1 /*
   2  * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "c1/c1_Compilation.hpp"
  27 #include "c1/c1_FrameMap.hpp"
  28 #include "c1/c1_Instruction.hpp"
  29 #include "c1/c1_LIRAssembler.hpp"
  30 #include "c1/c1_LIRGenerator.hpp"
  31 #include "c1/c1_Runtime1.hpp"
  32 #include "c1/c1_ValueStack.hpp"
  33 #include "ci/ciArray.hpp"
  34 #include "ci/ciObjArrayKlass.hpp"
  35 #include "ci/ciTypeArrayKlass.hpp"
  36 #include "runtime/sharedRuntime.hpp"
  37 #include "runtime/stubRoutines.hpp"
  38 #include "vmreg_x86.inline.hpp"
  39 
  40 #ifdef ASSERT
  41 #define __ gen()->lir(__FILE__, __LINE__)->
  42 #else
  43 #define __ gen()->lir()->
  44 #endif
  45 
  46 // Item will be loaded into a byte register; Intel only
  47 void LIRItem::load_byte_item() {
  48   load_item();
  49   LIR_Opr res = result();
  50 
  51   if (!res->is_virtual() || !_gen->is_vreg_flag_set(res, LIRGenerator::byte_reg)) {
  52     // make sure that it is a byte register
  53     assert(!value()->type()->is_float() && !value()->type()->is_double(),
  54            "can't load floats in byte register");
  55     LIR_Opr reg = _gen->rlock_byte(T_BYTE);
  56     __ move(res, reg);
  57 
  58     _result = reg;
  59   }
  60 }
  61 
  62 
  63 void LIRItem::load_nonconstant() {
  64   LIR_Opr r = value()->operand();
  65   if (r->is_constant()) {
  66     _result = r;
  67   } else {
  68     load_item();
  69   }
  70 }
  71 
  72 //--------------------------------------------------------------
  73 //               LIRGenerator
  74 //--------------------------------------------------------------
  75 
  76 
  77 LIR_Opr LIRGenerator::exceptionOopOpr() { return FrameMap::rax_oop_opr; }
  78 LIR_Opr LIRGenerator::exceptionPcOpr()  { return FrameMap::rdx_opr; }
  79 LIR_Opr LIRGenerator::divInOpr()        { return FrameMap::rax_opr; }
  80 LIR_Opr LIRGenerator::divOutOpr()       { return FrameMap::rax_opr; }
  81 LIR_Opr LIRGenerator::remOutOpr()       { return FrameMap::rdx_opr; }
  82 LIR_Opr LIRGenerator::shiftCountOpr()   { return FrameMap::rcx_opr; }
  83 LIR_Opr LIRGenerator::syncTempOpr()     { return FrameMap::rax_opr; }
  84 LIR_Opr LIRGenerator::getThreadTemp()   { return LIR_OprFact::illegalOpr; }
  85 
  86 
  87 LIR_Opr LIRGenerator::result_register_for(ValueType* type, bool callee) {
  88   LIR_Opr opr;
  89   switch (type->tag()) {
  90     case intTag:     opr = FrameMap::rax_opr;          break;
  91     case objectTag:  opr = FrameMap::rax_oop_opr;      break;
  92     case longTag:    opr = FrameMap::long0_opr;        break;
  93     case floatTag:   opr = UseSSE >= 1 ? FrameMap::xmm0_float_opr  : FrameMap::fpu0_float_opr;  break;
  94     case doubleTag:  opr = UseSSE >= 2 ? FrameMap::xmm0_double_opr : FrameMap::fpu0_double_opr;  break;
  95 
  96     case addressTag:
  97     default: ShouldNotReachHere(); return LIR_OprFact::illegalOpr;
  98   }
  99 
 100   assert(opr->type_field() == as_OprType(as_BasicType(type)), "type mismatch");
 101   return opr;
 102 }
 103 
 104 
 105 LIR_Opr LIRGenerator::rlock_byte(BasicType type) {
 106   LIR_Opr reg = new_register(T_INT);
 107   set_vreg_flag(reg, LIRGenerator::byte_reg);
 108   return reg;
 109 }
 110 
 111 
 112 //--------- loading items into registers --------------------------------
 113 
 114 
 115 // i486 instructions can inline constants
 116 bool LIRGenerator::can_store_as_constant(Value v, BasicType type) const {
 117   if (type == T_SHORT || type == T_CHAR) {
 118     // there is no immediate move of word values in asembler_i486.?pp
 119     return false;
 120   }
 121   Constant* c = v->as_Constant();
 122   if (c && c->state_before() == NULL) {
 123     // constants of any type can be stored directly, except for
 124     // unloaded object constants.
 125     return true;
 126   }
 127   return false;
 128 }
 129 
 130 
 131 bool LIRGenerator::can_inline_as_constant(Value v) const {
 132   if (v->type()->tag() == longTag) return false;
 133   return v->type()->tag() != objectTag ||
 134     (v->type()->is_constant() && v->type()->as_ObjectType()->constant_value()->is_null_object());
 135 }
 136 
 137 
 138 bool LIRGenerator::can_inline_as_constant(LIR_Const* c) const {
 139   if (c->type() == T_LONG) return false;
 140   return c->type() != T_OBJECT || c->as_jobject() == NULL;
 141 }
 142 
 143 
 144 LIR_Opr LIRGenerator::safepoint_poll_register() {
 145   return LIR_OprFact::illegalOpr;
 146 }
 147 
 148 
 149 LIR_Address* LIRGenerator::generate_address(LIR_Opr base, LIR_Opr index,
 150                                             int shift, int disp, BasicType type) {
 151   assert(base->is_register(), "must be");
 152   if (index->is_constant()) {
 153     return new LIR_Address(base,
 154                            (index->as_constant_ptr()->as_jint() << shift) + disp,
 155                            type);
 156   } else {
 157     return new LIR_Address(base, index, (LIR_Address::Scale)shift, disp, type);
 158   }
 159 }
 160 
 161 
 162 LIR_Address* LIRGenerator::emit_array_address(LIR_Opr array_opr, LIR_Opr index_opr,
 163                                               BasicType type, bool needs_card_mark) {
 164   int offset_in_bytes = arrayOopDesc::base_offset_in_bytes(type);
 165 
 166   LIR_Address* addr;
 167   if (index_opr->is_constant()) {
 168     int elem_size = type2aelembytes(type);
 169     addr = new LIR_Address(array_opr,
 170                            offset_in_bytes + index_opr->as_jint() * elem_size, type);
 171   } else {
 172 #ifdef _LP64
 173     if (index_opr->type() == T_INT) {
 174       LIR_Opr tmp = new_register(T_LONG);
 175       __ convert(Bytecodes::_i2l, index_opr, tmp);
 176       index_opr = tmp;
 177     }
 178 #endif // _LP64
 179     addr =  new LIR_Address(array_opr,
 180                             index_opr,
 181                             LIR_Address::scale(type),
 182                             offset_in_bytes, type);
 183   }
 184   if (needs_card_mark) {
 185     // This store will need a precise card mark, so go ahead and
 186     // compute the full adddres instead of computing once for the
 187     // store and again for the card mark.
