src/share/vm/opto/block.hpp

Wed, 07 Aug 2013 17:56:19 +0200

author
adlertz
date
Wed, 07 Aug 2013 17:56:19 +0200
changeset 5509
d1034bd8cefc
parent 4315
2aff40cb4703
child 5539
adb9a7d94cb5
permissions
-rw-r--r--

8022284: Hide internal data structure in PhaseCFG
Summary: Hide private node to block mapping using public interface
Reviewed-by: kvn, roland

duke@435 1 /*
mikael@4153 2 * Copyright (c) 1997, 2012, 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;
duke@435 108 public:
duke@435 109 // Nodes in this block, in order
duke@435 110 Node_List _nodes;
duke@435 111
duke@435 112 // Basic blocks have a Node which defines Control for all Nodes pinned in
duke@435 113 // this block. This Node is a RegionNode. Exception-causing Nodes
duke@435 114 // (division, subroutines) and Phi functions are always pinned. Later,
duke@435 115 // every Node will get pinned to some block.
duke@435 116 Node *head() const { return _nodes[0]; }
duke@435 117
duke@435 118 // CAUTION: num_preds() is ONE based, so that predecessor numbers match
duke@435 119 // input edges to Regions and Phis.
duke@435 120 uint num_preds() const { return head()->req(); }
duke@435 121 Node *pred(uint i) const { return head()->in(i); }
duke@435 122
duke@435 123 // Array of successor blocks, same size as projs array
duke@435 124 Block_Array _succs;
duke@435 125
duke@435 126 // Basic blocks have some number of Nodes which split control to all
duke@435 127 // following blocks. These Nodes are always Projections. The field in
duke@435 128 // the Projection and the block-ending Node determine which Block follows.
duke@435 129 uint _num_succs;
duke@435 130
duke@435 131 // Basic blocks also carry all sorts of good old fashioned DFS information
duke@435 132 // used to find loops, loop nesting depth, dominators, etc.
duke@435 133 uint _pre_order; // Pre-order DFS number
duke@435 134
duke@435 135 // Dominator tree
duke@435 136 uint _dom_depth; // Depth in dominator tree for fast LCA
duke@435 137 Block* _idom; // Immediate dominator block
duke@435 138
duke@435 139 CFGLoop *_loop; // Loop to which this block belongs
duke@435 140 uint _rpo; // Number in reverse post order walk
duke@435 141
duke@435 142 virtual bool is_block() { return true; }
rasbold@853 143 float succ_prob(uint i); // return probability of i'th successor
rasbold@853 144 int num_fall_throughs(); // How many fall-through candidate this block has
rasbold@853 145 void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code
rasbold@853 146 bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate
rasbold@853 147 Block* lone_fall_through(); // Return lone fall-through Block or null
duke@435 148
duke@435 149 Block* dom_lca(Block* that); // Compute LCA in dominator tree.
duke@435 150 #ifdef ASSERT
duke@435 151 bool dominates(Block* that) {
duke@435 152 int dom_diff = this->_dom_depth - that->_dom_depth;
duke@435 153 if (dom_diff > 0) return false;
duke@435 154 for (; dom_diff < 0; dom_diff++) that = that->_idom;
duke@435 155 return this == that;
duke@435 156 }
duke@435 157 #endif
duke@435 158
duke@435 159 // Report the alignment required by this block. Must be a power of 2.
duke@435 160 // The previous block will insert nops to get this alignment.
duke@435 161 uint code_alignment();
rasbold@853 162 uint compute_loop_alignment();
duke@435 163
duke@435 164 // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies.
duke@435 165 // It is currently also used to scale such frequencies relative to
duke@435 166 // FreqCountInvocations relative to the old value of 1500.
duke@435 167 #define BLOCK_FREQUENCY(f) ((f * (float) 1500) / FreqCountInvocations)
duke@435 168
duke@435 169 // Register Pressure (estimate) for Splitting heuristic
duke@435 170 uint _reg_pressure;
duke@435 171 uint _ihrp_index;
duke@435 172 uint _freg_pressure;
duke@435 173 uint _fhrp_index;
duke@435 174
duke@435 175 // Mark and visited bits for an LCA calculation in insert_anti_dependences.
duke@435 176 // Since they hold unique node indexes, they do not need reinitialization.
