1 /*
2 * Copyright (c) 2000, 2025, 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 #include "compiler/compileLog.hpp"
26 #include "compiler/oopMap.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/chaitin.hpp"
34 #include "opto/coalesce.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/idealGraphPrinter.hpp"
37 #include "opto/indexSet.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/opcodes.hpp"
42 #include "opto/rootnode.hpp"
43 #include "utilities/align.hpp"
44
45 #ifndef PRODUCT
46 void LRG::dump() const {
47 ttyLocker ttyl;
48 tty->print("%d ",num_regs());
49 _mask.dump();
50 if( _msize_valid ) {
51 if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size);
52 else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size());
53 } else {
54 tty->print(", #?(%d) ",_mask.Size());
55 }
56
57 tty->print("EffDeg: ");
58 if( _degree_valid ) tty->print( "%d ", _eff_degree );
59 else tty->print("? ");
60
61 if( is_multidef() ) {
62 tty->print("MultiDef ");
63 if (_defs != nullptr) {
64 tty->print("(");
65 for (int i = 0; i < _defs->length(); i++) {
66 tty->print("N%d ", _defs->at(i)->_idx);
67 }
68 tty->print(") ");
69 }
70 }
71 else if( _def == nullptr ) tty->print("Dead ");
72 else tty->print("Def: N%d ",_def->_idx);
73
74 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score());
75 // Flags
76 if( _is_oop ) tty->print("Oop ");
77 if( _is_float ) tty->print("Float ");
78 if( _is_vector ) tty->print("Vector ");
79 if( _is_predicate ) tty->print("Predicate ");
80 if( _is_scalable ) tty->print("Scalable ");
81 if( _was_spilled1 ) tty->print("Spilled ");
82 if( _was_spilled2 ) tty->print("Spilled2 ");
83 if( _direct_conflict ) tty->print("Direct_conflict ");
84 if( _fat_proj ) tty->print("Fat ");
85 if( _was_lo ) tty->print("Lo ");
86 if( _has_copy ) tty->print("Copy ");
87 if( _at_risk ) tty->print("Risk ");
88
89 if( _must_spill ) tty->print("Must_spill ");
90 if( _is_bound ) tty->print("Bound ");
91 if( _msize_valid ) {
92 if( _degree_valid && lo_degree() ) tty->print("Trivial ");
93 }
94
95 tty->cr();
96 }
97 #endif
98
99 // Compute score from cost and area. Low score is best to spill.
100 static double raw_score( double cost, double area ) {
101 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5;
102 }
103
104 double LRG::score() const {
105 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost.
106 // Bigger area lowers score, encourages spilling this live range.
107 // Bigger cost raise score, prevents spilling this live range.
108 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer
109 // to turn a divide by a constant into a multiply by the reciprical).
110 double score = raw_score( _cost, _area);
111
112 // Account for area. Basically, LRGs covering large areas are better
113 // to spill because more other LRGs get freed up.
114 if( _area == 0.0 ) // No area? Then no progress to spill
115 return 1e35;
116
117 if( _was_spilled2 ) // If spilled once before, we are unlikely
118 return score + 1e30; // to make progress again.
119
120 if( _cost >= _area*3.0 ) // Tiny area relative to cost
121 return score + 1e17; // Probably no progress to spill
122
123 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost
124 return score + 1e10; // Likely no progress to spill
125
126 return score;
127 }
128
129 #define NUMBUCKS 3
130
131 // Straight out of Tarjan's union-find algorithm
132 uint LiveRangeMap::find_compress(uint lrg) {
133 uint cur = lrg;
134 uint next = _uf_map.at(cur);
135 while (next != cur) { // Scan chain of equivalences
136 assert( next < cur, "always union smaller");
137 cur = next; // until find a fixed-point
138 next = _uf_map.at(cur);
139 }
140
141 // Core of union-find algorithm: update chain of
142 // equivalences to be equal to the root.
143 while (lrg != next) {
144 uint tmp = _uf_map.at(lrg);
145 _uf_map.at_put(lrg, next);
146 lrg = tmp;
147 }
148 return lrg;
149 }
150
151 // Reset the Union-Find map to identity
152 void LiveRangeMap::reset_uf_map(uint max_lrg_id) {
153 _max_lrg_id= max_lrg_id;
154 // Force the Union-Find mapping to be at least this large
155 _uf_map.at_put_grow(_max_lrg_id, 0);
156 // Initialize it to be the ID mapping.
157 for (uint i = 0; i < _max_lrg_id; ++i) {
158 _uf_map.at_put(i, i);
159 }
160 }
161
162 // Make all Nodes map directly to their final live range; no need for
163 // the Union-Find mapping after this call.
164 void LiveRangeMap::compress_uf_map_for_nodes() {
165 // For all Nodes, compress mapping
166 uint unique = _names.length();
167 for (uint i = 0; i < unique; ++i) {
168 uint lrg = _names.at(i);
169 uint compressed_lrg = find(lrg);
170 if (lrg != compressed_lrg) {
171 _names.at_put(i, compressed_lrg);
172 }
173 }
174 }
175
176 // Like Find above, but no path compress, so bad asymptotic behavior
177 uint LiveRangeMap::find_const(uint lrg) const {
178 if (!lrg) {
179 return lrg; // Ignore the zero LRG
180 }
181
182 // Off the end? This happens during debugging dumps when you got
183 // brand new live ranges but have not told the allocator yet.
184 if (lrg >= _max_lrg_id) {
185 return lrg;
186 }
187
188 uint next = _uf_map.at(lrg);
189 while (next != lrg) { // Scan chain of equivalences
190 assert(next < lrg, "always union smaller");
191 lrg = next; // until find a fixed-point
192 next = _uf_map.at(lrg);
193 }
194 return next;
195 }
196
197 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated)
198 : PhaseRegAlloc(unique, cfg, matcher,
199 #ifndef PRODUCT
200 print_chaitin_statistics
201 #else
202 nullptr
203 #endif
204 )
205 , _live(nullptr)
206 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0)
207 , _oldphi(unique)
208 #ifndef PRODUCT
209 , _trace_spilling(C->directive()->TraceSpillingOption)
210 #endif
211 , _lrg_map(Thread::current()->resource_area(), unique)
212 , _scheduling_info_generated(scheduling_info_generated)
213 , _sched_int_pressure(0, Matcher::int_pressure_limit())
214 , _sched_float_pressure(0, Matcher::float_pressure_limit())
215 , _scratch_int_pressure(0, Matcher::int_pressure_limit())
216 , _scratch_float_pressure(0, Matcher::float_pressure_limit())
217 {
218 Compile::TracePhase tp(_t_ctorChaitin);
219
220 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency());
221
222 // Build a list of basic blocks, sorted by frequency
223 // Experiment with sorting strategies to speed compilation
224 uint nr_blocks = _cfg.number_of_blocks();
225 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket
226 Block **buckets[NUMBUCKS]; // Array of buckets
227 uint buckcnt[NUMBUCKS]; // Array of bucket counters
228 double buckval[NUMBUCKS]; // Array of bucket value cutoffs
229
230 // The space which our buckets point into.
231 Block** start = NEW_RESOURCE_ARRAY(Block *, nr_blocks*NUMBUCKS);
232
233 for (uint i = 0; i < NUMBUCKS; i++) {
234 buckets[i] = &start[i*nr_blocks];
235 buckcnt[i] = 0;
236 // Bump by three orders of magnitude each time
237 cutoff *= 0.001;
238 buckval[i] = cutoff;
239 }
240
241 // Sort blocks into buckets
242 for (uint i = 0; i < nr_blocks; i++) {
243 for (uint j = 0; j < NUMBUCKS; j++) {
244 double bval = buckval[j];
245 Block* blk = _cfg.get_block(i);
246 if (j == NUMBUCKS - 1 || blk->_freq > bval) {
247 uint cnt = buckcnt[j];
248 // Assign block to end of list for appropriate bucket
249 buckets[j][cnt] = blk;
250 buckcnt[j] = cnt+1;
251 break; // kick out of inner loop
252 }
253 }
254 }
255
256 // Squash the partially filled buckets together into the first one.
257 static_assert(NUMBUCKS >= 2, "must"); // If this isn't true then it'll mess up the squashing.
258 Block** offset = &buckets[0][buckcnt[0]];
259 for (int i = 1; i < NUMBUCKS; i++) {
260 ::memmove(offset, buckets[i], buckcnt[i]*sizeof(Block*));
261 offset += buckcnt[i];
262 }
263 assert((&buckets[0][0] + nr_blocks) == offset, "should be");
264
265 // Free the now unused memory
266 FREE_RESOURCE_ARRAY(Block*, buckets[1], (NUMBUCKS-1)*nr_blocks);
267 // Finally, point the _blks to our memory
268 _blks = buckets[0];
269
270 #ifdef ASSERT
271 uint blkcnt = 0;
272 for (uint i = 0; i < NUMBUCKS; i++) {
273 blkcnt += buckcnt[i];
274 }
275 assert(blkcnt == nr_blocks, "Block array not totally filled");
276 #endif
277 }
278
279 // union 2 sets together.
280 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) {
281 uint src = _lrg_map.find(src_n);
282 uint dst = _lrg_map.find(dst_n);
283 assert(src, "");
284 assert(dst, "");
285 assert(src < _lrg_map.max_lrg_id(), "oob");
286 assert(dst < _lrg_map.max_lrg_id(), "oob");
287 assert(src < dst, "always union smaller");
288 _lrg_map.uf_map(dst, src);
289 }
290
291 void PhaseChaitin::new_lrg(const Node *x, uint lrg) {
292 // Make the Node->LRG mapping
293 _lrg_map.extend(x->_idx,lrg);
294 // Make the Union-Find mapping an identity function
295 _lrg_map.uf_extend(lrg, lrg);
296 }
297
298
299 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) {
300 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections");
301 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); )
302 int found_projs = 0;
303 uint cnt = orig->outcnt();
304 for (uint i = 0; i < cnt; i++) {
305 Node* proj = orig->raw_out(i);
306 if (proj->is_MachProj()) {
307 assert(proj->outcnt() == 0, "only kill projections are expected here");
308 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections");
309 found_projs++;
310 // Copy kill projections after the cloned node
311 Node* kills = proj->clone();
312 kills->set_req(0, copy);
313 b->insert_node(kills, idx++);
314 _cfg.map_node_to_block(kills, b);
315 new_lrg(kills, max_lrg_id++);
316 }
317 }
318 return found_projs;
319 }
320
321 // Renumber the live ranges to compact them. Makes the IFG smaller.
322 void PhaseChaitin::compact() {
323 Compile::TracePhase tp(_t_chaitinCompact);
324
325 // Current the _uf_map contains a series of short chains which are headed
326 // by a self-cycle. All the chains run from big numbers to little numbers.
327 // The Find() call chases the chains & shortens them for the next Find call.
328 // We are going to change this structure slightly. Numbers above a moving
329 // wave 'i' are unchanged. Numbers below 'j' point directly to their
330 // compacted live range with no further chaining. There are no chains or
331 // cycles below 'i', so the Find call no longer works.
332 uint j=1;
333 uint i;
334 for (i = 1; i < _lrg_map.max_lrg_id(); i++) {
335 uint lr = _lrg_map.uf_live_range_id(i);
336 // Ignore unallocated live ranges
337 if (!lr) {
338 continue;
339 }
340 assert(lr <= i, "");
341 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr));
342 }
343 // Now change the Node->LR mapping to reflect the compacted names
344 uint unique = _lrg_map.size();
345 for (i = 0; i < unique; i++) {
346 uint lrg_id = _lrg_map.live_range_id(i);
347 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id));
348 }
349
350 // Reset the Union-Find mapping
351 _lrg_map.reset_uf_map(j);
352 }
353
354 void PhaseChaitin::Register_Allocate() {
355
356 // Above the OLD FP (and in registers) are the incoming arguments. Stack
357 // slots in this area are called "arg_slots". Above the NEW FP (and in
358 // registers) is the outgoing argument area; above that is the spill/temp
359 // area. These are all "frame_slots". Arg_slots start at the zero
360 // stack_slots and count up to the known arg_size. Frame_slots start at
361 // the stack_slot #arg_size and go up. After allocation I map stack
362 // slots to actual offsets. Stack-slots in the arg_slot area are biased
363 // by the frame_size; stack-slots in the frame_slot area are biased by 0.