 188     LIR_Opr tmp = new_pointer_register();
 189     __ leal(LIR_OprFact::address(addr), tmp);
 190     return new LIR_Address(tmp, type);
 191   } else {
 192     return addr;
 193   }
 194 }
 195 
 196 
 197 LIR_Opr LIRGenerator::load_immediate(int x, BasicType type) {
 198   LIR_Opr r;
 199   if (type == T_LONG) {
 200     r = LIR_OprFact::longConst(x);
 201   } else if (type == T_INT) {
 202     r = LIR_OprFact::intConst(x);
 203   } else {
 204     ShouldNotReachHere();
 205   }
 206   return r;
 207 }
 208 
 209 void LIRGenerator::increment_counter(address counter, BasicType type, int step) {
 210   LIR_Opr pointer = new_pointer_register();
 211   __ move(LIR_OprFact::intptrConst(counter), pointer);
 212   LIR_Address* addr = new LIR_Address(pointer, type);
 213   increment_counter(addr, step);
 214 }
 215 
 216 
 217 void LIRGenerator::increment_counter(LIR_Address* addr, int step) {
 218   __ add((LIR_Opr)addr, LIR_OprFact::intConst(step), (LIR_Opr)addr);
 219 }
 220 
 221 void LIRGenerator::cmp_mem_int(LIR_Condition condition, LIR_Opr base, int disp, int c, CodeEmitInfo* info) {
 222   __ cmp_mem_int(condition, base, disp, c, info);
 223 }
 224 
 225 
 226 void LIRGenerator::cmp_reg_mem(LIR_Condition condition, LIR_Opr reg, LIR_Opr base, int disp, BasicType type, CodeEmitInfo* info) {
 227   __ cmp_reg_mem(condition, reg, new LIR_Address(base, disp, type), info);
 228 }
 229 
 230 
 231 void LIRGenerator::cmp_reg_mem(LIR_Condition condition, LIR_Opr reg, LIR_Opr base, LIR_Opr disp, BasicType type, CodeEmitInfo* info) {
 232   __ cmp_reg_mem(condition, reg, new LIR_Address(base, disp, type), info);
 233 }
 234 
 235 
 236 bool LIRGenerator::strength_reduce_multiply(LIR_Opr left, int c, LIR_Opr result, LIR_Opr tmp) {
 237   if (tmp->is_valid()) {
 238     if (is_power_of_2(c + 1)) {
 239       __ move(left, tmp);
 240       __ shift_left(left, log2_intptr(c + 1), left);
 241       __ sub(left, tmp, result);
 242       return true;
 243     } else if (is_power_of_2(c - 1)) {
 244       __ move(left, tmp);
 245       __ shift_left(left, log2_intptr(c - 1), left);
 246       __ add(left, tmp, result);
 247       return true;
 248     }
 249   }
 250   return false;
 251 }
 252 
 253 
 254 void LIRGenerator::store_stack_parameter (LIR_Opr item, ByteSize offset_from_sp) {
 255   BasicType type = item->type();
 256   __ store(item, new LIR_Address(FrameMap::rsp_opr, in_bytes(offset_from_sp), type));
 257 }
 258 
 259 //----------------------------------------------------------------------
 260 //             visitor functions
 261 //----------------------------------------------------------------------
 262 
 263 
 264 void LIRGenerator::do_StoreIndexed(StoreIndexed* x) {
 265   assert(x->is_pinned(),"");
 266   bool needs_range_check = x->compute_needs_range_check();
 267   bool use_length = x->length() != NULL;
 268   bool obj_store = x->elt_type() == T_ARRAY || x->elt_type() == T_OBJECT;
 269   bool needs_store_check = obj_store && (x->value()->as_Constant() == NULL ||
 270                                          !get_jobject_constant(x->value())->is_null_object() ||
 271                                          x->should_profile());
 272 
 273   LIRItem array(x->array(), this);
 274   LIRItem index(x->index(), this);
 275   LIRItem value(x->value(), this);
 276   LIRItem length(this);
 277 
 278   array.load_item();
 279   index.load_nonconstant();
 280 
 281   if (use_length && needs_range_check) {
 282     length.set_instruction(x->length());
 283     length.load_item();
 284 
 285   }
 286   if (needs_store_check) {
 287     value.load_item();
 288   } else {
 289     value.load_for_store(x->elt_type());
 290   }
 291 
 292   set_no_result(x);
 293 
 294   // the CodeEmitInfo must be duplicated for each different
 295   // LIR-instruction because spilling can occur anywhere between two
 296   // instructions and so the debug information must be different
 297   CodeEmitInfo* range_check_info = state_for(x);
 298   CodeEmitInfo* null_check_info = NULL;
 299   if (x->needs_null_check()) {
 300     null_check_info = new CodeEmitInfo(range_check_info);
 301   }
 302 
 303   LIR_Opr ary = array.result();
 304   ary = shenandoah_write_barrier(ary, null_check_info, x->needs_null_check());
 305   LIR_Opr val = value.result();
 306   if (obj_store && UseShenandoahGC) {
 307     if (! val->is_register()) {
 308       assert(val->is_constant(), "expect constant");
 309     } else {
 310       val = shenandoah_read_barrier(val, NULL, true);
 311     }
 312   }
 313 
 314   // emit array address setup early so it schedules better
 315   LIR_Address* array_addr = emit_array_address(ary, index.result(), x->elt_type(), obj_store);
 316 
 317   if (GenerateRangeChecks && needs_range_check) {
 318     if (use_length) {
 319       __ cmp(lir_cond_belowEqual, length.result(), index.result());
 320       __ branch(lir_cond_belowEqual, T_INT, new RangeCheckStub(range_check_info, index.result()));
 321     } else {
 322       array_range_check(ary, index.result(), null_check_info, range_check_info);
 323       // range_check also does the null check
 324       null_check_info = NULL;
 325     }
 326   }
 327 
 328   if (GenerateArrayStoreCheck && needs_store_check) {
 329     LIR_Opr tmp1 = new_register(objectType);
 330     LIR_Opr tmp2 = new_register(objectType);
 331     LIR_Opr tmp3 = new_register(objectType);
 332 
 333     CodeEmitInfo* store_check_info = new CodeEmitInfo(range_check_info);
 334     __ store_check(val, ary, tmp1, tmp2, tmp3, store_check_info, x->profiled_method(), x->profiled_bci());
 335   }
 336 
 337   if (obj_store) {
 338     // Needs GC write barriers.
 339     pre_barrier(LIR_OprFact::address(array_addr), LIR_OprFact::illegalOpr /* pre_val */,
 340                 true /* do_load */, false /* patch */, NULL);
 341     __ move(val, array_addr, null_check_info);
 342     // Seems to be a precise
 343     post_barrier(LIR_OprFact::address(array_addr), value.result());
 344   } else {
 345     __ move(val, array_addr, null_check_info);
 346   }
 347 }
 348 
 349 
 350 void LIRGenerator::do_MonitorEnter(MonitorEnter* x) {
 351   assert(x->is_pinned(),"");
 352   LIRItem obj(x->obj(), this);
 353   obj.load_item();
 354 
 355   set_no_result(x);
 356 
 357   // "lock" stores the address of the monitor stack slot, so this is not an oop
 358   LIR_Opr lock = new_register(T_INT);
 359   // Need a scratch register for biased locking on x86
 360   LIR_Opr scratch = LIR_OprFact::illegalOpr;
 361   if (UseBiasedLocking) {
 362     scratch = new_register(T_INT);
 363   }
 364 
 365   CodeEmitInfo* info_for_exception = NULL;
 366   if (x->needs_null_check()) {
 367     info_for_exception = state_for(x);
 368   }
 369   // this CodeEmitInfo must not have the xhandlers because here the
 370   // object is already locked (xhandlers expect object to be unlocked)
 371   CodeEmitInfo* info = state_for(x, x->state(), true);
 372   LIR_Opr obj_opr = obj.result();
 373   obj_opr = shenandoah_write_barrier(obj_opr, state_for(x), x->needs_null_check());
 374   monitor_enter(obj_opr, lock, syncTempOpr(), scratch,
 375                         x->monitor_no(), info_for_exception, info);
 376 }
 377 
 378 
 379 void LIRGenerator::do_MonitorExit(MonitorExit* x) {
 380   assert(x->is_pinned(),"");
 381 
 382   LIRItem obj(x->obj(), this);
 383   obj.dont_load_item();
 384 
 385   LIR_Opr lock = new_register(T_INT);
 386   LIR_Opr obj_temp = new_register(T_INT);
 387   set_no_result(x);
 388   monitor_exit(obj_temp, lock, syncTempOpr(), LIR_OprFact::illegalOpr, x->monitor_no());
 389 }
 390 
 391 
 392 // _ineg, _lneg, _fneg, _dneg
 393 void LIRGenerator::do_NegateOp(NegateOp* x) {
 394   LIRItem value(x->x(), this);
 395   value.set_destroys_register();
 396   value.load_item();
 397   LIR_Opr reg = rlock(x);
 398   __ negate(value.result(), reg);
 399 
 400   set_result(x, round_item(reg));
 401 }
 402 
 403 
 404 // for  _fadd, _fmul, _fsub, _fdiv, _frem
 405 //      _dadd, _dmul, _dsub, _ddiv, _drem
 406 void LIRGenerator::do_ArithmeticOp_FPU(ArithmeticOp* x) {
 407   LIRItem left(x->x(),  this);
 408   LIRItem right(x->y(), this);
 409   LIRItem* left_arg  = &left;
 410   LIRItem* right_arg = &right;
 411   assert(!left.is_stack() || !right.is_stack(), "can't both be memory operands");
 412   bool must_load_both = (x->op() == Bytecodes::_frem || x->op() == Bytecodes::_drem);
 413   if (left.is_register() || x->x()->type()->is_constant() || must_load_both) {
 414     left.load_item();
 415   } else {
 416     left.dont_load_item();
 417   }
 418 
 419   // do not load right operand if it is a constant.  only 0 and 1 are
 420   // loaded because there are special instructions for loading them
 421   // without memory access (not needed for SSE2 instructions)
 422   bool must_load_right = false;
 423   if (right.is_constant()) {
 424     LIR_Const* c = right.result()->as_constant_ptr();
 425     assert(c != NULL, "invalid constant");
 426     assert(c->type() == T_FLOAT || c->type() == T_DOUBLE, "invalid type");
 427 
 428     if (c->type() == T_FLOAT) {
 429       must_load_right = UseSSE < 1 && (c->is_one_float() || c->is_zero_float());
 430     } else {
 431       must_load_right = UseSSE < 2 && (c->is_one_double() || c->is_zero_double());
 432     }
 433   }
 434 
 435   if (must_load_both) {
 436     // frem and drem destroy also right operand, so move it to a new register
 437     right.set_destroys_register();
 438     right.load_item();
 439   } else if (right.is_register() || must_load_right) {
 440     right.load_item();
 441   } else {
 442     right.dont_load_item();
 443   }
 444   LIR_Opr reg = rlock(x);
 445   LIR_Opr tmp = LIR_OprFact::illegalOpr;
 446   if (x->is_strictfp() && (x->op() == Bytecodes::_dmul || x->op() == Bytecodes::_ddiv)) {
 447     tmp = new_register(T_DOUBLE);
 448   }
 449 
 450   if ((UseSSE >= 1 && x->op() == Bytecodes::_frem) || (UseSSE >= 2 && x->op() == Bytecodes::_drem)) {
 451     // special handling for frem and drem: no SSE instruction, so must use FPU with temporary fpu stack slots
 452     LIR_Opr fpu0, fpu1;
 453     if (x->op() == Bytecodes::_frem) {
 454       fpu0 = LIR_OprFact::single_fpu(0);
 455       fpu1 = LIR_OprFact::single_fpu(1);
 456     } else {
 457       fpu0 = LIR_OprFact::double_fpu(0);
 458       fpu1 = LIR_OprFact::double_fpu(1);
 459     }
 460     __ move(right.result(), fpu1); // order of left and right operand is important!
