stubGenerator_ppc.cpp revision 11155:2a22bacc38a7
1/* 2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2012, 2016 SAP SE. 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 "runtime/thread.inline.hpp" 41 42#define __ _masm-> 43 44#ifdef PRODUCT 45#define BLOCK_COMMENT(str) // nothing 46#else 47#define BLOCK_COMMENT(str) __ block_comment(str) 48#endif 49 50#if defined(ABI_ELFv2) 51#define STUB_ENTRY(name) StubRoutines::name() 52#else 53#define STUB_ENTRY(name) ((FunctionDescriptor*)StubRoutines::name())->entry() 54#endif 55 56class StubGenerator: public StubCodeGenerator { 57 private: 58 59 // Call stubs are used to call Java from C 60 // 61 // Arguments: 62 // 63 // R3 - call wrapper address : address 64 // R4 - result : intptr_t* 65 // R5 - result type : BasicType 66 // R6 - method : Method 67 // R7 - frame mgr entry point : address 68 // R8 - parameter block : intptr_t* 69 // R9 - parameter count in words : int 70 // R10 - thread : Thread* 71 // 72 address generate_call_stub(address& return_address) { 73 // Setup a new c frame, copy java arguments, call frame manager or 74 // native_entry, and process result. 75 76 StubCodeMark mark(this, "StubRoutines", "call_stub"); 77 78 address start = __ function_entry(); 79 80 // some sanity checks 81 assert((sizeof(frame::abi_minframe) % 16) == 0, "unaligned"); 82 assert((sizeof(frame::abi_reg_args) % 16) == 0, "unaligned"); 83 assert((sizeof(frame::spill_nonvolatiles) % 16) == 0, "unaligned"); 84 assert((sizeof(frame::parent_ijava_frame_abi) % 16) == 0, "unaligned"); 85 assert((sizeof(frame::entry_frame_locals) % 16) == 0, "unaligned"); 86 87 Register r_arg_call_wrapper_addr = R3; 88 Register r_arg_result_addr = R4; 89 Register r_arg_result_type = R5; 90 Register r_arg_method = R6; 91 Register r_arg_entry = R7; 92 Register r_arg_thread = R10; 93 94 Register r_temp = R24; 95 Register r_top_of_arguments_addr = R25; 96 Register r_entryframe_fp = R26; 97 98 { 99 // Stack on entry to call_stub: 100 // 101 // F1 [C_FRAME] 102 // ... 103 104 Register r_arg_argument_addr = R8; 105 Register r_arg_argument_count = R9; 106 Register r_frame_alignment_in_bytes = R27; 107 Register r_argument_addr = R28; 108 Register r_argumentcopy_addr = R29; 109 Register r_argument_size_in_bytes = R30; 110 Register r_frame_size = R23; 111 112 Label arguments_copied; 113 114 // Save LR/CR to caller's C_FRAME. 115 __ save_LR_CR(R0); 116 117 // Zero extend arg_argument_count. 118 __ clrldi(r_arg_argument_count, r_arg_argument_count, 32); 119 120 // Save non-volatiles GPRs to ENTRY_FRAME (not yet pushed, but it's safe). 121 __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14)); 122 123 // Keep copy of our frame pointer (caller's SP). 124 __ mr(r_entryframe_fp, R1_SP); 125 126 BLOCK_COMMENT("Push ENTRY_FRAME including arguments"); 127 // Push ENTRY_FRAME including arguments: 128 // 129 // F0 [TOP_IJAVA_FRAME_ABI] 130 // alignment (optional) 131 // [outgoing Java arguments] 132 // [ENTRY_FRAME_LOCALS] 133 // F1 [C_FRAME] 134 // ... 135 136 // calculate frame size 137 138 // unaligned size of arguments 139 __ sldi(r_argument_size_in_bytes, 140 r_arg_argument_count, Interpreter::logStackElementSize); 141 // arguments alignment (max 1 slot) 142 // FIXME: use round_to() here 143 __ andi_(r_frame_alignment_in_bytes, r_arg_argument_count, 1); 144 __ sldi(r_frame_alignment_in_bytes, 145 r_frame_alignment_in_bytes, Interpreter::logStackElementSize); 146 147 // size = unaligned size of arguments + top abi's size 148 __ addi(r_frame_size, r_argument_size_in_bytes, 149 frame::top_ijava_frame_abi_size); 150 // size += arguments alignment 151 __ add(r_frame_size, 152 r_frame_size, r_frame_alignment_in_bytes); 153 // size += size of call_stub locals 154 __ addi(r_frame_size, 155 r_frame_size, frame::entry_frame_locals_size); 156 157 // push ENTRY_FRAME 158 __ push_frame(r_frame_size, r_temp); 159 160 // initialize call_stub locals (step 1) 161 __ std(r_arg_call_wrapper_addr, 162 _entry_frame_locals_neg(call_wrapper_address), r_entryframe_fp); 163 __ std(r_arg_result_addr, 164 _entry_frame_locals_neg(result_address), r_entryframe_fp); 165 __ std(r_arg_result_type, 166 _entry_frame_locals_neg(result_type), r_entryframe_fp); 167 // we will save arguments_tos_address later 168 169 170 BLOCK_COMMENT("Copy Java arguments"); 171 // copy Java arguments 172 173 // Calculate top_of_arguments_addr which will be R17_tos (not prepushed) later. 174 // FIXME: why not simply use SP+frame::top_ijava_frame_size? 175 __ addi(r_top_of_arguments_addr, 176 R1_SP, frame::top_ijava_frame_abi_size); 177 __ add(r_top_of_arguments_addr, 178 r_top_of_arguments_addr, r_frame_alignment_in_bytes); 179 180 // any arguments to copy? 181 __ cmpdi(CCR0, r_arg_argument_count, 0); 182 __ beq(CCR0, arguments_copied); 183 184 // prepare loop and copy arguments in reverse order 185 { 186 // init CTR with arg_argument_count 187 __ mtctr(r_arg_argument_count); 188 189 // let r_argumentcopy_addr point to last outgoing Java arguments P 190 __ mr(r_argumentcopy_addr, r_top_of_arguments_addr); 191 192 // let r_argument_addr point to last incoming java argument 193 __ add(r_argument_addr, 194 r_arg_argument_addr, r_argument_size_in_bytes); 195 __ addi(r_argument_addr, r_argument_addr, -BytesPerWord); 196 197 // now loop while CTR > 0 and copy arguments 198 { 199 Label next_argument; 200 __ bind(next_argument); 201 202 __ ld(r_temp, 0, r_argument_addr); 203 // argument_addr--; 204 __ addi(r_argument_addr, r_argument_addr, -BytesPerWord); 205 __ std(r_temp, 0, r_argumentcopy_addr); 206 // argumentcopy_addr++; 207 __ addi(r_argumentcopy_addr, r_argumentcopy_addr, BytesPerWord); 208 209 __ bdnz(next_argument); 210 } 211 } 212 213 // Arguments copied, continue. 214 __ bind(arguments_copied); 215 } 216 217 { 218 BLOCK_COMMENT("Call frame manager or native entry."); 219 // Call frame manager or native entry. 220 Register r_new_arg_entry = R14; 221 assert_different_registers(r_new_arg_entry, r_top_of_arguments_addr, 222 r_arg_method, r_arg_thread); 223 224 __ mr(r_new_arg_entry, r_arg_entry); 225 226 // Register state on entry to frame manager / native entry: 227 // 228 // tos - intptr_t* sender tos (prepushed) Lesp = (SP) + copied_arguments_offset - 8 229 // R19_method - Method 230 // R16_thread - JavaThread* 231 232 // Tos must point to last argument - element_size. 233 const Register tos = R15_esp; 234 235 __ addi(tos, r_top_of_arguments_addr, -Interpreter::stackElementSize); 236 237 // initialize call_stub locals (step 2) 238 // now save tos as arguments_tos_address 239 __ std(tos, _entry_frame_locals_neg(arguments_tos_address), r_entryframe_fp); 240 241 // load argument registers for call 242 __ mr(R19_method, r_arg_method); 243 __ mr(R16_thread, r_arg_thread); 244 assert(tos != r_arg_method, "trashed r_arg_method"); 245 assert(tos != r_arg_thread && R19_method != r_arg_thread, "trashed r_arg_thread"); 246 247 // Set R15_prev_state to 0 for simplifying checks in callee. 248 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 249 // Stack on entry to frame manager / native entry: 250 // 251 // F0 [TOP_IJAVA_FRAME_ABI] 252 // alignment (optional) 253 // [outgoing Java arguments] 254 // [ENTRY_FRAME_LOCALS] 255 // F1 [C_FRAME] 256 // ... 257 // 258 259 // global toc register 260 __ load_const_optimized(R29_TOC, MacroAssembler::global_toc(), R11_scratch1); 261 // Remember the senderSP so we interpreter can pop c2i arguments off of the stack 262 // when called via a c2i. 263 264 // Pass initial_caller_sp to framemanager. 265 __ mr(R21_tmp1, R1_SP); 266 267 // Do a light-weight C-call here, r_new_arg_entry holds the address 268 // of the interpreter entry point (frame manager or native entry) 269 // and save runtime-value of LR in return_address. 270 assert(r_new_arg_entry != tos && r_new_arg_entry != R19_method && r_new_arg_entry != R16_thread, 271 "trashed r_new_arg_entry"); 272 return_address = __ call_stub(r_new_arg_entry); 273 } 274 275 { 276 BLOCK_COMMENT("Returned from frame manager or native entry."); 277 // Returned from frame manager or native entry. 278 // Now pop frame, process result, and return to caller. 279 280 // Stack on exit from frame manager / native entry: 281 // 282 // F0 [ABI] 283 // ... 284 // [ENTRY_FRAME_LOCALS] 285 // F1 [C_FRAME] 286 // ... 287 // 288 // Just pop the topmost frame ... 289 // 290 291 Label ret_is_object; 292 Label ret_is_long; 293 Label ret_is_float; 294 Label ret_is_double; 295 296 Register r_entryframe_fp = R30; 297 Register r_lr = R7_ARG5; 298 Register r_cr = R8_ARG6; 299 300 // Reload some volatile registers which we've spilled before the call 301 // to frame manager / native entry. 302 // Access all locals via frame pointer, because we know nothing about 303 // the topmost frame's size. 304 __ ld(r_entryframe_fp, _abi(callers_sp), R1_SP); 305 assert_different_registers(r_entryframe_fp, R3_RET, r_arg_result_addr, r_arg_result_type, r_cr, r_lr); 306 __ ld(r_arg_result_addr, 307 _entry_frame_locals_neg(result_address), r_entryframe_fp); 308 __ ld(r_arg_result_type, 309 _entry_frame_locals_neg(result_type), r_entryframe_fp); 310 __ ld(r_cr, _abi(cr), r_entryframe_fp); 311 __ ld(r_lr, _abi(lr), r_entryframe_fp); 312 313 // pop frame and restore non-volatiles, LR and CR 314 __ mr(R1_SP, r_entryframe_fp); 315 __ mtcr(r_cr); 316 __ mtlr(r_lr); 317 318 // Store result depending on type. Everything that is not 319 // T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE is treated as T_INT. 320 __ cmpwi(CCR0, r_arg_result_type, T_OBJECT); 321 __ cmpwi(CCR1, r_arg_result_type, T_LONG); 322 __ cmpwi(CCR5, r_arg_result_type, T_FLOAT); 323 __ cmpwi(CCR6, r_arg_result_type, T_DOUBLE); 324 325 // restore non-volatile registers 326 __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14)); 327 328 329 // Stack on exit from call_stub: 330 // 331 // 0 [C_FRAME] 332 // ... 333 // 334 // no call_stub frames left. 335 336 // All non-volatiles have been restored at this point!! 337 assert(R3_RET == R3, "R3_RET should be R3"); 338 339 __ beq(CCR0, ret_is_object); 340 __ beq(CCR1, ret_is_long); 341 __ beq(CCR5, ret_is_float); 342 __ beq(CCR6, ret_is_double); 343 344 // default: 345 __ stw(R3_RET, 0, r_arg_result_addr); 346 __ blr(); // return to caller 347 348 // case T_OBJECT: 349 __ bind(ret_is_object); 350 __ std(R3_RET, 0, r_arg_result_addr); 351 __ blr(); // return to caller 352 353 // case T_LONG: 354 __ bind(ret_is_long); 355 __ std(R3_RET, 0, r_arg_result_addr); 356 __ blr(); // return to caller 357 358 // case T_FLOAT: 359 __ bind(ret_is_float); 360 __ stfs(F1_RET, 0, r_arg_result_addr); 361 __ blr(); // return to caller 362 363 // case T_DOUBLE: 364 __ bind(ret_is_double); 365 __ stfd(F1_RET, 0, r_arg_result_addr); 366 __ blr(); // return to caller 367 } 368 369 return start; 370 } 371 372 // Return point for a Java call if there's an exception thrown in 373 // Java code. The exception is caught and transformed into a 374 // pending exception stored in JavaThread that can be tested from 375 // within the VM. 376 // 377 address generate_catch_exception() { 378 StubCodeMark mark(this, "StubRoutines", "catch_exception"); 379 380 address start = __ pc(); 381 382 // Registers alive 383 // 384 // R16_thread 385 // R3_ARG1 - address of pending exception 386 // R4_ARG2 - return address in call stub 387 388 const Register exception_file = R21_tmp1; 389 const Register exception_line = R22_tmp2; 390 391 __ load_const(exception_file, (void*)__FILE__); 392 __ load_const(exception_line, (void*)__LINE__); 393 394 __ std(R3_ARG1, in_bytes(JavaThread::pending_exception_offset()), R16_thread); 395 // store into `char *' 396 __ std(exception_file, in_bytes(JavaThread::exception_file_offset()), R16_thread); 397 // store into `int' 398 __ stw(exception_line, in_bytes(JavaThread::exception_line_offset()), R16_thread); 399 400 // complete return to VM 401 assert(StubRoutines::_call_stub_return_address != NULL, "must have been generated before"); 402 403 __ mtlr(R4_ARG2); 404 // continue in call stub 405 __ blr(); 406 407 return start; 408 } 409 410 // Continuation point for runtime calls returning with a pending 411 // exception. The pending exception check happened in the runtime 412 // or native call stub. The pending exception in Thread is 413 // converted into a Java-level exception. 414 // 415 // Read: 416 // 417 // LR: The pc the runtime library callee wants to return to. 418 // Since the exception occurred in the callee, the return pc 419 // from the point of view of Java is the exception pc. 420 // thread: Needed for method handles. 421 // 422 // Invalidate: 423 // 424 // volatile registers (except below). 425 // 426 // Update: 427 // 428 // R4_ARG2: exception 429 // 430 // (LR is unchanged and is live out). 