1 /*
   2  * Copyright (c) 1997, 2013, 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.inline.hpp"
  27 #include "compiler/disassembler.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 "memory/universe.hpp"
  33 #include "prims/methodHandles.hpp"
  34 #include "runtime/biasedLocking.hpp"
  35 #include "runtime/interfaceSupport.hpp"
  36 #include "runtime/objectMonitor.hpp"
  37 #include "runtime/os.hpp"
  38 #include "runtime/sharedRuntime.hpp"
  39 #include "runtime/stubRoutines.hpp"
  40 #include "utilities/macros.hpp"
  41 #if INCLUDE_ALL_GCS
  42 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  43 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
  44 #include "gc_implementation/g1/heapRegion.hpp"
  45 #endif // INCLUDE_ALL_GCS
  46 
  47 #ifdef PRODUCT
  48 #define BLOCK_COMMENT(str) /* nothing */
  49 #define STOP(error) stop(error)
  50 #else
  51 #define BLOCK_COMMENT(str) block_comment(str)
  52 #define STOP(error) block_comment(error); stop(error)
  53 #endif
  54 
  55 // Convert the raw encoding form into the form expected by the
  56 // constructor for Address.
  57 Address Address::make_raw(int base, int index, int scale, int disp, relocInfo::relocType disp_reloc) {
  58   assert(scale == 0, "not supported");
  59   RelocationHolder rspec;
  60   if (disp_reloc != relocInfo::none) {
  61     rspec = Relocation::spec_simple(disp_reloc);
  62   }
  63 
  64   Register rindex = as_Register(index);
  65   if (rindex != G0) {
  66     Address madr(as_Register(base), rindex);
  67     madr._rspec = rspec;
  68     return madr;
  69   } else {
  70     Address madr(as_Register(base), disp);
  71     madr._rspec = rspec;
  72     return madr;
  73   }
  74 }
  75 
  76 Address Argument::address_in_frame() const {
  77   // Warning: In LP64 mode disp will occupy more than 10 bits, but
  78   //          op codes such as ld or ldx, only access disp() to get
  79   //          their simm13 argument.
  80   int disp = ((_number - Argument::n_register_parameters + frame::memory_parameter_word_sp_offset) * BytesPerWord) + STACK_BIAS;
  81   if (is_in())
  82     return Address(FP, disp); // In argument.
  83   else
  84     return Address(SP, disp); // Out argument.
  85 }
  86 
  87 static const char* argumentNames[][2] = {
  88   {"A0","P0"}, {"A1","P1"}, {"A2","P2"}, {"A3","P3"}, {"A4","P4"},
  89   {"A5","P5"}, {"A6","P6"}, {"A7","P7"}, {"A8","P8"}, {"A9","P9"},
  90   {"A(n>9)","P(n>9)"}
  91 };
  92 
  93 const char* Argument::name() const {
  94   int nofArgs = sizeof argumentNames / sizeof argumentNames[0];
  95   int num = number();
  96   if (num >= nofArgs)  num = nofArgs - 1;
  97   return argumentNames[num][is_in() ? 1 : 0];
  98 }
  99 
 100 #ifdef ASSERT
 101 // On RISC, there's no benefit to verifying instruction boundaries.
 102 bool AbstractAssembler::pd_check_instruction_mark() { return false; }
 103 #endif
 104 
 105 // Patch instruction inst at offset inst_pos to refer to dest_pos
 106 // and return the resulting instruction.
 107 // We should have pcs, not offsets, but since all is relative, it will work out
 108 // OK.
 109 int MacroAssembler::patched_branch(int dest_pos, int inst, int inst_pos) {
 110   int m; // mask for displacement field
 111   int v; // new value for displacement field
 112   const int word_aligned_ones = -4;
 113   switch (inv_op(inst)) {
 114   default: ShouldNotReachHere();
 115   case call_op:    m = wdisp(word_aligned_ones, 0, 30);  v = wdisp(dest_pos, inst_pos, 30); break;
 116   case branch_op:
 117     switch (inv_op2(inst)) {
 118       case fbp_op2:    m = wdisp(  word_aligned_ones, 0, 19);  v = wdisp(  dest_pos, inst_pos, 19); break;
 119       case bp_op2:     m = wdisp(  word_aligned_ones, 0, 19);  v = wdisp(  dest_pos, inst_pos, 19); break;
 120       case fb_op2:     m = wdisp(  word_aligned_ones, 0, 22);  v = wdisp(  dest_pos, inst_pos, 22); break;
 121       case br_op2:     m = wdisp(  word_aligned_ones, 0, 22);  v = wdisp(  dest_pos, inst_pos, 22); break;
 122       case bpr_op2: {
 123         if (is_cbcond(inst)) {
 124           m = wdisp10(word_aligned_ones, 0);
 125           v = wdisp10(dest_pos, inst_pos);
 126         } else {
 127           m = wdisp16(word_aligned_ones, 0);
 128           v = wdisp16(dest_pos, inst_pos);
 129         }
 130         break;
 131       }
 132       default: ShouldNotReachHere();
 133     }
 134   }
 135   return  inst & ~m  |  v;
 136 }
 137 
 138 // Return the offset of the branch destionation of instruction inst
 139 // at offset pos.
 140 // Should have pcs, but since all is relative, it works out.
 141 int MacroAssembler::branch_destination(int inst, int pos) {
 142   int r;
 143   switch (inv_op(inst)) {
 144   default: ShouldNotReachHere();
 145   case call_op:        r = inv_wdisp(inst, pos, 30);  break;
 146   case branch_op:
 147     switch (inv_op2(inst)) {
 148       case fbp_op2:    r = inv_wdisp(  inst, pos, 19);  break;
 149       case bp_op2:     r = inv_wdisp(  inst, pos, 19);  break;
 150       case fb_op2:     r = inv_wdisp(  inst, pos, 22);  break;
 151       case br_op2:     r = inv_wdisp(  inst, pos, 22);  break;
 152       case bpr_op2: {
 153         if (is_cbcond(inst)) {
 154           r = inv_wdisp10(inst, pos);
 155         } else {
 156           r = inv_wdisp16(inst, pos);
 157         }
 158         break;
 159       }
 160       default: ShouldNotReachHere();
 161     }
 162   }
 163   return r;
 164 }
 165 
 166 void MacroAssembler::null_check(Register reg, int offset) {
 167   if (needs_explicit_null_check((intptr_t)offset)) {
 168     // provoke OS NULL exception if reg = NULL by
 169     // accessing M[reg] w/o changing any registers
 170     ld_ptr(reg, 0, G0);
 171   }
 172   else {
 173     // nothing to do, (later) access of M[reg + offset]
 174     // will provoke OS NULL exception if reg = NULL
 175   }
 176 }
 177 
 178 // Ring buffer jumps
 179 
 180 #ifndef PRODUCT
 181 void MacroAssembler::ret(  bool trace )   { if (trace) {
 182                                                     mov(I7, O7); // traceable register
 183                                                     JMP(O7, 2 * BytesPerInstWord);
 184                                                   } else {
 185                                                     jmpl( I7, 2 * BytesPerInstWord, G0 );
 186                                                   }
 187                                                 }
 188 
 189 void MacroAssembler::retl( bool trace )  { if (trace) JMP(O7, 2 * BytesPerInstWord);
 190                                                  else jmpl( O7, 2 * BytesPerInstWord, G0 ); }
 191 #endif /* PRODUCT */
 192 
 193 
 194 void MacroAssembler::jmp2(Register r1, Register r2, const char* file, int line ) {
 195   assert_not_delayed();
 196   // This can only be traceable if r1 & r2 are visible after a window save
 197   if (TraceJumps) {
 198 #ifndef PRODUCT
 199     save_frame(0);
 200     verify_thread();
 201     ld(G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()), O0);
 202     add(G2_thread, in_bytes(JavaThread::jmp_ring_offset()), O1);
 203     sll(O0, exact_log2(4*sizeof(intptr_t)), O2);
 204     add(O2, O1, O1);
 205 
 206     add(r1->after_save(), r2->after_save(), O2);
 207     set((intptr_t)file, O3);
 208     set(line, O4);
 209     Label L;
 210     // get nearby pc, store jmp target
 211     call(L, relocInfo::none);  // No relocation for call to pc+0x8
 212     delayed()->st(O2, O1, 0);
 213     bind(L);
 214 
 215     // store nearby pc
 216     st(O7, O1, sizeof(intptr_t));
 217     // store file
 218     st(O3, O1, 2*sizeof(intptr_t));
 219     // store line
 220     st(O4, O1, 3*sizeof(intptr_t));
 221     add(O0, 1, O0);
 222     and3(O0, JavaThread::jump_ring_buffer_size  - 1, O0);
 223     st(O0, G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()));
 224     restore();
 225 #endif /* PRODUCT */
 226   }
 227   jmpl(r1, r2, G0);
 228 }
 229 void MacroAssembler::jmp(Register r1, int offset, const char* file, int line ) {
 230   assert_not_delayed();
 231   // This can only be traceable if r1 is visible after a window save
 232   if (TraceJumps) {
 233 #ifndef PRODUCT
 234     save_frame(0);
 235     verify_thread();
 236     ld(G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()), O0);
 237     add(G2_thread, in_bytes(JavaThread::jmp_ring_offset()), O1);
 238     sll(O0, exact_log2(4*sizeof(intptr_t)), O2);
 239     add(O2, O1, O1);
 240 
 241     add(r1->after_save(), offset, O2);
 242     set((intptr_t)file, O3);
 243     set(line, O4);
 244     Label L;
 245     // get nearby pc, store jmp target
 246     call(L, relocInfo::none);  // No relocation for call to pc+0x8
 247     delayed()->st(O2, O1, 0);
 248     bind(L);
 249 
 250     // store nearby pc
 251     st(O7, O1, sizeof(intptr_t));
 252     // store file
 253     st(O3, O1, 2*sizeof(intptr_t));
 254     // store line
 255     st(O4, O1, 3*sizeof(intptr_t));
 256     add(O0, 1, O0);
 257     and3(O0, JavaThread::jump_ring_buffer_size  - 1, O0);
 258     st(O0, G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()));
 259     restore();
 260 #endif /* PRODUCT */
 261   }
 262   jmp(r1, offset);
 263 }
 264 
 265 // This code sequence is relocatable to any address, even on LP64.
 266 void MacroAssembler::jumpl(const AddressLiteral& addrlit, Register temp, Register d, int offset, const char* file, int line) {
 267   assert_not_delayed();
 268   // Force fixed length sethi because NativeJump and NativeFarCall don't handle
 269   // variable length instruction streams.
 270   patchable_sethi(addrlit, temp);
 271   Address a(temp, addrlit.low10() + offset);  // Add the offset to the displacement.
 272   if (TraceJumps) {
 273 #ifndef PRODUCT
 274     // Must do the add here so relocation can find the remainder of the
 275     // value to be relocated.
 276     add(a.base(), a.disp(), a.base(), addrlit.rspec(offset));
 277     save_frame(0);
 278     verify_thread();
 279     ld(G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()), O0);
 280     add(G2_thread, in_bytes(JavaThread::jmp_ring_offset()), O1);
 281     sll(O0, exact_log2(4*sizeof(intptr_t)), O2);
 282     add(O2, O1, O1);
 283 
 284     set((intptr_t)file, O3);
 285     set(line, O4);
 286     Label L;
 287 
 288     // get nearby pc, store jmp target
 289     call(L, relocInfo::none);  // No relocation for call to pc+0x8
 290     delayed()->st(a.base()->after_save(), O1, 0);
 291     bind(L);
 292 
 293     // store nearby pc
 294     st(O7, O1, sizeof(intptr_t));
 295     // store file
 296     st(O3, O1, 2*sizeof(intptr_t));
 297     // store line
 298     st(O4, O1, 3*sizeof(intptr_t));
 299     add(O0, 1, O0);
 300     and3(O0, JavaThread::jump_ring_buffer_size  - 1, O0);
 301     st(O0, G2_thread, in_bytes(JavaThread::jmp_ring_index_offset()));
 302     restore();
 303     jmpl(a.base(), G0, d);
 304 #else
 305     jmpl(a.base(), a.disp(), d);
 306 #endif /* PRODUCT */
 307   } else {
 308     jmpl(a.base(), a.disp(), d);
 309   }
 310 }
 311 
 312 void MacroAssembler::jump(const AddressLiteral& addrlit, Register temp, int offset, const char* file, int line) {
 313   jumpl(addrlit, temp, G0, offset, file, line);
 314 }
 315 
 316 
 317 // Conditional breakpoint (for assertion checks in assembly code)
 318 void MacroAssembler::breakpoint_trap(Condition c, CC cc) {
 319   trap(c, cc, G0, ST_RESERVED_FOR_USER_0);
 320 }
 321 
 322 // We want to use ST_BREAKPOINT here, but the debugger is confused by it.
 323 void MacroAssembler::breakpoint_trap() {
 324   trap(ST_RESERVED_FOR_USER_0);
 325 }
 326 
 327 // Write serialization page so VM thread can do a pseudo remote membar
 328 // We use the current thread pointer to calculate a thread specific
 329 // offset to write to within the page. This minimizes bus traffic
 330 // due to cache line collision.
 331 void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) {
 332   srl(thread, os::get_serialize_page_shift_count(), tmp2);
 333   if (Assembler::is_simm13(os::vm_page_size())) {
 334     and3(tmp2, (os::vm_page_size() - sizeof(int)), tmp2);
 335   }
 336   else {
 337     set((os::vm_page_size() - sizeof(int)), tmp1);
 338     and3(tmp2, tmp1, tmp2);
 339   }
 340   set(os::get_memory_serialize_page(), tmp1);
 341   st(G0, tmp1, tmp2);
 342 }
 343 
 344 
 345 
 346 void MacroAssembler::enter() {
 347   Unimplemented();
 348 }
 349 
 350 void MacroAssembler::leave() {
 351   Unimplemented();
 352 }
 353 
 354 // Calls to C land
 355 
 356 #ifdef ASSERT
 357 // a hook for debugging
 358 static Thread* reinitialize_thread() {
 359   return ThreadLocalStorage::thread();
 360 }
 361 #else
 362 #define reinitialize_thread ThreadLocalStorage::thread
 363 #endif
 364 
 365 #ifdef ASSERT
 366 address last_get_thread = NULL;
 367 #endif
 368 
 369 // call this when G2_thread is not known to be valid
 370 void MacroAssembler::get_thread() {
 371   save_frame(0);                // to avoid clobbering O0
 372   mov(G1, L0);                  // avoid clobbering G1
 373   mov(G5_method, L1);           // avoid clobbering G5
 374   mov(G3, L2);                  // avoid clobbering G3 also
 375   mov(G4, L5);                  // avoid clobbering G4
 376 #ifdef ASSERT
 377   AddressLiteral last_get_thread_addrlit(&last_get_thread);
 378   set(last_get_thread_addrlit, L3);
 379   rdpc(L4);
 380   inc(L4, 3 * BytesPerInstWord); // skip rdpc + inc + st_ptr to point L4 at call  st_ptr(L4, L3, 0);
 381 #endif
 382   call(CAST_FROM_FN_PTR(address, reinitialize_thread), relocInfo::runtime_call_type);
 383   delayed()->nop();
 384   mov(L0, G1);
 385   mov(L1, G5_method);
 386   mov(L2, G3);
 387   mov(L5, G4);
 388   restore(O0, 0, G2_thread);
 389 }
 390 
 391 static Thread* verify_thread_subroutine(Thread* gthread_value) {
 392   Thread* correct_value = ThreadLocalStorage::thread();
 393   guarantee(gthread_value == correct_value, "G2_thread value must be the thread");
 394   return correct_value;
 395 }
 396 
 397 void MacroAssembler::verify_thread() {
 398   if (VerifyThread) {
 399     // NOTE: this chops off the heads of the 64-bit O registers.
 400 #ifdef CC_INTERP
 401     save_frame(0);
 402 #else
 403     // make sure G2_thread contains the right value
 404     save_frame_and_mov(0, Lmethod, Lmethod);   // to avoid clobbering O0 (and propagate Lmethod for -Xprof)
 405     mov(G1, L1);                // avoid clobbering G1
 406     // G2 saved below
 407     mov(G3, L3);                // avoid clobbering G3
 408     mov(G4, L4);                // avoid clobbering G4
 409     mov(G5_method, L5);         // avoid clobbering G5_method
 410 #endif /* CC_INTERP */
 411 #if defined(COMPILER2) && !defined(_LP64)
 412     // Save & restore possible 64-bit Long arguments in G-regs
 413     srlx(G1,32,L0);
 414     srlx(G4,32,L6);
 415 #endif
 416     call(CAST_FROM_FN_PTR(address,verify_thread_subroutine), relocInfo::runtime_call_type);
 417     delayed()->mov(G2_thread, O0);
 418 
 419     mov(L1, G1);                // Restore G1
 420     // G2 restored below
 421     mov(L3, G3);                // restore G3
 422     mov(L4, G4);                // restore G4
 423     mov(L5, G5_method);         // restore G5_method
 424 #if defined(COMPILER2) && !defined(_LP64)
 425     // Save & restore possible 64-bit Long arguments in G-regs
 426     sllx(L0,32,G2);             // Move old high G1 bits high in G2
 427     srl(G1, 0,G1);              // Clear current high G1 bits
 428     or3 (G1,G2,G1);             // Recover 64-bit G1
 429     sllx(L6,32,G2);             // Move old high G4 bits high in G2
 430     srl(G4, 0,G4);              // Clear current high G4 bits
 431     or3 (G4,G2,G4);             // Recover 64-bit G4
 432 #endif
 433     restore(O0, 0, G2_thread);
 434   }
 435 }
 436 
 437 
 438 void MacroAssembler::save_thread(const Register thread_cache) {
 439   verify_thread();
 440   if (thread_cache->is_valid()) {
 441     assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile");
 442     mov(G2_thread, thread_cache);
 443   }
 444   if (VerifyThread) {
 445     // smash G2_thread, as if the VM were about to anyway
 446     set(0x67676767, G2_thread);
 447   }
 448 }
 449 
 450 
 451 void MacroAssembler::restore_thread(const Register thread_cache) {
 452   if (thread_cache->is_valid()) {
 453     assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile");
 454     mov(thread_cache, G2_thread);
 455     verify_thread();
 456   } else {
 457     // do it the slow way
 458     get_thread();
 459   }
 460 }
 461 
 462 
 463 // %%% maybe get rid of [re]set_last_Java_frame
 464 void MacroAssembler::set_last_Java_frame(Register last_java_sp, Register last_Java_pc) {
 465   assert_not_delayed();
 466   Address flags(G2_thread, JavaThread::frame_anchor_offset() +
 467                            JavaFrameAnchor::flags_offset());
 468   Address pc_addr(G2_thread, JavaThread::last_Java_pc_offset());
 469 
 470   // Always set last_Java_pc and flags first because once last_Java_sp is visible
 471   // has_last_Java_frame is true and users will look at the rest of the fields.
 472   // (Note: flags should always be zero before we get here so doesn't need to be set.)
 473 
 474 #ifdef ASSERT
 475   // Verify that flags was zeroed on return to Java
 476   Label PcOk;
 477   save_frame(0);                // to avoid clobbering O0
 478   ld_ptr(pc_addr, L0);
 479   br_null_short(L0, Assembler::pt, PcOk);
 480   STOP("last_Java_pc not zeroed before leaving Java");
 481   bind(PcOk);
 482 
 483   // Verify that flags was zeroed on return to Java
 484   Label FlagsOk;
 485   ld(flags, L0);
 486   tst(L0);
 487   br(Assembler::zero, false, Assembler::pt, FlagsOk);
 488   delayed() -> restore();
 489   STOP("flags not zeroed before leaving Java");
 490   bind(FlagsOk);
 491 #endif /* ASSERT */
 492   //
 493   // When returning from calling out from Java mode the frame anchor's last_Java_pc
 494   // will always be set to NULL. It is set here so that if we are doing a call to
 495   // native (not VM) that we capture the known pc and don't have to rely on the
 496   // native call having a standard frame linkage where we can find the pc.
 497 
 498   if (last_Java_pc->is_valid()) {
 499     st_ptr(last_Java_pc, pc_addr);
 500   }
 501 
 502 #ifdef _LP64
 503 #ifdef ASSERT
 504   // Make sure that we have an odd stack
 505   Label StackOk;
 506   andcc(last_java_sp, 0x01, G0);
 507   br(Assembler::notZero, false, Assembler::pt, StackOk);
 508   delayed()->nop();
 509   STOP("Stack Not Biased in set_last_Java_frame");
 510   bind(StackOk);
 511 #endif // ASSERT
 512   assert( last_java_sp != G4_scratch, "bad register usage in set_last_Java_frame");
 513   add( last_java_sp, STACK_BIAS, G4_scratch );
 514   st_ptr(G4_scratch, G2_thread, JavaThread::last_Java_sp_offset());
 515 #else
 516   st_ptr(last_java_sp, G2_thread, JavaThread::last_Java_sp_offset());
 517 #endif // _LP64
 518 }
 519 
 520 void MacroAssembler::reset_last_Java_frame(void) {
 521   assert_not_delayed();
 522 
 523   Address sp_addr(G2_thread, JavaThread::last_Java_sp_offset());
 524   Address pc_addr(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
 525   Address flags  (G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
 526 
 527 #ifdef ASSERT
 528   // check that it WAS previously set
 529 #ifdef CC_INTERP
 530     save_frame(0);
 531 #else
 532     save_frame_and_mov(0, Lmethod, Lmethod);     // Propagate Lmethod to helper frame for -Xprof
 533 #endif /* CC_INTERP */
 534     ld_ptr(sp_addr, L0);
 535     tst(L0);
 536     breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
 537     restore();
 538 #endif // ASSERT
 539 
 540   st_ptr(G0, sp_addr);
 541   // Always return last_Java_pc to zero
 542   st_ptr(G0, pc_addr);
 543   // Always null flags after return to Java
 544   st(G0, flags);
 545 }
 546 
 547 
 548 void MacroAssembler::call_VM_base(
 549   Register        oop_result,
 550   Register        thread_cache,
 551   Register        last_java_sp,
 552   address         entry_point,
 553   int             number_of_arguments,
 554   bool            check_exceptions)
 555 {
 556   assert_not_delayed();
 557 
 558   // determine last_java_sp register
 559   if (!last_java_sp->is_valid()) {
 560     last_java_sp = SP;
 561   }
 562   // debugging support
 563   assert(number_of_arguments >= 0   , "cannot have negative number of arguments");
 564 
 565   // 64-bit last_java_sp is biased!
