src/share/vm/opto/compile.cpp

Sat, 29 Sep 2012 06:40:00 -0400

author
coleenp
date
Sat, 29 Sep 2012 06:40:00 -0400
changeset 4142
d8ce2825b193
parent 4115
e626685e9f6c
child 4154
c3e799c37717
permissions
-rw-r--r--

8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
Summary: Capitalize these metadata types (and objArrayKlass)
Reviewed-by: stefank, twisti, kvn

     1 /*
     2  * Copyright (c) 1997, 2012, 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  */
    25 #include "precompiled.hpp"
    26 #include "asm/assembler.hpp"
    27 #include "classfile/systemDictionary.hpp"
    28 #include "code/exceptionHandlerTable.hpp"
    29 #include "code/nmethod.hpp"
    30 #include "compiler/compileLog.hpp"
    31 #include "compiler/oopMap.hpp"
    32 #include "opto/addnode.hpp"
    33 #include "opto/block.hpp"
    34 #include "opto/c2compiler.hpp"
    35 #include "opto/callGenerator.hpp"
    36 #include "opto/callnode.hpp"
    37 #include "opto/cfgnode.hpp"
    38 #include "opto/chaitin.hpp"
    39 #include "opto/compile.hpp"
    40 #include "opto/connode.hpp"
    41 #include "opto/divnode.hpp"
    42 #include "opto/escape.hpp"
    43 #include "opto/idealGraphPrinter.hpp"
    44 #include "opto/loopnode.hpp"
    45 #include "opto/machnode.hpp"
    46 #include "opto/macro.hpp"
    47 #include "opto/matcher.hpp"
    48 #include "opto/memnode.hpp"
    49 #include "opto/mulnode.hpp"
    50 #include "opto/node.hpp"
    51 #include "opto/opcodes.hpp"
    52 #include "opto/output.hpp"
    53 #include "opto/parse.hpp"
    54 #include "opto/phaseX.hpp"
    55 #include "opto/rootnode.hpp"
    56 #include "opto/runtime.hpp"
    57 #include "opto/stringopts.hpp"
    58 #include "opto/type.hpp"
    59 #include "opto/vectornode.hpp"
    60 #include "runtime/arguments.hpp"
    61 #include "runtime/signature.hpp"
    62 #include "runtime/stubRoutines.hpp"
    63 #include "runtime/timer.hpp"
    64 #include "utilities/copy.hpp"
    65 #ifdef TARGET_ARCH_MODEL_x86_32
    66 # include "adfiles/ad_x86_32.hpp"
    67 #endif
    68 #ifdef TARGET_ARCH_MODEL_x86_64
    69 # include "adfiles/ad_x86_64.hpp"
    70 #endif
    71 #ifdef TARGET_ARCH_MODEL_sparc
    72 # include "adfiles/ad_sparc.hpp"
    73 #endif
    74 #ifdef TARGET_ARCH_MODEL_zero
    75 # include "adfiles/ad_zero.hpp"
    76 #endif
    77 #ifdef TARGET_ARCH_MODEL_arm
    78 # include "adfiles/ad_arm.hpp"
    79 #endif
    80 #ifdef TARGET_ARCH_MODEL_ppc
    81 # include "adfiles/ad_ppc.hpp"
    82 #endif
    85 // -------------------- Compile::mach_constant_base_node -----------------------
    86 // Constant table base node singleton.
    87 MachConstantBaseNode* Compile::mach_constant_base_node() {
    88   if (_mach_constant_base_node == NULL) {
    89     _mach_constant_base_node = new (C) MachConstantBaseNode();
    90     _mach_constant_base_node->add_req(C->root());
    91   }
    92   return _mach_constant_base_node;
    93 }
    96 /// Support for intrinsics.
    98 // Return the index at which m must be inserted (or already exists).
    99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
   100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
   101 #ifdef ASSERT
   102   for (int i = 1; i < _intrinsics->length(); i++) {
   103     CallGenerator* cg1 = _intrinsics->at(i-1);
   104     CallGenerator* cg2 = _intrinsics->at(i);
   105     assert(cg1->method() != cg2->method()
   106            ? cg1->method()     < cg2->method()
   107            : cg1->is_virtual() < cg2->is_virtual(),
   108            "compiler intrinsics list must stay sorted");
   109   }
   110 #endif
   111   // Binary search sorted list, in decreasing intervals [lo, hi].
   112   int lo = 0, hi = _intrinsics->length()-1;
   113   while (lo <= hi) {
   114     int mid = (uint)(hi + lo) / 2;
   115     ciMethod* mid_m = _intrinsics->at(mid)->method();
   116     if (m < mid_m) {
   117       hi = mid-1;
   118     } else if (m > mid_m) {
   119       lo = mid+1;
   120     } else {
   121       // look at minor sort key
   122       bool mid_virt = _intrinsics->at(mid)->is_virtual();
   123       if (is_virtual < mid_virt) {
   124         hi = mid-1;
   125       } else if (is_virtual > mid_virt) {
   126         lo = mid+1;
   127       } else {
   128         return mid;  // exact match
   129       }
   130     }
   131   }
   132   return lo;  // inexact match
   133 }
   135 void Compile::register_intrinsic(CallGenerator* cg) {
   136   if (_intrinsics == NULL) {
   137     _intrinsics = new GrowableArray<CallGenerator*>(60);
   138   }
   139   // This code is stolen from ciObjectFactory::insert.
   140   // Really, GrowableArray should have methods for
   141   // insert_at, remove_at, and binary_search.
   142   int len = _intrinsics->length();
   143   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
   144   if (index == len) {
   145     _intrinsics->append(cg);
   146   } else {
   147 #ifdef ASSERT
   148     CallGenerator* oldcg = _intrinsics->at(index);
   149     assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
   150 #endif
   151     _intrinsics->append(_intrinsics->at(len-1));
   152     int pos;
   153     for (pos = len-2; pos >= index; pos--) {
   154       _intrinsics->at_put(pos+1,_intrinsics->at(pos));
   155     }
   156     _intrinsics->at_put(index, cg);
   157   }
   158   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
   159 }
   161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
   162   assert(m->is_loaded(), "don't try this on unloaded methods");
   163   if (_intrinsics != NULL) {
   164     int index = intrinsic_insertion_index(m, is_virtual);
   165     if (index < _intrinsics->length()
   166         && _intrinsics->at(index)->method() == m
   167         && _intrinsics->at(index)->is_virtual() == is_virtual) {
   168       return _intrinsics->at(index);
   169     }
   170   }
   171   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
   172   if (m->intrinsic_id() != vmIntrinsics::_none &&
   173       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
   174     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
   175     if (cg != NULL) {
   176       // Save it for next time:
   177       register_intrinsic(cg);
   178       return cg;
   179     } else {
   180       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
   181     }
   182   }
   183   return NULL;
   184 }
   186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
   187 // in library_call.cpp.
   190 #ifndef PRODUCT
   191 // statistics gathering...
   193 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
   194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
   196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
   197   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
   198   int oflags = _intrinsic_hist_flags[id];
   199   assert(flags != 0, "what happened?");
   200   if (is_virtual) {
   201     flags |= _intrinsic_virtual;
   202   }
   203   bool changed = (flags != oflags);
   204   if ((flags & _intrinsic_worked) != 0) {
   205     juint count = (_intrinsic_hist_count[id] += 1);
   206     if (count == 1) {
   207       changed = true;           // first time
   208     }
   209     // increment the overall count also:
   210     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
   211   }
   212   if (changed) {
   213     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
   214       // Something changed about the intrinsic's virtuality.
   215       if ((flags & _intrinsic_virtual) != 0) {
   216         // This is the first use of this intrinsic as a virtual call.
   217         if (oflags != 0) {
   218           // We already saw it as a non-virtual, so note both cases.
   219           flags |= _intrinsic_both;
   220         }
   221       } else if ((oflags & _intrinsic_both) == 0) {
   222         // This is the first use of this intrinsic as a non-virtual
   223         flags |= _intrinsic_both;
   224       }
   225     }
   226     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
   227   }
   228   // update the overall flags also:
   229   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
   230   return changed;
   231 }
   233 static char* format_flags(int flags, char* buf) {
   234   buf[0] = 0;
   235   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
   236   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
   237   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
   238   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
   239   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
   240   if (buf[0] == 0)  strcat(buf, ",");
   241   assert(buf[0] == ',', "must be");
   242   return &buf[1];
   243 }
   245 void Compile::print_intrinsic_statistics() {
   246   char flagsbuf[100];
   247   ttyLocker ttyl;
   248   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
   249   tty->print_cr("Compiler intrinsic usage:");
   250   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
   251   if (total == 0)  total = 1;  // avoid div0 in case of no successes
   252   #define PRINT_STAT_LINE(name, c, f) \
   253     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
   254   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
   255     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
   256     int   flags = _intrinsic_hist_flags[id];
   257     juint count = _intrinsic_hist_count[id];
   258     if ((flags | count) != 0) {
   259       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
   260     }
   261   }
   262   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
   263   if (xtty != NULL)  xtty->tail("statistics");
   264 }
   266 void Compile::print_statistics() {
   267   { ttyLocker ttyl;
   268     if (xtty != NULL)  xtty->head("statistics type='opto'");
   269     Parse::print_statistics();
   270     PhaseCCP::print_statistics();
   271     PhaseRegAlloc::print_statistics();
   272     Scheduling::print_statistics();
   273     PhasePeephole::print_statistics();
   274     PhaseIdealLoop::print_statistics();
   275     if (xtty != NULL)  xtty->tail("statistics");
   276   }
   277   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
   278     // put this under its own <statistics> element.
   279     print_intrinsic_statistics();
   280   }
   281 }
   282 #endif //PRODUCT
   284 // Support for bundling info
   285 Bundle* Compile::node_bundling(const Node *n) {
   286   assert(valid_bundle_info(n), "oob");
   287   return &_node_bundling_base[n->_idx];
   288 }
   290 bool Compile::valid_bundle_info(const Node *n) {
   291   return (_node_bundling_limit > n->_idx);
   292 }
   295 void Compile::gvn_replace_by(Node* n, Node* nn) {
   296   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
   297     Node* use = n->last_out(i);
   298     bool is_in_table = initial_gvn()->hash_delete(use);
   299     uint uses_found = 0;
   300     for (uint j = 0; j < use->len(); j++) {
   301       if (use->in(j) == n) {
   302         if (j < use->req())
   303           use->set_req(j, nn);
   304         else
   305           use->set_prec(j, nn);
   306         uses_found++;
   307       }
   308     }
   309     if (is_in_table) {
   310       // reinsert into table
   311       initial_gvn()->hash_find_insert(use);
   312     }
   313     record_for_igvn(use);
   314     i -= uses_found;    // we deleted 1 or more copies of this edge
   315   }
   316 }
   321 // Identify all nodes that are reachable from below, useful.
   322 // Use breadth-first pass that records state in a Unique_Node_List,
   323 // recursive traversal is slower.
   324 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
   325   int estimated_worklist_size = unique();
   326   useful.map( estimated_worklist_size, NULL );  // preallocate space
   328   // Initialize worklist
   329   if (root() != NULL)     { useful.push(root()); }
   330   // If 'top' is cached, declare it useful to preserve cached node
   331   if( cached_top_node() ) { useful.push(cached_top_node()); }
   333   // Push all useful nodes onto the list, breadthfirst
   334   for( uint next = 0; next < useful.size(); ++next ) {
   335     assert( next < unique(), "Unique useful nodes < total nodes");
   336     Node *n  = useful.at(next);
   337     uint max = n->len();
   338     for( uint i = 0; i < max; ++i ) {
   339       Node *m = n->in(i);
   340       if( m == NULL ) continue;
   341       useful.push(m);
   342     }
   343   }
   344 }
   346 // Disconnect all useless nodes by disconnecting those at the boundary.
   347 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
   348   uint next = 0;
   349   while (next < useful.size()) {
   350     Node *n = useful.at(next++);
   351     // Use raw traversal of out edges since this code removes out edges
   352     int max = n->outcnt();
   353     for (int j = 0; j < max; ++j) {
   354       Node* child = n->raw_out(j);
   355       if (! useful.member(child)) {
   356         assert(!child->is_top() || child != top(),
   357                "If top is cached in Compile object it is in useful list");
   358         // Only need to remove this out-edge to the useless node
   359         n->raw_del_out(j);
   360         --j;
   361         --max;
   362       }
   363     }
   364     if (n->outcnt() == 1 && n->has_special_unique_user()) {
   365       record_for_igvn(n->unique_out());
   366     }
   367   }
   368   // Remove useless macro and predicate opaq nodes
   369   for (int i = C->macro_count()-1; i >= 0; i--) {
   370     Node* n = C->macro_node(i);
   371     if (!useful.member(n)) {
   372       remove_macro_node(n);
   373     }
   374   }
   375   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
   376 }
   378 //------------------------------frame_size_in_words-----------------------------
   379 // frame_slots in units of words
   380 int Compile::frame_size_in_words() const {
   381   // shift is 0 in LP32 and 1 in LP64
   382   const int shift = (LogBytesPerWord - LogBytesPerInt);
   383   int words = _frame_slots >> shift;
   384   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
   385   return words;
   386 }
   388 // ============================================================================
   389 //------------------------------CompileWrapper---------------------------------
   390 class CompileWrapper : public StackObj {
   391   Compile *const _compile;
   392  public:
   393   CompileWrapper(Compile* compile);
   395   ~CompileWrapper();
   396 };
   398 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
   399   // the Compile* pointer is stored in the current ciEnv:
   400   ciEnv* env = compile->env();
   401   assert(env == ciEnv::current(), "must already be a ciEnv active");
   402   assert(env->compiler_data() == NULL, "compile already active?");
   403   env->set_compiler_data(compile);
   404   assert(compile == Compile::current(), "sanity");
   406   compile->set_type_dict(NULL);
   407   compile->set_type_hwm(NULL);
   408   compile->set_type_last_size(0);
   409   compile->set_last_tf(NULL, NULL);
   410   compile->set_indexSet_arena(NULL);
   411   compile->set_indexSet_free_block_list(NULL);
   412   compile->init_type_arena();
   413   Type::Initialize(compile);
   414   _compile->set_scratch_buffer_blob(NULL);
   415   _compile->begin_method();
   416 }
   417 CompileWrapper::~CompileWrapper() {
   418   _compile->end_method();
   419   if (_compile->scratch_buffer_blob() != NULL)
   420     BufferBlob::free(_compile->scratch_buffer_blob());
   421   _compile->env()->set_compiler_data(NULL);
   422 }
   425 //----------------------------print_compile_messages---------------------------
   426 void Compile::print_compile_messages() {
   427 #ifndef PRODUCT
   428   // Check if recompiling
   429   if (_subsume_loads == false && PrintOpto) {
   430     // Recompiling without allowing machine instructions to subsume loads
   431     tty->print_cr("*********************************************************");
   432     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
   433     tty->print_cr("*********************************************************");
   434   }
   435   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
   436     // Recompiling without escape analysis
   437     tty->print_cr("*********************************************************");
   438     tty->print_cr("** Bailout: Recompile without escape analysis          **");
   439     tty->print_cr("*********************************************************");
   440   }
   441   if (env()->break_at_compile()) {
   442     // Open the debugger when compiling this method.
