1 /*
2 * Copyright (c) 1997, 2009, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 class LoopTree;
26 class MachCallNode;
27 class MachSafePointNode;
28 class Matcher;
29 class PhaseCFG;
30 class PhaseLive;
31 class PhaseRegAlloc;
32 class PhaseChaitin;
33
34 #define OPTO_DEBUG_SPLIT_FREQ BLOCK_FREQUENCY(0.001)
35 #define OPTO_LRG_HIGH_FREQ BLOCK_FREQUENCY(0.25)
36
37 //------------------------------LRG--------------------------------------------
38 // Live-RanGe structure.
39 class LRG : public ResourceObj {
40 public:
41 enum { SPILL_REG=29999 }; // Register number of a spilled LRG
42
43 double _cost; // 2 for loads/1 for stores times block freq
44 double _area; // Sum of all simultaneously live values
45 double score() const; // Compute score from cost and area
46 double _maxfreq; // Maximum frequency of any def or use
47
48 Node *_def; // Check for multi-def live ranges
49 #ifndef PRODUCT
50 GrowableArray<Node*>* _defs;
51 #endif
52
53 uint _risk_bias; // Index of LRG which we want to avoid color
54 uint _copy_bias; // Index of LRG which we want to share color
55
56 uint _next; // Index of next LRG in linked list
57 uint _prev; // Index of prev LRG in linked list
58 private:
59 uint _reg; // Chosen register; undefined if mask is plural
60 public:
61 // Return chosen register for this LRG. Error if the LRG is not bound to
62 // a single register.
63 OptoReg::Name reg() const { return OptoReg::Name(_reg); }
64 void set_reg( OptoReg::Name r ) { _reg = r; }
65
66 private:
67 uint _eff_degree; // Effective degree: Sum of neighbors _num_regs
68 public:
69 int degree() const { assert( _degree_valid, "" ); return _eff_degree; }
70 // Degree starts not valid and any change to the IFG neighbor
71 // set makes it not valid.
72 void set_degree( uint degree ) { _eff_degree = degree; debug_only(_degree_valid = 1;) }
73 // Made a change that hammered degree
74 void invalid_degree() { debug_only(_degree_valid=0;) }
75 // Incrementally modify degree. If it was correct, it should remain correct
76 void inc_degree( uint mod ) { _eff_degree += mod; }
77 // Compute the degree between 2 live ranges
78 int compute_degree( LRG &l ) const;
79
80 private:
81 RegMask _mask; // Allowed registers for this LRG
82 uint _mask_size; // cache of _mask.Size();
83 public:
84 int compute_mask_size() const { return _mask.is_AllStack() ? 65535 : _mask.Size(); }
85 void set_mask_size( int size ) {
86 assert((size == 65535) || (size == (int)_mask.Size()), "");
87 _mask_size = size;
88 debug_only(_msize_valid=1;)
89 debug_only( if( _num_regs == 2 && !_fat_proj ) _mask.VerifyPairs(); )
90 }
91 void compute_set_mask_size() { set_mask_size(compute_mask_size()); }
92 int mask_size() const { assert( _msize_valid, "mask size not valid" );
93 return _mask_size; }
94 // Get the last mask size computed, even if it does not match the
95 // count of bits in the current mask.
96 int get_invalid_mask_size() const { return _mask_size; }
97 const RegMask &mask() const { return _mask; }
98 void set_mask( const RegMask &rm ) { _mask = rm; debug_only(_msize_valid=0;)}
99 void AND( const RegMask &rm ) { _mask.AND(rm); debug_only(_msize_valid=0;)}
100 void SUBTRACT( const RegMask &rm ) { _mask.SUBTRACT(rm); debug_only(_msize_valid=0;)}
101 void Clear() { _mask.Clear() ; debug_only(_msize_valid=1); _mask_size = 0; }
102 void Set_All() { _mask.Set_All(); debug_only(_msize_valid=1); _mask_size = RegMask::CHUNK_SIZE; }
103 void Insert( OptoReg::Name reg ) { _mask.Insert(reg); debug_only(_msize_valid=0;) }
104 void Remove( OptoReg::Name reg ) { _mask.Remove(reg); debug_only(_msize_valid=0;) }
105 void ClearToPairs() { _mask.ClearToPairs(); debug_only(_msize_valid=0;) }
106
107 // Number of registers this live range uses when it colors
108 private:
109 uint8 _num_regs; // 2 for Longs and Doubles, 1 for all else
110 // except _num_regs is kill count for fat_proj
111 public:
112 int num_regs() const { return _num_regs; }
113 void set_num_regs( int reg ) { assert( _num_regs == reg || !_num_regs, "" ); _num_regs = reg; }
114
115 private:
116 // Number of physical registers this live range uses when it colors
117 // Architecture and register-set dependent
118 uint8 _reg_pressure;
119 public:
120 void set_reg_pressure(int i) { _reg_pressure = i; }
121 int reg_pressure() const { return _reg_pressure; }
122
123 // How much 'wiggle room' does this live range have?