431 // 432 address generate_forward_exception() { 433 StubCodeMark mark(this, "StubRoutines", "forward_exception"); 434 address start = __ pc(); 435 436#if !defined(PRODUCT) 437 if (VerifyOops) { 438 // Get pending exception oop. 439 __ ld(R3_ARG1, 440 in_bytes(Thread::pending_exception_offset()), 441 R16_thread); 442 // Make sure that this code is only executed if there is a pending exception. 443 { 444 Label L; 445 __ cmpdi(CCR0, R3_ARG1, 0); 446 __ bne(CCR0, L); 447 __ stop("StubRoutines::forward exception: no pending exception (1)"); 448 __ bind(L); 449 } 450 __ verify_oop(R3_ARG1, "StubRoutines::forward exception: not an oop"); 451 } 452#endif 453 454 // Save LR/CR and copy exception pc (LR) into R4_ARG2. 455 __ save_LR_CR(R4_ARG2); 456 __ push_frame_reg_args(0, R0); 457 // Find exception handler. 458 __ call_VM_leaf(CAST_FROM_FN_PTR(address, 459 SharedRuntime::exception_handler_for_return_address), 460 R16_thread, 461 R4_ARG2); 462 // Copy handler's address. 463 __ mtctr(R3_RET); 464 __ pop_frame(); 465 __ restore_LR_CR(R0); 466 467 // Set up the arguments for the exception handler: 468 // - R3_ARG1: exception oop 469 // - R4_ARG2: exception pc. 470 471 // Load pending exception oop. 472 __ ld(R3_ARG1, 473 in_bytes(Thread::pending_exception_offset()), 474 R16_thread); 475 476 // The exception pc is the return address in the caller. 477 // Must load it into R4_ARG2. 478 __ mflr(R4_ARG2); 479 480#ifdef ASSERT 481 // Make sure exception is set. 482 { 483 Label L; 484 __ cmpdi(CCR0, R3_ARG1, 0); 485 __ bne(CCR0, L); 486 __ stop("StubRoutines::forward exception: no pending exception (2)"); 487 __ bind(L); 488 } 489#endif 490 491 // Clear the pending exception. 492 __ li(R0, 0); 493 __ std(R0, 494 in_bytes(Thread::pending_exception_offset()), 495 R16_thread); 496 // Jump to exception handler. 497 __ bctr(); 498 499 return start; 500 } 501 502#undef __ 503#define __ masm-> 504 // Continuation point for throwing of implicit exceptions that are 505 // not handled in the current activation. Fabricates an exception 506 // oop and initiates normal exception dispatching in this 507 // frame. Only callee-saved registers are preserved (through the 508 // normal register window / RegisterMap handling). If the compiler 509 // needs all registers to be preserved between the fault point and 510 // the exception handler then it must assume responsibility for that 511 // in AbstractCompiler::continuation_for_implicit_null_exception or 512 // continuation_for_implicit_division_by_zero_exception. All other 513 // implicit exceptions (e.g., NullPointerException or 514 // AbstractMethodError on entry) are either at call sites or 515 // otherwise assume that stack unwinding will be initiated, so 516 // caller saved registers were assumed volatile in the compiler. 517 // 518 // Note that we generate only this stub into a RuntimeStub, because 519 // it needs to be properly traversed and ignored during GC, so we 520 // change the meaning of the "__" macro within this method. 521 // 522 // Note: the routine set_pc_not_at_call_for_caller in 523 // SharedRuntime.cpp requires that this code be generated into a 524 // RuntimeStub. 525 address generate_throw_exception(const char* name, address runtime_entry, bool restore_saved_exception_pc, 526 Register arg1 = noreg, Register arg2 = noreg) { 527 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0); 528 MacroAssembler* masm = new MacroAssembler(&code); 529 530 OopMapSet* oop_maps = new OopMapSet(); 531 int frame_size_in_bytes = frame::abi_reg_args_size; 532 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0); 533 534 address start = __ pc(); 535 536 __ save_LR_CR(R11_scratch1); 537 538 // Push a frame. 539 __ push_frame_reg_args(0, R11_scratch1); 540 541 address frame_complete_pc = __ pc(); 542 543 if (restore_saved_exception_pc) { 544 __ unimplemented("StubGenerator::throw_exception with restore_saved_exception_pc", 74); 545 } 546 547 // Note that we always have a runtime stub frame on the top of 548 // stack by this point. Remember the offset of the instruction 549 // whose address will be moved to R11_scratch1. 550 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1); 551 552 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 553 554 __ mr(R3_ARG1, R16_thread); 555 if (arg1 != noreg) { 556 __ mr(R4_ARG2, arg1); 557 } 558 if (arg2 != noreg) { 559 __ mr(R5_ARG3, arg2); 560 } 561#if defined(ABI_ELFv2) 562 __ call_c(runtime_entry, relocInfo::none); 563#else 564 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, runtime_entry), relocInfo::none); 565#endif 566 567 // Set an oopmap for the call site. 568 oop_maps->add_gc_map((int)(gc_map_pc - start), map); 569 570 __ reset_last_Java_frame(); 571 572#ifdef ASSERT 573 // Make sure that this code is only executed if there is a pending 574 // exception. 575 { 576 Label L; 577 __ ld(R0, 578 in_bytes(Thread::pending_exception_offset()), 579 R16_thread); 580 __ cmpdi(CCR0, R0, 0); 581 __ bne(CCR0, L); 582 __ stop("StubRoutines::throw_exception: no pending exception"); 583 __ bind(L); 584 } 585#endif 586 587 // Pop frame. 588 __ pop_frame(); 589 590 __ restore_LR_CR(R11_scratch1); 591 592 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry()); 593 __ mtctr(R11_scratch1); 594 __ bctr(); 595 596 // Create runtime stub with OopMap. 597 RuntimeStub* stub = 598 RuntimeStub::new_runtime_stub(name, &code, 599 /*frame_complete=*/ (int)(frame_complete_pc - start), 600 frame_size_in_bytes/wordSize, 601 oop_maps, 602 false); 603 return stub->entry_point(); 604 } 605#undef __ 606#define __ _masm-> 607 608 // Generate G1 pre-write barrier for array. 609 // 610 // Input: 611 // from - register containing src address (only needed for spilling) 612 // to - register containing starting address 613 // count - register containing element count 614 // tmp - scratch register 615 // 616 // Kills: 617 // nothing 618 // 619 void gen_write_ref_array_pre_barrier(Register from, Register to, Register count, bool dest_uninitialized, Register Rtmp1, 620 Register preserve1 = noreg, Register preserve2 = noreg) { 621 BarrierSet* const bs = Universe::heap()->barrier_set(); 622 switch (bs->kind()) { 623 case BarrierSet::G1SATBCTLogging: 624 // With G1, don't generate the call if we statically know that the target in uninitialized 625 if (!dest_uninitialized) { 626 int spill_slots = 3; 627 if (preserve1 != noreg) { spill_slots++; } 628 if (preserve2 != noreg) { spill_slots++; } 629 const int frame_size = align_size_up(frame::abi_reg_args_size + spill_slots * BytesPerWord, frame::alignment_in_bytes); 630 Label filtered; 631 632 // Is marking active? 633 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 634 __ lwz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active()), R16_thread); 635 } else { 636 guarantee(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 637 __ lbz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::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(frame_size, R0); 644 int slot_nr = 0; 645 __ std(from, frame_size - (++slot_nr) * wordSize, R1_SP); 646 __ std(to, frame_size - (++slot_nr) * wordSize, R1_SP); 647 __ std(count, frame_size - (++slot_nr) * wordSize, R1_SP); 648 if (preserve1 != noreg) { __ std(preserve1, frame_size - (++slot_nr) * wordSize, R1_SP); } 649 if (preserve2 != noreg) { __ std(preserve2, frame_size - (++slot_nr) * wordSize, R1_SP); } 650 651 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), to, count); 652 653 slot_nr = 0; 654 __ ld(from, frame_size - (++slot_nr) * wordSize, R1_SP); 655 __ ld(to, frame_size - (++slot_nr) * wordSize, R1_SP); 656 __ ld(count, frame_size - (++slot_nr) * wordSize, R1_SP); 657 if (preserve1 != noreg) { __ ld(preserve1, frame_size - (++slot_nr) * wordSize, R1_SP); } 658 if (preserve2 != noreg) { __ ld(preserve2, frame_size - (++slot_nr) * wordSize, R1_SP); } 659 __ addi(R1_SP, R1_SP, frame_size); // pop_frame() 660 __ restore_LR_CR(R0); 661 662 __ bind(filtered); 663 } 664 break; 665 case BarrierSet::CardTableForRS: 666 case BarrierSet::CardTableExtension: 667 case BarrierSet::ModRef: 668 break; 669 default: 670 ShouldNotReachHere(); 671 } 672 } 673 674 // Generate CMS/G1 post-write barrier for array. 675 // 676 // Input: 677 // addr - register containing starting address 678 // count - register containing element count 679 // tmp - scratch register 680 // 681 // The input registers and R0 are overwritten. 682 // 683 void gen_write_ref_array_post_barrier(Register addr, Register count, Register tmp, Register preserve = noreg) { 684 BarrierSet* const bs = Universe::heap()->barrier_set(); 685 686 switch (bs->kind()) { 687 case BarrierSet::G1SATBCTLogging: 688 { 689 int spill_slots = (preserve != noreg) ? 1 : 0; 690 const int frame_size = align_size_up(frame::abi_reg_args_size + spill_slots * BytesPerWord, frame::alignment_in_bytes); 691 692 __ save_LR_CR(R0); 693 __ push_frame(frame_size, R0); 694 if (preserve != noreg) { __ std(preserve, frame_size - 1 * wordSize, R1_SP); } 695 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), addr, count); 696 if (preserve != noreg) { __ ld(preserve, frame_size - 1 * wordSize, R1_SP); } 697 __ addi(R1_SP, R1_SP, frame_size); // pop_frame(); 698 __ restore_LR_CR(R0); 699 } 700 break; 701 case BarrierSet::CardTableForRS: 702 case BarrierSet::CardTableExtension: 703 { 704 Label Lskip_loop, Lstore_loop; 705 if (UseConcMarkSweepGC) { 706 // TODO PPC port: contribute optimization / requires shared changes 707 __ release(); 708 } 709 710 CardTableModRefBS* const ct = barrier_set_cast<CardTableModRefBS>(bs); 711 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 712 assert_different_registers(addr, count, tmp); 713 714 __ sldi(count, count, LogBytesPerHeapOop); 715 __ addi(count, count, -BytesPerHeapOop); 716 __ add(count, addr, count); 717 // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.) 718 __ srdi(addr, addr, CardTableModRefBS::card_shift); 719 __ srdi(count, count, CardTableModRefBS::card_shift); 720 __ subf(count, addr, count); 721 assert_different_registers(R0, addr, count, tmp); 722 __ load_const(tmp, (address)ct->byte_map_base); 723 __ addic_(count, count, 1); 724 __ beq(CCR0, Lskip_loop); 725 __ li(R0, 0); 726 __ mtctr(count); 727 // Byte store loop 728 __ bind(Lstore_loop); 729 __ stbx(R0, tmp, addr); 730 __ addi(addr, addr, 1); 731 __ bdnz(Lstore_loop); 732 __ bind(Lskip_loop); 733 } 734 break; 735 case BarrierSet::ModRef: 736 break; 737 default: 738 ShouldNotReachHere(); 739 } 740 } 741 742 // Support for void zero_words_aligned8(HeapWord* to, size_t count) 743 // 744 // Arguments: 745 // to: 746 // count: 747 // 748 // Destroys: 749 // 750 address generate_zero_words_aligned8() { 751 StubCodeMark mark(this, "StubRoutines", "zero_words_aligned8"); 752 753 // Implemented as in ClearArray. 754 address start = __ function_entry(); 755 756 Register base_ptr_reg = R3_ARG1; // tohw (needs to be 8b aligned) 757 Register cnt_dwords_reg = R4_ARG2; // count (in dwords) 758 Register tmp1_reg = R5_ARG3; 759 Register tmp2_reg = R6_ARG4; 760 Register zero_reg = R7_ARG5; 761 762 // Procedure for large arrays (uses data cache block zero instruction). 763 Label dwloop, fast, fastloop, restloop, lastdword, done; 764 int cl_size = VM_Version::L1_data_cache_line_size(); 765 int cl_dwords = cl_size >> 3; 766 int 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#if !defined(PRODUCT) 829 // Wrapper which calls oopDesc::is_oop_or_null() 830 // Only called by MacroAssembler::verify_oop 831 static void verify_oop_helper(const char* message, oop o) { 832 if (!o->is_oop_or_null()) { 833 fatal("%s", message); 834 } 835 ++ StubRoutines::_verify_oop_count; 836 } 837#endif 838 839 // Return address of code to be called from code generated by 840 // MacroAssembler::verify_oop. 841 // 842 // Don't generate, rather use C++ code. 843 address generate_verify_oop() { 844 // this is actually a `FunctionDescriptor*'. 845 address start = 0; 846 847#if !defined(PRODUCT) 848 start = CAST_FROM_FN_PTR(address, verify_oop_helper); 849#endif 850 851 return start; 852 } 853 854 // Fairer handling of safepoints for native methods. 855 // 856 // Generate code which reads from the polling page. This special handling is needed as the 857 // linux-ppc64 kernel before 2.6.6 doesn't set si_addr on some segfaults in 64bit mode 858 // (cf. http://www.