src/share/vm/opto/block.hpp

Tue, 24 Dec 2013 11:48:39 -0800

author
mikael
date
Tue, 24 Dec 2013 11:48:39 -0800
changeset 6198
55fb97c4c58d
parent 5639
4b078f877b56
child 6503
a9becfeecd1b
permissions
-rw-r--r--

8029233: Update copyright year to match last edit in jdk8 hotspot repository for 2013
Summary: Copyright year updated for files modified during 2013
Reviewed-by: twisti, iveresov

duke@435 1 /*
mikael@6198 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
duke@435 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435 4 *
duke@435 5 * This code is free software; you can redistribute it and/or modify it
duke@435 6 * under the terms of the GNU General Public License version 2 only, as
duke@435 7 * published by the Free Software Foundation.
duke@435 8 *
duke@435 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@435 13 * accompanied this code).
duke@435 14 *
duke@435 15 * You should have received a copy of the GNU General Public License version
duke@435 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@435 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435 18 *
trims@1907 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907 20 * or visit www.oracle.com if you need additional information or have any
trims@1907 21 * questions.
duke@435 22 *
duke@435 23 */
duke@435 24
stefank@2314 25 #ifndef SHARE_VM_OPTO_BLOCK_HPP
stefank@2314 26 #define SHARE_VM_OPTO_BLOCK_HPP
stefank@2314 27
stefank@2314 28 #include "opto/multnode.hpp"
stefank@2314 29 #include "opto/node.hpp"
stefank@2314 30 #include "opto/phase.hpp"
stefank@2314 31
duke@435 32 // Optimization - Graph Style
duke@435 33
duke@435 34 class Block;
duke@435 35 class CFGLoop;
duke@435 36 class MachCallNode;
duke@435 37 class Matcher;
duke@435 38 class RootNode;
duke@435 39 class VectorSet;
duke@435 40 struct Tarjan;
duke@435 41
duke@435 42 //------------------------------Block_Array------------------------------------
duke@435 43 // Map dense integer indices to Blocks. Uses classic doubling-array trick.
duke@435 44 // Abstractly provides an infinite array of Block*'s, initialized to NULL.
duke@435 45 // Note that the constructor just zeros things, and since I use Arena
duke@435 46 // allocation I do not need a destructor to reclaim storage.
duke@435 47 class Block_Array : public ResourceObj {
never@3138 48 friend class VMStructs;
duke@435 49 uint _size; // allocated size, as opposed to formal limit
duke@435 50 debug_only(uint _limit;) // limit to formal domain
adlertz@5509 51 Arena *_arena; // Arena to allocate in
duke@435 52 protected:
duke@435 53 Block **_blocks;
duke@435 54 void grow( uint i ); // Grow array node to fit
duke@435 55
duke@435 56 public:
duke@435 57 Block_Array(Arena *a) : _arena(a), _size(OptoBlockListSize) {
duke@435 58 debug_only(_limit=0);
duke@435 59 _blocks = NEW_ARENA_ARRAY( a, Block *, OptoBlockListSize );
duke@435 60 for( int i = 0; i < OptoBlockListSize; i++ ) {
duke@435 61 _blocks[i] = NULL;
duke@435 62 }
duke@435 63 }
duke@435 64 Block *lookup( uint i ) const // Lookup, or NULL for not mapped
duke@435 65 { return (i<Max()) ? _blocks[i] : (Block*)NULL; }
duke@435 66 Block *operator[] ( uint i ) const // Lookup, or assert for not mapped
duke@435 67 { assert( i < Max(), "oob" ); return _blocks[i]; }
duke@435 68 // Extend the mapping: index i maps to Block *n.
duke@435 69 void map( uint i, Block *n ) { if( i>=Max() ) grow(i); _blocks[i] = n; }
duke@435 70 uint Max() const { debug_only(return _limit); return _size; }
duke@435 71 };
duke@435 72
duke@435 73
duke@435 74 class Block_List : public Block_Array {
never@3138 75 friend class VMStructs;
duke@435 76 public:
duke@435 77 uint _cnt;
duke@435 78 Block_List() : Block_Array(Thread::current()->resource_area()), _cnt(0) {}
adlertz@5509 79 void push( Block *b ) { map(_cnt++,b); }
duke@435 80 Block *pop() { return _blocks[--_cnt]; }
duke@435 81 Block *rpop() { Block *b = _blocks[0]; _blocks[0]=_blocks[--_cnt]; return b;}
duke@435 82 void remove( uint i );
duke@435 83 void insert( uint i, Block *n );
duke@435 84 uint size() const { return _cnt; }
duke@435 85 void reset() { _cnt = 0; }
rasbold@853 86 void print();
duke@435 87 };
duke@435 88
duke@435 89
duke@435 90 class CFGElement : public ResourceObj {
never@3138 91 friend class VMStructs;
duke@435 92 public:
duke@435 93 float _freq; // Execution frequency (estimate)
duke@435 94
duke@435 95 CFGElement() : _freq(0.0f) {}
duke@435 96 virtual bool is_block() { return false; }
duke@435 97 virtual bool is_loop() { return false; }
duke@435 98 Block* as_Block() { assert(is_block(), "must be block"); return (Block*)this; }
duke@435 99 CFGLoop* as_CFGLoop() { assert(is_loop(), "must be loop"); return (CFGLoop*)this; }
duke@435 100 };
duke@435 101
duke@435 102 //------------------------------Block------------------------------------------
duke@435 103 // This class defines a Basic Block.
duke@435 104 // Basic blocks are used during the output routines, and are not used during
duke@435 105 // any optimization pass. They are created late in the game.