duke@435 177 node_idx_t _raise_LCA_mark;
duke@435 178 void set_raise_LCA_mark(node_idx_t x) { _raise_LCA_mark = x; }
duke@435 179 node_idx_t raise_LCA_mark() const { return _raise_LCA_mark; }
duke@435 180 node_idx_t _raise_LCA_visited;
duke@435 181 void set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; }
duke@435 182 node_idx_t raise_LCA_visited() const { return _raise_LCA_visited; }
duke@435 183
duke@435 184 // Estimated size in bytes of first instructions in a loop.
duke@435 185 uint _first_inst_size;
duke@435 186 uint first_inst_size() const { return _first_inst_size; }
duke@435 187 void set_first_inst_size(uint s) { _first_inst_size = s; }
duke@435 188
duke@435 189 // Compute the size of first instructions in this block.
duke@435 190 uint compute_first_inst_size(uint& sum_size, uint inst_cnt, PhaseRegAlloc* ra);
duke@435 191
duke@435 192 // Compute alignment padding if the block needs it.
duke@435 193 // Align a loop if loop's padding is less or equal to padding limit
duke@435 194 // or the size of first instructions in the loop > padding.
duke@435 195 uint alignment_padding(int current_offset) {
duke@435 196 int block_alignment = code_alignment();
duke@435 197 int max_pad = block_alignment-relocInfo::addr_unit();
duke@435 198 if( max_pad > 0 ) {
duke@435 199 assert(is_power_of_2(max_pad+relocInfo::addr_unit()), "");
duke@435 200 int current_alignment = current_offset & max_pad;
duke@435 201 if( current_alignment != 0 ) {
duke@435 202 uint padding = (block_alignment-current_alignment) & max_pad;
rasbold@853 203 if( has_loop_alignment() &&
rasbold@853 204 padding > (uint)MaxLoopPad &&
rasbold@853 205 first_inst_size() <= padding ) {
rasbold@853 206 return 0;
duke@435 207 }
rasbold@853 208 return padding;
duke@435 209 }
duke@435 210 }
duke@435 211 return 0;
duke@435 212 }
duke@435 213
duke@435 214 // Connector blocks. Connector blocks are basic blocks devoid of
duke@435 215 // instructions, but may have relevant non-instruction Nodes, such as
duke@435 216 // Phis or MergeMems. Such blocks are discovered and marked during the
duke@435 217 // RemoveEmpty phase, and elided during Output.
duke@435 218 bool _connector;
duke@435 219 void set_connector() { _connector = true; }
duke@435 220 bool is_connector() const { return _connector; };
duke@435 221
rasbold@853 222 // Loop_alignment will be set for blocks which are at the top of loops.
rasbold@853 223 // The block layout pass may rotate loops such that the loop head may not
rasbold@853 224 // be the sequentially first block of the loop encountered in the linear
rasbold@853 225 // list of blocks. If the layout pass is not run, loop alignment is set
rasbold@853 226 // for each block which is the head of a loop.
rasbold@853 227 uint _loop_alignment;
rasbold@853 228 void set_loop_alignment(Block *loop_top) {
rasbold@853 229 uint new_alignment = loop_top->compute_loop_alignment();
rasbold@853 230 if (new_alignment > _loop_alignment) {
rasbold@853 231 _loop_alignment = new_alignment;
rasbold@853 232 }
rasbold@853 233 }
rasbold@853 234 uint loop_alignment() const { return _loop_alignment; }
rasbold@853 235 bool has_loop_alignment() const { return loop_alignment() > 0; }
rasbold@853 236
duke@435 237 // Create a new Block with given head Node.
duke@435 238 // Creates the (empty) predecessor arrays.
duke@435 239 Block( Arena *a, Node *headnode )
duke@435 240 : CFGElement(),
duke@435 241 _nodes(a),
duke@435 242 _succs(a),
duke@435 243 _num_succs(0),
duke@435 244 _pre_order(0),
duke@435 245 _idom(0),
duke@435 246 _loop(NULL),
duke@435 247 _reg_pressure(0),
duke@435 248 _ihrp_index(1),
duke@435 249 _freg_pressure(0),
duke@435 250 _fhrp_index(1),
duke@435 251 _raise_LCA_mark(0),
duke@435 252 _raise_LCA_visited(0),
duke@435 253 _first_inst_size(999999),
rasbold@853 254 _connector(false),
rasbold@853 255 _loop_alignment(0) {
duke@435 256 _nodes.push(headnode);
duke@435 257 }
duke@435 258
duke@435 259 // Index of 'end' Node
duke@435 260 uint end_idx() const {
duke@435 261 // %%%%% add a proj after every goto
duke@435 262 // so (last->is_block_proj() != last) always, then simplify this code
duke@435 263 // This will not give correct end_idx for block 0 when it only contains root.
