1 | /* |
2 | * Copyright (c) 1997, 2019, 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_OPTO_CHAITIN_HPP |
26 | #define SHARE_OPTO_CHAITIN_HPP |
27 | |
28 | #include "code/vmreg.hpp" |
29 | #include "memory/resourceArea.hpp" |
30 | #include "opto/connode.hpp" |
31 | #include "opto/live.hpp" |
32 | #include "opto/matcher.hpp" |
33 | #include "opto/phase.hpp" |
34 | #include "opto/regalloc.hpp" |
35 | #include "opto/regmask.hpp" |
36 | #include "opto/machnode.hpp" |
37 | |
38 | class Matcher; |
39 | class PhaseCFG; |
40 | class PhaseLive; |
41 | class PhaseRegAlloc; |
42 | class PhaseChaitin; |
43 | |
44 | #define OPTO_DEBUG_SPLIT_FREQ BLOCK_FREQUENCY(0.001) |
45 | #define OPTO_LRG_HIGH_FREQ BLOCK_FREQUENCY(0.25) |
46 | |
47 | //------------------------------LRG-------------------------------------------- |
48 | // Live-RanGe structure. |
49 | class LRG : public ResourceObj { |
50 | friend class VMStructs; |
51 | public: |
52 | static const uint AllStack_size = 0xFFFFF; // This mask size is used to tell that the mask of this LRG supports stack positions |
53 | enum { SPILL_REG=29999 }; // Register number of a spilled LRG |
54 | |
55 | double _cost; // 2 for loads/1 for stores times block freq |
56 | double _area; // Sum of all simultaneously live values |
57 | double score() const; // Compute score from cost and area |
58 | double _maxfreq; // Maximum frequency of any def or use |
59 | |
60 | Node *_def; // Check for multi-def live ranges |
61 | #ifndef PRODUCT |
62 | GrowableArray<Node*>* _defs; |
63 | #endif |
64 | |
65 | uint _risk_bias; // Index of LRG which we want to avoid color |
66 | uint _copy_bias; // Index of LRG which we want to share color |
67 | |
68 | uint _next; // Index of next LRG in linked list |
69 | uint _prev; // Index of prev LRG in linked list |
70 | private: |
71 | uint _reg; // Chosen register; undefined if mask is plural |
72 | public: |
73 | // Return chosen register for this LRG. Error if the LRG is not bound to |
74 | // a single register. |
75 | OptoReg::Name reg() const { return OptoReg::Name(_reg); } |
76 | void set_reg( OptoReg::Name r ) { _reg = r; } |
77 | |
78 | private: |
79 | uint _eff_degree; // Effective degree: Sum of neighbors _num_regs |
80 | public: |
81 | int degree() const { assert( _degree_valid , "" ); return _eff_degree; } |
82 | // Degree starts not valid and any change to the IFG neighbor |
83 | // set makes it not valid. |
84 | void set_degree( uint degree ) { |
85 | _eff_degree = degree; |
86 | debug_only(_degree_valid = 1;) |
87 | assert(!_mask.is_AllStack() || (_mask.is_AllStack() && lo_degree()), "_eff_degree can't be bigger than AllStack_size - _num_regs if the mask supports stack registers" ); |
88 | } |
89 | // Made a change that hammered degree |
90 | void invalid_degree() { debug_only(_degree_valid=0;) } |
91 | // Incrementally modify degree. If it was correct, it should remain correct |
92 | void inc_degree( uint mod ) { |
93 | _eff_degree += mod; |
94 | assert(!_mask.is_AllStack() || (_mask.is_AllStack() && lo_degree()), "_eff_degree can't be bigger than AllStack_size - _num_regs if the mask supports stack registers" ); |
95 | } |
96 | // Compute the degree between 2 live ranges |
97 | int compute_degree( LRG &l ) const; |
98 | bool mask_is_nonempty_and_up() const { |
99 | return mask().is_UP() && mask_size(); |
100 | } |
101 | bool is_float_or_vector() const { |
102 | return _is_float || _is_vector; |
103 | } |
104 | |
105 | private: |
106 | RegMask _mask; // Allowed registers for this LRG |
107 | uint _mask_size; // cache of _mask.Size(); |
108 | public: |
109 | int compute_mask_size() const { return _mask.is_AllStack() ? AllStack_size : _mask.Size(); } |
110 | void set_mask_size( int size ) { |
111 | assert((size == (int)AllStack_size) || (size == (int)_mask.Size()), "" ); |
112 | _mask_size = size; |
113 | #ifdef ASSERT |
114 | _msize_valid=1; |
115 | if (_is_vector) { |
116 | assert(!_fat_proj, "sanity" ); |
117 | assert(_mask.is_aligned_sets(_num_regs), "mask is not aligned, adjacent sets" ); |
