1 /*
   2  * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.inline.hpp"
  27 #include "code/compiledIC.hpp"
  28 #include "code/debugInfo.hpp"
  29 #include "code/debugInfoRec.hpp"
  30 #include "compiler/compileBroker.hpp"
  31 #include "compiler/oopMap.hpp"
  32 #include "memory/allocation.inline.hpp"
  33 #include "opto/callnode.hpp"
  34 #include "opto/cfgnode.hpp"
  35 #include "opto/locknode.hpp"
  36 #include "opto/machnode.hpp"
  37 #include "opto/output.hpp"
  38 #include "opto/regalloc.hpp"
  39 #include "opto/runtime.hpp"
  40 #include "opto/subnode.hpp"
  41 #include "opto/type.hpp"
  42 #include "runtime/handles.inline.hpp"
  43 #include "utilities/xmlstream.hpp"
  44 
  45 #ifndef PRODUCT
  46 #define DEBUG_ARG(x) , x
  47 #else
  48 #define DEBUG_ARG(x)
  49 #endif
  50 
  51 // Convert Nodes to instruction bits and pass off to the VM
  52 void Compile::Output() {
  53   // RootNode goes
  54   assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
  55 
  56   // The number of new nodes (mostly MachNop) is proportional to
  57   // the number of java calls and inner loops which are aligned.
  58   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
  59                             C->inner_loops()*(OptoLoopAlignment-1)),
  60                            "out of nodes before code generation" ) ) {
  61     return;
  62   }
  63   // Make sure I can find the Start Node
  64   Block *entry = _cfg->get_block(1);
  65   Block *broot = _cfg->get_root_block();
  66 
  67   const StartNode *start = entry->head()->as_Start();
  68 
  69   // Replace StartNode with prolog
  70   MachPrologNode *prolog = new (this) MachPrologNode();
  71   entry->map_node(prolog, 0);
  72   _cfg->map_node_to_block(prolog, entry);
  73   _cfg->unmap_node_from_block(start); // start is no longer in any block
  74 
  75   // Virtual methods need an unverified entry point
  76 
  77   if( is_osr_compilation() ) {
  78     if( PoisonOSREntry ) {
  79       // TODO: Should use a ShouldNotReachHereNode...
  80       _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
  81     }
  82   } else {
  83     if( _method && !_method->flags().is_static() ) {
  84       // Insert unvalidated entry point
  85       _cfg->insert( broot, 0, new (this) MachUEPNode() );
  86     }
  87 
  88   }
  89 
  90 
  91   // Break before main entry point
  92   if( (_method && _method->break_at_execute())
  93 #ifndef PRODUCT
  94     ||(OptoBreakpoint && is_method_compilation())
  95     ||(OptoBreakpointOSR && is_osr_compilation())
  96     ||(OptoBreakpointC2R && !_method)
  97 #endif
  98     ) {
  99     // checking for _method means that OptoBreakpoint does not apply to
 100     // runtime stubs or frame converters
 101     _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
 102   }
 103 
 104   // Insert epilogs before every return
 105   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
 106     Block* block = _cfg->get_block(i);
 107     if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
 108       Node* m = block->end();
 109       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
 110         MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 111         block->add_inst(epilog);
 112         _cfg->map_node_to_block(epilog, block);
 113       }
 114     }
 115   }
 116 
 117 # ifdef ENABLE_ZAP_DEAD_LOCALS
 118   if (ZapDeadCompiledLocals) {
 119     Insert_zap_nodes();
 120   }
 121 # endif
 122 
 123   uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
 124   blk_starts[0] = 0;
 125 
 126   // Initialize code buffer and process short branches.
 127   CodeBuffer* cb = init_buffer(blk_starts);
 128 
 129   if (cb == NULL || failing()) {
 130     return;
 131   }
 132 
 133   ScheduleAndBundle();
 134 
 135 #ifndef PRODUCT
 136   if (trace_opto_output()) {
 137     tty->print("\n---- After ScheduleAndBundle ----\n");
 138     for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
 139       tty->print("\nBB#%03d:\n", i);
 140       Block* block = _cfg->get_block(i);
 141       for (uint j = 0; j < block->number_of_nodes(); j++) {
 142         Node* n = block->get_node(j);
 143         OptoReg::Name reg = _regalloc->get_reg_first(n);
 144         tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
 145         n->dump();
 146       }
 147     }
 148   }
 149 #endif
 150 
 151   if (failing()) {
 152     return;
 153   }
 154 
 155   BuildOopMaps();
 156 
 157   if (failing())  {
 158     return;
 159   }
 160 
 161   fill_buffer(cb, blk_starts);
 162 }
 163 
 164 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
 165   // Determine if we need to generate a stack overflow check.
 166   // Do it if the method is not a stub function and
 167   // has java calls or has frame size > vm_page_size/8.
 168   // The debug VM checks that deoptimization doesn't trigger an
 169   // unexpected stack overflow (compiled method stack banging should
 170   // guarantee it doesn't happen) so we always need the stack bang in
 171   // a debug VM.
 172   return (UseStackBanging && stub_function() == NULL &&
 173           (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
 174            DEBUG_ONLY(|| true)));
 175 }
 176 
 177 bool Compile::need_register_stack_bang() const {
 178   // Determine if we need to generate a register stack overflow check.
 179   // This is only used on architectures which have split register
 180   // and memory stacks (ie. IA64).
 181   // Bang if the method is not a stub function and has java calls
 182   return (stub_function() == NULL && has_java_calls());
 183 }
 184 
 185 # ifdef ENABLE_ZAP_DEAD_LOCALS
 186 
 187 
 188 // In order to catch compiler oop-map bugs, we have implemented
 189 // a debugging mode called ZapDeadCompilerLocals.
 190 // This mode causes the compiler to insert a call to a runtime routine,
 191 // "zap_dead_locals", right before each place in compiled code
 192 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
 193 // The runtime routine checks that locations mapped as oops are really
 194 // oops, that locations mapped as values do not look like oops,
 195 // and that locations mapped as dead are not used later
 196 // (by zapping them to an invalid address).
 197 
 198 int Compile::_CompiledZap_count = 0;
 199 
 200 void Compile::Insert_zap_nodes() {
 201   bool skip = false;
 202 
 203 
 204   // Dink with static counts because code code without the extra
 205   // runtime calls is MUCH faster for debugging purposes
 206 
 207        if ( CompileZapFirst  ==  0  ) ; // nothing special
 208   else if ( CompileZapFirst  >  CompiledZap_count() )  skip = true;
 209   else if ( CompileZapFirst  == CompiledZap_count() )
 210     warning("starting zap compilation after skipping");
 211 
 212        if ( CompileZapLast  ==  -1  ) ; // nothing special
 213   else if ( CompileZapLast  <   CompiledZap_count() )  skip = true;
 214   else if ( CompileZapLast  ==  CompiledZap_count() )
 215     warning("about to compile last zap");
 216 
 217   ++_CompiledZap_count; // counts skipped zaps, too
 218 
 219   if ( skip )  return;
 220 
 221 
 222   if ( _method == NULL )
 223     return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
 224 
 225   // Insert call to zap runtime stub before every node with an oop map
 226   for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
 227     Block *b = _cfg->get_block(i);
 228     for ( uint j = 0;  j < b->number_of_nodes();  ++j ) {
 229       Node *n = b->get_node(j);
 230 
 231       // Determining if we should insert a zap-a-lot node in output.
 232       // We do that for all nodes that has oopmap info, except for calls
 233       // to allocation.  Calls to allocation passes in the old top-of-eden pointer
 234       // and expect the C code to reset it.  Hence, there can be no safepoints between
 235       // the inlined-allocation and the call to new_Java, etc.
 236       // We also cannot zap monitor calls, as they must hold the microlock
 237       // during the call to Zap, which also wants to grab the microlock.
 238       bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
 239       if ( insert ) { // it is MachSafePoint
 240         if ( !n->is_MachCall() ) {
 241           insert = false;
 242         } else if ( n->is_MachCall() ) {
 243           MachCallNode* call = n->as_MachCall();
 244           if (call->entry_point() == OptoRuntime::new_instance_Java() ||
 245               call->entry_point() == OptoRuntime::new_array_Java() ||
 246               call->entry_point() == OptoRuntime::multianewarray2_Java() ||
 247               call->entry_point() == OptoRuntime::multianewarray3_Java() ||
 248               call->entry_point() == OptoRuntime::multianewarray4_Java() ||
 249               call->entry_point() == OptoRuntime::multianewarray5_Java() ||
 250               call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
 251               call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
 252               ) {
 253             insert = false;
 254           }
 255         }
 256         if (insert) {
 257           Node *zap = call_zap_node(n->as_MachSafePoint(), i);
 258           b->insert_node(zap, j);
 259           _cfg->map_node_to_block(zap, b);
 260           ++j;
 261         }
 262       }
 263     }
 264   }
 265 }
 266 
 267 
 268 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
 269   const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
 270   CallStaticJavaNode* ideal_node =
 271     new (this) CallStaticJavaNode( tf,
 272          OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
 273                        "call zap dead locals stub", 0, TypePtr::BOTTOM);
 274   // We need to copy the OopMap from the site we're zapping at.
 275   // We have to make a copy, because the zap site might not be
 276   // a call site, and zap_dead is a call site.
 277   OopMap* clone = node_to_check->oop_map()->deep_copy();
 278 
 279   // Add the cloned OopMap to the zap node
 280   ideal_node->set_oop_map(clone);
 281   return _matcher->match_sfpt(ideal_node);
 282 }
 283 
 284 bool Compile::is_node_getting_a_safepoint( Node* n) {
 285   // This code duplicates the logic prior to the call of add_safepoint
 286   // below in this file.
 287   if( n->is_MachSafePoint() ) return true;
 288   return false;
 289 }
 290 
 291 # endif // ENABLE_ZAP_DEAD_LOCALS
 292 
 293 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 294 // of a loop. When aligning a loop we need to provide enough instructions
 295 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 296 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 297 // By default, the size is set to 999999 by Block's constructor so that
 298 // a loop will be aligned if the size is not reset here.
 299 //
 300 // Note: Mach instructions could contain several HW instructions
 301 // so the size is estimated only.
 302 //
 303 void Compile::compute_loop_first_inst_sizes() {
 304   // The next condition is used to gate the loop alignment optimization.
 305   // Don't aligned a loop if there are enough instructions at the head of a loop
 306   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 307   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 308   // equal to 11 bytes which is the largest address NOP instruction.
 309   if (MaxLoopPad < OptoLoopAlignment - 1) {
 310     uint last_block = _cfg->number_of_blocks() - 1;
 311     for (uint i = 1; i <= last_block; i++) {
 312       Block* block = _cfg->get_block(i);
 313       // Check the first loop's block which requires an alignment.
 314       if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
 315         uint sum_size = 0;
 316         uint inst_cnt = NumberOfLoopInstrToAlign;
 317         inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
 318 
 319         // Check subsequent fallthrough blocks if the loop's first
 320         // block(s) does not have enough instructions.
 321         Block *nb = block;
 322         while(inst_cnt > 0 &&
 323               i < last_block &&
 324               !_cfg->get_block(i + 1)->has_loop_alignment() &&
 325               !nb->has_successor(block)) {
 326           i++;
 327           nb = _cfg->get_block(i);
 328           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
 329         } // while( inst_cnt > 0 && i < last_block  )
 330 
 331         block->set_first_inst_size(sum_size);
 332       } // f( b->head()->is_Loop() )
 333     } // for( i <= last_block )
 334   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 335 }
 336 
 337 // The architecture description provides short branch variants for some long
 338 // branch instructions. Replace eligible long branches with short branches.
 339 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
 340   // Compute size of each block, method size, and relocation information size
 341   uint nblocks  = _cfg->number_of_blocks();
 342 
 343   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 344   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 345   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 346 
 347   // Collect worst case block paddings
 348   int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
 349   memset(block_worst_case_pad, 0, nblocks * sizeof(int));
 350 
 351   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 352   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 353 
 354   bool has_short_branch_candidate = false;
 355 
 356   // Initialize the sizes to 0
 357   code_size  = 0;          // Size in bytes of generated code
 358   stub_size  = 0;          // Size in bytes of all stub entries
 359   // Size in bytes of all relocation entries, including those in local stubs.
 360   // Start with 2-bytes of reloc info for the unvalidated entry point
 361   reloc_size = 1;          // Number of relocation entries
 362 
 363   // Make three passes.  The first computes pessimistic blk_starts,
 364   // relative jmp_offset and reloc_size information.  The second performs
 365   // short branch substitution using the pessimistic sizing.  The
 366   // third inserts nops where needed.
 367 
 368   // Step one, perform a pessimistic sizing pass.
 369   uint last_call_adr = (uint) -1;
 370   uint last_avoid_back_to_back_adr = (uint) -1;
 371   uint nop_size = (new (this) MachNopNode())->size(_regalloc);
 372   for (uint i = 0; i < nblocks; i++) { // For all blocks
 373     Block* block = _cfg->get_block(i);
 374 
 375     // During short branch replacement, we store the relative (to blk_starts)
 376     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 377     // This is so that we do not need to recompute sizes of all nodes when
 378     // we compute correct blk_starts in our next sizing pass.