364
365 _trip_cnt = 0;
366 _alternate = 0;
367 _matcher._allocation_started = true;
368
369 ResourceArea split_arena(mtCompiler, Arena::Tag::tag_regsplit); // Arena for Split local resources
370 ResourceArea live_arena(mtCompiler, Arena::Tag::tag_reglive); // Arena for liveness & IFG info
371 ResourceMark rm(&live_arena);
372
373 // Need live-ness for the IFG; need the IFG for coalescing. If the
374 // liveness is JUST for coalescing, then I can get some mileage by renaming
375 // all copy-related live ranges low and then using the max copy-related
376 // live range as a cut-off for LIVE and the IFG. In other words, I can
377 // build a subset of LIVE and IFG just for copies.
378 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false);
379
380 // Need IFG for coalescing and coloring
381 PhaseIFG ifg(&live_arena);
382 _ifg = &ifg;
383
384 // Come out of SSA world to the Named world. Assign (virtual) registers to
385 // Nodes. Use the same register for all inputs and the output of PhiNodes
386 // - effectively ending SSA form. This requires either coalescing live
387 // ranges or inserting copies. For the moment, we insert "virtual copies"
388 // - we pretend there is a copy prior to each Phi in predecessor blocks.
389 // We will attempt to coalesce such "virtual copies" before we manifest
390 // them for real.
391 de_ssa();
392
393 #ifdef ASSERT
394 // Verify the graph before RA.
395 verify(&live_arena);
396 #endif
397
398 {
399 Compile::TracePhase tp(_t_computeLive);
400 _live = nullptr; // Mark live as being not available
401 rm.reset_to_mark(); // Reclaim working storage
402 IndexSet::reset_memory(C, &live_arena);
403 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG
404 gather_lrg_masks( false ); // Collect LRG masks
405 live.compute(_lrg_map.max_lrg_id()); // Compute liveness
406 _live = &live; // Mark LIVE as being available
407 }
408
409 // Base pointers are currently "used" by instructions which define new
410 // derived pointers. This makes base pointers live up to the where the
411 // derived pointer is made, but not beyond. Really, they need to be live
412 // across any GC point where the derived value is live. So this code looks
413 // at all the GC points, and "stretches" the live range of any base pointer
414 // to the GC point.
415 if (stretch_base_pointer_live_ranges(&live_arena)) {
416 Compile::TracePhase tp("computeLive (sbplr)", _t_computeLive);
417 // Since some live range stretched, I need to recompute live
418 _live = nullptr;
419 rm.reset_to_mark(); // Reclaim working storage
420 IndexSet::reset_memory(C, &live_arena);
421 ifg.init(_lrg_map.max_lrg_id());
422 gather_lrg_masks(false);
423 live.compute(_lrg_map.max_lrg_id());
424 _live = &live;
425 }
426
427 C->print_method(PHASE_INITIAL_LIVENESS, 4);
428
429 // Create the interference graph using virtual copies
430 build_ifg_virtual(); // Include stack slots this time
431
432 // The IFG is/was triangular. I am 'squaring it up' so Union can run
433 // faster. Union requires a 'for all' operation which is slow on the
434 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' -
435 // meaning I can visit all the Nodes neighbors less than a Node in time
436 // O(# of neighbors), but I have to visit all the Nodes greater than a
437 // given Node and search them for an instance, i.e., time O(#MaxLRG)).
438 _ifg->SquareUp();
439
440 // Aggressive (but pessimistic) copy coalescing.
441 // This pass works on virtual copies. Any virtual copies which are not
442 // coalesced get manifested as actual copies
443 {
444 Compile::TracePhase tp(_t_chaitinCoalesce1);
445
446 PhaseAggressiveCoalesce coalesce(*this);
447 coalesce.coalesce_driver();
448 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do
449 // not match the Phi itself, insert a copy.
450 coalesce.insert_copies(_matcher);
451 if (C->failing()) {
452 return;
453 }
454 }
455
456 // After aggressive coalesce, attempt a first cut at coloring.
457 // To color, we need the IFG and for that we need LIVE.
458 {
459 Compile::TracePhase tp(_t_computeLive);
460 _live = nullptr;
461 rm.reset_to_mark(); // Reclaim working storage
462 IndexSet::reset_memory(C, &live_arena);
463 ifg.init(_lrg_map.max_lrg_id());
464 gather_lrg_masks( true );
465 live.compute(_lrg_map.max_lrg_id());
466 _live = &live;
467 }
468
469 C->print_method(PHASE_AGGRESSIVE_COALESCING, 4);
470
471 // Build physical interference graph
472 uint must_spill = 0;
473 must_spill = build_ifg_physical(&live_arena);
474 // If we have a guaranteed spill, might as well spill now
475 if (must_spill) {
476 if(!_lrg_map.max_lrg_id()) {
477 return;
478 }
479 // Bail out if unique gets too large (ie - unique > MaxNodeLimit)
480 C->check_node_count(10*must_spill, "out of nodes before split");
481 if (C->failing()) {
482 return;
483 }
484
485 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
486 if (C->failing()) {
487 return;
488 }
489 _lrg_map.set_max_lrg_id(new_max_lrg_id);
490 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
491 // or we failed to split
492 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split");
493 if (C->failing()) {
494 return;
495 }
496
497 NOT_PRODUCT(C->verify_graph_edges();)
498
499 compact(); // Compact LRGs; return new lower max lrg
500
501 {
502 Compile::TracePhase tp(_t_computeLive);
503 _live = nullptr;
504 rm.reset_to_mark(); // Reclaim working storage
505 IndexSet::reset_memory(C, &live_arena);
506 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph
507 gather_lrg_masks( true ); // Collect intersect mask
508 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE
509 _live = &live;
510 }
511
512 C->print_method(PHASE_INITIAL_SPILLING, 4);
513
514 build_ifg_physical(&live_arena);
515 _ifg->SquareUp();
516 _ifg->Compute_Effective_Degree();
517 // Only do conservative coalescing if requested
518 if (OptoCoalesce) {
519 Compile::TracePhase tp(_t_chaitinCoalesce2);
520 // Conservative (and pessimistic) copy coalescing of those spills
521 PhaseConservativeCoalesce coalesce(*this);
522 // If max live ranges greater than cutoff, don't color the stack.
523 // This cutoff can be larger than below since it is only done once.
524 coalesce.coalesce_driver();
525 }
526 _lrg_map.compress_uf_map_for_nodes();
527
528 if (OptoCoalesce) {
529 C->print_method(PHASE_CONSERVATIVE_COALESCING, 4);
530 }
531
532 #ifdef ASSERT
533 verify(&live_arena, true);
534 #endif
535 } else {
536 ifg.SquareUp();
537 ifg.Compute_Effective_Degree();
538 #ifdef ASSERT
539 set_was_low();
540 #endif
541 }
542
543 // Prepare for Simplify & Select
544 cache_lrg_info(); // Count degree of LRGs
545
546 // Simplify the InterFerence Graph by removing LRGs of low degree.
547 // LRGs of low degree are trivially colorable.
548 Simplify();
549
550 // Select colors by re-inserting LRGs back into the IFG in reverse order.
551 // Return whether or not something spills.
552 uint spills = Select( );
553
554 // If we spill, split and recycle the entire thing
555 while( spills ) {
556 if( _trip_cnt++ > 24 ) {
557 DEBUG_ONLY( dump_for_spill_split_recycle(); )
558 if( _trip_cnt > 27 ) {
559 C->record_method_not_compilable("failed spill-split-recycle sanity check");
560 return;
561 }
562 }
563
564 if (!_lrg_map.max_lrg_id()) {
565 return;
566 }
567 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
568 if (C->failing()) {
569 return;
570 }
571 _lrg_map.set_max_lrg_id(new_max_lrg_id);
572 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
573 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split");
574 if (C->failing()) {
575 return;
576 }
577
578 compact(); // Compact LRGs; return new lower max lrg
579
580 // Nuke the live-ness and interference graph and LiveRanGe info
581 {
582 Compile::TracePhase tp(_t_computeLive);
583 _live = nullptr;
584 rm.reset_to_mark(); // Reclaim working storage
585 IndexSet::reset_memory(C, &live_arena);
586 ifg.init(_lrg_map.max_lrg_id());
587
588 // Create LiveRanGe array.
589 // Intersect register masks for all USEs and DEFs
590 gather_lrg_masks(true);
591 live.compute(_lrg_map.max_lrg_id());
592 _live = &live;
593 }
594
595 C->print_method(PHASE_ITERATIVE_SPILLING, 4);
596
597 must_spill = build_ifg_physical(&live_arena);
598 _ifg->SquareUp();
599 _ifg->Compute_Effective_Degree();
600
601 // Only do conservative coalescing if requested
602 if (OptoCoalesce) {
603 Compile::TracePhase tp(_t_chaitinCoalesce3);
604 // Conservative (and pessimistic) copy coalescing
605 PhaseConservativeCoalesce coalesce(*this);
606 // Check for few live ranges determines how aggressive coalesce is.
607 coalesce.coalesce_driver();
608 }
609 _lrg_map.compress_uf_map_for_nodes();
610
611 if (OptoCoalesce) {
612 C->print_method(PHASE_CONSERVATIVE_COALESCING, 4);
613 }
614
615 #ifdef ASSERT
616 verify(&live_arena, true);
617 #endif
618 cache_lrg_info(); // Count degree of LRGs
619
620 // Simplify the InterFerence Graph by removing LRGs of low degree.
621 // LRGs of low degree are trivially colorable.
622 Simplify();
623
624 // Select colors by re-inserting LRGs back into the IFG in reverse order.
625 // Return whether or not something spills.
626 spills = Select();
627 }
628
629 C->print_method(PHASE_AFTER_ITERATIVE_SPILLING, 4);
630
631 // Count number of Simplify-Select trips per coloring success.
632 _allocator_attempts += _trip_cnt + 1;
633 _allocator_successes += 1;
634
635 // Peephole remove copies
636 post_allocate_copy_removal();
637
638 C->print_method(PHASE_POST_ALLOCATION_COPY_REMOVAL, 4);
639
640 // Merge multidefs if multiple defs representing the same value are used in a single block.
641 merge_multidefs();
642
643 C->print_method(PHASE_MERGE_MULTI_DEFS, 4);
644
645 #ifdef ASSERT
646 // Verify the graph after RA.
647 verify(&live_arena);
648 #endif
649
650 // max_reg is past the largest *register* used.
651 // Convert that to a frame_slot number.
652 if (_max_reg <= _matcher._new_SP) {
653 _framesize = C->out_preserve_stack_slots();
654 }
655 else {
656 _framesize = _max_reg -_matcher._new_SP;
657 }
658 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough");
659
660 // This frame must preserve the required fp alignment
661 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots());
662 assert(_framesize <= 1000000, "sanity check");
663 #ifndef PRODUCT
664 _total_framesize += _framesize;
665 if ((int)_framesize > _max_framesize) {
666 _max_framesize = _framesize;
667 }
668 #endif
669
670 // Convert CISC spills
671 fixup_spills();
672
673 C->print_method(PHASE_FIX_UP_SPILLS, 4);
674
675 // Log regalloc results
676 CompileLog* log = Compile::current()->log();
677 if (log != nullptr) {
678 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing());
679 }
680
681 if (C->failing()) {
682 return;
683 }
684
685 NOT_PRODUCT(C->verify_graph_edges();)
686
687 // Move important info out of the live_arena to longer lasting storage.
688 alloc_node_regs(_lrg_map.size());
689 for (uint i=0; i < _lrg_map.size(); i++) {
690 if (_lrg_map.live_range_id(i)) { // Live range associated with Node?
691 LRG &lrg = lrgs(_lrg_map.live_range_id(i));
692 if (!lrg.alive()) {
693 set_bad(i);
694 } else if ((lrg.num_regs() == 1 && !lrg.is_scalable()) ||
695 (lrg.is_scalable() && lrg.scalable_reg_slots() == 1)) {
696 set1(i, lrg.reg());
697 } else { // Must be a register-set
698 if (!lrg._fat_proj) { // Must be aligned adjacent register set
699 // Live ranges record the highest register in their mask.
700 // We want the low register for the AD file writer's convenience.
701 OptoReg::Name hi = lrg.reg(); // Get hi register
702 int num_regs = lrg.num_regs();
703 if (lrg.is_scalable() && OptoReg::is_stack(hi)) {
704 // For scalable vector registers, when they are allocated in physical
705 // registers, num_regs is RegMask::SlotsPerVecA for reg mask of scalable
706 // vector. If they are allocated on stack, we need to get the actual
707 // num_regs, which reflects the physical length of scalable registers.