118 | } else if (_num_regs == 2 && !_fat_proj) { |
119 | assert(_mask.is_aligned_pairs(), "mask is not aligned, adjacent pairs" ); |
120 | } |
121 | #endif |
122 | } |
123 | void compute_set_mask_size() { set_mask_size(compute_mask_size()); } |
124 | int mask_size() const { assert( _msize_valid, "mask size not valid" ); |
125 | return _mask_size; } |
126 | // Get the last mask size computed, even if it does not match the |
127 | // count of bits in the current mask. |
128 | int get_invalid_mask_size() const { return _mask_size; } |
129 | const RegMask &mask() const { return _mask; } |
130 | void set_mask( const RegMask &rm ) { _mask = rm; debug_only(_msize_valid=0;)} |
131 | void AND( const RegMask &rm ) { _mask.AND(rm); debug_only(_msize_valid=0;)} |
132 | void SUBTRACT( const RegMask &rm ) { _mask.SUBTRACT(rm); debug_only(_msize_valid=0;)} |
133 | void Clear() { _mask.Clear() ; debug_only(_msize_valid=1); _mask_size = 0; } |
134 | void Set_All() { _mask.Set_All(); debug_only(_msize_valid=1); _mask_size = RegMask::CHUNK_SIZE; } |
135 | |
136 | void Insert( OptoReg::Name reg ) { _mask.Insert(reg); debug_only(_msize_valid=0;) } |
137 | void Remove( OptoReg::Name reg ) { _mask.Remove(reg); debug_only(_msize_valid=0;) } |
138 | void clear_to_sets() { _mask.clear_to_sets(_num_regs); debug_only(_msize_valid=0;) } |
139 | |
140 | // Number of registers this live range uses when it colors |
141 | private: |
142 | uint16_t _num_regs; // 2 for Longs and Doubles, 1 for all else |
143 | // except _num_regs is kill count for fat_proj |
144 | public: |
145 | int num_regs() const { return _num_regs; } |
146 | void set_num_regs( int reg ) { assert( _num_regs == reg || !_num_regs, "" ); _num_regs = reg; } |
147 | |
148 | private: |
149 | // Number of physical registers this live range uses when it colors |
150 | // Architecture and register-set dependent |
151 | uint16_t _reg_pressure; |
152 | public: |
153 | void set_reg_pressure(int i) { _reg_pressure = i; } |
154 | int reg_pressure() const { return _reg_pressure; } |
155 | |
156 | // How much 'wiggle room' does this live range have? |
157 | // How many color choices can it make (scaled by _num_regs)? |
158 | int degrees_of_freedom() const { return mask_size() - _num_regs; } |
159 | // Bound LRGs have ZERO degrees of freedom. We also count |
160 | // must_spill as bound. |
161 | bool is_bound () const { return _is_bound; } |
162 | // Negative degrees-of-freedom; even with no neighbors this |
163 | // live range must spill. |
164 | bool not_free() const { return degrees_of_freedom() < 0; } |
165 | // Is this live range of "low-degree"? Trivially colorable? |
166 | bool lo_degree () const { return degree() <= degrees_of_freedom(); } |
167 | // Is this live range just barely "low-degree"? Trivially colorable? |
168 | bool just_lo_degree () const { return degree() == degrees_of_freedom(); } |
169 | |
170 | uint _is_oop:1, // Live-range holds an oop |
171 | _is_float:1, // True if in float registers |
172 | _is_vector:1, // True if in vector registers |
173 | _was_spilled1:1, // True if prior spilling on def |
174 | _was_spilled2:1, // True if twice prior spilling on def |
175 | _is_bound:1, // live range starts life with no |
176 | // degrees of freedom. |
177 | _direct_conflict:1, // True if def and use registers in conflict |
178 | _must_spill:1, // live range has lost all degrees of freedom |
179 | // If _fat_proj is set, live range does NOT require aligned, adjacent |
180 | // registers and has NO interferences. |
181 | // If _fat_proj is clear, live range requires num_regs() to be a power of |
182 | // 2, and it requires registers to form an aligned, adjacent set. |
183 | _fat_proj:1, // |
184 | _was_lo:1, // Was lo-degree prior to coalesce |
185 | _msize_valid:1, // _mask_size cache valid |
186 | _degree_valid:1, // _degree cache valid |
187 | _has_copy:1, // Adjacent to some copy instruction |
188 | _at_risk:1; // Simplify says this guy is at risk to spill |
