6743900: frequency based block layout

Summary: post-register allocation pass that drives block layout by edge frequencies
Reviewed-by: never, kvn

hotspot/src/share/vm/opto/block.hpp

@@ -75,6 +75,7 @@ public:
   void insert( uint i, Block *n );
   uint size() const { return _cnt; }
   void reset() { _cnt = 0; }
+  void print();
 };
 
 
@@ -129,7 +130,11 @@ class Block : public CFGElement {
   uint _rpo;                    // Number in reverse post order walk
 
   virtual bool is_block() { return true; }
-  float succ_prob(uint i); // return probability of i'th successor
+  float succ_prob(uint i);      // return probability of i'th successor
+  int num_fall_throughs();      // How many fall-through candidate this block has
+  void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code
+  bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate
+  Block* lone_fall_through();   // Return lone fall-through Block or null
 
   Block* dom_lca(Block* that);  // Compute LCA in dominator tree.
 #ifdef ASSERT
@@ -144,6 +149,7 @@ class Block : public CFGElement {
   // Report the alignment required by this block.  Must be a power of 2.
   // The previous block will insert nops to get this alignment.
   uint code_alignment();
+  uint compute_loop_alignment();
 
   // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies.
   // It is currently also used to scale such frequencies relative to
@@ -184,11 +190,12 @@ class Block : public CFGElement {
       int current_alignment = current_offset & max_pad;
       if( current_alignment != 0 ) {
         uint padding = (block_alignment-current_alignment) & max_pad;
-        if( !head()->is_Loop() ||
-            padding <= (uint)MaxLoopPad ||
-            first_inst_size() > padding ) {
-          return padding;
+        if( has_loop_alignment() &&
+            padding > (uint)MaxLoopPad &&
+            first_inst_size() <= padding ) {
+          return 0;
         }
+        return padding;
       }
     }
     return 0;
@@ -202,6 +209,21 @@ class Block : public CFGElement {
   void set_connector() { _connector = true; }
   bool is_connector() const { return _connector; };
 
+  // Loop_alignment will be set for blocks which are at the top of loops.
+  // The block layout pass may rotate loops such that the loop head may not
+  // be the sequentially first block of the loop encountered in the linear
+  // list of blocks.  If the layout pass is not run, loop alignment is set
+  // for each block which is the head of a loop.
+  uint _loop_alignment;
+  void set_loop_alignment(Block *loop_top) {
+    uint new_alignment = loop_top->compute_loop_alignment();
+    if (new_alignment > _loop_alignment) {
+      _loop_alignment = new_alignment;
+    }
+  }
+  uint loop_alignment() const { return _loop_alignment; }
+  bool has_loop_alignment() const { return loop_alignment() > 0; }
+
   // Create a new Block with given head Node.
   // Creates the (empty) predecessor arrays.
   Block( Arena *a, Node *headnode )
@@ -219,7 +241,8 @@ class Block : public CFGElement {
       _raise_LCA_mark(0),
       _raise_LCA_visited(0),
       _first_inst_size(999999),
-      _connector(false) {
+      _connector(false),
+      _loop_alignment(0) {
     _nodes.push(headnode);
   }
 
@@ -275,6 +298,16 @@ class Block : public CFGElement {
     return s;
   }
 
+  // Return true if b is a successor of this block
+  bool has_successor(Block* b) const {
+    for (uint i = 0; i < _num_succs; i++ ) {
+      if (non_connector_successor(i) == b) {
+        return true;
+      }
+    }
+    return false;
+  }
+
   // Successor block, after forwarding through connectors
   Block* non_connector_successor(int i) const {
     return _succs[i]->non_connector();
@@ -319,7 +352,6 @@ class PhaseCFG : public Phase {
 
   // I'll need a few machine-specific GotoNodes.  Clone from this one.
   MachNode *_goto;
-  void insert_goto_at(uint block_no, uint succ_no);
 
   Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);
   void verify_anti_dependences(Block* LCA, Node* load) {
@@ -379,10 +411,15 @@ class PhaseCFG : public Phase {
   // Compute the instruction global latency with a backwards walk
   void ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack);
 