 566   set_last_Java_frame(last_java_sp, noreg);
 567   if (VerifyThread)  mov(G2_thread, O0); // about to be smashed; pass early
 568   save_thread(thread_cache);
 569   // do the call
 570   call(entry_point, relocInfo::runtime_call_type);
 571   if (!VerifyThread)
 572     delayed()->mov(G2_thread, O0);  // pass thread as first argument
 573   else
 574     delayed()->nop();             // (thread already passed)
 575   restore_thread(thread_cache);
 576   reset_last_Java_frame();
 577 
 578   // check for pending exceptions. use Gtemp as scratch register.
 579   if (check_exceptions) {
 580     check_and_forward_exception(Gtemp);
 581   }
 582 
 583 #ifdef ASSERT
 584   set(badHeapWordVal, G3);
 585   set(badHeapWordVal, G4);
 586   set(badHeapWordVal, G5);
 587 #endif
 588 
 589   // get oop result if there is one and reset the value in the thread
 590   if (oop_result->is_valid()) {
 591     get_vm_result(oop_result);
 592   }
 593 }
 594 
 595 void MacroAssembler::check_and_forward_exception(Register scratch_reg)
 596 {
 597   Label L;
 598 
 599   check_and_handle_popframe(scratch_reg);
 600   check_and_handle_earlyret(scratch_reg);
 601 
 602   Address exception_addr(G2_thread, Thread::pending_exception_offset());
 603   ld_ptr(exception_addr, scratch_reg);
 604   br_null_short(scratch_reg, pt, L);
 605   // we use O7 linkage so that forward_exception_entry has the issuing PC
 606   call(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
 607   delayed()->nop();
 608   bind(L);
 609 }
 610 
 611 
 612 void MacroAssembler::check_and_handle_popframe(Register scratch_reg) {
 613 }
 614 
 615 
 616 void MacroAssembler::check_and_handle_earlyret(Register scratch_reg) {
 617 }
 618 
 619 
 620 void MacroAssembler::call_VM(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
 621   call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions);
 622 }
 623 
 624 
 625 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, bool check_exceptions) {
 626   // O0 is reserved for the thread
 627   mov(arg_1, O1);
 628   call_VM(oop_result, entry_point, 1, check_exceptions);
 629 }
 630 
 631 
 632 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) {
 633   // O0 is reserved for the thread
 634   mov(arg_1, O1);
 635   mov(arg_2, O2); assert(arg_2 != O1, "smashed argument");
 636   call_VM(oop_result, entry_point, 2, check_exceptions);
 637 }
 638 
 639 
 640 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, Register arg_3, bool check_exceptions) {
 641   // O0 is reserved for the thread
 642   mov(arg_1, O1);
 643   mov(arg_2, O2); assert(arg_2 != O1,                "smashed argument");
 644   mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument");
 645   call_VM(oop_result, entry_point, 3, check_exceptions);
 646 }
 647 
 648 
 649 
 650 // Note: The following call_VM overloadings are useful when a "save"
 651 // has already been performed by a stub, and the last Java frame is
 652 // the previous one.  In that case, last_java_sp must be passed as FP
 653 // instead of SP.
 654 
 655 
 656 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, int number_of_arguments, bool check_exceptions) {
 657   call_VM_base(oop_result, noreg, last_java_sp, entry_point, number_of_arguments, check_exceptions);
 658 }
 659 
 660 
 661 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, bool check_exceptions) {
 662   // O0 is reserved for the thread
 663   mov(arg_1, O1);
 664   call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
 665 }
 666 
 667 
 668 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) {
 669   // O0 is reserved for the thread
 670   mov(arg_1, O1);
 671   mov(arg_2, O2); assert(arg_2 != O1, "smashed argument");
 672   call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
 673 }
 674 
 675 
 676 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) {
 677   // O0 is reserved for the thread
 678   mov(arg_1, O1);
 679   mov(arg_2, O2); assert(arg_2 != O1,                "smashed argument");
 680   mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument");
 681   call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
 682 }
 683 
 684 
 685 
 686 void MacroAssembler::call_VM_leaf_base(Register thread_cache, address entry_point, int number_of_arguments) {
 687   assert_not_delayed();
 688   save_thread(thread_cache);
 689   // do the call
 690   call(entry_point, relocInfo::runtime_call_type);
 691   delayed()->nop();
 692   restore_thread(thread_cache);
 693 #ifdef ASSERT
 694   set(badHeapWordVal, G3);
 695   set(badHeapWordVal, G4);
 696   set(badHeapWordVal, G5);
 697 #endif
 698 }
 699 
 700 
 701 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, int number_of_arguments) {
 702   call_VM_leaf_base(thread_cache, entry_point, number_of_arguments);
 703 }
 704 
 705 
 706 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1) {
 707   mov(arg_1, O0);
 708   call_VM_leaf(thread_cache, entry_point, 1);
 709 }
 710 
 711 
 712 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2) {
 713   mov(arg_1, O0);
 714   mov(arg_2, O1); assert(arg_2 != O0, "smashed argument");
 715   call_VM_leaf(thread_cache, entry_point, 2);
 716 }
 717 
 718 
 719 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2, Register arg_3) {
 720   mov(arg_1, O0);
 721   mov(arg_2, O1); assert(arg_2 != O0,                "smashed argument");
 722   mov(arg_3, O2); assert(arg_3 != O0 && arg_3 != O1, "smashed argument");
 723   call_VM_leaf(thread_cache, entry_point, 3);
 724 }
 725 
 726 
 727 void MacroAssembler::get_vm_result(Register oop_result) {
 728   verify_thread();
 729   Address vm_result_addr(G2_thread, JavaThread::vm_result_offset());
 730   ld_ptr(    vm_result_addr, oop_result);
 731   st_ptr(G0, vm_result_addr);
 732   verify_oop(oop_result);
 733 }
 734 
 735 
 736 void MacroAssembler::get_vm_result_2(Register metadata_result) {
 737   verify_thread();
 738   Address vm_result_addr_2(G2_thread, JavaThread::vm_result_2_offset());
 739   ld_ptr(vm_result_addr_2, metadata_result);
 740   st_ptr(G0, vm_result_addr_2);
 741 }
 742 
 743 
 744 // We require that C code which does not return a value in vm_result will
 745 // leave it undisturbed.
 746 void MacroAssembler::set_vm_result(Register oop_result) {
 747   verify_thread();
 748   Address vm_result_addr(G2_thread, JavaThread::vm_result_offset());
 749   verify_oop(oop_result);
 750 
 751 # ifdef ASSERT
 752     // Check that we are not overwriting any other oop.
 753 #ifdef CC_INTERP
 754     save_frame(0);
 755 #else
 756     save_frame_and_mov(0, Lmethod, Lmethod);     // Propagate Lmethod for -Xprof
 757 #endif /* CC_INTERP */
 758     ld_ptr(vm_result_addr, L0);
 759     tst(L0);
 760     restore();
 761     breakpoint_trap(notZero, Assembler::ptr_cc);
 762     // }
 763 # endif
 764 
 765   st_ptr(oop_result, vm_result_addr);
 766 }
 767 
 768 
 769 void MacroAssembler::ic_call(address entry, bool emit_delay) {
 770   RelocationHolder rspec = virtual_call_Relocation::spec(pc());
 771   patchable_set((intptr_t)Universe::non_oop_word(), G5_inline_cache_reg);
 772   relocate(rspec);
 773   call(entry, relocInfo::none);
 774   if (emit_delay) {
 775     delayed()->nop();
 776   }
 777 }
 778 
 779 
 780 void MacroAssembler::card_table_write(jbyte* byte_map_base,
 781                                       Register tmp, Register obj) {
 782 #ifdef _LP64
 783   srlx(obj, CardTableModRefBS::card_shift, obj);
 784 #else
 785   srl(obj, CardTableModRefBS::card_shift, obj);
 786 #endif
 787   assert(tmp != obj, "need separate temp reg");
 788   set((address) byte_map_base, tmp);
 789   stb(G0, tmp, obj);
 790 }
 791 
 792 
 793 void MacroAssembler::internal_sethi(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) {
 794   address save_pc;
 795   int shiftcnt;
 796 #ifdef _LP64
 797 # ifdef CHECK_DELAY
 798   assert_not_delayed((char*) "cannot put two instructions in delay slot");
 799 # endif
 800   v9_dep();
 801   save_pc = pc();
 802 
 803   int msb32 = (int) (addrlit.value() >> 32);
 804   int lsb32 = (int) (addrlit.value());
 805 
 806   if (msb32 == 0 && lsb32 >= 0) {
 807     Assembler::sethi(lsb32, d, addrlit.rspec());
 808   }
 809   else if (msb32 == -1) {
 810     Assembler::sethi(~lsb32, d, addrlit.rspec());
 811     xor3(d, ~low10(~0), d);
 812   }
 813   else {
 814     Assembler::sethi(msb32, d, addrlit.rspec());  // msb 22-bits
 815     if (msb32 & 0x3ff)                            // Any bits?
 816       or3(d, msb32 & 0x3ff, d);                   // msb 32-bits are now in lsb 32
 817     if (lsb32 & 0xFFFFFC00) {                     // done?
 818       if ((lsb32 >> 20) & 0xfff) {                // Any bits set?
 819         sllx(d, 12, d);                           // Make room for next 12 bits
 820         or3(d, (lsb32 >> 20) & 0xfff, d);         // Or in next 12
 821         shiftcnt = 0;                             // We already shifted
 822       }
 823       else
 824         shiftcnt = 12;
 825       if ((lsb32 >> 10) & 0x3ff) {
 826         sllx(d, shiftcnt + 10, d);                // Make room for last 10 bits
 827         or3(d, (lsb32 >> 10) & 0x3ff, d);         // Or in next 10
 828         shiftcnt = 0;
 829       }
 830       else
 831         shiftcnt = 10;
 832       sllx(d, shiftcnt + 10, d);                  // Shift leaving disp field 0'd
 833     }
 834     else
 835       sllx(d, 32, d);
 836   }
 837   // Pad out the instruction sequence so it can be patched later.
 838   if (ForceRelocatable || (addrlit.rtype() != relocInfo::none &&
 839                            addrlit.rtype() != relocInfo::runtime_call_type)) {
 840     while (pc() < (save_pc + (7 * BytesPerInstWord)))
 841       nop();
 842   }
 843 #else
 844   Assembler::sethi(addrlit.value(), d, addrlit.rspec());
 845 #endif
 846 }
 847 
 848 
 849 void MacroAssembler::sethi(const AddressLiteral& addrlit, Register d) {
 850   internal_sethi(addrlit, d, false);
 851 }
 852 
 853 
 854 void MacroAssembler::patchable_sethi(const AddressLiteral& addrlit, Register d) {
 855   internal_sethi(addrlit, d, true);
 856 }
 857 
 858 
 859 int MacroAssembler::insts_for_sethi(address a, bool worst_case) {
 860 #ifdef _LP64
 861   if (worst_case)  return 7;
 862   intptr_t iaddr = (intptr_t) a;
 863   int msb32 = (int) (iaddr >> 32);
 864   int lsb32 = (int) (iaddr);
 865   int count;
 866   if (msb32 == 0 && lsb32 >= 0)
 867     count = 1;
 868   else if (msb32 == -1)
 869     count = 2;
 870   else {
 871     count = 2;
 872     if (msb32 & 0x3ff)
 873       count++;
 874     if (lsb32 & 0xFFFFFC00 ) {
 875       if ((lsb32 >> 20) & 0xfff)  count += 2;
 876       if ((lsb32 >> 10) & 0x3ff)  count += 2;
 877     }
 878   }
 879   return count;
 880 #else
 881   return 1;
 882 #endif
 883 }
 884 
 885 int MacroAssembler::worst_case_insts_for_set() {
 886   return insts_for_sethi(NULL, true) + 1;
 887 }
 888 
 889 
 890 // Keep in sync with MacroAssembler::insts_for_internal_set
 891 void MacroAssembler::internal_set(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) {
 892   intptr_t value = addrlit.value();
 893 
 894   if (!ForceRelocatable && addrlit.rspec().type() == relocInfo::none) {
 895     // can optimize
 896     if (-4096 <= value && value <= 4095) {
 897       or3(G0, value, d); // setsw (this leaves upper 32 bits sign-extended)
 898       return;
 899     }
 900     if (inv_hi22(hi22(value)) == value) {
 901       sethi(addrlit, d);
 902       return;
 903     }
 904   }
 905   assert_not_delayed((char*) "cannot put two instructions in delay slot");
 906   internal_sethi(addrlit, d, ForceRelocatable);
 907   if (ForceRelocatable || addrlit.rspec().type() != relocInfo::none || addrlit.low10() != 0) {
 908     add(d, addrlit.low10(), d, addrlit.rspec());
 909   }
 910 }
 911 
 912 // Keep in sync with MacroAssembler::internal_set
 913 int MacroAssembler::insts_for_internal_set(intptr_t value) {
 914   // can optimize
 915   if (-4096 <= value && value <= 4095) {
 916     return 1;
 917   }
 918   if (inv_hi22(hi22(value)) == value) {
 919     return insts_for_sethi((address) value);
 920   }
 921   int count = insts_for_sethi((address) value);
 922   AddressLiteral al(value);
 923   if (al.low10() != 0) {
 924     count++;
 925   }
 926   return count;
 927 }
 928 
 929 void MacroAssembler::set(const AddressLiteral& al, Register d) {
 930   internal_set(al, d, false);
 931 }
 932 
 933 void MacroAssembler::set(intptr_t value, Register d) {
 934   AddressLiteral al(value);
 935   internal_set(al, d, false);
 936 }
 937 
 938 void MacroAssembler::set(address addr, Register d, RelocationHolder const& rspec) {
 939   AddressLiteral al(addr, rspec);
 940   internal_set(al, d, false);
 941 }
 942 
 943 void MacroAssembler::patchable_set(const AddressLiteral& al, Register d) {
 944   internal_set(al, d, true);
 945 }
 946 
 947 void MacroAssembler::patchable_set(intptr_t value, Register d) {
 948   AddressLiteral al(value);
 949   internal_set(al, d, true);
 950 }
 951 
 952 
 953 void MacroAssembler::set64(jlong value, Register d, Register tmp) {
 954   assert_not_delayed();
 955   v9_dep();
 956 
 957   int hi = (int)(value >> 32);
 958   int lo = (int)(value & ~0);
 959   // (Matcher::isSimpleConstant64 knows about the following optimizations.)
 960   if (Assembler::is_simm13(lo) && value == lo) {
 961     or3(G0, lo, d);
 962   } else if (hi == 0) {
 963     Assembler::sethi(lo, d);   // hardware version zero-extends to upper 32
 964     if (low10(lo) != 0)
 965       or3(d, low10(lo), d);
 966   }
 967   else if (hi == -1) {
 968     Assembler::sethi(~lo, d);  // hardware version zero-extends to upper 32
 969     xor3(d, low10(lo) ^ ~low10(~0), d);
 970   }
 971   else if (lo == 0) {
 972     if (Assembler::is_simm13(hi)) {
 973       or3(G0, hi, d);
 974     } else {
 975       Assembler::sethi(hi, d);   // hardware version zero-extends to upper 32
 976       if (low10(hi) != 0)
 977         or3(d, low10(hi), d);
 978     }
 979     sllx(d, 32, d);
 980   }
 981   else {
 982     Assembler::sethi(hi, tmp);
 983     Assembler::sethi(lo,   d); // macro assembler version sign-extends
 984     if (low10(hi) != 0)
 985       or3 (tmp, low10(hi), tmp);
 986     if (low10(lo) != 0)
 987       or3 (  d, low10(lo),   d);
 988     sllx(tmp, 32, tmp);
 989     or3 (d, tmp, d);
 990   }
 991 }
 992 
 993 int MacroAssembler::insts_for_set64(jlong value) {
 994   v9_dep();
 995 
 996   int hi = (int) (value >> 32);
 997   int lo = (int) (value & ~0);
 998   int count = 0;
 999 
1000   // (Matcher::isSimpleConstant64 knows about the following optimizations.)
1001   if (Assembler::is_simm13(lo) && value == lo) {
1002     count++;
1003   } else if (hi == 0) {
1004     count++;
1005     if (low10(lo) != 0)
1006       count++;
1007   }
1008   else if (hi == -1) {
1009     count += 2;
1010   }
1011   else if (lo == 0) {
1012     if (Assembler::is_simm13(hi)) {
1013       count++;
1014     } else {
1015       count++;
1016       if (low10(hi) != 0)
1017         count++;
1018     }
1019     count++;
1020   }
1021   else {
1022     count += 2;
1023     if (low10(hi) != 0)
1024       count++;
1025     if (low10(lo) != 0)
1026       count++;
1027     count += 2;
1028   }
1029   return count;
1030 }
1031 
1032 // compute size in bytes of sparc frame, given
1033 // number of extraWords
1034 int MacroAssembler::total_frame_size_in_bytes(int extraWords) {
1035 
1036   int nWords = frame::memory_parameter_word_sp_offset;
1037 
1038   nWords += extraWords;
1039 
1040   if (nWords & 1) ++nWords; // round up to double-word
1041 
1042   return nWords * BytesPerWord;
1043 }
1044 
1045 
1046 // save_frame: given number of "extra" words in frame,
1047 // issue approp. save instruction (p 200, v8 manual)
1048 
1049 void MacroAssembler::save_frame(int extraWords) {
1050   int delta = -total_frame_size_in_bytes(extraWords);
1051   if (is_simm13(delta)) {
1052     save(SP, delta, SP);
1053   } else {
1054     set(delta, G3_scratch);
1055     save(SP, G3_scratch, SP);
1056   }
1057 }
1058 
1059 
1060 void MacroAssembler::save_frame_c1(int size_in_bytes) {
1061   if (is_simm13(-size_in_bytes)) {
1062     save(SP, -size_in_bytes, SP);
1063   } else {
1064     set(-size_in_bytes, G3_scratch);
1065     save(SP, G3_scratch, SP);
1066   }
1067 }
1068 
1069 
1070 void MacroAssembler::save_frame_and_mov(int extraWords,
1071                                         Register s1, Register d1,
1072                                         Register s2, Register d2) {
1073   assert_not_delayed();
1074 
1075   // The trick here is to use precisely the same memory word
1076   // that trap handlers also use to save the register.
1077   // This word cannot be used for any other purpose, but
1078   // it works fine to save the register's value, whether or not
1079   // an interrupt flushes register windows at any given moment!
1080   Address s1_addr;
1081   if (s1->is_valid() && (s1->is_in() || s1->is_local())) {
1082     s1_addr = s1->address_in_saved_window();
1083     st_ptr(s1, s1_addr);
1084   }
1085 
1086   Address s2_addr;
1087   if (s2->is_valid() && (s2->is_in() || s2->is_local())) {
1088     s2_addr = s2->address_in_saved_window();
1089     st_ptr(s2, s2_addr);
1090   }
1091 
1092   save_frame(extraWords);
1093 
1094   if (s1_addr.base() == SP) {
1095     ld_ptr(s1_addr.after_save(), d1);
1096   } else if (s1->is_valid()) {
1097     mov(s1->after_save(), d1);
1098   }
1099 
1100   if (s2_addr.base() == SP) {
1101     ld_ptr(s2_addr.after_save(), d2);
1102   } else if (s2->is_valid()) {
1103     mov(s2->after_save(), d2);
1104   }
1105 }
1106 
1107 
1108 AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) {
1109   assert(oop_recorder() != NULL, "this assembler needs a Recorder");
1110   int index = oop_recorder()->allocate_metadata_index(obj);
1111   RelocationHolder rspec = metadata_Relocation::spec(index);
1112   return AddressLiteral((address)obj, rspec);
1113 }
1114 
1115 AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) {
1116   assert(oop_recorder() != NULL, "this assembler needs a Recorder");
1117   int index = oop_recorder()->find_index(obj);
1118   RelocationHolder rspec = metadata_Relocation::spec(index);
1119   return AddressLiteral((address)obj, rspec);
1120 }
1121 
1122 
1123 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) {
1124   assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1125   assert(Universe::heap()->is_in_reserved(JNIHandles::resolve(obj)), "not an oop");
1126   int oop_index = oop_recorder()->find_index(obj);
1127   return AddressLiteral(obj, oop_Relocation::spec(oop_index));
1128 }
1129 
1130 void  MacroAssembler::set_narrow_oop(jobject obj, Register d) {
1131   assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1132   int oop_index = oop_recorder()->find_index(obj);
1133   RelocationHolder rspec = oop_Relocation::spec(oop_index);
1134 
1135   assert_not_delayed();
1136   // Relocation with special format (see relocInfo_sparc.hpp).
1137   relocate(rspec, 1);
1138   // Assembler::sethi(0x3fffff, d);
1139   emit_int32( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(0x3fffff) );
1140   // Don't add relocation for 'add'. Do patching during 'sethi' processing.
1141   add(d, 0x3ff, d);
1142 
1143 }
1144 
1145 void  MacroAssembler::set_narrow_klass(Klass* k, Register d) {
1146   assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1147   int klass_index = oop_recorder()->find_index(k);
1148   RelocationHolder rspec = metadata_Relocation::spec(klass_index);
1149   narrowOop encoded_k = Klass::encode_klass(k);
1150 
1151   assert_not_delayed();
1152   // Relocation with special format (see relocInfo_sparc.hpp).
