notZero, O3_scratch, L_skip);
2150           __ delayed()->nop(); // to avoid back-to-back cbcond instructions
2151         }
2152         __ br_null_short(O1_array, Assembler::pn, L_array_is_empty);
2153         __ BIND(L_skip);
2154       }
2155       __ null_check(O1_array, oopDesc::klass_offset_in_bytes());
2156       __ load_klass(O1_array, O2_array_klass);
2157 
2158       // Check the array type.
2159       Register O3_klass = O3_scratch;
2160       __ load_heap_oop(G3_amh_argument, O3_klass);  // this is a Class object!
2161       load_klass_from_Class(_masm, O3_klass, O4_scratch, G5_scratch);
2162 
2163       Label L_ok_array_klass, L_bad_array_klass, L_bad_array_length;
2164       __ check_klass_subtype(O2_array_klass, O3_klass, O4_scratch, G5_scratch, L_ok_array_klass);
2165       // If we get here, the type check failed!
2166       __ ba_short(L_bad_array_klass);
2167       __ BIND(L_ok_array_klass);
2168 
2169       // Check length.
2170       if (length_constant >= 0) {
2171         __ ldsw(Address(O1_array, length_offset), O4_scratch);
2172         __ cmp(O4_scratch, length_constant);
2173       } else {
2174         Register O3_vminfo = O3_scratch;
2175         load_conversion_vminfo(_masm, G3_amh_conversion, O3_vminfo);
2176         __ ldsw(Address(O1_array, length_offset), O4_scratch);
2177         __ cmp(O3_vminfo, O4_scratch);
2178       }
2179       __ br(Assembler::notEqual, false, Assembler::pn, L_bad_array_length);
2180       __ delayed()->nop();
2181 
2182       Register O2_argslot_limit = O2_scratch;
2183 
2184       // Array length checks out.  Now insert any required stack slots.
2185       if (length_constant == -1) {
2186         // Form a pointer to the end of the affected region.
2187         __ add(O0_argslot, Interpreter::stackElementSize, O2_argslot_limit);
2188         // 'stack_move' is negative number of words to insert
2189         // This number already accounts for elem_slots.
2190         Register O3_stack_move = O3_scratch;
2191         load_stack_move(_masm, G3_amh_conversion, O3_stack_move);
2192         __ cmp(O3_stack_move, 0);
2193         assert(stack_move_unit() < 0, "else change this comparison");
2194         __ br(Assembler::less, false, Assembler::pn, L_insert_arg_space);
2195         __ delayed()->nop();
2196         __ br(Assembler::equal, false, Assembler::pn, L_copy_args);
2197         __ delayed()->nop();
2198         // single argument case, with no array movement
2199         __ BIND(L_array_is_empty);
2200         remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2201                          O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2202         __ ba_short(L_args_done);  // no spreading to do
2203         __ BIND(L_insert_arg_space);
2204         // come here in the usual case, stack_move < 0 (2 or more spread arguments)
2205         // Live: O1_array, O2_argslot_limit, O3_stack_move
2206         insert_arg_slots(_masm, O3_stack_move,
2207                          O0_argslot, O4_scratch, G5_scratch, O1_scratch);
2208         // reload from rdx_argslot_limit since rax_argslot is now decremented
2209         __ ld_ptr(Address(O2_argslot_limit, -Interpreter::stackElementSize), O1_array);
2210       } else if (length_constant >= 1) {
2211         int new_slots = (length_constant * elem_slots) - array_slots;
2212         insert_arg_slots(_masm, new_slots * stack_move_unit(),
2213                          O0_argslot, O2_scratch, O3_scratch, O4_scratch);
2214       } else if (length_constant == 0) {
2215         __ BIND(L_array_is_empty);
2216         remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2217                          O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2218       } else {
2219         ShouldNotReachHere();
2220       }
2221 
2222       // Copy from the array to the new slots.
2223       // Note: Stack change code preserves integrity of O0_argslot pointer.
2224       // So even after slot insertions, O0_argslot still points to first argument.
2225       // Beware:  Arguments that are shallow on the stack are deep in the array,
2226       // and vice versa.  So a downward-growing stack (the usual) has to be copied
2227       // elementwise in reverse order from the source array.
2228       __ BIND(L_copy_args);
2229       if (length_constant == -1) {
2230         // [O0_argslot, O2_argslot_limit) is the area we are inserting into.
2231         // Array element [0] goes at O0_argslot_limit[-wordSize].
2232         Register O1_source = O1_array;
2233         __ add(Address(O1_array, elem0_offset), O1_source);
2234         Register O4_fill_ptr = O4_scratch;
2235         __ mov(O2_argslot_limit, O4_fill_ptr);
2236         Label L_loop;
2237         __ BIND(L_loop);
2238         __ add(O4_fill_ptr, -Interpreter::stackElementSize * elem_slots, O4_fill_ptr);
2239         move_typed_arg(_masm, elem_type, true,
2240                        Address(O1_source, 0), Address(O4_fill_ptr, 0),
2241                        O2_scratch);  // must be an even register for !_LP64 long moves (uses O2/O3)
2242         __ add(O1_source, type2aelembytes(elem_type), O1_source);
2243         __ cmp_and_brx_short(O4_fill_ptr, O0_argslot, Assembler::greaterUnsigned, Assembler::pt, L_loop);
2244       } else if (length_constant == 0) {
2245         // nothing to copy
2246       } else {
2247         int elem_offset = elem0_offset;
2248         int slot_offset = length_constant * Interpreter::stackElementSize;
2249         for (int index = 0; index < length_constant; index++) {
2250           slot_offset -= Interpreter::stackElementSize * elem_slots;  // fill backward
2251           move_typed_arg(_masm, elem_type, true,
2252                          Address(O1_array, elem_offset), Address(O0_argslot, slot_offset),
2253                          O2_scratch);  // must be an even register for !_LP64 long moves (uses O2/O3)
2254           elem_offset += type2aelembytes(elem_type);
2255         }
2256       }
2257       __ BIND(L_args_done);
2258 
2259       // Arguments are spread.  Move to next method handle.
2260       __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
2261       __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2262 
2263       __ BIND(L_bad_array_klass);
2264       assert(!vmarg.uses(O2_required), "must be different registers");
2265       __ load_heap_oop(Address(O2_array_klass, java_mirror_offset), O2_required);  // required class
2266       __ ld_ptr(       vmarg,                                       O1_actual);    // bad object
2267       __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2268       __ delayed()->mov(Bytecodes::_aaload,                         O0_code);      // who is complaining?
2269 
2270       __ bind(L_bad_array_length);
2271       assert(!vmarg.uses(O2_required), "must be different registers");
2272       __ mov(   G3_method_handle,                O2_required);  // required class
2273       __ ld_ptr(vmarg,                           O1_actual);    // bad object
2274       __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2275       __ delayed()->mov(Bytecodes::_arraylength, O0_code);      // who is complaining?
2276     }
2277     break;
2278 
2279   default:
2280     DEBUG_ONLY(tty->print_cr("bad ek=%d (%s)", (int)ek, entry_name(ek)));
2281     ShouldNotReachHere();
2282   }
2283   BLOCK_COMMENT(err_msg("} Entry %s", entry_name(ek)));
2284 
2285   address me_cookie = MethodHandleEntry::start_compiled_entry(_masm, interp_entry);
2286   __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
2287 
2288   init_entry(ek, MethodHandleEntry::finish_compiled_entry(_masm, me_cookie));
2289 }