1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * Copyright 2012, 2015 SAP AG. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "interpreter/interpreter.hpp" 29 #include "nativeInst_ppc.hpp" 30 #include "oops/instanceOop.hpp" 31 #include "oops/method.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "prims/methodHandles.hpp" 35 #include "runtime/frame.inline.hpp" 36 #include "runtime/handles.inline.hpp" 37 #include "runtime/sharedRuntime.hpp" 38 #include "runtime/stubCodeGenerator.hpp" 39 #include "runtime/stubRoutines.hpp" 40 #include "utilities/top.hpp" 41 #include "runtime/thread.inline.hpp" 42 43 #define __ _masm-> 44 45 #ifdef PRODUCT 46 #define BLOCK_COMMENT(str) // nothing 47 #else 48 #define BLOCK_COMMENT(str) __ block_comment(str) 49 #endif 50 51 class StubGenerator: public StubCodeGenerator { 52 private: 53 54 // Call stubs are used to call Java from C 55 // 56 // Arguments: 57 // 58 // R3 - call wrapper address : address 59 // R4 - result : intptr_t* 60 // R5 - result type : BasicType 61 // R6 - method : Method 62 // R7 - frame mgr entry point : address 63 // R8 - parameter block : intptr_t* 64 // R9 - parameter count in words : int 65 // R10 - thread : Thread* 66 // 67 address generate_call_stub(address& return_address) { 68 // Setup a new c frame, copy java arguments, call frame manager or 69 // native_entry, and process result. 70 71 StubCodeMark mark(this, "StubRoutines", "call_stub"); 72 73 address start = __ function_entry(); 74 75 // some sanity checks 76 assert((sizeof(frame::abi_minframe) % 16) == 0, "unaligned"); 77 assert((sizeof(frame::abi_reg_args) % 16) == 0, "unaligned"); 78 assert((sizeof(frame::spill_nonvolatiles) % 16) == 0, "unaligned"); 79 assert((sizeof(frame::parent_ijava_frame_abi) % 16) == 0, "unaligned"); 80 assert((sizeof(frame::entry_frame_locals) % 16) == 0, "unaligned"); 81 82 Register r_arg_call_wrapper_addr = R3; 83 Register r_arg_result_addr = R4; 84 Register r_arg_result_type = R5; 85 Register r_arg_method = R6; 86 Register r_arg_entry = R7; 87 Register r_arg_thread = R10; 88 89 Register r_temp = R24; 90 Register r_top_of_arguments_addr = R25; 91 Register r_entryframe_fp = R26; 92 93 { 94 // Stack on entry to call_stub: 95 // 96 // F1 [C_FRAME] 97 // ... 98 99 Register r_arg_argument_addr = R8; 100 Register r_arg_argument_count = R9; 101 Register r_frame_alignment_in_bytes = R27; 102 Register r_argument_addr = R28; 103 Register r_argumentcopy_addr = R29; 104 Register r_argument_size_in_bytes = R30; 105 Register r_frame_size = R23; 106 107 Label arguments_copied; 108 109 // Save LR/CR to caller's C_FRAME. 110 __ save_LR_CR(R0); 111 112 // Zero extend arg_argument_count. 113 __ clrldi(r_arg_argument_count, r_arg_argument_count, 32); 114 115 // Save non-volatiles GPRs to ENTRY_FRAME (not yet pushed, but it's safe). 116 __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14)); 117 118 // Keep copy of our frame pointer (caller's SP). 119 __ mr(r_entryframe_fp, R1_SP); 120 121 BLOCK_COMMENT("Push ENTRY_FRAME including arguments"); 122 // Push ENTRY_FRAME including arguments: 123 // 124 // F0 [TOP_IJAVA_FRAME_ABI] 125 // alignment (optional) 126 // [outgoing Java arguments] 127 // [ENTRY_FRAME_LOCALS] 128 // F1 [C_FRAME] 129 // ... 130 131 // calculate frame size 132 133 // unaligned size of arguments 134 __ sldi(r_argument_size_in_bytes, 135 r_arg_argument_count, Interpreter::logStackElementSize); 136 // arguments alignment (max 1 slot) 137 // FIXME: use round_to() here 138 __ andi_(r_frame_alignment_in_bytes, r_arg_argument_count, 1); 139 __ sldi(r_frame_alignment_in_bytes, 140 r_frame_alignment_in_bytes, Interpreter::logStackElementSize); 141 142 // size = unaligned size of arguments + top abi's size 143 __ addi(r_frame_size, r_argument_size_in_bytes, 144 frame::top_ijava_frame_abi_size); 145 // size += arguments alignment 146 __ add(r_frame_size, 147 r_frame_size, r_frame_alignment_in_bytes); 148 // size += size of call_stub locals 149 __ addi(r_frame_size, 150 r_frame_size, frame::entry_frame_locals_size); 151 152 // push ENTRY_FRAME 153 __ push_frame(r_frame_size, r_temp); 154 155 // initialize call_stub locals (step 1) 156 __ std(r_arg_call_wrapper_addr, 157 _entry_frame_locals_neg(call_wrapper_address), r_entryframe_fp); 158 __ std(r_arg_result_addr, 159 _entry_frame_locals_neg(result_address), r_entryframe_fp); 160 __ std(r_arg_result_type, 161 _entry_frame_locals_neg(result_type), r_entryframe_fp); 162 // we will save arguments_tos_address later 163 164 165 BLOCK_COMMENT("Copy Java arguments"); 166 // copy Java arguments 167 168 // Calculate top_of_arguments_addr which will be R17_tos (not prepushed) later. 169 // FIXME: why not simply use SP+frame::top_ijava_frame_size? 170 __ addi(r_top_of_arguments_addr, 171 R1_SP, frame::top_ijava_frame_abi_size); 172 __ add(r_top_of_arguments_addr, 173 r_top_of_arguments_addr, r_frame_alignment_in_bytes); 174 175 // any arguments to copy? 176 __ cmpdi(CCR0, r_arg_argument_count, 0); 177 __ beq(CCR0, arguments_copied); 178 179 // prepare loop and copy arguments in reverse order 180 { 181 // init CTR with arg_argument_count 182 __ mtctr(r_arg_argument_count); 183 184 // let r_argumentcopy_addr point to last outgoing Java arguments P 185 __ mr(r_argumentcopy_addr, r_top_of_arguments_addr); 186 187 // let r_argument_addr point to last incoming java argument 188 __ add(r_argument_addr, 189 r_arg_argument_addr, r_argument_size_in_bytes); 190 __ addi(r_argument_addr, r_argument_addr, -BytesPerWord); 191 192 // now loop while CTR > 0 and copy arguments 193 { 194 Label next_argument; 195 __ bind(next_argument); 196 197 __ ld(r_temp, 0, r_argument_addr); 198 // argument_addr--; 199 __ addi(r_argument_addr, r_argument_addr, -BytesPerWord); 200 __ std(r_temp, 0, r_argumentcopy_addr); 201 // argumentcopy_addr++; 202 __ addi(r_argumentcopy_addr, r_argumentcopy_addr, BytesPerWord); 203 204 __ bdnz(next_argument); 205 } 206 } 207 208 // Arguments copied, continue. 209 __ bind(arguments_copied); 210 } 211 212 { 213 BLOCK_COMMENT("Call frame manager or native entry."); 214 // Call frame manager or native entry. 215 Register r_new_arg_entry = R14; 216 assert_different_registers(r_new_arg_entry, r_top_of_arguments_addr, 217 r_arg_method, r_arg_thread); 218 219 __ mr(r_new_arg_entry, r_arg_entry); 220 221 // Register state on entry to frame manager / native entry: 222 // 223 // tos - intptr_t* sender tos (prepushed) Lesp = (SP) + copied_arguments_offset - 8 224 // R19_method - Method 225 // R16_thread - JavaThread* 226 227 // Tos must point to last argument - element_size. 228 #ifdef CC_INTERP 229 const Register tos = R17_tos; 230 #else 231 const Register tos = R15_esp; 232 #endif 233 __ addi(tos, r_top_of_arguments_addr, -Interpreter::stackElementSize); 234 235 // initialize call_stub locals (step 2) 236 // now save tos as arguments_tos_address 237 __ std(tos, _entry_frame_locals_neg(arguments_tos_address), r_entryframe_fp); 238 239 // load argument registers for call 240 __ mr(R19_method, r_arg_method); 241 __ mr(R16_thread, r_arg_thread); 242 assert(tos != r_arg_method, "trashed r_arg_method"); 243 assert(tos != r_arg_thread && R19_method != r_arg_thread, "trashed r_arg_thread"); 244 245 // Set R15_prev_state to 0 for simplifying checks in callee. 246 #ifdef CC_INTERP 247 __ li(R15_prev_state, 0); 248 #else 249 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 250 #endif 251 // Stack on entry to frame manager / native entry: 252 // 253 // F0 [TOP_IJAVA_FRAME_ABI] 254 // alignment (optional) 255 // [outgoing Java arguments] 256 // [ENTRY_FRAME_LOCALS] 257 // F1 [C_FRAME] 258 // ... 259 // 260 261 // global toc register 262 __ load_const(R29, MacroAssembler::global_toc(), R11_scratch1); 263 264 // Remember the senderSP so we interpreter can pop c2i arguments off of the stack 265 // when called via a c2i. 266 267 // Pass initial_caller_sp to framemanager. 268 __ mr(R21_tmp1, R1_SP); 269 270 // Do a light-weight C-call here, r_new_arg_entry holds the address 271 // of the interpreter entry point (frame manager or native entry) 272 // and save runtime-value of LR in return_address. 273 assert(r_new_arg_entry != tos && r_new_arg_entry != R19_method && r_new_arg_entry != R16_thread, 274 "trashed r_new_arg_entry"); 275 return_address = __ call_stub(r_new_arg_entry); 276 } 277 278 { 279 BLOCK_COMMENT("Returned from frame manager or native entry."); 280 // Returned from frame manager or native entry. 281 // Now pop frame, process result, and return to caller. 282 283 // Stack on exit from frame manager / native entry: 284 // 285 // F0 [ABI] 286 // ... 287 // [ENTRY_FRAME_LOCALS] 288 // F1 [C_FRAME] 289 // ... 290 // 291 // Just pop the topmost frame ... 