+  // Set loop alignment
+  void set_loop_alignment();
+
   // Remove empty basic blocks
-  void RemoveEmpty();
-  bool MoveToNext(Block* bx, uint b_index);
-  void MoveToEnd(Block* bx, uint b_index);
+  void remove_empty();
+  void fixup_flow();
+  bool move_to_next(Block* bx, uint b_index);
+  void move_to_end(Block* bx, uint b_index);
+  void insert_goto_at(uint block_no, uint succ_no);
 
   // Check for NeverBranch at block end.  This needs to become a GOTO to the
   // true target.  NeverBranch are treated as a conditional branch that always
@@ -413,7 +450,7 @@ class PhaseCFG : public Phase {
 };
 
 
-//------------------------------UnionFindInfo----------------------------------
+//------------------------------UnionFind--------------------------------------
 // Map Block indices to a block-index for a cfg-cover.
 // Array lookup in the optimized case.
 class UnionFind : public ResourceObj {
@@ -508,3 +545,166 @@ class CFGLoop : public CFGElement {
   void dump_tree() const;
 #endif
 };
+
+
+//----------------------------------CFGEdge------------------------------------
+// A edge between two basic blocks that will be embodied by a branch or a
+// fall-through.
+class CFGEdge : public ResourceObj {
+ private:
+  Block * _from;        // Source basic block
+  Block * _to;          // Destination basic block
+  float _freq;          // Execution frequency (estimate)
+  int   _state;
+  bool  _infrequent;
+  int   _from_pct;
+  int   _to_pct;
+
+  // Private accessors
+  int  from_pct() const { return _from_pct; }
+  int  to_pct()   const { return _to_pct;   }
+  int  from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; }
+  int  to_infrequent()   const { return to_pct()   < BlockLayoutMinDiamondPercentage; }
+
+ public:
+  enum {
+    open,               // initial edge state; unprocessed
+    connected,          // edge used to connect two traces together
+    interior            // edge is interior to trace (could be backedge)
+  };
+
+  CFGEdge(Block *from, Block *to, float freq, int from_pct, int to_pct) :
+    _from(from), _to(to), _freq(freq),
+    _from_pct(from_pct), _to_pct(to_pct), _state(open) {
+    _infrequent = from_infrequent() || to_infrequent();
+  }
+
+  float  freq() const { return _freq; }
+  Block* from() const { return _from; }
+  Block* to  () const { return _to;   }
+  int  infrequent() const { return _infrequent; }
+  int state() const { return _state; }
+
+  void set_state(int state) { _state = state; }
+
+#ifndef PRODUCT
+  void dump( ) const;
+#endif
+};
+
+
+//-----------------------------------Trace-------------------------------------
+// An ordered list of basic blocks.
+class Trace : public ResourceObj {
+ private:
+  uint _id;             // Unique Trace id (derived from initial block)
+  Block ** _next_list;  // Array mapping index to next block
+  Block ** _prev_list;  // Array mapping index to previous block
+  Block * _first;       // First block in the trace
+  Block * _last;        // Last block in the trace
+
+  // Return the block that follows "b" in the trace.
+  Block * next(Block *b) const { return _next_list[b->_pre_order]; }
+  void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; }
+
+  // Return the block that preceeds "b" in the trace.
+  Block * prev(Block *b) const { return _prev_list[b->_pre_order]; }
+  void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; }
+
+  // We've discovered a loop in this trace. Reset last to be "b", and first as
+  // the block following "b
+  void break_loop_after(Block *b) {
+    _last = b;
+    _first = next(b);
+    set_prev(_first, NULL);
+    set_next(_last, NULL);
+  }
+
+ public:
+
+  Trace(Block *b, Block **next_list, Block **prev_list) :
+    _first(b),
+    _last(b),
+    _next_list(next_list),
+    _prev_list(prev_list),
+    _id(b->_pre_order) {
+    set_next(b, NULL);
+    set_prev(b, NULL);
+  };
+
+  // Return the id number
+  uint id() const { return _id; }
+  void set_id(uint id) { _id = id; }
+
+  // Return the first block in the trace
+  Block * first_block() const { return _first; }
+
+  // Return the last block in the trace
+  Block * last_block() const { return _last; }
+
+  // Insert a trace in the middle of this one after b
+  void insert_after(Block *b, Trace *tr) {
+    set_next(tr->last_block(), next(b));
+    if (next(b) != NULL) {
+      set_prev(next(b), tr->last_block());
+    }
+
+    set_next(b, tr->first_block());
+    set_prev(tr->first_block(), b);
+
+    if (b == _last) {
+      _last = tr->last_block();
+    }
+  }
+
+  void insert_before(Block *b, Trace *tr) {
+    Block *p = prev(b);
+    assert(p != NULL, "use append instead");
+    insert_after(p, tr);
+  }
+
+  // Append another trace to this one.
+  void append(Trace *tr) {
+    insert_after(_last, tr);
+  }
+
+  // Append a block at the end of this trace
+  void append(Block *b) {
+    set_next(_last, b);
+    set_prev(b, _last);
+    _last = b;
+  }
+
+  // Adjust the the blocks in this trace
+  void fixup_blocks(PhaseCFG &cfg);
+  bool backedge(CFGEdge *e);
+
+#ifndef PRODUCT
+  void dump( ) const;
+#endif
+};
+
+//------------------------------PhaseBlockLayout-------------------------------
+// Rearrange blocks into some canonical order, based on edges and their frequencies
+class PhaseBlockLayout : public Phase {
+  PhaseCFG &_cfg;               // Control flow graph
+
+  GrowableArray<CFGEdge *> *edges;
+  Trace **traces;
+  Block **next;
+  Block **prev;
+  UnionFind *uf;
+
+  // Given a block, find its encompassing Trace
+  Trace * trace(Block *b) {
+    return traces[uf->Find_compress(b->_pre_order)];
+  }
+ public:
+  PhaseBlockLayout(PhaseCFG &cfg);
+
+  void find_edges();
+  void grow_traces();
+  void merge_traces(bool loose_connections);
+  void reorder_traces(int count);
+  void union_traces(Trace* from, Trace* to);
+};