   443     tty->print("### Breaking when compiling: ");
   444     method()->print_short_name();
   445     tty->cr();
   446     BREAKPOINT;
   447   }
   449   if( PrintOpto ) {
   450     if (is_osr_compilation()) {
   451       tty->print("[OSR]%3d", _compile_id);
   452     } else {
   453       tty->print("%3d", _compile_id);
   454     }
   455   }
   456 #endif
   457 }
   460 //-----------------------init_scratch_buffer_blob------------------------------
   461 // Construct a temporary BufferBlob and cache it for this compile.
   462 void Compile::init_scratch_buffer_blob(int const_size) {
   463   // If there is already a scratch buffer blob allocated and the
   464   // constant section is big enough, use it.  Otherwise free the
   465   // current and allocate a new one.
   466   BufferBlob* blob = scratch_buffer_blob();
   467   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
   468     // Use the current blob.
   469   } else {
   470     if (blob != NULL) {
   471       BufferBlob::free(blob);
   472     }
   474     ResourceMark rm;
   475     _scratch_const_size = const_size;
   476     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
   477     blob = BufferBlob::create("Compile::scratch_buffer", size);
   478     // Record the buffer blob for next time.
   479     set_scratch_buffer_blob(blob);
   480     // Have we run out of code space?
   481     if (scratch_buffer_blob() == NULL) {
   482       // Let CompilerBroker disable further compilations.
   483       record_failure("Not enough space for scratch buffer in CodeCache");
   484       return;
   485     }
   486   }
   488   // Initialize the relocation buffers
   489   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
   490   set_scratch_locs_memory(locs_buf);
   491 }
   494 //-----------------------scratch_emit_size-------------------------------------
   495 // Helper function that computes size by emitting code
   496 uint Compile::scratch_emit_size(const Node* n) {
   497   // Start scratch_emit_size section.
   498   set_in_scratch_emit_size(true);
   500   // Emit into a trash buffer and count bytes emitted.
   501   // This is a pretty expensive way to compute a size,
   502   // but it works well enough if seldom used.
   503   // All common fixed-size instructions are given a size
   504   // method by the AD file.
   505   // Note that the scratch buffer blob and locs memory are
   506   // allocated at the beginning of the compile task, and
   507   // may be shared by several calls to scratch_emit_size.
   508   // The allocation of the scratch buffer blob is particularly
   509   // expensive, since it has to grab the code cache lock.
   510   BufferBlob* blob = this->scratch_buffer_blob();
   511   assert(blob != NULL, "Initialize BufferBlob at start");
   512   assert(blob->size() > MAX_inst_size, "sanity");
   513   relocInfo* locs_buf = scratch_locs_memory();
   514   address blob_begin = blob->content_begin();
   515   address blob_end   = (address)locs_buf;
   516   assert(blob->content_contains(blob_end), "sanity");
   517   CodeBuffer buf(blob_begin, blob_end - blob_begin);
   518   buf.initialize_consts_size(_scratch_const_size);
   519   buf.initialize_stubs_size(MAX_stubs_size);
   520   assert(locs_buf != NULL, "sanity");
   521   int lsize = MAX_locs_size / 3;
   522   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
   523   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
   524   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
   526   // Do the emission.
   528   Label fakeL; // Fake label for branch instructions.
   529   Label*   saveL = NULL;
   530   uint save_bnum = 0;
   531   bool is_branch = n->is_MachBranch();
   532   if (is_branch) {
   533     MacroAssembler masm(&buf);
   534     masm.bind(fakeL);
   535     n->as_MachBranch()->save_label(&saveL, &save_bnum);
   536     n->as_MachBranch()->label_set(&fakeL, 0);
   537   }
   538   n->emit(buf, this->regalloc());
   539   if (is_branch) // Restore label.
   540     n->as_MachBranch()->label_set(saveL, save_bnum);
   542   // End scratch_emit_size section.
   543   set_in_scratch_emit_size(false);
   545   return buf.insts_size();
   546 }
   549 // ============================================================================
   550 //------------------------------Compile standard-------------------------------
   551 debug_only( int Compile::_debug_idx = 100000; )
   553 // Compile a method.  entry_bci is -1 for normal compilations and indicates
   554 // the continuation bci for on stack replacement.
   557 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
   558                 : Phase(Compiler),
   559                   _env(ci_env),
   560                   _log(ci_env->log()),
   561                   _compile_id(ci_env->compile_id()),
   562                   _save_argument_registers(false),
   563                   _stub_name(NULL),
   564                   _stub_function(NULL),
   565                   _stub_entry_point(NULL),
   566                   _method(target),
   567                   _entry_bci(osr_bci),
   568                   _initial_gvn(NULL),
   569                   _for_igvn(NULL),
   570                   _warm_calls(NULL),
   571                   _subsume_loads(subsume_loads),
   572                   _do_escape_analysis(do_escape_analysis),
   573                   _failure_reason(NULL),
   574                   _code_buffer("Compile::Fill_buffer"),
   575                   _orig_pc_slot(0),
   576                   _orig_pc_slot_offset_in_bytes(0),
   577                   _has_method_handle_invokes(false),
   578                   _mach_constant_base_node(NULL),
   579                   _node_bundling_limit(0),
   580                   _node_bundling_base(NULL),
   581                   _java_calls(0),
   582                   _inner_loops(0),
   583                   _scratch_const_size(-1),
   584                   _in_scratch_emit_size(false),
   585 #ifndef PRODUCT
   586                   _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
   587                   _printer(IdealGraphPrinter::printer()),
   588 #endif
   589                   _congraph(NULL) {
   590   C = this;
   592   CompileWrapper cw(this);
   593 #ifndef PRODUCT
   594   if (TimeCompiler2) {
   595     tty->print(" ");
   596     target->holder()->name()->print();
   597     tty->print(".");
   598     target->print_short_name();
   599     tty->print("  ");
   600   }
   601   TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
   602   TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
   603   bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
   604   if (!print_opto_assembly) {
   605     bool print_assembly = (PrintAssembly || _method->should_print_assembly());
   606     if (print_assembly && !Disassembler::can_decode()) {
   607       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
   608       print_opto_assembly = true;
   609     }
   610   }
   611   set_print_assembly(print_opto_assembly);
   612   set_parsed_irreducible_loop(false);
   613 #endif
   615   if (ProfileTraps) {
   616     // Make sure the method being compiled gets its own MDO,
   617     // so we can at least track the decompile_count().
   618     method()->ensure_method_data();
   619   }
   621   Init(::AliasLevel);
   624   print_compile_messages();
   626   if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
   627     _ilt = InlineTree::build_inline_tree_root();
   628   else
   629     _ilt = NULL;
   631   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
   632   assert(num_alias_types() >= AliasIdxRaw, "");
   634 #define MINIMUM_NODE_HASH  1023
   635   // Node list that Iterative GVN will start with
   636   Unique_Node_List for_igvn(comp_arena());
   637   set_for_igvn(&for_igvn);
   639   // GVN that will be run immediately on new nodes
   640   uint estimated_size = method()->code_size()*4+64;
   641   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
   642   PhaseGVN gvn(node_arena(), estimated_size);
   643   set_initial_gvn(&gvn);
   645   { // Scope for timing the parser
   646     TracePhase t3("parse", &_t_parser, true);
   648     // Put top into the hash table ASAP.
   649     initial_gvn()->transform_no_reclaim(top());
   651     // Set up tf(), start(), and find a CallGenerator.
   652     CallGenerator* cg = NULL;
   653     if (is_osr_compilation()) {
   654       const TypeTuple *domain = StartOSRNode::osr_domain();
   655       const TypeTuple *range = TypeTuple::make_range(method()->signature());
   656       init_tf(TypeFunc::make(domain, range));
   657       StartNode* s = new (this) StartOSRNode(root(), domain);
   658       initial_gvn()->set_type_bottom(s);
   659       init_start(s);
   660       cg = CallGenerator::for_osr(method(), entry_bci());
   661     } else {
   662       // Normal case.
   663       init_tf(TypeFunc::make(method()));
   664       StartNode* s = new (this) StartNode(root(), tf()->domain());
   665       initial_gvn()->set_type_bottom(s);
   666       init_start(s);
   667       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
   668         // With java.lang.ref.reference.get() we must go through the
   669         // intrinsic when G1 is enabled - even when get() is the root
   670         // method of the compile - so that, if necessary, the value in
   671         // the referent field of the reference object gets recorded by
   672         // the pre-barrier code.
   673         // Specifically, if G1 is enabled, the value in the referent
   674         // field is recorded by the G1 SATB pre barrier. This will
   675         // result in the referent being marked live and the reference
   676         // object removed from the list of discovered references during
   677         // reference processing.
   678         cg = find_intrinsic(method(), false);
   679       }
   680       if (cg == NULL) {
   681         float past_uses = method()->interpreter_invocation_count();
   682         float expected_uses = past_uses;
   683         cg = CallGenerator::for_inline(method(), expected_uses);
   684       }
   685     }
   686     if (failing())  return;
   687     if (cg == NULL) {
   688       record_method_not_compilable_all_tiers("cannot parse method");
   689       return;
   690     }
   691     JVMState* jvms = build_start_state(start(), tf());
   692     if ((jvms = cg->generate(jvms)) == NULL) {
   693       record_method_not_compilable("method parse failed");
   694       return;
   695     }
   696     GraphKit kit(jvms);
   698     if (!kit.stopped()) {
   699       // Accept return values, and transfer control we know not where.
   700       // This is done by a special, unique ReturnNode bound to root.
   701       return_values(kit.jvms());
   702     }
   704     if (kit.has_exceptions()) {
   705       // Any exceptions that escape from this call must be rethrown
   706       // to whatever caller is dynamically above us on the stack.
   707       // This is done by a special, unique RethrowNode bound to root.
   708       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
   709     }
   711     if (!failing() && has_stringbuilder()) {
   712       {
   713         // remove useless nodes to make the usage analysis simpler
   714         ResourceMark rm;
   715         PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
   716       }
   718       {
   719         ResourceMark rm;
   720         print_method("Before StringOpts", 3);
   721         PhaseStringOpts pso(initial_gvn(), &for_igvn);
   722         print_method("After StringOpts", 3);
   723       }
   725       // now inline anything that we skipped the first time around
   726       while (_late_inlines.length() > 0) {
   727         CallGenerator* cg = _late_inlines.pop();
   728         cg->do_late_inline();
   729         if (failing())  return;
   730       }
   731     }
   732     assert(_late_inlines.length() == 0, "should have been processed");
   734     print_method("Before RemoveUseless", 3);
   736     // Remove clutter produced by parsing.
   737     if (!failing()) {
   738       ResourceMark rm;
   739       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
   740     }
   741   }
   743   // Note:  Large methods are capped off in do_one_bytecode().
   744   if (failing())  return;
   746   // After parsing, node notes are no longer automagic.
   747   // They must be propagated by register_new_node_with_optimizer(),
   748   // clone(), or the like.
   749   set_default_node_notes(NULL);
   751   for (;;) {
   752     int successes = Inline_Warm();
   753     if (failing())  return;
   754     if (successes == 0)  break;
   755   }
   757   // Drain the list.
   758   Finish_Warm();
   759 #ifndef PRODUCT
   760   if (_printer) {
   761     _printer->print_inlining(this);
   762   }
   763 #endif
   765   if (failing())  return;
   766   NOT_PRODUCT( verify_graph_edges(); )
   768   // Now optimize
   769   Optimize();
   770   if (failing())  return;
   771   NOT_PRODUCT( verify_graph_edges(); )
   773 #ifndef PRODUCT
   774   if (PrintIdeal) {
   775     ttyLocker ttyl;  // keep the following output all in one block
   776     // This output goes directly to the tty, not the compiler log.
   777     // To enable tools to match it up with the compilation activity,
   778     // be sure to tag this tty output with the compile ID.
   779     if (xtty != NULL) {
   780       xtty->head("ideal compile_id='%d'%s", compile_id(),
   781                  is_osr_compilation()    ? " compile_kind='osr'" :
   782                  "");
   783     }
   784     root()->dump(9999);
   785     if (xtty != NULL) {
   786       xtty->tail("ideal");
   787     }
   788   }
   789 #endif
   791   // Now that we know the size of all the monitors we can add a fixed slot
   792   // for the original deopt pc.
   794   _orig_pc_slot =  fixed_slots();
   795   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
   796   set_fixed_slots(next_slot);
   798   // Now generate code
   799   Code_Gen();
   800   if (failing())  return;
   802   // Check if we want to skip execution of all compiled code.