189 | |
190 | |
191 | // Alive if non-zero, dead if zero |
192 | bool alive() const { return _def != NULL; } |
193 | bool is_multidef() const { return _def == NodeSentinel; } |
194 | bool is_singledef() const { return _def != NodeSentinel; } |
195 | |
196 | #ifndef PRODUCT |
197 | void dump( ) const; |
198 | #endif |
199 | }; |
200 | |
201 | //------------------------------IFG-------------------------------------------- |
202 | // InterFerence Graph |
203 | // An undirected graph implementation. Created with a fixed number of |
204 | // vertices. Edges can be added & tested. Vertices can be removed, then |
205 | // added back later with all edges intact. Can add edges between one vertex |
206 | // and a list of other vertices. Can union vertices (and their edges) |
207 | // together. The IFG needs to be really really fast, and also fairly |
208 | // abstract! It needs abstraction so I can fiddle with the implementation to |
209 | // get even more speed. |
210 | class PhaseIFG : public Phase { |
211 | friend class VMStructs; |
212 | // Current implementation: a triangular adjacency list. |
213 | |
214 | // Array of adjacency-lists, indexed by live-range number |
215 | IndexSet *_adjs; |
216 | |
217 | // Assertion bit for proper use of Squaring |
218 | bool _is_square; |
219 | |
220 | // Live range structure goes here |
221 | LRG *_lrgs; // Array of LRG structures |
222 | |
223 | public: |
224 | // Largest live-range number |
225 | uint _maxlrg; |
226 | |
227 | Arena *_arena; |
228 | |
229 | // Keep track of inserted and deleted Nodes |
230 | VectorSet *_yanked; |
231 | |
232 | PhaseIFG( Arena *arena ); |
233 | void init( uint maxlrg ); |
234 | |
235 | // Add edge between a and b. Returns true if actually addded. |
236 | int add_edge( uint a, uint b ); |
237 | |
238 | // Test for edge existance |
239 | int test_edge( uint a, uint b ) const; |
240 | |
241 | // Square-up matrix for faster Union |
242 | void SquareUp(); |
243 | |
244 | // Return number of LRG neighbors |
245 | uint neighbor_cnt( uint a ) const { return _adjs[a].count(); } |
246 | // Union edges of b into a on Squared-up matrix |
247 | void Union( uint a, uint b ); |
248 | // Test for edge in Squared-up matrix |
249 | int test_edge_sq( uint a, uint b ) const; |
250 | // Yank a Node and all connected edges from the IFG. Be prepared to |
251 | // re-insert the yanked Node in reverse order of yanking. Return a |
252 | // list of neighbors (edges) yanked. |
253 | IndexSet *remove_node( uint a ); |
254 | // Reinsert a yanked Node |
255 | void re_insert( uint a ); |
256 | // Return set of neighbors |
257 | IndexSet *neighbors( uint a ) const { return &_adjs[a]; } |
258 | |
259 | #ifndef PRODUCT |
260 | // Dump the IFG |
261 | void dump() const; |
262 | void stats() const; |
263 | void verify( const PhaseChaitin * ) const; |
264 | #endif |
265 | |
266 | //--------------- Live Range Accessors |
267 | LRG &lrgs(uint idx) const { assert(idx < _maxlrg, "oob" ); return _lrgs[idx]; } |
268 | |
269 | // Compute and set effective degree. Might be folded into SquareUp(). |
270 | void Compute_Effective_Degree(); |
271 | |
272 | // Compute effective degree as the sum of neighbors' _sizes. |
273 | int effective_degree( uint lidx ) const; |
274 | }; |
275 | |
276 | // The LiveRangeMap class is responsible for storing node to live range id mapping. |
277 | // Each node is mapped to a live range id (a virtual register). Nodes that are |
278 | // not considered for register allocation are given live range id 0. |
279 | class LiveRangeMap { |
280 | |
281 | private: |
282 | |
283 | uint _max_lrg_id; |
284 | |
285 | // Union-find map. Declared as a short for speed. |
286 | // Indexed by live-range number, it returns the compacted live-range number |
287 | LRG_List _uf_map; |
288 | |
289 | // Map from Nodes to live ranges |
290 | LRG_List _names; |
291 | |
292 | // Straight out of Tarjan's union-find algorithm |
293 | uint find_compress(const Node *node) { |