124 // How many color choices can it make (scaled by _num_regs)?
125 int degrees_of_freedom() const { return mask_size() - _num_regs; }
126 // Bound LRGs have ZERO degrees of freedom. We also count
127 // must_spill as bound.
128 bool is_bound () const { return _is_bound; }
129 // Negative degrees-of-freedom; even with no neighbors this
130 // live range must spill.
131 bool not_free() const { return degrees_of_freedom() < 0; }
132 // Is this live range of "low-degree"? Trivially colorable?
133 bool lo_degree () const { return degree() <= degrees_of_freedom(); }
134 // Is this live range just barely "low-degree"? Trivially colorable?
135 bool just_lo_degree () const { return degree() == degrees_of_freedom(); }
136
137 uint _is_oop:1, // Live-range holds an oop
138 _is_float:1, // True if in float registers
139 _was_spilled1:1, // True if prior spilling on def
140 _was_spilled2:1, // True if twice prior spilling on def
141 _is_bound:1, // live range starts life with no
142 // degrees of freedom.
143 _direct_conflict:1, // True if def and use registers in conflict
144 _must_spill:1, // live range has lost all degrees of freedom
145 // If _fat_proj is set, live range does NOT require aligned, adjacent
146 // registers and has NO interferences.
147 // If _fat_proj is clear, live range requires num_regs() to be a power of
148 // 2, and it requires registers to form an aligned, adjacent set.
149 _fat_proj:1, //
150 _was_lo:1, // Was lo-degree prior to coalesce
151 _msize_valid:1, // _mask_size cache valid
152 _degree_valid:1, // _degree cache valid
153 _has_copy:1, // Adjacent to some copy instruction
154 _at_risk:1; // Simplify says this guy is at risk to spill
155
156
157 // Alive if non-zero, dead if zero
158 bool alive() const { return _def != NULL; }
159 bool is_multidef() const { return _def == NodeSentinel; }
160 bool is_singledef() const { return _def != NodeSentinel; }
161
162 #ifndef PRODUCT
163 void dump( ) const;
164 #endif
165 };
166
167 //------------------------------LRG_List---------------------------------------
168 // Map Node indices to Live RanGe indices.
169 // Array lookup in the optimized case.
170 class LRG_List : public ResourceObj {
171 uint _cnt, _max;
172 uint* _lidxs;
173 ReallocMark _nesting; // assertion check for reallocations
174 public:
175 LRG_List( uint max );
176
177 uint lookup( uint nidx ) const {
178 return _lidxs[nidx];
179 }
180 uint operator[] (uint nidx) const { return lookup(nidx); }
181
182 void map( uint nidx, uint lidx ) {
183 assert( nidx < _cnt, "oob" );
184 _lidxs[nidx] = lidx;
185 }
186 void extend( uint nidx, uint lidx );
187
188 uint Size() const { return _cnt; }
189 };
190
191 //------------------------------IFG--------------------------------------------
192 // InterFerence Graph
193 // An undirected graph implementation. Created with a fixed number of
194 // vertices. Edges can be added & tested. Vertices can be removed, then
195 // added back later with all edges intact. Can add edges between one vertex
196 // and a list of other vertices. Can union vertices (and their edges)
197 // together. The IFG needs to be really really fast, and also fairly
198 // abstract! It needs abstraction so I can fiddle with the implementation to
199 // get even more speed.
200 class PhaseIFG : public Phase {
201 // Current implementation: a triangular adjacency list.