1153   relocate(rspec, 1);
1154   // Assembler::sethi(encoded_k, d);
1155   emit_int32( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(encoded_k) );
1156   // Don't add relocation for 'add'. Do patching during 'sethi' processing.
1157   add(d, low10(encoded_k), d);
1158 
1159 }
1160 
1161 void MacroAssembler::align(int modulus) {
1162   while (offset() % modulus != 0) nop();
1163 }
1164 
1165 void RegistersForDebugging::print(outputStream* s) {
1166   FlagSetting fs(Debugging, true);
1167   int j;
1168   for (j = 0; j < 8; ++j) {
1169     if (j != 6) { s->print("i%d = ", j); os::print_location(s, i[j]); }
1170     else        { s->print( "fp = "   ); os::print_location(s, i[j]); }
1171   }
1172   s->cr();
1173 
1174   for (j = 0;  j < 8;  ++j) {
1175     s->print("l%d = ", j); os::print_location(s, l[j]);
1176   }
1177   s->cr();
1178 
1179   for (j = 0; j < 8; ++j) {
1180     if (j != 6) { s->print("o%d = ", j); os::print_location(s, o[j]); }
1181     else        { s->print( "sp = "   ); os::print_location(s, o[j]); }
1182   }
1183   s->cr();
1184 
1185   for (j = 0; j < 8; ++j) {
1186     s->print("g%d = ", j); os::print_location(s, g[j]);
1187   }
1188   s->cr();
1189 
1190   // print out floats with compression
1191   for (j = 0; j < 32; ) {
1192     jfloat val = f[j];
1193     int last = j;
1194     for ( ;  last+1 < 32;  ++last ) {
1195       char b1[1024], b2[1024];
1196       sprintf(b1, "%f", val);
1197       sprintf(b2, "%f", f[last+1]);
1198       if (strcmp(b1, b2))
1199         break;
1200     }
1201     s->print("f%d", j);
1202     if ( j != last )  s->print(" - f%d", last);
1203     s->print(" = %f", val);
1204     s->fill_to(25);
1205     s->print_cr(" (0x%x)", val);
1206     j = last + 1;
1207   }
1208   s->cr();
1209 
1210   // and doubles (evens only)
1211   for (j = 0; j < 32; ) {
1212     jdouble val = d[j];
1213     int last = j;
1214     for ( ;  last+1 < 32;  ++last ) {
1215       char b1[1024], b2[1024];
1216       sprintf(b1, "%f", val);
1217       sprintf(b2, "%f", d[last+1]);
1218       if (strcmp(b1, b2))
1219         break;
1220     }
1221     s->print("d%d", 2 * j);
1222     if ( j != last )  s->print(" - d%d", last);
1223     s->print(" = %f", val);
1224     s->fill_to(30);
1225     s->print("(0x%x)", *(int*)&val);
1226     s->fill_to(42);
1227     s->print_cr("(0x%x)", *(1 + (int*)&val));
1228     j = last + 1;
1229   }
1230   s->cr();
1231 }
1232 
1233 void RegistersForDebugging::save_registers(MacroAssembler* a) {
1234   a->sub(FP, round_to(sizeof(RegistersForDebugging), sizeof(jdouble)) - STACK_BIAS, O0);
1235   a->flushw();
1236   int i;
1237   for (i = 0; i < 8; ++i) {
1238     a->ld_ptr(as_iRegister(i)->address_in_saved_window().after_save(), L1);  a->st_ptr( L1, O0, i_offset(i));
1239     a->ld_ptr(as_lRegister(i)->address_in_saved_window().after_save(), L1);  a->st_ptr( L1, O0, l_offset(i));
1240     a->st_ptr(as_oRegister(i)->after_save(), O0, o_offset(i));
1241     a->st_ptr(as_gRegister(i)->after_save(), O0, g_offset(i));
1242   }
1243   for (i = 0;  i < 32; ++i) {
1244     a->stf(FloatRegisterImpl::S, as_FloatRegister(i), O0, f_offset(i));
1245   }
1246   for (i = 0; i < 64; i += 2) {
1247     a->stf(FloatRegisterImpl::D, as_FloatRegister(i), O0, d_offset(i));
1248   }
1249 }
1250 
1251 void RegistersForDebugging::restore_registers(MacroAssembler* a, Register r) {
1252   for (int i = 1; i < 8;  ++i) {
1253     a->ld_ptr(r, g_offset(i), as_gRegister(i));
1254   }
1255   for (int j = 0; j < 32; ++j) {
1256     a->ldf(FloatRegisterImpl::S, O0, f_offset(j), as_FloatRegister(j));
1257   }
1258   for (int k = 0; k < 64; k += 2) {
1259     a->ldf(FloatRegisterImpl::D, O0, d_offset(k), as_FloatRegister(k));
1260   }
1261 }
1262 
1263 
1264 // pushes double TOS element of FPU stack on CPU stack; pops from FPU stack
1265 void MacroAssembler::push_fTOS() {
1266   // %%%%%% need to implement this
1267 }
1268 
1269 // pops double TOS element from CPU stack and pushes on FPU stack
1270 void MacroAssembler::pop_fTOS() {
1271   // %%%%%% need to implement this
1272 }
1273 
1274 void MacroAssembler::empty_FPU_stack() {
1275   // %%%%%% need to implement this
1276 }
1277 
1278 void MacroAssembler::_verify_oop(Register reg, const char* msg, const char * file, int line) {
1279   // plausibility check for oops
1280   if (!VerifyOops) return;
1281 
1282   if (reg == G0)  return;       // always NULL, which is always an oop
1283 
1284   BLOCK_COMMENT("verify_oop {");
1285   char buffer[64];
1286 #ifdef COMPILER1
1287   if (CommentedAssembly) {
1288     snprintf(buffer, sizeof(buffer), "verify_oop at %d", offset());
1289     block_comment(buffer);
1290   }
1291 #endif
1292 
1293   const char* real_msg = NULL;
1294   {
1295     ResourceMark rm;
1296     stringStream ss;
1297     ss.print("%s at offset %d (%s:%d)", msg, offset(), file, line);
1298     real_msg = code_string(ss.as_string());
1299   }
1300 
1301   // Call indirectly to solve generation ordering problem
1302   AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address());
1303 
1304   // Make some space on stack above the current register window.
1305   // Enough to hold 8 64-bit registers.
1306   add(SP,-8*8,SP);
1307 
1308   // Save some 64-bit registers; a normal 'save' chops the heads off
1309   // of 64-bit longs in the 32-bit build.
1310   stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8);
1311   stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8);
1312   mov(reg,O0); // Move arg into O0; arg might be in O7 which is about to be crushed
1313   stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8);
1314 
1315   // Size of set() should stay the same
1316   patchable_set((intptr_t)real_msg, O1);
1317   // Load address to call to into O7
1318   load_ptr_contents(a, O7);
1319   // Register call to verify_oop_subroutine
1320   callr(O7, G0);
1321   delayed()->nop();
1322   // recover frame size
1323   add(SP, 8*8,SP);
1324   BLOCK_COMMENT("} verify_oop");
1325 }
1326 
1327 void MacroAssembler::_verify_oop_addr(Address addr, const char* msg, const char * file, int line) {
1328   // plausibility check for oops
1329   if (!VerifyOops) return;
1330 
1331   const char* real_msg = NULL;
1332   {
1333     ResourceMark rm;
1334     stringStream ss;
1335     ss.print("%s at SP+%d (%s:%d)", msg, addr.disp(), file, line);
1336     real_msg = code_string(ss.as_string());
1337   }
1338 
1339   // Call indirectly to solve generation ordering problem
1340   AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address());
1341 
1342   // Make some space on stack above the current register window.
1343   // Enough to hold 8 64-bit registers.
1344   add(SP,-8*8,SP);
1345 
1346   // Save some 64-bit registers; a normal 'save' chops the heads off
1347   // of 64-bit longs in the 32-bit build.
1348   stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8);
1349   stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8);
1350   ld_ptr(addr.base(), addr.disp() + 8*8, O0); // Load arg into O0; arg might be in O7 which is about to be crushed
1351   stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8);
1352 
1353   // Size of set() should stay the same
1354   patchable_set((intptr_t)real_msg, O1);
1355   // Load address to call to into O7
1356   load_ptr_contents(a, O7);
1357   // Register call to verify_oop_subroutine
1358   callr(O7, G0);
1359   delayed()->nop();
1360   // recover frame size
1361   add(SP, 8*8,SP);
1362 }
1363 
1364 // side-door communication with signalHandler in os_solaris.cpp
1365 address MacroAssembler::_verify_oop_implicit_branch[3] = { NULL };
1366 
1367 // This macro is expanded just once; it creates shared code.  Contract:
1368 // receives an oop in O0.  Must restore O0 & O7 from TLS.  Must not smash ANY
1369 // registers, including flags.  May not use a register 'save', as this blows
1370 // the high bits of the O-regs if they contain Long values.  Acts as a 'leaf'
1371 // call.
1372 void MacroAssembler::verify_oop_subroutine() {
1373   // Leaf call; no frame.
1374   Label succeed, fail, null_or_fail;
1375 
1376   // O0 and O7 were saved already (O0 in O0's TLS home, O7 in O5's TLS home).
1377   // O0 is now the oop to be checked.  O7 is the return address.
1378   Register O0_obj = O0;
1379 
1380   // Save some more registers for temps.
1381   stx(O2,SP,frame::register_save_words*wordSize+STACK_BIAS+2*8);
1382   stx(O3,SP,frame::register_save_words*wordSize+STACK_BIAS+3*8);
1383   stx(O4,SP,frame::register_save_words*wordSize+STACK_BIAS+4*8);
1384   stx(O5,SP,frame::register_save_words*wordSize+STACK_BIAS+5*8);
1385 
1386   // Save flags
1387   Register O5_save_flags = O5;
1388   rdccr( O5_save_flags );
1389 
1390   { // count number of verifies
1391     Register O2_adr   = O2;
1392     Register O3_accum = O3;
1393     inc_counter(StubRoutines::verify_oop_count_addr(), O2_adr, O3_accum);
1394   }
1395 
1396   Register O2_mask = O2;
1397   Register O3_bits = O3;
1398   Register O4_temp = O4;
1399 
1400   // mark lower end of faulting range
1401   assert(_verify_oop_implicit_branch[0] == NULL, "set once");
1402   _verify_oop_implicit_branch[0] = pc();
1403 
1404   // We can't check the mark oop because it could be in the process of
1405   // locking or unlocking while this is running.
1406   set(Universe::verify_oop_mask (), O2_mask);
1407   set(Universe::verify_oop_bits (), O3_bits);
1408 
1409   // assert((obj & oop_mask) == oop_bits);
1410   and3(O0_obj, O2_mask, O4_temp);
1411   cmp_and_brx_short(O4_temp, O3_bits, notEqual, pn, null_or_fail);
1412 
1413   if ((NULL_WORD & Universe::verify_oop_mask()) == Universe::verify_oop_bits()) {
1414     // the null_or_fail case is useless; must test for null separately
1415     br_null_short(O0_obj, pn, succeed);
1416   }
1417 
1418   // Check the Klass* of this object for being in the right area of memory.
1419   // Cannot do the load in the delay above slot in case O0 is null
1420   load_klass(O0_obj, O0_obj);
1421   // assert((klass != NULL)
1422   br_null_short(O0_obj, pn, fail);
1423 
1424   wrccr( O5_save_flags ); // Restore CCR's
1425 
1426   // mark upper end of faulting range
1427   _verify_oop_implicit_branch[1] = pc();
1428 
1429   //-----------------------
1430   // all tests pass
1431   bind(succeed);
1432 
1433   // Restore prior 64-bit registers
1434   ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+0*8,O0);
1435   ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+1*8,O1);
1436   ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+2*8,O2);
1437   ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+3*8,O3);
1438   ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+4*8,O4);
1439   ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+5*8,O5);
1440 
1441   retl();                       // Leaf return; restore prior O7 in delay slot
1442   delayed()->ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+7*8,O7);
1443 
1444   //-----------------------
1445   bind(null_or_fail);           // nulls are less common but OK
1446   br_null(O0_obj, false, pt, succeed);
1447   delayed()->wrccr( O5_save_flags ); // Restore CCR's
1448 
1449   //-----------------------
1450   // report failure:
1451   bind(fail);
1452   _verify_oop_implicit_branch[2] = pc();
1453 
1454   wrccr( O5_save_flags ); // Restore CCR's
1455 
1456   save_frame(::round_to(sizeof(RegistersForDebugging) / BytesPerWord, 2));
1457 
1458   // stop_subroutine expects message pointer in I1.
1459   mov(I1, O1);
1460 
1461   // Restore prior 64-bit registers
1462   ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+0*8,I0);
1463   ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+1*8,I1);
1464   ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+2*8,I2);
1465   ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+3*8,I3);
1466   ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+4*8,I4);
1467   ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+5*8,I5);
1468 
1469   // factor long stop-sequence into subroutine to save space
1470   assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet");
1471 
1472   // call indirectly to solve generation ordering problem
1473   AddressLiteral al(StubRoutines::Sparc::stop_subroutine_entry_address());
1474   load_ptr_contents(al, O5);
1475   jmpl(O5, 0, O7);
1476   delayed()->nop();
1477 }
1478 
1479 
1480 void MacroAssembler::stop(const char* msg) {
1481   // save frame first to get O7 for return address
1482   // add one word to size in case struct is odd number of words long
1483   // It must be doubleword-aligned for storing doubles into it.
1484 
1485     save_frame(::round_to(sizeof(RegistersForDebugging) / BytesPerWord, 2));
1486 
1487     // stop_subroutine expects message pointer in I1.
1488     // Size of set() should stay the same
1489     patchable_set((intptr_t)msg, O1);
1490 
1491     // factor long stop-sequence into subroutine to save space
1492     assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet");
1493 
1494     // call indirectly to solve generation ordering problem
1495     AddressLiteral a(StubRoutines::Sparc::stop_subroutine_entry_address());
1496     load_ptr_contents(a, O5);
1497     jmpl(O5, 0, O7);
1498     delayed()->nop();
1499 
1500     breakpoint_trap();   // make stop actually stop rather than writing
1501                          // unnoticeable results in the output files.
1502 
1503     // restore(); done in callee to save space!
1504 }
1505 
1506 
1507 void MacroAssembler::warn(const char* msg) {
1508   save_frame(::round_to(sizeof(RegistersForDebugging) / BytesPerWord, 2));
1509   RegistersForDebugging::save_registers(this);
1510   mov(O0, L0);
1511   // Size of set() should stay the same
1512   patchable_set((intptr_t)msg, O0);
1513   call( CAST_FROM_FN_PTR(address, warning) );
1514   delayed()->nop();
1515 //  ret();
1516 //  delayed()->restore();
1517   RegistersForDebugging::restore_registers(this, L0);
1518   restore();
1519 }
1520 
1521 
1522 void MacroAssembler::untested(const char* what) {
1523   // We must be able to turn interactive prompting off
1524   // in order to run automated test scripts on the VM
1525   // Use the flag ShowMessageBoxOnError
1526 
1527   const char* b = NULL;
1528   {
1529     ResourceMark rm;
1530     stringStream ss;
1531     ss.print("untested: %s", what);
1532     b = code_string(ss.as_string());
1533   }
1534   if (ShowMessageBoxOnError) { STOP(b); }
1535   else                       { warn(b); }
1536 }
1537 
1538 
1539 void MacroAssembler::stop_subroutine() {
1540   RegistersForDebugging::save_registers(this);
1541 
1542   // for the sake of the debugger, stick a PC on the current frame
1543   // (this assumes that the caller has performed an extra "save")
1544   mov(I7, L7);
1545   add(O7, -7 * BytesPerInt, I7);
1546 
1547   save_frame(); // one more save to free up another O7 register
1548   mov(I0, O1); // addr of reg save area
1549 
1550   // We expect pointer to message in I1. Caller must set it up in O1
1551   mov(I1, O0); // get msg
1552   call (CAST_FROM_FN_PTR(address, MacroAssembler::debug), relocInfo::runtime_call_type);
1553   delayed()->nop();
1554 
1555   restore();
1556 
1557   RegistersForDebugging::restore_registers(this, O0);
1558 
1559   save_frame(0);
1560   call(CAST_FROM_FN_PTR(address,breakpoint));
1561   delayed()->nop();
1562   restore();
1563 
1564   mov(L7, I7);
1565   retl();
1566   delayed()->restore(); // see stop above
1567 }
1568 
1569 
1570 void MacroAssembler::debug(char* msg, RegistersForDebugging* regs) {
1571   if ( ShowMessageBoxOnError ) {
1572     JavaThread* thread = JavaThread::current();
1573     JavaThreadState saved_state = thread->thread_state();
1574     thread->set_thread_state(_thread_in_vm);
1575       {
1576         // In order to get locks work, we need to fake a in_VM state
1577         ttyLocker ttyl;
1578         ::tty->print_cr("EXECUTION STOPPED: %s\n", msg);
1579         if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
1580         BytecodeCounter::print();
1581         }
1582         if (os::message_box(msg, "Execution stopped, print registers?"))
1583           regs->print(::tty);
1584       }
1585     BREAKPOINT;
1586       ThreadStateTransition::transition(JavaThread::current(), _thread_in_vm, saved_state);
1587   }
1588   else {
1589      ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
1590   }
1591   assert(false, err_msg("DEBUG MESSAGE: %s", msg));
1592 }
1593 
1594 
1595 void MacroAssembler::calc_mem_param_words(Register Rparam_words, Register Rresult) {
1596   subcc( Rparam_words, Argument::n_register_parameters, Rresult); // how many mem words?
1597   Label no_extras;
1598   br( negative, true, pt, no_extras ); // if neg, clear reg
1599   delayed()->set(0, Rresult);          // annuled, so only if taken
1600   bind( no_extras );
1601 }
1602 
1603 
1604 void MacroAssembler::calc_frame_size(Register Rextra_words, Register Rresult) {
1605 #ifdef _LP64
1606   add(Rextra_words, frame::memory_parameter_word_sp_offset, Rresult);
1607 #else
1608   add(Rextra_words, frame::memory_parameter_word_sp_offset + 1, Rresult);
1609 #endif
1610   bclr(1, Rresult);
1611   sll(Rresult, LogBytesPerWord, Rresult);  // Rresult has total frame bytes
1612 }
1613 
1614 
1615 void MacroAssembler::calc_frame_size_and_save(Register Rextra_words, Register Rresult) {
1616   calc_frame_size(Rextra_words, Rresult);
1617   neg(Rresult);
1618   save(SP, Rresult, SP);
1619 }
1620 
1621 
1622 // ---------------------------------------------------------
1623 Assembler::RCondition cond2rcond(Assembler::Condition c) {
1624   switch (c) {
1625     /*case zero: */
1626     case Assembler::equal:        return Assembler::rc_z;
1627     case Assembler::lessEqual:    return Assembler::rc_lez;
1628     case Assembler::less:         return Assembler::rc_lz;
1629     /*case notZero:*/
1630     case Assembler::notEqual:     return Assembler::rc_nz;
1631     case Assembler::greater:      return Assembler::rc_gz;
1632     case Assembler::greaterEqual: return Assembler::rc_gez;
1633   }
1634   ShouldNotReachHere();
1635   return Assembler::rc_z;
1636 }
1637 
1638 // compares (32 bit) register with zero and branches.  NOT FOR USE WITH 64-bit POINTERS
1639 void MacroAssembler::cmp_zero_and_br(Condition c, Register s1, Label& L, bool a, Predict p) {
1640   tst(s1);
1641   br (c, a, p, L);
1642 }
1643 
1644 // Compares a pointer register with zero and branches on null.
1645 // Does a test & branch on 32-bit systems and a register-branch on 64-bit.
1646 void MacroAssembler::br_null( Register s1, bool a, Predict p, Label& L ) {
1647   assert_not_delayed();
1648 #ifdef _LP64
1649   bpr( rc_z, a, p, s1, L );
1650 #else
1651   tst(s1);
1652   br ( zero, a, p, L );
1653 #endif
1654 }
1655 
1656 void MacroAssembler::br_notnull( Register s1, bool a, Predict p, Label& L ) {
1657   assert_not_delayed();
1658 #ifdef _LP64
1659   bpr( rc_nz, a, p, s1, L );
1660 #else
1661   tst(s1);
1662   br ( notZero, a, p, L );
1663 #endif
1664 }
1665 
1666 // Compare registers and branch with nop in delay slot or cbcond without delay slot.
1667 
1668 // Compare integer (32 bit) values (icc only).
1669 void MacroAssembler::cmp_and_br_short(Register s1, Register s2, Condition c,
1670                                       Predict p, Label& L) {
1671   assert_not_delayed();
1672   if (use_cbcond(L)) {
1673     Assembler::cbcond(c, icc, s1, s2, L);
1674   } else {
1675     cmp(s1, s2);
1676     br(c, false, p, L);
1677     delayed()->nop();
1678   }
1679 }
1680 
1681 // Compare integer (32 bit) values (icc only).
1682 void MacroAssembler::cmp_and_br_short(Register s1, int simm13a, Condition c,
1683                                       Predict p, Label& L) {
1684   assert_not_delayed();
1685   if (is_simm(simm13a,5) && use_cbcond(L)) {
1686     Assembler::cbcond(c, icc, s1, simm13a, L);
1687   } else {
1688     cmp(s1, simm13a);
1689     br(c, false, p, L);
1690     delayed()->nop();
1691   }
1692 }
1693 
1694 // Branch that tests xcc in LP64 and icc in !LP64
1695 void MacroAssembler::cmp_and_brx_short(Register s1, Register s2, Condition c,
1696                                        Predict p, Label& L) {
1697   assert_not_delayed();
1698   if (use_cbcond(L)) {
1699     Assembler::cbcond(c, ptr_cc, s1, s2, L);
1700   } else {
1701     cmp(s1, s2);
1702     brx(c, false, p, L);
1703     delayed()->nop();
1704   }
1705 }
1706 
1707 // Branch that tests xcc in LP64 and icc in !LP64
1708 void MacroAssembler::cmp_and_brx_short(Register s1, int simm13a, Condition c,
1709                                        Predict p, Label& L) {
1710   assert_not_delayed();
1711   if (is_simm(simm13a,5) && use_cbcond(L)) {
1712     Assembler::cbcond(c, ptr_cc, s1, simm13a, L);
1713   } else {
1714     cmp(s1, simm13a);
1715     brx(c, false, p, L);
1716     delayed()->nop();
1717   }
1718 }
1719 
1720 // Short branch version for compares a pointer with zero.