hotspot/src/share/vm/opto/c2_globals.hpp

@@ -396,5 +396,15 @@
                                                                             \
   diagnostic(intx, DominatorSearchLimit, 1000,                              \
           "Iterations limit in Node::dominates")                            \
+                                                                            \
+  product(bool, BlockLayoutByFrequency, true,                               \
+          "Use edge frequencies to drive block ordering")                   \
+                                                                            \
+  product(intx, BlockLayoutMinDiamondPercentage, 20,                        \
+          "Miniumum %% of a successor (predecessor) for which block layout "\
+          "a will allow a fork (join) in a single chain")                   \
+                                                                            \
+  product(bool, BlockLayoutRotateLoops, false,                              \
+          "Allow back branches to be fall throughs in the block layour")    \
 
 C2_FLAGS(DECLARE_DEVELOPER_FLAG, DECLARE_PD_DEVELOPER_FLAG, DECLARE_PRODUCT_FLAG, DECLARE_PD_PRODUCT_FLAG, DECLARE_DIAGNOSTIC_FLAG, DECLARE_NOTPRODUCT_FLAG)

hotspot/src/share/vm/opto/compile.cpp

@@ -822,6 +822,7 @@ void Compile::Init(int aliaslevel) {
   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
   set_decompile_count(0);
 
+  set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
   // Compilation level related initialization
   if (env()->comp_level() == CompLevel_fast_compile) {
     set_num_loop_opts(Tier1LoopOptsCount);
@@ -1701,8 +1702,14 @@ void Compile::Code_Gen() {
   // are not adding any new instructions.  If any basic block is empty, we
   // can now safely remove it.
   {
-    NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); )
-    cfg.RemoveEmpty();
+    NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
+    cfg.remove_empty();
+    if (do_freq_based_layout()) {
+      PhaseBlockLayout layout(cfg);
+    } else {
+      cfg.set_loop_alignment();
+    }
+    cfg.fixup_flow();
   }
 
   // Perform any platform dependent postallocation verifications.

hotspot/src/share/vm/opto/compile.hpp

@@ -154,6 +154,7 @@ class Compile : public Phase {
   uint                  _decompile_count;       // Cumulative decompilation counts.
   bool                  _do_inlining;           // True if we intend to do inlining
   bool                  _do_scheduling;         // True if we intend to do scheduling
+  bool                  _do_freq_based_layout;  // True if we intend to do frequency based block layout
   bool                  _do_count_invocations;  // True if we generate code to count invocations
   bool                  _do_method_data_update; // True if we generate code to update methodDataOops
   int                   _AliasLevel;            // Locally-adjusted version of AliasLevel flag.
@@ -307,6 +308,8 @@ class Compile : public Phase {
   void          set_do_inlining(bool z)         { _do_inlining = z; }
   bool              do_scheduling() const       { return _do_scheduling; }
   void          set_do_scheduling(bool z)       { _do_scheduling = z; }
+  bool              do_freq_based_layout() const{ return _do_freq_based_layout; }
+  void          set_do_freq_based_layout(bool z){ _do_freq_based_layout = z; }
   bool              do_count_invocations() const{ return _do_count_invocations; }
   void          set_do_count_invocations(bool z){ _do_count_invocations = z; }
   bool              do_method_data_update() const { return _do_method_data_update; }