 461     __ move(left.result(), fpu0);
 462     __ rem (fpu0, fpu1, fpu0);
 463     __ move(fpu0, reg);
 464 
 465   } else {
 466     arithmetic_op_fpu(x->op(), reg, left.result(), right.result(), x->is_strictfp(), tmp);
 467   }
 468 
 469   set_result(x, round_item(reg));
 470 }
 471 
 472 
 473 // for  _ladd, _lmul, _lsub, _ldiv, _lrem
 474 void LIRGenerator::do_ArithmeticOp_Long(ArithmeticOp* x) {
 475   if (x->op() == Bytecodes::_ldiv || x->op() == Bytecodes::_lrem ) {
 476     // long division is implemented as a direct call into the runtime
 477     LIRItem left(x->x(), this);
 478     LIRItem right(x->y(), this);
 479 
 480     // the check for division by zero destroys the right operand
 481     right.set_destroys_register();
 482 
 483     BasicTypeList signature(2);
 484     signature.append(T_LONG);
 485     signature.append(T_LONG);
 486     CallingConvention* cc = frame_map()->c_calling_convention(&signature);
 487 
 488     // check for division by zero (destroys registers of right operand!)
 489     CodeEmitInfo* info = state_for(x);
 490 
 491     const LIR_Opr result_reg = result_register_for(x->type());
 492     left.load_item_force(cc->at(1));
 493     right.load_item();
 494 
 495     __ move(right.result(), cc->at(0));
 496 
 497     __ cmp(lir_cond_equal, right.result(), LIR_OprFact::longConst(0));
 498     __ branch(lir_cond_equal, T_LONG, new DivByZeroStub(info));
 499 
 500     address entry;
 501     switch (x->op()) {
 502     case Bytecodes::_lrem:
 503       entry = CAST_FROM_FN_PTR(address, SharedRuntime::lrem);
 504       break; // check if dividend is 0 is done elsewhere
 505     case Bytecodes::_ldiv:
 506       entry = CAST_FROM_FN_PTR(address, SharedRuntime::ldiv);
 507       break; // check if dividend is 0 is done elsewhere
 508     case Bytecodes::_lmul:
 509       entry = CAST_FROM_FN_PTR(address, SharedRuntime::lmul);
 510       break;
 511     default:
 512       ShouldNotReachHere();
 513     }
 514 
 515     LIR_Opr result = rlock_result(x);
 516     __ call_runtime_leaf(entry, getThreadTemp(), result_reg, cc->args());
 517     __ move(result_reg, result);
 518   } else if (x->op() == Bytecodes::_lmul) {
 519     // missing test if instr is commutative and if we should swap
 520     LIRItem left(x->x(), this);
 521     LIRItem right(x->y(), this);
 522 
 523     // right register is destroyed by the long mul, so it must be
 524     // copied to a new register.
 525     right.set_destroys_register();
 526 
 527     left.load_item();
 528     right.load_item();
 529 
 530     LIR_Opr reg = FrameMap::long0_opr;
 531     arithmetic_op_long(x->op(), reg, left.result(), right.result(), NULL);
 532     LIR_Opr result = rlock_result(x);
 533     __ move(reg, result);
 534   } else {
 535     // missing test if instr is commutative and if we should swap
 536     LIRItem left(x->x(), this);
 537     LIRItem right(x->y(), this);
 538 
 539     left.load_item();
 540     // don't load constants to save register
 541     right.load_nonconstant();
 542     rlock_result(x);
 543     arithmetic_op_long(x->op(), x->operand(), left.result(), right.result(), NULL);
 544   }
 545 }
 546 
 547 
 548 
 549 // for: _iadd, _imul, _isub, _idiv, _irem
 550 void LIRGenerator::do_ArithmeticOp_Int(ArithmeticOp* x) {
 551   if (x->op() == Bytecodes::_idiv || x->op() == Bytecodes::_irem) {
 552     // The requirements for division and modulo
 553     // input : rax,: dividend                         min_int
 554     //         reg: divisor   (may not be rax,/rdx)   -1
 555     //
 556     // output: rax,: quotient  (= rax, idiv reg)       min_int
 557     //         rdx: remainder (= rax, irem reg)       0
 558 
 559     // rax, and rdx will be destroyed
 560 
 561     // Note: does this invalidate the spec ???
 562     LIRItem right(x->y(), this);
 563     LIRItem left(x->x() , this);   // visit left second, so that the is_register test is valid
 564 
 565     // call state_for before load_item_force because state_for may
 566     // force the evaluation of other instructions that are needed for
 567     // correct debug info.  Otherwise the live range of the fix
 568     // register might be too long.
