1 /* 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.hpp" 27 #include "assembler_sparc.inline.hpp" 28 #include "gc_interface/collectedHeap.inline.hpp" 29 #include "interpreter/interpreter.hpp" 30 #include "memory/cardTableModRefBS.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "prims/methodHandles.hpp" 33 #include "runtime/biasedLocking.hpp" 34 #include "runtime/interfaceSupport.hpp" 35 #include "runtime/objectMonitor.hpp" 36 #include "runtime/os.hpp" 37 #include "runtime/sharedRuntime.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #ifndef SERIALGC 40 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 41 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 42 #include "gc_implementation/g1/heapRegion.hpp" 43 #endif 44 45 #ifdef PRODUCT 46 #define BLOCK_COMMENT(str) /* nothing */ 47 #else 48 #define BLOCK_COMMENT(str) block_comment(str) 49 #endif 50 51 // Convert the raw encoding form into the form expected by the 52 // constructor for Address. 53 Address Address::make_raw(int base, int index, int scale, int disp, bool disp_is_oop) { 54 assert(scale == 0, "not supported"); 55 RelocationHolder rspec; 56 if (disp_is_oop) { 57 rspec = Relocation::spec_simple(relocInfo::oop_type); 58 } 59 60 Register rindex = as_Register(index); 61 if (rindex != G0) { 62 Address madr(as_Register(base), rindex); 63 madr._rspec = rspec; 64 return madr; 65 } else { 66 Address madr(as_Register(base), disp); 67 madr._rspec = rspec; 68 return madr; 69 } 70 } 71 72 Address Argument::address_in_frame() const { 73 // Warning: In LP64 mode disp will occupy more than 10 bits, but 74 // op codes such as ld or ldx, only access disp() to get 75 // their simm13 argument. 76 int disp = ((_number - Argument::n_register_parameters + frame::memory_parameter_word_sp_offset) * BytesPerWord) + STACK_BIAS; 77 if (is_in()) 78 return Address(FP, disp); // In argument. 79 else 80 return Address(SP, disp); // Out argument. 81 } 82 83 static const char* argumentNames[][2] = { 84 {"A0","P0"}, {"A1","P1"}, {"A2","P2"}, {"A3","P3"}, {"A4","P4"}, 85 {"A5","P5"}, {"A6","P6"}, {"A7","P7"}, {"A8","P8"}, {"A9","P9"}, 86 {"A(n>9)","P(n>9)"} 87 }; 88 89 const char* Argument::name() const { 90 int nofArgs = sizeof argumentNames / sizeof argumentNames[0]; 91 int num = number(); 92 if (num >= nofArgs) num = nofArgs - 1; 93 return argumentNames[num][is_in() ? 1 : 0]; 94 } 95 96 void Assembler::print_instruction(int inst) { 97 const char* s; 98 switch (inv_op(inst)) { 99 default: s = "????"; break; 100 case call_op: s = "call"; break; 101 case branch_op: 102 switch (inv_op2(inst)) { 103 case fb_op2: s = "fb"; break; 104 case fbp_op2: s = "fbp"; break; 105 case br_op2: s = "br"; break; 106 case bp_op2: s = "bp"; break; 107 case cb_op2: s = "cb"; break; 108 case bpr_op2: { 109 if (is_cbc(inst)) { 110 s = is_cxb(inst) ? "cxb" : "cwb"; 111 } else { 112 s = "bpr"; 113 } 114 break; 115 } 116 default: s = "????"; break; 117 } 118 } 119 ::tty->print("%s", s); 120 } 121 122 123 // Patch instruction inst at offset inst_pos to refer to dest_pos 124 // and return the resulting instruction. 125 // We should have pcs, not offsets, but since all is relative, it will work out 126 // OK. 127 int Assembler::patched_branch(int dest_pos, int inst, int inst_pos) { 128 129 int m; // mask for displacement field 130 int v; // new value for displacement field 131 const int word_aligned_ones = -4; 132 switch (inv_op(inst)) { 133 default: ShouldNotReachHere(); 134 case call_op: m = wdisp(word_aligned_ones, 0, 30); v = wdisp(dest_pos, inst_pos, 30); break; 135 case branch_op: 136 switch (inv_op2(inst)) { 137 case fbp_op2: m = wdisp( word_aligned_ones, 0, 19); v = wdisp( dest_pos, inst_pos, 19); break; 138 case bp_op2: m = wdisp( word_aligned_ones, 0, 19); v = wdisp( dest_pos, inst_pos, 19); break; 139 case fb_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break; 140 case br_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break; 141 case cb_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break; 142 case bpr_op2: { 143 if (is_cbc(inst)) { 144 m = wdisp10(word_aligned_ones, 0); 145 v = wdisp10(dest_pos, inst_pos); 146 } else { 147 m = wdisp16(word_aligned_ones, 0); 148 v = wdisp16(dest_pos, inst_pos); 149 } 150 break; 151 } 152 default: ShouldNotReachHere(); 153 } 154 } 155 return inst & ~m | v; 156 } 157 158 // Return the offset of the branch destionation of instruction inst 159 // at offset pos. 160 // Should have pcs, but since all is relative, it works out. 161 int Assembler::branch_destination(int inst, int pos) { 162 int r; 163 switch (inv_op(inst)) { 164 default: ShouldNotReachHere(); 165 case call_op: r = inv_wdisp(inst, pos, 30); break; 166 case branch_op: 167 switch (inv_op2(inst)) { 168 case fbp_op2: r = inv_wdisp( inst, pos, 19); break; 169 case bp_op2: r = inv_wdisp( inst, pos, 19); break; 170 case fb_op2: r = inv_wdisp( inst, pos, 22); break; 171 case br_op2: r = inv_wdisp( inst, pos, 22); break; 172 case cb_op2: r = inv_wdisp( inst, pos, 22); break; 173 case bpr_op2: { 174 if (is_cbc(inst)) { 175 r = inv_wdisp10(inst, pos); 176 } else { 177 r = inv_wdisp16(inst, pos); 178 } 179 break; 180 } 181 default: ShouldNotReachHere(); 182 } 183 } 184 return r; 185 } 186 187 int AbstractAssembler::code_fill_byte() { 188 return 0x00; // illegal instruction 0x00000000 189 } 190 191 Assembler::Condition Assembler::reg_cond_to_cc_cond(Assembler::RCondition in) { 192 switch (in) { 193 case rc_z: return equal; 194 case rc_lez: return lessEqual; 195 case rc_lz: return less; 196 case rc_nz: return notEqual; 197 case rc_gz: return greater; 198 case rc_gez: return greaterEqual; 199 default: 200 ShouldNotReachHere(); 201 } 202 return equal; 203 } 204 205 // Generate a bunch 'o stuff (including v9's 206 #ifndef PRODUCT 207 void Assembler::test_v9() { 208 add( G0, G1, G2 ); 209 add( G3, 0, G4 ); 210 211 addcc( G5, G6, G7 ); 212 addcc( I0, 1, I1 ); 213 addc( I2, I3, I4 ); 214 addc( I5, -1, I6 ); 215 addccc( I7, L0, L1 ); 216 addccc( L2, (1 << 12) - 2, L3 ); 217 218 Label lbl1, lbl2, lbl3; 219 220 bind(lbl1); 221 222 bpr( rc_z, true, pn, L4, pc(), relocInfo::oop_type ); 223 delayed()->nop(); 224 bpr( rc_lez, false, pt, L5, lbl1); 225 delayed()->nop(); 226 227 fb( f_never, true, pc() + 4, relocInfo::none); 228 delayed()->nop(); 229 fb( f_notEqual, false, lbl2 ); 230 delayed()->nop(); 231 232 fbp( f_notZero, true, fcc0, pn, pc() - 4, relocInfo::none); 233 delayed()->nop(); 234 fbp( f_lessOrGreater, false, fcc1, pt, lbl3 ); 235 delayed()->nop(); 236 237 br( equal, true, pc() + 1024, relocInfo::none); 238 delayed()->nop(); 239 br( lessEqual, false, lbl1 ); 240 delayed()->nop(); 241 br( never, false, lbl1 ); 242 delayed()->nop(); 243 244 bp( less, true, icc, pn, pc(), relocInfo::none); 245 delayed()->nop(); 246 bp( lessEqualUnsigned, false, xcc, pt, lbl2 ); 247 delayed()->nop(); 248 249 call( pc(), relocInfo::none); 250 delayed()->nop(); 251 call( lbl3 ); 252 delayed()->nop(); 253 254 255 casa( L6, L7, O0 ); 256 casxa( O1, O2, O3, 0 ); 257 258 udiv( O4, O5, O7 ); 259 udiv( G0, (1 << 12) - 1, G1 ); 260 sdiv( G1, G2, G3 ); 261 sdiv( G4, -((1 << 12) - 1), G5 ); 262 udivcc( G6, G7, I0 ); 263 udivcc( I1, -((1 << 12) - 2), I2 ); 264 sdivcc( I3, I4, I5 ); 265 sdivcc( I6, -((1 << 12) - 0), I7 ); 266 267 done(); 268 retry(); 269 270 fadd( FloatRegisterImpl::S, F0, F1, F2 ); 271 fsub( FloatRegisterImpl::D, F34, F0, F62 ); 272 273 fcmp( FloatRegisterImpl::Q, fcc0, F0, F60); 274 fcmpe( FloatRegisterImpl::S, fcc1, F31, F30); 275 276 ftox( FloatRegisterImpl::D, F2, F4 ); 277 ftoi( FloatRegisterImpl::Q, F4, F8 ); 278 279 ftof( FloatRegisterImpl::S, FloatRegisterImpl::Q, F3, F12 ); 280 281 fxtof( FloatRegisterImpl::S, F4, F5 ); 282 fitof( FloatRegisterImpl::D, F6, F8 ); 283 284 fmov( FloatRegisterImpl::Q, F16, F20 ); 285 fneg( FloatRegisterImpl::S, F6, F7 ); 286 fabs( FloatRegisterImpl::D, F10, F12 ); 287 288 fmul( FloatRegisterImpl::Q, F24, F28, F32 ); 289 fmul( FloatRegisterImpl::S, FloatRegisterImpl::D, F8, F9, F14 ); 290 fdiv( FloatRegisterImpl::S, F10, F11, F12 ); 291 292 fsqrt( FloatRegisterImpl::S, F13, F14 ); 293 294 flush( L0, L1 ); 295 flush( L2, -1 ); 296 297 flushw(); 298 299 illtrap( (1 << 22) - 2); 300 301 impdep1( 17, (1 << 19) - 1 ); 302 impdep2( 3, 0 ); 303 304 jmpl( L3, L4, L5 ); 305 delayed()->nop(); 306 jmpl( L6, -1, L7, Relocation::spec_simple(relocInfo::none)); 307 delayed()->nop(); 308 309 310 ldf( FloatRegisterImpl::S, O0, O1, F15 ); 311 ldf( FloatRegisterImpl::D, O2, -1, F14 ); 312 313 314 ldfsr( O3, O4 ); 315 ldfsr( O5, -1 ); 316 ldxfsr( O6, O7 ); 317 ldxfsr( I0, -1 ); 318 319 ldfa( FloatRegisterImpl::D, I1, I2, 1, F16 ); 320 ldfa( FloatRegisterImpl::Q, I3, -1, F36 ); 321 322 ldsb( I4, I5, I6 ); 323 ldsb( I7, -1, G0 ); 324 ldsh( G1, G3, G4 ); 325 ldsh( G5, -1, G6 ); 326 ldsw( G7, L0, L1 ); 327 ldsw( L2, -1, L3 ); 328 ldub( L4, L5, L6 ); 329 ldub( L7, -1, O0 ); 330 lduh( O1, O2, O3 ); 331 lduh( O4, -1, O5 ); 332 lduw( O6, O7, G0 ); 333 lduw( G1, -1, G2 ); 334 ldx( G3, G4, G5 ); 335 ldx( G6, -1, G7 ); 336 ldd( I0, I1, I2 ); 337 ldd( I3, -1, I4 ); 338 339 ldsba( I5, I6, 2, I7 ); 340 ldsba( L0, -1, L1 ); 341 ldsha( L2, L3, 3, L4 ); 342 ldsha( L5, -1, L6 ); 343 ldswa( L7, O0, (1 << 8) - 1, O1 ); 344 ldswa( O2, -1, O3 ); 345 lduba( O4, O5, 0, O6 ); 346 lduba( O7, -1, I0 ); 347 lduha( I1, I2, 1, I3 ); 348 lduha( I4, -1, I5 ); 349 lduwa( I6, I7, 2, L0 ); 350 lduwa( L1, -1, L2 ); 351 ldxa( L3, L4, 3, L5 ); 352 ldxa( L6, -1, L7 ); 353 ldda( G0, G1, 4, G2 ); 354 ldda( G3, -1, G4 ); 355 356 ldstub( G5, G6, G7 ); 357 ldstub( O0, -1, O1 ); 358 359 ldstuba( O2, O3, 5, O4 ); 360 ldstuba( O5, -1, O6 ); 361 362 and3( I0, L0, O0 ); 363 and3( G7, -1, O7 ); 364 andcc( L2, I2, G2 ); 365 andcc( L4, -1, G4 ); 366 andn( I5, I6, I7 ); 367 andn( I6, -1, I7 ); 368 andncc( I5, I6, I7 ); 369 andncc( I7, -1, I6 ); 370 or3( I5, I6, I7 ); 371 or3( I7, -1, I6 ); 372 orcc( I5, I6, I7 ); 373 orcc( I7, -1, I6 ); 374 orn( I5, I6, I7 ); 375 orn( I7, -1, I6 ); 376 orncc( I5, I6, I7 ); 377 orncc( I7, -1, I6 ); 378 xor3( I5, I6, I7 ); 379 xor3( I7, -1, I6 ); 380 xorcc( I5, I6, I7 ); 381 xorcc( I7, -1, I6 ); 382 xnor( I5, I6, I7 ); 383 xnor( I7, -1, I6 ); 384 xnorcc( I5, I6, I7 ); 385 xnorcc( I7, -1, I6 ); 386 387 membar( Membar_mask_bits(StoreStore | LoadStore | StoreLoad | LoadLoad | Sync | MemIssue | Lookaside ) ); 388 membar( StoreStore ); 389 membar( LoadStore ); 390 membar( StoreLoad ); 391 membar( LoadLoad ); 392 membar( Sync ); 393 membar( MemIssue ); 394 membar( Lookaside ); 395 396 fmov( FloatRegisterImpl::S, f_ordered, true, fcc2, F16, F17 ); 397 fmov( FloatRegisterImpl::D, rc_lz, L5, F18, F20 ); 398 399 movcc( overflowClear, false, icc, I6, L4 ); 400 movcc( f_unorderedOrEqual, true, fcc2, (1 << 10) - 1, O0 ); 401 402 movr( rc_nz, I5, I6, I7 ); 403 movr( rc_gz, L1, -1, L2 ); 404 405 mulx( I5, I6, I7 ); 406 mulx( I7, -1, I6 ); 407 sdivx( I5, I6, I7 ); 408 sdivx( I7, -1, I6 ); 409 udivx( I5, I6, I7 ); 410 udivx( I7, -1, I6 ); 411 412 umul( I5, I6, I7 ); 413 umul( I7, -1, I6 ); 414 smul( I5, I6, I7 ); 415 smul( I7, -1, I6 ); 416 umulcc( I5, I6, I7 ); 417 umulcc( I7, -1, I6 ); 418 smulcc( I5, I6, I7 ); 419 smulcc( I7, -1, I6 ); 420 421 mulscc( I5, I6, I7 ); 422 mulscc( I7, -1, I6 ); 423 424 nop(); 425 426 427 popc( G0, G1); 428 popc( -1, G2); 429 430 prefetch( L1, L2, severalReads ); 431 prefetch( L3, -1, oneRead ); 432 prefetcha( O3, O2, 6, severalWritesAndPossiblyReads ); 433 prefetcha( G2, -1, oneWrite ); 434 435 rett( I7, I7); 436 delayed()->nop(); 437 rett( G0, -1, relocInfo::none); 438 delayed()->nop(); 439 440 save( I5, I6, I7 ); 441 save( I7, -1, I6 ); 442 restore( I5, I6, I7 ); 443 restore( I7, -1, I6 ); 444 445 saved(); 446 restored(); 447 448 sethi( 0xaaaaaaaa, I3, Relocation::spec_simple(relocInfo::none)); 449 450 sll( I5, I6, I7 ); 451 sll( I7, 31, I6 ); 452 srl( I5, I6, I7 ); 453 srl( I7, 0, I6 ); 454 sra( I5, I6, I7 ); 455 sra( I7, 30, I6 ); 456 sllx( I5, I6, I7 ); 457 sllx( I7, 63, I6 ); 458 srlx( I5, I6, I7 ); 459 srlx( I7, 0, I6 ); 460 srax( I5, I6, I7 ); 461 srax( I7, 62, I6 ); 462 463 sir( -1 ); 464 465 stbar(); 466 467 stf( FloatRegisterImpl::Q, F40, G0, I7 ); 468 stf( FloatRegisterImpl::S, F18, I3, -1 ); 469 470 stfsr( L1, L2 ); 471 stfsr( I7, -1 ); 472 stxfsr( I6, I5 ); 473 stxfsr( L4, -1 ); 474 475 stfa( FloatRegisterImpl::D, F22, I6, I7, 7 ); 476 stfa( FloatRegisterImpl::Q, F44, G0, -1 ); 477 478 stb( L5, O2, I7 ); 479 stb( I7, I6, -1 ); 480 sth( L5, O2, I7 ); 481 sth( I7, I6, -1 ); 482 stw( L5, O2, I7 ); 483 stw( I7, I6, -1 ); 484 stx( L5, O2, I7 ); 485 stx( I7, I6, -1 ); 486 std( L5, O2, I7 ); 487 std( I7, I6, -1 ); 488 489 stba( L5, O2, I7, 8 ); 490 stba( I7, I6, -1 ); 491 stha( L5, O2, I7, 9 ); 492 stha( I7, I6, -1 ); 493 stwa( L5, O2, I7, 0 ); 494 stwa( I7, I6, -1 ); 495 stxa( L5, O2, I7, 11 ); 496 stxa( I7, I6, -1 ); 497 stda( L5, O2, I7, 12 ); 498 stda( I7, I6, -1 ); 499 500 sub( I5, I6, I7 ); 501 sub( I7, -1, I6 ); 502 subcc( I5, I6, I7 ); 503 subcc( I7, -1, I6 ); 504 subc( I5, I6, I7 ); 505 subc( I7, -1, I6 ); 506 subccc( I5, I6, I7 ); 507 subccc( I7, -1, I6 ); 508 509 swap( I5, I6, I7 ); 510 swap( I7, -1, I6 ); 511 512 swapa( G0, G1, 13, G2 ); 513 swapa( I7, -1, I6 ); 514 515 taddcc( I5, I6, I7 ); 516 taddcc( I7, -1, I6 ); 517 taddcctv( I5, I6, I7 ); 518 taddcctv( I7, -1, I6 ); 519 520 tsubcc( I5, I6, I7 ); 521 tsubcc( I7, -1, I6 ); 522 tsubcctv( I5, I6, I7 ); 523 tsubcctv( I7, -1, I6 ); 524 525 trap( overflowClear, xcc, G0, G1 ); 526 trap( lessEqual, icc, I7, 17 ); 527 528 bind(lbl2); 529 bind(lbl3); 530 531 code()->decode(); 532 } 533 534 // Generate a bunch 'o stuff unique to V8 535 void Assembler::test_v8_onlys() { 536 Label lbl1; 537 538 cb( cp_0or1or2, false, pc() - 4, relocInfo::none); 539 delayed()->nop(); 540 cb( cp_never, true, lbl1); 541 delayed()->nop(); 542 543 cpop1(1, 2, 3, 4); 544 cpop2(5, 6, 7, 8); 545 546 ldc( I0, I1, 31); 547 ldc( I2, -1, 0); 548 549 lddc( I4, I4, 30); 550 lddc( I6, 0, 1 ); 551 552 ldcsr( L0, L1, 0); 553 ldcsr( L1, (1 << 12) - 1, 17 ); 554 555 stc( 31, L4, L5); 556 stc( 30, L6, -(1 << 12) ); 557 558 stdc( 0, L7, G0); 559 stdc( 1, G1, 0 ); 560 561 stcsr( 16, G2, G3); 562 stcsr( 17, G4, 1 ); 563 564 stdcq( 4, G5, G6); 565 stdcq( 5, G7, -1 ); 566 567 bind(lbl1); 568 569 code()->decode(); 570 } 571 #endif 572 573 // Implementation of MacroAssembler 574 575 void MacroAssembler::null_check(Register reg, int offset) { 576 if (needs_explicit_null_check((intptr_t)offset)) { 577 // provoke OS NULL exception if reg = NULL by 578 // accessing M[reg] w/o changing any registers 579 ld_ptr(reg, 0, G0); 580 } 581 else { 582 // nothing to do, (later) access of M[reg + offset] 583 // will provoke OS NULL exception if reg = NULL 584 } 585 } 586 587 // Ring buffer jumps 588 589 #ifndef PRODUCT 590 void MacroAssembler::ret( bool trace ) { if (trace) { 591 mov(I7, O7); // traceable register 592 JMP(O7, 2 * BytesPerInstWord); 593 } else { 594 jmpl( I7, 2 * BytesPerInstWord, G0 ); 595 } 596 } 597 598 void MacroAssembler::retl( bool trace ) { if (trace) JMP(O7, 2 * BytesPerInstWord); 599 else jmpl( O7, 2 * BytesPerInstWord, G0 ); } 600 #endif /* PRODUCT */ 601 602 603 void MacroAssembler::jmp2(Register r1, Register r2, const char* file, int line ) { 604 assert_not_delayed(); 605 // This can only be traceable if r1 & r2 are visible after a window save 606 if (TraceJumps) { 607 #ifndef PRODUCT 608 save_frame(0); 609 verify_thread(); 610 ld(G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()), O0); 611 add(G2_thread, in_bytes(JavaThread::jmp_ring_offset()), O1); 612 sll(O0, exact_log2(4*sizeof(intptr_t)), O2); 613 add(O2, O1, O1); 614 615 add(r1->after_save(), r2->after_save(), O2); 616 set((intptr_t)file, O3); 617 set(line, O4); 618 Label L; 619 // get nearby pc, store jmp target 620 call(L, relocInfo::none); // No relocation for call to pc+0x8 621 delayed()->st(O2, O1, 0); 622 bind(L); 623 624 // store nearby pc 625 st(O7, O1, sizeof(intptr_t)); 626 // store file 627 st(O3, O1, 2*sizeof(intptr_t)); 628 // store line 629 st(O4, O1, 3*sizeof(intptr_t)); 630 add(O0, 1, O0); 631 and3(O0, JavaThread::jump_ring_buffer_size - 1, O0); 632 st(O0, G2_thread, in_bytes(JavaThread::jmp_ring_index_offset())); 633 restore(); 634 #endif /* PRODUCT */ 635 } 636 jmpl(r1, r2, G0); 637 } 638 void MacroAssembler::jmp(Register r1, int offset, const char* file, int line ) { 639 assert_not_delayed(); 640 // This can only be traceable if r1 is visible after a window save 641 if (TraceJumps) { 642 #ifndef PRODUCT 643 save_frame(0); 644 verify_thread(); 645 ld(G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()), O0); 646 add(G2_thread, in_bytes(JavaThread::jmp_ring_offset()), O1); 647 sll(O0, exact_log2(4*sizeof(intptr_t)), O2); 648 add(O2, O1, O1); 649 650 add(r1->after_save(), offset, O2); 651 set((intptr_t)file, O3); 652 set(line, O4); 653 Label L; 654 // get nearby pc, store jmp target 655 call(L, relocInfo::none); // No relocation for call to pc+0x8 656 delayed()->st(O2, O1, 0); 657 bind(L); 658 659 // store nearby pc 660 st(O7, O1, sizeof(intptr_t)); 661 // store file 662 st(O3, O1, 2*sizeof(intptr_t)); 663 // store line 664 st(O4, O1, 3*sizeof(intptr_t)); 665 add(O0, 1, O0); 666 and3(O0, JavaThread::jump_ring_buffer_size - 1, O0); 667 st(O0, G2_thread, in_bytes(JavaThread::jmp_ring_index_offset())); 668 restore(); 669 #endif /* PRODUCT */ 670 } 671 jmp(r1, offset); 672 } 673 674 // This code sequence is relocatable to any address, even on LP64. 675 void MacroAssembler::jumpl(const AddressLiteral& addrlit, Register temp, Register d, int offset, const char* file, int line) { 676 assert_not_delayed(); 677 // Force fixed length sethi because NativeJump and NativeFarCall don't handle 678 // variable length instruction streams. 679 patchable_sethi(addrlit, temp); 680 Address a(temp, addrlit.low10() + offset); // Add the offset to the displacement. 681 if (TraceJumps) { 682 #ifndef PRODUCT 683 // Must do the add here so relocation can find the remainder of the 684 // value to be relocated. 685 add(a.base(), a.disp(), a.base(), addrlit.rspec(offset)); 686 save_frame(0); 687 verify_thread(); 688 ld(G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()), O0); 689 add(G2_thread, in_bytes(JavaThread::jmp_ring_offset()), O1); 690 sll(O0, exact_log2(4*sizeof(intptr_t)), O2); 691 add(O2, O1, O1); 692 693 set((intptr_t)file, O3); 694 set(line, O4); 695 Label L; 696 697 // get nearby pc, store jmp target 698 call(L, relocInfo::none); // No relocation for call to pc+0x8 699 delayed()->st(a.base()->after_save(), O1, 0); 700 bind(L); 701 702 // store nearby pc 703 st(O7, O1, sizeof(intptr_t)); 704 // store file 705 st(O3, O1, 2*sizeof(intptr_t)); 706 // store line 707 st(O4, O1, 3*sizeof(intptr_t)); 708 add(O0, 1, O0); 709 and3(O0, JavaThread::jump_ring_buffer_size - 1, O0); 710 st(O0, G2_thread, in_bytes(JavaThread::jmp_ring_index_offset())); 711 restore(); 712 jmpl(a.base(), G0, d); 713 #else 714 jmpl(a.base(), a.disp(), d); 715 #endif /* PRODUCT */ 716 } else { 717 jmpl(a.base(), a.disp(), d); 718 } 719 } 720 721 void MacroAssembler::jump(const AddressLiteral& addrlit, Register temp, int offset, const char* file, int line) { 722 jumpl(addrlit, temp, G0, offset, file, line); 723 } 724 725 726 // Convert to C varargs format 727 void MacroAssembler::set_varargs( Argument inArg, Register d ) { 728 // spill register-resident args to their memory slots 729 // (SPARC calling convention requires callers to have already preallocated these) 730 // Note that the inArg might in fact be an outgoing argument, 731 // if a leaf routine or stub does some tricky argument shuffling. 732 // This routine must work even though one of the saved arguments 733 // is in the d register (e.g., set_varargs(Argument(0, false), O0)). 734 for (Argument savePtr = inArg; 735 savePtr.is_register(); 736 savePtr = savePtr.successor()) { 737 st_ptr(savePtr.as_register(), savePtr.address_in_frame()); 738 } 739 // return the address of the first memory slot 740 Address a = inArg.address_in_frame(); 741 add(a.base(), a.disp(), d); 742 } 743 744 // Conditional breakpoint (for assertion checks in assembly code) 745 void MacroAssembler::breakpoint_trap(Condition c, CC cc) { 746 trap(c, cc, G0, ST_RESERVED_FOR_USER_0); 747 } 748 749 // We want to use ST_BREAKPOINT here, but the debugger is confused by it. 750 void MacroAssembler::breakpoint_trap() { 751 trap(ST_RESERVED_FOR_USER_0); 752 } 753 754 // flush windows (except current) using flushw instruction if avail. 755 void MacroAssembler::flush_windows() { 756 if (VM_Version::v9_instructions_work()) flushw(); 757 else flush_windows_trap(); 758 } 759 760 // Write serialization page so VM thread can do a pseudo remote membar 761 // We use the current thread pointer to calculate a thread specific 762 // offset to write to within the page. This minimizes bus traffic 763 // due to cache line collision. 