294 | uint lrg_id = find_compress(_names.at(node->_idx)); |
295 | _names.at_put(node->_idx, lrg_id); |
296 | return lrg_id; |
297 | } |
298 | |
299 | uint find_compress(uint lrg); |
300 | |
301 | public: |
302 | |
303 | const LRG_List& names() { |
304 | return _names; |
305 | } |
306 | |
307 | uint max_lrg_id() const { |
308 | return _max_lrg_id; |
309 | } |
310 | |
311 | void set_max_lrg_id(uint max_lrg_id) { |
312 | _max_lrg_id = max_lrg_id; |
313 | } |
314 | |
315 | uint size() const { |
316 | return _names.length(); |
317 | } |
318 | |
319 | uint live_range_id(uint idx) const { |
320 | return _names.at(idx); |
321 | } |
322 | |
323 | uint live_range_id(const Node *node) const { |
324 | return _names.at(node->_idx); |
325 | } |
326 | |
327 | uint uf_live_range_id(uint lrg_id) const { |
328 | return _uf_map.at(lrg_id); |
329 | } |
330 | |
331 | void map(uint idx, uint lrg_id) { |
332 | _names.at_put(idx, lrg_id); |
333 | } |
334 | |
335 | void uf_map(uint dst_lrg_id, uint src_lrg_id) { |
336 | _uf_map.at_put(dst_lrg_id, src_lrg_id); |
337 | } |
338 | |
339 | void extend(uint idx, uint lrg_id) { |
340 | _names.at_put_grow(idx, lrg_id); |
341 | } |
342 | |
343 | void uf_extend(uint dst_lrg_id, uint src_lrg_id) { |
344 | _uf_map.at_put_grow(dst_lrg_id, src_lrg_id); |
345 | } |
346 | |
347 | LiveRangeMap(Arena* arena, uint unique) |
348 | : _max_lrg_id(0) |
349 | , _uf_map(arena, unique, unique, 0) |
350 | , _names(arena, unique, unique, 0) {} |
351 | |
352 | uint find_id( const Node *n ) { |
353 | uint retval = live_range_id(n); |
354 | assert(retval == find(n),"Invalid node to lidx mapping" ); |
355 | return retval; |
356 | } |
357 | |
358 | // Reset the Union-Find map to identity |
359 | void reset_uf_map(uint max_lrg_id); |
360 | |
361 | // Make all Nodes map directly to their final live range; no need for |
362 | // the Union-Find mapping after this call. |
363 | void compress_uf_map_for_nodes(); |
364 | |
365 | uint find(uint lidx) { |
366 | uint uf_lidx = _uf_map.at(lidx); |
367 | return (uf_lidx == lidx) ? uf_lidx : find_compress(lidx); |
368 | } |
369 | |
370 | // Convert a Node into a Live Range Index - a lidx |
371 | uint find(const Node *node) { |
372 | uint lidx = live_range_id(node); |
373 | uint uf_lidx = _uf_map.at(lidx); |
374 | return (uf_lidx == lidx) ? uf_lidx : find_compress(node); |
375 | } |
376 | |
377 | // Like Find above, but no path compress, so bad asymptotic behavior |
378 | uint find_const(uint lrg) const; |
379 | |
380 | // Like Find above, but no path compress, so bad asymptotic behavior |
381 | uint find_const(const Node *node) const { |
382 | if(node->_idx >= (uint)_names.length()) { |
383 | return 0; // not mapped, usual for debug dump |
384 | } |
385 | return find_const(_names.at(node->_idx)); |
386 | } |
387 | }; |
388 | |
389 | //------------------------------Chaitin---------------------------------------- |
390 | // Briggs-Chaitin style allocation, mostly. |
391 | class PhaseChaitin : public PhaseRegAlloc { |
392 | friend class VMStructs; |
393 | |
394 | int _trip_cnt; |
395 | int _alternate; |
396 | |
397 | PhaseLive *_live; // Liveness, used in the interference graph |
398 | PhaseIFG *_ifg; // Interference graph (for original chunk) |
399 | VectorSet _spilled_once; // Nodes that have been spilled |
400 | VectorSet _spilled_twice; // Nodes that have been spilled twice |
401 | |
402 | // Combine the Live Range Indices for these 2 Nodes into a single live |
403 | // range. Future requests for any Node in either live range will |
404 | // return the live range index for the combined live range. |
405 | void Union( const Node *src, const Node *dst ); |
406 | |
407 | void new_lrg( const Node *x, uint lrg ); |
408 | |
409 | // Compact live ranges, removing unused ones. Return new maxlrg. |
410 | void compact(); |
411 | |
412 | uint _lo_degree; // Head of lo-degree LRGs list |
413 | uint _lo_stk_degree; // Head of lo-stk-degree LRGs list |
414 | uint _hi_degree; // Head of hi-degree LRGs list |
