1 /*
   2  * Copyright (c) 1998, 2014, 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/ad.hpp"
  34 #include "opto/callnode.hpp"
  35 #include "opto/cfgnode.hpp"
  36 #include "opto/locknode.hpp"
  37 #include "opto/machnode.hpp"
  38 #include "opto/optoreg.hpp"
  39 #include "opto/output.hpp"
  40 #include "opto/regalloc.hpp"
  41 #include "opto/runtime.hpp"
  42 #include "opto/subnode.hpp"
  43 #include "opto/type.hpp"
  44 #include "runtime/handles.inline.hpp"
  45 #include "utilities/xmlstream.hpp"
  46 
  47 #ifndef PRODUCT
  48 #define DEBUG_ARG(x) , x
  49 #else
  50 #define DEBUG_ARG(x)
  51 #endif
  52 
  53 // Convert Nodes to instruction bits and pass off to the VM
  54 void Compile::Output() {
  55   // RootNode goes
  56   assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
  57 
  58   // The number of new nodes (mostly MachNop) is proportional to
  59   // the number of java calls and inner loops which are aligned.
  60   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
  61                             C->inner_loops()*(OptoLoopAlignment-1)),
  62                            "out of nodes before code generation" ) ) {
  63     return;
  64   }
  65   // Make sure I can find the Start Node
  66   Block *entry = _cfg->get_block(1);
  67   Block *broot = _cfg->get_root_block();
  68 
  69   const StartNode *start = entry->head()->as_Start();
  70 
  71   // Replace StartNode with prolog
  72   MachPrologNode *prolog = new MachPrologNode();
  73   entry->map_node(prolog, 0);
  74   _cfg->map_node_to_block(prolog, entry);
  75   _cfg->unmap_node_from_block(start); // start is no longer in any block
  76 
  77   // Virtual methods need an unverified entry point
  78 
  79   if( is_osr_compilation() ) {
  80     if( PoisonOSREntry ) {
  81       // TODO: Should use a ShouldNotReachHereNode...
  82       _cfg->insert( broot, 0, new MachBreakpointNode() );
  83     }
  84   } else {
  85     if( _method && !_method->flags().is_static() ) {
  86       // Insert unvalidated entry point
  87       _cfg->insert( broot, 0, new MachUEPNode() );
  88     }
  89 
  90   }
  91 
  92 
  93   // Break before main entry point
  94   if( (_method && _method->break_at_execute())
  95 #ifndef PRODUCT
  96     ||(OptoBreakpoint && is_method_compilation())
  97     ||(OptoBreakpointOSR && is_osr_compilation())
  98     ||(OptoBreakpointC2R && !_method)
  99 #endif
 100     ) {
 101     // checking for _method means that OptoBreakpoint does not apply to
 102     // runtime stubs or frame converters
 103     _cfg->insert( entry, 1, new MachBreakpointNode() );
 104   }
 105 
 106   // Insert epilogs before every return
 107   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
 108     Block* block = _cfg->get_block(i);
 109     if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
 110       Node* m = block->end();
 111       if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
 112         MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
 113         block->add_inst(epilog);
 114         _cfg->map_node_to_block(epilog, block);
 115       }
 116     }
 117   }
 118 
 119 # ifdef ENABLE_ZAP_DEAD_LOCALS
 120   if (ZapDeadCompiledLocals) {
 121     Insert_zap_nodes();
 122   }
 123 # endif
 124 
 125   uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
 126   blk_starts[0] = 0;
 127 
 128   // Initialize code buffer and process short branches.
 129   CodeBuffer* cb = init_buffer(blk_starts);
 130 
 131   if (cb == NULL || failing()) {
 132     return;
 133   }
 134 
 135   ScheduleAndBundle();
 136 
 137 #ifndef PRODUCT
 138   if (trace_opto_output()) {
 139     tty->print("\n---- After ScheduleAndBundle ----\n");
 140     for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
 141       tty->print("\nBB#%03d:\n", i);
 142       Block* block = _cfg->get_block(i);
 143       for (uint j = 0; j < block->number_of_nodes(); j++) {
 144         Node* n = block->get_node(j);
 145         OptoReg::Name reg = _regalloc->get_reg_first(n);
 146         tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
 147         n->dump();
 148       }
 149     }
 150   }
 151 #endif
 152 
 153   if (failing()) {
 154     return;
 155   }
 156 
 157   BuildOopMaps();
 158 
 159   if (failing())  {
 160     return;
 161   }
 162 
 163   fill_buffer(cb, blk_starts);
 164 }
 165 
 166 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
 167   // Determine if we need to generate a stack overflow check.
 168   // Do it if the method is not a stub function and
 169   // has java calls or has frame size > vm_page_size/8.
 170   // The debug VM checks that deoptimization doesn't trigger an
 171   // unexpected stack overflow (compiled method stack banging should
 172   // guarantee it doesn't happen) so we always need the stack bang in
 173   // a debug VM.
 174   return (UseStackBanging && stub_function() == NULL &&
 175           (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
 176            DEBUG_ONLY(|| true)));
 177 }
 178 
 179 bool Compile::need_register_stack_bang() const {
 180   // Determine if we need to generate a register stack overflow check.
 181   // This is only used on architectures which have split register
 182   // and memory stacks (ie. IA64).
 183   // Bang if the method is not a stub function and has java calls
 184   return (stub_function() == NULL && has_java_calls());
 185 }
 186 
 187 # ifdef ENABLE_ZAP_DEAD_LOCALS
 188 
 189 
 190 // In order to catch compiler oop-map bugs, we have implemented
 191 // a debugging mode called ZapDeadCompilerLocals.
 192 // This mode causes the compiler to insert a call to a runtime routine,
 193 // "zap_dead_locals", right before each place in compiled code
 194 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
 195 // The runtime routine checks that locations mapped as oops are really
 196 // oops, that locations mapped as values do not look like oops,
 197 // and that locations mapped as dead are not used later
 198 // (by zapping them to an invalid address).
 199 
 200 int Compile::_CompiledZap_count = 0;
 201 
 202 void Compile::Insert_zap_nodes() {
 203   bool skip = false;
 204 
 205 
 206   // Dink with static counts because code code without the extra
 207   // runtime calls is MUCH faster for debugging purposes
 208 
 209        if ( CompileZapFirst  ==  0  ) ; // nothing special
 210   else if ( CompileZapFirst  >  CompiledZap_count() )  skip = true;
 211   else if ( CompileZapFirst  == CompiledZap_count() )
 212     warning("starting zap compilation after skipping");
 213 
 214        if ( CompileZapLast  ==  -1  ) ; // nothing special
 215   else if ( CompileZapLast  <   CompiledZap_count() )  skip = true;
 216   else if ( CompileZapLast  ==  CompiledZap_count() )
 217     warning("about to compile last zap");
 218 
 219   ++_CompiledZap_count; // counts skipped zaps, too
 220 
 221   if ( skip )  return;
 222 
 223 
 224   if ( _method == NULL )
 225     return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
 226 
 227   // Insert call to zap runtime stub before every node with an oop map
 228   for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
 229     Block *b = _cfg->get_block(i);
 230     for ( uint j = 0;  j < b->number_of_nodes();  ++j ) {
 231       Node *n = b->get_node(j);
 232 
 233       // Determining if we should insert a zap-a-lot node in output.
 234       // We do that for all nodes that has oopmap info, except for calls
 235       // to allocation.  Calls to allocation passes in the old top-of-eden pointer
 236       // and expect the C code to reset it.  Hence, there can be no safepoints between
 237       // the inlined-allocation and the call to new_Java, etc.
 238       // We also cannot zap monitor calls, as they must hold the microlock
 239       // during the call to Zap, which also wants to grab the microlock.
 240       bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
 241       if ( insert ) { // it is MachSafePoint
 242         if ( !n->is_MachCall() ) {
 243           insert = false;
 244         } else if ( n->is_MachCall() ) {
 245           MachCallNode* call = n->as_MachCall();
 246           if (call->entry_point() == OptoRuntime::new_instance_Java() ||
 247               call->entry_point() == OptoRuntime::new_array_Java() ||
 248               call->entry_point() == OptoRuntime::multianewarray2_Java() ||
 249               call->entry_point() == OptoRuntime::multianewarray3_Java() ||
 250               call->entry_point() == OptoRuntime::multianewarray4_Java() ||
 251               call->entry_point() == OptoRuntime::multianewarray5_Java() ||
 252               call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
 253               call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
 254               ) {
 255             insert = false;
 256           }
 257         }
 258         if (insert) {
 259           Node *zap = call_zap_node(n->as_MachSafePoint(), i);
 260           b->insert_node(zap, j);
 261           _cfg->map_node_to_block(zap, b);
 262           ++j;
 263         }
 264       }
 265     }
 266   }
 267 }
 268 
 269 
 270 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
 271   const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
 272   CallStaticJavaNode* ideal_node =
 273     new CallStaticJavaNode( tf,
 274          OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
 275                        "call zap dead locals stub", 0, TypePtr::BOTTOM);
 276   // We need to copy the OopMap from the site we're zapping at.
 277   // We have to make a copy, because the zap site might not be
 278   // a call site, and zap_dead is a call site.
 279   OopMap* clone = node_to_check->oop_map()->deep_copy();
 280 
 281   // Add the cloned OopMap to the zap node
 282   ideal_node->set_oop_map(clone);
 283   return _matcher->match_sfpt(ideal_node);
 284 }
 285 
 286 bool Compile::is_node_getting_a_safepoint( Node* n) {
 287   // This code duplicates the logic prior to the call of add_safepoint
 288   // below in this file.
 289   if( n->is_MachSafePoint() ) return true;
 290   return false;
 291 }
 292 
 293 # endif // ENABLE_ZAP_DEAD_LOCALS
 294 
 295 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 296 // of a loop. When aligning a loop we need to provide enough instructions
 297 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 298 // avoided if we have enough instructions in fetch buffer at the head of a loop.
 299 // By default, the size is set to 999999 by Block's constructor so that
 300 // a loop will be aligned if the size is not reset here.
 301 //
 302 // Note: Mach instructions could contain several HW instructions
 303 // so the size is estimated only.
 304 //
 305 void Compile::compute_loop_first_inst_sizes() {
 306   // The next condition is used to gate the loop alignment optimization.
 307   // Don't aligned a loop if there are enough instructions at the head of a loop
 308   // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
 309   // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
 310   // equal to 11 bytes which is the largest address NOP instruction.
 311   if (MaxLoopPad < OptoLoopAlignment - 1) {
 312     uint last_block = _cfg->number_of_blocks() - 1;
 313     for (uint i = 1; i <= last_block; i++) {
 314       Block* block = _cfg->get_block(i);
 315       // Check the first loop's block which requires an alignment.
 316       if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
 317         uint sum_size = 0;
 318         uint inst_cnt = NumberOfLoopInstrToAlign;
 319         inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
 320 
 321         // Check subsequent fallthrough blocks if the loop's first
 322         // block(s) does not have enough instructions.
 323         Block *nb = block;
 324         while(inst_cnt > 0 &&
 325               i < last_block &&
 326               !_cfg->get_block(i + 1)->has_loop_alignment() &&
 327               !nb->has_successor(block)) {
 328           i++;
 329           nb = _cfg->get_block(i);
 330           inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
 331         } // while( inst_cnt > 0 && i < last_block  )
 332 
 333         block->set_first_inst_size(sum_size);
 334       } // f( b->head()->is_Loop() )
 335     } // for( i <= last_block )
 336   } // if( MaxLoopPad < OptoLoopAlignment-1 )
 337 }
 338 
 339 // The architecture description provides short branch variants for some long
 340 // branch instructions. Replace eligible long branches with short branches.
 341 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
 342   // Compute size of each block, method size, and relocation information size
 343   uint nblocks  = _cfg->number_of_blocks();
 344 
 345   uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
 346   uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
 347   int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
 348 
 349   // Collect worst case block paddings
 350   int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
 351   memset(block_worst_case_pad, 0, nblocks * sizeof(int));
 352 
 353   DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
 354   DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
 355 
 356   bool has_short_branch_candidate = false;
 357 
 358   // Initialize the sizes to 0
 359   code_size  = 0;          // Size in bytes of generated code
 360   stub_size  = 0;          // Size in bytes of all stub entries
 361   // Size in bytes of all relocation entries, including those in local stubs.
 362   // Start with 2-bytes of reloc info for the unvalidated entry point
 363   reloc_size = 1;          // Number of relocation entries
 364 
 365   // Make three passes.  The first computes pessimistic blk_starts,
 366   // relative jmp_offset and reloc_size information.  The second performs
 367   // short branch substitution using the pessimistic sizing.  The
 368   // third inserts nops where needed.
 369 
 370   // Step one, perform a pessimistic sizing pass.
 371   uint last_call_adr = max_juint;
 372   uint last_avoid_back_to_back_adr = max_juint;
 373   uint nop_size = (new MachNopNode())->size(_regalloc);
 374   for (uint i = 0; i < nblocks; i++) { // For all blocks
 375     Block* block = _cfg->get_block(i);
 376 
 377     // During short branch replacement, we store the relative (to blk_starts)
 378     // offset of jump in jmp_offset, rather than the absolute offset of jump.
 379     // This is so that we do not need to recompute sizes of all nodes when
 380     // we compute correct blk_starts in our next sizing pass.