 379     jmp_offset[i] = 0;
 380     jmp_size[i]   = 0;
 381     jmp_nidx[i]   = -1;
 382     DEBUG_ONLY( jmp_target[i] = 0; )
 383     DEBUG_ONLY( jmp_rule[i]   = 0; )
 384 
 385     // Sum all instruction sizes to compute block size
 386     uint last_inst = block->number_of_nodes();
 387     uint blk_size = 0;
 388     for (uint j = 0; j < last_inst; j++) {
 389       Node* nj = block->get_node(j);
 390       // Handle machine instruction nodes
 391       if (nj->is_Mach()) {
 392         MachNode *mach = nj->as_Mach();
 393         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 394         reloc_size += mach->reloc();
 395         if (mach->is_MachCall()) {
 396           // add size information for trampoline stub
 397           // class CallStubImpl is platform-specific and defined in the *.ad files.
 398           stub_size  += CallStubImpl::size_call_trampoline();
 399           reloc_size += CallStubImpl::reloc_call_trampoline();
 400 
 401           MachCallNode *mcall = mach->as_MachCall();
 402           // This destination address is NOT PC-relative
 403 
 404           mcall->method_set((intptr_t)mcall->entry_point());
 405 
 406           if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
 407             stub_size  += CompiledStaticCall::to_interp_stub_size();
 408             reloc_size += CompiledStaticCall::reloc_to_interp_stub();
 409           }
 410         } else if (mach->is_MachSafePoint()) {
 411           // If call/safepoint are adjacent, account for possible
 412           // nop to disambiguate the two safepoints.
 413           // ScheduleAndBundle() can rearrange nodes in a block,
 414           // check for all offsets inside this block.
 415           if (last_call_adr >= blk_starts[i]) {
 416             blk_size += nop_size;
 417           }
 418         }
 419         if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 420           // Nop is inserted between "avoid back to back" instructions.
 421           // ScheduleAndBundle() can rearrange nodes in a block,
 422           // check for all offsets inside this block.
 423           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 424             blk_size += nop_size;
 425           }
 426         }
 427         if (mach->may_be_short_branch()) {
 428           if (!nj->is_MachBranch()) {
 429 #ifndef PRODUCT
 430             nj->dump(3);
 431 #endif
 432             Unimplemented();
 433           }
 434           assert(jmp_nidx[i] == -1, "block should have only one branch");
 435           jmp_offset[i] = blk_size;
 436           jmp_size[i]   = nj->size(_regalloc);
 437           jmp_nidx[i]   = j;
 438           has_short_branch_candidate = true;
 439         }
 440       }
 441       blk_size += nj->size(_regalloc);
 442       // Remember end of call offset
 443       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 444         last_call_adr = blk_starts[i]+blk_size;
 445       }
 446       // Remember end of avoid_back_to_back offset
 447       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
 448         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 449       }
 450     }
 451 
 452     // When the next block starts a loop, we may insert pad NOP
 453     // instructions.  Since we cannot know our future alignment,
 454     // assume the worst.
 455     if (i < nblocks - 1) {
 456       Block* nb = _cfg->get_block(i + 1);
 457       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 458       if (max_loop_pad > 0) {
 459         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 460         // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
 461         // If either is the last instruction in this block, bump by
 462         // max_loop_pad in lock-step with blk_size, so sizing
 463         // calculations in subsequent blocks still can conservatively
 464         // detect that it may the last instruction in this block.
 465         if (last_call_adr == blk_starts[i]+blk_size) {
 466           last_call_adr += max_loop_pad;
 467         }
 468         if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
 469           last_avoid_back_to_back_adr += max_loop_pad;
 470         }
 471         blk_size += max_loop_pad;
 472         block_worst_case_pad[i + 1] = max_loop_pad;
 473       }
 474     }
 475 
 476     // Save block size; update total method size
 477     blk_starts[i+1] = blk_starts[i]+blk_size;
 478   }
 479 
 480   // Step two, replace eligible long jumps.
 481   bool progress = true;
 482   uint last_may_be_short_branch_adr = (uint) -1;
 483   while (has_short_branch_candidate && progress) {
 484     progress = false;
 485     has_short_branch_candidate = false;
 486     int adjust_block_start = 0;
 487     for (uint i = 0; i < nblocks; i++) {
 488       Block* block = _cfg->get_block(i);
 489       int idx = jmp_nidx[i];
 490       MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
 491       if (mach != NULL && mach->may_be_short_branch()) {
 492 #ifdef ASSERT
 493         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
 494         int j;
 495         // Find the branch; ignore trailing NOPs.
 496         for (j = block->number_of_nodes()-1; j>=0; j--) {
 497           Node* n = block->get_node(j);
 498           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 499             break;
 500         }
 501         assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
 502 #endif
 503         int br_size = jmp_size[i];
 504         int br_offs = blk_starts[i] + jmp_offset[i];
 505 
 506         // This requires the TRUE branch target be in succs[0]
 507         uint bnum = block->non_connector_successor(0)->_pre_order;
 508         int offset = blk_starts[bnum] - br_offs;
 509         if (bnum > i) { // adjust following block's offset
 510           offset -= adjust_block_start;
 511         }
 512 
 513         // This block can be a loop header, account for the padding
 514         // in the previous block.
 515         int block_padding = block_worst_case_pad[i];
 516         assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
 517         // In the following code a nop could be inserted before
 518         // the branch which will increase the backward distance.
 519         bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
 520         assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
 521 
 522         if (needs_padding && offset <= 0)
 523           offset -= nop_size;
 524 
 525         if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
 526           // We've got a winner.  Replace this branch.
 527           MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
 528 
 529           // Update the jmp_size.
 530           int new_size = replacement->size(_regalloc);
 531           int diff     = br_size - new_size;
 532           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 533           // Conservatively take into account padding between
 534           // avoid_back_to_back branches. Previous branch could be
 535           // converted into avoid_back_to_back branch during next
 536           // rounds.
 537           if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 538             jmp_offset[i] += nop_size;
 539             diff -= nop_size;
 540           }
 541           adjust_block_start += diff;
 542           block->map_node(replacement, idx);
 543           mach->subsume_by(replacement, C);
 544           mach = replacement;
 545           progress = true;
 546 
 547           jmp_size[i] = new_size;
 548           DEBUG_ONLY( jmp_target[i] = bnum; );
 549           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 550         } else {
 551           // The jump distance is not short, try again during next iteration.
 552           has_short_branch_candidate = true;
 553         }
 554       } // (mach->may_be_short_branch())
 555       if (mach != NULL && (mach->may_be_short_branch() ||
 556                            mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
 557         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 558       }
 559       blk_starts[i+1] -= adjust_block_start;
 560     }
 561   }
 562 
 563 #ifdef ASSERT
 564   for (uint i = 0; i < nblocks; i++) { // For all blocks
 565     if (jmp_target[i] != 0) {
 566       int br_size = jmp_size[i];
 567       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 568       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 569         tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
 570       }
 571       assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 572     }
 573   }
 574 #endif
 575 
 576   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
 577   // after ScheduleAndBundle().
 578 
 579   // ------------------
 580   // Compute size for code buffer
 581   code_size = blk_starts[nblocks];
 582 
 583   // Relocation records
 584   reloc_size += 1;              // Relo entry for exception handler
 585 
 586   // Adjust reloc_size to number of record of relocation info
 587   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 588   // a relocation index.
 589   // The CodeBuffer will expand the locs array if this estimate is too low.
 590   reloc_size *= 10 / sizeof(relocInfo);
 591 }
 592 
 593 //------------------------------FillLocArray-----------------------------------
 594 // Create a bit of debug info and append it to the array.  The mapping is from
 595 // Java local or expression stack to constant, register or stack-slot.  For
 596 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 597 // entry has been taken care of and caller should skip it).
 598 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 599   // This should never have accepted Bad before
 600   assert(OptoReg::is_valid(regnum), "location must be valid");
 601   return (OptoReg::is_reg(regnum))
 602     ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 603     : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 604 }
 605 
 606 
 607 ObjectValue*
 608 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 609   for (int i = 0; i < objs->length(); i++) {
 610     assert(objs->at(i)->is_object(), "corrupt object cache");
 611     ObjectValue* sv = (ObjectValue*) objs->at(i);
 612     if (sv->id() == id) {
 613       return sv;
 614     }
 615   }
 616   // Otherwise..
 617   return NULL;
 618 }
 619 
 620 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 621                                      ObjectValue* sv ) {
 622   assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
 623   objs->append(sv);
 624 }
 625 
 626 
 627 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 628                             GrowableArray<ScopeValue*> *array,
 629                             GrowableArray<ScopeValue*> *objs ) {
 630   assert( local, "use _top instead of null" );
 631   if (array->length() != idx) {
 632     assert(array->length() == idx + 1, "Unexpected array count");
 633     // Old functionality:
 634     //   return
 635     // New functionality:
 636     //   Assert if the local is not top. In product mode let the new node
 637     //   override the old entry.
 638     assert(local == top(), "LocArray collision");
 639     if (local == top()) {
 640       return;
 641     }
 642     array->pop();
 643   }
 644   const Type *t = local->bottom_type();
 645 
 646   // Is it a safepoint scalar object node?
 647   if (local->is_SafePointScalarObject()) {
 648     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 649 
 650     ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
 651     if (sv == NULL) {
 652       ciKlass* cik = t->is_oopptr()->klass();
 653       assert(cik->is_instance_klass() ||
 654              cik->is_array_klass(), "Not supported allocation.");
 655       sv = new ObjectValue(spobj->_idx,
 656                            new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 657       Compile::set_sv_for_object_node(objs, sv);
 658 
 659       uint first_ind = spobj->first_index(sfpt->jvms());
 660       for (uint i = 0; i < spobj->n_fields(); i++) {
 661         Node* fld_node = sfpt->in(first_ind+i);
 662         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 663       }
 664     }
 665     array->append(sv);
 666     return;
 667   }
 668 
 669   // Grab the register number for the local
 670   OptoReg::Name regnum = _regalloc->get_reg_first(local);
 671   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 672     // Record the double as two float registers.
 673     // The register mask for such a value always specifies two adjacent
 674     // float registers, with the lower register number even.
 675     // Normally, the allocation of high and low words to these registers
 676     // is irrelevant, because nearly all operations on register pairs
 677     // (e.g., StoreD) treat them as a single unit.
 678     // Here, we assume in addition that the words in these two registers
 679     // stored "naturally" (by operations like StoreD and double stores
 680     // within the interpreter) such that the lower-numbered register
 681     // is written to the lower memory address.  This may seem like
 682     // a machine dependency, but it is not--it is a requirement on
 683     // the author of the <arch>.ad file to ensure that, for every
 684     // even/odd double-register pair to which a double may be allocated,
 685     // the word in the even single-register is stored to the first
 686     // memory word.  (Note that register numbers are completely
 687     // arbitrary, and are not tied to any machine-level encodings.)
 688 #ifdef _LP64
 689     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 690       array->append(new ConstantIntValue(0));
 691       array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
 692     } else if ( t->base() == Type::Long ) {
 693       array->append(new ConstantIntValue(0));
 694       array->append(new_loc_value( _regalloc, regnum, Location::lng ));
 695     } else if ( t->base() == Type::RawPtr ) {
 696       // jsr/ret return address which must be restored into a the full
 697       // width 64-bit stack slot.
 698       array->append(new_loc_value( _regalloc, regnum, Location::lng ));
 699     }
 700 #else //_LP64
 701 #ifdef SPARC
 702     if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
 703       // For SPARC we have to swap high and low words for
 704       // long values stored in a single-register (g0-g7).
 705       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));
 706       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
 707     } else
 708 #endif //SPARC
 709     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 710       // Repack the double/long as two jints.
 711       // The convention the interpreter uses is that the second local
 712       // holds the first raw word of the native double representation.
 713       // This is actually reasonable, since locals and stack arrays
 714       // grow downwards in all implementations.
 715       // (If, on some machine, the interpreter's Java locals or stack
 716       // were to grow upwards, the embedded doubles would be word-swapped.)
 717       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
 718       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));
 719     }
 720 #endif //_LP64
 721     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 722                OptoReg::is_reg(regnum) ) {
 723       array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
 724                                    ? Location::float_in_dbl : Location::normal ));
 725     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 726       array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
 727                                    ? Location::int_in_long : Location::normal ));
 728     } else if( t->base() == Type::NarrowOop ) {
 729       array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
 730     } else {
 731       array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
 732     }
 733     return;
 734   }
 735 
 736   // No register.  It must be constant data.