415 | uint _simplified; // Linked list head of simplified LRGs |
416 | |
417 | // Helper functions for Split() |
418 | uint split_DEF(Node *def, Block *b, int loc, uint max, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ); |
419 | uint split_USE(MachSpillCopyNode::SpillType spill_type, Node *def, Block *b, Node *use, uint useidx, uint max, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx ); |
420 | |
421 | //------------------------------clone_projs------------------------------------ |
422 | // After cloning some rematerialized instruction, clone any MachProj's that |
423 | // follow it. Example: Intel zero is XOR, kills flags. Sparc FP constants |
424 | // use G3 as an address temp. |
425 | int clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id); |
426 | |
427 | int clone_projs(Block* b, uint idx, Node* orig, Node* copy, LiveRangeMap& lrg_map) { |
428 | uint max_lrg_id = lrg_map.max_lrg_id(); |
429 | int found_projs = clone_projs(b, idx, orig, copy, max_lrg_id); |
430 | if (found_projs > 0) { |
431 | // max_lrg_id is updated during call above |
432 | lrg_map.set_max_lrg_id(max_lrg_id); |
433 | } |
434 | return found_projs; |
435 | } |
436 | |
437 | Node *split_Rematerialize(Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits, |
438 | int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru); |
439 | // True if lidx is used before any real register is def'd in the block |
440 | bool prompt_use( Block *b, uint lidx ); |
441 | Node *get_spillcopy_wide(MachSpillCopyNode::SpillType spill_type, Node *def, Node *use, uint uidx ); |
442 | // Insert the spill at chosen location. Skip over any intervening Proj's or |
443 | // Phis. Skip over a CatchNode and projs, inserting in the fall-through block |
444 | // instead. Update high-pressure indices. Create a new live range. |
445 | void insert_proj( Block *b, uint i, Node *spill, uint maxlrg ); |
446 | |
447 | bool is_high_pressure( Block *b, LRG *lrg, uint insidx ); |
448 | |
449 | uint _oldphi; // Node index which separates pre-allocation nodes |
450 | |
451 | Block **_blks; // Array of blocks sorted by frequency for coalescing |
452 | |
453 | float _high_frequency_lrg; // Frequency at which LRG will be spilled for debug info |
454 | |
455 | #ifndef PRODUCT |
456 | bool _trace_spilling; |
457 | #endif |
458 | |
459 | public: |
460 | PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool track_liveout_pressure); |
461 | ~PhaseChaitin() {} |
462 | |
463 | LiveRangeMap _lrg_map; |
464 | |
465 | LRG &lrgs(uint idx) const { return _ifg->lrgs(idx); } |
466 | |
467 | // Do all the real work of allocate |
468 | void Register_Allocate(); |
469 | |
470 | float high_frequency_lrg() const { return _high_frequency_lrg; } |
471 | |
472 | // Used when scheduling info generated, not in general register allocation |
473 | bool _scheduling_info_generated; |
474 | |
475 | void set_ifg(PhaseIFG &ifg) { _ifg = &ifg; } |
476 | void set_live(PhaseLive &live) { _live = &live; } |
477 | PhaseLive* get_live() { return _live; } |
478 | |
479 | // Populate the live range maps with ssa info for scheduling |
480 | void mark_ssa(); |
481 | |
482 | #ifndef PRODUCT |
483 | bool trace_spilling() const { return _trace_spilling; } |
484 | #endif |
485 | |
486 | private: |
487 | // De-SSA the world. Assign registers to Nodes. Use the same register for |
488 | // all inputs to a PhiNode, effectively coalescing live ranges. Insert |
489 | // copies as needed. |
490 | void de_ssa(); |
491 | |
492 | // Add edge between reg and everything in the vector. |
493 | // Use the RegMask information to trim the set of interferences. Return the |
494 | // count of edges added. |
495 | void interfere_with_live(uint lid, IndexSet* liveout); |
496 | #ifdef ASSERT |
497 | // Count register pressure for asserts |
498 | uint count_int_pressure(IndexSet* liveout); |
499 | uint count_float_pressure(IndexSet* liveout); |
500 | #endif |
501 | |
502 | // Build the interference graph using virtual registers only. |
503 | // Used for aggressive coalescing. |