 381     jmp_offset[i] = 0;
 382     jmp_size[i]   = 0;
 383     jmp_nidx[i]   = -1;
 384     DEBUG_ONLY( jmp_target[i] = 0; )
 385     DEBUG_ONLY( jmp_rule[i]   = 0; )
 386 
 387     // Sum all instruction sizes to compute block size
 388     uint last_inst = block->number_of_nodes();
 389     uint blk_size = 0;
 390     for (uint j = 0; j < last_inst; j++) {
 391       Node* nj = block->get_node(j);
 392       // Handle machine instruction nodes
 393       if (nj->is_Mach()) {
 394         MachNode *mach = nj->as_Mach();
 395         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 396         reloc_size += mach->reloc();
 397         if (mach->is_MachCall()) {
 398           // add size information for trampoline stub
 399           // class CallStubImpl is platform-specific and defined in the *.ad files.
 400           stub_size  += CallStubImpl::size_call_trampoline();
 401           reloc_size += CallStubImpl::reloc_call_trampoline();
 402 
 403           MachCallNode *mcall = mach->as_MachCall();
 404           // This destination address is NOT PC-relative
 405 
 406           mcall->method_set((intptr_t)mcall->entry_point());
 407 
 408           if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
 409             stub_size  += CompiledStaticCall::to_interp_stub_size();
 410             reloc_size += CompiledStaticCall::reloc_to_interp_stub();
 411           }
 412         } else if (mach->is_MachSafePoint()) {
 413           // If call/safepoint are adjacent, account for possible
 414           // nop to disambiguate the two safepoints.
 415           // ScheduleAndBundle() can rearrange nodes in a block,
 416           // check for all offsets inside this block.
 417           if (last_call_adr >= blk_starts[i]) {
 418             blk_size += nop_size;
 419           }
 420         }
 421         if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 422           // Nop is inserted between "avoid back to back" instructions.
 423           // ScheduleAndBundle() can rearrange nodes in a block,
 424           // check for all offsets inside this block.
 425           if (last_avoid_back_to_back_adr >= blk_starts[i]) {
 426             blk_size += nop_size;
 427           }
 428         }
 429         if (mach->may_be_short_branch()) {
 430           if (!nj->is_MachBranch()) {
 431 #ifndef PRODUCT
 432             nj->dump(3);
 433 #endif
 434             Unimplemented();
 435           }
 436           assert(jmp_nidx[i] == -1, "block should have only one branch");
 437           jmp_offset[i] = blk_size;
 438           jmp_size[i]   = nj->size(_regalloc);
 439           jmp_nidx[i]   = j;
 440           has_short_branch_candidate = true;
 441         }
 442       }
 443       blk_size += nj->size(_regalloc);
 444       // Remember end of call offset
 445       if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
 446         last_call_adr = blk_starts[i]+blk_size;
 447       }
 448       // Remember end of avoid_back_to_back offset
 449       if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
 450         last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
 451       }
 452     }
 453 
 454     // When the next block starts a loop, we may insert pad NOP
 455     // instructions.  Since we cannot know our future alignment,
 456     // assume the worst.
 457     if (i < nblocks - 1) {
 458       Block* nb = _cfg->get_block(i + 1);
 459       int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
 460       if (max_loop_pad > 0) {
 461         assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 462         // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
 463         // If either is the last instruction in this block, bump by
 464         // max_loop_pad in lock-step with blk_size, so sizing
 465         // calculations in subsequent blocks still can conservatively
 466         // detect that it may the last instruction in this block.
 467         if (last_call_adr == blk_starts[i]+blk_size) {
 468           last_call_adr += max_loop_pad;
 469         }
 470         if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
 471           last_avoid_back_to_back_adr += max_loop_pad;
 472         }
 473         blk_size += max_loop_pad;
 474         block_worst_case_pad[i + 1] = max_loop_pad;
 475       }
 476     }
 477 
 478     // Save block size; update total method size
 479     blk_starts[i+1] = blk_starts[i]+blk_size;
 480   }
 481 
 482   // Step two, replace eligible long jumps.
 483   bool progress = true;
 484   uint last_may_be_short_branch_adr = max_juint;
 485   while (has_short_branch_candidate && progress) {
 486     progress = false;
 487     has_short_branch_candidate = false;
 488     int adjust_block_start = 0;
 489     for (uint i = 0; i < nblocks; i++) {
 490       Block* block = _cfg->get_block(i);
 491       int idx = jmp_nidx[i];
 492       MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
 493       if (mach != NULL && mach->may_be_short_branch()) {
 494 #ifdef ASSERT
 495         assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
 496         int j;
 497         // Find the branch; ignore trailing NOPs.
 498         for (j = block->number_of_nodes()-1; j>=0; j--) {
 499           Node* n = block->get_node(j);
 500           if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
 501             break;
 502         }
 503         assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
 504 #endif
 505         int br_size = jmp_size[i];
 506         int br_offs = blk_starts[i] + jmp_offset[i];
 507 
 508         // This requires the TRUE branch target be in succs[0]
 509         uint bnum = block->non_connector_successor(0)->_pre_order;
 510         int offset = blk_starts[bnum] - br_offs;
 511         if (bnum > i) { // adjust following block's offset
 512           offset -= adjust_block_start;
 513         }
 514 
 515         // This block can be a loop header, account for the padding
 516         // in the previous block.
 517         int block_padding = block_worst_case_pad[i];
 518         assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
 519         // In the following code a nop could be inserted before
 520         // the branch which will increase the backward distance.
 521         bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
 522         assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
 523 
 524         if (needs_padding && offset <= 0)
 525           offset -= nop_size;
 526 
 527         if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
 528           // We've got a winner.  Replace this branch.
 529           MachNode* replacement = mach->as_MachBranch()->short_branch_version();
 530 
 531           // Update the jmp_size.
 532           int new_size = replacement->size(_regalloc);
 533           int diff     = br_size - new_size;
 534           assert(diff >= (int)nop_size, "short_branch size should be smaller");
 535           // Conservatively take into account padding between
 536           // avoid_back_to_back branches. Previous branch could be
 537           // converted into avoid_back_to_back branch during next
 538           // rounds.
 539           if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
 540             jmp_offset[i] += nop_size;
 541             diff -= nop_size;
 542           }
 543           adjust_block_start += diff;
 544           block->map_node(replacement, idx);
 545           mach->subsume_by(replacement, C);
 546           mach = replacement;
 547           progress = true;
 548 
 549           jmp_size[i] = new_size;
 550           DEBUG_ONLY( jmp_target[i] = bnum; );
 551           DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 552         } else {
 553           // The jump distance is not short, try again during next iteration.
 554           has_short_branch_candidate = true;
 555         }
 556       } // (mach->may_be_short_branch())
 557       if (mach != NULL && (mach->may_be_short_branch() ||
 558                            mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
 559         last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
 560       }
 561       blk_starts[i+1] -= adjust_block_start;
 562     }
 563   }
 564 
 565 #ifdef ASSERT
 566   for (uint i = 0; i < nblocks; i++) { // For all blocks
 567     if (jmp_target[i] != 0) {
 568       int br_size = jmp_size[i];
 569       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
 570       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
 571         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]);
 572       }
 573       assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 574     }
 575   }
 576 #endif
 577 
 578   // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
 579   // after ScheduleAndBundle().
 580 
 581   // ------------------
 582   // Compute size for code buffer
 583   code_size = blk_starts[nblocks];
 584 
 585   // Relocation records
 586   reloc_size += 1;              // Relo entry for exception handler
 587 
 588   // Adjust reloc_size to number of record of relocation info
 589   // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
 590   // a relocation index.
 591   // The CodeBuffer will expand the locs array if this estimate is too low.
 592   reloc_size *= 10 / sizeof(relocInfo);
 593 }
 594 
 595 //------------------------------FillLocArray-----------------------------------
 596 // Create a bit of debug info and append it to the array.  The mapping is from
 597 // Java local or expression stack to constant, register or stack-slot.  For
 598 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
 599 // entry has been taken care of and caller should skip it).
 600 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
 601   // This should never have accepted Bad before
 602   assert(OptoReg::is_valid(regnum), "location must be valid");
 603   return (OptoReg::is_reg(regnum))
 604     ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
 605     : new LocationValue(Location::new_stk_loc(l_type,  ra->reg2offset(regnum)));
 606 }
 607 
 608 
 609 ObjectValue*
 610 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
 611   for (int i = 0; i < objs->length(); i++) {
 612     assert(objs->at(i)->is_object(), "corrupt object cache");
 613     ObjectValue* sv = (ObjectValue*) objs->at(i);
 614     if (sv->id() == id) {
 615       return sv;
 616     }
 617   }
 618   // Otherwise..
 619   return NULL;
 620 }
 621 
 622 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
 623                                      ObjectValue* sv ) {
 624   assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
 625   objs->append(sv);
 626 }
 627 
 628 
 629 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
 630                             GrowableArray<ScopeValue*> *array,
 631                             GrowableArray<ScopeValue*> *objs ) {
 632   assert( local, "use _top instead of null" );
 633   if (array->length() != idx) {
 634     assert(array->length() == idx + 1, "Unexpected array count");
 635     // Old functionality:
 636     //   return
 637     // New functionality:
 638     //   Assert if the local is not top. In product mode let the new node
 639     //   override the old entry.
 640     assert(local == top(), "LocArray collision");
 641     if (local == top()) {
 642       return;
 643     }
 644     array->pop();
 645   }
 646   const Type *t = local->bottom_type();
 647 
 648   // Is it a safepoint scalar object node?
 649   if (local->is_SafePointScalarObject()) {
 650     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
 651 
 652     ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
 653     if (sv == NULL) {
 654       ciKlass* cik = t->is_oopptr()->klass();
 655       assert(cik->is_instance_klass() ||
 656              cik->is_array_klass(), "Not supported allocation.");
 657       sv = new ObjectValue(spobj->_idx,
 658                            new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 659       Compile::set_sv_for_object_node(objs, sv);
 660 
 661       uint first_ind = spobj->first_index(sfpt->jvms());
 662       for (uint i = 0; i < spobj->n_fields(); i++) {
 663         Node* fld_node = sfpt->in(first_ind+i);
 664         (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
 665       }
 666     }
 667     array->append(sv);
 668     return;
 669   }
 670 
 671   // Grab the register number for the local
 672   OptoReg::Name regnum = _regalloc->get_reg_first(local);
 673   if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
 674     // Record the double as two float registers.
 675     // The register mask for such a value always specifies two adjacent
 676     // float registers, with the lower register number even.
 677     // Normally, the allocation of high and low words to these registers
 678     // is irrelevant, because nearly all operations on register pairs
 679     // (e.g., StoreD) treat them as a single unit.
 680     // Here, we assume in addition that the words in these two registers
 681     // stored "naturally" (by operations like StoreD and double stores
 682     // within the interpreter) such that the lower-numbered register
 683     // is written to the lower memory address.  This may seem like
 684     // a machine dependency, but it is not--it is a requirement on
 685     // the author of the <arch>.ad file to ensure that, for every
 686     // even/odd double-register pair to which a double may be allocated,
 687     // the word in the even single-register is stored to the first
 688     // memory word.  (Note that register numbers are completely
 689     // arbitrary, and are not tied to any machine-level encodings.)
 690 #ifdef _LP64
 691     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
 692       array->append(new ConstantIntValue(0));
 693       array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
 694     } else if ( t->base() == Type::Long ) {
 695       array->append(new ConstantIntValue(0));
 696       array->append(new_loc_value( _regalloc, regnum, Location::lng ));
 697     } else if ( t->base() == Type::RawPtr ) {
 698       // jsr/ret return address which must be restored into a the full
 699       // width 64-bit stack slot.
 700       array->append(new_loc_value( _regalloc, regnum, Location::lng ));
 701     }
 702 #else //_LP64
 703 #ifdef SPARC
 704     if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
 705       // For SPARC we have to swap high and low words for
 706       // long values stored in a single-register (g0-g7).
 707       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));
 708       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
 709     } else
 710 #endif //SPARC
 711     if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
 712       // Repack the double/long as two jints.
 713       // The convention the interpreter uses is that the second local
 714       // holds the first raw word of the native double representation.
 715       // This is actually reasonable, since locals and stack arrays
 716       // grow downwards in all implementations.
 717       // (If, on some machine, the interpreter's Java locals or stack
 718       // were to grow upwards, the embedded doubles would be word-swapped.)
 719       array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
 720       array->append(new_loc_value( _regalloc,              regnum   , Location::normal ));
 721     }
 722 #endif //_LP64
 723     else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
 724                OptoReg::is_reg(regnum) ) {
 725       array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
 726                                    ? Location::float_in_dbl : Location::normal ));
 727     } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
 728       array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
 729                                    ? Location::int_in_long : Location::normal ));
 730     } else if( t->base() == Type::NarrowOop ) {
 731       array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
 732     } else {
 733       array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
 734     }
 735     return;
 736   }
 737 
 738   // No register.  It must be constant data.
 739   switch (t->base()) {
 740   case Type::Half:              // Second half of a double
 741     ShouldNotReachHere();       // Caller should skip 2nd halves
 742     break;
 743   case Type::AnyPtr:
 744     array->append(new ConstantOopWriteValue(NULL));
 745     break;
 746   case Type::AryPtr:
 747   case Type::InstPtr:          // fall through
 748     array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
 749     break;
 750   case Type::NarrowOop:
 751     if (t == TypeNarrowOop::NULL_PTR) {
 752       array->append(new ConstantOopWriteValue(NULL));
 753     } else {
 754       array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
 755     }
 756     break;
 757   case Type::Int:
 758     array->append(new ConstantIntValue(t->is_int()->get_con()));
 759     break;
 760   case Type::RawPtr:
 761     // A return address (T_ADDRESS).