708 num_regs = lrg.scalable_reg_slots();
709 }
710 if (num_regs == 1) {
711 set1(i, hi);
712 } else {
713 OptoReg::Name lo = OptoReg::add(hi, (1 - num_regs)); // Find lo
714 // We have to use pair [lo,lo+1] even for wide vectors/vmasks because
715 // the rest of code generation works only with pairs. It is safe
716 // since for registers encoding only 'lo' is used.
717 // Second reg from pair is used in ScheduleAndBundle with vector max
718 // size 8 which corresponds to registers pair.
719 // It is also used in BuildOopMaps but oop operations are not
720 // vectorized.
721 set2(i, lo);
722 }
723 } else { // Misaligned; extract 2 bits
724 OptoReg::Name hi = lrg.reg(); // Get hi register
725 lrg.Remove(hi); // Yank from mask
726 int lo = lrg.mask().find_first_elem(); // Find lo
727 set_pair(i, hi, lo);
728 }
729 }
730 if( lrg._is_oop ) _node_oops.set(i);
731 } else {
732 set_bad(i);
733 }
734 }
735
736 // Done!
737 _live = nullptr;
738 _ifg = nullptr;
739 C->set_indexSet_arena(nullptr); // ResourceArea is at end of scope
740 }
741
742 void PhaseChaitin::de_ssa() {
743 // Set initial Names for all Nodes. Most Nodes get the virtual register
744 // number. A few get the ZERO live range number. These do not
745 // get allocated, but instead rely on correct scheduling to ensure that
746 // only one instance is simultaneously live at a time.
747 uint lr_counter = 1;
748 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
749 Block* block = _cfg.get_block(i);
750 uint cnt = block->number_of_nodes();
751
752 // Handle all the normal Nodes in the block
753 for( uint j = 0; j < cnt; j++ ) {
754 Node *n = block->get_node(j);
755 // Pre-color to the zero live range, or pick virtual register
756 const RegMask &rm = n->out_RegMask();
757 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0);
758 }
759 }
760
761 // Reset the Union-Find mapping to be identity
762 _lrg_map.reset_uf_map(lr_counter);
763 }
764
765 void PhaseChaitin::mark_ssa() {
766 // Use ssa names to populate the live range maps or if no mask
767 // is available, use the 0 entry.
768 uint max_idx = 0;
769 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
770 Block* block = _cfg.get_block(i);
771 uint cnt = block->number_of_nodes();
772
773 // Handle all the normal Nodes in the block
774 for ( uint j = 0; j < cnt; j++ ) {
775 Node *n = block->get_node(j);
776 // Pre-color to the zero live range, or pick virtual register
777 const RegMask &rm = n->out_RegMask();
778 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0);
779 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx;
780 }
781 }
782 _lrg_map.set_max_lrg_id(max_idx+1);
783
784 // Reset the Union-Find mapping to be identity
785 _lrg_map.reset_uf_map(max_idx+1);
786 }
787
788
789 // Gather LiveRanGe information, including register masks. Modification of
790 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce.
791 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) {
792
793 // Nail down the frame pointer live range
794 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr));
795 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite
796
797 // For all blocks
798 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
799 Block* block = _cfg.get_block(i);
800
801 // For all instructions
802 for (uint j = 1; j < block->number_of_nodes(); j++) {
803 Node* n = block->get_node(j);
804 uint input_edge_start =1; // Skip control most nodes
805 bool is_machine_node = false;
806 if (n->is_Mach()) {
807 is_machine_node = true;
808 input_edge_start = n->as_Mach()->oper_input_base();
809 }
810 uint idx = n->is_Copy();
811
812 // Get virtual register number, same as LiveRanGe index
813 uint vreg = _lrg_map.live_range_id(n);
814 LRG& lrg = lrgs(vreg);
815 if (vreg) { // No vreg means un-allocable (e.g. memory)
816
817 // Check for float-vs-int live range (used in register-pressure
818 // calculations)
819 const Type *n_type = n->bottom_type();
820 if (n_type->is_floatingpoint()) {
821 lrg._is_float = 1;
822 }
823
824 // Check for twice prior spilling. Once prior spilling might have
825 // spilled 'soft', 2nd prior spill should have spilled 'hard' and
826 // further spilling is unlikely to make progress.
827 if (_spilled_once.test(n->_idx)) {
828 lrg._was_spilled1 = 1;
829 if (_spilled_twice.test(n->_idx)) {
830 lrg._was_spilled2 = 1;
831 }
832 }
833
834 #ifndef PRODUCT
835 // Collect bits not used by product code, but which may be useful for
836 // debugging.
837
838 // Collect has-copy bit
839 if (idx) {
840 lrg._has_copy = 1;
841 uint clidx = _lrg_map.live_range_id(n->in(idx));
842 LRG& copy_src = lrgs(clidx);
843 copy_src._has_copy = 1;
844 }
845
846 if (trace_spilling() && lrg._def != nullptr) {
847 // collect defs for MultiDef printing
848 if (lrg._defs == nullptr) {
849 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, nullptr);
850 lrg._defs->append(lrg._def);
851 }
852 lrg._defs->append(n);
853 }
854 #endif
855
856 // Check for a single def LRG; these can spill nicely
857 // via rematerialization. Flag as null for no def found
858 // yet, or 'n' for single def or -1 for many defs.
859 lrg._def = lrg._def ? NodeSentinel : n;
860
861 // Limit result register mask to acceptable registers
862 const RegMask &rm = n->out_RegMask();
863 lrg.AND( rm );
864
865 uint ireg = n->ideal_reg();
866 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP,
867 "oops must be in Op_RegP's" );
868
869 // Check for vector live range (only if vector register is used).
870 // On SPARC vector uses RegD which could be misaligned so it is not
871 // processes as vector in RA.
872 if (RegMask::is_vector(ireg)) {
873 lrg._is_vector = 1;
874 if (Matcher::implements_scalable_vector && ireg == Op_VecA) {
875 assert(Matcher::supports_scalable_vector(), "scalable vector should be supported");
876 lrg._is_scalable = 1;
877 // For scalable vector, when it is allocated in physical register,
878 // num_regs is RegMask::SlotsPerVecA for reg mask,
879 // which may not be the actual physical register size.
880 // If it is allocated in stack, we need to get the actual
881 // physical length of scalable vector register.
882 lrg.set_scalable_reg_slots(Matcher::scalable_vector_reg_size(T_FLOAT));
883 }
884 }
885
886 if (ireg == Op_RegVectMask) {
887 assert(Matcher::has_predicated_vectors(), "predicated vector should be supported");
888 lrg._is_predicate = 1;
889 if (Matcher::supports_scalable_vector()) {
890 lrg._is_scalable = 1;
891 // For scalable predicate, when it is allocated in physical register,
892 // num_regs is RegMask::SlotsPerRegVectMask for reg mask,
893 // which may not be the actual physical register size.
894 // If it is allocated in stack, we need to get the actual
895 // physical length of scalable predicate register.
896 lrg.set_scalable_reg_slots(Matcher::scalable_predicate_reg_slots());
897 }
898 }
899 assert(n_type->isa_vect() == nullptr || lrg._is_vector ||
900 ireg == Op_RegD || ireg == Op_RegL || ireg == Op_RegVectMask,
901 "vector must be in vector registers");
902
903 // Check for bound register masks
904 const RegMask &lrgmask = lrg.mask();
905 if (lrgmask.is_bound(ireg)) {
906 lrg._is_bound = 1;
907 }
908
909 // Check for maximum frequency value
910 if (lrg._maxfreq < block->_freq) {
911 lrg._maxfreq = block->_freq;
912 }
913
914 // Check for oop-iness, or long/double
915 // Check for multi-kill projection
916 switch (ireg) {
917 case MachProjNode::fat_proj:
918 // Fat projections have size equal to number of registers killed
919 lrg.set_num_regs(rm.Size());
920 lrg.set_reg_pressure(lrg.num_regs());
921 lrg._fat_proj = 1;
922 lrg._is_bound = 1;
923 break;
924 case Op_RegP:
925 #ifdef _LP64
926 lrg.set_num_regs(2); // Size is 2 stack words
927 #else
928 lrg.set_num_regs(1); // Size is 1 stack word
929 #endif
930 // Register pressure is tracked relative to the maximum values
931 // suggested for that platform, INTPRESSURE and FLOATPRESSURE,
932 // and relative to other types which compete for the same regs.
933 //
934 // The following table contains suggested values based on the
935 // architectures as defined in each .ad file.
936 // INTPRESSURE and FLOATPRESSURE may be tuned differently for
937 // compile-speed or performance.
938 // Note1:
939 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1
940 // since .ad registers are defined as high and low halves.
941 // These reg_pressure values remain compatible with the code
942 // in is_high_pressure() which relates get_invalid_mask_size(),
943 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE.
944 // Note2:
945 // SPARC -d32 has 24 registers available for integral values,
946 // but only 10 of these are safe for 64-bit longs.
947 // Using set_reg_pressure(2) for both int and long means
948 // the allocator will believe it can fit 26 longs into
949 // registers. Using 2 for longs and 1 for ints means the
950 // allocator will attempt to put 52 integers into registers.
951 // The settings below limit this problem to methods with
952 // many long values which are being run on 32-bit SPARC.
953 //
954 // ------------------- reg_pressure --------------------
955 // Each entry is reg_pressure_per_value,number_of_regs
956 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE
957 // IA32 2 1 1 1 1 6 6
958 // SPARC 2 2 2 2 2 48 (24) 52 (26)
959 // SPARCV9 2 2 2 2 2 48 (24) 52 (26)
960 // AMD64 1 1 1 1 1 14 15
961 // -----------------------------------------------------
962 lrg.set_reg_pressure(1); // normally one value per register
963 if( n_type->isa_oop_ptr() ) {
964 lrg._is_oop = 1;
965 }
966 break;
967 case Op_RegL: // Check for long or double
968 case Op_RegD:
969 lrg.set_num_regs(2);
970 // Define platform specific register pressure
971 #if defined(ARM32)
972 lrg.set_reg_pressure(2);
973 #elif defined(IA32)
974 if( ireg == Op_RegL ) {
975 lrg.set_reg_pressure(2);
976 } else {
977 lrg.set_reg_pressure(1);
978 }
979 #else
980 lrg.set_reg_pressure(1); // normally one value per register
981 #endif
982 // If this def of a double forces a mis-aligned double,
983 // flag as '_fat_proj' - really flag as allowing misalignment
984 // AND changes how we count interferences. A mis-aligned
985 // double can interfere with TWO aligned pairs, or effectively
986 // FOUR registers!