   803   {
   804 #ifndef PRODUCT
   805     if (OptoNoExecute) {
   806       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
   807       return;
   808     }
   809     TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
   810 #endif
   812     if (is_osr_compilation()) {
   813       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
   814       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
   815     } else {
   816       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
   817       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
   818     }
   820     env()->register_method(_method, _entry_bci,
   821                            &_code_offsets,
   822                            _orig_pc_slot_offset_in_bytes,
   823                            code_buffer(),
   824                            frame_size_in_words(), _oop_map_set,
   825                            &_handler_table, &_inc_table,
   826                            compiler,
   827                            env()->comp_level(),
   828                            has_unsafe_access(),
   829                            SharedRuntime::is_wide_vector(max_vector_size())
   830                            );
   831   }
   832 }
   834 //------------------------------Compile----------------------------------------
   835 // Compile a runtime stub
   836 Compile::Compile( ciEnv* ci_env,
   837                   TypeFunc_generator generator,
   838                   address stub_function,
   839                   const char *stub_name,
   840                   int is_fancy_jump,
   841                   bool pass_tls,
   842                   bool save_arg_registers,
   843                   bool return_pc )
   844   : Phase(Compiler),
   845     _env(ci_env),
   846     _log(ci_env->log()),
   847     _compile_id(-1),
   848     _save_argument_registers(save_arg_registers),
   849     _method(NULL),
   850     _stub_name(stub_name),
   851     _stub_function(stub_function),
   852     _stub_entry_point(NULL),
   853     _entry_bci(InvocationEntryBci),
   854     _initial_gvn(NULL),
   855     _for_igvn(NULL),
   856     _warm_calls(NULL),
   857     _orig_pc_slot(0),
   858     _orig_pc_slot_offset_in_bytes(0),
   859     _subsume_loads(true),
   860     _do_escape_analysis(false),
   861     _failure_reason(NULL),
   862     _code_buffer("Compile::Fill_buffer"),
   863     _has_method_handle_invokes(false),
   864     _mach_constant_base_node(NULL),
   865     _node_bundling_limit(0),
   866     _node_bundling_base(NULL),
   867     _java_calls(0),
   868     _inner_loops(0),
   869 #ifndef PRODUCT
   870     _trace_opto_output(TraceOptoOutput),
   871     _printer(NULL),
   872 #endif
   873     _congraph(NULL) {
   874   C = this;
   876 #ifndef PRODUCT
   877   TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
   878   TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
   879   set_print_assembly(PrintFrameConverterAssembly);
   880   set_parsed_irreducible_loop(false);
   881 #endif
   882   CompileWrapper cw(this);
   883   Init(/*AliasLevel=*/ 0);
   884   init_tf((*generator)());
   886   {
   887     // The following is a dummy for the sake of GraphKit::gen_stub
   888     Unique_Node_List for_igvn(comp_arena());
   889     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
   890     PhaseGVN gvn(Thread::current()->resource_area(),255);
   891     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
   892     gvn.transform_no_reclaim(top());
   894     GraphKit kit;
   895     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
   896   }
   898   NOT_PRODUCT( verify_graph_edges(); )
   899   Code_Gen();
   900   if (failing())  return;
   903   // Entry point will be accessed using compile->stub_entry_point();
   904   if (code_buffer() == NULL) {
   905     Matcher::soft_match_failure();
   906   } else {
   907     if (PrintAssembly && (WizardMode || Verbose))
   908       tty->print_cr("### Stub::%s", stub_name);
   910     if (!failing()) {
   911       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
   913       // Make the NMethod
   914       // For now we mark the frame as never safe for profile stackwalking
   915       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
   916                                                       code_buffer(),
   917                                                       CodeOffsets::frame_never_safe,
   918                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
   919                                                       frame_size_in_words(),
   920                                                       _oop_map_set,
   921                                                       save_arg_registers);
   922       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
   924       _stub_entry_point = rs->entry_point();
   925     }
   926   }
   927 }
   929 //------------------------------Init-------------------------------------------
   930 // Prepare for a single compilation
   931 void Compile::Init(int aliaslevel) {
   932   _unique  = 0;
   933   _regalloc = NULL;
   935   _tf      = NULL;  // filled in later
   936   _top     = NULL;  // cached later
   937   _matcher = NULL;  // filled in later
   938   _cfg     = NULL;  // filled in later
   940   set_24_bit_selection_and_mode(Use24BitFP, false);
   942   _node_note_array = NULL;
   943   _default_node_notes = NULL;
   945   _immutable_memory = NULL; // filled in at first inquiry
   947   // Globally visible Nodes
   948   // First set TOP to NULL to give safe behavior during creation of RootNode
   949   set_cached_top_node(NULL);
   950   set_root(new (this) RootNode());
   951   // Now that you have a Root to point to, create the real TOP
   952   set_cached_top_node( new (this) ConNode(Type::TOP) );
   953   set_recent_alloc(NULL, NULL);
   955   // Create Debug Information Recorder to record scopes, oopmaps, etc.
   956   env()->set_oop_recorder(new OopRecorder(env()->arena()));
   957   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
   958   env()->set_dependencies(new Dependencies(env()));
   960   _fixed_slots = 0;
   961   set_has_split_ifs(false);
   962   set_has_loops(has_method() && method()->has_loops()); // first approximation
   963   set_has_stringbuilder(false);
   964   _trap_can_recompile = false;  // no traps emitted yet
   965   _major_progress = true; // start out assuming good things will happen
   966   set_has_unsafe_access(false);
   967   set_max_vector_size(0);
   968   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
   969   set_decompile_count(0);
   971   set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
   972   set_num_loop_opts(LoopOptsCount);
   973   set_do_inlining(Inline);
   974   set_max_inline_size(MaxInlineSize);
   975   set_freq_inline_size(FreqInlineSize);
   976   set_do_scheduling(OptoScheduling);
   977   set_do_count_invocations(false);
   978   set_do_method_data_update(false);
   980   if (debug_info()->recording_non_safepoints()) {
   981     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
   982                         (comp_arena(), 8, 0, NULL));
   983     set_default_node_notes(Node_Notes::make(this));
   984   }
   986   // // -- Initialize types before each compile --
   987   // // Update cached type information
   988   // if( _method && _method->constants() )
   989   //   Type::update_loaded_types(_method, _method->constants());
   991   // Init alias_type map.
   992   if (!_do_escape_analysis && aliaslevel == 3)
   993     aliaslevel = 2;  // No unique types without escape analysis
   994   _AliasLevel = aliaslevel;
   995   const int grow_ats = 16;
   996   _max_alias_types = grow_ats;
   997   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
   998   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
   999   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
  1001     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
  1003   // Initialize the first few types.
  1004   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
  1005   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
  1006   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
  1007   _num_alias_types = AliasIdxRaw+1;
  1008   // Zero out the alias type cache.
  1009   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
  1010   // A NULL adr_type hits in the cache right away.  Preload the right answer.
  1011   probe_alias_cache(NULL)->_index = AliasIdxTop;
  1013   _intrinsics = NULL;
  1014   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
  1015   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
  1016   register_library_intrinsics();
  1019 //---------------------------init_start----------------------------------------
  1020 // Install the StartNode on this compile object.
  1021 void Compile::init_start(StartNode* s) {
  1022   if (failing())
  1023     return; // already failing
  1024   assert(s == start(), "");
  1027 StartNode* Compile::start() const {
  1028   assert(!failing(), "");
  1029   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
  1030     Node* start = root()->fast_out(i);
  1031     if( start->is_Start() )
  1032       return start->as_Start();
  1034   ShouldNotReachHere();
  1035   return NULL;
  1038 //-------------------------------immutable_memory-------------------------------------
  1039 // Access immutable memory
  1040 Node* Compile::immutable_memory() {
  1041   if (_immutable_memory != NULL) {
  1042     return _immutable_memory;
  1044   StartNode* s = start();
  1045   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
  1046     Node *p = s->fast_out(i);
  1047     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
  1048       _immutable_memory = p;
  1049       return _immutable_memory;
  1052   ShouldNotReachHere();
  1053   return NULL;
  1056 //----------------------set_cached_top_node------------------------------------
  1057 // Install the cached top node, and make sure Node::is_top works correctly.
  1058 void Compile::set_cached_top_node(Node* tn) {
  1059   if (tn != NULL)  verify_top(tn);
  1060   Node* old_top = _top;
  1061   _top = tn;
  1062   // Calling Node::setup_is_top allows the nodes the chance to adjust
  1063   // their _out arrays.
  1064   if (_top != NULL)     _top->setup_is_top();
  1065   if (old_top != NULL)  old_top->setup_is_top();
  1066   assert(_top == NULL || top()->is_top(), "");
  1069 #ifndef PRODUCT
  1070 void Compile::verify_top(Node* tn) const {
  1071   if (tn != NULL) {
  1072     assert(tn->is_Con(), "top node must be a constant");
  1073     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
  1074     assert(tn->in(0) != NULL, "must have live top node");
  1077 #endif
  1080 ///-------------------Managing Per-Node Debug & Profile Info-------------------
  1082 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
  1083   guarantee(arr != NULL, "");
  1084   int num_blocks = arr->length();
  1085   if (grow_by < num_blocks)  grow_by = num_blocks;
  1086   int num_notes = grow_by * _node_notes_block_size;
  1087   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
  1088   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
  1089   while (num_notes > 0) {
  1090     arr->append(notes);
  1091     notes     += _node_notes_block_size;
  1092     num_notes -= _node_notes_block_size;
  1094   assert(num_notes == 0, "exact multiple, please");
  1097 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
  1098   if (source == NULL || dest == NULL)  return false;
  1100   if (dest->is_Con())
  1101     return false;               // Do not push debug info onto constants.
  1103 #ifdef ASSERT
  1104   // Leave a bread crumb trail pointing to the original node:
  1105   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
  1106     dest->set_debug_orig(source);
  1108 #endif
  1110   if (node_note_array() == NULL)
  1111     return false;               // Not collecting any notes now.
  1113   // This is a copy onto a pre-existing node, which may already have notes.
  1114   // If both nodes have notes, do not overwrite any pre-existing notes.
  1115   Node_Notes* source_notes = node_notes_at(source->_idx);
  1116   if (source_notes == NULL || source_notes->is_clear())  return false;
  1117   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
  1118   if (dest_notes == NULL || dest_notes->is_clear()) {
  1119     return set_node_notes_at(dest->_idx, source_notes);
  1122   Node_Notes merged_notes = (*source_notes);
  1123   // The order of operations here ensures that dest notes will win...
  1124   merged_notes.update_from(dest_notes);
  1125   return set_node_notes_at(dest->_idx, &merged_notes);
  1129 //--------------------------allow_range_check_smearing-------------------------
  1130 // Gating condition for coalescing similar range checks.
  1131 // Sometimes we try 'speculatively' replacing a series of a range checks by a
  1132 // single covering check that is at least as strong as any of them.
  1133 // If the optimization succeeds, the simplified (strengthened) range check
  1134 // will always succeed.  If it fails, we will deopt, and then give up
  1135 // on the optimization.
  1136 bool Compile::allow_range_check_smearing() const {
  1137   // If this method has already thrown a range-check,
  1138   // assume it was because we already tried range smearing
  1139   // and it failed.
  1140   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
  1141   return !already_trapped;
  1145 //------------------------------flatten_alias_type-----------------------------
  1146 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
  1147   int offset = tj->offset();
  1148   TypePtr::PTR ptr = tj->ptr();
  1150   // Known instance (scalarizable allocation) alias only with itself.
  1151   bool is_known_inst = tj->isa_oopptr() != NULL &&
  1152                        tj->is_oopptr()->is_known_instance();
  1154   // Process weird unsafe references.
  1155   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
  1156     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
  1157     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
  1158     tj = TypeOopPtr::BOTTOM;
  1159     ptr = tj->ptr();
  1160     offset = tj->offset();
  1163   // Array pointers need some flattening
  1164   const TypeAryPtr *ta = tj->isa_aryptr();
  1165   if( ta && is_known_inst ) {
  1166     if ( offset != Type::OffsetBot &&
  1167          offset > arrayOopDesc::length_offset_in_bytes() ) {
  1168       offset = Type::OffsetBot; // Flatten constant access into array body only
  1169       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
  1171   } else if( ta && _AliasLevel >= 2 ) {
  1172     // For arrays indexed by constant indices, we flatten the alias
  1173     // space to include all of the array body.  Only the header, klass
  1174     // and array length can be accessed un-aliased.
  1175     if( offset != Type::OffsetBot ) {
  1176       if( ta->const_oop() ) { // MethodData* or Method*
  1177         offset = Type::OffsetBot;   // Flatten constant access into array body
  1178         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
  1179       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
  1180         // range is OK as-is.
  1181         tj = ta = TypeAryPtr::RANGE;
  1182       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
  1183         tj = TypeInstPtr::KLASS; // all klass loads look alike
  1184         ta = TypeAryPtr::RANGE; // generic ignored junk
  1185         ptr = TypePtr::BotPTR;
  1186       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
  1187         tj = TypeInstPtr::MARK;
  1188         ta = TypeAryPtr::RANGE; // generic ignored junk
  1189         ptr = TypePtr::BotPTR;
  1190       } else {                  // Random constant offset into array body
  1191         offset = Type::OffsetBot;   // Flatten constant access into array body
  1192         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
  1195     // Arrays of fixed size alias with arrays of unknown size.
  1196     if (ta->size() != TypeInt::POS) {
  1197       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
  1198       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
  1200     // Arrays of known objects become arrays of unknown objects.
  1201     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
  1202       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
  1203       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
  1205     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
  1206       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
  1207       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
  1209     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
  1210     // cannot be distinguished by bytecode alone.
  1211     if (ta->elem() == TypeInt::BOOL) {
  1212       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
  1213       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
  1214       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
  1216     // During the 2nd round of IterGVN, NotNull castings are removed.
  1217     // Make sure the Bottom and NotNull variants alias the same.
  1218     // Also, make sure exact and non-exact variants alias the same.
  1219     if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
  1220       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
  1224   // Oop pointers need some flattening
  1225   const TypeInstPtr *to = tj->isa_instptr();
  1226   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
  1227     ciInstanceKlass *k = to->klass()->as_instance_klass();
  1228     if( ptr == TypePtr::Constant ) {
  1229       if (to->klass() != ciEnv::current()->Class_klass() ||
  1230           offset < k->size_helper() * wordSize) {
  1231         // No constant oop pointers (such as Strings); they alias with
  1232         // unknown strings.