duke@435 264 int last_idx = _nodes.size() - 1;
duke@435 265 Node *last = _nodes[last_idx];
duke@435 266 assert(last->is_block_proj() == last || last->is_block_proj() == _nodes[last_idx - _num_succs], "");
duke@435 267 return (last->is_block_proj() == last) ? last_idx : (last_idx - _num_succs);
duke@435 268 }
duke@435 269
duke@435 270 // Basic blocks have a Node which ends them. This Node determines which
duke@435 271 // basic block follows this one in the program flow. This Node is either an
duke@435 272 // IfNode, a GotoNode, a JmpNode, or a ReturnNode.
duke@435 273 Node *end() const { return _nodes[end_idx()]; }
duke@435 274
duke@435 275 // Add an instruction to an existing block. It must go after the head
duke@435 276 // instruction and before the end instruction.
duke@435 277 void add_inst( Node *n ) { _nodes.insert(end_idx(),n); }
duke@435 278 // Find node in block
duke@435 279 uint find_node( const Node *n ) const;
duke@435 280 // Find and remove n from block list
duke@435 281 void find_remove( const Node *n );
duke@435 282
roland@3316 283 // helper function that adds caller save registers to MachProjNode
roland@3316 284 void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe);
duke@435 285 // Schedule a call next in the block
adlertz@5509 286 uint sched_call(Matcher &matcher, PhaseCFG* cfg, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call);
duke@435 287
duke@435 288 // Perform basic-block local scheduling
roland@3447 289 Node *select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot);
adlertz@5509 290 void set_next_call( Node *n, VectorSet &next_call, PhaseCFG* cfg);
adlertz@5509 291 void needed_for_next_call(Node *this_call, VectorSet &next_call, PhaseCFG* cfg);
roland@3447 292 bool schedule_local(PhaseCFG *cfg, Matcher &m, GrowableArray<int> &ready_cnt, VectorSet &next_call);
duke@435 293 // Cleanup if any code lands between a Call and his Catch
adlertz@5509 294 void call_catch_cleanup(PhaseCFG* cfg, Compile *C);
duke@435 295 // Detect implicit-null-check opportunities. Basically, find NULL checks
duke@435 296 // with suitable memory ops nearby. Use the memory op to do the NULL check.
duke@435 297 // I can generate a memory op if there is not one nearby.
duke@435 298 void implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons);
duke@435 299
duke@435 300 // Return the empty status of a block
duke@435 301 enum { not_empty, empty_with_goto, completely_empty };
duke@435 302 int is_Empty() const;
duke@435 303
duke@435 304 // Forward through connectors
duke@435 305 Block* non_connector() {
duke@435 306 Block* s = this;
duke@435 307 while (s->is_connector()) {
duke@435 308 s = s->_succs[0];
duke@435 309 }
duke@435 310 return s;
duke@435 311 }
duke@435 312
rasbold@853 313 // Return true if b is a successor of this block
rasbold@853 314 bool has_successor(Block* b) const {
rasbold@853 315 for (uint i = 0; i < _num_succs; i++ ) {
rasbold@853 316 if (non_connector_successor(i) == b) {
rasbold@853 317 return true;
rasbold@853 318 }
rasbold@853 319 }
rasbold@853 320 return false;
rasbold@853 321 }
rasbold@853 322
duke@435 323 // Successor block, after forwarding through connectors
duke@435 324 Block* non_connector_successor(int i) const {
duke@435 325 return _succs[i]->non_connector();
duke@435 326 }
duke@435 327
duke@435 328 // Examine block's code shape to predict if it is not commonly executed.
duke@435 329 bool has_uncommon_code() const;
duke@435 330
duke@435 331 // Use frequency calculations and code shape to predict if the block
duke@435 332 // is uncommon.