 762     assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
 763 #ifdef _LP64
 764     // Must be restored to the full-width 64-bit stack slot.
 765     array->append(new ConstantLongValue(t->is_ptr()->get_con()));
 766 #else
 767     array->append(new ConstantIntValue(t->is_ptr()->get_con()));
 768 #endif
 769     break;
 770   case Type::FloatCon: {
 771     float f = t->is_float_constant()->getf();
 772     array->append(new ConstantIntValue(jint_cast(f)));
 773     break;
 774   }
 775   case Type::DoubleCon: {
 776     jdouble d = t->is_double_constant()->getd();
 777 #ifdef _LP64
 778     array->append(new ConstantIntValue(0));
 779     array->append(new ConstantDoubleValue(d));
 780 #else
 781     // Repack the double as two jints.
 782     // The convention the interpreter uses is that the second local
 783     // holds the first raw word of the native double representation.
 784     // This is actually reasonable, since locals and stack arrays
 785     // grow downwards in all implementations.
 786     // (If, on some machine, the interpreter's Java locals or stack
 787     // were to grow upwards, the embedded doubles would be word-swapped.)
 788     jlong_accessor acc;
 789     acc.long_value = jlong_cast(d);
 790     array->append(new ConstantIntValue(acc.words[1]));
 791     array->append(new ConstantIntValue(acc.words[0]));
 792 #endif
 793     break;
 794   }
 795   case Type::Long: {
 796     jlong d = t->is_long()->get_con();
 797 #ifdef _LP64
 798     array->append(new ConstantIntValue(0));
 799     array->append(new ConstantLongValue(d));
 800 #else
 801     // Repack the long as two jints.
 802     // The convention the interpreter uses is that the second local
 803     // holds the first raw word of the native double representation.
 804     // This is actually reasonable, since locals and stack arrays
 805     // grow downwards in all implementations.
 806     // (If, on some machine, the interpreter's Java locals or stack
 807     // were to grow upwards, the embedded doubles would be word-swapped.)
 808     jlong_accessor acc;
 809     acc.long_value = d;
 810     array->append(new ConstantIntValue(acc.words[1]));
 811     array->append(new ConstantIntValue(acc.words[0]));
 812 #endif
 813     break;
 814   }
 815   case Type::Top:               // Add an illegal value here
 816     array->append(new LocationValue(Location()));
 817     break;
 818   default:
 819     ShouldNotReachHere();
 820     break;
 821   }
 822 }
 823 
 824 // Determine if this node starts a bundle
 825 bool Compile::starts_bundle(const Node *n) const {
 826   return (_node_bundling_limit > n->_idx &&
 827           _node_bundling_base[n->_idx].starts_bundle());
 828 }
 829 
 830 //--------------------------Process_OopMap_Node--------------------------------
 831 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
 832 
 833   // Handle special safepoint nodes for synchronization
 834   MachSafePointNode *sfn   = mach->as_MachSafePoint();
 835   MachCallNode      *mcall;
 836 
 837 #ifdef ENABLE_ZAP_DEAD_LOCALS
 838   assert( is_node_getting_a_safepoint(mach),  "logic does not match; false negative");
 839 #endif
 840 
 841   int safepoint_pc_offset = current_offset;
 842   bool is_method_handle_invoke = false;
 843   bool return_oop = false;
 844 
 845   // Add the safepoint in the DebugInfoRecorder
 846   if( !mach->is_MachCall() ) {
 847     mcall = NULL;
 848     debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
 849   } else {
 850     mcall = mach->as_MachCall();
 851 
 852     // Is the call a MethodHandle call?
 853     if (mcall->is_MachCallJava()) {
 854       if (mcall->as_MachCallJava()->_method_handle_invoke) {
 855         assert(has_method_handle_invokes(), "must have been set during call generation");
 856         is_method_handle_invoke = true;
 857       }
 858     }
 859 
 860     // Check if a call returns an object.
 861     if (mcall->returns_pointer()) {
 862       return_oop = true;
 863     }
 864     safepoint_pc_offset += mcall->ret_addr_offset();
 865     debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
 866   }
 867 
 868   // Loop over the JVMState list to add scope information
 869   // Do not skip safepoints with a NULL method, they need monitor info
 870   JVMState* youngest_jvms = sfn->jvms();
 871   int max_depth = youngest_jvms->depth();
 872 
 873   // Allocate the object pool for scalar-replaced objects -- the map from
 874   // small-integer keys (which can be recorded in the local and ostack
 875   // arrays) to descriptions of the object state.
 876   GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
 877 
 878   // Visit scopes from oldest to youngest.
 879   for (int depth = 1; depth <= max_depth; depth++) {
 880     JVMState* jvms = youngest_jvms->of_depth(depth);
 881     int idx;
 882     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
 883     // Safepoints that do not have method() set only provide oop-map and monitor info
 884     // to support GC; these do not support deoptimization.
 885     int num_locs = (method == NULL) ? 0 : jvms->loc_size();
 886     int num_exps = (method == NULL) ? 0 : jvms->stk_size();
 887     int num_mon  = jvms->nof_monitors();
 888     assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
 889            "JVMS local count must match that of the method");
 890 
 891     // Add Local and Expression Stack Information
 892 
 893     // Insert locals into the locarray
 894     GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
 895     for( idx = 0; idx < num_locs; idx++ ) {
 896       FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
 897     }
 898 
 899     // Insert expression stack entries into the exparray
 900     GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
 901     for( idx = 0; idx < num_exps; idx++ ) {
 902       FillLocArray( idx,  sfn, sfn->stack(jvms, idx), exparray, objs );
 903     }
 904 
 905     // Add in mappings of the monitors
 906     assert( !method ||
 907             !method->is_synchronized() ||
 908             method->is_native() ||
 909             num_mon > 0 ||
 910             !GenerateSynchronizationCode,
 911             "monitors must always exist for synchronized methods");
 912 
 913     // Build the growable array of ScopeValues for exp stack
 914     GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
 915 
 916     // Loop over monitors and insert into array
 917     for (idx = 0; idx < num_mon; idx++) {
 918       // Grab the node that defines this monitor
 919       Node* box_node = sfn->monitor_box(jvms, idx);
 920       Node* obj_node = sfn->monitor_obj(jvms, idx);
 921 
 922       // Create ScopeValue for object
 923       ScopeValue *scval = NULL;
 924 
 925       if (obj_node->is_SafePointScalarObject()) {
 926         SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
 927         scval = Compile::sv_for_node_id(objs, spobj->_idx);
 928         if (scval == NULL) {
 929           const Type *t = spobj->bottom_type();
 930           ciKlass* cik = t->is_oopptr()->klass();
 931           assert(cik->is_instance_klass() ||
 932                  cik->is_array_klass(), "Not supported allocation.");
 933           ObjectValue* sv = new ObjectValue(spobj->_idx,
 934                                             new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
 935           Compile::set_sv_for_object_node(objs, sv);
 936 
 937           uint first_ind = spobj->first_index(youngest_jvms);
 938           for (uint i = 0; i < spobj->n_fields(); i++) {
 939             Node* fld_node = sfn->in(first_ind+i);
 940             (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
 941           }
 942           scval = sv;
 943         }
 944       } else if (!obj_node->is_Con()) {
 945         OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
 946         if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
 947           scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
 948         } else {
 949           scval = new_loc_value( _regalloc, obj_reg, Location::oop );
 950         }
 951       } else {
 952         const TypePtr *tp = obj_node->get_ptr_type();
 953         scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
 954       }
 955 
 956       OptoReg::Name box_reg = BoxLockNode::reg(box_node);
 957       Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
 958       bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
 959       monarray->append(new MonitorValue(scval, basic_lock, eliminated));
 960     }
 961 
 962     // We dump the object pool first, since deoptimization reads it in first.
 963     debug_info()->dump_object_pool(objs);
 964 
 965     // Build first class objects to pass to scope
 966     DebugToken *locvals = debug_info()->create_scope_values(locarray);
 967     DebugToken *expvals = debug_info()->create_scope_values(exparray);
 968     DebugToken *monvals = debug_info()->create_monitor_values(monarray);
 969 
 970     // Make method available for all Safepoints
 971     ciMethod* scope_method = method ? method : _method;
 972     // Describe the scope here
 973     assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
 974     assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
 975     // Now we can describe the scope.
 976     debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
 977   } // End jvms loop
 978 
 979   // Mark the end of the scope set.
 980   debug_info()->end_safepoint(safepoint_pc_offset);
 981 }
 982 
 983 
 984 
 985 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
 986 class NonSafepointEmitter {
 987   Compile*  C;
 988   JVMState* _pending_jvms;
 989   int       _pending_offset;
 990 
 991   void emit_non_safepoint();
 992 
 993  public:
 994   NonSafepointEmitter(Compile* compile) {
 995     this->C = compile;
 996     _pending_jvms = NULL;
 997     _pending_offset = 0;
 998   }
 999 
1000   void observe_instruction(Node* n, int pc_offset) {
1001     if (!C->debug_info()->recording_non_safepoints())  return;
1002 
1003     Node_Notes* nn = C->node_notes_at(n->_idx);
1004     if (nn == NULL || nn->jvms() == NULL)  return;
1005     if (_pending_jvms != NULL &&
1006         _pending_jvms->same_calls_as(nn->jvms())) {
1007       // Repeated JVMS?  Stretch it up here.
1008       _pending_offset = pc_offset;
1009     } else {
1010       if (_pending_jvms != NULL &&
1011           _pending_offset < pc_offset) {
1012         emit_non_safepoint();
1013       }
1014       _pending_jvms = NULL;
1015       if (pc_offset > C->debug_info()->last_pc_offset()) {
1016         // This is the only way _pending_jvms can become non-NULL:
1017         _pending_jvms = nn->jvms();
1018         _pending_offset = pc_offset;
1019       }
1020     }
1021   }
1022 
1023   // Stay out of the way of real safepoints:
1024   void observe_safepoint(JVMState* jvms, int pc_offset) {
1025     if (_pending_jvms != NULL &&
1026         !_pending_jvms->same_calls_as(jvms) &&
1027         _pending_offset < pc_offset) {
1028       emit_non_safepoint();
1029     }
1030     _pending_jvms = NULL;
1031   }
1032 
1033   void flush_at_end() {
1034     if (_pending_jvms != NULL) {
1035       emit_non_safepoint();
1036     }
1037     _pending_jvms = NULL;
1038   }
1039 };
1040 
1041 void NonSafepointEmitter::emit_non_safepoint() {
1042   JVMState* youngest_jvms = _pending_jvms;
1043   int       pc_offset     = _pending_offset;
1044 
1045   // Clear it now:
1046   _pending_jvms = NULL;
1047 
1048   DebugInformationRecorder* debug_info = C->debug_info();
1049   assert(debug_info->recording_non_safepoints(), "sanity");
1050 
1051   debug_info->add_non_safepoint(pc_offset);
1052   int max_depth = youngest_jvms->depth();
1053 
1054   // Visit scopes from oldest to youngest.
1055   for (int depth = 1; depth <= max_depth; depth++) {
1056     JVMState* jvms = youngest_jvms->of_depth(depth);
1057     ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1058     assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1059     debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1060   }
1061 
1062   // Mark the end of the scope set.
1063   debug_info->end_non_safepoint(pc_offset);
1064 }
1065 
1066 //------------------------------init_buffer------------------------------------
1067 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1068 
1069   // Set the initially allocated size
1070   int  code_req   = initial_code_capacity;
1071   int  locs_req   = initial_locs_capacity;
1072   int  stub_req   = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1073   int  const_req  = initial_const_capacity;
1074 
1075   int  pad_req    = NativeCall::instruction_size;
1076   // The extra spacing after the code is necessary on some platforms.
1077   // Sometimes we need to patch in a jump after the last instruction,
1078   // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
1079 
1080   // Compute the byte offset where we can store the deopt pc.
1081   if (fixed_slots() != 0) {
1082     _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1083   }
1084 
1085   // Compute prolog code size
1086   _method_size = 0;
1087   _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1088 #if defined(IA64) && !defined(AIX)
1089   if (save_argument_registers()) {
1090     // 4815101: this is a stub with implicit and unknown precision fp args.
1091     // The usual spill mechanism can only generate stfd's in this case, which
1092     // doesn't work if the fp reg to spill contains a single-precision denorm.
1093     // Instead, we hack around the normal spill mechanism using stfspill's and
1094     // ldffill's in the MachProlog and MachEpilog emit methods.  We allocate
1095     // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1096     //
1097     // If we ever implement 16-byte 'registers' == stack slots, we can
1098     // get rid of this hack and have SpillCopy generate stfspill/ldffill
1099     // instead of stfd/stfs/ldfd/ldfs.
1100     _frame_slots += 8*(16/BytesPerInt);
1101   }
1102 #endif
1103   assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1104 
1105   if (has_mach_constant_base_node()) {
1106     uint add_size = 0;
1107     // Fill the constant table.
1108     // Note:  This must happen before shorten_branches.