292 // 293 294 Label ret_is_object; 295 Label ret_is_long; 296 Label ret_is_float; 297 Label ret_is_double; 298 299 Register r_entryframe_fp = R30; 300 Register r_lr = R7_ARG5; 301 Register r_cr = R8_ARG6; 302 303 // Reload some volatile registers which we've spilled before the call 304 // to frame manager / native entry. 305 // Access all locals via frame pointer, because we know nothing about 306 // the topmost frame's size. 307 __ ld(r_entryframe_fp, _abi(callers_sp), R1_SP); 308 assert_different_registers(r_entryframe_fp, R3_RET, r_arg_result_addr, r_arg_result_type, r_cr, r_lr); 309 __ ld(r_arg_result_addr, 310 _entry_frame_locals_neg(result_address), r_entryframe_fp); 311 __ ld(r_arg_result_type, 312 _entry_frame_locals_neg(result_type), r_entryframe_fp); 313 __ ld(r_cr, _abi(cr), r_entryframe_fp); 314 __ ld(r_lr, _abi(lr), r_entryframe_fp); 315 316 // pop frame and restore non-volatiles, LR and CR 317 __ mr(R1_SP, r_entryframe_fp); 318 __ mtcr(r_cr); 319 __ mtlr(r_lr); 320 321 // Store result depending on type. Everything that is not 322 // T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE is treated as T_INT. 323 __ cmpwi(CCR0, r_arg_result_type, T_OBJECT); 324 __ cmpwi(CCR1, r_arg_result_type, T_LONG); 325 __ cmpwi(CCR5, r_arg_result_type, T_FLOAT); 326 __ cmpwi(CCR6, r_arg_result_type, T_DOUBLE); 327 328 // restore non-volatile registers 329 __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14)); 330 331 332 // Stack on exit from call_stub: 333 // 334 // 0 [C_FRAME] 335 // ... 336 // 337 // no call_stub frames left. 338 339 // All non-volatiles have been restored at this point!! 340 assert(R3_RET == R3, "R3_RET should be R3"); 341 342 __ beq(CCR0, ret_is_object); 343 __ beq(CCR1, ret_is_long); 344 __ beq(CCR5, ret_is_float); 345 __ beq(CCR6, ret_is_double); 346 347 // default: 348 __ stw(R3_RET, 0, r_arg_result_addr); 349 __ blr(); // return to caller 350 351 // case T_OBJECT: 352 __ bind(ret_is_object); 353 __ std(R3_RET, 0, r_arg_result_addr); 354 __ blr(); // return to caller 355 356 // case T_LONG: 357 __ bind(ret_is_long); 358 __ std(R3_RET, 0, r_arg_result_addr); 359 __ blr(); // return to caller 360 361 // case T_FLOAT: 362 __ bind(ret_is_float); 363 __ stfs(F1_RET, 0, r_arg_result_addr); 364 __ blr(); // return to caller 365 366 // case T_DOUBLE: 367 __ bind(ret_is_double); 368 __ stfd(F1_RET, 0, r_arg_result_addr); 369 __ blr(); // return to caller 370 } 371 372 return start; 373 } 374 375 // Return point for a Java call if there's an exception thrown in 376 // Java code. The exception is caught and transformed into a 377 // pending exception stored in JavaThread that can be tested from 378 // within the VM. 379 // 380 address generate_catch_exception() { 381 StubCodeMark mark(this, "StubRoutines", "catch_exception"); 382 383 address start = __ pc(); 384 385 // Registers alive 386 // 387 // R16_thread 388 // R3_ARG1 - address of pending exception 389 // R4_ARG2 - return address in call stub 390 391 const Register exception_file = R21_tmp1; 392 const Register exception_line = R22_tmp2; 393 394 __ load_const(exception_file, (void*)__FILE__); 395 __ load_const(exception_line, (void*)__LINE__); 396 397 __ std(R3_ARG1, in_bytes(JavaThread::pending_exception_offset()), R16_thread); 398 // store into `char *' 399 __ std(exception_file, in_bytes(JavaThread::exception_file_offset()), R16_thread); 400 // store into `int' 401 __ stw(exception_line, in_bytes(JavaThread::exception_line_offset()), R16_thread); 402 403 // complete return to VM 404 assert(StubRoutines::_call_stub_return_address != NULL, "must have been generated before"); 405 406 __ mtlr(R4_ARG2); 407 // continue in call stub 408 __ blr(); 409 410 return start; 411 } 412 413 // Continuation point for runtime calls returning with a pending 414 // exception. The pending exception check happened in the runtime 415 // or native call stub. The pending exception in Thread is 416 // converted into a Java-level exception. 417 // 418 // Read: 419 // 420 // LR: The pc the runtime library callee wants to return to. 421 // Since the exception occurred in the callee, the return pc 422 // from the point of view of Java is the exception pc. 423 // thread: Needed for method handles. 424 // 425 // Invalidate: 426 // 427 // volatile registers (except below). 428 // 429 // Update: 430 // 431 // R4_ARG2: exception 432 // 433 // (LR is unchanged and is live out). 434 // 435 address generate_forward_exception() { 436 StubCodeMark mark(this, "StubRoutines", "forward_exception"); 437 address start = __ pc(); 438 439 #if !defined(PRODUCT) 440 if (VerifyOops) { 441 // Get pending exception oop. 442 __ ld(R3_ARG1, 443 in_bytes(Thread::pending_exception_offset()), 444 R16_thread); 445 // Make sure that this code is only executed if there is a pending exception. 446 { 447 Label L; 448 __ cmpdi(CCR0, R3_ARG1, 0); 449 __ bne(CCR0, L); 450 __ stop("StubRoutines::forward exception: no pending exception (1)"); 451 __ bind(L); 452 } 453 __ verify_oop(R3_ARG1, "StubRoutines::forward exception: not an oop"); 454 } 455 #endif 456 457 // Save LR/CR and copy exception pc (LR) into R4_ARG2. 458 __ save_LR_CR(R4_ARG2); 459 __ push_frame_reg_args(0, R0); 460 // Find exception handler. 461 __ call_VM_leaf(CAST_FROM_FN_PTR(address, 462 SharedRuntime::exception_handler_for_return_address), 463 R16_thread, 464 R4_ARG2); 465 // Copy handler's address. 466 __ mtctr(R3_RET); 467 __ pop_frame(); 468 __ restore_LR_CR(R0); 469 470 // Set up the arguments for the exception handler: 471 // - R3_ARG1: exception oop 472 // - R4_ARG2: exception pc. 473 474 // Load pending exception oop. 475 __ ld(R3_ARG1, 476 in_bytes(Thread::pending_exception_offset()), 477 R16_thread); 478 479 // The exception pc is the return address in the caller. 480 // Must load it into R4_ARG2. 481 __ mflr(R4_ARG2); 482 483 #ifdef ASSERT 484 // Make sure exception is set. 485 { 486 Label L; 487 __ cmpdi(CCR0, R3_ARG1, 0); 488 __ bne(CCR0, L); 489 __ stop("StubRoutines::forward exception: no pending exception (2)"); 490 __ bind(L); 491 } 492 #endif 493 494 // Clear the pending exception. 495 __ li(R0, 0); 496 __ std(R0, 497 in_bytes(Thread::pending_exception_offset()), 498 R16_thread); 499 // Jump to exception handler. 500 __ bctr(); 501 502 return start; 503 } 504 505 #undef __ 506 #define __ masm-> 507 // Continuation point for throwing of implicit exceptions that are 508 // not handled in the current activation. Fabricates an exception 509 // oop and initiates normal exception dispatching in this 510 // frame. Only callee-saved registers are preserved (through the 511 // normal register window / RegisterMap handling). If the compiler 512 // needs all registers to be preserved between the fault point and 513 // the exception handler then it must assume responsibility for that 514 // in AbstractCompiler::continuation_for_implicit_null_exception or 515 // continuation_for_implicit_division_by_zero_exception. All other 516 // implicit exceptions (e.g., NullPointerException or 517 // AbstractMethodError on entry) are either at call sites or 518 // otherwise assume that stack unwinding will be initiated, so 519 // caller saved registers were assumed volatile in the compiler. 520 // 521 // Note that we generate only this stub into a RuntimeStub, because 522 // it needs to be properly traversed and ignored during GC, so we 523 // change the meaning of the "__" macro within this method. 524 // 525 // Note: the routine set_pc_not_at_call_for_caller in 526 // SharedRuntime.cpp requires that this code be generated into a 527 // RuntimeStub. 528 address generate_throw_exception(const char* name, address runtime_entry, bool restore_saved_exception_pc, 529 Register arg1 = noreg, Register arg2 = noreg) { 530 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0); 531 MacroAssembler* masm = new MacroAssembler(&code); 532 533 OopMapSet* oop_maps = new OopMapSet(); 534 int frame_size_in_bytes = frame::abi_reg_args_size; 535 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0); 536 537 address start = __ pc(); 538 539 __ save_LR_CR(R11_scratch1); 540 541 // Push a frame. 542 __ push_frame_reg_args(0, R11_scratch1); 543 544 address frame_complete_pc = __ pc(); 545 546 if (restore_saved_exception_pc) { 547 __ unimplemented("StubGenerator::throw_exception with restore_saved_exception_pc", 74); 548 } 549 550 // Note that we always have a runtime stub frame on the top of 551 // stack by this point. Remember the offset of the instruction 552 // whose address will be moved to R11_scratch1. 