1 /*
   2  * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2012, 2026 SAP SE. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "asm/macroAssembler.inline.hpp"
  27 #include "code/compiledIC.hpp"
  28 #include "compiler/disassembler.hpp"
  29 #include "gc/shared/collectedHeap.inline.hpp"
  30 #include "gc/shared/barrierSet.hpp"
  31 #include "gc/shared/barrierSetAssembler.hpp"
  32 #include "interpreter/interpreter.hpp"
  33 #include "interpreter/interpreterRuntime.hpp"
  34 #include "memory/resourceArea.hpp"
  35 #include "nativeInst_ppc.hpp"
  36 #include "oops/compressedKlass.inline.hpp"
  37 #include "oops/compressedOops.inline.hpp"
  38 #include "oops/klass.inline.hpp"
  39 #include "oops/methodData.hpp"
  40 #include "prims/methodHandles.hpp"
  41 #include "register_ppc.hpp"
  42 #include "runtime/icache.hpp"
  43 #include "runtime/interfaceSupport.inline.hpp"
  44 #include "runtime/objectMonitor.hpp"
  45 #include "runtime/objectMonitorTable.hpp"
  46 #include "runtime/os.hpp"
  47 #include "runtime/safepoint.hpp"
  48 #include "runtime/safepointMechanism.hpp"
  49 #include "runtime/sharedRuntime.hpp"
  50 #include "runtime/stubRoutines.hpp"
  51 #include "runtime/vm_version.hpp"
  52 #include "utilities/macros.hpp"
  53 #include "utilities/powerOfTwo.hpp"
  54 
  55 #ifdef PRODUCT
  56 #define BLOCK_COMMENT(str) // nothing
  57 #else
  58 #define BLOCK_COMMENT(str) block_comment(str)
  59 #endif
  60 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
  61 
  62 #ifdef ASSERT
  63 // On RISC, there's no benefit to verifying instruction boundaries.
  64 bool AbstractAssembler::pd_check_instruction_mark() { return false; }
  65 #endif
  66 
  67 void MacroAssembler::ld_largeoffset_unchecked(Register d, int si31, Register a, int emit_filler_nop) {
  68   assert(Assembler::is_simm(si31, 31) && si31 >= 0, "si31 out of range");
  69   if (Assembler::is_simm(si31, 16)) {
  70     ld(d, si31, a);
  71     if (emit_filler_nop) nop();
  72   } else {
  73     const int hi = MacroAssembler::largeoffset_si16_si16_hi(si31);
  74     const int lo = MacroAssembler::largeoffset_si16_si16_lo(si31);
  75     addis(d, a, hi);
  76     ld(d, lo, d);
  77   }
  78 }
  79 
  80 void MacroAssembler::ld_largeoffset(Register d, int si31, Register a, int emit_filler_nop) {
  81   assert_different_registers(d, a);
  82   ld_largeoffset_unchecked(d, si31, a, emit_filler_nop);
  83 }
  84 
  85 void MacroAssembler::load_sized_value(Register dst, RegisterOrConstant offs, Register base,
  86                                       size_t size_in_bytes, bool is_signed) {
  87   switch (size_in_bytes) {
  88   case  8:              ld(dst, offs, base);                         break;
  89   case  4:  is_signed ? lwa(dst, offs, base) : lwz(dst, offs, base); break;
  90   case  2:  is_signed ? lha(dst, offs, base) : lhz(dst, offs, base); break;
  91   case  1:  lbz(dst, offs, base); if (is_signed) extsb(dst, dst);    break; // lba doesn't exist :(
  92   default:  ShouldNotReachHere();
  93   }
  94 }
  95 
  96 void MacroAssembler::store_sized_value(Register dst, RegisterOrConstant offs, Register base,
  97                                        size_t size_in_bytes) {
  98   switch (size_in_bytes) {
  99   case  8:  std(dst, offs, base); break;
 100   case  4:  stw(dst, offs, base); break;
 101   case  2:  sth(dst, offs, base); break;
 102   case  1:  stb(dst, offs, base); break;
 103   default:  ShouldNotReachHere();
 104   }
 105 }
 106 
 107 void MacroAssembler::align(int modulus, int max, int rem) {
 108   int padding = (rem + modulus - (offset() % modulus)) % modulus;
 109   if (padding > max) return;
 110   for (int c = (padding >> 2); c > 0; --c) { nop(); }
 111 }
 112 
 113 void MacroAssembler::align_prefix() {
 114   if (is_aligned(offset() + BytesPerInstWord, 64)) { nop(); }
 115 }
 116 
 117 // Issue instructions that calculate given TOC from global TOC.
 118 void MacroAssembler::calculate_address_from_global_toc(Register dst, address addr, bool hi16, bool lo16,
 119                                                        bool add_relocation, bool emit_dummy_addr,
 120                                                        bool add_addr_to_reloc) {
 121   int offset = -1;
 122   if (emit_dummy_addr) {
 123     offset = -128; // dummy address
 124   } else if (addr != (address)(intptr_t)-1) {
 125     offset = MacroAssembler::offset_to_global_toc(addr);
 126   }
 127 
 128   if (hi16) {
 129     addis(dst, R29_TOC, MacroAssembler::largeoffset_si16_si16_hi(offset));
 130   }
 131   if (lo16) {
 132     if (add_relocation) {
 133       // Relocate at the addi to avoid confusion with a load from the method's TOC.
 134       RelocationHolder rh = add_addr_to_reloc ?
 135           internal_word_Relocation::spec(addr) :
 136           internal_word_Relocation::spec_for_immediate();
 137       relocate(rh);
 138     }
 139     addi(dst, dst, MacroAssembler::largeoffset_si16_si16_lo(offset));
 140   }
 141 }
 142 
 143 address MacroAssembler::patch_calculate_address_from_global_toc_at(address a, address bound, address addr) {
 144   const int offset = MacroAssembler::offset_to_global_toc(addr);
 145 
 146   const address inst2_addr = a;
 147   const int inst2 = *(int *)inst2_addr;
 148 
 149   // The relocation points to the second instruction, the addi,
 150   // and the addi reads and writes the same register dst.
 151   const int dst = inv_rt_field(inst2);
 152   assert(is_addi(inst2) && inv_ra_field(inst2) == dst, "must be addi reading and writing dst");
 153 
 154   // Now, find the preceding addis which writes to dst.
 155   int inst1 = 0;
 156   address inst1_addr = inst2_addr - BytesPerInstWord;
 157   while (inst1_addr >= bound) {
 158     inst1 = *(int *) inst1_addr;
 159     if (is_addis(inst1) && inv_rt_field(inst1) == dst) {
 160       // Stop, found the addis which writes dst.
 161       break;
 162     }
 163     inst1_addr -= BytesPerInstWord;
 164   }
 165 
 166   assert(is_addis(inst1) && inv_ra_field(inst1) == 29 /* R29 */, "source must be global TOC");
 167   set_imm((int *)inst1_addr, MacroAssembler::largeoffset_si16_si16_hi(offset));
 168   set_imm((int *)inst2_addr, MacroAssembler::largeoffset_si16_si16_lo(offset));
 169   return inst1_addr;
 170 }
 171 
 172 address MacroAssembler::get_address_of_calculate_address_from_global_toc_at(address a, address bound) {
 173   const address inst2_addr = a;
 174   const int inst2 = *(int *)inst2_addr;
 175 
 176   // The relocation points to the second instruction, the addi,
 177   // and the addi reads and writes the same register dst.
 178   const int dst = inv_rt_field(inst2);
 179   assert(is_addi(inst2) && inv_ra_field(inst2) == dst, "must be addi reading and writing dst");
 180 
 181   // Now, find the preceding addis which writes to dst.
 182   int inst1 = 0;
 183   address inst1_addr = inst2_addr - BytesPerInstWord;
 184   while (inst1_addr >= bound) {
 185     inst1 = *(int *) inst1_addr;
 186     if (is_addis(inst1) && inv_rt_field(inst1) == dst) {
 187       // stop, found the addis which writes dst
 188       break;
 189     }
 190     inst1_addr -= BytesPerInstWord;
 191   }
 192 
 193   assert(is_addis(inst1) && inv_ra_field(inst1) == 29 /* R29 */, "source must be global TOC");
 194 
 195   int offset = (get_imm(inst1_addr, 0) << 16) + get_imm(inst2_addr, 0);
 196   // -1 is a special case
 197   if (offset == -1) {
 198     return (address)(intptr_t)-1;
 199   } else {
 200     return global_toc() + offset;
 201   }
 202 }
 203 
 204 #ifdef _LP64
 205 // Patch compressed oops or klass constants.
 206 // Assembler sequence is
 207 // 1) compressed oops:
 208 //    lis  rx = const.hi
 209 //    ori rx = rx | const.lo
 210 // 2) compressed klass:
 211 //    lis  rx = const.hi
 212 //    clrldi rx = rx & 0xFFFFffff // clearMS32b, optional
 213 //    ori rx = rx | const.lo
 214 // Clrldi will be passed by.
 215 address MacroAssembler::patch_set_narrow_oop(address a, address bound, narrowOop data) {
 216   assert(UseCompressedOops, "Should only patch compressed oops");
 217 
 218   const address inst2_addr = a;
 219   const int inst2 = *(int *)inst2_addr;
 220 
 221   // The relocation points to the second instruction, the ori,
 222   // and the ori reads and writes the same register dst.
 223   const int dst = inv_rta_field(inst2);
 224   assert(is_ori(inst2) && inv_rs_field(inst2) == dst, "must be ori reading and writing dst");
 225   // Now, find the preceding addis which writes to dst.
 226   int inst1 = 0;
 227   address inst1_addr = inst2_addr - BytesPerInstWord;
 228   bool inst1_found = false;
 229   while (inst1_addr >= bound) {
 230     inst1 = *(int *)inst1_addr;
 231     if (is_lis(inst1) && inv_rs_field(inst1) == dst) { inst1_found = true; break; }
 232     inst1_addr -= BytesPerInstWord;
 233   }
 234   assert(inst1_found, "inst is not lis");
 235 
 236   uint32_t data_value = CompressedOops::narrow_oop_value(data);
 237   int xc = (data_value >> 16) & 0xffff;
 238   int xd = (data_value >>  0) & 0xffff;
 239 
 240   set_imm((int *)inst1_addr, (short)(xc)); // see enc_load_con_narrow_hi/_lo
 241   set_imm((int *)inst2_addr,        (xd)); // unsigned int
 242   return inst1_addr;
 243 }
 244 
 245 // Get compressed oop constant.
 246 narrowOop MacroAssembler::get_narrow_oop(address a, address bound) {
 247   assert(UseCompressedOops, "Should only patch compressed oops");
 248 
 249   const address inst2_addr = a;
 250   const int inst2 = *(int *)inst2_addr;
 251 
 252   // The relocation points to the second instruction, the ori,
 253   // and the ori reads and writes the same register dst.
 254   const int dst = inv_rta_field(inst2);
 255   assert(is_ori(inst2) && inv_rs_field(inst2) == dst, "must be ori reading and writing dst");
 256   // Now, find the preceding lis which writes to dst.
 257   int inst1 = 0;
 258   address inst1_addr = inst2_addr - BytesPerInstWord;
 259   bool inst1_found = false;
 260 
 261   while (inst1_addr >= bound) {
 262     inst1 = *(int *) inst1_addr;
 263     if (is_lis(inst1) && inv_rs_field(inst1) == dst) { inst1_found = true; break;}
 264     inst1_addr -= BytesPerInstWord;
 265   }
 266   assert(inst1_found, "inst is not lis");
 267 
 268   uint xl = ((unsigned int) (get_imm(inst2_addr, 0) & 0xffff));
 269   uint xh = (((get_imm(inst1_addr, 0)) & 0xffff) << 16);
 270 
 271   return CompressedOops::narrow_oop_cast(xl | xh);
 272 }
 273 #endif // _LP64
 274 
 275 // Returns true if successful.
 276 bool MacroAssembler::load_const_from_method_toc(Register dst, AddressLiteral& a,
 277                                                 Register toc, bool fixed_size) {
 278   int toc_offset = 0;
 279   // Use RelocationHolder::none for the constant pool entry, otherwise
 280   // we will end up with a failing NativeCall::verify(x) where x is
 281   // the address of the constant pool entry.
 282   // FIXME: We should insert relocation information for oops at the constant
 283   // pool entries instead of inserting it at the loads; patching of a constant
 284   // pool entry should be less expensive.
 285   address const_address = address_constant((address)a.value(), RelocationHolder::none);
 286   if (const_address == nullptr) { return false; } // allocation failure
 287   // Relocate at the pc of the load.
 288   relocate(a.rspec());
 289   toc_offset = (int)(const_address - code()->consts()->start());
 290   ld_largeoffset_unchecked(dst, toc_offset, toc, fixed_size);
 291   return true;
 292 }
 293 
 294 bool MacroAssembler::is_load_const_from_method_toc_at(address a) {
 295   const address inst1_addr = a;
 296   const int inst1 = *(int *)inst1_addr;
 297 
 298    // The relocation points to the ld or the addis.
 299    return (is_ld(inst1)) ||
 300           (is_addis(inst1) && inv_ra_field(inst1) != 0);
 301 }
 302 
 303 int MacroAssembler::get_offset_of_load_const_from_method_toc_at(address a) {
 304   assert(is_load_const_from_method_toc_at(a), "must be load_const_from_method_toc");
 305 
 306   const address inst1_addr = a;
 307   const int inst1 = *(int *)inst1_addr;
 308 
 309   if (is_ld(inst1)) {
 310     return inv_d1_field(inst1);
 311   } else if (is_addis(inst1)) {
 312     const int dst = inv_rt_field(inst1);
 313 
 314     // Now, find the succeeding ld which reads and writes to dst.
 315     address inst2_addr = inst1_addr + BytesPerInstWord;
 316     int inst2 = 0;
 317     while (true) {
 318       inst2 = *(int *) inst2_addr;
 319       if (is_ld(inst2) && inv_ra_field(inst2) == dst && inv_rt_field(inst2) == dst) {
 320         // Stop, found the ld which reads and writes dst.
 321         break;
 322       }
 323       inst2_addr += BytesPerInstWord;
 324     }
 325     return (inv_d1_field(inst1) << 16) + inv_d1_field(inst2);
 326   }
 327   ShouldNotReachHere();
 328   return 0;
 329 }
 330 
 331 // Get the constant from a `load_const' sequence.
 332 long MacroAssembler::get_const(address a) {
 333   assert(is_load_const_at(a), "not a load of a constant");
 334   const int *p = (const int*) a;
 335   unsigned long x = (((unsigned long) (get_imm(a,0) & 0xffff)) << 48);
 336   if (is_ori(*(p+1))) {
 337     x |= (((unsigned long) (get_imm(a,1) & 0xffff)) << 32);
 338     x |= (((unsigned long) (get_imm(a,3) & 0xffff)) << 16);
 339     x |= (((unsigned long) (get_imm(a,4) & 0xffff)));
 340   } else if (is_lis(*(p+1))) {
 341     x |= (((unsigned long) (get_imm(a,2) & 0xffff)) << 32);
 342     x |= (((unsigned long) (get_imm(a,1) & 0xffff)) << 16);
 343     x |= (((unsigned long) (get_imm(a,3) & 0xffff)));
 344   } else {
 345     ShouldNotReachHere();
 346     return (long) 0;
 347   }
 348   return (long) x;
 349 }
 350 
 351 // Patch the 64 bit constant of a `load_const' sequence. This is a low
 352 // level procedure. It neither flushes the instruction cache nor is it
 353 // mt safe.
 354 void MacroAssembler::patch_const(address a, long x) {
 355   assert(is_load_const_at(a), "not a load of a constant");
 356   int *p = (int*) a;
 357   if (is_ori(*(p+1))) {
 358     set_imm(0 + p, (x >> 48) & 0xffff);
 359     set_imm(1 + p, (x >> 32) & 0xffff);
 360     set_imm(3 + p, (x >> 16) & 0xffff);
 361     set_imm(4 + p, x & 0xffff);
 362   } else if (is_lis(*(p+1))) {
 363     set_imm(0 + p, (x >> 48) & 0xffff);
 364     set_imm(2 + p, (x >> 32) & 0xffff);
 365     set_imm(1 + p, (x >> 16) & 0xffff);
 366     set_imm(3 + p, x & 0xffff);
 367   } else {
 368     ShouldNotReachHere();
 369   }
 370 }
 371 
 372 AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) {
 373   assert(oop_recorder() != nullptr, "this assembler needs a Recorder");
 374   int index = oop_recorder()->allocate_metadata_index(obj);
 375   RelocationHolder rspec = metadata_Relocation::spec(index);
 376   return AddressLiteral((address)obj, rspec);
 377 }
 378 
 379 AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) {
 380   assert(oop_recorder() != nullptr, "this assembler needs a Recorder");
 381   int index = oop_recorder()->find_index(obj);
 382   RelocationHolder rspec = metadata_Relocation::spec(index);
 383   return AddressLiteral((address)obj, rspec);
 384 }
 385 
 386 AddressLiteral MacroAssembler::allocate_oop_address(jobject obj) {
 387   assert(oop_recorder() != nullptr, "this assembler needs an OopRecorder");
 388   int oop_index = oop_recorder()->allocate_oop_index(obj);
 389   return AddressLiteral(address(obj), oop_Relocation::spec(oop_index));
 390 }
 391 
 392 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) {
 393   assert(oop_recorder() != nullptr, "this assembler needs an OopRecorder");
 394   int oop_index = oop_recorder()->find_index(obj);
 395   return AddressLiteral(address(obj), oop_Relocation::spec(oop_index));
 396 }
 397 
 398 #ifndef PRODUCT
 399 void MacroAssembler::pd_print_patched_instruction(address branch) {
 400   Unimplemented(); // TODO: PPC port
 401 }
 402 #endif // ndef PRODUCT
 403 
 404 // Conditional far branch for destinations encodable in 24+2 bits.
 405 void MacroAssembler::bc_far(int boint, int biint, Label& dest, int optimize) {
 406 
 407   // If requested by flag optimize, relocate the bc_far as a
 408   // runtime_call and prepare for optimizing it when the code gets
 409   // relocated.
 410   if (optimize == bc_far_optimize_on_relocate) {
 411     relocate(relocInfo::runtime_call_type);
 412   }
 413 
 414   // variant 2:
 415   //
 416   //    b!cxx SKIP
 417   //    bxx   DEST
 418   //  SKIP:
 419   //
 420 
 421   const int opposite_boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(boint)),
 422                                                 opposite_bcond(inv_boint_bcond(boint)));
 423 
 424   // We emit two branches.
 425   // First, a conditional branch which jumps around the far branch.
 426   const address not_taken_pc = pc() + 2 * BytesPerInstWord;
 427   const address bc_pc        = pc();
 428   bc(opposite_boint, biint, not_taken_pc);
 429 
 430   const int bc_instr = *(int*)bc_pc;
 431   assert(not_taken_pc == (address)inv_bd_field(bc_instr, (intptr_t)bc_pc), "postcondition");
 432   assert(opposite_boint == inv_bo_field(bc_instr), "postcondition");
 433   assert(boint == add_bhint_to_boint(opposite_bhint(inv_boint_bhint(inv_bo_field(bc_instr))),
 434                                      opposite_bcond(inv_boint_bcond(inv_bo_field(bc_instr)))),
 435          "postcondition");
 436   assert(biint == inv_bi_field(bc_instr), "postcondition");
 437 
 438   // Second, an unconditional far branch which jumps to dest.
 439   // Note: target(dest) remembers the current pc (see CodeSection::target)
 440   //       and returns the current pc if the label is not bound yet; when
 441   //       the label gets bound, the unconditional far branch will be patched.
 442   const address target_pc = target(dest);
 443   const address b_pc  = pc();
 444   b(target_pc);
 445 
 446   assert(not_taken_pc == pc(),                     "postcondition");
 447   assert(dest.is_bound() || target_pc == b_pc, "postcondition");
 448 }
 449 
 450 // 1 or 2 instructions
 451 void MacroAssembler::bc_far_optimized(int boint, int biint, Label& dest) {
 452   if (dest.is_bound() && is_within_range_of_bcxx(target(dest), pc())) {
 453     bc(boint, biint, dest);
 454   } else {
 455     bc_far(boint, biint, dest, MacroAssembler::bc_far_optimize_on_relocate);
 456   }
 457 }
 458 
 459 bool MacroAssembler::is_bc_far_at(address instruction_addr) {
 460   return is_bc_far_variant1_at(instruction_addr) ||
 461          is_bc_far_variant2_at(instruction_addr) ||
 462          is_bc_far_variant3_at(instruction_addr);
 463 }
 464 
 465 address MacroAssembler::get_dest_of_bc_far_at(address instruction_addr) {
 466   if (is_bc_far_variant1_at(instruction_addr)) {
 467     const address instruction_1_addr = instruction_addr;
 468     const int instruction_1 = *(int*)instruction_1_addr;
 469     return (address)inv_bd_field(instruction_1, (intptr_t)instruction_1_addr);
 470   } else if (is_bc_far_variant2_at(instruction_addr)) {
 471     const address instruction_2_addr = instruction_addr + 4;
 472     return bxx_destination(instruction_2_addr);
 473   } else if (is_bc_far_variant3_at(instruction_addr)) {
 474     return instruction_addr + 8;
 475   }
 476   // variant 4 ???
 477   ShouldNotReachHere();
 478   return nullptr;
 479 }
 480 void MacroAssembler::set_dest_of_bc_far_at(address instruction_addr, address dest) {
 481 
 482   if (is_bc_far_variant3_at(instruction_addr)) {
 483     // variant 3, far cond branch to the next instruction, already patched to nops:
 484     //
 485     //    nop
 486     //    nop
 487     //  SKIP/DEST:
 488     //
 489     return;
 490   }
 491 
 492   // first, extract boint and biint from the current branch
 493   int boint = 0;
 494   int biint = 0;
 495 
 496   ResourceMark rm;
 497   const int code_size = 2 * BytesPerInstWord;
 498   CodeBuffer buf(instruction_addr, code_size);
 499   MacroAssembler masm(&buf);
 500   if (is_bc_far_variant2_at(instruction_addr) && dest == instruction_addr + 8) {
 501     // Far branch to next instruction: Optimize it by patching nops (produce variant 3).
 502     masm.nop();
 503     masm.nop();
 504   } else {
 505     if (is_bc_far_variant1_at(instruction_addr)) {
 506       // variant 1, the 1st instruction contains the destination address:
 507       //
 508       //    bcxx  DEST
 509       //    nop
 510       //
 511       const int instruction_1 = *(int*)(instruction_addr);
 512       boint = inv_bo_field(instruction_1);
 513       biint = inv_bi_field(instruction_1);
 514     } else if (is_bc_far_variant2_at(instruction_addr)) {
 515       // variant 2, the 2nd instruction contains the destination address:
 516       //
 517       //    b!cxx SKIP
 518       //    bxx   DEST
 519       //  SKIP:
 520       //
 521       const int instruction_1 = *(int*)(instruction_addr);
 522       boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(inv_bo_field(instruction_1))),
 523           opposite_bcond(inv_boint_bcond(inv_bo_field(instruction_1))));
 524       biint = inv_bi_field(instruction_1);
 525     } else {
 526       // variant 4???
 527       ShouldNotReachHere();
 528     }
 529 
 530     // second, set the new branch destination and optimize the code
 531     if (dest != instruction_addr + 4 && // the bc_far is still unbound!
 532         masm.is_within_range_of_bcxx(dest, instruction_addr)) {
 533       // variant 1:
 534       //
 535       //    bcxx  DEST
 536       //    nop
 537       //
 538       masm.bc(boint, biint, dest);
 539       masm.nop();
 540     } else {
 541       // variant 2:
 542       //
 543       //    b!cxx SKIP
 544       //    bxx   DEST
 545       //  SKIP:
 546       //
 547       const int opposite_boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(boint)),
 548                                                     opposite_bcond(inv_boint_bcond(boint)));
 549       const address not_taken_pc = masm.pc() + 2 * BytesPerInstWord;
 550       masm.bc(opposite_boint, biint, not_taken_pc);
 551       masm.b(dest);
 552     }
 553   }
 554   ICache::ppc64_flush_icache_bytes(instruction_addr, code_size);
 555 }
 556 
 557 // Emit a NOT mt-safe patchable 64 bit absolute call/jump.
 558 void MacroAssembler::bxx64_patchable(address dest, relocInfo::relocType rt, bool link) {
 559   // get current pc
 560   uint64_t start_pc = (uint64_t) pc();
 561 
 562   const address pc_of_bl = (address) (start_pc + (6*BytesPerInstWord)); // bl is last
 563   const address pc_of_b  = (address) (start_pc + (0*BytesPerInstWord)); // b is first
 564 
 565   // relocate here
 566   if (rt != relocInfo::none) {
 567     relocate(rt);
 568   }
 569 
 570   if ( ReoptimizeCallSequences &&
 571        (( link && is_within_range_of_b(dest, pc_of_bl)) ||
 572         (!link && is_within_range_of_b(dest, pc_of_b)))) {
 573     // variant 2:
 574     // Emit an optimized, pc-relative call/jump.
 575 
 576     if (link) {
 577       // some padding
 578       nop();
 579       nop();
 580       nop();
 581       nop();
 582       nop();
 583       nop();
 584 
 585       // do the call
 586       assert(pc() == pc_of_bl, "just checking");
 587       bl(dest, relocInfo::none);
 588     } else {
 589       // do the jump
 590       assert(pc() == pc_of_b, "just checking");
 591       b(dest, relocInfo::none);
 592 
 593       // some padding
 594       nop();
 595       nop();
 596       nop();
 597       nop();
 598       nop();
 599       nop();
 600     }
 601 
 602     // Assert that we can identify the emitted call/jump.
 603     assert(is_bxx64_patchable_variant2_at((address)start_pc, link),
 604            "can't identify emitted call");
 605   } else {
 606     // variant 1:
 607     mr(R0, R11);  // spill R11 -> R0.
 608 
 609     // Load the destination address into CTR,
 610     // calculate destination relative to global toc.
 611     calculate_address_from_global_toc(R11, dest, true, true, false);
 612 
 613     mtctr(R11);
 614     mr(R11, R0);  // spill R11 <- R0.
 615     nop();
 616 
 617     // do the call/jump
 618     if (link) {
 619       bctrl();
 620     } else{
 621       bctr();
 622     }
 623     // Assert that we can identify the emitted call/jump.