1109     for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1110       Block* b = _cfg->get_block(i);
1111 
1112       for (uint j = 0; j < b->number_of_nodes(); j++) {
1113         Node* n = b->get_node(j);
1114 
1115         // If the node is a MachConstantNode evaluate the constant
1116         // value section.
1117         if (n->is_MachConstant()) {
1118           MachConstantNode* machcon = n->as_MachConstant();
1119           machcon->eval_constant(C);
1120         } else if (n->is_Mach()) {
1121           // On Power there are more nodes that issue constants.
1122           add_size += (n->as_Mach()->ins_num_consts() * 8);
1123         }
1124       }
1125     }
1126 
1127     // Calculate the offsets of the constants and the size of the
1128     // constant table (including the padding to the next section).
1129     constant_table().calculate_offsets_and_size();
1130     const_req = constant_table().size() + add_size;
1131   }
1132 
1133   // Initialize the space for the BufferBlob used to find and verify
1134   // instruction size in MachNode::emit_size()
1135   init_scratch_buffer_blob(const_req);
1136   if (failing())  return NULL; // Out of memory
1137 
1138   // Pre-compute the length of blocks and replace
1139   // long branches with short if machine supports it.
1140   shorten_branches(blk_starts, code_req, locs_req, stub_req);
1141 
1142   // nmethod and CodeBuffer count stubs & constants as part of method's code.
1143   // class HandlerImpl is platform-specific and defined in the *.ad files.
1144   int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1145   int deopt_handler_req     = HandlerImpl::size_deopt_handler()     + MAX_stubs_size; // add marginal slop for handler
1146   stub_req += MAX_stubs_size;   // ensure per-stub margin
1147   code_req += MAX_inst_size;    // ensure per-instruction margin
1148 
1149   if (StressCodeBuffers)
1150     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
1151 
1152   int total_req =
1153     const_req +
1154     code_req +
1155     pad_req +
1156     stub_req +
1157     exception_handler_req +
1158     deopt_handler_req;               // deopt handler
1159 
1160   if (has_method_handle_invokes())
1161     total_req += deopt_handler_req;  // deopt MH handler
1162 
1163   CodeBuffer* cb = code_buffer();
1164   cb->initialize(total_req, locs_req);
1165 
1166   // Have we run out of code space?
1167   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1168     C->record_failure("CodeCache is full");
1169     return NULL;
1170   }
1171   // Configure the code buffer.
1172   cb->initialize_consts_size(const_req);
1173   cb->initialize_stubs_size(stub_req);
1174   cb->initialize_oop_recorder(env()->oop_recorder());
1175 
1176   // fill in the nop array for bundling computations
1177   MachNode *_nop_list[Bundle::_nop_count];
1178   Bundle::initialize_nops(_nop_list);
1179 
1180   return cb;
1181 }
1182 
1183 //------------------------------fill_buffer------------------------------------
1184 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1185   // blk_starts[] contains offsets calculated during short branches processing,
1186   // offsets should not be increased during following steps.
1187 
1188   // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1189   // of a loop. It is used to determine the padding for loop alignment.
1190   compute_loop_first_inst_sizes();
1191 
1192   // Create oopmap set.
1193   _oop_map_set = new OopMapSet();
1194 
1195   // !!!!! This preserves old handling of oopmaps for now
1196   debug_info()->set_oopmaps(_oop_map_set);
1197 
1198   uint nblocks  = _cfg->number_of_blocks();
1199   // Count and start of implicit null check instructions
1200   uint inct_cnt = 0;
1201   uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1202 
1203   // Count and start of calls
1204   uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1205 
1206   uint  return_offset = 0;
1207   int nop_size = (new MachNopNode())->size(_regalloc);
1208 
1209   int previous_offset = 0;
1210   int current_offset  = 0;
1211   int last_call_offset = -1;
1212   int last_avoid_back_to_back_offset = -1;
1213 #ifdef ASSERT
1214   uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1215   uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1216   uint* jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
1217   uint* jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
1218 #endif
1219 
1220   // Create an array of unused labels, one for each basic block, if printing is enabled
1221 #ifndef PRODUCT
1222   int *node_offsets      = NULL;
1223   uint node_offset_limit = unique();
1224 
1225   if (print_assembly())
1226     node_offsets         = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1227 #endif
1228 
1229   NonSafepointEmitter non_safepoints(this);  // emit non-safepoints lazily
1230 
1231   // Emit the constant table.
1232   if (has_mach_constant_base_node()) {
1233     constant_table().emit(*cb);
1234   }
1235 
1236   // Create an array of labels, one for each basic block
1237   Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1238   for (uint i=0; i <= nblocks; i++) {
1239     blk_labels[i].init();
1240   }
1241 
1242   // ------------------
1243   // Now fill in the code buffer
1244   Node *delay_slot = NULL;
1245 
1246   for (uint i = 0; i < nblocks; i++) {
1247     Block* block = _cfg->get_block(i);
1248     Node* head = block->head();
1249 
1250     // If this block needs to start aligned (i.e, can be reached other
1251     // than by falling-thru from the previous block), then force the
1252     // start of a new bundle.
1253     if (Pipeline::requires_bundling() && starts_bundle(head)) {
1254       cb->flush_bundle(true);
1255     }
1256 
1257 #ifdef ASSERT
1258     if (!block->is_connector()) {
1259       stringStream st;
1260       block->dump_head(_cfg, &st);
1261       MacroAssembler(cb).block_comment(st.as_string());
1262     }
1263     jmp_target[i] = 0;
1264     jmp_offset[i] = 0;
1265     jmp_size[i]   = 0;
1266     jmp_rule[i]   = 0;
1267 #endif
1268     int blk_offset = current_offset;
1269 
1270     // Define the label at the beginning of the basic block
1271     MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1272 
1273     uint last_inst = block->number_of_nodes();
1274 
1275     // Emit block normally, except for last instruction.
1276     // Emit means "dump code bits into code buffer".
1277     for (uint j = 0; j<last_inst; j++) {
1278 
1279       // Get the node
1280       Node* n = block->get_node(j);
1281 
1282       // See if delay slots are supported
1283       if (valid_bundle_info(n) &&
1284           node_bundling(n)->used_in_unconditional_delay()) {
1285         assert(delay_slot == NULL, "no use of delay slot node");
1286         assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1287 
1288         delay_slot = n;
1289         continue;
1290       }
1291 
1292       // If this starts a new instruction group, then flush the current one
1293       // (but allow split bundles)
1294       if (Pipeline::requires_bundling() && starts_bundle(n))
1295         cb->flush_bundle(false);
1296 
1297       // The following logic is duplicated in the code ifdeffed for
1298       // ENABLE_ZAP_DEAD_LOCALS which appears above in this file.  It
1299       // should be factored out.  Or maybe dispersed to the nodes?
1300 
1301       // Special handling for SafePoint/Call Nodes
1302       bool is_mcall = false;
1303       if (n->is_Mach()) {
1304         MachNode *mach = n->as_Mach();
1305         is_mcall = n->is_MachCall();
1306         bool is_sfn = n->is_MachSafePoint();
1307 
1308         // If this requires all previous instructions be flushed, then do so
1309         if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1310           cb->flush_bundle(true);
1311           current_offset = cb->insts_size();
1312         }
1313 
1314         // A padding may be needed again since a previous instruction
1315         // could be moved to delay slot.
1316 
1317         // align the instruction if necessary
1318         int padding = mach->compute_padding(current_offset);
1319         // Make sure safepoint node for polling is distinct from a call's
1320         // return by adding a nop if needed.
1321         if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1322           padding = nop_size;
1323         }
1324         if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1325             current_offset == last_avoid_back_to_back_offset) {
1326           // Avoid back to back some instructions.
1327           padding = nop_size;
1328         }
1329 
1330         if(padding > 0) {
1331           assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1332           int nops_cnt = padding / nop_size;
1333           MachNode *nop = new MachNopNode(nops_cnt);
1334           block->insert_node(nop, j++);
1335           last_inst++;
1336           _cfg->map_node_to_block(nop, block);
1337           nop->emit(*cb, _regalloc);
1338           cb->flush_bundle(true);
1339           current_offset = cb->insts_size();
1340         }
1341 
1342         // Remember the start of the last call in a basic block
1343         if (is_mcall) {
1344           MachCallNode *mcall = mach->as_MachCall();
1345 
1346           // This destination address is NOT PC-relative
1347           mcall->method_set((intptr_t)mcall->entry_point());
1348 
1349           // Save the return address
1350           call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1351 
1352           if (mcall->is_MachCallLeaf()) {
1353             is_mcall = false;
1354             is_sfn = false;
1355           }
1356         }
1357 
1358         // sfn will be valid whenever mcall is valid now because of inheritance
1359         if (is_sfn || is_mcall) {
1360 
1361           // Handle special safepoint nodes for synchronization
1362           if (!is_mcall) {
1363             MachSafePointNode *sfn = mach->as_MachSafePoint();
1364             // !!!!! Stubs only need an oopmap right now, so bail out
1365             if (sfn->jvms()->method() == NULL) {
1366               // Write the oopmap directly to the code blob??!!
1367 #             ifdef ENABLE_ZAP_DEAD_LOCALS
1368               assert( !is_node_getting_a_safepoint(sfn),  "logic does not match; false positive");
1369 #             endif
1370               continue;
1371             }
1372           } // End synchronization
1373 
1374           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1375                                            current_offset);
1376           Process_OopMap_Node(mach, current_offset);
1377         } // End if safepoint
1378 
1379         // If this is a null check, then add the start of the previous instruction to the list
1380         else if( mach->is_MachNullCheck() ) {
1381           inct_starts[inct_cnt++] = previous_offset;
1382         }
1383 
1384         // If this is a branch, then fill in the label with the target BB's label
1385         else if (mach->is_MachBranch()) {
1386           // This requires the TRUE branch target be in succs[0]
1387           uint block_num = block->non_connector_successor(0)->_pre_order;
1388 
1389           // Try to replace long branch if delay slot is not used,
1390           // it is mostly for back branches since forward branch's
1391           // distance is not updated yet.
1392           bool delay_slot_is_used = valid_bundle_info(n) &&
1393                                     node_bundling(n)->use_unconditional_delay();
1394           if (!delay_slot_is_used && mach->may_be_short_branch()) {
1395            assert(delay_slot == NULL, "not expecting delay slot node");
1396            int br_size = n->size(_regalloc);
1397             int offset = blk_starts[block_num] - current_offset;
1398             if (block_num >= i) {
1399               // Current and following block's offset are not
1400               // finalized yet, adjust distance by the difference
1401               // between calculated and final offsets of current block.
1402               offset -= (blk_starts[i] - blk_offset);
1403             }
1404             // In the following code a nop could be inserted before
1405             // the branch which will increase the backward distance.
1406             bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1407             if (needs_padding && offset <= 0)
1408               offset -= nop_size;
1409 
1410             if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1411               // We've got a winner.  Replace this branch.
1412               MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1413 
1414               // Update the jmp_size.
1415               int new_size = replacement->size(_regalloc);
1416               assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1417               // Insert padding between avoid_back_to_back branches.
1418               if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1419                 MachNode *nop = new MachNopNode();
1420                 block->insert_node(nop, j++);
1421                 _cfg->map_node_to_block(nop, block);
1422                 last_inst++;
1423                 nop->emit(*cb, _regalloc);
1424                 cb->flush_bundle(true);
1425                 current_offset = cb->insts_size();
1426               }
1427 #ifdef ASSERT
1428               jmp_target[i] = block_num;
1429               jmp_offset[i] = current_offset - blk_offset;
1430               jmp_size[i]   = new_size;
1431               jmp_rule[i]   = mach->rule();
1432 #endif
1433               block->map_node(replacement, j);
1434               mach->subsume_by(replacement, C);
1435               n    = replacement;
1436               mach = replacement;
1437             }
1438           }
1439           mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1440         } else if (mach->ideal_Opcode() == Op_Jump) {
1441           for (uint h = 0; h < block->_num_succs; h++) {
1442             Block* succs_block = block->_succs[h];
1443             for (uint j = 1; j < succs_block->num_preds(); j++) {
1444               Node* jpn = succs_block->pred(j);
1445               if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1446                 uint block_num = succs_block->non_connector()->_pre_order;
1447                 Label *blkLabel = &blk_labels[block_num];
1448                 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1449               }
1450             }
1451           }
1452         }
1453 #ifdef ASSERT
1454         // Check that oop-store precedes the card-mark
1455         else if (mach->ideal_Opcode() == Op_StoreCM) {
1456           uint storeCM_idx = j;
1457           int count = 0;
1458           for (uint prec = mach->req(); prec < mach->len(); prec++) {
1459             Node *oop_store = mach->in(prec);  // Precedence edge
1460             if (oop_store == NULL) continue;
1461             count++;
1462             uint i4;
1463             for (i4 = 0; i4 < last_inst; ++i4) {
1464               if (block->get_node(i4) == oop_store) {
1465                 break;
1466               }
1467             }
1468             // Note: This test can provide a false failure if other precedence
1469             // edges have been added to the storeCMNode.
1470             assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1471           }
1472           assert(count > 0, "storeCM expects at least one precedence edge");
1473         }
1474 #endif
1475         else if (!n->is_Proj()) {
1476           // Remember the beginning of the previous instruction, in case
1477           // it's followed by a flag-kill and a null-check.  Happens on
1478           // Intel all the time, with add-to-memory kind of opcodes.