duke@435 106 class Block : public CFGElement {
never@3138 107 friend class VMStructs;
adlertz@5635 108
adlertz@5635 109 private:
duke@435 110 // Nodes in this block, in order
duke@435 111 Node_List _nodes;
duke@435 112
adlertz@5635 113 public:
adlertz@5635 114
adlertz@5635 115 // Get the node at index 'at_index', if 'at_index' is out of bounds return NULL
adlertz@5635 116 Node* get_node(uint at_index) const {
adlertz@5635 117 return _nodes[at_index];
adlertz@5635 118 }
adlertz@5635 119
adlertz@5635 120 // Get the number of nodes in this block
adlertz@5635 121 uint number_of_nodes() const {
adlertz@5635 122 return _nodes.size();
adlertz@5635 123 }
adlertz@5635 124
adlertz@5635 125 // Map a node 'node' to index 'to_index' in the block, if the index is out of bounds the size of the node list is increased
adlertz@5635 126 void map_node(Node* node, uint to_index) {
adlertz@5635 127 _nodes.map(to_index, node);
adlertz@5635 128 }
adlertz@5635 129
adlertz@5635 130 // Insert a node 'node' at index 'at_index', moving all nodes that are on a higher index one step, if 'at_index' is out of bounds we crash
adlertz@5635 131 void insert_node(Node* node, uint at_index) {
adlertz@5635 132 _nodes.insert(at_index, node);
adlertz@5635 133 }
adlertz@5635 134
adlertz@5635 135 // Remove a node at index 'at_index'
adlertz@5635 136 void remove_node(uint at_index) {
adlertz@5635 137 _nodes.remove(at_index);
adlertz@5635 138 }
adlertz@5635 139
adlertz@5635 140 // Push a node 'node' onto the node list
adlertz@5635 141 void push_node(Node* node) {
adlertz@5635 142 _nodes.push(node);
adlertz@5635 143 }
adlertz@5635 144
adlertz@5635 145 // Pop the last node off the node list
adlertz@5635 146 Node* pop_node() {
adlertz@5635 147 return _nodes.pop();
adlertz@5635 148 }
adlertz@5635 149
duke@435 150 // Basic blocks have a Node which defines Control for all Nodes pinned in
duke@435 151 // this block. This Node is a RegionNode. Exception-causing Nodes
duke@435 152 // (division, subroutines) and Phi functions are always pinned. Later,
duke@435 153 // every Node will get pinned to some block.
adlertz@5635 154 Node *head() const { return get_node(0); }
duke@435 155
duke@435 156 // CAUTION: num_preds() is ONE based, so that predecessor numbers match
duke@435 157 // input edges to Regions and Phis.
duke@435 158 uint num_preds() const { return head()->req(); }
duke@435 159 Node *pred(uint i) const { return head()->in(i); }
duke@435 160
duke@435 161 // Array of successor blocks, same size as projs array
duke@435 162 Block_Array _succs;
duke@435 163
duke@435 164 // Basic blocks have some number of Nodes which split control to all
duke@435 165 // following blocks. These Nodes are always Projections. The field in
duke@435 166 // the Projection and the block-ending Node determine which Block follows.
duke@435 167 uint _num_succs;
duke@435 168
duke@435 169 // Basic blocks also carry all sorts of good old fashioned DFS information
duke@435 170 // used to find loops, loop nesting depth, dominators, etc.
duke@435 171 uint _pre_order; // Pre-order DFS number
duke@435 172
duke@435 173 // Dominator tree
duke@435 174 uint _dom_depth; // Depth in dominator tree for fast LCA
duke@435 175 Block* _idom; // Immediate dominator block
duke@435 176
duke@435 177 CFGLoop *_loop; // Loop to which this block belongs
duke@435 178 uint _rpo; // Number in reverse post order walk
duke@435 179
duke@435 180 virtual bool is_block() { return true; }
rasbold@853 181 float succ_prob(uint i); // return probability of i'th successor
rasbold@853 182 int num_fall_throughs(); // How many fall-through candidate this block has
rasbold@853 183 void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code
rasbold@853 184 bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate
rasbold@853 185 Block* lone_fall_through(); // Return lone fall-through Block or null
duke@435 186
duke@435 187 Block* dom_lca(Block* that); // Compute LCA in dominator tree.
duke@435 188 #ifdef ASSERT
duke@435 189 bool dominates(Block* that) {
duke@435 190 int dom_diff = this->_dom_depth - that->_dom_depth;
duke@435 191 if (dom_diff > 0) return false;
duke@435 192 for (; dom_diff < 0; dom_diff++) that = that->_idom;
duke@435 193 return this == that;
duke@435 194 }
duke@435 195 #endif
duke@435 196
duke@435 197 // Report the alignment required by this block. Must be a power of 2.
duke@435 198 // The previous block will insert nops to get this alignment.
duke@435 199 uint code_alignment();
rasbold@853 200 uint compute_loop_alignment();
duke@435 201
duke@435 202 // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies.
duke@435 203 // It is currently also used to scale such frequencies relative to
duke@435 204 // FreqCountInvocations relative to the old value of 1500.
duke@435 205 #define BLOCK_FREQUENCY(f) ((f * (float) 1500) / FreqCountInvocations)
duke@435 206
duke@435 207 // Register Pressure (estimate) for Splitting heuristic
duke@435 208 uint _reg_pressure;
duke@435 209 uint _ihrp_index;
duke@435 210 uint _freg_pressure;
duke@435 211 uint _fhrp_index;
duke@435 212
duke@435 213 // Mark and visited bits for an LCA calculation in insert_anti_dependences.
