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 C->print_method(PHASE_INITIAL_LIVENESS, 4);
410
411 // Base pointers are currently "used" by instructions which define new
412 // derived pointers. This makes base pointers live up to the where the
413 // derived pointer is made, but not beyond. Really, they need to be live
414 // across any GC point where the derived value is live. So this code looks
415 // at all the GC points, and "stretches" the live range of any base pointer
416 // to the GC point.
417 if (stretch_base_pointer_live_ranges(&live_arena)) {
418 Compile::TracePhase tp("computeLive (sbplr)", _t_computeLive);
419 // Since some live range stretched, I need to recompute live
420 _live = nullptr;
421 rm.reset_to_mark(); // Reclaim working storage
422 IndexSet::reset_memory(C, &live_arena);
423 ifg.init(_lrg_map.max_lrg_id());
424 gather_lrg_masks(false);
425 live.compute(_lrg_map.max_lrg_id());
426 _live = &live;
427 C->print_method(PHASE_LIVE_RANGE_STRETCHING, 4);
428 }
429
430 // Create the interference graph using virtual copies
431 build_ifg_virtual(); // Include stack slots this time
432 if (C->failing()) {
433 return;
434 }
435
436 // The IFG is/was triangular. I am 'squaring it up' so Union can run
437 // faster. Union requires a 'for all' operation which is slow on the
438 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' -
439 // meaning I can visit all the Nodes neighbors less than a Node in time
440 // O(# of neighbors), but I have to visit all the Nodes greater than a
441 // given Node and search them for an instance, i.e., time O(#MaxLRG)).
442 _ifg->SquareUp();
443
444 // Aggressive (but pessimistic) copy coalescing.
445 // This pass works on virtual copies. Any virtual copies which are not
446 // coalesced get manifested as actual copies
447 {
448 Compile::TracePhase tp(_t_chaitinCoalesce1);
449
450 PhaseAggressiveCoalesce coalesce(*this);
451 coalesce.coalesce_driver();
452 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do
453 // not match the Phi itself, insert a copy.
454 coalesce.insert_copies(_matcher);
455 if (C->failing()) {
456 return;
457 }
458 }
459
460 // After aggressive coalesce, attempt a first cut at coloring.
461 // To color, we need the IFG and for that we need LIVE.
462 {
463 Compile::TracePhase tp(_t_computeLive);
464 _live = nullptr;
465 rm.reset_to_mark(); // Reclaim working storage
466 IndexSet::reset_memory(C, &live_arena);
467 ifg.init(_lrg_map.max_lrg_id());
468 gather_lrg_masks( true );
469 live.compute(_lrg_map.max_lrg_id());
470 _live = &live;
471 }
472
473 C->print_method(PHASE_AGGRESSIVE_COALESCING, 4);
474
475 // Build physical interference graph
476 uint must_spill = 0;
477 must_spill = build_ifg_physical(&live_arena);
478 if (C->failing()) {
479 return;
480 }
481 // If we have a guaranteed spill, might as well spill now
482 if (must_spill) {
483 if(!_lrg_map.max_lrg_id()) {
484 return;
485 }
486 // Bail out if unique gets too large (ie - unique > MaxNodeLimit)
487 C->check_node_count(10*must_spill, "out of nodes before split");
488 if (C->failing()) {
489 return;
490 }
491
492 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
493 if (C->failing()) {
494 return;
495 }
496 _lrg_map.set_max_lrg_id(new_max_lrg_id);
497 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
498 // or we failed to split
499 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split");
500 if (C->failing()) {
501 return;
502 }
503
504 NOT_PRODUCT(C->verify_graph_edges();)
505
506 compact(); // Compact LRGs; return new lower max lrg
507
508 {
509 Compile::TracePhase tp(_t_computeLive);
510 _live = nullptr;
511 rm.reset_to_mark(); // Reclaim working storage
512 IndexSet::reset_memory(C, &live_arena);
513 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph
514 gather_lrg_masks( true ); // Collect intersect mask
515 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE
516 _live = &live;
517 }
518
519 C->print_method(PHASE_INITIAL_SPILLING, 4);
520
521 build_ifg_physical(&live_arena);
522 if (C->failing()) {
523 return;
524 }
525 _ifg->SquareUp();
526 _ifg->Compute_Effective_Degree();
527 // Only do conservative coalescing if requested
528 if (OptoCoalesce) {
529 Compile::TracePhase tp(_t_chaitinCoalesce2);
530 // Conservative (and pessimistic) copy coalescing of those spills
531 PhaseConservativeCoalesce coalesce(*this);
532 // If max live ranges greater than cutoff, don't color the stack.
533 // This cutoff can be larger than below since it is only done once.
534 coalesce.coalesce_driver();
535 }
536 _lrg_map.compress_uf_map_for_nodes();
537
538 if (OptoCoalesce) {
539 C->print_method(PHASE_CONSERVATIVE_COALESCING, 4);
540 }
541
542 #ifdef ASSERT
543 verify(&live_arena, true);
544 #endif
545 } else {
546 ifg.SquareUp();
547 ifg.Compute_Effective_Degree();
548 #ifdef ASSERT
549 set_was_low();
550 #endif
551 }
552
553 // Prepare for Simplify & Select
554 cache_lrg_info(); // Count degree of LRGs
555
556 // Simplify the InterFerence Graph by removing LRGs of low degree.
557 // LRGs of low degree are trivially colorable.
558 Simplify();
559
560 // Select colors by re-inserting LRGs back into the IFG in reverse order.
561 // Return whether or not something spills.
562 uint spills = Select( );
563 if (C->failing()) {
564 return;
565 }
566
567 // If we spill, split and recycle the entire thing
568 while( spills ) {
569 if( _trip_cnt++ > 24 ) {
570 DEBUG_ONLY( dump_for_spill_split_recycle(); )
571 if( _trip_cnt > 27 ) {
572 C->record_method_not_compilable("failed spill-split-recycle sanity check");
573 return;
574 }
575 }
576
577 if (!_lrg_map.max_lrg_id()) {
578 return;
579 }
580 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
581 if (C->failing()) {
582 return;
583 }
584 _lrg_map.set_max_lrg_id(new_max_lrg_id);
585 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
586 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split");
587 if (C->failing()) {
588 return;
589 }
590
591 compact(); // Compact LRGs; return new lower max lrg
592
593 // Nuke the live-ness and interference graph and LiveRanGe info
594 {
595 Compile::TracePhase tp(_t_computeLive);
596 _live = nullptr;
597 rm.reset_to_mark(); // Reclaim working storage
598 IndexSet::reset_memory(C, &live_arena);
599 ifg.init(_lrg_map.max_lrg_id());
600
601 // Create LiveRanGe array.
602 // Intersect register masks for all USEs and DEFs
603 gather_lrg_masks(true);
604 live.compute(_lrg_map.max_lrg_id());
605 _live = &live;
606 }
607
608 C->print_method(PHASE_ITERATIVE_SPILLING, 4);
609
610 must_spill = build_ifg_physical(&live_arena);
611 if (C->failing()) {
612 return;
613 }
614 _ifg->SquareUp();
615 _ifg->Compute_Effective_Degree();
616
617 // Only do conservative coalescing if requested
618 if (OptoCoalesce) {
619 Compile::TracePhase tp(_t_chaitinCoalesce3);
620 // Conservative (and pessimistic) copy coalescing
621 PhaseConservativeCoalesce coalesce(*this);
622 // Check for few live ranges determines how aggressive coalesce is.
623 coalesce.coalesce_driver();
624 }
625 _lrg_map.compress_uf_map_for_nodes();
626
627 if (OptoCoalesce) {
628 C->print_method(PHASE_CONSERVATIVE_COALESCING, 4);
629 }
630
631 #ifdef ASSERT
632 verify(&live_arena, true);
633 #endif
634 cache_lrg_info(); // Count degree of LRGs
635
636 // Simplify the InterFerence Graph by removing LRGs of low degree.
637 // LRGs of low degree are trivially colorable.
638 Simplify();
639
640 // Select colors by re-inserting LRGs back into the IFG in reverse order.
641 // Return whether or not something spills.
642 spills = Select();
643 if (C->failing()) {
644 return;
645 }
646 }
647
648 C->print_method(PHASE_AFTER_ITERATIVE_SPILLING, 4);
649
650 // Count number of Simplify-Select trips per coloring success.