 569     CodeEmitInfo* info = state_for(x);
 570 
 571     left.load_item_force(divInOpr());
 572 
 573     right.load_item();
 574 
 575     LIR_Opr result = rlock_result(x);
 576     LIR_Opr result_reg;
 577     if (x->op() == Bytecodes::_idiv) {
 578       result_reg = divOutOpr();
 579     } else {
 580       result_reg = remOutOpr();
 581     }
 582 
 583     if (!ImplicitDiv0Checks) {
 584       __ cmp(lir_cond_equal, right.result(), LIR_OprFact::intConst(0));
 585       __ branch(lir_cond_equal, T_INT, new DivByZeroStub(info));
 586     }
 587     LIR_Opr tmp = FrameMap::rdx_opr; // idiv and irem use rdx in their implementation
 588     if (x->op() == Bytecodes::_irem) {
 589       __ irem(left.result(), right.result(), result_reg, tmp, info);
 590     } else if (x->op() == Bytecodes::_idiv) {
 591       __ idiv(left.result(), right.result(), result_reg, tmp, info);
 592     } else {
 593       ShouldNotReachHere();
 594     }
 595 
 596     __ move(result_reg, result);
 597   } else {
 598     // missing test if instr is commutative and if we should swap
 599     LIRItem left(x->x(),  this);
 600     LIRItem right(x->y(), this);
 601     LIRItem* left_arg = &left;
 602     LIRItem* right_arg = &right;
 603     if (x->is_commutative() && left.is_stack() && right.is_register()) {
 604       // swap them if left is real stack (or cached) and right is real register(not cached)
 605       left_arg = &right;
 606       right_arg = &left;
 607     }
 608 
 609     left_arg->load_item();
 610 
 611     // do not need to load right, as we can handle stack and constants
 612     if (x->op() == Bytecodes::_imul ) {
 613       // check if we can use shift instead
 614       bool use_constant = false;
 615       bool use_tmp = false;
 616       if (right_arg->is_constant()) {
 617         int iconst = right_arg->get_jint_constant();
 618         if (iconst > 0) {
 619           if (is_power_of_2(iconst)) {
 620             use_constant = true;
 621           } else if (is_power_of_2(iconst - 1) || is_power_of_2(iconst + 1)) {
 622             use_constant = true;
 623             use_tmp = true;
 624           }
 625         }
 626       }
 627       if (use_constant) {
 628         right_arg->dont_load_item();
 629       } else {
 630         right_arg->load_item();
 631       }
 632       LIR_Opr tmp = LIR_OprFact::illegalOpr;
 633       if (use_tmp) {
 634         tmp = new_register(T_INT);
 635       }
 636       rlock_result(x);
 637 
 638       arithmetic_op_int(x->op(), x->operand(), left_arg->result(), right_arg->result(), tmp);
 639     } else {
 640       right_arg->dont_load_item();
 641       rlock_result(x);
 642       LIR_Opr tmp = LIR_OprFact::illegalOpr;
 643       arithmetic_op_int(x->op(), x->operand(), left_arg->result(), right_arg->result(), tmp);
 644     }
 645   }
 646 }
 647 
 648 
 649 void LIRGenerator::do_ArithmeticOp(ArithmeticOp* x) {
 650   // when an operand with use count 1 is the left operand, then it is
 651   // likely that no move for 2-operand-LIR-form is necessary
 652   if (x->is_commutative() && x->y()->as_Constant() == NULL && x->x()->use_count() > x->y()->use_count()) {
 653     x->swap_operands();
 654   }
 655 
 656   ValueTag tag = x->type()->tag();
 657   assert(x->x()->type()->tag() == tag && x->y()->type()->tag() == tag, "wrong parameters");
 658   switch (tag) {
 659     case floatTag:
 660     case doubleTag:  do_ArithmeticOp_FPU(x);  return;
 661     case longTag:    do_ArithmeticOp_Long(x); return;
 662     case intTag:     do_ArithmeticOp_Int(x);  return;
 663   }
 664   ShouldNotReachHere();
 665 }
 666 
 667 
 668 // _ishl, _lshl, _ishr, _lshr, _iushr, _lushr
 669 void LIRGenerator::do_ShiftOp(ShiftOp* x) {
 670   // count must always be in rcx
 671   LIRItem value(x->x(), this);
 672   LIRItem count(x->y(), this);
 673 
 674   ValueTag elemType = x->type()->tag();
 675   bool must_load_count = !count.is_constant() || elemType == longTag;
 676   if (must_load_count) {
 677     // count for long must be in register
 678     count.load_item_force(shiftCountOpr());
 679   } else {
 680     count.dont_load_item();
 681   }
 682   value.load_item();
 683   LIR_Opr reg = rlock_result(x);
 684 
 685   shift_op(x->op(), reg, value.result(), count.result(), LIR_OprFact::illegalOpr);
 686 }
 687 
 688 
 689 // _iand, _land, _ior, _lor, _ixor, _lxor
 690 void LIRGenerator::do_LogicOp(LogicOp* x) {
 691   // when an operand with use count 1 is the left operand, then it is
 692   // likely that no move for 2-operand-LIR-form is necessary
 693   if (x->is_commutative() && x->y()->as_Constant() == NULL && x->x()->use_count() > x->y()->use_count()) {
 694     x->swap_operands();
 695   }
 696 
 697   LIRItem left(x->x(), this);
 698   LIRItem right(x->y(), this);
 699 
 700   left.load_item();
 701   right.load_nonconstant();
 702   LIR_Opr reg = rlock_result(x);
 703 
 704   logic_op(x->op(), reg, left.result(), right.result());
 705 }
 706 
 707 
 708 
 709 // _lcmp, _fcmpl, _fcmpg, _dcmpl, _dcmpg
 710 void LIRGenerator::do_CompareOp(CompareOp* x) {
 711   LIRItem left(x->x(), this);
 712   LIRItem right(x->y(), this);
 713   ValueTag tag = x->x()->type()->tag();
 714   if (tag == longTag) {
 715     left.set_destroys_register();
 716   }
 717   left.load_item();
 718   right.load_item();
 719   LIR_Opr reg = rlock_result(x);
 720 
 721   if (x->x()->type()->is_float_kind()) {
 722     Bytecodes::Code code = x->op();
 723     __ fcmp2int(left.result(), right.result(), reg, (code == Bytecodes::_fcmpl || code == Bytecodes::_dcmpl));
 724   } else if (x->x()->type()->tag() == longTag) {
 725     __ lcmp2int(left.result(), right.result(), reg);
 726   } else {
 727     Unimplemented();
 728   }
 729 }
 730 
 731 
 732 void LIRGenerator::do_CompareAndSwap(Intrinsic* x, ValueType* type) {
 733   assert(x->number_of_arguments() == 4, "wrong type");
 734   LIRItem obj   (x->argument_at(0), this);  // object
 735   LIRItem offset(x->argument_at(1), this);  // offset of field
 736   LIRItem cmp   (x->argument_at(2), this);  // value to compare with field
 737   LIRItem val   (x->argument_at(3), this);  // replace field with val if matches cmp
 738 
 739   assert(obj.type()->tag() == objectTag, "invalid type");
 740 
 741   // In 64bit the type can be long, sparc doesn't have this assert
 742   // assert(offset.type()->tag() == intTag, "invalid type");
 743 
 744   assert(cmp.type()->tag() == type->tag(), "invalid type");
 745   assert(val.type()->tag() == type->tag(), "invalid type");
 746 
 747   // get address of field
 748   obj.load_item();
 749   offset.load_nonconstant();
 750 
 751   if (type == objectType) {
 752     cmp.load_item_force(FrameMap::rax_oop_opr);
 753     val.load_item();
 754   } else if (type == intType) {
 755     cmp.load_item_force(FrameMap::rax_opr);
 756     val.load_item();
 757   } else if (type == longType) {
 758     cmp.load_item_force(FrameMap::long0_opr);
 759     val.load_item_force(FrameMap::long1_opr);
 760   } else {
 761     ShouldNotReachHere();
 762   }
 763 
 764   LIR_Opr addr = new_pointer_register();
 765   LIR_Address* a;
 766 
 767   LIR_Opr obj_op = obj.result();
 768   obj_op = shenandoah_write_barrier(obj_op, NULL, false);
 769 
 770   if(offset.result()->is_constant()) {
 771 #ifdef _LP64
 772     jlong c = offset.result()->as_jlong();
 773     if ((jlong)((jint)c) == c) {
 774       a = new LIR_Address(obj_op,
 775                           (jint)c,
 776                           as_BasicType(type));
 777     } else {
 778       LIR_Opr tmp = new_register(T_LONG);
 779       __ move(offset.result(), tmp);
 780       a = new LIR_Address(obj_op,
 781                           tmp,
 782                           as_BasicType(type));
 783     }
 784 #else
 785     a = new LIR_Address(obj_op,
 786                         offset.result()->as_jint(),
 787                         as_BasicType(type));
 788 #endif
 789   } else {
 790     a = new LIR_Address(obj_op,
 791                         offset.result(),
 792                         LIR_Address::times_1,
 793                         0,
 794                         as_BasicType(type));
 795   }
 796   __ leal(LIR_OprFact::address(a), addr);
 797 
 798   if (type == objectType) {  // Write-barrier needed for Object fields.
 799     // Do the pre-write barrier, if any.
 800     pre_barrier(addr, LIR_OprFact::illegalOpr /* pre_val */,
 801                 true /* do_load */, false /* patch */, NULL);
 802   }
 803 
 804   LIR_Opr ill = LIR_OprFact::illegalOpr;  // for convenience
 805 
 806   LIR_Opr val_op = val.result();
 807 
 808   if (type == objectType) {
 809     val_op = shenandoah_read_barrier(val_op, NULL, true);
 810     __ cas_obj(addr, cmp.result(), val_op, new_register(T_OBJECT), new_register(T_OBJECT));
 811   }
 812   else if (type == intType)
 813     __ cas_int(addr, cmp.result(), val_op, ill, ill);
 814   else if (type == longType)
 815     __ cas_long(addr, cmp.result(), val_op, ill, ill);
 816   else {
 817     ShouldNotReachHere();
 818   }
 819 
 820   // generate conditional move of boolean result
 821   LIR_Opr result = rlock_result(x);
 822   __ cmove(lir_cond_equal, LIR_OprFact::intConst(1), LIR_OprFact::intConst(0),
 823            result, as_BasicType(type));
 824   if (type == objectType) {   // Write-barrier needed for Object fields.