504 | void build_ifg_virtual( ); |
505 | |
506 | // used when computing the register pressure for each block in the CFG. This |
507 | // is done during IFG creation. |
508 | class Pressure { |
509 | // keeps track of the register pressure at the current |
510 | // instruction (used when stepping backwards in the block) |
511 | uint _current_pressure; |
512 | |
513 | // keeps track of the instruction index of the first low to high register pressure |
514 | // transition (starting from the top) in the block |
515 | // if high_pressure_index == 0 then the whole block is high pressure |
516 | // if high_pressure_index = b.end_idx() + 1 then the whole block is low pressure |
517 | uint _high_pressure_index; |
518 | |
519 | // stores the highest pressure we find |
520 | uint _final_pressure; |
521 | |
522 | // number of live ranges that constitute high register pressure |
523 | uint _high_pressure_limit; |
524 | |
525 | // initial pressure observed |
526 | uint _start_pressure; |
527 | |
528 | public: |
529 | |
530 | // lower the register pressure and look for a low to high pressure |
531 | // transition |
532 | void lower(LRG& lrg, uint& location) { |
533 | _current_pressure -= lrg.reg_pressure(); |
534 | if (_current_pressure == _high_pressure_limit) { |
535 | _high_pressure_index = location; |
536 | } |
537 | } |
538 | |
539 | // raise the pressure and store the pressure if it's the biggest |
540 | // pressure so far |
541 | void raise(LRG &lrg) { |
542 | _current_pressure += lrg.reg_pressure(); |
543 | if (_current_pressure > _final_pressure) { |
544 | _final_pressure = _current_pressure; |
545 | } |
546 | } |
547 | |
548 | void init(int limit) { |
549 | _current_pressure = 0; |
550 | _high_pressure_index = 0; |
551 | _final_pressure = 0; |
552 | _high_pressure_limit = limit; |
553 | _start_pressure = 0; |
554 | } |
555 | |
556 | uint high_pressure_index() const { |
557 | return _high_pressure_index; |
558 | } |
559 | |
560 | uint final_pressure() const { |
561 | return _final_pressure; |
562 | } |
563 | |
564 | uint start_pressure() const { |
565 | return _start_pressure; |
566 | } |
567 | |
568 | uint current_pressure() const { |
569 | return _current_pressure; |
570 | } |
571 | |
572 | uint high_pressure_limit() const { |
573 | return _high_pressure_limit; |
574 | } |
575 | |
576 | void lower_high_pressure_index() { |
577 | _high_pressure_index--; |
578 | } |
579 | |
580 | void set_high_pressure_index_to_block_start() { |
581 | _high_pressure_index = 0; |
582 | } |
583 | |
584 | void set_start_pressure(int value) { |
585 | _start_pressure = value; |
586 | _final_pressure = value; |
587 | } |
588 | |
589 | void set_current_pressure(int value) { |
590 | _current_pressure = value; |
591 | } |
592 | |
593 | void check_pressure_at_fatproj(uint fatproj_location, RegMask& fatproj_mask) { |
594 | // this pressure is only valid at this instruction, i.e. we don't need to lower |
595 | // the register pressure since the fat proj was never live before (going backwards) |
596 | uint new_pressure = current_pressure() + fatproj_mask.Size(); |
597 | if (new_pressure > final_pressure()) { |
598 | _final_pressure = new_pressure; |
599 | } |
600 | |
601 | // if we were at a low pressure and now and the fat proj is at high pressure, record the fat proj location |
602 | // as coming from a low to high (to low again) |
603 | if (current_pressure() <= high_pressure_limit() && new_pressure > high_pressure_limit()) { |
604 | _high_pressure_index = fatproj_location; |
605 | } |
606 | } |
607 | |
608 | Pressure(uint high_pressure_index, uint high_pressure_limit) |
609 | : _current_pressure(0) |
610 | , _high_pressure_index(high_pressure_index) |
611 | , _final_pressure(0) |
612 | , _high_pressure_limit(high_pressure_limit) |
613 | , _start_pressure(0) {} |
614 | }; |
615 | |
616 | void check_for_high_pressure_transition_at_fatproj(uint& block_reg_pressure, uint location, LRG& lrg, Pressure& pressure, const int op_regtype); |