651 _allocator_attempts += _trip_cnt + 1;
652 _allocator_successes += 1;
653
654 // Peephole remove copies
655 post_allocate_copy_removal();
656
657 C->print_method(PHASE_POST_ALLOCATION_COPY_REMOVAL, 4);
658
659 // Merge multidefs if multiple defs representing the same value are used in a single block.
660 merge_multidefs();
661
662 C->print_method(PHASE_MERGE_MULTI_DEFS, 4);
663
664 #ifdef ASSERT
665 // Verify the graph after RA.
666 verify(&live_arena);
667 #endif
668
669 // max_reg is past the largest *register* used.
670 // Convert that to a frame_slot number.
671 if (_max_reg <= _matcher._new_SP) {
672 _framesize = C->out_preserve_stack_slots();
673 }
674 else {
675 _framesize = _max_reg -_matcher._new_SP;
676 }
677 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough");
678
679 // This frame must preserve the required fp alignment
680 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots());
681 assert(_framesize <= 1000000, "sanity check");
682 #ifndef PRODUCT
683 _total_framesize += _framesize;
684 if ((int)_framesize > _max_framesize) {
685 _max_framesize = _framesize;
686 }
687 #endif
688
689 // Convert CISC spills
690 fixup_spills();
691
692 C->print_method(PHASE_FIX_UP_SPILLS, 4);
693
694 // Log regalloc results
695 CompileLog* log = Compile::current()->log();
696 if (log != nullptr) {
697 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing());
698 }
699
700 if (C->failing()) {
701 return;
702 }
703
704 NOT_PRODUCT(C->verify_graph_edges();)
705
706 // Move important info out of the live_arena to longer lasting storage.
707 alloc_node_regs(_lrg_map.size());
708 for (uint i=0; i < _lrg_map.size(); i++) {
709 if (_lrg_map.live_range_id(i)) { // Live range associated with Node?
710 LRG &lrg = lrgs(_lrg_map.live_range_id(i));
711 if (!lrg.alive()) {
712 set_bad(i);
713 } else if ((lrg.num_regs() == 1 && !lrg.is_scalable()) ||
714 (lrg.is_scalable() && lrg.scalable_reg_slots() == 1)) {
715 set1(i, lrg.reg());
716 } else { // Must be a register-set
717 if (!lrg._fat_proj) { // Must be aligned adjacent register set
718 // Live ranges record the highest register in their mask.
719 // We want the low register for the AD file writer's convenience.
720 OptoReg::Name hi = lrg.reg(); // Get hi register
721 int num_regs = lrg.num_regs();
722 if (lrg.is_scalable() && OptoReg::is_stack(hi)) {
723 // For scalable vector registers, when they are allocated in physical
724 // registers, num_regs is RegMask::SlotsPerVecA for reg mask of scalable
725 // vector. If they are allocated on stack, we need to get the actual
726 // num_regs, which reflects the physical length of scalable registers.
727 num_regs = lrg.scalable_reg_slots();
728 }
729 if (num_regs == 1) {
730 set1(i, hi);
731 } else {
732 OptoReg::Name lo = OptoReg::add(hi, (1 - num_regs)); // Find lo
733 // We have to use pair [lo,lo+1] even for wide vectors/vmasks because
734 // the rest of code generation works only with pairs. It is safe
735 // since for registers encoding only 'lo' is used.
736 // Second reg from pair is used in ScheduleAndBundle with vector max
737 // size 8 which corresponds to registers pair.
738 // It is also used in BuildOopMaps but oop operations are not
739 // vectorized.
740 set2(i, lo);
741 }
742 } else { // Misaligned; extract 2 bits
743 OptoReg::Name hi = lrg.reg(); // Get hi register
744 lrg.Remove(hi); // Yank from mask
745 int lo = lrg.mask().find_first_elem(); // Find lo
746 set_pair(i, hi, lo);
747 }
748 }
749 if( lrg._is_oop ) _node_oops.set(i);
750 } else {
751 set_bad(i);
752 }
753 }
754
755 // Done!
756 _live = nullptr;
757 _ifg = nullptr;
758 C->set_indexSet_arena(nullptr); // ResourceArea is at end of scope
759 }
760
761 void PhaseChaitin::de_ssa() {
762 // Set initial Names for all Nodes. Most Nodes get the virtual register
763 // number. A few get the ZERO live range number. These do not
764 // get allocated, but instead rely on correct scheduling to ensure that
765 // only one instance is simultaneously live at a time.
766 uint lr_counter = 1;
767 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
768 Block* block = _cfg.get_block(i);
769 uint cnt = block->number_of_nodes();
770
771 // Handle all the normal Nodes in the block
772 for( uint j = 0; j < cnt; j++ ) {
773 Node *n = block->get_node(j);
774 // Pre-color to the zero live range, or pick virtual register
775 const RegMask &rm = n->out_RegMask();
776 _lrg_map.map(n->_idx, !rm.is_Empty() ? lr_counter++ : 0);
777 }
778 }
779
780 // Reset the Union-Find mapping to be identity
781 _lrg_map.reset_uf_map(lr_counter);
782 }
783
784 void PhaseChaitin::mark_ssa() {
785 // Use ssa names to populate the live range maps or if no mask
786 // is available, use the 0 entry.
787 uint max_idx = 0;
788 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
789 Block* block = _cfg.get_block(i);
790 uint cnt = block->number_of_nodes();
791
792 // Handle all the normal Nodes in the block
793 for ( uint j = 0; j < cnt; j++ ) {
794 Node *n = block->get_node(j);
795 // Pre-color to the zero live range, or pick virtual register
796 const RegMask &rm = n->out_RegMask();
797 _lrg_map.map(n->_idx, !rm.is_Empty() ? n->_idx : 0);
798 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx;
799 }
800 }
801 _lrg_map.set_max_lrg_id(max_idx+1);
802
803 // Reset the Union-Find mapping to be identity
804 _lrg_map.reset_uf_map(max_idx+1);
805 }
806
807
808 // Gather LiveRanGe information, including register masks. Modification of
809 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce.
810 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) {
811
812 // Nail down the frame pointer live range
813 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr));
814 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite
815
816 // For all blocks
817 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
818 Block* block = _cfg.get_block(i);
819
820 // For all instructions
821 for (uint j = 1; j < block->number_of_nodes(); j++) {
822 Node* n = block->get_node(j);
823 uint input_edge_start =1; // Skip control most nodes
824 bool is_machine_node = false;
825 if (n->is_Mach()) {
826 is_machine_node = true;
827 input_edge_start = n->as_Mach()->oper_input_base();
828 }
829 uint idx = n->is_Copy();
830
831 // Get virtual register number, same as LiveRanGe index
832 uint vreg = _lrg_map.live_range_id(n);
833 LRG& lrg = lrgs(vreg);
834 if (vreg) { // No vreg means un-allocable (e.g. memory)
835
836 // Check for float-vs-int live range (used in register-pressure
837 // calculations)
838 const Type *n_type = n->bottom_type();
839 if (n_type->is_floatingpoint()) {
840 lrg._is_float = 1;
841 }
842
843 // Check for twice prior spilling. Once prior spilling might have
844 // spilled 'soft', 2nd prior spill should have spilled 'hard' and
845 // further spilling is unlikely to make progress.
846 if (_spilled_once.test(n->_idx)) {
847 lrg._was_spilled1 = 1;
848 if (_spilled_twice.test(n->_idx)) {
849 lrg._was_spilled2 = 1;
850 }
851 }
852
853 #ifndef PRODUCT
854 // Collect bits not used by product code, but which may be useful for
855 // debugging.
856
857 // Collect has-copy bit
858 if (idx) {
859 lrg._has_copy = 1;
860 uint clidx = _lrg_map.live_range_id(n->in(idx));
861 LRG& copy_src = lrgs(clidx);
862 copy_src._has_copy = 1;
863 }
864
865 if (trace_spilling() && lrg._def != nullptr) {
866 // collect defs for MultiDef printing
867 if (lrg._defs == nullptr) {
868 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, nullptr);
869 lrg._defs->append(lrg._def);
870 }
871 lrg._defs->append(n);
872 }
873 #endif
874
875 // Check for a single def LRG; these can spill nicely
876 // via rematerialization. Flag as null for no def found
877 // yet, or 'n' for single def or -1 for many defs.