hotspot/src/share/vm/opto/gcm.cpp

@@ -1319,11 +1319,33 @@ void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_
 //------------------------------Estimate_Block_Frequency-----------------------
 // Estimate block frequencies based on IfNode probabilities.
 void PhaseCFG::Estimate_Block_Frequency() {
-  int cnts = C->method() ? C->method()->interpreter_invocation_count() : 1;
-  // Most of our algorithms will die horribly if frequency can become
-  // negative so make sure cnts is a sane value.
-  if( cnts <= 0 ) cnts = 1;
-  float f = (float)cnts/(float)FreqCountInvocations;
+
+  // Force conditional branches leading to uncommon traps to be unlikely,
+  // not because we get to the uncommon_trap with less relative frequency,
+  // but because an uncommon_trap typically causes a deopt, so we only get
+  // there once.
+  if (C->do_freq_based_layout()) {
+    Block_List worklist;
+    Block* root_blk = _blocks[0];
+    for (uint i = 1; i < root_blk->num_preds(); i++) {
+      Block *pb = _bbs[root_blk->pred(i)->_idx];
+      if (pb->has_uncommon_code()) {
+        worklist.push(pb);
+      }
+    }
+    while (worklist.size() > 0) {
+      Block* uct = worklist.pop();
+      if (uct == _broot) continue;
+      for (uint i = 1; i < uct->num_preds(); i++) {
+        Block *pb = _bbs[uct->pred(i)->_idx];
+        if (pb->_num_succs == 1) {
+          worklist.push(pb);
+        } else if (pb->num_fall_throughs() == 2) {
+          pb->update_uncommon_branch(uct);
+        }
+      }
+    }
+  }
 
   // Create the loop tree and calculate loop depth.
   _root_loop = create_loop_tree();
@@ -1333,27 +1355,29 @@ void PhaseCFG::Estimate_Block_Frequency() {
   _root_loop->compute_freq();
 
   // Adjust all frequencies to be relative to a single method entry
-  _root_loop->_freq = f * 1.0;
+  _root_loop->_freq = 1.0;
   _root_loop->scale_freq();
 
   // force paths ending at uncommon traps to be infrequent
-  Block_List worklist;
-  Block* root_blk = _blocks[0];
-  for (uint i = 0; i < root_blk->num_preds(); i++) {
-    Block *pb = _bbs[root_blk->pred(i)->_idx];
-    if (pb->has_uncommon_code()) {
-      worklist.push(pb);
-    }
-  }
-  while (worklist.size() > 0) {
-    Block* uct = worklist.pop();
-    uct->_freq = PROB_MIN;
-    for (uint i = 0; i < uct->num_preds(); i++) {
-      Block *pb = _bbs[uct->pred(i)->_idx];
-      if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
+  if (!C->do_freq_based_layout()) {
+    Block_List worklist;
+    Block* root_blk = _blocks[0];
+    for (uint i = 1; i < root_blk->num_preds(); i++) {
+      Block *pb = _bbs[root_blk->pred(i)->_idx];
+      if (pb->has_uncommon_code()) {
         worklist.push(pb);
       }
     }
+    while (worklist.size() > 0) {
+      Block* uct = worklist.pop();
+      uct->_freq = PROB_MIN;
+      for (uint i = 1; i < uct->num_preds(); i++) {
+        Block *pb = _bbs[uct->pred(i)->_idx];
+        if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
+          worklist.push(pb);
+        }
+      }
+    }
   }
 
 #ifndef PRODUCT
@@ -1556,22 +1580,6 @@ void CFGLoop::compute_freq() {
     }
   }
 
