1 /*
2 * Copyright (c) 1997, 2010, 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 #ifndef SHARE_VM_OPTO_CHAITIN_HPP
26 #define SHARE_VM_OPTO_CHAITIN_HPP
27
28 #include "code/vmreg.hpp"
29 #include "libadt/port.hpp"
30 #include "memory/resourceArea.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/live.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/phase.hpp"
35 #include "opto/regalloc.hpp"
36 #include "opto/regmask.hpp"
37
38 class LoopTree;
39 class MachCallNode;
40 class MachSafePointNode;
41 class Matcher;
42 class PhaseCFG;
43 class PhaseLive;
44 class PhaseRegAlloc;
45 class PhaseChaitin;
46
47 #define OPTO_DEBUG_SPLIT_FREQ BLOCK_FREQUENCY(0.001)
48 #define OPTO_LRG_HIGH_FREQ BLOCK_FREQUENCY(0.25)
49
50 //------------------------------LRG--------------------------------------------
51 // Live-RanGe structure.
52 class LRG : public ResourceObj {
53 public:
54 enum { SPILL_REG=29999 }; // Register number of a spilled LRG
55
56 double _cost; // 2 for loads/1 for stores times block freq
57 double _area; // Sum of all simultaneously live values
58 double score() const; // Compute score from cost and area
59 double _maxfreq; // Maximum frequency of any def or use
60
61 Node *_def; // Check for multi-def live ranges
62 #ifndef PRODUCT
63 GrowableArray<Node*>* _defs;
64 #endif
65
66 uint _risk_bias; // Index of LRG which we want to avoid color
67 uint _copy_bias; // Index of LRG which we want to share color
68
69 uint _next; // Index of next LRG in linked list
70 uint _prev; // Index of prev LRG in linked list
71 private:
72 uint _reg; // Chosen register; undefined if mask is plural
73 public:
74 // Return chosen register for this LRG. Error if the LRG is not bound to
75 // a single register.
76 OptoReg::Name reg() const { return OptoReg::Name(_reg); }
77 void set_reg( OptoReg::Name r ) { _reg = r; }
78
79 private:
80 uint _eff_degree; // Effective degree: Sum of neighbors _num_regs
81 public:
82 int degree() const { assert( _degree_valid, "" ); return _eff_degree; }
83 // Degree starts not valid and any change to the IFG neighbor
84 // set makes it not valid.
85 void set_degree( uint degree ) { _eff_degree = degree; debug_only(_degree_valid = 1;) }
86 // Made a change that hammered degree
87 void invalid_degree() { debug_only(_degree_valid=0;) }
88 // Incrementally modify degree. If it was correct, it should remain correct
89 void inc_degree( uint mod ) { _eff_degree += mod; }
90 // Compute the degree between 2 live ranges
91 int compute_degree( LRG &l ) const;
92
93 private:
94 RegMask _mask; // Allowed registers for this LRG
95 uint _mask_size; // cache of _mask.Size();
96 public:
97 int compute_mask_size() const { return _mask.is_AllStack() ? 65535 : _mask.Size(); }
98 void set_mask_size( int size ) {
99 assert((size == 65535) || (size == (int)_mask.Size()), "");
100 _mask_size = size;
101 debug_only(_msize_valid=1;)
102 debug_only( if( _num_regs == 2 && !_fat_proj ) _mask.VerifyPairs(); )
103 }
104 void compute_set_mask_size() { set_mask_size(compute_mask_size()); }
105 int mask_size() const { assert( _msize_valid, "mask size not valid" );
106 return _mask_size; }
107 // Get the last mask size computed, even if it does not match the
108 // count of bits in the current mask.