1479           previous_offset = current_offset;
1480         }
1481 
1482         // Not an else-if!
1483         // If this is a trap based cmp then add its offset to the list.
1484         if (mach->is_TrapBasedCheckNode()) {
1485           inct_starts[inct_cnt++] = current_offset;
1486         }
1487       }
1488 
1489       // Verify that there is sufficient space remaining
1490       cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1491       if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1492         C->record_failure("CodeCache is full");
1493         return;
1494       }
1495 
1496       // Save the offset for the listing
1497 #ifndef PRODUCT
1498       if (node_offsets && n->_idx < node_offset_limit)
1499         node_offsets[n->_idx] = cb->insts_size();
1500 #endif
1501 
1502       // "Normal" instruction case
1503       DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1504       n->emit(*cb, _regalloc);
1505       current_offset  = cb->insts_size();
1506 
1507       // Above we only verified that there is enough space in the instruction section.
1508       // However, the instruction may emit stubs that cause code buffer expansion.
1509       // Bail out here if expansion failed due to a lack of code cache space.
1510       if (failing()) {
1511         return;
1512       }
1513 
1514 #ifdef ASSERT
1515       if (n->size(_regalloc) < (current_offset-instr_offset)) {
1516         n->dump();
1517         assert(false, "wrong size of mach node");
1518       }
1519 #endif
1520       non_safepoints.observe_instruction(n, current_offset);
1521 
1522       // mcall is last "call" that can be a safepoint
1523       // record it so we can see if a poll will directly follow it
1524       // in which case we'll need a pad to make the PcDesc sites unique
1525       // see  5010568. This can be slightly inaccurate but conservative
1526       // in the case that return address is not actually at current_offset.
1527       // This is a small price to pay.
1528 
1529       if (is_mcall) {
1530         last_call_offset = current_offset;
1531       }
1532 
1533       if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1534         // Avoid back to back some instructions.
1535         last_avoid_back_to_back_offset = current_offset;
1536       }
1537 
1538       // See if this instruction has a delay slot
1539       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1540         assert(delay_slot != NULL, "expecting delay slot node");
1541 
1542         // Back up 1 instruction
1543         cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1544 
1545         // Save the offset for the listing
1546 #ifndef PRODUCT
1547         if (node_offsets && delay_slot->_idx < node_offset_limit)
1548           node_offsets[delay_slot->_idx] = cb->insts_size();
1549 #endif
1550 
1551         // Support a SafePoint in the delay slot
1552         if (delay_slot->is_MachSafePoint()) {
1553           MachNode *mach = delay_slot->as_Mach();
1554           // !!!!! Stubs only need an oopmap right now, so bail out
1555           if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1556             // Write the oopmap directly to the code blob??!!
1557 #           ifdef ENABLE_ZAP_DEAD_LOCALS
1558             assert( !is_node_getting_a_safepoint(mach),  "logic does not match; false positive");
1559 #           endif
1560             delay_slot = NULL;
1561             continue;
1562           }
1563 
1564           int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1565           non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1566                                            adjusted_offset);
1567           // Generate an OopMap entry
1568           Process_OopMap_Node(mach, adjusted_offset);
1569         }
1570 
1571         // Insert the delay slot instruction
1572         delay_slot->emit(*cb, _regalloc);
1573 
1574         // Don't reuse it
1575         delay_slot = NULL;
1576       }
1577 
1578     } // End for all instructions in block
1579 
1580     // If the next block is the top of a loop, pad this block out to align
1581     // the loop top a little. Helps prevent pipe stalls at loop back branches.
1582     if (i < nblocks-1) {
1583       Block *nb = _cfg->get_block(i + 1);
1584       int padding = nb->alignment_padding(current_offset);
1585       if( padding > 0 ) {
1586         MachNode *nop = new MachNopNode(padding / nop_size);
1587         block->insert_node(nop, block->number_of_nodes());
1588         _cfg->map_node_to_block(nop, block);
1589         nop->emit(*cb, _regalloc);
1590         current_offset = cb->insts_size();
1591       }
1592     }
1593     // Verify that the distance for generated before forward
1594     // short branches is still valid.
1595     guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1596 
1597     // Save new block start offset
1598     blk_starts[i] = blk_offset;
1599   } // End of for all blocks
1600   blk_starts[nblocks] = current_offset;
1601 
1602   non_safepoints.flush_at_end();
1603 
1604   // Offset too large?
1605   if (failing())  return;
1606 
1607   // Define a pseudo-label at the end of the code
1608   MacroAssembler(cb).bind( blk_labels[nblocks] );
1609 
1610   // Compute the size of the first block
1611   _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1612 
1613   assert(cb->insts_size() < 500000, "method is unreasonably large");
1614 
1615 #ifdef ASSERT
1616   for (uint i = 0; i < nblocks; i++) { // For all blocks
1617     if (jmp_target[i] != 0) {
1618       int br_size = jmp_size[i];
1619       int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1620       if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1621         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]);
1622         assert(false, "Displacement too large for short jmp");
1623       }
1624     }
1625   }
1626 #endif
1627 
1628 #ifndef PRODUCT
1629   // Information on the size of the method, without the extraneous code
1630   Scheduling::increment_method_size(cb->insts_size());
1631 #endif
1632 
1633   // ------------------
1634   // Fill in exception table entries.
1635   FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1636 
1637   // Only java methods have exception handlers and deopt handlers
1638   // class HandlerImpl is platform-specific and defined in the *.ad files.
1639   if (_method) {
1640     // Emit the exception handler code.
1641     _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1642     if (failing()) {
1643       return; // CodeBuffer::expand failed
1644     }
1645     // Emit the deopt handler code.
1646     _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1647 
1648     // Emit the MethodHandle deopt handler code (if required).
1649     if (has_method_handle_invokes() && !failing()) {
1650       // We can use the same code as for the normal deopt handler, we
1651       // just need a different entry point address.
1652       _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1653     }
1654   }
1655 
1656   // One last check for failed CodeBuffer::expand:
1657   if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1658     C->record_failure("CodeCache is full");
1659     return;
1660   }
1661 
1662 #ifndef PRODUCT
1663   // Dump the assembly code, including basic-block numbers
1664   if (print_assembly()) {
1665     ttyLocker ttyl;  // keep the following output all in one block
1666     if (!VMThread::should_terminate()) {  // test this under the tty lock
1667       // This output goes directly to the tty, not the compiler log.
1668       // To enable tools to match it up with the compilation activity,
1669       // be sure to tag this tty output with the compile ID.
1670       if (xtty != NULL) {
1671         xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1672                    is_osr_compilation()    ? " compile_kind='osr'" :
1673                    "");
1674       }
1675       if (method() != NULL) {
1676         method()->print_metadata();
1677       }
1678       dump_asm(node_offsets, node_offset_limit);
1679       if (xtty != NULL) {
1680         xtty->tail("opto_assembly");
1681       }
1682     }
1683   }
1684 #endif
1685 
1686 }
1687 
1688 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1689   _inc_table.set_size(cnt);
1690 
1691   uint inct_cnt = 0;
1692   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1693     Block* block = _cfg->get_block(i);
1694     Node *n = NULL;
1695     int j;
1696 
1697     // Find the branch; ignore trailing NOPs.
1698     for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1699       n = block->get_node(j);
1700       if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1701         break;
1702       }
1703     }
1704 
1705     // If we didn't find anything, continue
1706     if (j < 0) {
1707       continue;
1708     }
1709 
1710     // Compute ExceptionHandlerTable subtable entry and add it
1711     // (skip empty blocks)
1712     if (n->is_Catch()) {
1713 
1714       // Get the offset of the return from the call
1715       uint call_return = call_returns[block->_pre_order];
1716 #ifdef ASSERT
1717       assert( call_return > 0, "no call seen for this basic block" );
1718       while (block->get_node(--j)->is_MachProj()) ;
1719       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1720 #endif
1721       // last instruction is a CatchNode, find it's CatchProjNodes
1722       int nof_succs = block->_num_succs;
1723       // allocate space
1724       GrowableArray<intptr_t> handler_bcis(nof_succs);
1725       GrowableArray<intptr_t> handler_pcos(nof_succs);
1726       // iterate through all successors
1727       for (int j = 0; j < nof_succs; j++) {
1728         Block* s = block->_succs[j];
1729         bool found_p = false;
1730         for (uint k = 1; k < s->num_preds(); k++) {
1731           Node* pk = s->pred(k);
1732           if (pk->is_CatchProj() && pk->in(0) == n) {
1733             const CatchProjNode* p = pk->as_CatchProj();
1734             found_p = true;
1735             // add the corresponding handler bci & pco information
1736             if (p->_con != CatchProjNode::fall_through_index) {
1737               // p leads to an exception handler (and is not fall through)
1738               assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1739               // no duplicates, please
1740               if (!handler_bcis.contains(p->handler_bci())) {
1741                 uint block_num = s->non_connector()->_pre_order;
1742                 handler_bcis.append(p->handler_bci());
1743                 handler_pcos.append(blk_labels[block_num].loc_pos());
1744               }
1745             }
1746           }
1747         }
1748         assert(found_p, "no matching predecessor found");
1749         // Note:  Due to empty block removal, one block may have
1750         // several CatchProj inputs, from the same Catch.
1751       }
1752 
1753       // Set the offset of the return from the call
1754       _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1755       continue;
1756     }
1757 
1758     // Handle implicit null exception table updates
1759     if (n->is_MachNullCheck()) {
1760       uint block_num = block->non_connector_successor(0)->_pre_order;
1761       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1762       continue;
1763     }
1764     // Handle implicit exception table updates: trap instructions.
1765     if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1766       uint block_num = block->non_connector_successor(0)->_pre_order;
1767       _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1768       continue;
1769     }
1770   } // End of for all blocks fill in exception table entries
1771 }
1772 
1773 // Static Variables
1774 #ifndef PRODUCT
1775 uint Scheduling::_total_nop_size = 0;
1776 uint Scheduling::_total_method_size = 0;
1777 uint Scheduling::_total_branches = 0;
1778 uint Scheduling::_total_unconditional_delays = 0;
1779 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1780 #endif
1781 
1782 // Initializer for class Scheduling
1783 
1784 Scheduling::Scheduling(Arena *arena, Compile &compile)
1785   : _arena(arena),
1786     _cfg(compile.cfg()),
1787     _regalloc(compile.regalloc()),
1788     _reg_node(arena),
1789     _bundle_instr_count(0),
1790     _bundle_cycle_number(0),
1791     _scheduled(arena),
1792     _available(arena),
1793     _next_node(NULL),
1794     _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1795     _pinch_free_list(arena)
1796 #ifndef PRODUCT
1797   , _branches(0)
1798   , _unconditional_delays(0)
1799 #endif
1800 {
1801   // Create a MachNopNode
1802   _nop = new MachNopNode();
1803 
1804   // Now that the nops are in the array, save the count
1805   // (but allow entries for the nops)
1806   _node_bundling_limit = compile.unique();
1807   uint node_max = _regalloc->node_regs_max_index();
1808 
1809   compile.set_node_bundling_limit(_node_bundling_limit);
1810 
1811   // This one is persistent within the Compile class
1812   _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1813 
1814   // Allocate space for fixed-size arrays
1815   _node_latency    = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1816   _uses            = NEW_ARENA_ARRAY(arena, short,          node_max);
1817   _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1818 
1819   // Clear the arrays
1820   memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1821   memset(_node_latency,       0, node_max * sizeof(unsigned short));
1822   memset(_uses,               0, node_max * sizeof(short));
1823   memset(_current_latency,    0, node_max * sizeof(unsigned short));
1824 
1825   // Clear the bundling information
1826   memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1827 
1828   // Get the last node
1829   Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1830 
1831   _next_node = block->get_node(block->number_of_nodes() - 1);
1832 }
1833 
1834 #ifndef PRODUCT
1835 // Scheduling destructor
1836 Scheduling::~Scheduling() {
1837   _total_branches             += _branches;
1838   _total_unconditional_delays += _unconditional_delays;
1839 }
1840 #endif
1841 
1842 // Step ahead "i" cycles
1843 void Scheduling::step(uint i) {
1844 
1845   Bundle *bundle = node_bundling(_next_node);
1846   bundle->set_starts_bundle();
1847 
1848   // Update the bundle record, but leave the flags information alone
1849   if (_bundle_instr_count > 0) {
1850     bundle->set_instr_count(_bundle_instr_count);
1851     bundle->set_resources_used(_bundle_use.resourcesUsed());
1852   }
1853 
1854   // Update the state information
1855   _bundle_instr_count = 0;
1856   _bundle_cycle_number += i;
1857   _bundle_use.step(i);
1858 }
1859 
1860 void Scheduling::step_and_clear() {
1861   Bundle *bundle = node_bundling(_next_node);
1862   bundle->set_starts_bundle();
1863 
1864   // Update the bundle record
1865   if (_bundle_instr_count > 0) {
1866     bundle->set_instr_count(_bundle_instr_count);
1867     bundle->set_resources_used(_bundle_use.resourcesUsed());
1868 
1869     _bundle_cycle_number += 1;
1870   }
1871 
1872   // Clear the bundling information
1873   _bundle_instr_count = 0;
1874   _bundle_use.reset();
1875 
1876   memcpy(_bundle_use_elements,
1877     Pipeline_Use::elaborated_elements,
1878     sizeof(Pipeline_Use::elaborated_elements));
1879 }
1880 
1881 // Perform instruction scheduling and bundling over the sequence of
1882 // instructions in backwards order.