617 | void add_input_to_liveout(Block* b, Node* n, IndexSet* liveout, double cost, Pressure& int_pressure, Pressure& float_pressure); |
618 | void compute_initial_block_pressure(Block* b, IndexSet* liveout, Pressure& int_pressure, Pressure& float_pressure, double cost); |
619 | bool remove_node_if_not_used(Block* b, uint location, Node* n, uint lid, IndexSet* liveout); |
620 | void assign_high_score_to_immediate_copies(Block* b, Node* n, LRG& lrg, uint next_inst, uint last_inst); |
621 | void remove_interference_from_copy(Block* b, uint location, uint lid_copy, IndexSet* liveout, double cost, Pressure& int_pressure, Pressure& float_pressure); |
622 | void remove_bound_register_from_interfering_live_ranges(LRG& lrg, IndexSet* liveout, uint& must_spill); |
623 | void check_for_high_pressure_block(Pressure& pressure); |
624 | void adjust_high_pressure_index(Block* b, uint& hrp_index, Pressure& pressure); |
625 | |
626 | // Build the interference graph using physical registers when available. |
627 | // That is, if 2 live ranges are simultaneously alive but in their |
628 | // acceptable register sets do not overlap, then they do not interfere. |
629 | uint build_ifg_physical( ResourceArea *a ); |
630 | |
631 | public: |
632 | // Gather LiveRanGe information, including register masks and base pointer/ |
633 | // derived pointer relationships. |
634 | void gather_lrg_masks( bool mod_cisc_masks ); |
635 | |
636 | // user visible pressure variables for scheduling |
637 | Pressure _sched_int_pressure; |
638 | Pressure _sched_float_pressure; |
639 | Pressure _scratch_int_pressure; |
640 | Pressure _scratch_float_pressure; |
641 | |
642 | // Pressure functions for user context |
643 | void lower_pressure(Block* b, uint location, LRG& lrg, IndexSet* liveout, Pressure& int_pressure, Pressure& float_pressure); |
644 | void raise_pressure(Block* b, LRG& lrg, Pressure& int_pressure, Pressure& float_pressure); |
645 | void compute_entry_block_pressure(Block* b); |
646 | void compute_exit_block_pressure(Block* b); |
647 | void print_pressure_info(Pressure& pressure, const char *str); |
648 | |
649 | private: |
650 | // Force the bases of derived pointers to be alive at GC points. |
651 | bool stretch_base_pointer_live_ranges( ResourceArea *a ); |
652 | // Helper to stretch above; recursively discover the base Node for |
653 | // a given derived Node. Easy for AddP-related machine nodes, but |
654 | // needs to be recursive for derived Phis. |
655 | Node *find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ); |
656 | |
657 | // Set the was-lo-degree bit. Conservative coalescing should not change the |
658 | // colorability of the graph. If any live range was of low-degree before |
659 | // coalescing, it should Simplify. This call sets the was-lo-degree bit. |
660 | void set_was_low(); |
661 | |
662 | // Init LRG caching of degree, numregs. Init lo_degree list. |
663 | void cache_lrg_info( ); |
664 | |
665 | // Simplify the IFG by removing LRGs of low degree |
666 | void Simplify(); |
667 | |
668 | // Select colors by re-inserting edges into the IFG. |
669 | // Return TRUE if any spills occurred. |
670 | uint Select( ); |
671 | // Helper function for select which allows biased coloring |
672 | OptoReg::Name choose_color( LRG &lrg, int chunk ); |
673 | // Helper function which implements biasing heuristic |
674 | OptoReg::Name bias_color( LRG &lrg, int chunk ); |
675 | |
676 | // Split uncolorable live ranges |
677 | // Return new number of live ranges |
678 | uint Split(uint maxlrg, ResourceArea* split_arena); |
679 | |
680 | // Set the 'spilled_once' or 'spilled_twice' flag on a node. |
681 | void set_was_spilled( Node *n ); |
682 | |
683 | // Convert ideal spill-nodes into machine loads & stores |
684 | // Set C->failing when fixup spills could not complete, node limit exceeded. |
685 | void fixup_spills(); |
686 | |
687 | // Post-Allocation peephole copy removal |
688 | void post_allocate_copy_removal(); |
689 | Node *skip_copies( Node *c ); |