 624     assert(is_bxx64_patchable_variant1b_at((address)start_pc, link),
 625            "can't identify emitted call");
 626   }
 627 
 628   // Assert that we can identify the emitted call/jump.
 629   assert(is_bxx64_patchable_at((address)start_pc, link),
 630          "can't identify emitted call");
 631   assert(get_dest_of_bxx64_patchable_at((address)start_pc, link) == dest,
 632          "wrong encoding of dest address");
 633 }
 634 
 635 // Identify a bxx64_patchable instruction.
 636 bool MacroAssembler::is_bxx64_patchable_at(address instruction_addr, bool link) {
 637   return is_bxx64_patchable_variant1b_at(instruction_addr, link)
 638     //|| is_bxx64_patchable_variant1_at(instruction_addr, link)
 639       || is_bxx64_patchable_variant2_at(instruction_addr, link);
 640 }
 641 
 642 // Does the call64_patchable instruction use a pc-relative encoding of
 643 // the call destination?
 644 bool MacroAssembler::is_bxx64_patchable_pcrelative_at(address instruction_addr, bool link) {
 645   // variant 2 is pc-relative
 646   return is_bxx64_patchable_variant2_at(instruction_addr, link);
 647 }
 648 
 649 // Identify variant 1.
 650 bool MacroAssembler::is_bxx64_patchable_variant1_at(address instruction_addr, bool link) {
 651   unsigned int* instr = (unsigned int*) instruction_addr;
 652   return (link ? is_bctrl(instr[6]) : is_bctr(instr[6])) // bctr[l]
 653       && is_mtctr(instr[5]) // mtctr
 654     && is_load_const_at(instruction_addr);
 655 }
 656 
 657 // Identify variant 1b: load destination relative to global toc.
 658 bool MacroAssembler::is_bxx64_patchable_variant1b_at(address instruction_addr, bool link) {
 659   unsigned int* instr = (unsigned int*) instruction_addr;
 660   return (link ? is_bctrl(instr[6]) : is_bctr(instr[6])) // bctr[l]
 661     && is_mtctr(instr[3]) // mtctr
 662     && is_calculate_address_from_global_toc_at(instruction_addr + 2*BytesPerInstWord, instruction_addr);
 663 }
 664 
 665 // Identify variant 2.
 666 bool MacroAssembler::is_bxx64_patchable_variant2_at(address instruction_addr, bool link) {
 667   unsigned int* instr = (unsigned int*) instruction_addr;
 668   if (link) {
 669     return is_bl (instr[6])  // bl dest is last
 670       && is_nop(instr[0])  // nop
 671       && is_nop(instr[1])  // nop
 672       && is_nop(instr[2])  // nop
 673       && is_nop(instr[3])  // nop
 674       && is_nop(instr[4])  // nop
 675       && is_nop(instr[5]); // nop
 676   } else {
 677     return is_b  (instr[0])  // b  dest is first
 678       && is_nop(instr[1])  // nop
 679       && is_nop(instr[2])  // nop
 680       && is_nop(instr[3])  // nop
 681       && is_nop(instr[4])  // nop
 682       && is_nop(instr[5])  // nop
 683       && is_nop(instr[6]); // nop
 684   }
 685 }
 686 
 687 // Set dest address of a bxx64_patchable instruction.
 688 void MacroAssembler::set_dest_of_bxx64_patchable_at(address instruction_addr, address dest, bool link) {
 689   ResourceMark rm;
 690   int code_size = MacroAssembler::bxx64_patchable_size;
 691   CodeBuffer buf(instruction_addr, code_size);
 692   MacroAssembler masm(&buf);
 693   masm.bxx64_patchable(dest, relocInfo::none, link);
 694   ICache::ppc64_flush_icache_bytes(instruction_addr, code_size);
 695 }
 696 
 697 // Get dest address of a bxx64_patchable instruction.
 698 address MacroAssembler::get_dest_of_bxx64_patchable_at(address instruction_addr, bool link) {
 699   if (is_bxx64_patchable_variant1_at(instruction_addr, link)) {
 700     return (address) (unsigned long) get_const(instruction_addr);
 701   } else if (is_bxx64_patchable_variant2_at(instruction_addr, link)) {
 702     unsigned int* instr = (unsigned int*) instruction_addr;
 703     if (link) {
 704       const int instr_idx = 6; // bl is last
 705       int branchoffset = branch_destination(instr[instr_idx], 0);
 706       return instruction_addr + branchoffset + instr_idx*BytesPerInstWord;
 707     } else {
 708       const int instr_idx = 0; // b is first
 709       int branchoffset = branch_destination(instr[instr_idx], 0);
 710       return instruction_addr + branchoffset + instr_idx*BytesPerInstWord;
 711     }
 712   // Load dest relative to global toc.
 713   } else if (is_bxx64_patchable_variant1b_at(instruction_addr, link)) {
 714     return get_address_of_calculate_address_from_global_toc_at(instruction_addr + 2*BytesPerInstWord,
 715                                                                instruction_addr);
 716   } else {
 717     ShouldNotReachHere();
 718     return nullptr;
 719   }
 720 }
 721 
 722 #ifdef ASSERT
 723 void MacroAssembler::clobber_volatile_gprs(Register excluded_register) {
 724   const int magic_number = 0x42;
 725 
 726   // Preserve stack pointer register (R1_SP) and system thread id register (R13);
 727   // although they're technically volatile
 728   for (int i = 2; i < 13; i++) {
 729     Register reg = as_Register(i);
 730     if (reg == excluded_register) {
 731       continue;
 732     }
 733 
 734     li(reg, magic_number);
 735   }
 736 }
 737 
 738 void MacroAssembler::clobber_nonvolatile_registers() {
 739   BLOCK_COMMENT("clobber nonvolatile registers {");
 740   static const Register regs[] = {
 741       R14,
 742       R15,
 743       // don't zap R16_thread
 744       R17,
 745       R18,
 746       R19,
 747       R20,
 748       R21,
 749       R22,
 750       R23,
 751       R24,
 752       R25,
 753       R26,
 754       R27,
 755       R28,
 756       // don't zap R29_TOC
 757       R30,
 758       R31
 759   };
 760   Register bad = regs[0];
 761   load_const_optimized(bad, 0xbad0101babe00000);
 762   for (int i = (sizeof(regs) / sizeof(Register)) - 1; i >= 0; i--) {
 763     addi(regs[i], bad, regs[i]->encoding());
 764   }
 765   BLOCK_COMMENT("} clobber nonvolatile registers");
 766 }
 767 #endif // ASSERT
 768 
 769 void MacroAssembler::clobber_carg_stack_slots(Register tmp) {
 770   const int magic_number = 0x43;
 771 
 772   li(tmp, magic_number);
 773   for (int m = 0; m <= 7; m++) {
 774     std(tmp, frame::native_abi_minframe_size + m * 8, R1_SP);
 775   }
 776 }
 777 
 778 void MacroAssembler::save_nonvolatile_registers(Register dst, int offset, bool include_fp_regs, bool include_vector_regs) {
 779   BLOCK_COMMENT("save_nonvolatile_registers {");
 780 
 781   for (int i = 14; i < 32; i++) {
 782     std(as_Register(i), offset, dst);
 783     offset += 8;
 784   }
 785 
 786   if (include_fp_regs) {
 787     for (int i = 14; i < 32; i++) {
 788       stfd(as_FloatRegister(i), offset, dst);
 789       offset += 8;
 790     }
 791   }
 792 
 793   if (include_vector_regs) {
 794     assert(is_aligned(offset, StackAlignmentInBytes), "should be");
 795     if (PowerArchitecturePPC64 >= 10) {
 796       for (int i = 20; i < 32; i += 2) {
 797         stxvp(as_VectorRegister(i)->to_vsr(), offset, dst);
 798         offset += 32;
 799       }
 800     } else {
 801       for (int i = 20; i < 32; i++) {
 802         stxv(as_VectorRegister(i)->to_vsr(), offset, dst);
 803         offset += 16;
 804       }
 805     }
 806   }
 807 
 808   BLOCK_COMMENT("} save_nonvolatile_registers ");
 809 }
 810 
 811 void MacroAssembler::restore_nonvolatile_registers(Register src, int offset, bool include_fp_regs, bool include_vector_regs) {
 812   BLOCK_COMMENT("restore_nonvolatile_registers {");
 813 
 814   for (int i = 14; i < 32; i++) {
 815     ld(as_Register(i), offset, src);
 816     offset += 8;
 817   }
 818 
 819   if (include_fp_regs) {
 820     for (int i = 14; i < 32; i++) {
 821       lfd(as_FloatRegister(i), offset, src);
 822       offset += 8;
 823     }
 824   }
 825 
 826   if (include_vector_regs) {
 827     assert(is_aligned(offset, StackAlignmentInBytes), "should be");
 828     if (PowerArchitecturePPC64 >= 10) {
 829       for (int i = 20; i < 32; i += 2) {
 830         lxvp(as_VectorRegister(i)->to_vsr(), offset, src);
 831         offset += 32;
 832       }
 833     } else {
 834       for (int i = 20; i < 32; i++) {
 835         lxv(as_VectorRegister(i)->to_vsr(), offset, src);
 836         offset += 16;
 837       }
 838     }
 839   }
 840 
 841   BLOCK_COMMENT("} restore_nonvolatile_registers");
 842 }
 843 
 844 // For verify_oops.
 845 void MacroAssembler::save_volatile_gprs(Register dst, int offset, bool include_fp_regs, bool include_R3_RET_reg) {
 846   std(R2,  offset, dst);   offset += 8;
 847   if (include_R3_RET_reg) {
 848     std(R3, offset, dst);  offset += 8;
 849   }
 850   std(R4,  offset, dst);   offset += 8;
 851   std(R5,  offset, dst);   offset += 8;
 852   std(R6,  offset, dst);   offset += 8;
 853   std(R7,  offset, dst);   offset += 8;
 854   std(R8,  offset, dst);   offset += 8;
 855   std(R9,  offset, dst);   offset += 8;
 856   std(R10, offset, dst);   offset += 8;
 857   std(R11, offset, dst);   offset += 8;
 858   std(R12, offset, dst);   offset += 8;
 859 
 860   if (include_fp_regs) {
 861     stfd(F0, offset, dst);   offset += 8;
 862     stfd(F1, offset, dst);   offset += 8;
 863     stfd(F2, offset, dst);   offset += 8;
 864     stfd(F3, offset, dst);   offset += 8;
 865     stfd(F4, offset, dst);   offset += 8;
 866     stfd(F5, offset, dst);   offset += 8;
 867     stfd(F6, offset, dst);   offset += 8;
 868     stfd(F7, offset, dst);   offset += 8;
 869     stfd(F8, offset, dst);   offset += 8;
 870     stfd(F9, offset, dst);   offset += 8;
 871     stfd(F10, offset, dst);  offset += 8;
 872     stfd(F11, offset, dst);  offset += 8;
 873     stfd(F12, offset, dst);  offset += 8;
 874     stfd(F13, offset, dst);
 875   }
 876 }
 877 
 878 // For verify_oops.
 879 void MacroAssembler::restore_volatile_gprs(Register src, int offset, bool include_fp_regs, bool include_R3_RET_reg) {
 880   ld(R2,  offset, src);   offset += 8;
 881   if (include_R3_RET_reg) {
 882     ld(R3,  offset, src);   offset += 8;
 883   }
 884   ld(R4,  offset, src);   offset += 8;
 885   ld(R5,  offset, src);   offset += 8;
 886   ld(R6,  offset, src);   offset += 8;
 887   ld(R7,  offset, src);   offset += 8;
 888   ld(R8,  offset, src);   offset += 8;
 889   ld(R9,  offset, src);   offset += 8;
 890   ld(R10, offset, src);   offset += 8;
 891   ld(R11, offset, src);   offset += 8;
 892   ld(R12, offset, src);   offset += 8;
 893 
 894   if (include_fp_regs) {
 895     lfd(F0, offset, src);   offset += 8;
 896     lfd(F1, offset, src);   offset += 8;
 897     lfd(F2, offset, src);   offset += 8;
 898     lfd(F3, offset, src);   offset += 8;
 899     lfd(F4, offset, src);   offset += 8;
 900     lfd(F5, offset, src);   offset += 8;
 901     lfd(F6, offset, src);   offset += 8;
 902     lfd(F7, offset, src);   offset += 8;
 903     lfd(F8, offset, src);   offset += 8;
 904     lfd(F9, offset, src);   offset += 8;
 905     lfd(F10, offset, src);  offset += 8;
 906     lfd(F11, offset, src);  offset += 8;
 907     lfd(F12, offset, src);  offset += 8;
 908     lfd(F13, offset, src);
 909   }
 910 }
 911 
 912 void MacroAssembler::save_LR(Register tmp) {
 913   mflr(tmp);
 914   std(tmp, _abi0(lr), R1_SP);
 915 }
 916 
 917 void MacroAssembler::restore_LR(Register tmp) {
 918   assert(tmp != R1_SP, "must be distinct");
 919   ld(tmp, _abi0(lr), R1_SP);
 920   mtlr(tmp);
 921 }
 922 
 923 void MacroAssembler::save_LR_CR(Register tmp) {
 924   mfcr(tmp);
 925   std(tmp, _abi0(cr), R1_SP);
 926   save_LR(tmp);
 927   // Tmp must contain lr on exit! (see return_addr and prolog in ppc64.ad)
 928 }
 929 
 930 void MacroAssembler::restore_LR_CR(Register tmp) {
 931   restore_LR(tmp);
 932   ld(tmp, _abi0(cr), R1_SP);
 933   mtcr(tmp);
 934 }
 935 
 936 address MacroAssembler::get_PC_trash_LR(Register result) {
 937   Label L;
 938   bl(L);
 939   bind(L);
 940   address lr_pc = pc();
 941   mflr(result);
 942   return lr_pc;
 943 }
 944 
 945 void MacroAssembler::resize_frame(Register offset, Register tmp) {
 946 #ifdef ASSERT
 947   assert_different_registers(offset, tmp, R1_SP);
 948   andi_(tmp, offset, frame::alignment_in_bytes-1);
 949   asm_assert_eq("resize_frame: unaligned");
 950 #endif
 951 
 952   // tmp <- *(SP)
 953   ld(tmp, _abi0(callers_sp), R1_SP);
 954   // addr <- SP + offset;
 955   // *(addr) <- tmp;
 956   // SP <- addr
 957   stdux(tmp, R1_SP, offset);
 958 }
 959 
 960 void MacroAssembler::resize_frame(int offset, Register tmp) {
 961   assert(is_simm(offset, 16), "too big an offset");
 962   assert_different_registers(tmp, R1_SP);
 963   assert((offset & (frame::alignment_in_bytes-1))==0, "resize_frame: unaligned");
 964   // tmp <- *(SP)
 965   ld(tmp, _abi0(callers_sp), R1_SP);
 966   // addr <- SP + offset;
 967   // *(addr) <- tmp;
 968   // SP <- addr
 969   stdu(tmp, offset, R1_SP);
 970 }
 971 
 972 void MacroAssembler::resize_frame_absolute(Register addr, Register tmp1, Register tmp2) {
 973   // (addr == tmp1) || (addr == tmp2) is allowed here!
 974   assert(tmp1 != tmp2, "must be distinct");
 975 
 976   // compute offset w.r.t. current stack pointer
 977   // tmp_1 <- addr - SP (!)
 978   subf(tmp1, R1_SP, addr);
 979 
 980   // atomically update SP keeping back link.
 981   resize_frame(tmp1/* offset */, tmp2/* tmp */);
 982 }
 983 
 984 void MacroAssembler::push_frame(Register bytes, Register tmp) {
 985 #ifdef ASSERT
 986   assert(bytes != R0, "r0 not allowed here");
 987   andi_(R0, bytes, frame::alignment_in_bytes-1);
 988   asm_assert_eq("push_frame(Reg, Reg): unaligned");
 989 #endif
 990   neg(tmp, bytes);
 991   stdux(R1_SP, R1_SP, tmp);
 992 }
 993 
 994 // Push a frame of size `bytes'.
 995 void MacroAssembler::push_frame(unsigned int bytes, Register tmp) {
 996   long offset = align_addr(bytes, frame::alignment_in_bytes);
 997   if (is_simm(-offset, 16)) {
 998     stdu(R1_SP, -offset, R1_SP);
 999   } else {
1000     load_const_optimized(tmp, -offset);
1001     stdux(R1_SP, R1_SP, tmp);
1002   }
1003 }
1004 
1005 // Push a frame of size `bytes' plus native_abi_reg_args on top.
1006 void MacroAssembler::push_frame_reg_args(unsigned int bytes, Register tmp) {
1007   push_frame(bytes + frame::native_abi_reg_args_size, tmp);
1008 }
1009 
1010 // Pop current C frame.
1011 void MacroAssembler::pop_frame() {
1012   ld(R1_SP, _abi0(callers_sp), R1_SP);
1013 }
1014 
1015 #if defined(ABI_ELFv2)
1016 address MacroAssembler::branch_to(Register r_function_entry, bool and_link) {
1017   // TODO(asmundak): make sure the caller uses R12 as function descriptor
1018   // most of the times.
1019   if (R12 != r_function_entry) {
1020     mr(R12, r_function_entry);
1021   }
1022   mtctr(R12);
1023   // Do a call or a branch.
1024   if (and_link) {
1025     bctrl();
1026   } else {
1027     bctr();
1028   }
1029   _last_calls_return_pc = pc();
1030 
1031   return _last_calls_return_pc;
1032 }
1033 
1034 // Call a C function via a function descriptor and use full C
1035 // calling conventions. Updates and returns _last_calls_return_pc.
1036 address MacroAssembler::call_c(Register r_function_entry) {
1037   return branch_to(r_function_entry, /*and_link=*/true);
1038 }
1039 
1040 // For tail calls: only branch, don't link, so callee returns to caller of this function.
1041 address MacroAssembler::call_c_and_return_to_caller(Register r_function_entry) {
1042   return branch_to(r_function_entry, /*and_link=*/false);
1043 }
1044 
1045 address MacroAssembler::call_c(address function_entry, relocInfo::relocType rt) {
1046   load_const(R12, function_entry, R0);
1047   return branch_to(R12,  /*and_link=*/true);
1048 }
1049 
1050 #else
1051 // Generic version of a call to C function via a function descriptor
1052 // with variable support for C calling conventions (TOC, ENV, etc.).
1053 // Updates and returns _last_calls_return_pc.
1054 address MacroAssembler::branch_to(Register function_descriptor, bool and_link, bool save_toc_before_call,
1055                                   bool restore_toc_after_call, bool load_toc_of_callee, bool load_env_of_callee) {
1056   // we emit standard ptrgl glue code here
1057   assert((function_descriptor != R0), "function_descriptor cannot be R0");
1058 
1059   // retrieve necessary entries from the function descriptor
1060   ld(R0, in_bytes(FunctionDescriptor::entry_offset()), function_descriptor);
1061   mtctr(R0);
1062 
1063   if (load_toc_of_callee) {
1064     ld(R2_TOC, in_bytes(FunctionDescriptor::toc_offset()), function_descriptor);
1065   }
1066   if (load_env_of_callee) {
1067     ld(R11, in_bytes(FunctionDescriptor::env_offset()), function_descriptor);
1068   } else if (load_toc_of_callee) {
1069     li(R11, 0);
1070   }
1071 
1072   // do a call or a branch
1073   if (and_link) {
1074     bctrl();
1075   } else {
1076     bctr();
1077   }
1078   _last_calls_return_pc = pc();
1079 
1080   return _last_calls_return_pc;
1081 }
1082 
1083 // Call a C function via a function descriptor and use full C calling
1084 // conventions.
1085 // We don't use the TOC in generated code, so there is no need to save
1086 // and restore its value.
1087 address MacroAssembler::call_c(Register fd) {
1088   return branch_to(fd, /*and_link=*/true,
1089                        /*save toc=*/false,
1090                        /*restore toc=*/false,
1091                        /*load toc=*/true,
1092                        /*load env=*/true);
1093 }
1094 
1095 address MacroAssembler::call_c_and_return_to_caller(Register fd) {
1096   return branch_to(fd, /*and_link=*/false,
1097                        /*save toc=*/false,
1098                        /*restore toc=*/false,
1099                        /*load toc=*/true,
1100                        /*load env=*/true);
1101 }
1102 
1103 address MacroAssembler::call_c(const FunctionDescriptor* fd, relocInfo::relocType rt) {
1104   if (rt != relocInfo::none) {
1105     // this call needs to be relocatable
1106     if (!ReoptimizeCallSequences
1107         || (rt != relocInfo::runtime_call_type && rt != relocInfo::none)
1108         || fd == nullptr   // support code-size estimation
1109         || !fd->is_friend_function()
1110         || fd->entry() == nullptr) {
1111       // it's not a friend function as defined by class FunctionDescriptor,
1112       // so do a full call-c here.
1113       load_const(R11, (address)fd, R0);
1114 
1115       bool has_env = (fd != nullptr && fd->env() != nullptr);
1116       return branch_to(R11, /*and_link=*/true,
1117                             /*save toc=*/false,
1118                             /*restore toc=*/false,
1119                             /*load toc=*/true,
1120                             /*load env=*/has_env);
1121     } else {
1122       // It's a friend function. Load the entry point and don't care about
1123       // toc and env. Use an optimizable call instruction, but ensure the
1124       // same code-size as in the case of a non-friend function.
1125       nop();
1126       nop();
1127       nop();
1128       bl64_patchable(fd->entry(), rt);
1129       _last_calls_return_pc = pc();
1130       return _last_calls_return_pc;
1131     }
1132   } else {
1133     // This call does not need to be relocatable, do more aggressive
1134     // optimizations.
1135     if (!ReoptimizeCallSequences
1136       || !fd->is_friend_function()) {
1137       // It's not a friend function as defined by class FunctionDescriptor,
1138       // so do a full call-c here.
1139       load_const(R11, (address)fd, R0);
1140       return branch_to(R11, /*and_link=*/true,
1141                             /*save toc=*/false,
1142                             /*restore toc=*/false,
1143                             /*load toc=*/true,
1144                             /*load env=*/true);
1145     } else {
1146       // it's a friend function, load the entry point and don't care about
1147       // toc and env.
1148       address dest = fd->entry();
1149       if (is_within_range_of_b(dest, pc())) {
1150         bl(dest);
1151       } else {
1152         bl64_patchable(dest, rt);
1153       }
1154       _last_calls_return_pc = pc();
1155       return _last_calls_return_pc;
1156     }
1157   }
1158 }
1159 
1160 // Call a C function.  All constants needed reside in TOC.
1161 //
1162 // Read the address to call from the TOC.
1163 // Read env from TOC, if fd specifies an env.
1164 // Read new TOC from TOC.
1165 address MacroAssembler::call_c_using_toc(const FunctionDescriptor* fd,
1166                                          relocInfo::relocType rt, Register toc) {
1167   if (!ReoptimizeCallSequences
1168     || (rt != relocInfo::runtime_call_type && rt != relocInfo::none)
1169     || !fd->is_friend_function()) {
1170     // It's not a friend function as defined by class FunctionDescriptor,
1171     // so do a full call-c here.
1172     assert(fd->entry() != nullptr, "function must be linked");
1173 
1174     AddressLiteral fd_entry(fd->entry());
1175     bool success = load_const_from_method_toc(R11, fd_entry, toc, /*fixed_size*/ true);
1176     mtctr(R11);
1177     if (fd->env() == nullptr) {
1178       li(R11, 0);
1179       nop();
1180     } else {
1181       AddressLiteral fd_env(fd->env());
1182       success = success && load_const_from_method_toc(R11, fd_env, toc, /*fixed_size*/ true);
1183     }
1184     AddressLiteral fd_toc(fd->toc());
1185     // Set R2_TOC (load from toc)
1186     success = success && load_const_from_method_toc(R2_TOC, fd_toc, toc, /*fixed_size*/ true);
1187     bctrl();
1188     _last_calls_return_pc = pc();
1189     if (!success) { return nullptr; }
1190   } else {
1191     // It's a friend function, load the entry point and don't care about
1192     // toc and env. Use an optimizable call instruction, but ensure the
1193     // same code-size as in the case of a non-friend function.
1194     nop();
1195     bl64_patchable(fd->entry(), rt);
1196     _last_calls_return_pc = pc();
1197   }
1198   return _last_calls_return_pc;
1199 }
1200 #endif // ABI_ELFv2
1201 
1202 bool MacroAssembler::ic_call(Register Rmethod_toc,
1203                              address target,
1204                              jint method_index,
1205                              bool scratch_emit,
1206                              bool fixed_size) {
1207   AddressLiteral target_al(target, virtual_call_Relocation::spec(pc(), method_index));
1208   DEBUG_ONLY(int ic_load_offset = offset());
1209 
1210   // Load a clear inline cache.
1211   AddressLiteral empty_ic((address) Universe::non_oop_word());
1212   bool success = load_const_from_method_toc(R19_inline_cache_reg, empty_ic, Rmethod_toc, fixed_size);
1213   if (!success) return false;
1214 
1215   assert(MacroAssembler::is_load_const_from_method_toc_at(addr_at(ic_load_offset)),
1216          "should be load from TOC");
1217 
1218   address call_pc = trampoline_call(target_al, Rmethod_toc, scratch_emit);
1219   return call_pc != nullptr;
1220 }
1221 
1222 address MacroAssembler::trampoline_call(AddressLiteral target,
1223                                         Register Rmethod_toc,
1224                                         bool scratch_emit) {
1225   // First, emit the trampoline stub
1226   if (!scratch_emit) {
1227     RelocationHolder rh = trampoline_stub_Relocation::spec(pc() /* of the bl below */);
1228 
1229     // Put the target's entry point as a constant into the constant pool.
1230     const address target_toc_addr = address_constant((address)target.value());
1231     if (target_toc_addr == nullptr) return nullptr;
1232 
1233     const int target_toc_offset = offset_to_method_toc(target_toc_addr);
1234     address stub = start_a_stub(64);
1235     if (stub == nullptr) return nullptr;
1236 
1237     // Annotate the stub with a relocation that points to the owning call instruction.
1238     relocate(rh);
1239     DEBUG_ONLY(int stub_start_offset = offset());
1240 
1241     // For java_to_interp stubs we use R11_scratch1 as scratch register
1242     // and in call trampoline stubs we use R12_scratch2. This way we
1243     // can distinguish them (see is_NativeCallTrampolineStub_at()).