1883 void Compile::ScheduleAndBundle() {
1884 
1885   // Don't optimize this if it isn't a method
1886   if (!_method)
1887     return;
1888 
1889   // Don't optimize this if scheduling is disabled
1890   if (!do_scheduling())
1891     return;
1892 
1893   // Scheduling code works only with pairs (16 bytes) maximum.
1894   if (max_vector_size() > 16)
1895     return;
1896 
1897   TracePhase tp("isched", &timers[_t_instrSched]);
1898 
1899   // Create a data structure for all the scheduling information
1900   Scheduling scheduling(Thread::current()->resource_area(), *this);
1901 
1902   // Walk backwards over each basic block, computing the needed alignment
1903   // Walk over all the basic blocks
1904   scheduling.DoScheduling();
1905 }
1906 
1907 // Compute the latency of all the instructions.  This is fairly simple,
1908 // because we already have a legal ordering.  Walk over the instructions
1909 // from first to last, and compute the latency of the instruction based
1910 // on the latency of the preceding instruction(s).
1911 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1912 #ifndef PRODUCT
1913   if (_cfg->C->trace_opto_output())
1914     tty->print("# -> ComputeLocalLatenciesForward\n");
1915 #endif
1916 
1917   // Walk over all the schedulable instructions
1918   for( uint j=_bb_start; j < _bb_end; j++ ) {
1919 
1920     // This is a kludge, forcing all latency calculations to start at 1.
1921     // Used to allow latency 0 to force an instruction to the beginning
1922     // of the bb
1923     uint latency = 1;
1924     Node *use = bb->get_node(j);
1925     uint nlen = use->len();
1926 
1927     // Walk over all the inputs
1928     for ( uint k=0; k < nlen; k++ ) {
1929       Node *def = use->in(k);
1930       if (!def)
1931         continue;
1932 
1933       uint l = _node_latency[def->_idx] + use->latency(k);
1934       if (latency < l)
1935         latency = l;
1936     }
1937 
1938     _node_latency[use->_idx] = latency;
1939 
1940 #ifndef PRODUCT
1941     if (_cfg->C->trace_opto_output()) {
1942       tty->print("# latency %4d: ", latency);
1943       use->dump();
1944     }
1945 #endif
1946   }
1947 
1948 #ifndef PRODUCT
1949   if (_cfg->C->trace_opto_output())
1950     tty->print("# <- ComputeLocalLatenciesForward\n");
1951 #endif
1952 
1953 } // end ComputeLocalLatenciesForward
1954 
1955 // See if this node fits into the present instruction bundle
1956 bool Scheduling::NodeFitsInBundle(Node *n) {
1957   uint n_idx = n->_idx;
1958 
1959   // If this is the unconditional delay instruction, then it fits
1960   if (n == _unconditional_delay_slot) {
1961 #ifndef PRODUCT
1962     if (_cfg->C->trace_opto_output())
1963       tty->print("#     NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1964 #endif
1965     return (true);
1966   }
1967 
1968   // If the node cannot be scheduled this cycle, skip it
1969   if (_current_latency[n_idx] > _bundle_cycle_number) {
1970 #ifndef PRODUCT
1971     if (_cfg->C->trace_opto_output())
1972       tty->print("#     NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1973         n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1974 #endif
1975     return (false);
1976   }
1977 
1978   const Pipeline *node_pipeline = n->pipeline();
1979 
1980   uint instruction_count = node_pipeline->instructionCount();
1981   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1982     instruction_count = 0;
1983   else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1984     instruction_count++;
1985 
1986   if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1987 #ifndef PRODUCT
1988     if (_cfg->C->trace_opto_output())
1989       tty->print("#     NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1990         n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1991 #endif
1992     return (false);
1993   }
1994 
1995   // Don't allow non-machine nodes to be handled this way
1996   if (!n->is_Mach() && instruction_count == 0)
1997     return (false);
1998 
1999   // See if there is any overlap
2000   uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2001 
2002   if (delay > 0) {
2003 #ifndef PRODUCT
2004     if (_cfg->C->trace_opto_output())
2005       tty->print("#     NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2006 #endif
2007     return false;
2008   }
2009 
2010 #ifndef PRODUCT
2011   if (_cfg->C->trace_opto_output())
2012     tty->print("#     NodeFitsInBundle [%4d]:  TRUE\n", n_idx);
2013 #endif
2014 
2015   return true;
2016 }
2017 
2018 Node * Scheduling::ChooseNodeToBundle() {
2019   uint siz = _available.size();
2020 
2021   if (siz == 0) {
2022 
2023 #ifndef PRODUCT
2024     if (_cfg->C->trace_opto_output())
2025       tty->print("#   ChooseNodeToBundle: NULL\n");
2026 #endif
2027     return (NULL);
2028   }
2029 
2030   // Fast path, if only 1 instruction in the bundle
2031   if (siz == 1) {
2032 #ifndef PRODUCT
2033     if (_cfg->C->trace_opto_output()) {
2034       tty->print("#   ChooseNodeToBundle (only 1): ");
2035       _available[0]->dump();
2036     }
2037 #endif
2038     return (_available[0]);
2039   }
2040 
2041   // Don't bother, if the bundle is already full
2042   if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2043     for ( uint i = 0; i < siz; i++ ) {
2044       Node *n = _available[i];
2045 
2046       // Skip projections, we'll handle them another way
2047       if (n->is_Proj())
2048         continue;
2049 
2050       // This presupposed that instructions are inserted into the
2051       // available list in a legality order; i.e. instructions that
2052       // must be inserted first are at the head of the list
2053       if (NodeFitsInBundle(n)) {
2054 #ifndef PRODUCT
2055         if (_cfg->C->trace_opto_output()) {
2056           tty->print("#   ChooseNodeToBundle: ");
2057           n->dump();
2058         }
2059 #endif
2060         return (n);
2061       }
2062     }
2063   }
2064 
2065   // Nothing fits in this bundle, choose the highest priority
2066 #ifndef PRODUCT
2067   if (_cfg->C->trace_opto_output()) {
2068     tty->print("#   ChooseNodeToBundle: ");
2069     _available[0]->dump();
2070   }
2071 #endif
2072 
2073   return _available[0];
2074 }
2075 
2076 void Scheduling::AddNodeToAvailableList(Node *n) {
2077   assert( !n->is_Proj(), "projections never directly made available" );
2078 #ifndef PRODUCT
2079   if (_cfg->C->trace_opto_output()) {
2080     tty->print("#   AddNodeToAvailableList: ");
2081     n->dump();
2082   }
2083 #endif
2084 
2085   int latency = _current_latency[n->_idx];
2086 
2087   // Insert in latency order (insertion sort)
2088   uint i;
2089   for ( i=0; i < _available.size(); i++ )
2090     if (_current_latency[_available[i]->_idx] > latency)
2091       break;
2092 
2093   // Special Check for compares following branches
2094   if( n->is_Mach() && _scheduled.size() > 0 ) {
2095     int op = n->as_Mach()->ideal_Opcode();
2096     Node *last = _scheduled[0];
2097     if( last->is_MachIf() && last->in(1) == n &&
2098         ( op == Op_CmpI ||
2099           op == Op_CmpU ||
2100           op == Op_CmpP ||
2101           op == Op_CmpF ||
2102           op == Op_CmpD ||
2103           op == Op_CmpL ) ) {
2104 
2105       // Recalculate position, moving to front of same latency
2106       for ( i=0 ; i < _available.size(); i++ )
2107         if (_current_latency[_available[i]->_idx] >= latency)
2108           break;
2109     }
2110   }
2111 
2112   // Insert the node in the available list
2113   _available.insert(i, n);
2114 
2115 #ifndef PRODUCT
2116   if (_cfg->C->trace_opto_output())
2117     dump_available();
2118 #endif
2119 }
2120 
2121 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2122   for ( uint i=0; i < n->len(); i++ ) {
2123     Node *def = n->in(i);
2124     if (!def) continue;
2125     if( def->is_Proj() )        // If this is a machine projection, then
2126       def = def->in(0);         // propagate usage thru to the base instruction
2127 
2128     if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2129       continue;
2130     }
2131 
2132     // Compute the latency
2133     uint l = _bundle_cycle_number + n->latency(i);
2134     if (_current_latency[def->_idx] < l)
2135       _current_latency[def->_idx] = l;
2136 
2137     // If this does not have uses then schedule it
2138     if ((--_uses[def->_idx]) == 0)
2139       AddNodeToAvailableList(def);
2140   }
2141 }
2142 
2143 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2144 #ifndef PRODUCT
2145   if (_cfg->C->trace_opto_output()) {
2146     tty->print("#   AddNodeToBundle: ");
2147     n->dump();
2148   }
2149 #endif
2150 
2151   // Remove this from the available list
2152   uint i;
2153   for (i = 0; i < _available.size(); i++)
2154     if (_available[i] == n)
2155       break;
2156   assert(i < _available.size(), "entry in _available list not found");
2157   _available.remove(i);
2158 
2159   // See if this fits in the current bundle
2160   const Pipeline *node_pipeline = n->pipeline();
2161   const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2162 
2163   // Check for instructions to be placed in the delay slot. We
2164   // do this before we actually schedule the current instruction,
2165   // because the delay slot follows the current instruction.
2166   if (Pipeline::_branch_has_delay_slot &&
2167       node_pipeline->hasBranchDelay() &&
2168       !_unconditional_delay_slot) {
2169 
2170     uint siz = _available.size();
2171 
2172     // Conditional branches can support an instruction that
2173     // is unconditionally executed and not dependent by the
2174     // branch, OR a conditionally executed instruction if
2175     // the branch is taken.  In practice, this means that
2176     // the first instruction at the branch target is
2177     // copied to the delay slot, and the branch goes to
2178     // the instruction after that at the branch target
2179     if ( n->is_MachBranch() ) {
2180 
2181       assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2182       assert( !n->is_Catch(),         "should not look for delay slot for Catch" );
2183 
2184 #ifndef PRODUCT
2185       _branches++;
2186 #endif
2187 
2188       // At least 1 instruction is on the available list
2189       // that is not dependent on the branch
2190       for (uint i = 0; i < siz; i++) {
2191         Node *d = _available[i];
2192         const Pipeline *avail_pipeline = d->pipeline();
2193 
2194         // Don't allow safepoints in the branch shadow, that will
2195         // cause a number of difficulties
2196         if ( avail_pipeline->instructionCount() == 1 &&
2197             !avail_pipeline->hasMultipleBundles() &&
2198             !avail_pipeline->hasBranchDelay() &&
2199             Pipeline::instr_has_unit_size() &&
2200             d->size(_regalloc) == Pipeline::instr_unit_size() &&
2201             NodeFitsInBundle(d) &&
2202             !node_bundling(d)->used_in_delay()) {
2203 
2204           if (d->is_Mach() && !d->is_MachSafePoint()) {
2205             // A node that fits in the delay slot was found, so we need to
2206             // set the appropriate bits in the bundle pipeline information so
2207             // that it correctly indicates resource usage.  Later, when we
2208             // attempt to add this instruction to the bundle, we will skip
2209             // setting the resource usage.
2210             _unconditional_delay_slot = d;
2211             node_bundling(n)->set_use_unconditional_delay();
2212             node_bundling(d)->set_used_in_unconditional_delay();
2213             _bundle_use.add_usage(avail_pipeline->resourceUse());
2214             _current_latency[d->_idx] = _bundle_cycle_number;
2215             _next_node = d;
2216             ++_bundle_instr_count;
2217 #ifndef PRODUCT
2218             _unconditional_delays++;
2219 #endif
2220             break;
2221           }
2222         }
2223       }
2224     }
2225 
2226     // No delay slot, add a nop to the usage
2227     if (!_unconditional_delay_slot) {
2228       // See if adding an instruction in the delay slot will overflow
2229       // the bundle.