 737   switch (t->base()) {
 738   case Type::Half:              // Second half of a double
 739     ShouldNotReachHere();       // Caller should skip 2nd halves
 740     break;
 741   case Type::AnyPtr:
 742     array->append(new ConstantOopWriteValue(NULL));
 743     break;
 744   case Type::AryPtr:
 745   case Type::InstPtr:          // fall through
 746     array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 747     break;
 748   case Type::NarrowOop:
 749     if (t == TypeNarrowOop::NULL_PTR) {
 750       array->append(new ConstantOopWriteValue(NULL));
 751     } else {
 752       array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 753     }
 754     break;
 755   case Type::Int:
 756     array->append(new ConstantIntValue(t->is_int()->get_con()));
 757     break;
 758   case Type::RawPtr:
 759     // A return address (T_ADDRESS).
 760     assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 761 #ifdef _LP64
 762     // Must be restored to the full-width 64-bit stack slot.
 763     array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 764 #else
 765     array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 766 #endif
 767     break;
 768   case Type::FloatCon: {
 769     float f = t->is_float_constant()->getf();
 770     array->append(new ConstantIntValue(jint_cast(f)));
 771     break;
 772   }
 773   case Type::DoubleCon: {
 774     jdouble d = t->is_double_constant()->getd();
 775 #ifdef _LP64
 776     array->append(new ConstantIntValue(0));
 777     array->append(new ConstantDoubleValue(d));
 778 #else
 779     // Repack the double as two jints.
 780     // The convention the interpreter uses is that the second local
 781     // holds the first raw word of the native double representation.
 782     // This is actually reasonable, since locals and stack arrays
 783     // grow downwards in all implementations.
 784     // (If, on some machine, the interpreter's Java locals or stack
 785     // were to grow upwards, the embedded doubles would be word-swapped.)
 786     jint   *dp = (jint*)&d;
 787     array->append(new ConstantIntValue(dp[1]));
 788     array->append(new ConstantIntValue(dp[0]));
 789 #endif
 790     break;
 791   }
 792   case Type::Long: {
 793     jlong d = t->is_long()->get_con();
 794 #ifdef _LP64
 795     array->append(new ConstantIntValue(0));
 796     array->append(new ConstantLongValue(d));
 797 #else
 798     // Repack the long as two jints.
 799     // The convention the interpreter uses is that the second local
 800     // holds the first raw word of the native double representation.
 801     // This is actually reasonable, since locals and stack arrays
 802     // grow downwards in all implementations.
 803     // (If, on some machine, the interpreter's Java locals or stack
 804     // were to grow upwards, the embedded doubles would be word-swapped.)
 805     jint *dp = (jint*)&d;
 806     array->append(new ConstantIntValue(dp[1]));
 807     array->append(new ConstantIntValue(dp[0]));
 808 #endif
 809     break;
 810   }
 811   case Type::Top:               // Add an illegal value here
 812     array->append(new LocationValue(Location()));
 813     break;
 814   default:
 815     ShouldNotReachHere();
 816     break;
 817   }
 818 }
 819 
 820 // Determine if this node starts a bundle
 821 bool Compile::starts_bundle(const Node *n) const {
 822   return (_node_bundling_limit > n->_idx &&
 823           _node_bundling_base[n->_idx].starts_bundle());
 824 }
 825 
 826 //--------------------------Process_OopMap_Node--------------------------------
 827 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
 828 
 829   // Handle special safepoint nodes for synchronization
 830   MachSafePointNode *sfn   = mach->as_MachSafePoint();
 831   MachCallNode      *mcall;
 832 
 833 #ifdef ENABLE_ZAP_DEAD_LOCALS
 834   assert( is_node_getting_a_safepoint(mach),  "logic does not match; false negative");
 835 #endif
 836 
 837   int safepoint_pc_offset = current_offset;
 838   bool is_method_handle_invoke = false;
 839   bool return_oop = false;
 840 
 841   // Add the safepoint in the DebugInfoRecorder
 842   if( !mach->is_MachCall() ) {
 843     mcall = NULL;
 844     debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
 845   } else {
 846     mcall = mach->as_MachCall();
 847 
 848     // Is the call a MethodHandle call?
 849     if (mcall->is_MachCallJava()) {
 850       if (mcall->as_MachCallJava()->_method_handle_invoke) {
 851         assert(has_method_handle_invokes(), "must have been set during call generation");
 852         is_method_handle_invoke = true;
 853       }
 854     }
 855 
 856     // Check if a call returns an object.
 857     if (mcall->return_value_is_used() &&
 858         mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
 859       return_oop = true;
 860     }
 861     safepoint_pc_offset += mcall->ret_addr_offset();
 862     debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
 863   }
 864 
 865   // Loop over the JVMState list to add scope information
 866   // Do not skip safepoints with a NULL method, they need monitor info
 867   JVMState* youngest_jvms = sfn->jvms();
 868   int max_depth = youngest_jvms->depth();
 869 
 870   // Allocate the object pool for scalar-replaced objects -- the map from
 871   // small-integer keys (which can be recorded in the local and ostack
 872   // arrays) to descriptions of the object state.
 873   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
 874 
 875   // Visit scopes from oldest to youngest.
 876   for (int depth = 1; depth <= max_depth; depth++) {
 877     JVMState* jvms = youngest_jvms->of_depth(depth);
 878     int idx;
 879     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
 880     // Safepoints that do not have method() set only provide oop-map and monitor info
 881     // to support GC; these do not support deoptimization.
 882     int num_locs = (method == NULL) ? 0 : jvms->loc_size();
 883     int num_exps = (method == NULL) ? 0 : jvms->stk_size();
 884     int num_mon  = jvms->nof_monitors();
 885     assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
 886            "JVMS local count must match that of the method");
 887 
 888     // Add Local and Expression Stack Information
 889 
 890     // Insert locals into the locarray
 891     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
 892     for( idx = 0; idx < num_locs; idx++ ) {
 893       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
 894     }
 895 
 896     // Insert expression stack entries into the exparray
 897     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
 898     for( idx = 0; idx < num_exps; idx++ ) {
 899       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
 900     }
 901 
 902     // Add in mappings of the monitors
 903     assert( !method ||
 904             !method->is_synchronized() ||
 905             method->is_native() ||
 906             num_mon > 0 ||
 907             !GenerateSynchronizationCode,
 908             "monitors must always exist for synchronized methods");
 909 
 910     // Build the growable array of ScopeValues for exp stack
 911     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
 912 
 913     // Loop over monitors and insert into array
 914     for (idx = 0; idx < num_mon; idx++) {
 915       // Grab the node that defines this monitor
 916       Node* box_node = sfn->monitor_box(jvms, idx);
 917       Node* obj_node = sfn->monitor_obj(jvms, idx);
 918 
 919       // Create ScopeValue for object
 920       ScopeValue *scval = NULL;
 921 
 922       if (obj_node->is_SafePointScalarObject()) {
 923         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
 924         scval = Compile::sv_for_node_id(objs, spobj->_idx);
 925         if (scval == NULL) {
 926           const Type *t = spobj->bottom_type();
 927           ciKlass* cik = t->is_oopptr()->klass();
 928           assert(cik->is_instance_klass() ||
 929                  cik->is_array_klass(), "Not supported allocation.");
 930           ObjectValue* sv = new ObjectValue(spobj->_idx,
 931                                             new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 932           Compile::set_sv_for_object_node(objs, sv);
 933 
 934           uint first_ind = spobj->first_index(youngest_jvms);
 935           for (uint i = 0; i < spobj->n_fields(); i++) {
 936             Node* fld_node = sfn->in(first_ind+i);
 937             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
 938           }
 939           scval = sv;
 940         }
 941       } else if (!obj_node->is_Con()) {
 942         OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
 943         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
 944           scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
 945         } else {
 946           scval = new_loc_value( _regalloc, obj_reg, Location::oop );
 947         }
 948       } else {
 949         const TypePtr *tp = obj_node->get_ptr_type();
 950         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
 951       }
 952 
 953       OptoReg::Name box_reg = BoxLockNode::reg(box_node);
 954       Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
 955       bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
 956       monarray->append(new MonitorValue(scval, basic_lock, eliminated));
 957     }
 958 
 959     // We dump the object pool first, since deoptimization reads it in first.
 960     debug_info()->dump_object_pool(objs);
 961 
 962     // Build first class objects to pass to scope
 963     DebugToken *locvals = debug_info()->create_scope_values(locarray);
 964     DebugToken *expvals = debug_info()->create_scope_values(exparray);
 965     DebugToken *monvals = debug_info()->create_monitor_values(monarray);
 966 
 967     // Make method available for all Safepoints
 968     ciMethod* scope_method = method ? method : _method;
 969     // Describe the scope here
 970     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
 971     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
 972     // Now we can describe the scope.
 973     debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
 974   } // End jvms loop
 975 
 976   // Mark the end of the scope set.
 977   debug_info()->end_safepoint(safepoint_pc_offset);
 978 }
 979 
 980 
 981 
 982 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
 983 class NonSafepointEmitter {
 984   Compile*  C;
 985   JVMState* _pending_jvms;
 986   int       _pending_offset;
 987 
 988   void emit_non_safepoint();
 989 
 990  public:
 991   NonSafepointEmitter(Compile* compile) {
 992     this->C = compile;
 993     _pending_jvms = NULL;
 994     _pending_offset = 0;
 995   }
 996 
 997   void observe_instruction(Node* n, int pc_offset) {
 998     if (!C->debug_info()->recording_non_safepoints())  return;
 999 
1000     Node_Notes* nn = C->node_notes_at(n->_idx);
1001     if (nn == NULL || nn->jvms() == NULL)  return;
1002     if (_pending_jvms != NULL &&
1003         _pending_jvms->same_calls_as(nn->jvms())) {
1004       // Repeated JVMS?  Stretch it up here.
1005       _pending_offset = pc_offset;
1006     } else {
1007       if (_pending_jvms != NULL &&
1008           _pending_offset < pc_offset) {
1009         emit_non_safepoint();
1010       }
1011       _pending_jvms = NULL;
1012       if (pc_offset > C->debug_info()->last_pc_offset()) {
1013         // This is the only way _pending_jvms can become non-NULL:
1014         _pending_jvms = nn->jvms();
1015         _pending_offset = pc_offset;
1016       }
1017     }
1018   }
1019 
1020   // Stay out of the way of real safepoints:
1021   void observe_safepoint(JVMState* jvms, int pc_offset) {
1022     if (_pending_jvms != NULL &&
1023         !_pending_jvms->same_calls_as(jvms) &&
1024         _pending_offset < pc_offset) {
1025       emit_non_safepoint();
1026     }
1027     _pending_jvms = NULL;
1028   }
1029 
1030   void flush_at_end() {
1031     if (_pending_jvms != NULL) {
1032       emit_non_safepoint();
1033     }
1034     _pending_jvms = NULL;
1035   }
1036 };
1037 
1038 void NonSafepointEmitter::emit_non_safepoint() {
1039   JVMState* youngest_jvms = _pending_jvms;
1040   int       pc_offset     = _pending_offset;
1041 
1042   // Clear it now:
1043   _pending_jvms = NULL;
1044 
1045   DebugInformationRecorder* debug_info = C->debug_info();
1046   assert(debug_info->recording_non_safepoints(), "sanity");
1047 
1048   debug_info->add_non_safepoint(pc_offset);
1049   int max_depth = youngest_jvms->depth();
1050 
1051   // Visit scopes from oldest to youngest.
1052   for (int depth = 1; depth <= max_depth; depth++) {
1053     JVMState* jvms = youngest_jvms->of_depth(depth);
1054     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1055     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1056     debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1057   }
1058 
1059   // Mark the end of the scope set.
1060   debug_info->end_non_safepoint(pc_offset);
1061 }
1062 
1063 //------------------------------init_buffer------------------------------------
1064 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1065 
1066   // Set the initially allocated size
1067   int  code_req   = initial_code_capacity;
1068   int  locs_req   = initial_locs_capacity;
1069   int  stub_req   = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1070   int  const_req  = initial_const_capacity;
1071 
1072   int  pad_req    = NativeCall::instruction_size;
1073   // The extra spacing after the code is necessary on some platforms.
1074   // Sometimes we need to patch in a jump after the last instruction,
1075   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1076 
1077   // Compute the byte offset where we can store the deopt pc.
1078   if (fixed_slots() != 0) {
1079     _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1080   }
1081 
1082   // Compute prolog code size
1083   _method_size = 0;
1084   _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1085 #if defined(IA64) && !defined(AIX)
1086   if (save_argument_registers()) {
1087     // 4815101: this is a stub with implicit and unknown precision fp args.
1088     // The usual spill mechanism can only generate stfd's in this case, which
1089     // doesn't work if the fp reg to spill contains a single-precision denorm.
1090     // Instead, we hack around the normal spill mechanism using stfspill's and
1091     // ldffill's in the MachProlog and MachEpilog emit methods.  We allocate
1092     // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1093     //
1094     // If we ever implement 16-byte 'registers' == stack slots, we can
1095     // get rid of this hack and have SpillCopy generate stfspill/ldffill
1096     // instead of stfd/stfs/ldfd/ldfs.