adlertz@5509 333 bool is_uncommon(PhaseCFG* cfg) const;
duke@435 334
duke@435 335 #ifndef PRODUCT
duke@435 336 // Debugging print of basic block
kvn@3049 337 void dump_bidx(const Block* orig, outputStream* st = tty) const;
adlertz@5509 338 void dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st = tty) const;
adlertz@5509 339 void dump_head(const PhaseCFG* cfg, outputStream* st = tty) const;
kvn@3049 340 void dump() const;
adlertz@5509 341 void dump(const PhaseCFG* cfg) const;
duke@435 342 #endif
duke@435 343 };
duke@435 344
duke@435 345
duke@435 346 //------------------------------PhaseCFG---------------------------------------
duke@435 347 // Build an array of Basic Block pointers, one per Node.
duke@435 348 class PhaseCFG : public Phase {
never@3138 349 friend class VMStructs;
duke@435 350 private:
adlertz@5509 351 // Arena for the blocks to be stored in
adlertz@5509 352 Arena* _block_arena;
adlertz@5509 353
adlertz@5509 354 // Map nodes to owning basic block
adlertz@5509 355 Block_Array _node_to_block_mapping;
adlertz@5509 356
duke@435 357 // Build a proper looking cfg. Return count of basic blocks
duke@435 358 uint build_cfg();
duke@435 359
duke@435 360 // Perform DFS search.
duke@435 361 // Setup 'vertex' as DFS to vertex mapping.
duke@435 362 // Setup 'semi' as vertex to DFS mapping.
duke@435 363 // Set 'parent' to DFS parent.
duke@435 364 uint DFS( Tarjan *tarjan );
duke@435 365
duke@435 366 // Helper function to insert a node into a block
duke@435 367 void schedule_node_into_block( Node *n, Block *b );
duke@435 368
kvn@1039 369 void replace_block_proj_ctrl( Node *n );
kvn@1036 370
duke@435 371 // Set the basic block for pinned Nodes
duke@435 372 void schedule_pinned_nodes( VectorSet &visited );
duke@435 373
duke@435 374 // I'll need a few machine-specific GotoNodes. Clone from this one.
duke@435 375 MachNode *_goto;
duke@435 376
duke@435 377 Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);
duke@435 378 void verify_anti_dependences(Block* LCA, Node* load) {
adlertz@5509 379 assert(LCA == get_block_for_node(load), "should already be scheduled");
duke@435 380 insert_anti_dependences(LCA, load, true);
duke@435 381 }
duke@435 382
duke@435 383 public:
adlertz@5509 384 PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher);
duke@435 385
duke@435 386 uint _num_blocks; // Count of basic blocks
duke@435 387 Block_List _blocks; // List of basic blocks
duke@435 388 RootNode *_root; // Root of whole program
duke@435 389 Block *_broot; // Basic block of root
duke@435 390 uint _rpo_ctr;
duke@435 391 CFGLoop* _root_loop;
kvn@1108 392 float _outer_loop_freq; // Outmost loop frequency
duke@435 393
adlertz@5509 394
adlertz@5509 395 // set which block this node should reside in
adlertz@5509 396 void map_node_to_block(const Node* node, Block* block) {
adlertz@5509 397 _node_to_block_mapping.map(node->_idx, block);
adlertz@5509 398 }
adlertz@5509 399
adlertz@5509 400 // removes the mapping from a node to a block
adlertz@5509 401 void unmap_node_from_block(const Node* node) {
adlertz@5509 402 _node_to_block_mapping.map(node->_idx, NULL);
adlertz@5509 403 }
adlertz@5509 404
adlertz@5509 405 // get the block in which this node resides
adlertz@5509 406 Block* get_block_for_node(const Node* node) const {
adlertz@5509 407 return _node_to_block_mapping[node->_idx];
adlertz@5509 408 }
adlertz@5509 409
adlertz@5509 410 // does this node reside in a block; return true
adlertz@5509 411 bool has_block(const Node* node) const {
adlertz@5509 412 return (_node_to_block_mapping.lookup(node->_idx) != NULL);
adlertz@5509 413 }
adlertz@5509 414
duke@435 415 // Per node latency estimation, valid only during GCM
kvn@2040 416 GrowableArray<uint> *_node_latency;
duke@435 417
duke@435 418 #ifndef PRODUCT
duke@435 419 bool _trace_opto_pipelining; // tracing flag
duke@435 420 #endif
duke@435 421
kvn@1268 422 #ifdef ASSERT
kvn@1268 423 Unique_Node_List _raw_oops;
kvn@1268 424 #endif
kvn@1268 425
duke@435 426 // Build dominators
duke@435 427 void Dominators();
duke@435 428
duke@435 429 // Estimate block frequencies based on IfNode probabilities
duke@435 430 void Estimate_Block_Frequency();
duke@435 431
duke@435 432 // Global Code Motion. See Click's PLDI95 paper. Place Nodes in specific
adlertz@5509 433 // basic blocks; i.e. _node_to_block_mapping now maps _idx for all Nodes to some Block.