553 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1); 554 555 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 556 557 __ mr(R3_ARG1, R16_thread); 558 if (arg1 != noreg) { 559 __ mr(R4_ARG2, arg1); 560 } 561 if (arg2 != noreg) { 562 __ mr(R5_ARG3, arg2); 563 } 564 #if defined(ABI_ELFv2) 565 __ call_c(runtime_entry, relocInfo::none); 566 #else 567 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, runtime_entry), relocInfo::none); 568 #endif 569 570 // Set an oopmap for the call site. 571 oop_maps->add_gc_map((int)(gc_map_pc - start), map); 572 573 __ reset_last_Java_frame(); 574 575 #ifdef ASSERT 576 // Make sure that this code is only executed if there is a pending 577 // exception. 578 { 579 Label L; 580 __ ld(R0, 581 in_bytes(Thread::pending_exception_offset()), 582 R16_thread); 583 __ cmpdi(CCR0, R0, 0); 584 __ bne(CCR0, L); 585 __ stop("StubRoutines::throw_exception: no pending exception"); 586 __ bind(L); 587 } 588 #endif 589 590 // Pop frame. 591 __ pop_frame(); 592 593 __ restore_LR_CR(R11_scratch1); 594 595 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry()); 596 __ mtctr(R11_scratch1); 597 __ bctr(); 598 599 // Create runtime stub with OopMap. 600 RuntimeStub* stub = 601 RuntimeStub::new_runtime_stub(name, &code, 602 /*frame_complete=*/ (int)(frame_complete_pc - start), 603 frame_size_in_bytes/wordSize, 604 oop_maps, 605 false); 606 return stub->entry_point(); 607 } 608 #undef __ 609 #define __ _masm-> 610 611 // Generate G1 pre-write barrier for array. 612 // 613 // Input: 614 // from - register containing src address (only needed for spilling) 615 // to - register containing starting address 616 // count - register containing element count 617 // tmp - scratch register 618 // 619 // Kills: 620 // nothing 621 // 622 void gen_write_ref_array_pre_barrier(Register from, Register to, Register count, bool dest_uninitialized, Register Rtmp1) { 623 BarrierSet* const bs = Universe::heap()->barrier_set(); 624 switch (bs->kind()) { 625 case BarrierSet::G1SATBCTLogging: 626 // With G1, don't generate the call if we statically know that the target in uninitialized 627 if (!dest_uninitialized) { 628 const int spill_slots = 4 * wordSize; 629 const int frame_size = frame::abi_reg_args_size + spill_slots; 630 Label filtered; 631 632 // Is marking active? 633 if (in_bytes(PtrQueue::byte_width_of_active()) == 4) { 634 __ lwz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread); 635 } else { 636 guarantee(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption"); 637 __ lbz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread); 638 } 639 __ cmpdi(CCR0, Rtmp1, 0); 640 __ beq(CCR0, filtered); 641 642 __ save_LR_CR(R0); 643 __ push_frame_reg_args(spill_slots, R0); 644 __ std(from, frame_size - 1 * wordSize, R1_SP); 645 __ std(to, frame_size - 2 * wordSize, R1_SP); 646 __ std(count, frame_size - 3 * wordSize, R1_SP); 647 648 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), to, count); 649 650 __ ld(from, frame_size - 1 * wordSize, R1_SP); 651 __ ld(to, frame_size - 2 * wordSize, R1_SP); 652 __ ld(count, frame_size - 3 * wordSize, R1_SP); 653 __ pop_frame(); 654 __ restore_LR_CR(R0); 655 656 __ bind(filtered); 657 } 658 break; 659 case BarrierSet::CardTableModRef: 660 case BarrierSet::CardTableExtension: 661 case BarrierSet::ModRef: 662 break; 663 default: 664 ShouldNotReachHere(); 665 } 666 } 667 668 // Generate CMS/G1 post-write barrier for array. 669 // 670 // Input: 671 // addr - register containing starting address 672 // count - register containing element count 673 // tmp - scratch register 674 // 675 // The input registers and R0 are overwritten. 676 // 677 void gen_write_ref_array_post_barrier(Register addr, Register count, Register tmp, bool branchToEnd) { 678 BarrierSet* const bs = Universe::heap()->barrier_set(); 679 680 switch (bs->kind()) { 681 case BarrierSet::G1SATBCTLogging: 682 { 683 if (branchToEnd) { 684 __ save_LR_CR(R0); 685 // We need this frame only to spill LR. 686 __ push_frame_reg_args(0, R0); 687 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), addr, count); 688 __ pop_frame(); 689 __ restore_LR_CR(R0); 690 } else { 691 // Tail call: fake call from stub caller by branching without linking. 692 address entry_point = (address)CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post); 693 __ mr_if_needed(R3_ARG1, addr); 694 __ mr_if_needed(R4_ARG2, count); 695 __ load_const(R11, entry_point, R0); 696 __ call_c_and_return_to_caller(R11); 697 } 698 } 699 break; 700 case BarrierSet::CardTableModRef: 701 case BarrierSet::CardTableExtension: 702 { 703 Label Lskip_loop, Lstore_loop; 704 if (UseConcMarkSweepGC) { 705 // TODO PPC port: contribute optimization / requires shared changes 706 __ release(); 707 } 708 709 CardTableModRefBS* const ct = barrier_set_cast<CardTableModRefBS>(bs); 710 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 711 assert_different_registers(addr, count, tmp); 712 713 __ sldi(count, count, LogBytesPerHeapOop); 714 __ addi(count, count, -BytesPerHeapOop); 715 __ add(count, addr, count); 716 // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.) 717 __ srdi(addr, addr, CardTableModRefBS::card_shift); 718 __ srdi(count, count, CardTableModRefBS::card_shift); 719 __ subf(count, addr, count); 720 assert_different_registers(R0, addr, count, tmp); 721 __ load_const(tmp, (address)ct->byte_map_base); 722 __ addic_(count, count, 1); 723 __ beq(CCR0, Lskip_loop); 724 __ li(R0, 0); 725 __ mtctr(count); 726 // Byte store loop 727 __ bind(Lstore_loop); 728 __ stbx(R0, tmp, addr); 729 __ addi(addr, addr, 1); 730 __ bdnz(Lstore_loop); 731 __ bind(Lskip_loop); 732 733 if (!branchToEnd) __ blr(); 734 } 735 break; 736 case BarrierSet::ModRef: 737 if (!branchToEnd) __ blr(); 738 break; 739 default: 740 ShouldNotReachHere(); 741 } 742 } 743 744 // Support for void zero_words_aligned8(HeapWord* to, size_t count) 745 // 746 // Arguments: 747 // to: 748 // count: 749 // 750 // Destroys: 751 // 752 address generate_zero_words_aligned8() { 753 StubCodeMark mark(this, "StubRoutines", "zero_words_aligned8"); 754 755 // Implemented as in ClearArray. 756 address start = __ function_entry(); 757 758 Register base_ptr_reg = R3_ARG1; // tohw (needs to be 8b aligned) 759 Register cnt_dwords_reg = R4_ARG2; // count (in dwords) 760 Register tmp1_reg = R5_ARG3; 761 Register tmp2_reg = R6_ARG4; 762 Register zero_reg = R7_ARG5; 763 764 // Procedure for large arrays (uses data cache block zero instruction). 765 Label dwloop, fast, fastloop, restloop, lastdword, done; 766 int cl_size=VM_Version::get_cache_line_size(), cl_dwords=cl_size>>3, cl_dwordaddr_bits=exact_log2(cl_dwords); 767 int min_dcbz=2; // Needs to be positive, apply dcbz only to at least min_dcbz cache lines. 768 769 // Clear up to 128byte boundary if long enough, dword_cnt=(16-(base>>3))%16. 770 __ dcbtst(base_ptr_reg); // Indicate write access to first cache line ... 771 __ andi(tmp2_reg, cnt_dwords_reg, 1); // to check if number of dwords is even. 772 __ srdi_(tmp1_reg, cnt_dwords_reg, 1); // number of double dwords 773 __ load_const_optimized(zero_reg, 0L); // Use as zero register. 774 775 __ cmpdi(CCR1, tmp2_reg, 0); // cnt_dwords even? 776 __ beq(CCR0, lastdword); // size <= 1 777 __ mtctr(tmp1_reg); // Speculatively preload counter for rest loop (>0). 778 __ cmpdi(CCR0, cnt_dwords_reg, (min_dcbz+1)*cl_dwords-1); // Big enough to ensure >=min_dcbz cache lines are included? 779 __ neg(tmp1_reg, base_ptr_reg); // bit 0..58: bogus, bit 57..60: (16-(base>>3))%16, bit 61..63: 000 780 781 __ blt(CCR0, restloop); // Too small. (<31=(2*cl_dwords)-1 is sufficient, but bigger performs better.) 782 __ rldicl_(tmp1_reg, tmp1_reg, 64-3, 64-cl_dwordaddr_bits); // Extract number of dwords to 128byte boundary=(16-(base>>3))%16. 783 784 __ beq(CCR0, fast); // already 128byte aligned 785 __ mtctr(tmp1_reg); // Set ctr to hit 128byte boundary (0<ctr<cnt). 786 __ subf(cnt_dwords_reg, tmp1_reg, cnt_dwords_reg); // rest (>0 since size>=256-8) 787 788 // Clear in first cache line dword-by-dword if not already 128byte aligned. 789 __ bind(dwloop); 790 __ std(zero_reg, 0, base_ptr_reg); // Clear 8byte aligned block. 791 __ addi(base_ptr_reg, base_ptr_reg, 8); 792 __ bdnz(dwloop); 793 794 // clear 128byte blocks 795 __ bind(fast); 796 __ srdi(tmp1_reg, cnt_dwords_reg, cl_dwordaddr_bits); // loop count for 128byte loop (>0 since size>=256-8) 797 __ andi(tmp2_reg, cnt_dwords_reg, 1); // to check if rest even 798 799 __ mtctr(tmp1_reg); // load counter 800 __ cmpdi(CCR1, tmp2_reg, 0); // rest even? 801 __ rldicl_(tmp1_reg, cnt_dwords_reg, 63, 65-cl_dwordaddr_bits); // rest in double dwords 802 803 __ bind(fastloop); 804 __ dcbz(base_ptr_reg); // Clear 128byte aligned block. 805 __ addi(base_ptr_reg, base_ptr_reg, cl_size); 806 __ bdnz(fastloop); 807 808 //__ dcbtst(base_ptr_reg); // Indicate write access to last cache line. 809 __ beq(CCR0, lastdword); // rest<=1 810 __ mtctr(tmp1_reg); // load counter 811 812 // Clear rest. 813 __ bind(restloop); 814 __ std(zero_reg, 0, base_ptr_reg); // Clear 8byte aligned block. 