-#if 0
-  // Raise frequency of the loop backedge block, in an effort
-  // to keep it empty.  Skip the method level "loop".
-  if (_parent != NULL) {
-    CFGElement* s = _members.at(_members.length() - 1);
-    if (s->is_block()) {
-      Block* bk = s->as_Block();
-      if (bk->_num_succs == 1 && bk->_succs[0] == hd) {
-        // almost any value >= 1.0f works
-        // FIXME: raw constant
-        bk->_freq = 1.05f;
-      }
-    }
-  }
-#endif
-
   // For all loops other than the outer, "method" loop,
   // sum and normalize the exit probability. The "method" loop
   // should keep the initial exit probability of 1, so that
@@ -1589,12 +1597,15 @@ void CFGLoop::compute_freq() {
     // the probability of exit per loop entry.
     for (int i = 0; i < _exits.length(); i++) {
       Block* et = _exits.at(i).get_target();
-      float new_prob = _exits.at(i).get_prob() / exits_sum;
+      float new_prob = 0.0f;
+      if (_exits.at(i).get_prob() > 0.0f) {
+        new_prob = _exits.at(i).get_prob() / exits_sum;
+      }
       BlockProbPair bpp(et, new_prob);
       _exits.at_put(i, bpp);
     }
 
-    // Save the total, but guard against unreasoable probability,
+    // Save the total, but guard against unreasonable probability,
     // as the value is used to estimate the loop trip count.
     // An infinite trip count would blur relative block
     // frequencies.
@@ -1688,6 +1699,137 @@ float Block::succ_prob(uint i) {
   return 0.0f;
 }
 
+//------------------------------num_fall_throughs-----------------------------
+// Return the number of fall-through candidates for a block
+int Block::num_fall_throughs() {
+  int eidx = end_idx();
+  Node *n = _nodes[eidx];  // Get ending Node
+
+  int op = n->Opcode();
+  if (n->is_Mach()) {
+    if (n->is_MachNullCheck()) {
+      // In theory, either side can fall-thru, for simplicity sake,
+      // let's say only the false branch can now.
+      return 1;
+    }
+    op = n->as_Mach()->ideal_Opcode();
+  }
+
+  // Switch on branch type
+  switch( op ) {
+  case Op_CountedLoopEnd:
+  case Op_If:
+    return 2;
+
+  case Op_Root:
+  case Op_Goto:
+    return 1;
+
+  case Op_Catch: {
+    for (uint i = 0; i < _num_succs; i++) {
+      const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
+      if (ci->_con == CatchProjNode::fall_through_index) {
+        return 1;
+      }
+    }
+    return 0;
+  }
+
+  case Op_Jump:
+  case Op_NeverBranch:
+  case Op_TailCall:
+  case Op_TailJump:
+  case Op_Return:
+  case Op_Halt:
+  case Op_Rethrow:
+    return 0;
+
+  default:
+    ShouldNotReachHere();
+  }
+
+  return 0;
+}
+
+//------------------------------succ_fall_through-----------------------------
+// Return true if a specific successor could be fall-through target.
+bool Block::succ_fall_through(uint i) {
+  int eidx = end_idx();
+  Node *n = _nodes[eidx];  // Get ending Node
+
+  int op = n->Opcode();
+  if (n->is_Mach()) {
+    if (n->is_MachNullCheck()) {
+      // In theory, either side can fall-thru, for simplicity sake,
+      // let's say only the false branch can now.
+      return _nodes[i + eidx + 1]->Opcode() == Op_IfFalse;
+    }
+    op = n->as_Mach()->ideal_Opcode();
+  }
+
+  // Switch on branch type
+  switch( op ) {
+  case Op_CountedLoopEnd:
+  case Op_If:
+  case Op_Root:
+  case Op_Goto:
+    return true;
+
+  case Op_Catch: {
+    const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
+    return ci->_con == CatchProjNode::fall_through_index;
+  }
+
+  case Op_Jump:
+  case Op_NeverBranch:
+  case Op_TailCall:
+  case Op_TailJump:
+  case Op_Return:
+  case Op_Halt:
+  case Op_Rethrow:
+    return false;
+
+  default:
+    ShouldNotReachHere();
+  }
+
+  return false;
+}
+
+//------------------------------update_uncommon_branch------------------------
+// Update the probability of a two-branch to be uncommon
+void Block::update_uncommon_branch(Block* ub) {
+  int eidx = end_idx();
+  Node *n = _nodes[eidx];  // Get ending Node
+
+  int op = n->as_Mach()->ideal_Opcode();
+
+  assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
+  assert(num_fall_throughs() == 2, "must be a two way branch block");
+
+  // Which successor is ub?
+  uint s;
+  for (s = 0; s <_num_succs; s++) {
+    if (_succs[s] == ub) break;
+  }
+  assert(s < 2, "uncommon successor must be found");
+
+  // If ub is the true path, make the proability small, else
+  // ub is the false path, and make the probability large
+  bool invert = (_nodes[s + eidx + 1]->Opcode() == Op_IfFalse);
+
+  // Get existing probability
+  float p = n->as_MachIf()->_prob;
+
+  if (invert) p = 1.0 - p;
+  if (p > PROB_MIN) {
+    p = PROB_MIN;
+  }
+  if (invert) p = 1.0 - p;
+
+  n->as_MachIf()->_prob = p;
+}
+
 //------------------------------update_succ_freq-------------------------------
 // Update the appropriate frequency associated with block 'b', a succesor of
 // a block in this loop.