109 int get_invalid_mask_size() const { return _mask_size; }
110 const RegMask &mask() const { return _mask; }
111 void set_mask( const RegMask &rm ) { _mask = rm; debug_only(_msize_valid=0;)}
112 void AND( const RegMask &rm ) { _mask.AND(rm); debug_only(_msize_valid=0;)}
113 void SUBTRACT( const RegMask &rm ) { _mask.SUBTRACT(rm); debug_only(_msize_valid=0;)}
114 void Clear() { _mask.Clear() ; debug_only(_msize_valid=1); _mask_size = 0; }
115 void Set_All() { _mask.Set_All(); debug_only(_msize_valid=1); _mask_size = RegMask::CHUNK_SIZE; }
116 void Insert( OptoReg::Name reg ) { _mask.Insert(reg); debug_only(_msize_valid=0;) }
117 void Remove( OptoReg::Name reg ) { _mask.Remove(reg); debug_only(_msize_valid=0;) }
118 void ClearToPairs() { _mask.ClearToPairs(); debug_only(_msize_valid=0;) }
119
120 // Number of registers this live range uses when it colors
121 private:
122 uint8 _num_regs; // 2 for Longs and Doubles, 1 for all else
123 // except _num_regs is kill count for fat_proj
124 public:
125 int num_regs() const { return _num_regs; }
126 void set_num_regs( int reg ) { assert( _num_regs == reg || !_num_regs, "" ); _num_regs = reg; }
127
128 private:
129 // Number of physical registers this live range uses when it colors
130 // Architecture and register-set dependent
131 uint8 _reg_pressure;
132 public:
133 void set_reg_pressure(int i) { _reg_pressure = i; }
134 int reg_pressure() const { return _reg_pressure; }
135
136 // How much 'wiggle room' does this live range have?
137 // How many color choices can it make (scaled by _num_regs)?
138 int degrees_of_freedom() const { return mask_size() - _num_regs; }
139 // Bound LRGs have ZERO degrees of freedom. We also count
140 // must_spill as bound.
141 bool is_bound () const { return _is_bound; }
142 // Negative degrees-of-freedom; even with no neighbors this
143 // live range must spill.
144 bool not_free() const { return degrees_of_freedom() < 0; }
145 // Is this live range of "low-degree"? Trivially colorable?
146 bool lo_degree () const { return degree() <= degrees_of_freedom(); }
147 // Is this live range just barely "low-degree"? Trivially colorable?
148 bool just_lo_degree () const { return degree() == degrees_of_freedom(); }
149
150 uint _is_oop:1, // Live-range holds an oop
151 _is_float:1, // True if in float registers
152 _was_spilled1:1, // True if prior spilling on def
153 _was_spilled2:1, // True if twice prior spilling on def
154 _is_bound:1, // live range starts life with no
155 // degrees of freedom.
156 _direct_conflict:1, // True if def and use registers in conflict
157 _must_spill:1, // live range has lost all degrees of freedom
158 // If _fat_proj is set, live range does NOT require aligned, adjacent
159 // registers and has NO interferences.
160 // If _fat_proj is clear, live range requires num_regs() to be a power of
161 // 2, and it requires registers to form an aligned, adjacent set.
162 _fat_proj:1, //
163 _was_lo:1, // Was lo-degree prior to coalesce
164 _msize_valid:1, // _mask_size cache valid
165 _degree_valid:1, // _degree cache valid
166 _has_copy:1, // Adjacent to some copy instruction
167 _at_risk:1; // Simplify says this guy is at risk to spill
168
169
170 // Alive if non-zero, dead if zero
171 bool alive() const { return _def != NULL; }
172 bool is_multidef() const { return _def == NodeSentinel; }
173 bool is_singledef() const { return _def != NodeSentinel; }
174
175 #ifndef PRODUCT
176 void dump( ) const;
177 #endif
178 };
179
180 //------------------------------LRG_List---------------------------------------
181 // Map Node indices to Live RanGe indices.
182 // Array lookup in the optimized case.
183 class LRG_List : public ResourceObj {
184 uint _cnt, _max;
185 uint* _lidxs;
186 ReallocMark _nesting; // assertion check for reallocations
187 public:
188 LRG_List( uint max );
189
190 uint lookup( uint nidx ) const {
191 return _lidxs[nidx];
192 }
193 uint operator[] (uint nidx) const { return lookup(nidx); }
194
195 void map( uint nidx, uint lidx ) {
196 assert( nidx < _cnt, "oob" );
197 _lidxs[nidx] = lidx;
198 }
199 void extend( uint nidx, uint lidx );
200
201 uint Size() const { return _cnt; }
202 };
203
204 //------------------------------IFG--------------------------------------------
205 // InterFerence Graph
206 // An undirected graph implementation. Created with a fixed number of
207 // vertices. Edges can be added & tested. Vertices can be removed, then
208 // added back later with all edges intact. Can add edges between one vertex
209 // and a list of other vertices. Can union vertices (and their edges)
210 // together. The IFG needs to be really really fast, and also fairly
211 // abstract! It needs abstraction so I can fiddle with the implementation to
212 // get even more speed.