  1233         assert(!is_known_inst, "not scalarizable allocation");
  1234         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
  1236     } else if( is_known_inst ) {
  1237       tj = to; // Keep NotNull and klass_is_exact for instance type
  1238     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
  1239       // During the 2nd round of IterGVN, NotNull castings are removed.
  1240       // Make sure the Bottom and NotNull variants alias the same.
  1241       // Also, make sure exact and non-exact variants alias the same.
  1242       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
  1244     // Canonicalize the holder of this field
  1245     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
  1246       // First handle header references such as a LoadKlassNode, even if the
  1247       // object's klass is unloaded at compile time (4965979).
  1248       if (!is_known_inst) { // Do it only for non-instance types
  1249         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
  1251     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
  1252       // Static fields are in the space above the normal instance
  1253       // fields in the java.lang.Class instance.
  1254       if (to->klass() != ciEnv::current()->Class_klass()) {
  1255         to = NULL;
  1256         tj = TypeOopPtr::BOTTOM;
  1257         offset = tj->offset();
  1259     } else {
  1260       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
  1261       if (!k->equals(canonical_holder) || tj->offset() != offset) {
  1262         if( is_known_inst ) {
  1263           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
  1264         } else {
  1265           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
  1271   // Klass pointers to object array klasses need some flattening
  1272   const TypeKlassPtr *tk = tj->isa_klassptr();
  1273   if( tk ) {
  1274     // If we are referencing a field within a Klass, we need
  1275     // to assume the worst case of an Object.  Both exact and
  1276     // inexact types must flatten to the same alias class so
  1277     // use NotNull as the PTR.
  1278     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
  1280       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
  1281                                    TypeKlassPtr::OBJECT->klass(),
  1282                                    offset);
  1285     ciKlass* klass = tk->klass();
  1286     if( klass->is_obj_array_klass() ) {
  1287       ciKlass* k = TypeAryPtr::OOPS->klass();
  1288       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
  1289         k = TypeInstPtr::BOTTOM->klass();
  1290       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
  1293     // Check for precise loads from the primary supertype array and force them
  1294     // to the supertype cache alias index.  Check for generic array loads from
  1295     // the primary supertype array and also force them to the supertype cache
  1296     // alias index.  Since the same load can reach both, we need to merge
  1297     // these 2 disparate memories into the same alias class.  Since the
  1298     // primary supertype array is read-only, there's no chance of confusion
  1299     // where we bypass an array load and an array store.
  1300     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
  1301     if (offset == Type::OffsetBot ||
  1302         (offset >= primary_supers_offset &&
  1303          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
  1304         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
  1305       offset = in_bytes(Klass::secondary_super_cache_offset());
  1306       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
  1310   // Flatten all Raw pointers together.
  1311   if (tj->base() == Type::RawPtr)
  1312     tj = TypeRawPtr::BOTTOM;
  1314   if (tj->base() == Type::AnyPtr)
  1315     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
  1317   // Flatten all to bottom for now
  1318   switch( _AliasLevel ) {
  1319   case 0:
  1320     tj = TypePtr::BOTTOM;
  1321     break;
  1322   case 1:                       // Flatten to: oop, static, field or array
  1323     switch (tj->base()) {
  1324     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
  1325     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
  1326     case Type::AryPtr:   // do not distinguish arrays at all
  1327     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
  1328     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
  1329     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
  1330     default: ShouldNotReachHere();
  1332     break;
  1333   case 2:                       // No collapsing at level 2; keep all splits
  1334   case 3:                       // No collapsing at level 3; keep all splits
  1335     break;
  1336   default:
  1337     Unimplemented();
  1340   offset = tj->offset();
  1341   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
  1343   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
  1344           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
  1345           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
  1346           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
  1347           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
  1348           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
  1349           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
  1350           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
  1351   assert( tj->ptr() != TypePtr::TopPTR &&
  1352           tj->ptr() != TypePtr::AnyNull &&
  1353           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
  1354 //    assert( tj->ptr() != TypePtr::Constant ||
  1355 //            tj->base() == Type::RawPtr ||
  1356 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
  1358   return tj;
  1361 void Compile::AliasType::Init(int i, const TypePtr* at) {
  1362   _index = i;
  1363   _adr_type = at;
  1364   _field = NULL;
  1365   _is_rewritable = true; // default
  1366   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
  1367   if (atoop != NULL && atoop->is_known_instance()) {
  1368     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
  1369     _general_index = Compile::current()->get_alias_index(gt);
  1370   } else {
  1371     _general_index = 0;
  1375 //---------------------------------print_on------------------------------------
  1376 #ifndef PRODUCT
  1377 void Compile::AliasType::print_on(outputStream* st) {
  1378   if (index() < 10)
  1379         st->print("@ <%d> ", index());
  1380   else  st->print("@ <%d>",  index());
  1381   st->print(is_rewritable() ? "   " : " RO");
  1382   int offset = adr_type()->offset();
  1383   if (offset == Type::OffsetBot)
  1384         st->print(" +any");
  1385   else  st->print(" +%-3d", offset);
  1386   st->print(" in ");
  1387   adr_type()->dump_on(st);
  1388   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
  1389   if (field() != NULL && tjp) {
  1390     if (tjp->klass()  != field()->holder() ||
  1391         tjp->offset() != field()->offset_in_bytes()) {
  1392       st->print(" != ");
  1393       field()->print();
  1394       st->print(" ***");
  1399 void print_alias_types() {
  1400   Compile* C = Compile::current();
  1401   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
  1402   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
  1403     C->alias_type(idx)->print_on(tty);
  1404     tty->cr();
  1407 #endif
  1410 //----------------------------probe_alias_cache--------------------------------
  1411 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
  1412   intptr_t key = (intptr_t) adr_type;
  1413   key ^= key >> logAliasCacheSize;
  1414   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
  1418 //-----------------------------grow_alias_types--------------------------------
  1419 void Compile::grow_alias_types() {
  1420   const int old_ats  = _max_alias_types; // how many before?
  1421   const int new_ats  = old_ats;          // how many more?
  1422   const int grow_ats = old_ats+new_ats;  // how many now?
  1423   _max_alias_types = grow_ats;
  1424   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
  1425   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
  1426   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
  1427   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
  1431 //--------------------------------find_alias_type------------------------------
  1432 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
  1433   if (_AliasLevel == 0)
  1434     return alias_type(AliasIdxBot);
  1436   AliasCacheEntry* ace = probe_alias_cache(adr_type);
  1437   if (ace->_adr_type == adr_type) {
  1438     return alias_type(ace->_index);
  1441   // Handle special cases.
  1442   if (adr_type == NULL)             return alias_type(AliasIdxTop);
  1443   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
  1445   // Do it the slow way.
  1446   const TypePtr* flat = flatten_alias_type(adr_type);
  1448 #ifdef ASSERT
  1449   assert(flat == flatten_alias_type(flat), "idempotent");
  1450   assert(flat != TypePtr::BOTTOM,     "cannot alias-analyze an untyped ptr");
  1451   if (flat->isa_oopptr() && !flat->isa_klassptr()) {
  1452     const TypeOopPtr* foop = flat->is_oopptr();
  1453     // Scalarizable allocations have exact klass always.
  1454     bool exact = !foop->klass_is_exact() || foop->is_known_instance();
  1455     const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
  1456     assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
  1458   assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
  1459 #endif
  1461   int idx = AliasIdxTop;
  1462   for (int i = 0; i < num_alias_types(); i++) {
  1463     if (alias_type(i)->adr_type() == flat) {
  1464       idx = i;
  1465       break;
  1469   if (idx == AliasIdxTop) {
  1470     if (no_create)  return NULL;
  1471     // Grow the array if necessary.
  1472     if (_num_alias_types == _max_alias_types)  grow_alias_types();
  1473     // Add a new alias type.
  1474     idx = _num_alias_types++;
  1475     _alias_types[idx]->Init(idx, flat);
  1476     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
  1477     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
  1478     if (flat->isa_instptr()) {
  1479       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
  1480           && flat->is_instptr()->klass() == env()->Class_klass())
  1481         alias_type(idx)->set_rewritable(false);
  1483     if (flat->isa_klassptr()) {
  1484       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
  1485         alias_type(idx)->set_rewritable(false);
  1486       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
  1487         alias_type(idx)->set_rewritable(false);
  1488       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
  1489         alias_type(idx)->set_rewritable(false);
  1490       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
  1491         alias_type(idx)->set_rewritable(false);
  1493     // %%% (We would like to finalize JavaThread::threadObj_offset(),
  1494     // but the base pointer type is not distinctive enough to identify
  1495     // references into JavaThread.)
  1497     // Check for final fields.
  1498     const TypeInstPtr* tinst = flat->isa_instptr();
  1499     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
  1500       ciField* field;
  1501       if (tinst->const_oop() != NULL &&
  1502           tinst->klass() == ciEnv::current()->Class_klass() &&
  1503           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
  1504         // static field
  1505         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
  1506         field = k->get_field_by_offset(tinst->offset(), true);
  1507       } else {
  1508         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
  1509         field = k->get_field_by_offset(tinst->offset(), false);
  1511       assert(field == NULL ||
  1512              original_field == NULL ||
  1513              (field->holder() == original_field->holder() &&
  1514               field->offset() == original_field->offset() &&
  1515               field->is_static() == original_field->is_static()), "wrong field?");
  1516       // Set field() and is_rewritable() attributes.
  1517       if (field != NULL)  alias_type(idx)->set_field(field);
  1521   // Fill the cache for next time.
  1522   ace->_adr_type = adr_type;
  1523   ace->_index    = idx;
  1524   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
  1526   // Might as well try to fill the cache for the flattened version, too.
  1527   AliasCacheEntry* face = probe_alias_cache(flat);
  1528   if (face->_adr_type == NULL) {
  1529     face->_adr_type = flat;
  1530     face->_index    = idx;
  1531     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
  1534   return alias_type(idx);
  1538 Compile::AliasType* Compile::alias_type(ciField* field) {
  1539   const TypeOopPtr* t;
  1540   if (field->is_static())
  1541     t = TypeInstPtr::make(field->holder()->java_mirror());
  1542   else
  1543     t = TypeOopPtr::make_from_klass_raw(field->holder());
  1544   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
  1545   assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
  1546   return atp;
  1550 //------------------------------have_alias_type--------------------------------
  1551 bool Compile::have_alias_type(const TypePtr* adr_type) {
  1552   AliasCacheEntry* ace = probe_alias_cache(adr_type);
  1553   if (ace->_adr_type == adr_type) {
  1554     return true;
  1557   // Handle special cases.
  1558   if (adr_type == NULL)             return true;
  1559   if (adr_type == TypePtr::BOTTOM)  return true;
  1561   return find_alias_type(adr_type, true, NULL) != NULL;
  1564 //-----------------------------must_alias--------------------------------------
  1565 // True if all values of the given address type are in the given alias category.
  1566 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
  1567   if (alias_idx == AliasIdxBot)         return true;  // the universal category
  1568   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
  1569   if (alias_idx == AliasIdxTop)         return false; // the empty category
  1570   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
  1572   // the only remaining possible overlap is identity
  1573   int adr_idx = get_alias_index(adr_type);
  1574   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
  1575   assert(adr_idx == alias_idx ||
  1576          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
  1577           && adr_type                       != TypeOopPtr::BOTTOM),
  1578          "should not be testing for overlap with an unsafe pointer");
  1579   return adr_idx == alias_idx;
  1582 //------------------------------can_alias--------------------------------------
  1583 // True if any values of the given address type are in the given alias category.
  1584 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
  1585   if (alias_idx == AliasIdxTop)         return false; // the empty category
  1586   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
  1587   if (alias_idx == AliasIdxBot)         return true;  // the universal category
  1588   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
  1590   // the only remaining possible overlap is identity
  1591   int adr_idx = get_alias_index(adr_type);
  1592   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
  1593   return adr_idx == alias_idx;
  1598 //---------------------------pop_warm_call-------------------------------------
  1599 WarmCallInfo* Compile::pop_warm_call() {
  1600   WarmCallInfo* wci = _warm_calls;
  1601   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
  1602   return wci;
  1605 //----------------------------Inline_Warm--------------------------------------
  1606 int Compile::Inline_Warm() {
  1607   // If there is room, try to inline some more warm call sites.
  1608   // %%% Do a graph index compaction pass when we think we're out of space?
  1609   if (!InlineWarmCalls)  return 0;
  1611   int calls_made_hot = 0;
  1612   int room_to_grow   = NodeCountInliningCutoff - unique();
  1613   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
  1614   int amount_grown   = 0;
  1615   WarmCallInfo* call;
  1616   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
  1617     int est_size = (int)call->size();
  1618     if (est_size > (room_to_grow - amount_grown)) {
  1619       // This one won't fit anyway.  Get rid of it.
  1620       call->make_cold();
  1621       continue;
  1623     call->make_hot();
  1624     calls_made_hot++;
  1625     amount_grown   += est_size;
  1626     amount_to_grow -= est_size;
  1629   if (calls_made_hot > 0)  set_major_progress();
  1630   return calls_made_hot;
  1634 //----------------------------Finish_Warm--------------------------------------
  1635 void Compile::Finish_Warm() {
  1636   if (!InlineWarmCalls)  return;
  1637   if (failing())  return;
  1638   if (warm_calls() == NULL)  return;
  1640   // Clean up loose ends, if we are out of space for inlining.
  1641   WarmCallInfo* call;
  1642   while ((call = pop_warm_call()) != NULL) {
  1643     call->make_cold();
  1647 //---------------------cleanup_loop_predicates-----------------------
  1648 // Remove the opaque nodes that protect the predicates so that all unused
  1649 // checks and uncommon_traps will be eliminated from the ideal graph
  1650 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
  1651   if (predicate_count()==0) return;
  1652   for (int i = predicate_count(); i > 0; i--) {
  1653     Node * n = predicate_opaque1_node(i-1);
  1654     assert(n->Opcode() == Op_Opaque1, "must be");
  1655     igvn.replace_node(n, n->in(1));
  1657   assert(predicate_count()==0, "should be clean!");
  1660 //------------------------------Optimize---------------------------------------
  1661 // Given a graph, optimize it.