duke@435 434 void GlobalCodeMotion( Matcher &m, uint unique, Node_List &proj_list );
duke@435 435
duke@435 436 // Compute the (backwards) latency of a node from the uses
duke@435 437 void latency_from_uses(Node *n);
duke@435 438
duke@435 439 // Compute the (backwards) latency of a node from a single use
duke@435 440 int latency_from_use(Node *n, const Node *def, Node *use);
duke@435 441
duke@435 442 // Compute the (backwards) latency of a node from the uses of this instruction
duke@435 443 void partial_latency_of_defs(Node *n);
duke@435 444
duke@435 445 // Schedule Nodes early in their basic blocks.
duke@435 446 bool schedule_early(VectorSet &visited, Node_List &roots);
duke@435 447
duke@435 448 // For each node, find the latest block it can be scheduled into
duke@435 449 // and then select the cheapest block between the latest and earliest
duke@435 450 // block to place the node.
duke@435 451 void schedule_late(VectorSet &visited, Node_List &stack);
duke@435 452
duke@435 453 // Pick a block between early and late that is a cheaper alternative
duke@435 454 // to late. Helper for schedule_late.
duke@435 455 Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self);
duke@435 456
duke@435 457 // Compute the instruction global latency with a backwards walk
duke@435 458 void ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack);
duke@435 459
rasbold@853 460 // Set loop alignment
rasbold@853 461 void set_loop_alignment();
rasbold@853 462
duke@435 463 // Remove empty basic blocks
rasbold@853 464 void remove_empty();
rasbold@853 465 void fixup_flow();
rasbold@853 466 bool move_to_next(Block* bx, uint b_index);
rasbold@853 467 void move_to_end(Block* bx, uint b_index);
rasbold@853 468 void insert_goto_at(uint block_no, uint succ_no);
duke@435 469
duke@435 470 // Check for NeverBranch at block end. This needs to become a GOTO to the
duke@435 471 // true target. NeverBranch are treated as a conditional branch that always
duke@435 472 // goes the same direction for most of the optimizer and are used to give a
duke@435 473 // fake exit path to infinite loops. At this late stage they need to turn
duke@435 474 // into Goto's so that when you enter the infinite loop you indeed hang.
duke@435 475 void convert_NeverBranch_to_Goto(Block *b);
duke@435 476
duke@435 477 CFGLoop* create_loop_tree();
duke@435 478
duke@435 479 // Insert a node into a block, and update the _bbs
duke@435 480 void insert( Block *b, uint idx, Node *n ) {
duke@435 481 b->_nodes.insert( idx, n );
adlertz@5509 482 map_node_to_block(n, b);
duke@435 483 }
duke@435 484
duke@435 485 #ifndef PRODUCT
duke@435 486 bool trace_opto_pipelining() const { return _trace_opto_pipelining; }
duke@435 487
duke@435 488 // Debugging print of CFG
duke@435 489 void dump( ) const; // CFG only
duke@435 490 void _dump_cfg( const Node *end, VectorSet &visited ) const;
duke@435 491 void verify() const;
duke@435 492 void dump_headers();
duke@435 493 #else
duke@435 494 bool trace_opto_pipelining() const { return false; }
duke@435 495 #endif
duke@435 496 };
duke@435 497
duke@435 498
rasbold@853 499 //------------------------------UnionFind--------------------------------------
duke@435 500 // Map Block indices to a block-index for a cfg-cover.
duke@435 501 // Array lookup in the optimized case.