764 void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) { 765 srl(thread, os::get_serialize_page_shift_count(), tmp2); 766 if (Assembler::is_simm13(os::vm_page_size())) { 767 and3(tmp2, (os::vm_page_size() - sizeof(int)), tmp2); 768 } 769 else { 770 set((os::vm_page_size() - sizeof(int)), tmp1); 771 and3(tmp2, tmp1, tmp2); 772 } 773 set(os::get_memory_serialize_page(), tmp1); 774 st(G0, tmp1, tmp2); 775 } 776 777 778 779 void MacroAssembler::enter() { 780 Unimplemented(); 781 } 782 783 void MacroAssembler::leave() { 784 Unimplemented(); 785 } 786 787 void MacroAssembler::mult(Register s1, Register s2, Register d) { 788 if(VM_Version::v9_instructions_work()) { 789 mulx (s1, s2, d); 790 } else { 791 smul (s1, s2, d); 792 } 793 } 794 795 void MacroAssembler::mult(Register s1, int simm13a, Register d) { 796 if(VM_Version::v9_instructions_work()) { 797 mulx (s1, simm13a, d); 798 } else { 799 smul (s1, simm13a, d); 800 } 801 } 802 803 804 #ifdef ASSERT 805 void MacroAssembler::read_ccr_v8_assert(Register ccr_save) { 806 const Register s1 = G3_scratch; 807 const Register s2 = G4_scratch; 808 Label get_psr_test; 809 // Get the condition codes the V8 way. 810 read_ccr_trap(s1); 811 mov(ccr_save, s2); 812 // This is a test of V8 which has icc but not xcc 813 // so mask off the xcc bits 814 and3(s2, 0xf, s2); 815 // Compare condition codes from the V8 and V9 ways. 816 subcc(s2, s1, G0); 817 br(Assembler::notEqual, true, Assembler::pt, get_psr_test); 818 delayed()->breakpoint_trap(); 819 bind(get_psr_test); 820 } 821 822 void MacroAssembler::write_ccr_v8_assert(Register ccr_save) { 823 const Register s1 = G3_scratch; 824 const Register s2 = G4_scratch; 825 Label set_psr_test; 826 // Write out the saved condition codes the V8 way 827 write_ccr_trap(ccr_save, s1, s2); 828 // Read back the condition codes using the V9 instruction 829 rdccr(s1); 830 mov(ccr_save, s2); 831 // This is a test of V8 which has icc but not xcc 832 // so mask off the xcc bits 833 and3(s2, 0xf, s2); 834 and3(s1, 0xf, s1); 835 // Compare the V8 way with the V9 way. 836 subcc(s2, s1, G0); 837 br(Assembler::notEqual, true, Assembler::pt, set_psr_test); 838 delayed()->breakpoint_trap(); 839 bind(set_psr_test); 840 } 841 #else 842 #define read_ccr_v8_assert(x) 843 #define write_ccr_v8_assert(x) 844 #endif // ASSERT 845 846 void MacroAssembler::read_ccr(Register ccr_save) { 847 if (VM_Version::v9_instructions_work()) { 848 rdccr(ccr_save); 849 // Test code sequence used on V8. Do not move above rdccr. 850 read_ccr_v8_assert(ccr_save); 851 } else { 852 read_ccr_trap(ccr_save); 853 } 854 } 855 856 void MacroAssembler::write_ccr(Register ccr_save) { 857 if (VM_Version::v9_instructions_work()) { 858 // Test code sequence used on V8. Do not move below wrccr. 859 write_ccr_v8_assert(ccr_save); 860 wrccr(ccr_save); 861 } else { 862 const Register temp_reg1 = G3_scratch; 863 const Register temp_reg2 = G4_scratch; 864 write_ccr_trap(ccr_save, temp_reg1, temp_reg2); 865 } 866 } 867 868 869 // Calls to C land 870 871 #ifdef ASSERT 872 // a hook for debugging 873 static Thread* reinitialize_thread() { 874 return ThreadLocalStorage::thread(); 875 } 876 #else 877 #define reinitialize_thread ThreadLocalStorage::thread 878 #endif 879 880 #ifdef ASSERT 881 address last_get_thread = NULL; 882 #endif 883 884 // call this when G2_thread is not known to be valid 885 void MacroAssembler::get_thread() { 886 save_frame(0); // to avoid clobbering O0 887 mov(G1, L0); // avoid clobbering G1 888 mov(G5_method, L1); // avoid clobbering G5 889 mov(G3, L2); // avoid clobbering G3 also 890 mov(G4, L5); // avoid clobbering G4 891 #ifdef ASSERT 892 AddressLiteral last_get_thread_addrlit(&last_get_thread); 893 set(last_get_thread_addrlit, L3); 894 inc(L4, get_pc(L4) + 2 * BytesPerInstWord); // skip getpc() code + inc + st_ptr to point L4 at call 895 st_ptr(L4, L3, 0); 896 #endif 897 call(CAST_FROM_FN_PTR(address, reinitialize_thread), relocInfo::runtime_call_type); 898 delayed()->nop(); 899 mov(L0, G1); 900 mov(L1, G5_method); 901 mov(L2, G3); 902 mov(L5, G4); 903 restore(O0, 0, G2_thread); 904 } 905 906 static Thread* verify_thread_subroutine(Thread* gthread_value) { 907 Thread* correct_value = ThreadLocalStorage::thread(); 908 guarantee(gthread_value == correct_value, "G2_thread value must be the thread"); 909 return correct_value; 910 } 911 912 void MacroAssembler::verify_thread() { 913 if (VerifyThread) { 914 // NOTE: this chops off the heads of the 64-bit O registers. 915 #ifdef CC_INTERP 916 save_frame(0); 917 #else 918 // make sure G2_thread contains the right value 919 save_frame_and_mov(0, Lmethod, Lmethod); // to avoid clobbering O0 (and propagate Lmethod for -Xprof) 920 mov(G1, L1); // avoid clobbering G1 921 // G2 saved below 922 mov(G3, L3); // avoid clobbering G3 923 mov(G4, L4); // avoid clobbering G4 924 mov(G5_method, L5); // avoid clobbering G5_method 925 #endif /* CC_INTERP */ 926 #if defined(COMPILER2) && !defined(_LP64) 927 // Save & restore possible 64-bit Long arguments in G-regs 928 srlx(G1,32,L0); 929 srlx(G4,32,L6); 930 #endif 931 call(CAST_FROM_FN_PTR(address,verify_thread_subroutine), relocInfo::runtime_call_type); 932 delayed()->mov(G2_thread, O0); 933 934 mov(L1, G1); // Restore G1 935 // G2 restored below 936 mov(L3, G3); // restore G3 937 mov(L4, G4); // restore G4 938 mov(L5, G5_method); // restore G5_method 939 #if defined(COMPILER2) && !defined(_LP64) 940 // Save & restore possible 64-bit Long arguments in G-regs 941 sllx(L0,32,G2); // Move old high G1 bits high in G2 942 srl(G1, 0,G1); // Clear current high G1 bits 943 or3 (G1,G2,G1); // Recover 64-bit G1 944 sllx(L6,32,G2); // Move old high G4 bits high in G2 945 srl(G4, 0,G4); // Clear current high G4 bits 946 or3 (G4,G2,G4); // Recover 64-bit G4 947 #endif 948 restore(O0, 0, G2_thread); 949 } 950 } 951 952 953 void MacroAssembler::save_thread(const Register thread_cache) { 954 verify_thread(); 955 if (thread_cache->is_valid()) { 956 assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile"); 957 mov(G2_thread, thread_cache); 958 } 959 if (VerifyThread) { 960 // smash G2_thread, as if the VM were about to anyway 961 set(0x67676767, G2_thread); 962 } 963 } 964 965 966 void MacroAssembler::restore_thread(const Register thread_cache) { 967 if (thread_cache->is_valid()) { 968 assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile"); 969 mov(thread_cache, G2_thread); 970 verify_thread(); 971 } else { 972 // do it the slow way 973 get_thread(); 974 } 975 } 976 977 978 // %%% maybe get rid of [re]set_last_Java_frame 979 void MacroAssembler::set_last_Java_frame(Register last_java_sp, Register last_Java_pc) { 980 assert_not_delayed(); 981 Address flags(G2_thread, JavaThread::frame_anchor_offset() + 982 JavaFrameAnchor::flags_offset()); 983 Address pc_addr(G2_thread, JavaThread::last_Java_pc_offset()); 984 985 // Always set last_Java_pc and flags first because once last_Java_sp is visible 986 // has_last_Java_frame is true and users will look at the rest of the fields. 987 // (Note: flags should always be zero before we get here so doesn't need to be set.) 988 989 #ifdef ASSERT 990 // Verify that flags was zeroed on return to Java 991 Label PcOk; 992 save_frame(0); // to avoid clobbering O0 993 ld_ptr(pc_addr, L0); 994 br_null(L0, false, Assembler::pt, PcOk); 995 stop("last_Java_pc not zeroed before leaving Java"); 996 bind(PcOk); 997 998 // Verify that flags was zeroed on return to Java 999 Label FlagsOk; 1000 ld(flags, L0); 1001 tst(L0); 1002 br(Assembler::zero, false, Assembler::pt, FlagsOk); 1003 delayed() -> restore(); 1004 stop("flags not zeroed before leaving Java"); 1005 bind(FlagsOk); 1006 #endif /* ASSERT */ 1007 // 1008 // When returning from calling out from Java mode the frame anchor's last_Java_pc 1009 // will always be set to NULL. It is set here so that if we are doing a call to 1010 // native (not VM) that we capture the known pc and don't have to rely on the 1011 // native call having a standard frame linkage where we can find the pc. 1012 1013 if (last_Java_pc->is_valid()) { 1014 st_ptr(last_Java_pc, pc_addr); 1015 } 1016 1017 #ifdef _LP64 1018 #ifdef ASSERT 1019 // Make sure that we have an odd stack 1020 Label StackOk; 1021 andcc(last_java_sp, 0x01, G0); 1022 br(Assembler::notZero, false, Assembler::pt, StackOk); 1023 delayed()->nop(); 1024 stop("Stack Not Biased in set_last_Java_frame"); 1025 bind(StackOk); 1026 #endif // ASSERT 1027 assert( last_java_sp != G4_scratch, "bad register usage in set_last_Java_frame"); 1028 add( last_java_sp, STACK_BIAS, G4_scratch ); 1029 st_ptr(G4_scratch, G2_thread, JavaThread::last_Java_sp_offset()); 1030 #else 1031 st_ptr(last_java_sp, G2_thread, JavaThread::last_Java_sp_offset()); 1032 #endif // _LP64 1033 } 1034 1035 void MacroAssembler::reset_last_Java_frame(void) { 1036 assert_not_delayed(); 1037 1038 Address sp_addr(G2_thread, JavaThread::last_Java_sp_offset()); 1039 Address pc_addr(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 1040 Address flags (G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 1041 1042 #ifdef ASSERT 1043 // check that it WAS previously set 1044 #ifdef CC_INTERP 1045 save_frame(0); 1046 #else 1047 save_frame_and_mov(0, Lmethod, Lmethod); // Propagate Lmethod to helper frame for -Xprof 1048 #endif /* CC_INTERP */ 1049 ld_ptr(sp_addr, L0); 1050 tst(L0); 1051 breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 1052 restore(); 1053 #endif // ASSERT 1054 1055 st_ptr(G0, sp_addr); 1056 // Always return last_Java_pc to zero 1057 st_ptr(G0, pc_addr); 1058 // Always null flags after return to Java 1059 st(G0, flags); 1060 } 1061 1062 1063 void MacroAssembler::call_VM_base( 1064 Register oop_result, 1065 Register thread_cache, 1066 Register last_java_sp, 1067 address entry_point, 1068 int number_of_arguments, 1069 bool check_exceptions) 1070 { 1071 assert_not_delayed(); 1072 1073 // determine last_java_sp register 1074 if (!last_java_sp->is_valid()) { 1075 last_java_sp = SP; 1076 } 1077 // debugging support 1078 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 1079 1080 // 64-bit last_java_sp is biased! 1081 set_last_Java_frame(last_java_sp, noreg); 1082 if (VerifyThread) mov(G2_thread, O0); // about to be smashed; pass early 1083 save_thread(thread_cache); 1084 // do the call 1085 call(entry_point, relocInfo::runtime_call_type); 1086 if (!VerifyThread) 1087 delayed()->mov(G2_thread, O0); // pass thread as first argument 1088 else 1089 delayed()->nop(); // (thread already passed) 1090 restore_thread(thread_cache); 1091 reset_last_Java_frame(); 1092 1093 // check for pending exceptions. use Gtemp as scratch register. 1094 if (check_exceptions) { 1095 check_and_forward_exception(Gtemp); 1096 } 1097 1098 #ifdef ASSERT 1099 set(badHeapWordVal, G3); 1100 set(badHeapWordVal, G4); 1101 set(badHeapWordVal, G5); 1102 #endif 1103 1104 // get oop result if there is one and reset the value in the thread 1105 if (oop_result->is_valid()) { 1106 get_vm_result(oop_result); 1107 } 1108 } 1109 1110 void MacroAssembler::check_and_forward_exception(Register scratch_reg) 1111 { 1112 Label L; 1113 1114 check_and_handle_popframe(scratch_reg); 1115 check_and_handle_earlyret(scratch_reg); 1116 1117 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 1118 ld_ptr(exception_addr, scratch_reg); 1119 br_null(scratch_reg,false,pt,L); 1120 // we use O7 linkage so that forward_exception_entry has the issuing PC 1121 call(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 1122 delayed()->nop(); 1123 bind(L); 1124 } 1125 1126 1127 void MacroAssembler::check_and_handle_popframe(Register scratch_reg) { 1128 } 1129 1130 1131 void MacroAssembler::check_and_handle_earlyret(Register scratch_reg) { 1132 } 1133 1134 1135 void MacroAssembler::call_VM(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 1136 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions); 1137 } 1138 1139 1140 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, bool check_exceptions) { 1141 // O0 is reserved for the thread 1142 mov(arg_1, O1); 1143 call_VM(oop_result, entry_point, 1, check_exceptions); 1144 } 1145 1146 1147 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) { 1148 // O0 is reserved for the thread 1149 mov(arg_1, O1); 1150 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 1151 call_VM(oop_result, entry_point, 2, check_exceptions); 1152 } 1153 1154 1155 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, Register arg_3, bool check_exceptions) { 1156 // O0 is reserved for the thread 1157 mov(arg_1, O1); 1158 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 1159 mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument"); 1160 call_VM(oop_result, entry_point, 3, check_exceptions); 1161 } 1162 1163 1164 1165 // Note: The following call_VM overloadings are useful when a "save" 1166 // has already been performed by a stub, and the last Java frame is 1167 // the previous one. In that case, last_java_sp must be passed as FP 1168 // instead of SP. 1169 1170 1171 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, int number_of_arguments, bool check_exceptions) { 1172 call_VM_base(oop_result, noreg, last_java_sp, entry_point, number_of_arguments, check_exceptions); 1173 } 1174 1175 1176 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, bool check_exceptions) { 1177 // O0 is reserved for the thread 1178 mov(arg_1, O1); 1179 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 1180 } 1181 1182 1183 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) { 1184 // O0 is reserved for the thread 1185 mov(arg_1, O1); 1186 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 1187 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 1188 } 1189 1190 1191 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) { 1192 // O0 is reserved for the thread 1193 mov(arg_1, O1); 1194 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument"); 1195 mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument"); 1196 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 1197 } 1198 1199 1200 1201 void MacroAssembler::call_VM_leaf_base(Register thread_cache, address entry_point, int number_of_arguments) { 1202 assert_not_delayed(); 1203 save_thread(thread_cache); 1204 // do the call 1205 call(entry_point, relocInfo::runtime_call_type); 1206 delayed()->nop(); 1207 restore_thread(thread_cache); 1208 #ifdef ASSERT 1209 set(badHeapWordVal, G3); 1210 set(badHeapWordVal, G4); 1211 set(badHeapWordVal, G5); 1212 #endif 1213 } 1214 1215 1216 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, int number_of_arguments) { 1217 call_VM_leaf_base(thread_cache, entry_point, number_of_arguments); 1218 } 1219 1220 1221 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1) { 1222 mov(arg_1, O0); 1223 call_VM_leaf(thread_cache, entry_point, 1); 1224 } 1225 1226 1227 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2) { 1228 mov(arg_1, O0); 1229 mov(arg_2, O1); assert(arg_2 != O0, "smashed argument"); 1230 call_VM_leaf(thread_cache, entry_point, 2); 1231 } 1232 1233 1234 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2, Register arg_3) { 1235 mov(arg_1, O0); 1236 mov(arg_2, O1); assert(arg_2 != O0, "smashed argument"); 1237 mov(arg_3, O2); assert(arg_3 != O0 && arg_3 != O1, "smashed argument"); 1238 call_VM_leaf(thread_cache, entry_point, 3); 1239 } 1240 1241 1242 void MacroAssembler::get_vm_result(Register oop_result) { 1243 verify_thread(); 1244 Address vm_result_addr(G2_thread, JavaThread::vm_result_offset()); 1245 ld_ptr( vm_result_addr, oop_result); 1246 st_ptr(G0, vm_result_addr); 1247 verify_oop(oop_result); 1248 } 1249 1250 1251 void MacroAssembler::get_vm_result_2(Register oop_result) { 1252 verify_thread(); 1253 Address vm_result_addr_2(G2_thread, JavaThread::vm_result_2_offset()); 1254 ld_ptr(vm_result_addr_2, oop_result); 1255 st_ptr(G0, vm_result_addr_2); 1256 verify_oop(oop_result); 1257 } 1258 1259 1260 // We require that C code which does not return a value in vm_result will 1261 // leave it undisturbed. 1262 void MacroAssembler::set_vm_result(Register oop_result) { 1263 verify_thread(); 1264 Address vm_result_addr(G2_thread, JavaThread::vm_result_offset()); 1265 verify_oop(oop_result); 1266 1267 # ifdef ASSERT 1268 // Check that we are not overwriting any other oop. 1269 #ifdef CC_INTERP 1270 save_frame(0); 1271 #else 1272 save_frame_and_mov(0, Lmethod, Lmethod); // Propagate Lmethod for -Xprof 1273 #endif /* CC_INTERP */ 1274 ld_ptr(vm_result_addr, L0); 1275 tst(L0); 1276 restore(); 1277 breakpoint_trap(notZero, Assembler::ptr_cc); 1278 // } 1279 # endif 1280 1281 st_ptr(oop_result, vm_result_addr); 1282 } 1283 1284 1285 void MacroAssembler::card_table_write(jbyte* byte_map_base, 1286 Register tmp, Register obj) { 1287 #ifdef _LP64 1288 srlx(obj, CardTableModRefBS::card_shift, obj); 1289 #else 1290 srl(obj, CardTableModRefBS::card_shift, obj); 1291 #endif 1292 assert(tmp != obj, "need separate temp reg"); 1293 set((address) byte_map_base, tmp); 1294 stb(G0, tmp, obj); 1295 } 1296 1297 1298 void MacroAssembler::internal_sethi(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) { 1299 address save_pc; 1300 int shiftcnt; 1301 #ifdef _LP64 1302 # ifdef CHECK_DELAY 1303 assert_not_delayed((char*) "cannot put two instructions in delay slot"); 1304 # endif 1305 v9_dep(); 1306 save_pc = pc(); 1307 1308 int msb32 = (int) (addrlit.value() >> 32); 1309 int lsb32 = (int) (addrlit.value()); 1310 1311 if (msb32 == 0 && lsb32 >= 0) { 1312 Assembler::sethi(lsb32, d, addrlit.rspec()); 1313 } 1314 else if (msb32 == -1) { 1315 Assembler::sethi(~lsb32, d, addrlit.rspec()); 1316 xor3(d, ~low10(~0), d); 1317 } 1318 else { 1319 Assembler::sethi(msb32, d, addrlit.rspec()); // msb 22-bits 1320 if (msb32 & 0x3ff) // Any bits? 1321 or3(d, msb32 & 0x3ff, d); // msb 32-bits are now in lsb 32 1322 if (lsb32 & 0xFFFFFC00) { // done? 1323 if ((lsb32 >> 20) & 0xfff) { // Any bits set? 1324 sllx(d, 12, d); // Make room for next 12 bits 1325 or3(d, (lsb32 >> 20) & 0xfff, d); // Or in next 12 1326 shiftcnt = 0; // We already shifted 1327 } 1328 else 1329 shiftcnt = 12; 1330 if ((lsb32 >> 10) & 0x3ff) { 1331 sllx(d, shiftcnt + 10, d); // Make room for last 10 bits 1332 or3(d, (lsb32 >> 10) & 0x3ff, d); // Or in next 10 1333 shiftcnt = 0; 1334 } 1335 else 1336 shiftcnt = 10; 1337 sllx(d, shiftcnt + 10, d); // Shift leaving disp field 0'd 1338 } 1339 else 1340 sllx(d, 32, d); 1341 } 1342 // Pad out the instruction sequence so it can be patched later. 1343 if (ForceRelocatable || (addrlit.rtype() != relocInfo::none && 1344 addrlit.rtype() != relocInfo::runtime_call_type)) { 1345 while (pc() < (save_pc + (7 * BytesPerInstWord))) 1346 nop(); 1347 } 1348 #else 1349 Assembler::sethi(addrlit.value(), d, addrlit.rspec()); 1350 #endif 1351 } 1352 1353 1354 void MacroAssembler::sethi(const AddressLiteral& addrlit, Register d) { 1355 internal_sethi(addrlit, d, false); 1356 } 1357 1358 1359 void MacroAssembler::patchable_sethi(const AddressLiteral& addrlit, Register d) { 1360 internal_sethi(addrlit, d, true); 1361 } 1362 1363 1364 int MacroAssembler::insts_for_sethi(address a, bool worst_case) { 1365 #ifdef _LP64 1366 if (worst_case) return 7; 1367 intptr_t iaddr = (intptr_t) a; 1368 int msb32 = (int) (iaddr >> 32); 1369 int lsb32 = (int) (iaddr); 1370 int count; 1371 if (msb32 == 0 && lsb32 >= 0) 1372 count = 1; 1373 else if (msb32 == -1) 1374 count = 2; 1375 else { 1376 count = 2; 1377 if (msb32 & 0x3ff) 1378 count++; 1379 if (lsb32 & 0xFFFFFC00 ) { 1380 if ((lsb32 >> 20) & 0xfff) count += 2; 1381 if ((lsb32 >> 10) & 0x3ff) count += 2; 1382 } 1383 } 1384 return count; 1385 #else 1386 return 1; 1387 #endif 1388 } 1389 1390 int MacroAssembler::worst_case_insts_for_set() { 1391 return insts_for_sethi(NULL, true) + 1; 1392 } 1393 1394 1395 // Keep in sync with MacroAssembler::insts_for_internal_set 1396 void MacroAssembler::internal_set(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) { 1397 intptr_t value = addrlit.value(); 1398 1399 if (!ForceRelocatable && addrlit.rspec().type() == relocInfo::none) { 1400 // can optimize 1401 if (-4096 <= value && value <= 4095) { 1402 or3(G0, value, d); // setsw (this leaves upper 32 bits sign-extended) 1403 return; 1404 } 1405 if (inv_hi22(hi22(value)) == value) { 1406 sethi(addrlit, d); 1407 return; 1408 } 1409 } 1410 assert_not_delayed((char*) "cannot put two instructions in delay slot"); 1411 internal_sethi(addrlit, d, ForceRelocatable); 1412 if (ForceRelocatable || addrlit.rspec().type() != relocInfo::none || addrlit.low10() != 0) { 1413 add(d, addrlit.low10(), d, addrlit.rspec()); 1414 } 1415 } 1416 1417 // Keep in sync with MacroAssembler::internal_set 1418 int MacroAssembler::insts_for_internal_set(intptr_t value) { 1419 // can optimize 1420 if (-4096 <= value && value <= 4095) { 1421 return 1; 1422 } 1423 if (inv_hi22(hi22(value)) == value) { 1424 return insts_for_sethi((address) value); 1425 } 1426 int count = insts_for_sethi((address) value); 1427 AddressLiteral al(value); 1428 if (al.low10() != 0) { 1429 count++; 1430 } 1431 return count; 1432 } 1433 1434 void MacroAssembler::set(const AddressLiteral& al, Register d) { 1435 internal_set(al, d, false); 1436 } 1437 1438 void MacroAssembler::set(intptr_t value, Register d) { 1439 AddressLiteral al(value); 1440 internal_set(al, d, false); 1441 } 1442 1443 void MacroAssembler::set(address addr, Register d, RelocationHolder const& rspec) { 1444 AddressLiteral al(addr, rspec); 1445 internal_set(al, d, false); 1446 } 1447 1448 void MacroAssembler::patchable_set(const AddressLiteral& al, Register d) { 1449 internal_set(al, d, true); 1450 } 1451 1452 void MacroAssembler::patchable_set(intptr_t value, Register d) { 1453 AddressLiteral al(value); 1454 internal_set(al, d, true); 1455 } 1456 1457 1458 void MacroAssembler::set64(jlong value, Register d, Register tmp) { 1459 assert_not_delayed(); 1460 v9_dep(); 1461 1462 int hi = (int)(value >> 32); 1463 int lo = (int)(value & ~0); 1464 // (Matcher::isSimpleConstant64 knows about the following optimizations.) 