878 lrg._def = lrg._def ? NodeSentinel : n;
879
880 // Limit result register mask to acceptable registers
881 const RegMask &rm = n->out_RegMask();
882 lrg.AND( rm );
883
884 uint ireg = n->ideal_reg();
885 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP,
886 "oops must be in Op_RegP's" );
887
888 // Check for vector live range (only if vector register is used).
889 // On SPARC vector uses RegD which could be misaligned so it is not
890 // processes as vector in RA.
891 if (RegMask::is_vector(ireg)) {
892 lrg._is_vector = 1;
893 if (Matcher::implements_scalable_vector && ireg == Op_VecA) {
894 assert(Matcher::supports_scalable_vector(), "scalable vector should be supported");
895 lrg._is_scalable = 1;
896 // For scalable vector, when it is allocated in physical register,
897 // num_regs is RegMask::SlotsPerVecA for reg mask,
898 // which may not be the actual physical register size.
899 // If it is allocated in stack, we need to get the actual
900 // physical length of scalable vector register.
901 lrg.set_scalable_reg_slots(Matcher::scalable_vector_reg_size(T_FLOAT));
902 }
903 }
904
905 if (ireg == Op_RegVectMask) {
906 assert(Matcher::has_predicated_vectors(), "predicated vector should be supported");
907 lrg._is_predicate = 1;
908 if (Matcher::supports_scalable_vector()) {
909 lrg._is_scalable = 1;
910 // For scalable predicate, when it is allocated in physical register,
911 // num_regs is RegMask::SlotsPerRegVectMask for reg mask,
912 // which may not be the actual physical register size.
913 // If it is allocated in stack, we need to get the actual
914 // physical length of scalable predicate register.
915 lrg.set_scalable_reg_slots(Matcher::scalable_predicate_reg_slots());
916 }
917 }
918 assert(n_type->isa_vect() == nullptr || lrg._is_vector ||
919 ireg == Op_RegD || ireg == Op_RegL || ireg == Op_RegVectMask,
920 "vector must be in vector registers");
921
922 // Check for bound register masks
923 const RegMask &lrgmask = lrg.mask();
924 if (lrgmask.is_bound(ireg)) {
925 lrg._is_bound = 1;
926 }
927
928 // Check for maximum frequency value
929 if (lrg._maxfreq < block->_freq) {
930 lrg._maxfreq = block->_freq;
931 }
932
933 // Check for oop-iness, or long/double
934 // Check for multi-kill projection
935 switch (ireg) {
936 case MachProjNode::fat_proj:
937 // Fat projections have size equal to number of registers killed
938 lrg.set_num_regs(rm.Size());
939 lrg.set_reg_pressure(lrg.num_regs());
940 lrg._fat_proj = 1;
941 lrg._is_bound = 1;
942 break;
943 case Op_RegP:
944 #ifdef _LP64
945 lrg.set_num_regs(2); // Size is 2 stack words
946 #else
947 lrg.set_num_regs(1); // Size is 1 stack word
948 #endif
949 // Register pressure is tracked relative to the maximum values
950 // suggested for that platform, INTPRESSURE and FLOATPRESSURE,
951 // and relative to other types which compete for the same regs.
952 //
953 // The following table contains suggested values based on the
954 // architectures as defined in each .ad file.
955 // INTPRESSURE and FLOATPRESSURE may be tuned differently for
956 // compile-speed or performance.
957 // Note1:
958 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1
959 // since .ad registers are defined as high and low halves.
960 // These reg_pressure values remain compatible with the code
961 // in is_high_pressure() which relates get_invalid_mask_size(),
962 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE.
963 // Note2:
964 // SPARC -d32 has 24 registers available for integral values,
965 // but only 10 of these are safe for 64-bit longs.
966 // Using set_reg_pressure(2) for both int and long means
967 // the allocator will believe it can fit 26 longs into
968 // registers. Using 2 for longs and 1 for ints means the
969 // allocator will attempt to put 52 integers into registers.
970 // The settings below limit this problem to methods with
971 // many long values which are being run on 32-bit SPARC.
972 //
973 // ------------------- reg_pressure --------------------
974 // Each entry is reg_pressure_per_value,number_of_regs
975 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE
976 // IA32 2 1 1 1 1 6 6
977 // SPARC 2 2 2 2 2 48 (24) 52 (26)
978 // SPARCV9 2 2 2 2 2 48 (24) 52 (26)
979 // AMD64 1 1 1 1 1 14 15
980 // -----------------------------------------------------
981 lrg.set_reg_pressure(1); // normally one value per register
982 if( n_type->isa_oop_ptr() ) {
983 lrg._is_oop = 1;
984 }
985 break;
986 case Op_RegL: // Check for long or double
987 case Op_RegD:
988 lrg.set_num_regs(2);
989 // Define platform specific register pressure
990 #if defined(ARM32)
991 lrg.set_reg_pressure(2);
992 #elif defined(IA32)
993 if( ireg == Op_RegL ) {
994 lrg.set_reg_pressure(2);
995 } else {
996 lrg.set_reg_pressure(1);
997 }
998 #else
999 lrg.set_reg_pressure(1); // normally one value per register
1000 #endif
1001 // If this def of a double forces a mis-aligned double,
1002 // flag as '_fat_proj' - really flag as allowing misalignment
1003 // AND changes how we count interferences. A mis-aligned
1004 // double can interfere with TWO aligned pairs, or effectively
1005 // FOUR registers!
1006 if (rm.is_misaligned_pair()) {
1007 lrg._fat_proj = 1;
1008 lrg._is_bound = 1;
1009 }
1010 break;
1011 case Op_RegVectMask:
1012 assert(Matcher::has_predicated_vectors(), "sanity");
1013 assert(RegMask::num_registers(Op_RegVectMask) == RegMask::SlotsPerRegVectMask, "sanity");
1014 lrg.set_num_regs(RegMask::SlotsPerRegVectMask);
1015 lrg.set_reg_pressure(1);
1016 break;
1017 case Op_RegF:
1018 case Op_RegI:
1019 case Op_RegN:
1020 case Op_RegFlags:
1021 case 0: // not an ideal register
1022 lrg.set_num_regs(1);
1023 lrg.set_reg_pressure(1);
1024 break;
1025 case Op_VecA:
1026 assert(Matcher::supports_scalable_vector(), "does not support scalable vector");
1027 assert(RegMask::num_registers(Op_VecA) == RegMask::SlotsPerVecA, "sanity");
1028 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecA), "vector should be aligned");
1029 lrg.set_num_regs(RegMask::SlotsPerVecA);
1030 lrg.set_reg_pressure(1);
1031 break;
1032 case Op_VecS:
1033 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity");
1034 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity");
1035 lrg.set_num_regs(RegMask::SlotsPerVecS);
1036 lrg.set_reg_pressure(1);
1037 break;
1038 case Op_VecD:
1039 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity");
1040 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity");
1041 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned");
1042 lrg.set_num_regs(RegMask::SlotsPerVecD);
1043 lrg.set_reg_pressure(1);
1044 break;
1045 case Op_VecX:
1046 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity");
1047 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity");
1048 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned");
1049 lrg.set_num_regs(RegMask::SlotsPerVecX);
1050 lrg.set_reg_pressure(1);
1051 break;
1052 case Op_VecY:
1053 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity");
1054 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity");
1055 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned");
1056 lrg.set_num_regs(RegMask::SlotsPerVecY);
1057 lrg.set_reg_pressure(1);
1058 break;
1059 case Op_VecZ:
1060 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity");
1061 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity");
1062 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned");
1063 lrg.set_num_regs(RegMask::SlotsPerVecZ);
1064 lrg.set_reg_pressure(1);
1065 break;
1066 default:
1067 ShouldNotReachHere();
1068 }
1069 }
1070
1071 // Now do the same for inputs
1072 uint cnt = n->req();
1073 // Setup for CISC SPILLING
1074 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable;
1075 if( UseCISCSpill && after_aggressive ) {
1076 inp = n->cisc_operand();
1077 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable )
1078 // Convert operand number to edge index number
1079 inp = n->as_Mach()->operand_index(inp);
1080 }
1081
1082 // Prepare register mask for each input
1083 for( uint k = input_edge_start; k < cnt; k++ ) {
1084 uint vreg = _lrg_map.live_range_id(n->in(k));
1085 if (!vreg) {
1086 continue;
1087 }
1088
1089 // If this instruction is CISC Spillable, add the flags
1090 // bit to its appropriate input
1091 if( UseCISCSpill && after_aggressive && inp == k ) {
1092 #ifndef PRODUCT
1093 if( TraceCISCSpill ) {
1094 tty->print(" use_cisc_RegMask: ");
1095 n->dump();
1096 }
1097 #endif
1098 n->as_Mach()->use_cisc_RegMask();
1099 }
1100
1101 if (is_machine_node && _scheduling_info_generated) {
1102 MachNode* cur_node = n->as_Mach();
1103 // this is cleaned up by register allocation
1104 if (k >= cur_node->num_opnds()) continue;
1105 }
1106
1107 LRG &lrg = lrgs(vreg);
1108 // // Testing for floating point code shape
1109 // Node *test = n->in(k);
1110 // if( test->is_Mach() ) {
1111 // MachNode *m = test->as_Mach();
1112 // int op = m->ideal_Opcode();
1113 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) {
1114 // int zzz = 1;
1115 // }
1116 // }
1117
1118 // Limit result register mask to acceptable registers.