1097     _frame_slots += 8*(16/BytesPerInt);
1098   }
1099 #endif
1100   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1101 
1102   if (has_mach_constant_base_node()) {
1103     uint add_size = 0;
1104     // Fill the constant table.
1105     // Note:  This must happen before shorten_branches.
1106     for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1107       Block* b = _cfg->get_block(i);
1108 
1109       for (uint j = 0; j < b->number_of_nodes(); j++) {
1110         Node* n = b->get_node(j);
1111 
1112         // If the node is a MachConstantNode evaluate the constant
1113         // value section.
1114         if (n->is_MachConstant()) {
1115           MachConstantNode* machcon = n->as_MachConstant();
1116           machcon->eval_constant(C);
1117         } else if (n->is_Mach()) {
1118           // On Power there are more nodes that issue constants.
1119           add_size += (n->as_Mach()->ins_num_consts() * 8);
1120         }
1121       }
1122     }
1123 
1124     // Calculate the offsets of the constants and the size of the
1125     // constant table (including the padding to the next section).
1126     constant_table().calculate_offsets_and_size();
1127     const_req = constant_table().size() + add_size;
1128   }
1129 
1130   // Initialize the space for the BufferBlob used to find and verify
1131   // instruction size in MachNode::emit_size()
1132   init_scratch_buffer_blob(const_req);
1133   if (failing())  return NULL; // Out of memory
1134 
1135   // Pre-compute the length of blocks and replace
1136   // long branches with short if machine supports it.
1137   shorten_branches(blk_starts, code_req, locs_req, stub_req);
1138 
1139   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1140   // class HandlerImpl is platform-specific and defined in the *.ad files.
1141   int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1142   int deopt_handler_req     = HandlerImpl::size_deopt_handler()     + MAX_stubs_size; // add marginal slop for handler
1143   stub_req += MAX_stubs_size;   // ensure per-stub margin
1144   code_req += MAX_inst_size;    // ensure per-instruction margin
1145 
1146   if (StressCodeBuffers)
1147     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
1148 
1149   int total_req =
1150     const_req +
1151     code_req +
1152     pad_req +
1153     stub_req +
1154     exception_handler_req +
1155     deopt_handler_req;               // deopt handler
1156 
1157   if (has_method_handle_invokes())
1158     total_req += deopt_handler_req;  // deopt MH handler
1159 
1160   CodeBuffer* cb = code_buffer();
1161   cb->initialize(total_req, locs_req);
1162 
1163   // Have we run out of code space?
1164   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1165     C->record_failure("CodeCache is full");
1166     return NULL;
1167   }
1168   // Configure the code buffer.
1169   cb->initialize_consts_size(const_req);
1170   cb->initialize_stubs_size(stub_req);
1171   cb->initialize_oop_recorder(env()->oop_recorder());
1172 
1173   // fill in the nop array for bundling computations
1174   MachNode *_nop_list[Bundle::_nop_count];
1175   Bundle::initialize_nops(_nop_list, this);
1176 
1177   return cb;
1178 }
1179 
1180 //------------------------------fill_buffer------------------------------------
1181 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1182   // blk_starts[] contains offsets calculated during short branches processing,
1183   // offsets should not be increased during following steps.
1184 
1185   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1186   // of a loop. It is used to determine the padding for loop alignment.
1187   compute_loop_first_inst_sizes();
1188 
1189   // Create oopmap set.
1190   _oop_map_set = new OopMapSet();
1191 
1192   // !!!!! This preserves old handling of oopmaps for now
1193   debug_info()->set_oopmaps(_oop_map_set);
1194 
1195   uint nblocks  = _cfg->number_of_blocks();
1196   // Count and start of implicit null check instructions
1197   uint inct_cnt = 0;
1198   uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1199 
1200   // Count and start of calls
1201   uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1202 
1203   uint  return_offset = 0;
1204   int nop_size = (new (this) MachNopNode())->size(_regalloc);
1205 
1206   int previous_offset = 0;
1207   int current_offset  = 0;
1208   int last_call_offset = -1;
1209   int last_avoid_back_to_back_offset = -1;
1210 #ifdef ASSERT
1211   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1212   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1213   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
1214   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
1215 #endif
1216 
1217   // Create an array of unused labels, one for each basic block, if printing is enabled
1218 #ifndef PRODUCT
1219   int *node_offsets      = NULL;
1220   uint node_offset_limit = unique();
1221 
1222   if (print_assembly())
1223     node_offsets         = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1224 #endif
1225 
1226   NonSafepointEmitter non_safepoints(this);  // emit non-safepoints lazily
1227 
1228   // Emit the constant table.
1229   if (has_mach_constant_base_node()) {
1230     constant_table().emit(*cb);
1231   }
1232 
1233   // Create an array of labels, one for each basic block
1234   Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1235   for (uint i=0; i <= nblocks; i++) {
1236     blk_labels[i].init();
1237   }
1238 
1239   // ------------------
1240   // Now fill in the code buffer
1241   Node *delay_slot = NULL;
1242 
1243   for (uint i = 0; i < nblocks; i++) {
1244     Block* block = _cfg->get_block(i);
1245     Node* head = block->head();
1246 
1247     // If this block needs to start aligned (i.e, can be reached other
1248     // than by falling-thru from the previous block), then force the
1249     // start of a new bundle.
1250     if (Pipeline::requires_bundling() && starts_bundle(head)) {
1251       cb->flush_bundle(true);
1252     }
1253 
1254 #ifdef ASSERT
1255     if (!block->is_connector()) {
1256       stringStream st;
1257       block->dump_head(_cfg, &st);
1258       MacroAssembler(cb).block_comment(st.as_string());
1259     }
1260     jmp_target[i] = 0;
1261     jmp_offset[i] = 0;
1262     jmp_size[i]   = 0;
1263     jmp_rule[i]   = 0;
1264 #endif
1265     int blk_offset = current_offset;
1266 
1267     // Define the label at the beginning of the basic block
1268     MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1269 
1270     uint last_inst = block->number_of_nodes();
1271 
1272     // Emit block normally, except for last instruction.
1273     // Emit means "dump code bits into code buffer".
1274     for (uint j = 0; j<last_inst; j++) {
1275 
1276       // Get the node
1277       Node* n = block->get_node(j);
1278 
1279       // See if delay slots are supported
1280       if (valid_bundle_info(n) &&
1281           node_bundling(n)->used_in_unconditional_delay()) {
1282         assert(delay_slot == NULL, "no use of delay slot node");
1283         assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1284 
1285         delay_slot = n;
1286         continue;
1287       }
1288 
1289       // If this starts a new instruction group, then flush the current one
1290       // (but allow split bundles)
1291       if (Pipeline::requires_bundling() && starts_bundle(n))
1292         cb->flush_bundle(false);
1293 
1294       // The following logic is duplicated in the code ifdeffed for
1295       // ENABLE_ZAP_DEAD_LOCALS which appears above in this file.  It
1296       // should be factored out.  Or maybe dispersed to the nodes?
1297 
1298       // Special handling for SafePoint/Call Nodes
1299       bool is_mcall = false;
1300       if (n->is_Mach()) {
1301         MachNode *mach = n->as_Mach();
1302         is_mcall = n->is_MachCall();
1303         bool is_sfn = n->is_MachSafePoint();
1304 
1305         // If this requires all previous instructions be flushed, then do so
1306         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1307           cb->flush_bundle(true);
1308           current_offset = cb->insts_size();
1309         }
1310 
1311         // A padding may be needed again since a previous instruction
1312         // could be moved to delay slot.
1313 
1314         // align the instruction if necessary
1315         int padding = mach->compute_padding(current_offset);
1316         // Make sure safepoint node for polling is distinct from a call's
1317         // return by adding a nop if needed.
1318         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1319           padding = nop_size;
1320         }
1321         if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1322             current_offset == last_avoid_back_to_back_offset) {
1323           // Avoid back to back some instructions.
1324           padding = nop_size;
1325         }
1326 
1327         if(padding > 0) {
1328           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1329           int nops_cnt = padding / nop_size;
1330           MachNode *nop = new (this) MachNopNode(nops_cnt);
1331           block->insert_node(nop, j++);
1332           last_inst++;
1333           _cfg->map_node_to_block(nop, block);
1334           nop->emit(*cb, _regalloc);
1335           cb->flush_bundle(true);
1336           current_offset = cb->insts_size();
1337         }
1338 
1339         // Remember the start of the last call in a basic block
1340         if (is_mcall) {
1341           MachCallNode *mcall = mach->as_MachCall();
1342 
1343           // This destination address is NOT PC-relative
1344           mcall->method_set((intptr_t)mcall->entry_point());
1345 
1346           // Save the return address
1347           call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1348 
1349           if (mcall->is_MachCallLeaf()) {
1350             is_mcall = false;
1351             is_sfn = false;
1352           }
1353         }
1354 
1355         // sfn will be valid whenever mcall is valid now because of inheritance
1356         if (is_sfn || is_mcall) {
1357 
1358           // Handle special safepoint nodes for synchronization
1359           if (!is_mcall) {
1360             MachSafePointNode *sfn = mach->as_MachSafePoint();
1361             // !!!!! Stubs only need an oopmap right now, so bail out
1362             if (sfn->jvms()->method() == NULL) {
1363               // Write the oopmap directly to the code blob??!!
1364 #             ifdef ENABLE_ZAP_DEAD_LOCALS
1365               assert( !is_node_getting_a_safepoint(sfn),  "logic does not match; false positive");
1366 #             endif
1367               continue;
1368             }
1369           } // End synchronization
1370 
1371           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1372                                            current_offset);
1373           Process_OopMap_Node(mach, current_offset);
1374         } // End if safepoint
1375 
1376         // If this is a null check, then add the start of the previous instruction to the list
1377         else if( mach->is_MachNullCheck() ) {
1378           inct_starts[inct_cnt++] = previous_offset;
1379         }
1380 
1381         // If this is a branch, then fill in the label with the target BB's label
1382         else if (mach->is_MachBranch()) {
1383           // This requires the TRUE branch target be in succs[0]
1384           uint block_num = block->non_connector_successor(0)->_pre_order;
1385 
1386           // Try to replace long branch if delay slot is not used,
1387           // it is mostly for back branches since forward branch's
1388           // distance is not updated yet.
1389           bool delay_slot_is_used = valid_bundle_info(n) &&
1390                                     node_bundling(n)->use_unconditional_delay();
1391           if (!delay_slot_is_used && mach->may_be_short_branch()) {
1392            assert(delay_slot == NULL, "not expecting delay slot node");
1393            int br_size = n->size(_regalloc);
1394             int offset = blk_starts[block_num] - current_offset;
1395             if (block_num >= i) {
1396               // Current and following block's offset are not
1397               // finalized yet, adjust distance by the difference
1398               // between calculated and final offsets of current block.
1399               offset -= (blk_starts[i] - blk_offset);
1400             }
1401             // In the following code a nop could be inserted before
1402             // the branch which will increase the backward distance.
1403             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1404             if (needs_padding && offset <= 0)
1405               offset -= nop_size;
1406 
1407             if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1408               // We've got a winner.  Replace this branch.
1409               MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1410 
1411               // Update the jmp_size.
1412               int new_size = replacement->size(_regalloc);
1413               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1414               // Insert padding between avoid_back_to_back branches.
1415               if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1416                 MachNode *nop = new (this) MachNopNode();
1417                 block->insert_node(nop, j++);
1418                 _cfg->map_node_to_block(nop, block);
1419                 last_inst++;
1420                 nop->emit(*cb, _regalloc);
1421                 cb->flush_bundle(true);
1422                 current_offset = cb->insts_size();
1423               }
1424 #ifdef ASSERT
1425               jmp_target[i] = block_num;
1426               jmp_offset[i] = current_offset - blk_offset;
1427               jmp_size[i]   = new_size;
1428               jmp_rule[i]   = mach->rule();
1429 #endif
1430               block->map_node(replacement, j);
1431               mach->subsume_by(replacement, C);
1432               n    = replacement;
1433               mach = replacement;
1434             }
1435           }
1436           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1437         } else if (mach->ideal_Opcode() == Op_Jump) {
1438           for (uint h = 0; h < block->_num_succs; h++) {
1439             Block* succs_block = block->_succs[h];
1440             for (uint j = 1; j < succs_block->num_preds(); j++) {
1441               Node* jpn = succs_block->pred(j);
1442               if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1443                 uint block_num = succs_block->non_connector()->_pre_order;
1444                 Label *blkLabel = &blk_labels[block_num];
1445                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1446               }
1447             }
1448           }
1449         }
1450 #ifdef ASSERT
1451         // Check that oop-store precedes the card-mark
1452         else if (mach->ideal_Opcode() == Op_StoreCM) {
1453           uint storeCM_idx = j;
1454           int count = 0;
1455           for (uint prec = mach->req(); prec < mach->len(); prec++) {
1456             Node *oop_store = mach->in(prec);  // Precedence edge
1457             if (oop_store == NULL) continue;
1458             count++;
1459             uint i4;
1460             for (i4 = 0; i4 < last_inst; ++i4) {
1461               if (block->get_node(i4) == oop_store) {
1462                 break;
1463               }
1464             }
1465             // Note: This test can provide a false failure if other precedence
1466             // edges have been added to the storeCMNode.