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.6), especially when we try 859 // to read from the safepoint polling page. 860 address generate_load_from_poll() { 861 StubCodeMark mark(this, "StubRoutines", "generate_load_from_poll"); 862 address start = __ function_entry(); 863 __ unimplemented("StubRoutines::verify_oop", 95); // TODO PPC port 864 return start; 865 } 866 867 // -XX:+OptimizeFill : convert fill/copy loops into intrinsic 868 // 869 // The code is implemented(ported from sparc) as we believe it benefits JVM98, however 870 // tracing(-XX:+TraceOptimizeFill) shows the intrinsic replacement doesn't happen at all! 871 // 872 // Source code in function is_range_check_if() shows that OptimizeFill relaxed the condition 873 // for turning on loop predication optimization, and hence the behavior of "array range check" 874 // and "loop invariant check" could be influenced, which potentially boosted JVM98. 875 // 876 // Generate stub for disjoint short fill. If "aligned" is true, the 877 // "to" address is assumed to be heapword aligned. 878 // 879 // Arguments for generated stub: 880 // to: R3_ARG1 881 // value: R4_ARG2 882 // count: R5_ARG3 treated as signed 883 // 884 address generate_fill(BasicType t, bool aligned, const char* name) { 885 StubCodeMark mark(this, "StubRoutines", name); 886 address start = __ function_entry(); 887 888 const Register to = R3_ARG1; // source array address 889 const Register value = R4_ARG2; // fill value 890 const Register count = R5_ARG3; // elements count 891 const Register temp = R6_ARG4; // temp register 892 893 //assert_clean_int(count, O3); // Make sure 'count' is clean int. 894 895 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; 896 Label L_fill_2_bytes, L_fill_4_bytes, L_fill_elements, L_fill_32_bytes; 897 898 int shift = -1; 899 switch (t) { 900 case T_BYTE: 901 shift = 2; 902 // Clone bytes (zero extend not needed because store instructions below ignore high order bytes). 903 __ rldimi(value, value, 8, 48); // 8 bit -> 16 bit 904 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element. 905 __ blt(CCR0, L_fill_elements); 906 __ rldimi(value, value, 16, 32); // 16 bit -> 32 bit 907 break; 908 case T_SHORT: 909 shift = 1; 910 // Clone bytes (zero extend not needed because store instructions below ignore high order bytes). 911 __ rldimi(value, value, 16, 32); // 16 bit -> 32 bit 912 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element. 913 __ blt(CCR0, L_fill_elements); 914 break; 915 case T_INT: 916 shift = 0; 917 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element. 918 __ blt(CCR0, L_fill_4_bytes); 919 break; 920 default: ShouldNotReachHere(); 921 } 922 923 if (!aligned && (t == T_BYTE || t == T_SHORT)) { 924 // Align source address at 4 bytes address boundary. 925 if (t == T_BYTE) { 926 // One byte misalignment happens only for byte arrays. 927 __ andi_(temp, to, 1); 928 __ beq(CCR0, L_skip_align1); 929 __ stb(value, 0, to); 930 __ addi(to, to, 1); 931 __ addi(count, count, -1); 932 __ bind(L_skip_align1); 933 } 934 // Two bytes misalignment happens only for byte and short (char) arrays. 935 __ andi_(temp, to, 2); 936 __ beq(CCR0, L_skip_align2); 937 __ sth(value, 0, to); 938 __ addi(to, to, 2); 939 __ addi(count, count, -(1 << (shift - 1))); 940 __ bind(L_skip_align2); 941 } 942 943 if (!aligned) { 944 // Align to 8 bytes, we know we are 4 byte aligned to start. 945 __ andi_(temp, to, 7); 946 __ beq(CCR0, L_fill_32_bytes); 947 __ stw(value, 0, to); 948 __ addi(to, to, 4); 949 __ addi(count, count, -(1 << shift)); 950 __ bind(L_fill_32_bytes); 951 } 952 953 __ li(temp, 8<<shift); // Prepare for 32 byte loop. 954 // Clone bytes int->long as above. 955 __ rldimi(value, value, 32, 0); // 32 bit -> 64 bit 956 957 Label L_check_fill_8_bytes; 958 // Fill 32-byte chunks. 959 __ subf_(count, temp, count); 960 __ blt(CCR0, L_check_fill_8_bytes); 961 962 Label L_fill_32_bytes_loop; 963 __ align(32); 964 __ bind(L_fill_32_bytes_loop); 965 966 __ std(value, 0, to); 967 __ std(value, 8, to); 968 __ subf_(count, temp, count); // Update count. 969 __ std(value, 16, to); 970 __ std(value, 24, to); 971 972 __ addi(to, to, 32); 973 __ bge(CCR0, L_fill_32_bytes_loop); 974 975 __ bind(L_check_fill_8_bytes); 976 __ add_(count, temp, count); 977 __ beq(CCR0, L_exit); 978 __ addic_(count, count, -(2 << shift)); 979 __ blt(CCR0, L_fill_4_bytes); 980 981 // 982 // Length is too short, just fill 8 bytes at a time. 983 // 984 Label L_fill_8_bytes_loop; 985 __ bind(L_fill_8_bytes_loop); 986 __ std(value, 0, to); 987 __ addic_(count, count, -(2 << shift)); 988 __ addi(to, to, 8); 989 __ bge(CCR0, L_fill_8_bytes_loop); 990 991 // Fill trailing 4 bytes. 992 __ bind(L_fill_4_bytes); 993 __ andi_(temp, count, 1<<shift); 994 __ beq(CCR0, L_fill_2_bytes); 995 996 __ stw(value, 0, to); 997 if (t == T_BYTE || t == T_SHORT) { 998 __ addi(to, to, 4); 999 // Fill trailing 2 bytes. 1000 __ bind(L_fill_2_bytes); 1001 __ andi_(temp, count, 1<<(shift-1)); 1002 __ beq(CCR0, L_fill_byte); 1003 __ sth(value, 0, to); 1004 if (t == T_BYTE) { 1005 __ addi(to, to, 2); 1006 // Fill trailing byte. 1007 __ bind(L_fill_byte); 1008 __ andi_(count, count, 1); 1009 __ beq(CCR0, L_exit); 1010 __ stb(value, 0, to); 1011 } else { 1012 __ bind(L_fill_byte); 1013 } 1014 } else { 1015 __ bind(L_fill_2_bytes); 1016 } 1017 __ bind(L_exit); 1018 __ blr(); 1019 1020 // Handle copies less than 8 bytes. Int is handled elsewhere. 1021 if (t == T_BYTE) { 1022 __ bind(L_fill_elements); 1023 Label L_fill_2, L_fill_4; 1024 __ andi_(temp, count, 1); 1025 __ beq(CCR0, L_fill_2); 1026 __ stb(value, 0, to); 1027 __ addi(to, to, 1); 1028 __ bind(L_fill_2); 1029 __ andi_(temp, count, 2); 1030 __ beq(CCR0, L_fill_4); 1031 __ stb(value, 0, to); 1032 __ stb(value, 0, to); 1033 __ addi(to, to, 2); 1034 __ bind(L_fill_4); 1035 __ andi_(temp, count, 4); 1036 __ beq(CCR0, L_exit); 1037 __ stb(value, 0, to); 1038 __ stb(value, 1, to); 1039 __ stb(value, 2, to); 1040 __ stb(value, 3, to); 1041 __ blr(); 1042 } 1043 1044 if (t == T_SHORT) { 1045 Label L_fill_2; 1046 __ bind(L_fill_elements); 1047 __ andi_(temp, count, 1); 1048 __ beq(CCR0, L_fill_2); 1049 __ sth(value, 0, to); 1050 __ addi(to, to, 2); 1051 __ bind(L_fill_2); 1052 __ andi_(temp, count, 2); 1053 __ beq(CCR0, L_exit); 1054 __ sth(value, 0, to); 1055 __ sth(value, 2, to); 1056 __ blr(); 1057 } 1058 return start; 1059 } 1060 1061 inline void assert_positive_int(Register count) { 1062#ifdef ASSERT 1063 __ srdi_(R0, count, 31); 1064 __ asm_assert_eq("missing zero extend", 0xAFFE); 1065#endif 1066 } 1067 1068 // Generate overlap test for array copy stubs. 1069 // 1070 // Input: 1071 // R3_ARG1 - from 1072 // R4_ARG2 - to 1073 // R5_ARG3 - element count 1074 // 1075 void array_overlap_test(address no_overlap_target, int log2_elem_size) { 1076 Register tmp1 = R6_ARG4; 1077 Register tmp2 = R7_ARG5; 1078 1079 assert_positive_int(R5_ARG3); 1080 1081 __ subf(tmp1, R3_ARG1, R4_ARG2); // distance in bytes 1082 __ sldi(tmp2, R5_ARG3, log2_elem_size); // size in bytes 1083 __ cmpld(CCR0, R3_ARG1, R4_ARG2); // Use unsigned comparison! 1084 __ cmpld(CCR1, tmp1, tmp2); 1085 __ crnand(CCR0, Assembler::less, CCR1, Assembler::less); 1086 // Overlaps if Src before dst and distance smaller than size. 1087 // Branch to forward copy routine otherwise (within range of 32kB). 1088 __ bc(Assembler::bcondCRbiIs1, Assembler::bi0(CCR0, Assembler::less), no_overlap_target); 1089 1090 // need to copy backwards 1091 } 1092 1093 // The guideline in the implementations of generate_disjoint_xxx_copy 1094 // (xxx=byte,short,int,long,oop) is to copy as many elements as possible with 1095 // single instructions, but to avoid alignment interrupts (see subsequent 1096 // comment). Furthermore, we try to minimize misaligned access, even 1097 // though they cause no alignment interrupt. 1098 // 1099 // In Big-Endian mode, the PowerPC architecture requires implementations to 1100 // handle automatically misaligned integer halfword and word accesses, 1101 // word-aligned integer doubleword accesses, and word-aligned floating-point 1102 // accesses. Other accesses may or may not generate an Alignment interrupt 1103 // depending on the implementation. 1104 // Alignment interrupt handling may require on the order of hundreds of cycles, 1105 // so every effort should be made to avoid misaligned memory values. 1106 // 1107 // 1108 // Generate stub for disjoint byte copy. If "aligned" is true, the 1109 // "from" and "to" addresses are assumed to be heapword aligned. 1110 // 1111 // Arguments for generated stub: 1112 // from: R3_ARG1 1113 // to: R4_ARG2 1114 // count: R5_ARG3 treated as signed 1115 // 1116 address generate_disjoint_byte_copy(bool aligned, const char * name) { 1117 StubCodeMark mark(this, "StubRoutines", name); 1118 address start = __ function_entry(); 1119 assert_positive_int(R5_ARG3); 1120 1121 Register tmp1 = R6_ARG4; 1122 Register tmp2 = R7_ARG5; 1123 Register tmp3 = R8_ARG6; 1124 Register tmp4 = R9_ARG7; 1125 1126 Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8, l_9; 1127 1128 // Don't try anything fancy if arrays don't have many elements. 1129 __ li(tmp3, 0); 1130 __ cmpwi(CCR0, R5_ARG3, 17); 1131 __ ble(CCR0, l_6); // copy 4 at a time 1132 1133 if (!aligned) { 1134 __ xorr(tmp1, R3_ARG1, R4_ARG2); 1135 __ andi_(tmp1, tmp1, 3); 1136 __ bne(CCR0, l_6); // If arrays don't have the same alignment mod 4, do 4 element copy. 1137 1138 // Copy elements if necessary to align to 4 bytes. 1139 __ neg(tmp1, R3_ARG1); // Compute distance to alignment boundary. 1140 __ andi_(tmp1, tmp1, 3); 1141 __ beq(CCR0, l_2); 1142 1143 __ subf(R5_ARG3, tmp1, R5_ARG3); 1144 __ bind(l_9); 1145 __ lbz(tmp2, 0, R3_ARG1); 1146 __ addic_(tmp1, tmp1, -1); 1147 __ stb(tmp2, 0, R4_ARG2); 1148 __ addi(R3_ARG1, R3_ARG1, 1); 1149 __ addi(R4_ARG2, R4_ARG2, 1); 1150 __ bne(CCR0, l_9); 1151 1152 __ bind(l_2); 1153 } 1154 1155 // copy 8 elements at a time 1156 __ xorr(tmp2, R3_ARG1, R4_ARG2); // skip if src & dest have differing alignment mod 8 1157 __ andi_(tmp1, tmp2, 7); 1158 __ bne(CCR0, l_7); // not same alignment -> to or from is aligned -> copy 8 1159 1160 // copy a 2-element word if necessary to align to 8 bytes 1161 __ andi_(R0, R3_ARG1, 7); 1162 __ beq(CCR0, l_7); 1163 1164 __ lwzx(tmp2, R3_ARG1, tmp3); 1165 __ addi(R5_ARG3, R5_ARG3, -4); 1166 __ stwx(tmp2, R4_ARG2, tmp3); 1167 { // FasterArrayCopy 1168 __ addi(R3_ARG1, R3_ARG1, 4); 1169 __ addi(R4_ARG2, R4_ARG2, 4); 1170 } 1171 __ bind(l_7); 1172 1173 { // FasterArrayCopy 1174 __ cmpwi(CCR0, R5_ARG3, 31); 1175 __ ble(CCR0, l_6); // copy 2 at a time if less than 32 elements remain 1176 1177 __ srdi(tmp1, R5_ARG3, 5); 1178 __ andi_(R5_ARG3, R5_ARG3, 31); 1179 __ mtctr(tmp1); 1180 1181 __ bind(l_8); 1182 // Use unrolled version for mass copying (copy 32 elements a time) 1183 // Load feeding store gets zero latency on Power6, however not on Power5. 1184 // Therefore, the following sequence is made for the good of both. 1185 __ ld(tmp1, 0, R3_ARG1); 1186 __ ld(tmp2, 8, R3_ARG1); 1187 __ ld(tmp3, 16, R3_ARG1); 1188 __ ld(tmp4, 24, R3_ARG1); 1189 __ std(tmp1, 0, R4_ARG2); 1190 __ std(tmp2, 8, R4_ARG2); 1191 __ std(tmp3, 16, R4_ARG2); 1192 __ std(tmp4, 24, R4_ARG2); 1193 __ addi(R3_ARG1, R3_ARG1, 32); 1194 __ addi(R4_ARG2, R4_ARG2, 32); 1195 __ bdnz(l_8); 1196 } 1197 1198 __ bind(l_6); 1199 1200 // copy 4 elements at a time 1201 __ cmpwi(CCR0, R5_ARG3, 4); 1202 __ blt(CCR0, l_1); 1203 __ srdi(tmp1, R5_ARG3, 2); 1204 __ mtctr(tmp1); // is > 0 1205 __ andi_(R5_ARG3, R5_ARG3, 3); 1206 1207 { // FasterArrayCopy 1208 __ addi(R3_ARG1, R3_ARG1, -4); 1209 __ addi(R4_ARG2, R4_ARG2, -4); 1210 __ bind(l_3); 1211 __ lwzu(tmp2, 4, R3_ARG1); 1212 __ stwu(tmp2, 4, R4_ARG2); 1213 __ bdnz(l_3); 1214 __ addi(R3_ARG1, R3_ARG1, 4); 1215 __ addi(R4_ARG2, R4_ARG2, 4); 1216 } 1217 1218 // do single element copy 1219 __ bind(l_1); 1220 __ cmpwi(CCR0, R5_ARG3, 0); 1221 __ beq(CCR0, l_4); 1222 1223 { // FasterArrayCopy 1224 __ mtctr(R5_ARG3); 1225 __ addi(R3_ARG1, R3_ARG1, -1); 1226 __ addi(R4_ARG2, R4_ARG2, -1); 1227 1228 __ bind(l_5); 1229 __ lbzu(tmp2, 1, R3_ARG1); 1230 __ stbu(tmp2, 1, R4_ARG2); 1231 __ bdnz(l_5); 1232 } 1233 1234 __ bind(l_4); 1235 __ li(R3_RET, 0); // return 0 1236 __ blr(); 1237 1238 return start; 1239 } 1240 1241 // Generate stub for conjoint byte copy. If "aligned" is true, the 1242 // "from" and "to" addresses are assumed to be heapword aligned. 1243 // 1244 // Arguments for generated stub: 1245 // from: R3_ARG1 1246 // to: R4_ARG2 1247 // count: R5_ARG3 treated as signed 1248 // 1249 address generate_conjoint_byte_copy(bool aligned, const char * name) { 1250 StubCodeMark mark(this, "StubRoutines", name); 1251 address start = __ function_entry(); 1252 assert_positive_int(R5_ARG3); 1253 1254 Register tmp1 = R6_ARG4; 1255 Register tmp2 = R7_ARG5; 1256 Register tmp3 = R8_ARG6; 1257 1258 address nooverlap_target = aligned ? 