987 if (rm.is_misaligned_pair()) {
988 lrg._fat_proj = 1;
989 lrg._is_bound = 1;
990 }
991 break;
992 case Op_RegVectMask:
993 assert(Matcher::has_predicated_vectors(), "sanity");
994 assert(RegMask::num_registers(Op_RegVectMask) == RegMask::SlotsPerRegVectMask, "sanity");
995 lrg.set_num_regs(RegMask::SlotsPerRegVectMask);
996 lrg.set_reg_pressure(1);
997 break;
998 case Op_RegF:
999 case Op_RegI:
1000 case Op_RegN:
1001 case Op_RegFlags:
1002 case 0: // not an ideal register
1003 lrg.set_num_regs(1);
1004 lrg.set_reg_pressure(1);
1005 break;
1006 case Op_VecA:
1007 assert(Matcher::supports_scalable_vector(), "does not support scalable vector");
1008 assert(RegMask::num_registers(Op_VecA) == RegMask::SlotsPerVecA, "sanity");
1009 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecA), "vector should be aligned");
1010 lrg.set_num_regs(RegMask::SlotsPerVecA);
1011 lrg.set_reg_pressure(1);
1012 break;
1013 case Op_VecS:
1014 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity");
1015 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity");
1016 lrg.set_num_regs(RegMask::SlotsPerVecS);
1017 lrg.set_reg_pressure(1);
1018 break;
1019 case Op_VecD:
1020 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity");
1021 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity");
1022 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned");
1023 lrg.set_num_regs(RegMask::SlotsPerVecD);
1024 lrg.set_reg_pressure(1);
1025 break;
1026 case Op_VecX:
1027 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity");
1028 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity");
1029 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned");
1030 lrg.set_num_regs(RegMask::SlotsPerVecX);
1031 lrg.set_reg_pressure(1);
1032 break;
1033 case Op_VecY:
1034 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity");
1035 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity");
1036 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned");
1037 lrg.set_num_regs(RegMask::SlotsPerVecY);
1038 lrg.set_reg_pressure(1);
1039 break;
1040 case Op_VecZ:
1041 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity");
1042 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity");
1043 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned");
1044 lrg.set_num_regs(RegMask::SlotsPerVecZ);
1045 lrg.set_reg_pressure(1);
1046 break;
1047 default:
1048 ShouldNotReachHere();
1049 }
1050 }
1051
1052 // Now do the same for inputs
1053 uint cnt = n->req();
1054 // Setup for CISC SPILLING
1055 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable;
1056 if( UseCISCSpill && after_aggressive ) {
1057 inp = n->cisc_operand();
1058 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable )
1059 // Convert operand number to edge index number
1060 inp = n->as_Mach()->operand_index(inp);
1061 }
1062
1063 // Prepare register mask for each input
1064 for( uint k = input_edge_start; k < cnt; k++ ) {
1065 uint vreg = _lrg_map.live_range_id(n->in(k));
1066 if (!vreg) {
1067 continue;
1068 }
1069
1070 // If this instruction is CISC Spillable, add the flags
1071 // bit to its appropriate input
1072 if( UseCISCSpill && after_aggressive && inp == k ) {
1073 #ifndef PRODUCT
1074 if( TraceCISCSpill ) {
1075 tty->print(" use_cisc_RegMask: ");
1076 n->dump();
1077 }
1078 #endif
1079 n->as_Mach()->use_cisc_RegMask();
1080 }
1081
1082 if (is_machine_node && _scheduling_info_generated) {
1083 MachNode* cur_node = n->as_Mach();
1084 // this is cleaned up by register allocation
1085 if (k >= cur_node->num_opnds()) continue;
1086 }
1087
1088 LRG &lrg = lrgs(vreg);
1089 // // Testing for floating point code shape
1090 // Node *test = n->in(k);
1091 // if( test->is_Mach() ) {
1092 // MachNode *m = test->as_Mach();
1093 // int op = m->ideal_Opcode();
1094 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) {
1095 // int zzz = 1;
1096 // }
1097 // }
1098
1099 // Limit result register mask to acceptable registers.
1100 // Do not limit registers from uncommon uses before
1101 // AggressiveCoalesce. This effectively pre-virtual-splits
1102 // around uncommon uses of common defs.
1103 const RegMask &rm = n->in_RegMask(k);
1104 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) {
1105 // Since we are BEFORE aggressive coalesce, leave the register
1106 // mask untrimmed by the call. This encourages more coalescing.
1107 // Later, AFTER aggressive, this live range will have to spill
1108 // but the spiller handles slow-path calls very nicely.
1109 } else {
1110 lrg.AND( rm );
1111 }
1112
1113 // Check for bound register masks
1114 const RegMask &lrgmask = lrg.mask();
1115 uint kreg = n->in(k)->ideal_reg();
1116 bool is_vect = RegMask::is_vector(kreg);
1117 assert(n->in(k)->bottom_type()->isa_vect() == nullptr || is_vect ||
1118 kreg == Op_RegD || kreg == Op_RegL || kreg == Op_RegVectMask,
1119 "vector must be in vector registers");
1120 if (lrgmask.is_bound(kreg))
1121 lrg._is_bound = 1;
1122
1123 // If this use of a double forces a mis-aligned double,
1124 // flag as '_fat_proj' - really flag as allowing misalignment
1125 // AND changes how we count interferences. A mis-aligned
1126 // double can interfere with TWO aligned pairs, or effectively
1127 // FOUR registers!
1128 #ifdef ASSERT
1129 if (is_vect && !_scheduling_info_generated) {
1130 if (lrg.num_regs() != 0) {
1131 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned");
1132 assert(!lrg._fat_proj, "sanity");
1133 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity");
1134 } else {
1135 assert(n->is_Phi(), "not all inputs processed only if Phi");
1136 }
1137 }
1138 #endif
1139 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) {
1140 lrg._fat_proj = 1;
1141 lrg._is_bound = 1;
1142 }
1143 // if the LRG is an unaligned pair, we will have to spill
1144 // so clear the LRG's register mask if it is not already spilled
1145 if (!is_vect && !n->is_SpillCopy() &&
1146 (lrg._def == nullptr || lrg.is_multidef() || !lrg._def->is_SpillCopy()) &&
1147 lrgmask.is_misaligned_pair()) {
1148 lrg.Clear();
1149 }
1150
1151 // Check for maximum frequency value
1152 if (lrg._maxfreq < block->_freq) {
1153 lrg._maxfreq = block->_freq;
1154 }
1155
1156 } // End for all allocated inputs
1157 } // end for all instructions
1158 } // end for all blocks
1159
1160 // Final per-liverange setup
1161 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) {
1162 LRG &lrg = lrgs(i2);
1163 assert(!lrg._is_vector || !lrg._fat_proj, "sanity");
1164 if (lrg.num_regs() > 1 && !lrg._fat_proj) {
1165 lrg.clear_to_sets();
1166 }
1167 lrg.compute_set_mask_size();
1168 if (lrg.not_free()) { // Handle case where we lose from the start
1169 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG));
1170 lrg._direct_conflict = 1;
1171 }
1172 lrg.set_degree(0); // no neighbors in IFG yet
1173 }
1174 }
1175
1176 // Set the was-lo-degree bit. Conservative coalescing should not change the
1177 // colorability of the graph. If any live range was of low-degree before
1178 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
1179 // The bit is checked in Simplify.
1180 void PhaseChaitin::set_was_low() {
1181 #ifdef ASSERT
1182 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1183 int size = lrgs(i).num_regs();
1184 uint old_was_lo = lrgs(i)._was_lo;
1185 lrgs(i)._was_lo = 0;
1186 if( lrgs(i).lo_degree() ) {
1187 lrgs(i)._was_lo = 1; // Trivially of low degree
1188 } else { // Else check the Brigg's assertion
1189 // Brigg's observation is that the lo-degree neighbors of a
1190 // hi-degree live range will not interfere with the color choices
1191 // of said hi-degree live range. The Simplify reverse-stack-coloring
1192 // order takes care of the details. Hence you do not have to count
1193 // low-degree neighbors when determining if this guy colors.
1194 int briggs_degree = 0;
1195 IndexSet *s = _ifg->neighbors(i);
1196 IndexSetIterator elements(s);
1197 uint lidx;
1198 while((lidx = elements.next()) != 0) {
1199 if( !lrgs(lidx).lo_degree() )
1200 briggs_degree += MAX2(size,lrgs(lidx).num_regs());
1201 }
1202 if( briggs_degree < lrgs(i).degrees_of_freedom() )
1203 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion
1204 }
1205 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease");
1206 }
1207 #endif
1208 }
1209
1210 // Compute cost/area ratio, in case we spill. Build the lo-degree list.
1211 void PhaseChaitin::cache_lrg_info( ) {
1212 Compile::TracePhase tp(_t_chaitinCacheLRG);
1213
1214 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1215 LRG &lrg = lrgs(i);
1216
1217 // Check for being of low degree: means we can be trivially colored.
1218 // Low degree, dead or must-spill guys just get to simplify right away
1219 if( lrg.lo_degree() ||
1220 !lrg.alive() ||
1221 lrg._must_spill ) {
1222 // Split low degree list into those guys that must get a
1223 // register and those that can go to register or stack.
1224 // The idea is LRGs that can go register or stack color first when
1225 // they have a good chance of getting a register. The register-only
1226 // lo-degree live ranges always get a register.
1227 OptoReg::Name hi_reg = lrg.mask().find_last_elem();
1228 if( OptoReg::is_stack(hi_reg)) { // Can go to stack?
1229 lrg._next = _lo_stk_degree;
1230 _lo_stk_degree = i;
1231 } else {
1232 lrg._next = _lo_degree;
1233 _lo_degree = i;
1234 }
1235 } else { // Else high degree
1236 lrgs(_hi_degree)._prev = i;
1237 lrg._next = _hi_degree;
1238 lrg._prev = 0;
1239 _hi_degree = i;
1240 }
1241 }
1242 }
1243
1244 // Simplify the IFG by removing LRGs of low degree.
1245 void PhaseChaitin::Simplify( ) {
1246 Compile::TracePhase tp(_t_chaitinSimplify);
1247
1248 while( 1 ) { // Repeat till simplified it all
1249 // May want to explore simplifying lo_degree before _lo_stk_degree.
1250 // This might result in more spills coloring into registers during
1251 // Select().
1252 while( _lo_degree || _lo_stk_degree ) {
1253 // If possible, pull from lo_stk first
1254 uint lo;
1255 if( _lo_degree ) {
1256 lo = _lo_degree;
1257 _lo_degree = lrgs(lo)._next;
1258 } else {
1259 lo = _lo_stk_degree;
1260 _lo_stk_degree = lrgs(lo)._next;
1261 }
1262
1263 // Put the simplified guy on the simplified list.
1264 lrgs(lo)._next = _simplified;
1265 _simplified = lo;
1266 // If this guy is "at risk" then mark his current neighbors
1267 if (lrgs(lo)._at_risk && !_ifg->neighbors(lo)->is_empty()) {
1268 IndexSetIterator elements(_ifg->neighbors(lo));
1269 uint datum;
1270 while ((datum = elements.next()) != 0) {
1271 lrgs(datum)._risk_bias = lo;
1272 }
1273 }
1274
1275 // Yank this guy from the IFG.
1276 IndexSet *adj = _ifg->remove_node(lo);
1277 if (adj->is_empty()) {
1278 continue;
1279 }
1280
1281 // If any neighbors' degrees fall below their number of
1282 // allowed registers, then put that neighbor on the low degree
1283 // list. Note that 'degree' can only fall and 'numregs' is
1284 // unchanged by this action. Thus the two are equal at most once,
1285 // so LRGs hit the lo-degree worklist at most once.
1286 IndexSetIterator elements(adj);
1287 uint neighbor;
1288 while ((neighbor = elements.next()) != 0) {
1289 LRG *n = &lrgs(neighbor);
1290 #ifdef ASSERT
1291 if (VerifyRegisterAllocator) {
1292 assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
1293 }
1294 #endif
1295
1296 // Check for just becoming of-low-degree just counting registers.
1297 // _must_spill live ranges are already on the low degree list.
1298 if (n->just_lo_degree() && !n->_must_spill) {
1299 assert(!_ifg->_yanked->test(neighbor), "Cannot move to lo degree twice");
1300 // Pull from hi-degree list
1301 uint prev = n->_prev;
1302 uint next = n->_next;
1303 if (prev) {
1304 lrgs(prev)._next = next;
1305 } else {
1306 _hi_degree = next;
1307 }
1308 lrgs(next)._prev = prev;
1309 n->_next = _lo_degree;
1310 _lo_degree = neighbor;
1311 }
1312 }
1313 } // End of while lo-degree/lo_stk_degree worklist not empty
1314
1315 // Check for got everything: is hi-degree list empty?
1316 if (!_hi_degree) break;
1317
1318 // Time to pick a potential spill guy
1319 uint lo_score = _hi_degree;
1320 double score = lrgs(lo_score).score();
1321 double area = lrgs(lo_score)._area;
1322 double cost = lrgs(lo_score)._cost;
1323 bool bound = lrgs(lo_score)._is_bound;
1324
1325 // Find cheapest guy
1326 debug_only( int lo_no_simplify=0; );
1327 for (uint i = _hi_degree; i; i = lrgs(i)._next) {
1328 assert(!_ifg->_yanked->test(i), "");
1329 // It's just vaguely possible to move hi-degree to lo-degree without
1330 // going through a just-lo-degree stage: If you remove a double from
1331 // a float live range it's degree will drop by 2 and you can skip the
1332 // just-lo-degree stage. It's very rare (shows up after 5000+ methods
1333 // in -Xcomp of Java2Demo). So just choose this guy to simplify next.