 825     // Seems to be precise
 826     post_barrier(addr, val_op);
 827   }
 828 }
 829 
 830 
 831 void LIRGenerator::do_MathIntrinsic(Intrinsic* x) {
 832   assert(x->number_of_arguments() == 1 || (x->number_of_arguments() == 2 && x->id() == vmIntrinsics::_dpow), "wrong type");
 833   LIRItem value(x->argument_at(0), this);
 834 
 835   bool use_fpu = false;
 836   if (UseSSE >= 2) {
 837     switch(x->id()) {
 838       case vmIntrinsics::_dsin:
 839       case vmIntrinsics::_dcos:
 840       case vmIntrinsics::_dtan:
 841       case vmIntrinsics::_dlog:
 842       case vmIntrinsics::_dlog10:
 843       case vmIntrinsics::_dexp:
 844       case vmIntrinsics::_dpow:
 845         use_fpu = true;
 846     }
 847   } else {
 848     value.set_destroys_register();
 849   }
 850 
 851   value.load_item();
 852 
 853   LIR_Opr calc_input = value.result();
 854   LIR_Opr calc_input2 = NULL;
 855   if (x->id() == vmIntrinsics::_dpow) {
 856     LIRItem extra_arg(x->argument_at(1), this);
 857     if (UseSSE < 2) {
 858       extra_arg.set_destroys_register();
 859     }
 860     extra_arg.load_item();
 861     calc_input2 = extra_arg.result();
 862   }
 863   LIR_Opr calc_result = rlock_result(x);
 864 
 865   // sin, cos, pow and exp need two free fpu stack slots, so register
 866   // two temporary operands
 867   LIR_Opr tmp1 = FrameMap::caller_save_fpu_reg_at(0);
 868   LIR_Opr tmp2 = FrameMap::caller_save_fpu_reg_at(1);
 869 
 870   if (use_fpu) {
 871     LIR_Opr tmp = FrameMap::fpu0_double_opr;
 872     int tmp_start = 1;
 873     if (calc_input2 != NULL) {
 874       __ move(calc_input2, tmp);
 875       tmp_start = 2;
 876       calc_input2 = tmp;
 877     }
 878     __ move(calc_input, tmp);
 879 
 880     calc_input = tmp;
 881     calc_result = tmp;
 882 
 883     tmp1 = FrameMap::caller_save_fpu_reg_at(tmp_start);
 884     tmp2 = FrameMap::caller_save_fpu_reg_at(tmp_start + 1);
 885   }
 886 
 887   switch(x->id()) {
 888     case vmIntrinsics::_dabs:   __ abs  (calc_input, calc_result, LIR_OprFact::illegalOpr); break;
 889     case vmIntrinsics::_dsqrt:  __ sqrt (calc_input, calc_result, LIR_OprFact::illegalOpr); break;
 890     case vmIntrinsics::_dsin:   __ sin  (calc_input, calc_result, tmp1, tmp2);              break;
 891     case vmIntrinsics::_dcos:   __ cos  (calc_input, calc_result, tmp1, tmp2);              break;
 892     case vmIntrinsics::_dtan:   __ tan  (calc_input, calc_result, tmp1, tmp2);              break;
 893     case vmIntrinsics::_dlog:   __ log  (calc_input, calc_result, tmp1);                    break;
 894     case vmIntrinsics::_dlog10: __ log10(calc_input, calc_result, tmp1);                    break;
 895     case vmIntrinsics::_dexp:   __ exp  (calc_input, calc_result,              tmp1, tmp2, FrameMap::rax_opr, FrameMap::rcx_opr, FrameMap::rdx_opr); break;
 896     case vmIntrinsics::_dpow:   __ pow  (calc_input, calc_input2, calc_result, tmp1, tmp2, FrameMap::rax_opr, FrameMap::rcx_opr, FrameMap::rdx_opr); break;
 897     default:                    ShouldNotReachHere();
 898   }
 899 
 900   if (use_fpu) {
 901     __ move(calc_result, x->operand());
 902   }
 903 }
 904 
 905 
 906 void LIRGenerator::do_ArrayCopy(Intrinsic* x) {
 907   assert(x->number_of_arguments() == 5, "wrong type");
 908 
 909   // Make all state_for calls early since they can emit code
 910   CodeEmitInfo* info = state_for(x, x->state());
 911 
 912   LIRItem src(x->argument_at(0), this);
 913   LIRItem src_pos(x->argument_at(1), this);
 914   LIRItem dst(x->argument_at(2), this);
 915   LIRItem dst_pos(x->argument_at(3), this);
 916   LIRItem length(x->argument_at(4), this);
 917 
 918   LIR_Opr dst_op = dst.result();
 919   dst_op = shenandoah_write_barrier(dst_op, info, x->arg_needs_null_check(2));
 920   LIR_Opr src_op = src.result();
 921   src_op = shenandoah_read_barrier(src_op, info, x->arg_needs_null_check(0));
 922 
 923   // operands for arraycopy must use fixed registers, otherwise
 924   // LinearScan will fail allocation (because arraycopy always needs a
 925   // call)
 926 
 927 #ifndef _LP64
 928   src_op = force_opr_to(src_op, FrameMap::rcx_oop_opr);
 929   src_pos.load_item_force (FrameMap::rdx_opr);
 930   dst_op = force_opr_to(dst_op, FrameMap::rax_oop_opr);
 931   dst_pos.load_item_force (FrameMap::rbx_opr);
 932   length.load_item_force  (FrameMap::rdi_opr);
 933   LIR_Opr tmp =           (FrameMap::rsi_opr);
 934 #else
 935 
 936   // The java calling convention will give us enough registers
 937   // so that on the stub side the args will be perfect already.
 938   // On the other slow/special case side we call C and the arg
 939   // positions are not similar enough to pick one as the best.