1244     Register reg_scratch = R12_scratch2;
1245 
1246     if (Rmethod_toc == noreg) {
1247       calculate_address_from_global_toc(reg_scratch, method_toc());
1248       Rmethod_toc = reg_scratch;
1249     }
1250 
1251     ld_largeoffset_unchecked(reg_scratch, target_toc_offset, Rmethod_toc, false);
1252     mtctr(reg_scratch);
1253     bctr();
1254 
1255     assert(target_toc_offset == NativeCallTrampolineStub_at(addr_at(stub_start_offset))->destination_toc_offset(),
1256            "encoded offset into the constant pool must match");
1257     assert((uint)(offset() - stub_start_offset) <= trampoline_stub_size, "should be good size");
1258     assert(is_NativeCallTrampolineStub_at(addr_at(stub_start_offset)), "doesn't look like a trampoline");
1259 
1260     // End the stub.
1261     end_a_stub();
1262   }
1263 
1264   // The call will be resolved / patched later.
1265   address call_pc = pc();
1266   relocate(target.rspec());
1267   bl(call_pc);
1268   return call_pc;
1269 }
1270 
1271 void MacroAssembler::post_call_nop() {
1272   // Make inline again when loom is always enabled.
1273   if (!Continuations::enabled()) {
1274     return;
1275   }
1276   // We use CMPI/CMPLI instructions to encode post call nops.
1277   // Refer to NativePostCallNop for details.
1278   relocate(post_call_nop_Relocation::spec());
1279   InlineSkippedInstructionsCounter skipCounter(this);
1280   Assembler::emit_int32(Assembler::CMPLI_OPCODE | Assembler::opp_u_field(1, 9, 9));
1281   assert(is_post_call_nop(*(int*)(pc() - 4)), "post call not not found");
1282 }
1283 
1284 int MacroAssembler::ic_check_size() {
1285   bool implicit_null_checks_available = ImplicitNullChecks && os::zero_page_read_protected(),
1286        use_fast_receiver_null_check   = implicit_null_checks_available || TrapBasedNullChecks,
1287        use_trap_based_null_check      = !implicit_null_checks_available && TrapBasedNullChecks;
1288 
1289   int num_ins;
1290   if (use_fast_receiver_null_check && TrapBasedICMissChecks) {
1291     num_ins = 3;
1292     if (use_trap_based_null_check) num_ins += 1;
1293   } else {
1294     num_ins = 7;
1295     if (!implicit_null_checks_available) num_ins += 2;
1296   }
1297 
1298   if (UseCompactObjectHeaders) num_ins++;
1299 
1300   return num_ins * BytesPerInstWord;
1301 }
1302 
1303 int MacroAssembler::ic_check(int end_alignment) {
1304   bool implicit_null_checks_available = ImplicitNullChecks && os::zero_page_read_protected(),
1305        use_fast_receiver_null_check   = implicit_null_checks_available || TrapBasedNullChecks,
1306        use_trap_based_null_check      = !implicit_null_checks_available && TrapBasedNullChecks;
1307 
1308   Register receiver = R3_ARG1;
1309   Register data = R19_inline_cache_reg;
1310   Register tmp1 = R11_scratch1;
1311   Register tmp2 = R12_scratch2;
1312 
1313   // The UEP of a code blob ensures that the VEP is padded. However, the padding of the UEP is placed
1314   // before the inline cache check, so we don't have to execute any nop instructions when dispatching
1315   // through the UEP, yet we can ensure that the VEP is aligned appropriately. That's why we align
1316   // before the inline cache check here, and not after
1317   align(end_alignment, end_alignment, end_alignment - ic_check_size());
1318 
1319   int uep_offset = offset();
1320 
1321   if (use_fast_receiver_null_check && TrapBasedICMissChecks) {
1322     // Fast version which uses SIGTRAP
1323 
1324     if (use_trap_based_null_check) {
1325       trap_null_check(receiver);
1326     }
1327     load_klass_no_decode(tmp1, receiver); // 2 instructions with UseCompactObjectHeaders
1328     ld(tmp2, in_bytes(CompiledICData::speculated_klass_offset()), data);
1329     trap_ic_miss_check(tmp1, tmp2);
1330 
1331   } else {
1332     // Slower version which doesn't use SIGTRAP
1333 
1334     // Load stub address using toc (fixed instruction size, unlike load_const_optimized)
1335     calculate_address_from_global_toc(tmp1, SharedRuntime::get_ic_miss_stub(),
1336                                       true, true, false); // 2 instructions
1337     mtctr(tmp1);
1338 
1339     if (!implicit_null_checks_available) {
1340       cmpdi(CR0, receiver, 0);
1341       beqctr(CR0);
1342     }
1343     load_klass_no_decode(tmp1, receiver); // 2 instructions with UseCompactObjectHeaders
1344     ld(tmp2, in_bytes(CompiledICData::speculated_klass_offset()), data);
1345     cmpd(CR0, tmp1, tmp2);
1346     bnectr(CR0);
1347   }
1348 
1349   assert((offset() % end_alignment) == 0, "Misaligned verified entry point");
1350 
1351   return uep_offset;
1352 }
1353 
1354 void MacroAssembler::call_VM_base(Register oop_result,
1355                                   Register last_java_sp,
1356                                   address  entry_point,
1357                                   bool     check_exceptions,
1358                                   Label*   last_java_pc) {
1359   BLOCK_COMMENT("call_VM {");
1360   // Determine last_java_sp register.
1361   if (!last_java_sp->is_valid()) {
1362     last_java_sp = R1_SP;
1363   }
1364   set_top_ijava_frame_at_SP_as_last_Java_frame(last_java_sp, R11_scratch1, last_java_pc);
1365 
1366   // ARG1 must hold thread address.
1367   mr(R3_ARG1, R16_thread);
1368   address return_pc = call_c(entry_point, relocInfo::none);
1369 
1370   reset_last_Java_frame();
1371 
1372   // Check for pending exceptions.
1373   if (check_exceptions) {
1374     // We don't check for exceptions here.
1375     ShouldNotReachHere();
1376   }
1377 
1378   // Get oop result if there is one and reset the value in the thread.
1379   if (oop_result->is_valid()) {
1380     get_vm_result_oop(oop_result);
1381   }
1382 
1383   _last_calls_return_pc = return_pc;
1384   BLOCK_COMMENT("} call_VM");
1385 }
1386 
1387 void MacroAssembler::call_VM_leaf_base(address entry_point) {
1388   BLOCK_COMMENT("call_VM_leaf {");
1389   call_c(entry_point);
1390   BLOCK_COMMENT("} call_VM_leaf");
1391 }
1392 
1393 void MacroAssembler::call_VM(Register oop_result, address entry_point, bool check_exceptions, Label* last_java_pc) {
1394   call_VM_base(oop_result, noreg, entry_point, check_exceptions, last_java_pc);
1395 }
1396 
1397 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1,
1398                              bool check_exceptions) {
1399   // R3_ARG1 is reserved for the thread.
1400   mr_if_needed(R4_ARG2, arg_1);
1401   call_VM(oop_result, entry_point, check_exceptions);
1402 }
1403 
1404 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2,
1405                              bool check_exceptions) {
1406   // R3_ARG1 is reserved for the thread
1407   assert_different_registers(arg_2, R4_ARG2);
1408   mr_if_needed(R4_ARG2, arg_1);
1409   mr_if_needed(R5_ARG3, arg_2);
1410   call_VM(oop_result, entry_point, check_exceptions);
1411 }
1412 
1413 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, Register arg_3,
1414                              bool check_exceptions) {
1415   // R3_ARG1 is reserved for the thread
1416   assert_different_registers(arg_2, R4_ARG2);
1417   assert_different_registers(arg_3, R4_ARG2, R5_ARG3);
1418   mr_if_needed(R4_ARG2, arg_1);
1419   mr_if_needed(R5_ARG3, arg_2);
1420   mr_if_needed(R6_ARG4, arg_3);
1421   call_VM(oop_result, entry_point, check_exceptions);
1422 }
1423 
1424 void MacroAssembler::call_VM_leaf(address entry_point) {
1425   call_VM_leaf_base(entry_point);
1426 }
1427 
1428 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1) {
1429   mr_if_needed(R3_ARG1, arg_1);
1430   call_VM_leaf(entry_point);
1431 }
1432 
1433 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2) {
1434   assert_different_registers(arg_2, R3_ARG1);
1435   mr_if_needed(R3_ARG1, arg_1);
1436   mr_if_needed(R4_ARG2, arg_2);
1437   call_VM_leaf(entry_point);
1438 }
1439 
1440 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2, Register arg_3) {
1441   assert_different_registers(arg_2, R3_ARG1);
1442   assert_different_registers(arg_3, R3_ARG1, R4_ARG2);
1443   mr_if_needed(R3_ARG1, arg_1);
1444   mr_if_needed(R4_ARG2, arg_2);
1445   mr_if_needed(R5_ARG3, arg_3);
1446   call_VM_leaf(entry_point);
1447 }
1448 
1449 // Check whether instruction is a read access to the polling page
1450 // which was emitted by load_from_polling_page(..).
1451 bool MacroAssembler::is_load_from_polling_page(int instruction, void* ucontext,
1452                                                address* polling_address_ptr) {
1453   if (!is_ld(instruction))
1454     return false; // It's not a ld. Fail.
1455 
1456   int rt = inv_rt_field(instruction);
1457   int ra = inv_ra_field(instruction);
1458   int ds = inv_ds_field(instruction);
1459   if (!(ds == 0 && ra != 0 && rt == 0)) {
1460     return false; // It's not a ld(r0, X, ra). Fail.
1461   }
1462 
1463   if (!ucontext) {
1464     // Set polling address.
1465     if (polling_address_ptr != nullptr) {
1466       *polling_address_ptr = nullptr;
1467     }
1468     return true; // No ucontext given. Can't check value of ra. Assume true.
1469   }
1470 
1471 #ifdef LINUX
1472   // Ucontext given. Check that register ra contains the address of
1473   // the safepoing polling page.
1474   ucontext_t* uc = (ucontext_t*) ucontext;
1475   // Set polling address.
1476   address addr = (address)uc->uc_mcontext.regs->gpr[ra] + (ssize_t)ds;
1477   if (polling_address_ptr != nullptr) {
1478     *polling_address_ptr = addr;
1479   }
1480   return SafepointMechanism::is_poll_address(addr);
1481 #else
1482   // Not on Linux, ucontext must be null.
1483   ShouldNotReachHere();
1484   return false;
1485 #endif
1486 }
1487 
1488 void MacroAssembler::bang_stack_with_offset(int offset) {
1489   // When increasing the stack, the old stack pointer will be written
1490   // to the new top of stack according to the PPC64 abi.
1491   // Therefore, stack banging is not necessary when increasing
1492   // the stack by <= os::vm_page_size() bytes.
1493   // When increasing the stack by a larger amount, this method is
1494   // called repeatedly to bang the intermediate pages.
1495 
1496   // Stack grows down, caller passes positive offset.
1497   assert(offset > 0, "must bang with positive offset");
1498 
1499   long stdoffset = -offset;
1500 
1501   if (is_simm(stdoffset, 16)) {
1502     // Signed 16 bit offset, a simple std is ok.
1503     if (UseLoadInstructionsForStackBangingPPC64) {
1504       ld(R0, (int)(signed short)stdoffset, R1_SP);
1505     } else {
1506       std(R0,(int)(signed short)stdoffset, R1_SP);
1507     }
1508   } else if (is_simm(stdoffset, 31)) {
1509     const int hi = MacroAssembler::largeoffset_si16_si16_hi(stdoffset);
1510     const int lo = MacroAssembler::largeoffset_si16_si16_lo(stdoffset);
1511 
1512     Register tmp = R11;
1513     addis(tmp, R1_SP, hi);
1514     if (UseLoadInstructionsForStackBangingPPC64) {
1515       ld(R0,  lo, tmp);
1516     } else {
1517       std(R0, lo, tmp);
1518     }
1519   } else {
1520     ShouldNotReachHere();
1521   }
1522 }
1523 
1524 // If instruction is a stack bang of the form
1525 //    std    R0,    x(Ry),       (see bang_stack_with_offset())
1526 //    stdu   R1_SP, x(R1_SP),    (see push_frame(), resize_frame())
1527 // or stdux  R1_SP, Rx, R1_SP    (see push_frame(), resize_frame())
1528 // return the banged address. Otherwise, return 0.
1529 address MacroAssembler::get_stack_bang_address(int instruction, void *ucontext) {
1530 #ifdef LINUX
1531   ucontext_t* uc = (ucontext_t*) ucontext;
1532   int rs = inv_rs_field(instruction);
1533   int ra = inv_ra_field(instruction);
1534   if (   (is_ld(instruction)   && rs == 0 &&  UseLoadInstructionsForStackBangingPPC64)
1535       || (is_std(instruction)  && rs == 0 && !UseLoadInstructionsForStackBangingPPC64)
1536       || (is_stdu(instruction) && rs == 1)) {
1537     int ds = inv_ds_field(instruction);
1538     // return banged address
1539     return ds+(address)uc->uc_mcontext.regs->gpr[ra];
1540   } else if (is_stdux(instruction) && rs == 1) {
1541     int rb = inv_rb_field(instruction);
1542     address sp = (address)uc->uc_mcontext.regs->gpr[1];
1543     long rb_val = (long)uc->uc_mcontext.regs->gpr[rb];
1544     return ra != 1 || rb_val >= 0 ? nullptr         // not a stack bang
1545                                   : sp + rb_val; // banged address
1546   }
1547   return nullptr; // not a stack bang
1548 #else
1549   // workaround not needed on !LINUX :-)
1550   ShouldNotCallThis();
1551   return nullptr;
1552 #endif
1553 }
1554 
1555 void MacroAssembler::reserved_stack_check(Register return_pc) {
1556   // Test if reserved zone needs to be enabled.
1557   Label no_reserved_zone_enabling;
1558 
1559   ld_ptr(R0, JavaThread::reserved_stack_activation_offset(), R16_thread);
1560   cmpld(CR0, R1_SP, R0);
1561   blt_predict_taken(CR0, no_reserved_zone_enabling);
1562 
1563   // Enable reserved zone again, throw stack overflow exception.
1564   push_frame_reg_args(0, R0);
1565   call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), R16_thread);
1566   pop_frame();
1567   mtlr(return_pc);
1568   load_const_optimized(R0, SharedRuntime::throw_delayed_StackOverflowError_entry());
1569   mtctr(R0);
1570   bctr();
1571 
1572   should_not_reach_here();
1573 
1574   bind(no_reserved_zone_enabling);
1575 }
1576 
1577 void MacroAssembler::getandsetd(Register dest_current_value, Register exchange_value, Register addr_base,
1578                                 bool cmpxchgx_hint) {
1579   Label retry;
1580   bind(retry);
1581   ldarx(dest_current_value, addr_base, cmpxchgx_hint);
1582   stdcx_(exchange_value, addr_base);
1583   if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1584     bne_predict_not_taken(CR0, retry); // StXcx_ sets CR0.
1585   } else {
1586     bne(                  CR0, retry); // StXcx_ sets CR0.
1587   }
1588 }
1589 
1590 void MacroAssembler::getandaddd(Register dest_current_value, Register inc_value, Register addr_base,
1591                                 Register tmp, bool cmpxchgx_hint) {
1592   Label retry;
1593   bind(retry);
1594   ldarx(dest_current_value, addr_base, cmpxchgx_hint);
1595   add(tmp, dest_current_value, inc_value);
1596   stdcx_(tmp, addr_base);
1597   if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1598     bne_predict_not_taken(CR0, retry); // StXcx_ sets CR0.
1599   } else {
1600     bne(                  CR0, retry); // StXcx_ sets CR0.
1601   }
1602 }
1603 
1604 // Word/sub-word atomic helper functions
1605 
1606 // Temps and addr_base are killed if size < 4 and processor does not support respective instructions.
1607 // Only signed types are supported with size < 4.
1608 // Atomic add always kills tmp1.
1609 void MacroAssembler::atomic_get_and_modify_generic(Register dest_current_value, Register exchange_value,
1610                                                    Register addr_base, Register tmp1, Register tmp2, Register tmp3,
1611                                                    bool cmpxchgx_hint, bool is_add, int size) {
1612   // Sub-word instructions are available since Power 8.
1613 
1614   Label retry;
1615   Register shift_amount = noreg,
1616            val32 = dest_current_value,
1617            modval = is_add ? tmp1 : exchange_value;
1618 
1619 
1620   // atomic emulation loop
1621   bind(retry);
1622 
1623   switch (size) {
1624     case 4: lwarx(val32, addr_base, cmpxchgx_hint); break;
1625     case 2: lharx(val32, addr_base, cmpxchgx_hint); break;
1626     case 1: lbarx(val32, addr_base, cmpxchgx_hint); break;
1627     default: ShouldNotReachHere();
1628   }
1629 
1630   if (is_add) { add(modval, dest_current_value, exchange_value); }
1631 
1632 
1633   switch (size) {
1634     case 4: stwcx_(modval, addr_base); break;
1635     case 2: sthcx_(modval, addr_base); break;
1636     case 1: stbcx_(modval, addr_base); break;
1637     default: ShouldNotReachHere();
1638   }
1639 
1640   if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1641     bne_predict_not_taken(CR0, retry); // StXcx_ sets CR0.
1642   } else {
1643     bne(                  CR0, retry); // StXcx_ sets CR0.
1644   }
1645 
1646   // l?arx zero-extends, but Java wants byte/short values sign-extended.
1647   if (size == 1) {
1648     extsb(dest_current_value, dest_current_value);
1649   } else if (size == 2) {
1650     extsh(dest_current_value, dest_current_value);
1651   };
1652 }
1653 
1654 // Temps, addr_base and exchange_value are killed if size < 4 and processor does not support respective instructions.
1655 // Only signed types are supported with size < 4.
1656 void MacroAssembler::cmpxchg_loop_body(ConditionRegister flag, Register dest_current_value,
1657                                        RegisterOrConstant compare_value, Register exchange_value,
1658                                        Register addr_base, Label &retry, Label &failed, bool cmpxchgx_hint, int size) {
1659   // Sub-word instructions are available since Power 8.
1660   Register shift_amount = noreg,
1661            val32 = dest_current_value,
1662            modval = exchange_value;
1663 
1664   // atomic emulation loop
1665   bind(retry);
1666 
1667   switch (size) {
1668     case 4: lwarx(val32, addr_base, cmpxchgx_hint); break;
1669     case 2: lharx(val32, addr_base, cmpxchgx_hint); break;
1670     case 1: lbarx(val32, addr_base, cmpxchgx_hint); break;
1671     default: ShouldNotReachHere();
1672   }
1673 
1674   if (size == 1) {
1675     extsb(dest_current_value, dest_current_value);
1676   } else if (size == 2) {
1677     extsh(dest_current_value, dest_current_value);
1678   };
1679 
1680   cmpw(flag, dest_current_value, compare_value);
1681   if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1682     bne_predict_not_taken(flag, failed);
1683   } else {
1684     bne(                  flag, failed);
1685   }
1686   // branch to done  => (flag == ne), (dest_current_value != compare_value)
1687   // fall through    => (flag == eq), (dest_current_value == compare_value)
1688 
1689   switch (size) {
1690     case 4: stwcx_(modval, addr_base); break;
1691     case 2: sthcx_(modval, addr_base); break;
1692     case 1: stbcx_(modval, addr_base); break;
1693     default: ShouldNotReachHere();
1694   }
1695 }
1696 
1697 // CmpxchgX sets condition register to cmpX(current, compare).
1698 void MacroAssembler::cmpxchg_generic(ConditionRegister flag, Register dest_current_value,
1699                                      RegisterOrConstant compare_value, Register exchange_value,
1700                                      Register addr_base, int semantics, bool cmpxchgx_hint, Register int_flag_success,
1701                                      Label* failed_ext, bool contention_hint, bool weak, int size) {
1702   Label retry;
1703   Label failed_int;
1704   Label& failed = (failed_ext != nullptr) ? *failed_ext : failed_int;
1705   Label done;
1706 
1707   // Save one branch if result is returned via register and
1708   // result register is different from the other ones.
1709   bool use_result_reg    = (int_flag_success != noreg);
1710   bool preset_result_reg = (int_flag_success != dest_current_value && int_flag_success != compare_value.register_or_noreg() &&
1711                             int_flag_success != exchange_value && int_flag_success != addr_base);
1712   assert(!weak || flag == CR0, "weak only supported with CR0");
1713   assert(int_flag_success == noreg || failed_ext == nullptr, "cannot have both");
1714   assert(size == 1 || size == 2 || size == 4, "unsupported");
1715 
1716   if (use_result_reg && preset_result_reg) {
1717     li(int_flag_success, 0); // preset (assume cas failed)
1718   }
1719 
1720   // Add simple guard in order to reduce risk of starving under high contention (recommended by IBM).
1721   if (contention_hint) { // Don't try to reserve if cmp fails.
1722     switch (size) {
1723       case 1: lbz(dest_current_value, 0, addr_base); extsb(dest_current_value, dest_current_value); break;
1724       case 2: lha(dest_current_value, 0, addr_base); break;
1725       case 4: lwz(dest_current_value, 0, addr_base); break;
1726       default: ShouldNotReachHere();
1727     }
1728     cmpw(flag, dest_current_value, compare_value);
1729     bne(flag, failed);
1730   }
1731 
1732   // release/fence semantics
1733   if (semantics & MemBarRel) {
1734     release();
1735   }
1736 
1737   cmpxchg_loop_body(flag, dest_current_value, compare_value, exchange_value, addr_base,
1738                     retry, failed, cmpxchgx_hint, size);
1739   if (!weak || use_result_reg || failed_ext) {
1740     if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1741       bne_predict_not_taken(CR0, weak ? failed : retry); // StXcx_ sets CR0.
1742     } else {
1743       bne(                  CR0, weak ? failed : retry); // StXcx_ sets CR0.
1744     }
1745   }
1746   // fall through    => (flag == eq), (dest_current_value == compare_value), (swapped)
1747 
1748   // Result in register (must do this at the end because int_flag_success can be the
1749   // same register as one above).
1750   if (use_result_reg) {
1751     li(int_flag_success, 1);
1752   }
1753 
1754   if (semantics & MemBarFenceAfter) {
1755     fence();
1756   } else if (semantics & MemBarAcq) {
1757     isync();
1758   }
1759 
1760   if (use_result_reg && !preset_result_reg) {
1761     b(done);
1762   }
1763 
1764   bind(failed_int);
1765   if (use_result_reg && !preset_result_reg) {
1766     li(int_flag_success, 0);
1767   }
1768 
1769   bind(done);
1770   // (flag == ne) => (dest_current_value != compare_value), (!swapped)
1771   // (flag == eq) => (dest_current_value == compare_value), ( swapped)
1772 }
1773 
1774 // Performs atomic compare exchange:
1775 //   if (compare_value == *addr_base)
1776 //     *addr_base = exchange_value
1777 //     int_flag_success = 1;
1778 //   else
1779 //     int_flag_success = 0;
1780 //
1781 // ConditionRegister flag       = cmp(compare_value, *addr_base)
1782 // Register dest_current_value  = *addr_base
1783 // Register compare_value       Used to compare with value in memory
1784 // Register exchange_value      Written to memory if compare_value == *addr_base
1785 // Register addr_base           The memory location to compareXChange
1786 // Register int_flag_success    Set to 1 if exchange_value was written to *addr_base
1787 //
1788 // To avoid the costly compare exchange the value is tested beforehand.
1789 // Several special cases exist to avoid that unnecessary information is generated.
1790 //
1791 void MacroAssembler::cmpxchgd(ConditionRegister flag, Register dest_current_value,
1792                               RegisterOrConstant compare_value, Register exchange_value,
1793                               Register addr_base,
1794                               int semantics, bool cmpxchgx_hint, Register int_flag_success,
1795                               Label* failed_ext, bool contention_hint, bool weak) {
1796   Label retry;
1797   Label failed_int;
1798   Label& failed = (failed_ext != nullptr) ? *failed_ext : failed_int;
1799   Label done;
1800 
1801   // Save one branch if result is returned via register and result register is different from the other ones.
1802   bool use_result_reg    = (int_flag_success!=noreg);
1803   bool preset_result_reg = (int_flag_success!=dest_current_value && int_flag_success!=compare_value.register_or_noreg() &&
1804                             int_flag_success!=exchange_value && int_flag_success!=addr_base);
1805   assert(!weak || flag == CR0, "weak only supported with CR0");
1806   assert(int_flag_success == noreg || failed_ext == nullptr, "cannot have both");
1807 
1808   if (use_result_reg && preset_result_reg) {
1809     li(int_flag_success, 0); // preset (assume cas failed)
1810   }
1811 
1812   // Add simple guard in order to reduce risk of starving under high contention (recommended by IBM).
1813   if (contention_hint) { // Don't try to reserve if cmp fails.
1814     ld(dest_current_value, 0, addr_base);
1815     cmpd(flag, dest_current_value, compare_value);
1816     bne(flag, failed);
1817   }
1818 
1819   // release/fence semantics
1820   if (semantics & MemBarRel) {
1821     release();
1822   }
1823 
1824   // atomic emulation loop
1825   bind(retry);
1826 
1827   ldarx(dest_current_value, addr_base, cmpxchgx_hint);
1828   cmpd(flag, dest_current_value, compare_value);
1829   if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1830     bne_predict_not_taken(flag, failed);
1831   } else {
1832     bne(                  flag, failed);
1833   }
1834 
1835   stdcx_(exchange_value, addr_base);
1836   if (!weak || use_result_reg || failed_ext) {
1837     if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1838       bne_predict_not_taken(CR0, weak ? failed : retry); // stXcx_ sets CR0
1839     } else {
1840       bne(                  CR0, weak ? failed : retry); // stXcx_ sets CR0
1841     }
1842   }
1843 
1844   // result in register (must do this at the end because int_flag_success can be the same register as one above)
1845   if (use_result_reg) {
1846     li(int_flag_success, 1);
1847   }
1848 
1849   if (semantics & MemBarFenceAfter) {
1850     fence();
1851   } else if (semantics & MemBarAcq) {
1852     isync();
1853   }
1854 
1855   if (use_result_reg && !preset_result_reg) {
1856     b(done);
1857   }
1858 
1859   bind(failed_int);
1860   if (use_result_reg && !preset_result_reg) {
1861     li(int_flag_success, 0);
1862   }
1863 
1864   bind(done);
1865   // (flag == ne) => (dest_current_value != compare_value), (!swapped)
1866   // (flag == eq) => (dest_current_value == compare_value), ( swapped)
1867 }
1868 
1869 // Look up the method for a megamorphic invokeinterface call.