1334 if( lrgs(i).lo_degree() ) {
1335 lo_score = i;
1336 break;
1337 }
1338 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; );
1339 double iscore = lrgs(i).score();
1340 double iarea = lrgs(i)._area;
1341 double icost = lrgs(i)._cost;
1342 bool ibound = lrgs(i)._is_bound;
1343
1344 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area
1345 // wins. Ties happen because all live ranges in question have spilled
1346 // a few times before and the spill-score adds a huge number which
1347 // washes out the low order bits. We are choosing the lesser of 2
1348 // evils; in this case pick largest area to spill.
1349 // Ties also happen when live ranges are defined and used only inside
1350 // one block. In which case their area is 0 and score set to max.
1351 // In such case choose bound live range over unbound to free registers
1352 // or with smaller cost to spill.
1353 if ( iscore < score ||
1354 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ||
1355 (iscore == score && iarea == area &&
1356 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) {
1357 lo_score = i;
1358 score = iscore;
1359 area = iarea;
1360 cost = icost;
1361 bound = ibound;
1362 }
1363 }
1364 LRG *lo_lrg = &lrgs(lo_score);
1365 // The live range we choose for spilling is either hi-degree, or very
1366 // rarely it can be low-degree. If we choose a hi-degree live range
1367 // there better not be any lo-degree choices.
1368 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" );
1369
1370 // Pull from hi-degree list
1371 uint prev = lo_lrg->_prev;
1372 uint next = lo_lrg->_next;
1373 if( prev ) lrgs(prev)._next = next;
1374 else _hi_degree = next;
1375 lrgs(next)._prev = prev;
1376 // Jam him on the lo-degree list, despite his high degree.
1377 // Maybe he'll get a color, and maybe he'll spill.
1378 // Only Select() will know.
1379 lrgs(lo_score)._at_risk = true;
1380 _lo_degree = lo_score;
1381 lo_lrg->_next = 0;
1382
1383 } // End of while not simplified everything
1384
1385 }
1386
1387 // Is 'reg' register legal for 'lrg'?
1388 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) {
1389 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) &&
1390 lrg.mask().Member(OptoReg::add(reg,-chunk))) {
1391 // RA uses OptoReg which represent the highest element of a registers set.
1392 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set
1393 // in which XMMd is used by RA to represent such vectors. A double value
1394 // uses [XMM,XMMb] pairs and XMMb is used by RA for it.
1395 // The register mask uses largest bits set of overlapping register sets.
1396 // On x86 with AVX it uses 8 bits for each XMM registers set.
1397 //
1398 // The 'lrg' already has cleared-to-set register mask (done in Select()
1399 // before calling choose_color()). Passing mask.Member(reg) check above
1400 // indicates that the size (num_regs) of 'reg' set is less or equal to
1401 // 'lrg' set size.
1402 // For set size 1 any register which is member of 'lrg' mask is legal.
1403 if (lrg.num_regs()==1)
1404 return true;
1405 // For larger sets only an aligned register with the same set size is legal.
1406 int mask = lrg.num_regs()-1;
1407 if ((reg&mask) == mask)
1408 return true;
1409 }
1410 return false;
1411 }
1412
1413 static OptoReg::Name find_first_set(LRG &lrg, RegMask mask, int chunk) {
1414 int num_regs = lrg.num_regs();
1415 OptoReg::Name assigned = mask.find_first_set(lrg, num_regs);
1416
1417 if (lrg.is_scalable()) {
1418 // a physical register is found
1419 if (chunk == 0 && OptoReg::is_reg(assigned)) {
1420 return assigned;
1421 }
1422
1423 // find available stack slots for scalable register
1424 if (lrg._is_vector) {
1425 num_regs = lrg.scalable_reg_slots();
1426 // if actual scalable vector register is exactly SlotsPerVecA * 32 bits
1427 if (num_regs == RegMask::SlotsPerVecA) {
1428 return assigned;
1429 }
1430
1431 // mask has been cleared out by clear_to_sets(SlotsPerVecA) before choose_color, but it
1432 // does not work for scalable size. We have to find adjacent scalable_reg_slots() bits
1433 // instead of SlotsPerVecA bits.
1434 assigned = mask.find_first_set(lrg, num_regs); // find highest valid reg
1435 while (OptoReg::is_valid(assigned) && RegMask::can_represent(assigned)) {
1436 // Verify the found reg has scalable_reg_slots() bits set.
1437 if (mask.is_valid_reg(assigned, num_regs)) {
1438 return assigned;
1439 } else {
1440 // Remove more for each iteration
1441 mask.Remove(assigned - num_regs + 1); // Unmask the lowest reg
1442 mask.clear_to_sets(RegMask::SlotsPerVecA); // Align by SlotsPerVecA bits
1443 assigned = mask.find_first_set(lrg, num_regs);
1444 }
1445 }
1446 return OptoReg::Bad; // will cause chunk change, and retry next chunk
1447 } else if (lrg._is_predicate) {
1448 assert(num_regs == RegMask::SlotsPerRegVectMask, "scalable predicate register");
1449 num_regs = lrg.scalable_reg_slots();
1450 mask.clear_to_sets(num_regs);
1451 return mask.find_first_set(lrg, num_regs);
1452 }
1453 }
1454
1455 return assigned;
1456 }
1457
1458 // Choose a color using the biasing heuristic
1459 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) {
1460
1461 // Check for "at_risk" LRG's
1462 uint risk_lrg = _lrg_map.find(lrg._risk_bias);
1463 if (risk_lrg != 0 && !_ifg->neighbors(risk_lrg)->is_empty()) {
1464 // Walk the colored neighbors of the "at_risk" candidate
1465 // Choose a color which is both legal and already taken by a neighbor
1466 // of the "at_risk" candidate in order to improve the chances of the
1467 // "at_risk" candidate of coloring
1468 IndexSetIterator elements(_ifg->neighbors(risk_lrg));
1469 uint datum;
1470 while ((datum = elements.next()) != 0) {
1471 OptoReg::Name reg = lrgs(datum).reg();
1472 // If this LRG's register is legal for us, choose it
1473 if (is_legal_reg(lrg, reg, chunk))
1474 return reg;
1475 }
1476 }
1477
1478 uint copy_lrg = _lrg_map.find(lrg._copy_bias);
1479 if (copy_lrg != 0) {
1480 // If he has a color,
1481 if(!_ifg->_yanked->test(copy_lrg)) {
1482 OptoReg::Name reg = lrgs(copy_lrg).reg();
1483 // And it is legal for you,
1484 if (is_legal_reg(lrg, reg, chunk))
1485 return reg;
1486 } else if( chunk == 0 ) {
1487 // Choose a color which is legal for him
1488 RegMask tempmask = lrg.mask();
1489 tempmask.AND(lrgs(copy_lrg).mask());
1490 tempmask.clear_to_sets(lrg.num_regs());
1491 OptoReg::Name reg = find_first_set(lrg, tempmask, chunk);
1492 if (OptoReg::is_valid(reg))
1493 return reg;
1494 }
1495 }
1496
1497 // If no bias info exists, just go with the register selection ordering
1498 if (lrg._is_vector || lrg.num_regs() == 2 || lrg.is_scalable()) {
1499 // Find an aligned set
1500 return OptoReg::add(find_first_set(lrg, lrg.mask(), chunk), chunk);
1501 }
1502
1503 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate
1504 // copy removal to remove many more copies, by preventing a just-assigned
1505 // register from being repeatedly assigned.
1506 OptoReg::Name reg = lrg.mask().find_first_elem();
1507 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) {
1508 // This 'Remove; find; Insert' idiom is an expensive way to find the
1509 // SECOND element in the mask.
1510 lrg.Remove(reg);
1511 OptoReg::Name reg2 = lrg.mask().find_first_elem();
1512 lrg.Insert(reg);
1513 if( OptoReg::is_reg(reg2))
1514 reg = reg2;
1515 }
1516 return OptoReg::add( reg, chunk );
1517 }
1518
1519 // Choose a color in the current chunk
1520 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) {
1521 assert( C->in_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP-1)), "must not allocate stack0 (inside preserve area)");
1522 assert(C->out_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP+0)), "must not allocate stack0 (inside preserve area)");
1523
1524 if( lrg.num_regs() == 1 || // Common Case
1525 !lrg._fat_proj ) // Aligned+adjacent pairs ok
1526 // Use a heuristic to "bias" the color choice
1527 return bias_color(lrg, chunk);
1528
1529 assert(!lrg._is_vector, "should be not vector here" );
1530 assert( lrg.num_regs() >= 2, "dead live ranges do not color" );
1531
1532 // Fat-proj case or misaligned double argument.
1533 assert(lrg.compute_mask_size() == lrg.num_regs() ||
1534 lrg.num_regs() == 2,"fat projs exactly color" );
1535 assert( !chunk, "always color in 1st chunk" );
1536 // Return the highest element in the set.
1537 return lrg.mask().find_last_elem();
1538 }
1539
1540 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted
1541 // in reverse order of removal. As long as nothing of hi-degree was yanked,
1542 // everything going back is guaranteed a color. Select that color. If some
1543 // hi-degree LRG cannot get a color then we record that we must spill.
1544 uint PhaseChaitin::Select( ) {
1545 Compile::TracePhase tp(_t_chaitinSelect);
1546
1547 uint spill_reg = LRG::SPILL_REG;
1548 _max_reg = OptoReg::Name(0); // Past max register used
1549 while( _simplified ) {
1550 // Pull next LRG from the simplified list - in reverse order of removal
1551 uint lidx = _simplified;
1552 LRG *lrg = &lrgs(lidx);
1553 _simplified = lrg->_next;
1554
1555 #ifndef PRODUCT
1556 if (trace_spilling()) {
1557 ttyLocker ttyl;
1558 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(),
1559 lrg->degrees_of_freedom());
1560 lrg->dump();
1561 }
1562 #endif
1563
1564 // Re-insert into the IFG
1565 _ifg->re_insert(lidx);
1566 if( !lrg->alive() ) continue;
1567 // capture allstackedness flag before mask is hacked
1568 const int is_allstack = lrg->mask().is_AllStack();
1569
1570 // Yeah, yeah, yeah, I know, I know. I can refactor this
1571 // to avoid the GOTO, although the refactored code will not
1572 // be much clearer. We arrive here IFF we have a stack-based
1573 // live range that cannot color in the current chunk, and it
1574 // has to move into the next free stack chunk.
1575 int chunk = 0; // Current chunk is first chunk
1576 retry_next_chunk:
1577
1578 // Remove neighbor colors
1579 IndexSet *s = _ifg->neighbors(lidx);
1580 debug_only(RegMask orig_mask = lrg->mask();)
1581
1582 if (!s->is_empty()) {
1583 IndexSetIterator elements(s);
1584 uint neighbor;
1585 while ((neighbor = elements.next()) != 0) {
1586 // Note that neighbor might be a spill_reg. In this case, exclusion
1587 // of its color will be a no-op, since the spill_reg chunk is in outer
1588 // space. Also, if neighbor is in a different chunk, this exclusion
1589 // will be a no-op. (Later on, if lrg runs out of possible colors in
1590 // its chunk, a new chunk of color may be tried, in which case
1591 // examination of neighbors is started again, at retry_next_chunk.)
1592 LRG &nlrg = lrgs(neighbor);
1593 OptoReg::Name nreg = nlrg.reg();
1594 // Only subtract masks in the same chunk
1595 if (nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE) {
1596 #ifndef PRODUCT
1597 uint size = lrg->mask().Size();
1598 RegMask rm = lrg->mask();
1599 #endif
1600 lrg->SUBTRACT(nlrg.mask());
1601 #ifndef PRODUCT
1602 if (trace_spilling() && lrg->mask().Size() != size) {
1603 ttyLocker ttyl;
1604 tty->print("L%d ", lidx);
1605 rm.dump();
1606 tty->print(" intersected L%d ", neighbor);
1607 nlrg.mask().dump();
1608 tty->print(" removed ");
1609 rm.SUBTRACT(lrg->mask());
1610 rm.dump();
1611 tty->print(" leaving ");
1612 lrg->mask().dump();
1613 tty->cr();
1614 }
1615 #endif
1616 }
1617 }
1618 }
1619 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness");
1620 // Aligned pairs need aligned masks
1621 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1622 if (lrg->num_regs() > 1 && !lrg->_fat_proj) {
1623 lrg->clear_to_sets();
1624 }
1625
1626 // Check if a color is available and if so pick the color
1627 OptoReg::Name reg = choose_color( *lrg, chunk );
1628
1629 //---------------
1630 // If we fail to color and the AllStack flag is set, trigger
1631 // a chunk-rollover event
1632 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) {
1633 // Bump register mask up to next stack chunk
1634 chunk += RegMask::CHUNK_SIZE;
1635 lrg->Set_All();
1636 goto retry_next_chunk;
1637 }
1638
1639 //---------------
1640 // Did we get a color?