duke@435 214 // Since they hold unique node indexes, they do not need reinitialization.
duke@435 215 node_idx_t _raise_LCA_mark;
duke@435 216 void set_raise_LCA_mark(node_idx_t x) { _raise_LCA_mark = x; }
duke@435 217 node_idx_t raise_LCA_mark() const { return _raise_LCA_mark; }
duke@435 218 node_idx_t _raise_LCA_visited;
duke@435 219 void set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; }
duke@435 220 node_idx_t raise_LCA_visited() const { return _raise_LCA_visited; }
duke@435 221
duke@435 222 // Estimated size in bytes of first instructions in a loop.
duke@435 223 uint _first_inst_size;
duke@435 224 uint first_inst_size() const { return _first_inst_size; }
duke@435 225 void set_first_inst_size(uint s) { _first_inst_size = s; }
duke@435 226
duke@435 227 // Compute the size of first instructions in this block.
duke@435 228 uint compute_first_inst_size(uint& sum_size, uint inst_cnt, PhaseRegAlloc* ra);
duke@435 229
duke@435 230 // Compute alignment padding if the block needs it.
duke@435 231 // Align a loop if loop's padding is less or equal to padding limit
duke@435 232 // or the size of first instructions in the loop > padding.
duke@435 233 uint alignment_padding(int current_offset) {
duke@435 234 int block_alignment = code_alignment();
duke@435 235 int max_pad = block_alignment-relocInfo::addr_unit();
duke@435 236 if( max_pad > 0 ) {
duke@435 237 assert(is_power_of_2(max_pad+relocInfo::addr_unit()), "");
duke@435 238 int current_alignment = current_offset & max_pad;
duke@435 239 if( current_alignment != 0 ) {
duke@435 240 uint padding = (block_alignment-current_alignment) & max_pad;
rasbold@853 241 if( has_loop_alignment() &&
rasbold@853 242 padding > (uint)MaxLoopPad &&
rasbold@853 243 first_inst_size() <= padding ) {
rasbold@853 244 return 0;
duke@435 245 }
rasbold@853 246 return padding;
duke@435 247 }
duke@435 248 }
duke@435 249 return 0;
duke@435 250 }
duke@435 251
duke@435 252 // Connector blocks. Connector blocks are basic blocks devoid of
duke@435 253 // instructions, but may have relevant non-instruction Nodes, such as
duke@435 254 // Phis or MergeMems. Such blocks are discovered and marked during the
duke@435 255 // RemoveEmpty phase, and elided during Output.
duke@435 256 bool _connector;
duke@435 257 void set_connector() { _connector = true; }
duke@435 258 bool is_connector() const { return _connector; };
duke@435 259
rasbold@853 260 // Loop_alignment will be set for blocks which are at the top of loops.
rasbold@853 261 // The block layout pass may rotate loops such that the loop head may not
rasbold@853 262 // be the sequentially first block of the loop encountered in the linear
rasbold@853 263 // list of blocks. If the layout pass is not run, loop alignment is set
rasbold@853 264 // for each block which is the head of a loop.
rasbold@853 265 uint _loop_alignment;
rasbold@853 266 void set_loop_alignment(Block *loop_top) {
rasbold@853 267 uint new_alignment = loop_top->compute_loop_alignment();
rasbold@853 268 if (new_alignment > _loop_alignment) {
rasbold@853 269 _loop_alignment = new_alignment;
rasbold@853 270 }
rasbold@853 271 }
rasbold@853 272 uint loop_alignment() const { return _loop_alignment; }
rasbold@853 273 bool has_loop_alignment() const { return loop_alignment() > 0; }
rasbold@853 274
duke@435 275 // Create a new Block with given head Node.
duke@435 276 // Creates the (empty) predecessor arrays.
duke@435 277 Block( Arena *a, Node *headnode )
duke@435 278 : CFGElement(),
duke@435 279 _nodes(a),
duke@435 280 _succs(a),
duke@435 281 _num_succs(0),
duke@435 282 _pre_order(0),
duke@435 283 _idom(0),
duke@435 284 _loop(NULL),
duke@435 285 _reg_pressure(0),
duke@435 286 _ihrp_index(1),
duke@435 287 _freg_pressure(0),
duke@435 288 _fhrp_index(1),
duke@435 289 _raise_LCA_mark(0),
duke@435 290 _raise_LCA_visited(0),
duke@435 291 _first_inst_size(999999),
rasbold@853 292 _connector(false),
rasbold@853 293 _loop_alignment(0) {
duke@435 294 _nodes.push(headnode);
duke@435 295 }
duke@435 296
duke@435 297 // Index of 'end' Node
duke@435 298 uint end_idx() const {
duke@435 299 // %%%%% add a proj after every goto
duke@435 300 // so (last->is_block_proj() != last) always, then simplify this code
duke@435 301 // This will not give correct end_idx for block 0 when it only contains root.
duke@435 302 int last_idx = _nodes.size() - 1;
duke@435 303 Node *last = _nodes[last_idx];
duke@435 304 assert(last->is_block_proj() == last || last->is_block_proj() == _nodes[last_idx - _num_succs], "");
duke@435 305 return (last->is_block_proj() == last) ? last_idx : (last_idx - _num_succs);
duke@435 306 }
duke@435 307
duke@435 308 // Basic blocks have a Node which ends them. This Node determines which
duke@435 309 // basic block follows this one in the program flow. This Node is either an
duke@435 310 // IfNode, a GotoNode, a JmpNode, or a ReturnNode.
duke@435 311 Node *end() const { return _nodes[end_idx()]; }
duke@435 312
duke@435 313 // Add an instruction to an existing block. It must go after the head
duke@435 314 // instruction and before the end instruction.
