Mercurial > hg > truffle
annotate src/share/vm/opto/block.hpp @ 3851:95134e034042
7063629: use cbcond in C2 generated code on T4
Summary: Use new short branch instruction in C2 generated code.
Reviewed-by: never
| author | kvn |
|---|---|
| date | Thu, 11 Aug 2011 12:08:11 -0700 |
| parents | f95d63e2154a |
| children | f6f3bb0ee072 |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 1972 | 2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
844
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
844
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
844
diff
changeset
|
21 * questions. |
| 0 | 22 * |
| 23 */ | |
| 24 | |
| 1972 | 25 #ifndef SHARE_VM_OPTO_BLOCK_HPP |
| 26 #define SHARE_VM_OPTO_BLOCK_HPP | |
| 27 | |
| 28 #include "opto/multnode.hpp" | |
| 29 #include "opto/node.hpp" | |
| 30 #include "opto/phase.hpp" | |
| 31 | |
| 0 | 32 // Optimization - Graph Style |
| 33 | |
| 34 class Block; | |
| 35 class CFGLoop; | |
| 36 class MachCallNode; | |
| 37 class Matcher; | |
| 38 class RootNode; | |
| 39 class VectorSet; | |
| 40 struct Tarjan; | |
| 41 | |
| 42 //------------------------------Block_Array------------------------------------ | |
| 43 // Map dense integer indices to Blocks. Uses classic doubling-array trick. | |
| 44 // Abstractly provides an infinite array of Block*'s, initialized to NULL. | |
| 45 // Note that the constructor just zeros things, and since I use Arena | |
| 46 // allocation I do not need a destructor to reclaim storage. | |
| 47 class Block_Array : public ResourceObj { | |
| 48 uint _size; // allocated size, as opposed to formal limit | |
| 49 debug_only(uint _limit;) // limit to formal domain | |
| 50 protected: | |
| 51 Block **_blocks; | |
| 52 void grow( uint i ); // Grow array node to fit | |
| 53 | |
| 54 public: | |
| 55 Arena *_arena; // Arena to allocate in | |
| 56 | |
| 57 Block_Array(Arena *a) : _arena(a), _size(OptoBlockListSize) { | |
| 58 debug_only(_limit=0); | |
| 59 _blocks = NEW_ARENA_ARRAY( a, Block *, OptoBlockListSize ); | |
| 60 for( int i = 0; i < OptoBlockListSize; i++ ) { | |
| 61 _blocks[i] = NULL; | |
| 62 } | |
| 63 } | |
| 64 Block *lookup( uint i ) const // Lookup, or NULL for not mapped | |
| 65 { return (i<Max()) ? _blocks[i] : (Block*)NULL; } | |
| 66 Block *operator[] ( uint i ) const // Lookup, or assert for not mapped | |
| 67 { assert( i < Max(), "oob" ); return _blocks[i]; } | |
| 68 // Extend the mapping: index i maps to Block *n. | |
| 69 void map( uint i, Block *n ) { if( i>=Max() ) grow(i); _blocks[i] = n; } | |
| 70 uint Max() const { debug_only(return _limit); return _size; } | |
| 71 }; | |
| 72 | |
| 73 | |
| 74 class Block_List : public Block_Array { | |
| 75 public: | |
| 76 uint _cnt; | |
| 77 Block_List() : Block_Array(Thread::current()->resource_area()), _cnt(0) {} | |
| 78 void push( Block *b ) { map(_cnt++,b); } | |
| 79 Block *pop() { return _blocks[--_cnt]; } | |
| 80 Block *rpop() { Block *b = _blocks[0]; _blocks[0]=_blocks[--_cnt]; return b;} | |
| 81 void remove( uint i ); | |
| 82 void insert( uint i, Block *n ); | |
| 83 uint size() const { return _cnt; } | |
| 84 void reset() { _cnt = 0; } | |
| 418 | 85 void print(); |
| 0 | 86 }; |
| 87 | |
| 88 | |
| 89 class CFGElement : public ResourceObj { | |
| 90 public: | |
| 91 float _freq; // Execution frequency (estimate) | |
| 92 | |
| 93 CFGElement() : _freq(0.0f) {} | |
| 94 virtual bool is_block() { return false; } | |
| 95 virtual bool is_loop() { return false; } | |
| 96 Block* as_Block() { assert(is_block(), "must be block"); return (Block*)this; } | |
| 97 CFGLoop* as_CFGLoop() { assert(is_loop(), "must be loop"); return (CFGLoop*)this; } | |
| 98 }; | |
| 99 | |
| 100 //------------------------------Block------------------------------------------ | |
| 101 // This class defines a Basic Block. | |