1119 // Do not limit registers from uncommon uses before
1120 // AggressiveCoalesce. This effectively pre-virtual-splits
1121 // around uncommon uses of common defs.
1122 const RegMask &rm = n->in_RegMask(k);
1123 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) {
1124 // Since we are BEFORE aggressive coalesce, leave the register
1125 // mask untrimmed by the call. This encourages more coalescing.
1126 // Later, AFTER aggressive, this live range will have to spill
1127 // but the spiller handles slow-path calls very nicely.
1128 } else {
1129 lrg.AND( rm );
1130 }
1131
1132 // Check for bound register masks
1133 const RegMask &lrgmask = lrg.mask();
1134 uint kreg = n->in(k)->ideal_reg();
1135 bool is_vect = RegMask::is_vector(kreg);
1136 assert(n->in(k)->bottom_type()->isa_vect() == nullptr || is_vect ||
1137 kreg == Op_RegD || kreg == Op_RegL || kreg == Op_RegVectMask,
1138 "vector must be in vector registers");
1139 if (lrgmask.is_bound(kreg))
1140 lrg._is_bound = 1;
1141
1142 // If this use of a double forces a mis-aligned double,
1143 // flag as '_fat_proj' - really flag as allowing misalignment
1144 // AND changes how we count interferences. A mis-aligned
1145 // double can interfere with TWO aligned pairs, or effectively
1146 // FOUR registers!
1147 #ifdef ASSERT
1148 if (is_vect && !_scheduling_info_generated) {
1149 if (lrg.num_regs() != 0) {
1150 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned");
1151 assert(!lrg._fat_proj, "sanity");
1152 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity");
1153 } else {
1154 assert(n->is_Phi(), "not all inputs processed only if Phi");
1155 }
1156 }
1157 #endif
1158 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) {
1159 lrg._fat_proj = 1;
1160 lrg._is_bound = 1;
1161 }
1162 // if the LRG is an unaligned pair, we will have to spill
1163 // so clear the LRG's register mask if it is not already spilled
1164 if (!is_vect && !n->is_SpillCopy() &&
1165 (lrg._def == nullptr || lrg.is_multidef() || !lrg._def->is_SpillCopy()) &&
1166 lrgmask.is_misaligned_pair()) {
1167 lrg.Clear();
1168 }
1169
1170 // Check for maximum frequency value
1171 if (lrg._maxfreq < block->_freq) {
1172 lrg._maxfreq = block->_freq;
1173 }
1174
1175 } // End for all allocated inputs
1176 } // end for all instructions
1177 } // end for all blocks
1178
1179 // Final per-liverange setup
1180 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) {
1181 LRG &lrg = lrgs(i2);
1182 assert(!lrg._is_vector || !lrg._fat_proj, "sanity");
1183 if (lrg.num_regs() > 1 && !lrg._fat_proj) {
1184 lrg.clear_to_sets();
1185 }
1186 lrg.compute_set_mask_size();
1187 if (lrg.not_free()) { // Handle case where we lose from the start
1188 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG));
1189 lrg._direct_conflict = 1;
1190 }
1191 lrg.set_degree(0); // no neighbors in IFG yet
1192 }
1193 }
1194
1195 // Set the was-lo-degree bit. Conservative coalescing should not change the
1196 // colorability of the graph. If any live range was of low-degree before
1197 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
1198 // The bit is checked in Simplify.
1199 void PhaseChaitin::set_was_low() {
1200 #ifdef ASSERT
1201 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1202 int size = lrgs(i).num_regs();
1203 uint old_was_lo = lrgs(i)._was_lo;
1204 lrgs(i)._was_lo = 0;
1205 if( lrgs(i).lo_degree() ) {
1206 lrgs(i)._was_lo = 1; // Trivially of low degree
1207 } else { // Else check the Brigg's assertion
1208 // Brigg's observation is that the lo-degree neighbors of a
1209 // hi-degree live range will not interfere with the color choices
1210 // of said hi-degree live range. The Simplify reverse-stack-coloring
1211 // order takes care of the details. Hence you do not have to count
1212 // low-degree neighbors when determining if this guy colors.
1213 int briggs_degree = 0;
1214 IndexSet *s = _ifg->neighbors(i);
1215 IndexSetIterator elements(s);
1216 uint lidx;
1217 while((lidx = elements.next()) != 0) {
1218 if( !lrgs(lidx).lo_degree() )
1219 briggs_degree += MAX2(size,lrgs(lidx).num_regs());
1220 }
1221 if( briggs_degree < lrgs(i).degrees_of_freedom() )
1222 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion
1223 }
1224 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease");
1225 }
1226 #endif
1227 }
1228
1229 // Compute cost/area ratio, in case we spill. Build the lo-degree list.
1230 void PhaseChaitin::cache_lrg_info( ) {
1231 Compile::TracePhase tp(_t_chaitinCacheLRG);
1232
1233 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1234 LRG &lrg = lrgs(i);
1235
1236 // Check for being of low degree: means we can be trivially colored.
1237 // Low degree, dead or must-spill guys just get to simplify right away
1238 if( lrg.lo_degree() ||
1239 !lrg.alive() ||
1240 lrg._must_spill ) {
1241 // Split low degree list into those guys that must get a
1242 // register and those that can go to register or stack.
1243 // The idea is LRGs that can go register or stack color first when
1244 // they have a good chance of getting a register. The register-only
1245 // lo-degree live ranges always get a register.
1246 OptoReg::Name hi_reg = lrg.mask().find_last_elem();
1247 if( OptoReg::is_stack(hi_reg)) { // Can go to stack?
1248 lrg._next = _lo_stk_degree;
1249 _lo_stk_degree = i;
1250 } else {
1251 lrg._next = _lo_degree;
1252 _lo_degree = i;
1253 }
1254 } else { // Else high degree
1255 lrgs(_hi_degree)._prev = i;
1256 lrg._next = _hi_degree;
1257 lrg._prev = 0;
1258 _hi_degree = i;
1259 }
1260 }
1261 }
1262
1263 // Simplify the IFG by removing LRGs of low degree.
1264 void PhaseChaitin::Simplify( ) {
1265 Compile::TracePhase tp(_t_chaitinSimplify);
1266
1267 while( 1 ) { // Repeat till simplified it all
1268 // May want to explore simplifying lo_degree before _lo_stk_degree.
1269 // This might result in more spills coloring into registers during
1270 // Select().
1271 while( _lo_degree || _lo_stk_degree ) {
1272 // If possible, pull from lo_stk first
1273 uint lo;
1274 if( _lo_degree ) {
1275 lo = _lo_degree;
1276 _lo_degree = lrgs(lo)._next;
1277 } else {
1278 lo = _lo_stk_degree;
1279 _lo_stk_degree = lrgs(lo)._next;
1280 }
1281
1282 // Put the simplified guy on the simplified list.