1870 // The target method is determined by <intf_klass, itable_index>.
1871 // The receiver klass is in recv_klass.
1872 // On success, the result will be in method_result, and execution falls through.
1873 // On failure, execution transfers to the given label.
1874 void MacroAssembler::lookup_interface_method(Register recv_klass,
1875                                              Register intf_klass,
1876                                              RegisterOrConstant itable_index,
1877                                              Register method_result,
1878                                              Register scan_temp,
1879                                              Register temp2,
1880                                              Label& L_no_such_interface,
1881                                              bool return_method) {
1882   assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
1883 
1884   // Compute start of first itableOffsetEntry (which is at the end of the vtable).
1885   int vtable_base = in_bytes(Klass::vtable_start_offset());
1886   int itentry_off = in_bytes(itableMethodEntry::method_offset());
1887   int logMEsize   = exact_log2(itableMethodEntry::size() * wordSize);
1888   int scan_step   = itableOffsetEntry::size() * wordSize;
1889   int log_vte_size= exact_log2(vtableEntry::size_in_bytes());
1890 
1891   lwz(scan_temp, in_bytes(Klass::vtable_length_offset()), recv_klass);
1892   // We should store the aligned, prescaled offset in the klass.
1893   // Then the next several instructions would fold away.
1894 
1895   sldi(scan_temp, scan_temp, log_vte_size);
1896   addi(scan_temp, scan_temp, vtable_base);
1897   add(scan_temp, recv_klass, scan_temp);
1898 
1899   // Adjust recv_klass by scaled itable_index, so we can free itable_index.
1900   if (return_method) {
1901     if (itable_index.is_register()) {
1902       Register itable_offset = itable_index.as_register();
1903       sldi(method_result, itable_offset, logMEsize);
1904       if (itentry_off) { addi(method_result, method_result, itentry_off); }
1905       add(method_result, method_result, recv_klass);
1906     } else {
1907       long itable_offset = (long)itable_index.as_constant();
1908       // static address, no relocation
1909       add_const_optimized(method_result, recv_klass, (itable_offset << logMEsize) + itentry_off, temp2);
1910     }
1911   }
1912 
1913   // for (scan = klass->itable(); scan->interface() != nullptr; scan += scan_step) {
1914   //   if (scan->interface() == intf) {
1915   //     result = (klass + scan->offset() + itable_index);
1916   //   }
1917   // }
1918   Label search, found_method;
1919 
1920   for (int peel = 1; peel >= 0; peel--) {
1921     // %%%% Could load both offset and interface in one ldx, if they were
1922     // in the opposite order. This would save a load.
1923     ld(temp2, in_bytes(itableOffsetEntry::interface_offset()), scan_temp);
1924 
1925     // Check that this entry is non-null. A null entry means that
1926     // the receiver class doesn't implement the interface, and wasn't the
1927     // same as when the caller was compiled.
1928     cmpd(CR0, temp2, intf_klass);
1929 
1930     if (peel) {
1931       beq(CR0, found_method);
1932     } else {
1933       bne(CR0, search);
1934       // (invert the test to fall through to found_method...)
1935     }
1936 
1937     if (!peel) break;
1938 
1939     bind(search);
1940 
1941     cmpdi(CR0, temp2, 0);
1942     beq(CR0, L_no_such_interface);
1943     addi(scan_temp, scan_temp, scan_step);
1944   }
1945 
1946   bind(found_method);
1947 
1948   // Got a hit.
1949   if (return_method) {
1950     int ito_offset = in_bytes(itableOffsetEntry::offset_offset());
1951     lwz(scan_temp, ito_offset, scan_temp);
1952     ldx(method_result, scan_temp, method_result);
1953   }
1954 }
1955 
1956 // virtual method calling
1957 void MacroAssembler::lookup_virtual_method(Register recv_klass,
1958                                            RegisterOrConstant vtable_index,
1959                                            Register method_result) {
1960 
1961   assert_different_registers(recv_klass, method_result, vtable_index.register_or_noreg());
1962 
1963   const ByteSize base = Klass::vtable_start_offset();
1964   assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1965 
1966   if (vtable_index.is_register()) {
1967     sldi(vtable_index.as_register(), vtable_index.as_register(), LogBytesPerWord);
1968     add(recv_klass, vtable_index.as_register(), recv_klass);
1969   } else {
1970     addi(recv_klass, recv_klass, vtable_index.as_constant() << LogBytesPerWord);
1971   }
1972   ld(R19_method, in_bytes(base + vtableEntry::method_offset()), recv_klass);
1973 }
1974 
1975 /////////////////////////////////////////// subtype checking ////////////////////////////////////////////
1976 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
1977                                                    Register super_klass,
1978                                                    Register temp1_reg,
1979                                                    Register temp2_reg,
1980                                                    Label* L_success,
1981                                                    Label* L_failure,
1982                                                    Label* L_slow_path,
1983                                                    RegisterOrConstant super_check_offset) {
1984 
1985   const Register check_cache_offset = temp1_reg;
1986   const Register cached_super       = temp2_reg;
1987 
1988   assert_different_registers(sub_klass, super_klass, check_cache_offset, cached_super);
1989 
1990   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1991   int sc_offset  = in_bytes(Klass::secondary_super_cache_offset());
1992 
1993   bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
1994   bool need_slow_path = (must_load_sco || super_check_offset.constant_or_zero() == sco_offset);
1995 
1996   Label L_fallthrough;
1997   int label_nulls = 0;
1998   if (L_success == nullptr)   { L_success   = &L_fallthrough; label_nulls++; }
1999   if (L_failure == nullptr)   { L_failure   = &L_fallthrough; label_nulls++; }
2000   if (L_slow_path == nullptr) { L_slow_path = &L_fallthrough; label_nulls++; }
2001   assert(label_nulls <= 1 ||
2002          (L_slow_path == &L_fallthrough && label_nulls <= 2 && !need_slow_path),
2003          "at most one null in the batch, usually");
2004 
2005   // If the pointers are equal, we are done (e.g., String[] elements).
2006   // This self-check enables sharing of secondary supertype arrays among
2007   // non-primary types such as array-of-interface. Otherwise, each such
2008   // type would need its own customized SSA.
2009   // We move this check to the front of the fast path because many
2010   // type checks are in fact trivially successful in this manner,
2011   // so we get a nicely predicted branch right at the start of the check.
2012   cmpd(CR0, sub_klass, super_klass);
2013   beq(CR0, *L_success);
2014 
2015   // Check the supertype display:
2016   if (must_load_sco) {
2017     // The super check offset is always positive...
2018     lwz(check_cache_offset, sco_offset, super_klass);
2019     super_check_offset = RegisterOrConstant(check_cache_offset);
2020     // super_check_offset is register.
2021     assert_different_registers(sub_klass, super_klass, cached_super, super_check_offset.as_register());
2022   }
2023   // The loaded value is the offset from Klass.
2024 
2025   ld(cached_super, super_check_offset, sub_klass);
2026   cmpd(CR0, cached_super, super_klass);
2027 
2028   // This check has worked decisively for primary supers.
2029   // Secondary supers are sought in the super_cache ('super_cache_addr').
2030   // (Secondary supers are interfaces and very deeply nested subtypes.)
2031   // This works in the same check above because of a tricky aliasing
2032   // between the super_cache and the primary super display elements.
2033   // (The 'super_check_addr' can address either, as the case requires.)
2034   // Note that the cache is updated below if it does not help us find
2035   // what we need immediately.
2036   // So if it was a primary super, we can just fail immediately.
2037   // Otherwise, it's the slow path for us (no success at this point).
2038 
2039 #define FINAL_JUMP(label) if (&(label) != &L_fallthrough) { b(label); }
2040 
2041   if (super_check_offset.is_register()) {
2042     beq(CR0, *L_success);
2043     cmpwi(CR0, super_check_offset.as_register(), sc_offset);
2044     if (L_failure == &L_fallthrough) {
2045       beq(CR0, *L_slow_path);
2046     } else {
2047       bne(CR0, *L_failure);
2048       FINAL_JUMP(*L_slow_path);
2049     }
2050   } else {
2051     if (super_check_offset.as_constant() == sc_offset) {
2052       // Need a slow path; fast failure is impossible.
2053       if (L_slow_path == &L_fallthrough) {
2054         beq(CR0, *L_success);
2055       } else {
2056         bne(CR0, *L_slow_path);
2057         FINAL_JUMP(*L_success);
2058       }
2059     } else {
2060       // No slow path; it's a fast decision.
2061       if (L_failure == &L_fallthrough) {
2062         beq(CR0, *L_success);
2063       } else {
2064         bne(CR0, *L_failure);
2065         FINAL_JUMP(*L_success);
2066       }
2067     }
2068   }
2069 
2070   bind(L_fallthrough);
2071 #undef FINAL_JUMP
2072 }
2073 
2074 void MacroAssembler::check_klass_subtype_slow_path_linear(Register sub_klass,
2075                                                           Register super_klass,
2076                                                           Register temp1_reg,
2077                                                           Register temp2_reg,
2078                                                           Label* L_success,
2079                                                           Register result_reg) {
2080   const Register array_ptr = temp1_reg; // current value from cache array
2081   const Register temp      = temp2_reg;
2082 
2083   assert_different_registers(sub_klass, super_klass, array_ptr, temp);
2084   assert(L_success == nullptr || result_reg == noreg, "can't have both");
2085 
2086   int source_offset = in_bytes(Klass::secondary_supers_offset());
2087   int target_offset = in_bytes(Klass::secondary_super_cache_offset());
2088 
2089   int length_offset = Array<Klass*>::length_offset_in_bytes();
2090   int base_offset   = Array<Klass*>::base_offset_in_bytes();
2091 
2092   Label hit, loop, failure, fallthru;
2093 
2094   ld(array_ptr, source_offset, sub_klass);
2095 
2096   // TODO: PPC port: assert(4 == arrayOopDesc::length_length_in_bytes(), "precondition violated.");
2097   lwz(temp, length_offset, array_ptr);
2098   cmpwi(CR0, temp, 0);
2099   beq(CR0, (L_success == nullptr) ? failure : fallthru); // indicate failure if length 0
2100 
2101   mtctr(temp); // load ctr
2102 
2103   bind(loop);
2104   // Oops in table are NO MORE compressed.
2105   ld(temp, base_offset, array_ptr);
2106   cmpd(CR0, temp, super_klass);
2107   beq(CR0, hit);
2108   addi(array_ptr, array_ptr, BytesPerWord);
2109   bdnz(loop);
2110 
2111   bind(failure);
2112   if (result_reg != noreg) {
2113     li(result_reg, 1); // load non-zero result (indicates a miss)
2114   } else if (L_success == nullptr) {
2115     crandc(CR0, Assembler::equal, CR0, Assembler::equal); // miss indicated by CR0.ne
2116   }
2117   b(fallthru);
2118 
2119   bind(hit);
2120   std(super_klass, target_offset, sub_klass); // save result to cache
2121   if (result_reg != noreg) {
2122     li(result_reg, 0); // load zero result (indicates a hit)
2123   } else if (L_success != nullptr) {
2124     b(*L_success);
2125   }
2126 
2127   bind(fallthru);
2128 }
2129 
2130 Register MacroAssembler::allocate_if_noreg(Register r,
2131                                   RegSetIterator<Register> &available_regs,
2132                                   RegSet &regs_to_push) {
2133   if (!r->is_valid()) {
2134     r = *available_regs++;
2135     regs_to_push += r;
2136   }
2137   return r;
2138 }
2139 
2140 void MacroAssembler::push_set(RegSet set)
2141 {
2142   int spill_offset = 0;
2143   for (RegSetIterator<Register> it = set.begin(); *it != noreg; ++it) {
2144     spill_offset += wordSize;
2145     std(*it, -spill_offset, R1_SP);
2146   }
2147 }
2148 
2149 void MacroAssembler::pop_set(RegSet set)
2150 {
2151   int spill_offset = 0;
2152   for (RegSetIterator<Register> it = set.begin(); *it != noreg; ++it) {
2153     spill_offset += wordSize;
2154     ld(*it, -spill_offset, R1_SP);
2155   }
2156 }
2157 
2158 void MacroAssembler::check_klass_subtype_slow_path_table(Register sub_klass,
2159                                                          Register super_klass,
2160                                                          Register temp1_reg,
2161                                                          Register temp2_reg,
2162                                                          Label* L_success,
2163                                                          Register result_reg) {
2164   RegSet temps = RegSet::of(temp1_reg, temp2_reg);
2165 
2166   assert_different_registers(sub_klass, super_klass, temp1_reg, temp2_reg, result_reg, R0);
2167 
2168   Register temp3_reg = noreg, temp4_reg = noreg;
2169   bool result_reg_provided = (result_reg != noreg); // otherwise, result will be in CR0
2170 
2171   BLOCK_COMMENT("check_klass_subtype_slow_path_table");
2172 
2173   RegSetIterator<Register> available_regs
2174     = (RegSet::range(R2, R12) - temps - sub_klass - super_klass).begin();
2175 
2176   RegSet pushed_regs;
2177 
2178   temp1_reg = allocate_if_noreg(temp1_reg, available_regs, pushed_regs);
2179   temp2_reg = allocate_if_noreg(temp2_reg, available_regs, pushed_regs);
2180   temp3_reg = allocate_if_noreg(temp3_reg, available_regs, pushed_regs);
2181   temp4_reg = allocate_if_noreg(temp4_reg, available_regs, pushed_regs);
2182   result_reg = allocate_if_noreg(result_reg, available_regs, pushed_regs);
2183 
2184   push_set(pushed_regs);
2185 
2186   lookup_secondary_supers_table_var(sub_klass, super_klass,
2187                                     temp1_reg, temp2_reg, temp3_reg, temp4_reg,
2188                                     result_reg);
2189 
2190   if (L_success != nullptr || !result_reg_provided) {
2191     // result_reg may get overwritten by pop_set
2192     cmpdi(CR0, result_reg, 0);
2193   }
2194 
2195   // Unspill the temp. registers:
2196   pop_set(pushed_regs);
2197 
2198   if (L_success != nullptr) {
2199     beq(CR0, *L_success);
2200   }
2201 }
2202 
2203 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
2204                                                    Register super_klass,
2205                                                    Register temp1_reg,
2206                                                    Register temp2_reg,
2207                                                    Label* L_success,
2208                                                    Register result_reg) {
2209   if (UseSecondarySupersTable) {
2210     check_klass_subtype_slow_path_table(sub_klass, super_klass, temp1_reg, temp2_reg, L_success, result_reg);
2211   } else {
2212     if (temp2_reg == noreg) temp2_reg = R0;
2213     check_klass_subtype_slow_path_linear(sub_klass, super_klass, temp1_reg, temp2_reg, L_success, result_reg);
2214   }
2215 }
2216 
2217 // Try fast path, then go to slow one if not successful
2218 void MacroAssembler::check_klass_subtype(Register sub_klass,
2219                          Register super_klass,
2220                          Register temp1_reg,
2221                          Register temp2_reg,
2222                          Label& L_success) {
2223   Label L_failure;
2224   check_klass_subtype_fast_path(sub_klass, super_klass, temp1_reg, temp2_reg, &L_success, &L_failure);
2225   check_klass_subtype_slow_path(sub_klass, super_klass, temp1_reg, temp2_reg, &L_success);
2226   bind(L_failure); // Fallthru if not successful.
2227 }
2228 
2229 // scans count pointer sized words at [addr] for occurrence of value,
2230 // generic (count must be >0)
2231 // iff found: CR0 eq, scratch == 0
2232 void MacroAssembler::repne_scan(Register addr, Register value, Register count, Register scratch) {
2233   Label Lloop, Lafter_loop, Lexit;
2234 
2235   srdi_(scratch, count, 1);
2236   beq(CR0, Lafter_loop);
2237   mtctr(scratch);
2238 
2239   bind(Lloop); // 2x unrolled
2240   ld(scratch, 0, addr);
2241   xor_(scratch, scratch, value);
2242   beq(CR0, Lexit);
2243   ld(scratch, 8, addr);
2244   xor_(scratch, scratch, value);
2245   beq(CR0, Lexit);
2246   addi(addr, addr, 2 * wordSize);
2247   bdnz(Lloop);
2248 
2249   bind(Lafter_loop);
2250   andi_(scratch, count, 1);
2251   beq(CR0, Lexit); // if taken: CR0 eq and scratch == 0
2252   ld(scratch, 0, addr);
2253   xor_(scratch, scratch, value);
2254 
2255   bind(Lexit);
2256 }
2257 
2258 // Ensure that the inline code and the stub are using the same registers.
2259 #define LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS                       \
2260 do {                                                                  \
2261   assert(r_super_klass  == R4_ARG2                                 && \
2262          r_array_base   == R3_ARG1                                 && \
2263          r_array_length == R7_ARG5                                 && \
2264          (r_array_index == R6_ARG4      || r_array_index == noreg) && \
2265          (r_sub_klass   == R5_ARG3      || r_sub_klass   == noreg) && \
2266          (r_bitmap      == R11_scratch1 || r_bitmap      == noreg) && \
2267          (result        == R8_ARG6      || result        == noreg), "registers must match ppc64.ad"); \
2268 } while(0)
2269 
2270 void MacroAssembler::lookup_secondary_supers_table_const(Register r_sub_klass,
2271                                                          Register r_super_klass,
2272                                                          Register temp1,
2273                                                          Register temp2,
2274                                                          Register temp3,
2275                                                          Register temp4,
2276                                                          Register result,
2277                                                          u1 super_klass_slot) {
2278   assert_different_registers(r_sub_klass, r_super_klass, temp1, temp2, temp3, temp4, result);
2279 
2280   Label L_done;
2281 
2282   BLOCK_COMMENT("lookup_secondary_supers_table_const {");
2283 
2284   const Register
2285     r_array_base   = temp1,
2286     r_array_length = temp2,
2287     r_array_index  = temp3,
2288     r_bitmap       = temp4;
2289 
2290   LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS; // Required for stub call below.
2291 
2292   ld(r_bitmap, in_bytes(Klass::secondary_supers_bitmap_offset()), r_sub_klass);
2293 
2294   // First check the bitmap to see if super_klass might be present. If
2295   // the bit is zero, we are certain that super_klass is not one of
2296   // the secondary supers.
2297   u1 bit = super_klass_slot;
2298   int shift_count = Klass::SECONDARY_SUPERS_TABLE_MASK - bit;
2299 
2300   // if (shift_count == 0) this is used for comparing with 0:
2301   sldi_(r_array_index, r_bitmap, shift_count);
2302 
2303   li(result, 1); // failure
2304   // We test the MSB of r_array_index, i.e. its sign bit
2305   bge(CR0, L_done);
2306 
2307   // We will consult the secondary-super array.
2308   ld(r_array_base, in_bytes(Klass::secondary_supers_offset()), r_sub_klass);
2309 
2310   // The value i in r_array_index is >= 1, so even though r_array_base
2311   // points to the length, we don't need to adjust it to point to the
2312   // data.
2313   assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "Adjust this code");
2314 
2315   // Get the first array index that can contain super_klass.
2316   if (bit != 0) {
2317     popcntd(r_array_index, r_array_index);
2318     // NB! r_array_index is off by 1. It is compensated by keeping r_array_base off by 1 word.
2319     sldi(r_array_index, r_array_index, LogBytesPerWord); // scale
2320     ldx(result, r_array_base, r_array_index);
2321   } else {
2322     // Actually use index 0, but r_array_base and r_array_index are off by 1 word
2323     // such that the sum is precise.
2324     ld(result, BytesPerWord, r_array_base);
2325     li(r_array_index, BytesPerWord); // for slow path (scaled)
2326   }
2327 
2328   xor_(result, result, r_super_klass);
2329   beq(CR0, L_done); // Found a match (result == 0)
2330 
2331   // Is there another entry to check? Consult the bitmap.
2332   testbitdi(CR0, /* temp */ r_array_length, r_bitmap, (bit + 1) & Klass::SECONDARY_SUPERS_TABLE_MASK);
2333   beq(CR0, L_done); // (result != 0)
2334 
2335   // Linear probe. Rotate the bitmap so that the next bit to test is
2336   // in Bit 2 for the look-ahead check in the slow path.
2337   if (bit != 0) {
2338     rldicl(r_bitmap, r_bitmap, 64 - bit, 0);
2339   }
2340 
2341   // Calls into the stub generated by lookup_secondary_supers_table_slow_path.
2342   // Arguments: r_super_klass, r_array_base, r_array_index, r_bitmap.
2343   // Kills: r_array_length.
2344   // Returns: result.
2345   address stub = StubRoutines::lookup_secondary_supers_table_slow_path_stub();
2346   Register r_stub_addr = r_array_length;
2347   add_const_optimized(r_stub_addr, R29_TOC, MacroAssembler::offset_to_global_toc(stub), R0);
2348   mtctr(r_stub_addr);
2349   bctrl();
2350 
2351   bind(L_done);
2352   BLOCK_COMMENT("} lookup_secondary_supers_table_const");
2353 
2354   if (VerifySecondarySupers) {
2355     verify_secondary_supers_table(r_sub_klass, r_super_klass, result,
2356                                   temp1, temp2, temp3);
2357   }
2358 }
2359 
2360 // At runtime, return 0 in result if r_super_klass is a superclass of
2361 // r_sub_klass, otherwise return nonzero. Use this version of
2362 // lookup_secondary_supers_table() if you don't know ahead of time
2363 // which superclass will be searched for. Used by interpreter and
2364 // runtime stubs. It is larger and has somewhat greater latency than
2365 // the version above, which takes a constant super_klass_slot.
2366 void MacroAssembler::lookup_secondary_supers_table_var(Register r_sub_klass,
2367                                                        Register r_super_klass,
2368                                                        Register temp1,
2369                                                        Register temp2,
2370                                                        Register temp3,
2371                                                        Register temp4,
2372                                                        Register result) {
2373   assert_different_registers(r_sub_klass, r_super_klass, temp1, temp2, temp3, temp4, result, R0);
2374 
2375   Label L_done;
2376 
2377   BLOCK_COMMENT("lookup_secondary_supers_table_var {");
2378 
2379   const Register
2380     r_array_base   = temp1,
2381     slot           = temp2,
2382     r_array_index  = temp3,
2383     r_bitmap       = temp4;
2384 
2385   lbz(slot, in_bytes(Klass::hash_slot_offset()), r_super_klass);
2386   ld(r_bitmap, in_bytes(Klass::secondary_supers_bitmap_offset()), r_sub_klass);
2387 
2388   li(result, 1); // Make sure that result is nonzero if the test below misses.
2389 
2390   // First check the bitmap to see if super_klass might be present. If
2391   // the bit is zero, we are certain that super_klass is not one of
2392   // the secondary supers.
2393   xori(R0, slot, Klass::SECONDARY_SUPERS_TABLE_SIZE - 1); // slot ^ 63 === 63 - slot (mod 64)
2394   sld_(r_array_index, r_bitmap, R0); // shift left by 63-slot
2395 
2396   // We test the MSB of r_array_index, i.e. its sign bit
2397   bge(CR0, L_done);
2398 
2399   // We will consult the secondary-super array.
2400   ld(r_array_base, in_bytes(Klass::secondary_supers_offset()), r_sub_klass);
2401 
2402   // The value i in r_array_index is >= 1, so even though r_array_base
2403   // points to the length, we don't need to adjust it to point to the data.
2404   assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "Adjust this code");
2405   assert(Array<Klass*>::length_offset_in_bytes() == 0, "Adjust this code");
2406 
2407   // Get the first array index that can contain super_klass into r_array_index.
2408   popcntd(r_array_index, r_array_index);
2409 
2410   // NB! r_array_index is off by 1. It is compensated by keeping r_array_base off by 1 word.
2411   sldi(r_array_index, r_array_index, LogBytesPerWord); // scale
2412 
2413   ldx(R0, r_array_base, r_array_index);
2414   xor_(result, R0, r_super_klass);
2415   beq(CR0, L_done); // found a match, result is 0 in this case
2416 
2417   // Linear probe. Rotate the bitmap so that the next bit to test is
2418   // in Bit 1.
2419   neg(R0, slot); // rotate right
2420   rldcl(r_bitmap, r_bitmap, R0, 0);
2421   Register temp = slot;
2422   andi_(temp, r_bitmap, 2);
2423   beq(CR0, L_done); // fail (result != 0)
2424 
2425   // The slot we just inspected is at secondary_supers[r_array_index - 1].
2426   // The next slot to be inspected, by the logic we're about to call,
2427   // is secondary_supers[r_array_index]. Bits 0 and 1 in the bitmap
2428   // have been checked.
2429   lookup_secondary_supers_table_slow_path(r_super_klass, r_array_base, r_array_index,
2430                                           r_bitmap, result, temp);
2431   // return whatever we got from slow path
2432 
2433   bind(L_done);
2434 
2435   BLOCK_COMMENT("} lookup_secondary_supers_table_var");
2436 
2437   if (VerifySecondarySupers) {
2438     verify_secondary_supers_table(r_sub_klass, r_super_klass, result,
2439                                   temp1, temp2, temp3);
2440   }
2441 }
2442 
2443 // Called by code generated by check_klass_subtype_slow_path
2444 // above. This is called when there is a collision in the hashed
2445 // lookup in the secondary supers array.
2446 void MacroAssembler::lookup_secondary_supers_table_slow_path(Register r_super_klass,
2447                                                              Register r_array_base,
2448                                                              Register r_array_index,
2449                                                              Register r_bitmap,
2450                                                              Register result,
2451                                                              Register temp1) {
2452   assert_different_registers(r_super_klass, r_array_base, r_array_index, r_bitmap, result, temp1);
2453 
2454   const Register
2455     r_array_length = temp1,
2456     r_sub_klass    = noreg;
2457 
2458   Label L_done;
2459 
2460   // Load the array length.
2461   lwa(r_array_length, Array<Klass*>::length_offset_in_bytes(), r_array_base);
2462   // And adjust the array base to point to the data.
2463   // NB! Effectively increments current slot index by 1.
2464   assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "");
2465   addi(r_array_base, r_array_base, Array<Klass*>::base_offset_in_bytes());
2466 
2467   // Linear probe
2468   Label L_huge;
2469 
2470   // The bitmap is full to bursting.