1721 
1722 void MacroAssembler::br_null_short(Register s1, Predict p, Label& L) {
1723   assert_not_delayed();
1724   if (use_cbcond(L)) {
1725     Assembler::cbcond(zero, ptr_cc, s1, 0, L);
1726     return;
1727   }
1728   br_null(s1, false, p, L);
1729   delayed()->nop();
1730 }
1731 
1732 void MacroAssembler::br_notnull_short(Register s1, Predict p, Label& L) {
1733   assert_not_delayed();
1734   if (use_cbcond(L)) {
1735     Assembler::cbcond(notZero, ptr_cc, s1, 0, L);
1736     return;
1737   }
1738   br_notnull(s1, false, p, L);
1739   delayed()->nop();
1740 }
1741 
1742 // Unconditional short branch
1743 void MacroAssembler::ba_short(Label& L) {
1744   if (use_cbcond(L)) {
1745     Assembler::cbcond(equal, icc, G0, G0, L);
1746     return;
1747   }
1748   br(always, false, pt, L);
1749   delayed()->nop();
1750 }
1751 
1752 // instruction sequences factored across compiler & interpreter
1753 
1754 
1755 void MacroAssembler::lcmp( Register Ra_hi, Register Ra_low,
1756                            Register Rb_hi, Register Rb_low,
1757                            Register Rresult) {
1758 
1759   Label check_low_parts, done;
1760 
1761   cmp(Ra_hi, Rb_hi );  // compare hi parts
1762   br(equal, true, pt, check_low_parts);
1763   delayed()->cmp(Ra_low, Rb_low); // test low parts
1764 
1765   // And, with an unsigned comparison, it does not matter if the numbers
1766   // are negative or not.
1767   // E.g., -2 cmp -1: the low parts are 0xfffffffe and 0xffffffff.
1768   // The second one is bigger (unsignedly).
1769 
1770   // Other notes:  The first move in each triplet can be unconditional
1771   // (and therefore probably prefetchable).
1772   // And the equals case for the high part does not need testing,
1773   // since that triplet is reached only after finding the high halves differ.
1774 
1775   mov(-1, Rresult);
1776   ba(done);
1777   delayed()->movcc(greater, false, icc,  1, Rresult);
1778 
1779   bind(check_low_parts);
1780 
1781   mov(                               -1, Rresult);
1782   movcc(equal,           false, icc,  0, Rresult);
1783   movcc(greaterUnsigned, false, icc,  1, Rresult);
1784 
1785   bind(done);
1786 }
1787 
1788 void MacroAssembler::lneg( Register Rhi, Register Rlow ) {
1789   subcc(  G0, Rlow, Rlow );
1790   subc(   G0, Rhi,  Rhi  );
1791 }
1792 
1793 void MacroAssembler::lshl( Register Rin_high,  Register Rin_low,
1794                            Register Rcount,
1795                            Register Rout_high, Register Rout_low,
1796                            Register Rtemp ) {
1797 
1798 
1799   Register Ralt_count = Rtemp;
1800   Register Rxfer_bits = Rtemp;
1801 
1802   assert( Ralt_count != Rin_high
1803       &&  Ralt_count != Rin_low
1804       &&  Ralt_count != Rcount
1805       &&  Rxfer_bits != Rin_low
1806       &&  Rxfer_bits != Rin_high
1807       &&  Rxfer_bits != Rcount
1808       &&  Rxfer_bits != Rout_low
1809       &&  Rout_low   != Rin_high,
1810         "register alias checks");
1811 
1812   Label big_shift, done;
1813 
1814   // This code can be optimized to use the 64 bit shifts in V9.
1815   // Here we use the 32 bit shifts.
1816 
1817   and3( Rcount, 0x3f, Rcount);     // take least significant 6 bits
1818   subcc(Rcount,   31, Ralt_count);
1819   br(greater, true, pn, big_shift);
1820   delayed()->dec(Ralt_count);
1821 
1822   // shift < 32 bits, Ralt_count = Rcount-31
1823 
1824   // We get the transfer bits by shifting right by 32-count the low
1825   // register. This is done by shifting right by 31-count and then by one
1826   // more to take care of the special (rare) case where count is zero
1827   // (shifting by 32 would not work).
1828 
1829   neg(Ralt_count);
1830 
1831   // The order of the next two instructions is critical in the case where
1832   // Rin and Rout are the same and should not be reversed.
1833 
1834   srl(Rin_low, Ralt_count, Rxfer_bits); // shift right by 31-count
1835   if (Rcount != Rout_low) {
1836     sll(Rin_low, Rcount, Rout_low); // low half
1837   }
1838   sll(Rin_high, Rcount, Rout_high);
1839   if (Rcount == Rout_low) {
1840     sll(Rin_low, Rcount, Rout_low); // low half
1841   }
1842   srl(Rxfer_bits, 1, Rxfer_bits ); // shift right by one more
1843   ba(done);
1844   delayed()->or3(Rout_high, Rxfer_bits, Rout_high);   // new hi value: or in shifted old hi part and xfer from low
1845 
1846   // shift >= 32 bits, Ralt_count = Rcount-32
1847   bind(big_shift);
1848   sll(Rin_low, Ralt_count, Rout_high  );
1849   clr(Rout_low);
1850 
1851   bind(done);
1852 }
1853 
1854 
1855 void MacroAssembler::lshr( Register Rin_high,  Register Rin_low,
1856                            Register Rcount,
1857                            Register Rout_high, Register Rout_low,
1858                            Register Rtemp ) {
1859 
1860   Register Ralt_count = Rtemp;
1861   Register Rxfer_bits = Rtemp;
1862 
1863   assert( Ralt_count != Rin_high
1864       &&  Ralt_count != Rin_low
1865       &&  Ralt_count != Rcount
1866       &&  Rxfer_bits != Rin_low
1867       &&  Rxfer_bits != Rin_high
1868       &&  Rxfer_bits != Rcount
1869       &&  Rxfer_bits != Rout_high
1870       &&  Rout_high  != Rin_low,
1871         "register alias checks");
1872 
1873   Label big_shift, done;
1874 
1875   // This code can be optimized to use the 64 bit shifts in V9.
1876   // Here we use the 32 bit shifts.
1877 
1878   and3( Rcount, 0x3f, Rcount);     // take least significant 6 bits
1879   subcc(Rcount,   31, Ralt_count);
1880   br(greater, true, pn, big_shift);
1881   delayed()->dec(Ralt_count);
1882 
1883   // shift < 32 bits, Ralt_count = Rcount-31
1884 
1885   // We get the transfer bits by shifting left by 32-count the high
1886   // register. This is done by shifting left by 31-count and then by one
1887   // more to take care of the special (rare) case where count is zero
1888   // (shifting by 32 would not work).
1889 
1890   neg(Ralt_count);
1891   if (Rcount != Rout_low) {
1892     srl(Rin_low, Rcount, Rout_low);
1893   }
1894 
1895   // The order of the next two instructions is critical in the case where
1896   // Rin and Rout are the same and should not be reversed.
1897 
1898   sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count
1899   sra(Rin_high,     Rcount, Rout_high ); // high half
1900   sll(Rxfer_bits,        1, Rxfer_bits); // shift left by one more
1901   if (Rcount == Rout_low) {
1902     srl(Rin_low, Rcount, Rout_low);
1903   }
1904   ba(done);
1905   delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high
1906 
1907   // shift >= 32 bits, Ralt_count = Rcount-32
1908   bind(big_shift);
1909 
1910   sra(Rin_high, Ralt_count, Rout_low);
1911   sra(Rin_high,         31, Rout_high); // sign into hi
1912 
1913   bind( done );
1914 }
1915 
1916 
1917 
1918 void MacroAssembler::lushr( Register Rin_high,  Register Rin_low,
1919                             Register Rcount,
1920                             Register Rout_high, Register Rout_low,
1921                             Register Rtemp ) {
1922 
1923   Register Ralt_count = Rtemp;
1924   Register Rxfer_bits = Rtemp;
1925 
1926   assert( Ralt_count != Rin_high
1927       &&  Ralt_count != Rin_low
1928       &&  Ralt_count != Rcount
1929       &&  Rxfer_bits != Rin_low
1930       &&  Rxfer_bits != Rin_high
1931       &&  Rxfer_bits != Rcount
1932       &&  Rxfer_bits != Rout_high
1933       &&  Rout_high  != Rin_low,
1934         "register alias checks");
1935 
1936   Label big_shift, done;
1937 
1938   // This code can be optimized to use the 64 bit shifts in V9.
1939   // Here we use the 32 bit shifts.
1940 
1941   and3( Rcount, 0x3f, Rcount);     // take least significant 6 bits
1942   subcc(Rcount,   31, Ralt_count);
1943   br(greater, true, pn, big_shift);
1944   delayed()->dec(Ralt_count);
1945 
1946   // shift < 32 bits, Ralt_count = Rcount-31
1947 
1948   // We get the transfer bits by shifting left by 32-count the high
1949   // register. This is done by shifting left by 31-count and then by one
1950   // more to take care of the special (rare) case where count is zero
1951   // (shifting by 32 would not work).
1952 
1953   neg(Ralt_count);
1954   if (Rcount != Rout_low) {
1955     srl(Rin_low, Rcount, Rout_low);
1956   }
1957 
1958   // The order of the next two instructions is critical in the case where
1959   // Rin and Rout are the same and should not be reversed.
1960 
1961   sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count
1962   srl(Rin_high,     Rcount, Rout_high ); // high half
1963   sll(Rxfer_bits,        1, Rxfer_bits); // shift left by one more
1964   if (Rcount == Rout_low) {
1965     srl(Rin_low, Rcount, Rout_low);
1966   }
1967   ba(done);
1968   delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high
1969 
1970   // shift >= 32 bits, Ralt_count = Rcount-32
1971   bind(big_shift);
1972 
1973   srl(Rin_high, Ralt_count, Rout_low);
1974   clr(Rout_high);
1975 
1976   bind( done );
1977 }
1978 
1979 #ifdef _LP64
1980 void MacroAssembler::lcmp( Register Ra, Register Rb, Register Rresult) {
1981   cmp(Ra, Rb);
1982   mov(-1, Rresult);
1983   movcc(equal,   false, xcc,  0, Rresult);
1984   movcc(greater, false, xcc,  1, Rresult);
1985 }
1986 #endif
1987 
1988 
1989 void MacroAssembler::load_sized_value(Address src, Register dst, size_t size_in_bytes, bool is_signed) {
1990   switch (size_in_bytes) {
1991   case  8:  ld_long(src, dst); break;
1992   case  4:  ld(     src, dst); break;
1993   case  2:  is_signed ? ldsh(src, dst) : lduh(src, dst); break;
1994   case  1:  is_signed ? ldsb(src, dst) : ldub(src, dst); break;
1995   default:  ShouldNotReachHere();
1996   }
1997 }
1998 
1999 void MacroAssembler::store_sized_value(Register src, Address dst, size_t size_in_bytes) {
2000   switch (size_in_bytes) {
2001   case  8:  st_long(src, dst); break;
2002   case  4:  st(     src, dst); break;
2003   case  2:  sth(    src, dst); break;
2004   case  1:  stb(    src, dst); break;
2005   default:  ShouldNotReachHere();
2006   }
2007 }
2008 
2009 
2010 void MacroAssembler::float_cmp( bool is_float, int unordered_result,
2011                                 FloatRegister Fa, FloatRegister Fb,
2012                                 Register Rresult) {
2013   if (is_float) {
2014     fcmp(FloatRegisterImpl::S, fcc0, Fa, Fb);
2015   } else {
2016     fcmp(FloatRegisterImpl::D, fcc0, Fa, Fb);
2017   }
2018 
2019   if (unordered_result == 1) {
2020     mov(                                    -1, Rresult);
2021     movcc(f_equal,              true, fcc0,  0, Rresult);
2022     movcc(f_unorderedOrGreater, true, fcc0,  1, Rresult);
2023   } else {
2024     mov(                                    -1, Rresult);
2025     movcc(f_equal,              true, fcc0,  0, Rresult);
2026     movcc(f_greater,            true, fcc0,  1, Rresult);
2027   }
2028 }
2029 
2030 
2031 void MacroAssembler::save_all_globals_into_locals() {
2032   mov(G1,L1);
2033   mov(G2,L2);
2034   mov(G3,L3);
2035   mov(G4,L4);
2036   mov(G5,L5);
2037   mov(G6,L6);
2038   mov(G7,L7);
2039 }
2040 
2041 void MacroAssembler::restore_globals_from_locals() {
2042   mov(L1,G1);
2043   mov(L2,G2);
2044   mov(L3,G3);
2045   mov(L4,G4);
2046   mov(L5,G5);
2047   mov(L6,G6);
2048   mov(L7,G7);
2049 }
2050 
2051 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
2052                                                       Register tmp,
2053                                                       int offset) {
2054   intptr_t value = *delayed_value_addr;
2055   if (value != 0)
2056     return RegisterOrConstant(value + offset);
2057 
2058   // load indirectly to solve generation ordering problem
2059   AddressLiteral a(delayed_value_addr);
2060   load_ptr_contents(a, tmp);
2061 
2062 #ifdef ASSERT
2063   tst(tmp);
2064   breakpoint_trap(zero, xcc);
2065 #endif
2066 
2067   if (offset != 0)
2068     add(tmp, offset, tmp);
2069 
2070   return RegisterOrConstant(tmp);
2071 }
2072 
2073 
2074 RegisterOrConstant MacroAssembler::regcon_andn_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) {
2075   assert(d.register_or_noreg() != G0, "lost side effect");
2076   if ((s2.is_constant() && s2.as_constant() == 0) ||
2077       (s2.is_register() && s2.as_register() == G0)) {
2078     // Do nothing, just move value.
2079     if (s1.is_register()) {
2080       if (d.is_constant())  d = temp;
2081       mov(s1.as_register(), d.as_register());
2082       return d;
2083     } else {
2084       return s1;
2085     }
2086   }
2087 
2088   if (s1.is_register()) {
2089     assert_different_registers(s1.as_register(), temp);
2090     if (d.is_constant())  d = temp;
2091     andn(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register());
2092     return d;
2093   } else {
2094     if (s2.is_register()) {
2095       assert_different_registers(s2.as_register(), temp);
2096       if (d.is_constant())  d = temp;
2097       set(s1.as_constant(), temp);
2098       andn(temp, s2.as_register(), d.as_register());
2099       return d;
2100     } else {
2101       intptr_t res = s1.as_constant() & ~s2.as_constant();
2102       return res;
2103     }
2104   }
2105 }
2106 
2107 RegisterOrConstant MacroAssembler::regcon_inc_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) {
2108   assert(d.register_or_noreg() != G0, "lost side effect");
2109   if ((s2.is_constant() && s2.as_constant() == 0) ||
2110       (s2.is_register() && s2.as_register() == G0)) {
2111     // Do nothing, just move value.
2112     if (s1.is_register()) {
2113       if (d.is_constant())  d = temp;
2114       mov(s1.as_register(), d.as_register());
2115       return d;
2116     } else {
2117       return s1;
2118     }
2119   }
2120 
2121   if (s1.is_register()) {
2122     assert_different_registers(s1.as_register(), temp);
2123     if (d.is_constant())  d = temp;
2124     add(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register());
2125     return d;
2126   } else {
2127     if (s2.is_register()) {
2128       assert_different_registers(s2.as_register(), temp);
2129       if (d.is_constant())  d = temp;
2130       add(s2.as_register(), ensure_simm13_or_reg(s1, temp), d.as_register());
2131       return d;
2132     } else {
2133       intptr_t res = s1.as_constant() + s2.as_constant();
2134       return res;
2135     }
2136   }
2137 }
2138 
2139 RegisterOrConstant MacroAssembler::regcon_sll_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) {
2140   assert(d.register_or_noreg() != G0, "lost side effect");
2141   if (!is_simm13(s2.constant_or_zero()))
2142     s2 = (s2.as_constant() & 0xFF);
2143   if ((s2.is_constant() && s2.as_constant() == 0) ||
2144       (s2.is_register() && s2.as_register() == G0)) {
2145     // Do nothing, just move value.
2146     if (s1.is_register()) {
2147       if (d.is_constant())  d = temp;
2148       mov(s1.as_register(), d.as_register());
2149       return d;
2150     } else {
2151       return s1;
2152     }
2153   }
2154 
2155   if (s1.is_register()) {
2156     assert_different_registers(s1.as_register(), temp);
2157     if (d.is_constant())  d = temp;
2158     sll_ptr(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register());
2159     return d;
2160   } else {
2161     if (s2.is_register()) {
2162       assert_different_registers(s2.as_register(), temp);
2163       if (d.is_constant())  d = temp;
2164       set(s1.as_constant(), temp);
2165       sll_ptr(temp, s2.as_register(), d.as_register());
2166       return d;
2167     } else {
2168       intptr_t res = s1.as_constant() << s2.as_constant();
2169       return res;
2170     }
2171   }
2172 }
2173 
2174 
2175 // Look up the method for a megamorphic invokeinterface call.
2176 // The target method is determined by <intf_klass, itable_index>.
2177 // The receiver klass is in recv_klass.
2178 // On success, the result will be in method_result, and execution falls through.
2179 // On failure, execution transfers to the given label.
2180 void MacroAssembler::lookup_interface_method(Register recv_klass,
2181                                              Register intf_klass,
2182                                              RegisterOrConstant itable_index,
2183                                              Register method_result,
2184                                              Register scan_temp,
2185                                              Register sethi_temp,
2186                                              Label& L_no_such_interface) {
2187   assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
2188   assert(itable_index.is_constant() || itable_index.as_register() == method_result,
2189          "caller must use same register for non-constant itable index as for method");
2190 
2191   Label L_no_such_interface_restore;
2192   bool did_save = false;
2193   if (scan_temp == noreg || sethi_temp == noreg) {
2194     Register recv_2 = recv_klass->is_global() ? recv_klass : L0;
2195     Register intf_2 = intf_klass->is_global() ? intf_klass : L1;
2196     assert(method_result->is_global(), "must be able to return value");
2197     scan_temp  = L2;
2198     sethi_temp = L3;
2199     save_frame_and_mov(0, recv_klass, recv_2, intf_klass, intf_2);
2200     recv_klass = recv_2;
2201     intf_klass = intf_2;
2202     did_save = true;
2203   }
2204 
2205   // Compute start of first itableOffsetEntry (which is at the end of the vtable)
2206   int vtable_base = InstanceKlass::vtable_start_offset() * wordSize;
2207   int scan_step   = itableOffsetEntry::size() * wordSize;
2208   int vte_size    = vtableEntry::size() * wordSize;
2209 
2210   lduw(recv_klass, InstanceKlass::vtable_length_offset() * wordSize, scan_temp);
2211   // %%% We should store the aligned, prescaled offset in the klassoop.
2212   // Then the next several instructions would fold away.
2213 
2214   int round_to_unit = ((HeapWordsPerLong > 1) ? BytesPerLong : 0);
2215   int itb_offset = vtable_base;
2216   if (round_to_unit != 0) {
2217     // hoist first instruction of round_to(scan_temp, BytesPerLong):
2218     itb_offset += round_to_unit - wordSize;
2219   }
2220   int itb_scale = exact_log2(vtableEntry::size() * wordSize);
2221   sll(scan_temp, itb_scale,  scan_temp);
2222   add(scan_temp, itb_offset, scan_temp);
2223   if (round_to_unit != 0) {
2224     // Round up to align_object_offset boundary
2225     // see code for InstanceKlass::start_of_itable!
2226     // Was: round_to(scan_temp, BytesPerLong);
2227     // Hoisted: add(scan_temp, BytesPerLong-1, scan_temp);
2228     and3(scan_temp, -round_to_unit, scan_temp);
2229   }
2230   add(recv_klass, scan_temp, scan_temp);
2231 
2232   // Adjust recv_klass by scaled itable_index, so we can free itable_index.
2233   RegisterOrConstant itable_offset = itable_index;
2234   itable_offset = regcon_sll_ptr(itable_index, exact_log2(itableMethodEntry::size() * wordSize), itable_offset);
2235   itable_offset = regcon_inc_ptr(itable_offset, itableMethodEntry::method_offset_in_bytes(), itable_offset);
2236   add(recv_klass, ensure_simm13_or_reg(itable_offset, sethi_temp), recv_klass);
2237 
2238   // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
2239   //   if (scan->interface() == intf) {
2240   //     result = (klass + scan->offset() + itable_index);
2241   //   }
2242   // }
2243   Label L_search, L_found_method;
2244 
2245   for (int peel = 1; peel >= 0; peel--) {
2246     // %%%% Could load both offset and interface in one ldx, if they were
2247     // in the opposite order.  This would save a load.
2248     ld_ptr(scan_temp, itableOffsetEntry::interface_offset_in_bytes(), method_result);
2249 
2250     // Check that this entry is non-null.  A null entry means that
2251     // the receiver class doesn't implement the interface, and wasn't the
2252     // same as when the caller was compiled.
2253     bpr(Assembler::rc_z, false, Assembler::pn, method_result, did_save ? L_no_such_interface_restore : L_no_such_interface);
2254     delayed()->cmp(method_result, intf_klass);
2255 
2256     if (peel) {
2257       brx(Assembler::equal,    false, Assembler::pt, L_found_method);
2258     } else {
2259       brx(Assembler::notEqual, false, Assembler::pn, L_search);
2260       // (invert the test to fall through to found_method...)