2230       if (!NodeFitsInBundle(_nop)) {
2231 #ifndef PRODUCT
2232         if (_cfg->C->trace_opto_output())
2233           tty->print("#  *** STEP(1 instruction for delay slot) ***\n");
2234 #endif
2235         step(1);
2236       }
2237 
2238       _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2239       _next_node = _nop;
2240       ++_bundle_instr_count;
2241     }
2242 
2243     // See if the instruction in the delay slot requires a
2244     // step of the bundles
2245     if (!NodeFitsInBundle(n)) {
2246 #ifndef PRODUCT
2247         if (_cfg->C->trace_opto_output())
2248           tty->print("#  *** STEP(branch won't fit) ***\n");
2249 #endif
2250         // Update the state information
2251         _bundle_instr_count = 0;
2252         _bundle_cycle_number += 1;
2253         _bundle_use.step(1);
2254     }
2255   }
2256 
2257   // Get the number of instructions
2258   uint instruction_count = node_pipeline->instructionCount();
2259   if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2260     instruction_count = 0;
2261 
2262   // Compute the latency information
2263   uint delay = 0;
2264 
2265   if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2266     int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2267     if (relative_latency < 0)
2268       relative_latency = 0;
2269 
2270     delay = _bundle_use.full_latency(relative_latency, node_usage);
2271 
2272     // Does not fit in this bundle, start a new one
2273     if (delay > 0) {
2274       step(delay);
2275 
2276 #ifndef PRODUCT
2277       if (_cfg->C->trace_opto_output())
2278         tty->print("#  *** STEP(%d) ***\n", delay);
2279 #endif
2280     }
2281   }
2282 
2283   // If this was placed in the delay slot, ignore it
2284   if (n != _unconditional_delay_slot) {
2285 
2286     if (delay == 0) {
2287       if (node_pipeline->hasMultipleBundles()) {
2288 #ifndef PRODUCT
2289         if (_cfg->C->trace_opto_output())
2290           tty->print("#  *** STEP(multiple instructions) ***\n");
2291 #endif
2292         step(1);
2293       }
2294 
2295       else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2296 #ifndef PRODUCT
2297         if (_cfg->C->trace_opto_output())
2298           tty->print("#  *** STEP(%d >= %d instructions) ***\n",
2299             instruction_count + _bundle_instr_count,
2300             Pipeline::_max_instrs_per_cycle);
2301 #endif
2302         step(1);
2303       }
2304     }
2305 
2306     if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2307       _bundle_instr_count++;
2308 
2309     // Set the node's latency
2310     _current_latency[n->_idx] = _bundle_cycle_number;
2311 
2312     // Now merge the functional unit information
2313     if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2314       _bundle_use.add_usage(node_usage);
2315 
2316     // Increment the number of instructions in this bundle
2317     _bundle_instr_count += instruction_count;
2318 
2319     // Remember this node for later
2320     if (n->is_Mach())
2321       _next_node = n;
2322   }
2323 
2324   // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2325   // not in the bb->_nodes array.  This happens for debug-info-only BoxLocks.
2326   // 'Schedule' them (basically ignore in the schedule) but do not insert them
2327   // into the block.  All other scheduled nodes get put in the schedule here.
2328   int op = n->Opcode();
2329   if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2330       (op != Op_Node &&         // Not an unused antidepedence node and
2331        // not an unallocated boxlock
2332        (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2333 
2334     // Push any trailing projections
2335     if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2336       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2337         Node *foi = n->fast_out(i);
2338         if( foi->is_Proj() )
2339           _scheduled.push(foi);
2340       }
2341     }
2342 
2343     // Put the instruction in the schedule list
2344     _scheduled.push(n);
2345   }
2346 
2347 #ifndef PRODUCT
2348   if (_cfg->C->trace_opto_output())
2349     dump_available();
2350 #endif
2351 
2352   // Walk all the definitions, decrementing use counts, and
2353   // if a definition has a 0 use count, place it in the available list.
2354   DecrementUseCounts(n,bb);
2355 }
2356 
2357 // This method sets the use count within a basic block.  We will ignore all
2358 // uses outside the current basic block.  As we are doing a backwards walk,
2359 // any node we reach that has a use count of 0 may be scheduled.  This also
2360 // avoids the problem of cyclic references from phi nodes, as long as phi
2361 // nodes are at the front of the basic block.  This method also initializes
2362 // the available list to the set of instructions that have no uses within this
2363 // basic block.
2364 void Scheduling::ComputeUseCount(const Block *bb) {
2365 #ifndef PRODUCT
2366   if (_cfg->C->trace_opto_output())
2367     tty->print("# -> ComputeUseCount\n");
2368 #endif
2369 
2370   // Clear the list of available and scheduled instructions, just in case
2371   _available.clear();
2372   _scheduled.clear();
2373 
2374   // No delay slot specified
2375   _unconditional_delay_slot = NULL;
2376 
2377 #ifdef ASSERT
2378   for( uint i=0; i < bb->number_of_nodes(); i++ )
2379     assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2380 #endif
2381 
2382   // Force the _uses count to never go to zero for unscheduable pieces
2383   // of the block
2384   for( uint k = 0; k < _bb_start; k++ )
2385     _uses[bb->get_node(k)->_idx] = 1;
2386   for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2387     _uses[bb->get_node(l)->_idx] = 1;
2388 
2389   // Iterate backwards over the instructions in the block.  Don't count the
2390   // branch projections at end or the block header instructions.
2391   for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2392     Node *n = bb->get_node(j);
2393     if( n->is_Proj() ) continue; // Projections handled another way
2394 
2395     // Account for all uses
2396     for ( uint k = 0; k < n->len(); k++ ) {
2397       Node *inp = n->in(k);
2398       if (!inp) continue;
2399       assert(inp != n, "no cycles allowed" );
2400       if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2401         if (inp->is_Proj()) { // Skip through Proj's
2402           inp = inp->in(0);
2403         }
2404         ++_uses[inp->_idx];     // Count 1 block-local use
2405       }
2406     }
2407 
2408     // If this instruction has a 0 use count, then it is available
2409     if (!_uses[n->_idx]) {
2410       _current_latency[n->_idx] = _bundle_cycle_number;
2411       AddNodeToAvailableList(n);
2412     }
2413 
2414 #ifndef PRODUCT
2415     if (_cfg->C->trace_opto_output()) {
2416       tty->print("#   uses: %3d: ", _uses[n->_idx]);
2417       n->dump();
2418     }
2419 #endif
2420   }
2421 
2422 #ifndef PRODUCT
2423   if (_cfg->C->trace_opto_output())
2424     tty->print("# <- ComputeUseCount\n");
2425 #endif
2426 }
2427 
2428 // This routine performs scheduling on each basic block in reverse order,
2429 // using instruction latencies and taking into account function unit
2430 // availability.
2431 void Scheduling::DoScheduling() {
2432 #ifndef PRODUCT
2433   if (_cfg->C->trace_opto_output())
2434     tty->print("# -> DoScheduling\n");
2435 #endif
2436 
2437   Block *succ_bb = NULL;
2438   Block *bb;
2439 
2440   // Walk over all the basic blocks in reverse order
2441   for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2442     bb = _cfg->get_block(i);
2443 
2444 #ifndef PRODUCT
2445     if (_cfg->C->trace_opto_output()) {
2446       tty->print("#  Schedule BB#%03d (initial)\n", i);
2447       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2448         bb->get_node(j)->dump();
2449       }
2450     }
2451 #endif
2452 
2453     // On the head node, skip processing
2454     if (bb == _cfg->get_root_block()) {
2455       continue;
2456     }
2457 
2458     // Skip empty, connector blocks
2459     if (bb->is_connector())
2460       continue;
2461 
2462     // If the following block is not the sole successor of
2463     // this one, then reset the pipeline information
2464     if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2465 #ifndef PRODUCT
2466       if (_cfg->C->trace_opto_output()) {
2467         tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2468                    _next_node->_idx, _bundle_instr_count);
2469       }
2470 #endif
2471       step_and_clear();
2472     }
2473 
2474     // Leave untouched the starting instruction, any Phis, a CreateEx node
2475     // or Top.  bb->get_node(_bb_start) is the first schedulable instruction.
2476     _bb_end = bb->number_of_nodes()-1;
2477     for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2478       Node *n = bb->get_node(_bb_start);
2479       // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2480       // Also, MachIdealNodes do not get scheduled
2481       if( !n->is_Mach() ) continue;     // Skip non-machine nodes
2482       MachNode *mach = n->as_Mach();
2483       int iop = mach->ideal_Opcode();
2484       if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2485       if( iop == Op_Con ) continue;      // Do not schedule Top
2486       if( iop == Op_Node &&     // Do not schedule PhiNodes, ProjNodes
2487           mach->pipeline() == MachNode::pipeline_class() &&
2488           !n->is_SpillCopy() && !n->is_MachMerge() )  // Breakpoints, Prolog, etc
2489         continue;
2490       break;                    // Funny loop structure to be sure...
2491     }
2492     // Compute last "interesting" instruction in block - last instruction we
2493     // might schedule.  _bb_end points just after last schedulable inst.  We
2494     // normally schedule conditional branches (despite them being forced last
2495     // in the block), because they have delay slots we can fill.  Calls all
2496     // have their delay slots filled in the template expansions, so we don't
2497     // bother scheduling them.
2498     Node *last = bb->get_node(_bb_end);
2499     // Ignore trailing NOPs.
2500     while (_bb_end > 0 && last->is_Mach() &&
2501            last->as_Mach()->ideal_Opcode() == Op_Con) {
2502       last = bb->get_node(--_bb_end);
2503     }
2504     assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2505     if( last->is_Catch() ||
2506        // Exclude unreachable path case when Halt node is in a separate block.
2507        (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2508       // There must be a prior call.  Skip it.
2509       while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2510         assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2511       }
2512     } else if( last->is_MachNullCheck() ) {
2513       // Backup so the last null-checked memory instruction is
2514       // outside the schedulable range. Skip over the nullcheck,
2515       // projection, and the memory nodes.
2516       Node *mem = last->in(1);
2517       do {
2518         _bb_end--;
2519       } while (mem != bb->get_node(_bb_end));
2520     } else {
2521       // Set _bb_end to point after last schedulable inst.
2522       _bb_end++;
2523     }
2524 
2525     assert( _bb_start <= _bb_end, "inverted block ends" );
2526 
2527     // Compute the register antidependencies for the basic block
2528     ComputeRegisterAntidependencies(bb);
2529     if (_cfg->C->failing())  return;  // too many D-U pinch points
2530 
2531     // Compute intra-bb latencies for the nodes
2532     ComputeLocalLatenciesForward(bb);
2533 
2534     // Compute the usage within the block, and set the list of all nodes
2535     // in the block that have no uses within the block.
2536     ComputeUseCount(bb);
2537 
2538     // Schedule the remaining instructions in the block
2539     while ( _available.size() > 0 ) {
2540       Node *n = ChooseNodeToBundle();
2541       guarantee(n != NULL, "no nodes available");
2542       AddNodeToBundle(n,bb);
2543     }
2544 
2545     assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2546 #ifdef ASSERT
2547     for( uint l = _bb_start; l < _bb_end; l++ ) {
2548       Node *n = bb->get_node(l);
2549       uint m;
2550       for( m = 0; m < _bb_end-_bb_start; m++ )
2551         if( _scheduled[m] == n )
2552           break;
2553       assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2554     }
2555 #endif
2556 
2557     // Now copy the instructions (in reverse order) back to the block
2558     for ( uint k = _bb_start; k < _bb_end; k++ )
2559       bb->map_node(_scheduled[_bb_end-k-1], k);
2560 
2561 #ifndef PRODUCT
2562     if (_cfg->C->trace_opto_output()) {
2563       tty->print("#  Schedule BB#%03d (final)\n", i);
2564       uint current = 0;
2565       for (uint j = 0; j < bb->number_of_nodes(); j++) {
2566         Node *n = bb->get_node(j);
2567         if( valid_bundle_info(n) ) {
2568           Bundle *bundle = node_bundling(n);
2569           if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2570             tty->print("*** Bundle: ");
2571             bundle->dump();
2572           }
2573           n->dump();
2574         }
2575       }
2576     }
2577 #endif
2578 #ifdef ASSERT
2579   verify_good_schedule(bb,"after block local scheduling");
2580 #endif
2581   }
2582 
2583 #ifndef PRODUCT
2584   if (_cfg->C->trace_opto_output())
2585     tty->print("# <- DoScheduling\n");
2586 #endif
2587 
2588   // Record final node-bundling array location
2589   _regalloc->C->set_node_bundling_base(_node_bundling_base);
2590 
2591 } // end DoScheduling
2592 
2593 // Verify that no live-range used in the block is killed in the block by a
2594 // wrong DEF.  This doesn't verify live-ranges that span blocks.
2595 
2596 // Check for edge existence.  Used to avoid adding redundant precedence edges.
2597 static bool edge_from_to( Node *from, Node *to ) {
2598   for( uint i=0; i<from->len(); i++ )
2599     if( from->in(i) == to )
2600       return true;
2601   return false;
2602 }
2603 
2604 #ifdef ASSERT
2605 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2606   // Check for bad kills
2607   if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2608     Node *prior_use = _reg_node[def];
2609     if( prior_use && !edge_from_to(prior_use,n) ) {
2610       tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2611       n->dump();
2612       tty->print_cr("...");
2613       prior_use->dump();
2614       assert(edge_from_to(prior_use,n),msg);
2615     }
2616     _reg_node.map(def,NULL); // Kill live USEs
2617   }
2618 }
2619 
2620 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2621 
2622   // Zap to something reasonable for the verify code
2623   _reg_node.clear();
2624 
2625   // Walk over the block backwards.  Check to make sure each DEF doesn't
2626   // kill a live value (other than the one it's supposed to).  Add each
2627   // USE to the live set.
2628   for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2629     Node *n = b->get_node(i);
2630     int n_op = n->Opcode();
2631     if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2632       // Fat-proj kills a slew of registers
2633       RegMask rm = n->out_RegMask();// Make local copy
2634       while( rm.is_NotEmpty() ) {
2635         OptoReg::Name kill = rm.find_first_elem();
2636         rm.Remove(kill);
2637         verify_do_def( n, kill, msg );
2638       }
2639     } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2640       // Get DEF'd registers the normal way
2641       verify_do_def( n, _regalloc->get_reg_first(n), msg );
2642       verify_do_def( n, _regalloc->get_reg_second(n), msg );
2643     }
2644 
2645     // Now make all USEs live
2646     for( uint i=1; i<n->req(); i++ ) {
2647       Node *def = n->in(i);
2648       assert(def != 0, "input edge required");
2649       OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2650       OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2651       if( OptoReg::is_valid(reg_lo) ) {
2652         assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2653         _reg_node.map(reg_lo,n);
2654       }
2655       if( OptoReg::is_valid(reg_hi) ) {
2656         assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2657         _reg_node.map(reg_hi,n);
2658       }
2659     }
2660 
2661   }
2662 
2663   // Zap to something reasonable for the Antidependence code
2664   _reg_node.clear();
2665 }
2666 #endif
2667 
2668 // Conditionally add precedence edges.  Avoid putting edges on Projs.