 940   // Also because the java calling convention is a "shifted" version
 941   // of the C convention we can process the java args trivially into C
 942   // args without worry of overwriting during the xfer
 943 
 944   src_op = force_opr_to(src_op, FrameMap::as_oop_opr(j_rarg0));
 945   src_pos.load_item_force (FrameMap::as_opr(j_rarg1));
 946   dst_op = force_opr_to(dst_op, FrameMap::as_oop_opr(j_rarg2));
 947   dst_pos.load_item_force (FrameMap::as_opr(j_rarg3));
 948   length.load_item_force  (FrameMap::as_opr(j_rarg4));
 949 
 950   LIR_Opr tmp =           FrameMap::as_opr(j_rarg5);
 951 #endif // LP64
 952 
 953   set_no_result(x);
 954 
 955   int flags;
 956   ciArrayKlass* expected_type;
 957   arraycopy_helper(x, &flags, &expected_type);
 958 
 959   __ arraycopy(src_op, src_pos.result(), dst_op, dst_pos.result(), length.result(), tmp, expected_type, flags, info); // does add_safepoint
 960 }
 961 
 962 void LIRGenerator::do_update_CRC32(Intrinsic* x) {
 963   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
 964   // Make all state_for calls early since they can emit code
 965   LIR_Opr result = rlock_result(x);
 966   int flags = 0;
 967   switch (x->id()) {
 968     case vmIntrinsics::_updateCRC32: {
 969       LIRItem crc(x->argument_at(0), this);
 970       LIRItem val(x->argument_at(1), this);
 971       // val is destroyed by update_crc32
 972       val.set_destroys_register();
 973       crc.load_item();
 974       val.load_item();
 975       __ update_crc32(crc.result(), val.result(), result);
 976       break;
 977     }
 978     case vmIntrinsics::_updateBytesCRC32:
 979     case vmIntrinsics::_updateByteBufferCRC32: {
 980       bool is_updateBytes = (x->id() == vmIntrinsics::_updateBytesCRC32);
 981 
 982       LIRItem crc(x->argument_at(0), this);
 983       LIRItem buf(x->argument_at(1), this);
 984       LIRItem off(x->argument_at(2), this);
 985       LIRItem len(x->argument_at(3), this);
 986       buf.load_item();
 987       off.load_nonconstant();
 988 
 989       LIR_Opr index = off.result();
 990       int offset = is_updateBytes ? arrayOopDesc::base_offset_in_bytes(T_BYTE) : 0;
 991       if(off.result()->is_constant()) {
 992         index = LIR_OprFact::illegalOpr;
 993        offset += off.result()->as_jint();
 994       }
 995       LIR_Opr base_op = buf.result();
 996 
 997 #ifndef _LP64
 998       if (!is_updateBytes) { // long b raw address
 999          base_op = new_register(T_INT);
1000          __ convert(Bytecodes::_l2i, buf.result(), base_op);
1001       }
1002 #else
1003       if (index->is_valid()) {
1004         LIR_Opr tmp = new_register(T_LONG);
1005         __ convert(Bytecodes::_i2l, index, tmp);
1006         index = tmp;
1007       }
1008 #endif
1009 
1010       if (is_updateBytes) {
1011         base_op = shenandoah_read_barrier(base_op, NULL, false);
1012       }
1013 
1014       LIR_Address* a = new LIR_Address(base_op,
1015                                        index,
1016                                        LIR_Address::times_1,
1017                                        offset,
1018                                        T_BYTE);
1019       BasicTypeList signature(3);
1020       signature.append(T_INT);
1021       signature.append(T_ADDRESS);
1022       signature.append(T_INT);
1023       CallingConvention* cc = frame_map()->c_calling_convention(&signature);
1024       const LIR_Opr result_reg = result_register_for(x->type());
1025 
1026       LIR_Opr addr = new_pointer_register();
1027       __ leal(LIR_OprFact::address(a), addr);
1028 
1029       crc.load_item_force(cc->at(0));
1030       __ move(addr, cc->at(1));
1031       len.load_item_force(cc->at(2));
1032 
1033       __ call_runtime_leaf(StubRoutines::updateBytesCRC32(), getThreadTemp(), result_reg, cc->args());
1034       __ move(result_reg, result);
1035 
1036       break;
1037     }
1038     default: {
1039       ShouldNotReachHere();
1040     }
1041   }
1042 }
1043 
1044 // _i2l, _i2f, _i2d, _l2i, _l2f, _l2d, _f2i, _f2l, _f2d, _d2i, _d2l, _d2f
1045 // _i2b, _i2c, _i2s
1046 LIR_Opr fixed_register_for(BasicType type) {
1047   switch (type) {
1048     case T_FLOAT:  return FrameMap::fpu0_float_opr;
1049     case T_DOUBLE: return FrameMap::fpu0_double_opr;
1050     case T_INT:    return FrameMap::rax_opr;
1051     case T_LONG:   return FrameMap::long0_opr;
1052     default:       ShouldNotReachHere(); return LIR_OprFact::illegalOpr;
1053   }
1054 }
1055 
1056 void LIRGenerator::do_Convert(Convert* x) {
1057   // flags that vary for the different operations and different SSE-settings
1058   bool fixed_input, fixed_result, round_result, needs_stub;
1059 
1060   switch (x->op()) {
1061     case Bytecodes::_i2l: // fall through
1062     case Bytecodes::_l2i: // fall through
1063     case Bytecodes::_i2b: // fall through
1064     case Bytecodes::_i2c: // fall through
1065     case Bytecodes::_i2s: fixed_input = false;       fixed_result = false;       round_result = false;      needs_stub = false; break;
1066 
1067     case Bytecodes::_f2d: fixed_input = UseSSE == 1; fixed_result = false;       round_result = false;      needs_stub = false; break;
1068     case Bytecodes::_d2f: fixed_input = false;       fixed_result = UseSSE == 1; round_result = UseSSE < 1; needs_stub = false; break;
1069     case Bytecodes::_i2f: fixed_input = false;       fixed_result = false;       round_result = UseSSE < 1; needs_stub = false; break;
1070     case Bytecodes::_i2d: fixed_input = false;       fixed_result = false;       round_result = false;      needs_stub = false; break;
1071     case Bytecodes::_f2i: fixed_input = false;       fixed_result = false;       round_result = false;      needs_stub = true;  break;
1072     case Bytecodes::_d2i: fixed_input = false;       fixed_result = false;       round_result = false;      needs_stub = true;  break;
1073     case Bytecodes::_l2f: fixed_input = false;       fixed_result = UseSSE >= 1; round_result = UseSSE < 1; needs_stub = false; break;
1074     case Bytecodes::_l2d: fixed_input = false;       fixed_result = UseSSE >= 2; round_result = UseSSE < 2; needs_stub = false; break;
1075     case Bytecodes::_f2l: fixed_input = true;        fixed_result = true;        round_result = false;      needs_stub = false; break;
1076     case Bytecodes::_d2l: fixed_input = true;        fixed_result = true;        round_result = false;      needs_stub = false; break;
1077     default: ShouldNotReachHere();
1078   }
1079 
1080   LIRItem value(x->value(), this);
1081   value.load_item();
1082   LIR_Opr input = value.result();
1083   LIR_Opr result = rlock(x);
1084 
1085   // arguments of lir_convert
1086   LIR_Opr conv_input = input;
1087   LIR_Opr conv_result = result;
1088   ConversionStub* stub = NULL;
1089 
1090   if (fixed_input) {
1091     conv_input = fixed_register_for(input->type());
1092     __ move(input, conv_input);
1093   }
1094 
1095   assert(fixed_result == false || round_result == false, "cannot set both");
1096   if (fixed_result) {
1097     conv_result = fixed_register_for(result->type());
1098   } else if (round_result) {
1099     result = new_register(result->type());
1100     set_vreg_flag(result, must_start_in_memory);
1101   }
1102 
1103   if (needs_stub) {
1104     stub = new ConversionStub(x->op(), conv_input, conv_result);
1105   }
1106 
1107   __ convert(x->op(), conv_input, conv_result, stub);
1108 
1109   if (result != conv_result) {
1110     __ move(conv_result, result);
1111   }
1112 
1113   assert(result->is_virtual(), "result must be virtual register");
1114   set_result(x, result);
1115 }
1116 
1117 
1118 void LIRGenerator::do_NewInstance(NewInstance* x) {
1119   print_if_not_loaded(x);
1120 
1121   CodeEmitInfo* info = state_for(x, x->state());
1122   LIR_Opr reg = result_register_for(x->type());
1123   new_instance(reg, x->klass(), x->is_unresolved(),
1124                        FrameMap::rcx_oop_opr,
1125                        FrameMap::rdi_oop_opr,
1126                        FrameMap::rsi_oop_opr,
1127                        LIR_OprFact::illegalOpr,
1128                        FrameMap::rdx_metadata_opr, info);
1129   LIR_Opr result = rlock_result(x);
1130   __ move(reg, result);
1131 }
1132 
1133 
1134 void LIRGenerator::do_NewTypeArray(NewTypeArray* x) {
1135   CodeEmitInfo* info = state_for(x, x->state());
1136 
1137   LIRItem length(x->length(), this);
1138   length.load_item_force(FrameMap::rbx_opr);
1139 
1140   LIR_Opr reg = result_register_for(x->type());
1141   LIR_Opr tmp1 = FrameMap::rcx_oop_opr;
1142   LIR_Opr tmp2 = FrameMap::rsi_oop_opr;
1143   LIR_Opr tmp3 = FrameMap::rdi_oop_opr;
1144   LIR_Opr tmp4 = reg;
1145   LIR_Opr klass_reg = FrameMap::rdx_metadata_opr;
1146   LIR_Opr len = length.result();
1147   BasicType elem_type = x->elt_type();
1148 
1149   __ metadata2reg(ciTypeArrayKlass::make(elem_type)->constant_encoding(), klass_reg);
1150 
1151   CodeStub* slow_path = new NewTypeArrayStub(klass_reg, len, reg, info);
1152   __ allocate_array(reg, len, tmp1, tmp2, tmp3, tmp4, elem_type, klass_reg, slow_path);
1153 
1154   LIR_Opr result = rlock_result(x);
1155   __ move(reg, result);
1156 }
1157 
1158 
1159 void LIRGenerator::do_NewObjectArray(NewObjectArray* x) {
1160   LIRItem length(x->length(), this);
1161   // in case of patching (i.e., object class is not yet loaded), we need to reexecute the instruction
1162   // and therefore provide the state before the parameters have been consumed
1163   CodeEmitInfo* patching_info = NULL;
1164   if (!x->klass()->is_loaded() || PatchALot) {
1165     patching_info =  state_for(x, x->state_before());
1166   }
1167 
1168   CodeEmitInfo* info = state_for(x, x->state());
1169 
1170   const LIR_Opr reg = result_register_for(x->type());
1171   LIR_Opr tmp1 = FrameMap::rcx_oop_opr;
1172   LIR_Opr tmp2 = FrameMap::rsi_oop_opr;
1173   LIR_Opr tmp3 = FrameMap::rdi_oop_opr;
1174   LIR_Opr tmp4 = reg;
1175   LIR_Opr klass_reg = FrameMap::rdx_metadata_opr;
1176 
1177   length.load_item_force(FrameMap::rbx_opr);
1178   LIR_Opr len = length.result();
1179 
1180   CodeStub* slow_path = new NewObjectArrayStub(klass_reg, len, reg, info);
1181   ciKlass* obj = (ciKlass*) ciObjArrayKlass::make(x->klass());
1182   if (obj == ciEnv::unloaded_ciobjarrayklass()) {
1183     BAILOUT("encountered unloaded_ciobjarrayklass due to out of memory error");
1184   }
1185   klass2reg_with_patching(klass_reg, obj, patching_info);
1186   __ allocate_array(reg, len, tmp1, tmp2, tmp3, tmp4, T_OBJECT, klass_reg, slow_path);
1187 
1188   LIR_Opr result = rlock_result(x);
1189   __ move(reg, result);
1190 }
1191 
1192 
1193 void LIRGenerator::do_NewMultiArray(NewMultiArray* x) {
1194   Values* dims = x->dims();
1195   int i = dims->length();
1196   LIRItemList* items = new LIRItemList(dims->length(), NULL);
1197   while (i-- > 0) {
1198     LIRItem* size = new LIRItem(dims->at(i), this);
1199     items->at_put(i, size);
1200   }
1201 
1202   // Evaluate state_for early since it may emit code.