213 class PhaseIFG : public Phase {
214 // Current implementation: a triangular adjacency list.
215
216 // Array of adjacency-lists, indexed by live-range number
217 IndexSet *_adjs;
218
219 // Assertion bit for proper use of Squaring
220 bool _is_square;
221
222 // Live range structure goes here
223 LRG *_lrgs; // Array of LRG structures
224
225 public:
226 // Largest live-range number
227 uint _maxlrg;
228
229 Arena *_arena;
230
231 // Keep track of inserted and deleted Nodes
232 VectorSet *_yanked;
233
234 PhaseIFG( Arena *arena );
235 void init( uint maxlrg );
236
237 // Add edge between a and b. Returns true if actually addded.
238 int add_edge( uint a, uint b );
239
240 // Add edge between a and everything in the vector
241 void add_vector( uint a, IndexSet *vec );
242
243 // Test for edge existance
244 int test_edge( uint a, uint b ) const;
245
246 // Square-up matrix for faster Union
247 void SquareUp();
248
249 // Return number of LRG neighbors
250 uint neighbor_cnt( uint a ) const { return _adjs[a].count(); }
251 // Union edges of b into a on Squared-up matrix
252 void Union( uint a, uint b );
253 // Test for edge in Squared-up matrix
254 int test_edge_sq( uint a, uint b ) const;
255 // Yank a Node and all connected edges from the IFG. Be prepared to
256 // re-insert the yanked Node in reverse order of yanking. Return a
257 // list of neighbors (edges) yanked.
258 IndexSet *remove_node( uint a );
259 // Reinsert a yanked Node
260 void re_insert( uint a );
261 // Return set of neighbors
262 IndexSet *neighbors( uint a ) const { return &_adjs[a]; }
263
264 #ifndef PRODUCT
265 // Dump the IFG
266 void dump() const;
267 void stats() const;
268 void verify( const PhaseChaitin * ) const;
269 #endif
270
271 //--------------- Live Range Accessors
272 LRG &lrgs(uint idx) const { assert(idx < _maxlrg, "oob"); return _lrgs[idx]; }
273
274 // Compute and set effective degree. Might be folded into SquareUp().
275 void Compute_Effective_Degree();
276
277 // Compute effective degree as the sum of neighbors' _sizes.
278 int effective_degree( uint lidx ) const;
279 };
280
281 // TEMPORARILY REPLACED WITH COMMAND LINE FLAG
282
283 //// !!!!! Magic Constants need to move into ad file
284 #ifdef SPARC
285 //#define FLOAT_PRESSURE 30 /* SFLT_REG_mask.Size() - 1 */
286 //#define INT_PRESSURE 23 /* NOTEMP_I_REG_mask.Size() - 1 */
287 #define FLOAT_INCREMENT(regs) regs
288 #else
289 //#define FLOAT_PRESSURE 6
290 //#define INT_PRESSURE 6
291 #define FLOAT_INCREMENT(regs) 1
292 #endif
293
294 //------------------------------Chaitin----------------------------------------
295 // Briggs-Chaitin style allocation, mostly.
296 class PhaseChaitin : public PhaseRegAlloc {
297
298 int _trip_cnt;
299 int _alternate;
300
301 uint _maxlrg; // Max live range number
302 LRG &lrgs(uint idx) const { return _ifg->lrgs(idx); }
303 PhaseLive *_live; // Liveness, used in the interference graph
304 PhaseIFG *_ifg; // Interference graph (for original chunk)
305 Node_List **_lrg_nodes; // Array of node; lists for lrgs which spill
306 VectorSet _spilled_once; // Nodes that have been spilled
307 VectorSet _spilled_twice; // Nodes that have been spilled twice
308
309 LRG_List _names; // Map from Nodes to Live RanGes
310
311 // Union-find map. Declared as a short for speed.
312 // Indexed by live-range number, it returns the compacted live-range number
313 LRG_List _uf_map;
314 // Reset the Union-Find map to identity
315 void reset_uf_map( uint maxlrg );
316 // Remove the need for the Union-Find mapping
317 void compress_uf_map_for_nodes( );
318
319 // Combine the Live Range Indices for these 2 Nodes into a single live
320 // range. Future requests for any Node in either live range will
321 // return the live range index for the combined live range.