1283 lrgs(lo)._next = _simplified;
1284 _simplified = lo;
1285 // If this guy is "at risk" then mark his current neighbors
1286 if (lrgs(lo)._at_risk && !_ifg->neighbors(lo)->is_empty()) {
1287 IndexSetIterator elements(_ifg->neighbors(lo));
1288 uint datum;
1289 while ((datum = elements.next()) != 0) {
1290 lrgs(datum)._risk_bias = lo;
1291 }
1292 }
1293
1294 // Yank this guy from the IFG.
1295 IndexSet *adj = _ifg->remove_node(lo);
1296 if (adj->is_empty()) {
1297 continue;
1298 }
1299
1300 // If any neighbors' degrees fall below their number of
1301 // allowed registers, then put that neighbor on the low degree
1302 // list. Note that 'degree' can only fall and 'numregs' is
1303 // unchanged by this action. Thus the two are equal at most once,
1304 // so LRGs hit the lo-degree worklist at most once.
1305 IndexSetIterator elements(adj);
1306 uint neighbor;
1307 while ((neighbor = elements.next()) != 0) {
1308 LRG *n = &lrgs(neighbor);
1309 #ifdef ASSERT
1310 if (VerifyRegisterAllocator) {
1311 assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
1312 }
1313 #endif
1314
1315 // Check for just becoming of-low-degree just counting registers.
1316 // _must_spill live ranges are already on the low degree list.
1317 if (n->just_lo_degree() && !n->_must_spill) {
1318 assert(!_ifg->_yanked->test(neighbor), "Cannot move to lo degree twice");
1319 // Pull from hi-degree list
1320 uint prev = n->_prev;
1321 uint next = n->_next;
1322 if (prev) {
1323 lrgs(prev)._next = next;
1324 } else {
1325 _hi_degree = next;
1326 }
1327 lrgs(next)._prev = prev;
1328 n->_next = _lo_degree;
1329 _lo_degree = neighbor;
1330 }
1331 }
1332 } // End of while lo-degree/lo_stk_degree worklist not empty
1333
1334 // Check for got everything: is hi-degree list empty?
1335 if (!_hi_degree) break;
1336
1337 // Time to pick a potential spill guy
1338 uint lo_score = _hi_degree;
1339 double score = lrgs(lo_score).score();
1340 double area = lrgs(lo_score)._area;
1341 double cost = lrgs(lo_score)._cost;
1342 bool bound = lrgs(lo_score)._is_bound;
1343
1344 // Find cheapest guy
1345 DEBUG_ONLY( int lo_no_simplify=0; );
1346 for (uint i = _hi_degree; i; i = lrgs(i)._next) {
1347 assert(!_ifg->_yanked->test(i), "");
1348 // It's just vaguely possible to move hi-degree to lo-degree without
1349 // going through a just-lo-degree stage: If you remove a double from
1350 // a float live range it's degree will drop by 2 and you can skip the
1351 // just-lo-degree stage. It's very rare (shows up after 5000+ methods
1352 // in -Xcomp of Java2Demo). So just choose this guy to simplify next.
1353 if( lrgs(i).lo_degree() ) {
1354 lo_score = i;
1355 break;
1356 }
1357 DEBUG_ONLY( if( lrgs(i)._was_lo ) lo_no_simplify=i; );
1358 double iscore = lrgs(i).score();
1359 double iarea = lrgs(i)._area;
1360 double icost = lrgs(i)._cost;
1361 bool ibound = lrgs(i)._is_bound;
1362
1363 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area
1364 // wins. Ties happen because all live ranges in question have spilled
1365 // a few times before and the spill-score adds a huge number which
1366 // washes out the low order bits. We are choosing the lesser of 2
1367 // evils; in this case pick largest area to spill.
1368 // Ties also happen when live ranges are defined and used only inside
1369 // one block. In which case their area is 0 and score set to max.
1370 // In such case choose bound live range over unbound to free registers
1371 // or with smaller cost to spill.
1372 if ( iscore < score ||
1373 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ||
1374 (iscore == score && iarea == area &&
1375 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) {
1376 lo_score = i;
1377 score = iscore;
1378 area = iarea;
1379 cost = icost;
1380 bound = ibound;
1381 }
1382 }
1383 LRG *lo_lrg = &lrgs(lo_score);
1384 // The live range we choose for spilling is either hi-degree, or very
1385 // rarely it can be low-degree. If we choose a hi-degree live range
1386 // there better not be any lo-degree choices.
1387 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" );
1388
1389 // Pull from hi-degree list
1390 uint prev = lo_lrg->_prev;
1391 uint next = lo_lrg->_next;
1392 if( prev ) lrgs(prev)._next = next;
1393 else _hi_degree = next;
1394 lrgs(next)._prev = prev;
1395 // Jam him on the lo-degree list, despite his high degree.
1396 // Maybe he'll get a color, and maybe he'll spill.
1397 // Only Select() will know.
1398 lrgs(lo_score)._at_risk = true;
1399 _lo_degree = lo_score;
1400 lo_lrg->_next = 0;
1401
1402 } // End of while not simplified everything
1403
1404 }
1405
1406 // Is 'reg' register legal for 'lrg'?
1407 static bool is_legal_reg(LRG& lrg, OptoReg::Name reg) {
1408 if (lrg.mask().can_represent(reg) && lrg.mask().Member(reg)) {
1409 // RA uses OptoReg which represent the highest element of a registers set.
1410 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set
1411 // in which XMMd is used by RA to represent such vectors. A double value
1412 // uses [XMM,XMMb] pairs and XMMb is used by RA for it.
1413 // The register mask uses largest bits set of overlapping register sets.
1414 // On x86 with AVX it uses 8 bits for each XMM registers set.
1415 //
1416 // The 'lrg' already has cleared-to-set register mask (done in Select()
1417 // before calling choose_color()). Passing mask.Member(reg) check above
1418 // indicates that the size (num_regs) of 'reg' set is less or equal to
1419 // 'lrg' set size.
1420 // For set size 1 any register which is member of 'lrg' mask is legal.
1421 if (lrg.num_regs()==1)
1422 return true;
1423 // For larger sets only an aligned register with the same set size is legal.
1424 int mask = lrg.num_regs()-1;
1425 if ((reg&mask) == mask)
1426 return true;
1427 }
1428 return false;
1429 }
1430
1431 static OptoReg::Name find_first_set(LRG& lrg, RegMask& mask) {
1432 int num_regs = lrg.num_regs();
1433 OptoReg::Name assigned = mask.find_first_set(lrg, num_regs);
1434
1435 if (lrg.is_scalable()) {
1436 // a physical register is found
1437 if (OptoReg::is_reg(assigned)) {
1438 assert(!lrg.mask().is_offset(),
1439 "offset register masks can only contain stack slots");
1440 return assigned;
1441 }
1442
1443 // find available stack slots for scalable register
1444 if (lrg._is_vector) {
1445 num_regs = lrg.scalable_reg_slots();
1446 // if actual scalable vector register is exactly SlotsPerVecA * 32 bits
1447 if (num_regs == RegMask::SlotsPerVecA) {
1448 return assigned;
1449 }
1450
1451 // mask has been cleared out by clear_to_sets(SlotsPerVecA) before choose_color, but it
1452 // does not work for scalable size. We have to find adjacent scalable_reg_slots() bits
1453 // instead of SlotsPerVecA bits.
1454 assigned = mask.find_first_set(lrg, num_regs); // find highest valid reg
1455 while (OptoReg::is_valid(assigned)) {
1456 assert(mask.can_represent(assigned), "sanity");
1457 // Verify the found reg has scalable_reg_slots() bits set.