1203   CodeEmitInfo* patching_info = NULL;
1204   if (!x->klass()->is_loaded() || PatchALot) {
1205     patching_info = state_for(x, x->state_before());
1206 
1207     // Cannot re-use same xhandlers for multiple CodeEmitInfos, so
1208     // clone all handlers (NOTE: Usually this is handled transparently
1209     // by the CodeEmitInfo cloning logic in CodeStub constructors but
1210     // is done explicitly here because a stub isn't being used).
1211     x->set_exception_handlers(new XHandlers(x->exception_handlers()));
1212   }
1213   CodeEmitInfo* info = state_for(x, x->state());
1214 
1215   i = dims->length();
1216   while (i-- > 0) {
1217     LIRItem* size = items->at(i);
1218     size->load_nonconstant();
1219 
1220     store_stack_parameter(size->result(), in_ByteSize(i*4));
1221   }
1222 
1223   LIR_Opr klass_reg = FrameMap::rax_metadata_opr;
1224   klass2reg_with_patching(klass_reg, x->klass(), patching_info);
1225 
1226   LIR_Opr rank = FrameMap::rbx_opr;
1227   __ move(LIR_OprFact::intConst(x->rank()), rank);
1228   LIR_Opr varargs = FrameMap::rcx_opr;
1229   __ move(FrameMap::rsp_opr, varargs);
1230   LIR_OprList* args = new LIR_OprList(3);
1231   args->append(klass_reg);
1232   args->append(rank);
1233   args->append(varargs);
1234   LIR_Opr reg = result_register_for(x->type());
1235   __ call_runtime(Runtime1::entry_for(Runtime1::new_multi_array_id),
1236                   LIR_OprFact::illegalOpr,
1237                   reg, args, info);
1238 
1239   LIR_Opr result = rlock_result(x);
1240   __ move(reg, result);
1241 }
1242 
1243 
1244 void LIRGenerator::do_BlockBegin(BlockBegin* x) {
1245   // nothing to do for now
1246 }
1247 
1248 
1249 void LIRGenerator::do_CheckCast(CheckCast* x) {
1250   LIRItem obj(x->obj(), this);
1251 
1252   CodeEmitInfo* patching_info = NULL;
1253   if (!x->klass()->is_loaded() || (PatchALot && !x->is_incompatible_class_change_check())) {
1254     // must do this before locking the destination register as an oop register,
1255     // and before the obj is loaded (the latter is for deoptimization)
1256     patching_info = state_for(x, x->state_before());
1257   }
1258   obj.load_item();
1259 
1260   // info for exceptions
1261   CodeEmitInfo* info_for_exception = state_for(x);
1262 
1263   CodeStub* stub;
1264   if (x->is_incompatible_class_change_check()) {
1265     assert(patching_info == NULL, "can't patch this");
1266     stub = new SimpleExceptionStub(Runtime1::throw_incompatible_class_change_error_id, LIR_OprFact::illegalOpr, info_for_exception);
1267   } else {
1268     stub = new SimpleExceptionStub(Runtime1::throw_class_cast_exception_id, obj.result(), info_for_exception);
1269   }
1270   LIR_Opr reg = rlock_result(x);
1271   LIR_Opr tmp3 = LIR_OprFact::illegalOpr;
1272   if (!x->klass()->is_loaded() || UseCompressedClassPointers) {
1273     tmp3 = new_register(objectType);
1274   }
1275   __ checkcast(reg, obj.result(), x->klass(),
1276                new_register(objectType), new_register(objectType), tmp3,
1277                x->direct_compare(), info_for_exception, patching_info, stub,
1278                x->profiled_method(), x->profiled_bci());
1279 }
1280 
1281 
1282 void LIRGenerator::do_InstanceOf(InstanceOf* x) {
1283   LIRItem obj(x->obj(), this);
1284 
1285   // result and test object may not be in same register
1286   LIR_Opr reg = rlock_result(x);
1287   CodeEmitInfo* patching_info = NULL;
1288   if ((!x->klass()->is_loaded() || PatchALot)) {
1289     // must do this before locking the destination register as an oop register
1290     patching_info = state_for(x, x->state_before());
1291   }
1292   obj.load_item();
1293   LIR_Opr tmp3 = LIR_OprFact::illegalOpr;
1294   if (!x->klass()->is_loaded() || UseCompressedClassPointers) {
1295     tmp3 = new_register(objectType);
1296   }
1297   __ instanceof(reg, obj.result(), x->klass(),
1298                 new_register(objectType), new_register(objectType), tmp3,
1299                 x->direct_compare(), patching_info, x->profiled_method(), x->profiled_bci());
1300 }
1301 
1302 
1303 void LIRGenerator::do_If(If* x) {
1304   assert(x->number_of_sux() == 2, "inconsistency");
1305   ValueTag tag = x->x()->type()->tag();
1306   bool is_safepoint = x->is_safepoint();
1307 
1308   If::Condition cond = x->cond();
1309 
1310   LIRItem xitem(x->x(), this);
1311   LIRItem yitem(x->y(), this);
1312   LIRItem* xin = &xitem;
1313   LIRItem* yin = &yitem;
1314 
1315   if (tag == longTag) {
1316     // for longs, only conditions "eql", "neq", "lss", "geq" are valid;
1317     // mirror for other conditions
1318     if (cond == If::gtr || cond == If::leq) {
1319       cond = Instruction::mirror(cond);
1320       xin = &yitem;
1321       yin = &xitem;
1322     }
1323     xin->set_destroys_register();
1324   }
1325   xin->load_item();
1326   if (tag == longTag && yin->is_constant() && yin->get_jlong_constant() == 0 && (cond == If::eql || cond == If::neq)) {
1327     // inline long zero
1328     yin->dont_load_item();
1329   } else if (tag == longTag || tag == floatTag || tag == doubleTag) {
1330     // longs cannot handle constants at right side
1331     yin->load_item();
1332   } else {
1333     yin->dont_load_item();
1334   }
1335 
1336   // add safepoint before generating condition code so it can be recomputed
1337   if (x->is_safepoint()) {
1338     // increment backedge counter if needed
1339     increment_backedge_counter(state_for(x, x->state_before()), x->profiled_bci());
1340     __ safepoint(LIR_OprFact::illegalOpr, state_for(x, x->state_before()));
1341   }
1342   set_no_result(x);
1343 
1344   LIR_Opr left = xin->result();
1345   LIR_Opr right = yin->result();
1346   if (tag == objectTag && UseShenandoahGC && x->y()->type() != objectNull) { // Don't need to resolve for ifnull.