815 __ std(zero_reg, 8, base_ptr_reg); // Clear 8byte aligned block. 816 __ addi(base_ptr_reg, base_ptr_reg, 16); 817 __ bdnz(restloop); 818 819 __ bind(lastdword); 820 __ beq(CCR1, done); 821 __ std(zero_reg, 0, base_ptr_reg); 822 __ bind(done); 823 __ blr(); // return 824 825 return start; 826 } 827 828 // The following routine generates a subroutine to throw an asynchronous 829 // UnknownError when an unsafe access gets a fault that could not be 830 // reasonably prevented by the programmer. (Example: SIGBUS/OBJERR.) 831 // 832 address generate_handler_for_unsafe_access() { 833 StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access"); 834 address start = __ function_entry(); 835 __ unimplemented("StubRoutines::handler_for_unsafe_access", 93); 836 return start; 837 } 838 839 #if !defined(PRODUCT) 840 // Wrapper which calls oopDesc::is_oop_or_null() 841 // Only called by MacroAssembler::verify_oop 842 static void verify_oop_helper(const char* message, oop o) { 843 if (!o->is_oop_or_null()) { 844 fatal(message); 845 } 846 ++ StubRoutines::_verify_oop_count; 847 } 848 #endif 849 850 // Return address of code to be called from code generated by 851 // MacroAssembler::verify_oop. 852 // 853 // Don't generate, rather use C++ code. 854 address generate_verify_oop() { 855 // this is actually a `FunctionDescriptor*'. 856 address start = 0; 857 858 #if !defined(PRODUCT) 859 start = CAST_FROM_FN_PTR(address, verify_oop_helper); 860 #endif 861 862 return start; 863 } 864 865 // Fairer handling of safepoints for native methods. 866 // 867 // Generate code which reads from the polling page. This special handling is needed as the 868 // linux-ppc64 kernel before 2.6.6 doesn't set si_addr on some segfaults in 64bit mode 869 // (cf. http://www.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.6), especially when we try 870 // to read from the safepoint polling page. 871 address generate_load_from_poll() { 872 StubCodeMark mark(this, "StubRoutines", "generate_load_from_poll"); 873 address start = __ function_entry(); 874 __ unimplemented("StubRoutines::verify_oop", 95); // TODO PPC port 875 return start; 876 } 877 878 // -XX:+OptimizeFill : convert fill/copy loops into intrinsic 879 // 880 // The code is implemented(ported from sparc) as we believe it benefits JVM98, however 881 // tracing(-XX:+TraceOptimizeFill) shows the intrinsic replacement doesn't happen at all! 882 // 883 // Source code in function is_range_check_if() shows that OptimizeFill relaxed the condition 884 // for turning on loop predication optimization, and hence the behavior of "array range check" 885 // and "loop invariant check" could be influenced, which potentially boosted JVM98. 886 // 887 // Generate stub for disjoint short fill. If "aligned" is true, the 888 // "to" address is assumed to be heapword aligned. 889 // 890 // Arguments for generated stub: 891 // to: R3_ARG1 892 // value: R4_ARG2 893 // count: R5_ARG3 treated as signed 894 // 895 address generate_fill(BasicType t, bool aligned, const char* name) { 896 StubCodeMark mark(this, "StubRoutines", name); 897 address start = __ function_entry(); 898 899 const Register to = R3_ARG1; // source array address 900 const Register value = R4_ARG2; // fill value 901 const Register count = R5_ARG3; // elements count 902 const Register temp = R6_ARG4; // temp register 903 904 //assert_clean_int(count, O3); // Make sure 'count' is clean int. 905 906 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; 907 Label L_fill_2_bytes, L_fill_4_bytes, L_fill_elements, L_fill_32_bytes; 908 909 int shift = -1; 910 switch (t) { 911 case T_BYTE: 912 shift = 2; 913 // Clone bytes (zero extend not needed because store instructions below ignore high order bytes). 914 __ rldimi(value, value, 8, 48); // 8 bit -> 16 bit 915 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element. 916 __ blt(CCR0, L_fill_elements); 917 __ rldimi(value, value, 16, 32); // 16 bit -> 32 bit 918 break; 919 case T_SHORT: 920 shift = 1; 921 // Clone bytes (zero extend not needed because store instructions below ignore high order bytes). 922 __ rldimi(value, value, 16, 32); // 16 bit -> 32 bit 923 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element. 924 __ blt(CCR0, L_fill_elements); 925 break; 926 case T_INT: 927 shift = 0; 928 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element. 929 __ blt(CCR0, L_fill_4_bytes); 930 break; 931 default: ShouldNotReachHere(); 932 } 933 934 if (!aligned && (t == T_BYTE || t == T_SHORT)) { 935 // Align source address at 4 bytes address boundary. 936 if (t == T_BYTE) { 937 // One byte misalignment happens only for byte arrays. 938 __ andi_(temp, to, 1); 939 __ beq(CCR0, L_skip_align1); 940 __ stb(value, 0, to); 941 __ addi(to, to, 1); 942 __ addi(count, count, -1); 943 __ bind(L_skip_align1); 944 } 945 // Two bytes misalignment happens only for byte and short (char) arrays. 946 __ andi_(temp, to, 2); 947 __ beq(CCR0, L_skip_align2); 948 __ sth(value, 0, to); 949 __ addi(to, to, 2); 950 __ addi(count, count, -(1 << (shift - 1))); 951 __ bind(L_skip_align2); 952 } 953 954 if (!aligned) { 955 // Align to 8 bytes, we know we are 4 byte aligned to start. 956 __ andi_(temp, to, 7); 957 __ beq(CCR0, L_fill_32_bytes); 958 __ stw(value, 0, to); 959 __ addi(to, to, 4); 960 __ addi(count, count, -(1 << shift)); 961 __ bind(L_fill_32_bytes); 962 } 963 964 __ li(temp, 8<<shift); // Prepare for 32 byte loop. 965 // Clone bytes int->long as above. 966 __ rldimi(value, value, 32, 0); // 32 bit -> 64 bit 967 968 Label L_check_fill_8_bytes; 969 // Fill 32-byte chunks. 970 __ subf_(count, temp, count); 971 __ blt(CCR0, L_check_fill_8_bytes); 972 973 Label L_fill_32_bytes_loop; 974 __ align(32); 975 __ bind(L_fill_32_bytes_loop); 976 977 __ std(value, 0, to); 978 __ std(value, 8, to); 979 __ subf_(count, temp, count); // Update count. 980 __ std(value, 16, to); 981 __ std(value, 24, to); 982 983 __ addi(to, to, 32); 984 __ bge(CCR0, L_fill_32_bytes_loop); 985 986 __ bind(L_check_fill_8_bytes); 987 __ add_(count, temp, count); 988 __ beq(CCR0, L_exit); 989 __ addic_(count, count, -(2 << shift)); 990 __ blt(CCR0, L_fill_4_bytes); 991 992 // 993 // Length is too short, just fill 8 bytes at a time. 994 // 995 Label L_fill_8_bytes_loop; 996 __ bind(L_fill_8_bytes_loop); 997 __ std(value, 0, to); 998 __ addic_(count, count, -(2 << shift)); 999 __ addi(to, to, 8); 1000 __ bge(CCR0, L_fill_8_bytes_loop); 1001 1002 // Fill trailing 4 bytes. 1003 __ bind(L_fill_4_bytes); 1004 __ andi_(temp, count, 1<<shift); 1005 __ beq(CCR0, L_fill_2_bytes); 1006 1007 __ stw(value, 0, to); 1008 if (t == T_BYTE || t == T_SHORT) { 1009 __ addi(to, to, 4); 1010 // Fill trailing 2 bytes. 1011 __ bind(L_fill_2_bytes); 1012 __ andi_(temp, count, 1<<(shift-1)); 1013 __ beq(CCR0, L_fill_byte); 1014 __ sth(value, 0, to); 1015 if (t == T_BYTE) { 1016 __ addi(to, to, 2); 1017 // Fill trailing byte. 1018 __ bind(L_fill_byte); 1019 __ andi_(count, count, 1); 1020 __ beq(CCR0, L_exit); 1021 __ stb(value, 0, to); 1022 } else { 1023 __ bind(L_fill_byte); 1024 } 1025 } else { 1026 __ bind(L_fill_2_bytes); 1027 } 1028 __ bind(L_exit); 1029 __ blr(); 1030 1031 // Handle copies less than 8 bytes. Int is handled elsewhere. 1032 if (t == T_BYTE) { 1033 __ bind(L_fill_elements); 1034 Label L_fill_2, L_fill_4; 1035 __ andi_(temp, count, 1); 1036 __ beq(CCR0, L_fill_2); 1037 __ stb(value, 0, to); 1038 __ addi(to, to, 1); 1039 __ bind(L_fill_2); 1040 __ andi_(temp, count, 2); 1041 __ beq(CCR0, L_fill_4); 1042 __ stb(value, 0, to); 1043 __ stb(value, 0, to); 1044 __ addi(to, to, 2); 1045 __ bind(L_fill_4); 1046 __ andi_(temp, count, 4); 1047 __ beq(CCR0, L_exit); 1048 __ stb(value, 0, to); 1049 __ stb(value, 1, to); 1050 __ stb(value, 2, to); 1051 __ stb(value, 3, to); 1052 __ blr(); 1053 } 1054 1055 if (t == T_SHORT) { 1056 Label L_fill_2; 1057 __ bind(L_fill_elements); 1058 __ andi_(temp, count, 1); 1059 __ beq(CCR0, L_fill_2); 1060 __ sth(value, 0, to); 1061 __ addi(to, to, 2); 1062 __ bind(L_fill_2); 1063 __ andi_(temp, count, 2); 1064 __ beq(CCR0, L_exit); 1065 __ sth(value, 0, to); 1066 __ sth(value, 2, to); 1067 __ blr(); 1068 } 1069 return start; 1070 } 1071 1072 1073 // Generate overlap test for array copy stubs. 1074 // 1075 // Input: 1076 // R3_ARG1 - from 1077 // R4_ARG2 - to 1078 // R5_ARG3 - element count 1079 // 1080 void array_overlap_test(address no_overlap_target, int log2_elem_size) { 1081 Register tmp1 = R6_ARG4; 1082 Register tmp2 = R7_ARG5; 1083 1084 Label l_overlap; 1085 #ifdef ASSERT 1086 __ srdi_(tmp2, R5_ARG3, 31); 1087 __ asm_assert_eq("missing zero extend", 0xAFFE); 1088 #endif 1089 1090 __ subf(tmp1, R3_ARG1, R4_ARG2); // distance in bytes 1091 __ sldi(tmp2, R5_ARG3, log2_elem_size); // size in bytes 1092 __ cmpld(CCR0, R3_ARG1, R4_ARG2); // Use unsigned comparison! 