  1662 void Compile::Optimize() {
  1663   TracePhase t1("optimizer", &_t_optimizer, true);
  1665 #ifndef PRODUCT
  1666   if (env()->break_at_compile()) {
  1667     BREAKPOINT;
  1670 #endif
  1672   ResourceMark rm;
  1673   int          loop_opts_cnt;
  1675   NOT_PRODUCT( verify_graph_edges(); )
  1677   print_method("After Parsing");
  1680   // Iterative Global Value Numbering, including ideal transforms
  1681   // Initialize IterGVN with types and values from parse-time GVN
  1682   PhaseIterGVN igvn(initial_gvn());
  1684     NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
  1685     igvn.optimize();
  1688   print_method("Iter GVN 1", 2);
  1690   if (failing())  return;
  1692   // Perform escape analysis
  1693   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
  1694     if (has_loops()) {
  1695       // Cleanup graph (remove dead nodes).
  1696       TracePhase t2("idealLoop", &_t_idealLoop, true);
  1697       PhaseIdealLoop ideal_loop( igvn, false, true );
  1698       if (major_progress()) print_method("PhaseIdealLoop before EA", 2);
  1699       if (failing())  return;
  1701     ConnectionGraph::do_analysis(this, &igvn);
  1703     if (failing())  return;
  1705     // Optimize out fields loads from scalar replaceable allocations.
  1706     igvn.optimize();
  1707     print_method("Iter GVN after EA", 2);
  1709     if (failing())  return;
  1711     if (congraph() != NULL && macro_count() > 0) {
  1712       NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
  1713       PhaseMacroExpand mexp(igvn);
  1714       mexp.eliminate_macro_nodes();
  1715       igvn.set_delay_transform(false);
  1717       igvn.optimize();
  1718       print_method("Iter GVN after eliminating allocations and locks", 2);
  1720       if (failing())  return;
  1724   // Loop transforms on the ideal graph.  Range Check Elimination,
  1725   // peeling, unrolling, etc.
  1727   // Set loop opts counter
  1728   loop_opts_cnt = num_loop_opts();
  1729   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
  1731       TracePhase t2("idealLoop", &_t_idealLoop, true);
  1732       PhaseIdealLoop ideal_loop( igvn, true );
  1733       loop_opts_cnt--;
  1734       if (major_progress()) print_method("PhaseIdealLoop 1", 2);
  1735       if (failing())  return;
  1737     // Loop opts pass if partial peeling occurred in previous pass
  1738     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
  1739       TracePhase t3("idealLoop", &_t_idealLoop, true);
  1740       PhaseIdealLoop ideal_loop( igvn, false );
  1741       loop_opts_cnt--;
  1742       if (major_progress()) print_method("PhaseIdealLoop 2", 2);
  1743       if (failing())  return;
  1745     // Loop opts pass for loop-unrolling before CCP
  1746     if(major_progress() && (loop_opts_cnt > 0)) {
  1747       TracePhase t4("idealLoop", &_t_idealLoop, true);
  1748       PhaseIdealLoop ideal_loop( igvn, false );
  1749       loop_opts_cnt--;
  1750       if (major_progress()) print_method("PhaseIdealLoop 3", 2);
  1752     if (!failing()) {
  1753       // Verify that last round of loop opts produced a valid graph
  1754       NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
  1755       PhaseIdealLoop::verify(igvn);
  1758   if (failing())  return;
  1760   // Conditional Constant Propagation;
  1761   PhaseCCP ccp( &igvn );
  1762   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
  1764     TracePhase t2("ccp", &_t_ccp, true);
  1765     ccp.do_transform();
  1767   print_method("PhaseCPP 1", 2);
  1769   assert( true, "Break here to ccp.dump_old2new_map()");
  1771   // Iterative Global Value Numbering, including ideal transforms
  1773     NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
  1774     igvn = ccp;
  1775     igvn.optimize();
  1778   print_method("Iter GVN 2", 2);
  1780   if (failing())  return;
  1782   // Loop transforms on the ideal graph.  Range Check Elimination,
  1783   // peeling, unrolling, etc.
  1784   if(loop_opts_cnt > 0) {
  1785     debug_only( int cnt = 0; );
  1786     while(major_progress() && (loop_opts_cnt > 0)) {
  1787       TracePhase t2("idealLoop", &_t_idealLoop, true);
  1788       assert( cnt++ < 40, "infinite cycle in loop optimization" );
  1789       PhaseIdealLoop ideal_loop( igvn, true);
  1790       loop_opts_cnt--;
  1791       if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
  1792       if (failing())  return;
  1797     // Verify that all previous optimizations produced a valid graph
  1798     // at least to this point, even if no loop optimizations were done.
  1799     NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
  1800     PhaseIdealLoop::verify(igvn);
  1804     NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
  1805     PhaseMacroExpand  mex(igvn);
  1806     if (mex.expand_macro_nodes()) {
  1807       assert(failing(), "must bail out w/ explicit message");
  1808       return;
  1812  } // (End scope of igvn; run destructor if necessary for asserts.)
  1814   // A method with only infinite loops has no edges entering loops from root
  1816     NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
  1817     if (final_graph_reshaping()) {
  1818       assert(failing(), "must bail out w/ explicit message");
  1819       return;
  1823   print_method("Optimize finished", 2);
  1827 //------------------------------Code_Gen---------------------------------------
  1828 // Given a graph, generate code for it
  1829 void Compile::Code_Gen() {
  1830   if (failing())  return;
  1832   // Perform instruction selection.  You might think we could reclaim Matcher
  1833   // memory PDQ, but actually the Matcher is used in generating spill code.
  1834   // Internals of the Matcher (including some VectorSets) must remain live
  1835   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
  1836   // set a bit in reclaimed memory.
  1838   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
  1839   // nodes.  Mapping is only valid at the root of each matched subtree.
  1840   NOT_PRODUCT( verify_graph_edges(); )
  1842   Node_List proj_list;
  1843   Matcher m(proj_list);
  1844   _matcher = &m;
  1846     TracePhase t2("matcher", &_t_matcher, true);
  1847     m.match();
  1849   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
  1850   // nodes.  Mapping is only valid at the root of each matched subtree.
  1851   NOT_PRODUCT( verify_graph_edges(); )
  1853   // If you have too many nodes, or if matching has failed, bail out
  1854   check_node_count(0, "out of nodes matching instructions");
  1855   if (failing())  return;
  1857   // Build a proper-looking CFG
  1858   PhaseCFG cfg(node_arena(), root(), m);
  1859   _cfg = &cfg;
  1861     NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
  1862     cfg.Dominators();
  1863     if (failing())  return;
  1865     NOT_PRODUCT( verify_graph_edges(); )
  1867     cfg.Estimate_Block_Frequency();
  1868     cfg.GlobalCodeMotion(m,unique(),proj_list);
  1869     if (failing())  return;
  1871     print_method("Global code motion", 2);
  1873     NOT_PRODUCT( verify_graph_edges(); )
  1875     debug_only( cfg.verify(); )
  1877   NOT_PRODUCT( verify_graph_edges(); )
  1879   PhaseChaitin regalloc(unique(),cfg,m);
  1880   _regalloc = &regalloc;
  1882     TracePhase t2("regalloc", &_t_registerAllocation, true);
  1883     // Perform any platform dependent preallocation actions.  This is used,
  1884     // for example, to avoid taking an implicit null pointer exception
  1885     // using the frame pointer on win95.
  1886     _regalloc->pd_preallocate_hook();
  1888     // Perform register allocation.  After Chaitin, use-def chains are
  1889     // no longer accurate (at spill code) and so must be ignored.
  1890     // Node->LRG->reg mappings are still accurate.
  1891     _regalloc->Register_Allocate();
  1893     // Bail out if the allocator builds too many nodes
  1894     if (failing())  return;
  1897   // Prior to register allocation we kept empty basic blocks in case the
  1898   // the allocator needed a place to spill.  After register allocation we
  1899   // are not adding any new instructions.  If any basic block is empty, we
  1900   // can now safely remove it.
  1902     NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
  1903     cfg.remove_empty();
  1904     if (do_freq_based_layout()) {
  1905       PhaseBlockLayout layout(cfg);
  1906     } else {
  1907       cfg.set_loop_alignment();
  1909     cfg.fixup_flow();
  1912   // Perform any platform dependent postallocation verifications.
  1913   debug_only( _regalloc->pd_postallocate_verify_hook(); )
  1915   // Apply peephole optimizations
  1916   if( OptoPeephole ) {
  1917     NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
  1918     PhasePeephole peep( _regalloc, cfg);
  1919     peep.do_transform();
  1922   // Convert Nodes to instruction bits in a buffer
  1924     // %%%% workspace merge brought two timers together for one job
  1925     TracePhase t2a("output", &_t_output, true);
  1926     NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
  1927     Output();
  1930   print_method("Final Code");
  1932   // He's dead, Jim.
  1933   _cfg     = (PhaseCFG*)0xdeadbeef;
  1934   _regalloc = (PhaseChaitin*)0xdeadbeef;
  1938 //------------------------------dump_asm---------------------------------------
  1939 // Dump formatted assembly
  1940 #ifndef PRODUCT
  1941 void Compile::dump_asm(int *pcs, uint pc_limit) {
  1942   bool cut_short = false;
  1943   tty->print_cr("#");
  1944   tty->print("#  ");  _tf->dump();  tty->cr();
  1945   tty->print_cr("#");
  1947   // For all blocks
  1948   int pc = 0x0;                 // Program counter
  1949   char starts_bundle = ' ';
  1950   _regalloc->dump_frame();
  1952   Node *n = NULL;
  1953   for( uint i=0; i<_cfg->_num_blocks; i++ ) {
  1954     if (VMThread::should_terminate()) { cut_short = true; break; }
  1955     Block *b = _cfg->_blocks[i];
  1956     if (b->is_connector() && !Verbose) continue;
  1957     n = b->_nodes[0];
  1958     if (pcs && n->_idx < pc_limit)
  1959       tty->print("%3.3x   ", pcs[n->_idx]);
  1960     else
  1961       tty->print("      ");
  1962     b->dump_head( &_cfg->_bbs );
  1963     if (b->is_connector()) {
  1964       tty->print_cr("        # Empty connector block");
  1965     } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
  1966       tty->print_cr("        # Block is sole successor of call");
  1969     // For all instructions
  1970     Node *delay = NULL;
  1971     for( uint j = 0; j<b->_nodes.size(); j++ ) {
  1972       if (VMThread::should_terminate()) { cut_short = true; break; }
  1973       n = b->_nodes[j];
  1974       if (valid_bundle_info(n)) {
  1975         Bundle *bundle = node_bundling(n);
  1976         if (bundle->used_in_unconditional_delay()) {
  1977           delay = n;
  1978           continue;
  1980         if (bundle->starts_bundle())
  1981           starts_bundle = '+';
  1984       if (WizardMode) n->dump();
  1986       if( !n->is_Region() &&    // Dont print in the Assembly
  1987           !n->is_Phi() &&       // a few noisely useless nodes
  1988           !n->is_Proj() &&
  1989           !n->is_MachTemp() &&
  1990           !n->is_SafePointScalarObject() &&
  1991           !n->is_Catch() &&     // Would be nice to print exception table targets
  1992           !n->is_MergeMem() &&  // Not very interesting
  1993           !n->is_top() &&       // Debug info table constants
  1994           !(n->is_Con() && !n->is_Mach())// Debug info table constants
  1995           ) {
  1996         if (pcs && n->_idx < pc_limit)
  1997           tty->print("%3.3x", pcs[n->_idx]);
  1998         else
  1999           tty->print("   ");
  2000         tty->print(" %c ", starts_bundle);
  2001         starts_bundle = ' ';
  2002         tty->print("\t");
  2003         n->format(_regalloc, tty);
  2004         tty->cr();
  2007       // If we have an instruction with a delay slot, and have seen a delay,
  2008       // then back up and print it
  2009       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
  2010         assert(delay != NULL, "no unconditional delay instruction");
  2011         if (WizardMode) delay->dump();
  2013         if (node_bundling(delay)->starts_bundle())
  2014           starts_bundle = '+';
  2015         if (pcs && n->_idx < pc_limit)
  2016           tty->print("%3.3x", pcs[n->_idx]);
  2017         else
  2018           tty->print("   ");
  2019         tty->print(" %c ", starts_bundle);
  2020         starts_bundle = ' ';
  2021         tty->print("\t");
  2022         delay->format(_regalloc, tty);
  2023         tty->print_cr("");
  2024         delay = NULL;
  2027       // Dump the exception table as well
  2028       if( n->is_Catch() && (Verbose || WizardMode) ) {
  2029         // Print the exception table for this offset
  2030         _handler_table.print_subtable_for(pc);
  2034     if (pcs && n->_idx < pc_limit)
  2035       tty->print_cr("%3.3x", pcs[n->_idx]);
  2036     else
  2037       tty->print_cr("");
  2039     assert(cut_short || delay == NULL, "no unconditional delay branch");
  2041   } // End of per-block dump
  2042   tty->print_cr("");
  2044   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
  2046 #endif
  2048 //------------------------------Final_Reshape_Counts---------------------------
  2049 // This class defines counters to help identify when a method
  2050 // may/must be executed using hardware with only 24-bit precision.