2471   // Implicit invariant: BITMAP_FULL implies (length > 0)
2472   cmpwi(CR0, r_array_length, (int32_t)Klass::SECONDARY_SUPERS_TABLE_SIZE - 2);
2473   bgt(CR0, L_huge);
2474 
2475   // NB! Our caller has checked bits 0 and 1 in the bitmap. The
2476   // current slot (at secondary_supers[r_array_index]) has not yet
2477   // been inspected, and r_array_index may be out of bounds if we
2478   // wrapped around the end of the array.
2479 
2480   { // This is conventional linear probing, but instead of terminating
2481     // when a null entry is found in the table, we maintain a bitmap
2482     // in which a 0 indicates missing entries.
2483     // The check above guarantees there are 0s in the bitmap, so the loop
2484     // eventually terminates.
2485 
2486 #ifdef ASSERT
2487     {
2488       // We should only reach here after having found a bit in the bitmap.
2489       // Invariant: array_length == popcount(bitmap)
2490       Label ok;
2491       cmpdi(CR0, r_array_length, 0);
2492       bgt(CR0, ok);
2493       stop("array_length must be positive");
2494       bind(ok);
2495     }
2496 #endif
2497 
2498     // Compute limit in r_array_length
2499     addi(r_array_length, r_array_length, -1);
2500     sldi(r_array_length, r_array_length, LogBytesPerWord);
2501 
2502     Label L_loop;
2503     bind(L_loop);
2504 
2505     // Check for wraparound.
2506     cmpd(CR0, r_array_index, r_array_length);
2507     isel_0(r_array_index, CR0, Assembler::greater);
2508 
2509     ldx(result, r_array_base, r_array_index);
2510     xor_(result, result, r_super_klass);
2511     beq(CR0, L_done); // success (result == 0)
2512 
2513     // look-ahead check (Bit 2); result is non-zero
2514     testbitdi(CR0, R0, r_bitmap, 2);
2515     beq(CR0, L_done); // fail (result != 0)
2516 
2517     rldicl(r_bitmap, r_bitmap, 64 - 1, 0);
2518     addi(r_array_index, r_array_index, BytesPerWord);
2519     b(L_loop);
2520   }
2521 
2522   { // Degenerate case: more than 64 secondary supers.
2523     // FIXME: We could do something smarter here, maybe a vectorized
2524     // comparison or a binary search, but is that worth any added
2525     // complexity?
2526     bind(L_huge);
2527     repne_scan(r_array_base, r_super_klass, r_array_length, result);
2528   }
2529 
2530   bind(L_done);
2531 }
2532 
2533 // Make sure that the hashed lookup and a linear scan agree.
2534 void MacroAssembler::verify_secondary_supers_table(Register r_sub_klass,
2535                                                    Register r_super_klass,
2536                                                    Register result,
2537                                                    Register temp1,
2538                                                    Register temp2,
2539                                                    Register temp3) {
2540   assert_different_registers(r_sub_klass, r_super_klass, result, temp1, temp2, temp3);
2541 
2542   const Register
2543     r_array_base   = temp1,
2544     r_array_length = temp2,
2545     r_array_index  = temp3,
2546     r_bitmap       = noreg; // unused
2547 
2548   BLOCK_COMMENT("verify_secondary_supers_table {");
2549 
2550   Label passed, failure;
2551 
2552   // We will consult the secondary-super array.
2553   ld(r_array_base, in_bytes(Klass::secondary_supers_offset()), r_sub_klass);
2554   // Load the array length.
2555   lwa(r_array_length, Array<Klass*>::length_offset_in_bytes(), r_array_base);
2556   // And adjust the array base to point to the data.
2557   addi(r_array_base, r_array_base, Array<Klass*>::base_offset_in_bytes());
2558 
2559   // convert !=0 to 1
2560   normalize_bool(result, R0, true);
2561   const Register linear_result = r_array_index; // reuse
2562   li(linear_result, 1);
2563   cmpdi(CR0, r_array_length, 0);
2564   ble(CR0, failure);
2565   repne_scan(r_array_base, r_super_klass, r_array_length, linear_result);
2566   bind(failure);
2567 
2568   // convert !=0 to 1
2569   normalize_bool(linear_result, R0, true);
2570 
2571   cmpd(CR0, result, linear_result);
2572   beq(CR0, passed);
2573 
2574   // report fatal error and terminate VM
2575 
2576   // Argument shuffle. Using stack to avoid clashes.
2577   std(r_super_klass, -8, R1_SP);
2578   std(r_sub_klass, -16, R1_SP);
2579   std(linear_result, -24, R1_SP);
2580   mr_if_needed(R6_ARG4, result);
2581   ld(R3_ARG1, -8, R1_SP);
2582   ld(R4_ARG2, -16, R1_SP);
2583   ld(R5_ARG3, -24, R1_SP);
2584 
2585   const char* msg = "mismatch";
2586   load_const_optimized(R7_ARG5, (intptr_t)msg, R0);
2587   call_VM_leaf(CAST_FROM_FN_PTR(address, Klass::on_secondary_supers_verification_failure));
2588   should_not_reach_here();
2589 
2590   bind(passed);
2591 
2592   BLOCK_COMMENT("} verify_secondary_supers_table");
2593 }
2594 
2595 void MacroAssembler::clinit_barrier(Register klass, Register thread, Label* L_fast_path, Label* L_slow_path) {
2596   assert(L_fast_path != nullptr || L_slow_path != nullptr, "at least one is required");
2597 
2598   Label L_check_thread, L_fallthrough;
2599   if (L_fast_path == nullptr) {
2600     L_fast_path = &L_fallthrough;
2601   } else if (L_slow_path == nullptr) {
2602     L_slow_path = &L_fallthrough;
2603   }
2604 
2605   // Fast path check: class is fully initialized
2606   lbz(R0, in_bytes(InstanceKlass::init_state_offset()), klass);
2607   // acquire by cmp-branch-isync if fully_initialized
2608   cmpwi(CR0, R0, InstanceKlass::fully_initialized);
2609   bne(CR0, L_check_thread);
2610   isync();
2611   b(*L_fast_path);
2612 
2613   // Fast path check: current thread is initializer thread
2614   bind(L_check_thread);
2615   ld(R0, in_bytes(InstanceKlass::init_thread_offset()), klass);
2616   cmpd(CR0, thread, R0);
2617   if (L_slow_path == &L_fallthrough) {
2618     beq(CR0, *L_fast_path);
2619   } else if (L_fast_path == &L_fallthrough) {
2620     bne(CR0, *L_slow_path);
2621   } else {
2622     Unimplemented();
2623   }
2624 
2625   bind(L_fallthrough);
2626 }
2627 
2628 RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot,
2629                                                    Register temp_reg,
2630                                                    int extra_slot_offset) {
2631   // cf. TemplateTable::prepare_invoke(), if (load_receiver).
2632   int stackElementSize = Interpreter::stackElementSize;
2633   int offset = extra_slot_offset * stackElementSize;
2634   if (arg_slot.is_constant()) {
2635     offset += arg_slot.as_constant() * stackElementSize;
2636     return offset;
2637   } else {
2638     assert(temp_reg != noreg, "must specify");
2639     sldi(temp_reg, arg_slot.as_register(), exact_log2(stackElementSize));
2640     if (offset != 0)
2641       addi(temp_reg, temp_reg, offset);
2642     return temp_reg;
2643   }
2644 }
2645 
2646 void MacroAssembler::tlab_allocate(
2647   Register obj,                      // result: pointer to object after successful allocation
2648   Register var_size_in_bytes,        // object size in bytes if unknown at compile time; invalid otherwise
2649   int      con_size_in_bytes,        // object size in bytes if   known at compile time
2650   Register t1,                       // temp register
2651   Label&   slow_case                 // continuation point if fast allocation fails
2652 ) {
2653   // make sure arguments make sense
2654   assert_different_registers(obj, var_size_in_bytes, t1);
2655   assert(0 <= con_size_in_bytes && is_simm16(con_size_in_bytes), "illegal object size");
2656   assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment");
2657 
2658   const Register new_top = t1;
2659   //verify_tlab(); not implemented
2660 
2661   ld(obj, in_bytes(JavaThread::tlab_top_offset()), R16_thread);
2662   ld(R0, in_bytes(JavaThread::tlab_end_offset()), R16_thread);
2663   if (var_size_in_bytes == noreg) {
2664     addi(new_top, obj, con_size_in_bytes);
2665   } else {
2666     add(new_top, obj, var_size_in_bytes);
2667   }
2668   cmpld(CR0, new_top, R0);
2669   bc_far_optimized(Assembler::bcondCRbiIs1, bi0(CR0, Assembler::greater), slow_case);
2670 
2671 #ifdef ASSERT
2672   // make sure new free pointer is properly aligned
2673   {
2674     Label L;
2675     andi_(R0, new_top, MinObjAlignmentInBytesMask);
2676     beq(CR0, L);
2677     stop("updated TLAB free is not properly aligned");
2678     bind(L);
2679   }
2680 #endif // ASSERT
2681 
2682   // update the tlab top pointer
2683   std(new_top, in_bytes(JavaThread::tlab_top_offset()), R16_thread);
2684   //verify_tlab(); not implemented
2685 }
2686 
2687 // "The box" is the space on the stack where we copy the object mark.
2688 void MacroAssembler::compiler_fast_lock_object(ConditionRegister flag, Register obj, Register box,
2689                                                Register tmp1, Register tmp2, Register tmp3) {
2690   assert_different_registers(obj, box, tmp1, tmp2, tmp3);
2691   assert(UseObjectMonitorTable || tmp3 == noreg, "tmp3 not needed");
2692   assert(flag == CR0, "bad condition register");
2693 
2694   // Handle inflated monitor.
2695   Label inflated;
2696   // Finish fast lock successfully. MUST reach to with flag == NE
2697   Label locked;
2698   // Finish fast lock unsuccessfully. MUST branch to with flag == EQ
2699   Label slow_path;
2700 
2701   if (UseObjectMonitorTable) {
2702     // Clear cache in case fast locking succeeds or we need to take the slow-path.
2703     li(tmp1, 0);
2704     std(tmp1, in_bytes(BasicObjectLock::lock_offset()) + BasicLock::object_monitor_cache_offset_in_bytes(), box);
2705   }
2706 
2707   if (DiagnoseSyncOnValueBasedClasses != 0) {
2708     load_klass(tmp1, obj);
2709     lbz(tmp1, in_bytes(Klass::misc_flags_offset()), tmp1);
2710     testbitdi(CR0, R0, tmp1, exact_log2(KlassFlags::_misc_is_value_based_class));
2711     bne(CR0, slow_path);
2712   }
2713 
2714   Register mark = tmp1;
2715 
2716   { // Fast locking
2717 
2718     // Push lock to the lock stack and finish successfully. MUST reach to with flag == EQ
2719     Label push;
2720 
2721     const Register top = tmp2;
2722 
2723     // Check if lock-stack is full.
2724     lwz(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
2725     cmplwi(CR0, top, LockStack::end_offset() - 1);
2726     bgt(CR0, slow_path);
2727 
2728     // The underflow check is elided. The recursive check will always fail
2729     // when the lock stack is empty because of the _bad_oop_sentinel field.
2730 
2731     // Check if recursive.
2732     subi(R0, top, oopSize);
2733     ldx(R0, R16_thread, R0);
2734     cmpd(CR0, obj, R0);
2735     beq(CR0, push);
2736 
2737     // Check for monitor (0b10) or locked (0b00).
2738     ld(mark, oopDesc::mark_offset_in_bytes(), obj);
2739     andi_(R0, mark, markWord::lock_mask_in_place);
2740     cmpldi(CR0, R0, markWord::unlocked_value);
2741     bgt(CR0, inflated);
2742     bne(CR0, slow_path);
2743 
2744     // Not inflated.
2745 
2746     // Try to lock. Transition lock bits 0b01 => 0b00
2747     assert(oopDesc::mark_offset_in_bytes() == 0, "required to avoid a lea");
2748     atomically_flip_locked_state(/* is_unlock */ false, obj, mark, slow_path, MacroAssembler::MemBarAcq);
2749 
2750     bind(push);
2751     // After successful lock, push object on lock-stack.
2752     stdx(obj, R16_thread, top);
2753     addi(top, top, oopSize);
2754     stw(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
2755     b(locked);
2756   }
2757 
2758   { // Handle inflated monitor.
2759     bind(inflated);
2760 
2761     // mark contains the tagged ObjectMonitor*.
2762     const uintptr_t monitor_tag = markWord::monitor_value;
2763     const Register monitor    = UseObjectMonitorTable ? tmp1 : noreg;
2764     const Register owner_addr = tmp2;
2765     const Register thread_id  = UseObjectMonitorTable ? tmp3 : tmp1;
2766     // Offsets into the current thread's object monitor cache (omc).
2767     const ByteSize thr_omc_offset     = JavaThread::om_cache_offset();
2768     const ByteSize omc_monitor_offset = OMCache::monitor_offset();
2769     const ByteSize omc_obj_offset     = OMCache::obj_offset();
2770 
2771     Label monitor_locked;
2772 
2773     if (!UseObjectMonitorTable) {
2774       // Compute owner address.
2775       addi(owner_addr, mark, in_bytes(ObjectMonitor::owner_offset()) - monitor_tag);
2776       mark = noreg;
2777     } else {
2778       const Register tmp3_bucket = tmp3;
2779       const Register tmp2_hash = tmp2;
2780       Label monitor_found;
2781 
2782       // Save the mark, we might need it to extract the hash.
2783       mr(tmp2_hash, mark);
2784 
2785       // Look for the monitor in the current thread's object monitor cache (omc).
2786 
2787       ld(R0, in_bytes(thr_omc_offset + omc_obj_offset), R16_thread);
2788       ld(monitor, in_bytes(thr_omc_offset + omc_monitor_offset), R16_thread);
2789       cmpd(CR0, R0, obj);
2790       beq(CR0, monitor_found);
2791 
2792       // Look for the monitor in the table.
2793 
2794       // Get the hash code.
2795       srdi(tmp2_hash, tmp2_hash, markWord::hash_shift);
2796 
2797       // Get the table and calculate the bucket's address
2798       int simm16_rest = load_const_optimized(tmp3, ObjectMonitorTable::current_table_address(), R0, true);
2799       ld_ptr(tmp3, simm16_rest, tmp3);
2800       ld(tmp1, in_bytes(ObjectMonitorTable::table_capacity_mask_offset()), tmp3);
2801       andr(tmp2_hash, tmp2_hash, tmp1);
2802       ld(tmp3_bucket, in_bytes(ObjectMonitorTable::table_buckets_offset()), tmp3);
2803 
2804       // Read the monitor from the bucket.
2805       sldi(tmp2_hash, tmp2_hash, LogBytesPerWord);
2806       ldx(monitor, tmp3_bucket, tmp2_hash);
2807 
2808       // Check if the monitor in the bucket is special (empty, tombstone or removed).
2809       cmpldi(CR0, monitor, ObjectMonitorTable::SpecialPointerValues::below_is_special);
2810       blt(CR0, slow_path);
2811 
2812       // Check if object matches.
2813       ld(tmp3, in_bytes(ObjectMonitor::object_offset()), monitor);
2814       BarrierSetAssembler* bs_asm = BarrierSet::barrier_set()->barrier_set_assembler();
2815       bs_asm->try_peek_weak_handle_in_nmethod(this, tmp3, tmp3, tmp2, slow_path);
2816       cmpd(CR0, tmp3, obj);
2817       bne(CR0, slow_path);
2818 
2819       // Store the monitor in the current thread's object monitor cache (omc).
2820       std(monitor, in_bytes(thr_omc_offset + omc_monitor_offset), R16_thread);
2821       std(obj, in_bytes(thr_omc_offset + omc_obj_offset), R16_thread);
2822 
2823       bind(monitor_found);
2824 
2825       // Compute owner address.
2826       addi(owner_addr, monitor, in_bytes(ObjectMonitor::owner_offset()));
2827     }
2828 
2829     // Try to CAS owner (no owner => current thread's _monitor_owner_id).
2830     assert_different_registers(thread_id, monitor, owner_addr, box, R0);
2831     ld(thread_id, in_bytes(JavaThread::monitor_owner_id_offset()), R16_thread);
2832     cmpxchgd(/*flag=*/CR0,
2833             /*current_value=*/R0,
2834             /*compare_value=*/(intptr_t)0,
2835             /*exchange_value=*/thread_id,
2836             /*where=*/owner_addr,
2837             MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq,
2838             MacroAssembler::cmpxchgx_hint_acquire_lock());
2839     beq(CR0, monitor_locked);
2840 
2841     // Check if recursive.
2842     cmpd(CR0, R0, thread_id);
2843     bne(CR0, slow_path);
2844 
2845     // Recursive.
2846     if (!UseObjectMonitorTable) {
2847       assert_different_registers(tmp1, owner_addr);
2848       ld(tmp1, in_bytes(ObjectMonitor::recursions_offset() - ObjectMonitor::owner_offset()), owner_addr);
2849       addi(tmp1, tmp1, 1);
2850       std(tmp1, in_bytes(ObjectMonitor::recursions_offset() - ObjectMonitor::owner_offset()), owner_addr);
2851     } else {
2852       assert_different_registers(tmp2, monitor);
2853       ld(tmp2, in_bytes(ObjectMonitor::recursions_offset()), monitor);
2854       addi(tmp2, tmp2, 1);
2855       std(tmp2, in_bytes(ObjectMonitor::recursions_offset()), monitor);
2856     }
2857 
2858     bind(monitor_locked);
2859     if (UseObjectMonitorTable) {
2860       // Cache the monitor for unlock.
2861       std(monitor, BasicLock::object_monitor_cache_offset_in_bytes(), box);
2862     }
2863   }
2864 
2865   bind(locked);
2866 
2867 #ifdef ASSERT
2868   // Check that locked label is reached with flag == EQ.
2869   Label flag_correct;
2870   beq(CR0, flag_correct);
2871   stop("Fast Lock Flag != EQ");
2872 #endif
2873   bind(slow_path);
2874 #ifdef ASSERT
2875   // Check that slow_path label is reached with flag == NE.
2876   bne(CR0, flag_correct);
2877   stop("Fast Lock Flag != NE");
2878   bind(flag_correct);
2879 #endif
2880   // C2 uses the value of flag (NE vs EQ) to determine the continuation.
2881 }
2882 
2883 void MacroAssembler::compiler_fast_unlock_object(ConditionRegister flag, Register obj, Register box,
2884                                                  Register tmp1, Register tmp2, Register tmp3) {
2885   assert_different_registers(obj, tmp1, tmp2, tmp3);
2886   assert(flag == CR0, "bad condition register");
2887 
2888   // Handle inflated monitor.
2889   Label inflated, inflated_load_monitor;
2890   // Finish fast unlock successfully. MUST reach to with flag == EQ.
2891   Label unlocked;
2892   // Finish fast unlock unsuccessfully. MUST branch to with flag == NE.
2893   Label slow_path;
2894 
2895   const Register mark = tmp1;
2896   const Register top = tmp2;
2897   const Register t = tmp3;
2898 
2899   { // Fast unlock
2900     Label push_and_slow;
2901 
2902     // Check if obj is top of lock-stack.
2903     lwz(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
2904     subi(top, top, oopSize);
2905     ldx(t, R16_thread, top);
2906     cmpd(CR0, obj, t);
2907     // Top of lock stack was not obj. Must be monitor.
2908     bne(CR0, inflated_load_monitor);
2909 
2910     // Pop lock-stack.
2911     DEBUG_ONLY(li(t, 0);)
2912     DEBUG_ONLY(stdx(t, R16_thread, top);)
2913     stw(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
2914 
2915     // The underflow check is elided. The recursive check will always fail
2916     // when the lock stack is empty because of the _bad_oop_sentinel field.
2917 
2918     // Check if recursive.
2919     subi(t, top, oopSize);
2920     ldx(t, R16_thread, t);
2921     cmpd(CR0, obj, t);
2922     beq(CR0, unlocked);
2923 
2924     // Not recursive.
2925 
2926     // Check for monitor (0b10).
2927     ld(mark, oopDesc::mark_offset_in_bytes(), obj);
2928     andi_(t, mark, markWord::monitor_value);
2929     if (!UseObjectMonitorTable) {
2930       bne(CR0, inflated);
2931     } else {
2932       bne(CR0, push_and_slow);
2933     }
2934 
2935 #ifdef ASSERT
2936     // Check header not unlocked (0b01).
2937     Label not_unlocked;
2938     andi_(t, mark, markWord::unlocked_value);
2939     beq(CR0, not_unlocked);
2940     stop("fast_unlock already unlocked");
2941     bind(not_unlocked);
2942 #endif
2943 
2944     // Try to unlock. Transition lock bits 0b00 => 0b01
2945     atomically_flip_locked_state(/* is_unlock */ true, obj, mark, push_and_slow, MacroAssembler::MemBarRel);
2946     b(unlocked);
2947 
2948     bind(push_and_slow);
2949     // Restore lock-stack and handle the unlock in runtime.
2950     DEBUG_ONLY(stdx(obj, R16_thread, top);)
2951     addi(top, top, oopSize);
2952     stw(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
2953     b(slow_path);
2954   }
2955 
2956   { // Handle inflated monitor.
2957     bind(inflated_load_monitor);
2958     ld(mark, oopDesc::mark_offset_in_bytes(), obj);
2959 #ifdef ASSERT
2960     andi_(t, mark, markWord::monitor_value);
2961     bne(CR0, inflated);
2962     stop("Fast Unlock not monitor");
2963 #endif
2964 
2965     bind(inflated);
2966 
2967 #ifdef ASSERT
2968     Label check_done;
2969     subi(top, top, oopSize);
2970     cmplwi(CR0, top, in_bytes(JavaThread::lock_stack_base_offset()));
2971     blt(CR0, check_done);
2972     ldx(t, R16_thread, top);
2973     cmpd(CR0, obj, t);
2974     bne(CR0, inflated);
2975     stop("Fast Unlock lock on stack");
2976     bind(check_done);
2977 #endif
2978 
2979     // mark contains the tagged ObjectMonitor*.
2980     const Register monitor = mark;
2981     const uintptr_t monitor_tag = markWord::monitor_value;
2982 
2983     if (!UseObjectMonitorTable) {
2984       // Untag the monitor.
2985       subi(monitor, mark, monitor_tag);
2986     } else {
2987       ld(monitor, BasicLock::object_monitor_cache_offset_in_bytes(), box);
2988       // null check with Flags == NE, no valid pointer below alignof(ObjectMonitor*)
2989       cmpldi(CR0, monitor, checked_cast<uint8_t>(alignof(ObjectMonitor*)));
2990       blt(CR0, slow_path);
2991     }
2992 
2993     const Register recursions = tmp2;
2994     Label not_recursive;
2995 
2996     // Check if recursive.
2997     ld(recursions, in_bytes(ObjectMonitor::recursions_offset()), monitor);
2998     addic_(recursions, recursions, -1);
2999     blt(CR0, not_recursive);
3000 
3001     // Recursive unlock.
3002     std(recursions, in_bytes(ObjectMonitor::recursions_offset()), monitor);
3003     crorc(CR0, Assembler::equal, CR0, Assembler::equal);
3004     b(unlocked);
3005 
3006     bind(not_recursive);
3007 
3008     // Set owner to null.
3009     // Release to satisfy the JMM
3010     release();
3011     li(t, 0);
3012     std(t, in_bytes(ObjectMonitor::owner_offset()), monitor);
3013     // We need a full fence after clearing owner to avoid stranding.
3014     // StoreLoad achieves this.
3015     membar(StoreLoad);
3016 
3017     // Check if the entry_list is empty.
3018     ld(t, in_bytes(ObjectMonitor::entry_list_offset()), monitor);
3019     cmpdi(CR0, t, 0);
3020     beq(CR0, unlocked); // If so we are done.
3021 
3022     // Check if there is a successor.
3023     ld(t, in_bytes(ObjectMonitor::succ_offset()), monitor);
3024     cmpdi(CR0, t, 0);
3025     // Invert equal bit
3026     crnand(flag, Assembler::equal, flag, Assembler::equal);
3027     beq(CR0, unlocked); // If there is a successor we are done.
3028 
3029     // Save the monitor pointer in the current thread, so we can try
3030     // to reacquire the lock in SharedRuntime::monitor_exit_helper().
3031     std(monitor, in_bytes(JavaThread::unlocked_inflated_monitor_offset()), R16_thread);
3032     b(slow_path); // flag == NE
3033   }
3034 
3035   bind(unlocked);
3036 
3037 #ifdef ASSERT
3038   // Check that unlocked label is reached with flag == EQ.
3039   Label flag_correct;
3040   beq(CR0, flag_correct);
3041   stop("Fast Lock Flag != EQ");
3042 #endif
3043   bind(slow_path);
3044 #ifdef ASSERT
3045   // Check that slow_path label is reached with flag == NE.
3046   bne(CR0, flag_correct);
3047   stop("Fast Lock Flag != NE");
3048   bind(flag_correct);
3049 #endif
3050   // C2 uses the value of flag (NE vs EQ) to determine the continuation.
3051 }
3052 
3053 void MacroAssembler::safepoint_poll(Label& slow_path, Register temp, bool at_return, bool in_nmethod) {
3054   ld(temp, in_bytes(JavaThread::polling_word_offset()), R16_thread);
3055 
3056   if (at_return) {
3057     if (in_nmethod) {
3058       if (UseSIGTRAP) {
3059         // Use Signal Handler.
3060         relocate(relocInfo::poll_return_type);
3061         td(traptoGreaterThanUnsigned, R1_SP, temp);
3062       } else {
3063         cmpld(CR0, R1_SP, temp);
3064         // Stub may be out of range for short conditional branch.
3065         bc_far_optimized(Assembler::bcondCRbiIs1, bi0(CR0, Assembler::greater), slow_path);
3066       }
3067     } else { // Not in nmethod.
3068       // Frame still on stack, need to get fp.
3069       Register fp = R0;
3070       ld(fp, _abi0(callers_sp), R1_SP);
3071       cmpld(CR0, fp, temp);
3072       bgt(CR0, slow_path);
3073     }
3074   } else { // Normal safepoint poll. Not at return.
3075     assert(!in_nmethod, "should use load_from_polling_page");
3076     andi_(temp, temp, SafepointMechanism::poll_bit());
3077     bne(CR0, slow_path);
3078   }
3079 }
3080 
3081 void MacroAssembler::jump_to_polling_page_return_handler_blob(int safepoint_offset, bool fixed_size) {
3082   assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
3083          "polling page return stub not created yet");
3084   address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
3085 
3086   // Determine saved exception pc using pc relative address computation.