1465 if (Assembler::is_simm13(lo) && value == lo) { 1466 or3(G0, lo, d); 1467 } else if (hi == 0) { 1468 Assembler::sethi(lo, d); // hardware version zero-extends to upper 32 1469 if (low10(lo) != 0) 1470 or3(d, low10(lo), d); 1471 } 1472 else if (hi == -1) { 1473 Assembler::sethi(~lo, d); // hardware version zero-extends to upper 32 1474 xor3(d, low10(lo) ^ ~low10(~0), d); 1475 } 1476 else if (lo == 0) { 1477 if (Assembler::is_simm13(hi)) { 1478 or3(G0, hi, d); 1479 } else { 1480 Assembler::sethi(hi, d); // hardware version zero-extends to upper 32 1481 if (low10(hi) != 0) 1482 or3(d, low10(hi), d); 1483 } 1484 sllx(d, 32, d); 1485 } 1486 else { 1487 Assembler::sethi(hi, tmp); 1488 Assembler::sethi(lo, d); // macro assembler version sign-extends 1489 if (low10(hi) != 0) 1490 or3 (tmp, low10(hi), tmp); 1491 if (low10(lo) != 0) 1492 or3 ( d, low10(lo), d); 1493 sllx(tmp, 32, tmp); 1494 or3 (d, tmp, d); 1495 } 1496 } 1497 1498 int MacroAssembler::insts_for_set64(jlong value) { 1499 v9_dep(); 1500 1501 int hi = (int) (value >> 32); 1502 int lo = (int) (value & ~0); 1503 int count = 0; 1504 1505 // (Matcher::isSimpleConstant64 knows about the following optimizations.) 1506 if (Assembler::is_simm13(lo) && value == lo) { 1507 count++; 1508 } else if (hi == 0) { 1509 count++; 1510 if (low10(lo) != 0) 1511 count++; 1512 } 1513 else if (hi == -1) { 1514 count += 2; 1515 } 1516 else if (lo == 0) { 1517 if (Assembler::is_simm13(hi)) { 1518 count++; 1519 } else { 1520 count++; 1521 if (low10(hi) != 0) 1522 count++; 1523 } 1524 count++; 1525 } 1526 else { 1527 count += 2; 1528 if (low10(hi) != 0) 1529 count++; 1530 if (low10(lo) != 0) 1531 count++; 1532 count += 2; 1533 } 1534 return count; 1535 } 1536 1537 // compute size in bytes of sparc frame, given 1538 // number of extraWords 1539 int MacroAssembler::total_frame_size_in_bytes(int extraWords) { 1540 1541 int nWords = frame::memory_parameter_word_sp_offset; 1542 1543 nWords += extraWords; 1544 1545 if (nWords & 1) ++nWords; // round up to double-word 1546 1547 return nWords * BytesPerWord; 1548 } 1549 1550 1551 // save_frame: given number of "extra" words in frame, 1552 // issue approp. save instruction (p 200, v8 manual) 1553 1554 void MacroAssembler::save_frame(int extraWords) { 1555 int delta = -total_frame_size_in_bytes(extraWords); 1556 if (is_simm13(delta)) { 1557 save(SP, delta, SP); 1558 } else { 1559 set(delta, G3_scratch); 1560 save(SP, G3_scratch, SP); 1561 } 1562 } 1563 1564 1565 void MacroAssembler::save_frame_c1(int size_in_bytes) { 1566 if (is_simm13(-size_in_bytes)) { 1567 save(SP, -size_in_bytes, SP); 1568 } else { 1569 set(-size_in_bytes, G3_scratch); 1570 save(SP, G3_scratch, SP); 1571 } 1572 } 1573 1574 1575 void MacroAssembler::save_frame_and_mov(int extraWords, 1576 Register s1, Register d1, 1577 Register s2, Register d2) { 1578 assert_not_delayed(); 1579 1580 // The trick here is to use precisely the same memory word 1581 // that trap handlers also use to save the register. 1582 // This word cannot be used for any other purpose, but 1583 // it works fine to save the register's value, whether or not 1584 // an interrupt flushes register windows at any given moment! 1585 Address s1_addr; 1586 if (s1->is_valid() && (s1->is_in() || s1->is_local())) { 1587 s1_addr = s1->address_in_saved_window(); 1588 st_ptr(s1, s1_addr); 1589 } 1590 1591 Address s2_addr; 1592 if (s2->is_valid() && (s2->is_in() || s2->is_local())) { 1593 s2_addr = s2->address_in_saved_window(); 1594 st_ptr(s2, s2_addr); 1595 } 1596 1597 save_frame(extraWords); 1598 1599 if (s1_addr.base() == SP) { 1600 ld_ptr(s1_addr.after_save(), d1); 1601 } else if (s1->is_valid()) { 1602 mov(s1->after_save(), d1); 1603 } 1604 1605 if (s2_addr.base() == SP) { 1606 ld_ptr(s2_addr.after_save(), d2); 1607 } else if (s2->is_valid()) { 1608 mov(s2->after_save(), d2); 1609 } 1610 } 1611 1612 1613 AddressLiteral MacroAssembler::allocate_oop_address(jobject obj) { 1614 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); 1615 int oop_index = oop_recorder()->allocate_index(obj); 1616 return AddressLiteral(obj, oop_Relocation::spec(oop_index)); 1617 } 1618 1619 1620 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) { 1621 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); 1622 int oop_index = oop_recorder()->find_index(obj); 1623 return AddressLiteral(obj, oop_Relocation::spec(oop_index)); 1624 } 1625 1626 void MacroAssembler::set_narrow_oop(jobject obj, Register d) { 1627 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); 1628 int oop_index = oop_recorder()->find_index(obj); 1629 RelocationHolder rspec = oop_Relocation::spec(oop_index); 1630 1631 assert_not_delayed(); 1632 // Relocation with special format (see relocInfo_sparc.hpp). 1633 relocate(rspec, 1); 1634 // Assembler::sethi(0x3fffff, d); 1635 emit_long( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(0x3fffff) ); 1636 // Don't add relocation for 'add'. Do patching during 'sethi' processing. 1637 add(d, 0x3ff, d); 1638 1639 } 1640 1641 1642 void MacroAssembler::align(int modulus) { 1643 while (offset() % modulus != 0) nop(); 1644 } 1645 1646 1647 void MacroAssembler::safepoint() { 1648 relocate(breakpoint_Relocation::spec(breakpoint_Relocation::safepoint)); 1649 } 1650 1651 1652 void RegistersForDebugging::print(outputStream* s) { 1653 int j; 1654 for ( j = 0; j < 8; ++j ) 1655 if ( j != 6 ) s->print_cr("i%d = 0x%.16lx", j, i[j]); 1656 else s->print_cr( "fp = 0x%.16lx", i[j]); 1657 s->cr(); 1658 1659 for ( j = 0; j < 8; ++j ) 1660 s->print_cr("l%d = 0x%.16lx", j, l[j]); 1661 s->cr(); 1662 1663 for ( j = 0; j < 8; ++j ) 1664 if ( j != 6 ) s->print_cr("o%d = 0x%.16lx", j, o[j]); 1665 else s->print_cr( "sp = 0x%.16lx", o[j]); 1666 s->cr(); 1667 1668 for ( j = 0; j < 8; ++j ) 1669 s->print_cr("g%d = 0x%.16lx", j, g[j]); 1670 s->cr(); 1671 1672 // print out floats with compression 1673 for (j = 0; j < 32; ) { 1674 jfloat val = f[j]; 1675 int last = j; 1676 for ( ; last+1 < 32; ++last ) { 1677 char b1[1024], b2[1024]; 1678 sprintf(b1, "%f", val); 1679 sprintf(b2, "%f", f[last+1]); 1680 if (strcmp(b1, b2)) 1681 break; 1682 } 1683 s->print("f%d", j); 1684 if ( j != last ) s->print(" - f%d", last); 1685 s->print(" = %f", val); 1686 s->fill_to(25); 1687 s->print_cr(" (0x%x)", val); 1688 j = last + 1; 1689 } 1690 s->cr(); 1691 1692 // and doubles (evens only) 1693 for (j = 0; j < 32; ) { 1694 jdouble val = d[j]; 1695 int last = j; 1696 for ( ; last+1 < 32; ++last ) { 1697 char b1[1024], b2[1024]; 1698 sprintf(b1, "%f", val); 1699 sprintf(b2, "%f", d[last+1]); 1700 if (strcmp(b1, b2)) 1701 break; 1702 } 1703 s->print("d%d", 2 * j); 1704 if ( j != last ) s->print(" - d%d", last); 1705 s->print(" = %f", val); 1706 s->fill_to(30); 1707 s->print("(0x%x)", *(int*)&val); 1708 s->fill_to(42); 1709 s->print_cr("(0x%x)", *(1 + (int*)&val)); 1710 j = last + 1; 1711 } 1712 s->cr(); 1713 } 1714 1715 void RegistersForDebugging::save_registers(MacroAssembler* a) { 1716 a->sub(FP, round_to(sizeof(RegistersForDebugging), sizeof(jdouble)) - STACK_BIAS, O0); 1717 a->flush_windows(); 1718 int i; 1719 for (i = 0; i < 8; ++i) { 1720 a->ld_ptr(as_iRegister(i)->address_in_saved_window().after_save(), L1); a->st_ptr( L1, O0, i_offset(i)); 1721 a->ld_ptr(as_lRegister(i)->address_in_saved_window().after_save(), L1); a->st_ptr( L1, O0, l_offset(i)); 1722 a->st_ptr(as_oRegister(i)->after_save(), O0, o_offset(i)); 1723 a->st_ptr(as_gRegister(i)->after_save(), O0, g_offset(i)); 1724 } 1725 for (i = 0; i < 32; ++i) { 1726 a->stf(FloatRegisterImpl::S, as_FloatRegister(i), O0, f_offset(i)); 1727 } 1728 for (i = 0; i < (VM_Version::v9_instructions_work() ? 64 : 32); i += 2) { 1729 a->stf(FloatRegisterImpl::D, as_FloatRegister(i), O0, d_offset(i)); 1730 } 1731 } 1732 1733 void RegistersForDebugging::restore_registers(MacroAssembler* a, Register r) { 1734 for (int i = 1; i < 8; ++i) { 1735 a->ld_ptr(r, g_offset(i), as_gRegister(i)); 1736 } 1737 for (int j = 0; j < 32; ++j) { 1738 a->ldf(FloatRegisterImpl::S, O0, f_offset(j), as_FloatRegister(j)); 1739 } 1740 for (int k = 0; k < (VM_Version::v9_instructions_work() ? 64 : 32); k += 2) { 1741 a->ldf(FloatRegisterImpl::D, O0, d_offset(k), as_FloatRegister(k)); 1742 } 1743 } 1744 1745 1746 // pushes double TOS element of FPU stack on CPU stack; pops from FPU stack 1747 void MacroAssembler::push_fTOS() { 1748 // %%%%%% need to implement this 1749 } 1750 1751 // pops double TOS element from CPU stack and pushes on FPU stack 1752 void MacroAssembler::pop_fTOS() { 1753 // %%%%%% need to implement this 1754 } 1755 1756 void MacroAssembler::empty_FPU_stack() { 1757 // %%%%%% need to implement this 1758 } 1759 1760 void MacroAssembler::_verify_oop(Register reg, const char* msg, const char * file, int line) { 1761 // plausibility check for oops 1762 if (!VerifyOops) return; 1763 1764 if (reg == G0) return; // always NULL, which is always an oop 1765 1766 BLOCK_COMMENT("verify_oop {"); 1767 char buffer[64]; 1768 #ifdef COMPILER1 1769 if (CommentedAssembly) { 1770 snprintf(buffer, sizeof(buffer), "verify_oop at %d", offset()); 1771 block_comment(buffer); 1772 } 1773 #endif 1774 1775 int len = strlen(file) + strlen(msg) + 1 + 4; 1776 sprintf(buffer, "%d", line); 1777 len += strlen(buffer); 1778 sprintf(buffer, " at offset %d ", offset()); 1779 len += strlen(buffer); 1780 char * real_msg = new char[len]; 1781 sprintf(real_msg, "%s%s(%s:%d)", msg, buffer, file, line); 1782 1783 // Call indirectly to solve generation ordering problem 1784 AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address()); 1785 1786 // Make some space on stack above the current register window. 1787 // Enough to hold 8 64-bit registers. 1788 add(SP,-8*8,SP); 1789 1790 // Save some 64-bit registers; a normal 'save' chops the heads off 1791 // of 64-bit longs in the 32-bit build. 1792 stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8); 1793 stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8); 1794 mov(reg,O0); // Move arg into O0; arg might be in O7 which is about to be crushed 1795 stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8); 1796 1797 set((intptr_t)real_msg, O1); 1798 // Load address to call to into O7 1799 load_ptr_contents(a, O7); 1800 // Register call to verify_oop_subroutine 1801 callr(O7, G0); 1802 delayed()->nop(); 1803 // recover frame size 1804 add(SP, 8*8,SP); 1805 BLOCK_COMMENT("} verify_oop"); 1806 } 1807 1808 void MacroAssembler::_verify_oop_addr(Address addr, const char* msg, const char * file, int line) { 1809 // plausibility check for oops 1810 if (!VerifyOops) return; 1811 1812 char buffer[64]; 1813 sprintf(buffer, "%d", line); 1814 int len = strlen(file) + strlen(msg) + 1 + 4 + strlen(buffer); 1815 sprintf(buffer, " at SP+%d ", addr.disp()); 1816 len += strlen(buffer); 1817 char * real_msg = new char[len]; 1818 sprintf(real_msg, "%s at SP+%d (%s:%d)", msg, addr.disp(), file, line); 1819 1820 // Call indirectly to solve generation ordering problem 1821 AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address()); 1822 1823 // Make some space on stack above the current register window. 1824 // Enough to hold 8 64-bit registers. 1825 add(SP,-8*8,SP); 1826 1827 // Save some 64-bit registers; a normal 'save' chops the heads off 1828 // of 64-bit longs in the 32-bit build. 1829 stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8); 1830 stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8); 1831 ld_ptr(addr.base(), addr.disp() + 8*8, O0); // Load arg into O0; arg might be in O7 which is about to be crushed 1832 stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8); 1833 1834 set((intptr_t)real_msg, O1); 1835 // Load address to call to into O7 1836 load_ptr_contents(a, O7); 1837 // Register call to verify_oop_subroutine 1838 callr(O7, G0); 1839 delayed()->nop(); 1840 // recover frame size 1841 add(SP, 8*8,SP); 1842 } 1843 1844 // side-door communication with signalHandler in os_solaris.cpp 1845 address MacroAssembler::_verify_oop_implicit_branch[3] = { NULL }; 1846 1847 // This macro is expanded just once; it creates shared code. Contract: 1848 // receives an oop in O0. Must restore O0 & O7 from TLS. Must not smash ANY 1849 // registers, including flags. May not use a register 'save', as this blows 1850 // the high bits of the O-regs if they contain Long values. Acts as a 'leaf' 1851 // call. 1852 void MacroAssembler::verify_oop_subroutine() { 1853 assert( VM_Version::v9_instructions_work(), "VerifyOops not supported for V8" ); 1854 1855 // Leaf call; no frame. 1856 Label succeed, fail, null_or_fail; 1857 1858 // O0 and O7 were saved already (O0 in O0's TLS home, O7 in O5's TLS home). 1859 // O0 is now the oop to be checked. O7 is the return address. 1860 Register O0_obj = O0; 1861 1862 // Save some more registers for temps. 1863 stx(O2,SP,frame::register_save_words*wordSize+STACK_BIAS+2*8); 1864 stx(O3,SP,frame::register_save_words*wordSize+STACK_BIAS+3*8); 1865 stx(O4,SP,frame::register_save_words*wordSize+STACK_BIAS+4*8); 1866 stx(O5,SP,frame::register_save_words*wordSize+STACK_BIAS+5*8); 1867 1868 // Save flags 1869 Register O5_save_flags = O5; 1870 rdccr( O5_save_flags ); 1871 1872 { // count number of verifies 1873 Register O2_adr = O2; 1874 Register O3_accum = O3; 1875 inc_counter(StubRoutines::verify_oop_count_addr(), O2_adr, O3_accum); 1876 } 1877 1878 Register O2_mask = O2; 1879 Register O3_bits = O3; 1880 Register O4_temp = O4; 1881 1882 // mark lower end of faulting range 1883 assert(_verify_oop_implicit_branch[0] == NULL, "set once"); 1884 _verify_oop_implicit_branch[0] = pc(); 1885 1886 // We can't check the mark oop because it could be in the process of 1887 // locking or unlocking while this is running. 1888 set(Universe::verify_oop_mask (), O2_mask); 1889 set(Universe::verify_oop_bits (), O3_bits); 1890 1891 // assert((obj & oop_mask) == oop_bits); 1892 and3(O0_obj, O2_mask, O4_temp); 1893 cmp_and_brx(O4_temp, O3_bits, notEqual, false, pn, null_or_fail); 1894 1895 if ((NULL_WORD & Universe::verify_oop_mask()) == Universe::verify_oop_bits()) { 1896 // the null_or_fail case is useless; must test for null separately 1897 br_null(O0_obj, false, pn, succeed); 1898 } 1899 1900 // Check the klassOop of this object for being in the right area of memory. 1901 // Cannot do the load in the delay above slot in case O0 is null 1902 load_klass(O0_obj, O0_obj); 1903 // assert((klass & klass_mask) == klass_bits); 1904 if( Universe::verify_klass_mask() != Universe::verify_oop_mask() ) 1905 set(Universe::verify_klass_mask(), O2_mask); 1906 if( Universe::verify_klass_bits() != Universe::verify_oop_bits() ) 1907 set(Universe::verify_klass_bits(), O3_bits); 1908 and3(O0_obj, O2_mask, O4_temp); 1909 cmp_and_brx(O4_temp, O3_bits, notEqual, false, pn, fail); 1910 // Check the klass's klass 1911 load_klass(O0_obj, O0_obj); 1912 and3(O0_obj, O2_mask, O4_temp); 1913 cmp(O4_temp, O3_bits); 1914 brx(notEqual, false, pn, fail); 1915 delayed()->wrccr( O5_save_flags ); // Restore CCR's 1916 1917 // mark upper end of faulting range 1918 _verify_oop_implicit_branch[1] = pc(); 1919 1920 //----------------------- 1921 // all tests pass 1922 bind(succeed); 1923 1924 // Restore prior 64-bit registers 1925 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+0*8,O0); 1926 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+1*8,O1); 1927 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+2*8,O2); 1928 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+3*8,O3); 1929 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+4*8,O4); 1930 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+5*8,O5); 1931 1932 retl(); // Leaf return; restore prior O7 in delay slot 1933 delayed()->ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+7*8,O7); 1934 1935 //----------------------- 1936 bind(null_or_fail); // nulls are less common but OK 1937 br_null(O0_obj, false, pt, succeed, false); 1938 delayed()->wrccr( O5_save_flags ); // Restore CCR's 1939 1940 //----------------------- 1941 // report failure: 1942 bind(fail); 1943 _verify_oop_implicit_branch[2] = pc(); 1944 1945 wrccr( O5_save_flags ); // Restore CCR's 1946 1947 save_frame(::round_to(sizeof(RegistersForDebugging) / BytesPerWord, 2)); 1948 1949 // stop_subroutine expects message pointer in I1. 1950 mov(I1, O1); 1951 1952 // Restore prior 64-bit registers 1953 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+0*8,I0); 1954 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+1*8,I1); 1955 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+2*8,I2); 1956 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+3*8,I3); 1957 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+4*8,I4); 1958 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+5*8,I5); 1959 1960 // factor long stop-sequence into subroutine to save space 1961 assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet"); 1962 1963 // call indirectly to solve generation ordering problem 1964 AddressLiteral al(StubRoutines::Sparc::stop_subroutine_entry_address()); 1965 load_ptr_contents(al, O5); 1966 jmpl(O5, 0, O7); 1967 delayed()->nop(); 1968 } 1969 1970 1971 void MacroAssembler::stop(const char* msg) { 1972 // save frame first to get O7 for return address 1973 // add one word to size in case struct is odd number of words long 1974 // It must be doubleword-aligned for storing doubles into it. 1975 1976 save_frame(::round_to(sizeof(RegistersForDebugging) / BytesPerWord, 2)); 1977 1978 // stop_subroutine expects message pointer in I1. 1979 set((intptr_t)msg, O1); 1980 1981 // factor long stop-sequence into subroutine to save space 1982 assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet"); 1983 1984 // call indirectly to solve generation ordering problem 1985 AddressLiteral a(StubRoutines::Sparc::stop_subroutine_entry_address()); 1986 load_ptr_contents(a, O5); 1987 jmpl(O5, 0, O7); 1988 delayed()->nop(); 1989 1990 breakpoint_trap(); // make stop actually stop rather than writing 1991 // unnoticeable results in the output files. 1992 1993 // restore(); done in callee to save space! 1994 } 1995 1996 1997 void MacroAssembler::warn(const char* msg) { 1998 save_frame(::round_to(sizeof(RegistersForDebugging) / BytesPerWord, 2)); 1999 RegistersForDebugging::save_registers(this); 2000 mov(O0, L0); 2001 set((intptr_t)msg, O0); 2002 call( CAST_FROM_FN_PTR(address, warning) ); 2003 delayed()->nop(); 2004 // ret(); 2005 // delayed()->restore(); 2006 RegistersForDebugging::restore_registers(this, L0); 2007 restore(); 2008 } 2009 2010 2011 void MacroAssembler::untested(const char* what) { 2012 // We must be able to turn interactive prompting off 2013 // in order to run automated test scripts on the VM 2014 // Use the flag ShowMessageBoxOnError 2015 2016 char* b = new char[1024]; 2017 sprintf(b, "untested: %s", what); 2018 2019 if ( ShowMessageBoxOnError ) stop(b); 2020 else warn(b); 2021 } 2022 2023 2024 void MacroAssembler::stop_subroutine() { 2025 RegistersForDebugging::save_registers(this); 2026 2027 // for the sake of the debugger, stick a PC on the current frame 2028 // (this assumes that the caller has performed an extra "save") 2029 mov(I7, L7); 2030 add(O7, -7 * BytesPerInt, I7); 2031 2032 save_frame(); // one more save to free up another O7 register 2033 mov(I0, O1); // addr of reg save area 2034 2035 // We expect pointer to message in I1. Caller must set it up in O1 2036 mov(I1, O0); // get msg 2037 call (CAST_FROM_FN_PTR(address, MacroAssembler::debug), relocInfo::runtime_call_type); 2038 delayed()->nop(); 2039 2040 restore(); 2041 2042 RegistersForDebugging::restore_registers(this, O0); 2043 2044 save_frame(0); 2045 call(CAST_FROM_FN_PTR(address,breakpoint)); 2046 delayed()->nop(); 2047 restore(); 2048 2049 mov(L7, I7); 2050 retl(); 2051 delayed()->restore(); // see stop above 2052 } 2053 2054 2055 void MacroAssembler::debug(char* msg, RegistersForDebugging* regs) { 2056 if ( ShowMessageBoxOnError ) { 2057 JavaThreadState saved_state = JavaThread::current()->thread_state(); 2058 JavaThread::current()->set_thread_state(_thread_in_vm); 2059 { 2060 // In order to get locks work, we need to fake a in_VM state 2061 ttyLocker ttyl; 2062 ::tty->print_cr("EXECUTION STOPPED: %s\n", msg); 2063 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 2064 ::tty->print_cr("Interpreter::bytecode_counter = %d", BytecodeCounter::counter_value()); 2065 } 2066 if (os::message_box(msg, "Execution stopped, print registers?")) 2067 regs->print(::tty); 2068 } 2069 ThreadStateTransition::transition(JavaThread::current(), _thread_in_vm, saved_state); 2070 } 2071 else 2072 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); 2073 assert(false, err_msg("DEBUG MESSAGE: %s", msg)); 2074 } 2075 2076 2077 #ifndef PRODUCT 2078 void MacroAssembler::test() { 2079 ResourceMark rm; 2080 2081 CodeBuffer cb("test", 10000, 10000); 2082 MacroAssembler* a = new MacroAssembler(&cb); 2083 VM_Version::allow_all(); 2084 a->test_v9(); 2085 a->test_v8_onlys(); 2086 VM_Version::revert(); 2087 2088 StubRoutines::Sparc::test_stop_entry()(); 2089 } 2090 #endif 2091 2092 2093 void MacroAssembler::calc_mem_param_words(Register Rparam_words, Register Rresult) { 2094 subcc( Rparam_words, Argument::n_register_parameters, Rresult); // how many mem words? 