adlertz@5635 315 void add_inst( Node *n ) { insert_node(n, end_idx()); }
duke@435 316 // Find node in block
duke@435 317 uint find_node( const Node *n ) const;
duke@435 318 // Find and remove n from block list
duke@435 319 void find_remove( const Node *n );
duke@435 320
duke@435 321 // Return the empty status of a block
duke@435 322 enum { not_empty, empty_with_goto, completely_empty };
duke@435 323 int is_Empty() const;
duke@435 324
duke@435 325 // Forward through connectors
duke@435 326 Block* non_connector() {
duke@435 327 Block* s = this;
duke@435 328 while (s->is_connector()) {
duke@435 329 s = s->_succs[0];
duke@435 330 }
duke@435 331 return s;
duke@435 332 }
duke@435 333
rasbold@853 334 // Return true if b is a successor of this block
rasbold@853 335 bool has_successor(Block* b) const {
rasbold@853 336 for (uint i = 0; i < _num_succs; i++ ) {
rasbold@853 337 if (non_connector_successor(i) == b) {
rasbold@853 338 return true;
rasbold@853 339 }
rasbold@853 340 }
rasbold@853 341 return false;
rasbold@853 342 }
rasbold@853 343
duke@435 344 // Successor block, after forwarding through connectors
duke@435 345 Block* non_connector_successor(int i) const {
duke@435 346 return _succs[i]->non_connector();
duke@435 347 }
duke@435 348
duke@435 349 // Examine block's code shape to predict if it is not commonly executed.
duke@435 350 bool has_uncommon_code() const;
duke@435 351
duke@435 352 #ifndef PRODUCT
duke@435 353 // Debugging print of basic block
kvn@3049 354 void dump_bidx(const Block* orig, outputStream* st = tty) const;
adlertz@5509 355 void dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st = tty) const;
adlertz@5509 356 void dump_head(const PhaseCFG* cfg, outputStream* st = tty) const;
kvn@3049 357 void dump() const;
adlertz@5509 358 void dump(const PhaseCFG* cfg) const;
duke@435 359 #endif
duke@435 360 };
duke@435 361
duke@435 362
duke@435 363 //------------------------------PhaseCFG---------------------------------------
duke@435 364 // Build an array of Basic Block pointers, one per Node.
duke@435 365 class PhaseCFG : public Phase {
never@3138 366 friend class VMStructs;
duke@435 367 private:
adlertz@5539 368
adlertz@5539 369 // Root of whole program
adlertz@5539 370 RootNode* _root;
adlertz@5539 371
adlertz@5539 372 // The block containing the root node
adlertz@5539 373 Block* _root_block;
adlertz@5539 374
adlertz@5539 375 // List of basic blocks that are created during CFG creation
adlertz@5539 376 Block_List _blocks;
adlertz@5539 377
adlertz@5539 378 // Count of basic blocks
adlertz@5539 379 uint _number_of_blocks;
adlertz@5539 380
adlertz@5509 381 // Arena for the blocks to be stored in
adlertz@5509 382 Arena* _block_arena;
adlertz@5509 383
adlertz@5539 384 // The matcher for this compilation
adlertz@5539 385 Matcher& _matcher;
adlertz@5539 386
adlertz@5509 387 // Map nodes to owning basic block
adlertz@5509 388 Block_Array _node_to_block_mapping;
adlertz@5509 389
adlertz@5539 390 // Loop from the root
adlertz@5539 391 CFGLoop* _root_loop;
adlertz@5539 392
adlertz@5539 393 // Outmost loop frequency
adlertz@5539 394 float _outer_loop_frequency;
adlertz@5539 395
adlertz@5539 396 // Per node latency estimation, valid only during GCM
adlertz@5539 397 GrowableArray<uint>* _node_latency;
adlertz@5539 398
duke@435 399 // Build a proper looking cfg. Return count of basic blocks
duke@435 400 uint build_cfg();
duke@435 401
adlertz@5539 402 // Build the dominator tree so that we know where we can move instructions
adlertz@5539 403 void build_dominator_tree();
adlertz@5539 404
adlertz@5539 405 // Estimate block frequencies based on IfNode probabilities, so that we know where we want to move instructions
adlertz@5539 406 void estimate_block_frequency();
adlertz@5539 407
adlertz@5539 408 // Global Code Motion. See Click's PLDI95 paper. Place Nodes in specific
adlertz@5539 409 // basic blocks; i.e. _node_to_block_mapping now maps _idx for all Nodes to some Block.
adlertz@5539 410 // Move nodes to ensure correctness from GVN and also try to move nodes out of loops.
adlertz@5539 411 void global_code_motion();
adlertz@5539 412
adlertz@5539 413 // Schedule Nodes early in their basic blocks.
adlertz@5539 414 bool schedule_early(VectorSet &visited, Node_List &roots);
adlertz@5539 415
adlertz@5539 416 // For each node, find the latest block it can be scheduled into
adlertz@5539 417 // and then select the cheapest block between the latest and earliest
adlertz@5539 418 // block to place the node.
adlertz@5539 419 void schedule_late(VectorSet &visited, Node_List &stack);
adlertz@5539 420
adlertz@5539 421 // Compute the (backwards) latency of a node from a single use
adlertz@5539 422 int latency_from_use(Node *n, const Node *def, Node *use);
adlertz@5539 423
adlertz@5539 424 // Compute the (backwards) latency of a node from the uses of this instruction
adlertz@5539 425 void partial_latency_of_defs(Node *n);
adlertz@5539 426
adlertz@5539 427 // Compute the instruction global latency with a backwards walk
adlertz@5539 428 void compute_latencies_backwards(VectorSet &visited, Node_List &stack);
adlertz@5539 429
adlertz@5539 430 // Pick a block between early and late that is a cheaper alternative
adlertz@5539 431 // to late. Helper for schedule_late.