1093 __ cmpld(CCR1, tmp1, tmp2); 1094 __ crand(CCR0, Assembler::less, CCR1, Assembler::less); 1095 __ blt(CCR0, l_overlap); // Src before dst and distance smaller than size. 1096 1097 // need to copy forwards 1098 if (__ is_within_range_of_b(no_overlap_target, __ pc())) { 1099 __ b(no_overlap_target); 1100 } else { 1101 __ load_const(tmp1, no_overlap_target, tmp2); 1102 __ mtctr(tmp1); 1103 __ bctr(); 1104 } 1105 1106 __ bind(l_overlap); 1107 // need to copy backwards 1108 } 1109 1110 // The guideline in the implementations of generate_disjoint_xxx_copy 1111 // (xxx=byte,short,int,long,oop) is to copy as many elements as possible with 1112 // single instructions, but to avoid alignment interrupts (see subsequent 1113 // comment). Furthermore, we try to minimize misaligned access, even 1114 // though they cause no alignment interrupt. 1115 // 1116 // In Big-Endian mode, the PowerPC architecture requires implementations to 1117 // handle automatically misaligned integer halfword and word accesses, 1118 // word-aligned integer doubleword accesses, and word-aligned floating-point 1119 // accesses. Other accesses may or may not generate an Alignment interrupt 1120 // depending on the implementation. 1121 // Alignment interrupt handling may require on the order of hundreds of cycles, 1122 // so every effort should be made to avoid misaligned memory values. 1123 // 1124 // 1125 // Generate stub for disjoint byte copy. If "aligned" is true, the 1126 // "from" and "to" addresses are assumed to be heapword aligned. 1127 // 1128 // Arguments for generated stub: 1129 // from: R3_ARG1 1130 // to: R4_ARG2 1131 // count: R5_ARG3 treated as signed 1132 // 1133 address generate_disjoint_byte_copy(bool aligned, const char * name) { 1134 StubCodeMark mark(this, "StubRoutines", name); 1135 address start = __ function_entry(); 1136 1137 Register tmp1 = R6_ARG4; 1138 Register tmp2 = R7_ARG5; 1139 Register tmp3 = R8_ARG6; 1140 Register tmp4 = R9_ARG7; 1141 1142 1143 Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8, l_9; 1144 // Don't try anything fancy if arrays don't have many elements. 1145 __ li(tmp3, 0); 1146 __ cmpwi(CCR0, R5_ARG3, 17); 1147 __ ble(CCR0, l_6); // copy 4 at a time 1148 1149 if (!aligned) { 1150 __ xorr(tmp1, R3_ARG1, R4_ARG2); 1151 __ andi_(tmp1, tmp1, 3); 1152 __ bne(CCR0, l_6); // If arrays don't have the same alignment mod 4, do 4 element copy. 1153 1154 // Copy elements if necessary to align to 4 bytes. 1155 __ neg(tmp1, R3_ARG1); // Compute distance to alignment boundary. 1156 __ andi_(tmp1, tmp1, 3); 1157 __ beq(CCR0, l_2); 1158 1159 __ subf(R5_ARG3, tmp1, R5_ARG3); 1160 __ bind(l_9); 1161 __ lbz(tmp2, 0, R3_ARG1); 1162 __ addic_(tmp1, tmp1, -1); 1163 __ stb(tmp2, 0, R4_ARG2); 1164 __ addi(R3_ARG1, R3_ARG1, 1); 1165 __ addi(R4_ARG2, R4_ARG2, 1); 1166 __ bne(CCR0, l_9); 1167 1168 __ bind(l_2); 1169 } 1170 1171 // copy 8 elements at a time 1172 __ xorr(tmp2, R3_ARG1, R4_ARG2); // skip if src & dest have differing alignment mod 8 1173 __ andi_(tmp1, tmp2, 7); 1174 __ bne(CCR0, l_7); // not same alignment -> to or from is aligned -> copy 8 1175 1176 // copy a 2-element word if necessary to align to 8 bytes 1177 __ andi_(R0, R3_ARG1, 7); 1178 __ beq(CCR0, l_7); 1179 1180 __ lwzx(tmp2, R3_ARG1, tmp3); 1181 __ addi(R5_ARG3, R5_ARG3, -4); 1182 __ stwx(tmp2, R4_ARG2, tmp3); 1183 { // FasterArrayCopy 1184 __ addi(R3_ARG1, R3_ARG1, 4); 1185 __ addi(R4_ARG2, R4_ARG2, 4); 1186 } 1187 __ bind(l_7); 1188 1189 { // FasterArrayCopy 1190 __ cmpwi(CCR0, R5_ARG3, 31); 1191 __ ble(CCR0, l_6); // copy 2 at a time if less than 32 elements remain 1192 1193 __ srdi(tmp1, R5_ARG3, 5); 1194 __ andi_(R5_ARG3, R5_ARG3, 31); 1195 __ mtctr(tmp1); 1196 1197 __ bind(l_8); 1198 // Use unrolled version for mass copying (copy 32 elements a time) 1199 // Load feeding store gets zero latency on Power6, however not on Power5. 1200 // Therefore, the following sequence is made for the good of both. 1201 __ ld(tmp1, 0, R3_ARG1); 1202 __ ld(tmp2, 8, R3_ARG1); 1203 __ ld(tmp3, 16, R3_ARG1); 1204 __ ld(tmp4, 24, R3_ARG1); 1205 __ std(tmp1, 0, R4_ARG2); 1206 __ std(tmp2, 8, R4_ARG2); 1207 __ std(tmp3, 16, R4_ARG2); 1208 __ std(tmp4, 24, R4_ARG2); 1209 __ addi(R3_ARG1, R3_ARG1, 32); 1210 __ addi(R4_ARG2, R4_ARG2, 32); 1211 __ bdnz(l_8); 1212 } 1213 1214 __ bind(l_6); 1215 1216 // copy 4 elements at a time 1217 __ cmpwi(CCR0, R5_ARG3, 4); 1218 __ blt(CCR0, l_1); 1219 __ srdi(tmp1, R5_ARG3, 2); 1220 __ mtctr(tmp1); // is > 0 1221 __ andi_(R5_ARG3, R5_ARG3, 3); 1222 1223 { // FasterArrayCopy 1224 __ addi(R3_ARG1, R3_ARG1, -4); 1225 __ addi(R4_ARG2, R4_ARG2, -4); 1226 __ bind(l_3); 1227 __ lwzu(tmp2, 4, R3_ARG1); 1228 __ stwu(tmp2, 4, R4_ARG2); 1229 __ bdnz(l_3); 1230 __ addi(R3_ARG1, R3_ARG1, 4); 1231 __ addi(R4_ARG2, R4_ARG2, 4); 1232 } 1233 1234 // do single element copy 1235 __ bind(l_1); 1236 __ cmpwi(CCR0, R5_ARG3, 0); 1237 __ beq(CCR0, l_4); 1238 1239 { // FasterArrayCopy 1240 __ mtctr(R5_ARG3); 1241 __ addi(R3_ARG1, R3_ARG1, -1); 1242 __ addi(R4_ARG2, R4_ARG2, -1); 1243 1244 __ bind(l_5); 1245 __ lbzu(tmp2, 1, R3_ARG1); 1246 __ stbu(tmp2, 1, R4_ARG2); 1247 __ bdnz(l_5); 1248 } 1249 1250 __ bind(l_4); 1251 __ blr(); 1252 1253 return start; 1254 } 1255 1256 // Generate stub for conjoint byte copy. If "aligned" is true, the 1257 // "from" and "to" addresses are assumed to be heapword aligned. 1258 // 1259 // Arguments for generated stub: 1260 // from: R3_ARG1 1261 // to: R4_ARG2 1262 // count: R5_ARG3 treated as signed 1263 // 1264 address generate_conjoint_byte_copy(bool aligned, const char * name) { 1265 StubCodeMark mark(this, "StubRoutines", name); 1266 address start = __ function_entry(); 1267 1268 Register tmp1 = R6_ARG4; 1269 Register tmp2 = R7_ARG5; 1270 Register tmp3 = R8_ARG6; 1271 1272 #if defined(ABI_ELFv2) 1273 address nooverlap_target = aligned ? 1274 StubRoutines::arrayof_jbyte_disjoint_arraycopy() : 1275 StubRoutines::jbyte_disjoint_arraycopy(); 1276 #else 1277 address nooverlap_target = aligned ? 1278 ((FunctionDescriptor*)StubRoutines::arrayof_jbyte_disjoint_arraycopy())->entry() : 1279 ((FunctionDescriptor*)StubRoutines::jbyte_disjoint_arraycopy())->entry(); 1280 #endif 1281 1282 array_overlap_test(nooverlap_target, 0); 1283 // Do reverse copy. We assume the case of actual overlap is rare enough 1284 // that we don't have to optimize it. 1285 Label l_1, l_2; 1286 1287 __ b(l_2); 1288 __ bind(l_1); 1289 __ stbx(tmp1, R4_ARG2, R5_ARG3); 1290 __ bind(l_2); 1291 __ addic_(R5_ARG3, R5_ARG3, -1); 1292 __ lbzx(tmp1, R3_ARG1, R5_ARG3); 1293 __ bge(CCR0, l_1); 1294 1295 __ blr(); 1296 1297 return start; 1298 } 1299 1300 // Generate stub for disjoint short copy. If "aligned" is true, the 1301 // "from" and "to" addresses are assumed to be heapword aligned. 1302 // 1303 // Arguments for generated stub: 1304 // from: R3_ARG1 1305 // to: R4_ARG2 1306 // elm.count: R5_ARG3 treated as signed 1307 // 1308 // Strategy for aligned==true: 1309 // 1310 // If length <= 9: 1311 // 1. copy 2 elements at a time (l_6) 1312 // 2. copy last element if original element count was odd (l_1) 1313 // 1314 // If length > 9: 1315 // 1. copy 4 elements at a time until less than 4 elements are left (l_7) 1316 // 2. copy 2 elements at a time until less than 2 elements are left (l_6) 1317 // 3. copy last element if one was left in step 2. (l_1) 1318 // 1319 // 1320 // Strategy for aligned==false: 1321 // 1322 // If length <= 9: same as aligned==true case, but NOTE: load/stores 1323 // can be unaligned (see comment below) 1324 // 1325 // If length > 9: 1326 // 1. continue with step 6. if the alignment of from and to mod 4 1327 // is different. 1328 // 2. align from and to to 4 bytes by copying 1 element if necessary 1329 // 3. at l_2 from and to are 4 byte aligned; continue with 1330 // 5. if they cannot be aligned to 8 bytes because they have 1331 // got different alignment mod 8. 1332 // 4. at this point we know that both, from and to, have the same 1333 // alignment mod 8, now copy one element if necessary to get 1334 // 8 byte alignment of from and to. 1335 // 5. copy 4 elements at a time until less than 4 elements are 1336 // left; depending on step 3. all load/stores are aligned or 1337 // either all loads or all stores are unaligned. 1338 // 6. copy 2 elements at a time until less than 2 elements are 1339 // left (l_6); arriving here from step 1., there is a chance 1340 // that all accesses are unaligned. 1341 // 7. copy last element if one was left in step 6. (l_1) 1342 // 1343 // There are unaligned data accesses using integer load/store 1344 // instructions in this stub. POWER allows such accesses. 