| 102 // Basic blocks are used during the output routines, and are not used during | |
| 103 // any optimization pass. They are created late in the game. | |
| 104 class Block : public CFGElement { | |
| 105 public: | |
| 106 // Nodes in this block, in order | |
| 107 Node_List _nodes; | |
| 108 | |
| 109 // Basic blocks have a Node which defines Control for all Nodes pinned in | |
| 110 // this block. This Node is a RegionNode. Exception-causing Nodes | |
| 111 // (division, subroutines) and Phi functions are always pinned. Later, | |
| 112 // every Node will get pinned to some block. | |
| 113 Node *head() const { return _nodes[0]; } | |
| 114 | |
| 115 // CAUTION: num_preds() is ONE based, so that predecessor numbers match | |
| 116 // input edges to Regions and Phis. | |
| 117 uint num_preds() const { return head()->req(); } | |
| 118 Node *pred(uint i) const { return head()->in(i); } | |
| 119 | |
| 120 // Array of successor blocks, same size as projs array | |
| 121 Block_Array _succs; | |
| 122 | |
| 123 // Basic blocks have some number of Nodes which split control to all | |
| 124 // following blocks. These Nodes are always Projections. The field in | |
| 125 // the Projection and the block-ending Node determine which Block follows. | |
| 126 uint _num_succs; | |
| 127 | |
| 128 // Basic blocks also carry all sorts of good old fashioned DFS information | |
| 129 // used to find loops, loop nesting depth, dominators, etc. | |
| 130 uint _pre_order; // Pre-order DFS number | |
| 131 | |
| 132 // Dominator tree | |
| 133 uint _dom_depth; // Depth in dominator tree for fast LCA | |
| 134 Block* _idom; // Immediate dominator block | |
| 135 | |
| 136 CFGLoop *_loop; // Loop to which this block belongs | |
| 137 uint _rpo; // Number in reverse post order walk | |
| 138 | |
| 139 virtual bool is_block() { return true; } | |
| 418 | 140 float succ_prob(uint i); // return probability of i'th successor |
| 141 int num_fall_throughs(); // How many fall-through candidate this block has | |
| 142 void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code | |
| 143 bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate | |
| 144 Block* lone_fall_through(); // Return lone fall-through Block or null | |
| 0 | 145 |
| 146 Block* dom_lca(Block* that); // Compute LCA in dominator tree. | |
| 147 #ifdef ASSERT | |
| 148 bool dominates(Block* that) { | |
| 149 int dom_diff = this->_dom_depth - that->_dom_depth; | |
| 150 if (dom_diff > 0) return false; | |
| 151 for (; dom_diff < 0; dom_diff++) that = that->_idom; | |
| 152 return this == that; | |
| 153 } | |
| 154 #endif | |
| 155 | |
| 156 // Report the alignment required by this block. Must be a power of 2. | |
| 157 // The previous block will insert nops to get this alignment. | |
| 158 uint code_alignment(); | |
| 418 | 159 uint compute_loop_alignment(); |
| 0 | 160 |
| 161 // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies. | |
| 162 // It is currently also used to scale such frequencies relative to | |
| 163 // FreqCountInvocations relative to the old value of 1500. | |
| 164 #define BLOCK_FREQUENCY(f) ((f * (float) 1500) / FreqCountInvocations) | |
| 165 | |
| 166 // Register Pressure (estimate) for Splitting heuristic | |
| 167 uint _reg_pressure; | |
| 168 uint _ihrp_index; | |
| 169 uint _freg_pressure; | |
| 170 uint _fhrp_index; | |
| 171 | |
| 172 // Mark and visited bits for an LCA calculation in insert_anti_dependences. | |
| 173 // Since they hold unique node indexes, they do not need reinitialization. | |
| 174 node_idx_t _raise_LCA_mark; | |
| 175 void set_raise_LCA_mark(node_idx_t x) { _raise_LCA_mark = x; } | |
| 176 node_idx_t raise_LCA_mark() const { return _raise_LCA_mark; } | |
| 177 node_idx_t _raise_LCA_visited; | |
| 178 void set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; } | |