2261     }
2262     delayed()->add(scan_temp, scan_step, scan_temp);
2263 
2264     if (!peel)  break;
2265 
2266     bind(L_search);
2267   }
2268 
2269   bind(L_found_method);
2270 
2271   // Got a hit.
2272   int ito_offset = itableOffsetEntry::offset_offset_in_bytes();
2273   // scan_temp[-scan_step] points to the vtable offset we need
2274   ito_offset -= scan_step;
2275   lduw(scan_temp, ito_offset, scan_temp);
2276   ld_ptr(recv_klass, scan_temp, method_result);
2277 
2278   if (did_save) {
2279     Label L_done;
2280     ba(L_done);
2281     delayed()->restore();
2282 
2283     bind(L_no_such_interface_restore);
2284     ba(L_no_such_interface);
2285     delayed()->restore();
2286 
2287     bind(L_done);
2288   }
2289 }
2290 
2291 
2292 // virtual method calling
2293 void MacroAssembler::lookup_virtual_method(Register recv_klass,
2294                                            RegisterOrConstant vtable_index,
2295                                            Register method_result) {
2296   assert_different_registers(recv_klass, method_result, vtable_index.register_or_noreg());
2297   Register sethi_temp = method_result;
2298   const int base = (InstanceKlass::vtable_start_offset() * wordSize +
2299                     // method pointer offset within the vtable entry:
2300                     vtableEntry::method_offset_in_bytes());
2301   RegisterOrConstant vtable_offset = vtable_index;
2302   // Each of the following three lines potentially generates an instruction.
2303   // But the total number of address formation instructions will always be
2304   // at most two, and will often be zero.  In any case, it will be optimal.
2305   // If vtable_index is a register, we will have (sll_ptr N,x; inc_ptr B,x; ld_ptr k,x).
2306   // If vtable_index is a constant, we will have at most (set B+X<<N,t; ld_ptr k,t).
2307   vtable_offset = regcon_sll_ptr(vtable_index, exact_log2(vtableEntry::size() * wordSize), vtable_offset);
2308   vtable_offset = regcon_inc_ptr(vtable_offset, base, vtable_offset, sethi_temp);
2309   Address vtable_entry_addr(recv_klass, ensure_simm13_or_reg(vtable_offset, sethi_temp));
2310   ld_ptr(vtable_entry_addr, method_result);
2311 }
2312 
2313 
2314 void MacroAssembler::check_klass_subtype(Register sub_klass,
2315                                          Register super_klass,
2316                                          Register temp_reg,
2317                                          Register temp2_reg,
2318                                          Label& L_success) {
2319   Register sub_2 = sub_klass;
2320   Register sup_2 = super_klass;
2321   if (!sub_2->is_global())  sub_2 = L0;
2322   if (!sup_2->is_global())  sup_2 = L1;
2323   bool did_save = false;
2324   if (temp_reg == noreg || temp2_reg == noreg) {
2325     temp_reg = L2;
2326     temp2_reg = L3;
2327     save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2);
2328     sub_klass = sub_2;
2329     super_klass = sup_2;
2330     did_save = true;
2331   }
2332   Label L_failure, L_pop_to_failure, L_pop_to_success;
2333   check_klass_subtype_fast_path(sub_klass, super_klass,
2334                                 temp_reg, temp2_reg,
2335                                 (did_save ? &L_pop_to_success : &L_success),
2336                                 (did_save ? &L_pop_to_failure : &L_failure), NULL);
2337 
2338   if (!did_save)
2339     save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2);
2340   check_klass_subtype_slow_path(sub_2, sup_2,
2341                                 L2, L3, L4, L5,
2342                                 NULL, &L_pop_to_failure);
2343 
2344   // on success:
2345   bind(L_pop_to_success);
2346   restore();
2347   ba_short(L_success);
2348 
2349   // on failure:
2350   bind(L_pop_to_failure);
2351   restore();
2352   bind(L_failure);
2353 }
2354 
2355 
2356 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
2357                                                    Register super_klass,
2358                                                    Register temp_reg,
2359                                                    Register temp2_reg,
2360                                                    Label* L_success,
2361                                                    Label* L_failure,
2362                                                    Label* L_slow_path,
2363                                         RegisterOrConstant super_check_offset) {
2364   int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
2365   int sco_offset = in_bytes(Klass::super_check_offset_offset());
2366 
2367   bool must_load_sco  = (super_check_offset.constant_or_zero() == -1);
2368   bool need_slow_path = (must_load_sco ||
2369                          super_check_offset.constant_or_zero() == sco_offset);
2370 
2371   assert_different_registers(sub_klass, super_klass, temp_reg);
2372   if (super_check_offset.is_register()) {
2373     assert_different_registers(sub_klass, super_klass, temp_reg,
2374                                super_check_offset.as_register());
2375   } else if (must_load_sco) {
2376     assert(temp2_reg != noreg, "supply either a temp or a register offset");
2377   }
2378 
2379   Label L_fallthrough;
2380   int label_nulls = 0;
2381   if (L_success == NULL)   { L_success   = &L_fallthrough; label_nulls++; }
2382   if (L_failure == NULL)   { L_failure   = &L_fallthrough; label_nulls++; }
2383   if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
2384   assert(label_nulls <= 1 ||
2385          (L_slow_path == &L_fallthrough && label_nulls <= 2 && !need_slow_path),
2386          "at most one NULL in the batch, usually");
2387 
2388   // If the pointers are equal, we are done (e.g., String[] elements).
2389   // This self-check enables sharing of secondary supertype arrays among
2390   // non-primary types such as array-of-interface.  Otherwise, each such
2391   // type would need its own customized SSA.
2392   // We move this check to the front of the fast path because many
2393   // type checks are in fact trivially successful in this manner,
2394   // so we get a nicely predicted branch right at the start of the check.
2395   cmp(super_klass, sub_klass);
2396   brx(Assembler::equal, false, Assembler::pn, *L_success);
2397   delayed()->nop();
2398 
2399   // Check the supertype display:
2400   if (must_load_sco) {
2401     // The super check offset is always positive...
2402     lduw(super_klass, sco_offset, temp2_reg);
2403     super_check_offset = RegisterOrConstant(temp2_reg);
2404     // super_check_offset is register.
2405     assert_different_registers(sub_klass, super_klass, temp_reg, super_check_offset.as_register());
2406   }
2407   ld_ptr(sub_klass, super_check_offset, temp_reg);
2408   cmp(super_klass, temp_reg);
2409 
2410   // This check has worked decisively for primary supers.
2411   // Secondary supers are sought in the super_cache ('super_cache_addr').
2412   // (Secondary supers are interfaces and very deeply nested subtypes.)
2413   // This works in the same check above because of a tricky aliasing
2414   // between the super_cache and the primary super display elements.
2415   // (The 'super_check_addr' can address either, as the case requires.)
2416   // Note that the cache is updated below if it does not help us find
2417   // what we need immediately.
2418   // So if it was a primary super, we can just fail immediately.
2419   // Otherwise, it's the slow path for us (no success at this point).
2420 
2421   // Hacked ba(), which may only be used just before L_fallthrough.
2422 #define FINAL_JUMP(label)            \
2423   if (&(label) != &L_fallthrough) {  \
2424     ba(label);  delayed()->nop();    \
2425   }
2426 
2427   if (super_check_offset.is_register()) {
2428     brx(Assembler::equal, false, Assembler::pn, *L_success);
2429     delayed()->cmp(super_check_offset.as_register(), sc_offset);
2430 
2431     if (L_failure == &L_fallthrough) {
2432       brx(Assembler::equal, false, Assembler::pt, *L_slow_path);
2433       delayed()->nop();
2434     } else {
2435       brx(Assembler::notEqual, false, Assembler::pn, *L_failure);
2436       delayed()->nop();
2437       FINAL_JUMP(*L_slow_path);
2438     }
2439   } else if (super_check_offset.as_constant() == sc_offset) {
2440     // Need a slow path; fast failure is impossible.
2441     if (L_slow_path == &L_fallthrough) {
2442       brx(Assembler::equal, false, Assembler::pt, *L_success);
2443       delayed()->nop();
2444     } else {
2445       brx(Assembler::notEqual, false, Assembler::pn, *L_slow_path);
2446       delayed()->nop();
2447       FINAL_JUMP(*L_success);
2448     }
2449   } else {
2450     // No slow path; it's a fast decision.
2451     if (L_failure == &L_fallthrough) {
2452       brx(Assembler::equal, false, Assembler::pt, *L_success);
2453       delayed()->nop();
2454     } else {
2455       brx(Assembler::notEqual, false, Assembler::pn, *L_failure);
2456       delayed()->nop();
2457       FINAL_JUMP(*L_success);
2458     }
2459   }
2460 
2461   bind(L_fallthrough);
2462 
2463 #undef FINAL_JUMP
2464 }
2465 
2466 
2467 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
2468                                                    Register super_klass,
2469                                                    Register count_temp,
2470                                                    Register scan_temp,
2471                                                    Register scratch_reg,
2472                                                    Register coop_reg,
2473                                                    Label* L_success,
2474                                                    Label* L_failure) {
2475   assert_different_registers(sub_klass, super_klass,
2476                              count_temp, scan_temp, scratch_reg, coop_reg);
2477 
2478   Label L_fallthrough, L_loop;
2479   int label_nulls = 0;
2480   if (L_success == NULL)   { L_success   = &L_fallthrough; label_nulls++; }
2481   if (L_failure == NULL)   { L_failure   = &L_fallthrough; label_nulls++; }
2482   assert(label_nulls <= 1, "at most one NULL in the batch");
2483 
2484   // a couple of useful fields in sub_klass:
2485   int ss_offset = in_bytes(Klass::secondary_supers_offset());
2486   int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
2487 
2488   // Do a linear scan of the secondary super-klass chain.
2489   // This code is rarely used, so simplicity is a virtue here.
2490 
2491 #ifndef PRODUCT
2492   int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
2493   inc_counter((address) pst_counter, count_temp, scan_temp);
2494 #endif
2495 
2496   // We will consult the secondary-super array.
2497   ld_ptr(sub_klass, ss_offset, scan_temp);
2498 
2499   Register search_key = super_klass;
2500 
2501   // Load the array length.  (Positive movl does right thing on LP64.)
2502   lduw(scan_temp, Array<Klass*>::length_offset_in_bytes(), count_temp);
2503 
2504   // Check for empty secondary super list
2505   tst(count_temp);
2506 
2507   // In the array of super classes elements are pointer sized.
2508   int element_size = wordSize;
2509 
2510   // Top of search loop
2511   bind(L_loop);
2512   br(Assembler::equal, false, Assembler::pn, *L_failure);
2513   delayed()->add(scan_temp, element_size, scan_temp);
2514 
2515   // Skip the array header in all array accesses.
2516   int elem_offset = Array<Klass*>::base_offset_in_bytes();
2517   elem_offset -= element_size;   // the scan pointer was pre-incremented also
2518 
2519   // Load next super to check
2520     ld_ptr( scan_temp, elem_offset, scratch_reg );
2521 
2522   // Look for Rsuper_klass on Rsub_klass's secondary super-class-overflow list
2523   cmp(scratch_reg, search_key);
2524 
2525   // A miss means we are NOT a subtype and need to keep looping
2526   brx(Assembler::notEqual, false, Assembler::pn, L_loop);
2527   delayed()->deccc(count_temp); // decrement trip counter in delay slot
2528 
2529   // Success.  Cache the super we found and proceed in triumph.
2530   st_ptr(super_klass, sub_klass, sc_offset);
2531 
2532   if (L_success != &L_fallthrough) {
2533     ba(*L_success);
2534     delayed()->nop();
2535   }
2536 
2537   bind(L_fallthrough);
2538 }
2539 
2540 
2541 RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot,
2542                                                    Register temp_reg,
2543                                                    int extra_slot_offset) {
2544   // cf. TemplateTable::prepare_invoke(), if (load_receiver).
2545   int stackElementSize = Interpreter::stackElementSize;
2546   int offset = extra_slot_offset * stackElementSize;
2547   if (arg_slot.is_constant()) {
2548     offset += arg_slot.as_constant() * stackElementSize;
2549     return offset;
2550   } else {
2551     assert(temp_reg != noreg, "must specify");
2552     sll_ptr(arg_slot.as_register(), exact_log2(stackElementSize), temp_reg);
2553     if (offset != 0)
2554       add(temp_reg, offset, temp_reg);
2555     return temp_reg;
2556   }
2557 }
2558 
2559 
2560 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
2561                                          Register temp_reg,
2562                                          int extra_slot_offset) {
2563   return Address(Gargs, argument_offset(arg_slot, temp_reg, extra_slot_offset));
2564 }
2565 
2566 
2567 void MacroAssembler::biased_locking_enter(Register obj_reg, Register mark_reg,
2568                                           Register temp_reg,
2569                                           Label& done, Label* slow_case,
2570                                           BiasedLockingCounters* counters) {
2571   assert(UseBiasedLocking, "why call this otherwise?");
2572 
2573   if (PrintBiasedLockingStatistics) {
2574     assert_different_registers(obj_reg, mark_reg, temp_reg, O7);
2575     if (counters == NULL)
2576       counters = BiasedLocking::counters();
2577   }
2578 
2579   Label cas_label;
2580 
2581   // Biased locking
2582   // See whether the lock is currently biased toward our thread and
2583   // whether the epoch is still valid
2584   // Note that the runtime guarantees sufficient alignment of JavaThread
2585   // pointers to allow age to be placed into low bits
2586   assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
2587   and3(mark_reg, markOopDesc::biased_lock_mask_in_place, temp_reg);
2588   cmp_and_brx_short(temp_reg, markOopDesc::biased_lock_pattern, Assembler::notEqual, Assembler::pn, cas_label);
2589 
2590   load_klass(obj_reg, temp_reg);
2591   ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg);
2592   or3(G2_thread, temp_reg, temp_reg);
2593   xor3(mark_reg, temp_reg, temp_reg);
2594   andcc(temp_reg, ~((int) markOopDesc::age_mask_in_place), temp_reg);
2595   if (counters != NULL) {
2596     cond_inc(Assembler::equal, (address) counters->biased_lock_entry_count_addr(), mark_reg, temp_reg);
2597     // Reload mark_reg as we may need it later
2598     ld_ptr(Address(obj_reg, oopDesc::mark_offset_in_bytes()), mark_reg);
2599   }
2600   brx(Assembler::equal, true, Assembler::pt, done);
2601   delayed()->nop();
2602 
2603   Label try_revoke_bias;
2604   Label try_rebias;
2605   Address mark_addr = Address(obj_reg, oopDesc::mark_offset_in_bytes());
2606   assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2607 
2608   // At this point we know that the header has the bias pattern and
2609   // that we are not the bias owner in the current epoch. We need to
2610   // figure out more details about the state of the header in order to
2611   // know what operations can be legally performed on the object's
2612   // header.
2613 
2614   // If the low three bits in the xor result aren't clear, that means
2615   // the prototype header is no longer biased and we have to revoke
2616   // the bias on this object.
2617   btst(markOopDesc::biased_lock_mask_in_place, temp_reg);
2618   brx(Assembler::notZero, false, Assembler::pn, try_revoke_bias);
2619 
2620   // Biasing is still enabled for this data type. See whether the
2621   // epoch of the current bias is still valid, meaning that the epoch
2622   // bits of the mark word are equal to the epoch bits of the
2623   // prototype header. (Note that the prototype header's epoch bits
2624   // only change at a safepoint.) If not, attempt to rebias the object
2625   // toward the current thread. Note that we must be absolutely sure
2626   // that the current epoch is invalid in order to do this because
2627   // otherwise the manipulations it performs on the mark word are
2628   // illegal.
2629   delayed()->btst(markOopDesc::epoch_mask_in_place, temp_reg);
2630   brx(Assembler::notZero, false, Assembler::pn, try_rebias);
2631 
2632   // The epoch of the current bias is still valid but we know nothing
2633   // about the owner; it might be set or it might be clear. Try to
2634   // acquire the bias of the object using an atomic operation. If this
2635   // fails we will go in to the runtime to revoke the object's bias.
2636   // Note that we first construct the presumed unbiased header so we
2637   // don't accidentally blow away another thread's valid bias.
2638   delayed()->and3(mark_reg,
2639                   markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place,
2640                   mark_reg);
2641   or3(G2_thread, mark_reg, temp_reg);
2642   cas_ptr(mark_addr.base(), mark_reg, temp_reg);
2643   // If the biasing toward our thread failed, this means that
2644   // another thread succeeded in biasing it toward itself and we
2645   // need to revoke that bias. The revocation will occur in the
2646   // interpreter runtime in the slow case.
2647   cmp(mark_reg, temp_reg);
2648   if (counters != NULL) {
2649     cond_inc(Assembler::zero, (address) counters->anonymously_biased_lock_entry_count_addr(), mark_reg, temp_reg);
2650   }
2651   if (slow_case != NULL) {
2652     brx(Assembler::notEqual, true, Assembler::pn, *slow_case);
2653     delayed()->nop();
2654   }
2655   ba_short(done);
2656 
2657   bind(try_rebias);
2658   // At this point we know the epoch has expired, meaning that the
2659   // current "bias owner", if any, is actually invalid. Under these
2660   // circumstances _only_, we are allowed to use the current header's
2661   // value as the comparison value when doing the cas to acquire the
2662   // bias in the current epoch. In other words, we allow transfer of
2663   // the bias from one thread to another directly in this situation.
2664   //
2665   // FIXME: due to a lack of registers we currently blow away the age
2666   // bits in this situation. Should attempt to preserve them.
2667   load_klass(obj_reg, temp_reg);
2668   ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg);
2669   or3(G2_thread, temp_reg, temp_reg);
2670   cas_ptr(mark_addr.base(), mark_reg, temp_reg);
2671   // If the biasing toward our thread failed, this means that
2672   // another thread succeeded in biasing it toward itself and we
2673   // need to revoke that bias. The revocation will occur in the
2674   // interpreter runtime in the slow case.
2675   cmp(mark_reg, temp_reg);
2676   if (counters != NULL) {
2677     cond_inc(Assembler::zero, (address) counters->rebiased_lock_entry_count_addr(), mark_reg, temp_reg);
2678   }
2679   if (slow_case != NULL) {
2680     brx(Assembler::notEqual, true, Assembler::pn, *slow_case);
2681     delayed()->nop();
2682   }
2683   ba_short(done);
2684 
2685   bind(try_revoke_bias);
2686   // The prototype mark in the klass doesn't have the bias bit set any
2687   // more, indicating that objects of this data type are not supposed
2688   // to be biased any more. We are going to try to reset the mark of
2689   // this object to the prototype value and fall through to the
2690   // CAS-based locking scheme. Note that if our CAS fails, it means
2691   // that another thread raced us for the privilege of revoking the
2692   // bias of this particular object, so it's okay to continue in the
2693   // normal locking code.
2694   //
2695   // FIXME: due to a lack of registers we currently blow away the age
2696   // bits in this situation. Should attempt to preserve them.
2697   load_klass(obj_reg, temp_reg);
2698   ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg);
2699   cas_ptr(mark_addr.base(), mark_reg, temp_reg);
2700   // Fall through to the normal CAS-based lock, because no matter what
2701   // the result of the above CAS, some thread must have succeeded in
2702   // removing the bias bit from the object's header.
2703   if (counters != NULL) {
2704     cmp(mark_reg, temp_reg);
2705     cond_inc(Assembler::zero, (address) counters->revoked_lock_entry_count_addr(), mark_reg, temp_reg);
2706   }
2707 
2708   bind(cas_label);
2709 }
2710 
2711 void MacroAssembler::biased_locking_exit (Address mark_addr, Register temp_reg, Label& done,
2712                                           bool allow_delay_slot_filling) {
2713   // Check for biased locking unlock case, which is a no-op
2714   // Note: we do not have to check the thread ID for two reasons.
2715   // First, the interpreter checks for IllegalMonitorStateException at
2716   // a higher level. Second, if the bias was revoked while we held the
2717   // lock, the object could not be rebiased toward another thread, so
2718   // the bias bit would be clear.
2719   ld_ptr(mark_addr, temp_reg);
2720   and3(temp_reg, markOopDesc::biased_lock_mask_in_place, temp_reg);
2721   cmp(temp_reg, markOopDesc::biased_lock_pattern);
2722   brx(Assembler::equal, allow_delay_slot_filling, Assembler::pt, done);
2723   delayed();
2724   if (!allow_delay_slot_filling) {
2725     nop();
2726   }
2727 }
2728 
2729 
2730 // compiler_lock_object() and compiler_unlock_object() are direct transliterations
2731 // of i486.ad fast_lock() and fast_unlock().  See those methods for detailed comments.
2732 // The code could be tightened up considerably.
2733 //
2734 // box->dhw disposition - post-conditions at DONE_LABEL.
2735 // -   Successful inflated lock:  box->dhw != 0.
2736 //     Any non-zero value suffices.
2737 //     Consider G2_thread, rsp, boxReg, or unused_mark()
2738 // -   Successful Stack-lock: box->dhw == mark.
2739 //     box->dhw must contain the displaced mark word value
2740 // -   Failure -- icc.ZFlag == 0 and box->dhw is undefined.
2741 //     The slow-path fast_enter() and slow_enter() operators
2742 //     are responsible for setting box->dhw = NonZero (typically ::unused_mark).
2743 // -   Biased: box->dhw is undefined
2744 //
2745 // SPARC refworkload performance - specifically jetstream and scimark - are
2746 // extremely sensitive to the size of the code emitted by compiler_lock_object
2747 // and compiler_unlock_object.  Critically, the key factor is code size, not path
2748 // length.  (Simply experiments to pad CLO with unexecuted NOPs demonstrte the
2749 // effect).