1345 // 1346 // According to the manuals (PowerISA_V2.06_PUBLIC, Book II, 1347 // Chapter 2: Effect of Operand Placement on Performance) unaligned 1348 // integer load/stores have good performance. Only unaligned 1349 // floating point load/stores can have poor performance. 1350 // 1351 // TODO: 1352 // 1353 // 1. check if aligning the backbranch target of loops is beneficial 1354 // 1355 address generate_disjoint_short_copy(bool aligned, const char * name) { 1356 StubCodeMark mark(this, "StubRoutines", name); 1357 1358 Register tmp1 = R6_ARG4; 1359 Register tmp2 = R7_ARG5; 1360 Register tmp3 = R8_ARG6; 1361 Register tmp4 = R9_ARG7; 1362 1363 address start = __ function_entry(); 1364 1365 Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8; 1366 // don't try anything fancy if arrays don't have many elements 1367 __ li(tmp3, 0); 1368 __ cmpwi(CCR0, R5_ARG3, 9); 1369 __ ble(CCR0, l_6); // copy 2 at a time 1370 1371 if (!aligned) { 1372 __ xorr(tmp1, R3_ARG1, R4_ARG2); 1373 __ andi_(tmp1, tmp1, 3); 1374 __ bne(CCR0, l_6); // if arrays don't have the same alignment mod 4, do 2 element copy 1375 1376 // At this point it is guaranteed that both, from and to have the same alignment mod 4. 1377 1378 // Copy 1 element if necessary to align to 4 bytes. 1379 __ andi_(tmp1, R3_ARG1, 3); 1380 __ beq(CCR0, l_2); 1381 1382 __ lhz(tmp2, 0, R3_ARG1); 1383 __ addi(R3_ARG1, R3_ARG1, 2); 1384 __ sth(tmp2, 0, R4_ARG2); 1385 __ addi(R4_ARG2, R4_ARG2, 2); 1386 __ addi(R5_ARG3, R5_ARG3, -1); 1387 __ bind(l_2); 1388 1389 // At this point the positions of both, from and to, are at least 4 byte aligned. 1390 1391 // Copy 4 elements at a time. 1392 // Align to 8 bytes, but only if both, from and to, have same alignment mod 8. 1393 __ xorr(tmp2, R3_ARG1, R4_ARG2); 1394 __ andi_(tmp1, tmp2, 7); 1395 __ bne(CCR0, l_7); // not same alignment mod 8 -> copy 4, either from or to will be unaligned 1396 1397 // Copy a 2-element word if necessary to align to 8 bytes. 1398 __ andi_(R0, R3_ARG1, 7); 1399 __ beq(CCR0, l_7); 1400 1401 __ lwzx(tmp2, R3_ARG1, tmp3); 1402 __ addi(R5_ARG3, R5_ARG3, -2); 1403 __ stwx(tmp2, R4_ARG2, tmp3); 1404 { // FasterArrayCopy 1405 __ addi(R3_ARG1, R3_ARG1, 4); 1406 __ addi(R4_ARG2, R4_ARG2, 4); 1407 } 1408 } 1409 1410 __ bind(l_7); 1411 1412 // Copy 4 elements at a time; either the loads or the stores can 1413 // be unaligned if aligned == false. 1414 1415 { // FasterArrayCopy 1416 __ cmpwi(CCR0, R5_ARG3, 15); 1417 __ ble(CCR0, l_6); // copy 2 at a time if less than 16 elements remain 1418 1419 __ srdi(tmp1, R5_ARG3, 4); 1420 __ andi_(R5_ARG3, R5_ARG3, 15); 1421 __ mtctr(tmp1); 1422 1423 __ bind(l_8); 1424 // Use unrolled version for mass copying (copy 16 elements a time). 1425 // Load feeding store gets zero latency on Power6, however not on Power5. 1426 // Therefore, the following sequence is made for the good of both. 1427 __ ld(tmp1, 0, R3_ARG1); 1428 __ ld(tmp2, 8, R3_ARG1); 1429 __ ld(tmp3, 16, R3_ARG1); 1430 __ ld(tmp4, 24, R3_ARG1); 1431 __ std(tmp1, 0, R4_ARG2); 1432 __ std(tmp2, 8, R4_ARG2); 1433 __ std(tmp3, 16, R4_ARG2); 1434 __ std(tmp4, 24, R4_ARG2); 1435 __ addi(R3_ARG1, R3_ARG1, 32); 1436 __ addi(R4_ARG2, R4_ARG2, 32); 1437 __ bdnz(l_8); 1438 } 1439 __ bind(l_6); 1440 1441 // copy 2 elements at a time 1442 { // FasterArrayCopy 1443 __ cmpwi(CCR0, R5_ARG3, 2); 1444 __ blt(CCR0, l_1); 1445 __ srdi(tmp1, R5_ARG3, 1); 1446 __ andi_(R5_ARG3, R5_ARG3, 1); 1447 1448 __ addi(R3_ARG1, R3_ARG1, -4); 1449 __ addi(R4_ARG2, R4_ARG2, -4); 1450 __ mtctr(tmp1); 1451 1452 __ bind(l_3); 1453 __ lwzu(tmp2, 4, R3_ARG1); 1454 __ stwu(tmp2, 4, R4_ARG2); 1455 __ bdnz(l_3); 1456 1457 __ addi(R3_ARG1, R3_ARG1, 4); 1458 __ addi(R4_ARG2, R4_ARG2, 4); 1459 } 1460 1461 // do single element copy 1462 __ bind(l_1); 1463 __ cmpwi(CCR0, R5_ARG3, 0); 1464 __ beq(CCR0, l_4); 1465 1466 { // FasterArrayCopy 1467 __ mtctr(R5_ARG3); 1468 __ addi(R3_ARG1, R3_ARG1, -2); 1469 __ addi(R4_ARG2, R4_ARG2, -2); 1470 1471 __ bind(l_5); 1472 __ lhzu(tmp2, 2, R3_ARG1); 1473 __ sthu(tmp2, 2, R4_ARG2); 1474 __ bdnz(l_5); 1475 } 1476 __ bind(l_4); 1477 __ blr(); 1478 1479 return start; 1480 } 1481 1482 // Generate stub for conjoint short copy. If "aligned" is true, the 1483 // "from" and "to" addresses are assumed to be heapword aligned. 1484 // 1485 // Arguments for generated stub: 1486 // from: R3_ARG1 1487 // to: R4_ARG2 1488 // count: R5_ARG3 treated as signed 1489 // 1490 address generate_conjoint_short_copy(bool aligned, const char * name) { 1491 StubCodeMark mark(this, "StubRoutines", name); 1492 address start = __ function_entry(); 1493 1494 Register tmp1 = R6_ARG4; 1495 Register tmp2 = R7_ARG5; 1496 Register tmp3 = R8_ARG6; 1497 1498 #if defined(ABI_ELFv2) 1499 address nooverlap_target = aligned ? 1500 StubRoutines::arrayof_jshort_disjoint_arraycopy() : 1501 StubRoutines::jshort_disjoint_arraycopy(); 1502 #else 1503 address nooverlap_target = aligned ? 1504 ((FunctionDescriptor*)StubRoutines::arrayof_jshort_disjoint_arraycopy())->entry() : 1505 ((FunctionDescriptor*)StubRoutines::jshort_disjoint_arraycopy())->entry(); 1506 #endif 1507 1508 array_overlap_test(nooverlap_target, 1); 1509 1510 Label l_1, l_2; 1511 __ sldi(tmp1, R5_ARG3, 1); 1512 __ b(l_2); 1513 __ bind(l_1); 1514 __ sthx(tmp2, R4_ARG2, tmp1); 1515 __ bind(l_2); 1516 __ addic_(tmp1, tmp1, -2); 1517 __ lhzx(tmp2, R3_ARG1, tmp1); 1518 __ bge(CCR0, l_1); 1519 1520 __ blr(); 1521 1522 return start; 1523 } 1524 1525 // Generate core code for disjoint int copy (and oop copy on 32-bit). If "aligned" 1526 // is true, the "from" and "to" addresses are assumed to be heapword aligned. 1527 // 1528 // Arguments: 1529 // from: R3_ARG1 1530 // to: R4_ARG2 1531 // count: R5_ARG3 treated as signed 1532 // 1533 void generate_disjoint_int_copy_core(bool aligned) { 1534 Register tmp1 = R6_ARG4; 1535 Register tmp2 = R7_ARG5; 1536 Register tmp3 = R8_ARG6; 1537 Register tmp4 = R0; 1538 1539 Label l_1, l_2, l_3, l_4, l_5, l_6; 1540 // for short arrays, just do single element copy 1541 __ li(tmp3, 0); 1542 __ cmpwi(CCR0, R5_ARG3, 5); 1543 __ ble(CCR0, l_2); 1544 1545 if (!aligned) { 1546 // check if arrays have same alignment mod 8. 1547 __ xorr(tmp1, R3_ARG1, R4_ARG2); 1548 __ andi_(R0, tmp1, 7); 1549 // Not the same alignment, but ld and std just need to be 4 byte aligned. 1550 __ bne(CCR0, l_4); // to OR from is 8 byte aligned -> copy 2 at a time 1551 1552 // copy 1 element to align to and from on an 8 byte boundary 1553 __ andi_(R0, R3_ARG1, 7); 1554 __ beq(CCR0, l_4); 1555 1556 __ lwzx(tmp2, R3_ARG1, tmp3); 1557 __ addi(R5_ARG3, R5_ARG3, -1); 1558 __ stwx(tmp2, R4_ARG2, tmp3); 1559 { // FasterArrayCopy 1560 __ addi(R3_ARG1, R3_ARG1, 4); 1561 __ addi(R4_ARG2, R4_ARG2, 4); 1562 } 1563 __ bind(l_4); 1564 } 1565 1566 { // FasterArrayCopy 1567 __ cmpwi(CCR0, R5_ARG3, 7); 1568 __ ble(CCR0, l_2); // copy 1 at a time if less than 8 elements remain 1569 1570 __ srdi(tmp1, R5_ARG3, 3); 1571 __ andi_(R5_ARG3, R5_ARG3, 7); 1572 __ mtctr(tmp1); 1573 1574 __ bind(l_6); 1575 // Use unrolled version for mass copying (copy 8 elements a time). 1576 // Load feeding store gets zero latency on power6, however not on power 5. 1577 // Therefore, the following sequence is made for the good of both. 1578 __ ld(tmp1, 0, R3_ARG1); 1579 __ ld(tmp2, 8, R3_ARG1); 1580 __ ld(tmp3, 16, R3_ARG1); 1581 __ ld(tmp4, 24, R3_ARG1); 1582 __ std(tmp1, 0, R4_ARG2); 1583 __ std(tmp2, 8, R4_ARG2); 1584 __ std(tmp3, 16, R4_ARG2); 1585 __ std(tmp4, 24, R4_ARG2); 1586 __ addi(R3_ARG1, R3_ARG1, 32); 1587 __ addi(R4_ARG2, R4_ARG2, 32); 1588 __ bdnz(l_6); 1589 } 1590 1591 // copy 1 element at a time 1592 __ bind(l_2); 1593 __ cmpwi(CCR0, R5_ARG3, 0); 1594 __ beq(CCR0, l_1); 1595 1596 { // FasterArrayCopy 1597 __ mtctr(R5_ARG3); 1598 __ addi(R3_ARG1, R3_ARG1, -4); 1599 __ addi(R4_ARG2, R4_ARG2, -4); 1600 1601 __ bind(l_3); 1602 __ lwzu(tmp2, 4, R3_ARG1); 1603 __ stwu(tmp2, 4, R4_ARG2); 1604 __ bdnz(l_3); 1605 } 1606 1607 __ bind(l_1); 1608 return; 1609 } 1610 1611 // Generate stub for disjoint int copy. If "aligned" is true, the 1612 // "from" and "to" addresses are assumed to be heapword aligned. 1613 // 1614 // Arguments for generated stub: 1615 // from: R3_ARG1 1616 // to: R4_ARG2 1617 // count: R5_ARG3 treated as signed 1618 // 1619 address generate_disjoint_int_copy(bool aligned, const char * name) { 1620 StubCodeMark mark(this, "StubRoutines", name); 1621 address start = __ function_entry(); 1622 generate_disjoint_int_copy_core(aligned); 1623 __ blr(); 1624 return start; 1625 } 1626 1627 // Generate core code for conjoint int copy (and oop copy on 1628 // 32-bit). If "aligned" is true, the "from" and "to" addresses 1629 // are assumed to be heapword aligned. 