  2051 struct Final_Reshape_Counts : public StackObj {
  2052   int  _call_count;             // count non-inlined 'common' calls
  2053   int  _float_count;            // count float ops requiring 24-bit precision
  2054   int  _double_count;           // count double ops requiring more precision
  2055   int  _java_call_count;        // count non-inlined 'java' calls
  2056   int  _inner_loop_count;       // count loops which need alignment
  2057   VectorSet _visited;           // Visitation flags
  2058   Node_List _tests;             // Set of IfNodes & PCTableNodes
  2060   Final_Reshape_Counts() :
  2061     _call_count(0), _float_count(0), _double_count(0),
  2062     _java_call_count(0), _inner_loop_count(0),
  2063     _visited( Thread::current()->resource_area() ) { }
  2065   void inc_call_count  () { _call_count  ++; }
  2066   void inc_float_count () { _float_count ++; }
  2067   void inc_double_count() { _double_count++; }
  2068   void inc_java_call_count() { _java_call_count++; }
  2069   void inc_inner_loop_count() { _inner_loop_count++; }
  2071   int  get_call_count  () const { return _call_count  ; }
  2072   int  get_float_count () const { return _float_count ; }
  2073   int  get_double_count() const { return _double_count; }
  2074   int  get_java_call_count() const { return _java_call_count; }
  2075   int  get_inner_loop_count() const { return _inner_loop_count; }
  2076 };
  2078 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
  2079   ciInstanceKlass *k = tp->klass()->as_instance_klass();
  2080   // Make sure the offset goes inside the instance layout.
  2081   return k->contains_field_offset(tp->offset());
  2082   // Note that OffsetBot and OffsetTop are very negative.
  2085 // Eliminate trivially redundant StoreCMs and accumulate their
  2086 // precedence edges.
  2087 static void eliminate_redundant_card_marks(Node* n) {
  2088   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
  2089   if (n->in(MemNode::Address)->outcnt() > 1) {
  2090     // There are multiple users of the same address so it might be
  2091     // possible to eliminate some of the StoreCMs
  2092     Node* mem = n->in(MemNode::Memory);
  2093     Node* adr = n->in(MemNode::Address);
  2094     Node* val = n->in(MemNode::ValueIn);
  2095     Node* prev = n;
  2096     bool done = false;
  2097     // Walk the chain of StoreCMs eliminating ones that match.  As
  2098     // long as it's a chain of single users then the optimization is
  2099     // safe.  Eliminating partially redundant StoreCMs would require
  2100     // cloning copies down the other paths.
  2101     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
  2102       if (adr == mem->in(MemNode::Address) &&
  2103           val == mem->in(MemNode::ValueIn)) {
  2104         // redundant StoreCM
  2105         if (mem->req() > MemNode::OopStore) {
  2106           // Hasn't been processed by this code yet.
  2107           n->add_prec(mem->in(MemNode::OopStore));
  2108         } else {
  2109           // Already converted to precedence edge
  2110           for (uint i = mem->req(); i < mem->len(); i++) {
  2111             // Accumulate any precedence edges
  2112             if (mem->in(i) != NULL) {
  2113               n->add_prec(mem->in(i));
  2116           // Everything above this point has been processed.
  2117           done = true;
  2119         // Eliminate the previous StoreCM
  2120         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
  2121         assert(mem->outcnt() == 0, "should be dead");
  2122         mem->disconnect_inputs(NULL);
  2123       } else {
  2124         prev = mem;
  2126       mem = prev->in(MemNode::Memory);
  2131 //------------------------------final_graph_reshaping_impl----------------------
  2132 // Implement items 1-5 from final_graph_reshaping below.
  2133 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
  2135   if ( n->outcnt() == 0 ) return; // dead node
  2136   uint nop = n->Opcode();
  2138   // Check for 2-input instruction with "last use" on right input.
  2139   // Swap to left input.  Implements item (2).
  2140   if( n->req() == 3 &&          // two-input instruction
  2141       n->in(1)->outcnt() > 1 && // left use is NOT a last use
  2142       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
  2143       n->in(2)->outcnt() == 1 &&// right use IS a last use
  2144       !n->in(2)->is_Con() ) {   // right use is not a constant
  2145     // Check for commutative opcode
  2146     switch( nop ) {
  2147     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
  2148     case Op_MaxI:  case Op_MinI:
  2149     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
  2150     case Op_AndL:  case Op_XorL:  case Op_OrL:
  2151     case Op_AndI:  case Op_XorI:  case Op_OrI: {
  2152       // Move "last use" input to left by swapping inputs
  2153       n->swap_edges(1, 2);
  2154       break;
  2156     default:
  2157       break;
  2161 #ifdef ASSERT
  2162   if( n->is_Mem() ) {
  2163     Compile* C = Compile::current();
  2164     int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
  2165     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
  2166             // oop will be recorded in oop map if load crosses safepoint
  2167             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
  2168                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
  2169             "raw memory operations should have control edge");
  2171 #endif
  2172   // Count FPU ops and common calls, implements item (3)
  2173   switch( nop ) {
  2174   // Count all float operations that may use FPU
  2175   case Op_AddF:
  2176   case Op_SubF:
  2177   case Op_MulF:
  2178   case Op_DivF:
  2179   case Op_NegF:
  2180   case Op_ModF:
  2181   case Op_ConvI2F:
  2182   case Op_ConF:
  2183   case Op_CmpF:
  2184   case Op_CmpF3:
  2185   // case Op_ConvL2F: // longs are split into 32-bit halves
  2186     frc.inc_float_count();
  2187     break;
  2189   case Op_ConvF2D:
  2190   case Op_ConvD2F:
  2191     frc.inc_float_count();
  2192     frc.inc_double_count();
  2193     break;
  2195   // Count all double operations that may use FPU
  2196   case Op_AddD:
  2197   case Op_SubD:
  2198   case Op_MulD:
  2199   case Op_DivD:
  2200   case Op_NegD:
  2201   case Op_ModD:
  2202   case Op_ConvI2D:
  2203   case Op_ConvD2I:
  2204   // case Op_ConvL2D: // handled by leaf call
  2205   // case Op_ConvD2L: // handled by leaf call
  2206   case Op_ConD:
  2207   case Op_CmpD:
  2208   case Op_CmpD3:
  2209     frc.inc_double_count();
  2210     break;
  2211   case Op_Opaque1:              // Remove Opaque Nodes before matching
  2212   case Op_Opaque2:              // Remove Opaque Nodes before matching
  2213     n->subsume_by(n->in(1));
  2214     break;
  2215   case Op_CallStaticJava:
  2216   case Op_CallJava:
  2217   case Op_CallDynamicJava:
  2218     frc.inc_java_call_count(); // Count java call site;
  2219   case Op_CallRuntime:
  2220   case Op_CallLeaf:
  2221   case Op_CallLeafNoFP: {
  2222     assert( n->is_Call(), "" );
  2223     CallNode *call = n->as_Call();
  2224     // Count call sites where the FP mode bit would have to be flipped.
  2225     // Do not count uncommon runtime calls:
  2226     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
  2227     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
  2228     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
  2229       frc.inc_call_count();   // Count the call site
  2230     } else {                  // See if uncommon argument is shared
  2231       Node *n = call->in(TypeFunc::Parms);
  2232       int nop = n->Opcode();
  2233       // Clone shared simple arguments to uncommon calls, item (1).
  2234       if( n->outcnt() > 1 &&
  2235           !n->is_Proj() &&
  2236           nop != Op_CreateEx &&
  2237           nop != Op_CheckCastPP &&
  2238           nop != Op_DecodeN &&
  2239           !n->is_Mem() ) {
  2240         Node *x = n->clone();
  2241         call->set_req( TypeFunc::Parms, x );
  2244     break;
  2247   case Op_StoreD:
  2248   case Op_LoadD:
  2249   case Op_LoadD_unaligned:
  2250     frc.inc_double_count();
  2251     goto handle_mem;
  2252   case Op_StoreF:
  2253   case Op_LoadF:
  2254     frc.inc_float_count();
  2255     goto handle_mem;
  2257   case Op_StoreCM:
  2259       // Convert OopStore dependence into precedence edge
  2260       Node* prec = n->in(MemNode::OopStore);
  2261       n->del_req(MemNode::OopStore);
  2262       n->add_prec(prec);
  2263       eliminate_redundant_card_marks(n);
  2266     // fall through
  2268   case Op_StoreB:
  2269   case Op_StoreC:
  2270   case Op_StorePConditional:
  2271   case Op_StoreI:
  2272   case Op_StoreL:
  2273   case Op_StoreIConditional:
  2274   case Op_StoreLConditional:
  2275   case Op_CompareAndSwapI:
  2276   case Op_CompareAndSwapL:
  2277   case Op_CompareAndSwapP:
  2278   case Op_CompareAndSwapN:
  2279   case Op_GetAndAddI:
  2280   case Op_GetAndAddL:
  2281   case Op_GetAndSetI:
  2282   case Op_GetAndSetL:
  2283   case Op_GetAndSetP:
  2284   case Op_GetAndSetN:
  2285   case Op_StoreP:
  2286   case Op_StoreN:
  2287   case Op_LoadB:
  2288   case Op_LoadUB:
  2289   case Op_LoadUS:
  2290   case Op_LoadI:
  2291   case Op_LoadUI2L:
  2292   case Op_LoadKlass:
  2293   case Op_LoadNKlass:
  2294   case Op_LoadL:
  2295   case Op_LoadL_unaligned:
  2296   case Op_LoadPLocked:
  2297   case Op_LoadP:
  2298   case Op_LoadN:
  2299   case Op_LoadRange:
  2300   case Op_LoadS: {
  2301   handle_mem:
  2302 #ifdef ASSERT
  2303     if( VerifyOptoOopOffsets ) {
  2304       assert( n->is_Mem(), "" );
  2305       MemNode *mem  = (MemNode*)n;
  2306       // Check to see if address types have grounded out somehow.
  2307       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
  2308       assert( !tp || oop_offset_is_sane(tp), "" );
  2310 #endif
  2311     break;
  2314   case Op_AddP: {               // Assert sane base pointers
  2315     Node *addp = n->in(AddPNode::Address);
  2316     assert( !addp->is_AddP() ||
  2317             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
  2318             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
  2319             "Base pointers must match" );
  2320 #ifdef _LP64
  2321     if (UseCompressedOops &&
  2322         addp->Opcode() == Op_ConP &&
  2323         addp == n->in(AddPNode::Base) &&
  2324         n->in(AddPNode::Offset)->is_Con()) {
  2325       // Use addressing with narrow klass to load with offset on x86.
  2326       // On sparc loading 32-bits constant and decoding it have less
  2327       // instructions (4) then load 64-bits constant (7).
  2328       // Do this transformation here since IGVN will convert ConN back to ConP.
  2329       const Type* t = addp->bottom_type();
  2330       if (t->isa_oopptr()) {
  2331         Node* nn = NULL;
  2333         // Look for existing ConN node of the same exact type.
  2334         Compile* C = Compile::current();
  2335         Node* r  = C->root();
  2336         uint cnt = r->outcnt();
  2337         for (uint i = 0; i < cnt; i++) {
  2338           Node* m = r->raw_out(i);
  2339           if (m!= NULL && m->Opcode() == Op_ConN &&
  2340               m->bottom_type()->make_ptr() == t) {
  2341             nn = m;
  2342             break;
  2345         if (nn != NULL) {
  2346           // Decode a narrow oop to match address
  2347           // [R12 + narrow_oop_reg<<3 + offset]
  2348           nn = new (C) DecodeNNode(nn, t);
  2349           n->set_req(AddPNode::Base, nn);
  2350           n->set_req(AddPNode::Address, nn);
  2351           if (addp->outcnt() == 0) {
  2352             addp->disconnect_inputs(NULL);
  2357 #endif
  2358     break;
  2361 #ifdef _LP64
  2362   case Op_CastPP:
  2363     if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
  2364       Compile* C = Compile::current();
  2365       Node* in1 = n->in(1);
  2366       const Type* t = n->bottom_type();
  2367       Node* new_in1 = in1->clone();
  2368       new_in1->as_DecodeN()->set_type(t);
  2370       if (!Matcher::narrow_oop_use_complex_address()) {
  2371         //
  2372         // x86, ARM and friends can handle 2 adds in addressing mode
  2373         // and Matcher can fold a DecodeN node into address by using
  2374         // a narrow oop directly and do implicit NULL check in address:
  2375         //
  2376         // [R12 + narrow_oop_reg<<3 + offset]
  2377         // NullCheck narrow_oop_reg
  2378         //
  2379         // On other platforms (Sparc) we have to keep new DecodeN node and
  2380         // use it to do implicit NULL check in address:
  2381         //
  2382         // decode_not_null narrow_oop_reg, base_reg
  2383         // [base_reg + offset]
  2384         // NullCheck base_reg
  2385         //
  2386         // Pin the new DecodeN node to non-null path on these platform (Sparc)
  2387         // to keep the information to which NULL check the new DecodeN node
  2388         // corresponds to use it as value in implicit_null_check().
  2389         //
  2390         new_in1->set_req(0, n->in(0));
  2393       n->subsume_by(new_in1);
  2394       if (in1->outcnt() == 0) {
  2395         in1->disconnect_inputs(NULL);
  2398     break;
  2400   case Op_CmpP:
  2401     // Do this transformation here to preserve CmpPNode::sub() and
  2402     // other TypePtr related Ideal optimizations (for example, ptr nullness).
  2403     if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
  2404       Node* in1 = n->in(1);
  2405       Node* in2 = n->in(2);
  2406       if (!in1->is_DecodeN()) {
  2407         in2 = in1;
  2408         in1 = n->in(2);
  2410       assert(in1->is_DecodeN(), "sanity");
  2412       Compile* C = Compile::current();
  2413       Node* new_in2 = NULL;
  2414       if (in2->is_DecodeN()) {
  2415         new_in2 = in2->in(1);
  2416       } else if (in2->Opcode() == Op_ConP) {
  2417         const Type* t = in2->bottom_type();
  2418         if (t == TypePtr::NULL_PTR) {
  2419           // Don't convert CmpP null check into CmpN if compressed
  2420           // oops implicit null check is not generated.
  2421           // This will allow to generate normal oop implicit null check.
  2422           if (Matcher::gen_narrow_oop_implicit_null_checks())
  2423             new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
  2424           //
  2425           // This transformation together with CastPP transformation above
  2426           // will generated code for implicit NULL checks for compressed oops.