1259 STUB_ENTRY(arrayof_jbyte_disjoint_arraycopy) : 1260 STUB_ENTRY(jbyte_disjoint_arraycopy); 1261 1262 array_overlap_test(nooverlap_target, 0); 1263 // Do reverse copy. We assume the case of actual overlap is rare enough 1264 // that we don't have to optimize it. 1265 Label l_1, l_2; 1266 1267 __ b(l_2); 1268 __ bind(l_1); 1269 __ stbx(tmp1, R4_ARG2, R5_ARG3); 1270 __ bind(l_2); 1271 __ addic_(R5_ARG3, R5_ARG3, -1); 1272 __ lbzx(tmp1, R3_ARG1, R5_ARG3); 1273 __ bge(CCR0, l_1); 1274 1275 __ li(R3_RET, 0); // return 0 1276 __ blr(); 1277 1278 return start; 1279 } 1280 1281 // Generate stub for disjoint short copy. If "aligned" is true, the 1282 // "from" and "to" addresses are assumed to be heapword aligned. 1283 // 1284 // Arguments for generated stub: 1285 // from: R3_ARG1 1286 // to: R4_ARG2 1287 // elm.count: R5_ARG3 treated as signed 1288 // 1289 // Strategy for aligned==true: 1290 // 1291 // If length <= 9: 1292 // 1. copy 2 elements at a time (l_6) 1293 // 2. copy last element if original element count was odd (l_1) 1294 // 1295 // If length > 9: 1296 // 1. copy 4 elements at a time until less than 4 elements are left (l_7) 1297 // 2. copy 2 elements at a time until less than 2 elements are left (l_6) 1298 // 3. copy last element if one was left in step 2. (l_1) 1299 // 1300 // 1301 // Strategy for aligned==false: 1302 // 1303 // If length <= 9: same as aligned==true case, but NOTE: load/stores 1304 // can be unaligned (see comment below) 1305 // 1306 // If length > 9: 1307 // 1. continue with step 6. if the alignment of from and to mod 4 1308 // is different. 1309 // 2. align from and to to 4 bytes by copying 1 element if necessary 1310 // 3. at l_2 from and to are 4 byte aligned; continue with 1311 // 5. if they cannot be aligned to 8 bytes because they have 1312 // got different alignment mod 8. 1313 // 4. at this point we know that both, from and to, have the same 1314 // alignment mod 8, now copy one element if necessary to get 1315 // 8 byte alignment of from and to. 1316 // 5. copy 4 elements at a time until less than 4 elements are 1317 // left; depending on step 3. all load/stores are aligned or 1318 // either all loads or all stores are unaligned. 1319 // 6. copy 2 elements at a time until less than 2 elements are 1320 // left (l_6); arriving here from step 1., there is a chance 1321 // that all accesses are unaligned. 1322 // 7. copy last element if one was left in step 6. (l_1) 1323 // 1324 // There are unaligned data accesses using integer load/store 1325 // instructions in this stub. POWER allows such accesses. 1326 // 1327 // According to the manuals (PowerISA_V2.06_PUBLIC, Book II, 1328 // Chapter 2: Effect of Operand Placement on Performance) unaligned 1329 // integer load/stores have good performance. Only unaligned 1330 // floating point load/stores can have poor performance. 1331 // 1332 // TODO: 1333 // 1334 // 1. check if aligning the backbranch target of loops is beneficial 1335 // 1336 address generate_disjoint_short_copy(bool aligned, const char * name) { 1337 StubCodeMark mark(this, "StubRoutines", name); 1338 1339 Register tmp1 = R6_ARG4; 1340 Register tmp2 = R7_ARG5; 1341 Register tmp3 = R8_ARG6; 1342 Register tmp4 = R9_ARG7; 1343 1344 address start = __ function_entry(); 1345 assert_positive_int(R5_ARG3); 1346 1347 Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8; 1348 1349 // don't try anything fancy if arrays don't have many elements 1350 __ li(tmp3, 0); 1351 __ cmpwi(CCR0, R5_ARG3, 9); 1352 __ ble(CCR0, l_6); // copy 2 at a time 1353 1354 if (!aligned) { 1355 __ xorr(tmp1, R3_ARG1, R4_ARG2); 1356 __ andi_(tmp1, tmp1, 3); 1357 __ bne(CCR0, l_6); // if arrays don't have the same alignment mod 4, do 2 element copy 1358 1359 // At this point it is guaranteed that both, from and to have the same alignment mod 4. 1360 1361 // Copy 1 element if necessary to align to 4 bytes. 1362 __ andi_(tmp1, R3_ARG1, 3); 1363 __ beq(CCR0, l_2); 1364 1365 __ lhz(tmp2, 0, R3_ARG1); 1366 __ addi(R3_ARG1, R3_ARG1, 2); 1367 __ sth(tmp2, 0, R4_ARG2); 1368 __ addi(R4_ARG2, R4_ARG2, 2); 1369 __ addi(R5_ARG3, R5_ARG3, -1); 1370 __ bind(l_2); 1371 1372 // At this point the positions of both, from and to, are at least 4 byte aligned. 1373 1374 // Copy 4 elements at a time. 1375 // Align to 8 bytes, but only if both, from and to, have same alignment mod 8. 1376 __ xorr(tmp2, R3_ARG1, R4_ARG2); 1377 __ andi_(tmp1, tmp2, 7); 1378 __ bne(CCR0, l_7); // not same alignment mod 8 -> copy 4, either from or to will be unaligned 1379 1380 // Copy a 2-element word if necessary to align to 8 bytes. 1381 __ andi_(R0, R3_ARG1, 7); 1382 __ beq(CCR0, l_7); 1383 1384 __ lwzx(tmp2, R3_ARG1, tmp3); 1385 __ addi(R5_ARG3, R5_ARG3, -2); 1386 __ stwx(tmp2, R4_ARG2, tmp3); 1387 { // FasterArrayCopy 1388 __ addi(R3_ARG1, R3_ARG1, 4); 1389 __ addi(R4_ARG2, R4_ARG2, 4); 1390 } 1391 } 1392 1393 __ bind(l_7); 1394 1395 // Copy 4 elements at a time; either the loads or the stores can 1396 // be unaligned if aligned == false. 1397 1398 { // FasterArrayCopy 1399 __ cmpwi(CCR0, R5_ARG3, 15); 1400 __ ble(CCR0, l_6); // copy 2 at a time if less than 16 elements remain 1401 1402 __ srdi(tmp1, R5_ARG3, 4); 1403 __ andi_(R5_ARG3, R5_ARG3, 15); 1404 __ mtctr(tmp1); 1405 1406 __ bind(l_8); 1407 // Use unrolled version for mass copying (copy 16 elements a time). 1408 // Load feeding store gets zero latency on Power6, however not on Power5. 1409 // Therefore, the following sequence is made for the good of both. 1410 __ ld(tmp1, 0, R3_ARG1); 1411 __ ld(tmp2, 8, R3_ARG1); 1412 __ ld(tmp3, 16, R3_ARG1); 1413 __ ld(tmp4, 24, R3_ARG1); 1414 __ std(tmp1, 0, R4_ARG2); 1415 __ std(tmp2, 8, R4_ARG2); 1416 __ std(tmp3, 16, R4_ARG2); 1417 __ std(tmp4, 24, R4_ARG2); 1418 __ addi(R3_ARG1, R3_ARG1, 32); 1419 __ addi(R4_ARG2, R4_ARG2, 32); 1420 __ bdnz(l_8); 1421 } 1422 __ bind(l_6); 1423 1424 // copy 2 elements at a time 1425 { // FasterArrayCopy 1426 __ cmpwi(CCR0, R5_ARG3, 2); 1427 __ blt(CCR0, l_1); 1428 __ srdi(tmp1, R5_ARG3, 1); 1429 __ andi_(R5_ARG3, R5_ARG3, 1); 1430 1431 __ addi(R3_ARG1, R3_ARG1, -4); 1432 __ addi(R4_ARG2, R4_ARG2, -4); 1433 __ mtctr(tmp1); 1434 1435 __ bind(l_3); 1436 __ lwzu(tmp2, 4, R3_ARG1); 1437 __ stwu(tmp2, 4, R4_ARG2); 1438 __ bdnz(l_3); 1439 1440 __ addi(R3_ARG1, R3_ARG1, 4); 1441 __ addi(R4_ARG2, R4_ARG2, 4); 1442 } 1443 1444 // do single element copy 1445 __ bind(l_1); 1446 __ cmpwi(CCR0, R5_ARG3, 0); 1447 __ beq(CCR0, l_4); 1448 1449 { // FasterArrayCopy 1450 __ mtctr(R5_ARG3); 1451 __ addi(R3_ARG1, R3_ARG1, -2); 1452 __ addi(R4_ARG2, R4_ARG2, -2); 1453 1454 __ bind(l_5); 1455 __ lhzu(tmp2, 2, R3_ARG1); 1456 __ sthu(tmp2, 2, R4_ARG2); 1457 __ bdnz(l_5); 1458 } 1459 __ bind(l_4); 1460 __ li(R3_RET, 0); // return 0 1461 __ blr(); 1462 1463 return start; 1464 } 1465 1466 // Generate stub for conjoint short copy. If "aligned" is true, the 1467 // "from" and "to" addresses are assumed to be heapword aligned. 1468 // 1469 // Arguments for generated stub: 1470 // from: R3_ARG1 1471 // to: R4_ARG2 1472 // count: R5_ARG3 treated as signed 1473 // 1474 address generate_conjoint_short_copy(bool aligned, const char * name) { 1475 StubCodeMark mark(this, "StubRoutines", name); 1476 address start = __ function_entry(); 1477 assert_positive_int(R5_ARG3); 1478 1479 Register tmp1 = R6_ARG4; 1480 Register tmp2 = R7_ARG5; 1481 Register tmp3 = R8_ARG6; 1482 1483 address nooverlap_target = aligned ? 1484 STUB_ENTRY(arrayof_jshort_disjoint_arraycopy) : 1485 STUB_ENTRY(jshort_disjoint_arraycopy); 1486 1487 array_overlap_test(nooverlap_target, 1); 1488 1489 Label l_1, l_2; 1490 __ sldi(tmp1, R5_ARG3, 1); 1491 __ b(l_2); 1492 __ bind(l_1); 1493 __ sthx(tmp2, R4_ARG2, tmp1); 1494 __ bind(l_2); 1495 __ addic_(tmp1, tmp1, -2); 1496 __ lhzx(tmp2, R3_ARG1, tmp1); 1497 __ bge(CCR0, l_1); 1498 1499 __ li(R3_RET, 0); // return 0 1500 __ blr(); 1501 1502 return start; 1503 } 1504 1505 // Generate core code for disjoint int copy (and oop copy on 32-bit). If "aligned" 1506 // is true, the "from" and "to" addresses are assumed to be heapword aligned. 1507 // 1508 // Arguments: 1509 // from: R3_ARG1 1510 // to: R4_ARG2 1511 // count: R5_ARG3 treated as signed 1512 // 1513 void generate_disjoint_int_copy_core(bool aligned) { 1514 Register tmp1 = R6_ARG4; 1515 Register tmp2 = R7_ARG5; 1516 Register tmp3 = R8_ARG6; 1517 Register tmp4 = R0; 1518 1519 Label l_1, l_2, l_3, l_4, l_5, l_6; 1520 1521 // for short arrays, just do single element copy 1522 __ li(tmp3, 0); 1523 __ cmpwi(CCR0, R5_ARG3, 5); 1524 __ ble(CCR0, l_2); 1525 1526 if (!aligned) { 1527 // check if arrays have same alignment mod 8. 1528 __ xorr(tmp1, R3_ARG1, R4_ARG2); 1529 __ andi_(R0, tmp1, 7); 1530 // Not the same alignment, but ld and std just need to be 4 byte aligned. 1531 __ bne(CCR0, l_4); // to OR from is 8 byte aligned -> copy 2 at a time 1532 1533 // copy 1 element to align to and from on an 8 byte boundary 1534 __ andi_(R0, R3_ARG1, 7); 1535 __ beq(CCR0, l_4); 1536 1537 __ lwzx(tmp2, R3_ARG1, tmp3); 1538 __ addi(R5_ARG3, R5_ARG3, -1); 1539 __ stwx(tmp2, R4_ARG2, tmp3); 1540 { // FasterArrayCopy 1541 __ addi(R3_ARG1, R3_ARG1, 4); 1542 __ addi(R4_ARG2, R4_ARG2, 4); 1543 } 1544 __ bind(l_4); 1545 } 1546 1547 { // FasterArrayCopy 1548 __ cmpwi(CCR0, R5_ARG3, 7); 1549 __ ble(CCR0, l_2); // copy 1 at a time if less than 8 elements remain 1550 1551 __ srdi(tmp1, R5_ARG3, 3); 1552 __ andi_(R5_ARG3, R5_ARG3, 7); 1553 __ mtctr(tmp1); 1554 1555 __ bind(l_6); 1556 // Use unrolled version for mass copying (copy 8 elements a time). 1557 // Load feeding store gets zero latency on power6, however not on power 5. 1558 // Therefore, the following sequence is made for the good of both. 1559 __ ld(tmp1, 0, R3_ARG1); 1560 __ ld(tmp2, 8, R3_ARG1); 1561 __ ld(tmp3, 16, R3_ARG1); 1562 __ ld(tmp4, 24, R3_ARG1); 1563 __ std(tmp1, 0, R4_ARG2); 1564 __ std(tmp2, 8, R4_ARG2); 1565 __ std(tmp3, 16, R4_ARG2); 1566 __ std(tmp4, 24, R4_ARG2); 1567 __ addi(R3_ARG1, R3_ARG1, 32); 1568 __ addi(R4_ARG2, R4_ARG2, 32); 1569 __ bdnz(l_6); 1570 } 1571 1572 // copy 1 element at a time 1573 __ bind(l_2); 1574 __ cmpwi(CCR0, R5_ARG3, 0); 1575 __ beq(CCR0, l_1); 1576 1577 { // FasterArrayCopy 1578 __ mtctr(R5_ARG3); 1579 __ addi(R3_ARG1, R3_ARG1, -4); 1580 __ addi(R4_ARG2, R4_ARG2, -4); 1581 1582 __ bind(l_3); 1583 __ lwzu(tmp2, 4, R3_ARG1); 1584 __ stwu(tmp2, 4, R4_ARG2); 1585 __ bdnz(l_3); 1586 } 1587 1588 __ bind(l_1); 1589 return; 1590 } 1591 1592 // Generate stub for disjoint int copy. If "aligned" is true, the 1593 // "from" and "to" addresses are assumed to be heapword aligned. 1594 // 1595 // Arguments for generated stub: 1596 // from: R3_ARG1 1597 // to: R4_ARG2 1598 // count: R5_ARG3 treated as signed 1599 // 1600 address generate_disjoint_int_copy(bool aligned, const char * name) { 1601 StubCodeMark mark(this, "StubRoutines", name); 1602 address start = __ function_entry(); 1603 assert_positive_int(R5_ARG3); 1604 generate_disjoint_int_copy_core(aligned); 1605 __ li(R3_RET, 0); // return 0 1606 __ blr(); 1607 return start; 1608 } 1609 1610 // Generate core code for conjoint int copy (and oop copy on 1611 // 32-bit). If "aligned" is true, the "from" and "to" addresses 1612 // are assumed to be heapword aligned. 1613 // 1614 // Arguments: 1615 // from: R3_ARG1 1616 // to: R4_ARG2 1617 // count: R5_ARG3 treated as signed 1618 // 1619 void generate_conjoint_int_copy_core(bool aligned) { 1620 // Do reverse copy. We assume the case of actual overlap is rare enough 1621 // that we don't have to optimize it. 1622 1623 Label l_1, l_2, l_3, l_4, l_5, l_6; 1624 1625 Register tmp1 = R6_ARG4; 1626 Register tmp2 = R7_ARG5; 1627 Register tmp3 = R8_ARG6; 1628 Register tmp4 = R0; 1629 1630 { // FasterArrayCopy 1631 __ cmpwi(CCR0, R5_ARG3, 0); 1632 __ beq(CCR0, l_6); 1633 1634 __ sldi(R5_ARG3, R5_ARG3, 2); 1635 __ add(R3_ARG1, R3_ARG1, R5_ARG3); 1636 __ add(R4_ARG2, R4_ARG2, R5_ARG3); 1637 __ srdi(R5_ARG3, R5_ARG3, 2); 1638 1639 __ cmpwi(CCR0, R5_ARG3, 7); 1640 __ ble(CCR0, l_5); // copy 1 at a time if less than 8 elements remain 1641 1642 __ srdi(tmp1, R5_ARG3, 3); 1643 __ andi(R5_ARG3, R5_ARG3, 7); 1644 __ mtctr(tmp1); 1645 1646 __ bind(l_4); 1647 // Use unrolled version for mass copying (copy 4 elements a time). 1648 // Load feeding store gets zero latency on Power6, however not on Power5. 1649 // Therefore, the following sequence is made for the good of both. 1650 __ addi(R3_ARG1, R3_ARG1, -32); 1651 __ addi(R4_ARG2, R4_ARG2, -32); 1652 __ ld(tmp4, 24, R3_ARG1); 1653 __ ld(tmp3, 16, R3_ARG1); 1654 __ ld(tmp2, 8, R3_ARG1); 1655 __ ld(tmp1, 0, R3_ARG1); 1656 __ std(tmp4, 24, R4_ARG2); 1657 __ std(tmp3, 16, R4_ARG2); 1658 __ std(tmp2, 8, R4_ARG2); 1659 __ std(tmp1, 0, R4_ARG2); 1660 __ bdnz(l_4); 1661 1662 __ cmpwi(CCR0, R5_ARG3, 0); 1663 __ beq(CCR0, l_6); 1664 1665 __ bind(l_5); 1666 __ mtctr(R5_ARG3); 1667 __ bind(l_3); 1668 __ lwz(R0, -4, R3_ARG1); 1669 __ stw(R0, -4, R4_ARG2); 1670 __ addi(R3_ARG1, R3_ARG1, -4); 1671 __ addi(R4_ARG2, R4_ARG2, -4); 1672 __ bdnz(l_3); 1673 1674 __ bind(l_6); 1675 } 1676 } 1677 1678 // Generate stub for conjoint int copy. If "aligned" is true, the 1679 // "from" and "to" addresses are assumed to be heapword aligned. 