1641 else if( OptoReg::is_valid(reg)) {
1642 #ifndef PRODUCT
1643 RegMask avail_rm = lrg->mask();
1644 #endif
1645
1646 // Record selected register
1647 lrg->set_reg(reg);
1648
1649 if( reg >= _max_reg ) // Compute max register limit
1650 _max_reg = OptoReg::add(reg,1);
1651 // Fold reg back into normal space
1652 reg = OptoReg::add(reg,-chunk);
1653
1654 // If the live range is not bound, then we actually had some choices
1655 // to make. In this case, the mask has more bits in it than the colors
1656 // chosen. Restrict the mask to just what was picked.
1657 int n_regs = lrg->num_regs();
1658 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1659 if (n_regs == 1 || !lrg->_fat_proj) {
1660 if (Matcher::supports_scalable_vector()) {
1661 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecA, "sanity");
1662 } else {
1663 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity");
1664 }
1665 lrg->Clear(); // Clear the mask
1666 lrg->Insert(reg); // Set regmask to match selected reg
1667 // For vectors and pairs, also insert the low bit of the pair
1668 // We always choose the high bit, then mask the low bits by register size
1669 if (lrg->is_scalable() && OptoReg::is_stack(lrg->reg())) { // stack
1670 n_regs = lrg->scalable_reg_slots();
1671 }
1672 for (int i = 1; i < n_regs; i++) {
1673 lrg->Insert(OptoReg::add(reg,-i));
1674 }
1675 lrg->set_mask_size(n_regs);
1676 } else { // Else fatproj
1677 // mask must be equal to fatproj bits, by definition
1678 }
1679 #ifndef PRODUCT
1680 if (trace_spilling()) {
1681 ttyLocker ttyl;
1682 tty->print("L%d selected ", lidx);
1683 lrg->mask().dump();
1684 tty->print(" from ");
1685 avail_rm.dump();
1686 tty->cr();
1687 }
1688 #endif
1689 // Note that reg is the highest-numbered register in the newly-bound mask.
1690 } // end color available case
1691
1692 //---------------
1693 // Live range is live and no colors available
1694 else {
1695 assert( lrg->alive(), "" );
1696 assert( !lrg->_fat_proj || lrg->is_multidef() ||
1697 lrg->_def->outcnt() > 0, "fat_proj cannot spill");
1698 assert( !orig_mask.is_AllStack(), "All Stack does not spill" );
1699
1700 // Assign the special spillreg register
1701 lrg->set_reg(OptoReg::Name(spill_reg++));
1702 // Do not empty the regmask; leave mask_size lying around
1703 // for use during Spilling
1704 #ifndef PRODUCT
1705 if( trace_spilling() ) {
1706 ttyLocker ttyl;
1707 tty->print("L%d spilling with neighbors: ", lidx);
1708 s->dump();
1709 debug_only(tty->print(" original mask: "));
1710 debug_only(orig_mask.dump());
1711 dump_lrg(lidx);
1712 }
1713 #endif
1714 } // end spill case
1715
1716 }
1717
1718 return spill_reg-LRG::SPILL_REG; // Return number of spills
1719 }
1720
1721 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
1722 void PhaseChaitin::set_was_spilled( Node *n ) {
1723 if( _spilled_once.test_set(n->_idx) )
1724 _spilled_twice.set(n->_idx);
1725 }
1726
1727 // Convert Ideal spill instructions into proper FramePtr + offset Loads and
1728 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are.
1729 void PhaseChaitin::fixup_spills() {
1730 // This function does only cisc spill work.
1731 if( !UseCISCSpill ) return;
1732
1733 Compile::TracePhase tp(_t_fixupSpills);
1734
1735 // Grab the Frame Pointer
1736 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr);
1737
1738 // For all blocks
1739 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1740 Block* block = _cfg.get_block(i);
1741
1742 // For all instructions in block
1743 uint last_inst = block->end_idx();
1744 for (uint j = 1; j <= last_inst; j++) {
1745 Node* n = block->get_node(j);
1746
1747 // Dead instruction???
1748 assert( n->outcnt() != 0 ||// Nothing dead after post alloc
1749 C->top() == n || // Or the random TOP node
1750 n->is_Proj(), // Or a fat-proj kill node
1751 "No dead instructions after post-alloc" );
1752
1753 int inp = n->cisc_operand();
1754 if( inp != AdlcVMDeps::Not_cisc_spillable ) {
1755 // Convert operand number to edge index number
1756 MachNode *mach = n->as_Mach();
1757 inp = mach->operand_index(inp);
1758 Node *src = n->in(inp); // Value to load or store
1759 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src));
1760 OptoReg::Name src_reg = lrg_cisc.reg();
1761 // Doubles record the HIGH register of an adjacent pair.
1762 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs());
1763 if( OptoReg::is_stack(src_reg) ) { // If input is on stack
1764 // This is a CISC Spill, get stack offset and construct new node
1765 #ifndef PRODUCT
1766 if( TraceCISCSpill ) {
1767 tty->print(" reg-instr: ");
1768 n->dump();
1769 }
1770 #endif
1771 int stk_offset = reg2offset(src_reg);
1772 // Bailout if we might exceed node limit when spilling this instruction
1773 C->check_node_count(0, "out of nodes fixing spills");
1774 if (C->failing()) return;
1775 // Transform node
1776 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach();
1777 cisc->set_req(inp,fp); // Base register is frame pointer
1778 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) {
1779 assert( cisc->oper_input_base() == 2, "Only adding one edge");
1780 cisc->ins_req(1,src); // Requires a memory edge
1781 } else {
1782 // There is no space reserved for a memory edge before the inputs for
1783 // instructions which have "stackSlotX" parameter instead of "memory".
1784 // For example, "MoveF2I_stack_reg". We always need a memory edge from
1785 // src to cisc, else we might schedule cisc before src, loading from a
1786 // spill location before storing the spill. On some platforms, we land
1787 // in this else case because mach->oper_input_base() > 1, i.e. we have
1788 // multiple inputs. In some rare cases there are even multiple memory
1789 // operands, before and after spilling.
1790 // (e.g. spilling "addFPR24_reg_mem" to "addFPR24_mem_cisc")
1791 // In either case, there is no space in the inputs for the memory edge
1792 // so we add an additional precedence / memory edge.
1793 cisc->add_prec(src);
1794 }
1795 block->map_node(cisc, j); // Insert into basic block
1796 n->subsume_by(cisc, C); // Correct graph
1797 //
1798 ++_used_cisc_instructions;
1799 #ifndef PRODUCT
1800 if( TraceCISCSpill ) {
1801 tty->print(" cisc-instr: ");
1802 cisc->dump();
1803 }
1804 #endif
1805 } else {
1806 #ifndef PRODUCT
1807 if( TraceCISCSpill ) {
1808 tty->print(" using reg-instr: ");
1809 n->dump();
1810 }
1811 #endif
1812 ++_unused_cisc_instructions; // input can be on stack
1813 }
1814 }
1815
1816 } // End of for all instructions
1817
1818 } // End of for all blocks
1819 }
1820
1821 // Helper to stretch above; recursively discover the base Node for a
1822 // given derived Node. Easy for AddP-related machine nodes, but needs
1823 // to be recursive for derived Phis.
1824 Node* PhaseChaitin::find_base_for_derived(Node** derived_base_map, Node* derived, uint& maxlrg) {
1825 // See if already computed; if so return it
1826 if (derived_base_map[derived->_idx]) {
1827 return derived_base_map[derived->_idx];
1828 }
1829
1830 #ifdef ASSERT
1831 if (derived->is_Mach() && derived->as_Mach()->ideal_Opcode() == Op_VerifyVectorAlignment) {
1832 // Bypass the verification node
1833 Node* base = find_base_for_derived(derived_base_map, derived->in(1), maxlrg);
1834 derived_base_map[derived->_idx] = base;
1835 return base;
1836 }
1837 #endif
1838
1839 // See if this happens to be a base.
1840 // NOTE: we use TypePtr instead of TypeOopPtr because we can have
1841 // pointers derived from null! These are always along paths that
1842 // can't happen at run-time but the optimizer cannot deduce it so
1843 // we have to handle it gracefully.
1844 assert(!derived->bottom_type()->isa_narrowoop() ||
1845 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity");
1846 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1847 // If its an OOP with a non-zero offset, then it is derived.
1848 if( tj == nullptr || tj->_offset == 0 ) {
1849 derived_base_map[derived->_idx] = derived;
1850 return derived;
1851 }
1852 // Derived is null+offset? Base is null!
1853 if( derived->is_Con() ) {
1854 Node *base = _matcher.mach_null();
1855 assert(base != nullptr, "sanity");
1856 if (base->in(0) == nullptr) {
1857 // Initialize it once and make it shared:
1858 // set control to _root and place it into Start block
1859 // (where top() node is placed).
1860 base->init_req(0, _cfg.get_root_node());
1861 Block *startb = _cfg.get_block_for_node(C->top());
1862 uint node_pos = startb->find_node(C->top());
1863 startb->insert_node(base, node_pos);
1864 _cfg.map_node_to_block(base, startb);
1865 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet");
1866
1867 // The loadConP0 might have projection nodes depending on architecture
1868 // Add the projection nodes to the CFG
1869 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) {
1870 Node* use = base->fast_out(i);
1871 if (use->is_MachProj()) {
1872 startb->insert_node(use, ++node_pos);
1873 _cfg.map_node_to_block(use, startb);
1874 new_lrg(use, maxlrg++);
1875 }
1876 }
1877 }
1878 if (_lrg_map.live_range_id(base) == 0) {
1879 new_lrg(base, maxlrg++);
1880 }
1881 assert(base->in(0) == _cfg.get_root_node() && _cfg.get_block_for_node(base) == _cfg.get_block_for_node(C->top()), "base null should be shared");
1882 derived_base_map[derived->_idx] = base;
1883 return base;
1884 }
1885
1886 // Check for AddP-related opcodes
1887 if (!derived->is_Phi()) {
1888 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name());
1889 Node *base = derived->in(AddPNode::Base);
1890 derived_base_map[derived->_idx] = base;
1891 return base;
1892 }
1893
1894 // Recursively find bases for Phis.
1895 // First check to see if we can avoid a base Phi here.
1896 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg);
1897 uint i;
1898 for( i = 2; i < derived->req(); i++ )
1899 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg))
1900 break;
1901 // Went to the end without finding any different bases?
1902 if( i == derived->req() ) { // No need for a base Phi here
1903 derived_base_map[derived->_idx] = base;
1904 return base;
1905 }
1906
1907 // Now we see we need a base-Phi here to merge the bases
1908 const Type *t = base->bottom_type();
1909 base = new PhiNode( derived->in(0), t );
1910 for( i = 1; i < derived->req(); i++ ) {
1911 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg));
1912 t = t->meet(base->in(i)->bottom_type());
1913 }
1914 base->as_Phi()->set_type(t);
1915
1916 // Search the current block for an existing base-Phi
1917 Block *b = _cfg.get_block_for_node(derived);
1918 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi
1919 Node *phi = b->get_node(i);
1920 if( !phi->is_Phi() ) { // Found end of Phis with no match?
1921 b->insert_node(base, i); // Must insert created Phi here as base
1922 _cfg.map_node_to_block(base, b);
1923 new_lrg(base,maxlrg++);
1924 break;
1925 }
1926 // See if Phi matches.
1927 uint j;
1928 for( j = 1; j < base->req(); j++ )
1929 if( phi->in(j) != base->in(j) &&
1930 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different nulls
1931 break;
1932 if( j == base->req() ) { // All inputs match?
1933 base = phi; // Then use existing 'phi' and drop 'base'
1934 break;
1935 }
1936 }
1937
1938
1939 // Cache info for later passes
1940 derived_base_map[derived->_idx] = base;
1941 return base;
1942 }
1943
1944 // At each Safepoint, insert extra debug edges for each pair of derived value/
1945 // base pointer that is live across the Safepoint for oopmap building. The
1946 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the
1947 // required edge set.
1948 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) {
1949 int must_recompute_live = false;
1950 uint maxlrg = _lrg_map.max_lrg_id();
1951 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique());
1952 memset( derived_base_map, 0, sizeof(Node*)*C->unique() );
1953
1954 // For all blocks in RPO do...