1630 // 1631 // Arguments: 1632 // from: R3_ARG1 1633 // to: R4_ARG2 1634 // count: R5_ARG3 treated as signed 1635 // 1636 void generate_conjoint_int_copy_core(bool aligned) { 1637 // Do reverse copy. We assume the case of actual overlap is rare enough 1638 // that we don't have to optimize it. 1639 1640 Label l_1, l_2, l_3, l_4, l_5, l_6; 1641 1642 Register tmp1 = R6_ARG4; 1643 Register tmp2 = R7_ARG5; 1644 Register tmp3 = R8_ARG6; 1645 Register tmp4 = R0; 1646 1647 { // FasterArrayCopy 1648 __ cmpwi(CCR0, R5_ARG3, 0); 1649 __ beq(CCR0, l_6); 1650 1651 __ sldi(R5_ARG3, R5_ARG3, 2); 1652 __ add(R3_ARG1, R3_ARG1, R5_ARG3); 1653 __ add(R4_ARG2, R4_ARG2, R5_ARG3); 1654 __ srdi(R5_ARG3, R5_ARG3, 2); 1655 1656 __ cmpwi(CCR0, R5_ARG3, 7); 1657 __ ble(CCR0, l_5); // copy 1 at a time if less than 8 elements remain 1658 1659 __ srdi(tmp1, R5_ARG3, 3); 1660 __ andi(R5_ARG3, R5_ARG3, 7); 1661 __ mtctr(tmp1); 1662 1663 __ bind(l_4); 1664 // Use unrolled version for mass copying (copy 4 elements a time). 1665 // Load feeding store gets zero latency on Power6, however not on Power5. 1666 // Therefore, the following sequence is made for the good of both. 1667 __ addi(R3_ARG1, R3_ARG1, -32); 1668 __ addi(R4_ARG2, R4_ARG2, -32); 1669 __ ld(tmp4, 24, R3_ARG1); 1670 __ ld(tmp3, 16, R3_ARG1); 1671 __ ld(tmp2, 8, R3_ARG1); 1672 __ ld(tmp1, 0, R3_ARG1); 1673 __ std(tmp4, 24, R4_ARG2); 1674 __ std(tmp3, 16, R4_ARG2); 1675 __ std(tmp2, 8, R4_ARG2); 1676 __ std(tmp1, 0, R4_ARG2); 1677 __ bdnz(l_4); 1678 1679 __ cmpwi(CCR0, R5_ARG3, 0); 1680 __ beq(CCR0, l_6); 1681 1682 __ bind(l_5); 1683 __ mtctr(R5_ARG3); 1684 __ bind(l_3); 1685 __ lwz(R0, -4, R3_ARG1); 1686 __ stw(R0, -4, R4_ARG2); 1687 __ addi(R3_ARG1, R3_ARG1, -4); 1688 __ addi(R4_ARG2, R4_ARG2, -4); 1689 __ bdnz(l_3); 1690 1691 __ bind(l_6); 1692 } 1693 } 1694 1695 // Generate stub for conjoint int copy. If "aligned" is true, the 1696 // "from" and "to" addresses are assumed to be heapword aligned. 1697 // 1698 // Arguments for generated stub: 1699 // from: R3_ARG1 1700 // to: R4_ARG2 1701 // count: R5_ARG3 treated as signed 1702 // 1703 address generate_conjoint_int_copy(bool aligned, const char * name) { 1704 StubCodeMark mark(this, "StubRoutines", name); 1705 address start = __ function_entry(); 1706 1707 #if defined(ABI_ELFv2) 1708 address nooverlap_target = aligned ? 1709 StubRoutines::arrayof_jint_disjoint_arraycopy() : 1710 StubRoutines::jint_disjoint_arraycopy(); 1711 #else 1712 address nooverlap_target = aligned ? 1713 ((FunctionDescriptor*)StubRoutines::arrayof_jint_disjoint_arraycopy())->entry() : 1714 ((FunctionDescriptor*)StubRoutines::jint_disjoint_arraycopy())->entry(); 1715 #endif 1716 1717 array_overlap_test(nooverlap_target, 2); 1718 1719 generate_conjoint_int_copy_core(aligned); 1720 1721 __ blr(); 1722 1723 return start; 1724 } 1725 1726 // Generate core code for disjoint long copy (and oop copy on 1727 // 64-bit). If "aligned" is true, the "from" and "to" addresses 1728 // are assumed to be heapword aligned. 1729 // 1730 // Arguments: 1731 // from: R3_ARG1 1732 // to: R4_ARG2 1733 // count: R5_ARG3 treated as signed 1734 // 1735 void generate_disjoint_long_copy_core(bool aligned) { 1736 Register tmp1 = R6_ARG4; 1737 Register tmp2 = R7_ARG5; 1738 Register tmp3 = R8_ARG6; 1739 Register tmp4 = R0; 1740 1741 Label l_1, l_2, l_3, l_4; 1742 1743 { // FasterArrayCopy 1744 __ cmpwi(CCR0, R5_ARG3, 3); 1745 __ ble(CCR0, l_3); // copy 1 at a time if less than 4 elements remain 1746 1747 __ srdi(tmp1, R5_ARG3, 2); 1748 __ andi_(R5_ARG3, R5_ARG3, 3); 1749 __ mtctr(tmp1); 1750 1751 __ bind(l_4); 1752 // Use unrolled version for mass copying (copy 4 elements a time). 1753 // Load feeding store gets zero latency on Power6, however not on Power5. 1754 // Therefore, the following sequence is made for the good of both. 1755 __ ld(tmp1, 0, R3_ARG1); 1756 __ ld(tmp2, 8, R3_ARG1); 1757 __ ld(tmp3, 16, R3_ARG1); 1758 __ ld(tmp4, 24, R3_ARG1); 1759 __ std(tmp1, 0, R4_ARG2); 1760 __ std(tmp2, 8, R4_ARG2); 1761 __ std(tmp3, 16, R4_ARG2); 1762 __ std(tmp4, 24, R4_ARG2); 1763 __ addi(R3_ARG1, R3_ARG1, 32); 1764 __ addi(R4_ARG2, R4_ARG2, 32); 1765 __ bdnz(l_4); 1766 } 1767 1768 // copy 1 element at a time 1769 __ bind(l_3); 1770 __ cmpwi(CCR0, R5_ARG3, 0); 1771 __ beq(CCR0, l_1); 1772 1773 { // FasterArrayCopy 1774 __ mtctr(R5_ARG3); 1775 __ addi(R3_ARG1, R3_ARG1, -8); 1776 __ addi(R4_ARG2, R4_ARG2, -8); 1777 1778 __ bind(l_2); 1779 __ ldu(R0, 8, R3_ARG1); 1780 __ stdu(R0, 8, R4_ARG2); 1781 __ bdnz(l_2); 1782 1783 } 1784 __ bind(l_1); 1785 } 1786 1787 // Generate stub for disjoint long copy. If "aligned" is true, the 1788 // "from" and "to" addresses are assumed to be heapword aligned. 1789 // 1790 // Arguments for generated stub: 1791 // from: R3_ARG1 1792 // to: R4_ARG2 1793 // count: R5_ARG3 treated as signed 1794 // 1795 address generate_disjoint_long_copy(bool aligned, const char * name) { 1796 StubCodeMark mark(this, "StubRoutines", name); 1797 address start = __ function_entry(); 1798 generate_disjoint_long_copy_core(aligned); 1799 __ blr(); 1800 1801 return start; 1802 } 1803 1804 // Generate core code for conjoint long copy (and oop copy on 1805 // 64-bit). If "aligned" is true, the "from" and "to" addresses 1806 // are assumed to be heapword aligned. 1807 // 1808 // Arguments: 1809 // from: R3_ARG1 1810 // to: R4_ARG2 1811 // count: R5_ARG3 treated as signed 1812 // 1813 void generate_conjoint_long_copy_core(bool aligned) { 1814 Register tmp1 = R6_ARG4; 1815 Register tmp2 = R7_ARG5; 1816 Register tmp3 = R8_ARG6; 1817 Register tmp4 = R0; 1818 1819 Label l_1, l_2, l_3, l_4, l_5; 1820 1821 __ cmpwi(CCR0, R5_ARG3, 0); 1822 __ beq(CCR0, l_1); 1823 1824 { // FasterArrayCopy 1825 __ sldi(R5_ARG3, R5_ARG3, 3); 1826 __ add(R3_ARG1, R3_ARG1, R5_ARG3); 1827 __ add(R4_ARG2, R4_ARG2, R5_ARG3); 1828 __ srdi(R5_ARG3, R5_ARG3, 3); 1829 1830 __ cmpwi(CCR0, R5_ARG3, 3); 1831 __ ble(CCR0, l_5); // copy 1 at a time if less than 4 elements remain 1832 1833 __ srdi(tmp1, R5_ARG3, 2); 1834 __ andi(R5_ARG3, R5_ARG3, 3); 1835 __ mtctr(tmp1); 1836 1837 __ bind(l_4); 1838 // Use unrolled version for mass copying (copy 4 elements a time). 1839 // Load feeding store gets zero latency on Power6, however not on Power5. 1840 // Therefore, the following sequence is made for the good of both. 1841 __ addi(R3_ARG1, R3_ARG1, -32); 1842 __ addi(R4_ARG2, R4_ARG2, -32); 1843 __ ld(tmp4, 24, R3_ARG1); 1844 __ ld(tmp3, 16, R3_ARG1); 1845 __ ld(tmp2, 8, R3_ARG1); 1846 __ ld(tmp1, 0, R3_ARG1); 1847 __ std(tmp4, 24, R4_ARG2); 1848 __ std(tmp3, 16, R4_ARG2); 1849 __ std(tmp2, 8, R4_ARG2); 1850 __ std(tmp1, 0, R4_ARG2); 1851 __ bdnz(l_4); 1852 1853 __ cmpwi(CCR0, R5_ARG3, 0); 1854 __ beq(CCR0, l_1); 1855 1856 __ bind(l_5); 1857 __ mtctr(R5_ARG3); 1858 __ bind(l_3); 1859 __ ld(R0, -8, R3_ARG1); 1860 __ std(R0, -8, R4_ARG2); 1861 __ addi(R3_ARG1, R3_ARG1, -8); 1862 __ addi(R4_ARG2, R4_ARG2, -8); 1863 __ bdnz(l_3); 1864 1865 } 1866 __ bind(l_1); 1867 } 1868 1869 // Generate stub for conjoint long copy. If "aligned" is true, the 1870 // "from" and "to" addresses are assumed to be heapword aligned. 1871 // 1872 // Arguments for generated stub: 1873 // from: R3_ARG1 1874 // to: R4_ARG2 1875 // count: R5_ARG3 treated as signed 1876 // 1877 address generate_conjoint_long_copy(bool aligned, const char * name) { 1878 StubCodeMark mark(this, "StubRoutines", name); 1879 address start = __ function_entry(); 1880 1881 #if defined(ABI_ELFv2) 1882 address nooverlap_target = aligned ? 1883 StubRoutines::arrayof_jlong_disjoint_arraycopy() : 1884 StubRoutines::jlong_disjoint_arraycopy(); 1885 #else 1886 address nooverlap_target = aligned ? 1887 ((FunctionDescriptor*)StubRoutines::arrayof_jlong_disjoint_arraycopy())->entry() : 1888 ((FunctionDescriptor*)StubRoutines::jlong_disjoint_arraycopy())->entry(); 1889 #endif 1890 1891 array_overlap_test(nooverlap_target, 3); 1892 generate_conjoint_long_copy_core(aligned); 1893 1894 __ blr(); 1895 1896 return start; 1897 } 1898 1899 // Generate stub for conjoint oop copy. If "aligned" is true, the 1900 // "from" and "to" addresses are assumed to be heapword aligned. 1901 // 1902 // Arguments for generated stub: 1903 // from: R3_ARG1 1904 // to: R4_ARG2 1905 // count: R5_ARG3 treated as signed 1906 // dest_uninitialized: G1 support 1907 // 1908 address generate_conjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) { 1909 StubCodeMark mark(this, "StubRoutines", name); 1910 1911 address start = __ function_entry(); 1912 1913 #if defined(ABI_ELFv2) 1914 address nooverlap_target = aligned ? 1915 StubRoutines::arrayof_oop_disjoint_arraycopy() : 1916 StubRoutines::oop_disjoint_arraycopy(); 1917 #else 1918 address nooverlap_target = aligned ? 