2669 static void add_prec_edge_from_to( Node *from, Node *to ) {
2670   if( from->is_Proj() ) {       // Put precedence edge on Proj's input
2671     assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2672     from = from->in(0);
2673   }
2674   if( from != to &&             // No cycles (for things like LD L0,[L0+4] )
2675       !edge_from_to( from, to ) ) // Avoid duplicate edge
2676     from->add_prec(to);
2677 }
2678 
2679 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2680   if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2681     return;
2682 
2683   Node *pinch = _reg_node[def_reg]; // Get pinch point
2684   if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2685       is_def ) {    // Check for a true def (not a kill)
2686     _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2687     return;
2688   }
2689 
2690   Node *kill = def;             // Rename 'def' to more descriptive 'kill'
2691   debug_only( def = (Node*)0xdeadbeef; )
2692 
2693   // After some number of kills there _may_ be a later def
2694   Node *later_def = NULL;
2695 
2696   // Finding a kill requires a real pinch-point.
2697   // Check for not already having a pinch-point.
2698   // Pinch points are Op_Node's.
2699   if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2700     later_def = pinch;            // Must be def/kill as optimistic pinch-point
2701     if ( _pinch_free_list.size() > 0) {
2702       pinch = _pinch_free_list.pop();
2703     } else {
2704       pinch = new Node(1); // Pinch point to-be
2705     }
2706     if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2707       _cfg->C->record_method_not_compilable("too many D-U pinch points");
2708       return;
2709     }
2710     _cfg->map_node_to_block(pinch, b);      // Pretend it's valid in this block (lazy init)
2711     _reg_node.map(def_reg,pinch); // Record pinch-point
2712     //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2713     if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2714       pinch->init_req(0, _cfg->C->top());     // set not NULL for the next call
2715       add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2716       later_def = NULL;           // and no later def
2717     }
2718     pinch->set_req(0,later_def);  // Hook later def so we can find it
2719   } else {                        // Else have valid pinch point
2720     if( pinch->in(0) )            // If there is a later-def
2721       later_def = pinch->in(0);   // Get it
2722   }
2723 
2724   // Add output-dependence edge from later def to kill
2725   if( later_def )               // If there is some original def
2726     add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2727 
2728   // See if current kill is also a use, and so is forced to be the pinch-point.
2729   if( pinch->Opcode() == Op_Node ) {
2730     Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2731     for( uint i=1; i<uses->req(); i++ ) {
2732       if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2733           _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2734         // Yes, found a use/kill pinch-point
2735         pinch->set_req(0,NULL);  //
2736         pinch->replace_by(kill); // Move anti-dep edges up
2737         pinch = kill;
2738         _reg_node.map(def_reg,pinch);
2739         return;
2740       }
2741     }
2742   }
2743 
2744   // Add edge from kill to pinch-point
2745   add_prec_edge_from_to(kill,pinch);
2746 }
2747 
2748 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2749   if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2750     return;
2751   Node *pinch = _reg_node[use_reg]; // Get pinch point
2752   // Check for no later def_reg/kill in block
2753   if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2754       // Use has to be block-local as well
2755       _cfg->get_block_for_node(use) == b) {
2756     if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2757         pinch->req() == 1 ) {   // pinch not yet in block?
2758       pinch->del_req(0);        // yank pointer to later-def, also set flag
2759       // Insert the pinch-point in the block just after the last use
2760       b->insert_node(pinch, b->find_node(use) + 1);
2761       _bb_end++;                // Increase size scheduled region in block
2762     }
2763 
2764     add_prec_edge_from_to(pinch,use);
2765   }
2766 }
2767 
2768 // We insert antidependences between the reads and following write of
2769 // allocated registers to prevent illegal code motion. Hopefully, the
2770 // number of added references should be fairly small, especially as we
2771 // are only adding references within the current basic block.
2772 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2773 
2774 #ifdef ASSERT
2775   verify_good_schedule(b,"before block local scheduling");
2776 #endif
2777 
2778   // A valid schedule, for each register independently, is an endless cycle
2779   // of: a def, then some uses (connected to the def by true dependencies),
2780   // then some kills (defs with no uses), finally the cycle repeats with a new
2781   // def.  The uses are allowed to float relative to each other, as are the
2782   // kills.  No use is allowed to slide past a kill (or def).  This requires
2783   // antidependencies between all uses of a single def and all kills that
2784   // follow, up to the next def.  More edges are redundant, because later defs
2785   // & kills are already serialized with true or antidependencies.  To keep
2786   // the edge count down, we add a 'pinch point' node if there's more than
2787   // one use or more than one kill/def.
2788 
2789   // We add dependencies in one bottom-up pass.
2790 
2791   // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2792 
2793   // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2794   // register.  If not, we record the DEF/KILL in _reg_node, the
2795   // register-to-def mapping.  If there is a prior DEF/KILL, we insert a
2796   // "pinch point", a new Node that's in the graph but not in the block.
2797   // We put edges from the prior and current DEF/KILLs to the pinch point.
2798   // We put the pinch point in _reg_node.  If there's already a pinch point
2799   // we merely add an edge from the current DEF/KILL to the pinch point.
2800 
2801   // After doing the DEF/KILLs, we handle USEs.  For each used register, we
2802   // put an edge from the pinch point to the USE.
2803 
2804   // To be expedient, the _reg_node array is pre-allocated for the whole
2805   // compilation.  _reg_node is lazily initialized; it either contains a NULL,
2806   // or a valid def/kill/pinch-point, or a leftover node from some prior
2807   // block.  Leftover node from some prior block is treated like a NULL (no
2808   // prior def, so no anti-dependence needed).  Valid def is distinguished by
2809   // it being in the current block.
2810   bool fat_proj_seen = false;
2811   uint last_safept = _bb_end-1;
2812   Node* end_node         = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2813   Node* last_safept_node = end_node;
2814   for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2815     Node *n = b->get_node(i);
2816     int is_def = n->outcnt();   // def if some uses prior to adding precedence edges
2817     if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2818       // Fat-proj kills a slew of registers
2819       // This can add edges to 'n' and obscure whether or not it was a def,
2820       // hence the is_def flag.
2821       fat_proj_seen = true;
2822       RegMask rm = n->out_RegMask();// Make local copy
2823       while( rm.is_NotEmpty() ) {
2824         OptoReg::Name kill = rm.find_first_elem();
2825         rm.Remove(kill);
2826         anti_do_def( b, n, kill, is_def );
2827       }
2828     } else {
2829       // Get DEF'd registers the normal way
2830       anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2831       anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2832     }
2833 
2834     // Kill projections on a branch should appear to occur on the
2835     // branch, not afterwards, so grab the masks from the projections
2836     // and process them.
2837     if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2838       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2839         Node* use = n->fast_out(i);
2840         if (use->is_Proj()) {
2841           RegMask rm = use->out_RegMask();// Make local copy
2842           while( rm.is_NotEmpty() ) {
2843             OptoReg::Name kill = rm.find_first_elem();
2844             rm.Remove(kill);
2845             anti_do_def( b, n, kill, false );
2846           }
2847         }
2848       }
2849     }
2850 
2851     // Check each register used by this instruction for a following DEF/KILL
2852     // that must occur afterward and requires an anti-dependence edge.
2853     for( uint j=0; j<n->req(); j++ ) {
2854       Node *def = n->in(j);
2855       if( def ) {
2856         assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2857         anti_do_use( b, n, _regalloc->get_reg_first(def) );
2858         anti_do_use( b, n, _regalloc->get_reg_second(def) );
2859       }
2860     }
2861     // Do not allow defs of new derived values to float above GC
2862     // points unless the base is definitely available at the GC point.
2863 
2864     Node *m = b->get_node(i);
2865 
2866     // Add precedence edge from following safepoint to use of derived pointer
2867     if( last_safept_node != end_node &&
2868         m != last_safept_node) {
2869       for (uint k = 1; k < m->req(); k++) {
2870         const Type *t = m->in(k)->bottom_type();
2871         if( t->isa_oop_ptr() &&
2872             t->is_ptr()->offset() != 0 ) {
2873           last_safept_node->add_prec( m );
2874           break;
2875         }
2876       }
2877     }
2878 
2879     if( n->jvms() ) {           // Precedence edge from derived to safept
2880       // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2881       if( b->get_node(last_safept) != last_safept_node ) {
2882         last_safept = b->find_node(last_safept_node);
2883       }
2884       for( uint j=last_safept; j > i; j-- ) {
2885         Node *mach = b->get_node(j);
2886         if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2887           mach->add_prec( n );
2888       }
2889       last_safept = i;
2890       last_safept_node = m;
2891     }
2892   }
2893 
2894   if (fat_proj_seen) {
2895     // Garbage collect pinch nodes that were not consumed.
2896     // They are usually created by a fat kill MachProj for a call.
2897     garbage_collect_pinch_nodes();
2898   }
2899 }
2900 
2901 // Garbage collect pinch nodes for reuse by other blocks.
2902 //
2903 // The block scheduler's insertion of anti-dependence
2904 // edges creates many pinch nodes when the block contains
2905 // 2 or more Calls.  A pinch node is used to prevent a
2906 // combinatorial explosion of edges.  If a set of kills for a
2907 // register is anti-dependent on a set of uses (or defs), rather
2908 // than adding an edge in the graph between each pair of kill
2909 // and use (or def), a pinch is inserted between them:
2910 //
2911 //            use1   use2  use3
2912 //                \   |   /
2913 //                 \  |  /
2914 //                  pinch
2915 //                 /  |  \
2916 //                /   |   \
2917 //            kill1 kill2 kill3
2918 //
2919 // One pinch node is created per register killed when
2920 // the second call is encountered during a backwards pass
2921 // over the block.  Most of these pinch nodes are never
2922 // wired into the graph because the register is never
2923 // used or def'ed in the block.
2924 //
2925 void Scheduling::garbage_collect_pinch_nodes() {
2926 #ifndef PRODUCT
2927     if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2928 #endif
2929     int trace_cnt = 0;
2930     for (uint k = 0; k < _reg_node.Size(); k++) {
2931       Node* pinch = _reg_node[k];
2932       if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2933           // no predecence input edges
2934           (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2935         cleanup_pinch(pinch);
2936         _pinch_free_list.push(pinch);
2937         _reg_node.map(k, NULL);
2938 #ifndef PRODUCT
2939         if (_cfg->C->trace_opto_output()) {
2940           trace_cnt++;
2941           if (trace_cnt > 40) {
2942             tty->print("\n");
2943             trace_cnt = 0;
2944           }
2945           tty->print(" %d", pinch->_idx);
2946         }
2947 #endif
2948       }
2949     }
2950 #ifndef PRODUCT
2951     if (_cfg->C->trace_opto_output()) tty->print("\n");
2952 #endif
2953 }
2954 
2955 // Clean up a pinch node for reuse.
2956 void Scheduling::cleanup_pinch( Node *pinch ) {
2957   assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2958 
2959   for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2960     Node* use = pinch->last_out(i);
2961     uint uses_found = 0;
2962     for (uint j = use->req(); j < use->len(); j++) {
2963       if (use->in(j) == pinch) {
2964         use->rm_prec(j);
2965         uses_found++;
2966       }
2967     }
2968     assert(uses_found > 0, "must be a precedence edge");
2969     i -= uses_found;    // we deleted 1 or more copies of this edge
2970   }
2971   // May have a later_def entry
2972   pinch->set_req(0, NULL);
2973 }
2974 
2975 #ifndef PRODUCT
2976 
2977 void Scheduling::dump_available() const {
2978   tty->print("#Availist  ");
2979   for (uint i = 0; i < _available.size(); i++)
2980     tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2981   tty->cr();
2982 }
2983 
2984 // Print Scheduling Statistics
2985 void Scheduling::print_statistics() {
2986   // Print the size added by nops for bundling
2987   tty->print("Nops added %d bytes to total of %d bytes",
2988     _total_nop_size, _total_method_size);
2989   if (_total_method_size > 0)
2990     tty->print(", for %.2f%%",
2991       ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2992   tty->print("\n");
2993 
2994   // Print the number of branch shadows filled
2995   if (Pipeline::_branch_has_delay_slot) {
2996     tty->print("Of %d branches, %d had unconditional delay slots filled",
2997       _total_branches, _total_unconditional_delays);
2998     if (_total_branches > 0)
2999       tty->print(", for %.2f%%",
3000         ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3001     tty->print("\n");
3002   }
3003 
3004   uint total_instructions = 0, total_bundles = 0;
3005 
3006   for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3007     uint bundle_count   = _total_instructions_per_bundle[i];
3008     total_instructions += bundle_count * i;
3009     total_bundles      += bundle_count;
3010   }
3011 
3012   if (total_bundles > 0)
3013     tty->print("Average ILP (excluding nops) is %.2f\n",
3014       ((double)total_instructions) / ((double)total_bundles));
3015 }
3016 #endif