1458 if (mask.is_valid_reg(assigned, num_regs)) {
1459 return assigned;
1460 } else {
1461 // Remove more for each iteration
1462 mask.Remove(assigned - num_regs + 1); // Unmask the lowest reg
1463 mask.clear_to_sets(RegMask::SlotsPerVecA); // Align by SlotsPerVecA bits
1464 assigned = mask.find_first_set(lrg, num_regs);
1465 }
1466 }
1467 return OptoReg::Bad; // will cause chunk change, and retry next chunk
1468 } else if (lrg._is_predicate) {
1469 assert(num_regs == RegMask::SlotsPerRegVectMask, "scalable predicate register");
1470 num_regs = lrg.scalable_reg_slots();
1471 mask.clear_to_sets(num_regs);
1472 return mask.find_first_set(lrg, num_regs);
1473 }
1474 }
1475
1476 return assigned;
1477 }
1478
1479 // Choose a color using the biasing heuristic
1480 OptoReg::Name PhaseChaitin::bias_color(LRG& lrg) {
1481
1482 // Check for "at_risk" LRG's
1483 uint risk_lrg = _lrg_map.find(lrg._risk_bias);
1484 if (risk_lrg != 0 && !_ifg->neighbors(risk_lrg)->is_empty()) {
1485 // Walk the colored neighbors of the "at_risk" candidate
1486 // Choose a color which is both legal and already taken by a neighbor
1487 // of the "at_risk" candidate in order to improve the chances of the
1488 // "at_risk" candidate of coloring
1489 IndexSetIterator elements(_ifg->neighbors(risk_lrg));
1490 uint datum;
1491 while ((datum = elements.next()) != 0) {
1492 OptoReg::Name reg = lrgs(datum).reg();
1493 // If this LRG's register is legal for us, choose it
1494 if (is_legal_reg(lrg, reg)) {
1495 return reg;
1496 }
1497 }
1498 }
1499
1500 uint copy_lrg = _lrg_map.find(lrg._copy_bias);
1501 if (copy_lrg != 0) {
1502 // If he has a color,
1503 if(!_ifg->_yanked->test(copy_lrg)) {
1504 OptoReg::Name reg = lrgs(copy_lrg).reg();
1505 // And it is legal for you,
1506 if (is_legal_reg(lrg, reg)) {
1507 return reg;
1508 }
1509 } else if (!lrg.mask().is_offset()) {
1510 // Choose a color which is legal for him
1511 ResourceMark rm(C->regmask_arena());
1512 RegMask tempmask(lrg.mask(), C->regmask_arena());
1513 tempmask.AND(lrgs(copy_lrg).mask());
1514 tempmask.clear_to_sets(lrg.num_regs());
1515 OptoReg::Name reg = find_first_set(lrg, tempmask);
1516 if (OptoReg::is_valid(reg))
1517 return reg;
1518 }
1519 }
1520
1521 // If no bias info exists, just go with the register selection ordering
1522 if (lrg._is_vector || lrg.num_regs() == 2 || lrg.is_scalable()) {
1523 // Find an aligned set
1524 ResourceMark rm(C->regmask_arena());
1525 RegMask tempmask(lrg.mask(), C->regmask_arena());
1526 return find_first_set(lrg, tempmask);
1527 }
1528
1529 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate
1530 // copy removal to remove many more copies, by preventing a just-assigned
1531 // register from being repeatedly assigned.
1532 OptoReg::Name reg = lrg.mask().find_first_elem();
1533 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) {
1534 // This 'Remove; find; Insert' idiom is an expensive way to find the
1535 // SECOND element in the mask.
1536 lrg.Remove(reg);
1537 OptoReg::Name reg2 = lrg.mask().find_first_elem();
1538 lrg.Insert(reg);
1539 if (OptoReg::is_reg(reg2)) {
1540 reg = reg2;
1541 }
1542 }
1543 return reg;
1544 }
1545
1546 // Choose a color in the current chunk
1547 OptoReg::Name PhaseChaitin::choose_color(LRG& lrg) {
1548 assert(C->in_preserve_stack_slots() == 0 || lrg.mask().is_offset() || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP - 1)), "must not allocate stack0 (inside preserve area)");
1549 assert(C->out_preserve_stack_slots() == 0 || lrg.mask().is_offset() || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP + 0)), "must not allocate stack0 (inside preserve area)");
1550
1551 if( lrg.num_regs() == 1 || // Common Case
1552 !lrg._fat_proj ) // Aligned+adjacent pairs ok
1553 // Use a heuristic to "bias" the color choice
1554 return bias_color(lrg);
1555
1556 assert(!lrg._is_vector, "should be not vector here" );
1557 assert( lrg.num_regs() >= 2, "dead live ranges do not color" );
1558
1559 // Fat-proj case or misaligned double argument.
1560 assert(lrg.compute_mask_size() == lrg.num_regs() ||
1561 lrg.num_regs() == 2,"fat projs exactly color" );
1562 assert(!lrg.mask().is_offset(), "always color in 1st chunk");
1563 // Return the highest element in the set.
1564 return lrg.mask().find_last_elem();
1565 }
1566
1567 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted
1568 // in reverse order of removal. As long as nothing of hi-degree was yanked,
1569 // everything going back is guaranteed a color. Select that color. If some
1570 // hi-degree LRG cannot get a color then we record that we must spill.
1571 uint PhaseChaitin::Select( ) {
1572 Compile::TracePhase tp(_t_chaitinSelect);
1573
1574 uint spill_reg = LRG::SPILL_REG;
1575 _max_reg = OptoReg::Name(0); // Past max register used
1576 while( _simplified ) {
1577 // Pull next LRG from the simplified list - in reverse order of removal
1578 uint lidx = _simplified;
1579 LRG *lrg = &lrgs(lidx);
1580 _simplified = lrg->_next;
1581
1582 #ifndef PRODUCT
1583 if (trace_spilling()) {
1584 ttyLocker ttyl;
1585 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(),
1586 lrg->degrees_of_freedom());
1587 lrg->dump();
1588 }
1589 #endif
1590
1591 // Re-insert into the IFG
1592 _ifg->re_insert(lidx);
1593 if( !lrg->alive() ) continue;
1594 // capture infinitestackedness flag before mask is hacked
1595 const int is_infinite_stack = lrg->mask().is_infinite_stack();
1596
1597 // Yeah, yeah, yeah, I know, I know. I can refactor this
1598 // to avoid the GOTO, although the refactored code will not
1599 // be much clearer. We arrive here IFF we have a stack-based
1600 // live range that cannot color in the current chunk, and it
1601 // has to move into the next free stack chunk.
1602 retry_next_chunk:
1603
1604 // Remove neighbor colors
1605 IndexSet *s = _ifg->neighbors(lidx);
1606 #ifndef PRODUCT
1607 ResourceMark rm(C->regmask_arena());
1608 RegMask orig_mask(lrg->mask(), C->regmask_arena());
1609 #endif
1610
1611 if (!s->is_empty()) {
1612 IndexSetIterator elements(s);
1613 uint neighbor;
1614 while ((neighbor = elements.next()) != 0) {
1615 LRG &nlrg = lrgs(neighbor);
1616 OptoReg::Name nreg = nlrg.reg();
1617 // The neighbor might be a spill_reg. In this case, exclusion of its
1618 // color will be a no-op, since the spill_reg is in outer space. In
1619 // this case, do not exclude the corresponding mask. Later on, if lrg
1620 // runs out of possible colors in its chunk, a new chunk of color may
1621 // be tried, in which case examination of neighbors is started again,
1622 // at retry_next_chunk.
1623 if (nreg < LRG::SPILL_REG) {
1624 #ifndef PRODUCT
1625 uint size = lrg->mask().Size();
1626 ResourceMark rm(C->regmask_arena());
1627 RegMask trace_mask(lrg->mask(), C->regmask_arena());
1628 #endif
1629 lrg->SUBTRACT_inner(nlrg.mask());
1630 #ifndef PRODUCT
1631 if (trace_spilling() && lrg->mask().Size() != size) {
1632 ttyLocker ttyl;
1633 tty->print("L%d ", lidx);
1634 trace_mask.dump();
1635 tty->print(" intersected L%d ", neighbor);
1636 nlrg.mask().dump();
1637 tty->print(" removed ");
1638 trace_mask.SUBTRACT(lrg->mask());
1639 trace_mask.dump();
1640 tty->print(" leaving ");
1641 lrg->mask().dump();
1642 tty->cr();
1643 }
1644 #endif
1645 }
1646 }
1647 }
1648 //assert(is_infinite_stack == lrg->mask().is_infinite_stack(), "nbrs must not change InfiniteStackedness");
1649 // Aligned pairs need aligned masks
1650 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1651 if (lrg->num_regs() > 1 && !lrg->_fat_proj) {
1652 lrg->clear_to_sets();
1653 }
1654
1655 // Check if a color is available and if so pick the color
1656 OptoReg::Name reg = choose_color(*lrg);
1657
1658 //---------------
1659 // If we fail to color and the infinite flag is set, we must trigger
1660 // a chunk-rollover event and continue searching for a color in the next set
1661 // of slots (which are all necessarily stack slots, as registers are only in
1662 // the initial chunk)
1663 if (!OptoReg::is_valid(reg) && is_infinite_stack) {
1664 // Bump register mask up to next stack chunk
1665 bool success = lrg->rollover();
1666 if (!success) {
1667 // We should never get here in practice. Bail out in product,
1668 // assert in debug.