1919 ((FunctionDescriptor*)StubRoutines::arrayof_oop_disjoint_arraycopy())->entry() : 1920 ((FunctionDescriptor*)StubRoutines::oop_disjoint_arraycopy())->entry(); 1921 #endif 1922 1923 gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7); 1924 1925 // Save arguments. 1926 __ mr(R9_ARG7, R4_ARG2); 1927 __ mr(R10_ARG8, R5_ARG3); 1928 1929 if (UseCompressedOops) { 1930 array_overlap_test(nooverlap_target, 2); 1931 generate_conjoint_int_copy_core(aligned); 1932 } else { 1933 array_overlap_test(nooverlap_target, 3); 1934 generate_conjoint_long_copy_core(aligned); 1935 } 1936 1937 gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1, /*branchToEnd*/ false); 1938 return start; 1939 } 1940 1941 // Generate stub for disjoint oop copy. If "aligned" is true, the 1942 // "from" and "to" addresses are assumed to be heapword aligned. 1943 // 1944 // Arguments for generated stub: 1945 // from: R3_ARG1 1946 // to: R4_ARG2 1947 // count: R5_ARG3 treated as signed 1948 // dest_uninitialized: G1 support 1949 // 1950 address generate_disjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) { 1951 StubCodeMark mark(this, "StubRoutines", name); 1952 address start = __ function_entry(); 1953 1954 gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7); 1955 1956 // save some arguments, disjoint_long_copy_core destroys them. 1957 // needed for post barrier 1958 __ mr(R9_ARG7, R4_ARG2); 1959 __ mr(R10_ARG8, R5_ARG3); 1960 1961 if (UseCompressedOops) { 1962 generate_disjoint_int_copy_core(aligned); 1963 } else { 1964 generate_disjoint_long_copy_core(aligned); 1965 } 1966 1967 gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1, /*branchToEnd*/ false); 1968 1969 return start; 1970 } 1971 1972 void generate_arraycopy_stubs() { 1973 // Note: the disjoint stubs must be generated first, some of 1974 // the conjoint stubs use them. 1975 1976 // non-aligned disjoint versions 1977 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, "jbyte_disjoint_arraycopy"); 1978 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, "jshort_disjoint_arraycopy"); 1979 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_copy(false, "jint_disjoint_arraycopy"); 1980 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_copy(false, "jlong_disjoint_arraycopy"); 1981 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy", false); 1982 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy_uninit", true); 1983 1984 // aligned disjoint versions 1985 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(true, "arrayof_jbyte_disjoint_arraycopy"); 1986 StubRoutines::_arrayof_jshort_disjoint_arraycopy = generate_disjoint_short_copy(true, "arrayof_jshort_disjoint_arraycopy"); 1987 StubRoutines::_arrayof_jint_disjoint_arraycopy = generate_disjoint_int_copy(true, "arrayof_jint_disjoint_arraycopy"); 1988 StubRoutines::_arrayof_jlong_disjoint_arraycopy = generate_disjoint_long_copy(true, "arrayof_jlong_disjoint_arraycopy"); 1989 StubRoutines::_arrayof_oop_disjoint_arraycopy = generate_disjoint_oop_copy(true, "arrayof_oop_disjoint_arraycopy", false); 1990 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(true, "oop_disjoint_arraycopy_uninit", true); 1991 1992 // non-aligned conjoint versions 1993 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, "jbyte_arraycopy"); 1994 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, "jshort_arraycopy"); 1995 StubRoutines::_jint_arraycopy = generate_conjoint_int_copy(false, "jint_arraycopy"); 1996 StubRoutines::_jlong_arraycopy = generate_conjoint_long_copy(false, "jlong_arraycopy"); 1997 StubRoutines::_oop_arraycopy = generate_conjoint_oop_copy(false, "oop_arraycopy", false); 1998 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_oop_copy(false, "oop_arraycopy_uninit", true); 1999 2000 // aligned conjoint versions 2001 StubRoutines::_arrayof_jbyte_arraycopy = generate_conjoint_byte_copy(true, "arrayof_jbyte_arraycopy"); 2002 StubRoutines::_arrayof_jshort_arraycopy = generate_conjoint_short_copy(true, "arrayof_jshort_arraycopy"); 2003 StubRoutines::_arrayof_jint_arraycopy = generate_conjoint_int_copy(true, "arrayof_jint_arraycopy"); 2004 StubRoutines::_arrayof_jlong_arraycopy = generate_conjoint_long_copy(true, "arrayof_jlong_arraycopy"); 2005 StubRoutines::_arrayof_oop_arraycopy = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", false); 2006 StubRoutines::_arrayof_oop_arraycopy_uninit = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", true); 2007 2008 // fill routines 2009 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill"); 2010 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill"); 2011 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill"); 2012 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill"); 2013 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill"); 2014 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill"); 2015 } 2016 2017 // Safefetch stubs. 2018 void generate_safefetch(const char* name, int size, address* entry, address* fault_pc, address* continuation_pc) { 2019 // safefetch signatures: 2020 // int SafeFetch32(int* adr, int errValue); 2021 // intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue); 2022 // 2023 // arguments: 2024 // R3_ARG1 = adr 2025 // R4_ARG2 = errValue 2026 // 2027 // result: 2028 // R3_RET = *adr or errValue 2029 2030 StubCodeMark mark(this, "StubRoutines", name); 2031 2032 // Entry point, pc or function descriptor. 2033 *entry = __ function_entry(); 2034 2035 // Load *adr into R4_ARG2, may fault. 2036 *fault_pc = __ pc(); 2037 switch (size) { 2038 case 4: 2039 // int32_t, signed extended 2040 __ lwa(R4_ARG2, 0, R3_ARG1); 2041 break; 2042 case 8: 2043 // int64_t 2044 __ ld(R4_ARG2, 0, R3_ARG1); 2045 break; 2046 default: 2047 ShouldNotReachHere(); 2048 } 2049 2050 // return errValue or *adr 2051 *continuation_pc = __ pc(); 2052 __ mr(R3_RET, R4_ARG2); 2053 __ blr(); 2054 } 2055 2056 // Initialization 2057 void generate_initial() { 2058 // Generates all stubs and initializes the entry points 2059 2060 // Entry points that exist in all platforms. 2061 // Note: This is code that could be shared among different platforms - however the 2062 // benefit seems to be smaller than the disadvantage of having a 2063 // much more complicated generator structure. See also comment in 2064 // stubRoutines.hpp. 2065 2066 StubRoutines::_forward_exception_entry = generate_forward_exception(); 2067 StubRoutines::_call_stub_entry = generate_call_stub(StubRoutines::_call_stub_return_address); 2068 StubRoutines::_catch_exception_entry = generate_catch_exception(); 2069 2070 // Build this early so it's available for the interpreter. 2071 StubRoutines::_throw_StackOverflowError_entry = 2072 generate_throw_exception("StackOverflowError throw_exception", 2073 CAST_FROM_FN_PTR(address, SharedRuntime::throw_StackOverflowError), false); 2074 } 2075 2076 void generate_all() { 2077 // Generates all stubs and initializes the entry points 2078 2079 // These entry points require SharedInfo::stack0 to be set up in 2080 // non-core builds 2081 StubRoutines::_throw_AbstractMethodError_entry = generate_throw_exception("AbstractMethodError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError), false); 2082 // Handle IncompatibleClassChangeError in itable stubs. 2083 StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError), false); 2084 StubRoutines::_throw_NullPointerException_at_call_entry= generate_throw_exception("NullPointerException at call throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_NullPointerException_at_call), false); 2085 2086 StubRoutines::_handler_for_unsafe_access_entry = generate_handler_for_unsafe_access(); 2087 2088 // support for verify_oop (must happen after universe_init) 2089 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop(); 2090 2091 // arraycopy stubs used by compilers 2092 generate_arraycopy_stubs(); 2093 2094 if (UseAESIntrinsics) { 2095 guarantee(!UseAESIntrinsics, "not yet implemented."); 2096 } 2097 2098 // Safefetch stubs. 2099 generate_safefetch("SafeFetch32", sizeof(int), &StubRoutines::_safefetch32_entry, 2100 &StubRoutines::_safefetch32_fault_pc, 2101 &StubRoutines::_safefetch32_continuation_pc); 2102 generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry, 2103 &StubRoutines::_safefetchN_fault_pc, 2104 &StubRoutines::_safefetchN_continuation_pc); 2105 } 2106 2107 public: 2108 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) { 2109 // replace the standard masm with a special one: 2110 _masm = new MacroAssembler(code); 2111 if (all) { 2112 generate_all(); 2113 } else { 2114 generate_initial(); 2115 } 2116 } 2117 }; 2118 2119 void StubGenerator_generate(CodeBuffer* code, bool all) { 2120 StubGenerator g(code, all); 2121 }