1680 // 1681 // Arguments for generated stub: 1682 // from: R3_ARG1 1683 // to: R4_ARG2 1684 // count: R5_ARG3 treated as signed 1685 // 1686 address generate_conjoint_int_copy(bool aligned, const char * name) { 1687 StubCodeMark mark(this, "StubRoutines", name); 1688 address start = __ function_entry(); 1689 assert_positive_int(R5_ARG3); 1690 address nooverlap_target = aligned ? 1691 STUB_ENTRY(arrayof_jint_disjoint_arraycopy) : 1692 STUB_ENTRY(jint_disjoint_arraycopy); 1693 1694 array_overlap_test(nooverlap_target, 2); 1695 1696 generate_conjoint_int_copy_core(aligned); 1697 1698 __ li(R3_RET, 0); // return 0 1699 __ blr(); 1700 1701 return start; 1702 } 1703 1704 // Generate core code for disjoint long copy (and oop copy on 1705 // 64-bit). If "aligned" is true, the "from" and "to" addresses 1706 // are assumed to be heapword aligned. 1707 // 1708 // Arguments: 1709 // from: R3_ARG1 1710 // to: R4_ARG2 1711 // count: R5_ARG3 treated as signed 1712 // 1713 void generate_disjoint_long_copy_core(bool aligned) { 1714 Register tmp1 = R6_ARG4; 1715 Register tmp2 = R7_ARG5; 1716 Register tmp3 = R8_ARG6; 1717 Register tmp4 = R0; 1718 1719 Label l_1, l_2, l_3, l_4; 1720 1721 { // FasterArrayCopy 1722 __ cmpwi(CCR0, R5_ARG3, 3); 1723 __ ble(CCR0, l_3); // copy 1 at a time if less than 4 elements remain 1724 1725 __ srdi(tmp1, R5_ARG3, 2); 1726 __ andi_(R5_ARG3, R5_ARG3, 3); 1727 __ mtctr(tmp1); 1728 1729 __ bind(l_4); 1730 // Use unrolled version for mass copying (copy 4 elements a time). 1731 // Load feeding store gets zero latency on Power6, however not on Power5. 1732 // Therefore, the following sequence is made for the good of both. 1733 __ ld(tmp1, 0, R3_ARG1); 1734 __ ld(tmp2, 8, R3_ARG1); 1735 __ ld(tmp3, 16, R3_ARG1); 1736 __ ld(tmp4, 24, R3_ARG1); 1737 __ std(tmp1, 0, R4_ARG2); 1738 __ std(tmp2, 8, R4_ARG2); 1739 __ std(tmp3, 16, R4_ARG2); 1740 __ std(tmp4, 24, R4_ARG2); 1741 __ addi(R3_ARG1, R3_ARG1, 32); 1742 __ addi(R4_ARG2, R4_ARG2, 32); 1743 __ bdnz(l_4); 1744 } 1745 1746 // copy 1 element at a time 1747 __ bind(l_3); 1748 __ cmpwi(CCR0, R5_ARG3, 0); 1749 __ beq(CCR0, l_1); 1750 1751 { // FasterArrayCopy 1752 __ mtctr(R5_ARG3); 1753 __ addi(R3_ARG1, R3_ARG1, -8); 1754 __ addi(R4_ARG2, R4_ARG2, -8); 1755 1756 __ bind(l_2); 1757 __ ldu(R0, 8, R3_ARG1); 1758 __ stdu(R0, 8, R4_ARG2); 1759 __ bdnz(l_2); 1760 1761 } 1762 __ bind(l_1); 1763 } 1764 1765 // Generate stub for disjoint long copy. If "aligned" is true, the 1766 // "from" and "to" addresses are assumed to be heapword aligned. 1767 // 1768 // Arguments for generated stub: 1769 // from: R3_ARG1 1770 // to: R4_ARG2 1771 // count: R5_ARG3 treated as signed 1772 // 1773 address generate_disjoint_long_copy(bool aligned, const char * name) { 1774 StubCodeMark mark(this, "StubRoutines", name); 1775 address start = __ function_entry(); 1776 assert_positive_int(R5_ARG3); 1777 generate_disjoint_long_copy_core(aligned); 1778 __ li(R3_RET, 0); // return 0 1779 __ blr(); 1780 1781 return start; 1782 } 1783 1784 // Generate core code for conjoint long copy (and oop copy on 1785 // 64-bit). If "aligned" is true, the "from" and "to" addresses 1786 // are assumed to be heapword aligned. 1787 // 1788 // Arguments: 1789 // from: R3_ARG1 1790 // to: R4_ARG2 1791 // count: R5_ARG3 treated as signed 1792 // 1793 void generate_conjoint_long_copy_core(bool aligned) { 1794 Register tmp1 = R6_ARG4; 1795 Register tmp2 = R7_ARG5; 1796 Register tmp3 = R8_ARG6; 1797 Register tmp4 = R0; 1798 1799 Label l_1, l_2, l_3, l_4, l_5; 1800 1801 __ cmpwi(CCR0, R5_ARG3, 0); 1802 __ beq(CCR0, l_1); 1803 1804 { // FasterArrayCopy 1805 __ sldi(R5_ARG3, R5_ARG3, 3); 1806 __ add(R3_ARG1, R3_ARG1, R5_ARG3); 1807 __ add(R4_ARG2, R4_ARG2, R5_ARG3); 1808 __ srdi(R5_ARG3, R5_ARG3, 3); 1809 1810 __ cmpwi(CCR0, R5_ARG3, 3); 1811 __ ble(CCR0, l_5); // copy 1 at a time if less than 4 elements remain 1812 1813 __ srdi(tmp1, R5_ARG3, 2); 1814 __ andi(R5_ARG3, R5_ARG3, 3); 1815 __ mtctr(tmp1); 1816 1817 __ bind(l_4); 1818 // Use unrolled version for mass copying (copy 4 elements a time). 1819 // Load feeding store gets zero latency on Power6, however not on Power5. 1820 // Therefore, the following sequence is made for the good of both. 1821 __ addi(R3_ARG1, R3_ARG1, -32); 1822 __ addi(R4_ARG2, R4_ARG2, -32); 1823 __ ld(tmp4, 24, R3_ARG1); 1824 __ ld(tmp3, 16, R3_ARG1); 1825 __ ld(tmp2, 8, R3_ARG1); 1826 __ ld(tmp1, 0, R3_ARG1); 1827 __ std(tmp4, 24, R4_ARG2); 1828 __ std(tmp3, 16, R4_ARG2); 1829 __ std(tmp2, 8, R4_ARG2); 1830 __ std(tmp1, 0, R4_ARG2); 1831 __ bdnz(l_4); 1832 1833 __ cmpwi(CCR0, R5_ARG3, 0); 1834 __ beq(CCR0, l_1); 1835 1836 __ bind(l_5); 1837 __ mtctr(R5_ARG3); 1838 __ bind(l_3); 1839 __ ld(R0, -8, R3_ARG1); 1840 __ std(R0, -8, R4_ARG2); 1841 __ addi(R3_ARG1, R3_ARG1, -8); 1842 __ addi(R4_ARG2, R4_ARG2, -8); 1843 __ bdnz(l_3); 1844 1845 } 1846 __ bind(l_1); 1847 } 1848 1849 // Generate stub for conjoint long copy. If "aligned" is true, the 1850 // "from" and "to" addresses are assumed to be heapword aligned. 1851 // 1852 // Arguments for generated stub: 1853 // from: R3_ARG1 1854 // to: R4_ARG2 1855 // count: R5_ARG3 treated as signed 1856 // 1857 address generate_conjoint_long_copy(bool aligned, const char * name) { 1858 StubCodeMark mark(this, "StubRoutines", name); 1859 address start = __ function_entry(); 1860 assert_positive_int(R5_ARG3); 1861 address nooverlap_target = aligned ? 1862 STUB_ENTRY(arrayof_jlong_disjoint_arraycopy) : 1863 STUB_ENTRY(jlong_disjoint_arraycopy); 1864 1865 array_overlap_test(nooverlap_target, 3); 1866 generate_conjoint_long_copy_core(aligned); 1867 1868 __ li(R3_RET, 0); // return 0 1869 __ blr(); 1870 1871 return start; 1872 } 1873 1874 // Generate stub for conjoint oop copy. If "aligned" is true, the 1875 // "from" and "to" addresses are assumed to be heapword aligned. 1876 // 1877 // Arguments for generated stub: 1878 // from: R3_ARG1 1879 // to: R4_ARG2 1880 // count: R5_ARG3 treated as signed 1881 // dest_uninitialized: G1 support 1882 // 1883 address generate_conjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) { 1884 StubCodeMark mark(this, "StubRoutines", name); 1885 1886 address start = __ function_entry(); 1887 assert_positive_int(R5_ARG3); 1888 address nooverlap_target = aligned ? 1889 STUB_ENTRY(arrayof_oop_disjoint_arraycopy) : 1890 STUB_ENTRY(oop_disjoint_arraycopy); 1891 1892 gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7); 1893 1894 // Save arguments. 1895 __ mr(R9_ARG7, R4_ARG2); 1896 __ mr(R10_ARG8, R5_ARG3); 1897 1898 if (UseCompressedOops) { 1899 array_overlap_test(nooverlap_target, 2); 1900 generate_conjoint_int_copy_core(aligned); 1901 } else { 1902 array_overlap_test(nooverlap_target, 3); 1903 generate_conjoint_long_copy_core(aligned); 1904 } 1905 1906 gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1); 1907 __ li(R3_RET, 0); // return 0 1908 __ blr(); 1909 return start; 1910 } 1911 1912 // Generate stub for disjoint oop copy. If "aligned" is true, the 1913 // "from" and "to" addresses are assumed to be heapword aligned. 1914 // 1915 // Arguments for generated stub: 1916 // from: R3_ARG1 1917 // to: R4_ARG2 1918 // count: R5_ARG3 treated as signed 1919 // dest_uninitialized: G1 support 1920 // 1921 address generate_disjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) { 1922 StubCodeMark mark(this, "StubRoutines", name); 1923 address start = __ function_entry(); 1924 assert_positive_int(R5_ARG3); 1925 gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7); 1926 1927 // save some arguments, disjoint_long_copy_core destroys them. 1928 // needed for post barrier 1929 __ mr(R9_ARG7, R4_ARG2); 1930 __ mr(R10_ARG8, R5_ARG3); 1931 1932 if (UseCompressedOops) { 1933 generate_disjoint_int_copy_core(aligned); 1934 } else { 1935 generate_disjoint_long_copy_core(aligned); 1936 } 1937 1938 gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1); 1939 __ li(R3_RET, 0); // return 0 1940 __ blr(); 1941 1942 return start; 1943 } 1944 1945 1946 // Helper for generating a dynamic type check. 1947 // Smashes only the given temp registers. 1948 void generate_type_check(Register sub_klass, 1949 Register super_check_offset, 1950 Register super_klass, 1951 Register temp, 1952 Label& L_success) { 1953 assert_different_registers(sub_klass, super_check_offset, super_klass); 1954 1955 BLOCK_COMMENT("type_check:"); 1956 1957 Label L_miss; 1958 1959 __ check_klass_subtype_fast_path(sub_klass, super_klass, temp, R0, &L_success, &L_miss, NULL, 1960 super_check_offset); 1961 __ check_klass_subtype_slow_path(sub_klass, super_klass, temp, R0, &L_success, NULL); 1962 1963 // Fall through on failure! 1964 __ bind(L_miss); 1965 } 1966 1967 1968 // Generate stub for checked oop copy. 1969 // 1970 // Arguments for generated stub: 1971 // from: R3 1972 // to: R4 1973 // count: R5 treated as signed 1974 // ckoff: R6 (super_check_offset) 1975 // ckval: R7 (super_klass) 1976 // ret: R3 zero for success; (-1^K) where K is partial transfer count 1977 // 1978 address generate_checkcast_copy(const char *name, bool dest_uninitialized) { 1979 1980 const Register R3_from = R3_ARG1; // source array address 1981 const Register R4_to = R4_ARG2; // destination array address 1982 const Register R5_count = R5_ARG3; // elements count 1983 const Register R6_ckoff = R6_ARG4; // super_check_offset 1984 const Register R7_ckval = R7_ARG5; // super_klass 1985 1986 const Register R8_offset = R8_ARG6; // loop var, with stride wordSize 1987 const Register R9_remain = R9_ARG7; // loop var, with stride -1 1988 const Register R10_oop = R10_ARG8; // actual oop copied 1989 const Register R11_klass = R11_scratch1; // oop._klass 1990 const Register R12_tmp = R12_scratch2; 1991 1992 const Register R2_minus1 = R2; 1993 1994 //__ align(CodeEntryAlignment); 1995 StubCodeMark mark(this, "StubRoutines", name); 1996 address start = __ function_entry(); 1997 1998 // Assert that int is 64 bit sign extended and arrays are not conjoint. 1999#ifdef ASSERT 2000 { 2001 assert_positive_int(R5_ARG3); 2002 const Register tmp1 = R11_scratch1, tmp2 = R12_scratch2; 2003 Label no_overlap; 2004 __ subf(tmp1, R3_ARG1, R4_ARG2); // distance in bytes 2005 __ sldi(tmp2, R5_ARG3, LogBytesPerHeapOop); // size in bytes 2006 __ cmpld(CCR0, R3_ARG1, R4_ARG2); // Use unsigned comparison! 2007 __ cmpld(CCR1, tmp1, tmp2); 2008 __ crnand(CCR0, Assembler::less, CCR1, Assembler::less); 2009 // Overlaps if Src before dst and distance smaller than size. 2010 // Branch to forward copy routine otherwise. 2011 __ blt(CCR0, no_overlap); 2012 __ stop("overlap in checkcast_copy", 0x9543); 2013 __ bind(no_overlap); 2014 } 2015#endif 2016 2017 gen_write_ref_array_pre_barrier(R3_from, R4_to, R5_count, dest_uninitialized, R12_tmp, /* preserve: */ R6_ckoff, R7_ckval); 2018 2019 //inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr, R12_tmp, R3_RET); 2020 2021 Label load_element, store_element, store_null, success, do_card_marks; 2022 __ or_(R9_remain, R5_count, R5_count); // Initialize loop index, and test it. 2023 __ li(R8_offset, 0); // Offset from start of arrays. 2024 __ li(R2_minus1, -1); 2025 __ bne(CCR0, load_element); 2026 2027 // Empty array: Nothing to do. 2028 __ li(R3_RET, 0); // Return 0 on (trivial) success. 2029 __ blr(); 2030 2031 // ======== begin loop ======== 2032 // (Entry is load_element.) 2033 __ align(OptoLoopAlignment); 2034 __ bind(store_element); 2035 if (UseCompressedOops) { 2036 __ encode_heap_oop_not_null(R10_oop); 2037 __ bind(store_null); 2038 __ stw(R10_oop, R8_offset, R4_to); 2039 } else { 2040 __ bind(store_null); 2041 __ std(R10_oop, R8_offset, R4_to); 2042 } 2043 2044 __ addi(R8_offset, R8_offset, heapOopSize); // Step to next offset. 2045 __ add_(R9_remain, R2_minus1, R9_remain); // Decrement the count. 2046 __ beq(CCR0, success); 2047 2048 // ======== loop entry is here ======== 2049 __ bind(load_element); 2050 __ load_heap_oop(R10_oop, R8_offset, R3_from, &store_null); // Load the oop. 2051 2052 __ load_klass(R11_klass, R10_oop); // Query the object klass. 2053 2054 generate_type_check(R11_klass, R6_ckoff, R7_ckval, R12_tmp, 2055 // Branch to this on success: 2056 store_element); 2057 // ======== end loop ======== 2058 2059 // It was a real error; we must depend on the caller to finish the job. 2060 // Register R9_remain has number of *remaining* oops, R5_count number of *total* oops. 2061 // Emit GC store barriers for the oops we have copied (R5_count minus R9_remain), 2062 // and report their number to the caller. 