1955 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1956 Block* block = _cfg.get_block(i);
1957 // Note use of deep-copy constructor. I cannot hammer the original
1958 // liveout bits, because they are needed by the following coalesce pass.
1959 IndexSet liveout(_live->live(block));
1960
1961 for (uint j = block->end_idx() + 1; j > 1; j--) {
1962 Node* n = block->get_node(j - 1);
1963
1964 // Pre-split compares of loop-phis. Loop-phis form a cycle we would
1965 // like to see in the same register. Compare uses the loop-phi and so
1966 // extends its live range BUT cannot be part of the cycle. If this
1967 // extended live range overlaps with the update of the loop-phi value
1968 // we need both alive at the same time -- which requires at least 1
1969 // copy. But because Intel has only 2-address registers we end up with
1970 // at least 2 copies, one before the loop-phi update instruction and
1971 // one after. Instead we split the input to the compare just after the
1972 // phi.
1973 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) {
1974 Node *phi = n->in(1);
1975 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) {
1976 Block *phi_block = _cfg.get_block_for_node(phi);
1977 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) {
1978 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI];
1979 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask);
1980 insert_proj( phi_block, 1, spill, maxlrg++ );
1981 n->set_req(1,spill);
1982 must_recompute_live = true;
1983 }
1984 }
1985 }
1986
1987 // Get value being defined
1988 uint lidx = _lrg_map.live_range_id(n);
1989 // Ignore the occasional brand-new live range
1990 if (lidx && lidx < _lrg_map.max_lrg_id()) {
1991 // Remove from live-out set
1992 liveout.remove(lidx);
1993
1994 // Copies do not define a new value and so do not interfere.
1995 // Remove the copies source from the liveout set before interfering.
1996 uint idx = n->is_Copy();
1997 if (idx) {
1998 liveout.remove(_lrg_map.live_range_id(n->in(idx)));
1999 }
2000 }
2001
2002 // Found a safepoint?
2003 JVMState *jvms = n->jvms();
2004 if (jvms && !liveout.is_empty()) {
2005 // Now scan for a live derived pointer
2006 IndexSetIterator elements(&liveout);
2007 uint neighbor;
2008 while ((neighbor = elements.next()) != 0) {
2009 // Find reaching DEF for base and derived values
2010 // This works because we are still in SSA during this call.
2011 Node *derived = lrgs(neighbor)._def;
2012 const TypePtr *tj = derived->bottom_type()->isa_ptr();
2013 assert(!derived->bottom_type()->isa_narrowoop() ||
2014 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity");
2015 // If its an OOP with a non-zero offset, then it is derived.
2016 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) {
2017 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg);
2018 assert(base->_idx < _lrg_map.size(), "");
2019 // Add reaching DEFs of derived pointer and base pointer as a
2020 // pair of inputs
2021 n->add_req(derived);
2022 n->add_req(base);
2023
2024 // See if the base pointer is already live to this point.
2025 // Since I'm working on the SSA form, live-ness amounts to
2026 // reaching def's. So if I find the base's live range then
2027 // I know the base's def reaches here.
2028 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or
2029 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND
2030 (_lrg_map.live_range_id(base) > 0) && // not a constant
2031 _cfg.get_block_for_node(base) != block) { // base not def'd in blk)
2032 // Base pointer is not currently live. Since I stretched
2033 // the base pointer to here and it crosses basic-block
2034 // boundaries, the global live info is now incorrect.
2035 // Recompute live.
2036 must_recompute_live = true;
2037 } // End of if base pointer is not live to debug info
2038 }
2039 } // End of scan all live data for derived ptrs crossing GC point
2040 } // End of if found a GC point
2041
2042 // Make all inputs live
2043 if (!n->is_Phi()) { // Phi function uses come from prior block
2044 for (uint k = 1; k < n->req(); k++) {
2045 uint lidx = _lrg_map.live_range_id(n->in(k));
2046 if (lidx < _lrg_map.max_lrg_id()) {
2047 liveout.insert(lidx);
2048 }
2049 }
2050 }
2051
2052 } // End of forall instructions in block
2053 liveout.clear(); // Free the memory used by liveout.
2054
2055 } // End of forall blocks
2056 _lrg_map.set_max_lrg_id(maxlrg);
2057
2058 // If I created a new live range I need to recompute live
2059 if (maxlrg != _ifg->_maxlrg) {
2060 must_recompute_live = true;
2061 }
2062
2063 return must_recompute_live != 0;
2064 }
2065
2066 // Extend the node to LRG mapping
2067
2068 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) {
2069 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node));
2070 }
2071
2072 #ifndef PRODUCT
2073 void PhaseChaitin::dump(const Node* n) const {
2074 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0;
2075 tty->print("L%d",r);
2076 if (r && n->Opcode() != Op_Phi) {
2077 if( _node_regs ) { // Got a post-allocation copy of allocation?
2078 tty->print("[");
2079 OptoReg::Name second = get_reg_second(n);
2080 if( OptoReg::is_valid(second) ) {
2081 if( OptoReg::is_reg(second) )
2082 tty->print("%s:",Matcher::regName[second]);
2083 else
2084 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second));
2085 }
2086 OptoReg::Name first = get_reg_first(n);
2087 if( OptoReg::is_reg(first) )
2088 tty->print("%s]",Matcher::regName[first]);
2089 else
2090 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first));
2091 } else
2092 n->out_RegMask().dump();
2093 }
2094 tty->print("/N%d\t",n->_idx);
2095 tty->print("%s === ", n->Name());
2096 uint k;
2097 for (k = 0; k < n->req(); k++) {
2098 Node *m = n->in(k);
2099 if (!m) {
2100 tty->print("_ ");
2101 }
2102 else {
2103 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2104 tty->print("L%d",r);
2105 // Data MultiNode's can have projections with no real registers.
2106 // Don't die while dumping them.
2107 int op = n->Opcode();
2108 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) {
2109 if( _node_regs ) {
2110 tty->print("[");
2111 OptoReg::Name second = get_reg_second(n->in(k));
2112 if( OptoReg::is_valid(second) ) {
2113 if( OptoReg::is_reg(second) )
2114 tty->print("%s:",Matcher::regName[second]);
2115 else
2116 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer),
2117 reg2offset_unchecked(second));
2118 }
2119 OptoReg::Name first = get_reg_first(n->in(k));
2120 if( OptoReg::is_reg(first) )
2121 tty->print("%s]",Matcher::regName[first]);
2122 else
2123 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer),
2124 reg2offset_unchecked(first));
2125 } else
2126 n->in_RegMask(k).dump();
2127 }
2128 tty->print("/N%d ",m->_idx);
2129 }
2130 }
2131 if( k < n->len() && n->in(k) ) tty->print("| ");
2132 for( ; k < n->len(); k++ ) {
2133 Node *m = n->in(k);
2134 if(!m) {
2135 break;
2136 }
2137 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2138 tty->print("L%d",r);
2139 tty->print("/N%d ",m->_idx);
2140 }
2141 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty);
2142 else n->dump_spec(tty);
2143 if( _spilled_once.test(n->_idx ) ) {
2144 tty->print(" Spill_1");
2145 if( _spilled_twice.test(n->_idx ) )
2146 tty->print(" Spill_2");
2147 }
2148 tty->print("\n");
2149 }
2150
2151 void PhaseChaitin::dump(const Block* b) const {
2152 b->dump_head(&_cfg);
2153
2154 // For all instructions
2155 for( uint j = 0; j < b->number_of_nodes(); j++ )
2156 dump(b->get_node(j));
2157 // Print live-out info at end of block
2158 if( _live ) {
2159 tty->print("Liveout: ");
2160 IndexSet *live = _live->live(b);
2161 IndexSetIterator elements(live);
2162 tty->print("{");
2163 uint i;
2164 while ((i = elements.next()) != 0) {
2165 tty->print("L%d ", _lrg_map.find_const(i));
2166 }
2167 tty->print_cr("}");
2168 }
2169 tty->print("\n");
2170 }
2171
2172 void PhaseChaitin::dump() const {
2173 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n",
2174 _matcher._new_SP, _framesize );
2175
2176 // For all blocks
2177 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2178 dump(_cfg.get_block(i));
2179 }
2180 // End of per-block dump
2181 tty->print("\n");
2182
2183 if (!_ifg) {
2184 tty->print("(No IFG.)\n");
2185 return;
2186 }
2187
2188 // Dump LRG array
2189 tty->print("--- Live RanGe Array ---\n");
2190 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) {
2191 tty->print("L%d: ",i2);
2192 if (i2 < _ifg->_maxlrg) {
2193 lrgs(i2).dump();
2194 }
2195 else {
2196 tty->print_cr("new LRG");
2197 }
2198 }
2199 tty->cr();
2200
2201 // Dump lo-degree list
2202 tty->print("Lo degree: ");
2203 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next )
2204 tty->print("L%d ",i3);
2205 tty->cr();
2206
2207 // Dump lo-stk-degree list
2208 tty->print("Lo stk degree: ");
2209 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next )
2210 tty->print("L%d ",i4);
2211 tty->cr();
2212
2213 // Dump lo-degree list
2214 tty->print("Hi degree: ");
2215 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next )
2216 tty->print("L%d ",i5);
2217 tty->cr();
2218 }
2219
2220 void PhaseChaitin::dump_degree_lists() const {
2221 // Dump lo-degree list
2222 tty->print("Lo degree: ");
2223 for( uint i = _lo_degree; i; i = lrgs(i)._next )
2224 tty->print("L%d ",i);
2225 tty->cr();
2226
2227 // Dump lo-stk-degree list
2228 tty->print("Lo stk degree: ");
2229 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next )
2230 tty->print("L%d ",i2);
2231 tty->cr();
2232
2233 // Dump lo-degree list
2234 tty->print("Hi degree: ");
2235 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next )
2236 tty->print("L%d ",i3);
2237 tty->cr();
2238 }
2239
2240 void PhaseChaitin::dump_simplified() const {
2241 tty->print("Simplified: ");
2242 for( uint i = _simplified; i; i = lrgs(i)._next )
2243 tty->print("L%d ",i);
2244 tty->cr();
2245 }
2246
2247 static char *print_reg(OptoReg::Name reg, const PhaseChaitin* pc, char* buf, size_t buf_size) {
2248 if ((int)reg < 0)
2249 os::snprintf_checked(buf, buf_size, "<OptoReg::%d>", (int)reg);
2250 else if (OptoReg::is_reg(reg))
2251 strcpy(buf, Matcher::regName[reg]);
2252 else
2253 os::snprintf_checked(buf, buf_size, "%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer),
2254 pc->reg2offset(reg));
2255 return buf+strlen(buf);
2256 }
2257
2258 // Dump a register name into a buffer. Be intelligent if we get called
2259 // before allocation is complete.