2095 Label no_extras; 2096 br( negative, true, pt, no_extras ); // if neg, clear reg 2097 delayed()->set(0, Rresult); // annuled, so only if taken 2098 bind( no_extras ); 2099 } 2100 2101 2102 void MacroAssembler::calc_frame_size(Register Rextra_words, Register Rresult) { 2103 #ifdef _LP64 2104 add(Rextra_words, frame::memory_parameter_word_sp_offset, Rresult); 2105 #else 2106 add(Rextra_words, frame::memory_parameter_word_sp_offset + 1, Rresult); 2107 #endif 2108 bclr(1, Rresult); 2109 sll(Rresult, LogBytesPerWord, Rresult); // Rresult has total frame bytes 2110 } 2111 2112 2113 void MacroAssembler::calc_frame_size_and_save(Register Rextra_words, Register Rresult) { 2114 calc_frame_size(Rextra_words, Rresult); 2115 neg(Rresult); 2116 save(SP, Rresult, SP); 2117 } 2118 2119 2120 // --------------------------------------------------------- 2121 Assembler::RCondition cond2rcond(Assembler::Condition c) { 2122 switch (c) { 2123 /*case zero: */ 2124 case Assembler::equal: return Assembler::rc_z; 2125 case Assembler::lessEqual: return Assembler::rc_lez; 2126 case Assembler::less: return Assembler::rc_lz; 2127 /*case notZero:*/ 2128 case Assembler::notEqual: return Assembler::rc_nz; 2129 case Assembler::greater: return Assembler::rc_gz; 2130 case Assembler::greaterEqual: return Assembler::rc_gez; 2131 } 2132 ShouldNotReachHere(); 2133 return Assembler::rc_z; 2134 } 2135 2136 // compares (32 bit) register with zero and branches. NOT FOR USE WITH 64-bit POINTERS 2137 void MacroAssembler::br_zero(Register s1, Label& L) { 2138 assert_not_delayed(); 2139 if (use_cbc(L)) { 2140 Assembler::cbc(zero, icc, s1, 0, L); 2141 } else { 2142 tst(s1); 2143 br (zero, false, pt, L); 2144 delayed()->nop(); 2145 } 2146 } 2147 2148 // Compares a pointer register with zero and branches on null. 2149 // Does a test & branch on 32-bit systems and a register-branch on 64-bit. 2150 void MacroAssembler::br_null( Register s1, bool a, Predict p, Label& L, bool emit_delayed_nop ) { 2151 assert_not_delayed(); 2152 if (emit_delayed_nop && use_cbc(L)) { 2153 Assembler::cbc(zero, ptr_cc, s1, 0, L); 2154 return; 2155 } 2156 #ifdef _LP64 2157 bpr( rc_z, a, p, s1, L ); 2158 #else 2159 tst(s1); 2160 br ( zero, a, p, L ); 2161 #endif 2162 // Some callers can fill the delay slot. 2163 if (emit_delayed_nop) { 2164 delayed()->nop(); 2165 } 2166 } 2167 2168 void MacroAssembler::br_notnull( Register s1, bool a, Predict p, Label& L, bool emit_delayed_nop ) { 2169 assert_not_delayed(); 2170 if (emit_delayed_nop && use_cbc(L)) { 2171 Assembler::cbc(notZero, ptr_cc, s1, 0, L); 2172 return; 2173 } 2174 #ifdef _LP64 2175 bpr( rc_nz, a, p, s1, L ); 2176 #else 2177 tst(s1); 2178 br ( notZero, a, p, L ); 2179 #endif 2180 // Some callers can fill the delay slot. 2181 if (emit_delayed_nop) { 2182 delayed()->nop(); 2183 } 2184 } 2185 2186 void MacroAssembler::br_on_reg_cond( RCondition rc, bool a, Predict p, 2187 Register s1, address d, 2188 relocInfo::relocType rt ) { 2189 assert_not_delayed(); 2190 if (VM_Version::v9_instructions_work()) { 2191 bpr(rc, a, p, s1, d, rt); 2192 } else { 2193 tst(s1); 2194 br(reg_cond_to_cc_cond(rc), a, p, d, rt); 2195 } 2196 } 2197 2198 void MacroAssembler::br_on_reg_cond( RCondition rc, bool a, Predict p, 2199 Register s1, Label& L, bool emit_delayed_nop ) { 2200 assert_not_delayed(); 2201 if (emit_delayed_nop && use_cbc(L)) { 2202 // Use xcc to have the same result as bpr (it tests all 64 bits). 2203 Assembler::cbc(reg_cond_to_cc_cond(rc), xcc, s1, 0, L); 2204 return; 2205 } 2206 if (VM_Version::v9_instructions_work()) { 2207 bpr(rc, a, p, s1, L); 2208 } else { 2209 tst(s1); 2210 br(reg_cond_to_cc_cond(rc), a, p, L); 2211 } 2212 // Some callers can fill the delay slot. 2213 if (emit_delayed_nop) { 2214 delayed()->nop(); 2215 } 2216 } 2217 2218 // Compare registers and branch with nop in delay slot or cbcond without delay slot. 2219 void MacroAssembler::cmp_and_br(Register s1, Register s2, Condition c, 2220 bool a, Predict p, Label& L) { 2221 assert_not_delayed(); 2222 if (use_cbc(L)) { 2223 Assembler::cbc(c, icc, s1, s2, L); 2224 } else { 2225 cmp(s1, s2); 2226 br(c, a, p, L); 2227 delayed()->nop(); 2228 } 2229 } 2230 2231 void MacroAssembler::cmp_and_br(Register s1, int simm13a, Condition c, 2232 bool a, Predict p, Label& L) { 2233 assert_not_delayed(); 2234 if (is_simm(simm13a,5) && use_cbc(L)) { 2235 Assembler::cbc(c, icc, s1, simm13a, L); 2236 } else { 2237 cmp(s1, simm13a); 2238 br(c, a, p, L); 2239 delayed()->nop(); 2240 } 2241 } 2242 2243 // Branch that tests xcc in LP64 and icc in !LP64 2244 void MacroAssembler::cmp_and_brx(Register s1, Register s2, Condition c, 2245 bool a, Predict p, Label& L) { 2246 assert_not_delayed(); 2247 if (use_cbc(L)) { 2248 Assembler::cbc(c, ptr_cc, s1, s2, L); 2249 } else { 2250 cmp(s1, s2); 2251 brx(c, a, p, L); 2252 delayed()->nop(); 2253 } 2254 } 2255 2256 void MacroAssembler::cmp_and_brx(Register s1, int simm13a, Condition c, 2257 bool a, Predict p, Label& L) { 2258 assert_not_delayed(); 2259 if (is_simm(simm13a,5) && use_cbc(L)) { 2260 Assembler::cbc(c, ptr_cc, s1, simm13a, L); 2261 } else { 2262 cmp(s1, simm13a); 2263 brx(c, a, p, L); 2264 delayed()->nop(); 2265 } 2266 } 2267 2268 // instruction sequences factored across compiler & interpreter 2269 2270 2271 void MacroAssembler::lcmp( Register Ra_hi, Register Ra_low, 2272 Register Rb_hi, Register Rb_low, 2273 Register Rresult) { 2274 2275 Label check_low_parts, done; 2276 2277 cmp(Ra_hi, Rb_hi ); // compare hi parts 2278 br(equal, true, pt, check_low_parts); 2279 delayed()->cmp(Ra_low, Rb_low); // test low parts 2280 2281 // And, with an unsigned comparison, it does not matter if the numbers 2282 // are negative or not. 2283 // E.g., -2 cmp -1: the low parts are 0xfffffffe and 0xffffffff. 2284 // The second one is bigger (unsignedly). 2285 2286 // Other notes: The first move in each triplet can be unconditional 2287 // (and therefore probably prefetchable). 2288 // And the equals case for the high part does not need testing, 2289 // since that triplet is reached only after finding the high halves differ. 2290 2291 if (VM_Version::v9_instructions_work()) { 2292 mov(-1, Rresult); 2293 ba(done, false); delayed()-> movcc(greater, false, icc, 1, Rresult); 2294 } else { 2295 br(less, true, pt, done); delayed()-> set(-1, Rresult); 2296 br(greater, true, pt, done); delayed()-> set( 1, Rresult); 2297 } 2298 2299 bind( check_low_parts ); 2300 2301 if (VM_Version::v9_instructions_work()) { 2302 mov( -1, Rresult); 2303 movcc(equal, false, icc, 0, Rresult); 2304 movcc(greaterUnsigned, false, icc, 1, Rresult); 2305 } else { 2306 set(-1, Rresult); 2307 br(equal, true, pt, done); delayed()->set( 0, Rresult); 2308 br(greaterUnsigned, true, pt, done); delayed()->set( 1, Rresult); 2309 } 2310 bind( done ); 2311 } 2312 2313 void MacroAssembler::lneg( Register Rhi, Register Rlow ) { 2314 subcc( G0, Rlow, Rlow ); 2315 subc( G0, Rhi, Rhi ); 2316 } 2317 2318 void MacroAssembler::lshl( Register Rin_high, Register Rin_low, 2319 Register Rcount, 2320 Register Rout_high, Register Rout_low, 2321 Register Rtemp ) { 2322 2323 2324 Register Ralt_count = Rtemp; 2325 Register Rxfer_bits = Rtemp; 2326 2327 assert( Ralt_count != Rin_high 2328 && Ralt_count != Rin_low 2329 && Ralt_count != Rcount 2330 && Rxfer_bits != Rin_low 2331 && Rxfer_bits != Rin_high 2332 && Rxfer_bits != Rcount 2333 && Rxfer_bits != Rout_low 2334 && Rout_low != Rin_high, 2335 "register alias checks"); 2336 2337 Label big_shift, done; 2338 2339 // This code can be optimized to use the 64 bit shifts in V9. 2340 // Here we use the 32 bit shifts. 2341 2342 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits 2343 subcc(Rcount, 31, Ralt_count); 2344 br(greater, true, pn, big_shift); 2345 delayed()->dec(Ralt_count); 2346 2347 // shift < 32 bits, Ralt_count = Rcount-31 2348 2349 // We get the transfer bits by shifting right by 32-count the low 2350 // register. This is done by shifting right by 31-count and then by one 2351 // more to take care of the special (rare) case where count is zero 2352 // (shifting by 32 would not work). 2353 2354 neg(Ralt_count); 2355 2356 // The order of the next two instructions is critical in the case where 2357 // Rin and Rout are the same and should not be reversed. 2358 2359 srl(Rin_low, Ralt_count, Rxfer_bits); // shift right by 31-count 2360 if (Rcount != Rout_low) { 2361 sll(Rin_low, Rcount, Rout_low); // low half 2362 } 2363 sll(Rin_high, Rcount, Rout_high); 2364 if (Rcount == Rout_low) { 2365 sll(Rin_low, Rcount, Rout_low); // low half 2366 } 2367 srl(Rxfer_bits, 1, Rxfer_bits ); // shift right by one more 2368 ba(done, false); 2369 delayed()->or3(Rout_high, Rxfer_bits, Rout_high); // new hi value: or in shifted old hi part and xfer from low 2370 2371 // shift >= 32 bits, Ralt_count = Rcount-32 2372 bind(big_shift); 2373 sll(Rin_low, Ralt_count, Rout_high ); 2374 clr(Rout_low); 2375 2376 bind(done); 2377 } 2378 2379 2380 void MacroAssembler::lshr( Register Rin_high, Register Rin_low, 2381 Register Rcount, 2382 Register Rout_high, Register Rout_low, 2383 Register Rtemp ) { 2384 2385 Register Ralt_count = Rtemp; 2386 Register Rxfer_bits = Rtemp; 2387 2388 assert( Ralt_count != Rin_high 2389 && Ralt_count != Rin_low 2390 && Ralt_count != Rcount 2391 && Rxfer_bits != Rin_low 2392 && Rxfer_bits != Rin_high 2393 && Rxfer_bits != Rcount 2394 && Rxfer_bits != Rout_high 2395 && Rout_high != Rin_low, 2396 "register alias checks"); 2397 2398 Label big_shift, done; 2399 2400 // This code can be optimized to use the 64 bit shifts in V9. 2401 // Here we use the 32 bit shifts. 2402 2403 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits 2404 subcc(Rcount, 31, Ralt_count); 2405 br(greater, true, pn, big_shift); 2406 delayed()->dec(Ralt_count); 2407 2408 // shift < 32 bits, Ralt_count = Rcount-31 2409 2410 // We get the transfer bits by shifting left by 32-count the high 2411 // register. This is done by shifting left by 31-count and then by one 2412 // more to take care of the special (rare) case where count is zero 2413 // (shifting by 32 would not work). 2414 2415 neg(Ralt_count); 2416 if (Rcount != Rout_low) { 2417 srl(Rin_low, Rcount, Rout_low); 2418 } 2419 2420 // The order of the next two instructions is critical in the case where 2421 // Rin and Rout are the same and should not be reversed. 2422 2423 sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count 2424 sra(Rin_high, Rcount, Rout_high ); // high half 2425 sll(Rxfer_bits, 1, Rxfer_bits); // shift left by one more 2426 if (Rcount == Rout_low) { 2427 srl(Rin_low, Rcount, Rout_low); 2428 } 2429 ba(done, false); 2430 delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high 2431 2432 // shift >= 32 bits, Ralt_count = Rcount-32 2433 bind(big_shift); 2434 2435 sra(Rin_high, Ralt_count, Rout_low); 2436 sra(Rin_high, 31, Rout_high); // sign into hi 2437 2438 bind( done ); 2439 } 2440 2441 2442 2443 void MacroAssembler::lushr( Register Rin_high, Register Rin_low, 2444 Register Rcount, 2445 Register Rout_high, Register Rout_low, 2446 Register Rtemp ) { 2447 2448 Register Ralt_count = Rtemp; 2449 Register Rxfer_bits = Rtemp; 2450 2451 assert( Ralt_count != Rin_high 2452 && Ralt_count != Rin_low 2453 && Ralt_count != Rcount 2454 && Rxfer_bits != Rin_low 2455 && Rxfer_bits != Rin_high 2456 && Rxfer_bits != Rcount 2457 && Rxfer_bits != Rout_high 2458 && Rout_high != Rin_low, 2459 "register alias checks"); 2460 2461 Label big_shift, done; 2462 2463 // This code can be optimized to use the 64 bit shifts in V9. 2464 // Here we use the 32 bit shifts. 2465 2466 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits 2467 subcc(Rcount, 31, Ralt_count); 2468 br(greater, true, pn, big_shift); 2469 delayed()->dec(Ralt_count); 2470 2471 // shift < 32 bits, Ralt_count = Rcount-31 2472 2473 // We get the transfer bits by shifting left by 32-count the high 2474 // register. This is done by shifting left by 31-count and then by one 2475 // more to take care of the special (rare) case where count is zero 2476 // (shifting by 32 would not work). 2477 2478 neg(Ralt_count); 2479 if (Rcount != Rout_low) { 2480 srl(Rin_low, Rcount, Rout_low); 2481 } 2482 2483 // The order of the next two instructions is critical in the case where 2484 // Rin and Rout are the same and should not be reversed. 2485 2486 sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count 2487 srl(Rin_high, Rcount, Rout_high ); // high half 2488 sll(Rxfer_bits, 1, Rxfer_bits); // shift left by one more 2489 if (Rcount == Rout_low) { 2490 srl(Rin_low, Rcount, Rout_low); 2491 } 2492 ba(done, false); 2493 delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high 2494 2495 // shift >= 32 bits, Ralt_count = Rcount-32 2496 bind(big_shift); 2497 2498 srl(Rin_high, Ralt_count, Rout_low); 2499 clr(Rout_high); 2500 2501 bind( done ); 2502 } 2503 2504 #ifdef _LP64 2505 void MacroAssembler::lcmp( Register Ra, Register Rb, Register Rresult) { 2506 cmp(Ra, Rb); 2507 mov(-1, Rresult); 2508 movcc(equal, false, xcc, 0, Rresult); 2509 movcc(greater, false, xcc, 1, Rresult); 2510 } 2511 #endif 2512 2513 2514 void MacroAssembler::load_sized_value(Address src, Register dst, size_t size_in_bytes, bool is_signed) { 2515 switch (size_in_bytes) { 2516 case 8: ld_long(src, dst); break; 2517 case 4: ld( src, dst); break; 2518 case 2: is_signed ? ldsh(src, dst) : lduh(src, dst); break; 2519 case 1: is_signed ? ldsb(src, dst) : ldub(src, dst); break; 2520 default: ShouldNotReachHere(); 2521 } 2522 } 2523 2524 void MacroAssembler::store_sized_value(Register src, Address dst, size_t size_in_bytes) { 2525 switch (size_in_bytes) { 2526 case 8: st_long(src, dst); break; 2527 case 4: st( src, dst); break; 2528 case 2: sth( src, dst); break; 2529 case 1: stb( src, dst); break; 2530 default: ShouldNotReachHere(); 2531 } 2532 } 2533 2534 2535 void MacroAssembler::float_cmp( bool is_float, int unordered_result, 2536 FloatRegister Fa, FloatRegister Fb, 2537 Register Rresult) { 2538 2539 fcmp(is_float ? FloatRegisterImpl::S : FloatRegisterImpl::D, fcc0, Fa, Fb); 2540 2541 Condition lt = unordered_result == -1 ? f_unorderedOrLess : f_less; 2542 Condition eq = f_equal; 2543 Condition gt = unordered_result == 1 ? f_unorderedOrGreater : f_greater; 2544 2545 if (VM_Version::v9_instructions_work()) { 2546 2547 mov(-1, Rresult); 2548 movcc(eq, true, fcc0, 0, Rresult); 2549 movcc(gt, true, fcc0, 1, Rresult); 2550 2551 } else { 2552 Label done; 2553 2554 set( -1, Rresult ); 2555 //fb(lt, true, pn, done); delayed()->set( -1, Rresult ); 2556 fb( eq, true, pn, done); delayed()->set( 0, Rresult ); 2557 fb( gt, true, pn, done); delayed()->set( 1, Rresult ); 2558 2559 bind (done); 2560 } 2561 } 2562 2563 2564 void MacroAssembler::fneg( FloatRegisterImpl::Width w, FloatRegister s, FloatRegister d) 2565 { 2566 if (VM_Version::v9_instructions_work()) { 2567 Assembler::fneg(w, s, d); 2568 } else { 2569 if (w == FloatRegisterImpl::S) { 2570 Assembler::fneg(w, s, d); 2571 } else if (w == FloatRegisterImpl::D) { 2572 // number() does a sanity check on the alignment. 2573 assert(((s->encoding(FloatRegisterImpl::D) & 1) == 0) && 2574 ((d->encoding(FloatRegisterImpl::D) & 1) == 0), "float register alignment check"); 2575 2576 Assembler::fneg(FloatRegisterImpl::S, s, d); 2577 Assembler::fmov(FloatRegisterImpl::S, s->successor(), d->successor()); 2578 } else { 2579 assert(w == FloatRegisterImpl::Q, "Invalid float register width"); 2580 2581 // number() does a sanity check on the alignment. 2582 assert(((s->encoding(FloatRegisterImpl::D) & 3) == 0) && 2583 ((d->encoding(FloatRegisterImpl::D) & 3) == 0), "float register alignment check"); 2584 2585 Assembler::fneg(FloatRegisterImpl::S, s, d); 2586 Assembler::fmov(FloatRegisterImpl::S, s->successor(), d->successor()); 2587 Assembler::fmov(FloatRegisterImpl::S, s->successor()->successor(), d->successor()->successor()); 2588 Assembler::fmov(FloatRegisterImpl::S, s->successor()->successor()->successor(), d->successor()->successor()->successor()); 2589 } 2590 } 2591 } 2592 2593 void MacroAssembler::fmov( FloatRegisterImpl::Width w, FloatRegister s, FloatRegister d) 2594 { 2595 if (VM_Version::v9_instructions_work()) { 2596 Assembler::fmov(w, s, d); 2597 } else { 2598 if (w == FloatRegisterImpl::S) { 2599 Assembler::fmov(w, s, d); 2600 } else if (w == FloatRegisterImpl::D) { 2601 // number() does a sanity check on the alignment. 2602 assert(((s->encoding(FloatRegisterImpl::D) & 1) == 0) && 2603 ((d->encoding(FloatRegisterImpl::D) & 1) == 0), "float register alignment check"); 2604 2605 Assembler::fmov(FloatRegisterImpl::S, s, d); 2606 Assembler::fmov(FloatRegisterImpl::S, s->successor(), d->successor()); 2607 } else { 2608 assert(w == FloatRegisterImpl::Q, "Invalid float register width"); 2609 2610 // number() does a sanity check on the alignment. 2611 assert(((s->encoding(FloatRegisterImpl::D) & 3) == 0) && 2612 ((d->encoding(FloatRegisterImpl::D) & 3) == 0), "float register alignment check"); 2613 2614 Assembler::fmov(FloatRegisterImpl::S, s, d); 2615 Assembler::fmov(FloatRegisterImpl::S, s->successor(), d->successor()); 2616 Assembler::fmov(FloatRegisterImpl::S, s->successor()->successor(), d->successor()->successor()); 2617 Assembler::fmov(FloatRegisterImpl::S, s->successor()->successor()->successor(), d->successor()->successor()->successor()); 2618 } 2619 } 2620 } 2621 2622 void MacroAssembler::fabs( FloatRegisterImpl::Width w, FloatRegister s, FloatRegister d) 2623 { 2624 if (VM_Version::v9_instructions_work()) { 2625 Assembler::fabs(w, s, d); 2626 } else { 2627 if (w == FloatRegisterImpl::S) { 2628 Assembler::fabs(w, s, d); 2629 } else if (w == FloatRegisterImpl::D) { 2630 // number() does a sanity check on the alignment. 2631 assert(((s->encoding(FloatRegisterImpl::D) & 1) == 0) && 2632 ((d->encoding(FloatRegisterImpl::D) & 1) == 0), "float register alignment check"); 2633 2634 Assembler::fabs(FloatRegisterImpl::S, s, d); 2635 Assembler::fmov(FloatRegisterImpl::S, s->successor(), d->successor()); 2636 } else { 2637 assert(w == FloatRegisterImpl::Q, "Invalid float register width"); 2638 2639 // number() does a sanity check on the alignment. 2640 assert(((s->encoding(FloatRegisterImpl::D) & 3) == 0) && 2641 ((d->encoding(FloatRegisterImpl::D) & 3) == 0), "float register alignment check"); 2642 2643 Assembler::fabs(FloatRegisterImpl::S, s, d); 2644 Assembler::fmov(FloatRegisterImpl::S, s->successor(), d->successor()); 2645 Assembler::fmov(FloatRegisterImpl::S, s->successor()->successor(), d->successor()->successor()); 2646 Assembler::fmov(FloatRegisterImpl::S, s->successor()->successor()->successor(), d->successor()->successor()->successor()); 2647 } 2648 } 2649 } 2650 2651 void MacroAssembler::save_all_globals_into_locals() { 2652 mov(G1,L1); 2653 mov(G2,L2); 2654 mov(G3,L3); 2655 mov(G4,L4); 2656 mov(G5,L5); 2657 mov(G6,L6); 2658 mov(G7,L7); 2659 } 2660 2661 void MacroAssembler::restore_globals_from_locals() { 2662 mov(L1,G1); 2663 mov(L2,G2); 2664 mov(L3,G3); 2665 mov(L4,G4); 2666 mov(L5,G5); 2667 mov(L6,G6); 2668 mov(L7,G7); 2669 } 2670 2671 // Use for 64 bit operation. 2672 void MacroAssembler::casx_under_lock(Register top_ptr_reg, Register top_reg, Register ptr_reg, address lock_addr, bool use_call_vm) 2673 { 2674 // store ptr_reg as the new top value 2675 #ifdef _LP64 2676 casx(top_ptr_reg, top_reg, ptr_reg); 2677 #else 2678 cas_under_lock(top_ptr_reg, top_reg, ptr_reg, lock_addr, use_call_vm); 2679 #endif // _LP64 2680 } 2681 2682 // [RGV] This routine does not handle 64 bit operations. 2683 // use casx_under_lock() or casx directly!!! 2684 void MacroAssembler::cas_under_lock(Register top_ptr_reg, Register top_reg, Register ptr_reg, address lock_addr, bool use_call_vm) 2685 { 2686 // store ptr_reg as the new top value 2687 if (VM_Version::v9_instructions_work()) { 2688 cas(top_ptr_reg, top_reg, ptr_reg); 2689 } else { 2690 2691 // If the register is not an out nor global, it is not visible 2692 // after the save. Allocate a register for it, save its 2693 // value in the register save area (the save may not flush 2694 // registers to the save area). 