2750 
2751 
2752 void MacroAssembler::compiler_lock_object(Register Roop, Register Rmark,
2753                                           Register Rbox, Register Rscratch,
2754                                           BiasedLockingCounters* counters,
2755                                           bool try_bias) {
2756    Address mark_addr(Roop, oopDesc::mark_offset_in_bytes());
2757 
2758    verify_oop(Roop);
2759    Label done ;
2760 
2761    if (counters != NULL) {
2762      inc_counter((address) counters->total_entry_count_addr(), Rmark, Rscratch);
2763    }
2764 
2765    if (EmitSync & 1) {
2766      mov(3, Rscratch);
2767      st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2768      cmp(SP, G0);
2769      return ;
2770    }
2771 
2772    if (EmitSync & 2) {
2773 
2774      // Fetch object's markword
2775      ld_ptr(mark_addr, Rmark);
2776 
2777      if (try_bias) {
2778         biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters);
2779      }
2780 
2781      // Save Rbox in Rscratch to be used for the cas operation
2782      mov(Rbox, Rscratch);
2783 
2784      // set Rmark to markOop | markOopDesc::unlocked_value
2785      or3(Rmark, markOopDesc::unlocked_value, Rmark);
2786 
2787      // Initialize the box.  (Must happen before we update the object mark!)
2788      st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes());
2789 
2790      // compare object markOop with Rmark and if equal exchange Rscratch with object markOop
2791      assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2792      cas_ptr(mark_addr.base(), Rmark, Rscratch);
2793 
2794      // if compare/exchange succeeded we found an unlocked object and we now have locked it
2795      // hence we are done
2796      cmp(Rmark, Rscratch);
2797 #ifdef _LP64
2798      sub(Rscratch, STACK_BIAS, Rscratch);
2799 #endif
2800      brx(Assembler::equal, false, Assembler::pt, done);
2801      delayed()->sub(Rscratch, SP, Rscratch);  //pull next instruction into delay slot
2802 
2803      // we did not find an unlocked object so see if this is a recursive case
2804      // sub(Rscratch, SP, Rscratch);
2805      assert(os::vm_page_size() > 0xfff, "page size too small - change the constant");
2806      andcc(Rscratch, 0xfffff003, Rscratch);
2807      st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2808      bind (done);
2809      return ;
2810    }
2811 
2812    Label Egress ;
2813 
2814    if (EmitSync & 256) {
2815       Label IsInflated ;
2816 
2817       ld_ptr(mark_addr, Rmark);           // fetch obj->mark
2818       // Triage: biased, stack-locked, neutral, inflated
2819       if (try_bias) {
2820         biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters);
2821         // Invariant: if control reaches this point in the emitted stream
2822         // then Rmark has not been modified.
2823       }
2824 
2825       // Store mark into displaced mark field in the on-stack basic-lock "box"
2826       // Critically, this must happen before the CAS
2827       // Maximize the ST-CAS distance to minimize the ST-before-CAS penalty.
2828       st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes());
2829       andcc(Rmark, 2, G0);
2830       brx(Assembler::notZero, false, Assembler::pn, IsInflated);
2831       delayed()->
2832 
2833       // Try stack-lock acquisition.
2834       // Beware: the 1st instruction is in a delay slot
2835       mov(Rbox,  Rscratch);
2836       or3(Rmark, markOopDesc::unlocked_value, Rmark);
2837       assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2838       cas_ptr(mark_addr.base(), Rmark, Rscratch);
2839       cmp(Rmark, Rscratch);
2840       brx(Assembler::equal, false, Assembler::pt, done);
2841       delayed()->sub(Rscratch, SP, Rscratch);
2842 
2843       // Stack-lock attempt failed - check for recursive stack-lock.
2844       // See the comments below about how we might remove this case.
2845 #ifdef _LP64
2846       sub(Rscratch, STACK_BIAS, Rscratch);
2847 #endif
2848       assert(os::vm_page_size() > 0xfff, "page size too small - change the constant");
2849       andcc(Rscratch, 0xfffff003, Rscratch);
2850       br(Assembler::always, false, Assembler::pt, done);
2851       delayed()-> st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2852 
2853       bind(IsInflated);
2854       if (EmitSync & 64) {
2855          // If m->owner != null goto IsLocked
2856          // Pessimistic form: Test-and-CAS vs CAS
2857          // The optimistic form avoids RTS->RTO cache line upgrades.
2858          ld_ptr(Rmark, ObjectMonitor::owner_offset_in_bytes() - 2, Rscratch);
2859          andcc(Rscratch, Rscratch, G0);
2860          brx(Assembler::notZero, false, Assembler::pn, done);
2861          delayed()->nop();
2862          // m->owner == null : it's unlocked.
2863       }
2864 
2865       // Try to CAS m->owner from null to Self
2866       // Invariant: if we acquire the lock then _recursions should be 0.
2867       add(Rmark, ObjectMonitor::owner_offset_in_bytes()-2, Rmark);
2868       mov(G2_thread, Rscratch);
2869       cas_ptr(Rmark, G0, Rscratch);
2870       cmp(Rscratch, G0);
2871       // Intentional fall-through into done
2872    } else {
2873       // Aggressively avoid the Store-before-CAS penalty
2874       // Defer the store into box->dhw until after the CAS
2875       Label IsInflated, Recursive ;
2876 
2877 // Anticipate CAS -- Avoid RTS->RTO upgrade
2878 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads);
2879 
2880       ld_ptr(mark_addr, Rmark);           // fetch obj->mark
2881       // Triage: biased, stack-locked, neutral, inflated
2882 
2883       if (try_bias) {
2884         biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters);
2885         // Invariant: if control reaches this point in the emitted stream
2886         // then Rmark has not been modified.
2887       }
2888       andcc(Rmark, 2, G0);
2889       brx(Assembler::notZero, false, Assembler::pn, IsInflated);
2890       delayed()->                         // Beware - dangling delay-slot
2891 
2892       // Try stack-lock acquisition.
2893       // Transiently install BUSY (0) encoding in the mark word.
2894       // if the CAS of 0 into the mark was successful then we execute:
2895       //   ST box->dhw  = mark   -- save fetched mark in on-stack basiclock box
2896       //   ST obj->mark = box    -- overwrite transient 0 value
2897       // This presumes TSO, of course.
2898 
2899       mov(0, Rscratch);
2900       or3(Rmark, markOopDesc::unlocked_value, Rmark);
2901       assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2902       cas_ptr(mark_addr.base(), Rmark, Rscratch);
2903 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads);
2904       cmp(Rscratch, Rmark);
2905       brx(Assembler::notZero, false, Assembler::pn, Recursive);
2906       delayed()->st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes());
2907       if (counters != NULL) {
2908         cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch);
2909       }
2910       ba(done);
2911       delayed()->st_ptr(Rbox, mark_addr);
2912 
2913       bind(Recursive);
2914       // Stack-lock attempt failed - check for recursive stack-lock.
2915       // Tests show that we can remove the recursive case with no impact
2916       // on refworkload 0.83.  If we need to reduce the size of the code
2917       // emitted by compiler_lock_object() the recursive case is perfect
2918       // candidate.
2919       //
2920       // A more extreme idea is to always inflate on stack-lock recursion.
2921       // This lets us eliminate the recursive checks in compiler_lock_object
2922       // and compiler_unlock_object and the (box->dhw == 0) encoding.
2923       // A brief experiment - requiring changes to synchronizer.cpp, interpreter,
2924       // and showed a performance *increase*.  In the same experiment I eliminated
2925       // the fast-path stack-lock code from the interpreter and always passed
2926       // control to the "slow" operators in synchronizer.cpp.
2927 
2928       // RScratch contains the fetched obj->mark value from the failed CAS.
2929 #ifdef _LP64
2930       sub(Rscratch, STACK_BIAS, Rscratch);
2931 #endif
2932       sub(Rscratch, SP, Rscratch);
2933       assert(os::vm_page_size() > 0xfff, "page size too small - change the constant");
2934       andcc(Rscratch, 0xfffff003, Rscratch);
2935       if (counters != NULL) {
2936         // Accounting needs the Rscratch register
2937         st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2938         cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch);
2939         ba_short(done);
2940       } else {
2941         ba(done);
2942         delayed()->st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2943       }
2944 
2945       bind   (IsInflated);
2946       if (EmitSync & 64) {
2947          // If m->owner != null goto IsLocked
2948          // Test-and-CAS vs CAS
2949          // Pessimistic form avoids futile (doomed) CAS attempts
2950          // The optimistic form avoids RTS->RTO cache line upgrades.
2951          ld_ptr(Rmark, ObjectMonitor::owner_offset_in_bytes() - 2, Rscratch);
2952          andcc(Rscratch, Rscratch, G0);
2953          brx(Assembler::notZero, false, Assembler::pn, done);
2954          delayed()->nop();
2955          // m->owner == null : it's unlocked.
2956       }
2957 
2958       // Try to CAS m->owner from null to Self
2959       // Invariant: if we acquire the lock then _recursions should be 0.
2960       add(Rmark, ObjectMonitor::owner_offset_in_bytes()-2, Rmark);
2961       mov(G2_thread, Rscratch);
2962       cas_ptr(Rmark, G0, Rscratch);
2963       cmp(Rscratch, G0);
2964       // ST box->displaced_header = NonZero.
2965       // Any non-zero value suffices:
2966       //    unused_mark(), G2_thread, RBox, RScratch, rsp, etc.
2967       st_ptr(Rbox, Rbox, BasicLock::displaced_header_offset_in_bytes());
2968       // Intentional fall-through into done
2969    }
2970 
2971    bind   (done);
2972 }
2973 
2974 void MacroAssembler::compiler_unlock_object(Register Roop, Register Rmark,
2975                                             Register Rbox, Register Rscratch,
2976                                             bool try_bias) {
2977    Address mark_addr(Roop, oopDesc::mark_offset_in_bytes());
2978 
2979    Label done ;
2980 
2981    if (EmitSync & 4) {
2982      cmp(SP, G0);
2983      return ;
2984    }
2985 
2986    if (EmitSync & 8) {
2987      if (try_bias) {
2988         biased_locking_exit(mark_addr, Rscratch, done);
2989      }
2990 
2991      // Test first if it is a fast recursive unlock
2992      ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rmark);
2993      br_null_short(Rmark, Assembler::pt, done);
2994 
2995      // Check if it is still a light weight lock, this is is true if we see
2996      // the stack address of the basicLock in the markOop of the object
2997      assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2998      cas_ptr(mark_addr.base(), Rbox, Rmark);
2999      ba(done);
3000      delayed()->cmp(Rbox, Rmark);
3001      bind(done);
3002      return ;
3003    }
3004 
3005    // Beware ... If the aggregate size of the code emitted by CLO and CUO is
3006    // is too large performance rolls abruptly off a cliff.
3007    // This could be related to inlining policies, code cache management, or
3008    // I$ effects.
3009    Label LStacked ;
3010 
3011    if (try_bias) {
3012       // TODO: eliminate redundant LDs of obj->mark
3013       biased_locking_exit(mark_addr, Rscratch, done);
3014    }
3015 
3016    ld_ptr(Roop, oopDesc::mark_offset_in_bytes(), Rmark);
3017    ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rscratch);
3018    andcc(Rscratch, Rscratch, G0);
3019    brx(Assembler::zero, false, Assembler::pn, done);
3020    delayed()->nop();      // consider: relocate fetch of mark, above, into this DS
3021    andcc(Rmark, 2, G0);
3022    brx(Assembler::zero, false, Assembler::pt, LStacked);
3023    delayed()->nop();
3024 
3025    // It's inflated
3026    // Conceptually we need a #loadstore|#storestore "release" MEMBAR before
3027    // the ST of 0 into _owner which releases the lock.  This prevents loads
3028    // and stores within the critical section from reordering (floating)
3029    // past the store that releases the lock.  But TSO is a strong memory model
3030    // and that particular flavor of barrier is a noop, so we can safely elide it.
3031    // Note that we use 1-0 locking by default for the inflated case.  We
3032    // close the resultant (and rare) race by having contented threads in
3033    // monitorenter periodically poll _owner.
3034    ld_ptr(Rmark, ObjectMonitor::owner_offset_in_bytes() - 2, Rscratch);
3035    ld_ptr(Rmark, ObjectMonitor::recursions_offset_in_bytes() - 2, Rbox);
3036    xor3(Rscratch, G2_thread, Rscratch);
3037    orcc(Rbox, Rscratch, Rbox);
3038    brx(Assembler::notZero, false, Assembler::pn, done);
3039    delayed()->
3040    ld_ptr(Rmark, ObjectMonitor::EntryList_offset_in_bytes() - 2, Rscratch);
3041    ld_ptr(Rmark, ObjectMonitor::cxq_offset_in_bytes() - 2, Rbox);
3042    orcc(Rbox, Rscratch, G0);
3043    if (EmitSync & 65536) {
3044       Label LSucc ;
3045       brx(Assembler::notZero, false, Assembler::pn, LSucc);
3046       delayed()->nop();
3047       ba(done);
3048       delayed()->st_ptr(G0, Rmark, ObjectMonitor::owner_offset_in_bytes() - 2);
3049 
3050       bind(LSucc);
3051       st_ptr(G0, Rmark, ObjectMonitor::owner_offset_in_bytes() - 2);
3052       if (os::is_MP()) { membar (StoreLoad); }
3053       ld_ptr(Rmark, ObjectMonitor::succ_offset_in_bytes() - 2, Rscratch);
3054       andcc(Rscratch, Rscratch, G0);
3055       brx(Assembler::notZero, false, Assembler::pt, done);
3056       delayed()->andcc(G0, G0, G0);
3057       add(Rmark, ObjectMonitor::owner_offset_in_bytes()-2, Rmark);
3058       mov(G2_thread, Rscratch);
3059       cas_ptr(Rmark, G0, Rscratch);
3060       // invert icc.zf and goto done
3061       br_notnull(Rscratch, false, Assembler::pt, done);
3062       delayed()->cmp(G0, G0);
3063       ba(done);
3064       delayed()->cmp(G0, 1);
3065    } else {
3066       brx(Assembler::notZero, false, Assembler::pn, done);
3067       delayed()->nop();
3068       ba(done);
3069       delayed()->st_ptr(G0, Rmark, ObjectMonitor::owner_offset_in_bytes() - 2);
3070    }
3071 
3072    bind   (LStacked);
3073    // Consider: we could replace the expensive CAS in the exit
3074    // path with a simple ST of the displaced mark value fetched from
3075    // the on-stack basiclock box.  That admits a race where a thread T2
3076    // in the slow lock path -- inflating with monitor M -- could race a
3077    // thread T1 in the fast unlock path, resulting in a missed wakeup for T2.
3078    // More precisely T1 in the stack-lock unlock path could "stomp" the
3079    // inflated mark value M installed by T2, resulting in an orphan
3080    // object monitor M and T2 becoming stranded.  We can remedy that situation
3081    // by having T2 periodically poll the object's mark word using timed wait
3082    // operations.  If T2 discovers that a stomp has occurred it vacates
3083    // the monitor M and wakes any other threads stranded on the now-orphan M.
3084    // In addition the monitor scavenger, which performs deflation,
3085    // would also need to check for orpan monitors and stranded threads.
3086    //
3087    // Finally, inflation is also used when T2 needs to assign a hashCode
3088    // to O and O is stack-locked by T1.  The "stomp" race could cause
3089    // an assigned hashCode value to be lost.  We can avoid that condition
3090    // and provide the necessary hashCode stability invariants by ensuring
3091    // that hashCode generation is idempotent between copying GCs.
3092    // For example we could compute the hashCode of an object O as
3093    // O's heap address XOR some high quality RNG value that is refreshed
3094    // at GC-time.  The monitor scavenger would install the hashCode
3095    // found in any orphan monitors.  Again, the mechanism admits a
3096    // lost-update "stomp" WAW race but detects and recovers as needed.
3097    //
3098    // A prototype implementation showed excellent results, although
3099    // the scavenger and timeout code was rather involved.
3100 
3101    cas_ptr(mark_addr.base(), Rbox, Rscratch);
3102    cmp(Rbox, Rscratch);
3103    // Intentional fall through into done ...
3104 
3105    bind(done);
3106 }
3107 
3108 
3109 
3110 void MacroAssembler::print_CPU_state() {
3111   // %%%%% need to implement this
3112 }
3113 
3114 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
3115   // %%%%% need to implement this
3116 }
3117 
3118 void MacroAssembler::push_IU_state() {
3119   // %%%%% need to implement this
3120 }
3121 
3122 
3123 void MacroAssembler::pop_IU_state() {
3124   // %%%%% need to implement this
3125 }
3126 
3127 
3128 void MacroAssembler::push_FPU_state() {
3129   // %%%%% need to implement this
3130 }
3131 
3132 
3133 void MacroAssembler::pop_FPU_state() {
3134   // %%%%% need to implement this
3135 }
3136 
3137 
3138 void MacroAssembler::push_CPU_state() {
3139   // %%%%% need to implement this
3140 }
3141 
3142 
3143 void MacroAssembler::pop_CPU_state() {
3144   // %%%%% need to implement this
3145 }
3146 
3147 
3148 
3149 void MacroAssembler::verify_tlab() {
3150 #ifdef ASSERT
3151   if (UseTLAB && VerifyOops) {
3152     Label next, next2, ok;
3153     Register t1 = L0;
3154     Register t2 = L1;
3155     Register t3 = L2;
3156 
3157     save_frame(0);
3158     ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1);
3159     ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t2);
3160     or3(t1, t2, t3);
3161     cmp_and_br_short(t1, t2, Assembler::greaterEqual, Assembler::pn, next);
3162     STOP("assert(top >= start)");
3163     should_not_reach_here();
3164 
3165     bind(next);
3166     ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1);
3167     ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t2);
3168     or3(t3, t2, t3);
3169     cmp_and_br_short(t1, t2, Assembler::lessEqual, Assembler::pn, next2);
3170     STOP("assert(top <= end)");
3171     should_not_reach_here();
3172 
3173     bind(next2);
3174     and3(t3, MinObjAlignmentInBytesMask, t3);
3175     cmp_and_br_short(t3, 0, Assembler::lessEqual, Assembler::pn, ok);
3176     STOP("assert(aligned)");
3177     should_not_reach_here();
3178 
3179     bind(ok);
3180     restore();
3181   }
3182 #endif
3183 }
3184 
3185 
3186 void MacroAssembler::eden_allocate(
3187   Register obj,                        // result: pointer to object after successful allocation
3188   Register var_size_in_bytes,          // object size in bytes if unknown at compile time; invalid otherwise
3189   int      con_size_in_bytes,          // object size in bytes if   known at compile time
3190   Register t1,                         // temp register
3191   Register t2,                         // temp register
3192   Label&   slow_case                   // continuation point if fast allocation fails
3193 ){
3194   // make sure arguments make sense
3195   assert_different_registers(obj, var_size_in_bytes, t1, t2);
3196   assert(0 <= con_size_in_bytes && Assembler::is_simm13(con_size_in_bytes), "illegal object size");
3197   assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment");
3198 
3199   if (CMSIncrementalMode || !Universe::heap()->supports_inline_contig_alloc()) {
3200     // No allocation in the shared eden.
3201     ba(slow_case);
3202     delayed()->nop();
3203   } else {
3204     // get eden boundaries
3205     // note: we need both top & top_addr!
3206     const Register top_addr = t1;
3207     const Register end      = t2;
3208 
3209     CollectedHeap* ch = Universe::heap();
3210     set((intx)ch->top_addr(), top_addr);
3211     intx delta = (intx)ch->end_addr() - (intx)ch->top_addr();
3212     ld_ptr(top_addr, delta, end);
3213     ld_ptr(top_addr, 0, obj);
3214 
3215     // try to allocate
3216     Label retry;
3217     bind(retry);
3218 #ifdef ASSERT
3219     // make sure eden top is properly aligned
3220     {
3221       Label L;
3222       btst(MinObjAlignmentInBytesMask, obj);
3223       br(Assembler::zero, false, Assembler::pt, L);
3224       delayed()->nop();
3225       STOP("eden top is not properly aligned");
3226       bind(L);
3227     }
3228 #endif // ASSERT
3229     const Register free = end;
3230     sub(end, obj, free);                                   // compute amount of free space
3231     if (var_size_in_bytes->is_valid()) {
3232       // size is unknown at compile time
3233       cmp(free, var_size_in_bytes);
3234       br(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case
3235       delayed()->add(obj, var_size_in_bytes, end);
3236     } else {
3237       // size is known at compile time
3238       cmp(free, con_size_in_bytes);
3239       br(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case
3240       delayed()->add(obj, con_size_in_bytes, end);
3241     }
3242     // Compare obj with the value at top_addr; if still equal, swap the value of
3243     // end with the value at top_addr. If not equal, read the value at top_addr
3244     // into end.