duke@435 502 class UnionFind : public ResourceObj {
duke@435 503 uint _cnt, _max;
duke@435 504 uint* _indices;
duke@435 505 ReallocMark _nesting; // assertion check for reallocations
duke@435 506 public:
duke@435 507 UnionFind( uint max );
duke@435 508 void reset( uint max ); // Reset to identity map for [0..max]
duke@435 509
duke@435 510 uint lookup( uint nidx ) const {
duke@435 511 return _indices[nidx];
duke@435 512 }
duke@435 513 uint operator[] (uint nidx) const { return lookup(nidx); }
duke@435 514
duke@435 515 void map( uint from_idx, uint to_idx ) {
duke@435 516 assert( from_idx < _cnt, "oob" );
duke@435 517 _indices[from_idx] = to_idx;
duke@435 518 }
duke@435 519 void extend( uint from_idx, uint to_idx );
duke@435 520
duke@435 521 uint Size() const { return _cnt; }
duke@435 522
duke@435 523 uint Find( uint idx ) {
duke@435 524 assert( idx < 65536, "Must fit into uint");
duke@435 525 uint uf_idx = lookup(idx);
duke@435 526 return (uf_idx == idx) ? uf_idx : Find_compress(idx);
duke@435 527 }
duke@435 528 uint Find_compress( uint idx );
duke@435 529 uint Find_const( uint idx ) const;
duke@435 530 void Union( uint idx1, uint idx2 );
duke@435 531
duke@435 532 };
duke@435 533
duke@435 534 //----------------------------BlockProbPair---------------------------
duke@435 535 // Ordered pair of Node*.
duke@435 536 class BlockProbPair VALUE_OBJ_CLASS_SPEC {
duke@435 537 protected:
duke@435 538 Block* _target; // block target
duke@435 539 float _prob; // probability of edge to block
duke@435 540 public:
duke@435 541 BlockProbPair() : _target(NULL), _prob(0.0) {}
duke@435 542 BlockProbPair(Block* b, float p) : _target(b), _prob(p) {}
duke@435 543
duke@435 544 Block* get_target() const { return _target; }
duke@435 545 float get_prob() const { return _prob; }
duke@435 546 };
duke@435 547
duke@435 548 //------------------------------CFGLoop-------------------------------------------
duke@435 549 class CFGLoop : public CFGElement {
never@3138 550 friend class VMStructs;
duke@435 551 int _id;
duke@435 552 int _depth;
duke@435 553 CFGLoop *_parent; // root of loop tree is the method level "pseudo" loop, it's parent is null
duke@435 554 CFGLoop *_sibling; // null terminated list
duke@435 555 CFGLoop *_child; // first child, use child's sibling to visit all immediately nested loops
duke@435 556 GrowableArray<CFGElement*> _members; // list of members of loop
duke@435 557 GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities
duke@435 558 float _exit_prob; // probability any loop exit is taken on a single loop iteration
duke@435 559 void update_succ_freq(Block* b, float freq);
duke@435 560
duke@435 561 public:
duke@435 562 CFGLoop(int id) :
duke@435 563 CFGElement(),
duke@435 564 _id(id),
duke@435 565 _depth(0),
duke@435 566 _parent(NULL),
duke@435 567 _sibling(NULL),
duke@435 568 _child(NULL),
duke@435 569 _exit_prob(1.0f) {}
duke@435 570 CFGLoop* parent() { return _parent; }
adlertz@5509 571 void push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg);
duke@435 572 void add_member(CFGElement *s) { _members.push(s); }
duke@435 573 void add_nested_loop(CFGLoop* cl);
duke@435 574 Block* head() {
duke@435 575 assert(_members.at(0)->is_block(), "head must be a block");
duke@435 576 Block* hd = _members.at(0)->as_Block();
duke@435 577 assert(hd->_loop == this, "just checking");
duke@435 578 assert(hd->head()->is_Loop(), "must begin with loop head node");
duke@435 579 return hd;
duke@435 580 }
duke@435 581 Block* backedge_block(); // Return the block on the backedge of the loop (else NULL)
duke@435 582 void compute_loop_depth(int depth);
duke@435 583 void compute_freq(); // compute frequency with loop assuming head freq 1.0f
duke@435 584 void scale_freq(); // scale frequency by loop trip count (including outer loops)
kvn@1108 585 float outer_loop_freq() const; // frequency of outer loop
duke@435 586 bool in_loop_nest(Block* b);
duke@435 587 float trip_count() const { return 1.0f / _exit_prob; }
duke@435 588 virtual bool is_loop() { return true; }
duke@435 589 int id() { return _id; }
duke@435 590
duke@435 591 #ifndef PRODUCT
duke@435 592 void dump( ) const;
duke@435 593 void dump_tree() const;
duke@435 594 #endif
duke@435 595 };
rasbold@853 596
rasbold@853 597
rasbold@853 598 //----------------------------------CFGEdge------------------------------------
rasbold@853 599 // A edge between two basic blocks that will be embodied by a branch or a
rasbold@853 600 // fall-through.