1467             assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1468           }
1469           assert(count > 0, "storeCM expects at least one precedence edge");
1470         }
1471 #endif
1472         else if (!n->is_Proj()) {
1473           // Remember the beginning of the previous instruction, in case
1474           // it's followed by a flag-kill and a null-check.  Happens on
1475           // Intel all the time, with add-to-memory kind of opcodes.
1476           previous_offset = current_offset;
1477         }
1478 
1479         // Not an else-if!
1480         // If this is a trap based cmp then add its offset to the list.
1481         if (mach->is_TrapBasedCheckNode()) {
1482           inct_starts[inct_cnt++] = current_offset;
1483         }
1484       }
1485 
1486       // Verify that there is sufficient space remaining
1487       cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1488       if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1489         C->record_failure("CodeCache is full");
1490         return;
1491       }
1492 
1493       // Save the offset for the listing
1494 #ifndef PRODUCT
1495       if (node_offsets && n->_idx < node_offset_limit)
1496         node_offsets[n->_idx] = cb->insts_size();
1497 #endif
1498 
1499       // "Normal" instruction case
1500       DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1501       n->emit(*cb, _regalloc);
1502       current_offset  = cb->insts_size();
1503 
1504 #ifdef ASSERT
1505       if (n->size(_regalloc) < (current_offset-instr_offset)) {
1506         n->dump();
1507         assert(false, "wrong size of mach node");
1508       }
1509 #endif
1510       non_safepoints.observe_instruction(n, current_offset);
1511 
1512       // mcall is last "call" that can be a safepoint
1513       // record it so we can see if a poll will directly follow it
1514       // in which case we'll need a pad to make the PcDesc sites unique
1515       // see  5010568. This can be slightly inaccurate but conservative
1516       // in the case that return address is not actually at current_offset.
1517       // This is a small price to pay.
1518 
1519       if (is_mcall) {
1520         last_call_offset = current_offset;
1521       }
1522 
1523       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1524         // Avoid back to back some instructions.
1525         last_avoid_back_to_back_offset = current_offset;
1526       }
1527 
1528       // See if this instruction has a delay slot
1529       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1530         assert(delay_slot != NULL, "expecting delay slot node");
1531 
1532         // Back up 1 instruction
1533         cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1534 
1535         // Save the offset for the listing
1536 #ifndef PRODUCT
1537         if (node_offsets && delay_slot->_idx < node_offset_limit)
1538           node_offsets[delay_slot->_idx] = cb->insts_size();
1539 #endif
1540 
1541         // Support a SafePoint in the delay slot
1542         if (delay_slot->is_MachSafePoint()) {
1543           MachNode *mach = delay_slot->as_Mach();
1544           // !!!!! Stubs only need an oopmap right now, so bail out
1545           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1546             // Write the oopmap directly to the code blob??!!
1547 #           ifdef ENABLE_ZAP_DEAD_LOCALS
1548             assert( !is_node_getting_a_safepoint(mach),  "logic does not match; false positive");
1549 #           endif
1550             delay_slot = NULL;
1551             continue;
1552           }
1553 
1554           int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1555           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1556                                            adjusted_offset);
1557           // Generate an OopMap entry
1558           Process_OopMap_Node(mach, adjusted_offset);
1559         }
1560 
1561         // Insert the delay slot instruction
1562         delay_slot->emit(*cb, _regalloc);
1563 
1564         // Don't reuse it
1565         delay_slot = NULL;
1566       }
1567 
1568     } // End for all instructions in block
1569 
1570     // If the next block is the top of a loop, pad this block out to align
1571     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1572     if (i < nblocks-1) {
1573       Block *nb = _cfg->get_block(i + 1);
1574       int padding = nb->alignment_padding(current_offset);
1575       if( padding > 0 ) {
1576         MachNode *nop = new (this) MachNopNode(padding / nop_size);
1577         block->insert_node(nop, block->number_of_nodes());
1578         _cfg->map_node_to_block(nop, block);
1579         nop->emit(*cb, _regalloc);
1580         current_offset = cb->insts_size();
1581       }
1582     }
1583     // Verify that the distance for generated before forward
1584     // short branches is still valid.
1585     guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1586 
1587     // Save new block start offset
1588     blk_starts[i] = blk_offset;
1589   } // End of for all blocks
1590   blk_starts[nblocks] = current_offset;
1591 
1592   non_safepoints.flush_at_end();
1593 
1594   // Offset too large?
1595   if (failing())  return;
1596 
1597   // Define a pseudo-label at the end of the code
1598   MacroAssembler(cb).bind( blk_labels[nblocks] );
1599 
1600   // Compute the size of the first block
1601   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1602 
1603   assert(cb->insts_size() < 500000, "method is unreasonably large");
1604 
1605 #ifdef ASSERT
1606   for (uint i = 0; i < nblocks; i++) { // For all blocks
1607     if (jmp_target[i] != 0) {
1608       int br_size = jmp_size[i];
1609       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1610       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1611         tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
1612         assert(false, "Displacement too large for short jmp");
1613       }
1614     }
1615   }
1616 #endif
1617 
1618 #ifndef PRODUCT
1619   // Information on the size of the method, without the extraneous code
1620   Scheduling::increment_method_size(cb->insts_size());
1621 #endif
1622 
1623   // ------------------
1624   // Fill in exception table entries.
1625   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1626 
1627   // Only java methods have exception handlers and deopt handlers
1628   // class HandlerImpl is platform-specific and defined in the *.ad files.
1629   if (_method) {
1630     // Emit the exception handler code.
1631     _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1632     // Emit the deopt handler code.
1633     _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1634 
1635     // Emit the MethodHandle deopt handler code (if required).
1636     if (has_method_handle_invokes()) {
1637       // We can use the same code as for the normal deopt handler, we
1638       // just need a different entry point address.
1639       _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1640     }
1641   }
1642 
1643   // One last check for failed CodeBuffer::expand:
1644   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1645     C->record_failure("CodeCache is full");
1646     return;
1647   }
1648 
1649 #ifndef PRODUCT
1650   // Dump the assembly code, including basic-block numbers
1651   if (print_assembly()) {
1652     ttyLocker ttyl;  // keep the following output all in one block
1653     if (!VMThread::should_terminate()) {  // test this under the tty lock
1654       // This output goes directly to the tty, not the compiler log.
1655       // To enable tools to match it up with the compilation activity,
1656       // be sure to tag this tty output with the compile ID.
1657       if (xtty != NULL) {
1658         xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1659                    is_osr_compilation()    ? " compile_kind='osr'" :
1660                    "");
1661       }
1662       if (method() != NULL) {
1663         method()->print_metadata();
1664       }
1665       dump_asm(node_offsets, node_offset_limit);
1666       if (xtty != NULL) {
1667         xtty->tail("opto_assembly");
1668       }
1669     }
1670   }
1671 #endif
1672 
1673 }
1674 
1675 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1676   _inc_table.set_size(cnt);
1677 
1678   uint inct_cnt = 0;
1679   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1680     Block* block = _cfg->get_block(i);
1681     Node *n = NULL;
1682     int j;
1683 
1684     // Find the branch; ignore trailing NOPs.
1685     for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1686       n = block->get_node(j);
1687       if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1688         break;
1689       }
1690     }
1691 
1692     // If we didn't find anything, continue
1693     if (j < 0) {
1694       continue;
1695     }
1696 
1697     // Compute ExceptionHandlerTable subtable entry and add it
1698     // (skip empty blocks)
1699     if (n->is_Catch()) {
1700 
1701       // Get the offset of the return from the call
1702       uint call_return = call_returns[block->_pre_order];
1703 #ifdef ASSERT
1704       assert( call_return > 0, "no call seen for this basic block" );
1705       while (block->get_node(--j)->is_MachProj()) ;
1706       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1707 #endif
1708       // last instruction is a CatchNode, find it's CatchProjNodes
1709       int nof_succs = block->_num_succs;
1710       // allocate space
1711       GrowableArray<intptr_t> handler_bcis(nof_succs);
1712       GrowableArray<intptr_t> handler_pcos(nof_succs);
1713       // iterate through all successors
1714       for (int j = 0; j < nof_succs; j++) {
1715         Block* s = block->_succs[j];
1716         bool found_p = false;
1717         for (uint k = 1; k < s->num_preds(); k++) {
1718           Node* pk = s->pred(k);
1719           if (pk->is_CatchProj() && pk->in(0) == n) {
1720             const CatchProjNode* p = pk->as_CatchProj();
1721             found_p = true;
1722             // add the corresponding handler bci & pco information
1723             if (p->_con != CatchProjNode::fall_through_index) {
1724               // p leads to an exception handler (and is not fall through)
1725               assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1726               // no duplicates, please
1727               if (!handler_bcis.contains(p->handler_bci())) {
1728                 uint block_num = s->non_connector()->_pre_order;
1729                 handler_bcis.append(p->handler_bci());
1730                 handler_pcos.append(blk_labels[block_num].loc_pos());
1731               }
1732             }
1733           }
1734         }
1735         assert(found_p, "no matching predecessor found");
1736         // Note:  Due to empty block removal, one block may have
1737         // several CatchProj inputs, from the same Catch.
1738       }
1739 
1740       // Set the offset of the return from the call
1741       _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1742       continue;
1743     }
1744 
1745     // Handle implicit null exception table updates
1746     if (n->is_MachNullCheck()) {
1747       uint block_num = block->non_connector_successor(0)->_pre_order;
1748       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1749       continue;
1750     }
1751     // Handle implicit exception table updates: trap instructions.
1752     if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1753       uint block_num = block->non_connector_successor(0)->_pre_order;
1754       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1755       continue;
1756     }
1757   } // End of for all blocks fill in exception table entries
1758 }
1759 
1760 // Static Variables
1761 #ifndef PRODUCT
1762 uint Scheduling::_total_nop_size = 0;
1763 uint Scheduling::_total_method_size = 0;
1764 uint Scheduling::_total_branches = 0;
1765 uint Scheduling::_total_unconditional_delays = 0;
1766 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1767 #endif
1768 
1769 // Initializer for class Scheduling
1770 
1771 Scheduling::Scheduling(Arena *arena, Compile &compile)
1772   : _arena(arena),
1773     _cfg(compile.cfg()),
1774     _regalloc(compile.regalloc()),
1775     _reg_node(arena),
1776     _bundle_instr_count(0),
1777     _bundle_cycle_number(0),
1778     _scheduled(arena),
1779     _available(arena),
1780     _next_node(NULL),
1781     _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1782     _pinch_free_list(arena)
1783 #ifndef PRODUCT
1784   , _branches(0)
1785   , _unconditional_delays(0)
1786 #endif
1787 {
1788   // Create a MachNopNode
1789   _nop = new (&compile) MachNopNode();
1790 
1791   // Now that the nops are in the array, save the count
1792   // (but allow entries for the nops)
1793   _node_bundling_limit = compile.unique();
1794   uint node_max = _regalloc->node_regs_max_index();
1795 
1796   compile.set_node_bundling_limit(_node_bundling_limit);
1797 
1798   // This one is persistent within the Compile class
1799   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1800 
1801   // Allocate space for fixed-size arrays
1802   _node_latency    = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1803   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
1804   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1805 
1806   // Clear the arrays
1807   memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1808   memset(_node_latency,       0, node_max * sizeof(unsigned short));
1809   memset(_uses,               0, node_max * sizeof(short));
1810   memset(_current_latency,    0, node_max * sizeof(unsigned short));
1811 
1812   // Clear the bundling information
1813   memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1814 
1815   // Get the last node
1816   Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1817 
1818   _next_node = block->get_node(block->number_of_nodes() - 1);
1819 }
1820 
1821 #ifndef PRODUCT
1822 // Scheduling destructor
1823 Scheduling::~Scheduling() {
1824   _total_branches             += _branches;
1825   _total_unconditional_delays += _unconditional_delays;
1826 }
1827 #endif
1828 
1829 // Step ahead "i" cycles
1830 void Scheduling::step(uint i) {
1831 
1832   Bundle *bundle = node_bundling(_next_node);
1833   bundle->set_starts_bundle();
1834 
1835   // Update the bundle record, but leave the flags information alone
1836   if (_bundle_instr_count > 0) {
1837     bundle->set_instr_count(_bundle_instr_count);
1838     bundle->set_resources_used(_bundle_use.resourcesUsed());
1839   }
1840 
1841   // Update the state information
1842   _bundle_instr_count = 0;
1843   _bundle_cycle_number += i;
1844   _bundle_use.step(i);
1845 }
1846 
1847 void Scheduling::step_and_clear() {
1848   Bundle *bundle = node_bundling(_next_node);
1849   bundle->set_starts_bundle();
1850 
1851   // Update the bundle record
1852   if (_bundle_instr_count > 0) {
1853     bundle->set_instr_count(_bundle_instr_count);
1854     bundle->set_resources_used(_bundle_use.resourcesUsed());
1855 
1856     _bundle_cycle_number += 1;
1857   }
1858 
1859   // Clear the bundling information
1860   _bundle_instr_count = 0;
1861   _bundle_use.reset();
1862 
1863   memcpy(_bundle_use_elements,
1864     Pipeline_Use::elaborated_elements,
1865     sizeof(Pipeline_Use::elaborated_elements));
1866 }
1867 
1868 // Perform instruction scheduling and bundling over the sequence of
1869 // instructions in backwards order.
