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