  2427           //
  2428           // The original code after Optimize()
  2429           //
  2430           //    LoadN memory, narrow_oop_reg
  2431           //    decode narrow_oop_reg, base_reg
  2432           //    CmpP base_reg, NULL
  2433           //    CastPP base_reg // NotNull
  2434           //    Load [base_reg + offset], val_reg
  2435           //
  2436           // after these transformations will be
  2437           //
  2438           //    LoadN memory, narrow_oop_reg
  2439           //    CmpN narrow_oop_reg, NULL
  2440           //    decode_not_null narrow_oop_reg, base_reg
  2441           //    Load [base_reg + offset], val_reg
  2442           //
  2443           // and the uncommon path (== NULL) will use narrow_oop_reg directly
  2444           // since narrow oops can be used in debug info now (see the code in
  2445           // final_graph_reshaping_walk()).
  2446           //
  2447           // At the end the code will be matched to
  2448           // on x86:
  2449           //
  2450           //    Load_narrow_oop memory, narrow_oop_reg
  2451           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
  2452           //    NullCheck narrow_oop_reg
  2453           //
  2454           // and on sparc:
  2455           //
  2456           //    Load_narrow_oop memory, narrow_oop_reg
  2457           //    decode_not_null narrow_oop_reg, base_reg
  2458           //    Load [base_reg + offset], val_reg
  2459           //    NullCheck base_reg
  2460           //
  2461         } else if (t->isa_oopptr()) {
  2462           new_in2 = ConNode::make(C, t->make_narrowoop());
  2465       if (new_in2 != NULL) {
  2466         Node* cmpN = new (C) CmpNNode(in1->in(1), new_in2);
  2467         n->subsume_by( cmpN );
  2468         if (in1->outcnt() == 0) {
  2469           in1->disconnect_inputs(NULL);
  2471         if (in2->outcnt() == 0) {
  2472           in2->disconnect_inputs(NULL);
  2476     break;
  2478   case Op_DecodeN:
  2479     assert(!n->in(1)->is_EncodeP(), "should be optimized out");
  2480     // DecodeN could be pinned when it can't be fold into
  2481     // an address expression, see the code for Op_CastPP above.
  2482     assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
  2483     break;
  2485   case Op_EncodeP: {
  2486     Node* in1 = n->in(1);
  2487     if (in1->is_DecodeN()) {
  2488       n->subsume_by(in1->in(1));
  2489     } else if (in1->Opcode() == Op_ConP) {
  2490       Compile* C = Compile::current();
  2491       const Type* t = in1->bottom_type();
  2492       if (t == TypePtr::NULL_PTR) {
  2493         n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
  2494       } else if (t->isa_oopptr()) {
  2495         n->subsume_by(ConNode::make(C, t->make_narrowoop()));
  2498     if (in1->outcnt() == 0) {
  2499       in1->disconnect_inputs(NULL);
  2501     break;
  2504   case Op_Proj: {
  2505     if (OptimizeStringConcat) {
  2506       ProjNode* p = n->as_Proj();
  2507       if (p->_is_io_use) {
  2508         // Separate projections were used for the exception path which
  2509         // are normally removed by a late inline.  If it wasn't inlined
  2510         // then they will hang around and should just be replaced with
  2511         // the original one.
  2512         Node* proj = NULL;
  2513         // Replace with just one
  2514         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
  2515           Node *use = i.get();
  2516           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
  2517             proj = use;
  2518             break;
  2521         assert(proj != NULL, "must be found");
  2522         p->subsume_by(proj);
  2525     break;
  2528   case Op_Phi:
  2529     if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
  2530       // The EncodeP optimization may create Phi with the same edges
  2531       // for all paths. It is not handled well by Register Allocator.
  2532       Node* unique_in = n->in(1);
  2533       assert(unique_in != NULL, "");
  2534       uint cnt = n->req();
  2535       for (uint i = 2; i < cnt; i++) {
  2536         Node* m = n->in(i);
  2537         assert(m != NULL, "");
  2538         if (unique_in != m)
  2539           unique_in = NULL;
  2541       if (unique_in != NULL) {
  2542         n->subsume_by(unique_in);
  2545     break;
  2547 #endif
  2549   case Op_ModI:
  2550     if (UseDivMod) {
  2551       // Check if a%b and a/b both exist
  2552       Node* d = n->find_similar(Op_DivI);
  2553       if (d) {
  2554         // Replace them with a fused divmod if supported
  2555         Compile* C = Compile::current();
  2556         if (Matcher::has_match_rule(Op_DivModI)) {
  2557           DivModINode* divmod = DivModINode::make(C, n);
  2558           d->subsume_by(divmod->div_proj());
  2559           n->subsume_by(divmod->mod_proj());
  2560         } else {
  2561           // replace a%b with a-((a/b)*b)
  2562           Node* mult = new (C) MulINode(d, d->in(2));
  2563           Node* sub  = new (C) SubINode(d->in(1), mult);
  2564           n->subsume_by( sub );
  2568     break;
  2570   case Op_ModL:
  2571     if (UseDivMod) {
  2572       // Check if a%b and a/b both exist
  2573       Node* d = n->find_similar(Op_DivL);
  2574       if (d) {
  2575         // Replace them with a fused divmod if supported
  2576         Compile* C = Compile::current();
  2577         if (Matcher::has_match_rule(Op_DivModL)) {
  2578           DivModLNode* divmod = DivModLNode::make(C, n);
  2579           d->subsume_by(divmod->div_proj());
  2580           n->subsume_by(divmod->mod_proj());
  2581         } else {
  2582           // replace a%b with a-((a/b)*b)
  2583           Node* mult = new (C) MulLNode(d, d->in(2));
  2584           Node* sub  = new (C) SubLNode(d->in(1), mult);
  2585           n->subsume_by( sub );
  2589     break;
  2591   case Op_LoadVector:
  2592   case Op_StoreVector:
  2593     break;
  2595   case Op_PackB:
  2596   case Op_PackS:
  2597   case Op_PackI:
  2598   case Op_PackF:
  2599   case Op_PackL:
  2600   case Op_PackD:
  2601     if (n->req()-1 > 2) {
  2602       // Replace many operand PackNodes with a binary tree for matching
  2603       PackNode* p = (PackNode*) n;
  2604       Node* btp = p->binary_tree_pack(Compile::current(), 1, n->req());
  2605       n->subsume_by(btp);
  2607     break;
  2608   case Op_Loop:
  2609   case Op_CountedLoop:
  2610     if (n->as_Loop()->is_inner_loop()) {
  2611       frc.inc_inner_loop_count();
  2613     break;
  2614   case Op_LShiftI:
  2615   case Op_RShiftI:
  2616   case Op_URShiftI:
  2617   case Op_LShiftL:
  2618   case Op_RShiftL:
  2619   case Op_URShiftL:
  2620     if (Matcher::need_masked_shift_count) {
  2621       // The cpu's shift instructions don't restrict the count to the
  2622       // lower 5/6 bits. We need to do the masking ourselves.
  2623       Node* in2 = n->in(2);
  2624       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
  2625       const TypeInt* t = in2->find_int_type();
  2626       if (t != NULL && t->is_con()) {
  2627         juint shift = t->get_con();
  2628         if (shift > mask) { // Unsigned cmp
  2629           Compile* C = Compile::current();
  2630           n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask)));
  2632       } else {
  2633         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
  2634           Compile* C = Compile::current();
  2635           Node* shift = new (C) AndINode(in2, ConNode::make(C, TypeInt::make(mask)));
  2636           n->set_req(2, shift);
  2639       if (in2->outcnt() == 0) { // Remove dead node
  2640         in2->disconnect_inputs(NULL);
  2643     break;
  2644   default:
  2645     assert( !n->is_Call(), "" );
  2646     assert( !n->is_Mem(), "" );
  2647     break;
  2650   // Collect CFG split points
  2651   if (n->is_MultiBranch())
  2652     frc._tests.push(n);
  2655 //------------------------------final_graph_reshaping_walk---------------------
  2656 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
  2657 // requires that the walk visits a node's inputs before visiting the node.
  2658 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
  2659   ResourceArea *area = Thread::current()->resource_area();
  2660   Unique_Node_List sfpt(area);
  2662   frc._visited.set(root->_idx); // first, mark node as visited
  2663   uint cnt = root->req();
  2664   Node *n = root;
  2665   uint  i = 0;
  2666   while (true) {
  2667     if (i < cnt) {
  2668       // Place all non-visited non-null inputs onto stack
  2669       Node* m = n->in(i);
  2670       ++i;
  2671       if (m != NULL && !frc._visited.test_set(m->_idx)) {
  2672         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
  2673           sfpt.push(m);
  2674         cnt = m->req();
  2675         nstack.push(n, i); // put on stack parent and next input's index
  2676         n = m;
  2677         i = 0;
  2679     } else {
  2680       // Now do post-visit work
  2681       final_graph_reshaping_impl( n, frc );
  2682       if (nstack.is_empty())
  2683         break;             // finished
  2684       n = nstack.node();   // Get node from stack
  2685       cnt = n->req();
  2686       i = nstack.index();
  2687       nstack.pop();        // Shift to the next node on stack
  2691   // Skip next transformation if compressed oops are not used.
  2692   if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
  2693     return;
  2695   // Go over safepoints nodes to skip DecodeN nodes for debug edges.
  2696   // It could be done for an uncommon traps or any safepoints/calls
  2697   // if the DecodeN node is referenced only in a debug info.
  2698   while (sfpt.size() > 0) {
  2699     n = sfpt.pop();
  2700     JVMState *jvms = n->as_SafePoint()->jvms();
  2701     assert(jvms != NULL, "sanity");
  2702     int start = jvms->debug_start();
  2703     int end   = n->req();
  2704     bool is_uncommon = (n->is_CallStaticJava() &&
  2705                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
  2706     for (int j = start; j < end; j++) {
  2707       Node* in = n->in(j);
  2708       if (in->is_DecodeN()) {
  2709         bool safe_to_skip = true;
  2710         if (!is_uncommon ) {
  2711           // Is it safe to skip?
  2712           for (uint i = 0; i < in->outcnt(); i++) {
  2713             Node* u = in->raw_out(i);
  2714             if (!u->is_SafePoint() ||
  2715                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
  2716               safe_to_skip = false;
  2720         if (safe_to_skip) {
  2721           n->set_req(j, in->in(1));
  2723         if (in->outcnt() == 0) {
  2724           in->disconnect_inputs(NULL);
  2731 //------------------------------final_graph_reshaping--------------------------
  2732 // Final Graph Reshaping.
  2733 //
  2734 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
  2735 //     and not commoned up and forced early.  Must come after regular
  2736 //     optimizations to avoid GVN undoing the cloning.  Clone constant
  2737 //     inputs to Loop Phis; these will be split by the allocator anyways.
  2738 //     Remove Opaque nodes.
  2739 // (2) Move last-uses by commutative operations to the left input to encourage
  2740 //     Intel update-in-place two-address operations and better register usage
  2741 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
  2742 //     calls canonicalizing them back.
  2743 // (3) Count the number of double-precision FP ops, single-precision FP ops
  2744 //     and call sites.  On Intel, we can get correct rounding either by
  2745 //     forcing singles to memory (requires extra stores and loads after each
  2746 //     FP bytecode) or we can set a rounding mode bit (requires setting and
  2747 //     clearing the mode bit around call sites).  The mode bit is only used
  2748 //     if the relative frequency of single FP ops to calls is low enough.
  2749 //     This is a key transform for SPEC mpeg_audio.
  2750 // (4) Detect infinite loops; blobs of code reachable from above but not
  2751 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
  2752 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
  2753 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
  2754 //     Detection is by looking for IfNodes where only 1 projection is
  2755 //     reachable from below or CatchNodes missing some targets.
  2756 // (5) Assert for insane oop offsets in debug mode.
  2758 bool Compile::final_graph_reshaping() {
  2759   // an infinite loop may have been eliminated by the optimizer,
  2760   // in which case the graph will be empty.
  2761   if (root()->req() == 1) {
  2762     record_method_not_compilable("trivial infinite loop");
  2763     return true;
  2766   Final_Reshape_Counts frc;
  2768   // Visit everybody reachable!
  2769   // Allocate stack of size C->unique()/2 to avoid frequent realloc
  2770   Node_Stack nstack(unique() >> 1);
  2771   final_graph_reshaping_walk(nstack, root(), frc);
  2773   // Check for unreachable (from below) code (i.e., infinite loops).
  2774   for( uint i = 0; i < frc._tests.size(); i++ ) {
  2775     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
  2776     // Get number of CFG targets.
  2777     // Note that PCTables include exception targets after calls.
  2778     uint required_outcnt = n->required_outcnt();
  2779     if (n->outcnt() != required_outcnt) {
  2780       // Check for a few special cases.  Rethrow Nodes never take the
  2781       // 'fall-thru' path, so expected kids is 1 less.
  2782       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
  2783         if (n->in(0)->in(0)->is_Call()) {
  2784           CallNode *call = n->in(0)->in(0)->as_Call();
  2785           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
  2786             required_outcnt--;      // Rethrow always has 1 less kid
  2787           } else if (call->req() > TypeFunc::Parms &&
  2788                      call->is_CallDynamicJava()) {
  2789             // Check for null receiver. In such case, the optimizer has
  2790             // detected that the virtual call will always result in a null
  2791             // pointer exception. The fall-through projection of this CatchNode
  2792             // will not be populated.
  2793             Node *arg0 = call->in(TypeFunc::Parms);
  2794             if (arg0->is_Type() &&
  2795                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
  2796               required_outcnt--;
  2798           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
  2799                      call->req() > TypeFunc::Parms+1 &&
  2800                      call->is_CallStaticJava()) {
  2801             // Check for negative array length. In such case, the optimizer has
  2802             // detected that the allocation attempt will always result in an
  2803             // exception. There is no fall-through projection of this CatchNode .
  2804             Node *arg1 = call->in(TypeFunc::Parms+1);
  2805             if (arg1->is_Type() &&
  2806                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
  2807               required_outcnt--;
  2812       // Recheck with a better notion of 'required_outcnt'
  2813       if (n->outcnt() != required_outcnt) {
  2814         record_method_not_compilable("malformed control flow");
  2815         return true;            // Not all targets reachable!
  2818     // Check that I actually visited all kids.  Unreached kids
  2819     // must be infinite loops.