322 void Union( const Node *src, const Node *dst );
323
324 void new_lrg( const Node *x, uint lrg );
325
326 // Compact live ranges, removing unused ones. Return new maxlrg.
327 void compact();
328
329 uint _lo_degree; // Head of lo-degree LRGs list
330 uint _lo_stk_degree; // Head of lo-stk-degree LRGs list
331 uint _hi_degree; // Head of hi-degree LRGs list
332 uint _simplified; // Linked list head of simplified LRGs
333
334 // Helper functions for Split()
335 uint split_DEF( Node *def, Block *b, int loc, uint max, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx );
336 uint split_USE( Node *def, Block *b, Node *use, uint useidx, uint max, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx );
337 int clone_projs( Block *b, uint idx, Node *con, Node *copy, uint &maxlrg );
338 Node *split_Rematerialize(Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits,
339 int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru);
340 // True if lidx is used before any real register is def'd in the block
341 bool prompt_use( Block *b, uint lidx );
342 Node *get_spillcopy_wide( Node *def, Node *use, uint uidx );
343 // Insert the spill at chosen location. Skip over any intervening Proj's or
344 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block
345 // instead. Update high-pressure indices. Create a new live range.
346 void insert_proj( Block *b, uint i, Node *spill, uint maxlrg );
347
348 bool is_high_pressure( Block *b, LRG *lrg, uint insidx );
349
350 uint _oldphi; // Node index which separates pre-allocation nodes
351
352 Block **_blks; // Array of blocks sorted by frequency for coalescing
353
354 float _high_frequency_lrg; // Frequency at which LRG will be spilled for debug info
355
356 #ifndef PRODUCT
357 bool _trace_spilling;
358 #endif
359
360 public:
361 PhaseChaitin( uint unique, PhaseCFG &cfg, Matcher &matcher );
362 ~PhaseChaitin() {}
363
364 // Convert a Node into a Live Range Index - a lidx
365 uint Find( const Node *n ) {
366 uint lidx = n2lidx(n);
367 uint uf_lidx = _uf_map[lidx];
368 return (uf_lidx == lidx) ? uf_lidx : Find_compress(n);
369 }
370 uint Find_const( uint lrg ) const;
371 uint Find_const( const Node *n ) const;
372
373 // Do all the real work of allocate
374 void Register_Allocate();
375
376 uint n2lidx( const Node *n ) const { return _names[n->_idx]; }
377
378 float high_frequency_lrg() const { return _high_frequency_lrg; }
379
380 #ifndef PRODUCT
381 bool trace_spilling() const { return _trace_spilling; }
382 #endif
383
384 private:
385 // De-SSA the world. Assign registers to Nodes. Use the same register for
386 // all inputs to a PhiNode, effectively coalescing live ranges. Insert
387 // copies as needed.
388 void de_ssa();
389 uint Find_compress( const Node *n );
390 uint Find( uint lidx ) {
391 uint uf_lidx = _uf_map[lidx];
392 return (uf_lidx == lidx) ? uf_lidx : Find_compress(lidx);
393 }
394 uint Find_compress( uint lidx );
395
396 uint Find_id( const Node *n ) {
397 uint retval = n2lidx(n);
398 assert(retval == Find(n),"Invalid node to lidx mapping");
399 return retval;
400 }
401
402 // Add edge between reg and everything in the vector.
403 // Same as _ifg->add_vector(reg,live) EXCEPT use the RegMask
404 // information to trim the set of interferences. Return the
405 // count of edges added.
406 void interfere_with_live( uint reg, IndexSet *live );
407 // Count register pressure for asserts
408 uint count_int_pressure( IndexSet *liveout );
409 uint count_float_pressure( IndexSet *liveout );
410
411 // Build the interference graph using virtual registers only.
412 // Used for aggressive coalescing.
413 void build_ifg_virtual( );
414
415 // Build the interference graph using physical registers when available.
416 // That is, if 2 live ranges are simultaneously alive but in their
417 // acceptable register sets do not overlap, then they do not interfere.
418 uint build_ifg_physical( ResourceArea *a );
419
420 // Gather LiveRanGe information, including register masks and base pointer/
421 // derived pointer relationships.
422 void gather_lrg_masks( bool mod_cisc_masks );
423
424 // Force the bases of derived pointers to be alive at GC points.