hotspot/src/share/vm/opto/output.cpp

@@ -263,7 +263,7 @@ bool Compile::is_node_getting_a_safepoint( Node* n) {
 # endif // ENABLE_ZAP_DEAD_LOCALS
 
 //------------------------------compute_loop_first_inst_sizes------------------
-// Compute the size of first NumberOfLoopInstrToAlign instructions at head
+// Compute the size of first NumberOfLoopInstrToAlign instructions at the top
 // of a loop. When aligning a loop we need to provide enough instructions
 // in cpu's fetch buffer to feed decoders. The loop alignment could be
 // avoided if we have enough instructions in fetch buffer at the head of a loop.
@@ -284,34 +284,23 @@ void Compile::compute_loop_first_inst_sizes() {
     for( uint i=1; i <= last_block; i++ ) {
       Block *b = _cfg->_blocks[i];
       // Check the first loop's block which requires an alignment.
-      if( b->head()->is_Loop() &&
-          b->code_alignment() > (uint)relocInfo::addr_unit() ) {
+      if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {
         uint sum_size = 0;
         uint inst_cnt = NumberOfLoopInstrToAlign;
-        inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt,
-                                              _regalloc);
-        // Check the next fallthrough block if first loop's block does not have
-        // enough instructions.
-        if( inst_cnt > 0 && i < last_block ) {
-          // First, check if the first loop's block contains whole loop.
-          // LoopNode::LoopBackControl == 2.
-          Block *bx = _cfg->_bbs[b->pred(2)->_idx];
-          // Skip connector blocks (with limit in case of irreducible loops).
-          int search_limit = 16;
-          while( bx->is_connector() && search_limit-- > 0) {
-            bx = _cfg->_bbs[bx->pred(1)->_idx];
-          }
-          if( bx != b ) { // loop body is in several blocks.
-            Block *nb = NULL;
-            while( inst_cnt > 0 && i < last_block && nb != bx &&
-                  !_cfg->_blocks[i+1]->head()->is_Loop() ) {
-              i++;
-              nb = _cfg->_blocks[i];
-              inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt,
-                                                      _regalloc);
-            } // while( inst_cnt > 0 && i < last_block  )
-          } // if( bx != b )
-        } // if( inst_cnt > 0 && i < last_block )
+        inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
+
+        // Check subsequent fallthrough blocks if the loop's first
+        // block(s) does not have enough instructions.
+        Block *nb = b;
+        while( inst_cnt > 0 &&
+               i < last_block &&
+               !_cfg->_blocks[i+1]->has_loop_alignment() &&
+               !nb->has_successor(b) ) {
+          i++;
+          nb = _cfg->_blocks[i];
+          inst_cnt  = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
+        } // while( inst_cnt > 0 && i < last_block  )
+
         b->set_first_inst_size(sum_size);
       } // f( b->head()->is_Loop() )
     } // for( i <= last_block )
@@ -512,7 +501,7 @@ void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, i
       // Get the size of the block
       uint blk_size = adr - blk_starts[i];
 
-      // When the next block starts a loop, we may insert pad NOP
+      // When the next block is the top of a loop, we may insert pad NOP
       // instructions.
       Block *nb = _cfg->_blocks[i+1];
       int current_offset = blk_starts[i] + blk_size;
@@ -1382,8 +1371,8 @@ void Compile::Fill_buffer() {
 
