1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "jvm.h" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "compiler/disassembler.hpp" 29 #include "gc/shared/cardTable.hpp" 30 #include "gc/shared/cardTableModRefBS.hpp" 31 #include "gc/shared/collectedHeap.inline.hpp" 32 #include "interpreter/interpreter.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "memory/universe.hpp" 35 #include "oops/klass.inline.hpp" 36 #include "prims/methodHandles.hpp" 37 #include "runtime/biasedLocking.hpp" 38 #include "runtime/interfaceSupport.hpp" 39 #include "runtime/objectMonitor.hpp" 40 #include "runtime/os.inline.hpp" 41 #include "runtime/safepoint.hpp" 42 #include "runtime/safepointMechanism.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 #include "runtime/stubRoutines.hpp" 45 #include "utilities/align.hpp" 46 #include "utilities/macros.hpp" 47 #if INCLUDE_ALL_GCS 48 #include "gc/g1/g1CardTable.hpp" 49 #include "gc/g1/g1CollectedHeap.inline.hpp" 50 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 51 #include "gc/g1/heapRegion.hpp" 52 #endif // INCLUDE_ALL_GCS 53 #ifdef COMPILER2 54 #include "opto/intrinsicnode.hpp" 55 #endif 56 57 #ifdef PRODUCT 58 #define BLOCK_COMMENT(str) /* nothing */ 59 #define STOP(error) stop(error) 60 #else 61 #define BLOCK_COMMENT(str) block_comment(str) 62 #define STOP(error) block_comment(error); stop(error) 63 #endif 64 65 // Convert the raw encoding form into the form expected by the 66 // constructor for Address. 67 Address Address::make_raw(int base, int index, int scale, int disp, relocInfo::relocType disp_reloc) { 68 assert(scale == 0, "not supported"); 69 RelocationHolder rspec; 70 if (disp_reloc != relocInfo::none) { 71 rspec = Relocation::spec_simple(disp_reloc); 72 } 73 74 Register rindex = as_Register(index); 75 if (rindex != G0) { 76 Address madr(as_Register(base), rindex); 77 madr._rspec = rspec; 78 return madr; 79 } else { 80 Address madr(as_Register(base), disp); 81 madr._rspec = rspec; 82 return madr; 83 } 84 } 85 86 Address Argument::address_in_frame() const { 87 // Warning: In LP64 mode disp will occupy more than 10 bits, but 88 // op codes such as ld or ldx, only access disp() to get 89 // their simm13 argument. 90 int disp = ((_number - Argument::n_register_parameters + frame::memory_parameter_word_sp_offset) * BytesPerWord) + STACK_BIAS; 91 if (is_in()) 92 return Address(FP, disp); // In argument. 93 else 94 return Address(SP, disp); // Out argument. 95 } 96 97 static const char* argumentNames[][2] = { 98 {"A0","P0"}, {"A1","P1"}, {"A2","P2"}, {"A3","P3"}, {"A4","P4"}, 99 {"A5","P5"}, {"A6","P6"}, {"A7","P7"}, {"A8","P8"}, {"A9","P9"}, 100 {"A(n>9)","P(n>9)"} 101 }; 102 103 const char* Argument::name() const { 104 int nofArgs = sizeof argumentNames / sizeof argumentNames[0]; 105 int num = number(); 106 if (num >= nofArgs) num = nofArgs - 1; 107 return argumentNames[num][is_in() ? 1 : 0]; 108 } 109 110 #ifdef ASSERT 111 // On RISC, there's no benefit to verifying instruction boundaries. 112 bool AbstractAssembler::pd_check_instruction_mark() { return false; } 113 #endif 114 115 // Patch instruction inst at offset inst_pos to refer to dest_pos 116 // and return the resulting instruction. 117 // We should have pcs, not offsets, but since all is relative, it will work out 118 // OK. 119 int MacroAssembler::patched_branch(int dest_pos, int inst, int inst_pos) { 120 int m; // mask for displacement field 121 int v; // new value for displacement field 122 const int word_aligned_ones = -4; 123 switch (inv_op(inst)) { 124 default: ShouldNotReachHere(); 125 case call_op: m = wdisp(word_aligned_ones, 0, 30); v = wdisp(dest_pos, inst_pos, 30); break; 126 case branch_op: 127 switch (inv_op2(inst)) { 128 case fbp_op2: m = wdisp( word_aligned_ones, 0, 19); v = wdisp( dest_pos, inst_pos, 19); break; 129 case bp_op2: m = wdisp( word_aligned_ones, 0, 19); v = wdisp( dest_pos, inst_pos, 19); break; 130 case fb_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break; 131 case br_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break; 132 case bpr_op2: { 133 if (is_cbcond(inst)) { 134 m = wdisp10(word_aligned_ones, 0); 135 v = wdisp10(dest_pos, inst_pos); 136 } else { 137 m = wdisp16(word_aligned_ones, 0); 138 v = wdisp16(dest_pos, inst_pos); 139 } 140 break; 141 } 142 default: ShouldNotReachHere(); 143 } 144 } 145 return inst & ~m | v; 146 } 147 148 // Return the offset of the branch destionation of instruction inst 149 // at offset pos. 150 // Should have pcs, but since all is relative, it works out. 151 int MacroAssembler::branch_destination(int inst, int pos) { 152 int r; 153 switch (inv_op(inst)) { 154 default: ShouldNotReachHere(); 155 case call_op: r = inv_wdisp(inst, pos, 30); break; 156 case branch_op: 157 switch (inv_op2(inst)) { 158 case fbp_op2: r = inv_wdisp( inst, pos, 19); break; 159 case bp_op2: r = inv_wdisp( inst, pos, 19); break; 160 case fb_op2: r = inv_wdisp( inst, pos, 22); break; 161 case br_op2: r = inv_wdisp( inst, pos, 22); break; 162 case bpr_op2: { 163 if (is_cbcond(inst)) { 164 r = inv_wdisp10(inst, pos); 165 } else { 166 r = inv_wdisp16(inst, pos); 167 } 168 break; 169 } 170 default: ShouldNotReachHere(); 171 } 172 } 173 return r; 174 } 175 176 void MacroAssembler::null_check(Register reg, int offset) { 177 if (needs_explicit_null_check((intptr_t)offset)) { 178 // provoke OS NULL exception if reg = NULL by 179 // accessing M[reg] w/o changing any registers 180 ld_ptr(reg, 0, G0); 181 } 182 else { 183 // nothing to do, (later) access of M[reg + offset] 184 // will provoke OS NULL exception if reg = NULL 185 } 186 } 187 188 // Ring buffer jumps 189 190 191 void MacroAssembler::jmp2(Register r1, Register r2, const char* file, int line ) { 192 assert_not_delayed(); 193 jmpl(r1, r2, G0); 194 } 195 void MacroAssembler::jmp(Register r1, int offset, const char* file, int line ) { 196 assert_not_delayed(); 197 jmp(r1, offset); 198 } 199 200 // This code sequence is relocatable to any address, even on LP64. 201 void MacroAssembler::jumpl(const AddressLiteral& addrlit, Register temp, Register d, int offset, const char* file, int line) { 202 assert_not_delayed(); 203 // Force fixed length sethi because NativeJump and NativeFarCall don't handle 204 // variable length instruction streams. 205 patchable_sethi(addrlit, temp); 206 Address a(temp, addrlit.low10() + offset); // Add the offset to the displacement. 207 jmpl(a.base(), a.disp(), d); 208 } 209 210 void MacroAssembler::jump(const AddressLiteral& addrlit, Register temp, int offset, const char* file, int line) { 211 jumpl(addrlit, temp, G0, offset, file, line); 212 } 213 214 215 // Conditional breakpoint (for assertion checks in assembly code) 216 void MacroAssembler::breakpoint_trap(Condition c, CC cc) { 217 trap(c, cc, G0, ST_RESERVED_FOR_USER_0); 218 } 219 220 // We want to use ST_BREAKPOINT here, but the debugger is confused by it. 221 void MacroAssembler::breakpoint_trap() { 222 trap(ST_RESERVED_FOR_USER_0); 223 } 224 225 // Write serialization page so VM thread can do a pseudo remote membar 226 // We use the current thread pointer to calculate a thread specific 227 // offset to write to within the page. This minimizes bus traffic 228 // due to cache line collision. 229 void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) { 230 srl(thread, os::get_serialize_page_shift_count(), tmp2); 231 if (Assembler::is_simm13(os::vm_page_size())) { 232 and3(tmp2, (os::vm_page_size() - sizeof(int)), tmp2); 233 } 234 else { 235 set((os::vm_page_size() - sizeof(int)), tmp1); 236 and3(tmp2, tmp1, tmp2); 237 } 238 set(os::get_memory_serialize_page(), tmp1); 239 st(G0, tmp1, tmp2); 240 } 241 242 243 void MacroAssembler::safepoint_poll(Label& slow_path, bool a, Register thread_reg, Register temp_reg) { 244 if (SafepointMechanism::uses_thread_local_poll()) { 245 ldx(Address(thread_reg, Thread::polling_page_offset()), temp_reg, 0); 246 // Armed page has poll bit set. 247 and3(temp_reg, SafepointMechanism::poll_bit(), temp_reg); 248 br_notnull(temp_reg, a, Assembler::pn, slow_path); 249 } else { 250 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 251 252 load_contents(sync_state, temp_reg); 253 cmp(temp_reg, SafepointSynchronize::_not_synchronized); 254 br(Assembler::notEqual, a, Assembler::pn, slow_path); 255 } 256 } 257 258 void MacroAssembler::enter() { 259 Unimplemented(); 260 } 261 262 void MacroAssembler::leave() { 263 Unimplemented(); 264 } 265 266 // Calls to C land 267 268 #ifdef ASSERT 269 // a hook for debugging 270 static Thread* reinitialize_thread() { 271 return Thread::current(); 272 } 273 #else 274 #define reinitialize_thread Thread::current 275 #endif 276 277 #ifdef ASSERT 278 address last_get_thread = NULL; 279 #endif 280 281 // call this when G2_thread is not known to be valid 282 void MacroAssembler::get_thread() { 283 save_frame(0); // to avoid clobbering O0 284 mov(G1, L0); // avoid clobbering G1 285 mov(G5_method, L1); // avoid clobbering G5 286 mov(G3, L2); // avoid clobbering G3 also 287 mov(G4, L5); // avoid clobbering G4 288 #ifdef ASSERT 289 AddressLiteral last_get_thread_addrlit(&last_get_thread); 290 set(last_get_thread_addrlit, L3); 291 rdpc(L4); 292 inc(L4, 3 * BytesPerInstWord); // skip rdpc + inc + st_ptr to point L4 at call st_ptr(L4, L3, 0); 293 #endif 294 call(CAST_FROM_FN_PTR(address, reinitialize_thread), relocInfo::runtime_call_type); 295 delayed()->nop(); 296 mov(L0, G1); 297 mov(L1, G5_method); 298 mov(L2, G3); 299 mov(L5, G4); 300 restore(O0, 0, G2_thread); 301 } 302 303 static Thread* verify_thread_subroutine(Thread* gthread_value) { 304 Thread* correct_value = Thread::current(); 305 guarantee(gthread_value == correct_value, "G2_thread value must be the thread"); 306 return correct_value; 307 } 308 309 void MacroAssembler::verify_thread() { 310 if (VerifyThread) { 311 // NOTE: this chops off the heads of the 64-bit O registers. 312 // make sure G2_thread contains the right value 313 save_frame_and_mov(0, Lmethod, Lmethod); // to avoid clobbering O0 (and propagate Lmethod) 314 mov(G1, L1); // avoid clobbering G1 315 // G2 saved below 316 mov(G3, L3); // avoid clobbering G3 317 mov(G4, L4); // avoid clobbering G4 318 mov(G5_method, L5); // avoid clobbering G5_method 319 call(CAST_FROM_FN_PTR(address,verify_thread_subroutine), relocInfo::runtime_call_type); 320 delayed()->mov(G2_thread, O0); 321 322 mov(L1, G1); // Restore G1 323 // G2 restored below 324 mov(L3, G3); // restore G3 325 mov(L4, G4); // restore G4 326 mov(L5, G5_method); // restore G5_method 327 restore(O0, 0, G2_thread); 328 } 329 } 330 331 332 void MacroAssembler::save_thread(const Register thread_cache) { 333 verify_thread(); 334 if (thread_cache->is_valid()) { 335 assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile"); 336 mov(G2_thread, thread_cache); 337 } 338 if (VerifyThread) { 339 // smash G2_thread, as if the VM were about to anyway 340 set(0x67676767, G2_thread); 341 } 342 } 343 344 345 void MacroAssembler::restore_thread(const Register thread_cache) { 346 if (thread_cache->is_valid()) { 347 assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile"); 348 mov(thread_cache, G2_thread); 349 verify_thread(); 350 } else { 351 // do it the slow way 352 get_thread(); 353 } 354 } 355 356 357 // %%% maybe get rid of [re]set_last_Java_frame 358 void MacroAssembler::set_last_Java_frame(Register last_java_sp, Register last_Java_pc) { 359 assert_not_delayed(); 360 Address flags(G2_thread, JavaThread::frame_anchor_offset() + 361 JavaFrameAnchor::flags_offset()); 362 Address pc_addr(G2_thread, JavaThread::last_Java_pc_offset()); 363 364 // Always set last_Java_pc and flags first because once last_Java_sp is visible 365 // has_last_Java_frame is true and users will look at the rest of the fields. 366 // (Note: flags should always be zero before we get here so doesn't need to be set.) 367 368 #ifdef ASSERT 369 // Verify that flags was zeroed on return to Java 370 Label PcOk; 371 save_frame(0); // to avoid clobbering O0 372 ld_ptr(pc_addr, L0); 373 br_null_short(L0, Assembler::pt, PcOk); 374 STOP("last_Java_pc not zeroed before leaving Java"); 375 bind(PcOk); 376 377 // Verify that flags was zeroed on return to Java 378 Label FlagsOk; 379 ld(flags, L0); 380 tst(L0); 381 br(Assembler::zero, false, Assembler::pt, FlagsOk); 382 delayed() -> restore(); 383 STOP("flags not zeroed before leaving Java"); 384 bind(FlagsOk); 385 #endif /* ASSERT */ 386 // 387 // When returning from calling out from Java mode the frame anchor's last_Java_pc 388 // will always be set to NULL. It is set here so that if we are doing a call to 389 // native (not VM) that we capture the known pc and don't have to rely on the 390 // native call having a standard frame linkage where we can find the pc. 391 392 if (last_Java_pc->is_valid()) { 393 st_ptr(last_Java_pc, pc_addr); 394 } 395 396 #ifdef ASSERT 397 // Make sure that we have an odd stack 398 Label StackOk; 399 andcc(last_java_sp, 0x01, G0); 400 br(Assembler::notZero, false, Assembler::pt, StackOk); 401 delayed()->nop(); 402 STOP("Stack Not Biased in set_last_Java_frame"); 403 bind(StackOk); 404 #endif // ASSERT 405 assert( last_java_sp != G4_scratch, "bad register usage in set_last_Java_frame"); 406 add( last_java_sp, STACK_BIAS, G4_scratch ); 407 st_ptr(G4_scratch, G2_thread, JavaThread::last_Java_sp_offset()); 408 } 409 410 void MacroAssembler::reset_last_Java_frame(void) { 411 assert_not_delayed(); 412 413 Address sp_addr(G2_thread, JavaThread::last_Java_sp_offset()); 414 Address pc_addr(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 415 Address flags (G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 416 417 #ifdef ASSERT 418 // check that it WAS previously set 419 save_frame_and_mov(0, Lmethod, Lmethod); // Propagate Lmethod to helper frame 420 ld_ptr(sp_addr, L0); 421 tst(L0); 422 breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 423 restore(); 424 #endif // ASSERT 425 426 st_ptr(G0, sp_addr); 427 // Always return last_Java_pc to zero 428 st_ptr(G0, pc_addr); 429 // Always null flags after return to Java 430 st(G0, flags); 431 } 432 433 434 void MacroAssembler::call_VM_base( 435 Register oop_result, 436 Register thread_cache, 437 Register last_java_sp, 438 address entry_point, 439 int number_of_arguments, 440 bool check_exceptions) 441 { 442 assert_not_delayed(); 443 444 // determine last_java_sp register 445 if (!last_java_sp->is_valid()) { 446 last_java_sp = SP; 447 } 448 // debugging support 449 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 450 451 // 64-bit last_java_sp is biased! 452 set_last_Java_frame(last_java_sp, noreg); 453 if (VerifyThread) mov(G2_thread, O0); // about to be smashed; pass early 454 save_thread(thread_cache); 455 // do the call 456 call(entry_point, relocInfo::runtime_call_type); 457 if (!VerifyThread) 458 delayed()->mov(G2_thread, O0); // pass thread as first argument 459 else 460 delayed()->nop(); // (thread already passed) 461 restore_thread(thread_cache); 462 reset_last_Java_frame(); 463 464 // check for pending exceptions. use Gtemp as scratch register. 465 if (check_exceptions) { 466 check_and_forward_exception(Gtemp); 467 } 468 469 #ifdef ASSERT 470 set(badHeapWordVal, G3); 471 set(badHeapWordVal, G4); 472 set(badHeapWordVal, G5); 473 #endif 474 475 // get oop result if there is one and reset the value in the thread 476 if (oop_result->is_valid()) { 477 get_vm_result(oop_result); 478 } 479 } 480 481 void MacroAssembler::check_and_forward_exception(Register scratch_reg) 482 { 483 Label L; 484 485 check_and_handle_popframe(scratch_reg); 486 check_and_handle_earlyret(scratch_reg); 487 488 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 489 ld_ptr(exception_addr, scratch_reg); 490 br_null_short(scratch_reg, pt, L); 491 // we use O7 linkage so that forward_exception_entry has the issuing PC 492 call(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 493 delayed()->nop(); 494 bind(L); 495 } 496 497 498 void MacroAssembler::check_and_handle_popframe(Register scratch_reg) { 499 } 500 501 502 void MacroAssembler::check_and_handle_earlyret(Register scratch_reg) { 503 } 504 505 506 void MacroAssembler::call_VM(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 507 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions); 508 } 509 510 511 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, bool check_exceptions) { 512 // O0 is reserved for the thread 513 mov(arg_1, O1); 514 call_VM(oop_result, entry_point, 1, check_exceptions); 515 } 516 517 518 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) { 519 // O0 is reserved for the thread 520 mov(arg_1, O1); 521 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 522 call_VM(oop_result, entry_point, 2, check_exceptions); 523 } 524 525 526 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, Register arg_3, bool check_exceptions) { 527 // O0 is reserved for the thread 528 mov(arg_1, O1); 529 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 530 mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument"); 531 call_VM(oop_result, entry_point, 3, check_exceptions); 532 } 533 534 535 536 // Note: The following call_VM overloadings are useful when a "save" 537 // has already been performed by a stub, and the last Java frame is 538 // the previous one. In that case, last_java_sp must be passed as FP 539 // instead of SP. 540 541 542 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, int number_of_arguments, bool check_exceptions) { 543 call_VM_base(oop_result, noreg, last_java_sp, entry_point, number_of_arguments, check_exceptions); 544 } 545 546 547 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, bool check_exceptions) { 548 // O0 is reserved for the thread 549 mov(arg_1, O1); 550 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 551 } 552 553 554 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) { 555 // O0 is reserved for the thread 556 mov(arg_1, O1); 557 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 558 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 559 } 560 561 562 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, Register arg_2, Register arg_3, bool check_exceptions) { 563 // O0 is reserved for the thread 564 mov(arg_1, O1); 565 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 566 mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument"); 567 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 568 } 569 570 571 572 void MacroAssembler::call_VM_leaf_base(Register thread_cache, address entry_point, int number_of_arguments) { 573 assert_not_delayed(); 574 save_thread(thread_cache); 575 // do the call 576 call(entry_point, relocInfo::runtime_call_type); 577 delayed()->nop(); 578 restore_thread(thread_cache); 579 #ifdef ASSERT 580 set(badHeapWordVal, G3); 581 set(badHeapWordVal, G4); 582 set(badHeapWordVal, G5); 583 #endif 584 } 585 586 587 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, int number_of_arguments) { 588 call_VM_leaf_base(thread_cache, entry_point, number_of_arguments); 589 } 590 591 592 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1) { 593 mov(arg_1, O0); 594 call_VM_leaf(thread_cache, entry_point, 1); 595 } 596 597 598 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2) { 599 mov(arg_1, O0); 600 mov(arg_2, O1); assert(arg_2 != O0, "smashed argument"); 601 call_VM_leaf(thread_cache, entry_point, 2); 602 } 603 604 605 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2, Register arg_3) { 606 mov(arg_1, O0); 607 mov(arg_2, O1); assert(arg_2 != O0, "smashed argument"); 608 mov(arg_3, O2); assert(arg_3 != O0 && arg_3 != O1, "smashed argument"); 609 call_VM_leaf(thread_cache, entry_point, 3); 610 } 611 612 613 void MacroAssembler::get_vm_result(Register oop_result) { 614 verify_thread(); 615 Address vm_result_addr(G2_thread, JavaThread::vm_result_offset()); 616 ld_ptr( vm_result_addr, oop_result); 617 st_ptr(G0, vm_result_addr); 618 verify_oop(oop_result); 619 } 620 621 622 void MacroAssembler::get_vm_result_2(Register metadata_result) { 623 verify_thread(); 624 Address vm_result_addr_2(G2_thread, JavaThread::vm_result_2_offset()); 625 ld_ptr(vm_result_addr_2, metadata_result); 626 st_ptr(G0, vm_result_addr_2); 627 } 628 629 630 // We require that C code which does not return a value in vm_result will 631 // leave it undisturbed. 632 void MacroAssembler::set_vm_result(Register oop_result) { 633 verify_thread(); 634 Address vm_result_addr(G2_thread, JavaThread::vm_result_offset()); 635 verify_oop(oop_result); 636 637 # ifdef ASSERT 638 // Check that we are not overwriting any other oop. 639 save_frame_and_mov(0, Lmethod, Lmethod); // Propagate Lmethod 640 ld_ptr(vm_result_addr, L0); 641 tst(L0); 642 restore(); 643 breakpoint_trap(notZero, Assembler::ptr_cc); 644 // } 645 # endif 646 647 st_ptr(oop_result, vm_result_addr); 648 } 649 650 651 void MacroAssembler::ic_call(address entry, bool emit_delay, jint method_index) { 652 RelocationHolder rspec = virtual_call_Relocation::spec(pc(), method_index); 653 patchable_set((intptr_t)Universe::non_oop_word(), G5_inline_cache_reg); 654 relocate(rspec); 655 call(entry, relocInfo::none); 656 if (emit_delay) { 657 delayed()->nop(); 658 } 659 } 660 661 void MacroAssembler::card_table_write(jbyte* byte_map_base, 662 Register tmp, Register obj) { 663 srlx(obj, CardTable::card_shift, obj); 664 assert(tmp != obj, "need separate temp reg"); 665 set((address) byte_map_base, tmp); 666 stb(G0, tmp, obj); 667 } 668 669 670 void MacroAssembler::internal_sethi(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) { 671 address save_pc; 672 int shiftcnt; 673 #ifdef VALIDATE_PIPELINE 674 assert_no_delay("Cannot put two instructions in delay-slot."); 675 #endif 676 v9_dep(); 677 save_pc = pc(); 678 679 int msb32 = (int) (addrlit.value() >> 32); 680 int lsb32 = (int) (addrlit.value()); 681 682 if (msb32 == 0 && lsb32 >= 0) { 683 Assembler::sethi(lsb32, d, addrlit.rspec()); 684 } 685 else if (msb32 == -1) { 686 Assembler::sethi(~lsb32, d, addrlit.rspec()); 687 xor3(d, ~low10(~0), d); 688 } 689 else { 690 Assembler::sethi(msb32, d, addrlit.rspec()); // msb 22-bits 691 if (msb32 & 0x3ff) // Any bits? 692 or3(d, msb32 & 0x3ff, d); // msb 32-bits are now in lsb 32 693 if (lsb32 & 0xFFFFFC00) { // done? 694 if ((lsb32 >> 20) & 0xfff) { // Any bits set? 695 sllx(d, 12, d); // Make room for next 12 bits 696 or3(d, (lsb32 >> 20) & 0xfff, d); // Or in next 12 697 shiftcnt = 0; // We already shifted 698 } 699 else 700 shiftcnt = 12; 701 if ((lsb32 >> 10) & 0x3ff) { 702 sllx(d, shiftcnt + 10, d); // Make room for last 10 bits 703 or3(d, (lsb32 >> 10) & 0x3ff, d); // Or in next 10 704 shiftcnt = 0; 705 } 706 else 707 shiftcnt = 10; 708 sllx(d, shiftcnt + 10, d); // Shift leaving disp field 0'd 709 } 710 else 711 sllx(d, 32, d); 712 } 713 // Pad out the instruction sequence so it can be patched later. 