1669 assert(false, "the next available stack slots should be within the "
1670 "OptoRegPair range");
1671 C->record_method_not_compilable(
1672 "chunk-rollover outside of OptoRegPair range");
1673 return -1;
1674 }
1675 goto retry_next_chunk;
1676 }
1677
1678 //---------------
1679 // Did we get a color?
1680 else if (OptoReg::is_valid(reg)) {
1681 #ifndef PRODUCT
1682 ResourceMark rm(C->regmask_arena());
1683 RegMask avail_rm(lrg->mask(), C->regmask_arena());
1684 #endif
1685
1686 // Record selected register
1687 lrg->set_reg(reg);
1688
1689 if (reg >= _max_reg) { // Compute max register limit
1690 _max_reg = OptoReg::add(reg, 1);
1691 }
1692
1693 // If the live range is not bound, then we actually had some choices
1694 // to make. In this case, the mask has more bits in it than the colors
1695 // chosen. Restrict the mask to just what was picked.
1696 int n_regs = lrg->num_regs();
1697 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1698 if (n_regs == 1 || !lrg->_fat_proj) {
1699 if (Matcher::supports_scalable_vector()) {
1700 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecA, "sanity");
1701 } else {
1702 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity");
1703 }
1704 lrg->Clear(); // Clear the mask
1705 lrg->Insert(reg); // Set regmask to match selected reg
1706 // For vectors and pairs, also insert the low bit of the pair
1707 // We always choose the high bit, then mask the low bits by register size
1708 if (lrg->is_scalable() && OptoReg::is_stack(lrg->reg())) { // stack
1709 n_regs = lrg->scalable_reg_slots();
1710 }
1711 for (int i = 1; i < n_regs; i++) {
1712 lrg->Insert(OptoReg::add(reg,-i));
1713 }
1714 lrg->set_mask_size(n_regs);
1715 } else { // Else fatproj
1716 // mask must be equal to fatproj bits, by definition
1717 }
1718 #ifndef PRODUCT
1719 if (trace_spilling()) {
1720 ttyLocker ttyl;
1721 tty->print("L%d selected ", lidx);
1722 lrg->mask().dump();
1723 tty->print(" from ");
1724 avail_rm.dump();
1725 tty->cr();
1726 }
1727 #endif
1728 // Note that reg is the highest-numbered register in the newly-bound mask.
1729 } // end color available case
1730
1731 //---------------
1732 // Live range is live and no colors available
1733 else {
1734 assert( lrg->alive(), "" );
1735 assert( !lrg->_fat_proj || lrg->is_multidef() ||
1736 lrg->_def->outcnt() > 0, "fat_proj cannot spill");
1737 assert( !orig_mask.is_infinite_stack(), "infinite stack does not spill" );
1738
1739 // Assign the special spillreg register
1740 lrg->set_reg(OptoReg::Name(spill_reg++));
1741 // Do not empty the regmask; leave mask_size lying around
1742 // for use during Spilling
1743 #ifndef PRODUCT
1744 if( trace_spilling() ) {
1745 ttyLocker ttyl;
1746 tty->print("L%d spilling with neighbors: ", lidx);
1747 s->dump();
1748 DEBUG_ONLY(tty->print(" original mask: "));
1749 DEBUG_ONLY(orig_mask.dump());
1750 dump_lrg(lidx);
1751 }
1752 #endif
1753 } // end spill case
1754
1755 }
1756
1757 return spill_reg-LRG::SPILL_REG; // Return number of spills
1758 }
1759
1760 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
1761 void PhaseChaitin::set_was_spilled( Node *n ) {
1762 if( _spilled_once.test_set(n->_idx) )
1763 _spilled_twice.set(n->_idx);
1764 }
1765
1766 // Convert Ideal spill instructions into proper FramePtr + offset Loads and
1767 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are.
1768 void PhaseChaitin::fixup_spills() {
1769 // This function does only cisc spill work.
1770 if( !UseCISCSpill ) return;
1771
1772 Compile::TracePhase tp(_t_fixupSpills);
1773
1774 // Grab the Frame Pointer
1775 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr);
1776
1777 // For all blocks
1778 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1779 Block* block = _cfg.get_block(i);
1780
1781 // For all instructions in block
1782 uint last_inst = block->end_idx();
1783 for (uint j = 1; j <= last_inst; j++) {
1784 Node* n = block->get_node(j);
1785
1786 // Dead instruction???
1787 assert( n->outcnt() != 0 ||// Nothing dead after post alloc
1788 C->top() == n || // Or the random TOP node
1789 n->is_Proj(), // Or a fat-proj kill node
1790 "No dead instructions after post-alloc" );
1791
1792 int inp = n->cisc_operand();
1793 if( inp != AdlcVMDeps::Not_cisc_spillable ) {
1794 // Convert operand number to edge index number
1795 MachNode *mach = n->as_Mach();
1796 inp = mach->operand_index(inp);
1797 Node *src = n->in(inp); // Value to load or store
1798 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src));
1799 OptoReg::Name src_reg = lrg_cisc.reg();
1800 // Doubles record the HIGH register of an adjacent pair.
1801 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs());
1802 if( OptoReg::is_stack(src_reg) ) { // If input is on stack
1803 // This is a CISC Spill, get stack offset and construct new node
1804 #ifndef PRODUCT
1805 if( TraceCISCSpill ) {
1806 tty->print(" reg-instr: ");
1807 n->dump();
1808 }
1809 #endif
1810 int stk_offset = reg2offset(src_reg);
1811 // Bailout if we might exceed node limit when spilling this instruction
1812 C->check_node_count(0, "out of nodes fixing spills");
1813 if (C->failing()) return;
1814 // Transform node
1815 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach();
1816 cisc->set_req(inp,fp); // Base register is frame pointer
1817 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) {
1818 assert( cisc->oper_input_base() == 2, "Only adding one edge");
1819 cisc->ins_req(1,src); // Requires a memory edge
1820 } else {
1821 // There is no space reserved for a memory edge before the inputs for
1822 // instructions which have "stackSlotX" parameter instead of "memory".
1823 // For example, "MoveF2I_stack_reg". We always need a memory edge from
1824 // src to cisc, else we might schedule cisc before src, loading from a
1825 // spill location before storing the spill. On some platforms, we land
1826 // in this else case because mach->oper_input_base() > 1, i.e. we have
1827 // multiple inputs. In some rare cases there are even multiple memory
1828 // operands, before and after spilling.
1829 // (e.g. spilling "addFPR24_reg_mem" to "addFPR24_mem_cisc")
1830 // In either case, there is no space in the inputs for the memory edge
1831 // so we add an additional precedence / memory edge.
1832 cisc->add_prec(src);
1833 }
1834 block->map_node(cisc, j); // Insert into basic block
1835 n->subsume_by(cisc, C); // Correct graph
1836 //
1837 ++_used_cisc_instructions;
1838 #ifndef PRODUCT
1839 if( TraceCISCSpill ) {
1840 tty->print(" cisc-instr: ");
1841 cisc->dump();
1842 }
1843 #endif
1844 } else {
1845 #ifndef PRODUCT
1846 if( TraceCISCSpill ) {
1847 tty->print(" using reg-instr: ");
1848 n->dump();
1849 }
1850 #endif
1851 ++_unused_cisc_instructions; // input can be on stack
1852 }
1853 }
1854
1855 } // End of for all instructions
1856
1857 } // End of for all blocks
1858 }
1859
1860 // Helper to stretch above; recursively discover the base Node for a
1861 // given derived Node. Easy for AddP-related machine nodes, but needs
1862 // to be recursive for derived Phis.