3245     cas_ptr(top_addr, obj, end);
3246     // if someone beat us on the allocation, try again, otherwise continue
3247     cmp(obj, end);
3248     brx(Assembler::notEqual, false, Assembler::pn, retry);
3249     delayed()->mov(end, obj);                              // nop if successfull since obj == end
3250 
3251 #ifdef ASSERT
3252     // make sure eden top is properly aligned
3253     {
3254       Label L;
3255       const Register top_addr = t1;
3256 
3257       set((intx)ch->top_addr(), top_addr);
3258       ld_ptr(top_addr, 0, top_addr);
3259       btst(MinObjAlignmentInBytesMask, top_addr);
3260       br(Assembler::zero, false, Assembler::pt, L);
3261       delayed()->nop();
3262       STOP("eden top is not properly aligned");
3263       bind(L);
3264     }
3265 #endif // ASSERT
3266   }
3267 }
3268 
3269 
3270 void MacroAssembler::tlab_allocate(
3271   Register obj,                        // result: pointer to object after successful allocation
3272   Register var_size_in_bytes,          // object size in bytes if unknown at compile time; invalid otherwise
3273   int      con_size_in_bytes,          // object size in bytes if   known at compile time
3274   Register t1,                         // temp register
3275   Label&   slow_case                   // continuation point if fast allocation fails
3276 ){
3277   // make sure arguments make sense
3278   assert_different_registers(obj, var_size_in_bytes, t1);
3279   assert(0 <= con_size_in_bytes && is_simm13(con_size_in_bytes), "illegal object size");
3280   assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment");
3281 
3282   const Register free  = t1;
3283 
3284   verify_tlab();
3285 
3286   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), obj);
3287 
3288   // calculate amount of free space
3289   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), free);
3290   sub(free, obj, free);
3291 
3292   Label done;
3293   if (var_size_in_bytes == noreg) {
3294     cmp(free, con_size_in_bytes);
3295   } else {
3296     cmp(free, var_size_in_bytes);
3297   }
3298   br(Assembler::less, false, Assembler::pn, slow_case);
3299   // calculate the new top pointer
3300   if (var_size_in_bytes == noreg) {
3301     delayed()->add(obj, con_size_in_bytes, free);
3302   } else {
3303     delayed()->add(obj, var_size_in_bytes, free);
3304   }
3305 
3306   bind(done);
3307 
3308 #ifdef ASSERT
3309   // make sure new free pointer is properly aligned
3310   {
3311     Label L;
3312     btst(MinObjAlignmentInBytesMask, free);
3313     br(Assembler::zero, false, Assembler::pt, L);
3314     delayed()->nop();
3315     STOP("updated TLAB free is not properly aligned");
3316     bind(L);
3317   }
3318 #endif // ASSERT
3319 
3320   // update the tlab top pointer
3321   st_ptr(free, G2_thread, in_bytes(JavaThread::tlab_top_offset()));
3322   verify_tlab();
3323 }
3324 
3325 
3326 void MacroAssembler::tlab_refill(Label& retry, Label& try_eden, Label& slow_case) {
3327   Register top = O0;
3328   Register t1 = G1;
3329   Register t2 = G3;
3330   Register t3 = O1;
3331   assert_different_registers(top, t1, t2, t3, G4, G5 /* preserve G4 and G5 */);
3332   Label do_refill, discard_tlab;
3333 
3334   if (CMSIncrementalMode || !Universe::heap()->supports_inline_contig_alloc()) {
3335     // No allocation in the shared eden.
3336     ba(slow_case);
3337     delayed()->nop();
3338   }
3339 
3340   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), top);
3341   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t1);
3342   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_refill_waste_limit_offset()), t2);
3343 
3344   // calculate amount of free space
3345   sub(t1, top, t1);
3346   srl_ptr(t1, LogHeapWordSize, t1);
3347 
3348   // Retain tlab and allocate object in shared space if
3349   // the amount free in the tlab is too large to discard.
3350   cmp(t1, t2);
3351   brx(Assembler::lessEqual, false, Assembler::pt, discard_tlab);
3352 
3353   // increment waste limit to prevent getting stuck on this slow path
3354   delayed()->add(t2, ThreadLocalAllocBuffer::refill_waste_limit_increment(), t2);
3355   st_ptr(t2, G2_thread, in_bytes(JavaThread::tlab_refill_waste_limit_offset()));
3356   if (TLABStats) {
3357     // increment number of slow_allocations
3358     ld(G2_thread, in_bytes(JavaThread::tlab_slow_allocations_offset()), t2);
3359     add(t2, 1, t2);
3360     stw(t2, G2_thread, in_bytes(JavaThread::tlab_slow_allocations_offset()));
3361   }
3362   ba(try_eden);
3363   delayed()->nop();
3364 
3365   bind(discard_tlab);
3366   if (TLABStats) {
3367     // increment number of refills
3368     ld(G2_thread, in_bytes(JavaThread::tlab_number_of_refills_offset()), t2);
3369     add(t2, 1, t2);
3370     stw(t2, G2_thread, in_bytes(JavaThread::tlab_number_of_refills_offset()));
3371     // accumulate wastage
3372     ld(G2_thread, in_bytes(JavaThread::tlab_fast_refill_waste_offset()), t2);
3373     add(t2, t1, t2);
3374     stw(t2, G2_thread, in_bytes(JavaThread::tlab_fast_refill_waste_offset()));
3375   }
3376 
3377   // if tlab is currently allocated (top or end != null) then
3378   // fill [top, end + alignment_reserve) with array object
3379   br_null_short(top, Assembler::pn, do_refill);
3380 
3381   set((intptr_t)markOopDesc::prototype()->copy_set_hash(0x2), t2);
3382   st_ptr(t2, top, oopDesc::mark_offset_in_bytes()); // set up the mark word
3383   // set klass to intArrayKlass
3384   sub(t1, typeArrayOopDesc::header_size(T_INT), t1);
3385   add(t1, ThreadLocalAllocBuffer::alignment_reserve(), t1);
3386   sll_ptr(t1, log2_intptr(HeapWordSize/sizeof(jint)), t1);
3387   st(t1, top, arrayOopDesc::length_offset_in_bytes());
3388   set((intptr_t)Universe::intArrayKlassObj_addr(), t2);
3389   ld_ptr(t2, 0, t2);
3390   // store klass last.  concurrent gcs assumes klass length is valid if
3391   // klass field is not null.
3392   store_klass(t2, top);
3393   verify_oop(top);
3394 
3395   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t1);
3396   sub(top, t1, t1); // size of tlab's allocated portion
3397   incr_allocated_bytes(t1, t2, t3);
3398 
3399   // refill the tlab with an eden allocation
3400   bind(do_refill);
3401   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_size_offset()), t1);
3402   sll_ptr(t1, LogHeapWordSize, t1);
3403   // allocate new tlab, address returned in top
3404   eden_allocate(top, t1, 0, t2, t3, slow_case);
3405 
3406   st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_start_offset()));
3407   st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_top_offset()));
3408 #ifdef ASSERT
3409   // check that tlab_size (t1) is still valid
3410   {
3411     Label ok;
3412     ld_ptr(G2_thread, in_bytes(JavaThread::tlab_size_offset()), t2);
3413     sll_ptr(t2, LogHeapWordSize, t2);
3414     cmp_and_br_short(t1, t2, Assembler::equal, Assembler::pt, ok);
3415     STOP("assert(t1 == tlab_size)");
3416     should_not_reach_here();
3417 
3418     bind(ok);
3419   }
3420 #endif // ASSERT
3421   add(top, t1, top); // t1 is tlab_size
3422   sub(top, ThreadLocalAllocBuffer::alignment_reserve_in_bytes(), top);
3423   st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_end_offset()));
3424   verify_tlab();
3425   ba(retry);
3426   delayed()->nop();
3427 }
3428 
3429 void MacroAssembler::incr_allocated_bytes(RegisterOrConstant size_in_bytes,
3430                                           Register t1, Register t2) {
3431   // Bump total bytes allocated by this thread
3432   assert(t1->is_global(), "must be global reg"); // so all 64 bits are saved on a context switch
3433   assert_different_registers(size_in_bytes.register_or_noreg(), t1, t2);
3434   // v8 support has gone the way of the dodo
3435   ldx(G2_thread, in_bytes(JavaThread::allocated_bytes_offset()), t1);
3436   add(t1, ensure_simm13_or_reg(size_in_bytes, t2), t1);
3437   stx(t1, G2_thread, in_bytes(JavaThread::allocated_bytes_offset()));
3438 }
3439 
3440 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
3441   switch (cond) {
3442     // Note some conditions are synonyms for others
3443     case Assembler::never:                return Assembler::always;
3444     case Assembler::zero:                 return Assembler::notZero;
3445     case Assembler::lessEqual:            return Assembler::greater;
3446     case Assembler::less:                 return Assembler::greaterEqual;
3447     case Assembler::lessEqualUnsigned:    return Assembler::greaterUnsigned;
3448     case Assembler::lessUnsigned:         return Assembler::greaterEqualUnsigned;
3449     case Assembler::negative:             return Assembler::positive;
3450     case Assembler::overflowSet:          return Assembler::overflowClear;
3451     case Assembler::always:               return Assembler::never;
3452     case Assembler::notZero:              return Assembler::zero;
3453     case Assembler::greater:              return Assembler::lessEqual;
3454     case Assembler::greaterEqual:         return Assembler::less;
3455     case Assembler::greaterUnsigned:      return Assembler::lessEqualUnsigned;
3456     case Assembler::greaterEqualUnsigned: return Assembler::lessUnsigned;
3457     case Assembler::positive:             return Assembler::negative;
3458     case Assembler::overflowClear:        return Assembler::overflowSet;
3459   }
3460 
3461   ShouldNotReachHere(); return Assembler::overflowClear;
3462 }
3463 
3464 void MacroAssembler::cond_inc(Assembler::Condition cond, address counter_ptr,
3465                               Register Rtmp1, Register Rtmp2 /*, Register Rtmp3, Register Rtmp4 */) {
3466   Condition negated_cond = negate_condition(cond);
3467   Label L;
3468   brx(negated_cond, false, Assembler::pt, L);
3469   delayed()->nop();
3470   inc_counter(counter_ptr, Rtmp1, Rtmp2);
3471   bind(L);
3472 }
3473 
3474 void MacroAssembler::inc_counter(address counter_addr, Register Rtmp1, Register Rtmp2) {
3475   AddressLiteral addrlit(counter_addr);
3476   sethi(addrlit, Rtmp1);                 // Move hi22 bits into temporary register.
3477   Address addr(Rtmp1, addrlit.low10());  // Build an address with low10 bits.
3478   ld(addr, Rtmp2);
3479   inc(Rtmp2);
3480   st(Rtmp2, addr);
3481 }
3482 
3483 void MacroAssembler::inc_counter(int* counter_addr, Register Rtmp1, Register Rtmp2) {
3484   inc_counter((address) counter_addr, Rtmp1, Rtmp2);
3485 }
3486 
3487 SkipIfEqual::SkipIfEqual(
3488     MacroAssembler* masm, Register temp, const bool* flag_addr,
3489     Assembler::Condition condition) {
3490   _masm = masm;
3491   AddressLiteral flag(flag_addr);
3492   _masm->sethi(flag, temp);
3493   _masm->ldub(temp, flag.low10(), temp);
3494   _masm->tst(temp);
3495   _masm->br(condition, false, Assembler::pt, _label);
3496   _masm->delayed()->nop();
3497 }
3498 
3499 SkipIfEqual::~SkipIfEqual() {
3500   _masm->bind(_label);
3501 }
3502 
3503 
3504 // Writes to stack successive pages until offset reached to check for
3505 // stack overflow + shadow pages.  This clobbers tsp and scratch.
3506 void MacroAssembler::bang_stack_size(Register Rsize, Register Rtsp,
3507                                      Register Rscratch) {
3508   // Use stack pointer in temp stack pointer
3509   mov(SP, Rtsp);
3510 
3511   // Bang stack for total size given plus stack shadow page size.
3512   // Bang one page at a time because a large size can overflow yellow and
3513   // red zones (the bang will fail but stack overflow handling can't tell that
3514   // it was a stack overflow bang vs a regular segv).
3515   int offset = os::vm_page_size();
3516   Register Roffset = Rscratch;
3517 
3518   Label loop;
3519   bind(loop);
3520   set((-offset)+STACK_BIAS, Rscratch);
3521   st(G0, Rtsp, Rscratch);
3522   set(offset, Roffset);
3523   sub(Rsize, Roffset, Rsize);
3524   cmp(Rsize, G0);
3525   br(Assembler::greater, false, Assembler::pn, loop);
3526   delayed()->sub(Rtsp, Roffset, Rtsp);
3527 
3528   // Bang down shadow pages too.
3529   // The -1 because we already subtracted 1 page.
3530   for (int i = 0; i< StackShadowPages-1; i++) {
3531     set((-i*offset)+STACK_BIAS, Rscratch);
3532     st(G0, Rtsp, Rscratch);
3533   }
3534 }
3535 
3536 ///////////////////////////////////////////////////////////////////////////////////
3537 #if INCLUDE_ALL_GCS
3538 
3539 static address satb_log_enqueue_with_frame = NULL;
3540 static u_char* satb_log_enqueue_with_frame_end = NULL;
3541 
3542 static address satb_log_enqueue_frameless = NULL;
3543 static u_char* satb_log_enqueue_frameless_end = NULL;
3544 
3545 static int EnqueueCodeSize = 128 DEBUG_ONLY( + 256); // Instructions?
3546 
3547 static void generate_satb_log_enqueue(bool with_frame) {
3548   BufferBlob* bb = BufferBlob::create("enqueue_with_frame", EnqueueCodeSize);
3549   CodeBuffer buf(bb);
3550   MacroAssembler masm(&buf);
3551 
3552 #define __ masm.
3553 
3554   address start = __ pc();
3555   Register pre_val;
3556 
3557   Label refill, restart;
3558   if (with_frame) {
3559     __ save_frame(0);
3560     pre_val = I0;  // Was O0 before the save.
3561   } else {
3562     pre_val = O0;
3563   }
3564 
3565   int satb_q_index_byte_offset =
3566     in_bytes(JavaThread::satb_mark_queue_offset() +
3567              PtrQueue::byte_offset_of_index());
3568 
3569   int satb_q_buf_byte_offset =
3570     in_bytes(JavaThread::satb_mark_queue_offset() +
3571              PtrQueue::byte_offset_of_buf());
3572 
3573   assert(in_bytes(PtrQueue::byte_width_of_index()) == sizeof(intptr_t) &&
3574          in_bytes(PtrQueue::byte_width_of_buf()) == sizeof(intptr_t),
3575          "check sizes in assembly below");
3576 
3577   __ bind(restart);
3578 
3579   // Load the index into the SATB buffer. PtrQueue::_index is a size_t
3580   // so ld_ptr is appropriate.
3581   __ ld_ptr(G2_thread, satb_q_index_byte_offset, L0);
3582 
3583   // index == 0?
3584   __ cmp_and_brx_short(L0, G0, Assembler::equal, Assembler::pn, refill);
3585 
3586   __ ld_ptr(G2_thread, satb_q_buf_byte_offset, L1);
3587   __ sub(L0, oopSize, L0);
3588 
3589   __ st_ptr(pre_val, L1, L0);  // [_buf + index] := I0
3590   if (!with_frame) {
3591     // Use return-from-leaf
3592     __ retl();
3593     __ delayed()->st_ptr(L0, G2_thread, satb_q_index_byte_offset);
3594   } else {
3595     // Not delayed.
3596     __ st_ptr(L0, G2_thread, satb_q_index_byte_offset);
3597   }
3598   if (with_frame) {
3599     __ ret();
3600     __ delayed()->restore();
3601   }
3602   __ bind(refill);
3603 
3604   address handle_zero =
3605     CAST_FROM_FN_PTR(address,
3606                      &SATBMarkQueueSet::handle_zero_index_for_thread);
3607   // This should be rare enough that we can afford to save all the
3608   // scratch registers that the calling context might be using.
3609   __ mov(G1_scratch, L0);
3610   __ mov(G3_scratch, L1);
3611   __ mov(G4, L2);
3612   // We need the value of O0 above (for the write into the buffer), so we
3613   // save and restore it.
3614   __ mov(O0, L3);
3615   // Since the call will overwrite O7, we save and restore that, as well.
3616   __ mov(O7, L4);
3617   __ call_VM_leaf(L5, handle_zero, G2_thread);
3618   __ mov(L0, G1_scratch);
3619   __ mov(L1, G3_scratch);
3620   __ mov(L2, G4);
3621   __ mov(L3, O0);
3622   __ br(Assembler::always, /*annul*/false, Assembler::pt, restart);
3623   __ delayed()->mov(L4, O7);
3624 
3625   if (with_frame) {
3626     satb_log_enqueue_with_frame = start;
3627     satb_log_enqueue_with_frame_end = __ pc();
3628   } else {
3629     satb_log_enqueue_frameless = start;
3630     satb_log_enqueue_frameless_end = __ pc();
3631   }
3632 
3633 #undef __
3634 }
3635 
3636 static inline void generate_satb_log_enqueue_if_necessary(bool with_frame) {
3637   if (with_frame) {
3638     if (satb_log_enqueue_with_frame == 0) {
3639       generate_satb_log_enqueue(with_frame);
3640       assert(satb_log_enqueue_with_frame != 0, "postcondition.");
3641       if (G1SATBPrintStubs) {
3642         tty->print_cr("Generated with-frame satb enqueue:");
3643         Disassembler::decode((u_char*)satb_log_enqueue_with_frame,
3644                              satb_log_enqueue_with_frame_end,
3645                              tty);
3646       }
3647     }
3648   } else {
3649     if (satb_log_enqueue_frameless == 0) {
3650       generate_satb_log_enqueue(with_frame);
3651       assert(satb_log_enqueue_frameless != 0, "postcondition.");
3652       if (G1SATBPrintStubs) {
3653         tty->print_cr("Generated frameless satb enqueue:");
3654         Disassembler::decode((u_char*)satb_log_enqueue_frameless,
3655                              satb_log_enqueue_frameless_end,
3656                              tty);
3657       }
3658     }
3659   }
3660 }
3661 
3662 void MacroAssembler::g1_write_barrier_pre(Register obj,
3663                                           Register index,
3664                                           int offset,
3665                                           Register pre_val,
3666                                           Register tmp,
3667                                           bool preserve_o_regs) {
3668   Label filtered;
3669 
3670   if (obj == noreg) {
3671     // We are not loading the previous value so make
3672     // sure that we don't trash the value in pre_val
3673     // with the code below.
3674     assert_different_registers(pre_val, tmp);
3675   } else {
3676     // We will be loading the previous value
3677     // in this code so...
3678     assert(offset == 0 || index == noreg, "choose one");
3679     assert(pre_val == noreg, "check this code");
3680   }
3681 
3682   // Is marking active?
3683   if (in_bytes(PtrQueue::byte_width_of_active()) == 4) {
3684     ld(G2,
3685        in_bytes(JavaThread::satb_mark_queue_offset() +
3686                 PtrQueue::byte_offset_of_active()),
3687        tmp);
3688   } else {
3689     guarantee(in_bytes(PtrQueue::byte_width_of_active()) == 1,
3690               "Assumption");
3691     ldsb(G2,
3692          in_bytes(JavaThread::satb_mark_queue_offset() +
3693                   PtrQueue::byte_offset_of_active()),
3694          tmp);
3695   }
3696 
3697   // Is marking active?
3698   cmp_and_br_short(tmp, G0, Assembler::equal, Assembler::pt, filtered);
3699 
3700   // Do we need to load the previous value?
3701   if (obj != noreg) {
3702     // Load the previous value...
3703     if (index == noreg) {
3704       if (Assembler::is_simm13(offset)) {
3705         load_heap_oop(obj, offset, tmp);
3706       } else {
3707         set(offset, tmp);
3708         load_heap_oop(obj, tmp, tmp);
3709       }
3710     } else {
3711       load_heap_oop(obj, index, tmp);
3712     }
3713     // Previous value has been loaded into tmp
3714     pre_val = tmp;
3715   }
3716 
3717   assert(pre_val != noreg, "must have a real register");
3718 
3719   // Is the previous value null?
3720   cmp_and_brx_short(pre_val, G0, Assembler::equal, Assembler::pt, filtered);
3721 
3722   // OK, it's not filtered, so we'll need to call enqueue.  In the normal
3723   // case, pre_val will be a scratch G-reg, but there are some cases in
3724   // which it's an O-reg.  In the first case, do a normal call.  In the
3725   // latter, do a save here and call the frameless version.
3726 
3727   guarantee(pre_val->is_global() || pre_val->is_out(),
3728             "Or we need to think harder.");
3729 
3730   if (pre_val->is_global() && !preserve_o_regs) {
3731     generate_satb_log_enqueue_if_necessary(true); // with frame
3732 
3733     call(satb_log_enqueue_with_frame);
3734     delayed()->mov(pre_val, O0);
3735   } else {
3736     generate_satb_log_enqueue_if_necessary(false); // frameless
3737 
3738     save_frame(0);
3739     call(satb_log_enqueue_frameless);
3740     delayed()->mov(pre_val->after_save(), O0);
3741     restore();
3742   }
3743 
3744   bind(filtered);
3745 }
3746 
3747 static address dirty_card_log_enqueue = 0;
3748 static u_char* dirty_card_log_enqueue_end = 0;
3749 
3750 // This gets to assume that o0 contains the object address.
3751 static void generate_dirty_card_log_enqueue(jbyte* byte_map_base) {
3752   BufferBlob* bb = BufferBlob::create("dirty_card_enqueue", EnqueueCodeSize*2);
3753   CodeBuffer buf(bb);
3754   MacroAssembler masm(&buf);
3755 #define __ masm.
3756   address start = __ pc();
3757 
3758   Label not_already_dirty, restart, refill;
3759 
3760 #ifdef _LP64
3761   __ srlx(O0, CardTableModRefBS::card_shift, O0);
3762 #else
3763   __ srl(O0, CardTableModRefBS::card_shift, O0);
3764 #endif
3765   AddressLiteral addrlit(byte_map_base);
3766   __ set(addrlit, O1); // O1 := <card table base>
3767   __ ldub(O0, O1, O2); // O2 := [O0 + O1]
3768 
3769   assert(CardTableModRefBS::dirty_card_val() == 0, "otherwise check this code");
3770   __ cmp_and_br_short(O2, G0, Assembler::notEqual, Assembler::pt, not_already_dirty);
3771 
3772   // We didn't take the branch, so we're already dirty: return.
3773   // Use return-from-leaf
3774   __ retl();
3775   __ delayed()->nop();
3776 
3777   // Not dirty.
3778   __ bind(not_already_dirty);
3779 
3780   // Get O0 + O1 into a reg by itself
3781   __ add(O0, O1, O3);
3782 
3783   // First, dirty it.
3784   __ stb(G0, O3, G0);  // [cardPtr] := 0  (i.e., dirty).