714 if (ForceRelocatable || (addrlit.rtype() != relocInfo::none && 715 addrlit.rtype() != relocInfo::runtime_call_type)) { 716 while (pc() < (save_pc + (7 * BytesPerInstWord))) 717 nop(); 718 } 719 } 720 721 722 void MacroAssembler::sethi(const AddressLiteral& addrlit, Register d) { 723 internal_sethi(addrlit, d, false); 724 } 725 726 727 void MacroAssembler::patchable_sethi(const AddressLiteral& addrlit, Register d) { 728 internal_sethi(addrlit, d, true); 729 } 730 731 732 int MacroAssembler::insts_for_sethi(address a, bool worst_case) { 733 if (worst_case) return 7; 734 intptr_t iaddr = (intptr_t) a; 735 int msb32 = (int) (iaddr >> 32); 736 int lsb32 = (int) (iaddr); 737 int count; 738 if (msb32 == 0 && lsb32 >= 0) 739 count = 1; 740 else if (msb32 == -1) 741 count = 2; 742 else { 743 count = 2; 744 if (msb32 & 0x3ff) 745 count++; 746 if (lsb32 & 0xFFFFFC00 ) { 747 if ((lsb32 >> 20) & 0xfff) count += 2; 748 if ((lsb32 >> 10) & 0x3ff) count += 2; 749 } 750 } 751 return count; 752 } 753 754 int MacroAssembler::worst_case_insts_for_set() { 755 return insts_for_sethi(NULL, true) + 1; 756 } 757 758 759 // Keep in sync with MacroAssembler::insts_for_internal_set 760 void MacroAssembler::internal_set(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) { 761 intptr_t value = addrlit.value(); 762 763 if (!ForceRelocatable && addrlit.rspec().type() == relocInfo::none) { 764 // can optimize 765 if (-4096 <= value && value <= 4095) { 766 or3(G0, value, d); // setsw (this leaves upper 32 bits sign-extended) 767 return; 768 } 769 if (inv_hi22(hi22(value)) == value) { 770 sethi(addrlit, d); 771 return; 772 } 773 } 774 assert_no_delay("Cannot put two instructions in delay-slot."); 775 internal_sethi(addrlit, d, ForceRelocatable); 776 if (ForceRelocatable || addrlit.rspec().type() != relocInfo::none || addrlit.low10() != 0) { 777 add(d, addrlit.low10(), d, addrlit.rspec()); 778 } 779 } 780 781 // Keep in sync with MacroAssembler::internal_set 782 int MacroAssembler::insts_for_internal_set(intptr_t value) { 783 // can optimize 784 if (-4096 <= value && value <= 4095) { 785 return 1; 786 } 787 if (inv_hi22(hi22(value)) == value) { 788 return insts_for_sethi((address) value); 789 } 790 int count = insts_for_sethi((address) value); 791 AddressLiteral al(value); 792 if (al.low10() != 0) { 793 count++; 794 } 795 return count; 796 } 797 798 void MacroAssembler::set(const AddressLiteral& al, Register d) { 799 internal_set(al, d, false); 800 } 801 802 void MacroAssembler::set(intptr_t value, Register d) { 803 AddressLiteral al(value); 804 internal_set(al, d, false); 805 } 806 807 void MacroAssembler::set(address addr, Register d, RelocationHolder const& rspec) { 808 AddressLiteral al(addr, rspec); 809 internal_set(al, d, false); 810 } 811 812 void MacroAssembler::patchable_set(const AddressLiteral& al, Register d) { 813 internal_set(al, d, true); 814 } 815 816 void MacroAssembler::patchable_set(intptr_t value, Register d) { 817 AddressLiteral al(value); 818 internal_set(al, d, true); 819 } 820 821 822 void MacroAssembler::set64(jlong value, Register d, Register tmp) { 823 assert_not_delayed(); 824 v9_dep(); 825 826 int hi = (int)(value >> 32); 827 int lo = (int)(value & ~0); 828 int bits_33to2 = (int)((value >> 2) & ~0); 829 // (Matcher::isSimpleConstant64 knows about the following optimizations.) 830 if (Assembler::is_simm13(lo) && value == lo) { 831 or3(G0, lo, d); 832 } else if (hi == 0) { 833 Assembler::sethi(lo, d); // hardware version zero-extends to upper 32 834 if (low10(lo) != 0) 835 or3(d, low10(lo), d); 836 } 837 else if ((hi >> 2) == 0) { 838 Assembler::sethi(bits_33to2, d); // hardware version zero-extends to upper 32 839 sllx(d, 2, d); 840 if (low12(lo) != 0) 841 or3(d, low12(lo), d); 842 } 843 else if (hi == -1) { 844 Assembler::sethi(~lo, d); // hardware version zero-extends to upper 32 845 xor3(d, low10(lo) ^ ~low10(~0), d); 846 } 847 else if (lo == 0) { 848 if (Assembler::is_simm13(hi)) { 849 or3(G0, hi, d); 850 } else { 851 Assembler::sethi(hi, d); // hardware version zero-extends to upper 32 852 if (low10(hi) != 0) 853 or3(d, low10(hi), d); 854 } 855 sllx(d, 32, d); 856 } 857 else { 858 Assembler::sethi(hi, tmp); 859 Assembler::sethi(lo, d); // macro assembler version sign-extends 860 if (low10(hi) != 0) 861 or3 (tmp, low10(hi), tmp); 862 if (low10(lo) != 0) 863 or3 ( d, low10(lo), d); 864 sllx(tmp, 32, tmp); 865 or3 (d, tmp, d); 866 } 867 } 868 869 int MacroAssembler::insts_for_set64(jlong value) { 870 v9_dep(); 871 872 int hi = (int) (value >> 32); 873 int lo = (int) (value & ~0); 874 int count = 0; 875 876 // (Matcher::isSimpleConstant64 knows about the following optimizations.) 877 if (Assembler::is_simm13(lo) && value == lo) { 878 count++; 879 } else if (hi == 0) { 880 count++; 881 if (low10(lo) != 0) 882 count++; 883 } 884 else if (hi == -1) { 885 count += 2; 886 } 887 else if (lo == 0) { 888 if (Assembler::is_simm13(hi)) { 889 count++; 890 } else { 891 count++; 892 if (low10(hi) != 0) 893 count++; 894 } 895 count++; 896 } 897 else { 898 count += 2; 899 if (low10(hi) != 0) 900 count++; 901 if (low10(lo) != 0) 902 count++; 903 count += 2; 904 } 905 return count; 906 } 907 908 // compute size in bytes of sparc frame, given 909 // number of extraWords 910 int MacroAssembler::total_frame_size_in_bytes(int extraWords) { 911 912 int nWords = frame::memory_parameter_word_sp_offset; 913 914 nWords += extraWords; 915 916 if (nWords & 1) ++nWords; // round up to double-word 917 918 return nWords * BytesPerWord; 919 } 920 921 922 // save_frame: given number of "extra" words in frame, 923 // issue approp. save instruction (p 200, v8 manual) 924 925 void MacroAssembler::save_frame(int extraWords) { 926 int delta = -total_frame_size_in_bytes(extraWords); 927 if (is_simm13(delta)) { 928 save(SP, delta, SP); 929 } else { 930 set(delta, G3_scratch); 931 save(SP, G3_scratch, SP); 932 } 933 } 934 935 936 void MacroAssembler::save_frame_c1(int size_in_bytes) { 937 if (is_simm13(-size_in_bytes)) { 938 save(SP, -size_in_bytes, SP); 939 } else { 940 set(-size_in_bytes, G3_scratch); 941 save(SP, G3_scratch, SP); 942 } 943 } 944 945 946 void MacroAssembler::save_frame_and_mov(int extraWords, 947 Register s1, Register d1, 948 Register s2, Register d2) { 949 assert_not_delayed(); 950 951 // The trick here is to use precisely the same memory word 952 // that trap handlers also use to save the register. 953 // This word cannot be used for any other purpose, but 954 // it works fine to save the register's value, whether or not 955 // an interrupt flushes register windows at any given moment! 956 Address s1_addr; 957 if (s1->is_valid() && (s1->is_in() || s1->is_local())) { 958 s1_addr = s1->address_in_saved_window(); 959 st_ptr(s1, s1_addr); 960 } 961 962 Address s2_addr; 963 if (s2->is_valid() && (s2->is_in() || s2->is_local())) { 964 s2_addr = s2->address_in_saved_window(); 965 st_ptr(s2, s2_addr); 966 } 967 968 save_frame(extraWords); 969 970 if (s1_addr.base() == SP) { 971 ld_ptr(s1_addr.after_save(), d1); 972 } else if (s1->is_valid()) { 973 mov(s1->after_save(), d1); 974 } 975 976 if (s2_addr.base() == SP) { 977 ld_ptr(s2_addr.after_save(), d2); 978 } else if (s2->is_valid()) { 979 mov(s2->after_save(), d2); 980 } 981 } 982 983 984 AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) { 985 assert(oop_recorder() != NULL, "this assembler needs a Recorder"); 986 int index = oop_recorder()->allocate_metadata_index(obj); 987 RelocationHolder rspec = metadata_Relocation::spec(index); 988 return AddressLiteral((address)obj, rspec); 989 } 990 991 AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) { 992 assert(oop_recorder() != NULL, "this assembler needs a Recorder"); 993 int index = oop_recorder()->find_index(obj); 994 RelocationHolder rspec = metadata_Relocation::spec(index); 995 return AddressLiteral((address)obj, rspec); 996 } 997 998 999 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) { 1000 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); 1001 assert(Universe::heap()->is_in_reserved(JNIHandles::resolve(obj)), "not an oop"); 1002 int oop_index = oop_recorder()->find_index(obj); 1003 return AddressLiteral(obj, oop_Relocation::spec(oop_index)); 1004 } 1005 1006 void MacroAssembler::set_narrow_oop(jobject obj, Register d) { 1007 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); 1008 int oop_index = oop_recorder()->find_index(obj); 1009 RelocationHolder rspec = oop_Relocation::spec(oop_index); 1010 1011 assert_not_delayed(); 1012 // Relocation with special format (see relocInfo_sparc.hpp). 1013 relocate(rspec, 1); 1014 // Assembler::sethi(0x3fffff, d); 1015 emit_int32( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(0x3fffff) ); 1016 // Don't add relocation for 'add'. Do patching during 'sethi' processing. 1017 add(d, 0x3ff, d); 1018 1019 } 1020 1021 void MacroAssembler::set_narrow_klass(Klass* k, Register d) { 1022 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); 1023 int klass_index = oop_recorder()->find_index(k); 1024 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 1025 narrowOop encoded_k = Klass::encode_klass(k); 1026 1027 assert_not_delayed(); 1028 // Relocation with special format (see relocInfo_sparc.hpp). 1029 relocate(rspec, 1); 1030 // Assembler::sethi(encoded_k, d); 1031 emit_int32( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(encoded_k) ); 1032 // Don't add relocation for 'add'. Do patching during 'sethi' processing. 1033 add(d, low10(encoded_k), d); 1034 1035 } 1036 1037 void MacroAssembler::align(int modulus) { 1038 while (offset() % modulus != 0) nop(); 1039 } 1040 1041 void RegistersForDebugging::print(outputStream* s) { 1042 FlagSetting fs(Debugging, true); 1043 int j; 1044 for (j = 0; j < 8; ++j) { 1045 if (j != 6) { s->print("i%d = ", j); os::print_location(s, i[j]); } 1046 else { s->print( "fp = " ); os::print_location(s, i[j]); } 1047 } 1048 s->cr(); 1049 1050 for (j = 0; j < 8; ++j) { 1051 s->print("l%d = ", j); os::print_location(s, l[j]); 1052 } 1053 s->cr(); 1054 1055 for (j = 0; j < 8; ++j) { 1056 if (j != 6) { s->print("o%d = ", j); os::print_location(s, o[j]); } 1057 else { s->print( "sp = " ); os::print_location(s, o[j]); } 1058 } 1059 s->cr(); 1060 1061 for (j = 0; j < 8; ++j) { 1062 s->print("g%d = ", j); os::print_location(s, g[j]); 1063 } 1064 s->cr(); 1065 1066 // print out floats with compression 1067 for (j = 0; j < 32; ) { 1068 jfloat val = f[j]; 1069 int last = j; 1070 for ( ; last+1 < 32; ++last ) { 1071 char b1[1024], b2[1024]; 1072 sprintf(b1, "%f", val); 1073 sprintf(b2, "%f", f[last+1]); 1074 if (strcmp(b1, b2)) 1075 break; 1076 } 1077 s->print("f%d", j); 1078 if ( j != last ) s->print(" - f%d", last); 1079 s->print(" = %f", val); 1080 s->fill_to(25); 1081 s->print_cr(" (0x%x)", *(int*)&val); 1082 j = last + 1; 1083 } 1084 s->cr(); 1085 1086 // and doubles (evens only) 1087 for (j = 0; j < 32; ) { 1088 jdouble val = d[j]; 1089 int last = j; 1090 for ( ; last+1 < 32; ++last ) { 1091 char b1[1024], b2[1024]; 1092 sprintf(b1, "%f", val); 1093 sprintf(b2, "%f", d[last+1]); 1094 if (strcmp(b1, b2)) 1095 break; 1096 } 1097 s->print("d%d", 2 * j); 1098 if ( j != last ) s->print(" - d%d", last); 1099 s->print(" = %f", val); 1100 s->fill_to(30); 1101 s->print("(0x%x)", *(int*)&val); 1102 s->fill_to(42); 1103 s->print_cr("(0x%x)", *(1 + (int*)&val)); 1104 j = last + 1; 1105 } 1106 s->cr(); 1107 } 1108 1109 void RegistersForDebugging::save_registers(MacroAssembler* a) { 1110 a->sub(FP, align_up(sizeof(RegistersForDebugging), sizeof(jdouble)) - STACK_BIAS, O0); 1111 a->flushw(); 1112 int i; 1113 for (i = 0; i < 8; ++i) { 1114 a->ld_ptr(as_iRegister(i)->address_in_saved_window().after_save(), L1); a->st_ptr( L1, O0, i_offset(i)); 1115 a->ld_ptr(as_lRegister(i)->address_in_saved_window().after_save(), L1); a->st_ptr( L1, O0, l_offset(i)); 1116 a->st_ptr(as_oRegister(i)->after_save(), O0, o_offset(i)); 1117 a->st_ptr(as_gRegister(i)->after_save(), O0, g_offset(i)); 1118 } 1119 for (i = 0; i < 32; ++i) { 1120 a->stf(FloatRegisterImpl::S, as_FloatRegister(i), O0, f_offset(i)); 1121 } 1122 for (i = 0; i < 64; i += 2) { 1123 a->stf(FloatRegisterImpl::D, as_FloatRegister(i), O0, d_offset(i)); 1124 } 1125 } 1126 1127 void RegistersForDebugging::restore_registers(MacroAssembler* a, Register r) { 1128 for (int i = 1; i < 8; ++i) { 1129 a->ld_ptr(r, g_offset(i), as_gRegister(i)); 1130 } 1131 for (int j = 0; j < 32; ++j) { 1132 a->ldf(FloatRegisterImpl::S, O0, f_offset(j), as_FloatRegister(j)); 1133 } 1134 for (int k = 0; k < 64; k += 2) { 1135 a->ldf(FloatRegisterImpl::D, O0, d_offset(k), as_FloatRegister(k)); 1136 } 1137 } 1138 1139 1140 // pushes double TOS element of FPU stack on CPU stack; pops from FPU stack 1141 void MacroAssembler::push_fTOS() { 1142 // %%%%%% need to implement this 1143 } 1144 1145 // pops double TOS element from CPU stack and pushes on FPU stack 1146 void MacroAssembler::pop_fTOS() { 1147 // %%%%%% need to implement this 1148 } 1149 1150 void MacroAssembler::empty_FPU_stack() { 1151 // %%%%%% need to implement this 1152 } 1153 1154 void MacroAssembler::_verify_oop(Register reg, const char* msg, const char * file, int line) { 1155 // plausibility check for oops 1156 if (!VerifyOops) return; 1157 1158 if (reg == G0) return; // always NULL, which is always an oop 1159 1160 BLOCK_COMMENT("verify_oop {"); 1161 char buffer[64]; 1162 #ifdef COMPILER1 1163 if (CommentedAssembly) { 1164 snprintf(buffer, sizeof(buffer), "verify_oop at %d", offset()); 1165 block_comment(buffer); 1166 } 1167 #endif 1168 1169 const char* real_msg = NULL; 1170 { 1171 ResourceMark rm; 1172 stringStream ss; 1173 ss.print("%s at offset %d (%s:%d)", msg, offset(), file, line); 1174 real_msg = code_string(ss.as_string()); 1175 } 1176 1177 // Call indirectly to solve generation ordering problem 1178 AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address()); 1179 1180 // Make some space on stack above the current register window. 1181 // Enough to hold 8 64-bit registers. 1182 add(SP,-8*8,SP); 1183 1184 // Save some 64-bit registers; a normal 'save' chops the heads off 1185 // of 64-bit longs in the 32-bit build. 1186 stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8); 1187 stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8); 1188 mov(reg,O0); // Move arg into O0; arg might be in O7 which is about to be crushed 1189 stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8); 1190 1191 // Size of set() should stay the same 1192 patchable_set((intptr_t)real_msg, O1); 1193 // Load address to call to into O7 1194 load_ptr_contents(a, O7); 1195 // Register call to verify_oop_subroutine 1196 callr(O7, G0); 1197 delayed()->nop(); 1198 // recover frame size 1199 add(SP, 8*8,SP); 1200 BLOCK_COMMENT("} verify_oop"); 1201 } 1202 1203 void MacroAssembler::_verify_oop_addr(Address addr, const char* msg, const char * file, int line) { 1204 // plausibility check for oops 1205 if (!VerifyOops) return; 1206 1207 const char* real_msg = NULL; 1208 { 1209 ResourceMark rm; 1210 stringStream ss; 1211 ss.print("%s at SP+%d (%s:%d)", msg, addr.disp(), file, line); 1212 real_msg = code_string(ss.as_string()); 1213 } 1214 1215 // Call indirectly to solve generation ordering problem 1216 AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address()); 1217 1218 // Make some space on stack above the current register window. 1219 // Enough to hold 8 64-bit registers. 1220 add(SP,-8*8,SP); 1221 1222 // Save some 64-bit registers; a normal 'save' chops the heads off 1223 // of 64-bit longs in the 32-bit build. 1224 stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8); 1225 stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8); 1226 ld_ptr(addr.base(), addr.disp() + 8*8, O0); // Load arg into O0; arg might be in O7 which is about to be crushed 1227 stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8); 1228 1229 // Size of set() should stay the same 1230 patchable_set((intptr_t)real_msg, O1); 1231 // Load address to call to into O7 1232 load_ptr_contents(a, O7); 1233 // Register call to verify_oop_subroutine 1234 callr(O7, G0); 1235 delayed()->nop(); 1236 // recover frame size 1237 add(SP, 8*8,SP); 1238 } 1239 1240 // side-door communication with signalHandler in os_solaris.cpp 1241 address MacroAssembler::_verify_oop_implicit_branch[3] = { NULL }; 1242 1243 // This macro is expanded just once; it creates shared code. Contract: 1244 // receives an oop in O0. Must restore O0 & O7 from TLS. Must not smash ANY 1245 // registers, including flags. May not use a register 'save', as this blows 1246 // the high bits of the O-regs if they contain Long values. Acts as a 'leaf' 1247 // call. 1248 void MacroAssembler::verify_oop_subroutine() { 1249 // Leaf call; no frame. 1250 Label succeed, fail, null_or_fail; 1251 1252 // O0 and O7 were saved already (O0 in O0's TLS home, O7 in O5's TLS home). 1253 // O0 is now the oop to be checked. O7 is the return address. 1254 Register O0_obj = O0; 1255 1256 // Save some more registers for temps. 1257 stx(O2,SP,frame::register_save_words*wordSize+STACK_BIAS+2*8); 1258 stx(O3,SP,frame::register_save_words*wordSize+STACK_BIAS+3*8); 1259 stx(O4,SP,frame::register_save_words*wordSize+STACK_BIAS+4*8); 1260 stx(O5,SP,frame::register_save_words*wordSize+STACK_BIAS+5*8); 1261 1262 // Save flags 1263 Register O5_save_flags = O5; 1264 rdccr( O5_save_flags ); 1265 1266 { // count number of verifies 1267 Register O2_adr = O2; 1268 Register O3_accum = O3; 1269 inc_counter(StubRoutines::verify_oop_count_addr(), O2_adr, O3_accum); 1270 } 1271 1272 Register O2_mask = O2; 1273 Register O3_bits = O3; 1274 Register O4_temp = O4; 1275 1276 // mark lower end of faulting range 1277 assert(_verify_oop_implicit_branch[0] == NULL, "set once"); 1278 _verify_oop_implicit_branch[0] = pc(); 1279 1280 // We can't check the mark oop because it could be in the process of 1281 // locking or unlocking while this is running. 1282 set(Universe::verify_oop_mask (), O2_mask); 1283 set(Universe::verify_oop_bits (), O3_bits); 1284 1285 // assert((obj & oop_mask) == oop_bits); 1286 and3(O0_obj, O2_mask, O4_temp); 1287 cmp_and_brx_short(O4_temp, O3_bits, notEqual, pn, null_or_fail); 1288 1289 if ((NULL_WORD & Universe::verify_oop_mask()) == Universe::verify_oop_bits()) { 1290 // the null_or_fail case is useless; must test for null separately 1291 br_null_short(O0_obj, pn, succeed); 1292 } 1293 1294 // Check the Klass* of this object for being in the right area of memory. 1295 // Cannot do the load in the delay above slot in case O0 is null 1296 load_klass(O0_obj, O0_obj); 1297 // assert((klass != NULL) 1298 br_null_short(O0_obj, pn, fail); 1299 1300 wrccr( O5_save_flags ); // Restore CCR's 1301 1302 // mark upper end of faulting range 1303 _verify_oop_implicit_branch[1] = pc(); 1304 1305 //----------------------- 1306 // all tests pass 1307 bind(succeed); 1308 1309 // Restore prior 64-bit registers 1310 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+0*8,O0); 1311 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+1*8,O1); 1312 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+2*8,O2); 1313 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+3*8,O3); 1314 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+4*8,O4); 1315 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+5*8,O5); 1316 1317 retl(); // Leaf return; restore prior O7 in delay slot 1318 delayed()->ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+7*8,O7); 1319 1320 //----------------------- 1321 bind(null_or_fail); // nulls are less common but OK 1322 br_null(O0_obj, false, pt, succeed); 1323 delayed()->wrccr( O5_save_flags ); // Restore CCR's 1324 1325 //----------------------- 1326 // report failure: 1327 bind(fail); 1328 _verify_oop_implicit_branch[2] = pc(); 1329 1330 wrccr( O5_save_flags ); // Restore CCR's 1331 1332 save_frame(align_up(sizeof(RegistersForDebugging) / BytesPerWord, 2)); 1333 1334 // stop_subroutine expects message pointer in I1. 1335 mov(I1, O1); 1336 1337 // Restore prior 64-bit registers 1338 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+0*8,I0); 1339 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+1*8,I1); 1340 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+2*8,I2); 1341 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+3*8,I3); 1342 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+4*8,I4); 1343 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+5*8,I5); 1344 1345 // factor long stop-sequence into subroutine to save space 1346 assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet"); 1347 1348 // call indirectly to solve generation ordering problem 1349 AddressLiteral al(StubRoutines::Sparc::stop_subroutine_entry_address()); 1350 load_ptr_contents(al, O5); 1351 jmpl(O5, 0, O7); 1352 delayed()->nop(); 1353 } 1354 1355 1356 void MacroAssembler::stop(const char* msg) { 1357 // save frame first to get O7 for return address 1358 // add one word to size in case struct is odd number of words long 1359 // It must be doubleword-aligned for storing doubles into it. 1360 1361 save_frame(align_up(sizeof(RegistersForDebugging) / BytesPerWord, 2)); 1362 1363 // stop_subroutine expects message pointer in I1. 1364 // Size of set() should stay the same 1365 patchable_set((intptr_t)msg, O1); 1366 1367 // factor long stop-sequence into subroutine to save space 1368 assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet"); 1369 1370 // call indirectly to solve generation ordering problem 1371 AddressLiteral a(StubRoutines::Sparc::stop_subroutine_entry_address()); 1372 load_ptr_contents(a, O5); 1373 jmpl(O5, 0, O7); 1374 delayed()->nop(); 1375 1376 breakpoint_trap(); // make stop actually stop rather than writing 1377 // unnoticeable results in the output files. 1378 1379 // restore(); done in callee to save space! 