2695 2696 Register top_ptr_reg_after_save; 2697 Register top_reg_after_save; 2698 Register ptr_reg_after_save; 2699 2700 if (top_ptr_reg->is_out() || top_ptr_reg->is_global()) { 2701 top_ptr_reg_after_save = top_ptr_reg->after_save(); 2702 } else { 2703 Address reg_save_addr = top_ptr_reg->address_in_saved_window(); 2704 top_ptr_reg_after_save = L0; 2705 st(top_ptr_reg, reg_save_addr); 2706 } 2707 2708 if (top_reg->is_out() || top_reg->is_global()) { 2709 top_reg_after_save = top_reg->after_save(); 2710 } else { 2711 Address reg_save_addr = top_reg->address_in_saved_window(); 2712 top_reg_after_save = L1; 2713 st(top_reg, reg_save_addr); 2714 } 2715 2716 if (ptr_reg->is_out() || ptr_reg->is_global()) { 2717 ptr_reg_after_save = ptr_reg->after_save(); 2718 } else { 2719 Address reg_save_addr = ptr_reg->address_in_saved_window(); 2720 ptr_reg_after_save = L2; 2721 st(ptr_reg, reg_save_addr); 2722 } 2723 2724 const Register& lock_reg = L3; 2725 const Register& lock_ptr_reg = L4; 2726 const Register& value_reg = L5; 2727 const Register& yield_reg = L6; 2728 const Register& yieldall_reg = L7; 2729 2730 save_frame(); 2731 2732 if (top_ptr_reg_after_save == L0) { 2733 ld(top_ptr_reg->address_in_saved_window().after_save(), top_ptr_reg_after_save); 2734 } 2735 2736 if (top_reg_after_save == L1) { 2737 ld(top_reg->address_in_saved_window().after_save(), top_reg_after_save); 2738 } 2739 2740 if (ptr_reg_after_save == L2) { 2741 ld(ptr_reg->address_in_saved_window().after_save(), ptr_reg_after_save); 2742 } 2743 2744 Label(retry_get_lock); 2745 Label(not_same); 2746 Label(dont_yield); 2747 2748 assert(lock_addr, "lock_address should be non null for v8"); 2749 set((intptr_t)lock_addr, lock_ptr_reg); 2750 // Initialize yield counter 2751 mov(G0,yield_reg); 2752 mov(G0, yieldall_reg); 2753 set(StubRoutines::Sparc::locked, lock_reg); 2754 2755 bind(retry_get_lock); 2756 cmp_and_br(yield_reg, V8AtomicOperationUnderLockSpinCount, Assembler::less, false, Assembler::pt, dont_yield); 2757 2758 if(use_call_vm) { 2759 Untested("Need to verify global reg consistancy"); 2760 call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::yield_all), yieldall_reg); 2761 } else { 2762 // Save the regs and make space for a C call 2763 save(SP, -96, SP); 2764 save_all_globals_into_locals(); 2765 call(CAST_FROM_FN_PTR(address,os::yield_all)); 2766 delayed()->mov(yieldall_reg, O0); 2767 restore_globals_from_locals(); 2768 restore(); 2769 } 2770 2771 // reset the counter 2772 mov(G0,yield_reg); 2773 add(yieldall_reg, 1, yieldall_reg); 2774 2775 bind(dont_yield); 2776 // try to get lock 2777 swap(lock_ptr_reg, 0, lock_reg); 2778 2779 // did we get the lock? 2780 cmp(lock_reg, StubRoutines::Sparc::unlocked); 2781 br(Assembler::notEqual, true, Assembler::pn, retry_get_lock); 2782 delayed()->add(yield_reg,1,yield_reg); 2783 2784 // yes, got lock. do we have the same top? 2785 ld(top_ptr_reg_after_save, 0, value_reg); 2786 cmp_and_br(value_reg, top_reg_after_save, Assembler::notEqual, false, Assembler::pt, not_same); 2787 2788 // yes, same top. 2789 st(ptr_reg_after_save, top_ptr_reg_after_save, 0); 2790 membar(Assembler::StoreStore); 2791 2792 bind(not_same); 2793 mov(value_reg, ptr_reg_after_save); 2794 st(lock_reg, lock_ptr_reg, 0); // unlock 2795 2796 restore(); 2797 } 2798 } 2799 2800 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 2801 Register tmp, 2802 int offset) { 2803 intptr_t value = *delayed_value_addr; 2804 if (value != 0) 2805 return RegisterOrConstant(value + offset); 2806 2807 // load indirectly to solve generation ordering problem 2808 AddressLiteral a(delayed_value_addr); 2809 load_ptr_contents(a, tmp); 2810 2811 #ifdef ASSERT 2812 tst(tmp); 2813 breakpoint_trap(zero, xcc); 2814 #endif 2815 2816 if (offset != 0) 2817 add(tmp, offset, tmp); 2818 2819 return RegisterOrConstant(tmp); 2820 } 2821 2822 2823 RegisterOrConstant MacroAssembler::regcon_andn_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) { 2824 assert(d.register_or_noreg() != G0, "lost side effect"); 2825 if ((s2.is_constant() && s2.as_constant() == 0) || 2826 (s2.is_register() && s2.as_register() == G0)) { 2827 // Do nothing, just move value. 2828 if (s1.is_register()) { 2829 if (d.is_constant()) d = temp; 2830 mov(s1.as_register(), d.as_register()); 2831 return d; 2832 } else { 2833 return s1; 2834 } 2835 } 2836 2837 if (s1.is_register()) { 2838 assert_different_registers(s1.as_register(), temp); 2839 if (d.is_constant()) d = temp; 2840 andn(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register()); 2841 return d; 2842 } else { 2843 if (s2.is_register()) { 2844 assert_different_registers(s2.as_register(), temp); 2845 if (d.is_constant()) d = temp; 2846 set(s1.as_constant(), temp); 2847 andn(temp, s2.as_register(), d.as_register()); 2848 return d; 2849 } else { 2850 intptr_t res = s1.as_constant() & ~s2.as_constant(); 2851 return res; 2852 } 2853 } 2854 } 2855 2856 RegisterOrConstant MacroAssembler::regcon_inc_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) { 2857 assert(d.register_or_noreg() != G0, "lost side effect"); 2858 if ((s2.is_constant() && s2.as_constant() == 0) || 2859 (s2.is_register() && s2.as_register() == G0)) { 2860 // Do nothing, just move value. 2861 if (s1.is_register()) { 2862 if (d.is_constant()) d = temp; 2863 mov(s1.as_register(), d.as_register()); 2864 return d; 2865 } else { 2866 return s1; 2867 } 2868 } 2869 2870 if (s1.is_register()) { 2871 assert_different_registers(s1.as_register(), temp); 2872 if (d.is_constant()) d = temp; 2873 add(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register()); 2874 return d; 2875 } else { 2876 if (s2.is_register()) { 2877 assert_different_registers(s2.as_register(), temp); 2878 if (d.is_constant()) d = temp; 2879 add(s2.as_register(), ensure_simm13_or_reg(s1, temp), d.as_register()); 2880 return d; 2881 } else { 2882 intptr_t res = s1.as_constant() + s2.as_constant(); 2883 return res; 2884 } 2885 } 2886 } 2887 2888 RegisterOrConstant MacroAssembler::regcon_sll_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) { 2889 assert(d.register_or_noreg() != G0, "lost side effect"); 2890 if (!is_simm13(s2.constant_or_zero())) 2891 s2 = (s2.as_constant() & 0xFF); 2892 if ((s2.is_constant() && s2.as_constant() == 0) || 2893 (s2.is_register() && s2.as_register() == G0)) { 2894 // Do nothing, just move value. 2895 if (s1.is_register()) { 2896 if (d.is_constant()) d = temp; 2897 mov(s1.as_register(), d.as_register()); 2898 return d; 2899 } else { 2900 return s1; 2901 } 2902 } 2903 2904 if (s1.is_register()) { 2905 assert_different_registers(s1.as_register(), temp); 2906 if (d.is_constant()) d = temp; 2907 sll_ptr(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register()); 2908 return d; 2909 } else { 2910 if (s2.is_register()) { 2911 assert_different_registers(s2.as_register(), temp); 2912 if (d.is_constant()) d = temp; 2913 set(s1.as_constant(), temp); 2914 sll_ptr(temp, s2.as_register(), d.as_register()); 2915 return d; 2916 } else { 2917 intptr_t res = s1.as_constant() << s2.as_constant(); 2918 return res; 2919 } 2920 } 2921 } 2922 2923 2924 // Look up the method for a megamorphic invokeinterface call. 2925 // The target method is determined by <intf_klass, itable_index>. 2926 // The receiver klass is in recv_klass. 2927 // On success, the result will be in method_result, and execution falls through. 2928 // On failure, execution transfers to the given label. 2929 void MacroAssembler::lookup_interface_method(Register recv_klass, 2930 Register intf_klass, 2931 RegisterOrConstant itable_index, 2932 Register method_result, 2933 Register scan_temp, 2934 Register sethi_temp, 2935 Label& L_no_such_interface) { 2936 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); 2937 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 2938 "caller must use same register for non-constant itable index as for method"); 2939 2940 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 2941 int vtable_base = instanceKlass::vtable_start_offset() * wordSize; 2942 int scan_step = itableOffsetEntry::size() * wordSize; 2943 int vte_size = vtableEntry::size() * wordSize; 2944 2945 lduw(recv_klass, instanceKlass::vtable_length_offset() * wordSize, scan_temp); 2946 // %%% We should store the aligned, prescaled offset in the klassoop. 2947 // Then the next several instructions would fold away. 2948 2949 int round_to_unit = ((HeapWordsPerLong > 1) ? BytesPerLong : 0); 2950 int itb_offset = vtable_base; 2951 if (round_to_unit != 0) { 2952 // hoist first instruction of round_to(scan_temp, BytesPerLong): 2953 itb_offset += round_to_unit - wordSize; 2954 } 2955 int itb_scale = exact_log2(vtableEntry::size() * wordSize); 2956 sll(scan_temp, itb_scale, scan_temp); 2957 add(scan_temp, itb_offset, scan_temp); 2958 if (round_to_unit != 0) { 2959 // Round up to align_object_offset boundary 2960 // see code for instanceKlass::start_of_itable! 2961 // Was: round_to(scan_temp, BytesPerLong); 2962 // Hoisted: add(scan_temp, BytesPerLong-1, scan_temp); 2963 and3(scan_temp, -round_to_unit, scan_temp); 2964 } 2965 add(recv_klass, scan_temp, scan_temp); 2966 2967 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 2968 RegisterOrConstant itable_offset = itable_index; 2969 itable_offset = regcon_sll_ptr(itable_index, exact_log2(itableMethodEntry::size() * wordSize), itable_offset); 2970 itable_offset = regcon_inc_ptr(itable_offset, itableMethodEntry::method_offset_in_bytes(), itable_offset); 2971 add(recv_klass, ensure_simm13_or_reg(itable_offset, sethi_temp), recv_klass); 2972 2973 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 2974 // if (scan->interface() == intf) { 2975 // result = (klass + scan->offset() + itable_index); 2976 // } 2977 // } 2978 Label search, found_method; 2979 2980 for (int peel = 1; peel >= 0; peel--) { 2981 // %%%% Could load both offset and interface in one ldx, if they were 2982 // in the opposite order. This would save a load. 2983 ld_ptr(scan_temp, itableOffsetEntry::interface_offset_in_bytes(), method_result); 2984 2985 // Check that this entry is non-null. A null entry means that 2986 // the receiver class doesn't implement the interface, and wasn't the 2987 // same as when the caller was compiled. 2988 bpr(Assembler::rc_z, false, Assembler::pn, method_result, L_no_such_interface); 2989 delayed()->cmp(method_result, intf_klass); 2990 2991 if (peel) { 2992 brx(Assembler::equal, false, Assembler::pt, found_method); 2993 } else { 2994 brx(Assembler::notEqual, false, Assembler::pn, search); 2995 // (invert the test to fall through to found_method...) 2996 } 2997 delayed()->add(scan_temp, scan_step, scan_temp); 2998 2999 if (!peel) break; 3000 3001 bind(search); 3002 } 3003 3004 bind(found_method); 3005 3006 // Got a hit. 3007 int ito_offset = itableOffsetEntry::offset_offset_in_bytes(); 3008 // scan_temp[-scan_step] points to the vtable offset we need 3009 ito_offset -= scan_step; 3010 lduw(scan_temp, ito_offset, scan_temp); 3011 ld_ptr(recv_klass, scan_temp, method_result); 3012 } 3013 3014 3015 void MacroAssembler::check_klass_subtype(Register sub_klass, 3016 Register super_klass, 3017 Register temp_reg, 3018 Register temp2_reg, 3019 Label& L_success) { 3020 Label L_failure, L_pop_to_failure; 3021 check_klass_subtype_fast_path(sub_klass, super_klass, 3022 temp_reg, temp2_reg, 3023 &L_success, &L_failure, NULL); 3024 Register sub_2 = sub_klass; 3025 Register sup_2 = super_klass; 3026 if (!sub_2->is_global()) sub_2 = L0; 3027 if (!sup_2->is_global()) sup_2 = L1; 3028 3029 save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2); 3030 check_klass_subtype_slow_path(sub_2, sup_2, 3031 L2, L3, L4, L5, 3032 NULL, &L_pop_to_failure); 3033 3034 // on success: 3035 restore(); 3036 ba(L_success); 3037 3038 // on failure: 3039 bind(L_pop_to_failure); 3040 restore(); 3041 bind(L_failure); 3042 } 3043 3044 3045 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 3046 Register super_klass, 3047 Register temp_reg, 3048 Register temp2_reg, 3049 Label* L_success, 3050 Label* L_failure, 3051 Label* L_slow_path, 3052 RegisterOrConstant super_check_offset) { 3053 int sc_offset = (klassOopDesc::header_size() * HeapWordSize + 3054 Klass::secondary_super_cache_offset_in_bytes()); 3055 int sco_offset = (klassOopDesc::header_size() * HeapWordSize + 3056 Klass::super_check_offset_offset_in_bytes()); 3057 3058 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 3059 bool need_slow_path = (must_load_sco || 3060 super_check_offset.constant_or_zero() == sco_offset); 3061 3062 assert_different_registers(sub_klass, super_klass, temp_reg); 3063 if (super_check_offset.is_register()) { 3064 assert_different_registers(sub_klass, super_klass, temp_reg, 3065 super_check_offset.as_register()); 3066 } else if (must_load_sco) { 3067 assert(temp2_reg != noreg, "supply either a temp or a register offset"); 3068 } 3069 3070 Label L_fallthrough; 3071 int label_nulls = 0; 3072 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 3073 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 3074 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 3075 assert(label_nulls <= 1 || 3076 (L_slow_path == &L_fallthrough && label_nulls <= 2 && !need_slow_path), 3077 "at most one NULL in the batch, usually"); 3078 3079 // If the pointers are equal, we are done (e.g., String[] elements). 3080 // This self-check enables sharing of secondary supertype arrays among 3081 // non-primary types such as array-of-interface. Otherwise, each such 3082 // type would need its own customized SSA. 3083 // We move this check to the front of the fast path because many 3084 // type checks are in fact trivially successful in this manner, 3085 // so we get a nicely predicted branch right at the start of the check. 3086 cmp(super_klass, sub_klass); 3087 brx(Assembler::equal, false, Assembler::pn, *L_success); 3088 delayed()->nop(); 3089 3090 // Check the supertype display: 3091 if (must_load_sco) { 3092 // The super check offset is always positive... 3093 lduw(super_klass, sco_offset, temp2_reg); 3094 super_check_offset = RegisterOrConstant(temp2_reg); 3095 // super_check_offset is register. 3096 assert_different_registers(sub_klass, super_klass, temp_reg, super_check_offset.as_register()); 3097 } 3098 ld_ptr(sub_klass, super_check_offset, temp_reg); 3099 cmp(super_klass, temp_reg); 3100 3101 // This check has worked decisively for primary supers. 3102 // Secondary supers are sought in the super_cache ('super_cache_addr'). 3103 // (Secondary supers are interfaces and very deeply nested subtypes.) 3104 // This works in the same check above because of a tricky aliasing 3105 // between the super_cache and the primary super display elements. 3106 // (The 'super_check_addr' can address either, as the case requires.) 3107 // Note that the cache is updated below if it does not help us find 3108 // what we need immediately. 3109 // So if it was a primary super, we can just fail immediately. 3110 // Otherwise, it's the slow path for us (no success at this point). 3111 3112 // Hacked ba(), which may only be used just before L_fallthrough. 3113 #define FINAL_JUMP(label) \ 3114 if (&(label) != &L_fallthrough) { \ 3115 ba(label, false); \ 3116 delayed()->nop(); \ 3117 } 3118 3119 if (super_check_offset.is_register()) { 3120 brx(Assembler::equal, false, Assembler::pn, *L_success); 3121 delayed()->cmp(super_check_offset.as_register(), sc_offset); 3122 3123 if (L_failure == &L_fallthrough) { 3124 brx(Assembler::equal, false, Assembler::pt, *L_slow_path); 3125 delayed()->nop(); 3126 } else { 3127 brx(Assembler::notEqual, false, Assembler::pn, *L_failure); 3128 delayed()->nop(); 3129 FINAL_JUMP(*L_slow_path); 3130 } 3131 } else if (super_check_offset.as_constant() == sc_offset) { 3132 // Need a slow path; fast failure is impossible. 3133 if (L_slow_path == &L_fallthrough) { 3134 brx(Assembler::equal, false, Assembler::pt, *L_success); 3135 delayed()->nop(); 3136 } else { 3137 brx(Assembler::notEqual, false, Assembler::pn, *L_slow_path); 3138 delayed()->nop(); 3139 FINAL_JUMP(*L_success); 3140 } 3141 } else { 3142 // No slow path; it's a fast decision. 3143 if (L_failure == &L_fallthrough) { 3144 brx(Assembler::equal, false, Assembler::pt, *L_success); 3145 delayed()->nop(); 3146 } else { 3147 brx(Assembler::notEqual, false, Assembler::pn, *L_failure); 3148 delayed()->nop(); 3149 FINAL_JUMP(*L_success); 3150 } 3151 } 3152 3153 bind(L_fallthrough); 3154 3155 #undef FINAL_JUMP 3156 } 3157 3158 3159 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 3160 Register super_klass, 3161 Register count_temp, 3162 Register scan_temp, 3163 Register scratch_reg, 3164 Register coop_reg, 3165 Label* L_success, 3166 Label* L_failure) { 3167 assert_different_registers(sub_klass, super_klass, 3168 count_temp, scan_temp, scratch_reg, coop_reg); 3169 3170 Label L_fallthrough, L_loop; 3171 int label_nulls = 0; 3172 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 3173 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 3174 assert(label_nulls <= 1, "at most one NULL in the batch"); 3175 3176 // a couple of useful fields in sub_klass: 3177 int ss_offset = (klassOopDesc::header_size() * HeapWordSize + 3178 Klass::secondary_supers_offset_in_bytes()); 3179 int sc_offset = (klassOopDesc::header_size() * HeapWordSize + 3180 Klass::secondary_super_cache_offset_in_bytes()); 3181 3182 // Do a linear scan of the secondary super-klass chain. 3183 // This code is rarely used, so simplicity is a virtue here. 3184 3185 #ifndef PRODUCT 3186 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 3187 inc_counter((address) pst_counter, count_temp, scan_temp); 3188 #endif 3189 3190 // We will consult the secondary-super array. 3191 ld_ptr(sub_klass, ss_offset, scan_temp); 3192 3193 // Compress superclass if necessary. 3194 Register search_key = super_klass; 3195 bool decode_super_klass = false; 3196 if (UseCompressedOops) { 3197 if (coop_reg != noreg) { 3198 encode_heap_oop_not_null(super_klass, coop_reg); 3199 search_key = coop_reg; 3200 } else { 3201 encode_heap_oop_not_null(super_klass); 3202 decode_super_klass = true; // scarce temps! 3203 } 3204 // The superclass is never null; it would be a basic system error if a null 3205 // pointer were to sneak in here. Note that we have already loaded the 3206 // Klass::super_check_offset from the super_klass in the fast path, 3207 // so if there is a null in that register, we are already in the afterlife. 3208 } 3209 3210 // Load the array length. (Positive movl does right thing on LP64.) 3211 lduw(scan_temp, arrayOopDesc::length_offset_in_bytes(), count_temp); 3212 3213 // Check for empty secondary super list 3214 tst(count_temp); 3215 3216 // Top of search loop 3217 bind(L_loop); 3218 br(Assembler::equal, false, Assembler::pn, *L_failure); 3219 delayed()->add(scan_temp, heapOopSize, scan_temp); 3220 assert(heapOopSize != 0, "heapOopSize should be initialized"); 3221 3222 // Skip the array header in all array accesses. 3223 int elem_offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT); 3224 elem_offset -= heapOopSize; // the scan pointer was pre-incremented also 3225 3226 // Load next super to check 3227 if (UseCompressedOops) { 3228 // Don't use load_heap_oop; we don't want to decode the element. 3229 lduw( scan_temp, elem_offset, scratch_reg ); 3230 } else { 3231 ld_ptr( scan_temp, elem_offset, scratch_reg ); 3232 } 3233 3234 // Look for Rsuper_klass on Rsub_klass's secondary super-class-overflow list 3235 cmp(scratch_reg, search_key); 3236 3237 // A miss means we are NOT a subtype and need to keep looping 3238 brx(Assembler::notEqual, false, Assembler::pn, L_loop); 3239 delayed()->deccc(count_temp); // decrement trip counter in delay slot 3240 3241 // Falling out the bottom means we found a hit; we ARE a subtype 3242 if (decode_super_klass) decode_heap_oop(super_klass); 3243 3244 // Success. Cache the super we found and proceed in triumph. 3245 st_ptr(super_klass, sub_klass, sc_offset); 3246 3247 if (L_success != &L_fallthrough) { 3248 ba(*L_success, false); 3249 delayed()->nop(); 3250 } 3251 3252 bind(L_fallthrough); 3253 } 3254 3255 3256 void MacroAssembler::check_method_handle_type(Register mtype_reg, Register mh_reg, 3257 Register temp_reg, 3258 Label& wrong_method_type) { 3259 assert_different_registers(mtype_reg, mh_reg, temp_reg); 3260 // compare method type against that of the receiver 3261 RegisterOrConstant mhtype_offset = delayed_value(java_lang_invoke_MethodHandle::type_offset_in_bytes, temp_reg); 3262 load_heap_oop(mh_reg, mhtype_offset, temp_reg); 3263 cmp_and_brx(temp_reg, mtype_reg, Assembler::notEqual, false, Assembler::pn, wrong_method_type); 3264 } 3265 3266 3267 // A method handle has a "vmslots" field which gives the size of its 3268 // argument list in JVM stack slots. This field is either located directly 3269 // in every method handle, or else is indirectly accessed through the 3270 // method handle's MethodType. This macro hides the distinction. 