3087   {
3088     Label next_pc;
3089     bl(next_pc);
3090     bind(next_pc);
3091   }
3092   int current_offset = offset();
3093 
3094   if (fixed_size) {
3095     // Code size must not depend on offsets.
3096     load_const32(R12, safepoint_offset - current_offset);
3097     mflr(R0);
3098     add(R12, R12, R0);
3099   } else {
3100     mflr(R12);
3101     add_const_optimized(R12, R12, safepoint_offset - current_offset);
3102   }
3103   std(R12, in_bytes(JavaThread::saved_exception_pc_offset()), R16_thread);
3104 
3105   add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(stub));
3106   mtctr(R0);
3107   bctr();
3108 }
3109 
3110 void MacroAssembler::resolve_jobject(Register value, Register tmp1, Register tmp2,
3111                                      MacroAssembler::PreservationLevel preservation_level) {
3112   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
3113   bs->resolve_jobject(this, value, tmp1, tmp2, preservation_level);
3114 }
3115 
3116 void MacroAssembler::resolve_global_jobject(Register value, Register tmp1, Register tmp2,
3117                                      MacroAssembler::PreservationLevel preservation_level) {
3118   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
3119   bs->resolve_global_jobject(this, value, tmp1, tmp2, preservation_level);
3120 }
3121 
3122 // Values for last_Java_pc, and last_Java_sp must comply to the rules
3123 // in frame_ppc.hpp.
3124 void MacroAssembler::set_last_Java_frame(Register last_Java_sp, Register last_Java_pc) {
3125   // Always set last_Java_pc and flags first because once last_Java_sp
3126   // is visible has_last_Java_frame is true and users will look at the
3127   // rest of the fields. (Note: flags should always be zero before we
3128   // get here so doesn't need to be set.)
3129 
3130   // Verify that last_Java_pc was zeroed on return to Java
3131   asm_assert_mem8_is_zero(in_bytes(JavaThread::last_Java_pc_offset()), R16_thread,
3132                           "last_Java_pc not zeroed before leaving Java");
3133 
3134   // When returning from calling out from Java mode the frame anchor's
3135   // last_Java_pc will always be set to null. It is set here so that
3136   // if we are doing a call to native (not VM) that we capture the
3137   // known pc and don't have to rely on the native call having a
3138   // standard frame linkage where we can find the pc.
3139   if (last_Java_pc != noreg)
3140     std(last_Java_pc, in_bytes(JavaThread::last_Java_pc_offset()), R16_thread);
3141 
3142   // Set last_Java_sp last.
3143   std(last_Java_sp, in_bytes(JavaThread::last_Java_sp_offset()), R16_thread);
3144 }
3145 
3146 void MacroAssembler::reset_last_Java_frame(bool check_last_java_sp) {
3147   if (check_last_java_sp) {
3148     asm_assert_mem8_isnot_zero(in_bytes(JavaThread::last_Java_sp_offset()),
3149                                R16_thread, "SP was not set, still zero");
3150   }
3151 
3152   BLOCK_COMMENT("reset_last_Java_frame {");
3153   li(R0, 0);
3154 
3155   // _last_Java_sp = 0
3156   std(R0, in_bytes(JavaThread::last_Java_sp_offset()), R16_thread);
3157 
3158   // _last_Java_pc = 0
3159   std(R0, in_bytes(JavaThread::last_Java_pc_offset()), R16_thread);
3160   BLOCK_COMMENT("} reset_last_Java_frame");
3161 }
3162 
3163 void MacroAssembler::set_top_ijava_frame_at_SP_as_last_Java_frame(Register sp, Register tmp1, Label* jpc) {
3164   assert_different_registers(sp, tmp1);
3165 
3166   if (jpc == nullptr || jpc->is_bound()) {
3167     load_const_optimized(tmp1, jpc == nullptr ? pc() : target(*jpc));
3168   } else {
3169     load_const(tmp1, *jpc, R12_scratch2);
3170   }
3171 
3172   set_last_Java_frame(/*sp=*/sp, /*pc=*/tmp1);
3173 }
3174 
3175 void MacroAssembler::get_vm_result_oop(Register oop_result) {
3176   // Read:
3177   //   R16_thread
3178   //   R16_thread->in_bytes(JavaThread::vm_result_oop_offset())
3179   //
3180   // Updated:
3181   //   oop_result
3182   //   R16_thread->in_bytes(JavaThread::vm_result_oop_offset())
3183 
3184   ld(oop_result, in_bytes(JavaThread::vm_result_oop_offset()), R16_thread);
3185   li(R0, 0);
3186   std(R0, in_bytes(JavaThread::vm_result_oop_offset()), R16_thread);
3187 
3188   verify_oop(oop_result, FILE_AND_LINE);
3189 }
3190 
3191 void MacroAssembler::get_vm_result_metadata(Register metadata_result) {
3192   // Read:
3193   //   R16_thread
3194   //   R16_thread->in_bytes(JavaThread::vm_result_metadata_offset())
3195   //
3196   // Updated:
3197   //   metadata_result
3198   //   R16_thread->in_bytes(JavaThread::vm_result_metadata_offset())
3199 
3200   ld(metadata_result, in_bytes(JavaThread::vm_result_metadata_offset()), R16_thread);
3201   li(R0, 0);
3202   std(R0, in_bytes(JavaThread::vm_result_metadata_offset()), R16_thread);
3203 }
3204 
3205 Register MacroAssembler::encode_klass_not_null(Register dst, Register src) {
3206   Register current = (src != noreg) ? src : dst; // Klass is in dst if no src provided.
3207   if (CompressedKlassPointers::base() != nullptr) {
3208     // Use dst as temp if it is free.
3209     sub_const_optimized(dst, current, CompressedKlassPointers::base(), R0);
3210     current = dst;
3211   }
3212   if (CompressedKlassPointers::shift() != 0) {
3213     srdi(dst, current, CompressedKlassPointers::shift());
3214     current = dst;
3215   }
3216   return current;
3217 }
3218 
3219 void MacroAssembler::store_klass(Register dst_oop, Register klass, Register ck) {
3220   assert(!UseCompactObjectHeaders, "not with compact headers");
3221   Register compressedKlass = encode_klass_not_null(ck, klass);
3222   stw(compressedKlass, oopDesc::klass_offset_in_bytes(), dst_oop);
3223 }
3224 
3225 void MacroAssembler::store_klass_gap(Register dst_oop, Register val) {
3226   assert(!UseCompactObjectHeaders, "not with compact headers");
3227   if (val == noreg) {
3228     val = R0;
3229     li(val, 0);
3230   }
3231   stw(val, oopDesc::klass_gap_offset_in_bytes(), dst_oop);
3232 }
3233 
3234 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
3235   assert(dst != R0, "Dst reg may not be R0, as R0 is used here.");
3236   if (src == noreg) src = dst;
3237   Register shifted_src = src;
3238   if (CompressedKlassPointers::shift() != 0 ||
3239       (CompressedKlassPointers::base() == nullptr && src != dst)) {  // Move required.
3240     shifted_src = dst;
3241     sldi(shifted_src, src, CompressedKlassPointers::shift());
3242   }
3243   if (CompressedKlassPointers::base() != nullptr) {
3244     add_const_optimized(dst, shifted_src, CompressedKlassPointers::base(), R0);
3245   }
3246 }
3247 
3248 void MacroAssembler::load_klass_no_decode(Register dst, Register src) {
3249   if (UseCompactObjectHeaders) {
3250     load_narrow_klass_compact(dst, src);
3251   } else {
3252     lwz(dst, oopDesc::klass_offset_in_bytes(), src);
3253   }
3254 }
3255 
3256 void MacroAssembler::load_klass(Register dst, Register src) {
3257   load_klass_no_decode(dst, src);
3258   decode_klass_not_null(dst);
3259 }
3260 
3261 // Loads the obj's Klass* into dst.
3262 // Preserves all registers (incl src, rscratch1 and rscratch2).
3263 // Input:
3264 // src - the oop we want to load the klass from.
3265 // dst - output nklass.
3266 void MacroAssembler::load_narrow_klass_compact(Register dst, Register src) {
3267   assert(UseCompactObjectHeaders, "expects UseCompactObjectHeaders");
3268   ld(dst, oopDesc::mark_offset_in_bytes(), src);
3269   srdi(dst, dst, markWord::klass_shift);
3270 }
3271 
3272 void MacroAssembler::cmp_klass(ConditionRegister dst, Register obj, Register klass, Register tmp, Register tmp2) {
3273   assert_different_registers(obj, klass, tmp);
3274   if (UseCompactObjectHeaders) {
3275     load_narrow_klass_compact(tmp, obj);
3276   } else {
3277     lwz(tmp, oopDesc::klass_offset_in_bytes(), obj);
3278   }
3279   Register encoded_klass = encode_klass_not_null(tmp2, klass);
3280   cmpw(dst, tmp, encoded_klass);
3281 }
3282 
3283 void MacroAssembler::cmp_klasses_from_objects(ConditionRegister dst, Register obj1, Register obj2, Register tmp1, Register tmp2) {
3284   if (UseCompactObjectHeaders) {
3285     load_narrow_klass_compact(tmp1, obj1);
3286     load_narrow_klass_compact(tmp2, obj2);
3287     cmpw(dst, tmp1, tmp2);
3288   } else {
3289     lwz(tmp1, oopDesc::klass_offset_in_bytes(), obj1);
3290     lwz(tmp2, oopDesc::klass_offset_in_bytes(), obj2);
3291     cmpw(dst, tmp1, tmp2);
3292   }
3293 }
3294 
3295 void MacroAssembler::load_klass_check_null(Register dst, Register src, Label* is_null) {
3296   null_check(src, oopDesc::klass_offset_in_bytes(), is_null);
3297   load_klass(dst, src);
3298 }
3299 
3300 // ((OopHandle)result).resolve();
3301 void MacroAssembler::resolve_oop_handle(Register result, Register tmp1, Register tmp2,
3302                                         MacroAssembler::PreservationLevel preservation_level) {
3303   access_load_at(T_OBJECT, IN_NATIVE, result, noreg, result, tmp1, tmp2, preservation_level);
3304 }
3305 
3306 void MacroAssembler::resolve_weak_handle(Register result, Register tmp1, Register tmp2,
3307                                          MacroAssembler::PreservationLevel preservation_level) {
3308   Label resolved;
3309 
3310   // A null weak handle resolves to null.
3311   cmpdi(CR0, result, 0);
3312   beq(CR0, resolved);
3313 
3314   access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, result, noreg, result, tmp1, tmp2,
3315                  preservation_level);
3316   bind(resolved);
3317 }
3318 
3319 void MacroAssembler::load_method_holder(Register holder, Register method) {
3320   ld(holder, in_bytes(Method::const_offset()), method);
3321   ld(holder, in_bytes(ConstMethod::constants_offset()), holder);
3322   ld(holder, ConstantPool::pool_holder_offset(), holder);
3323 }
3324 
3325 // Clear Array
3326 // For very short arrays. tmp == R0 is allowed.
3327 void MacroAssembler::clear_memory_unrolled(Register base_ptr, int cnt_dwords, Register tmp, int offset) {
3328   if (cnt_dwords > 0) { li(tmp, 0); }
3329   for (int i = 0; i < cnt_dwords; ++i) { std(tmp, offset + i * 8, base_ptr); }
3330 }
3331 
3332 // Version for constant short array length. Kills base_ptr. tmp == R0 is allowed.
3333 void MacroAssembler::clear_memory_constlen(Register base_ptr, int cnt_dwords, Register tmp) {
3334   if (cnt_dwords < 8) {
3335     clear_memory_unrolled(base_ptr, cnt_dwords, tmp);
3336     return;
3337   }
3338 
3339   Label loop;
3340   const long loopcnt   = cnt_dwords >> 1,
3341              remainder = cnt_dwords & 1;
3342 
3343   li(tmp, loopcnt);
3344   mtctr(tmp);
3345   li(tmp, 0);
3346   bind(loop);
3347     std(tmp, 0, base_ptr);
3348     std(tmp, 8, base_ptr);
3349     addi(base_ptr, base_ptr, 16);
3350     bdnz(loop);
3351   if (remainder) { std(tmp, 0, base_ptr); }
3352 }
3353 
3354 // Kills both input registers. tmp == R0 is allowed.
3355 void MacroAssembler::clear_memory_doubleword(Register base_ptr, Register cnt_dwords, Register tmp, long const_cnt) {
3356   // Procedure for large arrays (uses data cache block zero instruction).
3357     Label startloop, fast, fastloop, small_rest, restloop, done;
3358     const int cl_size         = VM_Version::L1_data_cache_line_size(),
3359               cl_dwords       = cl_size >> 3,
3360               cl_dw_addr_bits = exact_log2(cl_dwords),
3361               dcbz_min        = 1,  // Min count of dcbz executions, needs to be >0.
3362               min_cnt         = ((dcbz_min + 1) << cl_dw_addr_bits) - 1;
3363 
3364   if (const_cnt >= 0) {
3365     // Constant case.
3366     if (const_cnt < min_cnt) {
3367       clear_memory_constlen(base_ptr, const_cnt, tmp);
3368       return;
3369     }
3370     load_const_optimized(cnt_dwords, const_cnt, tmp);
3371   } else {
3372     // cnt_dwords already loaded in register. Need to check size.
3373     cmpdi(CR1, cnt_dwords, min_cnt); // Big enough? (ensure >= dcbz_min lines included).
3374     blt(CR1, small_rest);
3375   }
3376     rldicl_(tmp, base_ptr, 64-3, 64-cl_dw_addr_bits); // Extract dword offset within first cache line.
3377     beq(CR0, fast);                                  // Already 128byte aligned.
3378 
3379     subfic(tmp, tmp, cl_dwords);
3380     mtctr(tmp);                        // Set ctr to hit 128byte boundary (0<ctr<cl_dwords).
3381     subf(cnt_dwords, tmp, cnt_dwords); // rest.
3382     li(tmp, 0);
3383 
3384   bind(startloop);                     // Clear at the beginning to reach 128byte boundary.
3385     std(tmp, 0, base_ptr);             // Clear 8byte aligned block.
3386     addi(base_ptr, base_ptr, 8);
3387     bdnz(startloop);
3388 
3389   bind(fast);                                  // Clear 128byte blocks.
3390     srdi(tmp, cnt_dwords, cl_dw_addr_bits);    // Loop count for 128byte loop (>0).
3391     andi(cnt_dwords, cnt_dwords, cl_dwords-1); // Rest in dwords.
3392     mtctr(tmp);                                // Load counter.
3393 
3394   bind(fastloop);
3395     dcbz(base_ptr);                    // Clear 128byte aligned block.
3396     addi(base_ptr, base_ptr, cl_size);
3397     bdnz(fastloop);
3398 
3399   bind(small_rest);
3400     cmpdi(CR0, cnt_dwords, 0);        // size 0?
3401     beq(CR0, done);                   // rest == 0
3402     li(tmp, 0);
3403     mtctr(cnt_dwords);                 // Load counter.
3404 
3405   bind(restloop);                      // Clear rest.
3406     std(tmp, 0, base_ptr);             // Clear 8byte aligned block.
3407     addi(base_ptr, base_ptr, 8);
3408     bdnz(restloop);
3409 
3410   bind(done);
3411 }
3412 
3413 /////////////////////////////////////////// String intrinsics ////////////////////////////////////////////
3414 
3415 // Helpers for Intrinsic Emitters
3416 //
3417 // Revert the byte order of a 32bit value in a register
3418 //   src: 0x44556677
3419 //   dst: 0x77665544
3420 // Three steps to obtain the result:
3421 //  1) Rotate src (as doubleword) left 5 bytes. That puts the leftmost byte of the src word
3422 //     into the rightmost byte position. Afterwards, everything left of the rightmost byte is cleared.
3423 //     This value initializes dst.
3424 //  2) Rotate src (as word) left 3 bytes. That puts the rightmost byte of the src word into the leftmost
3425 //     byte position. Furthermore, byte 5 is rotated into byte 6 position where it is supposed to go.
3426 //     This value is mask inserted into dst with a [0..23] mask of 1s.
3427 //  3) Rotate src (as word) left 1 byte. That puts byte 6 into byte 5 position.
3428 //     This value is mask inserted into dst with a [8..15] mask of 1s.
3429 void MacroAssembler::load_reverse_32(Register dst, Register src) {
3430   assert_different_registers(dst, src);
3431 
3432   rldicl(dst, src, (4+1)*8, 56);       // Rotate byte 4 into position 7 (rightmost), clear all to the left.
3433   rlwimi(dst, src,     3*8,  0, 23);   // Insert byte 5 into position 6, 7 into 4, leave pos 7 alone.
3434   rlwimi(dst, src,     1*8,  8, 15);   // Insert byte 6 into position 5, leave the rest alone.
3435 }
3436 
3437 // Calculate the column addresses of the crc32 lookup table into distinct registers.
3438 // This loop-invariant calculation is moved out of the loop body, reducing the loop
3439 // body size from 20 to 16 instructions.
3440 // Returns the offset that was used to calculate the address of column tc3.
3441 // Due to register shortage, setting tc3 may overwrite table. With the return offset
3442 // at hand, the original table address can be easily reconstructed.
3443 int MacroAssembler::crc32_table_columns(Register table, Register tc0, Register tc1, Register tc2, Register tc3) {
3444 
3445   // Point to 4 byte folding tables (byte-reversed version for Big Endian)
3446   // Layout: See StubRoutines::ppc::generate_crc_constants.
3447 #ifdef VM_LITTLE_ENDIAN
3448   const int ix0 = 3 * CRC32_TABLE_SIZE;
3449   const int ix1 = 2 * CRC32_TABLE_SIZE;
3450   const int ix2 = 1 * CRC32_TABLE_SIZE;
3451   const int ix3 = 0 * CRC32_TABLE_SIZE;
3452 #else
3453   const int ix0 = 1 * CRC32_TABLE_SIZE;
3454   const int ix1 = 2 * CRC32_TABLE_SIZE;
3455   const int ix2 = 3 * CRC32_TABLE_SIZE;
3456   const int ix3 = 4 * CRC32_TABLE_SIZE;
3457 #endif
3458   assert_different_registers(table, tc0, tc1, tc2);
3459   assert(table == tc3, "must be!");
3460 
3461   addi(tc0, table, ix0);
3462   addi(tc1, table, ix1);
3463   addi(tc2, table, ix2);
3464   if (ix3 != 0) addi(tc3, table, ix3);
3465 
3466   return ix3;
3467 }
3468 
3469 /**
3470  * uint32_t crc;
3471  * table[crc & 0xFF] ^ (crc >> 8);
3472  */
3473 void MacroAssembler::fold_byte_crc32(Register crc, Register val, Register table, Register tmp) {
3474   assert_different_registers(crc, table, tmp);
3475   assert_different_registers(val, table);
3476 
3477   if (crc == val) {                   // Must rotate first to use the unmodified value.
3478     rlwinm(tmp, val, 2, 24-2, 31-2);  // Insert (rightmost) byte 7 of val, shifted left by 2, into byte 6..7 of tmp, clear the rest.
3479                                       // As we use a word (4-byte) instruction, we have to adapt the mask bit positions.
3480     srwi(crc, crc, 8);                // Unsigned shift, clear leftmost 8 bits.
3481   } else {
3482     srwi(crc, crc, 8);                // Unsigned shift, clear leftmost 8 bits.
3483     rlwinm(tmp, val, 2, 24-2, 31-2);  // Insert (rightmost) byte 7 of val, shifted left by 2, into byte 6..7 of tmp, clear the rest.
3484   }
3485   lwzx(tmp, table, tmp);
3486   xorr(crc, crc, tmp);
3487 }
3488 
3489 /**
3490  * Emits code to update CRC-32 with a byte value according to constants in table.
3491  *
3492  * @param [in,out]crc   Register containing the crc.
3493  * @param [in]val       Register containing the byte to fold into the CRC.
3494  * @param [in]table     Register containing the table of crc constants.
3495  *
3496  * uint32_t crc;
3497  * val = crc_table[(val ^ crc) & 0xFF];
3498  * crc = val ^ (crc >> 8);
3499  */
3500 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
3501   BLOCK_COMMENT("update_byte_crc32:");
3502   xorr(val, val, crc);
3503   fold_byte_crc32(crc, val, table, val);
3504 }
3505 
3506 /**
3507  * @param crc   register containing existing CRC (32-bit)
3508  * @param buf   register pointing to input byte buffer (byte*)
3509  * @param len   register containing number of bytes
3510  * @param table register pointing to CRC table
3511  */
3512 void MacroAssembler::update_byteLoop_crc32(Register crc, Register buf, Register len, Register table,
3513                                            Register data, bool loopAlignment) {
3514   assert_different_registers(crc, buf, len, table, data);
3515 
3516   Label L_mainLoop, L_done;
3517   const int mainLoop_stepping  = 1;
3518   const int mainLoop_alignment = loopAlignment ? 32 : 4; // (InputForNewCode > 4 ? InputForNewCode : 32) : 4;
3519 
3520   // Process all bytes in a single-byte loop.
3521   clrldi_(len, len, 32);                         // Enforce 32 bit. Anything to do?
3522   beq(CR0, L_done);
3523 
3524   mtctr(len);
3525   align(mainLoop_alignment);
3526   BIND(L_mainLoop);
3527     lbz(data, 0, buf);                           // Byte from buffer, zero-extended.
3528     addi(buf, buf, mainLoop_stepping);           // Advance buffer position.
3529     update_byte_crc32(crc, data, table);
3530     bdnz(L_mainLoop);                            // Iterate.
3531 
3532   bind(L_done);
3533 }
3534 
3535 /**
3536  * Emits code to update CRC-32 with a 4-byte value according to constants in table
3537  * Implementation according to jdk/src/share/native/java/util/zip/zlib-1.2.8/crc32.c
3538  */
3539 // A note on the lookup table address(es):
3540 // The implementation uses 4 table columns (byte-reversed versions for Big Endian).
3541 // To save the effort of adding the column offset to the table address each time
3542 // a table element is looked up, it is possible to pass the pre-calculated
3543 // column addresses.
3544 // Uses R9..R12 as work register. Must be saved/restored by caller, if necessary.
3545 void MacroAssembler::update_1word_crc32(Register crc, Register buf, Register table, int bufDisp, int bufInc,
3546                                         Register t0,  Register t1,  Register t2,  Register t3,
3547                                         Register tc0, Register tc1, Register tc2, Register tc3) {
3548   assert_different_registers(crc, t3);
3549 
3550   // XOR crc with next four bytes of buffer.
3551   lwz(t3, bufDisp, buf);
3552   if (bufInc != 0) {
3553     addi(buf, buf, bufInc);
3554   }
3555   xorr(t3, t3, crc);
3556 
3557   // Chop crc into 4 single-byte pieces, shifted left 2 bits, to form the table indices.
3558   rlwinm(t0, t3,  2,         24-2, 31-2);  // ((t1 >>  0) & 0xff) << 2
3559   rlwinm(t1, t3,  32+(2- 8), 24-2, 31-2);  // ((t1 >>  8) & 0xff) << 2
3560   rlwinm(t2, t3,  32+(2-16), 24-2, 31-2);  // ((t1 >> 16) & 0xff) << 2
3561   rlwinm(t3, t3,  32+(2-24), 24-2, 31-2);  // ((t1 >> 24) & 0xff) << 2
3562 
3563   // Use the pre-calculated column addresses.
3564   // Load pre-calculated table values.
3565   lwzx(t0, tc0, t0);
3566   lwzx(t1, tc1, t1);
3567   lwzx(t2, tc2, t2);
3568   lwzx(t3, tc3, t3);
3569 
3570   // Calculate new crc from table values.
3571   xorr(t0,  t0, t1);
3572   xorr(t2,  t2, t3);
3573   xorr(crc, t0, t2);  // Now crc contains the final checksum value.
3574 }
3575 
3576 
3577 /**
3578  * @param crc             register containing existing CRC (32-bit)
3579  * @param buf             register pointing to input byte buffer (byte*)
3580  * @param len             register containing number of bytes
3581  * @param constants       register pointing to precomputed constants
3582  * @param t0-t6           temp registers
3583  */
3584 void MacroAssembler::kernel_crc32_vpmsum(Register crc, Register buf, Register len, Register constants,
3585                                          Register t0, Register t1, Register t2, Register t3,
3586                                          Register t4, Register t5, Register t6, bool invertCRC) {
3587   assert_different_registers(crc, buf, len, constants);
3588 
3589   Label L_tail;
3590 
3591   BLOCK_COMMENT("kernel_crc32_vpmsum {");
3592 
3593   if (invertCRC) {
3594     nand(crc, crc, crc);                      // 1s complement of crc
3595   }
3596 
3597   // Enforce 32 bit.
3598   clrldi(len, len, 32);
3599 
3600   // Align if we have enough bytes for the fast version.
3601   const int alignment = 16,
3602             threshold = 32;
3603   Register prealign = t0;
3604 
3605   neg(prealign, buf);
3606   addi(t1, len, -threshold);
3607   andi(prealign, prealign, alignment - 1);
3608   cmpw(CR0, t1, prealign);
3609   blt(CR0, L_tail); // len - prealign < threshold?
3610 
3611   subf(len, prealign, len);
3612   update_byteLoop_crc32(crc, buf, prealign, constants, t2, false);
3613 
3614   // Calculate from first aligned address as far as possible.
3615   addi(constants, constants, CRC32_TABLE_SIZE); // Point to vector constants.
3616   kernel_crc32_vpmsum_aligned(crc, buf, len, constants, t0, t1, t2, t3, t4, t5, t6);
3617   addi(constants, constants, -CRC32_TABLE_SIZE); // Point to table again.
3618 
3619   // Remaining bytes.
3620   BIND(L_tail);
3621   update_byteLoop_crc32(crc, buf, len, constants, t2, false);
3622 
3623   if (invertCRC) {
3624     nand(crc, crc, crc);                      // 1s complement of crc
3625   }
3626 
3627   BLOCK_COMMENT("} kernel_crc32_vpmsum");
3628 }
3629 
3630 /**
3631  * @param crc             register containing existing CRC (32-bit)
3632  * @param buf             register pointing to input byte buffer (byte*)
3633  * @param len             register containing number of bytes (will get updated to remaining bytes)
3634  * @param constants       register pointing to CRC table for 128-bit aligned memory
3635  * @param t0-t6           temp registers
3636  */
3637 void MacroAssembler::kernel_crc32_vpmsum_aligned(Register crc, Register buf, Register len, Register constants,
3638     Register t0, Register t1, Register t2, Register t3, Register t4, Register t5, Register t6) {
3639 
3640   // Save non-volatile vector registers (frameless).