| 179 node_idx_t raise_LCA_visited() const { return _raise_LCA_visited; } | |
| 180 | |
| 181 // Estimated size in bytes of first instructions in a loop. | |
| 182 uint _first_inst_size; | |
| 183 uint first_inst_size() const { return _first_inst_size; } | |
| 184 void set_first_inst_size(uint s) { _first_inst_size = s; } | |
| 185 | |
| 186 // Compute the size of first instructions in this block. | |
| 187 uint compute_first_inst_size(uint& sum_size, uint inst_cnt, PhaseRegAlloc* ra); | |
| 188 | |
| 189 // Compute alignment padding if the block needs it. | |
| 190 // Align a loop if loop's padding is less or equal to padding limit | |
| 191 // or the size of first instructions in the loop > padding. | |
| 192 uint alignment_padding(int current_offset) { | |
| 193 int block_alignment = code_alignment(); | |
| 194 int max_pad = block_alignment-relocInfo::addr_unit(); | |
| 195 if( max_pad > 0 ) { | |
| 196 assert(is_power_of_2(max_pad+relocInfo::addr_unit()), ""); | |
| 197 int current_alignment = current_offset & max_pad; | |
| 198 if( current_alignment != 0 ) { | |
| 199 uint padding = (block_alignment-current_alignment) & max_pad; | |
| 418 | 200 if( has_loop_alignment() && |
| 201 padding > (uint)MaxLoopPad && | |
| 202 first_inst_size() <= padding ) { | |
| 203 return 0; | |
| 0 | 204 } |
| 418 | 205 return padding; |
| 0 | 206 } |
| 207 } | |
| 208 return 0; | |
| 209 } | |
| 210 | |
| 211 // Connector blocks. Connector blocks are basic blocks devoid of | |
| 212 // instructions, but may have relevant non-instruction Nodes, such as | |
| 213 // Phis or MergeMems. Such blocks are discovered and marked during the | |
| 214 // RemoveEmpty phase, and elided during Output. | |
| 215 bool _connector; | |
| 216 void set_connector() { _connector = true; } | |
| 217 bool is_connector() const { return _connector; }; | |
| 218 | |
| 418 | 219 // Loop_alignment will be set for blocks which are at the top of loops. |
| 220 // The block layout pass may rotate loops such that the loop head may not | |
| 221 // be the sequentially first block of the loop encountered in the linear | |
| 222 // list of blocks. If the layout pass is not run, loop alignment is set | |
| 223 // for each block which is the head of a loop. | |
| 224 uint _loop_alignment; | |
| 225 void set_loop_alignment(Block *loop_top) { | |
| 226 uint new_alignment = loop_top->compute_loop_alignment(); | |
| 227 if (new_alignment > _loop_alignment) { | |
| 228 _loop_alignment = new_alignment; | |
| 229 } | |
| 230 } | |
| 231 uint loop_alignment() const { return _loop_alignment; } | |
| 232 bool has_loop_alignment() const { return loop_alignment() > 0; } | |
| 233 | |
| 0 | 234 // Create a new Block with given head Node. |
| 235 // Creates the (empty) predecessor arrays. | |
| 236 Block( Arena *a, Node *headnode ) | |
| 237 : CFGElement(), | |
| 238 _nodes(a), | |
| 239 _succs(a), | |
| 240 _num_succs(0), | |
| 241 _pre_order(0), | |
| 242 _idom(0), | |
| 243 _loop(NULL), | |
| 244 _reg_pressure(0), | |
| 245 _ihrp_index(1), | |
| 246 _freg_pressure(0), | |
| 247 _fhrp_index(1), | |
| 248 _raise_LCA_mark(0), | |
| 249 _raise_LCA_visited(0), | |
| 250 _first_inst_size(999999), | |
| 418 | 251 _connector(false), |
| 252 _loop_alignment(0) { | |
| 0 | 253 _nodes.push(headnode); |
| 254 } | |
| 255 | |
| 256 // Index of 'end' Node | |
| 257 uint end_idx() const { | |
| 258 // %%%%% add a proj after every goto | |
| 259 // so (last->is_block_proj() != last) always, then simplify this code | |
| 260 // This will not give correct end_idx for block 0 when it only contains root. | |
| 261 int last_idx = _nodes.size() - 1; | |
| 262 Node *last = _nodes[last_idx]; | |
| 263 assert(last->is_block_proj() == last || last->is_block_proj() == _nodes[last_idx - _num_succs], ""); | |
| 264 return (last->is_block_proj() == last) ? last_idx : (last_idx - _num_succs); | |