rasbold@853 601 class CFGEdge : public ResourceObj {
never@3138 602 friend class VMStructs;
rasbold@853 603 private:
rasbold@853 604 Block * _from; // Source basic block
rasbold@853 605 Block * _to; // Destination basic block
rasbold@853 606 float _freq; // Execution frequency (estimate)
rasbold@853 607 int _state;
rasbold@853 608 bool _infrequent;
rasbold@853 609 int _from_pct;
rasbold@853 610 int _to_pct;
rasbold@853 611
rasbold@853 612 // Private accessors
rasbold@853 613 int from_pct() const { return _from_pct; }
rasbold@853 614 int to_pct() const { return _to_pct; }
rasbold@853 615 int from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; }
rasbold@853 616 int to_infrequent() const { return to_pct() < BlockLayoutMinDiamondPercentage; }
rasbold@853 617
rasbold@853 618 public:
rasbold@853 619 enum {
rasbold@853 620 open, // initial edge state; unprocessed
rasbold@853 621 connected, // edge used to connect two traces together
rasbold@853 622 interior // edge is interior to trace (could be backedge)
rasbold@853 623 };
rasbold@853 624
rasbold@853 625 CFGEdge(Block *from, Block *to, float freq, int from_pct, int to_pct) :
rasbold@853 626 _from(from), _to(to), _freq(freq),
rasbold@853 627 _from_pct(from_pct), _to_pct(to_pct), _state(open) {
rasbold@853 628 _infrequent = from_infrequent() || to_infrequent();
rasbold@853 629 }
rasbold@853 630
rasbold@853 631 float freq() const { return _freq; }
rasbold@853 632 Block* from() const { return _from; }
rasbold@853 633 Block* to () const { return _to; }
rasbold@853 634 int infrequent() const { return _infrequent; }
rasbold@853 635 int state() const { return _state; }
rasbold@853 636
rasbold@853 637 void set_state(int state) { _state = state; }
rasbold@853 638
rasbold@853 639 #ifndef PRODUCT
rasbold@853 640 void dump( ) const;
rasbold@853 641 #endif
rasbold@853 642 };
rasbold@853 643
rasbold@853 644
rasbold@853 645 //-----------------------------------Trace-------------------------------------
rasbold@853 646 // An ordered list of basic blocks.
rasbold@853 647 class Trace : public ResourceObj {
rasbold@853 648 private:
rasbold@853 649 uint _id; // Unique Trace id (derived from initial block)
rasbold@853 650 Block ** _next_list; // Array mapping index to next block
rasbold@853 651 Block ** _prev_list; // Array mapping index to previous block
rasbold@853 652 Block * _first; // First block in the trace
rasbold@853 653 Block * _last; // Last block in the trace
rasbold@853 654
rasbold@853 655 // Return the block that follows "b" in the trace.
rasbold@853 656 Block * next(Block *b) const { return _next_list[b->_pre_order]; }
rasbold@853 657 void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; }
rasbold@853 658
twisti@1040 659 // Return the block that precedes "b" in the trace.