2063 __ subf_(R5_count, R9_remain, R5_count); 2064 __ nand(R3_RET, R5_count, R5_count); // report (-1^K) to caller 2065 __ bne(CCR0, do_card_marks); 2066 __ blr(); 2067 2068 __ bind(success); 2069 __ li(R3_RET, 0); 2070 2071 __ bind(do_card_marks); 2072 // Store check on R4_to[0..R5_count-1]. 2073 gen_write_ref_array_post_barrier(R4_to, R5_count, R12_tmp, /* preserve: */ R3_RET); 2074 __ blr(); 2075 return start; 2076 } 2077 2078 2079 // Generate 'unsafe' array copy stub. 2080 // Though just as safe as the other stubs, it takes an unscaled 2081 // size_t argument instead of an element count. 2082 // 2083 // Arguments for generated stub: 2084 // from: R3 2085 // to: R4 2086 // count: R5 byte count, treated as ssize_t, can be zero 2087 // 2088 // Examines the alignment of the operands and dispatches 2089 // to a long, int, short, or byte copy loop. 2090 // 2091 address generate_unsafe_copy(const char* name, 2092 address byte_copy_entry, 2093 address short_copy_entry, 2094 address int_copy_entry, 2095 address long_copy_entry) { 2096 2097 const Register R3_from = R3_ARG1; // source array address 2098 const Register R4_to = R4_ARG2; // destination array address 2099 const Register R5_count = R5_ARG3; // elements count (as long on PPC64) 2100 2101 const Register R6_bits = R6_ARG4; // test copy of low bits 2102 const Register R7_tmp = R7_ARG5; 2103 2104 //__ align(CodeEntryAlignment); 2105 StubCodeMark mark(this, "StubRoutines", name); 2106 address start = __ function_entry(); 2107 2108 // Bump this on entry, not on exit: 2109 //inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr, R6_bits, R7_tmp); 2110 2111 Label short_copy, int_copy, long_copy; 2112 2113 __ orr(R6_bits, R3_from, R4_to); 2114 __ orr(R6_bits, R6_bits, R5_count); 2115 __ andi_(R0, R6_bits, (BytesPerLong-1)); 2116 __ beq(CCR0, long_copy); 2117 2118 __ andi_(R0, R6_bits, (BytesPerInt-1)); 2119 __ beq(CCR0, int_copy); 2120 2121 __ andi_(R0, R6_bits, (BytesPerShort-1)); 2122 __ beq(CCR0, short_copy); 2123 2124 // byte_copy: 2125 __ b(byte_copy_entry); 2126 2127 __ bind(short_copy); 2128 __ srwi(R5_count, R5_count, LogBytesPerShort); 2129 __ b(short_copy_entry); 2130 2131 __ bind(int_copy); 2132 __ srwi(R5_count, R5_count, LogBytesPerInt); 2133 __ b(int_copy_entry); 2134 2135 __ bind(long_copy); 2136 __ srwi(R5_count, R5_count, LogBytesPerLong); 2137 __ b(long_copy_entry); 2138 2139 return start; 2140 } 2141 2142 2143 // Perform range checks on the proposed arraycopy. 2144 // Kills the two temps, but nothing else. 2145 // Also, clean the sign bits of src_pos and dst_pos. 2146 void arraycopy_range_checks(Register src, // source array oop 2147 Register src_pos, // source position 2148 Register dst, // destination array oop 2149 Register dst_pos, // destination position 2150 Register length, // length of copy 2151 Register temp1, Register temp2, 2152 Label& L_failed) { 2153 BLOCK_COMMENT("arraycopy_range_checks:"); 2154 2155 const Register array_length = temp1; // scratch 2156 const Register end_pos = temp2; // scratch 2157 2158 // if (src_pos + length > arrayOop(src)->length() ) FAIL; 2159 __ lwa(array_length, arrayOopDesc::length_offset_in_bytes(), src); 2160 __ add(end_pos, src_pos, length); // src_pos + length 2161 __ cmpd(CCR0, end_pos, array_length); 2162 __ bgt(CCR0, L_failed); 2163 2164 // if (dst_pos + length > arrayOop(dst)->length() ) FAIL; 2165 __ lwa(array_length, arrayOopDesc::length_offset_in_bytes(), dst); 2166 __ add(end_pos, dst_pos, length); // src_pos + length 2167 __ cmpd(CCR0, end_pos, array_length); 2168 __ bgt(CCR0, L_failed); 2169 2170 BLOCK_COMMENT("arraycopy_range_checks done"); 2171 } 2172 2173 2174 // 2175 // Generate generic array copy stubs 2176 // 2177 // Input: 2178 // R3 - src oop 2179 // R4 - src_pos 2180 // R5 - dst oop 2181 // R6 - dst_pos 2182 // R7 - element count 2183 // 2184 // Output: 2185 // R3 == 0 - success 2186 // R3 == -1 - need to call System.arraycopy 2187 // 2188 address generate_generic_copy(const char *name, 2189 address entry_jbyte_arraycopy, 2190 address entry_jshort_arraycopy, 2191 address entry_jint_arraycopy, 2192 address entry_oop_arraycopy, 2193 address entry_disjoint_oop_arraycopy, 2194 address entry_jlong_arraycopy, 2195 address entry_checkcast_arraycopy) { 2196 Label L_failed, L_objArray; 2197 2198 // Input registers 2199 const Register src = R3_ARG1; // source array oop 2200 const Register src_pos = R4_ARG2; // source position 2201 const Register dst = R5_ARG3; // destination array oop 2202 const Register dst_pos = R6_ARG4; // destination position 2203 const Register length = R7_ARG5; // elements count 2204 2205 // registers used as temp 2206 const Register src_klass = R8_ARG6; // source array klass 2207 const Register dst_klass = R9_ARG7; // destination array klass 2208 const Register lh = R10_ARG8; // layout handler 2209 const Register temp = R2; 2210 2211 //__ align(CodeEntryAlignment); 2212 StubCodeMark mark(this, "StubRoutines", name); 2213 address start = __ function_entry(); 2214 2215 // Bump this on entry, not on exit: 2216 //inc_counter_np(SharedRuntime::_generic_array_copy_ctr, lh, temp); 2217 2218 // In principle, the int arguments could be dirty. 2219 2220 //----------------------------------------------------------------------- 2221 // Assembler stubs will be used for this call to arraycopy 2222 // if the following conditions are met: 2223 // 2224 // (1) src and dst must not be null. 2225 // (2) src_pos must not be negative. 2226 // (3) dst_pos must not be negative. 2227 // (4) length must not be negative. 2228 // (5) src klass and dst klass should be the same and not NULL. 2229 // (6) src and dst should be arrays. 2230 // (7) src_pos + length must not exceed length of src. 2231 // (8) dst_pos + length must not exceed length of dst. 2232 BLOCK_COMMENT("arraycopy initial argument checks"); 2233 2234 __ cmpdi(CCR1, src, 0); // if (src == NULL) return -1; 2235 __ extsw_(src_pos, src_pos); // if (src_pos < 0) return -1; 2236 __ cmpdi(CCR5, dst, 0); // if (dst == NULL) return -1; 2237 __ cror(CCR1, Assembler::equal, CCR0, Assembler::less); 2238 __ extsw_(dst_pos, dst_pos); // if (src_pos < 0) return -1; 2239 __ cror(CCR5, Assembler::equal, CCR0, Assembler::less); 2240 __ extsw_(length, length); // if (length < 0) return -1; 2241 __ cror(CCR1, Assembler::equal, CCR5, Assembler::equal); 2242 __ cror(CCR1, Assembler::equal, CCR0, Assembler::less); 2243 __ beq(CCR1, L_failed); 2244 2245 BLOCK_COMMENT("arraycopy argument klass checks"); 2246 __ load_klass(src_klass, src); 2247 __ load_klass(dst_klass, dst); 2248 2249 // Load layout helper 2250 // 2251 // |array_tag| | header_size | element_type | |log2_element_size| 2252 // 32 30 24 16 8 2 0 2253 // 2254 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0 2255 // 2256 2257 int lh_offset = in_bytes(Klass::layout_helper_offset()); 2258 2259 // Load 32-bits signed value. Use br() instruction with it to check icc. 2260 __ lwz(lh, lh_offset, src_klass); 2261 2262 // Handle objArrays completely differently... 2263 jint objArray_lh = Klass::array_layout_helper(T_OBJECT); 2264 __ load_const_optimized(temp, objArray_lh, R0); 2265 __ cmpw(CCR0, lh, temp); 2266 __ beq(CCR0, L_objArray); 2267 2268 __ cmpd(CCR5, src_klass, dst_klass); // if (src->klass() != dst->klass()) return -1; 2269 __ cmpwi(CCR6, lh, Klass::_lh_neutral_value); // if (!src->is_Array()) return -1; 2270 2271 __ crnand(CCR5, Assembler::equal, CCR6, Assembler::less); 2272 __ beq(CCR5, L_failed); 2273 2274 // At this point, it is known to be a typeArray (array_tag 0x3). 2275#ifdef ASSERT 2276 { Label L; 2277 jint lh_prim_tag_in_place = (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift); 2278 __ load_const_optimized(temp, lh_prim_tag_in_place, R0); 2279 __ cmpw(CCR0, lh, temp); 2280 __ bge(CCR0, L); 2281 __ stop("must be a primitive array"); 2282 __ bind(L); 2283 } 2284#endif 2285 2286 arraycopy_range_checks(src, src_pos, dst, dst_pos, length, 2287 temp, dst_klass, L_failed); 2288 2289 // TypeArrayKlass 2290 // 2291 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize); 2292 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize); 2293 // 2294 2295 const Register offset = dst_klass; // array offset 2296 const Register elsize = src_klass; // log2 element size 2297 2298 __ rldicl(offset, lh, 64 - Klass::_lh_header_size_shift, 64 - exact_log2(Klass::_lh_header_size_mask + 1)); 2299 __ andi(elsize, lh, Klass::_lh_log2_element_size_mask); 2300 __ add(src, offset, src); // src array offset 2301 __ add(dst, offset, dst); // dst array offset 2302 2303 // Next registers should be set before the jump to corresponding stub. 2304 const Register from = R3_ARG1; // source array address 2305 const Register to = R4_ARG2; // destination array address 2306 const Register count = R5_ARG3; // elements count 2307 2308 // 'from', 'to', 'count' registers should be set in this order 2309 // since they are the same as 'src', 'src_pos', 'dst'. 2310 2311 BLOCK_COMMENT("scale indexes to element size"); 2312 __ sld(src_pos, src_pos, elsize); 2313 __ sld(dst_pos, dst_pos, elsize); 2314 __ add(from, src_pos, src); // src_addr 2315 __ add(to, dst_pos, dst); // dst_addr 2316 __ mr(count, length); // length 2317 2318 BLOCK_COMMENT("choose copy loop based on element size"); 2319 // Using conditional branches with range 32kB. 2320 const int bo = Assembler::bcondCRbiIs1, bi = Assembler::bi0(CCR0, Assembler::equal); 2321 __ cmpwi(CCR0, elsize, 0); 2322 __ bc(bo, bi, entry_jbyte_arraycopy); 2323 __ cmpwi(CCR0, elsize, LogBytesPerShort); 2324 __ bc(bo, bi, entry_jshort_arraycopy); 2325 __ cmpwi(CCR0, elsize, LogBytesPerInt); 2326 __ bc(bo, bi, entry_jint_arraycopy); 2327#ifdef ASSERT 2328 { Label L; 2329 __ cmpwi(CCR0, elsize, LogBytesPerLong); 2330 __ beq(CCR0, L); 2331 __ stop("must be long copy, but elsize is wrong"); 2332 __ bind(L); 2333 } 2334#endif 2335 __ b(entry_jlong_arraycopy); 2336 2337 // ObjArrayKlass 2338 __ bind(L_objArray); 2339 // live at this point: src_klass, dst_klass, src[_pos], dst[_pos], length 2340 2341 Label L_disjoint_plain_copy, L_checkcast_copy; 2342 // test array classes for subtyping 2343 __ cmpd(CCR0, src_klass, dst_klass); // usual case is exact equality 2344 __ bne(CCR0, L_checkcast_copy); 2345 2346 // Identically typed arrays can be copied without element-wise checks. 2347 arraycopy_range_checks(src, src_pos, dst, dst_pos, length, 2348 temp, lh, L_failed); 2349 2350 __ addi(src, src, arrayOopDesc::base_offset_in_bytes(T_OBJECT)); //src offset 2351 __ addi(dst, dst, arrayOopDesc::base_offset_in_bytes(T_OBJECT)); //dst offset 2352 __ sldi(src_pos, src_pos, LogBytesPerHeapOop); 2353 __ sldi(dst_pos, dst_pos, LogBytesPerHeapOop); 2354 __ add(from, src_pos, src); // src_addr 2355 __ add(to, dst_pos, dst); // dst_addr 2356 __ mr(count, length); // length 2357 __ b(entry_oop_arraycopy); 2358 2359 __ bind(L_checkcast_copy); 2360 // live at this point: src_klass, dst_klass 2361 { 2362 // Before looking at dst.length, make sure dst is also an objArray. 2363 __ lwz(temp, lh_offset, dst_klass); 2364 __ cmpw(CCR0, lh, temp); 2365 __ bne(CCR0, L_failed); 2366 2367 // It is safe to examine both src.length and dst.length. 2368 arraycopy_range_checks(src, src_pos, dst, dst_pos, length, 2369 temp, lh, L_failed); 2370 2371 // Marshal the base address arguments now, freeing registers. 2372 __ addi(src, src, arrayOopDesc::base_offset_in_bytes(T_OBJECT)); //src offset 2373 __ addi(dst, dst, arrayOopDesc::base_offset_in_bytes(T_OBJECT)); //dst offset 2374 __ sldi(src_pos, src_pos, LogBytesPerHeapOop); 2375 __ sldi(dst_pos, dst_pos, LogBytesPerHeapOop); 2376 __ add(from, src_pos, src); // src_addr 2377 __ add(to, dst_pos, dst); // dst_addr 2378 __ mr(count, length); // length 2379 2380 Register sco_temp = R6_ARG4; // This register is free now. 2381 assert_different_registers(from, to, count, sco_temp, 2382 dst_klass, src_klass); 2383 2384 // Generate the type check. 2385 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 2386 __ lwz(sco_temp, sco_offset, dst_klass); 2387 generate_type_check(src_klass, sco_temp, dst_klass, 2388 temp, L_disjoint_plain_copy); 2389 2390 // Fetch destination element klass from the ObjArrayKlass header. 