202
203 // Array of adjacency-lists, indexed by live-range number
204 IndexSet *_adjs;
205
206 // Assertion bit for proper use of Squaring
207 bool _is_square;
208
209 // Live range structure goes here
210 LRG *_lrgs; // Array of LRG structures
211
212 public:
213 // Largest live-range number
214 uint _maxlrg;
215
216 Arena *_arena;
217
218 // Keep track of inserted and deleted Nodes
219 VectorSet *_yanked;
220
221 PhaseIFG( Arena *arena );
222 void init( uint maxlrg );
223
224 // Add edge between a and b. Returns true if actually addded.
225 int add_edge( uint a, uint b );
226
227 // Add edge between a and everything in the vector
228 void add_vector( uint a, IndexSet *vec );
229
230 // Test for edge existance
231 int test_edge( uint a, uint b ) const;
232
233 // Square-up matrix for faster Union
234 void SquareUp();
235
236 // Return number of LRG neighbors
237 uint neighbor_cnt( uint a ) const { return _adjs[a].count(); }
238 // Union edges of b into a on Squared-up matrix
239 void Union( uint a, uint b );
240 // Test for edge in Squared-up matrix
241 int test_edge_sq( uint a, uint b ) const;
242 // Yank a Node and all connected edges from the IFG. Be prepared to
243 // re-insert the yanked Node in reverse order of yanking. Return a
244 // list of neighbors (edges) yanked.
245 IndexSet *remove_node( uint a );
246 // Reinsert a yanked Node
247 void re_insert( uint a );
248 // Return set of neighbors
249 IndexSet *neighbors( uint a ) const { return &_adjs[a]; }
250
251 #ifndef PRODUCT
252 // Dump the IFG
253 void dump() const;
254 void stats() const;
255 void verify( const PhaseChaitin * ) const;
256 #endif
257
258 //--------------- Live Range Accessors
259 LRG &lrgs(uint idx) const { assert(idx < _maxlrg, "oob"); return _lrgs[idx]; }
260
261 // Compute and set effective degree. Might be folded into SquareUp().
262 void Compute_Effective_Degree();
263
264 // Compute effective degree as the sum of neighbors' _sizes.
265 int effective_degree( uint lidx ) const;
266 };
267
268 // TEMPORARILY REPLACED WITH COMMAND LINE FLAG
269
270 //// !!!!! Magic Constants need to move into ad file
271 #ifdef SPARC
272 //#define FLOAT_PRESSURE 30 /* SFLT_REG_mask.Size() - 1 */
273 //#define INT_PRESSURE 23 /* NOTEMP_I_REG_mask.Size() - 1 */
274 #define FLOAT_INCREMENT(regs) regs
275 #else
276 //#define FLOAT_PRESSURE 6
277 //#define INT_PRESSURE 6
278 #define FLOAT_INCREMENT(regs) 1
279 #endif
280
281 //------------------------------Chaitin----------------------------------------
282 // Briggs-Chaitin style allocation, mostly.
283 class PhaseChaitin : public PhaseRegAlloc {
284
285 int _trip_cnt;
286 int _alternate;
287
288 uint _maxlrg; // Max live range number
289 LRG &lrgs(uint idx) const { return _ifg->lrgs(idx); }
290 PhaseLive *_live; // Liveness, used in the interference graph
291 PhaseIFG *_ifg; // Interference graph (for original chunk)
292 Node_List **_lrg_nodes; // Array of node; lists for lrgs which spill
293 VectorSet _spilled_once; // Nodes that have been spilled
294 VectorSet _spilled_twice; // Nodes that have been spilled twice
295
296 LRG_List _names; // Map from Nodes to Live RanGes
297
298 // Union-find map. Declared as a short for speed.
299 // Indexed by live-range number, it returns the compacted live-range number
300 LRG_List _uf_map;
301 // Reset the Union-Find map to identity
302 void reset_uf_map( uint maxlrg );
303 // Remove the need for the Union-Find mapping
304 void compress_uf_map_for_nodes( );
305
306 // Combine the Live Range Indices for these 2 Nodes into a single live
307 // range. Future requests for any Node in either live range will
308 // return the live range index for the combined live range.
309 void Union( const Node *src, const Node *dst );
310
311 void new_lrg( const Node *x, uint lrg );
312
313 // Compact live ranges, removing unused ones. Return new maxlrg.