adlertz@5539 432 Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self);
adlertz@5539 433
adlertz@5639 434 bool schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call);
adlertz@5639 435 void set_next_call(Block* block, Node* n, VectorSet& next_call);
adlertz@5639 436 void needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call);
adlertz@5639 437
adlertz@5639 438 // Perform basic-block local scheduling
adlertz@5639 439 Node* select(Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, VectorSet& next_call, uint sched_slot);
adlertz@5639 440
adlertz@5639 441 // Schedule a call next in the block
adlertz@5639 442 uint sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call);
adlertz@5639 443
adlertz@5639 444 // Cleanup if any code lands between a Call and his Catch
adlertz@5639 445 void call_catch_cleanup(Block* block);
adlertz@5639 446
adlertz@5639 447 Node* catch_cleanup_find_cloned_def(Block* use_blk, Node* def, Block* def_blk, int n_clone_idx);
adlertz@5639 448 void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx);
adlertz@5639 449
adlertz@5639 450 // Detect implicit-null-check opportunities. Basically, find NULL checks
adlertz@5639 451 // with suitable memory ops nearby. Use the memory op to do the NULL check.
adlertz@5639 452 // I can generate a memory op if there is not one nearby.
adlertz@5639 453 void implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons);
adlertz@5639 454
adlertz@5539 455 // Perform a Depth First Search (DFS).
duke@435 456 // Setup 'vertex' as DFS to vertex mapping.
duke@435 457 // Setup 'semi' as vertex to DFS mapping.
duke@435 458 // Set 'parent' to DFS parent.
adlertz@5539 459 uint do_DFS(Tarjan* tarjan, uint rpo_counter);
duke@435 460
duke@435 461 // Helper function to insert a node into a block
duke@435 462 void schedule_node_into_block( Node *n, Block *b );
duke@435 463
kvn@1039 464 void replace_block_proj_ctrl( Node *n );
kvn@1036 465
duke@435 466 // Set the basic block for pinned Nodes
duke@435 467 void schedule_pinned_nodes( VectorSet &visited );
duke@435 468
duke@435 469 // I'll need a few machine-specific GotoNodes. Clone from this one.
adlertz@5539 470 // Used when building the CFG and creating end nodes for blocks.
adlertz@5539 471 MachNode* _goto;
duke@435 472
duke@435 473 Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);
duke@435 474 void verify_anti_dependences(Block* LCA, Node* load) {
adlertz@5509 475 assert(LCA == get_block_for_node(load), "should already be scheduled");
duke@435 476 insert_anti_dependences(LCA, load, true);
duke@435 477 }
duke@435 478
adlertz@5539 479 bool move_to_next(Block* bx, uint b_index);
adlertz@5539 480 void move_to_end(Block* bx, uint b_index);
adlertz@5539 481
adlertz@5539 482 void insert_goto_at(uint block_no, uint succ_no);
adlertz@5539 483
adlertz@5539 484 // Check for NeverBranch at block end. This needs to become a GOTO to the
adlertz@5539 485 // true target. NeverBranch are treated as a conditional branch that always
adlertz@5539 486 // goes the same direction for most of the optimizer and are used to give a
adlertz@5539 487 // fake exit path to infinite loops. At this late stage they need to turn
adlertz@5539 488 // into Goto's so that when you enter the infinite loop you indeed hang.
adlertz@5539 489 void convert_NeverBranch_to_Goto(Block *b);
adlertz@5539 490
adlertz@5539 491 CFGLoop* create_loop_tree();
adlertz@5539 492
adlertz@5539 493 #ifndef PRODUCT
adlertz@5539 494 bool _trace_opto_pipelining; // tracing flag
adlertz@5539 495 #endif
adlertz@5539 496
duke@435 497 public:
adlertz@5509 498 PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher);
duke@435 499
adlertz@5539 500 void set_latency_for_node(Node* node, int latency) {
adlertz@5539 501 _node_latency->at_put_grow(node->_idx, latency);
adlertz@5539 502 }
duke@435 503
adlertz@5539 504 uint get_latency_for_node(Node* node) {
adlertz@5539 505 return _node_latency->at_grow(node->_idx);
adlertz@5539 506 }
adlertz@5539 507
adlertz@5539 508 // Get the outer most frequency
adlertz@5539 509 float get_outer_loop_frequency() const {
adlertz@5539 510 return _outer_loop_frequency;
adlertz@5539 511 }
adlertz@5539 512
adlertz@5539 513 // Get the root node of the CFG
adlertz@5539 514 RootNode* get_root_node() const {
adlertz@5539 515 return _root;
adlertz@5539 516 }
adlertz@5539 517
adlertz@5539 518 // Get the block of the root node
adlertz@5539 519 Block* get_root_block() const {
adlertz@5539 520 return _root_block;
adlertz@5539 521 }
adlertz@5539 522
adlertz@5539 523 // Add a block at a position and moves the later ones one step
adlertz@5539 524 void add_block_at(uint pos, Block* block) {
adlertz@5539 525 _blocks.insert(pos, block);
adlertz@5539 526 _number_of_blocks++;
adlertz@5539 527 }
adlertz@5539 528
adlertz@5539 529 // Adds a block to the top of the block list