3271 void MacroAssembler::load_method_handle_vmslots(Register vmslots_reg, Register mh_reg, 3272 Register temp_reg) { 3273 assert_different_registers(vmslots_reg, mh_reg, temp_reg); 3274 // load mh.type.form.vmslots 3275 if (java_lang_invoke_MethodHandle::vmslots_offset_in_bytes() != 0) { 3276 // hoist vmslots into every mh to avoid dependent load chain 3277 ld( Address(mh_reg, delayed_value(java_lang_invoke_MethodHandle::vmslots_offset_in_bytes, temp_reg)), vmslots_reg); 3278 } else { 3279 Register temp2_reg = vmslots_reg; 3280 load_heap_oop(Address(mh_reg, delayed_value(java_lang_invoke_MethodHandle::type_offset_in_bytes, temp_reg)), temp2_reg); 3281 load_heap_oop(Address(temp2_reg, delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, temp_reg)), temp2_reg); 3282 ld( Address(temp2_reg, delayed_value(java_lang_invoke_MethodTypeForm::vmslots_offset_in_bytes, temp_reg)), vmslots_reg); 3283 } 3284 } 3285 3286 3287 void MacroAssembler::jump_to_method_handle_entry(Register mh_reg, Register temp_reg, bool emit_delayed_nop) { 3288 assert(mh_reg == G3_method_handle, "caller must put MH object in G3"); 3289 assert_different_registers(mh_reg, temp_reg); 3290 3291 // pick out the interpreted side of the handler 3292 // NOTE: vmentry is not an oop! 3293 ld_ptr(mh_reg, delayed_value(java_lang_invoke_MethodHandle::vmentry_offset_in_bytes, temp_reg), temp_reg); 3294 3295 // off we go... 3296 ld_ptr(temp_reg, MethodHandleEntry::from_interpreted_entry_offset_in_bytes(), temp_reg); 3297 jmp(temp_reg, 0); 3298 3299 // for the various stubs which take control at this point, 3300 // see MethodHandles::generate_method_handle_stub 3301 3302 // Some callers can fill the delay slot. 3303 if (emit_delayed_nop) { 3304 delayed()->nop(); 3305 } 3306 } 3307 3308 3309 RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot, 3310 Register temp_reg, 3311 int extra_slot_offset) { 3312 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 3313 int stackElementSize = Interpreter::stackElementSize; 3314 int offset = extra_slot_offset * stackElementSize; 3315 if (arg_slot.is_constant()) { 3316 offset += arg_slot.as_constant() * stackElementSize; 3317 return offset; 3318 } else { 3319 assert(temp_reg != noreg, "must specify"); 3320 sll_ptr(arg_slot.as_register(), exact_log2(stackElementSize), temp_reg); 3321 if (offset != 0) 3322 add(temp_reg, offset, temp_reg); 3323 return temp_reg; 3324 } 3325 } 3326 3327 3328 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 3329 Register temp_reg, 3330 int extra_slot_offset) { 3331 return Address(Gargs, argument_offset(arg_slot, temp_reg, extra_slot_offset)); 3332 } 3333 3334 3335 void MacroAssembler::biased_locking_enter(Register obj_reg, Register mark_reg, 3336 Register temp_reg, 3337 Label& done, Label* slow_case, 3338 BiasedLockingCounters* counters) { 3339 assert(UseBiasedLocking, "why call this otherwise?"); 3340 3341 if (PrintBiasedLockingStatistics) { 3342 assert_different_registers(obj_reg, mark_reg, temp_reg, O7); 3343 if (counters == NULL) 3344 counters = BiasedLocking::counters(); 3345 } 3346 3347 Label cas_label; 3348 3349 // Biased locking 3350 // See whether the lock is currently biased toward our thread and 3351 // whether the epoch is still valid 3352 // Note that the runtime guarantees sufficient alignment of JavaThread 3353 // pointers to allow age to be placed into low bits 3354 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 3355 and3(mark_reg, markOopDesc::biased_lock_mask_in_place, temp_reg); 3356 cmp_and_brx(temp_reg, markOopDesc::biased_lock_pattern, Assembler::notEqual, false, Assembler::pn, cas_label); 3357 3358 load_klass(obj_reg, temp_reg); 3359 ld_ptr(Address(temp_reg, Klass::prototype_header_offset_in_bytes() + klassOopDesc::klass_part_offset_in_bytes()), temp_reg); 3360 or3(G2_thread, temp_reg, temp_reg); 3361 xor3(mark_reg, temp_reg, temp_reg); 3362 andcc(temp_reg, ~((int) markOopDesc::age_mask_in_place), temp_reg); 3363 if (counters != NULL) { 3364 cond_inc(Assembler::equal, (address) counters->biased_lock_entry_count_addr(), mark_reg, temp_reg); 3365 // Reload mark_reg as we may need it later 3366 ld_ptr(Address(obj_reg, oopDesc::mark_offset_in_bytes()), mark_reg); 3367 } 3368 brx(Assembler::equal, true, Assembler::pt, done); 3369 delayed()->nop(); 3370 3371 Label try_revoke_bias; 3372 Label try_rebias; 3373 Address mark_addr = Address(obj_reg, oopDesc::mark_offset_in_bytes()); 3374 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 3375 3376 // At this point we know that the header has the bias pattern and 3377 // that we are not the bias owner in the current epoch. We need to 3378 // figure out more details about the state of the header in order to 3379 // know what operations can be legally performed on the object's 3380 // header. 3381 3382 // If the low three bits in the xor result aren't clear, that means 3383 // the prototype header is no longer biased and we have to revoke 3384 // the bias on this object. 3385 btst(markOopDesc::biased_lock_mask_in_place, temp_reg); 3386 brx(Assembler::notZero, false, Assembler::pn, try_revoke_bias); 3387 3388 // Biasing is still enabled for this data type. See whether the 3389 // epoch of the current bias is still valid, meaning that the epoch 3390 // bits of the mark word are equal to the epoch bits of the 3391 // prototype header. (Note that the prototype header's epoch bits 3392 // only change at a safepoint.) If not, attempt to rebias the object 3393 // toward the current thread. Note that we must be absolutely sure 3394 // that the current epoch is invalid in order to do this because 3395 // otherwise the manipulations it performs on the mark word are 3396 // illegal. 3397 delayed()->btst(markOopDesc::epoch_mask_in_place, temp_reg); 3398 brx(Assembler::notZero, false, Assembler::pn, try_rebias); 3399 3400 // The epoch of the current bias is still valid but we know nothing 3401 // about the owner; it might be set or it might be clear. Try to 3402 // acquire the bias of the object using an atomic operation. If this 3403 // fails we will go in to the runtime to revoke the object's bias. 3404 // Note that we first construct the presumed unbiased header so we 3405 // don't accidentally blow away another thread's valid bias. 3406 delayed()->and3(mark_reg, 3407 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place, 3408 mark_reg); 3409 or3(G2_thread, mark_reg, temp_reg); 3410 casn(mark_addr.base(), mark_reg, temp_reg); 3411 // If the biasing toward our thread failed, this means that 3412 // another thread succeeded in biasing it toward itself and we 3413 // need to revoke that bias. The revocation will occur in the 3414 // interpreter runtime in the slow case. 3415 cmp(mark_reg, temp_reg); 3416 if (counters != NULL) { 3417 cond_inc(Assembler::zero, (address) counters->anonymously_biased_lock_entry_count_addr(), mark_reg, temp_reg); 3418 } 3419 if (slow_case != NULL) { 3420 brx(Assembler::notEqual, true, Assembler::pn, *slow_case); 3421 delayed()->nop(); 3422 } 3423 ba(done); 3424 3425 bind(try_rebias); 3426 // At this point we know the epoch has expired, meaning that the 3427 // current "bias owner", if any, is actually invalid. Under these 3428 // circumstances _only_, we are allowed to use the current header's 3429 // value as the comparison value when doing the cas to acquire the 3430 // bias in the current epoch. In other words, we allow transfer of 3431 // the bias from one thread to another directly in this situation. 3432 // 3433 // FIXME: due to a lack of registers we currently blow away the age 3434 // bits in this situation. Should attempt to preserve them. 3435 load_klass(obj_reg, temp_reg); 3436 ld_ptr(Address(temp_reg, Klass::prototype_header_offset_in_bytes() + klassOopDesc::klass_part_offset_in_bytes()), temp_reg); 3437 or3(G2_thread, temp_reg, temp_reg); 3438 casn(mark_addr.base(), mark_reg, temp_reg); 3439 // If the biasing toward our thread failed, this means that 3440 // another thread succeeded in biasing it toward itself and we 3441 // need to revoke that bias. The revocation will occur in the 3442 // interpreter runtime in the slow case. 3443 cmp(mark_reg, temp_reg); 3444 if (counters != NULL) { 3445 cond_inc(Assembler::zero, (address) counters->rebiased_lock_entry_count_addr(), mark_reg, temp_reg); 3446 } 3447 if (slow_case != NULL) { 3448 brx(Assembler::notEqual, true, Assembler::pn, *slow_case); 3449 delayed()->nop(); 3450 } 3451 ba(done); 3452 3453 bind(try_revoke_bias); 3454 // The prototype mark in the klass doesn't have the bias bit set any 3455 // more, indicating that objects of this data type are not supposed 3456 // to be biased any more. We are going to try to reset the mark of 3457 // this object to the prototype value and fall through to the 3458 // CAS-based locking scheme. Note that if our CAS fails, it means 3459 // that another thread raced us for the privilege of revoking the 3460 // bias of this particular object, so it's okay to continue in the 3461 // normal locking code. 3462 // 3463 // FIXME: due to a lack of registers we currently blow away the age 3464 // bits in this situation. Should attempt to preserve them. 3465 load_klass(obj_reg, temp_reg); 3466 ld_ptr(Address(temp_reg, Klass::prototype_header_offset_in_bytes() + klassOopDesc::klass_part_offset_in_bytes()), temp_reg); 3467 casn(mark_addr.base(), mark_reg, temp_reg); 3468 // Fall through to the normal CAS-based lock, because no matter what 3469 // the result of the above CAS, some thread must have succeeded in 3470 // removing the bias bit from the object's header. 3471 if (counters != NULL) { 3472 cmp(mark_reg, temp_reg); 3473 cond_inc(Assembler::zero, (address) counters->revoked_lock_entry_count_addr(), mark_reg, temp_reg); 3474 } 3475 3476 bind(cas_label); 3477 } 3478 3479 void MacroAssembler::biased_locking_exit (Address mark_addr, Register temp_reg, Label& done, 3480 bool allow_delay_slot_filling) { 3481 // Check for biased locking unlock case, which is a no-op 3482 // Note: we do not have to check the thread ID for two reasons. 3483 // First, the interpreter checks for IllegalMonitorStateException at 3484 // a higher level. Second, if the bias was revoked while we held the 3485 // lock, the object could not be rebiased toward another thread, so 3486 // the bias bit would be clear. 3487 ld_ptr(mark_addr, temp_reg); 3488 and3(temp_reg, markOopDesc::biased_lock_mask_in_place, temp_reg); 3489 cmp(temp_reg, markOopDesc::biased_lock_pattern); 3490 brx(Assembler::equal, allow_delay_slot_filling, Assembler::pt, done); 3491 delayed(); 3492 if (!allow_delay_slot_filling) { 3493 nop(); 3494 } 3495 } 3496 3497 3498 // CASN -- 32-64 bit switch hitter similar to the synthetic CASN provided by 3499 // Solaris/SPARC's "as". Another apt name would be cas_ptr() 3500 3501 void MacroAssembler::casn (Register addr_reg, Register cmp_reg, Register set_reg ) { 3502 casx_under_lock (addr_reg, cmp_reg, set_reg, (address)StubRoutines::Sparc::atomic_memory_operation_lock_addr()) ; 3503 } 3504 3505 3506 3507 // compiler_lock_object() and compiler_unlock_object() are direct transliterations 3508 // of i486.ad fast_lock() and fast_unlock(). See those methods for detailed comments. 3509 // The code could be tightened up considerably. 3510 // 3511 // box->dhw disposition - post-conditions at DONE_LABEL. 3512 // - Successful inflated lock: box->dhw != 0. 3513 // Any non-zero value suffices. 3514 // Consider G2_thread, rsp, boxReg, or unused_mark() 3515 // - Successful Stack-lock: box->dhw == mark. 3516 // box->dhw must contain the displaced mark word value 3517 // - Failure -- icc.ZFlag == 0 and box->dhw is undefined. 3518 // The slow-path fast_enter() and slow_enter() operators 3519 // are responsible for setting box->dhw = NonZero (typically ::unused_mark). 3520 // - Biased: box->dhw is undefined 3521 // 3522 // SPARC refworkload performance - specifically jetstream and scimark - are 3523 // extremely sensitive to the size of the code emitted by compiler_lock_object 3524 // and compiler_unlock_object. Critically, the key factor is code size, not path 3525 // length. (Simply experiments to pad CLO with unexecuted NOPs demonstrte the 3526 // effect). 3527 3528 3529 void MacroAssembler::compiler_lock_object(Register Roop, Register Rmark, 3530 Register Rbox, Register Rscratch, 3531 BiasedLockingCounters* counters, 3532 bool try_bias) { 3533 Address mark_addr(Roop, oopDesc::mark_offset_in_bytes()); 3534 3535 verify_oop(Roop); 3536 Label done ; 3537 3538 if (counters != NULL) { 3539 inc_counter((address) counters->total_entry_count_addr(), Rmark, Rscratch); 3540 } 3541 3542 if (EmitSync & 1) { 3543 mov (3, Rscratch) ; 3544 st_ptr (Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3545 cmp (SP, G0) ; 3546 return ; 3547 } 3548 3549 if (EmitSync & 2) { 3550 3551 // Fetch object's markword 3552 ld_ptr(mark_addr, Rmark); 3553 3554 if (try_bias) { 3555 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters); 3556 } 3557 3558 // Save Rbox in Rscratch to be used for the cas operation 3559 mov(Rbox, Rscratch); 3560 3561 // set Rmark to markOop | markOopDesc::unlocked_value 3562 or3(Rmark, markOopDesc::unlocked_value, Rmark); 3563 3564 // Initialize the box. (Must happen before we update the object mark!) 3565 st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3566 3567 // compare object markOop with Rmark and if equal exchange Rscratch with object markOop 3568 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 3569 casx_under_lock(mark_addr.base(), Rmark, Rscratch, 3570 (address)StubRoutines::Sparc::atomic_memory_operation_lock_addr()); 3571 3572 // if compare/exchange succeeded we found an unlocked object and we now have locked it 3573 // hence we are done 3574 cmp(Rmark, Rscratch); 3575 #ifdef _LP64 3576 sub(Rscratch, STACK_BIAS, Rscratch); 3577 #endif 3578 brx(Assembler::equal, false, Assembler::pt, done); 3579 delayed()->sub(Rscratch, SP, Rscratch); //pull next instruction into delay slot 3580 3581 // we did not find an unlocked object so see if this is a recursive case 3582 // sub(Rscratch, SP, Rscratch); 3583 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant"); 3584 andcc(Rscratch, 0xfffff003, Rscratch); 3585 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3586 bind (done) ; 3587 return ; 3588 } 3589 3590 Label Egress ; 3591 3592 if (EmitSync & 256) { 3593 Label IsInflated ; 3594 3595 ld_ptr (mark_addr, Rmark); // fetch obj->mark 3596 // Triage: biased, stack-locked, neutral, inflated 3597 if (try_bias) { 3598 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters); 3599 // Invariant: if control reaches this point in the emitted stream 3600 // then Rmark has not been modified. 3601 } 3602 3603 // Store mark into displaced mark field in the on-stack basic-lock "box" 3604 // Critically, this must happen before the CAS 3605 // Maximize the ST-CAS distance to minimize the ST-before-CAS penalty. 3606 st_ptr (Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3607 andcc (Rmark, 2, G0) ; 3608 brx (Assembler::notZero, false, Assembler::pn, IsInflated) ; 3609 delayed() -> 3610 3611 // Try stack-lock acquisition. 3612 // Beware: the 1st instruction is in a delay slot 3613 mov (Rbox, Rscratch); 3614 or3 (Rmark, markOopDesc::unlocked_value, Rmark); 3615 assert (mark_addr.disp() == 0, "cas must take a zero displacement"); 3616 casn (mark_addr.base(), Rmark, Rscratch) ; 3617 cmp (Rmark, Rscratch); 3618 brx (Assembler::equal, false, Assembler::pt, done); 3619 delayed()->sub(Rscratch, SP, Rscratch); 3620 3621 // Stack-lock attempt failed - check for recursive stack-lock. 3622 // See the comments below about how we might remove this case. 3623 #ifdef _LP64 3624 sub (Rscratch, STACK_BIAS, Rscratch); 3625 #endif 3626 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant"); 3627 andcc (Rscratch, 0xfffff003, Rscratch); 3628 br (Assembler::always, false, Assembler::pt, done) ; 3629 delayed()-> st_ptr (Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3630 3631 bind (IsInflated) ; 3632 if (EmitSync & 64) { 3633 // If m->owner != null goto IsLocked 3634 // Pessimistic form: Test-and-CAS vs CAS 3635 // The optimistic form avoids RTS->RTO cache line upgrades. 3636 ld_ptr (Rmark, ObjectMonitor::owner_offset_in_bytes() - 2, Rscratch); 3637 andcc (Rscratch, Rscratch, G0) ; 3638 brx (Assembler::notZero, false, Assembler::pn, done) ; 3639 delayed()->nop() ; 3640 // m->owner == null : it's unlocked. 3641 } 3642 3643 // Try to CAS m->owner from null to Self 3644 // Invariant: if we acquire the lock then _recursions should be 0. 3645 add (Rmark, ObjectMonitor::owner_offset_in_bytes()-2, Rmark) ; 3646 mov (G2_thread, Rscratch) ; 3647 casn (Rmark, G0, Rscratch) ; 3648 cmp (Rscratch, G0) ; 3649 // Intentional fall-through into done 3650 } else { 3651 // Aggressively avoid the Store-before-CAS penalty 3652 // Defer the store into box->dhw until after the CAS 3653 Label IsInflated, Recursive ; 3654 3655 // Anticipate CAS -- Avoid RTS->RTO upgrade 3656 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads) ; 3657 3658 ld_ptr (mark_addr, Rmark); // fetch obj->mark 3659 // Triage: biased, stack-locked, neutral, inflated 3660 3661 if (try_bias) { 3662 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters); 3663 // Invariant: if control reaches this point in the emitted stream 3664 // then Rmark has not been modified. 3665 } 3666 andcc (Rmark, 2, G0) ; 3667 brx (Assembler::notZero, false, Assembler::pn, IsInflated) ; 3668 delayed()-> // Beware - dangling delay-slot 3669 3670 // Try stack-lock acquisition. 3671 // Transiently install BUSY (0) encoding in the mark word. 3672 // if the CAS of 0 into the mark was successful then we execute: 3673 // ST box->dhw = mark -- save fetched mark in on-stack basiclock box 3674 // ST obj->mark = box -- overwrite transient 0 value 3675 // This presumes TSO, of course. 3676 3677 mov (0, Rscratch) ; 3678 or3 (Rmark, markOopDesc::unlocked_value, Rmark); 3679 assert (mark_addr.disp() == 0, "cas must take a zero displacement"); 3680 casn (mark_addr.base(), Rmark, Rscratch) ; 3681 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads) ; 3682 cmp (Rscratch, Rmark) ; 3683 brx (Assembler::notZero, false, Assembler::pn, Recursive) ; 3684 delayed() -> 3685 st_ptr (Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3686 if (counters != NULL) { 3687 cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch); 3688 } 3689 br (Assembler::always, false, Assembler::pt, done); 3690 delayed() -> 3691 st_ptr (Rbox, mark_addr) ; 3692 3693 bind (Recursive) ; 3694 // Stack-lock attempt failed - check for recursive stack-lock. 3695 // Tests show that we can remove the recursive case with no impact 3696 // on refworkload 0.83. If we need to reduce the size of the code 3697 // emitted by compiler_lock_object() the recursive case is perfect 3698 // candidate. 3699 // 3700 // A more extreme idea is to always inflate on stack-lock recursion. 3701 // This lets us eliminate the recursive checks in compiler_lock_object 3702 // and compiler_unlock_object and the (box->dhw == 0) encoding. 3703 // A brief experiment - requiring changes to synchronizer.cpp, interpreter, 3704 // and showed a performance *increase*. In the same experiment I eliminated 3705 // the fast-path stack-lock code from the interpreter and always passed 3706 // control to the "slow" operators in synchronizer.cpp. 3707 3708 // RScratch contains the fetched obj->mark value from the failed CASN. 3709 #ifdef _LP64 3710 sub (Rscratch, STACK_BIAS, Rscratch); 3711 #endif 3712 sub(Rscratch, SP, Rscratch); 3713 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant"); 3714 andcc (Rscratch, 0xfffff003, Rscratch); 3715 if (counters != NULL) { 3716 // Accounting needs the Rscratch register 3717 st_ptr (Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3718 cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch); 3719 br (Assembler::always, false, Assembler::pt, done) ; 3720 delayed()->nop() ; 3721 } else { 3722 br (Assembler::always, false, Assembler::pt, done) ; 3723 delayed()-> st_ptr (Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3724 } 3725 3726 bind (IsInflated) ; 3727 if (EmitSync & 64) { 3728 // If m->owner != null goto IsLocked 3729 // Test-and-CAS vs CAS 3730 // Pessimistic form avoids futile (doomed) CAS attempts 3731 // The optimistic form avoids RTS->RTO cache line upgrades. 3732 ld_ptr (Rmark, ObjectMonitor::owner_offset_in_bytes() - 2, Rscratch); 3733 andcc (Rscratch, Rscratch, G0) ; 3734 brx (Assembler::notZero, false, Assembler::pn, done) ; 3735 delayed()->nop() ; 3736 // m->owner == null : it's unlocked. 3737 } 3738 3739 // Try to CAS m->owner from null to Self 3740 // Invariant: if we acquire the lock then _recursions should be 0. 3741 add (Rmark, ObjectMonitor::owner_offset_in_bytes()-2, Rmark) ; 3742 mov (G2_thread, Rscratch) ; 3743 casn (Rmark, G0, Rscratch) ; 3744 cmp (Rscratch, G0) ; 3745 // ST box->displaced_header = NonZero. 3746 // Any non-zero value suffices: 3747 // unused_mark(), G2_thread, RBox, RScratch, rsp, etc. 3748 st_ptr (Rbox, Rbox, BasicLock::displaced_header_offset_in_bytes()); 3749 // Intentional fall-through into done 3750 } 3751 3752 bind (done) ; 3753 } 3754 3755 void MacroAssembler::compiler_unlock_object(Register Roop, Register Rmark, 3756 Register Rbox, Register Rscratch, 3757 bool try_bias) { 3758 Address mark_addr(Roop, oopDesc::mark_offset_in_bytes()); 3759 3760 Label done ; 3761 3762 if (EmitSync & 4) { 3763 cmp (SP, G0) ; 3764 return ; 3765 } 3766 3767 if (EmitSync & 8) { 3768 if (try_bias) { 3769 biased_locking_exit(mark_addr, Rscratch, done); 3770 } 3771 3772 // Test first if it is a fast recursive unlock 3773 ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rmark); 3774 br_null(Rmark, false, Assembler::pt, done); 3775 3776 // Check if it is still a light weight lock, this is is true if we see 3777 // the stack address of the basicLock in the markOop of the object 3778 assert(mark_addr.disp() == 0, "cas must take a zero displacement"); 3779 casx_under_lock(mark_addr.base(), Rbox, Rmark, 3780 (address)StubRoutines::Sparc::atomic_memory_operation_lock_addr()); 3781 ba(done, false); 3782 delayed()->cmp(Rbox, Rmark); 3783 bind (done) ; 3784 return ; 3785 } 3786 3787 // Beware ... If the aggregate size of the code emitted by CLO and CUO is 3788 // is too large performance rolls abruptly off a cliff. 3789 // This could be related to inlining policies, code cache management, or 3790 // I$ effects. 