| 265 } | |
| 266 | |
| 267 // Basic blocks have a Node which ends them. This Node determines which | |
| 268 // basic block follows this one in the program flow. This Node is either an | |
| 269 // IfNode, a GotoNode, a JmpNode, or a ReturnNode. | |
| 270 Node *end() const { return _nodes[end_idx()]; } | |
| 271 | |
| 272 // Add an instruction to an existing block. It must go after the head | |
| 273 // instruction and before the end instruction. | |
| 274 void add_inst( Node *n ) { _nodes.insert(end_idx(),n); } | |
| 275 // Find node in block | |
| 276 uint find_node( const Node *n ) const; | |
| 277 // Find and remove n from block list | |
| 278 void find_remove( const Node *n ); | |
| 279 | |
| 280 // Schedule a call next in the block | |
| 281 uint sched_call(Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call); | |
| 282 | |
| 283 // Perform basic-block local scheduling | |
| 284 Node *select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot); | |
| 285 void set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ); | |
| 286 void needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs); | |
| 287 bool schedule_local(PhaseCFG *cfg, Matcher &m, int *ready_cnt, VectorSet &next_call); | |
| 288 // Cleanup if any code lands between a Call and his Catch | |
| 289 void call_catch_cleanup(Block_Array &bbs); | |
| 290 // Detect implicit-null-check opportunities. Basically, find NULL checks | |
| 291 // with suitable memory ops nearby. Use the memory op to do the NULL check. | |
| 292 // I can generate a memory op if there is not one nearby. | |
| 293 void implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons); | |
| 294 | |
| 295 // Return the empty status of a block | |
| 296 enum { not_empty, empty_with_goto, completely_empty }; | |
| 297 int is_Empty() const; | |
| 298 | |
| 299 // Forward through connectors | |
| 300 Block* non_connector() { | |
| 301 Block* s = this; | |
| 302 while (s->is_connector()) { | |
| 303 s = s->_succs[0]; | |
| 304 } | |
| 305 return s; | |
| 306 } | |
| 307 | |
| 418 | 308 // Return true if b is a successor of this block |
| 309 bool has_successor(Block* b) const { | |
| 310 for (uint i = 0; i < _num_succs; i++ ) { | |
| 311 if (non_connector_successor(i) == b) { | |
| 312 return true; | |
| 313 } | |
| 314 } | |
| 315 return false; | |
| 316 } | |
| 317 | |
| 0 | 318 // Successor block, after forwarding through connectors |
| 319 Block* non_connector_successor(int i) const { | |
| 320 return _succs[i]->non_connector(); | |
| 321 } | |
| 322 | |
| 323 // Examine block's code shape to predict if it is not commonly executed. | |
| 324 bool has_uncommon_code() const; | |
| 325 | |
| 326 // Use frequency calculations and code shape to predict if the block | |
| 327 // is uncommon. | |
| 328 bool is_uncommon( Block_Array &bbs ) const; | |
| 329 | |
| 330 #ifndef PRODUCT | |
| 331 // Debugging print of basic block | |
| 3851 | 332 void dump_bidx(const Block* orig, outputStream* st = tty) const; |
| 333 void dump_pred(const Block_Array *bbs, Block* orig, outputStream* st = tty) const; | |
| 334 void dump_head( const Block_Array *bbs, outputStream* st = tty ) const; | |
| 335 void dump() const; | |
| 0 | 336 void dump( const Block_Array *bbs ) const; |
| 337 #endif | |
| 338 }; | |
| 339 | |
| 340 | |
| 341 //------------------------------PhaseCFG--------------------------------------- | |
| 342 // Build an array of Basic Block pointers, one per Node. | |
| 343 class PhaseCFG : public Phase { | |
| 344 private: | |
| 345 // Build a proper looking cfg. Return count of basic blocks | |
| 346 uint build_cfg(); | |
| 347 | |
| 348 // Perform DFS search. | |
| 349 // Setup 'vertex' as DFS to vertex mapping. | |
| 350 // Setup 'semi' as vertex to DFS mapping. | |
| 351 // Set 'parent' to DFS parent. | |