2260 char *PhaseChaitin::dump_register(const Node* n, char* buf, size_t buf_size) const {
2261 if( _node_regs ) {
2262 // Post allocation, use direct mappings, no LRG info available
2263 print_reg( get_reg_first(n), this, buf, buf_size);
2264 } else {
2265 uint lidx = _lrg_map.find_const(n); // Grab LRG number
2266 if( !_ifg ) {
2267 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2268 } else if( !lidx ) { // Special, not allocated value
2269 strcpy(buf,"Special");
2270 } else {
2271 if (lrgs(lidx)._is_vector) {
2272 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs()))
2273 print_reg( lrgs(lidx).reg(), this, buf, buf_size); // a bound machine register
2274 else
2275 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2276 } else if( (lrgs(lidx).num_regs() == 1)
2277 ? lrgs(lidx).mask().is_bound1()
2278 : lrgs(lidx).mask().is_bound_pair() ) {
2279 // Hah! We have a bound machine register
2280 print_reg( lrgs(lidx).reg(), this, buf, buf_size);
2281 } else {
2282 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2283 }
2284 }
2285 }
2286 return buf+strlen(buf);
2287 }
2288
2289 void PhaseChaitin::dump_for_spill_split_recycle() const {
2290 if( WizardMode && (PrintCompilation || PrintOpto) ) {
2291 // Display which live ranges need to be split and the allocator's state
2292 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt);
2293 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) {
2294 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
2295 tty->print("L%d: ", bidx);
2296 lrgs(bidx).dump();
2297 }
2298 }
2299 tty->cr();
2300 dump();
2301 }
2302 }
2303
2304 void PhaseChaitin::dump_frame() const {
2305 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer);
2306 const TypeTuple *domain = C->tf()->domain();
2307 const int argcnt = domain->cnt() - TypeFunc::Parms;
2308
2309 // Incoming arguments in registers dump
2310 for( int k = 0; k < argcnt; k++ ) {
2311 OptoReg::Name parmreg = _matcher._parm_regs[k].first();
2312 if( OptoReg::is_reg(parmreg)) {
2313 const char *reg_name = OptoReg::regname(parmreg);
2314 tty->print("#r%3.3d %s", parmreg, reg_name);
2315 parmreg = _matcher._parm_regs[k].second();
2316 if( OptoReg::is_reg(parmreg)) {
2317 tty->print(":%s", OptoReg::regname(parmreg));
2318 }
2319 tty->print(" : parm %d: ", k);
2320 domain->field_at(k + TypeFunc::Parms)->dump();
2321 tty->cr();
2322 }
2323 }
2324
2325 // Check for un-owned padding above incoming args
2326 OptoReg::Name reg = _matcher._new_SP;
2327 if( reg > _matcher._in_arg_limit ) {
2328 reg = OptoReg::add(reg, -1);
2329 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg));
2330 }
2331
2332 // Incoming argument area dump
2333 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots());
2334 while( reg > begin_in_arg ) {
2335 reg = OptoReg::add(reg, -1);
2336 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2337 int j;
2338 for( j = 0; j < argcnt; j++) {
2339 if( _matcher._parm_regs[j].first() == reg ||
2340 _matcher._parm_regs[j].second() == reg ) {
2341 tty->print("parm %d: ",j);
2342 domain->field_at(j + TypeFunc::Parms)->dump();
2343 tty->cr();
2344 break;
2345 }
2346 }
2347 if( j >= argcnt )
2348 tty->print_cr("HOLE, owned by SELF");
2349 }
2350
2351 // Old outgoing preserve area
2352 while( reg > _matcher._old_SP ) {
2353 reg = OptoReg::add(reg, -1);
2354 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg));
2355 }
2356
2357 // Old SP
2358 tty->print_cr("# -- Old %s -- Framesize: %d --",fp,
2359 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize);
2360
2361 // Preserve area dump
2362 int fixed_slots = C->fixed_slots();
2363 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots());
2364 OptoReg::Name return_addr = _matcher.return_addr();
2365
2366 reg = OptoReg::add(reg, -1);
2367 while (OptoReg::is_stack(reg)) {
2368 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2369 if (return_addr == reg) {
2370 tty->print_cr("return address");
2371 } else if (reg >= begin_in_preserve) {
2372 // Preserved slots are present on x86
2373 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word))
2374 tty->print_cr("saved fp register");
2375 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) &&
2376 VerifyStackAtCalls)
2377 tty->print_cr("0xBADB100D +VerifyStackAtCalls");
2378 else
2379 tty->print_cr("in_preserve");
2380 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) {
2381 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg));
2382 } else {
2383 tty->print_cr("pad2, stack alignment");
2384 }
2385 reg = OptoReg::add(reg, -1);
2386 }
2387
2388 // Spill area dump
2389 reg = OptoReg::add(_matcher._new_SP, _framesize );
2390 while( reg > _matcher._out_arg_limit ) {
2391 reg = OptoReg::add(reg, -1);
2392 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg));
2393 }
2394
2395 // Outgoing argument area dump
2396 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) {
2397 reg = OptoReg::add(reg, -1);
2398 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg));
2399 }
2400
2401 // Outgoing new preserve area
2402 while( reg > _matcher._new_SP ) {
2403 reg = OptoReg::add(reg, -1);
2404 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg));
2405 }
2406 tty->print_cr("#");
2407 }
2408
2409 void PhaseChaitin::dump_bb(uint pre_order) const {
2410 tty->print_cr("---dump of B%d---",pre_order);
2411 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2412 Block* block = _cfg.get_block(i);
2413 if (block->_pre_order == pre_order) {
2414 dump(block);
2415 }
2416 }
2417 }
2418
2419 void PhaseChaitin::dump_lrg(uint lidx, bool defs_only) const {
2420 tty->print_cr("---dump of L%d---",lidx);
2421
2422 if (_ifg) {
2423 if (lidx >= _lrg_map.max_lrg_id()) {
2424 tty->print("Attempt to print live range index beyond max live range.\n");
2425 return;
2426 }
2427 tty->print("L%d: ",lidx);
2428 if (lidx < _ifg->_maxlrg) {
2429 lrgs(lidx).dump();
2430 } else {
2431 tty->print_cr("new LRG");
2432 }
2433 }
2434 if( _ifg && lidx < _ifg->_maxlrg) {
2435 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx));
2436 _ifg->neighbors(lidx)->dump();
2437 tty->cr();
2438 }
2439 // For all blocks
2440 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2441 Block* block = _cfg.get_block(i);
2442 int dump_once = 0;
2443
2444 // For all instructions
2445 for( uint j = 0; j < block->number_of_nodes(); j++ ) {
2446 Node *n = block->get_node(j);
2447 if (_lrg_map.find_const(n) == lidx) {
2448 if (!dump_once++) {
2449 tty->cr();
2450 block->dump_head(&_cfg);
2451 }
2452 dump(n);
2453 continue;
2454 }
2455 if (!defs_only) {
2456 uint cnt = n->req();
2457 for( uint k = 1; k < cnt; k++ ) {
2458 Node *m = n->in(k);
2459 if (!m) {
2460 continue; // be robust in the dumper
2461 }
2462 if (_lrg_map.find_const(m) == lidx) {
2463 if (!dump_once++) {
2464 tty->cr();
2465 block->dump_head(&_cfg);
2466 }
2467 dump(n);
2468 }
2469 }
2470 }
2471 }
2472 } // End of per-block dump
2473 tty->cr();
2474 }
2475 #endif // not PRODUCT
2476
2477 #ifdef ASSERT
2478 // Verify that base pointers and derived pointers are still sane.
2479 void PhaseChaitin::verify_base_ptrs(ResourceArea* a) const {
2480 Unique_Node_List worklist(a);
2481 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2482 Block* block = _cfg.get_block(i);
2483 for (uint j = block->end_idx() + 1; j > 1; j--) {
2484 Node* n = block->get_node(j-1);
2485 if (n->is_Phi()) {
2486 break;
2487 }
2488 // Found a safepoint?
2489 if (n->is_MachSafePoint()) {
2490 MachSafePointNode* sfpt = n->as_MachSafePoint();
2491 JVMState* jvms = sfpt->jvms();
2492 if (jvms != nullptr) {
2493 // Now scan for a live derived pointer
2494 if (jvms->oopoff() < sfpt->req()) {
2495 // Check each derived/base pair
2496 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
2497 Node* check = sfpt->in(idx);
2498 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
2499 // search upwards through spills and spill phis for AddP
2500 worklist.clear();
2501 worklist.push(check);
2502 uint k = 0;
2503 while (k < worklist.size()) {
2504 check = worklist.at(k);
2505 assert(check, "Bad base or derived pointer");
2506 // See PhaseChaitin::find_base_for_derived() for all cases.
2507 int isc = check->is_Copy();
2508 if (isc) {
2509 worklist.push(check->in(isc));
2510 } else if (check->is_Phi()) {
2511 for (uint m = 1; m < check->req(); m++) {
2512 worklist.push(check->in(m));
2513 }
2514 } else if (check->is_Con()) {
2515 if (is_derived && check->bottom_type()->is_ptr()->_offset != 0) {
2516 // Derived is null+non-zero offset, base must be null.
2517 assert(check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad derived pointer");
2518 } else {
2519 assert(check->bottom_type()->is_ptr()->_offset == 0, "Bad base pointer");
2520 // Base either ConP(nullptr) or loadConP
2521 if (check->is_Mach()) {
2522 assert(check->as_Mach()->ideal_Opcode() == Op_ConP, "Bad base pointer");
2523 } else {
2524 assert(check->Opcode() == Op_ConP &&
2525 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad base pointer");
2526 }
2527 }
2528 } else if (check->bottom_type()->is_ptr()->_offset == 0) {
2529 if (check->is_Proj() || (check->is_Mach() &&
2530 (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
2531 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
2532 check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
2533 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
2534 #ifdef _LP64
2535 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP) ||
2536 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN) ||
2537 (UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass) ||
2538 #endif // _LP64
2539 check->as_Mach()->ideal_Opcode() == Op_LoadP ||
2540 check->as_Mach()->ideal_Opcode() == Op_LoadKlass))) {
2541 // Valid nodes
2542 } else {
2543 check->dump();
2544 assert(false, "Bad base or derived pointer");
2545 }
2546 } else {
2547 assert(is_derived, "Bad base pointer");
2548 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP, "Bad derived pointer");
2549 }
2550 k++;
2551 assert(k < 100000, "Derived pointer checking in infinite loop");
2552 } // End while
2553 }
2554 } // End of check for derived pointers
2555 } // End of Kcheck for debug info
2556 } // End of if found a safepoint
2557 } // End of forall instructions in block
2558 } // End of forall blocks
2559 }
2560
2561 // Verify that graphs and base pointers are still sane.
2562 void PhaseChaitin::verify(ResourceArea* a, bool verify_ifg) const {
2563 if (VerifyRegisterAllocator) {
2564 _cfg.verify();
2565 if (C->failing()) {
2566 return;
2567 }
2568 verify_base_ptrs(a);
2569 if (verify_ifg) {
2570 _ifg->verify(this);
2571 }
2572 }
2573 }
2574 #endif // ASSERT
2575
2576 int PhaseChaitin::_final_loads = 0;
2577 int PhaseChaitin::_final_stores = 0;
2578 int PhaseChaitin::_final_memoves= 0;
2579 int PhaseChaitin::_final_copies = 0;
2580 double PhaseChaitin::_final_load_cost = 0;
2581 double PhaseChaitin::_final_store_cost = 0;
2582 double PhaseChaitin::_final_memove_cost= 0;
2583 double PhaseChaitin::_final_copy_cost = 0;
2584 int PhaseChaitin::_conserv_coalesce = 0;
2585 int PhaseChaitin::_conserv_coalesce_pair = 0;
2586 int PhaseChaitin::_conserv_coalesce_trie = 0;
2587 int PhaseChaitin::_conserv_coalesce_quad = 0;
2588 int PhaseChaitin::_post_alloc = 0;
2589 int PhaseChaitin::_lost_opp_pp_coalesce = 0;
2590 int PhaseChaitin::_lost_opp_cflow_coalesce = 0;
2591 int PhaseChaitin::_used_cisc_instructions = 0;
2592 int PhaseChaitin::_unused_cisc_instructions = 0;
2593 int PhaseChaitin::_allocator_attempts = 0;
2594 int PhaseChaitin::_allocator_successes = 0;
2595
2596 #ifndef PRODUCT
2597 uint PhaseChaitin::_high_pressure = 0;
2598 uint PhaseChaitin::_low_pressure = 0;
2599
2600 void PhaseChaitin::print_chaitin_statistics() {
2601 tty->print_cr("Inserted %d spill loads, %d spill stores, %d mem-mem moves and %d copies.", _final_loads, _final_stores, _final_memoves, _final_copies);
2602 tty->print_cr("Total load cost= %6.0f, store cost = %6.0f, mem-mem cost = %5.2f, copy cost = %5.0f.", _final_load_cost, _final_store_cost, _final_memove_cost, _final_copy_cost);
2603 tty->print_cr("Adjusted spill cost = %7.0f.",
2604 _final_load_cost*4.0 + _final_store_cost * 2.0 +
2605 _final_copy_cost*1.0 + _final_memove_cost*12.0);
2606 tty->print("Conservatively coalesced %d copies, %d pairs",
2607 _conserv_coalesce, _conserv_coalesce_pair);
2608 if( _conserv_coalesce_trie || _conserv_coalesce_quad )
2609 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad);
2610 tty->print_cr(", %d post alloc.", _post_alloc);
2611 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce )
2612 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.",
2613 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce );
2614 if( _used_cisc_instructions || _unused_cisc_instructions )
2615 tty->print_cr("Used cisc instruction %d, remained in register %d",
2616 _used_cisc_instructions, _unused_cisc_instructions);
2617 if( _allocator_successes != 0 )
2618 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes);
2619 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure);
2620 }
2621 #endif // not PRODUCT