1380 } 1381 1382 1383 void MacroAssembler::warn(const char* msg) { 1384 save_frame(align_up(sizeof(RegistersForDebugging) / BytesPerWord, 2)); 1385 RegistersForDebugging::save_registers(this); 1386 mov(O0, L0); 1387 // Size of set() should stay the same 1388 patchable_set((intptr_t)msg, O0); 1389 call( CAST_FROM_FN_PTR(address, warning) ); 1390 delayed()->nop(); 1391 // ret(); 1392 // delayed()->restore(); 1393 RegistersForDebugging::restore_registers(this, L0); 1394 restore(); 1395 } 1396 1397 1398 void MacroAssembler::untested(const char* what) { 1399 // We must be able to turn interactive prompting off 1400 // in order to run automated test scripts on the VM 1401 // Use the flag ShowMessageBoxOnError 1402 1403 const char* b = NULL; 1404 { 1405 ResourceMark rm; 1406 stringStream ss; 1407 ss.print("untested: %s", what); 1408 b = code_string(ss.as_string()); 1409 } 1410 if (ShowMessageBoxOnError) { STOP(b); } 1411 else { warn(b); } 1412 } 1413 1414 1415 void MacroAssembler::unimplemented(const char* what) { 1416 const char* buf = NULL; 1417 { 1418 ResourceMark rm; 1419 stringStream ss; 1420 ss.print("unimplemented: %s", what); 1421 buf = code_string(ss.as_string()); 1422 } 1423 stop(buf); 1424 } 1425 1426 1427 void MacroAssembler::stop_subroutine() { 1428 RegistersForDebugging::save_registers(this); 1429 1430 // for the sake of the debugger, stick a PC on the current frame 1431 // (this assumes that the caller has performed an extra "save") 1432 mov(I7, L7); 1433 add(O7, -7 * BytesPerInt, I7); 1434 1435 save_frame(); // one more save to free up another O7 register 1436 mov(I0, O1); // addr of reg save area 1437 1438 // We expect pointer to message in I1. Caller must set it up in O1 1439 mov(I1, O0); // get msg 1440 call (CAST_FROM_FN_PTR(address, MacroAssembler::debug), relocInfo::runtime_call_type); 1441 delayed()->nop(); 1442 1443 restore(); 1444 1445 RegistersForDebugging::restore_registers(this, O0); 1446 1447 save_frame(0); 1448 call(CAST_FROM_FN_PTR(address,breakpoint)); 1449 delayed()->nop(); 1450 restore(); 1451 1452 mov(L7, I7); 1453 retl(); 1454 delayed()->restore(); // see stop above 1455 } 1456 1457 1458 void MacroAssembler::debug(char* msg, RegistersForDebugging* regs) { 1459 if ( ShowMessageBoxOnError ) { 1460 JavaThread* thread = JavaThread::current(); 1461 JavaThreadState saved_state = thread->thread_state(); 1462 thread->set_thread_state(_thread_in_vm); 1463 { 1464 // In order to get locks work, we need to fake a in_VM state 1465 ttyLocker ttyl; 1466 ::tty->print_cr("EXECUTION STOPPED: %s\n", msg); 1467 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 1468 BytecodeCounter::print(); 1469 } 1470 if (os::message_box(msg, "Execution stopped, print registers?")) 1471 regs->print(::tty); 1472 } 1473 BREAKPOINT; 1474 ThreadStateTransition::transition(JavaThread::current(), _thread_in_vm, saved_state); 1475 } 1476 else { 1477 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); 1478 } 1479 assert(false, "DEBUG MESSAGE: %s", msg); 1480 } 1481 1482 1483 void MacroAssembler::calc_mem_param_words(Register Rparam_words, Register Rresult) { 1484 subcc( Rparam_words, Argument::n_register_parameters, Rresult); // how many mem words? 1485 Label no_extras; 1486 br( negative, true, pt, no_extras ); // if neg, clear reg 1487 delayed()->set(0, Rresult); // annuled, so only if taken 1488 bind( no_extras ); 1489 } 1490 1491 1492 void MacroAssembler::calc_frame_size(Register Rextra_words, Register Rresult) { 1493 add(Rextra_words, frame::memory_parameter_word_sp_offset, Rresult); 1494 bclr(1, Rresult); 1495 sll(Rresult, LogBytesPerWord, Rresult); // Rresult has total frame bytes 1496 } 1497 1498 1499 void MacroAssembler::calc_frame_size_and_save(Register Rextra_words, Register Rresult) { 1500 calc_frame_size(Rextra_words, Rresult); 1501 neg(Rresult); 1502 save(SP, Rresult, SP); 1503 } 1504 1505 1506 // --------------------------------------------------------- 1507 Assembler::RCondition cond2rcond(Assembler::Condition c) { 1508 switch (c) { 1509 /*case zero: */ 1510 case Assembler::equal: return Assembler::rc_z; 1511 case Assembler::lessEqual: return Assembler::rc_lez; 1512 case Assembler::less: return Assembler::rc_lz; 1513 /*case notZero:*/ 1514 case Assembler::notEqual: return Assembler::rc_nz; 1515 case Assembler::greater: return Assembler::rc_gz; 1516 case Assembler::greaterEqual: return Assembler::rc_gez; 1517 } 1518 ShouldNotReachHere(); 1519 return Assembler::rc_z; 1520 } 1521 1522 // compares (32 bit) register with zero and branches. NOT FOR USE WITH 64-bit POINTERS 1523 void MacroAssembler::cmp_zero_and_br(Condition c, Register s1, Label& L, bool a, Predict p) { 1524 tst(s1); 1525 br (c, a, p, L); 1526 } 1527 1528 // Compares a pointer register with zero and branches on null. 1529 // Does a test & branch on 32-bit systems and a register-branch on 64-bit. 1530 void MacroAssembler::br_null( Register s1, bool a, Predict p, Label& L ) { 1531 assert_not_delayed(); 1532 bpr( rc_z, a, p, s1, L ); 1533 } 1534 1535 void MacroAssembler::br_notnull( Register s1, bool a, Predict p, Label& L ) { 1536 assert_not_delayed(); 1537 bpr( rc_nz, a, p, s1, L ); 1538 } 1539 1540 // Compare registers and branch with nop in delay slot or cbcond without delay slot. 1541 1542 // Compare integer (32 bit) values (icc only). 1543 void MacroAssembler::cmp_and_br_short(Register s1, Register s2, Condition c, 1544 Predict p, Label& L) { 1545 assert_not_delayed(); 1546 if (use_cbcond(L)) { 1547 Assembler::cbcond(c, icc, s1, s2, L); 1548 } else { 1549 cmp(s1, s2); 1550 br(c, false, p, L); 1551 delayed()->nop(); 1552 } 1553 } 1554 1555 // Compare integer (32 bit) values (icc only). 1556 void MacroAssembler::cmp_and_br_short(Register s1, int simm13a, Condition c, 1557 Predict p, Label& L) { 1558 assert_not_delayed(); 1559 if (is_simm(simm13a,5) && use_cbcond(L)) { 1560 Assembler::cbcond(c, icc, s1, simm13a, L); 1561 } else { 1562 cmp(s1, simm13a); 1563 br(c, false, p, L); 1564 delayed()->nop(); 1565 } 1566 } 1567 1568 // Branch that tests xcc in LP64 and icc in !LP64 1569 void MacroAssembler::cmp_and_brx_short(Register s1, Register s2, Condition c, 1570 Predict p, Label& L) { 1571 assert_not_delayed(); 1572 if (use_cbcond(L)) { 1573 Assembler::cbcond(c, ptr_cc, s1, s2, L); 1574 } else { 1575 cmp(s1, s2); 1576 brx(c, false, p, L); 1577 delayed()->nop(); 1578 } 1579 } 1580 1581 // Branch that tests xcc in LP64 and icc in !LP64 1582 void MacroAssembler::cmp_and_brx_short(Register s1, int simm13a, Condition c, 1583 Predict p, Label& L) { 1584 assert_not_delayed(); 1585 if (is_simm(simm13a,5) && use_cbcond(L)) { 1586 Assembler::cbcond(c, ptr_cc, s1, simm13a, L); 1587 } else { 1588 cmp(s1, simm13a); 1589 brx(c, false, p, L); 1590 delayed()->nop(); 1591 } 1592 } 1593 1594 // Short branch version for compares a pointer with zero. 1595 1596 void MacroAssembler::br_null_short(Register s1, Predict p, Label& L) { 1597 assert_not_delayed(); 1598 if (use_cbcond(L)) { 1599 Assembler::cbcond(zero, ptr_cc, s1, 0, L); 1600 } else { 1601 br_null(s1, false, p, L); 1602 delayed()->nop(); 1603 } 1604 } 1605 1606 void MacroAssembler::br_notnull_short(Register s1, Predict p, Label& L) { 1607 assert_not_delayed(); 1608 if (use_cbcond(L)) { 1609 Assembler::cbcond(notZero, ptr_cc, s1, 0, L); 1610 } else { 1611 br_notnull(s1, false, p, L); 1612 delayed()->nop(); 1613 } 1614 } 1615 1616 // Unconditional short branch 1617 void MacroAssembler::ba_short(Label& L) { 1618 assert_not_delayed(); 1619 if (use_cbcond(L)) { 1620 Assembler::cbcond(equal, icc, G0, G0, L); 1621 } else { 1622 br(always, false, pt, L); 1623 delayed()->nop(); 1624 } 1625 } 1626 1627 // Branch if 'icc' says zero or not (i.e. icc.z == 1|0). 1628 1629 void MacroAssembler::br_icc_zero(bool iszero, Predict p, Label &L) { 1630 assert_not_delayed(); 1631 Condition cf = (iszero ? Assembler::zero : Assembler::notZero); 1632 br(cf, false, p, L); 1633 delayed()->nop(); 1634 } 1635 1636 // instruction sequences factored across compiler & interpreter 1637 1638 1639 void MacroAssembler::lcmp( Register Ra_hi, Register Ra_low, 1640 Register Rb_hi, Register Rb_low, 1641 Register Rresult) { 1642 1643 Label check_low_parts, done; 1644 1645 cmp(Ra_hi, Rb_hi ); // compare hi parts 1646 br(equal, true, pt, check_low_parts); 1647 delayed()->cmp(Ra_low, Rb_low); // test low parts 1648 1649 // And, with an unsigned comparison, it does not matter if the numbers 1650 // are negative or not. 1651 // E.g., -2 cmp -1: the low parts are 0xfffffffe and 0xffffffff. 1652 // The second one is bigger (unsignedly). 1653 1654 // Other notes: The first move in each triplet can be unconditional 1655 // (and therefore probably prefetchable). 1656 // And the equals case for the high part does not need testing, 1657 // since that triplet is reached only after finding the high halves differ. 1658 1659 mov(-1, Rresult); 1660 ba(done); 1661 delayed()->movcc(greater, false, icc, 1, Rresult); 1662 1663 bind(check_low_parts); 1664 1665 mov( -1, Rresult); 1666 movcc(equal, false, icc, 0, Rresult); 1667 movcc(greaterUnsigned, false, icc, 1, Rresult); 1668 1669 bind(done); 1670 } 1671 1672 void MacroAssembler::lneg( Register Rhi, Register Rlow ) { 1673 subcc( G0, Rlow, Rlow ); 1674 subc( G0, Rhi, Rhi ); 1675 } 1676 1677 void MacroAssembler::lshl( Register Rin_high, Register Rin_low, 1678 Register Rcount, 1679 Register Rout_high, Register Rout_low, 1680 Register Rtemp ) { 1681 1682 1683 Register Ralt_count = Rtemp; 1684 Register Rxfer_bits = Rtemp; 1685 1686 assert( Ralt_count != Rin_high 1687 && Ralt_count != Rin_low 1688 && Ralt_count != Rcount 1689 && Rxfer_bits != Rin_low 1690 && Rxfer_bits != Rin_high 1691 && Rxfer_bits != Rcount 1692 && Rxfer_bits != Rout_low 1693 && Rout_low != Rin_high, 1694 "register alias checks"); 1695 1696 Label big_shift, done; 1697 1698 // This code can be optimized to use the 64 bit shifts in V9. 1699 // Here we use the 32 bit shifts. 1700 1701 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits 1702 subcc(Rcount, 31, Ralt_count); 1703 br(greater, true, pn, big_shift); 1704 delayed()->dec(Ralt_count); 1705 1706 // shift < 32 bits, Ralt_count = Rcount-31 1707 1708 // We get the transfer bits by shifting right by 32-count the low 1709 // register. This is done by shifting right by 31-count and then by one 1710 // more to take care of the special (rare) case where count is zero 1711 // (shifting by 32 would not work). 1712 1713 neg(Ralt_count); 1714 1715 // The order of the next two instructions is critical in the case where 1716 // Rin and Rout are the same and should not be reversed. 1717 1718 srl(Rin_low, Ralt_count, Rxfer_bits); // shift right by 31-count 1719 if (Rcount != Rout_low) { 1720 sll(Rin_low, Rcount, Rout_low); // low half 1721 } 1722 sll(Rin_high, Rcount, Rout_high); 1723 if (Rcount == Rout_low) { 1724 sll(Rin_low, Rcount, Rout_low); // low half 1725 } 1726 srl(Rxfer_bits, 1, Rxfer_bits ); // shift right by one more 1727 ba(done); 1728 delayed()->or3(Rout_high, Rxfer_bits, Rout_high); // new hi value: or in shifted old hi part and xfer from low 1729 1730 // shift >= 32 bits, Ralt_count = Rcount-32 1731 bind(big_shift); 1732 sll(Rin_low, Ralt_count, Rout_high ); 1733 clr(Rout_low); 1734 1735 bind(done); 1736 } 1737 1738 1739 void MacroAssembler::lshr( Register Rin_high, Register Rin_low, 1740 Register Rcount, 1741 Register Rout_high, Register Rout_low, 1742 Register Rtemp ) { 1743 1744 Register Ralt_count = Rtemp; 1745 Register Rxfer_bits = Rtemp; 1746 1747 assert( Ralt_count != Rin_high 1748 && Ralt_count != Rin_low 1749 && Ralt_count != Rcount 1750 && Rxfer_bits != Rin_low 1751 && Rxfer_bits != Rin_high 1752 && Rxfer_bits != Rcount 1753 && Rxfer_bits != Rout_high 1754 && Rout_high != Rin_low, 1755 "register alias checks"); 1756 1757 Label big_shift, done; 1758 1759 // This code can be optimized to use the 64 bit shifts in V9. 1760 // Here we use the 32 bit shifts. 1761 1762 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits 1763 subcc(Rcount, 31, Ralt_count); 1764 br(greater, true, pn, big_shift); 1765 delayed()->dec(Ralt_count); 1766 1767 // shift < 32 bits, Ralt_count = Rcount-31 1768 1769 // We get the transfer bits by shifting left by 32-count the high 1770 // register. This is done by shifting left by 31-count and then by one 1771 // more to take care of the special (rare) case where count is zero 1772 // (shifting by 32 would not work). 1773 1774 neg(Ralt_count); 1775 if (Rcount != Rout_low) { 1776 srl(Rin_low, Rcount, Rout_low); 1777 } 1778 1779 // The order of the next two instructions is critical in the case where 1780 // Rin and Rout are the same and should not be reversed. 1781 1782 sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count 1783 sra(Rin_high, Rcount, Rout_high ); // high half 1784 sll(Rxfer_bits, 1, Rxfer_bits); // shift left by one more 1785 if (Rcount == Rout_low) { 1786 srl(Rin_low, Rcount, Rout_low); 1787 } 1788 ba(done); 1789 delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high 1790 1791 // shift >= 32 bits, Ralt_count = Rcount-32 1792 bind(big_shift); 1793 1794 sra(Rin_high, Ralt_count, Rout_low); 1795 sra(Rin_high, 31, Rout_high); // sign into hi 1796 1797 bind( done ); 1798 } 1799 1800 1801 1802 void MacroAssembler::lushr( Register Rin_high, Register Rin_low, 1803 Register Rcount, 1804 Register Rout_high, Register Rout_low, 1805 Register Rtemp ) { 1806 1807 Register Ralt_count = Rtemp; 1808 Register Rxfer_bits = Rtemp; 1809 1810 assert( Ralt_count != Rin_high 1811 && Ralt_count != Rin_low 1812 && Ralt_count != Rcount 1813 && Rxfer_bits != Rin_low 1814 && Rxfer_bits != Rin_high 1815 && Rxfer_bits != Rcount 1816 && Rxfer_bits != Rout_high 1817 && Rout_high != Rin_low, 1818 "register alias checks"); 1819 1820 Label big_shift, done; 1821 1822 // This code can be optimized to use the 64 bit shifts in V9. 1823 // Here we use the 32 bit shifts. 1824 1825 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits 1826 subcc(Rcount, 31, Ralt_count); 1827 br(greater, true, pn, big_shift); 1828 delayed()->dec(Ralt_count); 1829 1830 // shift < 32 bits, Ralt_count = Rcount-31 1831 1832 // We get the transfer bits by shifting left by 32-count the high 1833 // register. This is done by shifting left by 31-count and then by one 1834 // more to take care of the special (rare) case where count is zero 1835 // (shifting by 32 would not work). 1836 1837 neg(Ralt_count); 1838 if (Rcount != Rout_low) { 1839 srl(Rin_low, Rcount, Rout_low); 1840 } 1841 1842 // The order of the next two instructions is critical in the case where 1843 // Rin and Rout are the same and should not be reversed. 1844 1845 sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count 1846 srl(Rin_high, Rcount, Rout_high ); // high half 1847 sll(Rxfer_bits, 1, Rxfer_bits); // shift left by one more 1848 if (Rcount == Rout_low) { 1849 srl(Rin_low, Rcount, Rout_low); 1850 } 1851 ba(done); 1852 delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high 1853 1854 // shift >= 32 bits, Ralt_count = Rcount-32 1855 bind(big_shift); 1856 1857 srl(Rin_high, Ralt_count, Rout_low); 1858 clr(Rout_high); 1859 1860 bind( done ); 1861 } 1862 1863 void MacroAssembler::lcmp( Register Ra, Register Rb, Register Rresult) { 1864 cmp(Ra, Rb); 1865 mov(-1, Rresult); 1866 movcc(equal, false, xcc, 0, Rresult); 1867 movcc(greater, false, xcc, 1, Rresult); 1868 } 1869 1870 1871 void MacroAssembler::load_sized_value(Address src, Register dst, size_t size_in_bytes, bool is_signed) { 1872 switch (size_in_bytes) { 1873 case 8: ld_long(src, dst); break; 1874 case 4: ld( src, dst); break; 1875 case 2: is_signed ? ldsh(src, dst) : lduh(src, dst); break; 1876 case 1: is_signed ? ldsb(src, dst) : ldub(src, dst); break; 1877 default: ShouldNotReachHere(); 1878 } 1879 } 1880 1881 void MacroAssembler::store_sized_value(Register src, Address dst, size_t size_in_bytes) { 1882 switch (size_in_bytes) { 1883 case 8: st_long(src, dst); break; 1884 case 4: st( src, dst); break; 1885 case 2: sth( src, dst); break; 1886 case 1: stb( src, dst); break; 1887 default: ShouldNotReachHere(); 1888 } 1889 } 1890 1891 1892 void MacroAssembler::float_cmp( bool is_float, int unordered_result, 1893 FloatRegister Fa, FloatRegister Fb, 1894 Register Rresult) { 1895 if (is_float) { 1896 fcmp(FloatRegisterImpl::S, fcc0, Fa, Fb); 1897 } else { 1898 fcmp(FloatRegisterImpl::D, fcc0, Fa, Fb); 1899 } 1900 1901 if (unordered_result == 1) { 1902 mov( -1, Rresult); 1903 movcc(f_equal, true, fcc0, 0, Rresult); 1904 movcc(f_unorderedOrGreater, true, fcc0, 1, Rresult); 1905 } else { 1906 mov( -1, Rresult); 1907 movcc(f_equal, true, fcc0, 0, Rresult); 1908 movcc(f_greater, true, fcc0, 1, Rresult); 1909 } 1910 } 1911 1912 1913 void MacroAssembler::save_all_globals_into_locals() { 1914 mov(G1,L1); 1915 mov(G2,L2); 1916 mov(G3,L3); 1917 mov(G4,L4); 1918 mov(G5,L5); 1919 mov(G6,L6); 1920 mov(G7,L7); 1921 } 1922 1923 void MacroAssembler::restore_globals_from_locals() { 1924 mov(L1,G1); 1925 mov(L2,G2); 1926 mov(L3,G3); 1927 mov(L4,G4); 1928 mov(L5,G5); 1929 mov(L6,G6); 1930 mov(L7,G7); 1931 } 1932 1933 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 1934 Register tmp, 1935 int offset) { 1936 intptr_t value = *delayed_value_addr; 1937 if (value != 0) 1938 return RegisterOrConstant(value + offset); 1939 1940 // load indirectly to solve generation ordering problem 1941 AddressLiteral a(delayed_value_addr); 1942 load_ptr_contents(a, tmp); 1943 1944 #ifdef ASSERT 1945 tst(tmp); 1946 breakpoint_trap(zero, xcc); 1947 #endif 1948 1949 if (offset != 0) 1950 add(tmp, offset, tmp); 1951 1952 return RegisterOrConstant(tmp); 1953 } 1954 1955 1956 RegisterOrConstant MacroAssembler::regcon_andn_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) { 1957 assert(d.register_or_noreg() != G0, "lost side effect"); 1958 if ((s2.is_constant() && s2.as_constant() == 0) || 1959 (s2.is_register() && s2.as_register() == G0)) { 1960 // Do nothing, just move value. 1961 if (s1.is_register()) { 1962 if (d.is_constant()) d = temp; 1963 mov(s1.as_register(), d.as_register()); 1964 return d; 1965 } else { 1966 return s1; 1967 } 1968 } 1969 1970 if (s1.is_register()) { 1971 assert_different_registers(s1.as_register(), temp); 1972 if (d.is_constant()) d = temp; 1973 andn(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register()); 1974 return d; 1975 } else { 1976 if (s2.is_register()) { 1977 assert_different_registers(s2.as_register(), temp); 1978 if (d.is_constant()) d = temp; 1979 set(s1.as_constant(), temp); 1980 andn(temp, s2.as_register(), d.as_register()); 1981 return d; 1982 } else { 1983 intptr_t res = s1.as_constant() & ~s2.as_constant(); 1984 return res; 1985 } 1986 } 1987 } 1988 1989 RegisterOrConstant MacroAssembler::regcon_inc_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) { 1990 assert(d.register_or_noreg() != G0, "lost side effect"); 1991 if ((s2.is_constant() && s2.as_constant() == 0) || 1992 (s2.is_register() && s2.as_register() == G0)) { 1993 // Do nothing, just move value. 1994 if (s1.is_register()) { 1995 if (d.is_constant()) d = temp; 1996 mov(s1.as_register(), d.as_register()); 1997 return d; 1998 } else { 1999 return s1; 2000 } 2001 } 2002 2003 if (s1.is_register()) { 2004 assert_different_registers(s1.as_register(), temp); 2005 if (d.is_constant()) d = temp; 2006 add(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register()); 2007 return d; 2008 } else { 2009 if (s2.is_register()) { 2010 assert_different_registers(s2.as_register(), temp); 2011 if (d.is_constant()) d = temp; 2012 add(s2.as_register(), ensure_simm13_or_reg(s1, temp), d.as_register()); 2013 return d; 2014 } else { 2015 intptr_t res = s1.as_constant() + s2.as_constant(); 2016 return res; 2017 } 2018 } 2019 } 2020 2021 RegisterOrConstant MacroAssembler::regcon_sll_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) { 2022 assert(d.register_or_noreg() != G0, "lost side effect"); 2023 if (!is_simm13(s2.constant_or_zero())) 2024 s2 = (s2.as_constant() & 0xFF); 2025 if ((s2.is_constant() && s2.as_constant() == 0) || 2026 (s2.is_register() && s2.as_register() == G0)) { 2027 // Do nothing, just move value. 2028 if (s1.is_register()) { 2029 if (d.is_constant()) d = temp; 2030 mov(s1.as_register(), d.as_register()); 2031 return d; 2032 } else { 2033 return s1; 2034 } 2035 } 2036 2037 if (s1.is_register()) { 2038 assert_different_registers(s1.as_register(), temp); 2039 if (d.is_constant()) d = temp; 2040 sll_ptr(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register()); 2041 return d; 2042 } else { 2043 if (s2.is_register()) { 2044 assert_different_registers(s2.as_register(), temp); 2045 if (d.is_constant()) d = temp; 2046 set(s1.as_constant(), temp); 2047 sll_ptr(temp, s2.as_register(), d.as_register()); 2048 return d; 2049 } else { 2050 intptr_t res = s1.as_constant() << s2.as_constant(); 2051 return res; 2052 } 2053 } 2054 } 2055 2056 2057 // Look up the method for a megamorphic invokeinterface call. 2058 // The target method is determined by <intf_klass, itable_index>. 2059 // The receiver klass is in recv_klass. 2060 // On success, the result will be in method_result, and execution falls through. 2061 // On failure, execution transfers to the given label. 2062 void MacroAssembler::lookup_interface_method(Register recv_klass, 2063 Register intf_klass, 2064 RegisterOrConstant itable_index, 2065 Register method_result, 2066 Register scan_temp, 2067 Register sethi_temp, 2068 Label& L_no_such_interface, 2069 bool return_method) { 2070 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); 2071 assert(!return_method || itable_index.is_constant() || itable_index.as_register() == method_result, 2072 "caller must use same register for non-constant itable index as for method"); 2073 2074 Label L_no_such_interface_restore; 2075 bool did_save = false; 2076 if (scan_temp == noreg || sethi_temp == noreg) { 2077 Register recv_2 = recv_klass->is_global() ? recv_klass : L0; 2078 Register intf_2 = intf_klass->is_global() ? intf_klass : L1; 2079 assert(method_result->is_global(), "must be able to return value"); 2080 scan_temp = L2; 2081 sethi_temp = L3; 2082 save_frame_and_mov(0, recv_klass, recv_2, intf_klass, intf_2); 2083 recv_klass = recv_2; 2084 intf_klass = intf_2; 2085 did_save = true; 2086 } 2087 2088 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 2089 int vtable_base = in_bytes(Klass::vtable_start_offset()); 2090 int scan_step = itableOffsetEntry::size() * wordSize; 2091 int vte_size = vtableEntry::size_in_bytes(); 2092 2093 lduw(recv_klass, in_bytes(Klass::vtable_length_offset()), scan_temp); 2094 // %%% We should store the aligned, prescaled offset in the klassoop. 