3641   Register offset = t1;
3642   int offsetInt = 0;
3643   offsetInt -= 16; li(offset, offsetInt); stvx(VR20, offset, R1_SP);
3644   offsetInt -= 16; li(offset, offsetInt); stvx(VR21, offset, R1_SP);
3645   offsetInt -= 16; li(offset, offsetInt); stvx(VR22, offset, R1_SP);
3646   offsetInt -= 16; li(offset, offsetInt); stvx(VR23, offset, R1_SP);
3647   offsetInt -= 16; li(offset, offsetInt); stvx(VR24, offset, R1_SP);
3648   offsetInt -= 16; li(offset, offsetInt); stvx(VR25, offset, R1_SP);
3649 #ifndef VM_LITTLE_ENDIAN
3650   offsetInt -= 16; li(offset, offsetInt); stvx(VR26, offset, R1_SP);
3651 #endif
3652   offsetInt -= 8; std(R14, offsetInt, R1_SP);
3653   offsetInt -= 8; std(R15, offsetInt, R1_SP);
3654 
3655   // Implementation uses an inner loop which uses between 256 and 16 * unroll_factor
3656   // bytes per iteration. The basic scheme is:
3657   // lvx: load vector (Big Endian needs reversal)
3658   // vpmsumw: carry-less 32 bit multiplications with constant representing a large CRC shift
3659   // vxor: xor partial results together to get unroll_factor2 vectors
3660 
3661   // Outer loop performs the CRC shifts needed to combine the unroll_factor2 vectors.
3662 
3663   // Using 16 * unroll_factor / unroll_factor_2 bytes for constants.
3664   const int unroll_factor = CRC32_UNROLL_FACTOR,
3665             unroll_factor2 = CRC32_UNROLL_FACTOR2;
3666 
3667   const int outer_consts_size = (unroll_factor2 - 1) * 16,
3668             inner_consts_size = (unroll_factor / unroll_factor2) * 16;
3669 
3670   // Support registers.
3671   Register offs[] = { noreg, t0, t1, t2, t3, t4, t5, t6 };
3672   Register num_bytes = R14,
3673            loop_count = R15,
3674            cur_const = crc; // will live in VCRC
3675   // Constant array for outer loop: unroll_factor2 - 1 registers,
3676   // Constant array for inner loop: unroll_factor / unroll_factor2 registers.
3677   VectorRegister consts0[] = { VR16, VR17, VR18, VR19, VR20, VR21, VR22 },
3678                  consts1[] = { VR23, VR24 };
3679   // Data register arrays: 2 arrays with unroll_factor2 registers.
3680   VectorRegister data0[] = { VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7 },
3681                  data1[] = { VR8, VR9, VR10, VR11, VR12, VR13, VR14, VR15 };
3682 
3683   VectorRegister VCRC = data0[0];
3684   VectorRegister Vc = VR25;
3685   VectorRegister swap_bytes = VR26; // Only for Big Endian.
3686 
3687   // We have at least 1 iteration (ensured by caller).
3688   Label L_outer_loop, L_inner_loop, L_last;
3689 
3690   // Set DSCR pre-fetch to deepest.
3691   if (VM_Version::has_mfdscr()) {
3692     load_const_optimized(t0, VM_Version::_dscr_val | 7);
3693     mtdscr(t0);
3694   }
3695 
3696   mtvrwz(VCRC, crc); // crc lives in VCRC, now
3697 
3698   for (int i = 1; i < unroll_factor2; ++i) {
3699     li(offs[i], 16 * i);
3700   }
3701 
3702   // Load consts for outer loop
3703   lvx(consts0[0], constants);
3704   for (int i = 1; i < unroll_factor2 - 1; ++i) {
3705     lvx(consts0[i], offs[i], constants);
3706   }
3707 
3708   load_const_optimized(num_bytes, 16 * unroll_factor);
3709 
3710   // Reuse data registers outside of the loop.
3711   VectorRegister Vtmp = data1[0];
3712   VectorRegister Vtmp2 = data1[1];
3713   VectorRegister zeroes = data1[2];
3714 
3715   vspltisb(Vtmp, 0);
3716   vsldoi(VCRC, Vtmp, VCRC, 8); // 96 bit zeroes, 32 bit CRC.
3717 
3718   // Load vector for vpermxor (to xor both 64 bit parts together)
3719   lvsl(Vtmp, buf);   // 000102030405060708090a0b0c0d0e0f
3720   vspltisb(Vc, 4);
3721   vsl(Vc, Vtmp, Vc); // 00102030405060708090a0b0c0d0e0f0
3722   xxspltd(Vc->to_vsr(), Vc->to_vsr(), 0);
3723   vor(Vc, Vtmp, Vc); // 001122334455667708192a3b4c5d6e7f
3724 
3725 #ifdef VM_LITTLE_ENDIAN
3726 #define BE_swap_bytes(x)
3727 #else
3728   vspltisb(Vtmp2, 0xf);
3729   vxor(swap_bytes, Vtmp, Vtmp2);
3730 #define BE_swap_bytes(x) vperm(x, x, x, swap_bytes)
3731 #endif
3732 
3733   cmpd(CR0, len, num_bytes);
3734   blt(CR0, L_last);
3735 
3736   addi(cur_const, constants, outer_consts_size); // Point to consts for inner loop
3737   load_const_optimized(loop_count, unroll_factor / (2 * unroll_factor2) - 1); // One double-iteration peeled off.
3738 
3739   // ********** Main loop start **********
3740   align(32);
3741   bind(L_outer_loop);
3742 
3743   // Begin of unrolled first iteration (no xor).
3744   lvx(data1[0], buf);
3745   for (int i = 1; i < unroll_factor2 / 2; ++i) {
3746     lvx(data1[i], offs[i], buf);
3747   }
3748   vpermxor(VCRC, VCRC, VCRC, Vc); // xor both halves to 64 bit result.
3749   lvx(consts1[0], cur_const);
3750   mtctr(loop_count);
3751   for (int i = 0; i < unroll_factor2 / 2; ++i) {
3752     BE_swap_bytes(data1[i]);
3753     if (i == 0) { vxor(data1[0], data1[0], VCRC); } // xor in previous CRC.
3754     lvx(data1[i + unroll_factor2 / 2], offs[i + unroll_factor2 / 2], buf);
3755     vpmsumw(data0[i], data1[i], consts1[0]);
3756   }
3757   addi(buf, buf, 16 * unroll_factor2);
3758   subf(len, num_bytes, len);
3759   lvx(consts1[1], offs[1], cur_const);
3760   addi(cur_const, cur_const, 32);
3761   // Begin of unrolled second iteration (head).
3762   for (int i = 0; i < unroll_factor2 / 2; ++i) {
3763     BE_swap_bytes(data1[i + unroll_factor2 / 2]);
3764     if (i == 0) { lvx(data1[0], buf); } else { lvx(data1[i], offs[i], buf); }
3765     vpmsumw(data0[i + unroll_factor2 / 2], data1[i + unroll_factor2 / 2], consts1[0]);
3766   }
3767   for (int i = 0; i < unroll_factor2 / 2; ++i) {
3768     BE_swap_bytes(data1[i]);
3769     lvx(data1[i + unroll_factor2 / 2], offs[i + unroll_factor2 / 2], buf);
3770     vpmsumw(data1[i], data1[i], consts1[1]);
3771   }
3772   addi(buf, buf, 16 * unroll_factor2);
3773 
3774   // Generate most performance relevant code. Loads + half of the vpmsumw have been generated.
3775   // Double-iteration allows using the 2 constant registers alternatingly.
3776   align(32);
3777   bind(L_inner_loop);
3778   for (int j = 1; j < 3; ++j) { // j < unroll_factor / unroll_factor2 - 1 for complete unrolling.
3779     if (j & 1) {
3780       lvx(consts1[0], cur_const);
3781     } else {
3782       lvx(consts1[1], offs[1], cur_const);
3783       addi(cur_const, cur_const, 32);
3784     }
3785     for (int i = 0; i < unroll_factor2; ++i) {
3786       int idx = i + unroll_factor2 / 2, inc = 0; // For modulo-scheduled input.
3787       if (idx >= unroll_factor2) { idx -= unroll_factor2; inc = 1; }
3788       BE_swap_bytes(data1[idx]);
3789       vxor(data0[i], data0[i], data1[i]);
3790       if (i == 0) lvx(data1[0], buf); else lvx(data1[i], offs[i], buf);
3791       vpmsumw(data1[idx], data1[idx], consts1[(j + inc) & 1]);
3792     }
3793     addi(buf, buf, 16 * unroll_factor2);
3794   }
3795   bdnz(L_inner_loop);
3796 
3797   addi(cur_const, constants, outer_consts_size); // Reset
3798 
3799   // Tail of last iteration (no loads).
3800   for (int i = 0; i < unroll_factor2 / 2; ++i) {
3801     BE_swap_bytes(data1[i + unroll_factor2 / 2]);
3802     vxor(data0[i], data0[i], data1[i]);
3803     vpmsumw(data1[i + unroll_factor2 / 2], data1[i + unroll_factor2 / 2], consts1[1]);
3804   }
3805   for (int i = 0; i < unroll_factor2 / 2; ++i) {
3806     vpmsumw(data0[i], data0[i], consts0[unroll_factor2 - 2 - i]); // First half of fixup shifts.
3807     vxor(data0[i + unroll_factor2 / 2], data0[i + unroll_factor2 / 2], data1[i + unroll_factor2 / 2]);
3808   }
3809 
3810   // Last data register is ok, other ones need fixup shift.
3811   for (int i = unroll_factor2 / 2; i < unroll_factor2 - 1; ++i) {
3812     vpmsumw(data0[i], data0[i], consts0[unroll_factor2 - 2 - i]);
3813   }
3814 
3815   // Combine to 128 bit result vector VCRC = data0[0].
3816   for (int i = 1; i < unroll_factor2; i<<=1) {
3817     for (int j = 0; j <= unroll_factor2 - 2*i; j+=2*i) {
3818       vxor(data0[j], data0[j], data0[j+i]);
3819     }
3820   }
3821   cmpd(CR0, len, num_bytes);
3822   bge(CR0, L_outer_loop);
3823 
3824   // Last chance with lower num_bytes.
3825   bind(L_last);
3826   srdi(loop_count, len, exact_log2(16 * 2 * unroll_factor2)); // Use double-iterations.
3827   // Point behind last const for inner loop.
3828   add_const_optimized(cur_const, constants, outer_consts_size + inner_consts_size);
3829   sldi(R0, loop_count, exact_log2(16 * 2)); // Bytes of constants to be used.
3830   clrrdi(num_bytes, len, exact_log2(16 * 2 * unroll_factor2));
3831   subf(cur_const, R0, cur_const); // Point to constant to be used first.
3832 
3833   addic_(loop_count, loop_count, -1); // One double-iteration peeled off.
3834   bgt(CR0, L_outer_loop);
3835   // ********** Main loop end **********
3836 
3837   // Restore DSCR pre-fetch value.
3838   if (VM_Version::has_mfdscr()) {
3839     load_const_optimized(t0, VM_Version::_dscr_val);
3840     mtdscr(t0);
3841   }
3842 
3843   // ********** Simple loop for remaining 16 byte blocks **********
3844   {
3845     Label L_loop, L_done;
3846 
3847     srdi_(t0, len, 4); // 16 bytes per iteration
3848     clrldi(len, len, 64-4);
3849     beq(CR0, L_done);
3850 
3851     // Point to const (same as last const for inner loop).
3852     add_const_optimized(cur_const, constants, outer_consts_size + inner_consts_size - 16);
3853     mtctr(t0);
3854     lvx(Vtmp2, cur_const);
3855 
3856     align(32);
3857     bind(L_loop);
3858 
3859     lvx(Vtmp, buf);
3860     addi(buf, buf, 16);
3861     vpermxor(VCRC, VCRC, VCRC, Vc); // xor both halves to 64 bit result.
3862     BE_swap_bytes(Vtmp);
3863     vxor(VCRC, VCRC, Vtmp);
3864     vpmsumw(VCRC, VCRC, Vtmp2);
3865     bdnz(L_loop);
3866 
3867     bind(L_done);
3868   }
3869   // ********** Simple loop end **********
3870 #undef BE_swap_bytes
3871 
3872   // Point to Barrett constants
3873   add_const_optimized(cur_const, constants, outer_consts_size + inner_consts_size);
3874 
3875   vspltisb(zeroes, 0);
3876 
3877   // Combine to 64 bit result.
3878   vpermxor(VCRC, VCRC, VCRC, Vc); // xor both halves to 64 bit result.
3879 
3880   // Reduce to 32 bit CRC: Remainder by multiply-high.
3881   lvx(Vtmp, cur_const);
3882   vsldoi(Vtmp2, zeroes, VCRC, 12);  // Extract high 32 bit.
3883   vpmsumd(Vtmp2, Vtmp2, Vtmp);      // Multiply by inverse long poly.
3884   vsldoi(Vtmp2, zeroes, Vtmp2, 12); // Extract high 32 bit.
3885   vsldoi(Vtmp, zeroes, Vtmp, 8);
3886   vpmsumd(Vtmp2, Vtmp2, Vtmp);      // Multiply quotient by long poly.
3887   vxor(VCRC, VCRC, Vtmp2);          // Remainder fits into 32 bit.
3888 
3889   // Move result. len is already updated.
3890   vsldoi(VCRC, VCRC, zeroes, 8);
3891   mfvrd(crc, VCRC);
3892 
3893   // Restore non-volatile Vector registers (frameless).
3894   offsetInt = 0;
3895   offsetInt -= 16; li(offset, offsetInt); lvx(VR20, offset, R1_SP);
3896   offsetInt -= 16; li(offset, offsetInt); lvx(VR21, offset, R1_SP);
3897   offsetInt -= 16; li(offset, offsetInt); lvx(VR22, offset, R1_SP);
3898   offsetInt -= 16; li(offset, offsetInt); lvx(VR23, offset, R1_SP);
3899   offsetInt -= 16; li(offset, offsetInt); lvx(VR24, offset, R1_SP);
3900   offsetInt -= 16; li(offset, offsetInt); lvx(VR25, offset, R1_SP);
3901 #ifndef VM_LITTLE_ENDIAN
3902   offsetInt -= 16; li(offset, offsetInt); lvx(VR26, offset, R1_SP);
3903 #endif
3904   offsetInt -= 8;  ld(R14, offsetInt, R1_SP);
3905   offsetInt -= 8;  ld(R15, offsetInt, R1_SP);
3906 }
3907 
3908 void MacroAssembler::crc32(Register crc, Register buf, Register len, Register t0, Register t1, Register t2,
3909                            Register t3, Register t4, Register t5, Register t6, Register t7, bool is_crc32c) {
3910   load_const_optimized(t0, is_crc32c ? StubRoutines::crc32c_table_addr()
3911                                      : StubRoutines::crc_table_addr()   , R0);
3912 
3913   kernel_crc32_vpmsum(crc, buf, len, t0, t1, t2, t3, t4, t5, t6, t7, !is_crc32c);
3914 }
3915 
3916 void MacroAssembler::kernel_crc32_singleByteReg(Register crc, Register val, Register table, bool invertCRC) {
3917   assert_different_registers(crc, val, table);
3918 
3919   BLOCK_COMMENT("kernel_crc32_singleByteReg:");
3920   if (invertCRC) {
3921     nand(crc, crc, crc);                // 1s complement of crc
3922   }
3923 
3924   update_byte_crc32(crc, val, table);
3925 
3926   if (invertCRC) {
3927     nand(crc, crc, crc);                // 1s complement of crc
3928   }
3929 }
3930 
3931 // dest_lo += src1 + src2
3932 // dest_hi += carry1 + carry2
3933 void MacroAssembler::add2_with_carry(Register dest_hi,
3934                                      Register dest_lo,
3935                                      Register src1, Register src2) {
3936   li(R0, 0);
3937   addc(dest_lo, dest_lo, src1);
3938   adde(dest_hi, dest_hi, R0);
3939   addc(dest_lo, dest_lo, src2);
3940   adde(dest_hi, dest_hi, R0);
3941 }
3942 
3943 // Multiply 64 bit by 64 bit first loop.
3944 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart,
3945                                            Register x_xstart,
3946                                            Register y, Register y_idx,
3947                                            Register z,
3948                                            Register carry,
3949                                            Register product_high, Register product,
3950                                            Register idx, Register kdx,
3951                                            Register tmp) {
3952   //  jlong carry, x[], y[], z[];
3953   //  for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx--, kdx--) {
3954   //    huge_128 product = y[idx] * x[xstart] + carry;
3955   //    z[kdx] = (jlong)product;
3956   //    carry  = (jlong)(product >>> 64);
3957   //  }
3958   //  z[xstart] = carry;
3959 
3960   Label L_first_loop, L_first_loop_exit;
3961   Label L_one_x, L_one_y, L_multiply;
3962 
3963   addic_(xstart, xstart, -1);
3964   blt(CR0, L_one_x);   // Special case: length of x is 1.
3965 
3966   // Load next two integers of x.
3967   sldi(tmp, xstart, LogBytesPerInt);
3968   ldx(x_xstart, x, tmp);
3969 #ifdef VM_LITTLE_ENDIAN
3970   rldicl(x_xstart, x_xstart, 32, 0);
3971 #endif
3972 
3973   align(32, 16);
3974   bind(L_first_loop);
3975 
3976   cmpdi(CR0, idx, 1);
3977   blt(CR0, L_first_loop_exit);
3978   addi(idx, idx, -2);
3979   beq(CR0, L_one_y);
3980 
3981   // Load next two integers of y.
3982   sldi(tmp, idx, LogBytesPerInt);
3983   ldx(y_idx, y, tmp);
3984 #ifdef VM_LITTLE_ENDIAN
3985   rldicl(y_idx, y_idx, 32, 0);
3986 #endif
3987 
3988 
3989   bind(L_multiply);
3990   multiply64(product_high, product, x_xstart, y_idx);
3991 
3992   li(tmp, 0);
3993   addc(product, product, carry);         // Add carry to result.
3994   adde(product_high, product_high, tmp); // Add carry of the last addition.
3995   addi(kdx, kdx, -2);
3996 
3997   // Store result.
3998 #ifdef VM_LITTLE_ENDIAN
3999   rldicl(product, product, 32, 0);
4000 #endif
4001   sldi(tmp, kdx, LogBytesPerInt);
4002   stdx(product, z, tmp);
4003   mr_if_needed(carry, product_high);
4004   b(L_first_loop);
4005 
4006 
4007   bind(L_one_y); // Load one 32 bit portion of y as (0,value).
4008 
4009   lwz(y_idx, 0, y);
4010   b(L_multiply);
4011 
4012 
4013   bind(L_one_x); // Load one 32 bit portion of x as (0,value).
4014 
4015   lwz(x_xstart, 0, x);
4016   b(L_first_loop);
4017 
4018   bind(L_first_loop_exit);
4019 }
4020 
4021 // Multiply 64 bit by 64 bit and add 128 bit.
4022 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y,
4023                                             Register z, Register yz_idx,
4024                                             Register idx, Register carry,
4025                                             Register product_high, Register product,
4026                                             Register tmp, int offset) {
4027 
4028   //  huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
4029   //  z[kdx] = (jlong)product;
4030 
4031   sldi(tmp, idx, LogBytesPerInt);
4032   if (offset) {
4033     addi(tmp, tmp, offset);
4034   }
4035   ldx(yz_idx, y, tmp);
4036 #ifdef VM_LITTLE_ENDIAN
4037   rldicl(yz_idx, yz_idx, 32, 0);
4038 #endif
4039 
4040   multiply64(product_high, product, x_xstart, yz_idx);
4041   ldx(yz_idx, z, tmp);
4042 #ifdef VM_LITTLE_ENDIAN
4043   rldicl(yz_idx, yz_idx, 32, 0);
4044 #endif
4045 
4046   add2_with_carry(product_high, product, carry, yz_idx);
4047 
4048   sldi(tmp, idx, LogBytesPerInt);
4049   if (offset) {
4050     addi(tmp, tmp, offset);
4051   }
4052 #ifdef VM_LITTLE_ENDIAN
4053   rldicl(product, product, 32, 0);
4054 #endif
4055   stdx(product, z, tmp);
4056 }
4057 
4058 // Multiply 128 bit by 128 bit. Unrolled inner loop.
4059 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart,
4060                                              Register y, Register z,
4061                                              Register yz_idx, Register idx, Register carry,
4062                                              Register product_high, Register product,
4063                                              Register carry2, Register tmp) {
4064 
4065   //  jlong carry, x[], y[], z[];
4066   //  int kdx = ystart+1;
4067   //  for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
4068   //    huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
4069   //    z[kdx+idx+1] = (jlong)product;
4070   //    jlong carry2 = (jlong)(product >>> 64);
4071   //    product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
4072   //    z[kdx+idx] = (jlong)product;
4073   //    carry = (jlong)(product >>> 64);
4074   //  }
4075   //  idx += 2;
4076   //  if (idx > 0) {
4077   //    product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
4078   //    z[kdx+idx] = (jlong)product;
4079   //    carry = (jlong)(product >>> 64);
4080   //  }
4081 
4082   Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
4083   const Register jdx = R0;
4084 
4085   // Scale the index.
4086   srdi_(jdx, idx, 2);
4087   beq(CR0, L_third_loop_exit);
4088   mtctr(jdx);
4089 
4090   align(32, 16);
4091   bind(L_third_loop);
4092 
4093   addi(idx, idx, -4);
4094 
4095   multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product_high, product, tmp, 8);
4096   mr_if_needed(carry2, product_high);
4097 
4098   multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product_high, product, tmp, 0);
4099   mr_if_needed(carry, product_high);
4100   bdnz(L_third_loop);
4101 
4102   bind(L_third_loop_exit);  // Handle any left-over operand parts.
4103 
4104   andi_(idx, idx, 0x3);
4105   beq(CR0, L_post_third_loop_done);
4106 
4107   Label L_check_1;
4108 
4109   addic_(idx, idx, -2);
4110   blt(CR0, L_check_1);
4111 
4112   multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product_high, product, tmp, 0);
4113   mr_if_needed(carry, product_high);
4114 
4115   bind(L_check_1);
4116 
4117   addi(idx, idx, 0x2);
4118   andi_(idx, idx, 0x1);
4119   addic_(idx, idx, -1);
4120   blt(CR0, L_post_third_loop_done);
4121 
4122   sldi(tmp, idx, LogBytesPerInt);
4123   lwzx(yz_idx, y, tmp);
4124   multiply64(product_high, product, x_xstart, yz_idx);
4125   lwzx(yz_idx, z, tmp);
4126 
4127   add2_with_carry(product_high, product, yz_idx, carry);
4128 
4129   sldi(tmp, idx, LogBytesPerInt);
4130   stwx(product, z, tmp);
4131   srdi(product, product, 32);
4132 
4133   sldi(product_high, product_high, 32);
4134   orr(product, product, product_high);
4135   mr_if_needed(carry, product);
4136 
4137   bind(L_post_third_loop_done);
4138 }   // multiply_128_x_128_loop
4139 
4140 void MacroAssembler::muladd(Register out, Register in,
4141                             Register offset, Register len, Register k,
4142                             Register tmp1, Register tmp2, Register carry) {
4143 
4144   // Labels
4145   Label LOOP, SKIP;
4146 
4147   // Make sure length is positive.
4148   cmpdi  (CR0,    len,     0);
4149 
4150   // Prepare variables
4151   subi   (offset,  offset,  4);
4152   li     (carry,   0);
4153   ble    (CR0,    SKIP);
4154 
4155   mtctr  (len);
4156   subi   (len,     len,     1    );
4157   sldi   (len,     len,     2    );
4158 
4159   // Main loop
4160   bind(LOOP);
4161   lwzx   (tmp1,    len,     in   );
4162   lwzx   (tmp2,    offset,  out  );
4163   mulld  (tmp1,    tmp1,    k    );
4164   add    (tmp2,    carry,   tmp2 );
4165   add    (tmp2,    tmp1,    tmp2 );
4166   stwx   (tmp2,    offset,  out  );
4167   srdi   (carry,   tmp2,    32   );
4168   subi   (offset,  offset,  4    );
4169   subi   (len,     len,     4    );
4170   bdnz   (LOOP);
4171   bind(SKIP);
4172 }
4173 
4174 void MacroAssembler::multiply_to_len(Register x, Register xlen,
4175                                      Register y, Register ylen,
4176                                      Register z,
4177                                      Register tmp1, Register tmp2,
4178                                      Register tmp3, Register tmp4,
4179                                      Register tmp5, Register tmp6,
4180                                      Register tmp7, Register tmp8,
4181                                      Register tmp9, Register tmp10,
4182                                      Register tmp11, Register tmp12,
4183                                      Register tmp13) {
4184 
4185   ShortBranchVerifier sbv(this);
4186 
4187   assert_different_registers(x, xlen, y, ylen, z,
4188                              tmp1, tmp2, tmp3, tmp4, tmp5, tmp6);
4189   assert_different_registers(x, xlen, y, ylen, z,
4190                              tmp1, tmp2, tmp3, tmp4, tmp5, tmp7);
4191   assert_different_registers(x, xlen, y, ylen, z,
4192                              tmp1, tmp2, tmp3, tmp4, tmp5, tmp8);
4193 
4194   const Register idx = tmp1;
4195   const Register kdx = tmp2;
4196   const Register xstart = tmp3;
4197 
4198   const Register y_idx = tmp4;
4199   const Register carry = tmp5;
4200   const Register product = tmp6;
4201   const Register product_high = tmp7;
4202   const Register x_xstart = tmp8;
4203   const Register tmp = tmp9;
4204 
4205   // First Loop.
4206   //
4207   //  final static long LONG_MASK = 0xffffffffL;
4208   //  int xstart = xlen - 1;
4209   //  int ystart = ylen - 1;
4210   //  long carry = 0;
4211   //  for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
4212   //    long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
4213   //    z[kdx] = (int)product;
4214   //    carry = product >>> 32;
4215   //  }
4216   //  z[xstart] = (int)carry;
4217 
4218   mr_if_needed(idx, ylen);        // idx = ylen
4219   add(kdx, xlen, ylen);           // kdx = xlen + ylen
4220   li(carry, 0);                   // carry = 0
4221 
4222   Label L_done;
4223 
4224   addic_(xstart, xlen, -1);
4225   blt(CR0, L_done);
4226 
4227   multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z,
4228                         carry, product_high, product, idx, kdx, tmp);
4229 
4230   Label L_second_loop;
4231 
4232   cmpdi(CR0, kdx, 0);
4233   beq(CR0, L_second_loop);
4234 
4235   Label L_carry;
4236 
4237   addic_(kdx, kdx, -1);
4238   beq(CR0, L_carry);
4239 
4240   // Store lower 32 bits of carry.