  2820     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
  2821       if (!frc._visited.test(n->fast_out(j)->_idx)) {
  2822         record_method_not_compilable("infinite loop");
  2823         return true;            // Found unvisited kid; must be unreach
  2827   // If original bytecodes contained a mixture of floats and doubles
  2828   // check if the optimizer has made it homogenous, item (3).
  2829   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
  2830       frc.get_float_count() > 32 &&
  2831       frc.get_double_count() == 0 &&
  2832       (10 * frc.get_call_count() < frc.get_float_count()) ) {
  2833     set_24_bit_selection_and_mode( false,  true );
  2836   set_java_calls(frc.get_java_call_count());
  2837   set_inner_loops(frc.get_inner_loop_count());
  2839   // No infinite loops, no reason to bail out.
  2840   return false;
  2843 //-----------------------------too_many_traps----------------------------------
  2844 // Report if there are too many traps at the current method and bci.
  2845 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
  2846 bool Compile::too_many_traps(ciMethod* method,
  2847                              int bci,
  2848                              Deoptimization::DeoptReason reason) {
  2849   ciMethodData* md = method->method_data();
  2850   if (md->is_empty()) {
  2851     // Assume the trap has not occurred, or that it occurred only
  2852     // because of a transient condition during start-up in the interpreter.
  2853     return false;
  2855   if (md->has_trap_at(bci, reason) != 0) {
  2856     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
  2857     // Also, if there are multiple reasons, or if there is no per-BCI record,
  2858     // assume the worst.
  2859     if (log())
  2860       log()->elem("observe trap='%s' count='%d'",
  2861                   Deoptimization::trap_reason_name(reason),
  2862                   md->trap_count(reason));
  2863     return true;
  2864   } else {
  2865     // Ignore method/bci and see if there have been too many globally.
  2866     return too_many_traps(reason, md);
  2870 // Less-accurate variant which does not require a method and bci.
  2871 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
  2872                              ciMethodData* logmd) {
  2873  if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
  2874     // Too many traps globally.
  2875     // Note that we use cumulative trap_count, not just md->trap_count.
  2876     if (log()) {
  2877       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
  2878       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
  2879                   Deoptimization::trap_reason_name(reason),
  2880                   mcount, trap_count(reason));
  2882     return true;
  2883   } else {
  2884     // The coast is clear.
  2885     return false;
  2889 //--------------------------too_many_recompiles--------------------------------
  2890 // Report if there are too many recompiles at the current method and bci.
  2891 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
  2892 // Is not eager to return true, since this will cause the compiler to use
  2893 // Action_none for a trap point, to avoid too many recompilations.
  2894 bool Compile::too_many_recompiles(ciMethod* method,
  2895                                   int bci,
  2896                                   Deoptimization::DeoptReason reason) {
  2897   ciMethodData* md = method->method_data();
  2898   if (md->is_empty()) {
  2899     // Assume the trap has not occurred, or that it occurred only
  2900     // because of a transient condition during start-up in the interpreter.
  2901     return false;
  2903   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
  2904   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
  2905   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
  2906   Deoptimization::DeoptReason per_bc_reason
  2907     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
  2908   if ((per_bc_reason == Deoptimization::Reason_none
  2909        || md->has_trap_at(bci, reason) != 0)
  2910       // The trap frequency measure we care about is the recompile count:
  2911       && md->trap_recompiled_at(bci)
  2912       && md->overflow_recompile_count() >= bc_cutoff) {
  2913     // Do not emit a trap here if it has already caused recompilations.
  2914     // Also, if there are multiple reasons, or if there is no per-BCI record,
  2915     // assume the worst.
  2916     if (log())
  2917       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
  2918                   Deoptimization::trap_reason_name(reason),
  2919                   md->trap_count(reason),
  2920                   md->overflow_recompile_count());
  2921     return true;
  2922   } else if (trap_count(reason) != 0
  2923              && decompile_count() >= m_cutoff) {
  2924     // Too many recompiles globally, and we have seen this sort of trap.
  2925     // Use cumulative decompile_count, not just md->decompile_count.
  2926     if (log())
  2927       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
  2928                   Deoptimization::trap_reason_name(reason),
  2929                   md->trap_count(reason), trap_count(reason),
  2930                   md->decompile_count(), decompile_count());
  2931     return true;
  2932   } else {
  2933     // The coast is clear.
  2934     return false;
  2939 #ifndef PRODUCT
  2940 //------------------------------verify_graph_edges---------------------------
  2941 // Walk the Graph and verify that there is a one-to-one correspondence
  2942 // between Use-Def edges and Def-Use edges in the graph.
  2943 void Compile::verify_graph_edges(bool no_dead_code) {
  2944   if (VerifyGraphEdges) {
  2945     ResourceArea *area = Thread::current()->resource_area();
  2946     Unique_Node_List visited(area);
  2947     // Call recursive graph walk to check edges
  2948     _root->verify_edges(visited);
  2949     if (no_dead_code) {
  2950       // Now make sure that no visited node is used by an unvisited node.
  2951       bool dead_nodes = 0;
  2952       Unique_Node_List checked(area);
  2953       while (visited.size() > 0) {
  2954         Node* n = visited.pop();
  2955         checked.push(n);
  2956         for (uint i = 0; i < n->outcnt(); i++) {
  2957           Node* use = n->raw_out(i);
  2958           if (checked.member(use))  continue;  // already checked
  2959           if (visited.member(use))  continue;  // already in the graph
  2960           if (use->is_Con())        continue;  // a dead ConNode is OK
  2961           // At this point, we have found a dead node which is DU-reachable.
  2962           if (dead_nodes++ == 0)
  2963             tty->print_cr("*** Dead nodes reachable via DU edges:");
  2964           use->dump(2);
  2965           tty->print_cr("---");
  2966           checked.push(use);  // No repeats; pretend it is now checked.
  2969       assert(dead_nodes == 0, "using nodes must be reachable from root");
  2973 #endif
  2975 // The Compile object keeps track of failure reasons separately from the ciEnv.
  2976 // This is required because there is not quite a 1-1 relation between the
  2977 // ciEnv and its compilation task and the Compile object.  Note that one
  2978 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
  2979 // to backtrack and retry without subsuming loads.  Other than this backtracking
  2980 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
  2981 // by the logic in C2Compiler.
  2982 void Compile::record_failure(const char* reason) {
  2983   if (log() != NULL) {
  2984     log()->elem("failure reason='%s' phase='compile'", reason);
  2986   if (_failure_reason == NULL) {
  2987     // Record the first failure reason.
  2988     _failure_reason = reason;
  2990   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
  2991     C->print_method(_failure_reason);
  2993   _root = NULL;  // flush the graph, too
  2996 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
  2997   : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
  2999   if (dolog) {
  3000     C = Compile::current();
  3001     _log = C->log();
  3002   } else {
  3003     C = NULL;
  3004     _log = NULL;
  3006   if (_log != NULL) {
  3007     _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
  3008     _log->stamp();
  3009     _log->end_head();
  3013 Compile::TracePhase::~TracePhase() {
  3014   if (_log != NULL) {
  3015     _log->done("phase nodes='%d'", C->unique());
  3019 //=============================================================================
  3020 // Two Constant's are equal when the type and the value are equal.
  3021 bool Compile::Constant::operator==(const Constant& other) {
  3022   if (type()          != other.type()         )  return false;
  3023   if (can_be_reused() != other.can_be_reused())  return false;
  3024   // For floating point values we compare the bit pattern.
  3025   switch (type()) {
  3026   case T_FLOAT:   return (_v._value.i == other._v._value.i);
  3027   case T_LONG:
  3028   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
  3029   case T_OBJECT:
  3030   case T_METADATA: return (_v._metadata == other._v._metadata);
  3031   case T_ADDRESS: return (_v._value.l == other._v._value.l);
  3032   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
  3033   default: ShouldNotReachHere();
  3035   return false;
  3038 static int type_to_size_in_bytes(BasicType t) {
  3039   switch (t) {
  3040   case T_LONG:    return sizeof(jlong  );
  3041   case T_FLOAT:   return sizeof(jfloat );
  3042   case T_DOUBLE:  return sizeof(jdouble);
  3043   case T_METADATA: return sizeof(Metadata*);
  3044     // We use T_VOID as marker for jump-table entries (labels) which
  3045     // need an internal word relocation.
  3046   case T_VOID:
  3047   case T_ADDRESS:
  3048   case T_OBJECT:  return sizeof(jobject);
  3051   ShouldNotReachHere();
  3052   return -1;
  3055 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
  3056   // sort descending
  3057   if (a->freq() > b->freq())  return -1;
  3058   if (a->freq() < b->freq())  return  1;
  3059   return 0;
  3062 void Compile::ConstantTable::calculate_offsets_and_size() {
  3063   // First, sort the array by frequencies.
  3064   _constants.sort(qsort_comparator);
  3066 #ifdef ASSERT
  3067   // Make sure all jump-table entries were sorted to the end of the
  3068   // array (they have a negative frequency).
  3069   bool found_void = false;
  3070   for (int i = 0; i < _constants.length(); i++) {
  3071     Constant con = _constants.at(i);
  3072     if (con.type() == T_VOID)
  3073       found_void = true;  // jump-tables
  3074     else
  3075       assert(!found_void, "wrong sorting");
  3077 #endif
  3079   int offset = 0;
  3080   for (int i = 0; i < _constants.length(); i++) {
  3081     Constant* con = _constants.adr_at(i);
  3083     // Align offset for type.
  3084     int typesize = type_to_size_in_bytes(con->type());
  3085     offset = align_size_up(offset, typesize);
  3086     con->set_offset(offset);   // set constant's offset
  3088     if (con->type() == T_VOID) {
  3089       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
  3090       offset = offset + typesize * n->outcnt();  // expand jump-table
  3091     } else {
  3092       offset = offset + typesize;
  3096   // Align size up to the next section start (which is insts; see
  3097   // CodeBuffer::align_at_start).
  3098   assert(_size == -1, "already set?");
  3099   _size = align_size_up(offset, CodeEntryAlignment);
  3102 void Compile::ConstantTable::emit(CodeBuffer& cb) {
  3103   MacroAssembler _masm(&cb);
  3104   for (int i = 0; i < _constants.length(); i++) {
  3105     Constant con = _constants.at(i);
  3106     address constant_addr;
  3107     switch (con.type()) {
  3108     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
  3109     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
  3110     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
  3111     case T_OBJECT: {
  3112       jobject obj = con.get_jobject();
  3113       int oop_index = _masm.oop_recorder()->find_index(obj);
  3114       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
  3115       break;
  3117     case T_ADDRESS: {
  3118       address addr = (address) con.get_jobject();
  3119       constant_addr = _masm.address_constant(addr);
  3120       break;
  3122     // We use T_VOID as marker for jump-table entries (labels) which
  3123     // need an internal word relocation.
  3124     case T_VOID: {
  3125       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
  3126       // Fill the jump-table with a dummy word.  The real value is
  3127       // filled in later in fill_jump_table.
  3128       address dummy = (address) n;
  3129       constant_addr = _masm.address_constant(dummy);
  3130       // Expand jump-table
  3131       for (uint i = 1; i < n->outcnt(); i++) {
  3132         address temp_addr = _masm.address_constant(dummy + i);
  3133         assert(temp_addr, "consts section too small");
  3135       break;
  3137     case T_METADATA: {
  3138       Metadata* obj = con.get_metadata();
  3139       int metadata_index = _masm.oop_recorder()->find_index(obj);
  3140       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
  3141       break;
  3143     default: ShouldNotReachHere();
  3145     assert(constant_addr, "consts section too small");
  3146     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
  3150 int Compile::ConstantTable::find_offset(Constant& con) const {
  3151   int idx = _constants.find(con);
  3152   assert(idx != -1, "constant must be in constant table");
  3153   int offset = _constants.at(idx).offset();
  3154   assert(offset != -1, "constant table not emitted yet?");
  3155   return offset;
  3158 void Compile::ConstantTable::add(Constant& con) {
  3159   if (con.can_be_reused()) {
  3160     int idx = _constants.find(con);
  3161     if (idx != -1 && _constants.at(idx).can_be_reused()) {
  3162       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
  3163       return;
  3166   (void) _constants.append(con);
  3169 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
  3170   Block* b = Compile::current()->cfg()->_bbs[n->_idx];
  3171   Constant con(type, value, b->_freq);
  3172   add(con);
  3173   return con;
  3176 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
  3177   Constant con(metadata);
  3178   add(con);
  3179   return con;
  3182 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
  3183   jvalue value;
  3184   BasicType type = oper->type()->basic_type();
  3185   switch (type) {
  3186   case T_LONG:    value.j = oper->constantL(); break;
  3187   case T_FLOAT:   value.f = oper->constantF(); break;
  3188   case T_DOUBLE:  value.d = oper->constantD(); break;
  3189   case T_OBJECT:
  3190   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
  3191   case T_METADATA: return add((Metadata*)oper->constant()); break;
  3192   default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
  3194   return add(n, type, value);
  3197 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
  3198   jvalue value;
  3199   // We can use the node pointer here to identify the right jump-table
  3200   // as this method is called from Compile::Fill_buffer right before
  3201   // the MachNodes are emitted and the jump-table is filled (means the
  3202   // MachNode pointers do not change anymore).
  3203   value.l = (jobject) n;
  3204   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
  3205   add(con);
  3206   return con;
  3209 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
  3210   // If called from Compile::scratch_emit_size do nothing.
  3211   if (Compile::current()->in_scratch_emit_size())  return;
  3213   assert(labels.is_nonempty(), "must be");
  3214   assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
  3216   // Since MachConstantNode::constant_offset() also contains
  3217   // table_base_offset() we need to subtract the table_base_offset()
  3218   // to get the plain offset into the constant table.
  3219   int offset = n->constant_offset() - table_base_offset();
  3221   MacroAssembler _masm(&cb);
  3222   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
  3224   for (uint i = 0; i < n->outcnt(); i++) {
  3225     address* constant_addr = &jump_table_base[i];
  3226     assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
  3227     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
  3228     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);

mercurial