314 void compact();
315
316 uint _lo_degree; // Head of lo-degree LRGs list
317 uint _lo_stk_degree; // Head of lo-stk-degree LRGs list
318 uint _hi_degree; // Head of hi-degree LRGs list
319 uint _simplified; // Linked list head of simplified LRGs
320
321 // Helper functions for Split()
322 uint split_DEF( Node *def, Block *b, int loc, uint max, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx );
323 uint split_USE( Node *def, Block *b, Node *use, uint useidx, uint max, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx );
324 int clone_projs( Block *b, uint idx, Node *con, Node *copy, uint &maxlrg );
325 Node *split_Rematerialize(Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits,
326 int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru);
327 // True if lidx is used before any real register is def'd in the block
328 bool prompt_use( Block *b, uint lidx );
329 Node *get_spillcopy_wide( Node *def, Node *use, uint uidx );
330 // Insert the spill at chosen location. Skip over any intervening Proj's or
331 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block
332 // instead. Update high-pressure indices. Create a new live range.
333 void insert_proj( Block *b, uint i, Node *spill, uint maxlrg );
334
335 bool is_high_pressure( Block *b, LRG *lrg, uint insidx );
336
337 uint _oldphi; // Node index which separates pre-allocation nodes
338
339 Block **_blks; // Array of blocks sorted by frequency for coalescing
340
341 float _high_frequency_lrg; // Frequency at which LRG will be spilled for debug info
342
343 #ifndef PRODUCT
344 bool _trace_spilling;
345 #endif
346
347 public:
348 PhaseChaitin( uint unique, PhaseCFG &cfg, Matcher &matcher );
349 ~PhaseChaitin() {}
350
351 // Convert a Node into a Live Range Index - a lidx
352 uint Find( const Node *n ) {
353 uint lidx = n2lidx(n);
354 uint uf_lidx = _uf_map[lidx];
355 return (uf_lidx == lidx) ? uf_lidx : Find_compress(n);
356 }
357 uint Find_const( uint lrg ) const;
358 uint Find_const( const Node *n ) const;
359
360 // Do all the real work of allocate
361 void Register_Allocate();
362
363 uint n2lidx( const Node *n ) const { return _names[n->_idx]; }
364
365 float high_frequency_lrg() const { return _high_frequency_lrg; }
366
367 #ifndef PRODUCT
368 bool trace_spilling() const { return _trace_spilling; }
369 #endif
370
371 private:
372 // De-SSA the world. Assign registers to Nodes. Use the same register for
373 // all inputs to a PhiNode, effectively coalescing live ranges. Insert
374 // copies as needed.
375 void de_ssa();
376 uint Find_compress( const Node *n );
377 uint Find( uint lidx ) {
378 uint uf_lidx = _uf_map[lidx];
379 return (uf_lidx == lidx) ? uf_lidx : Find_compress(lidx);
380 }
381 uint Find_compress( uint lidx );
382
383 uint Find_id( const Node *n ) {
384 uint retval = n2lidx(n);
385 assert(retval == Find(n),"Invalid node to lidx mapping");
386 return retval;
387 }
388
389 // Add edge between reg and everything in the vector.
390 // Same as _ifg->add_vector(reg,live) EXCEPT use the RegMask
391 // information to trim the set of interferences. Return the
392 // count of edges added.
393 void interfere_with_live( uint reg, IndexSet *live );
394 // Count register pressure for asserts
395 uint count_int_pressure( IndexSet *liveout );
396 uint count_float_pressure( IndexSet *liveout );
397
398 // Build the interference graph using virtual registers only.
399 // Used for aggressive coalescing.
400 void build_ifg_virtual( );
401
402 // Build the interference graph using physical registers when available.
403 // That is, if 2 live ranges are simultaneously alive but in their
404 // acceptable register sets do not overlap, then they do not interfere.
405 uint build_ifg_physical( ResourceArea *a );
406
407 // Gather LiveRanGe information, including register masks and base pointer/
408 // derived pointer relationships.
409 void gather_lrg_masks( bool mod_cisc_masks );
410
411 // Force the bases of derived pointers to be alive at GC points.