| 352 uint DFS( Tarjan *tarjan ); | |
| 353 | |
| 354 // Helper function to insert a node into a block | |
| 355 void schedule_node_into_block( Node *n, Block *b ); | |
| 356 | |
| 604 | 357 void replace_block_proj_ctrl( Node *n ); |
|
601
523ded093c31
6809798: SafePointScalarObject node placed into incorrect block during GCM
kvn
parents:
470
diff
changeset
|
358 |
| 0 | 359 // Set the basic block for pinned Nodes |
| 360 void schedule_pinned_nodes( VectorSet &visited ); | |
| 361 | |
| 362 // I'll need a few machine-specific GotoNodes. Clone from this one. | |
| 363 MachNode *_goto; | |
| 364 | |
| 365 Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false); | |
| 366 void verify_anti_dependences(Block* LCA, Node* load) { | |
| 367 assert(LCA == _bbs[load->_idx], "should already be scheduled"); | |
| 368 insert_anti_dependences(LCA, load, true); | |
| 369 } | |
| 370 | |
| 371 public: | |
| 372 PhaseCFG( Arena *a, RootNode *r, Matcher &m ); | |
| 373 | |
| 374 uint _num_blocks; // Count of basic blocks | |
| 375 Block_List _blocks; // List of basic blocks | |
| 376 RootNode *_root; // Root of whole program | |
| 377 Block_Array _bbs; // Map Nodes to owning Basic Block | |
| 378 Block *_broot; // Basic block of root | |
| 379 uint _rpo_ctr; | |
| 380 CFGLoop* _root_loop; | |
| 673 | 381 float _outer_loop_freq; // Outmost loop frequency |
| 0 | 382 |
| 383 // Per node latency estimation, valid only during GCM | |
| 1685 | 384 GrowableArray<uint> *_node_latency; |
| 0 | 385 |
| 386 #ifndef PRODUCT | |
| 387 bool _trace_opto_pipelining; // tracing flag | |
| 388 #endif | |
| 389 | |
|
833
acba6af809c8
6840775: Multiple JVM crashes seen with 1.6.0_10 through 1.6.0_14
kvn
parents:
673
diff
changeset
|
390 #ifdef ASSERT |
|
acba6af809c8
6840775: Multiple JVM crashes seen with 1.6.0_10 through 1.6.0_14
kvn
parents:
673
diff
changeset
|
391 Unique_Node_List _raw_oops; |
|
acba6af809c8
6840775: Multiple JVM crashes seen with 1.6.0_10 through 1.6.0_14
kvn
parents:
673
diff
changeset
|
392 #endif |
|
acba6af809c8
6840775: Multiple JVM crashes seen with 1.6.0_10 through 1.6.0_14
kvn
parents:
673
diff
changeset
|
393 |
| 0 | 394 // Build dominators |
| 395 void Dominators(); | |
| 396 | |
| 397 // Estimate block frequencies based on IfNode probabilities | |
| 398 void Estimate_Block_Frequency(); | |
| 399 | |
| 400 // Global Code Motion. See Click's PLDI95 paper. Place Nodes in specific | |
| 401 // basic blocks; i.e. _bbs now maps _idx for all Nodes to some Block. | |
| 402 void GlobalCodeMotion( Matcher &m, uint unique, Node_List &proj_list ); | |
| 403 | |
| 404 // Compute the (backwards) latency of a node from the uses | |
| 405 void latency_from_uses(Node *n); | |
| 406 | |
| 407 // Compute the (backwards) latency of a node from a single use | |
| 408 int latency_from_use(Node *n, const Node *def, Node *use); | |
| 409 | |
| 410 // Compute the (backwards) latency of a node from the uses of this instruction | |
| 411 void partial_latency_of_defs(Node *n); | |
| 412 | |
| 413 // Schedule Nodes early in their basic blocks. | |
| 414 bool schedule_early(VectorSet &visited, Node_List &roots); | |
| 415 | |
| 416 // For each node, find the latest block it can be scheduled into | |
| 417 // and then select the cheapest block between the latest and earliest | |
| 418 // block to place the node. | |
| 419 void schedule_late(VectorSet &visited, Node_List &stack); | |
| 420 | |
| 421 // Pick a block between early and late that is a cheaper alternative | |
| 422 // to late. Helper for schedule_late. | |
| 423 Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self); | |
| 424 | |
| 425 // Compute the instruction global latency with a backwards walk | |
| 426 void ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack); | |
| 427 | |
| 418 | 428 // Set loop alignment |