1870 void Compile::ScheduleAndBundle() {
1871 
1872   // Don't optimize this if it isn't a method
1873   if (!_method)
1874     return;
1875 
1876   // Don't optimize this if scheduling is disabled
1877   if (!do_scheduling())
1878     return;
1879 
1880   // Scheduling code works only with pairs (8 bytes) maximum.
1881   if (max_vector_size() > 8)
1882     return;
1883 
1884   NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1885 
1886   // Create a data structure for all the scheduling information
1887   Scheduling scheduling(Thread::current()->resource_area(), *this);
1888 
1889   // Walk backwards over each basic block, computing the needed alignment
1890   // Walk over all the basic blocks
1891   scheduling.DoScheduling();
1892 }
1893 
1894 // Compute the latency of all the instructions.  This is fairly simple,
1895 // because we already have a legal ordering.  Walk over the instructions
1896 // from first to last, and compute the latency of the instruction based
1897 // on the latency of the preceding instruction(s).
1898 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1899 #ifndef PRODUCT
1900   if (_cfg->C->trace_opto_output())
1901     tty->print("# -> ComputeLocalLatenciesForward\n");
1902 #endif
1903 
1904   // Walk over all the schedulable instructions
1905   for( uint j=_bb_start; j < _bb_end; j++ ) {
1906 
1907     // This is a kludge, forcing all latency calculations to start at 1.
1908     // Used to allow latency 0 to force an instruction to the beginning
1909     // of the bb
1910     uint latency = 1;
1911     Node *use = bb->get_node(j);
1912     uint nlen = use->len();
1913 
1914     // Walk over all the inputs
1915     for ( uint k=0; k < nlen; k++ ) {
1916       Node *def = use->in(k);
1917       if (!def)
1918         continue;
1919 
1920       uint l = _node_latency[def->_idx] + use->latency(k);
1921       if (latency < l)
1922         latency = l;
1923     }
1924 
1925     _node_latency[use->_idx] = latency;
1926 
1927 #ifndef PRODUCT
1928     if (_cfg->C->trace_opto_output()) {
1929       tty->print("# latency %4d: ", latency);
1930       use->dump();
1931     }
1932 #endif
1933   }
1934 
1935 #ifndef PRODUCT
1936   if (_cfg->C->trace_opto_output())
1937     tty->print("# <- ComputeLocalLatenciesForward\n");
1938 #endif
1939 
1940 } // end ComputeLocalLatenciesForward
1941 
1942 // See if this node fits into the present instruction bundle
1943 bool Scheduling::NodeFitsInBundle(Node *n) {
1944   uint n_idx = n->_idx;
1945 
1946   // If this is the unconditional delay instruction, then it fits
1947   if (n == _unconditional_delay_slot) {
1948 #ifndef PRODUCT
1949     if (_cfg->C->trace_opto_output())
1950       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1951 #endif
1952     return (true);
1953   }
1954 
1955   // If the node cannot be scheduled this cycle, skip it
1956   if (_current_latency[n_idx] > _bundle_cycle_number) {
1957 #ifndef PRODUCT
1958     if (_cfg->C->trace_opto_output())
1959       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1960         n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1961 #endif
1962     return (false);
1963   }
1964 
1965   const Pipeline *node_pipeline = n->pipeline();
1966 
1967   uint instruction_count = node_pipeline->instructionCount();
1968   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1969     instruction_count = 0;
1970   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1971     instruction_count++;
1972 
1973   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1974 #ifndef PRODUCT
1975     if (_cfg->C->trace_opto_output())
1976       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1977         n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1978 #endif
1979     return (false);
1980   }
1981 
1982   // Don't allow non-machine nodes to be handled this way
1983   if (!n->is_Mach() && instruction_count == 0)
1984     return (false);
1985 
1986   // See if there is any overlap
1987   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1988 
1989   if (delay > 0) {
1990 #ifndef PRODUCT
1991     if (_cfg->C->trace_opto_output())
1992       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1993 #endif
1994     return false;
1995   }
1996 
1997 #ifndef PRODUCT
1998   if (_cfg->C->trace_opto_output())
1999     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2000 #endif
2001 
2002   return true;
2003 }
2004 
2005 Node * Scheduling::ChooseNodeToBundle() {
2006   uint siz = _available.size();
2007 
2008   if (siz == 0) {
2009 
2010 #ifndef PRODUCT
2011     if (_cfg->C->trace_opto_output())
2012       tty->print("#   ChooseNodeToBundle: NULL\n");
2013 #endif
2014     return (NULL);
2015   }
2016 
2017   // Fast path, if only 1 instruction in the bundle
2018   if (siz == 1) {
2019 #ifndef PRODUCT
2020     if (_cfg->C->trace_opto_output()) {
2021       tty->print("#   ChooseNodeToBundle (only 1): ");
2022       _available[0]->dump();
2023     }
2024 #endif
2025     return (_available[0]);
2026   }
2027 
2028   // Don't bother, if the bundle is already full
2029   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2030     for ( uint i = 0; i < siz; i++ ) {
2031       Node *n = _available[i];
2032 
2033       // Skip projections, we'll handle them another way
2034       if (n->is_Proj())
2035         continue;
2036 
2037       // This presupposed that instructions are inserted into the
2038       // available list in a legality order; i.e. instructions that
2039       // must be inserted first are at the head of the list
2040       if (NodeFitsInBundle(n)) {
2041 #ifndef PRODUCT
2042         if (_cfg->C->trace_opto_output()) {
2043           tty->print("#   ChooseNodeToBundle: ");
2044           n->dump();
2045         }
2046 #endif
2047         return (n);
2048       }
2049     }
2050   }
2051 
2052   // Nothing fits in this bundle, choose the highest priority
2053 #ifndef PRODUCT
2054   if (_cfg->C->trace_opto_output()) {
2055     tty->print("#   ChooseNodeToBundle: ");
2056     _available[0]->dump();
2057   }
2058 #endif
2059 
2060   return _available[0];
2061 }
2062 
2063 void Scheduling::AddNodeToAvailableList(Node *n) {
2064   assert( !n->is_Proj(), "projections never directly made available" );
2065 #ifndef PRODUCT
2066   if (_cfg->C->trace_opto_output()) {
2067     tty->print("#   AddNodeToAvailableList: ");
2068     n->dump();
2069   }
2070 #endif
2071 
2072   int latency = _current_latency[n->_idx];
2073 
2074   // Insert in latency order (insertion sort)
2075   uint i;
2076   for ( i=0; i < _available.size(); i++ )
2077     if (_current_latency[_available[i]->_idx] > latency)
2078       break;
2079 
2080   // Special Check for compares following branches
2081   if( n->is_Mach() && _scheduled.size() > 0 ) {
2082     int op = n->as_Mach()->ideal_Opcode();
2083     Node *last = _scheduled[0];
2084     if( last->is_MachIf() && last->in(1) == n &&
2085         ( op == Op_CmpI ||
2086           op == Op_CmpU ||
2087           op == Op_CmpP ||
2088           op == Op_CmpF ||
2089           op == Op_CmpD ||
2090           op == Op_CmpL ) ) {
2091 
2092       // Recalculate position, moving to front of same latency
2093       for ( i=0 ; i < _available.size(); i++ )
2094         if (_current_latency[_available[i]->_idx] >= latency)
2095           break;
2096     }
2097   }
2098 
2099   // Insert the node in the available list
2100   _available.insert(i, n);
2101 
2102 #ifndef PRODUCT
2103   if (_cfg->C->trace_opto_output())
2104     dump_available();
2105 #endif
2106 }
2107 
2108 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2109   for ( uint i=0; i < n->len(); i++ ) {
2110     Node *def = n->in(i);
2111     if (!def) continue;
2112     if( def->is_Proj() )        // If this is a machine projection, then
2113       def = def->in(0);         // propagate usage thru to the base instruction
2114 
2115     if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2116       continue;
2117     }
2118 
2119     // Compute the latency
2120     uint l = _bundle_cycle_number + n->latency(i);
2121     if (_current_latency[def->_idx] < l)
2122       _current_latency[def->_idx] = l;
2123 
2124     // If this does not have uses then schedule it
2125     if ((--_uses[def->_idx]) == 0)
2126       AddNodeToAvailableList(def);
2127   }
2128 }
2129 
2130 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2131 #ifndef PRODUCT
2132   if (_cfg->C->trace_opto_output()) {
2133     tty->print("#   AddNodeToBundle: ");
2134     n->dump();
2135   }
2136 #endif
2137 
2138   // Remove this from the available list
2139   uint i;
2140   for (i = 0; i < _available.size(); i++)
2141     if (_available[i] == n)
2142       break;
2143   assert(i < _available.size(), "entry in _available list not found");
2144   _available.remove(i);
2145 
2146   // See if this fits in the current bundle
2147   const Pipeline *node_pipeline = n->pipeline();
2148   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2149 
2150   // Check for instructions to be placed in the delay slot. We
2151   // do this before we actually schedule the current instruction,
2152   // because the delay slot follows the current instruction.
2153   if (Pipeline::_branch_has_delay_slot &&
2154       node_pipeline->hasBranchDelay() &&
2155       !_unconditional_delay_slot) {
2156 
2157     uint siz = _available.size();
2158 
2159     // Conditional branches can support an instruction that
2160     // is unconditionally executed and not dependent by the
2161     // branch, OR a conditionally executed instruction if
2162     // the branch is taken.  In practice, this means that
2163     // the first instruction at the branch target is
2164     // copied to the delay slot, and the branch goes to
2165     // the instruction after that at the branch target
2166     if ( n->is_MachBranch() ) {
2167 
2168       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2169       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );
2170 
2171 #ifndef PRODUCT
2172       _branches++;
2173 #endif
2174 
2175       // At least 1 instruction is on the available list
2176       // that is not dependent on the branch
2177       for (uint i = 0; i < siz; i++) {
2178         Node *d = _available[i];
2179         const Pipeline *avail_pipeline = d->pipeline();
2180 
2181         // Don't allow safepoints in the branch shadow, that will
2182         // cause a number of difficulties
2183         if ( avail_pipeline->instructionCount() == 1 &&
2184             !avail_pipeline->hasMultipleBundles() &&
2185             !avail_pipeline->hasBranchDelay() &&
2186             Pipeline::instr_has_unit_size() &&
2187             d->size(_regalloc) == Pipeline::instr_unit_size() &&
2188             NodeFitsInBundle(d) &&
2189             !node_bundling(d)->used_in_delay()) {
2190 
2191           if (d->is_Mach() && !d->is_MachSafePoint()) {
2192             // A node that fits in the delay slot was found, so we need to
2193             // set the appropriate bits in the bundle pipeline information so
2194             // that it correctly indicates resource usage.  Later, when we
2195             // attempt to add this instruction to the bundle, we will skip
2196             // setting the resource usage.
2197             _unconditional_delay_slot = d;
2198             node_bundling(n)->set_use_unconditional_delay();
2199             node_bundling(d)->set_used_in_unconditional_delay();
2200             _bundle_use.add_usage(avail_pipeline->resourceUse());
2201             _current_latency[d->_idx] = _bundle_cycle_number;
2202             _next_node = d;
2203             ++_bundle_instr_count;
2204 #ifndef PRODUCT
2205             _unconditional_delays++;
2206 #endif
2207             break;
2208           }
2209         }
2210       }
2211     }
2212 
2213     // No delay slot, add a nop to the usage
2214     if (!_unconditional_delay_slot) {
2215       // See if adding an instruction in the delay slot will overflow
2216       // the bundle.