2095 // Then the next several instructions would fold away. 2096 2097 int itb_offset = vtable_base; 2098 int itb_scale = exact_log2(vtableEntry::size_in_bytes()); 2099 sll(scan_temp, itb_scale, scan_temp); 2100 add(scan_temp, itb_offset, scan_temp); 2101 add(recv_klass, scan_temp, scan_temp); 2102 2103 if (return_method) { 2104 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 2105 RegisterOrConstant itable_offset = itable_index; 2106 itable_offset = regcon_sll_ptr(itable_index, exact_log2(itableMethodEntry::size() * wordSize), itable_offset); 2107 itable_offset = regcon_inc_ptr(itable_offset, itableMethodEntry::method_offset_in_bytes(), itable_offset); 2108 add(recv_klass, ensure_simm13_or_reg(itable_offset, sethi_temp), recv_klass); 2109 } 2110 2111 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 2112 // if (scan->interface() == intf) { 2113 // result = (klass + scan->offset() + itable_index); 2114 // } 2115 // } 2116 Label L_search, L_found_method; 2117 2118 for (int peel = 1; peel >= 0; peel--) { 2119 // %%%% Could load both offset and interface in one ldx, if they were 2120 // in the opposite order. This would save a load. 2121 ld_ptr(scan_temp, itableOffsetEntry::interface_offset_in_bytes(), method_result); 2122 2123 // Check that this entry is non-null. A null entry means that 2124 // the receiver class doesn't implement the interface, and wasn't the 2125 // same as when the caller was compiled. 2126 bpr(Assembler::rc_z, false, Assembler::pn, method_result, did_save ? L_no_such_interface_restore : L_no_such_interface); 2127 delayed()->cmp(method_result, intf_klass); 2128 2129 if (peel) { 2130 brx(Assembler::equal, false, Assembler::pt, L_found_method); 2131 } else { 2132 brx(Assembler::notEqual, false, Assembler::pn, L_search); 2133 // (invert the test to fall through to found_method...) 2134 } 2135 delayed()->add(scan_temp, scan_step, scan_temp); 2136 2137 if (!peel) break; 2138 2139 bind(L_search); 2140 } 2141 2142 bind(L_found_method); 2143 2144 if (return_method) { 2145 // Got a hit. 2146 int ito_offset = itableOffsetEntry::offset_offset_in_bytes(); 2147 // scan_temp[-scan_step] points to the vtable offset we need 2148 ito_offset -= scan_step; 2149 lduw(scan_temp, ito_offset, scan_temp); 2150 ld_ptr(recv_klass, scan_temp, method_result); 2151 } 2152 2153 if (did_save) { 2154 Label L_done; 2155 ba(L_done); 2156 delayed()->restore(); 2157 2158 bind(L_no_such_interface_restore); 2159 ba(L_no_such_interface); 2160 delayed()->restore(); 2161 2162 bind(L_done); 2163 } 2164 } 2165 2166 2167 // virtual method calling 2168 void MacroAssembler::lookup_virtual_method(Register recv_klass, 2169 RegisterOrConstant vtable_index, 2170 Register method_result) { 2171 assert_different_registers(recv_klass, method_result, vtable_index.register_or_noreg()); 2172 Register sethi_temp = method_result; 2173 const int base = in_bytes(Klass::vtable_start_offset()) + 2174 // method pointer offset within the vtable entry: 2175 vtableEntry::method_offset_in_bytes(); 2176 RegisterOrConstant vtable_offset = vtable_index; 2177 // Each of the following three lines potentially generates an instruction. 2178 // But the total number of address formation instructions will always be 2179 // at most two, and will often be zero. In any case, it will be optimal. 2180 // If vtable_index is a register, we will have (sll_ptr N,x; inc_ptr B,x; ld_ptr k,x). 2181 // If vtable_index is a constant, we will have at most (set B+X<<N,t; ld_ptr k,t). 2182 vtable_offset = regcon_sll_ptr(vtable_index, exact_log2(vtableEntry::size_in_bytes()), vtable_offset); 2183 vtable_offset = regcon_inc_ptr(vtable_offset, base, vtable_offset, sethi_temp); 2184 Address vtable_entry_addr(recv_klass, ensure_simm13_or_reg(vtable_offset, sethi_temp)); 2185 ld_ptr(vtable_entry_addr, method_result); 2186 } 2187 2188 2189 void MacroAssembler::check_klass_subtype(Register sub_klass, 2190 Register super_klass, 2191 Register temp_reg, 2192 Register temp2_reg, 2193 Label& L_success) { 2194 Register sub_2 = sub_klass; 2195 Register sup_2 = super_klass; 2196 if (!sub_2->is_global()) sub_2 = L0; 2197 if (!sup_2->is_global()) sup_2 = L1; 2198 bool did_save = false; 2199 if (temp_reg == noreg || temp2_reg == noreg) { 2200 temp_reg = L2; 2201 temp2_reg = L3; 2202 save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2); 2203 sub_klass = sub_2; 2204 super_klass = sup_2; 2205 did_save = true; 2206 } 2207 Label L_failure, L_pop_to_failure, L_pop_to_success; 2208 check_klass_subtype_fast_path(sub_klass, super_klass, 2209 temp_reg, temp2_reg, 2210 (did_save ? &L_pop_to_success : &L_success), 2211 (did_save ? &L_pop_to_failure : &L_failure), NULL); 2212 2213 if (!did_save) 2214 save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2); 2215 check_klass_subtype_slow_path(sub_2, sup_2, 2216 L2, L3, L4, L5, 2217 NULL, &L_pop_to_failure); 2218 2219 // on success: 2220 bind(L_pop_to_success); 2221 restore(); 2222 ba_short(L_success); 2223 2224 // on failure: 2225 bind(L_pop_to_failure); 2226 restore(); 2227 bind(L_failure); 2228 } 2229 2230 2231 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 2232 Register super_klass, 2233 Register temp_reg, 2234 Register temp2_reg, 2235 Label* L_success, 2236 Label* L_failure, 2237 Label* L_slow_path, 2238 RegisterOrConstant super_check_offset) { 2239 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 2240 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 2241 2242 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 2243 bool need_slow_path = (must_load_sco || 2244 super_check_offset.constant_or_zero() == sco_offset); 2245 2246 assert_different_registers(sub_klass, super_klass, temp_reg); 2247 if (super_check_offset.is_register()) { 2248 assert_different_registers(sub_klass, super_klass, temp_reg, 2249 super_check_offset.as_register()); 2250 } else if (must_load_sco) { 2251 assert(temp2_reg != noreg, "supply either a temp or a register offset"); 2252 } 2253 2254 Label L_fallthrough; 2255 int label_nulls = 0; 2256 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 2257 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 2258 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 2259 assert(label_nulls <= 1 || 2260 (L_slow_path == &L_fallthrough && label_nulls <= 2 && !need_slow_path), 2261 "at most one NULL in the batch, usually"); 2262 2263 // If the pointers are equal, we are done (e.g., String[] elements). 2264 // This self-check enables sharing of secondary supertype arrays among 2265 // non-primary types such as array-of-interface. Otherwise, each such 2266 // type would need its own customized SSA. 2267 // We move this check to the front of the fast path because many 2268 // type checks are in fact trivially successful in this manner, 2269 // so we get a nicely predicted branch right at the start of the check. 2270 cmp(super_klass, sub_klass); 2271 brx(Assembler::equal, false, Assembler::pn, *L_success); 2272 delayed()->nop(); 2273 2274 // Check the supertype display: 2275 if (must_load_sco) { 2276 // The super check offset is always positive... 2277 lduw(super_klass, sco_offset, temp2_reg); 2278 super_check_offset = RegisterOrConstant(temp2_reg); 2279 // super_check_offset is register. 2280 assert_different_registers(sub_klass, super_klass, temp_reg, super_check_offset.as_register()); 2281 } 2282 ld_ptr(sub_klass, super_check_offset, temp_reg); 2283 cmp(super_klass, temp_reg); 2284 2285 // This check has worked decisively for primary supers. 2286 // Secondary supers are sought in the super_cache ('super_cache_addr'). 2287 // (Secondary supers are interfaces and very deeply nested subtypes.) 2288 // This works in the same check above because of a tricky aliasing 2289 // between the super_cache and the primary super display elements. 2290 // (The 'super_check_addr' can address either, as the case requires.) 2291 // Note that the cache is updated below if it does not help us find 2292 // what we need immediately. 2293 // So if it was a primary super, we can just fail immediately. 2294 // Otherwise, it's the slow path for us (no success at this point). 2295 2296 // Hacked ba(), which may only be used just before L_fallthrough. 2297 #define FINAL_JUMP(label) \ 2298 if (&(label) != &L_fallthrough) { \ 2299 ba(label); delayed()->nop(); \ 2300 } 2301 2302 if (super_check_offset.is_register()) { 2303 brx(Assembler::equal, false, Assembler::pn, *L_success); 2304 delayed()->cmp(super_check_offset.as_register(), sc_offset); 2305 2306 if (L_failure == &L_fallthrough) { 2307 brx(Assembler::equal, false, Assembler::pt, *L_slow_path); 2308 delayed()->nop(); 2309 } else { 2310 brx(Assembler::notEqual, false, Assembler::pn, *L_failure); 2311 delayed()->nop(); 2312 FINAL_JUMP(*L_slow_path); 2313 } 2314 } else if (super_check_offset.as_constant() == sc_offset) { 2315 // Need a slow path; fast failure is impossible. 2316 if (L_slow_path == &L_fallthrough) { 2317 brx(Assembler::equal, false, Assembler::pt, *L_success); 2318 delayed()->nop(); 2319 } else { 2320 brx(Assembler::notEqual, false, Assembler::pn, *L_slow_path); 2321 delayed()->nop(); 2322 FINAL_JUMP(*L_success); 2323 } 2324 } else { 2325 // No slow path; it's a fast decision. 2326 if (L_failure == &L_fallthrough) { 2327 brx(Assembler::equal, false, Assembler::pt, *L_success); 2328 delayed()->nop(); 2329 } else { 2330 brx(Assembler::notEqual, false, Assembler::pn, *L_failure); 2331 delayed()->nop(); 2332 FINAL_JUMP(*L_success); 2333 } 2334 } 2335 2336 bind(L_fallthrough); 2337 2338 #undef FINAL_JUMP 2339 } 2340 2341 2342 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 2343 Register super_klass, 2344 Register count_temp, 2345 Register scan_temp, 2346 Register scratch_reg, 2347 Register coop_reg, 2348 Label* L_success, 2349 Label* L_failure) { 2350 assert_different_registers(sub_klass, super_klass, 2351 count_temp, scan_temp, scratch_reg, coop_reg); 2352 2353 Label L_fallthrough, L_loop; 2354 int label_nulls = 0; 2355 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 2356 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 2357 assert(label_nulls <= 1, "at most one NULL in the batch"); 2358 2359 // a couple of useful fields in sub_klass: 2360 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 2361 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 2362 2363 // Do a linear scan of the secondary super-klass chain. 2364 // This code is rarely used, so simplicity is a virtue here. 2365 2366 #ifndef PRODUCT 2367 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 2368 inc_counter((address) pst_counter, count_temp, scan_temp); 2369 #endif 2370 2371 // We will consult the secondary-super array. 2372 ld_ptr(sub_klass, ss_offset, scan_temp); 2373 2374 Register search_key = super_klass; 2375 2376 // Load the array length. (Positive movl does right thing on LP64.) 2377 lduw(scan_temp, Array<Klass*>::length_offset_in_bytes(), count_temp); 2378 2379 // Check for empty secondary super list 2380 tst(count_temp); 2381 2382 // In the array of super classes elements are pointer sized. 2383 int element_size = wordSize; 2384 2385 // Top of search loop 2386 bind(L_loop); 2387 br(Assembler::equal, false, Assembler::pn, *L_failure); 2388 delayed()->add(scan_temp, element_size, scan_temp); 2389 2390 // Skip the array header in all array accesses. 2391 int elem_offset = Array<Klass*>::base_offset_in_bytes(); 2392 elem_offset -= element_size; // the scan pointer was pre-incremented also 2393 2394 // Load next super to check 2395 ld_ptr( scan_temp, elem_offset, scratch_reg ); 2396 2397 // Look for Rsuper_klass on Rsub_klass's secondary super-class-overflow list 2398 cmp(scratch_reg, search_key); 2399 2400 // A miss means we are NOT a subtype and need to keep looping 2401 brx(Assembler::notEqual, false, Assembler::pn, L_loop); 2402 delayed()->deccc(count_temp); // decrement trip counter in delay slot 2403 2404 // Success. Cache the super we found and proceed in triumph. 2405 st_ptr(super_klass, sub_klass, sc_offset); 2406 2407 if (L_success != &L_fallthrough) { 2408 ba(*L_success); 2409 delayed()->nop(); 2410 } 2411 2412 bind(L_fallthrough); 2413 } 2414 2415 2416 RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot, 2417 Register temp_reg, 2418 int extra_slot_offset) { 2419 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 2420 int stackElementSize = Interpreter::stackElementSize; 2421 int offset = extra_slot_offset * stackElementSize; 2422 if (arg_slot.is_constant()) { 2423 offset += arg_slot.as_constant() * stackElementSize; 2424 return offset; 2425 } else { 2426 assert(temp_reg != noreg, "must specify"); 2427 sll_ptr(arg_slot.as_register(), exact_log2(stackElementSize), temp_reg); 2428 if (offset != 0) 2429 add(temp_reg, offset, temp_reg); 2430 return temp_reg; 2431 } 2432 } 2433 2434 2435 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 2436 Register temp_reg, 2437 int extra_slot_offset) { 2438 return Address(Gargs, argument_offset(arg_slot, temp_reg, extra_slot_offset)); 2439 } 2440 2441 2442 void MacroAssembler::biased_locking_enter(Register obj_reg, Register mark_reg, 2443 Register temp_reg, 2444 Label& done, Label* slow_case, 2445 BiasedLockingCounters* counters) { 2446 assert(UseBiasedLocking, "why call this otherwise?"); 2447 2448 if (PrintBiasedLockingStatistics) { 2449 assert_different_registers(obj_reg, mark_reg, temp_reg, O7); 2450 if (counters == NULL) 2451 counters = BiasedLocking::counters(); 2452 } 2453 2454 Label cas_label; 2455 2456 // Biased locking 2457 // See whether the lock is currently biased toward our thread and 2458 // whether the epoch is still valid 2459 // Note that the runtime guarantees sufficient alignment of JavaThread 2460 // pointers to allow age to be placed into low bits 2461 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 2462 and3(mark_reg, markOopDesc::biased_lock_mask_in_place, temp_reg); 2463 cmp_and_brx_short(temp_reg, markOopDesc::biased_lock_pattern, Assembler::notEqual, Assembler::pn, cas_label); 2464 2465 load_klass(obj_reg, temp_reg); 2466 ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg); 2467 or3(G2_thread, temp_reg, temp_reg); 2468 xor3(mark_reg, temp_reg, temp_reg); 2469 andcc(temp_reg, ~((int) markOopDesc::age_mask_in_place), temp_reg); 2470 if (counters != NULL) { 2471 cond_inc(Assembler::equal, (address) counters->biased_lock_entry_count_addr(), mark_reg, temp_reg); 2472 // Reload mark_reg as we may need it later 2473 ld_ptr(Address(obj_reg, oopDesc::mark_offset_in_bytes()), mark_reg); 2474 } 2475 brx(Assembler::equal, true, Assembler::pt, done); 2476 delayed()->nop(); 2477 2478 Label try_revoke_bias; 2479 Label try_rebias; 2480 Address mark_addr = Address(obj_reg, oopDesc::mark_offset_in_bytes()); 2481 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 2482 2483 // At this point we know that the header has the bias pattern and 2484 // that we are not the bias owner in the current epoch. We need to 2485 // figure out more details about the state of the header in order to 2486 // know what operations can be legally performed on the object's 2487 // header. 2488 2489 // If the low three bits in the xor result aren't clear, that means 2490 // the prototype header is no longer biased and we have to revoke 2491 // the bias on this object. 2492 btst(markOopDesc::biased_lock_mask_in_place, temp_reg); 2493 brx(Assembler::notZero, false, Assembler::pn, try_revoke_bias); 2494 2495 // Biasing is still enabled for this data type. See whether the 2496 // epoch of the current bias is still valid, meaning that the epoch 2497 // bits of the mark word are equal to the epoch bits of the 2498 // prototype header. (Note that the prototype header's epoch bits 2499 // only change at a safepoint.) If not, attempt to rebias the object 2500 // toward the current thread. Note that we must be absolutely sure 2501 // that the current epoch is invalid in order to do this because 2502 // otherwise the manipulations it performs on the mark word are 2503 // illegal. 2504 delayed()->btst(markOopDesc::epoch_mask_in_place, temp_reg); 2505 brx(Assembler::notZero, false, Assembler::pn, try_rebias); 2506 2507 // The epoch of the current bias is still valid but we know nothing 2508 // about the owner; it might be set or it might be clear. Try to 2509 // acquire the bias of the object using an atomic operation. If this 2510 // fails we will go in to the runtime to revoke the object's bias. 2511 // Note that we first construct the presumed unbiased header so we 2512 // don't accidentally blow away another thread's valid bias. 2513 delayed()->and3(mark_reg, 2514 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place, 2515 mark_reg); 2516 or3(G2_thread, mark_reg, temp_reg); 2517 cas_ptr(mark_addr.base(), mark_reg, temp_reg); 2518 // If the biasing toward our thread failed, this means that 2519 // another thread succeeded in biasing it toward itself and we 2520 // need to revoke that bias. The revocation will occur in the 2521 // interpreter runtime in the slow case. 2522 cmp(mark_reg, temp_reg); 2523 if (counters != NULL) { 2524 cond_inc(Assembler::zero, (address) counters->anonymously_biased_lock_entry_count_addr(), mark_reg, temp_reg); 2525 } 2526 if (slow_case != NULL) { 2527 brx(Assembler::notEqual, true, Assembler::pn, *slow_case); 2528 delayed()->nop(); 2529 } 2530 ba_short(done); 2531 2532 bind(try_rebias); 2533 // At this point we know the epoch has expired, meaning that the 2534 // current "bias owner", if any, is actually invalid. Under these 2535 // circumstances _only_, we are allowed to use the current header's 2536 // value as the comparison value when doing the cas to acquire the 2537 // bias in the current epoch. In other words, we allow transfer of 2538 // the bias from one thread to another directly in this situation. 2539 // 2540 // FIXME: due to a lack of registers we currently blow away the age 2541 // bits in this situation. Should attempt to preserve them. 2542 load_klass(obj_reg, temp_reg); 2543 ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg); 2544 or3(G2_thread, temp_reg, temp_reg); 2545 cas_ptr(mark_addr.base(), mark_reg, temp_reg); 2546 // If the biasing toward our thread failed, this means that 2547 // another thread succeeded in biasing it toward itself and we 2548 // need to revoke that bias. The revocation will occur in the 2549 // interpreter runtime in the slow case. 2550 cmp(mark_reg, temp_reg); 2551 if (counters != NULL) { 2552 cond_inc(Assembler::zero, (address) counters->rebiased_lock_entry_count_addr(), mark_reg, temp_reg); 2553 } 2554 if (slow_case != NULL) { 2555 brx(Assembler::notEqual, true, Assembler::pn, *slow_case); 2556 delayed()->nop(); 2557 } 2558 ba_short(done); 2559 2560 bind(try_revoke_bias); 2561 // The prototype mark in the klass doesn't have the bias bit set any 2562 // more, indicating that objects of this data type are not supposed 2563 // to be biased any more. We are going to try to reset the mark of 2564 // this object to the prototype value and fall through to the 2565 // CAS-based locking scheme. Note that if our CAS fails, it means 2566 // that another thread raced us for the privilege of revoking the 2567 // bias of this particular object, so it's okay to continue in the 2568 // normal locking code. 2569 // 2570 // FIXME: due to a lack of registers we currently blow away the age 2571 // bits in this situation. Should attempt to preserve them. 2572 load_klass(obj_reg, temp_reg); 2573 ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg); 2574 cas_ptr(mark_addr.base(), mark_reg, temp_reg); 2575 // Fall through to the normal CAS-based lock, because no matter what 2576 // the result of the above CAS, some thread must have succeeded in 2577 // removing the bias bit from the object's header. 2578 if (counters != NULL) { 2579 cmp(mark_reg, temp_reg); 2580 cond_inc(Assembler::zero, (address) counters->revoked_lock_entry_count_addr(), mark_reg, temp_reg); 2581 } 2582 2583 bind(cas_label); 2584 } 2585 2586 void MacroAssembler::biased_locking_exit (Address mark_addr, Register temp_reg, Label& done, 2587 bool allow_delay_slot_filling) { 2588 // Check for biased locking unlock case, which is a no-op 2589 // Note: we do not have to check the thread ID for two reasons. 2590 // First, the interpreter checks for IllegalMonitorStateException at 2591 // a higher level. Second, if the bias was revoked while we held the 2592 // lock, the object could not be rebiased toward another thread, so 2593 // the bias bit would be clear. 2594 ld_ptr(mark_addr, temp_reg); 2595 and3(temp_reg, markOopDesc::biased_lock_mask_in_place, temp_reg); 2596 cmp(temp_reg, markOopDesc::biased_lock_pattern); 2597 brx(Assembler::equal, allow_delay_slot_filling, Assembler::pt, done); 2598 delayed(); 2599 if (!allow_delay_slot_filling) { 2600 nop(); 2601 } 2602 } 2603 2604 2605 // compiler_lock_object() and compiler_unlock_object() are direct transliterations 2606 // of i486.ad fast_lock() and fast_unlock(). See those methods for detailed comments. 2607 // The code could be tightened up considerably. 2608 // 2609 // box->dhw disposition - post-conditions at DONE_LABEL. 2610 // - Successful inflated lock: box->dhw != 0. 2611 // Any non-zero value suffices. 2612 // Consider G2_thread, rsp, boxReg, or markOopDesc::unused_mark() 2613 // - Successful Stack-lock: box->dhw == mark. 2614 // box->dhw must contain the displaced mark word value 2615 // - Failure -- icc.ZFlag == 0 and box->dhw is undefined. 2616 // The slow-path fast_enter() and slow_enter() operators 2617 // are responsible for setting box->dhw = NonZero (typically markOopDesc::unused_mark()). 2618 // - Biased: box->dhw is undefined 2619 // 2620 // SPARC refworkload performance - specifically jetstream and scimark - are 2621 // extremely sensitive to the size of the code emitted by compiler_lock_object 2622 // and compiler_unlock_object. Critically, the key factor is code size, not path 2623 // length. (Simply experiments to pad CLO with unexecuted NOPs demonstrte the 2624 // effect). 