3791 Label LStacked ; 3792 3793 if (try_bias) { 3794 // TODO: eliminate redundant LDs of obj->mark 3795 biased_locking_exit(mark_addr, Rscratch, done); 3796 } 3797 3798 ld_ptr (Roop, oopDesc::mark_offset_in_bytes(), Rmark) ; 3799 ld_ptr (Rbox, BasicLock::displaced_header_offset_in_bytes(), Rscratch); 3800 andcc (Rscratch, Rscratch, G0); 3801 brx (Assembler::zero, false, Assembler::pn, done); 3802 delayed()-> nop() ; // consider: relocate fetch of mark, above, into this DS 3803 andcc (Rmark, 2, G0) ; 3804 brx (Assembler::zero, false, Assembler::pt, LStacked) ; 3805 delayed()-> nop() ; 3806 3807 // It's inflated 3808 // Conceptually we need a #loadstore|#storestore "release" MEMBAR before 3809 // the ST of 0 into _owner which releases the lock. This prevents loads 3810 // and stores within the critical section from reordering (floating) 3811 // past the store that releases the lock. But TSO is a strong memory model 3812 // and that particular flavor of barrier is a noop, so we can safely elide it. 3813 // Note that we use 1-0 locking by default for the inflated case. We 3814 // close the resultant (and rare) race by having contented threads in 3815 // monitorenter periodically poll _owner. 3816 ld_ptr (Rmark, ObjectMonitor::owner_offset_in_bytes() - 2, Rscratch); 3817 ld_ptr (Rmark, ObjectMonitor::recursions_offset_in_bytes() - 2, Rbox); 3818 xor3 (Rscratch, G2_thread, Rscratch) ; 3819 orcc (Rbox, Rscratch, Rbox) ; 3820 brx (Assembler::notZero, false, Assembler::pn, done) ; 3821 delayed()-> 3822 ld_ptr (Rmark, ObjectMonitor::EntryList_offset_in_bytes() - 2, Rscratch); 3823 ld_ptr (Rmark, ObjectMonitor::cxq_offset_in_bytes() - 2, Rbox); 3824 orcc (Rbox, Rscratch, G0) ; 3825 if (EmitSync & 65536) { 3826 Label LSucc ; 3827 brx (Assembler::notZero, false, Assembler::pn, LSucc) ; 3828 delayed()->nop() ; 3829 ba (done, false) ; 3830 delayed()-> 3831 st_ptr (G0, Rmark, ObjectMonitor::owner_offset_in_bytes() - 2); 3832 3833 bind (LSucc) ; 3834 st_ptr (G0, Rmark, ObjectMonitor::owner_offset_in_bytes() - 2); 3835 if (os::is_MP()) { membar (StoreLoad) ; } 3836 ld_ptr (Rmark, ObjectMonitor::succ_offset_in_bytes() - 2, Rscratch); 3837 andcc (Rscratch, Rscratch, G0) ; 3838 brx (Assembler::notZero, false, Assembler::pt, done) ; 3839 delayed()-> andcc (G0, G0, G0) ; 3840 add (Rmark, ObjectMonitor::owner_offset_in_bytes()-2, Rmark) ; 3841 mov (G2_thread, Rscratch) ; 3842 casn (Rmark, G0, Rscratch) ; 3843 // invert icc.zf and goto done 3844 br_notnull(Rscratch, false, Assembler::pt, done, false) ; 3845 delayed() -> cmp (G0, G0) ; 3846 ba (done, false); 3847 delayed() -> cmp (G0, 1) ; 3848 } else { 3849 brx (Assembler::notZero, false, Assembler::pn, done) ; 3850 delayed()->nop() ; 3851 ba (done, false) ; 3852 delayed()-> 3853 st_ptr (G0, Rmark, ObjectMonitor::owner_offset_in_bytes() - 2); 3854 } 3855 3856 bind (LStacked) ; 3857 // Consider: we could replace the expensive CAS in the exit 3858 // path with a simple ST of the displaced mark value fetched from 3859 // the on-stack basiclock box. That admits a race where a thread T2 3860 // in the slow lock path -- inflating with monitor M -- could race a 3861 // thread T1 in the fast unlock path, resulting in a missed wakeup for T2. 3862 // More precisely T1 in the stack-lock unlock path could "stomp" the 3863 // inflated mark value M installed by T2, resulting in an orphan 3864 // object monitor M and T2 becoming stranded. We can remedy that situation 3865 // by having T2 periodically poll the object's mark word using timed wait 3866 // operations. If T2 discovers that a stomp has occurred it vacates 3867 // the monitor M and wakes any other threads stranded on the now-orphan M. 3868 // In addition the monitor scavenger, which performs deflation, 3869 // would also need to check for orpan monitors and stranded threads. 3870 // 3871 // Finally, inflation is also used when T2 needs to assign a hashCode 3872 // to O and O is stack-locked by T1. The "stomp" race could cause 3873 // an assigned hashCode value to be lost. We can avoid that condition 3874 // and provide the necessary hashCode stability invariants by ensuring 3875 // that hashCode generation is idempotent between copying GCs. 3876 // For example we could compute the hashCode of an object O as 3877 // O's heap address XOR some high quality RNG value that is refreshed 3878 // at GC-time. The monitor scavenger would install the hashCode 3879 // found in any orphan monitors. Again, the mechanism admits a 3880 // lost-update "stomp" WAW race but detects and recovers as needed. 3881 // 3882 // A prototype implementation showed excellent results, although 3883 // the scavenger and timeout code was rather involved. 3884 3885 casn (mark_addr.base(), Rbox, Rscratch) ; 3886 cmp (Rbox, Rscratch); 3887 // Intentional fall through into done ... 3888 3889 bind (done) ; 3890 } 3891 3892 3893 3894 void MacroAssembler::print_CPU_state() { 3895 // %%%%% need to implement this 3896 } 3897 3898 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 3899 // %%%%% need to implement this 3900 } 3901 3902 void MacroAssembler::push_IU_state() { 3903 // %%%%% need to implement this 3904 } 3905 3906 3907 void MacroAssembler::pop_IU_state() { 3908 // %%%%% need to implement this 3909 } 3910 3911 3912 void MacroAssembler::push_FPU_state() { 3913 // %%%%% need to implement this 3914 } 3915 3916 3917 void MacroAssembler::pop_FPU_state() { 3918 // %%%%% need to implement this 3919 } 3920 3921 3922 void MacroAssembler::push_CPU_state() { 3923 // %%%%% need to implement this 3924 } 3925 3926 3927 void MacroAssembler::pop_CPU_state() { 3928 // %%%%% need to implement this 3929 } 3930 3931 3932 3933 void MacroAssembler::verify_tlab() { 3934 #ifdef ASSERT 3935 if (UseTLAB && VerifyOops) { 3936 Label next, next2, ok; 3937 Register t1 = L0; 3938 Register t2 = L1; 3939 Register t3 = L2; 3940 3941 save_frame(0); 3942 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1); 3943 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t2); 3944 or3(t1, t2, t3); 3945 cmp_and_br(t1, t2, Assembler::greaterEqual, false, Assembler::pn, next); 3946 stop("assert(top >= start)"); 3947 should_not_reach_here(); 3948 3949 bind(next); 3950 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1); 3951 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t2); 3952 or3(t3, t2, t3); 3953 cmp_and_br(t1, t2, Assembler::lessEqual, false, Assembler::pn, next2); 3954 stop("assert(top <= end)"); 3955 should_not_reach_here(); 3956 3957 bind(next2); 3958 and3(t3, MinObjAlignmentInBytesMask, t3); 3959 cmp_and_br(t3, 0, Assembler::lessEqual, false, Assembler::pn, ok); 3960 stop("assert(aligned)"); 3961 should_not_reach_here(); 3962 3963 bind(ok); 3964 restore(); 3965 } 3966 #endif 3967 } 3968 3969 3970 void MacroAssembler::eden_allocate( 3971 Register obj, // result: pointer to object after successful allocation 3972 Register var_size_in_bytes, // object size in bytes if unknown at compile time; invalid otherwise 3973 int con_size_in_bytes, // object size in bytes if known at compile time 3974 Register t1, // temp register 3975 Register t2, // temp register 3976 Label& slow_case // continuation point if fast allocation fails 3977 ){ 3978 // make sure arguments make sense 3979 assert_different_registers(obj, var_size_in_bytes, t1, t2); 3980 assert(0 <= con_size_in_bytes && Assembler::is_simm13(con_size_in_bytes), "illegal object size"); 3981 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment"); 3982 3983 if (CMSIncrementalMode || !Universe::heap()->supports_inline_contig_alloc()) { 3984 // No allocation in the shared eden. 3985 ba(slow_case); 3986 } else { 3987 // get eden boundaries 3988 // note: we need both top & top_addr! 3989 const Register top_addr = t1; 3990 const Register end = t2; 3991 3992 CollectedHeap* ch = Universe::heap(); 3993 set((intx)ch->top_addr(), top_addr); 3994 intx delta = (intx)ch->end_addr() - (intx)ch->top_addr(); 3995 ld_ptr(top_addr, delta, end); 3996 ld_ptr(top_addr, 0, obj); 3997 3998 // try to allocate 3999 Label retry; 4000 bind(retry); 4001 #ifdef ASSERT 4002 // make sure eden top is properly aligned 4003 { 4004 Label L; 4005 btst(MinObjAlignmentInBytesMask, obj); 4006 br(Assembler::zero, false, Assembler::pt, L); 4007 delayed()->nop(); 4008 stop("eden top is not properly aligned"); 4009 bind(L); 4010 } 4011 #endif // ASSERT 4012 const Register free = end; 4013 sub(end, obj, free); // compute amount of free space 4014 if (var_size_in_bytes->is_valid()) { 4015 // size is unknown at compile time 4016 cmp(free, var_size_in_bytes); 4017 br(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case 4018 delayed()->add(obj, var_size_in_bytes, end); 4019 } else { 4020 // size is known at compile time 4021 cmp(free, con_size_in_bytes); 4022 br(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case 4023 delayed()->add(obj, con_size_in_bytes, end); 4024 } 4025 // Compare obj with the value at top_addr; if still equal, swap the value of 4026 // end with the value at top_addr. If not equal, read the value at top_addr 4027 // into end. 4028 casx_under_lock(top_addr, obj, end, (address)StubRoutines::Sparc::atomic_memory_operation_lock_addr()); 4029 // if someone beat us on the allocation, try again, otherwise continue 4030 cmp(obj, end); 4031 brx(Assembler::notEqual, false, Assembler::pn, retry); 4032 delayed()->mov(end, obj); // nop if successfull since obj == end 4033 4034 #ifdef ASSERT 4035 // make sure eden top is properly aligned 4036 { 4037 Label L; 4038 const Register top_addr = t1; 4039 4040 set((intx)ch->top_addr(), top_addr); 4041 ld_ptr(top_addr, 0, top_addr); 4042 btst(MinObjAlignmentInBytesMask, top_addr); 4043 br(Assembler::zero, false, Assembler::pt, L); 4044 delayed()->nop(); 4045 stop("eden top is not properly aligned"); 4046 bind(L); 4047 } 4048 #endif // ASSERT 4049 } 4050 } 4051 4052 4053 void MacroAssembler::tlab_allocate( 4054 Register obj, // result: pointer to object after successful allocation 4055 Register var_size_in_bytes, // object size in bytes if unknown at compile time; invalid otherwise 4056 int con_size_in_bytes, // object size in bytes if known at compile time 4057 Register t1, // temp register 4058 Label& slow_case // continuation point if fast allocation fails 4059 ){ 4060 // make sure arguments make sense 4061 assert_different_registers(obj, var_size_in_bytes, t1); 4062 assert(0 <= con_size_in_bytes && is_simm13(con_size_in_bytes), "illegal object size"); 4063 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment"); 4064 4065 const Register free = t1; 4066 4067 verify_tlab(); 4068 4069 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), obj); 4070 4071 // calculate amount of free space 4072 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), free); 4073 sub(free, obj, free); 4074 4075 Label done; 4076 if (var_size_in_bytes == noreg) { 4077 cmp(free, con_size_in_bytes); 4078 } else { 4079 cmp(free, var_size_in_bytes); 4080 } 4081 br(Assembler::less, false, Assembler::pn, slow_case); 4082 // calculate the new top pointer 4083 if (var_size_in_bytes == noreg) { 4084 delayed()->add(obj, con_size_in_bytes, free); 4085 } else { 4086 delayed()->add(obj, var_size_in_bytes, free); 4087 } 4088 4089 bind(done); 4090 4091 #ifdef ASSERT 4092 // make sure new free pointer is properly aligned 4093 { 4094 Label L; 4095 btst(MinObjAlignmentInBytesMask, free); 4096 br(Assembler::zero, false, Assembler::pt, L); 4097 delayed()->nop(); 4098 stop("updated TLAB free is not properly aligned"); 4099 bind(L); 4100 } 4101 #endif // ASSERT 4102 4103 // update the tlab top pointer 4104 st_ptr(free, G2_thread, in_bytes(JavaThread::tlab_top_offset())); 4105 verify_tlab(); 4106 } 4107 4108 4109 void MacroAssembler::tlab_refill(Label& retry, Label& try_eden, Label& slow_case) { 4110 Register top = O0; 4111 Register t1 = G1; 4112 Register t2 = G3; 4113 Register t3 = O1; 4114 assert_different_registers(top, t1, t2, t3, G4, G5 /* preserve G4 and G5 */); 4115 Label do_refill, discard_tlab; 4116 4117 if (CMSIncrementalMode || !Universe::heap()->supports_inline_contig_alloc()) { 4118 // No allocation in the shared eden. 4119 ba(slow_case); 4120 } 4121 4122 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), top); 4123 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t1); 4124 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_refill_waste_limit_offset()), t2); 4125 4126 // calculate amount of free space 4127 sub(t1, top, t1); 4128 srl_ptr(t1, LogHeapWordSize, t1); 4129 4130 // Retain tlab and allocate object in shared space if 4131 // the amount free in the tlab is too large to discard. 4132 cmp(t1, t2); 4133 brx(Assembler::lessEqual, false, Assembler::pt, discard_tlab); 4134 4135 // increment waste limit to prevent getting stuck on this slow path 4136 delayed()->add(t2, ThreadLocalAllocBuffer::refill_waste_limit_increment(), t2); 4137 st_ptr(t2, G2_thread, in_bytes(JavaThread::tlab_refill_waste_limit_offset())); 4138 if (TLABStats) { 4139 // increment number of slow_allocations 4140 ld(G2_thread, in_bytes(JavaThread::tlab_slow_allocations_offset()), t2); 4141 add(t2, 1, t2); 4142 stw(t2, G2_thread, in_bytes(JavaThread::tlab_slow_allocations_offset())); 4143 } 4144 ba(try_eden); 4145 4146 bind(discard_tlab); 4147 if (TLABStats) { 4148 // increment number of refills 4149 ld(G2_thread, in_bytes(JavaThread::tlab_number_of_refills_offset()), t2); 4150 add(t2, 1, t2); 4151 stw(t2, G2_thread, in_bytes(JavaThread::tlab_number_of_refills_offset())); 4152 // accumulate wastage 4153 ld(G2_thread, in_bytes(JavaThread::tlab_fast_refill_waste_offset()), t2); 4154 add(t2, t1, t2); 4155 stw(t2, G2_thread, in_bytes(JavaThread::tlab_fast_refill_waste_offset())); 4156 } 4157 4158 // if tlab is currently allocated (top or end != null) then 4159 // fill [top, end + alignment_reserve) with array object 4160 br_null(top, false, Assembler::pn, do_refill); 4161 4162 set((intptr_t)markOopDesc::prototype()->copy_set_hash(0x2), t2); 4163 st_ptr(t2, top, oopDesc::mark_offset_in_bytes()); // set up the mark word 4164 // set klass to intArrayKlass 4165 sub(t1, typeArrayOopDesc::header_size(T_INT), t1); 4166 add(t1, ThreadLocalAllocBuffer::alignment_reserve(), t1); 4167 sll_ptr(t1, log2_intptr(HeapWordSize/sizeof(jint)), t1); 4168 st(t1, top, arrayOopDesc::length_offset_in_bytes()); 4169 set((intptr_t)Universe::intArrayKlassObj_addr(), t2); 4170 ld_ptr(t2, 0, t2); 4171 // store klass last. concurrent gcs assumes klass length is valid if 4172 // klass field is not null. 4173 store_klass(t2, top); 4174 verify_oop(top); 4175 4176 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t1); 4177 sub(top, t1, t1); // size of tlab's allocated portion 4178 incr_allocated_bytes(t1, t2, t3); 4179 4180 // refill the tlab with an eden allocation 4181 bind(do_refill); 4182 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_size_offset()), t1); 4183 sll_ptr(t1, LogHeapWordSize, t1); 4184 // allocate new tlab, address returned in top 4185 eden_allocate(top, t1, 0, t2, t3, slow_case); 4186 4187 st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_start_offset())); 4188 st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_top_offset())); 4189 #ifdef ASSERT 4190 // check that tlab_size (t1) is still valid 4191 { 4192 Label ok; 4193 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_size_offset()), t2); 4194 sll_ptr(t2, LogHeapWordSize, t2); 4195 cmp_and_br(t1, t2, Assembler::equal, false, Assembler::pn, ok); 4196 stop("assert(t1 == tlab_size)"); 4197 should_not_reach_here(); 4198 4199 bind(ok); 4200 } 4201 #endif // ASSERT 4202 add(top, t1, top); // t1 is tlab_size 4203 sub(top, ThreadLocalAllocBuffer::alignment_reserve_in_bytes(), top); 4204 st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_end_offset())); 4205 verify_tlab(); 4206 ba(retry); 4207 } 4208 4209 void MacroAssembler::incr_allocated_bytes(RegisterOrConstant size_in_bytes, 4210 Register t1, Register t2) { 4211 // Bump total bytes allocated by this thread 4212 assert(t1->is_global(), "must be global reg"); // so all 64 bits are saved on a context switch 4213 assert_different_registers(size_in_bytes.register_or_noreg(), t1, t2); 4214 // v8 support has gone the way of the dodo 4215 ldx(G2_thread, in_bytes(JavaThread::allocated_bytes_offset()), t1); 4216 add(t1, ensure_simm13_or_reg(size_in_bytes, t2), t1); 4217 stx(t1, G2_thread, in_bytes(JavaThread::allocated_bytes_offset())); 4218 } 4219 4220 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 4221 switch (cond) { 4222 // Note some conditions are synonyms for others 4223 case Assembler::never: return Assembler::always; 4224 case Assembler::zero: return Assembler::notZero; 4225 case Assembler::lessEqual: return Assembler::greater; 4226 case Assembler::less: return Assembler::greaterEqual; 4227 case Assembler::lessEqualUnsigned: return Assembler::greaterUnsigned; 4228 case Assembler::lessUnsigned: return Assembler::greaterEqualUnsigned; 4229 case Assembler::negative: return Assembler::positive; 4230 case Assembler::overflowSet: return Assembler::overflowClear; 4231 case Assembler::always: return Assembler::never; 4232 case Assembler::notZero: return Assembler::zero; 4233 case Assembler::greater: return Assembler::lessEqual; 4234 case Assembler::greaterEqual: return Assembler::less; 4235 case Assembler::greaterUnsigned: return Assembler::lessEqualUnsigned; 4236 case Assembler::greaterEqualUnsigned: return Assembler::lessUnsigned; 4237 case Assembler::positive: return Assembler::negative; 4238 case Assembler::overflowClear: return Assembler::overflowSet; 4239 } 4240 4241 ShouldNotReachHere(); return Assembler::overflowClear; 4242 } 4243 4244 void MacroAssembler::cond_inc(Assembler::Condition cond, address counter_ptr, 4245 Register Rtmp1, Register Rtmp2 /*, Register Rtmp3, Register Rtmp4 */) { 4246 Condition negated_cond = negate_condition(cond); 4247 Label L; 4248 brx(negated_cond, false, Assembler::pt, L); 4249 delayed()->nop(); 4250 inc_counter(counter_ptr, Rtmp1, Rtmp2); 4251 bind(L); 4252 } 4253 4254 void MacroAssembler::inc_counter(address counter_addr, Register Rtmp1, Register Rtmp2) { 4255 AddressLiteral addrlit(counter_addr); 4256 sethi(addrlit, Rtmp1); // Move hi22 bits into temporary register. 4257 Address addr(Rtmp1, addrlit.low10()); // Build an address with low10 bits. 4258 ld(addr, Rtmp2); 4259 inc(Rtmp2); 4260 st(Rtmp2, addr); 4261 } 4262 4263 void MacroAssembler::inc_counter(int* counter_addr, Register Rtmp1, Register Rtmp2) { 4264 inc_counter((address) counter_addr, Rtmp1, Rtmp2); 4265 } 4266 4267 SkipIfEqual::SkipIfEqual( 4268 MacroAssembler* masm, Register temp, const bool* flag_addr, 4269 Assembler::Condition condition) { 4270 _masm = masm; 4271 AddressLiteral flag(flag_addr); 4272 _masm->sethi(flag, temp); 4273 _masm->ldub(temp, flag.low10(), temp); 4274 _masm->tst(temp); 4275 _masm->br(condition, false, Assembler::pt, _label); 4276 _masm->delayed()->nop(); 4277 } 4278 4279 SkipIfEqual::~SkipIfEqual() { 4280 _masm->bind(_label); 4281 } 4282 4283 4284 // Writes to stack successive pages until offset reached to check for 4285 // stack overflow + shadow pages. This clobbers tsp and scratch. 4286 void MacroAssembler::bang_stack_size(Register Rsize, Register Rtsp, 4287 Register Rscratch) { 4288 // Use stack pointer in temp stack pointer 4289 mov(SP, Rtsp); 4290 4291 // Bang stack for total size given plus stack shadow page size. 4292 // Bang one page at a time because a large size can overflow yellow and 4293 // red zones (the bang will fail but stack overflow handling can't tell that 4294 // it was a stack overflow bang vs a regular segv). 4295 int offset = os::vm_page_size(); 4296 Register Roffset = Rscratch; 4297 4298 Label loop; 4299 bind(loop); 4300 set((-offset)+STACK_BIAS, Rscratch); 4301 st(G0, Rtsp, Rscratch); 4302 set(offset, Roffset); 4303 sub(Rsize, Roffset, Rsize); 4304 cmp(Rsize, G0); 4305 br(Assembler::greater, false, Assembler::pn, loop); 4306 delayed()->sub(Rtsp, Roffset, Rtsp); 4307 4308 // Bang down shadow pages too. 4309 // The -1 because we already subtracted 1 page. 4310 for (int i = 0; i< StackShadowPages-1; i++) { 4311 set((-i*offset)+STACK_BIAS, Rscratch); 4312 st(G0, Rtsp, Rscratch); 4313 } 4314 } 4315 4316 /////////////////////////////////////////////////////////////////////////////////// 4317 #ifndef SERIALGC 4318 4319 static address satb_log_enqueue_with_frame = NULL; 4320 static u_char* satb_log_enqueue_with_frame_end = NULL; 4321 4322 static address satb_log_enqueue_frameless = NULL; 4323 static u_char* satb_log_enqueue_frameless_end = NULL; 4324 4325 static int EnqueueCodeSize = 128 DEBUG_ONLY( + 256); // Instructions? 4326 4327 static void generate_satb_log_enqueue(bool with_frame) { 4328 BufferBlob* bb = BufferBlob::create("enqueue_with_frame", EnqueueCodeSize); 4329 CodeBuffer buf(bb); 4330 MacroAssembler masm(&buf); 4331 4332 #define __ masm. 4333 4334 address start = __ pc(); 4335 Register pre_val; 4336 4337 Label refill, restart; 4338 if (with_frame) { 4339 __ save_frame(0); 4340 pre_val = I0; // Was O0 before the save. 4341 } else { 4342 pre_val = O0; 4343 } 4344 int satb_q_index_byte_offset = 4345 in_bytes(JavaThread::satb_mark_queue_offset() + 4346 PtrQueue::byte_offset_of_index()); 4347 int satb_q_buf_byte_offset = 4348 in_bytes(JavaThread::satb_mark_queue_offset() + 4349 PtrQueue::byte_offset_of_buf()); 4350 assert(in_bytes(PtrQueue::byte_width_of_index()) == sizeof(intptr_t) && 4351 in_bytes(PtrQueue::byte_width_of_buf()) == sizeof(intptr_t), 4352 "check sizes in assembly below"); 4353 4354 __ bind(restart); 4355 __ ld_ptr(G2_thread, satb_q_index_byte_offset, L0); 4356 4357 __ br_on_reg_cond(Assembler::rc_z, /*annul*/false, Assembler::pn, L0, refill, false); 4358 // If the branch is taken, no harm in executing this in the delay slot. 4359 __ delayed()->ld_ptr(G2_thread, satb_q_buf_byte_offset, L1); 4360 __ sub(L0, oopSize, L0); 4361 4362 __ st_ptr(pre_val, L1, L0); // [_buf + index] := I0 4363 if (!with_frame) { 4364 // Use return-from-leaf 4365 __ retl(); 4366 __ delayed()->st_ptr(L0, G2_thread, satb_q_index_byte_offset); 4367 } else { 4368 // Not delayed. 