2217       if (!NodeFitsInBundle(_nop)) {
2218 #ifndef PRODUCT
2219         if (_cfg->C->trace_opto_output())
2220           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");
2221 #endif
2222         step(1);
2223       }
2224 
2225       _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2226       _next_node = _nop;
2227       ++_bundle_instr_count;
2228     }
2229 
2230     // See if the instruction in the delay slot requires a
2231     // step of the bundles
2232     if (!NodeFitsInBundle(n)) {
2233 #ifndef PRODUCT
2234         if (_cfg->C->trace_opto_output())
2235           tty->print("#  *** STEP(branch won't fit) ***\n");
2236 #endif
2237         // Update the state information
2238         _bundle_instr_count = 0;
2239         _bundle_cycle_number += 1;
2240         _bundle_use.step(1);
2241     }
2242   }
2243 
2244   // Get the number of instructions
2245   uint instruction_count = node_pipeline->instructionCount();
2246   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2247     instruction_count = 0;
2248 
2249   // Compute the latency information
2250   uint delay = 0;
2251 
2252   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2253     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2254     if (relative_latency < 0)
2255       relative_latency = 0;
2256 
2257     delay = _bundle_use.full_latency(relative_latency, node_usage);
2258 
2259     // Does not fit in this bundle, start a new one
2260     if (delay > 0) {
2261       step(delay);
2262 
2263 #ifndef PRODUCT
2264       if (_cfg->C->trace_opto_output())
2265         tty->print("#  *** STEP(%d) ***\n", delay);
2266 #endif
2267     }
2268   }
2269 
2270   // If this was placed in the delay slot, ignore it
2271   if (n != _unconditional_delay_slot) {
2272 
2273     if (delay == 0) {
2274       if (node_pipeline->hasMultipleBundles()) {
2275 #ifndef PRODUCT
2276         if (_cfg->C->trace_opto_output())
2277           tty->print("#  *** STEP(multiple instructions) ***\n");
2278 #endif
2279         step(1);
2280       }
2281 
2282       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2283 #ifndef PRODUCT
2284         if (_cfg->C->trace_opto_output())
2285           tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2286             instruction_count + _bundle_instr_count,
2287             Pipeline::_max_instrs_per_cycle);
2288 #endif
2289         step(1);
2290       }
2291     }
2292 
2293     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2294       _bundle_instr_count++;
2295 
2296     // Set the node's latency
2297     _current_latency[n->_idx] = _bundle_cycle_number;
2298 
2299     // Now merge the functional unit information
2300     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2301       _bundle_use.add_usage(node_usage);
2302 
2303     // Increment the number of instructions in this bundle
2304     _bundle_instr_count += instruction_count;
2305 
2306     // Remember this node for later
2307     if (n->is_Mach())
2308       _next_node = n;
2309   }
2310 
2311   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2312   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2313   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2314   // into the block.  All other scheduled nodes get put in the schedule here.
2315   int op = n->Opcode();
2316   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2317       (op != Op_Node &&         // Not an unused antidepedence node and
2318        // not an unallocated boxlock
2319        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2320 
2321     // Push any trailing projections
2322     if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2323       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2324         Node *foi = n->fast_out(i);
2325         if( foi->is_Proj() )
2326           _scheduled.push(foi);
2327       }
2328     }
2329 
2330     // Put the instruction in the schedule list
2331     _scheduled.push(n);
2332   }
2333 
2334 #ifndef PRODUCT
2335   if (_cfg->C->trace_opto_output())
2336     dump_available();
2337 #endif
2338 
2339   // Walk all the definitions, decrementing use counts, and
2340   // if a definition has a 0 use count, place it in the available list.
2341   DecrementUseCounts(n,bb);
2342 }
2343 
2344 // This method sets the use count within a basic block.  We will ignore all
2345 // uses outside the current basic block.  As we are doing a backwards walk,
2346 // any node we reach that has a use count of 0 may be scheduled.  This also
2347 // avoids the problem of cyclic references from phi nodes, as long as phi
2348 // nodes are at the front of the basic block.  This method also initializes
2349 // the available list to the set of instructions that have no uses within this
2350 // basic block.
2351 void Scheduling::ComputeUseCount(const Block *bb) {
2352 #ifndef PRODUCT
2353   if (_cfg->C->trace_opto_output())
2354     tty->print("# -> ComputeUseCount\n");
2355 #endif
2356 
2357   // Clear the list of available and scheduled instructions, just in case
2358   _available.clear();
2359   _scheduled.clear();
2360 
2361   // No delay slot specified
2362   _unconditional_delay_slot = NULL;
2363 
2364 #ifdef ASSERT
2365   for( uint i=0; i < bb->number_of_nodes(); i++ )
2366     assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2367 #endif
2368 
2369   // Force the _uses count to never go to zero for unscheduable pieces
2370   // of the block
2371   for( uint k = 0; k < _bb_start; k++ )
2372     _uses[bb->get_node(k)->_idx] = 1;
2373   for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2374     _uses[bb->get_node(l)->_idx] = 1;
2375 
2376   // Iterate backwards over the instructions in the block.  Don't count the
2377   // branch projections at end or the block header instructions.
2378   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2379     Node *n = bb->get_node(j);
2380     if( n->is_Proj() ) continue; // Projections handled another way
2381 
2382     // Account for all uses
2383     for ( uint k = 0; k < n->len(); k++ ) {
2384       Node *inp = n->in(k);
2385       if (!inp) continue;
2386       assert(inp != n, "no cycles allowed" );
2387       if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2388         if (inp->is_Proj()) { // Skip through Proj's
2389           inp = inp->in(0);
2390         }
2391         ++_uses[inp->_idx];     // Count 1 block-local use
2392       }
2393     }
2394 
2395     // If this instruction has a 0 use count, then it is available
2396     if (!_uses[n->_idx]) {
2397       _current_latency[n->_idx] = _bundle_cycle_number;
2398       AddNodeToAvailableList(n);
2399     }
2400 
2401 #ifndef PRODUCT
2402     if (_cfg->C->trace_opto_output()) {
2403       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2404       n->dump();
2405     }
2406 #endif
2407   }
2408 
2409 #ifndef PRODUCT
2410   if (_cfg->C->trace_opto_output())
2411     tty->print("# <- ComputeUseCount\n");
2412 #endif
2413 }
2414 
2415 // This routine performs scheduling on each basic block in reverse order,
2416 // using instruction latencies and taking into account function unit
2417 // availability.
2418 void Scheduling::DoScheduling() {
2419 #ifndef PRODUCT
2420   if (_cfg->C->trace_opto_output())
2421     tty->print("# -> DoScheduling\n");
2422 #endif
2423 
2424   Block *succ_bb = NULL;
2425   Block *bb;
2426 
2427   // Walk over all the basic blocks in reverse order
2428   for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2429     bb = _cfg->get_block(i);
2430 
2431 #ifndef PRODUCT
2432     if (_cfg->C->trace_opto_output()) {
2433       tty->print("#  Schedule BB#%03d (initial)\n", i);
2434       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2435         bb->get_node(j)->dump();
2436       }
2437     }
2438 #endif
2439 
2440     // On the head node, skip processing
2441     if (bb == _cfg->get_root_block()) {
2442       continue;
2443     }
2444 
2445     // Skip empty, connector blocks
2446     if (bb->is_connector())
2447       continue;
2448 
2449     // If the following block is not the sole successor of
2450     // this one, then reset the pipeline information
2451     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2452 #ifndef PRODUCT
2453       if (_cfg->C->trace_opto_output()) {
2454         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2455                    _next_node->_idx, _bundle_instr_count);
2456       }
2457 #endif
2458       step_and_clear();
2459     }
2460 
2461     // Leave untouched the starting instruction, any Phis, a CreateEx node
2462     // or Top.  bb->get_node(_bb_start) is the first schedulable instruction.
2463     _bb_end = bb->number_of_nodes()-1;
2464     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2465       Node *n = bb->get_node(_bb_start);
2466       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2467       // Also, MachIdealNodes do not get scheduled
2468       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2469       MachNode *mach = n->as_Mach();
2470       int iop = mach->ideal_Opcode();
2471       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2472       if( iop == Op_Con ) continue;      // Do not schedule Top
2473       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2474           mach->pipeline() == MachNode::pipeline_class() &&
2475           !n->is_SpillCopy() )  // Breakpoints, Prolog, etc
2476         continue;
2477       break;                    // Funny loop structure to be sure...
2478     }
2479     // Compute last "interesting" instruction in block - last instruction we
2480     // might schedule.  _bb_end points just after last schedulable inst.  We
2481     // normally schedule conditional branches (despite them being forced last
2482     // in the block), because they have delay slots we can fill.  Calls all
2483     // have their delay slots filled in the template expansions, so we don't
2484     // bother scheduling them.
2485     Node *last = bb->get_node(_bb_end);
2486     // Ignore trailing NOPs.
2487     while (_bb_end > 0 && last->is_Mach() &&
2488            last->as_Mach()->ideal_Opcode() == Op_Con) {
2489       last = bb->get_node(--_bb_end);
2490     }
2491     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2492     if( last->is_Catch() ||
2493        // Exclude unreachable path case when Halt node is in a separate block.
2494        (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2495       // There must be a prior call.  Skip it.
2496       while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2497         assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2498       }
2499     } else if( last->is_MachNullCheck() ) {
2500       // Backup so the last null-checked memory instruction is
2501       // outside the schedulable range. Skip over the nullcheck,
2502       // projection, and the memory nodes.
2503       Node *mem = last->in(1);
2504       do {
2505         _bb_end--;
2506       } while (mem != bb->get_node(_bb_end));
2507     } else {
2508       // Set _bb_end to point after last schedulable inst.
2509       _bb_end++;
2510     }
2511 
2512     assert( _bb_start <= _bb_end, "inverted block ends" );
2513 
2514     // Compute the register antidependencies for the basic block
2515     ComputeRegisterAntidependencies(bb);
2516     if (_cfg->C->failing())  return;  // too many D-U pinch points
2517 
2518     // Compute intra-bb latencies for the nodes
2519     ComputeLocalLatenciesForward(bb);
2520 
2521     // Compute the usage within the block, and set the list of all nodes
2522     // in the block that have no uses within the block.
2523     ComputeUseCount(bb);
2524 
2525     // Schedule the remaining instructions in the block
2526     while ( _available.size() > 0 ) {
2527       Node *n = ChooseNodeToBundle();
2528       guarantee(n != NULL, "no nodes available");
2529       AddNodeToBundle(n,bb);
2530     }
2531 
2532     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2533 #ifdef ASSERT
2534     for( uint l = _bb_start; l < _bb_end; l++ ) {
2535       Node *n = bb->get_node(l);
2536       uint m;
2537       for( m = 0; m < _bb_end-_bb_start; m++ )
2538         if( _scheduled[m] == n )
2539           break;
2540       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2541     }
2542 #endif
2543 
2544     // Now copy the instructions (in reverse order) back to the block
2545     for ( uint k = _bb_start; k < _bb_end; k++ )
2546       bb->map_node(_scheduled[_bb_end-k-1], k);
2547 
2548 #ifndef PRODUCT
2549     if (_cfg->C->trace_opto_output()) {
2550       tty->print("#  Schedule BB#%03d (final)\n", i);
2551       uint current = 0;
2552       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2553         Node *n = bb->get_node(j);
2554         if( valid_bundle_info(n) ) {
2555           Bundle *bundle = node_bundling(n);
2556           if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2557             tty->print("*** Bundle: ");
2558             bundle->dump();
2559           }
2560           n->dump();
2561         }
2562       }
2563     }
2564 #endif
2565 #ifdef ASSERT
2566   verify_good_schedule(bb,"after block local scheduling");
2567 #endif
2568   }
2569 
2570 #ifndef PRODUCT
2571   if (_cfg->C->trace_opto_output())
2572     tty->print("# <- DoScheduling\n");
2573 #endif
2574 
2575   // Record final node-bundling array location
2576   _regalloc->C->set_node_bundling_base(_node_bundling_base);
2577 
2578 } // end DoScheduling
2579 
2580 // Verify that no live-range used in the block is killed in the block by a
2581 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2582 
2583 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2584 static bool edge_from_to( Node *from, Node *to ) {
2585   for( uint i=0; i<from->len(); i++ )
2586     if( from->in(i) == to )
2587       return true;
2588   return false;
2589 }
2590 
2591 #ifdef ASSERT
2592 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2593   // Check for bad kills
2594   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2595     Node *prior_use = _reg_node[def];
2596     if( prior_use && !edge_from_to(prior_use,n) ) {
2597       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2598       n->dump();
2599       tty->print_cr("...");
2600       prior_use->dump();
2601       assert(edge_from_to(prior_use,n),msg);
2602     }
2603     _reg_node.map(def,NULL); // Kill live USEs
2604   }
2605 }
2606 
2607 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2608 
2609   // Zap to something reasonable for the verify code
2610   _reg_node.clear();
2611 
2612   // Walk over the block backwards.  Check to make sure each DEF doesn't
2613   // kill a live value (other than the one it's supposed to).  Add each
2614   // USE to the live set.
2615   for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2616     Node *n = b->get_node(i);
2617     int n_op = n->Opcode();
2618     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2619       // Fat-proj kills a slew of registers
2620       RegMask rm = n->out_RegMask();// Make local copy
2621       while( rm.is_NotEmpty() ) {
2622         OptoReg::Name kill = rm.find_first_elem();
2623         rm.Remove(kill);
2624         verify_do_def( n, kill, msg );
2625       }
2626     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2627       // Get DEF'd registers the normal way
2628       verify_do_def( n, _regalloc->get_reg_first(n), msg );
2629       verify_do_def( n, _regalloc->get_reg_second(n), msg );
2630     }
2631 
2632     // Now make all USEs live
2633     for( uint i=1; i<n->req(); i++ ) {
2634       Node *def = n->in(i);
2635       assert(def != 0, "input edge required");
2636       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2637       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2638       if( OptoReg::is_valid(reg_lo) ) {
2639         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2640         _reg_node.map(reg_lo,n);
2641       }
2642       if( OptoReg::is_valid(reg_hi) ) {
2643         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2644         _reg_node.map(reg_hi,n);
2645       }
2646     }
2647 
2648   }
2649 
2650   // Zap to something reasonable for the Antidependence code
2651   _reg_node.clear();
2652 }
2653 #endif
2654 
2655 // Conditionally add precedence edges.  Avoid putting edges on Projs.