     } // End for all instructions in block
 
-    // If the next block _starts_ a loop, pad this block out to align
-    // the loop start a little. Helps prevent pipe stalls at loop starts
+    // If the next block is the top of a loop, pad this block out to align
+    // the loop top a little. Helps prevent pipe stalls at loop back branches.
     int nop_size = (new (this) MachNopNode())->size(_regalloc);
     if( i<_cfg->_num_blocks-1 ) {
       Block *nb = _cfg->_blocks[i+1];

hotspot/src/share/vm/opto/phase.cpp

@@ -46,7 +46,7 @@ elapsedTimer Phase::_t_output;
 #ifndef PRODUCT
 elapsedTimer Phase::_t_graphReshaping;
 elapsedTimer Phase::_t_scheduler;
-elapsedTimer Phase::_t_removeEmptyBlocks;
+elapsedTimer Phase::_t_blockOrdering;
 elapsedTimer Phase::_t_macroExpand;
 elapsedTimer Phase::_t_peephole;
 elapsedTimer Phase::_t_codeGeneration;
@@ -128,7 +128,7 @@ void Phase::print_timers() {
     tty->print_cr ("      subtotal     : %3.3f sec,  %3.2f %%", regalloc_subtotal, percent_of_regalloc);
   }
   tty->print_cr ("    macroExpand  : %3.3f sec", Phase::_t_macroExpand.seconds());
-  tty->print_cr ("    removeEmpty  : %3.3f sec", Phase::_t_removeEmptyBlocks.seconds());
+  tty->print_cr ("    blockOrdering: %3.3f sec", Phase::_t_blockOrdering.seconds());
   tty->print_cr ("    peephole     : %3.3f sec", Phase::_t_peephole.seconds());
   tty->print_cr ("    codeGen      : %3.3f sec", Phase::_t_codeGeneration.seconds());
   tty->print_cr ("    install_code : %3.3f sec", Phase::_t_registerMethod.seconds());
@@ -137,7 +137,7 @@ void Phase::print_timers() {
     (DoEscapeAnalysis ? Phase::_t_escapeAnalysis.seconds() : 0.0) +
     Phase::_t_optimizer.seconds() + Phase::_t_graphReshaping.seconds() +
     Phase::_t_matcher.seconds() + Phase::_t_scheduler.seconds() +
-    Phase::_t_registerAllocation.seconds() + Phase::_t_removeEmptyBlocks.seconds() +
+    Phase::_t_registerAllocation.seconds() + Phase::_t_blockOrdering.seconds() +
     Phase::_t_macroExpand.seconds() + Phase::_t_peephole.seconds() +
     Phase::_t_codeGeneration.seconds() + Phase::_t_registerMethod.seconds();
   double percent_of_method_compile = ((phase_subtotal == 0.0) ? 0.0 : phase_subtotal / Phase::_t_methodCompilation.seconds()) * 100.0;

hotspot/src/share/vm/opto/phase.hpp

@@ -40,16 +40,12 @@ public:
     Optimistic,                 // Optimistic analysis phase
     GVN,                        // Pessimistic global value numbering phase
     Ins_Select,                 // Instruction selection phase
-    Copy_Elimination,           // Copy Elimination
-    Dead_Code_Elimination,      // DCE and compress Nodes
-    Conditional_Constant,       // Conditional Constant Propagation
     CFG,                        // Build a CFG
-    DefUse,                     // Build Def->Use chains
+    BlockLayout,                // Linear ordering of blocks
     Register_Allocation,        // Register allocation, duh
     LIVE,                       // Dragon-book LIVE range problem
     Interference_Graph,         // Building the IFG
     Coalesce,                   // Coalescing copies
-    Conditional_CProp,          // Conditional Constant Propagation
     Ideal_Loop,                 // Find idealized trip-counted loops
     Macro_Expand,               // Expand macro nodes
     Peephole,                   // Apply peephole optimizations
@@ -80,7 +76,7 @@ protected:
 #ifndef PRODUCT
   static elapsedTimer _t_graphReshaping;
   static elapsedTimer _t_scheduler;
-  static elapsedTimer _t_removeEmptyBlocks;
+  static elapsedTimer _t_blockOrdering;
   static elapsedTimer _t_macroExpand;
   static elapsedTimer _t_peephole;
   static elapsedTimer _t_codeGeneration;