3785 
3786   int dirty_card_q_index_byte_offset =
3787     in_bytes(JavaThread::dirty_card_queue_offset() +
3788              PtrQueue::byte_offset_of_index());
3789   int dirty_card_q_buf_byte_offset =
3790     in_bytes(JavaThread::dirty_card_queue_offset() +
3791              PtrQueue::byte_offset_of_buf());
3792   __ bind(restart);
3793 
3794   // Load the index into the update buffer. PtrQueue::_index is
3795   // a size_t so ld_ptr is appropriate here.
3796   __ ld_ptr(G2_thread, dirty_card_q_index_byte_offset, L0);
3797 
3798   // index == 0?
3799   __ cmp_and_brx_short(L0, G0, Assembler::equal, Assembler::pn, refill);
3800 
3801   __ ld_ptr(G2_thread, dirty_card_q_buf_byte_offset, L1);
3802   __ sub(L0, oopSize, L0);
3803 
3804   __ st_ptr(O3, L1, L0);  // [_buf + index] := I0
3805   // Use return-from-leaf
3806   __ retl();
3807   __ delayed()->st_ptr(L0, G2_thread, dirty_card_q_index_byte_offset);
3808 
3809   __ bind(refill);
3810   address handle_zero =
3811     CAST_FROM_FN_PTR(address,
3812                      &DirtyCardQueueSet::handle_zero_index_for_thread);
3813   // This should be rare enough that we can afford to save all the
3814   // scratch registers that the calling context might be using.
3815   __ mov(G1_scratch, L3);
3816   __ mov(G3_scratch, L5);
3817   // We need the value of O3 above (for the write into the buffer), so we
3818   // save and restore it.
3819   __ mov(O3, L6);
3820   // Since the call will overwrite O7, we save and restore that, as well.
3821   __ mov(O7, L4);
3822 
3823   __ call_VM_leaf(L7_thread_cache, handle_zero, G2_thread);
3824   __ mov(L3, G1_scratch);
3825   __ mov(L5, G3_scratch);
3826   __ mov(L6, O3);
3827   __ br(Assembler::always, /*annul*/false, Assembler::pt, restart);
3828   __ delayed()->mov(L4, O7);
3829 
3830   dirty_card_log_enqueue = start;
3831   dirty_card_log_enqueue_end = __ pc();
3832   // XXX Should have a guarantee here about not going off the end!
3833   // Does it already do so?  Do an experiment...
3834 
3835 #undef __
3836 
3837 }
3838 
3839 static inline void
3840 generate_dirty_card_log_enqueue_if_necessary(jbyte* byte_map_base) {
3841   if (dirty_card_log_enqueue == 0) {
3842     generate_dirty_card_log_enqueue(byte_map_base);
3843     assert(dirty_card_log_enqueue != 0, "postcondition.");
3844     if (G1SATBPrintStubs) {
3845       tty->print_cr("Generated dirty_card enqueue:");
3846       Disassembler::decode((u_char*)dirty_card_log_enqueue,
3847                            dirty_card_log_enqueue_end,
3848                            tty);
3849     }
3850   }
3851 }
3852 
3853 
3854 void MacroAssembler::g1_write_barrier_post(Register store_addr, Register new_val, Register tmp) {
3855 
3856   Label filtered;
3857   MacroAssembler* post_filter_masm = this;
3858 
3859   if (new_val == G0) return;
3860 
3861   G1SATBCardTableModRefBS* bs = (G1SATBCardTableModRefBS*) Universe::heap()->barrier_set();
3862   assert(bs->kind() == BarrierSet::G1SATBCT ||
3863          bs->kind() == BarrierSet::G1SATBCTLogging, "wrong barrier");
3864 
3865   if (G1RSBarrierRegionFilter) {
3866     xor3(store_addr, new_val, tmp);
3867 #ifdef _LP64
3868     srlx(tmp, HeapRegion::LogOfHRGrainBytes, tmp);
3869 #else
3870     srl(tmp, HeapRegion::LogOfHRGrainBytes, tmp);
3871 #endif
3872 
3873     // XXX Should I predict this taken or not?  Does it matter?
3874     cmp_and_brx_short(tmp, G0, Assembler::equal, Assembler::pt, filtered);
3875   }
3876 
3877   // If the "store_addr" register is an "in" or "local" register, move it to
3878   // a scratch reg so we can pass it as an argument.
3879   bool use_scr = !(store_addr->is_global() || store_addr->is_out());
3880   // Pick a scratch register different from "tmp".
3881   Register scr = (tmp == G1_scratch ? G3_scratch : G1_scratch);
3882   // Make sure we use up the delay slot!
3883   if (use_scr) {
3884     post_filter_masm->mov(store_addr, scr);
3885   } else {
3886     post_filter_masm->nop();
3887   }
3888   generate_dirty_card_log_enqueue_if_necessary(bs->byte_map_base);
3889   save_frame(0);
3890   call(dirty_card_log_enqueue);
3891   if (use_scr) {
3892     delayed()->mov(scr, O0);
3893   } else {
3894     delayed()->mov(store_addr->after_save(), O0);
3895   }
3896   restore();
3897 
3898   bind(filtered);
3899 }
3900 
3901 #endif // INCLUDE_ALL_GCS
3902 ///////////////////////////////////////////////////////////////////////////////////
3903 
3904 void MacroAssembler::card_write_barrier_post(Register store_addr, Register new_val, Register tmp) {
3905   // If we're writing constant NULL, we can skip the write barrier.
3906   if (new_val == G0) return;
3907   CardTableModRefBS* bs = (CardTableModRefBS*) Universe::heap()->barrier_set();
3908   assert(bs->kind() == BarrierSet::CardTableModRef ||
3909          bs->kind() == BarrierSet::CardTableExtension, "wrong barrier");
3910   card_table_write(bs->byte_map_base, tmp, store_addr);
3911 }
3912 
3913 void MacroAssembler::load_klass(Register src_oop, Register klass) {
3914   // The number of bytes in this code is used by
3915   // MachCallDynamicJavaNode::ret_addr_offset()
3916   // if this changes, change that.
3917   if (UseCompressedClassPointers) {
3918     lduw(src_oop, oopDesc::klass_offset_in_bytes(), klass);
3919     decode_klass_not_null(klass);
3920   } else {
3921     ld_ptr(src_oop, oopDesc::klass_offset_in_bytes(), klass);
3922   }
3923 }
3924 
3925 void MacroAssembler::store_klass(Register klass, Register dst_oop) {
3926   if (UseCompressedClassPointers) {
3927     assert(dst_oop != klass, "not enough registers");
3928     encode_klass_not_null(klass);
3929     st(klass, dst_oop, oopDesc::klass_offset_in_bytes());
3930   } else {
3931     st_ptr(klass, dst_oop, oopDesc::klass_offset_in_bytes());
3932   }
3933 }
3934 
3935 void MacroAssembler::store_klass_gap(Register s, Register d) {
3936   if (UseCompressedClassPointers) {
3937     assert(s != d, "not enough registers");
3938     st(s, d, oopDesc::klass_gap_offset_in_bytes());
3939   }
3940 }
3941 
3942 void MacroAssembler::load_heap_oop(const Address& s, Register d) {
3943   if (UseCompressedOops) {
3944     lduw(s, d);
3945     decode_heap_oop(d);
3946   } else {
3947     ld_ptr(s, d);
3948   }
3949 }
3950 
3951 void MacroAssembler::load_heap_oop(Register s1, Register s2, Register d) {
3952    if (UseCompressedOops) {
3953     lduw(s1, s2, d);
3954     decode_heap_oop(d, d);
3955   } else {
3956     ld_ptr(s1, s2, d);
3957   }
3958 }
3959 
3960 void MacroAssembler::load_heap_oop(Register s1, int simm13a, Register d) {
3961    if (UseCompressedOops) {
3962     lduw(s1, simm13a, d);
3963     decode_heap_oop(d, d);
3964   } else {
3965     ld_ptr(s1, simm13a, d);
3966   }
3967 }
3968 
3969 void MacroAssembler::load_heap_oop(Register s1, RegisterOrConstant s2, Register d) {
3970   if (s2.is_constant())  load_heap_oop(s1, s2.as_constant(), d);
3971   else                   load_heap_oop(s1, s2.as_register(), d);
3972 }
3973 
3974 void MacroAssembler::store_heap_oop(Register d, Register s1, Register s2) {
3975   if (UseCompressedOops) {
3976     assert(s1 != d && s2 != d, "not enough registers");
3977     encode_heap_oop(d);
3978     st(d, s1, s2);
3979   } else {
3980     st_ptr(d, s1, s2);
3981   }
3982 }
3983 
3984 void MacroAssembler::store_heap_oop(Register d, Register s1, int simm13a) {
3985   if (UseCompressedOops) {
3986     assert(s1 != d, "not enough registers");
3987     encode_heap_oop(d);
3988     st(d, s1, simm13a);
3989   } else {
3990     st_ptr(d, s1, simm13a);
3991   }
3992 }
3993 
3994 void MacroAssembler::store_heap_oop(Register d, const Address& a, int offset) {
3995   if (UseCompressedOops) {
3996     assert(a.base() != d, "not enough registers");
3997     encode_heap_oop(d);
3998     st(d, a, offset);
3999   } else {
4000     st_ptr(d, a, offset);
4001   }
4002 }
4003 
4004 
4005 void MacroAssembler::encode_heap_oop(Register src, Register dst) {
4006   assert (UseCompressedOops, "must be compressed");
4007   assert (Universe::heap() != NULL, "java heap should be initialized");
4008   assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4009   verify_oop(src);
4010   if (Universe::narrow_oop_base() == NULL) {
4011     srlx(src, LogMinObjAlignmentInBytes, dst);
4012     return;
4013   }
4014   Label done;
4015   if (src == dst) {
4016     // optimize for frequent case src == dst
4017     bpr(rc_nz, true, Assembler::pt, src, done);
4018     delayed() -> sub(src, G6_heapbase, dst); // annuled if not taken
4019     bind(done);
4020     srlx(src, LogMinObjAlignmentInBytes, dst);
4021   } else {
4022     bpr(rc_z, false, Assembler::pn, src, done);
4023     delayed() -> mov(G0, dst);
4024     // could be moved before branch, and annulate delay,
4025     // but may add some unneeded work decoding null
4026     sub(src, G6_heapbase, dst);
4027     srlx(dst, LogMinObjAlignmentInBytes, dst);
4028     bind(done);
4029   }
4030 }
4031 
4032 
4033 void MacroAssembler::encode_heap_oop_not_null(Register r) {
4034   assert (UseCompressedOops, "must be compressed");
4035   assert (Universe::heap() != NULL, "java heap should be initialized");
4036   assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4037   verify_oop(r);
4038   if (Universe::narrow_oop_base() != NULL)
4039     sub(r, G6_heapbase, r);
4040   srlx(r, LogMinObjAlignmentInBytes, r);
4041 }
4042 
4043 void MacroAssembler::encode_heap_oop_not_null(Register src, Register dst) {
4044   assert (UseCompressedOops, "must be compressed");
4045   assert (Universe::heap() != NULL, "java heap should be initialized");
4046   assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4047   verify_oop(src);
4048   if (Universe::narrow_oop_base() == NULL) {
4049     srlx(src, LogMinObjAlignmentInBytes, dst);
4050   } else {
4051     sub(src, G6_heapbase, dst);
4052     srlx(dst, LogMinObjAlignmentInBytes, dst);
4053   }
4054 }
4055 
4056 // Same algorithm as oops.inline.hpp decode_heap_oop.
4057 void  MacroAssembler::decode_heap_oop(Register src, Register dst) {
4058   assert (UseCompressedOops, "must be compressed");
4059   assert (Universe::heap() != NULL, "java heap should be initialized");
4060   assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4061   sllx(src, LogMinObjAlignmentInBytes, dst);
4062   if (Universe::narrow_oop_base() != NULL) {
4063     Label done;
4064     bpr(rc_nz, true, Assembler::pt, dst, done);
4065     delayed() -> add(dst, G6_heapbase, dst); // annuled if not taken
4066     bind(done);
4067   }
4068   verify_oop(dst);
4069 }
4070 
4071 void  MacroAssembler::decode_heap_oop_not_null(Register r) {
4072   // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4073   // pd_code_size_limit.
4074   // Also do not verify_oop as this is called by verify_oop.
4075   assert (UseCompressedOops, "must be compressed");
4076   assert (Universe::heap() != NULL, "java heap should be initialized");
4077   assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4078   sllx(r, LogMinObjAlignmentInBytes, r);
4079   if (Universe::narrow_oop_base() != NULL)
4080     add(r, G6_heapbase, r);
4081 }
4082 
4083 void  MacroAssembler::decode_heap_oop_not_null(Register src, Register dst) {
4084   // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4085   // pd_code_size_limit.
4086   // Also do not verify_oop as this is called by verify_oop.
4087   assert (UseCompressedOops, "must be compressed");
4088   assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4089   sllx(src, LogMinObjAlignmentInBytes, dst);
4090   if (Universe::narrow_oop_base() != NULL)
4091     add(dst, G6_heapbase, dst);
4092 }
4093 
4094 void MacroAssembler::encode_klass_not_null(Register r) {
4095   assert (UseCompressedClassPointers, "must be compressed");
4096   assert(Universe::narrow_klass_base() != NULL, "narrow_klass_base should be initialized");
4097   assert(r != G6_heapbase, "bad register choice");
4098   set((intptr_t)Universe::narrow_klass_base(), G6_heapbase);
4099   sub(r, G6_heapbase, r);
4100   if (Universe::narrow_klass_shift() != 0) {
4101     assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
4102     srlx(r, LogKlassAlignmentInBytes, r);
4103   }
4104   reinit_heapbase();
4105 }
4106 
4107 void MacroAssembler::encode_klass_not_null(Register src, Register dst) {
4108   if (src == dst) {
4109     encode_klass_not_null(src);
4110   } else {
4111     assert (UseCompressedClassPointers, "must be compressed");
4112     assert(Universe::narrow_klass_base() != NULL, "narrow_klass_base should be initialized");
4113     set((intptr_t)Universe::narrow_klass_base(), dst);
4114     sub(src, dst, dst);
4115     if (Universe::narrow_klass_shift() != 0) {
4116       srlx(dst, LogKlassAlignmentInBytes, dst);
4117     }
4118   }
4119 }
4120 
4121 // Function instr_size_for_decode_klass_not_null() counts the instructions
4122 // generated by decode_klass_not_null() and reinit_heapbase().  Hence, if
4123 // the instructions they generate change, then this method needs to be updated.
4124 int MacroAssembler::instr_size_for_decode_klass_not_null() {
4125   assert (UseCompressedClassPointers, "only for compressed klass ptrs");
4126   // set + add + set
4127   int num_instrs = insts_for_internal_set((intptr_t)Universe::narrow_klass_base()) + 1 +
4128     insts_for_internal_set((intptr_t)Universe::narrow_ptrs_base());
4129   if (Universe::narrow_klass_shift() == 0) {
4130     return num_instrs * BytesPerInstWord;
4131   } else { // sllx
4132     return (num_instrs + 1) * BytesPerInstWord;
4133   }
4134 }
4135 
4136 // !!! If the instructions that get generated here change then function
4137 // instr_size_for_decode_klass_not_null() needs to get updated.
4138 void  MacroAssembler::decode_klass_not_null(Register r) {
4139   // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4140   // pd_code_size_limit.
4141   assert (UseCompressedClassPointers, "must be compressed");
4142   assert(Universe::narrow_klass_base() != NULL, "narrow_klass_base should be initialized");
4143   assert(r != G6_heapbase, "bad register choice");
4144   set((intptr_t)Universe::narrow_klass_base(), G6_heapbase);
4145   if (Universe::narrow_klass_shift() != 0)
4146     sllx(r, LogKlassAlignmentInBytes, r);
4147   add(r, G6_heapbase, r);
4148   reinit_heapbase();
4149 }
4150 
4151 void  MacroAssembler::decode_klass_not_null(Register src, Register dst) {
4152   if (src == dst) {
4153     decode_klass_not_null(src);
4154   } else {
4155     // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4156     // pd_code_size_limit.
4157     assert (UseCompressedClassPointers, "must be compressed");
4158     assert(Universe::narrow_klass_base() != NULL, "narrow_klass_base should be initialized");
4159     if (Universe::narrow_klass_shift() != 0) {
4160       assert((src != G6_heapbase) && (dst != G6_heapbase), "bad register choice");
4161       set((intptr_t)Universe::narrow_klass_base(), G6_heapbase);
4162       sllx(src, LogKlassAlignmentInBytes, dst);
4163       add(dst, G6_heapbase, dst);
4164       reinit_heapbase();
4165     } else {
4166       set((intptr_t)Universe::narrow_klass_base(), dst);
4167       add(src, dst, dst);
4168     }
4169   }
4170 }
4171 
4172 void MacroAssembler::reinit_heapbase() {
4173   if (UseCompressedOops || UseCompressedClassPointers) {
4174     if (Universe::heap() != NULL) {
4175       set((intptr_t)Universe::narrow_ptrs_base(), G6_heapbase);
4176     } else {
4177       AddressLiteral base(Universe::narrow_ptrs_base_addr());
4178       load_ptr_contents(base, G6_heapbase);
4179     }
4180   }
4181 }
4182 
4183 // Compare char[] arrays aligned to 4 bytes.
4184 void MacroAssembler::char_arrays_equals(Register ary1, Register ary2,
4185                                         Register limit, Register result,
4186                                         Register chr1, Register chr2, Label& Ldone) {
4187   Label Lvector, Lloop;
4188   assert(chr1 == result, "should be the same");
4189 
4190   // Note: limit contains number of bytes (2*char_elements) != 0.
4191   andcc(limit, 0x2, chr1); // trailing character ?
4192   br(Assembler::zero, false, Assembler::pt, Lvector);
4193   delayed()->nop();
4194 
4195   // compare the trailing char
4196   sub(limit, sizeof(jchar), limit);
4197   lduh(ary1, limit, chr1);
4198   lduh(ary2, limit, chr2);
4199   cmp(chr1, chr2);
4200   br(Assembler::notEqual, true, Assembler::pt, Ldone);
4201   delayed()->mov(G0, result);     // not equal
4202 
4203   // only one char ?
4204   cmp_zero_and_br(zero, limit, Ldone, true, Assembler::pn);
4205   delayed()->add(G0, 1, result); // zero-length arrays are equal
4206 
4207   // word by word compare, dont't need alignment check
4208   bind(Lvector);
4209   // Shift ary1 and ary2 to the end of the arrays, negate limit
4210   add(ary1, limit, ary1);
4211   add(ary2, limit, ary2);
4212   neg(limit, limit);
4213 
4214   lduw(ary1, limit, chr1);
4215   bind(Lloop);
4216   lduw(ary2, limit, chr2);
4217   cmp(chr1, chr2);
4218   br(Assembler::notEqual, true, Assembler::pt, Ldone);
4219   delayed()->mov(G0, result);     // not equal
4220   inccc(limit, 2*sizeof(jchar));
4221   // annul LDUW if branch is not taken to prevent access past end of array
4222   br(Assembler::notZero, true, Assembler::pt, Lloop);
4223   delayed()->lduw(ary1, limit, chr1); // hoisted
4224 
4225   // Caller should set it:
4226   // add(G0, 1, result); // equals
4227 }
4228 
4229 // Use BIS for zeroing (count is in bytes).
4230 void MacroAssembler::bis_zeroing(Register to, Register count, Register temp, Label& Ldone) {
4231   assert(UseBlockZeroing && VM_Version::has_block_zeroing(), "only works with BIS zeroing");
4232   Register end = count;
4233   int cache_line_size = VM_Version::prefetch_data_size();
4234   // Minimum count when BIS zeroing can be used since
4235   // it needs membar which is expensive.
4236   int block_zero_size  = MAX2(cache_line_size*3, (int)BlockZeroingLowLimit);
4237 
4238   Label small_loop;
4239   // Check if count is negative (dead code) or zero.
4240   // Note, count uses 64bit in 64 bit VM.
4241   cmp_and_brx_short(count, 0, Assembler::lessEqual, Assembler::pn, Ldone);
4242 
4243   // Use BIS zeroing only for big arrays since it requires membar.
4244   if (Assembler::is_simm13(block_zero_size)) { // < 4096
4245     cmp(count, block_zero_size);
4246   } else {
4247     set(block_zero_size, temp);
4248     cmp(count, temp);
4249   }
4250   br(Assembler::lessUnsigned, false, Assembler::pt, small_loop);
4251   delayed()->add(to, count, end);
4252 
4253   // Note: size is >= three (32 bytes) cache lines.
4254 
4255   // Clean the beginning of space up to next cache line.
4256   for (int offs = 0; offs < cache_line_size; offs += 8) {
4257     stx(G0, to, offs);
4258   }
4259 
4260   // align to next cache line
4261   add(to, cache_line_size, to);
4262   and3(to, -cache_line_size, to);
4263 
4264   // Note: size left >= two (32 bytes) cache lines.
4265 
4266   // BIS should not be used to zero tail (64 bytes)
4267   // to avoid zeroing a header of the following object.
4268   sub(end, (cache_line_size*2)-8, end);
4269 
4270   Label bis_loop;
4271   bind(bis_loop);
4272   stxa(G0, to, G0, Assembler::ASI_ST_BLKINIT_PRIMARY);
4273   add(to, cache_line_size, to);
4274   cmp_and_brx_short(to, end, Assembler::lessUnsigned, Assembler::pt, bis_loop);
4275 
4276   // BIS needs membar.
4277   membar(Assembler::StoreLoad);
4278 
4279   add(end, (cache_line_size*2)-8, end); // restore end
4280   cmp_and_brx_short(to, end, Assembler::greaterEqualUnsigned, Assembler::pn, Ldone);
4281 
4282   // Clean the tail.
4283   bind(small_loop);
4284   stx(G0, to, 0);
4285   add(to, 8, to);
4286   cmp_and_brx_short(to, end, Assembler::lessUnsigned, Assembler::pt, small_loop);
4287   nop(); // Separate short branches
4288 }