2656 static void add_prec_edge_from_to( Node *from, Node *to ) {
2657   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
2658     assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2659     from = from->in(0);
2660   }
2661   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
2662       !edge_from_to( from, to ) ) // Avoid duplicate edge
2663     from->add_prec(to);
2664 }
2665 
2666 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2667   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2668     return;
2669 
2670   Node *pinch = _reg_node[def_reg]; // Get pinch point
2671   if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2672       is_def ) {    // Check for a true def (not a kill)
2673     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2674     return;
2675   }
2676 
2677   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
2678   debug_only( def = (Node*)0xdeadbeef; )
2679 
2680   // After some number of kills there _may_ be a later def
2681   Node *later_def = NULL;
2682 
2683   // Finding a kill requires a real pinch-point.
2684   // Check for not already having a pinch-point.
2685   // Pinch points are Op_Node's.
2686   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2687     later_def = pinch;            // Must be def/kill as optimistic pinch-point
2688     if ( _pinch_free_list.size() > 0) {
2689       pinch = _pinch_free_list.pop();
2690     } else {
2691       pinch = new (_cfg->C) Node(1); // Pinch point to-be
2692     }
2693     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2694       _cfg->C->record_method_not_compilable("too many D-U pinch points");
2695       return;
2696     }
2697     _cfg->map_node_to_block(pinch, b);      // Pretend it's valid in this block (lazy init)
2698     _reg_node.map(def_reg,pinch); // Record pinch-point
2699     //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2700     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2701       pinch->init_req(0, _cfg->C->top());     // set not NULL for the next call
2702       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2703       later_def = NULL;           // and no later def
2704     }
2705     pinch->set_req(0,later_def);  // Hook later def so we can find it
2706   } else {                        // Else have valid pinch point
2707     if( pinch->in(0) )            // If there is a later-def
2708       later_def = pinch->in(0);   // Get it
2709   }
2710 
2711   // Add output-dependence edge from later def to kill
2712   if( later_def )               // If there is some original def
2713     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2714 
2715   // See if current kill is also a use, and so is forced to be the pinch-point.
2716   if( pinch->Opcode() == Op_Node ) {
2717     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2718     for( uint i=1; i<uses->req(); i++ ) {
2719       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2720           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2721         // Yes, found a use/kill pinch-point
2722         pinch->set_req(0,NULL);  //
2723         pinch->replace_by(kill); // Move anti-dep edges up
2724         pinch = kill;
2725         _reg_node.map(def_reg,pinch);
2726         return;
2727       }
2728     }
2729   }
2730 
2731   // Add edge from kill to pinch-point
2732   add_prec_edge_from_to(kill,pinch);
2733 }
2734 
2735 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2736   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2737     return;
2738   Node *pinch = _reg_node[use_reg]; // Get pinch point
2739   // Check for no later def_reg/kill in block
2740   if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2741       // Use has to be block-local as well
2742       _cfg->get_block_for_node(use) == b) {
2743     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2744         pinch->req() == 1 ) {   // pinch not yet in block?
2745       pinch->del_req(0);        // yank pointer to later-def, also set flag
2746       // Insert the pinch-point in the block just after the last use
2747       b->insert_node(pinch, b->find_node(use) + 1);
2748       _bb_end++;                // Increase size scheduled region in block
2749     }
2750 
2751     add_prec_edge_from_to(pinch,use);
2752   }
2753 }
2754 
2755 // We insert antidependences between the reads and following write of
2756 // allocated registers to prevent illegal code motion. Hopefully, the
2757 // number of added references should be fairly small, especially as we
2758 // are only adding references within the current basic block.
2759 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2760 
2761 #ifdef ASSERT
2762   verify_good_schedule(b,"before block local scheduling");
2763 #endif
2764 
2765   // A valid schedule, for each register independently, is an endless cycle
2766   // of: a def, then some uses (connected to the def by true dependencies),
2767   // then some kills (defs with no uses), finally the cycle repeats with a new
2768   // def.  The uses are allowed to float relative to each other, as are the
2769   // kills.  No use is allowed to slide past a kill (or def).  This requires
2770   // antidependencies between all uses of a single def and all kills that
2771   // follow, up to the next def.  More edges are redundant, because later defs
2772   // & kills are already serialized with true or antidependencies.  To keep
2773   // the edge count down, we add a 'pinch point' node if there's more than
2774   // one use or more than one kill/def.
2775 
2776   // We add dependencies in one bottom-up pass.
2777 
2778   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2779 
2780   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2781   // register.  If not, we record the DEF/KILL in _reg_node, the
2782   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
2783   // "pinch point", a new Node that's in the graph but not in the block.
2784   // We put edges from the prior and current DEF/KILLs to the pinch point.
2785   // We put the pinch point in _reg_node.  If there's already a pinch point
2786   // we merely add an edge from the current DEF/KILL to the pinch point.
2787 
2788   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
2789   // put an edge from the pinch point to the USE.
2790 
2791   // To be expedient, the _reg_node array is pre-allocated for the whole
2792   // compilation.  _reg_node is lazily initialized; it either contains a NULL,
2793   // or a valid def/kill/pinch-point, or a leftover node from some prior
2794   // block.  Leftover node from some prior block is treated like a NULL (no
2795   // prior def, so no anti-dependence needed).  Valid def is distinguished by
2796   // it being in the current block.
2797   bool fat_proj_seen = false;
2798   uint last_safept = _bb_end-1;
2799   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2800   Node* last_safept_node = end_node;
2801   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2802     Node *n = b->get_node(i);
2803     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
2804     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2805       // Fat-proj kills a slew of registers
2806       // This can add edges to 'n' and obscure whether or not it was a def,
2807       // hence the is_def flag.
2808       fat_proj_seen = true;
2809       RegMask rm = n->out_RegMask();// Make local copy
2810       while( rm.is_NotEmpty() ) {
2811         OptoReg::Name kill = rm.find_first_elem();
2812         rm.Remove(kill);
2813         anti_do_def( b, n, kill, is_def );
2814       }
2815     } else {
2816       // Get DEF'd registers the normal way
2817       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2818       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2819     }
2820 
2821     // Kill projections on a branch should appear to occur on the
2822     // branch, not afterwards, so grab the masks from the projections
2823     // and process them.
2824     if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2825       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2826         Node* use = n->fast_out(i);
2827         if (use->is_Proj()) {
2828           RegMask rm = use->out_RegMask();// Make local copy
2829           while( rm.is_NotEmpty() ) {
2830             OptoReg::Name kill = rm.find_first_elem();
2831             rm.Remove(kill);
2832             anti_do_def( b, n, kill, false );
2833           }
2834         }
2835       }
2836     }
2837 
2838     // Check each register used by this instruction for a following DEF/KILL
2839     // that must occur afterward and requires an anti-dependence edge.
2840     for( uint j=0; j<n->req(); j++ ) {
2841       Node *def = n->in(j);
2842       if( def ) {
2843         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2844         anti_do_use( b, n, _regalloc->get_reg_first(def) );
2845         anti_do_use( b, n, _regalloc->get_reg_second(def) );
2846       }
2847     }
2848     // Do not allow defs of new derived values to float above GC
2849     // points unless the base is definitely available at the GC point.
2850 
2851     Node *m = b->get_node(i);
2852 
2853     // Add precedence edge from following safepoint to use of derived pointer
2854     if( last_safept_node != end_node &&
2855         m != last_safept_node) {
2856       for (uint k = 1; k < m->req(); k++) {
2857         const Type *t = m->in(k)->bottom_type();
2858         if( t->isa_oop_ptr() &&
2859             t->is_ptr()->offset() != 0 ) {
2860           last_safept_node->add_prec( m );
2861           break;
2862         }
2863       }
2864     }
2865 
2866     if( n->jvms() ) {           // Precedence edge from derived to safept
2867       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2868       if( b->get_node(last_safept) != last_safept_node ) {
2869         last_safept = b->find_node(last_safept_node);
2870       }
2871       for( uint j=last_safept; j > i; j-- ) {
2872         Node *mach = b->get_node(j);
2873         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2874           mach->add_prec( n );
2875       }
2876       last_safept = i;
2877       last_safept_node = m;
2878     }
2879   }
2880 
2881   if (fat_proj_seen) {
2882     // Garbage collect pinch nodes that were not consumed.
2883     // They are usually created by a fat kill MachProj for a call.
2884     garbage_collect_pinch_nodes();
2885   }
2886 }
2887 
2888 // Garbage collect pinch nodes for reuse by other blocks.
2889 //
2890 // The block scheduler's insertion of anti-dependence
2891 // edges creates many pinch nodes when the block contains
2892 // 2 or more Calls.  A pinch node is used to prevent a
2893 // combinatorial explosion of edges.  If a set of kills for a
2894 // register is anti-dependent on a set of uses (or defs), rather
2895 // than adding an edge in the graph between each pair of kill
2896 // and use (or def), a pinch is inserted between them:
2897 //
2898 //            use1   use2  use3
2899 //                \   |   /
2900 //                 \  |  /
2901 //                  pinch
2902 //                 /  |  \
2903 //                /   |   \
2904 //            kill1 kill2 kill3
2905 //
2906 // One pinch node is created per register killed when
2907 // the second call is encountered during a backwards pass
2908 // over the block.  Most of these pinch nodes are never
2909 // wired into the graph because the register is never
2910 // used or def'ed in the block.
2911 //
2912 void Scheduling::garbage_collect_pinch_nodes() {
2913 #ifndef PRODUCT
2914     if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2915 #endif
2916     int trace_cnt = 0;
2917     for (uint k = 0; k < _reg_node.Size(); k++) {
2918       Node* pinch = _reg_node[k];
2919       if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2920           // no predecence input edges
2921           (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2922         cleanup_pinch(pinch);
2923         _pinch_free_list.push(pinch);
2924         _reg_node.map(k, NULL);
2925 #ifndef PRODUCT
2926         if (_cfg->C->trace_opto_output()) {
2927           trace_cnt++;
2928           if (trace_cnt > 40) {
2929             tty->print("\n");
2930             trace_cnt = 0;
2931           }
2932           tty->print(" %d", pinch->_idx);
2933         }
2934 #endif
2935       }
2936     }
2937 #ifndef PRODUCT
2938     if (_cfg->C->trace_opto_output()) tty->print("\n");
2939 #endif
2940 }
2941 
2942 // Clean up a pinch node for reuse.
2943 void Scheduling::cleanup_pinch( Node *pinch ) {
2944   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2945 
2946   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2947     Node* use = pinch->last_out(i);
2948     uint uses_found = 0;
2949     for (uint j = use->req(); j < use->len(); j++) {
2950       if (use->in(j) == pinch) {
2951         use->rm_prec(j);
2952         uses_found++;
2953       }
2954     }
2955     assert(uses_found > 0, "must be a precedence edge");
2956     i -= uses_found;    // we deleted 1 or more copies of this edge
2957   }
2958   // May have a later_def entry
2959   pinch->set_req(0, NULL);
2960 }
2961 
2962 #ifndef PRODUCT
2963 
2964 void Scheduling::dump_available() const {
2965   tty->print("#Availist  ");
2966   for (uint i = 0; i < _available.size(); i++)
2967     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2968   tty->cr();
2969 }
2970 
2971 // Print Scheduling Statistics
2972 void Scheduling::print_statistics() {
2973   // Print the size added by nops for bundling
2974   tty->print("Nops added %d bytes to total of %d bytes",
2975     _total_nop_size, _total_method_size);
2976   if (_total_method_size > 0)
2977     tty->print(", for %.2f%%",
2978       ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2979   tty->print("\n");
2980 
2981   // Print the number of branch shadows filled
2982   if (Pipeline::_branch_has_delay_slot) {
2983     tty->print("Of %d branches, %d had unconditional delay slots filled",
2984       _total_branches, _total_unconditional_delays);
2985     if (_total_branches > 0)
2986       tty->print(", for %.2f%%",
2987         ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2988     tty->print("\n");
2989   }
2990 
2991   uint total_instructions = 0, total_bundles = 0;
2992 
2993   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2994     uint bundle_count   = _total_instructions_per_bundle[i];
2995     total_instructions += bundle_count * i;
2996     total_bundles      += bundle_count;
2997   }
2998 
2999   if (total_bundles > 0)
3000     tty->print("Average ILP (excluding nops) is %.2f\n",
3001       ((double)total_instructions) / ((double)total_bundles));
3002 }
3003 #endif