| 429 void set_loop_alignment(); | |
| 430 | |
| 0 | 431 // Remove empty basic blocks |
| 418 | 432 void remove_empty(); |
| 433 void fixup_flow(); | |
| 434 bool move_to_next(Block* bx, uint b_index); | |
| 435 void move_to_end(Block* bx, uint b_index); | |
| 436 void insert_goto_at(uint block_no, uint succ_no); | |
| 0 | 437 |
| 438 // Check for NeverBranch at block end. This needs to become a GOTO to the | |
| 439 // true target. NeverBranch are treated as a conditional branch that always | |
| 440 // goes the same direction for most of the optimizer and are used to give a | |
| 441 // fake exit path to infinite loops. At this late stage they need to turn | |
| 442 // into Goto's so that when you enter the infinite loop you indeed hang. | |
| 443 void convert_NeverBranch_to_Goto(Block *b); | |
| 444 | |
| 445 CFGLoop* create_loop_tree(); | |
| 446 | |
| 447 // Insert a node into a block, and update the _bbs | |
| 448 void insert( Block *b, uint idx, Node *n ) { | |
| 449 b->_nodes.insert( idx, n ); | |
| 450 _bbs.map( n->_idx, b ); | |
| 451 } | |
| 452 | |
| 453 #ifndef PRODUCT | |
| 454 bool trace_opto_pipelining() const { return _trace_opto_pipelining; } | |
| 455 | |
| 456 // Debugging print of CFG | |
| 457 void dump( ) const; // CFG only | |
| 458 void _dump_cfg( const Node *end, VectorSet &visited ) const; | |
| 459 void verify() const; | |
| 460 void dump_headers(); | |
| 461 #else | |
| 462 bool trace_opto_pipelining() const { return false; } | |
| 463 #endif | |
| 464 }; | |
| 465 | |
| 466 | |
| 418 | 467 //------------------------------UnionFind-------------------------------------- |
| 0 | 468 // Map Block indices to a block-index for a cfg-cover. |
| 469 // Array lookup in the optimized case. | |
| 470 class UnionFind : public ResourceObj { | |
| 471 uint _cnt, _max; | |
| 472 uint* _indices; | |
| 473 ReallocMark _nesting; // assertion check for reallocations | |
| 474 public: | |
| 475 UnionFind( uint max ); | |
| 476 void reset( uint max ); // Reset to identity map for [0..max] | |
| 477 | |
| 478 uint lookup( uint nidx ) const { | |
| 479 return _indices[nidx]; | |
| 480 } | |
| 481 uint operator[] (uint nidx) const { return lookup(nidx); } | |
| 482 | |
| 483 void map( uint from_idx, uint to_idx ) { | |
| 484 assert( from_idx < _cnt, "oob" ); | |
| 485 _indices[from_idx] = to_idx; | |
| 486 } | |
| 487 void extend( uint from_idx, uint to_idx ); | |
| 488 | |
| 489 uint Size() const { return _cnt; } | |
| 490 | |
| 491 uint Find( uint idx ) { | |
| 492 assert( idx < 65536, "Must fit into uint"); | |
| 493 uint uf_idx = lookup(idx); | |
| 494 return (uf_idx == idx) ? uf_idx : Find_compress(idx); | |
| 495 } | |
| 496 uint Find_compress( uint idx ); | |
| 497 uint Find_const( uint idx ) const; | |
| 498 void Union( uint idx1, uint idx2 ); | |
| 499 | |
| 500 }; | |
| 501 | |
| 502 //----------------------------BlockProbPair--------------------------- | |
| 503 // Ordered pair of Node*. | |
| 504 class BlockProbPair VALUE_OBJ_CLASS_SPEC { | |
| 505 protected: | |
| 506 Block* _target; // block target | |
| 507 float _prob; // probability of edge to block | |
| 508 public: | |
| 509 BlockProbPair() : _target(NULL), _prob(0.0) {} | |
| 510 BlockProbPair(Block* b, float p) : _target(b), _prob(p) {} | |
| 511 | |
| 512 Block* get_target() const { return _target; } | |
| 513 float get_prob() const { return _prob; } | |
| 514 }; | |
| 515 | |
| 516 //------------------------------CFGLoop------------------------------------------- | |
| 517 class CFGLoop : public CFGElement { | |
| 518 int _id; | |
| 519 int _depth; | |
| 520 CFGLoop *_parent; // root of loop tree is the method level "pseudo" loop, it's parent is null | |
| 521 CFGLoop *_sibling; // null terminated list | |
| 522 CFGLoop *_child; // first child, use child's sibling to visit all immediately nested loops | |