adlertz@5539 530 void add_block(Block* block) {
adlertz@5539 531 _blocks.push(block);
adlertz@5539 532 _number_of_blocks++;
adlertz@5539 533 }
adlertz@5539 534
adlertz@5539 535 // Clear the list of blocks
adlertz@5539 536 void clear_blocks() {
adlertz@5539 537 _blocks.reset();
adlertz@5539 538 _number_of_blocks = 0;
adlertz@5539 539 }
adlertz@5539 540
adlertz@5539 541 // Get the block at position pos in _blocks
adlertz@5539 542 Block* get_block(uint pos) const {
adlertz@5539 543 return _blocks[pos];
adlertz@5539 544 }
adlertz@5539 545
adlertz@5539 546 // Number of blocks
adlertz@5539 547 uint number_of_blocks() const {
adlertz@5539 548 return _number_of_blocks;
adlertz@5539 549 }
adlertz@5509 550
adlertz@5509 551 // set which block this node should reside in
adlertz@5509 552 void map_node_to_block(const Node* node, Block* block) {
adlertz@5509 553 _node_to_block_mapping.map(node->_idx, block);
adlertz@5509 554 }
adlertz@5509 555
adlertz@5509 556 // removes the mapping from a node to a block
adlertz@5509 557 void unmap_node_from_block(const Node* node) {
adlertz@5509 558 _node_to_block_mapping.map(node->_idx, NULL);
adlertz@5509 559 }
adlertz@5509 560
adlertz@5509 561 // get the block in which this node resides
adlertz@5509 562 Block* get_block_for_node(const Node* node) const {
adlertz@5509 563 return _node_to_block_mapping[node->_idx];
adlertz@5509 564 }
adlertz@5509 565
adlertz@5509 566 // does this node reside in a block; return true
adlertz@5509 567 bool has_block(const Node* node) const {
adlertz@5509 568 return (_node_to_block_mapping.lookup(node->_idx) != NULL);
adlertz@5509 569 }
adlertz@5509 570
adlertz@5639 571 // Use frequency calculations and code shape to predict if the block
adlertz@5639 572 // is uncommon.
adlertz@5639 573 bool is_uncommon(const Block* block);
adlertz@5639 574
kvn@1268 575 #ifdef ASSERT
kvn@1268 576 Unique_Node_List _raw_oops;
kvn@1268 577 #endif
kvn@1268 578
adlertz@5539 579 // Do global code motion by first building dominator tree and estimate block frequency
adlertz@5539 580 // Returns true on success
adlertz@5539 581 bool do_global_code_motion();
duke@435 582
duke@435 583 // Compute the (backwards) latency of a node from the uses
duke@435 584 void latency_from_uses(Node *n);
duke@435 585
rasbold@853 586 // Set loop alignment
rasbold@853 587 void set_loop_alignment();
rasbold@853 588
duke@435 589 // Remove empty basic blocks
adlertz@5539 590 void remove_empty_blocks();
rasbold@853 591 void fixup_flow();
duke@435 592
adlertz@5539 593 // Insert a node into a block at index and map the node to the block
adlertz@5539 594 void insert(Block *b, uint idx, Node *n) {
adlertz@5635 595 b->insert_node(n , idx);
adlertz@5509 596 map_node_to_block(n, b);
duke@435 597 }
duke@435 598
duke@435 599 #ifndef PRODUCT
duke@435 600 bool trace_opto_pipelining() const { return _trace_opto_pipelining; }
duke@435 601
duke@435 602 // Debugging print of CFG
duke@435 603 void dump( ) const; // CFG only
duke@435 604 void _dump_cfg( const Node *end, VectorSet &visited ) const;
duke@435 605 void verify() const;
duke@435 606 void dump_headers();
duke@435 607 #else
duke@435 608 bool trace_opto_pipelining() const { return false; }
duke@435 609 #endif
duke@435 610 };
duke@435 611
duke@435 612
rasbold@853 613 //------------------------------UnionFind--------------------------------------
duke@435 614 // Map Block indices to a block-index for a cfg-cover.
duke@435 615 // Array lookup in the optimized case.
duke@435 616 class UnionFind : public ResourceObj {
duke@435 617 uint _cnt, _max;
duke@435 618 uint* _indices;
duke@435 619 ReallocMark _nesting; // assertion check for reallocations
duke@435 620 public:
duke@435 621 UnionFind( uint max );
duke@435 622 void reset( uint max ); // Reset to identity map for [0..max]
duke@435 623
duke@435 624 uint lookup( uint nidx ) const {
duke@435 625 return _indices[nidx];
duke@435 626 }
duke@435 627 uint operator[] (uint nidx) const { return lookup(nidx); }
duke@435 628
duke@435 629 void map( uint from_idx, uint to_idx ) {
duke@435 630 assert( from_idx < _cnt, "oob" );
duke@435 631 _indices[from_idx] = to_idx;
duke@435 632 }
duke@435 633 void extend( uint from_idx, uint to_idx );
duke@435 634
duke@435 635 uint Size() const { return _cnt; }
duke@435 636
duke@435 637 uint Find( uint idx ) {
duke@435 638 assert( idx < 65536, "Must fit into uint");
duke@435 639 uint uf_idx = lookup(idx);
duke@435 640 return (uf_idx == idx) ? uf_idx : Find_compress(idx);
duke@435 641 }
duke@435 642 uint Find_compress( uint idx );
duke@435 643 uint Find_const( uint idx ) const;
duke@435 644 void Union( uint idx1, uint idx2 );
duke@435 645
duke@435 646 };
duke@435 647
duke@435 648 //----------------------------BlockProbPair---------------------------
duke@435 649 // Ordered pair of Node*.