rasbold@853 660 Block * prev(Block *b) const { return _prev_list[b->_pre_order]; }
rasbold@853 661 void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; }
rasbold@853 662
rasbold@853 663 // We've discovered a loop in this trace. Reset last to be "b", and first as
rasbold@853 664 // the block following "b
rasbold@853 665 void break_loop_after(Block *b) {
rasbold@853 666 _last = b;
rasbold@853 667 _first = next(b);
rasbold@853 668 set_prev(_first, NULL);
rasbold@853 669 set_next(_last, NULL);
rasbold@853 670 }
rasbold@853 671
rasbold@853 672 public:
rasbold@853 673
rasbold@853 674 Trace(Block *b, Block **next_list, Block **prev_list) :
rasbold@853 675 _first(b),
rasbold@853 676 _last(b),
rasbold@853 677 _next_list(next_list),
rasbold@853 678 _prev_list(prev_list),
rasbold@853 679 _id(b->_pre_order) {
rasbold@853 680 set_next(b, NULL);
rasbold@853 681 set_prev(b, NULL);
rasbold@853 682 };
rasbold@853 683
rasbold@853 684 // Return the id number
rasbold@853 685 uint id() const { return _id; }
rasbold@853 686 void set_id(uint id) { _id = id; }
rasbold@853 687
rasbold@853 688 // Return the first block in the trace
rasbold@853 689 Block * first_block() const { return _first; }
rasbold@853 690
rasbold@853 691 // Return the last block in the trace
rasbold@853 692 Block * last_block() const { return _last; }
rasbold@853 693
rasbold@853 694 // Insert a trace in the middle of this one after b
rasbold@853 695 void insert_after(Block *b, Trace *tr) {
rasbold@853 696 set_next(tr->last_block(), next(b));
rasbold@853 697 if (next(b) != NULL) {
rasbold@853 698 set_prev(next(b), tr->last_block());
rasbold@853 699 }
rasbold@853 700
rasbold@853 701 set_next(b, tr->first_block());
rasbold@853 702 set_prev(tr->first_block(), b);
rasbold@853 703
rasbold@853 704 if (b == _last) {
rasbold@853 705 _last = tr->last_block();
rasbold@853 706 }
rasbold@853 707 }
rasbold@853 708
rasbold@853 709 void insert_before(Block *b, Trace *tr) {
rasbold@853 710 Block *p = prev(b);
rasbold@853 711 assert(p != NULL, "use append instead");
rasbold@853 712 insert_after(p, tr);
rasbold@853 713 }
rasbold@853 714
rasbold@853 715 // Append another trace to this one.
rasbold@853 716 void append(Trace *tr) {
rasbold@853 717 insert_after(_last, tr);
rasbold@853 718 }
rasbold@853 719
rasbold@853 720 // Append a block at the end of this trace
rasbold@853 721 void append(Block *b) {
rasbold@853 722 set_next(_last, b);
rasbold@853 723 set_prev(b, _last);
rasbold@853 724 _last = b;
rasbold@853 725 }
rasbold@853 726
rasbold@853 727 // Adjust the the blocks in this trace
rasbold@853 728 void fixup_blocks(PhaseCFG &cfg);
rasbold@853 729 bool backedge(CFGEdge *e);
rasbold@853 730
rasbold@853 731 #ifndef PRODUCT
rasbold@853 732 void dump( ) const;
rasbold@853 733 #endif
rasbold@853 734 };
rasbold@853 735
rasbold@853 736 //------------------------------PhaseBlockLayout-------------------------------
rasbold@853 737 // Rearrange blocks into some canonical order, based on edges and their frequencies
rasbold@853 738 class PhaseBlockLayout : public Phase {
never@3138 739 friend class VMStructs;
rasbold@853 740 PhaseCFG &_cfg; // Control flow graph
rasbold@853 741
rasbold@853 742 GrowableArray<CFGEdge *> *edges;
rasbold@853 743 Trace **traces;
rasbold@853 744 Block **next;
rasbold@853 745 Block **prev;
rasbold@853 746 UnionFind *uf;
rasbold@853 747
rasbold@853 748 // Given a block, find its encompassing Trace
rasbold@853 749 Trace * trace(Block *b) {
rasbold@853 750 return traces[uf->Find_compress(b->_pre_order)];
rasbold@853 751 }
rasbold@853 752 public:
rasbold@853 753 PhaseBlockLayout(PhaseCFG &cfg);
rasbold@853 754
rasbold@853 755 void find_edges();
rasbold@853 756 void grow_traces();
rasbold@853 757 void merge_traces(bool loose_connections);
rasbold@853 758 void reorder_traces(int count);
rasbold@853 759 void union_traces(Trace* from, Trace* to);
rasbold@853 760 };
stefank@2314 761
stefank@2314 762 #endif // SHARE_VM_OPTO_BLOCK_HPP

mercurial