1347     left = shenandoah_write_barrier(left, NULL, true);
1348     right = shenandoah_read_barrier(right, NULL, true);
1349   }
1350   __ cmp(lir_cond(cond), left, right);
1351   // Generate branch profiling. Profiling code doesn't kill flags.
1352   profile_branch(x, cond);
1353   move_to_phi(x->state());
1354   if (x->x()->type()->is_float_kind()) {
1355     __ branch(lir_cond(cond), right->type(), x->tsux(), x->usux());
1356   } else {
1357     __ branch(lir_cond(cond), right->type(), x->tsux());
1358   }
1359   assert(x->default_sux() == x->fsux(), "wrong destination above");
1360   __ jump(x->default_sux());
1361 }
1362 
1363 
1364 LIR_Opr LIRGenerator::getThreadPointer() {
1365 #ifdef _LP64
1366   return FrameMap::as_pointer_opr(r15_thread);
1367 #else
1368   LIR_Opr result = new_register(T_INT);
1369   __ get_thread(result);
1370   return result;
1371 #endif //
1372 }
1373 
1374 void LIRGenerator::trace_block_entry(BlockBegin* block) {
1375   store_stack_parameter(LIR_OprFact::intConst(block->block_id()), in_ByteSize(0));
1376   LIR_OprList* args = new LIR_OprList();
1377   address func = CAST_FROM_FN_PTR(address, Runtime1::trace_block_entry);
1378   __ call_runtime_leaf(func, LIR_OprFact::illegalOpr, LIR_OprFact::illegalOpr, args);
1379 }
1380 
1381 
1382 void LIRGenerator::volatile_field_store(LIR_Opr value, LIR_Address* address,
1383                                         CodeEmitInfo* info) {
1384   if (address->type() == T_LONG) {
1385     address = new LIR_Address(address->base(),
1386                               address->index(), address->scale(),
1387                               address->disp(), T_DOUBLE);
1388     // Transfer the value atomically by using FP moves.  This means
1389     // the value has to be moved between CPU and FPU registers.  It
1390     // always has to be moved through spill slot since there's no
1391     // quick way to pack the value into an SSE register.
1392     LIR_Opr temp_double = new_register(T_DOUBLE);
1393     LIR_Opr spill = new_register(T_LONG);
1394     set_vreg_flag(spill, must_start_in_memory);
1395     __ move(value, spill);
1396     __ volatile_move(spill, temp_double, T_LONG);
1397     __ volatile_move(temp_double, LIR_OprFact::address(address), T_LONG, info);
1398   } else {
1399     __ store(value, address, info);
1400   }
1401 }
1402 
1403 
1404 
1405 void LIRGenerator::volatile_field_load(LIR_Address* address, LIR_Opr result,
1406                                        CodeEmitInfo* info) {
1407   if (address->type() == T_LONG) {
1408     address = new LIR_Address(address->base(),
1409                               address->index(), address->scale(),
1410                               address->disp(), T_DOUBLE);
1411     // Transfer the value atomically by using FP moves.  This means
1412     // the value has to be moved between CPU and FPU registers.  In
1413     // SSE0 and SSE1 mode it has to be moved through spill slot but in
1414     // SSE2+ mode it can be moved directly.
1415     LIR_Opr temp_double = new_register(T_DOUBLE);
1416     __ volatile_move(LIR_OprFact::address(address), temp_double, T_LONG, info);
1417     __ volatile_move(temp_double, result, T_LONG);
1418     if (UseSSE < 2) {
1419       // no spill slot needed in SSE2 mode because xmm->cpu register move is possible
1420       set_vreg_flag(result, must_start_in_memory);
1421     }
1422   } else {
1423     __ load(address, result, info);
1424   }
1425 }
1426 
1427 void LIRGenerator::get_Object_unsafe(LIR_Opr dst, LIR_Opr src, LIR_Opr offset,
1428                                      BasicType type, bool is_volatile) {
1429   src = shenandoah_read_barrier(src, NULL, false);
1430   if (is_volatile && type == T_LONG) {
1431     LIR_Address* addr = new LIR_Address(src, offset, T_DOUBLE);
1432     LIR_Opr tmp = new_register(T_DOUBLE);
1433     __ load(addr, tmp);
1434     LIR_Opr spill = new_register(T_LONG);
1435     set_vreg_flag(spill, must_start_in_memory);
1436     __ move(tmp, spill);
1437     __ move(spill, dst);
1438   } else {
1439     LIR_Address* addr = new LIR_Address(src, offset, type);
1440     __ load(addr, dst);
1441   }
1442 }
1443 
1444 
1445 void LIRGenerator::put_Object_unsafe(LIR_Opr src, LIR_Opr offset, LIR_Opr data,
1446                                      BasicType type, bool is_volatile) {
1447   src = shenandoah_write_barrier(src, NULL, false);
1448   if (is_volatile && type == T_LONG) {
1449     LIR_Address* addr = new LIR_Address(src, offset, T_DOUBLE);
1450     LIR_Opr tmp = new_register(T_DOUBLE);
1451     LIR_Opr spill = new_register(T_DOUBLE);
1452     set_vreg_flag(spill, must_start_in_memory);
1453     __ move(data, spill);
1454     __ move(spill, tmp);
1455     __ move(tmp, addr);
1456   } else {
1457     LIR_Address* addr = new LIR_Address(src, offset, type);
1458     bool is_obj = (type == T_ARRAY || type == T_OBJECT);
1459     if (is_obj) {
1460       // Do the pre-write barrier, if any.
1461       pre_barrier(LIR_OprFact::address(addr), LIR_OprFact::illegalOpr /* pre_val */,
1462                   true /* do_load */, false /* patch */, NULL);
1463       data = shenandoah_read_barrier(data, NULL, true);
1464       __ move(data, addr);
1465       assert(src->is_register(), "must be register");
1466       // Seems to be a precise address
1467       post_barrier(LIR_OprFact::address(addr), data);
1468     } else {
1469       __ move(data, addr);
1470     }
1471   }
1472 }
1473 
1474 void LIRGenerator::do_UnsafeGetAndSetObject(UnsafeGetAndSetObject* x) {
1475   BasicType type = x->basic_type();
1476   LIRItem src(x->object(), this);
1477   LIRItem off(x->offset(), this);
1478   LIRItem value(x->value(), this);
1479 
1480   src.load_item();
1481   value.load_item();
1482   off.load_nonconstant();
1483 
1484   LIR_Opr dst = rlock_result(x, type);
1485   LIR_Opr data = value.result();
1486   bool is_obj = (type == T_ARRAY || type == T_OBJECT);
1487   LIR_Opr offset = off.result();
1488 
1489   assert (type == T_INT || (!x->is_add() && is_obj) LP64_ONLY( || type == T_LONG ), "unexpected type");
1490 
1491   LIR_Opr src_op = src.result();
1492   src_op = shenandoah_write_barrier(src_op, NULL, false);
1493   if (is_obj) {
1494     data = shenandoah_read_barrier(data, NULL, true);
1495   }
1496 
1497   LIR_Address* addr;
1498   if (offset->is_constant()) {
1499 #ifdef _LP64
1500     jlong c = offset->as_jlong();
1501     if ((jlong)((jint)c) == c) {
1502       addr = new LIR_Address(src_op, (jint)c, type);
1503     } else {
1504       LIR_Opr tmp = new_register(T_LONG);
1505       __ move(offset, tmp);
1506       addr = new LIR_Address(src_op, tmp, type);
1507     }
1508 #else
1509     addr = new LIR_Address(src_op, offset->as_jint(), type);
1510 #endif
1511   } else {
1512     addr = new LIR_Address(src_op, offset, type);
1513   }
1514 
1515   // Because we want a 2-arg form of xchg and xadd
1516   __ move(data, dst);
1517 
1518   if (x->is_add()) {
1519     __ xadd(LIR_OprFact::address(addr), dst, dst, LIR_OprFact::illegalOpr);
1520   } else {
1521     if (is_obj) {
1522       // Do the pre-write barrier, if any.
1523       pre_barrier(LIR_OprFact::address(addr), LIR_OprFact::illegalOpr /* pre_val */,
1524                   true /* do_load */, false /* patch */, NULL);
1525     }
1526     __ xchg(LIR_OprFact::address(addr), dst, dst, LIR_OprFact::illegalOpr);
1527     if (is_obj) {
1528       // Seems to be a precise address
1529       post_barrier(LIR_OprFact::address(addr), data);
1530     }
1531   }
1532 }