duke@435 650 class BlockProbPair VALUE_OBJ_CLASS_SPEC {
duke@435 651 protected:
duke@435 652 Block* _target; // block target
duke@435 653 float _prob; // probability of edge to block
duke@435 654 public:
duke@435 655 BlockProbPair() : _target(NULL), _prob(0.0) {}
duke@435 656 BlockProbPair(Block* b, float p) : _target(b), _prob(p) {}
duke@435 657
duke@435 658 Block* get_target() const { return _target; }
duke@435 659 float get_prob() const { return _prob; }
duke@435 660 };
duke@435 661
duke@435 662 //------------------------------CFGLoop-------------------------------------------
duke@435 663 class CFGLoop : public CFGElement {
never@3138 664 friend class VMStructs;
duke@435 665 int _id;
duke@435 666 int _depth;
duke@435 667 CFGLoop *_parent; // root of loop tree is the method level "pseudo" loop, it's parent is null
duke@435 668 CFGLoop *_sibling; // null terminated list
duke@435 669 CFGLoop *_child; // first child, use child's sibling to visit all immediately nested loops
duke@435 670 GrowableArray<CFGElement*> _members; // list of members of loop
duke@435 671 GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities
duke@435 672 float _exit_prob; // probability any loop exit is taken on a single loop iteration
duke@435 673 void update_succ_freq(Block* b, float freq);
duke@435 674
duke@435 675 public:
duke@435 676 CFGLoop(int id) :
duke@435 677 CFGElement(),
duke@435 678 _id(id),
duke@435 679 _depth(0),
duke@435 680 _parent(NULL),
duke@435 681 _sibling(NULL),
duke@435 682 _child(NULL),
duke@435 683 _exit_prob(1.0f) {}
duke@435 684 CFGLoop* parent() { return _parent; }
adlertz@5509 685 void push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg);
duke@435 686 void add_member(CFGElement *s) { _members.push(s); }
duke@435 687 void add_nested_loop(CFGLoop* cl);
duke@435 688 Block* head() {
duke@435 689 assert(_members.at(0)->is_block(), "head must be a block");
duke@435 690 Block* hd = _members.at(0)->as_Block();
duke@435 691 assert(hd->_loop == this, "just checking");
duke@435 692 assert(hd->head()->is_Loop(), "must begin with loop head node");
duke@435 693 return hd;
duke@435 694 }
duke@435 695 Block* backedge_block(); // Return the block on the backedge of the loop (else NULL)
duke@435 696 void compute_loop_depth(int depth);
duke@435 697 void compute_freq(); // compute frequency with loop assuming head freq 1.0f
duke@435 698 void scale_freq(); // scale frequency by loop trip count (including outer loops)
kvn@1108 699 float outer_loop_freq() const; // frequency of outer loop
duke@435 700 bool in_loop_nest(Block* b);
duke@435 701 float trip_count() const { return 1.0f / _exit_prob; }
duke@435 702 virtual bool is_loop() { return true; }
duke@435 703 int id() { return _id; }
duke@435 704
duke@435 705 #ifndef PRODUCT
duke@435 706 void dump( ) const;
duke@435 707 void dump_tree() const;
duke@435 708 #endif
duke@435 709 };
rasbold@853 710
rasbold@853 711
rasbold@853 712 //----------------------------------CFGEdge------------------------------------
rasbold@853 713 // A edge between two basic blocks that will be embodied by a branch or a
rasbold@853 714 // fall-through.
rasbold@853 715 class CFGEdge : public ResourceObj {
never@3138 716 friend class VMStructs;
rasbold@853 717 private:
rasbold@853 718 Block * _from; // Source basic block
rasbold@853 719 Block * _to; // Destination basic block
rasbold@853 720 float _freq; // Execution frequency (estimate)
rasbold@853 721 int _state;
rasbold@853 722 bool _infrequent;
rasbold@853 723 int _from_pct;
rasbold@853 724 int _to_pct;
rasbold@853 725
rasbold@853 726 // Private accessors
rasbold@853 727 int from_pct() const { return _from_pct; }
rasbold@853 728 int to_pct() const { return _to_pct; }
rasbold@853 729 int from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; }
rasbold@853 730 int to_infrequent() const { return to_pct() < BlockLayoutMinDiamondPercentage; }
rasbold@853 731
rasbold@853 732 public:
rasbold@853 733 enum {
rasbold@853 734 open, // initial edge state; unprocessed
rasbold@853 735 connected, // edge used to connect two traces together
rasbold@853 736 interior // edge is interior to trace (could be backedge)
rasbold@853 737 };
rasbold@853 738
rasbold@853 739 CFGEdge(Block *from, Block *to, float freq, int from_pct, int to_pct) :
rasbold@853 740 _from(from), _to(to), _freq(freq),
rasbold@853 741 _from_pct(from_pct), _to_pct(to_pct), _state(open) {
rasbold@853 742 _infrequent = from_infrequent() || to_infrequent();
rasbold@853 743 }
rasbold@853 744
rasbold@853 745 float freq() const { return _freq; }
rasbold@853 746 Block* from() const { return _from; }
rasbold@853 747 Block* to () const { return _to; }
rasbold@853 748 int infrequent() const { return _infrequent; }
rasbold@853 749 int state() const { return _state; }
rasbold@853 750
rasbold@853 751 void set_state(int state) { _state = state; }
rasbold@853 752
rasbold@853 753 #ifndef PRODUCT
rasbold@853 754 void dump( ) const;
rasbold@853 755 #endif
rasbold@853 756 };
rasbold@853 757
rasbold@853 758
rasbold@853 759 //-----------------------------------Trace-------------------------------------
rasbold@853 760 // An ordered list of basic blocks.