| 523 GrowableArray<CFGElement*> _members; // list of members of loop | |
| 524 GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities | |
| 525 float _exit_prob; // probability any loop exit is taken on a single loop iteration | |
| 526 void update_succ_freq(Block* b, float freq); | |
| 527 | |
| 528 public: | |
| 529 CFGLoop(int id) : | |
| 530 CFGElement(), | |
| 531 _id(id), | |
| 532 _depth(0), | |
| 533 _parent(NULL), | |
| 534 _sibling(NULL), | |
| 535 _child(NULL), | |
| 536 _exit_prob(1.0f) {} | |
| 537 CFGLoop* parent() { return _parent; } | |
| 538 void push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk); | |
| 539 void add_member(CFGElement *s) { _members.push(s); } | |
| 540 void add_nested_loop(CFGLoop* cl); | |
| 541 Block* head() { | |
| 542 assert(_members.at(0)->is_block(), "head must be a block"); | |
| 543 Block* hd = _members.at(0)->as_Block(); | |
| 544 assert(hd->_loop == this, "just checking"); | |
| 545 assert(hd->head()->is_Loop(), "must begin with loop head node"); | |
| 546 return hd; | |
| 547 } | |
| 548 Block* backedge_block(); // Return the block on the backedge of the loop (else NULL) | |
| 549 void compute_loop_depth(int depth); | |
| 550 void compute_freq(); // compute frequency with loop assuming head freq 1.0f | |
| 551 void scale_freq(); // scale frequency by loop trip count (including outer loops) | |
| 673 | 552 float outer_loop_freq() const; // frequency of outer loop |
| 0 | 553 bool in_loop_nest(Block* b); |
| 554 float trip_count() const { return 1.0f / _exit_prob; } | |
| 555 virtual bool is_loop() { return true; } | |
| 556 int id() { return _id; } | |
| 557 | |
| 558 #ifndef PRODUCT | |
| 559 void dump( ) const; | |
| 560 void dump_tree() const; | |
| 561 #endif | |
| 562 }; | |
| 418 | 563 |
| 564 | |
| 565 //----------------------------------CFGEdge------------------------------------ | |
| 566 // A edge between two basic blocks that will be embodied by a branch or a | |
| 567 // fall-through. | |
| 568 class CFGEdge : public ResourceObj { | |
| 569 private: | |
| 570 Block * _from; // Source basic block | |
| 571 Block * _to; // Destination basic block | |
| 572 float _freq; // Execution frequency (estimate) | |
| 573 int _state; | |
| 574 bool _infrequent; | |
| 575 int _from_pct; | |
| 576 int _to_pct; | |
| 577 | |
| 578 // Private accessors | |
| 579 int from_pct() const { return _from_pct; } | |
| 580 int to_pct() const { return _to_pct; } | |
| 581 int from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; } | |
| 582 int to_infrequent() const { return to_pct() < BlockLayoutMinDiamondPercentage; } | |
| 583 | |
| 584 public: | |
| 585 enum { | |
| 586 open, // initial edge state; unprocessed | |
| 587 connected, // edge used to connect two traces together | |
| 588 interior // edge is interior to trace (could be backedge) | |
| 589 }; | |
| 590 | |
| 591 CFGEdge(Block *from, Block *to, float freq, int from_pct, int to_pct) : | |
| 592 _from(from), _to(to), _freq(freq), | |
| 593 _from_pct(from_pct), _to_pct(to_pct), _state(open) { | |
| 594 _infrequent = from_infrequent() || to_infrequent(); | |
| 595 } | |
| 596 | |
| 597 float freq() const { return _freq; } | |
| 598 Block* from() const { return _from; } | |
| 599 Block* to () const { return _to; } | |
| 600 int infrequent() const { return _infrequent; } | |
| 601 int state() const { return _state; } | |
| 602 | |
| 603 void set_state(int state) { _state = state; } | |
| 604 | |
| 605 #ifndef PRODUCT | |
| 606 void dump( ) const; | |
| 607 #endif | |
| 608 }; | |
| 609 | |
| 610 | |
| 611 //-----------------------------------Trace------------------------------------- | |
| 612 // An ordered list of basic blocks. | |
| 613 class Trace : public ResourceObj { | |
| 614 private: | |
| 615 uint _id; // Unique Trace id (derived from initial block) | |