2625 2626 2627 void MacroAssembler::compiler_lock_object(Register Roop, Register Rmark, 2628 Register Rbox, Register Rscratch, 2629 BiasedLockingCounters* counters, 2630 bool try_bias) { 2631 Address mark_addr(Roop, oopDesc::mark_offset_in_bytes()); 2632 2633 verify_oop(Roop); 2634 Label done ; 2635 2636 if (counters != NULL) { 2637 inc_counter((address) counters->total_entry_count_addr(), Rmark, Rscratch); 2638 } 2639 2640 if (EmitSync & 1) { 2641 mov(3, Rscratch); 2642 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2643 cmp(SP, G0); 2644 return ; 2645 } 2646 2647 if (EmitSync & 2) { 2648 2649 // Fetch object's markword 2650 ld_ptr(mark_addr, Rmark); 2651 2652 if (try_bias) { 2653 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters); 2654 } 2655 2656 // Save Rbox in Rscratch to be used for the cas operation 2657 mov(Rbox, Rscratch); 2658 2659 // set Rmark to markOop | markOopDesc::unlocked_value 2660 or3(Rmark, markOopDesc::unlocked_value, Rmark); 2661 2662 // Initialize the box. (Must happen before we update the object mark!) 2663 st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2664 2665 // compare object markOop with Rmark and if equal exchange Rscratch with object markOop 2666 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 2667 cas_ptr(mark_addr.base(), Rmark, Rscratch); 2668 2669 // if compare/exchange succeeded we found an unlocked object and we now have locked it 2670 // hence we are done 2671 cmp(Rmark, Rscratch); 2672 sub(Rscratch, STACK_BIAS, Rscratch); 2673 brx(Assembler::equal, false, Assembler::pt, done); 2674 delayed()->sub(Rscratch, SP, Rscratch); //pull next instruction into delay slot 2675 2676 // we did not find an unlocked object so see if this is a recursive case 2677 // sub(Rscratch, SP, Rscratch); 2678 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant"); 2679 andcc(Rscratch, 0xfffff003, Rscratch); 2680 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2681 bind (done); 2682 return ; 2683 } 2684 2685 Label Egress ; 2686 2687 if (EmitSync & 256) { 2688 Label IsInflated ; 2689 2690 ld_ptr(mark_addr, Rmark); // fetch obj->mark 2691 // Triage: biased, stack-locked, neutral, inflated 2692 if (try_bias) { 2693 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters); 2694 // Invariant: if control reaches this point in the emitted stream 2695 // then Rmark has not been modified. 2696 } 2697 2698 // Store mark into displaced mark field in the on-stack basic-lock "box" 2699 // Critically, this must happen before the CAS 2700 // Maximize the ST-CAS distance to minimize the ST-before-CAS penalty. 2701 st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2702 andcc(Rmark, 2, G0); 2703 brx(Assembler::notZero, false, Assembler::pn, IsInflated); 2704 delayed()-> 2705 2706 // Try stack-lock acquisition. 2707 // Beware: the 1st instruction is in a delay slot 2708 mov(Rbox, Rscratch); 2709 or3(Rmark, markOopDesc::unlocked_value, Rmark); 2710 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 2711 cas_ptr(mark_addr.base(), Rmark, Rscratch); 2712 cmp(Rmark, Rscratch); 2713 brx(Assembler::equal, false, Assembler::pt, done); 2714 delayed()->sub(Rscratch, SP, Rscratch); 2715 2716 // Stack-lock attempt failed - check for recursive stack-lock. 2717 // See the comments below about how we might remove this case. 2718 sub(Rscratch, STACK_BIAS, Rscratch); 2719 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant"); 2720 andcc(Rscratch, 0xfffff003, Rscratch); 2721 br(Assembler::always, false, Assembler::pt, done); 2722 delayed()-> st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2723 2724 bind(IsInflated); 2725 if (EmitSync & 64) { 2726 // If m->owner != null goto IsLocked 2727 // Pessimistic form: Test-and-CAS vs CAS 2728 // The optimistic form avoids RTS->RTO cache line upgrades. 2729 ld_ptr(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rscratch); 2730 andcc(Rscratch, Rscratch, G0); 2731 brx(Assembler::notZero, false, Assembler::pn, done); 2732 delayed()->nop(); 2733 // m->owner == null : it's unlocked. 2734 } 2735 2736 // Try to CAS m->owner from null to Self 2737 // Invariant: if we acquire the lock then _recursions should be 0. 2738 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark); 2739 mov(G2_thread, Rscratch); 2740 cas_ptr(Rmark, G0, Rscratch); 2741 cmp(Rscratch, G0); 2742 // Intentional fall-through into done 2743 } else { 2744 // Aggressively avoid the Store-before-CAS penalty 2745 // Defer the store into box->dhw until after the CAS 2746 Label IsInflated, Recursive ; 2747 2748 // Anticipate CAS -- Avoid RTS->RTO upgrade 2749 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads); 2750 2751 ld_ptr(mark_addr, Rmark); // fetch obj->mark 2752 // Triage: biased, stack-locked, neutral, inflated 2753 2754 if (try_bias) { 2755 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters); 2756 // Invariant: if control reaches this point in the emitted stream 2757 // then Rmark has not been modified. 2758 } 2759 andcc(Rmark, 2, G0); 2760 brx(Assembler::notZero, false, Assembler::pn, IsInflated); 2761 delayed()-> // Beware - dangling delay-slot 2762 2763 // Try stack-lock acquisition. 2764 // Transiently install BUSY (0) encoding in the mark word. 2765 // if the CAS of 0 into the mark was successful then we execute: 2766 // ST box->dhw = mark -- save fetched mark in on-stack basiclock box 2767 // ST obj->mark = box -- overwrite transient 0 value 2768 // This presumes TSO, of course. 2769 2770 mov(0, Rscratch); 2771 or3(Rmark, markOopDesc::unlocked_value, Rmark); 2772 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 2773 cas_ptr(mark_addr.base(), Rmark, Rscratch); 2774 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads); 2775 cmp(Rscratch, Rmark); 2776 brx(Assembler::notZero, false, Assembler::pn, Recursive); 2777 delayed()->st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2778 if (counters != NULL) { 2779 cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch); 2780 } 2781 ba(done); 2782 delayed()->st_ptr(Rbox, mark_addr); 2783 2784 bind(Recursive); 2785 // Stack-lock attempt failed - check for recursive stack-lock. 2786 // Tests show that we can remove the recursive case with no impact 2787 // on refworkload 0.83. If we need to reduce the size of the code 2788 // emitted by compiler_lock_object() the recursive case is perfect 2789 // candidate. 2790 // 2791 // A more extreme idea is to always inflate on stack-lock recursion. 2792 // This lets us eliminate the recursive checks in compiler_lock_object 2793 // and compiler_unlock_object and the (box->dhw == 0) encoding. 2794 // A brief experiment - requiring changes to synchronizer.cpp, interpreter, 2795 // and showed a performance *increase*. In the same experiment I eliminated 2796 // the fast-path stack-lock code from the interpreter and always passed 2797 // control to the "slow" operators in synchronizer.cpp. 2798 2799 // RScratch contains the fetched obj->mark value from the failed CAS. 2800 sub(Rscratch, STACK_BIAS, Rscratch); 2801 sub(Rscratch, SP, Rscratch); 2802 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant"); 2803 andcc(Rscratch, 0xfffff003, Rscratch); 2804 if (counters != NULL) { 2805 // Accounting needs the Rscratch register 2806 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2807 cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch); 2808 ba_short(done); 2809 } else { 2810 ba(done); 2811 delayed()->st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2812 } 2813 2814 bind (IsInflated); 2815 2816 // Try to CAS m->owner from null to Self 2817 // Invariant: if we acquire the lock then _recursions should be 0. 2818 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark); 2819 mov(G2_thread, Rscratch); 2820 cas_ptr(Rmark, G0, Rscratch); 2821 andcc(Rscratch, Rscratch, G0); // set ICCs for done: icc.zf iff success 2822 // set icc.zf : 1=success 0=failure 2823 // ST box->displaced_header = NonZero. 2824 // Any non-zero value suffices: 2825 // markOopDesc::unused_mark(), G2_thread, RBox, RScratch, rsp, etc. 2826 st_ptr(Rbox, Rbox, BasicLock::displaced_header_offset_in_bytes()); 2827 // Intentional fall-through into done 2828 } 2829 2830 bind (done); 2831 } 2832 2833 void MacroAssembler::compiler_unlock_object(Register Roop, Register Rmark, 2834 Register Rbox, Register Rscratch, 2835 bool try_bias) { 2836 Address mark_addr(Roop, oopDesc::mark_offset_in_bytes()); 2837 2838 Label done ; 2839 2840 if (EmitSync & 4) { 2841 cmp(SP, G0); 2842 return ; 2843 } 2844 2845 if (EmitSync & 8) { 2846 if (try_bias) { 2847 biased_locking_exit(mark_addr, Rscratch, done); 2848 } 2849 2850 // Test first if it is a fast recursive unlock 2851 ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rmark); 2852 br_null_short(Rmark, Assembler::pt, done); 2853 2854 // Check if it is still a light weight lock, this is is true if we see 2855 // the stack address of the basicLock in the markOop of the object 2856 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 2857 cas_ptr(mark_addr.base(), Rbox, Rmark); 2858 ba(done); 2859 delayed()->cmp(Rbox, Rmark); 2860 bind(done); 2861 return ; 2862 } 2863 2864 // Beware ... If the aggregate size of the code emitted by CLO and CUO is 2865 // is too large performance rolls abruptly off a cliff. 2866 // This could be related to inlining policies, code cache management, or 2867 // I$ effects. 2868 Label LStacked ; 2869 2870 if (try_bias) { 2871 // TODO: eliminate redundant LDs of obj->mark 2872 biased_locking_exit(mark_addr, Rscratch, done); 2873 } 2874 2875 ld_ptr(Roop, oopDesc::mark_offset_in_bytes(), Rmark); 2876 ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rscratch); 2877 andcc(Rscratch, Rscratch, G0); 2878 brx(Assembler::zero, false, Assembler::pn, done); 2879 delayed()->nop(); // consider: relocate fetch of mark, above, into this DS 2880 andcc(Rmark, 2, G0); 2881 brx(Assembler::zero, false, Assembler::pt, LStacked); 2882 delayed()->nop(); 2883 2884 // It's inflated 2885 // Conceptually we need a #loadstore|#storestore "release" MEMBAR before 2886 // the ST of 0 into _owner which releases the lock. This prevents loads 2887 // and stores within the critical section from reordering (floating) 2888 // past the store that releases the lock. But TSO is a strong memory model 2889 // and that particular flavor of barrier is a noop, so we can safely elide it. 2890 // Note that we use 1-0 locking by default for the inflated case. We 2891 // close the resultant (and rare) race by having contended threads in 2892 // monitorenter periodically poll _owner. 2893 2894 if (EmitSync & 1024) { 2895 // Emit code to check that _owner == Self 2896 // We could fold the _owner test into subsequent code more efficiently 2897 // than using a stand-alone check, but since _owner checking is off by 2898 // default we don't bother. We also might consider predicating the 2899 // _owner==Self check on Xcheck:jni or running on a debug build. 2900 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), Rscratch); 2901 orcc(Rscratch, G0, G0); 2902 brx(Assembler::notZero, false, Assembler::pn, done); 2903 delayed()->nop(); 2904 } 2905 2906 if (EmitSync & 512) { 2907 // classic lock release code absent 1-0 locking 2908 // m->Owner = null; 2909 // membar #storeload 2910 // if (m->cxq|m->EntryList) == null goto Success 2911 // if (m->succ != null) goto Success 2912 // if CAS (&m->Owner,0,Self) != 0 goto Success 2913 // goto SlowPath 2914 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)), Rbox); 2915 orcc(Rbox, G0, G0); 2916 brx(Assembler::notZero, false, Assembler::pn, done); 2917 delayed()->nop(); 2918 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2919 if (os::is_MP()) { membar(StoreLoad); } 2920 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)), Rscratch); 2921 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)), Rbox); 2922 orcc(Rbox, Rscratch, G0); 2923 brx(Assembler::zero, false, Assembler::pt, done); 2924 delayed()-> 2925 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), Rscratch); 2926 andcc(Rscratch, Rscratch, G0); 2927 brx(Assembler::notZero, false, Assembler::pt, done); 2928 delayed()->andcc(G0, G0, G0); 2929 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark); 2930 mov(G2_thread, Rscratch); 2931 cas_ptr(Rmark, G0, Rscratch); 2932 cmp(Rscratch, G0); 2933 // invert icc.zf and goto done 2934 brx(Assembler::notZero, false, Assembler::pt, done); 2935 delayed()->cmp(G0, G0); 2936 br(Assembler::always, false, Assembler::pt, done); 2937 delayed()->cmp(G0, 1); 2938 } else { 2939 // 1-0 form : avoids CAS and MEMBAR in the common case 2940 // Do not bother to ratify that m->Owner == Self. 2941 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)), Rbox); 2942 orcc(Rbox, G0, G0); 2943 brx(Assembler::notZero, false, Assembler::pn, done); 2944 delayed()-> 2945 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)), Rscratch); 2946 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)), Rbox); 2947 orcc(Rbox, Rscratch, G0); 2948 if (EmitSync & 16384) { 2949 // As an optional optimization, if (EntryList|cxq) != null and _succ is null then 2950 // we should transfer control directly to the slow-path. 2951 // This test makes the reacquire operation below very infrequent. 2952 // The logic is equivalent to : 2953 // if (cxq|EntryList) == null : Owner=null; goto Success 2954 // if succ == null : goto SlowPath 2955 // Owner=null; membar #storeload 2956 // if succ != null : goto Success 2957 // if CAS(&Owner,null,Self) != null goto Success 2958 // goto SlowPath 2959 brx(Assembler::zero, true, Assembler::pt, done); 2960 delayed()-> 2961 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2962 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), Rscratch); 2963 andcc(Rscratch, Rscratch, G0) ; 2964 brx(Assembler::zero, false, Assembler::pt, done); 2965 delayed()->orcc(G0, 1, G0); 2966 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2967 } else { 2968 brx(Assembler::zero, false, Assembler::pt, done); 2969 delayed()-> 2970 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2971 } 2972 if (os::is_MP()) { membar(StoreLoad); } 2973 // Check that _succ is (or remains) non-zero 2974 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), Rscratch); 2975 andcc(Rscratch, Rscratch, G0); 2976 brx(Assembler::notZero, false, Assembler::pt, done); 2977 delayed()->andcc(G0, G0, G0); 2978 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark); 2979 mov(G2_thread, Rscratch); 2980 cas_ptr(Rmark, G0, Rscratch); 2981 cmp(Rscratch, G0); 2982 // invert icc.zf and goto done 2983 // A slightly better v8+/v9 idiom would be the following: 2984 // movrnz Rscratch,1,Rscratch 2985 // ba done 2986 // xorcc Rscratch,1,G0 2987 // In v8+ mode the idiom would be valid IFF Rscratch was a G or O register 2988 brx(Assembler::notZero, false, Assembler::pt, done); 2989 delayed()->cmp(G0, G0); 2990 br(Assembler::always, false, Assembler::pt, done); 2991 delayed()->cmp(G0, 1); 2992 } 2993 2994 bind (LStacked); 2995 // Consider: we could replace the expensive CAS in the exit 2996 // path with a simple ST of the displaced mark value fetched from 2997 // the on-stack basiclock box. That admits a race where a thread T2 2998 // in the slow lock path -- inflating with monitor M -- could race a 2999 // thread T1 in the fast unlock path, resulting in a missed wakeup for T2. 3000 // More precisely T1 in the stack-lock unlock path could "stomp" the 3001 // inflated mark value M installed by T2, resulting in an orphan 3002 // object monitor M and T2 becoming stranded. We can remedy that situation 3003 // by having T2 periodically poll the object's mark word using timed wait 3004 // operations. If T2 discovers that a stomp has occurred it vacates 3005 // the monitor M and wakes any other threads stranded on the now-orphan M. 3006 // In addition the monitor scavenger, which performs deflation, 3007 // would also need to check for orpan monitors and stranded threads. 3008 // 3009 // Finally, inflation is also used when T2 needs to assign a hashCode 3010 // to O and O is stack-locked by T1. The "stomp" race could cause 3011 // an assigned hashCode value to be lost. We can avoid that condition 3012 // and provide the necessary hashCode stability invariants by ensuring 3013 // that hashCode generation is idempotent between copying GCs. 3014 // For example we could compute the hashCode of an object O as 3015 // O's heap address XOR some high quality RNG value that is refreshed 3016 // at GC-time. The monitor scavenger would install the hashCode 3017 // found in any orphan monitors. Again, the mechanism admits a 3018 // lost-update "stomp" WAW race but detects and recovers as needed. 3019 // 3020 // A prototype implementation showed excellent results, although 3021 // the scavenger and timeout code was rather involved. 3022 3023 cas_ptr(mark_addr.base(), Rbox, Rscratch); 3024 cmp(Rbox, Rscratch); 3025 // Intentional fall through into done ... 3026 3027 bind(done); 3028 } 3029 3030 3031 3032 void MacroAssembler::print_CPU_state() { 3033 // %%%%% need to implement this 3034 } 3035 3036 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 3037 // %%%%% need to implement this 3038 } 3039 3040 void MacroAssembler::push_IU_state() { 3041 // %%%%% need to implement this 3042 } 3043 3044 3045 void MacroAssembler::pop_IU_state() { 3046 // %%%%% need to implement this 3047 } 3048 3049 3050 void MacroAssembler::push_FPU_state() { 3051 // %%%%% need to implement this 3052 } 3053 3054 3055 void MacroAssembler::pop_FPU_state() { 3056 // %%%%% need to implement this 3057 } 3058 3059 3060 void MacroAssembler::push_CPU_state() { 3061 // %%%%% need to implement this 3062 } 3063 3064 3065 void MacroAssembler::pop_CPU_state() { 3066 // %%%%% need to implement this 3067 } 3068 3069 3070 3071 void MacroAssembler::verify_tlab() { 3072 #ifdef ASSERT 3073 if (UseTLAB && VerifyOops) { 3074 Label next, next2, ok; 3075 Register t1 = L0; 3076 Register t2 = L1; 3077 Register t3 = L2; 3078 3079 save_frame(0); 3080 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1); 3081 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t2); 3082 or3(t1, t2, t3); 3083 cmp_and_br_short(t1, t2, Assembler::greaterEqual, Assembler::pn, next); 3084 STOP("assert(top >= start)"); 3085 should_not_reach_here(); 3086 3087 bind(next); 3088 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1); 3089 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t2); 3090 or3(t3, t2, t3); 3091 cmp_and_br_short(t1, t2, Assembler::lessEqual, Assembler::pn, next2); 3092 STOP("assert(top <= end)"); 3093 should_not_reach_here(); 3094 3095 bind(next2); 3096 and3(t3, MinObjAlignmentInBytesMask, t3); 3097 cmp_and_br_short(t3, 0, Assembler::lessEqual, Assembler::pn, ok); 3098 STOP("assert(aligned)"); 3099 should_not_reach_here(); 3100 3101 bind(ok); 3102 restore(); 3103 } 3104 #endif 3105 } 3106 3107 3108 void MacroAssembler::eden_allocate( 3109 Register obj, // result: pointer to object after successful allocation 3110 Register var_size_in_bytes, // object size in bytes if unknown at compile time; invalid otherwise 3111 int con_size_in_bytes, // object size in bytes if known at compile time 3112 Register t1, // temp register 3113 Register t2, // temp register 3114 Label& slow_case // continuation point if fast allocation fails 3115 ){ 3116 // make sure arguments make sense 3117 assert_different_registers(obj, var_size_in_bytes, t1, t2); 3118 assert(0 <= con_size_in_bytes && Assembler::is_simm13(con_size_in_bytes), "illegal object size"); 3119 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment"); 3120 3121 if (!Universe::heap()->supports_inline_contig_alloc()) { 3122 // No allocation in the shared eden. 3123 ba(slow_case); 3124 delayed()->nop(); 3125 } else { 3126 // get eden boundaries 3127 // note: we need both top & top_addr! 3128 const Register top_addr = t1; 3129 const Register end = t2; 3130 3131 CollectedHeap* ch = Universe::heap(); 3132 set((intx)ch->top_addr(), top_addr); 3133 intx delta = (intx)ch->end_addr() - (intx)ch->top_addr(); 3134 ld_ptr(top_addr, delta, end); 3135 ld_ptr(top_addr, 0, obj); 3136 3137 // try to allocate 3138 Label retry; 3139 bind(retry); 3140 #ifdef ASSERT 3141 // make sure eden top is properly aligned 3142 { 3143 Label L; 3144 btst(MinObjAlignmentInBytesMask, obj); 3145 br(Assembler::zero, false, Assembler::pt, L); 3146 delayed()->nop(); 3147 STOP("eden top is not properly aligned"); 3148 bind(L); 3149 } 3150 #endif // ASSERT 3151 const Register free = end; 3152 sub(end, obj, free); // compute amount of free space 3153 if (var_size_in_bytes->is_valid()) { 3154 // size is unknown at compile time 3155 cmp(free, var_size_in_bytes); 3156 brx(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case 3157 delayed()->add(obj, var_size_in_bytes, end); 3158 } else { 3159 // size is known at compile time 3160 cmp(free, con_size_in_bytes); 3161 brx(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case 3162 delayed()->add(obj, con_size_in_bytes, end); 3163 } 3164 // Compare obj with the value at top_addr; if still equal, swap the value of 3165 // end with the value at top_addr. If not equal, read the value at top_addr 3166 // into end. 3167 cas_ptr(top_addr, obj, end); 3168 // if someone beat us on the allocation, try again, otherwise continue 3169 cmp(obj, end); 3170 brx(Assembler::notEqual, false, Assembler::pn, retry); 3171 delayed()->mov(end, obj); // nop if successfull since obj == end 3172 3173 #ifdef ASSERT 3174 // make sure eden top is properly aligned 3175 { 3176 Label L; 3177 const Register top_addr = t1; 3178 3179 set((intx)ch->top_addr(), top_addr); 3180 ld_ptr(top_addr, 0, top_addr); 3181 btst(MinObjAlignmentInBytesMask, top_addr); 3182 br(Assembler::zero, false, Assembler::pt, L); 3183 delayed()->nop(); 3184 STOP("eden top is not properly aligned"); 3185 bind(L); 3186 } 3187 #endif // ASSERT 3188 } 3189 } 3190 3191 3192 void MacroAssembler::tlab_allocate( 3193 Register obj, // result: pointer to object after successful allocation 3194 Register var_size_in_bytes, // object size in bytes if unknown at compile time; invalid otherwise 3195 int con_size_in_bytes, // object size in bytes if known at compile time 3196 Register t1, // temp register 3197 Label& slow_case // continuation point if fast allocation fails 3198 ){ 3199 // make sure arguments make sense 3200 assert_different_registers(obj, var_size_in_bytes, t1); 3201 assert(0 <= con_size_in_bytes && is_simm13(con_size_in_bytes), "illegal object size"); 3202 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment"); 3203 3204 const Register free = t1; 3205 3206 verify_tlab(); 3207 3208 