425 bool stretch_base_pointer_live_ranges( ResourceArea *a );
426 // Helper to stretch above; recursively discover the base Node for
427 // a given derived Node. Easy for AddP-related machine nodes, but
428 // needs to be recursive for derived Phis.
429 Node *find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg );
430
431 // Set the was-lo-degree bit. Conservative coalescing should not change the
432 // colorability of the graph. If any live range was of low-degree before
433 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
434 void set_was_low();
435
436 // Split live-ranges that must spill due to register conflicts (as opposed
437 // to capacity spills). Typically these are things def'd in a register
438 // and used on the stack or vice-versa.
439 void pre_spill();
440
441 // Init LRG caching of degree, numregs. Init lo_degree list.
442 void cache_lrg_info( );
443
444 // Simplify the IFG by removing LRGs of low degree with no copies
445 void Pre_Simplify();
446
447 // Simplify the IFG by removing LRGs of low degree
448 void Simplify();
449
450 // Select colors by re-inserting edges into the IFG.
451 // Return TRUE if any spills occurred.
452 uint Select( );
453 // Helper function for select which allows biased coloring
454 OptoReg::Name choose_color( LRG &lrg, int chunk );
455 // Helper function which implements biasing heuristic
456 OptoReg::Name bias_color( LRG &lrg, int chunk );
457
458 // Split uncolorable live ranges
459 // Return new number of live ranges
460 uint Split( uint maxlrg );
461
462 // Copy 'was_spilled'-edness from one Node to another.
463 void copy_was_spilled( Node *src, Node *dst );
464 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
465 void set_was_spilled( Node *n );
466
467 // Convert ideal spill-nodes into machine loads & stores
468 // Set C->failing when fixup spills could not complete, node limit exceeded.
469 void fixup_spills();
470
471 // Post-Allocation peephole copy removal
472 void post_allocate_copy_removal();
473 Node *skip_copies( Node *c );
474 // Replace the old node with the current live version of that value
475 // and yank the old value if it's dead.
476 int replace_and_yank_if_dead( Node *old, OptoReg::Name nreg,
477 Block *current_block, Node_List& value, Node_List& regnd ) {
478 Node* v = regnd[nreg];
479 assert(v->outcnt() != 0, "no dead values");
480 old->replace_by(v);
481 return yank_if_dead(old, current_block, &value, ®nd);
482 }
483
484 int yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd );
485 int elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs );
486 int use_prior_register( Node *copy, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd );
487 bool may_be_copy_of_callee( Node *def ) const;
488
489 // If nreg already contains the same constant as val then eliminate it
490 bool eliminate_copy_of_constant(Node* val, Node* n,
491 Block *current_block, Node_List& value, Node_List ®nd,
492 OptoReg::Name nreg, OptoReg::Name nreg2);
493 // Extend the node to LRG mapping
494 void add_reference( const Node *node, const Node *old_node);
495
496 private:
497
498 static int _final_loads, _final_stores, _final_copies, _final_memoves;
499 static double _final_load_cost, _final_store_cost, _final_copy_cost, _final_memove_cost;
500 static int _conserv_coalesce, _conserv_coalesce_pair;
501 static int _conserv_coalesce_trie, _conserv_coalesce_quad;
502 static int _post_alloc;
503 static int _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce;
504 static int _used_cisc_instructions, _unused_cisc_instructions;
505 static int _allocator_attempts, _allocator_successes;
506
507 #ifndef PRODUCT
508 static uint _high_pressure, _low_pressure;
509
510 void dump() const;
511 void dump( const Node *n ) const;
512 void dump( const Block * b ) const;
513 void dump_degree_lists() const;
514 void dump_simplified() const;
515 void dump_lrg( uint lidx ) const;
516 void dump_bb( uint pre_order ) const;
517
518 // Verify that base pointers and derived pointers are still sane
519 void verify_base_ptrs( ResourceArea *a ) const;
520
521 void verify( ResourceArea *a, bool verify_ifg = false ) const;
522
523 void dump_for_spill_split_recycle() const;
524
525 public:
526 void dump_frame() const;
527 char *dump_register( const Node *n, char *buf ) const;
528 private:
529 static void print_chaitin_statistics();
530 #endif
531 friend class PhaseCoalesce;
532 friend class PhaseAggressiveCoalesce;
533 friend class PhaseConservativeCoalesce;
534 };
535
536 #endif // SHARE_VM_OPTO_CHAITIN_HPP