| 616 Block ** _next_list; // Array mapping index to next block | |
| 617 Block ** _prev_list; // Array mapping index to previous block | |
| 618 Block * _first; // First block in the trace | |
| 619 Block * _last; // Last block in the trace | |
| 620 | |
| 621 // Return the block that follows "b" in the trace. | |
| 622 Block * next(Block *b) const { return _next_list[b->_pre_order]; } | |
| 623 void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; } | |
| 624 | |
| 605 | 625 // Return the block that precedes "b" in the trace. |
| 418 | 626 Block * prev(Block *b) const { return _prev_list[b->_pre_order]; } |
| 627 void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; } | |
| 628 | |
| 629 // We've discovered a loop in this trace. Reset last to be "b", and first as | |
| 630 // the block following "b | |
| 631 void break_loop_after(Block *b) { | |
| 632 _last = b; | |
| 633 _first = next(b); | |
| 634 set_prev(_first, NULL); | |
| 635 set_next(_last, NULL); | |
| 636 } | |
| 637 | |
| 638 public: | |
| 639 | |
| 640 Trace(Block *b, Block **next_list, Block **prev_list) : | |
| 641 _first(b), | |
| 642 _last(b), | |
| 643 _next_list(next_list), | |
| 644 _prev_list(prev_list), | |
| 645 _id(b->_pre_order) { | |
| 646 set_next(b, NULL); | |
| 647 set_prev(b, NULL); | |
| 648 }; | |
| 649 | |
| 650 // Return the id number | |
| 651 uint id() const { return _id; } | |
| 652 void set_id(uint id) { _id = id; } | |
| 653 | |
| 654 // Return the first block in the trace | |
| 655 Block * first_block() const { return _first; } | |
| 656 | |
| 657 // Return the last block in the trace | |
| 658 Block * last_block() const { return _last; } | |
| 659 | |
| 660 // Insert a trace in the middle of this one after b | |
| 661 void insert_after(Block *b, Trace *tr) { | |
| 662 set_next(tr->last_block(), next(b)); | |
| 663 if (next(b) != NULL) { | |
| 664 set_prev(next(b), tr->last_block()); | |
| 665 } | |
| 666 | |
| 667 set_next(b, tr->first_block()); | |
| 668 set_prev(tr->first_block(), b); | |
| 669 | |
| 670 if (b == _last) { | |
| 671 _last = tr->last_block(); | |
| 672 } | |
| 673 } | |
| 674 | |
| 675 void insert_before(Block *b, Trace *tr) { | |
| 676 Block *p = prev(b); | |
| 677 assert(p != NULL, "use append instead"); | |
| 678 insert_after(p, tr); | |
| 679 } | |
| 680 | |
| 681 // Append another trace to this one. | |
| 682 void append(Trace *tr) { | |
| 683 insert_after(_last, tr); | |
| 684 } | |
| 685 | |
| 686 // Append a block at the end of this trace | |
| 687 void append(Block *b) { | |
| 688 set_next(_last, b); | |
| 689 set_prev(b, _last); | |
| 690 _last = b; | |
| 691 } | |
| 692 | |
| 693 // Adjust the the blocks in this trace | |
| 694 void fixup_blocks(PhaseCFG &cfg); | |
| 695 bool backedge(CFGEdge *e); | |
| 696 | |
| 697 #ifndef PRODUCT | |
| 698 void dump( ) const; | |
| 699 #endif | |
| 700 }; | |
| 701 | |
| 702 //------------------------------PhaseBlockLayout------------------------------- | |
| 703 // Rearrange blocks into some canonical order, based on edges and their frequencies | |
| 704 class PhaseBlockLayout : public Phase { | |
| 705 PhaseCFG &_cfg; // Control flow graph | |
| 706 | |
| 707 GrowableArray<CFGEdge *> *edges; | |
| 708 Trace **traces; | |
| 709 Block **next; | |
| 710 Block **prev; | |
| 711 UnionFind *uf; | |
| 712 | |
| 713 // Given a block, find its encompassing Trace | |
| 714 Trace * trace(Block *b) { | |
| 715 return traces[uf->Find_compress(b->_pre_order)]; | |
| 716 } | |
| 717 public: | |
| 718 PhaseBlockLayout(PhaseCFG &cfg); | |
| 719 | |
| 720 void find_edges(); | |
| 721 void grow_traces(); | |
| 722 void merge_traces(bool loose_connections); | |
| 723 void reorder_traces(int count); | |
| 724 void union_traces(Trace* from, Trace* to); | |
| 725 }; | |
| 1972 | 726 |
| 727 #endif // SHARE_VM_OPTO_BLOCK_HPP |