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), obj); 3209 3210 // calculate amount of free space 3211 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), free); 3212 sub(free, obj, free); 3213 3214 Label done; 3215 if (var_size_in_bytes == noreg) { 3216 cmp(free, con_size_in_bytes); 3217 } else { 3218 cmp(free, var_size_in_bytes); 3219 } 3220 br(Assembler::less, false, Assembler::pn, slow_case); 3221 // calculate the new top pointer 3222 if (var_size_in_bytes == noreg) { 3223 delayed()->add(obj, con_size_in_bytes, free); 3224 } else { 3225 delayed()->add(obj, var_size_in_bytes, free); 3226 } 3227 3228 bind(done); 3229 3230 #ifdef ASSERT 3231 // make sure new free pointer is properly aligned 3232 { 3233 Label L; 3234 btst(MinObjAlignmentInBytesMask, free); 3235 br(Assembler::zero, false, Assembler::pt, L); 3236 delayed()->nop(); 3237 STOP("updated TLAB free is not properly aligned"); 3238 bind(L); 3239 } 3240 #endif // ASSERT 3241 3242 // update the tlab top pointer 3243 st_ptr(free, G2_thread, in_bytes(JavaThread::tlab_top_offset())); 3244 verify_tlab(); 3245 } 3246 3247 void MacroAssembler::zero_memory(Register base, Register index) { 3248 assert_different_registers(base, index); 3249 Label loop; 3250 bind(loop); 3251 subcc(index, HeapWordSize, index); 3252 brx(Assembler::greaterEqual, true, Assembler::pt, loop); 3253 delayed()->st_ptr(G0, base, index); 3254 } 3255 3256 void MacroAssembler::incr_allocated_bytes(RegisterOrConstant size_in_bytes, 3257 Register t1, Register t2) { 3258 // Bump total bytes allocated by this thread 3259 assert(t1->is_global(), "must be global reg"); // so all 64 bits are saved on a context switch 3260 assert_different_registers(size_in_bytes.register_or_noreg(), t1, t2); 3261 // v8 support has gone the way of the dodo 3262 ldx(G2_thread, in_bytes(JavaThread::allocated_bytes_offset()), t1); 3263 add(t1, ensure_simm13_or_reg(size_in_bytes, t2), t1); 3264 stx(t1, G2_thread, in_bytes(JavaThread::allocated_bytes_offset())); 3265 } 3266 3267 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 3268 switch (cond) { 3269 // Note some conditions are synonyms for others 3270 case Assembler::never: return Assembler::always; 3271 case Assembler::zero: return Assembler::notZero; 3272 case Assembler::lessEqual: return Assembler::greater; 3273 case Assembler::less: return Assembler::greaterEqual; 3274 case Assembler::lessEqualUnsigned: return Assembler::greaterUnsigned; 3275 case Assembler::lessUnsigned: return Assembler::greaterEqualUnsigned; 3276 case Assembler::negative: return Assembler::positive; 3277 case Assembler::overflowSet: return Assembler::overflowClear; 3278 case Assembler::always: return Assembler::never; 3279 case Assembler::notZero: return Assembler::zero; 3280 case Assembler::greater: return Assembler::lessEqual; 3281 case Assembler::greaterEqual: return Assembler::less; 3282 case Assembler::greaterUnsigned: return Assembler::lessEqualUnsigned; 3283 case Assembler::greaterEqualUnsigned: return Assembler::lessUnsigned; 3284 case Assembler::positive: return Assembler::negative; 3285 case Assembler::overflowClear: return Assembler::overflowSet; 3286 } 3287 3288 ShouldNotReachHere(); return Assembler::overflowClear; 3289 } 3290 3291 void MacroAssembler::cond_inc(Assembler::Condition cond, address counter_ptr, 3292 Register Rtmp1, Register Rtmp2 /*, Register Rtmp3, Register Rtmp4 */) { 3293 Condition negated_cond = negate_condition(cond); 3294 Label L; 3295 brx(negated_cond, false, Assembler::pt, L); 3296 delayed()->nop(); 3297 inc_counter(counter_ptr, Rtmp1, Rtmp2); 3298 bind(L); 3299 } 3300 3301 void MacroAssembler::inc_counter(address counter_addr, Register Rtmp1, Register Rtmp2) { 3302 AddressLiteral addrlit(counter_addr); 3303 sethi(addrlit, Rtmp1); // Move hi22 bits into temporary register. 3304 Address addr(Rtmp1, addrlit.low10()); // Build an address with low10 bits. 3305 ld(addr, Rtmp2); 3306 inc(Rtmp2); 3307 st(Rtmp2, addr); 3308 } 3309 3310 void MacroAssembler::inc_counter(int* counter_addr, Register Rtmp1, Register Rtmp2) { 3311 inc_counter((address) counter_addr, Rtmp1, Rtmp2); 3312 } 3313 3314 SkipIfEqual::SkipIfEqual( 3315 MacroAssembler* masm, Register temp, const bool* flag_addr, 3316 Assembler::Condition condition) { 3317 _masm = masm; 3318 AddressLiteral flag(flag_addr); 3319 _masm->sethi(flag, temp); 3320 _masm->ldub(temp, flag.low10(), temp); 3321 _masm->tst(temp); 3322 _masm->br(condition, false, Assembler::pt, _label); 3323 _masm->delayed()->nop(); 3324 } 3325 3326 SkipIfEqual::~SkipIfEqual() { 3327 _masm->bind(_label); 3328 } 3329 3330 3331 // Writes to stack successive pages until offset reached to check for 3332 // stack overflow + shadow pages. This clobbers tsp and scratch. 3333 void MacroAssembler::bang_stack_size(Register Rsize, Register Rtsp, 3334 Register Rscratch) { 3335 // Use stack pointer in temp stack pointer 3336 mov(SP, Rtsp); 3337 3338 // Bang stack for total size given plus stack shadow page size. 3339 // Bang one page at a time because a large size can overflow yellow and 3340 // red zones (the bang will fail but stack overflow handling can't tell that 3341 // it was a stack overflow bang vs a regular segv). 3342 int offset = os::vm_page_size(); 3343 Register Roffset = Rscratch; 3344 3345 Label loop; 3346 bind(loop); 3347 set((-offset)+STACK_BIAS, Rscratch); 3348 st(G0, Rtsp, Rscratch); 3349 set(offset, Roffset); 3350 sub(Rsize, Roffset, Rsize); 3351 cmp(Rsize, G0); 3352 br(Assembler::greater, false, Assembler::pn, loop); 3353 delayed()->sub(Rtsp, Roffset, Rtsp); 3354 3355 // Bang down shadow pages too. 3356 // At this point, (tmp-0) is the last address touched, so don't 3357 // touch it again. (It was touched as (tmp-pagesize) but then tmp 3358 // was post-decremented.) Skip this address by starting at i=1, and 3359 // touch a few more pages below. N.B. It is important to touch all 3360 // the way down to and including i=StackShadowPages. 3361 for (int i = 1; i < JavaThread::stack_shadow_zone_size() / os::vm_page_size(); i++) { 3362 set((-i*offset)+STACK_BIAS, Rscratch); 3363 st(G0, Rtsp, Rscratch); 3364 } 3365 } 3366 3367 void MacroAssembler::reserved_stack_check() { 3368 // testing if reserved zone needs to be enabled 3369 Label no_reserved_zone_enabling; 3370 3371 ld_ptr(G2_thread, JavaThread::reserved_stack_activation_offset(), G4_scratch); 3372 cmp_and_brx_short(SP, G4_scratch, Assembler::lessUnsigned, Assembler::pt, no_reserved_zone_enabling); 3373 3374 call_VM_leaf(L0, CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), G2_thread); 3375 3376 AddressLiteral stub(StubRoutines::throw_delayed_StackOverflowError_entry()); 3377 jump_to(stub, G4_scratch); 3378 delayed()->restore(); 3379 3380 should_not_reach_here(); 3381 3382 bind(no_reserved_zone_enabling); 3383 } 3384 3385 /////////////////////////////////////////////////////////////////////////////////// 3386 #if INCLUDE_ALL_GCS 3387 3388 static address satb_log_enqueue_with_frame = NULL; 3389 static u_char* satb_log_enqueue_with_frame_end = NULL; 3390 3391 static address satb_log_enqueue_frameless = NULL; 3392 static u_char* satb_log_enqueue_frameless_end = NULL; 3393 3394 static int EnqueueCodeSize = 128 DEBUG_ONLY( + 256); // Instructions? 3395 3396 static void generate_satb_log_enqueue(bool with_frame) { 3397 BufferBlob* bb = BufferBlob::create("enqueue_with_frame", EnqueueCodeSize); 3398 CodeBuffer buf(bb); 3399 MacroAssembler masm(&buf); 3400 3401 #define __ masm. 3402 3403 address start = __ pc(); 3404 Register pre_val; 3405 3406 Label refill, restart; 3407 if (with_frame) { 3408 __ save_frame(0); 3409 pre_val = I0; // Was O0 before the save. 3410 } else { 3411 pre_val = O0; 3412 } 3413 3414 int satb_q_index_byte_offset = 3415 in_bytes(JavaThread::satb_mark_queue_offset() + 3416 SATBMarkQueue::byte_offset_of_index()); 3417 3418 int satb_q_buf_byte_offset = 3419 in_bytes(JavaThread::satb_mark_queue_offset() + 3420 SATBMarkQueue::byte_offset_of_buf()); 3421 3422 assert(in_bytes(SATBMarkQueue::byte_width_of_index()) == sizeof(intptr_t) && 3423 in_bytes(SATBMarkQueue::byte_width_of_buf()) == sizeof(intptr_t), 3424 "check sizes in assembly below"); 3425 3426 __ bind(restart); 3427 3428 // Load the index into the SATB buffer. SATBMarkQueue::_index is a size_t 3429 // so ld_ptr is appropriate. 3430 __ ld_ptr(G2_thread, satb_q_index_byte_offset, L0); 3431 3432 // index == 0? 3433 __ cmp_and_brx_short(L0, G0, Assembler::equal, Assembler::pn, refill); 3434 3435 __ ld_ptr(G2_thread, satb_q_buf_byte_offset, L1); 3436 __ sub(L0, oopSize, L0); 3437 3438 __ st_ptr(pre_val, L1, L0); // [_buf + index] := I0 3439 if (!with_frame) { 3440 // Use return-from-leaf 3441 __ retl(); 3442 __ delayed()->st_ptr(L0, G2_thread, satb_q_index_byte_offset); 3443 } else { 3444 // Not delayed. 3445 __ st_ptr(L0, G2_thread, satb_q_index_byte_offset); 3446 } 3447 if (with_frame) { 3448 __ ret(); 3449 __ delayed()->restore(); 3450 } 3451 __ bind(refill); 3452 3453 address handle_zero = 3454 CAST_FROM_FN_PTR(address, 3455 &SATBMarkQueueSet::handle_zero_index_for_thread); 3456 // This should be rare enough that we can afford to save all the 3457 // scratch registers that the calling context might be using. 3458 __ mov(G1_scratch, L0); 3459 __ mov(G3_scratch, L1); 3460 __ mov(G4, L2); 3461 // We need the value of O0 above (for the write into the buffer), so we 3462 // save and restore it. 3463 __ mov(O0, L3); 3464 // Since the call will overwrite O7, we save and restore that, as well. 3465 __ mov(O7, L4); 3466 __ call_VM_leaf(L5, handle_zero, G2_thread); 3467 __ mov(L0, G1_scratch); 3468 __ mov(L1, G3_scratch); 3469 __ mov(L2, G4); 3470 __ mov(L3, O0); 3471 __ br(Assembler::always, /*annul*/false, Assembler::pt, restart); 3472 __ delayed()->mov(L4, O7); 3473 3474 if (with_frame) { 3475 satb_log_enqueue_with_frame = start; 3476 satb_log_enqueue_with_frame_end = __ pc(); 3477 } else { 3478 satb_log_enqueue_frameless = start; 3479 satb_log_enqueue_frameless_end = __ pc(); 3480 } 3481 3482 #undef __ 3483 } 3484 3485 void MacroAssembler::g1_write_barrier_pre(Register obj, 3486 Register index, 3487 int offset, 3488 Register pre_val, 3489 Register tmp, 3490 bool preserve_o_regs) { 3491 Label filtered; 3492 3493 if (obj == noreg) { 3494 // We are not loading the previous value so make 3495 // sure that we don't trash the value in pre_val 3496 // with the code below. 3497 assert_different_registers(pre_val, tmp); 3498 } else { 3499 // We will be loading the previous value 3500 // in this code so... 3501 assert(offset == 0 || index == noreg, "choose one"); 3502 assert(pre_val == noreg, "check this code"); 3503 } 3504 3505 // Is marking active? 3506 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 3507 ld(G2, 3508 in_bytes(JavaThread::satb_mark_queue_offset() + 3509 SATBMarkQueue::byte_offset_of_active()), 3510 tmp); 3511 } else { 3512 guarantee(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, 3513 "Assumption"); 3514 ldsb(G2, 3515 in_bytes(JavaThread::satb_mark_queue_offset() + 3516 SATBMarkQueue::byte_offset_of_active()), 3517 tmp); 3518 } 3519 3520 // Is marking active? 3521 cmp_and_br_short(tmp, G0, Assembler::equal, Assembler::pt, filtered); 3522 3523 // Do we need to load the previous value? 3524 if (obj != noreg) { 3525 // Load the previous value... 3526 if (index == noreg) { 3527 if (Assembler::is_simm13(offset)) { 3528 load_heap_oop(obj, offset, tmp); 3529 } else { 3530 set(offset, tmp); 3531 load_heap_oop(obj, tmp, tmp); 3532 } 3533 } else { 3534 load_heap_oop(obj, index, tmp); 3535 } 3536 // Previous value has been loaded into tmp 3537 pre_val = tmp; 3538 } 3539 3540 assert(pre_val != noreg, "must have a real register"); 3541 3542 // Is the previous value null? 3543 cmp_and_brx_short(pre_val, G0, Assembler::equal, Assembler::pt, filtered); 3544 3545 // OK, it's not filtered, so we'll need to call enqueue. In the normal 3546 // case, pre_val will be a scratch G-reg, but there are some cases in 3547 // which it's an O-reg. In the first case, do a normal call. In the 3548 // latter, do a save here and call the frameless version. 3549 3550 guarantee(pre_val->is_global() || pre_val->is_out(), 3551 "Or we need to think harder."); 3552 3553 if (pre_val->is_global() && !preserve_o_regs) { 3554 call(satb_log_enqueue_with_frame); 3555 delayed()->mov(pre_val, O0); 3556 } else { 3557 save_frame(0); 3558 call(satb_log_enqueue_frameless); 3559 delayed()->mov(pre_val->after_save(), O0); 3560 restore(); 3561 } 3562 3563 bind(filtered); 3564 } 3565 3566 static address dirty_card_log_enqueue = 0; 3567 static u_char* dirty_card_log_enqueue_end = 0; 3568 3569 // This gets to assume that o0 contains the object address. 3570 static void generate_dirty_card_log_enqueue(jbyte* byte_map_base) { 3571 BufferBlob* bb = BufferBlob::create("dirty_card_enqueue", EnqueueCodeSize*2); 3572 CodeBuffer buf(bb); 3573 MacroAssembler masm(&buf); 3574 #define __ masm. 3575 address start = __ pc(); 3576 3577 Label not_already_dirty, restart, refill, young_card; 3578 3579 __ srlx(O0, CardTable::card_shift, O0); 3580 AddressLiteral addrlit(byte_map_base); 3581 __ set(addrlit, O1); // O1 := <card table base> 3582 __ ldub(O0, O1, O2); // O2 := [O0 + O1] 3583 3584 __ cmp_and_br_short(O2, G1CardTable::g1_young_card_val(), Assembler::equal, Assembler::pt, young_card); 3585 3586 __ membar(Assembler::Membar_mask_bits(Assembler::StoreLoad)); 3587 __ ldub(O0, O1, O2); // O2 := [O0 + O1] 3588 3589 assert(CardTable::dirty_card_val() == 0, "otherwise check this code"); 3590 __ cmp_and_br_short(O2, G0, Assembler::notEqual, Assembler::pt, not_already_dirty); 3591 3592 __ bind(young_card); 3593 // We didn't take the branch, so we're already dirty: return. 3594 // Use return-from-leaf 3595 __ retl(); 3596 __ delayed()->nop(); 3597 3598 // Not dirty. 3599 __ bind(not_already_dirty); 3600 3601 // Get O0 + O1 into a reg by itself 3602 __ add(O0, O1, O3); 3603 3604 // First, dirty it. 3605 __ stb(G0, O3, G0); // [cardPtr] := 0 (i.e., dirty). 3606 3607 int dirty_card_q_index_byte_offset = 3608 in_bytes(JavaThread::dirty_card_queue_offset() + 3609 DirtyCardQueue::byte_offset_of_index()); 3610 int dirty_card_q_buf_byte_offset = 3611 in_bytes(JavaThread::dirty_card_queue_offset() + 3612 DirtyCardQueue::byte_offset_of_buf()); 3613 __ bind(restart); 3614 3615 // Load the index into the update buffer. DirtyCardQueue::_index is 3616 // a size_t so ld_ptr is appropriate here. 3617 __ ld_ptr(G2_thread, dirty_card_q_index_byte_offset, L0); 3618 3619 // index == 0? 3620 __ cmp_and_brx_short(L0, G0, Assembler::equal, Assembler::pn, refill); 3621 3622 __ ld_ptr(G2_thread, dirty_card_q_buf_byte_offset, L1); 3623 __ sub(L0, oopSize, L0); 3624 3625 __ st_ptr(O3, L1, L0); // [_buf + index] := I0 3626 // Use return-from-leaf 3627 __ retl(); 3628 __ delayed()->st_ptr(L0, G2_thread, dirty_card_q_index_byte_offset); 3629 3630 __ bind(refill); 3631 address handle_zero = 3632 CAST_FROM_FN_PTR(address, 3633 &DirtyCardQueueSet::handle_zero_index_for_thread); 3634 // This should be rare enough that we can afford to save all the 3635 // scratch registers that the calling context might be using. 3636 __ mov(G1_scratch, L3); 3637 __ mov(G3_scratch, L5); 3638 // We need the value of O3 above (for the write into the buffer), so we 3639 // save and restore it. 3640 __ mov(O3, L6); 3641 // Since the call will overwrite O7, we save and restore that, as well. 3642 __ mov(O7, L4); 3643 3644 __ call_VM_leaf(L7_thread_cache, handle_zero, G2_thread); 3645 __ mov(L3, G1_scratch); 3646 __ mov(L5, G3_scratch); 3647 __ mov(L6, O3); 3648 __ br(Assembler::always, /*annul*/false, Assembler::pt, restart); 3649 __ delayed()->mov(L4, O7); 3650 3651 dirty_card_log_enqueue = start; 3652 dirty_card_log_enqueue_end = __ pc(); 3653 // XXX Should have a guarantee here about not going off the end! 3654 // Does it already do so? Do an experiment... 3655 3656 #undef __ 3657 3658 } 3659 3660 void MacroAssembler::g1_write_barrier_post(Register store_addr, Register new_val, Register tmp) { 3661 3662 Label filtered; 3663 MacroAssembler* post_filter_masm = this; 3664 3665 if (new_val == G0) return; 3666 3667 G1SATBCardTableLoggingModRefBS* bs = 3668 barrier_set_cast<G1SATBCardTableLoggingModRefBS>(Universe::heap()->barrier_set()); 3669 CardTable* ct = bs->card_table(); 3670 3671 if (G1RSBarrierRegionFilter) { 3672 xor3(store_addr, new_val, tmp); 3673 srlx(tmp, HeapRegion::LogOfHRGrainBytes, tmp); 3674 3675 // XXX Should I predict this taken or not? Does it matter? 3676 cmp_and_brx_short(tmp, G0, Assembler::equal, Assembler::pt, filtered); 3677 } 3678 3679 // If the "store_addr" register is an "in" or "local" register, move it to 3680 // a scratch reg so we can pass it as an argument. 3681 bool use_scr = !(store_addr->is_global() || store_addr->is_out()); 3682 // Pick a scratch register different from "tmp". 3683 Register scr = (tmp == G1_scratch ? G3_scratch : G1_scratch); 3684 // Make sure we use up the delay slot! 3685 if (use_scr) { 3686 post_filter_masm->mov(store_addr, scr); 3687 } else { 3688 post_filter_masm->nop(); 3689 } 3690 save_frame(0); 3691 call(dirty_card_log_enqueue); 3692 if (use_scr) { 3693 delayed()->mov(scr, O0); 3694 } else { 3695 delayed()->mov(store_addr->after_save(), O0); 3696 } 3697 restore(); 3698 3699 bind(filtered); 3700 } 3701 3702 // Called from init_globals() after universe_init() and before interpreter_init() 3703 void g1_barrier_stubs_init() { 3704 CollectedHeap* heap = Universe::heap(); 3705 if (heap->kind() == CollectedHeap::G1CollectedHeap) { 3706 // Only needed for G1 3707 if (dirty_card_log_enqueue == 0) { 3708 G1SATBCardTableLoggingModRefBS* bs = 3709 barrier_set_cast<G1SATBCardTableLoggingModRefBS>(heap->barrier_set()); 3710 CardTable *ct = bs->card_table(); 3711 generate_dirty_card_log_enqueue(ct->byte_map_base()); 3712 assert(dirty_card_log_enqueue != 0, "postcondition."); 3713 } 3714 if (satb_log_enqueue_with_frame == 0) { 3715 generate_satb_log_enqueue(true); 3716 assert(satb_log_enqueue_with_frame != 0, "postcondition."); 3717 } 3718 if (satb_log_enqueue_frameless == 0) { 3719 generate_satb_log_enqueue(false); 3720 assert(satb_log_enqueue_frameless != 0, "postcondition."); 3721 } 3722 } 3723 } 3724 3725 #endif // INCLUDE_ALL_GCS 3726 /////////////////////////////////////////////////////////////////////////////////// 3727 3728 void MacroAssembler::card_write_barrier_post(Register store_addr, Register new_val, Register tmp) { 3729 // If we're writing constant NULL, we can skip the write barrier. 3730 if (new_val == G0) return; 3731 CardTableModRefBS* bs = 3732 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3733 CardTable* ct = bs->card_table(); 3734 3735 assert(bs->kind() == BarrierSet::CardTableModRef, "wrong barrier"); 3736 card_table_write(ct->byte_map_base(), tmp, store_addr); 3737 } 3738 3739 // ((OopHandle)result).resolve(); 3740 void MacroAssembler::resolve_oop_handle(Register result) { 3741 // OopHandle::resolve is an indirection. 3742 ld_ptr(result, 0, result); 3743 } 3744 3745 void MacroAssembler::load_mirror(Register mirror, Register method) { 3746 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 3747 ld_ptr(method, in_bytes(Method::const_offset()), mirror); 3748 ld_ptr(mirror, in_bytes(ConstMethod::constants_offset()), mirror); 3749 ld_ptr(mirror, ConstantPool::pool_holder_offset_in_bytes(), mirror); 3750 ld_ptr(mirror, mirror_offset, mirror); 3751 resolve_oop_handle(mirror); 3752 } 3753 3754 void MacroAssembler::load_klass(Register src_oop, Register klass) { 3755 // The number of bytes in this code is used by 3756 // MachCallDynamicJavaNode::ret_addr_offset() 3757 // if this changes, change that. 3758 if (UseCompressedClassPointers) { 3759 lduw(src_oop, oopDesc::klass_offset_in_bytes(), klass); 3760 decode_klass_not_null(klass); 3761 } else { 3762 ld_ptr(src_oop, oopDesc::klass_offset_in_bytes(), klass); 3763 } 3764 } 3765 3766 void MacroAssembler::store_klass(Register klass, Register dst_oop) { 3767 if (UseCompressedClassPointers) { 3768 assert(dst_oop != klass, "not enough registers"); 3769 encode_klass_not_null(klass); 3770 st(klass, dst_oop, oopDesc::klass_offset_in_bytes()); 3771 } else { 3772 st_ptr(klass, dst_oop, oopDesc::klass_offset_in_bytes()); 3773 } 3774 } 3775 3776 void MacroAssembler::store_klass_gap(Register s, Register d) { 3777 if (UseCompressedClassPointers) { 3778 assert(s != d, "not enough registers"); 3779 st(s, d, oopDesc::klass_gap_offset_in_bytes()); 3780 } 3781 } 3782 3783 void MacroAssembler::load_heap_oop(const Address& s, Register d) { 3784 if (UseCompressedOops) { 3785 lduw(s, d); 3786 decode_heap_oop(d); 3787 } else { 3788 ld_ptr(s, d); 3789 } 3790 } 3791 3792 void MacroAssembler::load_heap_oop(Register s1, Register s2, Register d) { 3793 if (UseCompressedOops) { 3794 lduw(s1, s2, d); 3795 decode_heap_oop(d, d); 3796 } else { 3797 ld_ptr(s1, s2, d); 3798 } 3799 } 3800 3801 void MacroAssembler::load_heap_oop(Register s1, int simm13a, Register d) { 3802 if (UseCompressedOops) { 3803 lduw(s1, simm13a, d); 3804 decode_heap_oop(d, d); 3805 } else { 3806 ld_ptr(s1, simm13a, d); 3807 } 3808 } 3809 3810 void MacroAssembler::load_heap_oop(Register s1, RegisterOrConstant s2, Register d) { 3811 if (s2.is_constant()) load_heap_oop(s1, s2.as_constant(), d); 3812 else load_heap_oop(s1, s2.as_register(), d); 3813 } 3814 3815 void MacroAssembler::store_heap_oop(Register d, Register s1, Register s2) { 3816 if (UseCompressedOops) { 3817 assert(s1 != d && s2 != d, "not enough registers"); 3818 encode_heap_oop(d); 3819 st(d, s1, s2); 3820 } else { 3821 st_ptr(d, s1, s2); 3822 } 3823 } 3824 3825 void MacroAssembler::store_heap_oop(Register d, Register s1, int simm13a) { 3826 if (UseCompressedOops) { 3827 assert(s1 != d, "not enough registers"); 3828 encode_heap_oop(d); 3829 st(d, s1, simm13a); 3830 } else { 3831 st_ptr(d, s1, simm13a); 3832 } 3833 } 3834 3835 void MacroAssembler::store_heap_oop(Register d, const Address& a, int offset) { 3836 if (UseCompressedOops) { 3837 assert(a.base() != d, "not enough registers"); 3838 encode_heap_oop(d); 3839 st(d, a, offset); 3840 } else { 3841 st_ptr(d, a, offset); 3842 } 3843 } 3844 3845 3846 void MacroAssembler::encode_heap_oop(Register src, Register dst) { 3847 assert (UseCompressedOops, "must be compressed"); 3848 assert (Universe::heap() != NULL, "java heap should be initialized"); 3849 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 3850 verify_oop(src); 3851 if (Universe::narrow_oop_base() == NULL) { 3852 srlx(src, LogMinObjAlignmentInBytes, dst); 3853 return; 3854 } 3855 Label done; 3856 if (src == dst) { 3857 // optimize for frequent case src == dst 3858 bpr(rc_nz, true, Assembler::pt, src, done); 3859 delayed() -> sub(src, G6_heapbase, dst); // annuled if not taken 3860 bind(done); 3861 srlx(src, LogMinObjAlignmentInBytes, dst); 3862 } else { 3863 bpr(rc_z, false, Assembler::pn, src, done); 3864 delayed() -> mov(G0, dst); 3865 // could be moved before branch, and annulate delay, 3866 // but may add some unneeded work decoding null 3867 sub(src, G6_heapbase, dst); 3868 srlx(dst, LogMinObjAlignmentInBytes, dst); 3869 bind(done); 3870 } 3871 } 3872 3873 3874 void MacroAssembler::encode_heap_oop_not_null(Register r) { 3875 assert (UseCompressedOops, "must be compressed"); 3876 assert (Universe::heap() != NULL, "java heap should be initialized"); 3877 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 3878 verify_oop(r); 3879 if (Universe::narrow_oop_base() != NULL) 3880 sub(r, G6_heapbase, r); 3881 srlx(r, LogMinObjAlignmentInBytes, r); 3882 } 3883 3884 void MacroAssembler::encode_heap_oop_not_null(Register src, Register dst) { 3885 assert (UseCompressedOops, "must be compressed"); 3886 assert (Universe::heap() != NULL, "java heap should be initialized"); 3887 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 3888 verify_oop(src); 3889 if (Universe::narrow_oop_base() == NULL) { 3890 srlx(src, LogMinObjAlignmentInBytes, dst); 3891 } else { 3892 sub(src, G6_heapbase, dst); 3893 srlx(dst, LogMinObjAlignmentInBytes, dst); 3894 } 3895 } 3896 3897 // Same algorithm as oops.inline.hpp decode_heap_oop. 3898 void MacroAssembler::decode_heap_oop(Register src, Register dst) { 3899 assert (UseCompressedOops, "must be compressed"); 3900 assert (Universe::heap() != NULL, "java heap should be initialized"); 3901 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 3902 sllx(src, LogMinObjAlignmentInBytes, dst); 3903 if (Universe::narrow_oop_base() != NULL) { 3904 Label done; 3905 bpr(rc_nz, true, Assembler::pt, dst, done); 3906 delayed() -> add(dst, G6_heapbase, dst); // annuled if not taken 3907 bind(done); 3908 } 3909 verify_oop(dst); 3910 } 3911 3912 void MacroAssembler::decode_heap_oop_not_null(Register r) { 3913 // Do not add assert code to this unless you change vtableStubs_sparc.cpp 3914 // pd_code_size_limit. 