412 bool stretch_base_pointer_live_ranges( ResourceArea *a );
413 // Helper to stretch above; recursively discover the base Node for
414 // a given derived Node. Easy for AddP-related machine nodes, but
415 // needs to be recursive for derived Phis.
416 Node *find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg );
417
418 // Set the was-lo-degree bit. Conservative coalescing should not change the
419 // colorability of the graph. If any live range was of low-degree before
420 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
421 void set_was_low();
422
423 // Split live-ranges that must spill due to register conflicts (as opposed
424 // to capacity spills). Typically these are things def'd in a register
425 // and used on the stack or vice-versa.
426 void pre_spill();
427
428 // Init LRG caching of degree, numregs. Init lo_degree list.
429 void cache_lrg_info( );
430
431 // Simplify the IFG by removing LRGs of low degree with no copies
432 void Pre_Simplify();
433
434 // Simplify the IFG by removing LRGs of low degree
435 void Simplify();
436
437 // Select colors by re-inserting edges into the IFG.
438 // Return TRUE if any spills occurred.
439 uint Select( );
440 // Helper function for select which allows biased coloring
441 OptoReg::Name choose_color( LRG &lrg, int chunk );
442 // Helper function which implements biasing heuristic
443 OptoReg::Name bias_color( LRG &lrg, int chunk );
444
445 // Split uncolorable live ranges
446 // Return new number of live ranges
447 uint Split( uint maxlrg );
448
449 // Copy 'was_spilled'-edness from one Node to another.
450 void copy_was_spilled( Node *src, Node *dst );
451 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
452 void set_was_spilled( Node *n );
453
454 // Convert ideal spill-nodes into machine loads & stores
455 // Set C->failing when fixup spills could not complete, node limit exceeded.
456 void fixup_spills();
457
458 // Post-Allocation peephole copy removal
459 void post_allocate_copy_removal();
460 Node *skip_copies( Node *c );
461 // Replace the old node with the current live version of that value
462 // and yank the old value if it's dead.
463 int replace_and_yank_if_dead( Node *old, OptoReg::Name nreg,
464 Block *current_block, Node_List& value, Node_List& regnd ) {
465 Node* v = regnd[nreg];
466 assert(v->outcnt() != 0, "no dead values");
467 old->replace_by(v);
468 return yank_if_dead(old, current_block, &value, ®nd);
469 }
470
471 int yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd );
472 int elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs );
473 int use_prior_register( Node *copy, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd );
474 bool may_be_copy_of_callee( Node *def ) const;
475
476 // If nreg already contains the same constant as val then eliminate it
477 bool eliminate_copy_of_constant(Node* val, Node* n,
478 Block *current_block, Node_List& value, Node_List ®nd,
479 OptoReg::Name nreg, OptoReg::Name nreg2);
480 // Extend the node to LRG mapping
481 void add_reference( const Node *node, const Node *old_node);
482
483 private:
484
485 static int _final_loads, _final_stores, _final_copies, _final_memoves;
486 static double _final_load_cost, _final_store_cost, _final_copy_cost, _final_memove_cost;
487 static int _conserv_coalesce, _conserv_coalesce_pair;
488 static int _conserv_coalesce_trie, _conserv_coalesce_quad;
489 static int _post_alloc;
490 static int _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce;
491 static int _used_cisc_instructions, _unused_cisc_instructions;
492 static int _allocator_attempts, _allocator_successes;
493
494 #ifndef PRODUCT
495 static uint _high_pressure, _low_pressure;
496
497 void dump() const;
498 void dump( const Node *n ) const;
499 void dump( const Block * b ) const;
500 void dump_degree_lists() const;
501 void dump_simplified() const;
502 void dump_lrg( uint lidx ) const;
503 void dump_bb( uint pre_order ) const;
504
505 // Verify that base pointers and derived pointers are still sane
506 void verify_base_ptrs( ResourceArea *a ) const;
507
508 void verify( ResourceArea *a, bool verify_ifg = false ) const;
509
510 void dump_for_spill_split_recycle() const;
511
512 public:
513 void dump_frame() const;
514 char *dump_register( const Node *n, char *buf ) const;
515 private:
516 static void print_chaitin_statistics();
517 #endif
518 friend class PhaseCoalesce;
519 friend class PhaseAggressiveCoalesce;
520 friend class PhaseConservativeCoalesce;
521 };