2391 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 2392 2393 // The checkcast_copy loop needs two extra arguments: 2394 __ ld(R7_ARG5, ek_offset, dst_klass); // dest elem klass 2395 __ lwz(R6_ARG4, sco_offset, R7_ARG5); // sco of elem klass 2396 __ b(entry_checkcast_arraycopy); 2397 } 2398 2399 __ bind(L_disjoint_plain_copy); 2400 __ b(entry_disjoint_oop_arraycopy); 2401 2402 __ bind(L_failed); 2403 __ li(R3_RET, -1); // return -1 2404 __ blr(); 2405 return start; 2406 } 2407 2408 // Arguments for generated stub (little endian only): 2409 // R3_ARG1 - source byte array address 2410 // R4_ARG2 - destination byte array address 2411 // R5_ARG3 - round key array 2412 address generate_aescrypt_encryptBlock() { 2413 assert(UseAES, "need AES instructions and misaligned SSE support"); 2414 StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock"); 2415 2416 address start = __ function_entry(); 2417 2418 Label L_doLast; 2419 2420 Register from = R3_ARG1; // source array address 2421 Register to = R4_ARG2; // destination array address 2422 Register key = R5_ARG3; // round key array 2423 2424 Register keylen = R8; 2425 Register temp = R9; 2426 Register keypos = R10; 2427 Register hex = R11; 2428 Register fifteen = R12; 2429 2430 VectorRegister vRet = VR0; 2431 2432 VectorRegister vKey1 = VR1; 2433 VectorRegister vKey2 = VR2; 2434 VectorRegister vKey3 = VR3; 2435 VectorRegister vKey4 = VR4; 2436 2437 VectorRegister fromPerm = VR5; 2438 VectorRegister keyPerm = VR6; 2439 VectorRegister toPerm = VR7; 2440 VectorRegister fSplt = VR8; 2441 2442 VectorRegister vTmp1 = VR9; 2443 VectorRegister vTmp2 = VR10; 2444 VectorRegister vTmp3 = VR11; 2445 VectorRegister vTmp4 = VR12; 2446 2447 VectorRegister vLow = VR13; 2448 VectorRegister vHigh = VR14; 2449 2450 __ li (hex, 16); 2451 __ li (fifteen, 15); 2452 __ vspltisb (fSplt, 0x0f); 2453 2454 // load unaligned from[0-15] to vsRet 2455 __ lvx (vRet, from); 2456 __ lvx (vTmp1, fifteen, from); 2457 __ lvsl (fromPerm, from); 2458 __ vxor (fromPerm, fromPerm, fSplt); 2459 __ vperm (vRet, vRet, vTmp1, fromPerm); 2460 2461 // load keylen (44 or 52 or 60) 2462 __ lwz (keylen, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT), key); 2463 2464 // to load keys 2465 __ lvsr (keyPerm, key); 2466 __ vxor (vTmp2, vTmp2, vTmp2); 2467 __ vspltisb (vTmp2, -16); 2468 __ vrld (keyPerm, keyPerm, vTmp2); 2469 __ vrld (keyPerm, keyPerm, vTmp2); 2470 __ vsldoi (keyPerm, keyPerm, keyPerm, -8); 2471 2472 // load the 1st round key to vKey1 2473 __ li (keypos, 0); 2474 __ lvx (vKey1, keypos, key); 2475 __ addi (keypos, keypos, 16); 2476 __ lvx (vTmp1, keypos, key); 2477 __ vperm (vKey1, vTmp1, vKey1, keyPerm); 2478 2479 // 1st round 2480 __ vxor (vRet, vRet, vKey1); 2481 2482 // load the 2nd round key to vKey1 2483 __ addi (keypos, keypos, 16); 2484 __ lvx (vTmp2, keypos, key); 2485 __ vperm (vKey1, vTmp2, vTmp1, keyPerm); 2486 2487 // load the 3rd round key to vKey2 2488 __ addi (keypos, keypos, 16); 2489 __ lvx (vTmp1, keypos, key); 2490 __ vperm (vKey2, vTmp1, vTmp2, keyPerm); 2491 2492 // load the 4th round key to vKey3 2493 __ addi (keypos, keypos, 16); 2494 __ lvx (vTmp2, keypos, key); 2495 __ vperm (vKey3, vTmp2, vTmp1, keyPerm); 2496 2497 // load the 5th round key to vKey4 2498 __ addi (keypos, keypos, 16); 2499 __ lvx (vTmp1, keypos, key); 2500 __ vperm (vKey4, vTmp1, vTmp2, keyPerm); 2501 2502 // 2nd - 5th rounds 2503 __ vcipher (vRet, vRet, vKey1); 2504 __ vcipher (vRet, vRet, vKey2); 2505 __ vcipher (vRet, vRet, vKey3); 2506 __ vcipher (vRet, vRet, vKey4); 2507 2508 // load the 6th round key to vKey1 2509 __ addi (keypos, keypos, 16); 2510 __ lvx (vTmp2, keypos, key); 2511 __ vperm (vKey1, vTmp2, vTmp1, keyPerm); 2512 2513 // load the 7th round key to vKey2 2514 __ addi (keypos, keypos, 16); 2515 __ lvx (vTmp1, keypos, key); 2516 __ vperm (vKey2, vTmp1, vTmp2, keyPerm); 2517 2518 // load the 8th round key to vKey3 2519 __ addi (keypos, keypos, 16); 2520 __ lvx (vTmp2, keypos, key); 2521 __ vperm (vKey3, vTmp2, vTmp1, keyPerm); 2522 2523 // load the 9th round key to vKey4 2524 __ addi (keypos, keypos, 16); 2525 __ lvx (vTmp1, keypos, key); 2526 __ vperm (vKey4, vTmp1, vTmp2, keyPerm); 2527 2528 // 6th - 9th rounds 2529 __ vcipher (vRet, vRet, vKey1); 2530 __ vcipher (vRet, vRet, vKey2); 2531 __ vcipher (vRet, vRet, vKey3); 2532 __ vcipher (vRet, vRet, vKey4); 2533 2534 // load the 10th round key to vKey1 2535 __ addi (keypos, keypos, 16); 2536 __ lvx (vTmp2, keypos, key); 2537 __ vperm (vKey1, vTmp2, vTmp1, keyPerm); 2538 2539 // load the 11th round key to vKey2 2540 __ addi (keypos, keypos, 16); 2541 __ lvx (vTmp1, keypos, key); 2542 __ vperm (vKey2, vTmp1, vTmp2, keyPerm); 2543 2544 // if all round keys are loaded, skip next 4 rounds 2545 __ cmpwi (CCR0, keylen, 44); 2546 __ beq (CCR0, L_doLast); 2547 2548 // 10th - 11th rounds 2549 __ vcipher (vRet, vRet, vKey1); 2550 __ vcipher (vRet, vRet, vKey2); 2551 2552 // load the 12th round key to vKey1 2553 __ addi (keypos, keypos, 16); 2554 __ lvx (vTmp2, keypos, key); 2555 __ vperm (vKey1, vTmp2, vTmp1, keyPerm); 2556 2557 // load the 13th round key to vKey2 2558 __ addi (keypos, keypos, 16); 2559 __ lvx (vTmp1, keypos, key); 2560 __ vperm (vKey2, vTmp1, vTmp2, keyPerm); 2561 2562 // if all round keys are loaded, skip next 2 rounds 2563 __ cmpwi (CCR0, keylen, 52); 2564 __ beq (CCR0, L_doLast); 2565 2566 // 12th - 13th rounds 2567 __ vcipher (vRet, vRet, vKey1); 2568 __ vcipher (vRet, vRet, vKey2); 2569 2570 // load the 14th round key to vKey1 2571 __ addi (keypos, keypos, 16); 2572 __ lvx (vTmp2, keypos, key); 2573 __ vperm (vKey1, vTmp2, vTmp1, keyPerm); 2574 2575 // load the 15th round key to vKey2 2576 __ addi (keypos, keypos, 16); 2577 __ lvx (vTmp1, keypos, key); 2578 __ vperm (vKey2, vTmp1, vTmp2, keyPerm); 2579 2580 __ bind(L_doLast); 2581 2582 // last two rounds 2583 __ vcipher (vRet, vRet, vKey1); 2584 __ vcipherlast (vRet, vRet, vKey2); 2585 2586 __ neg (temp, to); 2587 __ lvsr (toPerm, temp); 2588 __ vspltisb (vTmp2, -1); 2589 __ vxor (vTmp1, vTmp1, vTmp1); 2590 __ vperm (vTmp2, vTmp2, vTmp1, toPerm); 2591 __ vxor (toPerm, toPerm, fSplt); 2592 __ lvx (vTmp1, to); 2593 __ vperm (vRet, vRet, vRet, toPerm); 2594 __ vsel (vTmp1, vTmp1, vRet, vTmp2); 2595 __ lvx (vTmp4, fifteen, to); 2596 __ stvx (vTmp1, to); 2597 __ vsel (vRet, vRet, vTmp4, vTmp2); 2598 __ stvx (vRet, fifteen, to); 2599 2600 __ blr(); 2601 return start; 2602 } 2603 2604 // Arguments for generated stub (little endian only): 2605 // R3_ARG1 - source byte array address 2606 // R4_ARG2 - destination byte array address 2607 // R5_ARG3 - K (key) in little endian int array 2608 address generate_aescrypt_decryptBlock() { 2609 assert(UseAES, "need AES instructions and misaligned SSE support"); 2610 StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock"); 2611 2612 address start = __ function_entry(); 2613 2614 Label L_doLast; 2615 Label L_do44; 2616 Label L_do52; 2617 Label L_do60; 2618 2619 Register from = R3_ARG1; // source array address 2620 Register to = R4_ARG2; // destination array address 2621 Register key = R5_ARG3; // round key array 2622 2623 Register keylen = R8; 2624 Register temp = R9; 2625 Register keypos = R10; 2626 Register hex = R11; 2627 Register fifteen = R12; 2628 2629 VectorRegister vRet = VR0; 2630 2631 VectorRegister vKey1 = VR1; 2632 VectorRegister vKey2 = VR2; 2633 VectorRegister vKey3 = VR3; 2634 VectorRegister vKey4 = VR4; 2635 VectorRegister vKey5 = VR5; 2636 2637 VectorRegister fromPerm = VR6; 2638 VectorRegister keyPerm = VR7; 2639 VectorRegister toPerm = VR8; 2640 VectorRegister fSplt = VR9; 2641 2642 VectorRegister vTmp1 = VR10; 2643 VectorRegister vTmp2 = VR11; 2644 VectorRegister vTmp3 = VR12; 2645 VectorRegister vTmp4 = VR13; 2646 2647 VectorRegister vLow = VR14; 2648 VectorRegister vHigh = VR15; 2649 2650 __ li (hex, 16); 2651 __ li (fifteen, 15); 2652 __ vspltisb (fSplt, 0x0f); 2653 2654 // load unaligned from[0-15] to vsRet 2655 __ lvx (vRet, from); 2656 __ lvx (vTmp1, fifteen, from); 2657 __ lvsl (fromPerm, from); 2658 __ vxor (fromPerm, fromPerm, fSplt); 2659 __ vperm (vRet, vRet, vTmp1, fromPerm); // align [and byte swap in LE] 2660 2661 // load keylen (44 or 52 or 60) 2662 __ lwz (keylen, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT), key); 2663 2664 // to load keys 2665 __ lvsr (keyPerm, key); 2666 __ vxor (vTmp2, vTmp2, vTmp2); 2667 __ vspltisb (vTmp2, -16); 2668 __ vrld (keyPerm, keyPerm, vTmp2); 2669 __ vrld (keyPerm, keyPerm, vTmp2); 2670 __ vsldoi (keyPerm, keyPerm, keyPerm, -8); 2671 2672 __ cmpwi (CCR0, keylen, 44); 2673 __ beq (CCR0, L_do44); 2674 2675 __ cmpwi (CCR0, keylen, 52); 2676 __ beq (CCR0, L_do52); 2677 2678 // load the 15th round key to vKey11 2679 __ li (keypos, 240); 2680 __ lvx (vTmp1, keypos, key); 2681 __ addi (keypos, keypos, -16); 2682 __ lvx (vTmp2, keypos, key); 2683 __ vperm (vKey1, vTmp1, vTmp2, keyPerm); 2684 2685 // load the 14th round key to vKey10 2686 __ addi (keypos, keypos, -16); 2687 __ lvx (vTmp1, keypos, key); 2688 __ vperm (vKey2, vTmp2, vTmp1, keyPerm); 2689 2690 // load the 13th round key to vKey10 2691 __ addi (keypos, keypos, -16); 2692 __ lvx (vTmp2, keypos, key); 2693 __ vperm (vKey3, vTmp1, vTmp2, keyPerm); 2694 2695 // load the 12th round key to vKey10 2696 __ addi (keypos, keypos, -16); 2697 __ lvx (vTmp1, keypos, key); 2698 __ vperm (vKey4, vTmp2, vTmp1, keyPerm); 2699 2700 // load the 11th round key to vKey10 2701 __ addi (keypos, keypos, -16); 2702 __ lvx (vTmp2, keypos, key); 2703 __ vperm (vKey5, vTmp1, vTmp2, keyPerm); 2704 2705 // 1st - 5th rounds 2706 __ vxor (vRet, vRet, vKey1); 2707 __ vncipher (vRet, vRet, vKey2); 2708 __ vncipher (vRet, vRet, vKey3); 2709 __ vncipher (vRet, vRet, vKey4); 2710 __ vncipher (vRet, vRet, vKey5); 2711 2712 __ b (L_doLast); 2713 2714 __ bind (L_do52); 2715 2716 // load the 13th round key to vKey11 2717 __ li (keypos, 208); 2718 __ lvx (vTmp1, keypos, key); 2719 __ addi (keypos, keypos, -16); 2720 __ lvx (vTmp2, keypos, key); 2721 __ vperm (vKey1, vTmp1, vTmp2, keyPerm); 2722 2723 // load the 12th round key to vKey10 2724 __ addi (keypos, keypos, -16); 2725 __ lvx (vTmp1, keypos, key); 2726 __ vperm (vKey2, vTmp2, vTmp1, keyPerm); 2727 2728 // load the 11th round key to vKey10 2729 __ addi (keypos, keypos, -16); 2730 __ lvx (vTmp2, keypos, key); 2731 __ vperm (vKey3, vTmp1, vTmp2, keyPerm); 2732 2733 // 1st - 3rd rounds 2734 __ vxor (vRet, vRet, vKey1); 2735 __ vncipher (vRet, vRet, vKey2); 2736 __ vncipher (vRet, vRet, vKey3); 2737 2738 __ b (L_doLast); 2739 2740 __ bind (L_do44); 2741 2742 // load the 11th round key to vKey11 2743 __ li (keypos, 176); 2744 __ lvx (vTmp1, keypos, key); 2745 __ addi (keypos, keypos, -16); 2746 __ lvx (vTmp2, keypos, key); 2747 __ vperm (vKey1, vTmp1, vTmp2, keyPerm); 2748 2749 // 1st round 2750 __ vxor (vRet, vRet, vKey1); 2751 2752 __ bind (L_doLast); 2753 2754 // load the 10th round key to vKey10 2755 __ addi (keypos, keypos, -16); 2756 __ lvx (vTmp1, keypos, key); 2757 __ vperm (vKey1, vTmp2, vTmp1, keyPerm); 2758 2759 // load the 9th round key to vKey10 2760 __ addi (keypos, keypos, -16); 2761 __ lvx (vTmp2, keypos, key); 2762 __ vperm (vKey2, vTmp1, vTmp2, keyPerm); 2763 2764 // load the 8th round key to vKey10 2765 __ addi (keypos, keypos, -16); 2766 __ lvx (vTmp1, keypos, key); 2767 __ vperm (vKey3, vTmp2, vTmp1, keyPerm); 2768 2769 // load the 7th round key to vKey10 2770 __ addi (keypos, keypos, -16); 2771 __ lvx (vTmp2, keypos, key); 2772 __ vperm (vKey4, vTmp1, vTmp2, keyPerm); 2773 2774 // load the 6th round key to vKey10 2775 __ addi (keypos, keypos, -16); 2776 __ lvx (vTmp1, keypos, key); 2777 __ vperm (vKey5, vTmp2, vTmp1, keyPerm); 2778 2779 // last 10th - 6th rounds 2780 __ vncipher (vRet, vRet, vKey1); 2781 __ vncipher (vRet, vRet, vKey2); 2782 __ vncipher (vRet, vRet, vKey3); 2783 __ vncipher (vRet, vRet, vKey4); 2784 __ vncipher (vRet, vRet, vKey5); 2785 2786 // load the 5th round key to vKey10 2787 __ addi (keypos, keypos, -16); 2788 __ lvx (vTmp2, keypos, key); 2789 __ vperm (vKey1, vTmp1, vTmp2, keyPerm); 2790 2791 // load the 4th round key to vKey10 2792 __ addi (keypos, keypos, -16); 2793 __ lvx (vTmp1, keypos, key); 2794 __ vperm (vKey2, vTmp2, vTmp1, keyPerm); 2795 2796 // load the 3rd round key to vKey10 2797 __ addi (keypos, keypos, -16); 2798 __ lvx (vTmp2, keypos, key); 2799 __ vperm (vKey3, vTmp1, vTmp2, keyPerm); 2800 2801 // load the 2nd round key to vKey10 2802 __ addi (keypos, keypos, -16); 2803 __ lvx (vTmp1, keypos, key); 2804 __ vperm (vKey4, vTmp2, vTmp1, keyPerm); 2805 2806 // load the 1st round key to vKey10 2807 __ addi (keypos, keypos, -16); 2808 __ lvx (vTmp2, keypos, key); 2809 __ vperm (vKey5, vTmp1, vTmp2, keyPerm); 2810 2811 // last 5th - 1th rounds 2812 __ vncipher (vRet, vRet, vKey1); 2813 __ vncipher (vRet, vRet, vKey2); 2814 __ vncipher (vRet, vRet, vKey3); 2815 __ vncipher (vRet, vRet, vKey4); 2816 __ vncipherlast (vRet, vRet, vKey5); 2817 2818 __ neg (temp, to); 2819 __ lvsr (toPerm, temp); 2820 __ vspltisb (vTmp2, -1); 