4241   sldi(tmp, kdx, LogBytesPerInt);
4242   stwx(carry, z, tmp);
4243   srdi(carry, carry, 32);
4244   addi(kdx, kdx, -1);
4245 
4246 
4247   bind(L_carry);
4248 
4249   // Store upper 32 bits of carry.
4250   sldi(tmp, kdx, LogBytesPerInt);
4251   stwx(carry, z, tmp);
4252 
4253   // Second and third (nested) loops.
4254   //
4255   //  for (int i = xstart-1; i >= 0; i--) { // Second loop
4256   //    carry = 0;
4257   //    for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
4258   //      long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
4259   //                     (z[k] & LONG_MASK) + carry;
4260   //      z[k] = (int)product;
4261   //      carry = product >>> 32;
4262   //    }
4263   //    z[i] = (int)carry;
4264   //  }
4265   //
4266   //  i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
4267 
4268   bind(L_second_loop);
4269 
4270   li(carry, 0);                   // carry = 0;
4271 
4272   addic_(xstart, xstart, -1);     // i = xstart-1;
4273   blt(CR0, L_done);
4274 
4275   Register zsave = tmp10;
4276 
4277   mr(zsave, z);
4278 
4279 
4280   Label L_last_x;
4281 
4282   sldi(tmp, xstart, LogBytesPerInt);
4283   add(z, z, tmp);                 // z = z + k - j
4284   addi(z, z, 4);
4285   addic_(xstart, xstart, -1);     // i = xstart-1;
4286   blt(CR0, L_last_x);
4287 
4288   sldi(tmp, xstart, LogBytesPerInt);
4289   ldx(x_xstart, x, tmp);
4290 #ifdef VM_LITTLE_ENDIAN
4291   rldicl(x_xstart, x_xstart, 32, 0);
4292 #endif
4293 
4294 
4295   Label L_third_loop_prologue;
4296 
4297   bind(L_third_loop_prologue);
4298 
4299   Register xsave = tmp11;
4300   Register xlensave = tmp12;
4301   Register ylensave = tmp13;
4302 
4303   mr(xsave, x);
4304   mr(xlensave, xstart);
4305   mr(ylensave, ylen);
4306 
4307 
4308   multiply_128_x_128_loop(x_xstart, y, z, y_idx, ylen,
4309                           carry, product_high, product, x, tmp);
4310 
4311   mr(z, zsave);
4312   mr(x, xsave);
4313   mr(xlen, xlensave);   // This is the decrement of the loop counter!
4314   mr(ylen, ylensave);
4315 
4316   addi(tmp3, xlen, 1);
4317   sldi(tmp, tmp3, LogBytesPerInt);
4318   stwx(carry, z, tmp);
4319   addic_(tmp3, tmp3, -1);
4320   blt(CR0, L_done);
4321 
4322   srdi(carry, carry, 32);
4323   sldi(tmp, tmp3, LogBytesPerInt);
4324   stwx(carry, z, tmp);
4325   b(L_second_loop);
4326 
4327   // Next infrequent code is moved outside loops.
4328   bind(L_last_x);
4329 
4330   lwz(x_xstart, 0, x);
4331   b(L_third_loop_prologue);
4332 
4333   bind(L_done);
4334 }   // multiply_to_len
4335 
4336 void MacroAssembler::increment_mem64(Register base, RegisterOrConstant ind_or_offs, int val, Register tmp) {
4337   ld(tmp, ind_or_offs, base);
4338   addi(tmp, tmp, val);
4339   std(tmp, ind_or_offs, base);
4340 }
4341 
4342 // Handle the receiver type profile update given the "recv" klass.
4343 //
4344 // Normally updates the ReceiverData (RD) that starts at "mdp" + "mdp_offset".
4345 // If there are no matching or claimable receiver entries in RD, updates
4346 // the polymorphic counter.
4347 //
4348 // This code expected to run by either the interpreter or JIT-ed code, without
4349 // extra synchronization. For safety, receiver cells are claimed atomically, which
4350 // avoids grossly misrepresenting the profiles under concurrent updates. For speed,
4351 // counter updates are not atomic.
4352 //
4353 void MacroAssembler::profile_receiver_type(Register recv, Register mdp, int mdp_offset, Register tmp1, Register tmp2) {
4354   assert_different_registers(recv, mdp, tmp1, tmp2);
4355 
4356   int base_receiver_offset   = in_bytes(ReceiverTypeData::receiver_offset(0));
4357   int poly_count_offset      = in_bytes(CounterData::count_offset());
4358   int receiver_step          = in_bytes(ReceiverTypeData::receiver_offset(1)) - base_receiver_offset;
4359   int receiver_to_count_step = in_bytes(ReceiverTypeData::receiver_count_offset(0)) - base_receiver_offset;
4360 
4361   // Adjust for MDP offsets.
4362   base_receiver_offset += mdp_offset;
4363   poly_count_offset    += mdp_offset;
4364 
4365 #ifdef ASSERT
4366   // We are about to walk the MDO slots without asking for offsets.
4367   // Check that our math hits all the right spots.
4368   for (uint c = 0; c < ReceiverTypeData::row_limit(); c++) {
4369     int real_recv_offset  = mdp_offset + in_bytes(ReceiverTypeData::receiver_offset(c));
4370     int real_count_offset = mdp_offset + in_bytes(ReceiverTypeData::receiver_count_offset(c));
4371     int offset = base_receiver_offset + receiver_step*c;
4372     int count_offset = offset + receiver_to_count_step;
4373     assert(offset == real_recv_offset, "receiver slot math");
4374     assert(count_offset == real_count_offset, "receiver count math");
4375   }
4376   int real_poly_count_offset = mdp_offset + in_bytes(CounterData::count_offset());
4377   assert(poly_count_offset == real_poly_count_offset, "poly counter math");
4378 #endif
4379 
4380   // Corner case: no profile table. Increment poly counter and exit.
4381   if (ReceiverTypeData::row_limit() == 0) {
4382     increment_mem64(mdp, poly_count_offset, DataLayout::counter_increment, tmp1);
4383     return;
4384   }
4385 
4386   Label L_loop_search_receiver, L_loop_search_empty;
4387   Label L_restart, L_found_recv, L_found_empty, L_count_update;
4388   Register offset = tmp1, count = tmp2;
4389 
4390   // The code here recognizes three major cases:
4391   //   A. Fastest: receiver found in the table
4392   //   B. Fast: no receiver in the table, and the table is full
4393   //   C. Slow: no receiver in the table, free slots in the table
4394   //
4395   // The case A performance is most important, as perfectly-behaved code would end up
4396   // there, especially with larger TypeProfileWidth. The case B performance is
4397   // important as well, this is where bulk of code would land for normally megamorphic
4398   // cases. The case C performance is not essential, its job is to deal with installation
4399   // races, we optimize for code density instead. Case C needs to make sure that receiver
4400   // rows are only claimed once. This makes sure we never overwrite a row for another
4401   // receiver and never duplicate the receivers in the list, making profile type-accurate.
4402   //
4403   // It is very tempting to handle these cases in a single loop, and claim the first slot
4404   // without checking the rest of the table. But, profiling code should tolerate free slots
4405   // in the table, as class unloading can clear them. After such cleanup, the receiver
4406   // we need might be _after_ the free slot. Therefore, we need to let at least full scan
4407   // to complete, before trying to install new slots. Splitting the code in several tight
4408   // loops also helpfully optimizes for cases A and B.
4409   //
4410   // This code is effectively:
4411   //
4412   // restart:
4413   //   // Fastest: receiver is already installed
4414   //   for (i = 0; i < receiver_count(); i++) {
4415   //     if (receiver(i) == recv) goto found_recv(i);
4416   //   }
4417   //
4418   //   // Fast: no receiver, but profile is not full
4419   //   for (i = 0; i < receiver_count(); i++) {
4420   //     if (receiver(i) == null) goto found_null(i);
4421   //   }
4422   //
4423   //   // Slow: profile is full, polymorphic case
4424   //   count++;
4425   //   return
4426   //
4427   //   // Slow: try to install receiver
4428   // found_null(i):
4429   //   CAS(&receiver(i), null, recv);
4430   //   goto restart
4431   //
4432   // found_recv(i):
4433   //   *receiver_count(i)++
4434   //
4435 
4436   if (count != noreg) {
4437     li(count, ReceiverTypeData::row_limit());
4438   }
4439 
4440   bind(L_restart);
4441 
4442   // Fastest: receiver is already installed
4443   if (count != noreg) {
4444     mtctr(count);
4445   } else {
4446     li(R0, ReceiverTypeData::row_limit());
4447     mtctr(R0);
4448   }
4449   li(offset, base_receiver_offset);
4450   bind(L_loop_search_receiver);
4451     ldx(R0, offset, mdp);
4452     cmpd(CR0, R0, recv);
4453     beq(CR0, L_found_recv);
4454     addi(offset, offset, receiver_step);
4455   bdnz(L_loop_search_receiver);
4456 
4457   // Fast: no receiver, but profile is not full
4458   if (count != noreg) {
4459     mtctr(count);
4460   } else {
4461     li(R0, ReceiverTypeData::row_limit());
4462     mtctr(R0);
4463   }
4464   li(offset, base_receiver_offset);
4465   bind(L_loop_search_empty);
4466     ldx(R0, offset, mdp);
4467     cmpdi(CR0, R0, 0);
4468     beq(CR0, L_found_empty);
4469     addi(offset, offset, receiver_step);
4470   bdnz(L_loop_search_empty);
4471 
4472   // Slow: Receiver is not found and table is full.
4473   // Increment polymorphic counter instead of receiver slot.
4474   li(offset, poly_count_offset);
4475   b(L_count_update);
4476 
4477   // Slowest: try to install receiver
4478   bind(L_found_empty);
4479 
4480   // Atomically swing receiver slot: null -> recv.
4481   {
4482     Register receiver_addr = offset;
4483     add(receiver_addr, mdp, offset); // kills offset
4484     cmpxchgd(CR0, R0, RegisterOrConstant(0), recv, receiver_addr, MemBarNone, cmpxchgx_hint_atomic_update(),
4485              noreg, nullptr, /* check without ldarx first */ false, /* weak */ true);
4486   }
4487 
4488   // CAS success means the slot now has the receiver we want. CAS failure means
4489   // something had claimed the slot concurrently: it can be the same receiver we want,
4490   // or something else. Since this is a slow path, we can optimize for code density,
4491   // and just restart the search from the beginning.
4492   b(L_restart);
4493 
4494   // Found a receiver, convert its slot offset to corresponding count offset.
4495   bind(L_found_recv);
4496   addi(offset, offset, receiver_to_count_step);
4497 
4498   // Finally, update the counter
4499   bind(L_count_update);
4500   increment_mem64(mdp, offset, DataLayout::counter_increment, /* temp */ (count != noreg) ? count : recv);
4501 }
4502 
4503 #ifdef ASSERT
4504 void MacroAssembler::asm_assert(AsmAssertCond cond, const char *msg) {
4505   Label ok;
4506   switch (cond) {
4507   case eq:
4508     beq(CR0, ok);
4509     break;
4510   case ne:
4511     bne(CR0, ok);
4512     break;
4513   case ge:
4514     bge(CR0, ok);
4515     break;
4516   case gt:
4517     bgt(CR0, ok);
4518     break;
4519   case lt:
4520     blt(CR0, ok);
4521     break;
4522   case le:
4523     ble(CR0, ok);
4524     break;
4525   default:
4526     assert(false, "unknown cond:%d", cond);
4527   }
4528   stop(msg);
4529   bind(ok);
4530 }
4531 
4532 void MacroAssembler::asm_assert_mems_zero(AsmAssertCond cond, int size, int mem_offset,
4533                                           Register mem_base, const char* msg) {
4534   switch (size) {
4535     case 4:
4536       lwz(R0, mem_offset, mem_base);
4537       cmpwi(CR0, R0, 0);
4538       break;
4539     case 8:
4540       ld(R0, mem_offset, mem_base);
4541       cmpdi(CR0, R0, 0);
4542       break;
4543     default:
4544       ShouldNotReachHere();
4545   }
4546   asm_assert(cond, msg);
4547 }
4548 #endif // ASSERT
4549 
4550 void MacroAssembler::verify_coop(Register coop, const char* msg) {
4551   if (!VerifyOops) { return; }
4552   if (UseCompressedOops) { decode_heap_oop(coop); }
4553   verify_oop(coop, msg);
4554   if (UseCompressedOops) { encode_heap_oop(coop, coop); }
4555 }
4556 
4557 // READ: oop. KILL: R0. Volatile floats perhaps.
4558 void MacroAssembler::verify_oop(Register oop, const char* msg) {
4559   if (!VerifyOops) {
4560     return;
4561   }
4562 
4563   address/* FunctionDescriptor** */fd = StubRoutines::verify_oop_subroutine_entry_address();
4564   const Register tmp = R11; // Will be preserved.
4565   const int nbytes_save = MacroAssembler::num_volatile_regs * 8;
4566 
4567   BLOCK_COMMENT("verify_oop {");
4568 
4569   save_volatile_gprs(R1_SP, -nbytes_save); // except R0
4570 
4571   mr_if_needed(R4_ARG2, oop);
4572   save_LR_CR(tmp); // save in old frame
4573   push_frame_reg_args(nbytes_save, tmp);
4574   // load FunctionDescriptor** / entry_address *
4575   load_const_optimized(tmp, fd, R0);
4576   // load FunctionDescriptor* / entry_address
4577   ld(tmp, 0, tmp);
4578   load_const_optimized(R3_ARG1, (address)msg, R0);
4579   // Call destination for its side effect.
4580   call_c(tmp);
4581 
4582   pop_frame();
4583   restore_LR_CR(tmp);
4584   restore_volatile_gprs(R1_SP, -nbytes_save); // except R0
4585 
4586   BLOCK_COMMENT("} verify_oop");
4587 }
4588 
4589 void MacroAssembler::verify_oop_addr(RegisterOrConstant offs, Register base, const char* msg) {
4590   if (!VerifyOops) {
4591     return;
4592   }
4593 
4594   address/* FunctionDescriptor** */fd = StubRoutines::verify_oop_subroutine_entry_address();
4595   const Register tmp = R11; // Will be preserved.
4596   const int nbytes_save = MacroAssembler::num_volatile_regs * 8;
4597   save_volatile_gprs(R1_SP, -nbytes_save); // except R0
4598 
4599   ld(R4_ARG2, offs, base);
4600   save_LR_CR(tmp); // save in old frame
4601   push_frame_reg_args(nbytes_save, tmp);
4602   // load FunctionDescriptor** / entry_address *
4603   load_const_optimized(tmp, fd, R0);
4604   // load FunctionDescriptor* / entry_address
4605   ld(tmp, 0, tmp);
4606   load_const_optimized(R3_ARG1, (address)msg, R0);
4607   // Call destination for its side effect.
4608   call_c(tmp);
4609 
4610   pop_frame();
4611   restore_LR_CR(tmp);
4612   restore_volatile_gprs(R1_SP, -nbytes_save); // except R0
4613 }
4614 
4615 // Call a C-function that prints output.
4616 void MacroAssembler::stop(int type, const char* msg) {
4617   bool msg_present = (msg != nullptr);
4618 
4619 #ifndef PRODUCT
4620   block_comment(err_msg("stop(type %d): %s {", type, msg_present ? msg : "null"));
4621 #else
4622   block_comment("stop {");
4623 #endif
4624 
4625   if (msg_present) {
4626     type |= stop_msg_present;
4627   }
4628   tdi_unchecked(traptoUnconditional, 0/*reg 0*/, type);
4629   if (msg_present) {
4630     emit_int64((uintptr_t)msg);
4631   }
4632 
4633   block_comment("} stop;");
4634 }
4635 
4636 #ifndef PRODUCT
4637 // Write pattern 0x0101010101010101 in memory region [low-before, high+after].
4638 // Val, addr are temp registers.
4639 // If low == addr, addr is killed.
4640 // High is preserved.
4641 void MacroAssembler::zap_from_to(Register low, int before, Register high, int after, Register val, Register addr) {
4642   if (!ZapMemory) return;
4643 
4644   assert_different_registers(low, val);
4645 
4646   BLOCK_COMMENT("zap memory region {");
4647   load_const_optimized(val, 0x0101010101010101);
4648   int size = before + after;
4649   if (low == high && size < 5 && size > 0) {
4650     int offset = -before*BytesPerWord;
4651     for (int i = 0; i < size; ++i) {
4652       std(val, offset, low);
4653       offset += (1*BytesPerWord);
4654     }
4655   } else {
4656     addi(addr, low, -before*BytesPerWord);
4657     assert_different_registers(high, val);
4658     if (after) addi(high, high, after * BytesPerWord);
4659     Label loop;
4660     bind(loop);
4661     std(val, 0, addr);
4662     addi(addr, addr, 8);
4663     cmpd(CR6, addr, high);
4664     ble(CR6, loop);
4665     if (after) addi(high, high, -after * BytesPerWord);  // Correct back to old value.
4666   }
4667   BLOCK_COMMENT("} zap memory region");
4668 }
4669 
4670 #endif // !PRODUCT
4671 
4672 void MacroAssembler::cache_wb(Address line) {
4673   assert(line.index() == noreg, "index should be noreg");
4674   assert(line.disp() == 0, "displacement should be 0");
4675   assert(VM_Version::supports_data_cache_line_flush(), "CPU or OS does not support flush to persistent memory");
4676   // Data Cache Store, not really a flush, so it works like a sync of cache
4677   // line and persistent mem, i.e. copying the cache line to persistent whilst
4678   // not invalidating the cache line.
4679   dcbst(line.base());
4680 }
4681 
4682 void MacroAssembler::cache_wbsync(bool is_presync) {
4683   assert(VM_Version::supports_data_cache_line_flush(), "CPU or OS does not support sync related to persistent memory");
4684   // We only need a post sync barrier. Post means _after_ a cache line flush or
4685   // store instruction, pre means a barrier emitted before such a instructions.
4686   if (!is_presync) {
4687     fence();
4688   }
4689 }
4690 
4691 void MacroAssembler::push_cont_fastpath() {
4692   if (!Continuations::enabled()) return;
4693 
4694   Label done;
4695   ld_ptr(R0, JavaThread::cont_fastpath_offset(), R16_thread);
4696   cmpld(CR0, R1_SP, R0);
4697   ble(CR0, done);          // if (SP <= _cont_fastpath) goto done;
4698   st_ptr(R1_SP, JavaThread::cont_fastpath_offset(), R16_thread);
4699   bind(done);
4700 }
4701 
4702 void MacroAssembler::pop_cont_fastpath() {
4703   if (!Continuations::enabled()) return;
4704 
4705   Label done;
4706   ld_ptr(R0, JavaThread::cont_fastpath_offset(), R16_thread);
4707   cmpld(CR0, R1_SP, R0);
4708   blt(CR0, done);          // if (SP < _cont_fastpath) goto done;
4709   li(R0, 0);
4710   st_ptr(R0, JavaThread::cont_fastpath_offset(), R16_thread);
4711   bind(done);
4712 }
4713 
4714 // Function to flip between unlocked and locked state (fast locking).
4715 // Branches to failed if the state is not as expected with CR0 NE.
4716 // Falls through upon success with CR0 EQ.
4717 // This requires fewer instructions and registers and is easier to use than the
4718 // cmpxchg based implementation.
4719 void MacroAssembler::atomically_flip_locked_state(bool is_unlock, Register obj, Register tmp, Label& failed, int semantics) {
4720   assert_different_registers(obj, tmp, R0);
4721   Label retry;
4722 
4723   if (semantics & MemBarRel) {
4724     release();
4725   }
4726 
4727   bind(retry);
4728   STATIC_ASSERT(markWord::locked_value == 0); // Or need to change this!
4729   if (!is_unlock) {
4730     ldarx(tmp, obj, MacroAssembler::cmpxchgx_hint_acquire_lock());
4731     xori(tmp, tmp, markWord::unlocked_value); // flip unlocked bit
4732     andi_(R0, tmp, markWord::lock_mask_in_place);
4733     bne(CR0, failed); // failed if new header doesn't contain locked_value (which is 0)
4734   } else {
4735     ldarx(tmp, obj, MacroAssembler::cmpxchgx_hint_release_lock());
4736     andi_(R0, tmp, markWord::lock_mask_in_place);
4737     bne(CR0, failed); // failed if old header doesn't contain locked_value (which is 0)
4738     ori(tmp, tmp, markWord::unlocked_value); // set unlocked bit
4739   }
4740   stdcx_(tmp, obj);
4741   bne(CR0, retry);
4742 
4743   if (semantics & MemBarFenceAfter) {
4744     fence();
4745   } else if (semantics & MemBarAcq) {
4746     isync();
4747   }
4748 }
4749 
4750 // Implements fast-locking.
4751 //
4752 //  - obj: the object to be locked
4753 //  - t1, t2: temporary register
4754 void MacroAssembler::fast_lock(Register box, Register obj, Register t1, Register t2, Label& slow) {
4755   assert_different_registers(box, obj, t1, t2, R0);
4756 
4757   Label push;
4758   const Register t = R0;
4759 
4760   if (UseObjectMonitorTable) {
4761     // Clear cache in case fast locking succeeds or we need to take the slow-path.
4762     li(t, 0);
4763     std(t, in_bytes(BasicObjectLock::lock_offset()) + BasicLock::object_monitor_cache_offset_in_bytes(), box);
4764   }
4765 
4766   if (DiagnoseSyncOnValueBasedClasses != 0) {
4767     load_klass(t1, obj);
4768     lbz(t1, in_bytes(Klass::misc_flags_offset()), t1);
4769     testbitdi(CR0, R0, t1, exact_log2(KlassFlags::_misc_is_value_based_class));
4770     bne(CR0, slow);
4771   }
4772 
4773   const Register top = t1;
4774   const Register mark = t2;
4775 
4776   // Check if the lock-stack is full.
4777   lwz(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4778   cmplwi(CR0, top, LockStack::end_offset());
4779   bge(CR0, slow);
4780 
4781   // The underflow check is elided. The recursive check will always fail
4782   // when the lock stack is empty because of the _bad_oop_sentinel field.
4783 
4784   // Check for recursion.
4785   subi(t, top, oopSize);
4786   ldx(t, R16_thread, t);
4787   cmpd(CR0, obj, t);
4788   beq(CR0, push);
4789 
4790   // Check header for monitor (0b10) or locked (0b00).
4791   ld(mark, oopDesc::mark_offset_in_bytes(), obj);
4792   xori(t, mark, markWord::unlocked_value);
4793   andi_(t, t, markWord::lock_mask_in_place);
4794   bne(CR0, slow);
4795 
4796   // Try to lock. Transition lock bits 0b01 => 0b00
4797   atomically_flip_locked_state(/* is_unlock */ false, obj, mark, slow, MacroAssembler::MemBarAcq);
4798 
4799   bind(push);
4800   // After successful lock, push object on lock-stack
4801   stdx(obj, R16_thread, top);
4802   addi(top, top, oopSize);
4803   stw(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4804 }
4805 
4806 // Implements fast-unlocking.
4807 //
4808 // - obj: the object to be unlocked
4809 //  - t1: temporary register
4810 void MacroAssembler::fast_unlock(Register obj, Register t1, Label& slow) {
4811   assert_different_registers(obj, t1);
4812 
4813 #ifdef ASSERT
4814   {
4815     // The following checks rely on the fact that LockStack is only ever modified by
4816     // its owning thread, even if the lock got inflated concurrently; removal of LockStack
4817     // entries after inflation will happen delayed in that case.
4818 
4819     // Check for lock-stack underflow.
4820     Label stack_ok;
4821     lwz(t1, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4822     cmplwi(CR0, t1, LockStack::start_offset());
4823     bge(CR0, stack_ok);
4824     stop("Lock-stack underflow");
4825     bind(stack_ok);
4826   }
4827 #endif
4828 
4829   Label unlocked, push_and_slow;
4830   const Register top = t1;
4831   const Register mark = R0;
4832   Register t = R0;
4833 
4834   // Check if obj is top of lock-stack.
4835   lwz(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4836   subi(top, top, oopSize);
4837   ldx(t, R16_thread, top);
4838   cmpd(CR0, obj, t);
4839   bne(CR0, slow);
4840 
4841   // Pop lock-stack.
4842   DEBUG_ONLY(li(t, 0);)
4843   DEBUG_ONLY(stdx(t, R16_thread, top);)
4844   stw(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4845 
4846   // The underflow check is elided. The recursive check will always fail
4847   // when the lock stack is empty because of the _bad_oop_sentinel field.
4848 
4849   // Check if recursive.
4850   subi(t, top, oopSize);
4851   ldx(t, R16_thread, t);
4852   cmpd(CR0, obj, t);
4853   beq(CR0, unlocked);
4854 
4855   // Use top as tmp
4856   t = top;
4857 
4858   // Not recursive. Check header for monitor (0b10).
4859   ld(mark, oopDesc::mark_offset_in_bytes(), obj);
4860   andi_(t, mark, markWord::monitor_value);
4861   bne(CR0, push_and_slow);
4862 
4863 #ifdef ASSERT
4864   // Check header not unlocked (0b01).
4865   Label not_unlocked;
4866   andi_(t, mark, markWord::unlocked_value);
4867   beq(CR0, not_unlocked);
4868   stop("fast_unlock already unlocked");
4869   bind(not_unlocked);
4870 #endif
4871 
4872   // Try to unlock. Transition lock bits 0b00 => 0b01
4873   atomically_flip_locked_state(/* is_unlock */ true, obj, t, push_and_slow, MacroAssembler::MemBarRel);
4874   b(unlocked);
4875 
4876   bind(push_and_slow);
4877 
4878   // Restore lock-stack and handle the unlock in runtime.
4879   lwz(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4880   DEBUG_ONLY(stdx(obj, R16_thread, top);)
4881   addi(top, top, oopSize);
4882   stw(top, in_bytes(JavaThread::lock_stack_top_offset()), R16_thread);
4883   b(slow);
4884 
4885   bind(unlocked);
4886 }