690 | // Replace the old node with the current live version of that value |
691 | // and yank the old value if it's dead. |
692 | int replace_and_yank_if_dead( Node *old, OptoReg::Name nreg, |
693 | Block *current_block, Node_List& value, Node_List& regnd ) { |
694 | Node* v = regnd[nreg]; |
695 | assert(v->outcnt() != 0, "no dead values" ); |
696 | old->replace_by(v); |
697 | return yank_if_dead(old, current_block, &value, ®nd); |
698 | } |
699 | |
700 | int yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { |
701 | return yank_if_dead_recurse(old, old, current_block, value, regnd); |
702 | } |
703 | int yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block, |
704 | Node_List *value, Node_List *regnd); |
705 | int yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ); |
706 | int elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ); |
707 | int use_prior_register( Node *copy, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ); |
708 | bool may_be_copy_of_callee( Node *def ) const; |
709 | |
710 | // If nreg already contains the same constant as val then eliminate it |
711 | bool eliminate_copy_of_constant(Node* val, Node* n, |
712 | Block *current_block, Node_List& value, Node_List ®nd, |
713 | OptoReg::Name nreg, OptoReg::Name nreg2); |
714 | // Extend the node to LRG mapping |
715 | void add_reference( const Node *node, const Node *old_node); |
716 | |
717 | // Record the first use of a def in the block for a register. |
718 | class RegDefUse { |
719 | Node* _def; |
720 | Node* _first_use; |
721 | public: |
722 | RegDefUse() : _def(NULL), _first_use(NULL) { } |
723 | Node* def() const { return _def; } |
724 | Node* first_use() const { return _first_use; } |
725 | |
726 | void update(Node* def, Node* use) { |
727 | if (_def != def) { |
728 | _def = def; |
729 | _first_use = use; |
730 | } |
731 | } |
732 | void clear() { |
733 | _def = NULL; |
734 | _first_use = NULL; |
735 | } |
736 | }; |
737 | typedef GrowableArray<RegDefUse> RegToDefUseMap; |
738 | int possibly_merge_multidef(Node *n, uint k, Block *block, RegToDefUseMap& reg2defuse); |
739 | |
740 | // Merge nodes that are a part of a multidef lrg and produce the same value within a block. |
741 | void merge_multidefs(); |
742 | |
743 | private: |
744 | |
745 | static int _final_loads, _final_stores, _final_copies, _final_memoves; |
746 | static double _final_load_cost, _final_store_cost, _final_copy_cost, _final_memove_cost; |
747 | static int _conserv_coalesce, _conserv_coalesce_pair; |
748 | static int _conserv_coalesce_trie, _conserv_coalesce_quad; |
749 | static int _post_alloc; |
750 | static int _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce; |
751 | static int _used_cisc_instructions, _unused_cisc_instructions; |
752 | static int _allocator_attempts, _allocator_successes; |
753 | |
754 | #ifndef PRODUCT |
755 | static uint _high_pressure, _low_pressure; |
756 | |
757 | void dump() const; |
758 | void dump( const Node *n ) const; |
759 | void dump( const Block * b ) const; |
760 | void dump_degree_lists() const; |
761 | void dump_simplified() const; |
762 | void dump_lrg( uint lidx, bool defs_only) const; |
763 | void dump_lrg( uint lidx) const { |
764 | // dump defs and uses by default |
765 | dump_lrg(lidx, false); |
766 | } |
767 | void dump_bb( uint pre_order ) const; |
768 | |
769 | // Verify that base pointers and derived pointers are still sane |
770 | void verify_base_ptrs( ResourceArea *a ) const; |
771 | |
772 | void verify( ResourceArea *a, bool verify_ifg = false ) const; |
773 | |
774 | void dump_for_spill_split_recycle() const; |
775 | |
776 | public: |
777 | void dump_frame() const; |
778 | char *dump_register( const Node *n, char *buf ) const; |
779 | private: |
780 | static void print_chaitin_statistics(); |
781 | #endif |
782 | friend class PhaseCoalesce; |
783 | friend class PhaseAggressiveCoalesce; |
784 | friend class PhaseConservativeCoalesce; |
785 | }; |
786 | |
787 | #endif // SHARE_OPTO_CHAITIN_HPP |
788 | |