2821 __ vxor (vTmp1, vTmp1, vTmp1); 2822 __ vperm (vTmp2, vTmp2, vTmp1, toPerm); 2823 __ vxor (toPerm, toPerm, fSplt); 2824 __ lvx (vTmp1, to); 2825 __ vperm (vRet, vRet, vRet, toPerm); 2826 __ vsel (vTmp1, vTmp1, vRet, vTmp2); 2827 __ lvx (vTmp4, fifteen, to); 2828 __ stvx (vTmp1, to); 2829 __ vsel (vRet, vRet, vTmp4, vTmp2); 2830 __ stvx (vRet, fifteen, to); 2831 2832 __ blr(); 2833 return start; 2834 } 2835 2836 void generate_arraycopy_stubs() { 2837 // Note: the disjoint stubs must be generated first, some of 2838 // the conjoint stubs use them. 2839 2840 // non-aligned disjoint versions 2841 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, "jbyte_disjoint_arraycopy"); 2842 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, "jshort_disjoint_arraycopy"); 2843 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_copy(false, "jint_disjoint_arraycopy"); 2844 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_copy(false, "jlong_disjoint_arraycopy"); 2845 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy", false); 2846 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy_uninit", true); 2847 2848 // aligned disjoint versions 2849 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(true, "arrayof_jbyte_disjoint_arraycopy"); 2850 StubRoutines::_arrayof_jshort_disjoint_arraycopy = generate_disjoint_short_copy(true, "arrayof_jshort_disjoint_arraycopy"); 2851 StubRoutines::_arrayof_jint_disjoint_arraycopy = generate_disjoint_int_copy(true, "arrayof_jint_disjoint_arraycopy"); 2852 StubRoutines::_arrayof_jlong_disjoint_arraycopy = generate_disjoint_long_copy(true, "arrayof_jlong_disjoint_arraycopy"); 2853 StubRoutines::_arrayof_oop_disjoint_arraycopy = generate_disjoint_oop_copy(true, "arrayof_oop_disjoint_arraycopy", false); 2854 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(true, "oop_disjoint_arraycopy_uninit", true); 2855 2856 // non-aligned conjoint versions 2857 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, "jbyte_arraycopy"); 2858 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, "jshort_arraycopy"); 2859 StubRoutines::_jint_arraycopy = generate_conjoint_int_copy(false, "jint_arraycopy"); 2860 StubRoutines::_jlong_arraycopy = generate_conjoint_long_copy(false, "jlong_arraycopy"); 2861 StubRoutines::_oop_arraycopy = generate_conjoint_oop_copy(false, "oop_arraycopy", false); 2862 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_oop_copy(false, "oop_arraycopy_uninit", true); 2863 2864 // aligned conjoint versions 2865 StubRoutines::_arrayof_jbyte_arraycopy = generate_conjoint_byte_copy(true, "arrayof_jbyte_arraycopy"); 2866 StubRoutines::_arrayof_jshort_arraycopy = generate_conjoint_short_copy(true, "arrayof_jshort_arraycopy"); 2867 StubRoutines::_arrayof_jint_arraycopy = generate_conjoint_int_copy(true, "arrayof_jint_arraycopy"); 2868 StubRoutines::_arrayof_jlong_arraycopy = generate_conjoint_long_copy(true, "arrayof_jlong_arraycopy"); 2869 StubRoutines::_arrayof_oop_arraycopy = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", false); 2870 StubRoutines::_arrayof_oop_arraycopy_uninit = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", true); 2871 2872 // special/generic versions 2873 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy", false); 2874 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", true); 2875 2876 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy", 2877 STUB_ENTRY(jbyte_arraycopy), 2878 STUB_ENTRY(jshort_arraycopy), 2879 STUB_ENTRY(jint_arraycopy), 2880 STUB_ENTRY(jlong_arraycopy)); 2881 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy", 2882 STUB_ENTRY(jbyte_arraycopy), 2883 STUB_ENTRY(jshort_arraycopy), 2884 STUB_ENTRY(jint_arraycopy), 2885 STUB_ENTRY(oop_arraycopy), 2886 STUB_ENTRY(oop_disjoint_arraycopy), 2887 STUB_ENTRY(jlong_arraycopy), 2888 STUB_ENTRY(checkcast_arraycopy)); 2889 2890 // fill routines 2891 if (OptimizeFill) { 2892 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill"); 2893 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill"); 2894 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill"); 2895 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill"); 2896 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill"); 2897 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill"); 2898 } 2899 } 2900 2901 // Safefetch stubs. 2902 void generate_safefetch(const char* name, int size, address* entry, address* fault_pc, address* continuation_pc) { 2903 // safefetch signatures: 2904 // int SafeFetch32(int* adr, int errValue); 2905 // intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue); 2906 // 2907 // arguments: 2908 // R3_ARG1 = adr 2909 // R4_ARG2 = errValue 2910 // 2911 // result: 2912 // R3_RET = *adr or errValue 2913 2914 StubCodeMark mark(this, "StubRoutines", name); 2915 2916 // Entry point, pc or function descriptor. 2917 *entry = __ function_entry(); 2918 2919 // Load *adr into R4_ARG2, may fault. 2920 *fault_pc = __ pc(); 2921 switch (size) { 2922 case 4: 2923 // int32_t, signed extended 2924 __ lwa(R4_ARG2, 0, R3_ARG1); 2925 break; 2926 case 8: 2927 // int64_t 2928 __ ld(R4_ARG2, 0, R3_ARG1); 2929 break; 2930 default: 2931 ShouldNotReachHere(); 2932 } 2933 2934 // return errValue or *adr 2935 *continuation_pc = __ pc(); 2936 __ mr(R3_RET, R4_ARG2); 2937 __ blr(); 2938 } 2939 2940 // Stub for BigInteger::multiplyToLen() 2941 // 2942 // Arguments: 2943 // 2944 // Input: 2945 // R3 - x address 2946 // R4 - x length 2947 // R5 - y address 2948 // R6 - y length 2949 // R7 - z address 2950 // R8 - z length 2951 // 2952 address generate_multiplyToLen() { 2953 2954 StubCodeMark mark(this, "StubRoutines", "multiplyToLen"); 2955 2956 address start = __ function_entry(); 2957 2958 const Register x = R3; 2959 const Register xlen = R4; 2960 const Register y = R5; 2961 const Register ylen = R6; 2962 const Register z = R7; 2963 const Register zlen = R8; 2964 2965 const Register tmp1 = R2; // TOC not used. 2966 const Register tmp2 = R9; 2967 const Register tmp3 = R10; 2968 const Register tmp4 = R11; 2969 const Register tmp5 = R12; 2970 2971 // non-volatile regs 2972 const Register tmp6 = R31; 2973 const Register tmp7 = R30; 2974 const Register tmp8 = R29; 2975 const Register tmp9 = R28; 2976 const Register tmp10 = R27; 2977 const Register tmp11 = R26; 2978 const Register tmp12 = R25; 2979 const Register tmp13 = R24; 2980 2981 BLOCK_COMMENT("Entry:"); 2982 2983 // C2 does not respect int to long conversion for stub calls. 2984 __ clrldi(xlen, xlen, 32); 2985 __ clrldi(ylen, ylen, 32); 2986 __ clrldi(zlen, zlen, 32); 2987 2988 // Save non-volatile regs (frameless). 2989 int current_offs = 8; 2990 __ std(R24, -current_offs, R1_SP); current_offs += 8; 2991 __ std(R25, -current_offs, R1_SP); current_offs += 8; 2992 __ std(R26, -current_offs, R1_SP); current_offs += 8; 2993 __ std(R27, -current_offs, R1_SP); current_offs += 8; 2994 __ std(R28, -current_offs, R1_SP); current_offs += 8; 2995 __ std(R29, -current_offs, R1_SP); current_offs += 8; 2996 __ std(R30, -current_offs, R1_SP); current_offs += 8; 2997 __ std(R31, -current_offs, R1_SP); 2998 2999 __ multiply_to_len(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, 3000 tmp6, tmp7, tmp8, tmp9, tmp10, tmp11, tmp12, tmp13); 3001 3002 // Restore non-volatile regs. 3003 current_offs = 8; 3004 __ ld(R24, -current_offs, R1_SP); current_offs += 8; 3005 __ ld(R25, -current_offs, R1_SP); current_offs += 8; 3006 __ ld(R26, -current_offs, R1_SP); current_offs += 8; 3007 __ ld(R27, -current_offs, R1_SP); current_offs += 8; 3008 __ ld(R28, -current_offs, R1_SP); current_offs += 8; 3009 __ ld(R29, -current_offs, R1_SP); current_offs += 8; 3010 __ ld(R30, -current_offs, R1_SP); current_offs += 8; 3011 __ ld(R31, -current_offs, R1_SP); 3012 3013 __ blr(); // Return to caller. 3014 3015 return start; 3016 } 3017 3018 /** 3019 * Arguments: 3020 * 3021 * Inputs: 3022 * R3_ARG1 - int crc 3023 * R4_ARG2 - byte* buf 3024 * R5_ARG3 - int length (of buffer) 3025 * 3026 * scratch: 3027 * R2, R6-R12 3028 * 3029 * Ouput: 3030 * R3_RET - int crc result 3031 */ 3032 // Compute CRC32 function. 3033 address generate_CRC32_updateBytes(const char* name) { 3034 __ align(CodeEntryAlignment); 3035 StubCodeMark mark(this, "StubRoutines", name); 3036 address start = __ function_entry(); // Remember stub start address (is rtn value). 3037 3038 // arguments to kernel_crc32: 3039 const Register crc = R3_ARG1; // Current checksum, preset by caller or result from previous call. 3040 const Register data = R4_ARG2; // source byte array 3041 const Register dataLen = R5_ARG3; // #bytes to process 3042 const Register table = R6_ARG4; // crc table address 3043 3044 const Register t0 = R2; 3045 const Register t1 = R7; 3046 const Register t2 = R8; 3047 const Register t3 = R9; 3048 const Register tc0 = R10; 3049 const Register tc1 = R11; 3050 const Register tc2 = R12; 3051 3052 BLOCK_COMMENT("Stub body {"); 3053 assert_different_registers(crc, data, dataLen, table); 3054 3055 StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table); 3056 3057 __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, table); 3058 3059 BLOCK_COMMENT("return"); 3060 __ mr_if_needed(R3_RET, crc); // Updated crc is function result. No copying required (R3_ARG1 == R3_RET). 3061 __ blr(); 3062 3063 BLOCK_COMMENT("} Stub body"); 3064 return start; 3065 } 3066 3067 // Initialization 3068 void generate_initial() { 3069 // Generates all stubs and initializes the entry points 3070 3071 // Entry points that exist in all platforms. 3072 // Note: This is code that could be shared among different platforms - however the 3073 // benefit seems to be smaller than the disadvantage of having a 3074 // much more complicated generator structure. See also comment in 3075 // stubRoutines.hpp. 3076 3077 StubRoutines::_forward_exception_entry = generate_forward_exception(); 3078 StubRoutines::_call_stub_entry = generate_call_stub(StubRoutines::_call_stub_return_address); 3079 StubRoutines::_catch_exception_entry = generate_catch_exception(); 3080 3081 // Build this early so it's available for the interpreter. 3082 StubRoutines::_throw_StackOverflowError_entry = 3083 generate_throw_exception("StackOverflowError throw_exception", 3084 CAST_FROM_FN_PTR(address, SharedRuntime::throw_StackOverflowError), false); 3085 3086 // CRC32 Intrinsics. 3087 if (UseCRC32Intrinsics) { 3088 StubRoutines::_crc_table_adr = (address)StubRoutines::ppc64::_crc_table; 3089 StubRoutines::_updateBytesCRC32 = generate_CRC32_updateBytes("CRC32_updateBytes"); 3090 } 3091 } 3092 3093 void generate_all() { 3094 // Generates all stubs and initializes the entry points 3095 3096 // These entry points require SharedInfo::stack0 to be set up in 3097 // non-core builds 3098 StubRoutines::_throw_AbstractMethodError_entry = generate_throw_exception("AbstractMethodError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError), false); 3099 // Handle IncompatibleClassChangeError in itable stubs. 3100 StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError), false); 3101 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); 3102 3103 // support for verify_oop (must happen after universe_init) 3104 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop(); 3105 3106 // arraycopy stubs used by compilers 3107 generate_arraycopy_stubs(); 3108 3109 // Safefetch stubs. 3110 generate_safefetch("SafeFetch32", sizeof(int), &StubRoutines::_safefetch32_entry, 3111 &StubRoutines::_safefetch32_fault_pc, 3112 &StubRoutines::_safefetch32_continuation_pc); 3113 generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry, 3114 &StubRoutines::_safefetchN_fault_pc, 3115 &StubRoutines::_safefetchN_continuation_pc); 3116 3117#ifdef COMPILER2 3118 if (UseMultiplyToLenIntrinsic) { 3119 StubRoutines::_multiplyToLen = generate_multiplyToLen(); 3120 } 3121#endif 3122 3123 if (UseMontgomeryMultiplyIntrinsic) { 3124 StubRoutines::_montgomeryMultiply 3125 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply); 3126 } 3127 if (UseMontgomerySquareIntrinsic) { 3128 StubRoutines::_montgomerySquare 3129 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square); 3130 } 3131 3132 if (UseAESIntrinsics) { 3133 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock(); 3134 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock(); 3135 } 3136 3137 } 3138 3139 public: 3140 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) { 3141 // replace the standard masm with a special one: 3142 _masm = new MacroAssembler(code); 3143 if (all) { 3144 generate_all(); 3145 } else { 3146 generate_initial(); 3147 } 3148 } 3149}; 3150 3151void StubGenerator_generate(CodeBuffer* code, bool all) { 3152 StubGenerator g(code, all); 3153} 3154