4369 __ st_ptr(L0, G2_thread, satb_q_index_byte_offset); 4370 } 4371 if (with_frame) { 4372 __ ret(); 4373 __ delayed()->restore(); 4374 } 4375 __ bind(refill); 4376 4377 address handle_zero = 4378 CAST_FROM_FN_PTR(address, 4379 &SATBMarkQueueSet::handle_zero_index_for_thread); 4380 // This should be rare enough that we can afford to save all the 4381 // scratch registers that the calling context might be using. 4382 __ mov(G1_scratch, L0); 4383 __ mov(G3_scratch, L1); 4384 __ mov(G4, L2); 4385 // We need the value of O0 above (for the write into the buffer), so we 4386 // save and restore it. 4387 __ mov(O0, L3); 4388 // Since the call will overwrite O7, we save and restore that, as well. 4389 __ mov(O7, L4); 4390 __ call_VM_leaf(L5, handle_zero, G2_thread); 4391 __ mov(L0, G1_scratch); 4392 __ mov(L1, G3_scratch); 4393 __ mov(L2, G4); 4394 __ mov(L3, O0); 4395 __ br(Assembler::always, /*annul*/false, Assembler::pt, restart); 4396 __ delayed()->mov(L4, O7); 4397 4398 if (with_frame) { 4399 satb_log_enqueue_with_frame = start; 4400 satb_log_enqueue_with_frame_end = __ pc(); 4401 } else { 4402 satb_log_enqueue_frameless = start; 4403 satb_log_enqueue_frameless_end = __ pc(); 4404 } 4405 4406 #undef __ 4407 } 4408 4409 static inline void generate_satb_log_enqueue_if_necessary(bool with_frame) { 4410 if (with_frame) { 4411 if (satb_log_enqueue_with_frame == 0) { 4412 generate_satb_log_enqueue(with_frame); 4413 assert(satb_log_enqueue_with_frame != 0, "postcondition."); 4414 if (G1SATBPrintStubs) { 4415 tty->print_cr("Generated with-frame satb enqueue:"); 4416 Disassembler::decode((u_char*)satb_log_enqueue_with_frame, 4417 satb_log_enqueue_with_frame_end, 4418 tty); 4419 } 4420 } 4421 } else { 4422 if (satb_log_enqueue_frameless == 0) { 4423 generate_satb_log_enqueue(with_frame); 4424 assert(satb_log_enqueue_frameless != 0, "postcondition."); 4425 if (G1SATBPrintStubs) { 4426 tty->print_cr("Generated frameless satb enqueue:"); 4427 Disassembler::decode((u_char*)satb_log_enqueue_frameless, 4428 satb_log_enqueue_frameless_end, 4429 tty); 4430 } 4431 } 4432 } 4433 } 4434 4435 void MacroAssembler::g1_write_barrier_pre(Register obj, 4436 Register index, 4437 int offset, 4438 Register pre_val, 4439 Register tmp, 4440 bool preserve_o_regs) { 4441 Label filtered; 4442 4443 if (obj == noreg) { 4444 // We are not loading the previous value so make 4445 // sure that we don't trash the value in pre_val 4446 // with the code below. 4447 assert_different_registers(pre_val, tmp); 4448 } else { 4449 // We will be loading the previous value 4450 // in this code so... 4451 assert(offset == 0 || index == noreg, "choose one"); 4452 assert(pre_val == noreg, "check this code"); 4453 } 4454 4455 // Is marking active? 4456 if (in_bytes(PtrQueue::byte_width_of_active()) == 4) { 4457 ld(G2, 4458 in_bytes(JavaThread::satb_mark_queue_offset() + 4459 PtrQueue::byte_offset_of_active()), 4460 tmp); 4461 } else { 4462 guarantee(in_bytes(PtrQueue::byte_width_of_active()) == 1, 4463 "Assumption"); 4464 ldsb(G2, 4465 in_bytes(JavaThread::satb_mark_queue_offset() + 4466 PtrQueue::byte_offset_of_active()), 4467 tmp); 4468 } 4469 4470 // Check on whether to annul. 4471 br_on_reg_cond(rc_z, /*annul*/false, Assembler::pt, tmp, filtered); 4472 4473 // Do we need to load the previous value? 4474 if (obj != noreg) { 4475 // Load the previous value... 4476 if (index == noreg) { 4477 if (Assembler::is_simm13(offset)) { 4478 load_heap_oop(obj, offset, tmp); 4479 } else { 4480 set(offset, tmp); 4481 load_heap_oop(obj, tmp, tmp); 4482 } 4483 } else { 4484 load_heap_oop(obj, index, tmp); 4485 } 4486 // Previous value has been loaded into tmp 4487 pre_val = tmp; 4488 } 4489 4490 assert(pre_val != noreg, "must have a real register"); 4491 4492 // Is the previous value null? 4493 // Check on whether to annul. 4494 br_on_reg_cond(rc_z, /*annul*/false, Assembler::pt, pre_val, filtered); 4495 4496 // OK, it's not filtered, so we'll need to call enqueue. In the normal 4497 // case, pre_val will be a scratch G-reg, but there are some cases in 4498 // which it's an O-reg. In the first case, do a normal call. In the 4499 // latter, do a save here and call the frameless version. 4500 4501 guarantee(pre_val->is_global() || pre_val->is_out(), 4502 "Or we need to think harder."); 4503 4504 if (pre_val->is_global() && !preserve_o_regs) { 4505 generate_satb_log_enqueue_if_necessary(true); // with frame 4506 4507 call(satb_log_enqueue_with_frame); 4508 delayed()->mov(pre_val, O0); 4509 } else { 4510 generate_satb_log_enqueue_if_necessary(false); // frameless 4511 4512 save_frame(0); 4513 call(satb_log_enqueue_frameless); 4514 delayed()->mov(pre_val->after_save(), O0); 4515 restore(); 4516 } 4517 4518 bind(filtered); 4519 } 4520 4521 static jint num_ct_writes = 0; 4522 static jint num_ct_writes_filtered_in_hr = 0; 4523 static jint num_ct_writes_filtered_null = 0; 4524 static G1CollectedHeap* g1 = NULL; 4525 4526 static Thread* count_ct_writes(void* filter_val, void* new_val) { 4527 Atomic::inc(&num_ct_writes); 4528 if (filter_val == NULL) { 4529 Atomic::inc(&num_ct_writes_filtered_in_hr); 4530 } else if (new_val == NULL) { 4531 Atomic::inc(&num_ct_writes_filtered_null); 4532 } else { 4533 if (g1 == NULL) { 4534 g1 = G1CollectedHeap::heap(); 4535 } 4536 } 4537 if ((num_ct_writes % 1000000) == 0) { 4538 jint num_ct_writes_filtered = 4539 num_ct_writes_filtered_in_hr + 4540 num_ct_writes_filtered_null; 4541 4542 tty->print_cr("%d potential CT writes: %5.2f%% filtered\n" 4543 " (%5.2f%% intra-HR, %5.2f%% null).", 4544 num_ct_writes, 4545 100.0*(float)num_ct_writes_filtered/(float)num_ct_writes, 4546 100.0*(float)num_ct_writes_filtered_in_hr/ 4547 (float)num_ct_writes, 4548 100.0*(float)num_ct_writes_filtered_null/ 4549 (float)num_ct_writes); 4550 } 4551 return Thread::current(); 4552 } 4553 4554 static address dirty_card_log_enqueue = 0; 4555 static u_char* dirty_card_log_enqueue_end = 0; 4556 4557 // This gets to assume that o0 contains the object address. 4558 static void generate_dirty_card_log_enqueue(jbyte* byte_map_base) { 4559 BufferBlob* bb = BufferBlob::create("dirty_card_enqueue", EnqueueCodeSize*2); 4560 CodeBuffer buf(bb); 4561 MacroAssembler masm(&buf); 4562 #define __ masm. 4563 address start = __ pc(); 4564 4565 Label not_already_dirty, restart, refill; 4566 4567 #ifdef _LP64 4568 __ srlx(O0, CardTableModRefBS::card_shift, O0); 4569 #else 4570 __ srl(O0, CardTableModRefBS::card_shift, O0); 4571 #endif 4572 AddressLiteral addrlit(byte_map_base); 4573 __ set(addrlit, O1); // O1 := <card table base> 4574 __ ldub(O0, O1, O2); // O2 := [O0 + O1] 4575 4576 __ br_on_reg_cond(Assembler::rc_nz, /*annul*/false, Assembler::pt, 4577 O2, not_already_dirty, false); 4578 // Get O1 + O2 into a reg by itself -- useful in the take-the-branch 4579 // case, harmless if not. 4580 __ delayed()->add(O0, O1, O3); 4581 4582 // We didn't take the branch, so we're already dirty: return. 4583 // Use return-from-leaf 4584 __ retl(); 4585 __ delayed()->nop(); 4586 4587 // Not dirty. 4588 __ bind(not_already_dirty); 4589 // First, dirty it. 4590 __ stb(G0, O3, G0); // [cardPtr] := 0 (i.e., dirty). 4591 int dirty_card_q_index_byte_offset = 4592 in_bytes(JavaThread::dirty_card_queue_offset() + 4593 PtrQueue::byte_offset_of_index()); 4594 int dirty_card_q_buf_byte_offset = 4595 in_bytes(JavaThread::dirty_card_queue_offset() + 4596 PtrQueue::byte_offset_of_buf()); 4597 __ bind(restart); 4598 __ ld_ptr(G2_thread, dirty_card_q_index_byte_offset, L0); 4599 4600 __ br_on_reg_cond(Assembler::rc_z, /*annul*/false, Assembler::pn, 4601 L0, refill, false); 4602 // If the branch is taken, no harm in executing this in the delay slot. 4603 __ delayed()->ld_ptr(G2_thread, dirty_card_q_buf_byte_offset, L1); 4604 __ sub(L0, oopSize, L0); 4605 4606 __ st_ptr(O3, L1, L0); // [_buf + index] := I0 4607 // Use return-from-leaf 4608 __ retl(); 4609 __ delayed()->st_ptr(L0, G2_thread, dirty_card_q_index_byte_offset); 4610 4611 __ bind(refill); 4612 address handle_zero = 4613 CAST_FROM_FN_PTR(address, 4614 &DirtyCardQueueSet::handle_zero_index_for_thread); 4615 // This should be rare enough that we can afford to save all the 4616 // scratch registers that the calling context might be using. 4617 __ mov(G1_scratch, L3); 4618 __ mov(G3_scratch, L5); 4619 // We need the value of O3 above (for the write into the buffer), so we 4620 // save and restore it. 4621 __ mov(O3, L6); 4622 // Since the call will overwrite O7, we save and restore that, as well. 4623 __ mov(O7, L4); 4624 4625 __ call_VM_leaf(L7_thread_cache, handle_zero, G2_thread); 4626 __ mov(L3, G1_scratch); 4627 __ mov(L5, G3_scratch); 4628 __ mov(L6, O3); 4629 __ br(Assembler::always, /*annul*/false, Assembler::pt, restart); 4630 __ delayed()->mov(L4, O7); 4631 4632 dirty_card_log_enqueue = start; 4633 dirty_card_log_enqueue_end = __ pc(); 4634 // XXX Should have a guarantee here about not going off the end! 4635 // Does it already do so? Do an experiment... 4636 4637 #undef __ 4638 4639 } 4640 4641 static inline void 4642 generate_dirty_card_log_enqueue_if_necessary(jbyte* byte_map_base) { 4643 if (dirty_card_log_enqueue == 0) { 4644 generate_dirty_card_log_enqueue(byte_map_base); 4645 assert(dirty_card_log_enqueue != 0, "postcondition."); 4646 if (G1SATBPrintStubs) { 4647 tty->print_cr("Generated dirty_card enqueue:"); 4648 Disassembler::decode((u_char*)dirty_card_log_enqueue, 4649 dirty_card_log_enqueue_end, 4650 tty); 4651 } 4652 } 4653 } 4654 4655 4656 void MacroAssembler::g1_write_barrier_post(Register store_addr, Register new_val, Register tmp) { 4657 4658 Label filtered; 4659 MacroAssembler* post_filter_masm = this; 4660 4661 if (new_val == G0) return; 4662 4663 G1SATBCardTableModRefBS* bs = (G1SATBCardTableModRefBS*) Universe::heap()->barrier_set(); 4664 assert(bs->kind() == BarrierSet::G1SATBCT || 4665 bs->kind() == BarrierSet::G1SATBCTLogging, "wrong barrier"); 4666 if (G1RSBarrierRegionFilter) { 4667 xor3(store_addr, new_val, tmp); 4668 #ifdef _LP64 4669 srlx(tmp, HeapRegion::LogOfHRGrainBytes, tmp); 4670 #else 4671 srl(tmp, HeapRegion::LogOfHRGrainBytes, tmp); 4672 #endif 4673 4674 if (G1PrintCTFilterStats) { 4675 guarantee(tmp->is_global(), "Or stats won't work..."); 4676 // This is a sleazy hack: I'm temporarily hijacking G2, which I 4677 // promise to restore. 4678 mov(new_val, G2); 4679 save_frame(0); 4680 mov(tmp, O0); 4681 mov(G2, O1); 4682 // Save G-regs that target may use. 4683 mov(G1, L1); 4684 mov(G2, L2); 4685 mov(G3, L3); 4686 mov(G4, L4); 4687 mov(G5, L5); 4688 call(CAST_FROM_FN_PTR(address, &count_ct_writes)); 4689 delayed()->nop(); 4690 mov(O0, G2); 4691 // Restore G-regs that target may have used. 4692 mov(L1, G1); 4693 mov(L3, G3); 4694 mov(L4, G4); 4695 mov(L5, G5); 4696 restore(G0, G0, G0); 4697 } 4698 // XXX Should I predict this taken or not? Does it mattern? 4699 br_on_reg_cond(rc_z, /*annul*/false, Assembler::pt, tmp, filtered); 4700 } 4701 4702 // If the "store_addr" register is an "in" or "local" register, move it to 4703 // a scratch reg so we can pass it as an argument. 4704 bool use_scr = !(store_addr->is_global() || store_addr->is_out()); 4705 // Pick a scratch register different from "tmp". 4706 Register scr = (tmp == G1_scratch ? G3_scratch : G1_scratch); 4707 // Make sure we use up the delay slot! 4708 if (use_scr) { 4709 post_filter_masm->mov(store_addr, scr); 4710 } else { 4711 post_filter_masm->nop(); 4712 } 4713 generate_dirty_card_log_enqueue_if_necessary(bs->byte_map_base); 4714 save_frame(0); 4715 call(dirty_card_log_enqueue); 4716 if (use_scr) { 4717 delayed()->mov(scr, O0); 4718 } else { 4719 delayed()->mov(store_addr->after_save(), O0); 4720 } 4721 restore(); 4722 4723 bind(filtered); 4724 4725 } 4726 4727 #endif // SERIALGC 4728 /////////////////////////////////////////////////////////////////////////////////// 4729 4730 void MacroAssembler::card_write_barrier_post(Register store_addr, Register new_val, Register tmp) { 4731 // If we're writing constant NULL, we can skip the write barrier. 4732 if (new_val == G0) return; 4733 CardTableModRefBS* bs = (CardTableModRefBS*) Universe::heap()->barrier_set(); 4734 assert(bs->kind() == BarrierSet::CardTableModRef || 4735 bs->kind() == BarrierSet::CardTableExtension, "wrong barrier"); 4736 card_table_write(bs->byte_map_base, tmp, store_addr); 4737 } 4738 4739 void MacroAssembler::load_klass(Register src_oop, Register klass) { 4740 // The number of bytes in this code is used by 4741 // MachCallDynamicJavaNode::ret_addr_offset() 4742 // if this changes, change that. 4743 if (UseCompressedOops) { 4744 lduw(src_oop, oopDesc::klass_offset_in_bytes(), klass); 4745 decode_heap_oop_not_null(klass); 4746 } else { 4747 ld_ptr(src_oop, oopDesc::klass_offset_in_bytes(), klass); 4748 } 4749 } 4750 4751 void MacroAssembler::store_klass(Register klass, Register dst_oop) { 4752 if (UseCompressedOops) { 4753 assert(dst_oop != klass, "not enough registers"); 4754 encode_heap_oop_not_null(klass); 4755 st(klass, dst_oop, oopDesc::klass_offset_in_bytes()); 4756 } else { 4757 st_ptr(klass, dst_oop, oopDesc::klass_offset_in_bytes()); 4758 } 4759 } 4760 4761 void MacroAssembler::store_klass_gap(Register s, Register d) { 4762 if (UseCompressedOops) { 4763 assert(s != d, "not enough registers"); 4764 st(s, d, oopDesc::klass_gap_offset_in_bytes()); 4765 } 4766 } 4767 4768 void MacroAssembler::load_heap_oop(const Address& s, Register d) { 4769 if (UseCompressedOops) { 4770 lduw(s, d); 4771 decode_heap_oop(d); 4772 } else { 4773 ld_ptr(s, d); 4774 } 4775 } 4776 4777 void MacroAssembler::load_heap_oop(Register s1, Register s2, Register d) { 4778 if (UseCompressedOops) { 4779 lduw(s1, s2, d); 4780 decode_heap_oop(d, d); 4781 } else { 4782 ld_ptr(s1, s2, d); 4783 } 4784 } 4785 4786 void MacroAssembler::load_heap_oop(Register s1, int simm13a, Register d) { 4787 if (UseCompressedOops) { 4788 lduw(s1, simm13a, d); 4789 decode_heap_oop(d, d); 4790 } else { 4791 ld_ptr(s1, simm13a, d); 4792 } 4793 } 4794 4795 void MacroAssembler::load_heap_oop(Register s1, RegisterOrConstant s2, Register d) { 4796 if (s2.is_constant()) load_heap_oop(s1, s2.as_constant(), d); 4797 else load_heap_oop(s1, s2.as_register(), d); 4798 } 4799 4800 void MacroAssembler::store_heap_oop(Register d, Register s1, Register s2) { 4801 if (UseCompressedOops) { 4802 assert(s1 != d && s2 != d, "not enough registers"); 4803 encode_heap_oop(d); 4804 st(d, s1, s2); 4805 } else { 4806 st_ptr(d, s1, s2); 4807 } 4808 } 4809 4810 void MacroAssembler::store_heap_oop(Register d, Register s1, int simm13a) { 4811 if (UseCompressedOops) { 4812 assert(s1 != d, "not enough registers"); 4813 encode_heap_oop(d); 4814 st(d, s1, simm13a); 4815 } else { 4816 st_ptr(d, s1, simm13a); 4817 } 4818 } 4819 4820 void MacroAssembler::store_heap_oop(Register d, const Address& a, int offset) { 4821 if (UseCompressedOops) { 4822 assert(a.base() != d, "not enough registers"); 4823 encode_heap_oop(d); 4824 st(d, a, offset); 4825 } else { 4826 st_ptr(d, a, offset); 4827 } 4828 } 4829 4830 4831 void MacroAssembler::encode_heap_oop(Register src, Register dst) { 4832 assert (UseCompressedOops, "must be compressed"); 4833 assert (Universe::heap() != NULL, "java heap should be initialized"); 4834 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 4835 verify_oop(src); 4836 if (Universe::narrow_oop_base() == NULL) { 4837 srlx(src, LogMinObjAlignmentInBytes, dst); 4838 return; 4839 } 4840 Label done; 4841 if (src == dst) { 4842 // optimize for frequent case src == dst 4843 bpr(rc_nz, true, Assembler::pt, src, done); 4844 delayed() -> sub(src, G6_heapbase, dst); // annuled if not taken 4845 bind(done); 4846 srlx(src, LogMinObjAlignmentInBytes, dst); 4847 } else { 4848 bpr(rc_z, false, Assembler::pn, src, done); 4849 delayed() -> mov(G0, dst); 4850 // could be moved before branch, and annulate delay, 4851 // but may add some unneeded work decoding null 4852 sub(src, G6_heapbase, dst); 4853 srlx(dst, LogMinObjAlignmentInBytes, dst); 4854 bind(done); 4855 } 4856 } 4857 4858 4859 void MacroAssembler::encode_heap_oop_not_null(Register r) { 4860 assert (UseCompressedOops, "must be compressed"); 4861 assert (Universe::heap() != NULL, "java heap should be initialized"); 4862 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 4863 verify_oop(r); 4864 if (Universe::narrow_oop_base() != NULL) 4865 sub(r, G6_heapbase, r); 4866 srlx(r, LogMinObjAlignmentInBytes, r); 4867 } 4868 4869 void MacroAssembler::encode_heap_oop_not_null(Register src, Register dst) { 4870 assert (UseCompressedOops, "must be compressed"); 4871 assert (Universe::heap() != NULL, "java heap should be initialized"); 4872 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 4873 verify_oop(src); 4874 if (Universe::narrow_oop_base() == NULL) { 4875 srlx(src, LogMinObjAlignmentInBytes, dst); 4876 } else { 4877 sub(src, G6_heapbase, dst); 4878 srlx(dst, LogMinObjAlignmentInBytes, dst); 4879 } 4880 } 4881 4882 // Same algorithm as oops.inline.hpp decode_heap_oop. 4883 void MacroAssembler::decode_heap_oop(Register src, Register dst) { 4884 assert (UseCompressedOops, "must be compressed"); 4885 assert (Universe::heap() != NULL, "java heap should be initialized"); 4886 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 4887 sllx(src, LogMinObjAlignmentInBytes, dst); 4888 if (Universe::narrow_oop_base() != NULL) { 4889 Label done; 4890 bpr(rc_nz, true, Assembler::pt, dst, done); 4891 delayed() -> add(dst, G6_heapbase, dst); // annuled if not taken 4892 bind(done); 4893 } 4894 verify_oop(dst); 4895 } 4896 4897 void MacroAssembler::decode_heap_oop_not_null(Register r) { 4898 // Do not add assert code to this unless you change vtableStubs_sparc.cpp 4899 // pd_code_size_limit. 4900 // Also do not verify_oop as this is called by verify_oop. 4901 assert (UseCompressedOops, "must be compressed"); 4902 assert (Universe::heap() != NULL, "java heap should be initialized"); 4903 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 4904 sllx(r, LogMinObjAlignmentInBytes, r); 4905 if (Universe::narrow_oop_base() != NULL) 4906 add(r, G6_heapbase, r); 4907 } 4908 4909 void MacroAssembler::decode_heap_oop_not_null(Register src, Register dst) { 4910 // Do not add assert code to this unless you change vtableStubs_sparc.cpp 4911 // pd_code_size_limit. 4912 // Also do not verify_oop as this is called by verify_oop. 4913 assert (UseCompressedOops, "must be compressed"); 4914 assert (Universe::heap() != NULL, "java heap should be initialized"); 4915 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 4916 sllx(src, LogMinObjAlignmentInBytes, dst); 4917 if (Universe::narrow_oop_base() != NULL) 4918 add(dst, G6_heapbase, dst); 4919 } 4920 4921 void MacroAssembler::reinit_heapbase() { 4922 if (UseCompressedOops) { 4923 // call indirectly to solve generation ordering problem 4924 AddressLiteral base(Universe::narrow_oop_base_addr()); 4925 load_ptr_contents(base, G6_heapbase); 4926 } 4927 } 4928 4929 // Compare char[] arrays aligned to 4 bytes. 4930 void MacroAssembler::char_arrays_equals(Register ary1, Register ary2, 4931 Register limit, Register result, 4932 Register chr1, Register chr2, Label& Ldone) { 4933 Label Lvector, Lloop; 4934 assert(chr1 == result, "should be the same"); 4935 4936 // Note: limit contains number of bytes (2*char_elements) != 0. 4937 andcc(limit, 0x2, chr1); // trailing character ? 4938 br(Assembler::zero, false, Assembler::pt, Lvector); 4939 delayed()->nop(); 4940 4941 // compare the trailing char 4942 sub(limit, sizeof(jchar), limit); 4943 lduh(ary1, limit, chr1); 4944 lduh(ary2, limit, chr2); 4945 cmp(chr1, chr2); 4946 br(Assembler::notEqual, true, Assembler::pt, Ldone); 4947 delayed()->mov(G0, result); // not equal 4948 4949 // only one char ? 4950 bpr(rc_z, true, Assembler::pn, limit, Ldone); 4951 delayed()->add(G0, 1, result); // zero-length arrays are equal 4952 4953 // word by word compare, dont't need alignment check 4954 bind(Lvector); 4955 // Shift ary1 and ary2 to the end of the arrays, negate limit 4956 add(ary1, limit, ary1); 4957 add(ary2, limit, ary2); 4958 neg(limit, limit); 4959 4960 lduw(ary1, limit, chr1); 4961 bind(Lloop); 4962 lduw(ary2, limit, chr2); 4963 cmp(chr1, chr2); 4964 br(Assembler::notEqual, true, Assembler::pt, Ldone); 4965 delayed()->mov(G0, result); // not equal 4966 inccc(limit, 2*sizeof(jchar)); 4967 // annul LDUW if branch is not taken to prevent access past end of array 4968 br(Assembler::notZero, true, Assembler::pt, Lloop); 4969 delayed()->lduw(ary1, limit, chr1); // hoisted 4970 4971 // Caller should set it: 4972 // add(G0, 1, result); // equals 4973 }