rasbold@853 761 class Trace : public ResourceObj {
rasbold@853 762 private:
rasbold@853 763 uint _id; // Unique Trace id (derived from initial block)
rasbold@853 764 Block ** _next_list; // Array mapping index to next block
rasbold@853 765 Block ** _prev_list; // Array mapping index to previous block
rasbold@853 766 Block * _first; // First block in the trace
rasbold@853 767 Block * _last; // Last block in the trace
rasbold@853 768
rasbold@853 769 // Return the block that follows "b" in the trace.
rasbold@853 770 Block * next(Block *b) const { return _next_list[b->_pre_order]; }
rasbold@853 771 void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; }
rasbold@853 772
twisti@1040 773 // Return the block that precedes "b" in the trace.
rasbold@853 774 Block * prev(Block *b) const { return _prev_list[b->_pre_order]; }
rasbold@853 775 void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; }
rasbold@853 776
rasbold@853 777 // We've discovered a loop in this trace. Reset last to be "b", and first as
rasbold@853 778 // the block following "b
rasbold@853 779 void break_loop_after(Block *b) {
rasbold@853 780 _last = b;
rasbold@853 781 _first = next(b);
rasbold@853 782 set_prev(_first, NULL);
rasbold@853 783 set_next(_last, NULL);
rasbold@853 784 }
rasbold@853 785
rasbold@853 786 public:
rasbold@853 787
rasbold@853 788 Trace(Block *b, Block **next_list, Block **prev_list) :
rasbold@853 789 _first(b),
rasbold@853 790 _last(b),
rasbold@853 791 _next_list(next_list),
rasbold@853 792 _prev_list(prev_list),
rasbold@853 793 _id(b->_pre_order) {
rasbold@853 794 set_next(b, NULL);
rasbold@853 795 set_prev(b, NULL);
rasbold@853 796 };
rasbold@853 797
rasbold@853 798 // Return the id number
rasbold@853 799 uint id() const { return _id; }
rasbold@853 800 void set_id(uint id) { _id = id; }
rasbold@853 801
rasbold@853 802 // Return the first block in the trace
rasbold@853 803 Block * first_block() const { return _first; }
rasbold@853 804
rasbold@853 805 // Return the last block in the trace
rasbold@853 806 Block * last_block() const { return _last; }
rasbold@853 807
rasbold@853 808 // Insert a trace in the middle of this one after b
rasbold@853 809 void insert_after(Block *b, Trace *tr) {
rasbold@853 810 set_next(tr->last_block(), next(b));
rasbold@853 811 if (next(b) != NULL) {
rasbold@853 812 set_prev(next(b), tr->last_block());
rasbold@853 813 }
rasbold@853 814
rasbold@853 815 set_next(b, tr->first_block());
rasbold@853 816 set_prev(tr->first_block(), b);
rasbold@853 817
rasbold@853 818 if (b == _last) {
rasbold@853 819 _last = tr->last_block();
rasbold@853 820 }
rasbold@853 821 }
rasbold@853 822
rasbold@853 823 void insert_before(Block *b, Trace *tr) {
rasbold@853 824 Block *p = prev(b);
rasbold@853 825 assert(p != NULL, "use append instead");
rasbold@853 826 insert_after(p, tr);
rasbold@853 827 }
rasbold@853 828
rasbold@853 829 // Append another trace to this one.
rasbold@853 830 void append(Trace *tr) {
rasbold@853 831 insert_after(_last, tr);
rasbold@853 832 }
rasbold@853 833
rasbold@853 834 // Append a block at the end of this trace
rasbold@853 835 void append(Block *b) {
rasbold@853 836 set_next(_last, b);
rasbold@853 837 set_prev(b, _last);
rasbold@853 838 _last = b;
rasbold@853 839 }
rasbold@853 840
rasbold@853 841 // Adjust the the blocks in this trace
rasbold@853 842 void fixup_blocks(PhaseCFG &cfg);
rasbold@853 843 bool backedge(CFGEdge *e);
rasbold@853 844
rasbold@853 845 #ifndef PRODUCT
rasbold@853 846 void dump( ) const;
rasbold@853 847 #endif
rasbold@853 848 };
rasbold@853 849
rasbold@853 850 //------------------------------PhaseBlockLayout-------------------------------
rasbold@853 851 // Rearrange blocks into some canonical order, based on edges and their frequencies
rasbold@853 852 class PhaseBlockLayout : public Phase {
never@3138 853 friend class VMStructs;
rasbold@853 854 PhaseCFG &_cfg; // Control flow graph
rasbold@853 855
rasbold@853 856 GrowableArray<CFGEdge *> *edges;
rasbold@853 857 Trace **traces;
rasbold@853 858 Block **next;
rasbold@853 859 Block **prev;
rasbold@853 860 UnionFind *uf;
rasbold@853 861
rasbold@853 862 // Given a block, find its encompassing Trace
rasbold@853 863 Trace * trace(Block *b) {
rasbold@853 864 return traces[uf->Find_compress(b->_pre_order)];
rasbold@853 865 }
rasbold@853 866 public:
rasbold@853 867 PhaseBlockLayout(PhaseCFG &cfg);
rasbold@853 868
rasbold@853 869 void find_edges();
rasbold@853 870 void grow_traces();
rasbold@853 871 void merge_traces(bool loose_connections);
rasbold@853 872 void reorder_traces(int count);
rasbold@853 873 void union_traces(Trace* from, Trace* to);
rasbold@853 874 };
stefank@2314 875
stefank@2314 876 #endif // SHARE_VM_OPTO_BLOCK_HPP

mercurial