3915 // Also do not verify_oop as this is called by verify_oop. 3916 assert (UseCompressedOops, "must be compressed"); 3917 assert (Universe::heap() != NULL, "java heap should be initialized"); 3918 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 3919 sllx(r, LogMinObjAlignmentInBytes, r); 3920 if (Universe::narrow_oop_base() != NULL) 3921 add(r, G6_heapbase, r); 3922 } 3923 3924 void MacroAssembler::decode_heap_oop_not_null(Register src, Register dst) { 3925 // Do not add assert code to this unless you change vtableStubs_sparc.cpp 3926 // pd_code_size_limit. 3927 // Also do not verify_oop as this is called by verify_oop. 3928 assert (UseCompressedOops, "must be compressed"); 3929 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 3930 sllx(src, LogMinObjAlignmentInBytes, dst); 3931 if (Universe::narrow_oop_base() != NULL) 3932 add(dst, G6_heapbase, dst); 3933 } 3934 3935 void MacroAssembler::encode_klass_not_null(Register r) { 3936 assert (UseCompressedClassPointers, "must be compressed"); 3937 if (Universe::narrow_klass_base() != NULL) { 3938 assert(r != G6_heapbase, "bad register choice"); 3939 set((intptr_t)Universe::narrow_klass_base(), G6_heapbase); 3940 sub(r, G6_heapbase, r); 3941 if (Universe::narrow_klass_shift() != 0) { 3942 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 3943 srlx(r, LogKlassAlignmentInBytes, r); 3944 } 3945 reinit_heapbase(); 3946 } else { 3947 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong"); 3948 srlx(r, Universe::narrow_klass_shift(), r); 3949 } 3950 } 3951 3952 void MacroAssembler::encode_klass_not_null(Register src, Register dst) { 3953 if (src == dst) { 3954 encode_klass_not_null(src); 3955 } else { 3956 assert (UseCompressedClassPointers, "must be compressed"); 3957 if (Universe::narrow_klass_base() != NULL) { 3958 set((intptr_t)Universe::narrow_klass_base(), dst); 3959 sub(src, dst, dst); 3960 if (Universe::narrow_klass_shift() != 0) { 3961 srlx(dst, LogKlassAlignmentInBytes, dst); 3962 } 3963 } else { 3964 // shift src into dst 3965 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong"); 3966 srlx(src, Universe::narrow_klass_shift(), dst); 3967 } 3968 } 3969 } 3970 3971 // Function instr_size_for_decode_klass_not_null() counts the instructions 3972 // generated by decode_klass_not_null() and reinit_heapbase(). Hence, if 3973 // the instructions they generate change, then this method needs to be updated. 3974 int MacroAssembler::instr_size_for_decode_klass_not_null() { 3975 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 3976 int num_instrs = 1; // shift src,dst or add 3977 if (Universe::narrow_klass_base() != NULL) { 3978 // set + add + set 3979 num_instrs += insts_for_internal_set((intptr_t)Universe::narrow_klass_base()) + 3980 insts_for_internal_set((intptr_t)Universe::narrow_ptrs_base()); 3981 if (Universe::narrow_klass_shift() != 0) { 3982 num_instrs += 1; // sllx 3983 } 3984 } 3985 return num_instrs * BytesPerInstWord; 3986 } 3987 3988 // !!! If the instructions that get generated here change then function 3989 // instr_size_for_decode_klass_not_null() needs to get updated. 3990 void MacroAssembler::decode_klass_not_null(Register r) { 3991 // Do not add assert code to this unless you change vtableStubs_sparc.cpp 3992 // pd_code_size_limit. 3993 assert (UseCompressedClassPointers, "must be compressed"); 3994 if (Universe::narrow_klass_base() != NULL) { 3995 assert(r != G6_heapbase, "bad register choice"); 3996 set((intptr_t)Universe::narrow_klass_base(), G6_heapbase); 3997 if (Universe::narrow_klass_shift() != 0) 3998 sllx(r, LogKlassAlignmentInBytes, r); 3999 add(r, G6_heapbase, r); 4000 reinit_heapbase(); 4001 } else { 4002 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong"); 4003 sllx(r, Universe::narrow_klass_shift(), r); 4004 } 4005 } 4006 4007 void MacroAssembler::decode_klass_not_null(Register src, Register dst) { 4008 if (src == dst) { 4009 decode_klass_not_null(src); 4010 } else { 4011 // Do not add assert code to this unless you change vtableStubs_sparc.cpp 4012 // pd_code_size_limit. 4013 assert (UseCompressedClassPointers, "must be compressed"); 4014 if (Universe::narrow_klass_base() != NULL) { 4015 if (Universe::narrow_klass_shift() != 0) { 4016 assert((src != G6_heapbase) && (dst != G6_heapbase), "bad register choice"); 4017 set((intptr_t)Universe::narrow_klass_base(), G6_heapbase); 4018 sllx(src, LogKlassAlignmentInBytes, dst); 4019 add(dst, G6_heapbase, dst); 4020 reinit_heapbase(); 4021 } else { 4022 set((intptr_t)Universe::narrow_klass_base(), dst); 4023 add(src, dst, dst); 4024 } 4025 } else { 4026 // shift/mov src into dst. 4027 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong"); 4028 sllx(src, Universe::narrow_klass_shift(), dst); 4029 } 4030 } 4031 } 4032 4033 void MacroAssembler::reinit_heapbase() { 4034 if (UseCompressedOops || UseCompressedClassPointers) { 4035 if (Universe::heap() != NULL) { 4036 set((intptr_t)Universe::narrow_ptrs_base(), G6_heapbase); 4037 } else { 4038 AddressLiteral base(Universe::narrow_ptrs_base_addr()); 4039 load_ptr_contents(base, G6_heapbase); 4040 } 4041 } 4042 } 4043 4044 #ifdef COMPILER2 4045 4046 // Compress char[] to byte[] by compressing 16 bytes at once. Return 0 on failure. 4047 void MacroAssembler::string_compress_16(Register src, Register dst, Register cnt, Register result, 4048 Register tmp1, Register tmp2, Register tmp3, Register tmp4, 4049 FloatRegister ftmp1, FloatRegister ftmp2, FloatRegister ftmp3, Label& Ldone) { 4050 Label Lloop, Lslow; 4051 assert(UseVIS >= 3, "VIS3 is required"); 4052 assert_different_registers(src, dst, cnt, tmp1, tmp2, tmp3, tmp4, result); 4053 assert_different_registers(ftmp1, ftmp2, ftmp3); 4054 4055 // Check if cnt >= 8 (= 16 bytes) 4056 cmp(cnt, 8); 4057 br(Assembler::less, false, Assembler::pn, Lslow); 4058 delayed()->mov(cnt, result); // copy count 4059 4060 // Check for 8-byte alignment of src and dst 4061 or3(src, dst, tmp1); 4062 andcc(tmp1, 7, G0); 4063 br(Assembler::notZero, false, Assembler::pn, Lslow); 4064 delayed()->nop(); 4065 4066 // Set mask for bshuffle instruction 4067 Register mask = tmp4; 4068 set(0x13579bdf, mask); 4069 bmask(mask, G0, G0); 4070 4071 // Set mask to 0xff00 ff00 ff00 ff00 to check for non-latin1 characters 4072 Assembler::sethi(0xff00fc00, mask); // mask = 0x0000 0000 ff00 fc00 4073 add(mask, 0x300, mask); // mask = 0x0000 0000 ff00 ff00 4074 sllx(mask, 32, tmp1); // tmp1 = 0xff00 ff00 0000 0000 4075 or3(mask, tmp1, mask); // mask = 0xff00 ff00 ff00 ff00 4076 4077 // Load first 8 bytes 4078 ldx(src, 0, tmp1); 4079 4080 bind(Lloop); 4081 // Load next 8 bytes 4082 ldx(src, 8, tmp2); 4083 4084 // Check for non-latin1 character by testing if the most significant byte of a char is set. 4085 // Although we have to move the data between integer and floating point registers, this is 4086 // still faster than the corresponding VIS instructions (ford/fand/fcmpd). 4087 or3(tmp1, tmp2, tmp3); 4088 btst(tmp3, mask); 4089 // annul zeroing if branch is not taken to preserve original count 4090 brx(Assembler::notZero, true, Assembler::pn, Ldone); 4091 delayed()->mov(G0, result); // 0 - failed 4092 4093 // Move bytes into float register 4094 movxtod(tmp1, ftmp1); 4095 movxtod(tmp2, ftmp2); 4096 4097 // Compress by copying one byte per char from ftmp1 and ftmp2 to ftmp3 4098 bshuffle(ftmp1, ftmp2, ftmp3); 4099 stf(FloatRegisterImpl::D, ftmp3, dst, 0); 4100 4101 // Increment addresses and decrement count 4102 inc(src, 16); 4103 inc(dst, 8); 4104 dec(cnt, 8); 4105 4106 cmp(cnt, 8); 4107 // annul LDX if branch is not taken to prevent access past end of string 4108 br(Assembler::greaterEqual, true, Assembler::pt, Lloop); 4109 delayed()->ldx(src, 0, tmp1); 4110 4111 // Fallback to slow version 4112 bind(Lslow); 4113 } 4114 4115 // Compress char[] to byte[]. Return 0 on failure. 4116 void MacroAssembler::string_compress(Register src, Register dst, Register cnt, Register result, Register tmp, Label& Ldone) { 4117 Label Lloop; 4118 assert_different_registers(src, dst, cnt, tmp, result); 4119 4120 lduh(src, 0, tmp); 4121 4122 bind(Lloop); 4123 inc(src, sizeof(jchar)); 4124 cmp(tmp, 0xff); 4125 // annul zeroing if branch is not taken to preserve original count 4126 br(Assembler::greater, true, Assembler::pn, Ldone); // don't check xcc 4127 delayed()->mov(G0, result); // 0 - failed 4128 deccc(cnt); 4129 stb(tmp, dst, 0); 4130 inc(dst); 4131 // annul LDUH if branch is not taken to prevent access past end of string 4132 br(Assembler::notZero, true, Assembler::pt, Lloop); 4133 delayed()->lduh(src, 0, tmp); // hoisted 4134 } 4135 4136 // Inflate byte[] to char[] by inflating 16 bytes at once. 4137 void MacroAssembler::string_inflate_16(Register src, Register dst, Register cnt, Register tmp, 4138 FloatRegister ftmp1, FloatRegister ftmp2, FloatRegister ftmp3, FloatRegister ftmp4, Label& Ldone) { 4139 Label Lloop, Lslow; 4140 assert(UseVIS >= 3, "VIS3 is required"); 4141 assert_different_registers(src, dst, cnt, tmp); 4142 assert_different_registers(ftmp1, ftmp2, ftmp3, ftmp4); 4143 4144 // Check if cnt >= 8 (= 16 bytes) 4145 cmp(cnt, 8); 4146 br(Assembler::less, false, Assembler::pn, Lslow); 4147 delayed()->nop(); 4148 4149 // Check for 8-byte alignment of src and dst 4150 or3(src, dst, tmp); 4151 andcc(tmp, 7, G0); 4152 br(Assembler::notZero, false, Assembler::pn, Lslow); 4153 // Initialize float register to zero 4154 FloatRegister zerof = ftmp4; 4155 delayed()->fzero(FloatRegisterImpl::D, zerof); 4156 4157 // Load first 8 bytes 4158 ldf(FloatRegisterImpl::D, src, 0, ftmp1); 4159 4160 bind(Lloop); 4161 inc(src, 8); 4162 dec(cnt, 8); 4163 4164 // Inflate the string by interleaving each byte from the source array 4165 // with a zero byte and storing the result in the destination array. 4166 fpmerge(zerof, ftmp1->successor(), ftmp2); 4167 stf(FloatRegisterImpl::D, ftmp2, dst, 8); 4168 fpmerge(zerof, ftmp1, ftmp3); 4169 stf(FloatRegisterImpl::D, ftmp3, dst, 0); 4170 4171 inc(dst, 16); 4172 4173 cmp(cnt, 8); 4174 // annul LDX if branch is not taken to prevent access past end of string 4175 br(Assembler::greaterEqual, true, Assembler::pt, Lloop); 4176 delayed()->ldf(FloatRegisterImpl::D, src, 0, ftmp1); 4177 4178 // Fallback to slow version 4179 bind(Lslow); 4180 } 4181 4182 // Inflate byte[] to char[]. 4183 void MacroAssembler::string_inflate(Register src, Register dst, Register cnt, Register tmp, Label& Ldone) { 4184 Label Loop; 4185 assert_different_registers(src, dst, cnt, tmp); 4186 4187 ldub(src, 0, tmp); 4188 bind(Loop); 4189 inc(src); 4190 deccc(cnt); 4191 sth(tmp, dst, 0); 4192 inc(dst, sizeof(jchar)); 4193 // annul LDUB if branch is not taken to prevent access past end of string 4194 br(Assembler::notZero, true, Assembler::pt, Loop); 4195 delayed()->ldub(src, 0, tmp); // hoisted 4196 } 4197 4198 void MacroAssembler::string_compare(Register str1, Register str2, 4199 Register cnt1, Register cnt2, 4200 Register tmp1, Register tmp2, 4201 Register result, int ae) { 4202 Label Ldone, Lloop; 4203 assert_different_registers(str1, str2, cnt1, cnt2, tmp1, result); 4204 int stride1, stride2; 4205 4206 // Note: Making use of the fact that compareTo(a, b) == -compareTo(b, a) 4207 // we interchange str1 and str2 in the UL case and negate the result. 4208 // Like this, str1 is always latin1 encoded, expect for the UU case. 4209 4210 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 4211 srl(cnt2, 1, cnt2); 4212 } 4213 4214 // See if the lengths are different, and calculate min in cnt1. 4215 // Save diff in case we need it for a tie-breaker. 4216 Label Lskip; 4217 Register diff = tmp1; 4218 subcc(cnt1, cnt2, diff); 4219 br(Assembler::greater, true, Assembler::pt, Lskip); 4220 // cnt2 is shorter, so use its count: 4221 delayed()->mov(cnt2, cnt1); 4222 bind(Lskip); 4223 4224 // Rename registers 4225 Register limit1 = cnt1; 4226 Register limit2 = limit1; 4227 Register chr1 = result; 4228 Register chr2 = cnt2; 4229 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 4230 // We need an additional register to keep track of two limits 4231 assert_different_registers(str1, str2, cnt1, cnt2, tmp1, tmp2, result); 4232 limit2 = tmp2; 4233 } 4234 4235 // Is the minimum length zero? 4236 cmp(limit1, (int)0); // use cast to resolve overloading ambiguity 4237 br(Assembler::equal, true, Assembler::pn, Ldone); 4238 // result is difference in lengths 4239 if (ae == StrIntrinsicNode::UU) { 4240 delayed()->sra(diff, 1, result); // Divide by 2 to get number of chars 4241 } else { 4242 delayed()->mov(diff, result); 4243 } 4244 4245 // Load first characters 4246 if (ae == StrIntrinsicNode::LL) { 4247 stride1 = stride2 = sizeof(jbyte); 4248 ldub(str1, 0, chr1); 4249 ldub(str2, 0, chr2); 4250 } else if (ae == StrIntrinsicNode::UU) { 4251 stride1 = stride2 = sizeof(jchar); 4252 lduh(str1, 0, chr1); 4253 lduh(str2, 0, chr2); 4254 } else { 4255 stride1 = sizeof(jbyte); 4256 stride2 = sizeof(jchar); 4257 ldub(str1, 0, chr1); 4258 lduh(str2, 0, chr2); 4259 } 4260 4261 // Compare first characters 4262 subcc(chr1, chr2, chr1); 4263 br(Assembler::notZero, false, Assembler::pt, Ldone); 4264 assert(chr1 == result, "result must be pre-placed"); 4265 delayed()->nop(); 4266 4267 // Check if the strings start at same location 4268 cmp(str1, str2); 4269 brx(Assembler::equal, true, Assembler::pn, Ldone); 4270 delayed()->mov(G0, result); // result is zero 4271 4272 // We have no guarantee that on 64 bit the higher half of limit is 0 4273 signx(limit1); 4274 4275 // Get limit 4276 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 4277 sll(limit1, 1, limit2); 4278 subcc(limit2, stride2, chr2); 4279 } 4280 subcc(limit1, stride1, chr1); 4281 br(Assembler::zero, true, Assembler::pn, Ldone); 4282 // result is difference in lengths 4283 if (ae == StrIntrinsicNode::UU) { 4284 delayed()->sra(diff, 1, result); // Divide by 2 to get number of chars 4285 } else { 4286 delayed()->mov(diff, result); 4287 } 4288 4289 // Shift str1 and str2 to the end of the arrays, negate limit 4290 add(str1, limit1, str1); 4291 add(str2, limit2, str2); 4292 neg(chr1, limit1); // limit1 = -(limit1-stride1) 4293 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 4294 neg(chr2, limit2); // limit2 = -(limit2-stride2) 4295 } 4296 4297 // Compare the rest of the characters 4298 load_sized_value(Address(str1, limit1), chr1, (ae == StrIntrinsicNode::UU) ? 2 : 1, false); 4299 4300 bind(Lloop); 4301 load_sized_value(Address(str2, limit2), chr2, (ae == StrIntrinsicNode::LL) ? 1 : 2, false); 4302 4303 subcc(chr1, chr2, chr1); 4304 br(Assembler::notZero, false, Assembler::pt, Ldone); 4305 assert(chr1 == result, "result must be pre-placed"); 4306 delayed()->inccc(limit1, stride1); 4307 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 4308 inccc(limit2, stride2); 4309 } 4310 4311 // annul LDUB if branch is not taken to prevent access past end of string 4312 br(Assembler::notZero, true, Assembler::pt, Lloop); 4313 delayed()->load_sized_value(Address(str1, limit1), chr1, (ae == StrIntrinsicNode::UU) ? 2 : 1, false); 4314 4315 // If strings are equal up to min length, return the length difference. 4316 if (ae == StrIntrinsicNode::UU) { 4317 // Divide by 2 to get number of chars 4318 sra(diff, 1, result); 4319 } else { 4320 mov(diff, result); 4321 } 4322 4323 // Otherwise, return the difference between the first mismatched chars. 4324 bind(Ldone); 4325 if(ae == StrIntrinsicNode::UL) { 4326 // Negate result (see note above) 4327 neg(result); 4328 } 4329 } 4330 4331 void MacroAssembler::array_equals(bool is_array_equ, Register ary1, Register ary2, 4332 Register limit, Register tmp, Register result, bool is_byte) { 4333 Label Ldone, Lloop, Lremaining; 4334 assert_different_registers(ary1, ary2, limit, tmp, result); 4335 4336 int length_offset = arrayOopDesc::length_offset_in_bytes(); 4337 int base_offset = arrayOopDesc::base_offset_in_bytes(is_byte ? T_BYTE : T_CHAR); 4338 assert(base_offset % 8 == 0, "Base offset must be 8-byte aligned"); 4339 4340 if (is_array_equ) { 4341 // return true if the same array 4342 cmp(ary1, ary2); 4343 brx(Assembler::equal, true, Assembler::pn, Ldone); 4344 delayed()->mov(1, result); // equal 4345 4346 br_null(ary1, true, Assembler::pn, Ldone); 4347 delayed()->clr(result); // not equal 4348 4349 br_null(ary2, true, Assembler::pn, Ldone); 4350 delayed()->clr(result); // not equal 4351 4352 // load the lengths of arrays 4353 ld(Address(ary1, length_offset), limit); 4354 ld(Address(ary2, length_offset), tmp); 4355 4356 // return false if the two arrays are not equal length 4357 cmp(limit, tmp); 4358 br(Assembler::notEqual, true, Assembler::pn, Ldone); 4359 delayed()->clr(result); // not equal 4360 } 4361 4362 cmp_zero_and_br(Assembler::zero, limit, Ldone, true, Assembler::pn); 4363 delayed()->mov(1, result); // zero-length arrays are equal 4364 4365 if (is_array_equ) { 4366 // load array addresses 4367 add(ary1, base_offset, ary1); 4368 add(ary2, base_offset, ary2); 4369 // set byte count 4370 if (!is_byte) { 4371 sll(limit, exact_log2(sizeof(jchar)), limit); 4372 } 4373 } else { 4374 // We have no guarantee that on 64 bit the higher half of limit is 0 4375 signx(limit); 4376 } 4377 4378 #ifdef ASSERT 4379 // Sanity check for doubleword (8-byte) alignment of ary1 and ary2. 4380 // Guaranteed on 64-bit systems (see arrayOopDesc::header_size_in_bytes()). 4381 Label Laligned; 4382 or3(ary1, ary2, tmp); 4383 andcc(tmp, 7, tmp); 4384 br_null_short(tmp, Assembler::pn, Laligned); 4385 STOP("First array element is not 8-byte aligned."); 4386 should_not_reach_here(); 4387 bind(Laligned); 4388 #endif 4389 4390 // Shift ary1 and ary2 to the end of the arrays, negate limit 4391 add(ary1, limit, ary1); 4392 add(ary2, limit, ary2); 4393 neg(limit, limit); 4394 4395 // MAIN LOOP 4396 // Load and compare array elements of size 'byte_width' until the elements are not 4397 // equal or we reached the end of the arrays. If the size of the arrays is not a 4398 // multiple of 'byte_width', we simply read over the end of the array, bail out and 4399 // compare the remaining bytes below by skipping the garbage bytes. 4400 ldx(ary1, limit, result); 4401 bind(Lloop); 4402 ldx(ary2, limit, tmp); 4403 inccc(limit, 8); 4404 // Bail out if we reached the end (but still do the comparison) 4405 br(Assembler::positive, false, Assembler::pn, Lremaining); 4406 delayed()->cmp(result, tmp); 4407 // Check equality of elements 4408 brx(Assembler::equal, false, Assembler::pt, target(Lloop)); 4409 delayed()->ldx(ary1, limit, result); 4410 4411 ba(Ldone); 4412 delayed()->clr(result); // not equal 4413 4414 // TAIL COMPARISON 4415 // We got here because we reached the end of the arrays. 'limit' is the number of 4416 // garbage bytes we may have compared by reading over the end of the arrays. Shift 4417 // out the garbage and compare the remaining elements. 4418 bind(Lremaining); 4419 // Optimistic shortcut: elements potentially including garbage are equal 4420 brx(Assembler::equal, true, Assembler::pt, target(Ldone)); 4421 delayed()->mov(1, result); // equal 4422 // Shift 'limit' bytes to the right and compare 4423 sll(limit, 3, limit); // bytes to bits 4424 srlx(result, limit, result); 4425 srlx(tmp, limit, tmp); 4426 cmp(result, tmp); 4427 clr(result); 4428 movcc(Assembler::equal, false, xcc, 1, result); 4429 4430 bind(Ldone); 4431 } 4432 4433 void MacroAssembler::has_negatives(Register inp, Register size, Register result, Register t2, Register t3, Register t4, Register t5) { 4434 4435 // test for negative bytes in input string of a given size 4436 // result 1 if found, 0 otherwise. 4437 4438 Label Lcore, Ltail, Lreturn, Lcore_rpt; 4439 4440 assert_different_registers(inp, size, t2, t3, t4, t5, result); 4441 4442 Register i = result; // result used as integer index i until very end 4443 Register lmask = t2; // t2 is aliased to lmask 4444 4445 // INITIALIZATION 4446 // =========================================================== 4447 // initialize highbits mask -> lmask = 0x8080808080808080 (8B/64b) 4448 // compute unaligned offset -> i 4449 // compute core end index -> t5 4450 Assembler::sethi(0x80808000, t2); //! sethi macro fails to emit optimal 4451 add(t2, 0x80, t2); 4452 sllx(t2, 32, t3); 4453 or3(t3, t2, lmask); // 0x8080808080808080 -> lmask 4454 sra(size,0,size); 4455 andcc(inp, 0x7, i); // unaligned offset -> i 4456 br(Assembler::zero, true, Assembler::pn, Lcore); // starts 8B aligned? 4457 delayed()->add(size, -8, t5); // (annuled) core end index -> t5 4458 4459 // =========================================================== 4460 4461 // UNALIGNED HEAD 4462 // =========================================================== 4463 // * unaligned head handling: grab aligned 8B containing unaligned inp(ut) 4464 // * obliterate (ignore) bytes outside string by shifting off reg ends 4465 // * compare with bitmask, short circuit return true if one or more high 4466 // bits set. 4467 cmp(size, 0); 4468 br(Assembler::zero, true, Assembler::pn, Lreturn); // short-circuit? 4469 delayed()->mov(0,result); // annuled so i not clobbered for following 4470 neg(i, t4); 4471 add(i, size, t5); 4472 ldx(inp, t4, t3); // raw aligned 8B containing unaligned head -> t3 4473 mov(8, t4); 4474 sub(t4, t5, t4); 4475 sra(t4, 31, t5); 4476 andn(t4, t5, t5); 4477 add(i, t5, t4); 4478 sll(t5, 3, t5); 4479 sll(t4, 3, t4); // # bits to shift right, left -> t5,t4 4480 srlx(t3, t5, t3); 4481 sllx(t3, t4, t3); // bytes outside string in 8B header obliterated -> t3 4482 andcc(lmask, t3, G0); 4483 brx(Assembler::notZero, true, Assembler::pn, Lreturn); // short circuit? 4484 delayed()->mov(1,result); // annuled so i not clobbered for following 4485 add(size, -8, t5); // core end index -> t5 4486 mov(8, t4); 4487 sub(t4, i, i); // # bytes examined in unalgn head (<8) -> i 4488 // =========================================================== 4489 4490 // ALIGNED CORE 4491 // =========================================================== 4492 // * iterate index i over aligned 8B sections of core, comparing with 4493 // bitmask, short circuit return true if one or more high bits set 4494 // t5 contains core end index/loop limit which is the index 4495 // of the MSB of last (unaligned) 8B fully contained in the string. 4496 // inp contains address of first byte in string/array 4497 // lmask contains 8B high bit mask for comparison 4498 // i contains next index to be processed (adr. inp+i is on 8B boundary) 4499 bind(Lcore); 4500 cmp_and_br_short(i, t5, Assembler::greater, Assembler::pn, Ltail); 4501 bind(Lcore_rpt); 4502 ldx(inp, i, t3); 4503 andcc(t3, lmask, G0); 4504 brx(Assembler::notZero, true, Assembler::pn, Lreturn); 4505 delayed()->mov(1, result); // annuled so i not clobbered for following 4506 add(i, 8, i); 4507 cmp_and_br_short(i, t5, Assembler::lessEqual, Assembler::pn, Lcore_rpt); 4508 // =========================================================== 4509 4510 // ALIGNED TAIL (<8B) 4511 // =========================================================== 4512 // handle aligned tail of 7B or less as complete 8B, obliterating end of 4513 // string bytes by shifting them off end, compare what's left with bitmask 4514 // inp contains address of first byte in string/array 4515 // lmask contains 8B high bit mask for comparison 4516 // i contains next index to be processed (adr. inp+i is on 8B boundary) 4517 bind(Ltail); 4518 subcc(size, i, t4); // # of remaining bytes in string -> t4 4519 // return 0 if no more remaining bytes 4520 br(Assembler::lessEqual, true, Assembler::pn, Lreturn); 4521 delayed()->mov(0, result); // annuled so i not clobbered for following 4522 ldx(inp, i, t3); // load final 8B (aligned) containing tail -> t3 4523 mov(8, t5); 4524 sub(t5, t4, t4); 4525 mov(0, result); // ** i clobbered at this point 4526 sll(t4, 3, t4); // bits beyond end of string -> t4 4527 srlx(t3, t4, t3); // bytes beyond end now obliterated -> t3 4528 andcc(lmask, t3, G0); 4529 movcc(Assembler::notZero, false, xcc, 1, result); 4530 bind(Lreturn); 4531 } 4532 4533 #endif 4534 4535 4536 // Use BIS for zeroing (count is in bytes). 4537 void MacroAssembler::bis_zeroing(Register to, Register count, Register temp, Label& Ldone) { 4538 assert(UseBlockZeroing && VM_Version::has_blk_zeroing(), "only works with BIS zeroing"); 4539 Register end = count; 4540 int cache_line_size = VM_Version::prefetch_data_size(); 4541 assert(cache_line_size > 0, "cache line size should be known for this code"); 4542 // Minimum count when BIS zeroing can be used since 4543 // it needs membar which is expensive. 4544 int block_zero_size = MAX2(cache_line_size*3, (int)BlockZeroingLowLimit); 4545 4546 Label small_loop; 4547 // Check if count is negative (dead code) or zero. 4548 // Note, count uses 64bit in 64 bit VM. 4549 cmp_and_brx_short(count, 0, Assembler::lessEqual, Assembler::pn, Ldone); 4550 4551 // Use BIS zeroing only for big arrays since it requires membar. 4552 if (Assembler::is_simm13(block_zero_size)) { // < 4096 4553 cmp(count, block_zero_size); 4554 } else { 4555 set(block_zero_size, temp); 4556 cmp(count, temp); 4557 } 4558 br(Assembler::lessUnsigned, false, Assembler::pt, small_loop); 4559 delayed()->add(to, count, end); 4560 4561 // Note: size is >= three (32 bytes) cache lines. 4562 4563 // Clean the beginning of space up to next cache line. 4564 for (int offs = 0; offs < cache_line_size; offs += 8) { 4565 stx(G0, to, offs); 4566 } 4567 4568 // align to next cache line 4569 add(to, cache_line_size, to); 4570 and3(to, -cache_line_size, to); 4571 4572 // Note: size left >= two (32 bytes) cache lines. 4573 4574 // BIS should not be used to zero tail (64 bytes) 4575 // to avoid zeroing a header of the following object. 4576 sub(end, (cache_line_size*2)-8, end); 4577 4578 Label bis_loop; 4579 bind(bis_loop); 4580 stxa(G0, to, G0, Assembler::ASI_ST_BLKINIT_PRIMARY); 4581 add(to, cache_line_size, to); 4582 cmp_and_brx_short(to, end, Assembler::lessUnsigned, Assembler::pt, bis_loop); 4583 4584 // BIS needs membar. 4585 membar(Assembler::StoreLoad); 4586 4587 add(end, (cache_line_size*2)-8, end); // restore end 4588 cmp_and_brx_short(to, end, Assembler::greaterEqualUnsigned, Assembler::pn, Ldone); 4589 4590 // Clean the tail. 4591 bind(small_loop); 4592 stx(G0, to, 0); 4593 add(to, 8, to); 4594 cmp_and_brx_short(to, end, Assembler::lessUnsigned, Assembler::pt, small_loop); 4595 nop(); // Separate short branches 4596 } 4597 4598 /** 4599 * Update CRC-32[C] with a byte value according to constants in table 4600 * 4601 * @param [in,out]crc Register containing the crc. 4602 * @param [in]val Register containing the byte to fold into the CRC. 4603 * @param [in]table Register containing the table of crc constants. 4604 * 4605 * uint32_t crc; 4606 * val = crc_table[(val ^ crc) & 0xFF]; 4607 * crc = val ^ (crc >> 8); 4608 */ 4609 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 4610 xor3(val, crc, val); 4611 and3(val, 0xFF, val); 4612 sllx(val, 2, val); 4613 lduw(table, val, val); 4614 srlx(crc, 8, crc); 4615 xor3(val, crc, crc); 4616 } 4617 4618 // Reverse byte order of lower 32 bits, assuming upper 32 bits all zeros 4619 void MacroAssembler::reverse_bytes_32(Register src, Register dst, Register tmp) { 4620 srlx(src, 24, dst); 4621 4622 sllx(src, 32+8, tmp); 4623 srlx(tmp, 32+24, tmp); 4624 sllx(tmp, 8, tmp); 4625 or3(dst, tmp, dst); 4626 4627 sllx(src, 32+16, tmp); 4628 srlx(tmp, 32+24, tmp); 4629 sllx(tmp, 16, tmp); 4630 or3(dst, tmp, dst); 4631 4632 sllx(src, 32+24, tmp); 4633 srlx(tmp, 32, tmp); 4634 or3(dst, tmp, dst); 4635 } 4636 4637 void MacroAssembler::movitof_revbytes(Register src, FloatRegister dst, Register tmp1, Register tmp2) { 4638 reverse_bytes_32(src, tmp1, tmp2); 4639 movxtod(tmp1, dst); 4640 } 4641 4642 void MacroAssembler::movftoi_revbytes(FloatRegister src, Register dst, Register tmp1, Register tmp2) { 4643 movdtox(src, tmp1); 4644 reverse_bytes_32(tmp1, dst, tmp2); 4645 } 4646 4647 void MacroAssembler::fold_128bit_crc32(Register xcrc_hi, Register xcrc_lo, Register xK_hi, Register xK_lo, Register xtmp_hi, Register xtmp_lo, Register buf, int offset) { 4648 xmulx(xcrc_hi, xK_hi, xtmp_lo); 4649 xmulxhi(xcrc_hi, xK_hi, xtmp_hi); 4650 xmulxhi(xcrc_lo, xK_lo, xcrc_hi); 4651 xmulx(xcrc_lo, xK_lo, xcrc_lo); 4652 xor3(xcrc_lo, xtmp_lo, xcrc_lo); 4653 xor3(xcrc_hi, xtmp_hi, xcrc_hi); 4654 ldxl(buf, G0, xtmp_lo); 4655 inc(buf, 8); 4656 ldxl(buf, G0, xtmp_hi); 4657 inc(buf, 8); 4658 xor3(xcrc_lo, xtmp_lo, xcrc_lo); 4659 xor3(xcrc_hi, xtmp_hi, xcrc_hi); 4660 } 4661 4662 void MacroAssembler::fold_128bit_crc32(Register xcrc_hi, Register xcrc_lo, Register xK_hi, Register xK_lo, Register xtmp_hi, Register xtmp_lo, Register xbuf_hi, Register xbuf_lo) { 4663 mov(xcrc_lo, xtmp_lo); 4664 mov(xcrc_hi, xtmp_hi); 4665 xmulx(xtmp_hi, xK_hi, xtmp_lo); 4666 xmulxhi(xtmp_hi, xK_hi, xtmp_hi); 4667 xmulxhi(xcrc_lo, xK_lo, xcrc_hi); 4668 xmulx(xcrc_lo, xK_lo, xcrc_lo); 4669 xor3(xcrc_lo, xbuf_lo, xcrc_lo); 4670 xor3(xcrc_hi, xbuf_hi, xcrc_hi); 4671 xor3(xcrc_lo, xtmp_lo, xcrc_lo); 4672 xor3(xcrc_hi, xtmp_hi, xcrc_hi); 4673 } 4674 4675 void MacroAssembler::fold_8bit_crc32(Register xcrc, Register table, Register xtmp, Register tmp) { 4676 and3(xcrc, 0xFF, tmp); 4677 sllx(tmp, 2, tmp); 4678 lduw(table, tmp, xtmp); 4679 srlx(xcrc, 8, xcrc); 4680 xor3(xtmp, xcrc, xcrc); 4681 } 4682 4683 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 4684 and3(crc, 0xFF, tmp); 4685 srlx(crc, 8, crc); 4686 sllx(tmp, 2, tmp); 4687 lduw(table, tmp, tmp); 4688 xor3(tmp, crc, crc); 4689 } 4690 4691 #define CRC32_TMP_REG_NUM 18 4692 4693 #define CRC32_CONST_64 0x163cd6124 4694 #define CRC32_CONST_96 0x0ccaa009e 4695 #define CRC32_CONST_160 0x1751997d0 4696 #define CRC32_CONST_480 0x1c6e41596 4697 #define CRC32_CONST_544 0x154442bd4 4698 4699 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table) { 4700 4701 Label L_cleanup_loop, L_cleanup_check, L_align_loop, L_align_check; 4702 Label L_main_loop_prologue; 4703 Label L_fold_512b, L_fold_512b_loop, L_fold_128b; 4704 Label L_fold_tail, L_fold_tail_loop; 4705 Label L_8byte_fold_loop, L_8byte_fold_check; 4706 4707 const Register tmp[CRC32_TMP_REG_NUM] = {L0, L1, L2, L3, L4, L5, L6, G1, I0, I1, I2, I3, I4, I5, I7, O4, O5, G3}; 4708 4709 Register const_64 = tmp[CRC32_TMP_REG_NUM-1]; 4710 Register const_96 = tmp[CRC32_TMP_REG_NUM-1]; 4711 Register const_160 = tmp[CRC32_TMP_REG_NUM-2]; 4712 Register const_480 = tmp[CRC32_TMP_REG_NUM-1]; 4713 Register const_544 = tmp[CRC32_TMP_REG_NUM-2]; 4714 4715 set(ExternalAddress(StubRoutines::crc_table_addr()), table); 4716 4717 not1(crc); // ~c 4718 clruwu(crc); // clear upper 32 bits of crc 4719 4720 // Check if below cutoff, proceed directly to cleanup code 4721 mov(31, G4); 4722 cmp_and_br_short(len, G4, Assembler::lessEqualUnsigned, Assembler::pt, L_cleanup_check); 4723 4724 // Align buffer to 8 byte boundry 4725 mov(8, O5); 4726 and3(buf, 0x7, O4); 4727 sub(O5, O4, O5); 4728 and3(O5, 0x7, O5); 4729 sub(len, O5, len); 4730 ba(L_align_check); 4731 delayed()->nop(); 4732 4733 // Alignment loop, table look up method for up to 7 bytes 4734 bind(L_align_loop); 4735 ldub(buf, 0, O4); 4736 inc(buf); 4737 dec(O5); 4738 xor3(O4, crc, O4); 4739 and3(O4, 0xFF, O4); 4740 sllx(O4, 2, O4); 4741 lduw(table, O4, O4); 4742 srlx(crc, 8, crc); 4743 xor3(O4, crc, crc); 4744 bind(L_align_check); 4745 nop(); 4746 cmp_and_br_short(O5, 0, Assembler::notEqual, Assembler::pt, L_align_loop); 4747 4748 // Aligned on 64-bit (8-byte) boundry at this point 4749 // Check if still above cutoff (31-bytes) 4750 mov(31, G4); 4751 cmp_and_br_short(len, G4, Assembler::lessEqualUnsigned, Assembler::pt, L_cleanup_check); 4752 // At least 32 bytes left to process 4753 4754 // Free up registers by storing them to FP registers 4755 for (int i = 0; i < CRC32_TMP_REG_NUM; i++) { 4756 movxtod(tmp[i], as_FloatRegister(2*i)); 4757 } 4758 4759 // Determine which loop to enter 4760 // Shared prologue 4761 ldxl(buf, G0, tmp[0]); 4762 inc(buf, 8); 4763 ldxl(buf, G0, tmp[1]); 4764 inc(buf, 8); 4765 xor3(tmp[0], crc, tmp[0]); // Fold CRC into first few bytes 4766 and3(crc, 0, crc); // Clear out the crc register 4767 // Main loop needs 128-bytes at least 4768 mov(128, G4); 4769 mov(64, tmp[2]); 4770 cmp_and_br_short(len, G4, Assembler::greaterEqualUnsigned, Assembler::pt, L_main_loop_prologue); 4771 // Less than 64 bytes 4772 nop(); 4773 cmp_and_br_short(len, tmp[2], Assembler::lessUnsigned, Assembler::pt, L_fold_tail); 4774 // Between 64 and 127 bytes 4775 set64(CRC32_CONST_96, const_96, tmp[8]); 4776 set64(CRC32_CONST_160, const_160, tmp[9]); 4777 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[2], tmp[3], buf, 0); 4778 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[4], tmp[5], buf, 16); 4779 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[6], tmp[7], buf, 32); 4780 dec(len, 48); 4781 ba(L_fold_tail); 4782 delayed()->nop(); 4783 4784 bind(L_main_loop_prologue); 4785 for (int i = 2; i < 8; i++) { 4786 ldxl(buf, G0, tmp[i]); 4787 inc(buf, 8); 4788 } 4789 4790 // Fold total 512 bits of polynomial on each iteration, 4791 // 128 bits per each of 4 parallel streams 4792 set64(CRC32_CONST_480, const_480, tmp[8]); 4793 set64(CRC32_CONST_544, const_544, tmp[9]); 4794 4795 mov(128, G4); 4796 bind(L_fold_512b_loop); 4797 fold_128bit_crc32(tmp[1], tmp[0], const_480, const_544, tmp[9], tmp[8], buf, 0); 4798 fold_128bit_crc32(tmp[3], tmp[2], const_480, const_544, tmp[11], tmp[10], buf, 16); 4799 fold_128bit_crc32(tmp[5], tmp[4], const_480, const_544, tmp[13], tmp[12], buf, 32); 4800 fold_128bit_crc32(tmp[7], tmp[6], const_480, const_544, tmp[15], tmp[14], buf, 64); 4801 dec(len, 64); 4802 cmp_and_br_short(len, G4, Assembler::greaterEqualUnsigned, Assembler::pt, L_fold_512b_loop); 4803 4804 // Fold 512 bits to 128 bits 4805 bind(L_fold_512b); 4806 set64(CRC32_CONST_96, const_96, tmp[8]); 4807 set64(CRC32_CONST_160, const_160, tmp[9]); 4808 4809 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[8], tmp[9], tmp[3], tmp[2]); 4810 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[8], tmp[9], tmp[5], tmp[4]); 4811 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[8], tmp[9], tmp[7], tmp[6]); 4812 dec(len, 48); 4813 4814 // Fold the rest of 128 bits data chunks 4815 bind(L_fold_tail); 4816 mov(32, G4); 4817 cmp_and_br_short(len, G4, Assembler::lessEqualUnsigned, Assembler::pt, L_fold_128b); 4818 4819 set64(CRC32_CONST_96, const_96, tmp[8]); 4820 set64(CRC32_CONST_160, const_160, tmp[9]); 4821 4822 bind(L_fold_tail_loop); 4823 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[2], tmp[3], buf, 0); 4824 sub(len, 16, len); 4825 cmp_and_br_short(len, G4, Assembler::greaterEqualUnsigned, Assembler::pt, L_fold_tail_loop); 4826 4827 // Fold the 128 bits in tmps 0 - 1 into tmp 1 4828 bind(L_fold_128b); 4829 4830 set64(CRC32_CONST_64, const_64, tmp[4]); 4831 4832 xmulx(const_64, tmp[0], tmp[2]); 4833 xmulxhi(const_64, tmp[0], tmp[3]); 4834 4835 srl(tmp[2], G0, tmp[4]); 4836 xmulx(const_64, tmp[4], tmp[4]); 4837 4838 srlx(tmp[2], 32, tmp[2]); 4839 sllx(tmp[3], 32, tmp[3]); 4840 or3(tmp[2], tmp[3], tmp[2]); 4841 4842 xor3(tmp[4], tmp[1], tmp[4]); 4843 xor3(tmp[4], tmp[2], tmp[1]); 4844 dec(len, 8); 4845 4846 // Use table lookup for the 8 bytes left in tmp[1] 4847 dec(len, 8); 4848 4849 // 8 8-bit folds to compute 32-bit CRC. 4850 for (int j = 0; j < 4; j++) { 4851 fold_8bit_crc32(tmp[1], table, tmp[2], tmp[3]); 4852 } 4853 srl(tmp[1], G0, crc); // move 32 bits to general register 4854 for (int j = 0; j < 4; j++) { 4855 fold_8bit_crc32(crc, table, tmp[3]); 4856 } 4857 4858 bind(L_8byte_fold_check); 4859 4860 // Restore int registers saved in FP registers 4861 for (int i = 0; i < CRC32_TMP_REG_NUM; i++) { 4862 movdtox(as_FloatRegister(2*i), tmp[i]); 4863 } 4864 4865 ba(L_cleanup_check); 4866 delayed()->nop(); 4867 4868 // Table look-up method for the remaining few bytes 4869 bind(L_cleanup_loop); 4870 ldub(buf, 0, O4); 4871 inc(buf); 4872 dec(len); 4873 xor3(O4, crc, O4); 4874 and3(O4, 0xFF, O4); 4875 sllx(O4, 2, O4); 4876 lduw(table, O4, O4); 4877 srlx(crc, 8, crc); 4878 xor3(O4, crc, crc); 4879 bind(L_cleanup_check); 4880 nop(); 4881 cmp_and_br_short(len, 0, Assembler::greaterUnsigned, Assembler::pt, L_cleanup_loop); 4882 4883 not1(crc); 4884 } 4885 4886 #define CHUNK_LEN 128 /* 128 x 8B = 1KB */ 4887 #define CHUNK_K1 0x1307a0206 /* reverseBits(pow(x, CHUNK_LEN*8*8*3 - 32) mod P(x)) << 1 */ 4888 #define CHUNK_K2 0x1a0f717c4 /* reverseBits(pow(x, CHUNK_LEN*8*8*2 - 32) mod P(x)) << 1 */ 4889 #define CHUNK_K3 0x0170076fa /* reverseBits(pow(x, CHUNK_LEN*8*8*1 - 32) mod P(x)) << 1 */ 4890 4891 void MacroAssembler::kernel_crc32c(Register crc, Register buf, Register len, Register table) { 4892 4893 Label L_crc32c_head, L_crc32c_aligned; 4894 Label L_crc32c_parallel, L_crc32c_parallel_loop; 4895 Label L_crc32c_serial, L_crc32c_x32_loop, L_crc32c_x8, L_crc32c_x8_loop; 4896 Label L_crc32c_done, L_crc32c_tail, L_crc32c_return; 4897 4898 set(ExternalAddress(StubRoutines::crc32c_table_addr()), table); 4899 4900 cmp_and_br_short(len, 0, Assembler::lessEqual, Assembler::pn, L_crc32c_return); 4901 4902 // clear upper 32 bits of crc 4903 clruwu(crc); 4904 4905 and3(buf, 7, G4); 4906 cmp_and_brx_short(G4, 0, Assembler::equal, Assembler::pt, L_crc32c_aligned); 4907 4908 mov(8, G1); 4909 sub(G1, G4, G4); 4910 4911 // ------ process the misaligned head (7 bytes or less) ------ 4912 bind(L_crc32c_head); 4913 4914 // crc = (crc >>> 8) ^ byteTable[(crc ^ b) & 0xFF]; 4915 ldub(buf, 0, G1); 4916 update_byte_crc32(crc, G1, table); 4917 4918 inc(buf); 4919 dec(len); 4920 cmp_and_br_short(len, 0, Assembler::equal, Assembler::pn, L_crc32c_return); 4921 dec(G4); 4922 cmp_and_br_short(G4, 0, Assembler::greater, Assembler::pt, L_crc32c_head); 4923 4924 // ------ process the 8-byte-aligned body ------ 4925 bind(L_crc32c_aligned); 4926 nop(); 4927 cmp_and_br_short(len, 8, Assembler::less, Assembler::pn, L_crc32c_tail); 4928 4929 // reverse the byte order of lower 32 bits to big endian, and move to FP side 4930 movitof_revbytes(crc, F0, G1, G3); 4931 4932 set(CHUNK_LEN*8*4, G4); 4933 cmp_and_br_short(len, G4, Assembler::less, Assembler::pt, L_crc32c_serial); 4934 4935 // ------ process four 1KB chunks in parallel ------ 4936 bind(L_crc32c_parallel); 4937 4938 fzero(FloatRegisterImpl::D, F2); 4939 fzero(FloatRegisterImpl::D, F4); 4940 fzero(FloatRegisterImpl::D, F6); 4941 4942 mov(CHUNK_LEN - 1, G4); 4943 bind(L_crc32c_parallel_loop); 4944 // schedule ldf's ahead of crc32c's to hide the load-use latency 4945 ldf(FloatRegisterImpl::D, buf, 0, F8); 4946 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*8, F10); 4947 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*16, F12); 4948 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*24, F14); 4949 crc32c(F0, F8, F0); 4950 crc32c(F2, F10, F2); 4951 crc32c(F4, F12, F4); 4952 crc32c(F6, F14, F6); 4953 inc(buf, 8); 4954 dec(G4); 4955 cmp_and_br_short(G4, 0, Assembler::greater, Assembler::pt, L_crc32c_parallel_loop); 4956 4957 ldf(FloatRegisterImpl::D, buf, 0, F8); 4958 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*8, F10); 4959 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*16, F12); 4960 crc32c(F0, F8, F0); 4961 crc32c(F2, F10, F2); 4962 crc32c(F4, F12, F4); 4963 4964 inc(buf, CHUNK_LEN*24); 4965 ldfl(FloatRegisterImpl::D, buf, G0, F14); // load in little endian 4966 inc(buf, 8); 4967 4968 prefetch(buf, 0, Assembler::severalReads); 4969 prefetch(buf, CHUNK_LEN*8, Assembler::severalReads); 4970 prefetch(buf, CHUNK_LEN*16, Assembler::severalReads); 4971 prefetch(buf, CHUNK_LEN*24, Assembler::severalReads); 4972 4973 // move to INT side, and reverse the byte order of lower 32 bits to little endian 4974 movftoi_revbytes(F0, O4, G1, G4); 4975 movftoi_revbytes(F2, O5, G1, G4); 4976 movftoi_revbytes(F4, G5, G1, G4); 4977 4978 // combine the results of 4 chunks 4979 set64(CHUNK_K1, G3, G1); 4980 xmulx(O4, G3, O4); 4981 set64(CHUNK_K2, G3, G1); 4982 xmulx(O5, G3, O5); 4983 set64(CHUNK_K3, G3, G1); 4984 xmulx(G5, G3, G5); 4985 4986 movdtox(F14, G4); 4987 xor3(O4, O5, O5); 4988 xor3(G5, O5, O5); 4989 xor3(G4, O5, O5); 4990 4991 // reverse the byte order to big endian, via stack, and move to FP side 4992 // TODO: use new revb instruction 4993 add(SP, -8, G1); 4994 srlx(G1, 3, G1); 4995 sllx(G1, 3, G1); 4996 stx(O5, G1, G0); 4997 ldfl(FloatRegisterImpl::D, G1, G0, F2); // load in little endian 4998 4999 crc32c(F6, F2, F0); 5000 5001 set(CHUNK_LEN*8*4, G4); 5002 sub(len, G4, len); 5003 cmp_and_br_short(len, G4, Assembler::greaterEqual, Assembler::pt, L_crc32c_parallel); 5004 nop(); 5005 cmp_and_br_short(len, 0, Assembler::equal, Assembler::pt, L_crc32c_done); 5006 5007 bind(L_crc32c_serial); 5008 5009 mov(32, G4); 5010 cmp_and_br_short(len, G4, Assembler::less, Assembler::pn, L_crc32c_x8); 5011 5012 // ------ process 32B chunks ------ 5013 bind(L_crc32c_x32_loop); 5014 ldf(FloatRegisterImpl::D, buf, 0, F2); 5015 crc32c(F0, F2, F0); 5016 ldf(FloatRegisterImpl::D, buf, 8, F2); 5017 crc32c(F0, F2, F0); 5018 ldf(FloatRegisterImpl::D, buf, 16, F2); 5019 crc32c(F0, F2, F0); 5020 ldf(FloatRegisterImpl::D, buf, 24, F2); 5021 inc(buf, 32); 5022 crc32c(F0, F2, F0); 5023 dec(len, 32); 5024 cmp_and_br_short(len, G4, Assembler::greaterEqual, Assembler::pt, L_crc32c_x32_loop); 5025 5026 bind(L_crc32c_x8); 5027 nop(); 5028 cmp_and_br_short(len, 8, Assembler::less, Assembler::pt, L_crc32c_done); 5029 5030 // ------ process 8B chunks ------ 5031 bind(L_crc32c_x8_loop); 5032 ldf(FloatRegisterImpl::D, buf, 0, F2); 5033 inc(buf, 8); 5034 crc32c(F0, F2, F0); 5035 dec(len, 8); 5036 cmp_and_br_short(len, 8, Assembler::greaterEqual, Assembler::pt, L_crc32c_x8_loop); 5037 5038 bind(L_crc32c_done); 5039 5040 // move to INT side, and reverse the byte order of lower 32 bits to little endian 5041 movftoi_revbytes(F0, crc, G1, G3); 5042 5043 cmp_and_br_short(len, 0, Assembler::equal, Assembler::pt, L_crc32c_return); 5044 5045 // ------ process the misaligned tail (7 bytes or less) ------ 5046 bind(L_crc32c_tail); 5047 5048 // crc = (crc >>> 8) ^ byteTable[(crc ^ b) & 0xFF]; 5049 ldub(buf, 0, G1); 5050 update_byte_crc32(crc, G1, table); 5051 5052 inc(buf); 5053 dec(len); 5054 cmp_and_br_short(len, 0, Assembler::greater, Assembler::pt, L_crc32c_tail); 5055 5056 bind(L_crc32c_return); 5057 nop(); 5058 }