1863 Node* PhaseChaitin::find_base_for_derived(Node** derived_base_map, Node* derived, uint& maxlrg) {
1864 // See if already computed; if so return it
1865 if (derived_base_map[derived->_idx]) {
1866 return derived_base_map[derived->_idx];
1867 }
1868
1869 #ifdef ASSERT
1870 if (derived->is_Mach() && derived->as_Mach()->ideal_Opcode() == Op_VerifyVectorAlignment) {
1871 // Bypass the verification node
1872 Node* base = find_base_for_derived(derived_base_map, derived->in(1), maxlrg);
1873 derived_base_map[derived->_idx] = base;
1874 return base;
1875 }
1876 #endif
1877
1878 // See if this happens to be a base.
1879 // NOTE: we use TypePtr instead of TypeOopPtr because we can have
1880 // pointers derived from null! These are always along paths that
1881 // can't happen at run-time but the optimizer cannot deduce it so
1882 // we have to handle it gracefully.
1883 assert(!derived->bottom_type()->isa_narrowoop() ||
1884 derived->bottom_type()->make_ptr()->is_ptr()->offset() == 0, "sanity");
1885 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1886 // If its an OOP with a non-zero offset, then it is derived.
1887 if (tj == nullptr || tj->offset() == 0) {
1888 derived_base_map[derived->_idx] = derived;
1889 return derived;
1890 }
1891 // Derived is null+offset? Base is null!
1892 if( derived->is_Con() ) {
1893 Node *base = _matcher.mach_null();
1894 assert(base != nullptr, "sanity");
1895 if (base->in(0) == nullptr) {
1896 // Initialize it once and make it shared:
1897 // set control to _root and place it into Start block
1898 // (where top() node is placed).
1899 base->init_req(0, _cfg.get_root_node());
1900 Block *startb = _cfg.get_block_for_node(C->top());
1901 uint node_pos = startb->find_node(C->top());
1902 startb->insert_node(base, node_pos);
1903 _cfg.map_node_to_block(base, startb);
1904 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet");
1905
1906 // The loadConP0 might have projection nodes depending on architecture
1907 // Add the projection nodes to the CFG
1908 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) {
1909 Node* use = base->fast_out(i);
1910 if (use->is_MachProj()) {
1911 startb->insert_node(use, ++node_pos);
1912 _cfg.map_node_to_block(use, startb);
1913 new_lrg(use, maxlrg++);
1914 }
1915 }
1916 }
1917 if (_lrg_map.live_range_id(base) == 0) {
1918 new_lrg(base, maxlrg++);
1919 }
1920 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");
1921 derived_base_map[derived->_idx] = base;
1922 return base;
1923 }
1924
1925 // Check for AddP-related opcodes
1926 if (!derived->is_Phi()) {
1927 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name());
1928 Node *base = derived->in(AddPNode::Base);
1929 derived_base_map[derived->_idx] = base;
1930 return base;
1931 }
1932
1933 // Recursively find bases for Phis.
1934 // First check to see if we can avoid a base Phi here.
1935 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg);
1936 uint i;
1937 for( i = 2; i < derived->req(); i++ )
1938 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg))
1939 break;
1940 // Went to the end without finding any different bases?
1941 if( i == derived->req() ) { // No need for a base Phi here
1942 derived_base_map[derived->_idx] = base;
1943 return base;
1944 }
1945
1946 // Now we see we need a base-Phi here to merge the bases
1947 const Type *t = base->bottom_type();
1948 base = new PhiNode( derived->in(0), t );
1949 for( i = 1; i < derived->req(); i++ ) {
1950 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg));
1951 t = t->meet(base->in(i)->bottom_type());
1952 }
1953 base->as_Phi()->set_type(t);
1954
1955 // Search the current block for an existing base-Phi
1956 Block *b = _cfg.get_block_for_node(derived);
1957 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi
1958 Node *phi = b->get_node(i);
1959 if( !phi->is_Phi() ) { // Found end of Phis with no match?
1960 b->insert_node(base, i); // Must insert created Phi here as base
1961 _cfg.map_node_to_block(base, b);
1962 new_lrg(base,maxlrg++);
1963 break;
1964 }
1965 // See if Phi matches.
1966 uint j;
1967 for( j = 1; j < base->req(); j++ )
1968 if( phi->in(j) != base->in(j) &&
1969 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different nulls
1970 break;
1971 if( j == base->req() ) { // All inputs match?
1972 base = phi; // Then use existing 'phi' and drop 'base'
1973 break;
1974 }
1975 }
1976
1977
1978 // Cache info for later passes
1979 derived_base_map[derived->_idx] = base;
1980 return base;
1981 }
1982
1983 // At each Safepoint, insert extra debug edges for each pair of derived value/
1984 // base pointer that is live across the Safepoint for oopmap building. The
1985 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the
1986 // required edge set.
1987 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) {
1988 int must_recompute_live = false;
1989 uint maxlrg = _lrg_map.max_lrg_id();
1990 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique());
1991 memset( derived_base_map, 0, sizeof(Node*)*C->unique() );
1992
1993 // For all blocks in RPO do...
1994 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1995 Block* block = _cfg.get_block(i);
1996 // Note use of deep-copy constructor. I cannot hammer the original
1997 // liveout bits, because they are needed by the following coalesce pass.
1998 IndexSet liveout(_live->live(block));
1999
2000 for (uint j = block->end_idx() + 1; j > 1; j--) {
2001 Node* n = block->get_node(j - 1);
2002
2003 // Pre-split compares of loop-phis. Loop-phis form a cycle we would
2004 // like to see in the same register. Compare uses the loop-phi and so
2005 // extends its live range BUT cannot be part of the cycle. If this
2006 // extended live range overlaps with the update of the loop-phi value
2007 // we need both alive at the same time -- which requires at least 1
2008 // copy. But because Intel has only 2-address registers we end up with
2009 // at least 2 copies, one before the loop-phi update instruction and
2010 // one after. Instead we split the input to the compare just after the
2011 // phi.
2012 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) {
2013 Node *phi = n->in(1);
2014 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) {
2015 Block *phi_block = _cfg.get_block_for_node(phi);
2016 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) {
2017 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI];
2018 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask);
2019 insert_proj( phi_block, 1, spill, maxlrg++ );
2020 n->set_req(1,spill);
2021 must_recompute_live = true;
2022 }
2023 }
2024 }
2025
2026 // Get value being defined
2027 uint lidx = _lrg_map.live_range_id(n);
2028 // Ignore the occasional brand-new live range
2029 if (lidx && lidx < _lrg_map.max_lrg_id()) {
2030 // Remove from live-out set
2031 liveout.remove(lidx);
2032
2033 // Copies do not define a new value and so do not interfere.
2034 // Remove the copies source from the liveout set before interfering.
2035 uint idx = n->is_Copy();
2036 if (idx) {
2037 liveout.remove(_lrg_map.live_range_id(n->in(idx)));
2038 }
2039 }
2040
2041 // Found a safepoint?
2042 JVMState *jvms = n->jvms();
2043 if (jvms && !liveout.is_empty()) {
2044 // Now scan for a live derived pointer
2045 IndexSetIterator elements(&liveout);
2046 uint neighbor;
2047 while ((neighbor = elements.next()) != 0) {
2048 // Find reaching DEF for base and derived values
2049 // This works because we are still in SSA during this call.
2050 Node *derived = lrgs(neighbor)._def;
2051 const TypePtr *tj = derived->bottom_type()->isa_ptr();
2052 assert(!derived->bottom_type()->isa_narrowoop() ||
2053 derived->bottom_type()->make_ptr()->is_ptr()->offset() == 0, "sanity");
2054 // If its an OOP with a non-zero offset, then it is derived.
2055 if (tj && tj->offset() != 0 && tj->isa_oop_ptr()) {
2056 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg);
2057 assert(base->_idx < _lrg_map.size(), "");
2058 // Add reaching DEFs of derived pointer and base pointer as a
2059 // pair of inputs
2060 n->add_req(derived);
2061 n->add_req(base);
2062
2063 // See if the base pointer is already live to this point.
2064 // Since I'm working on the SSA form, live-ness amounts to
2065 // reaching def's. So if I find the base's live range then
2066 // I know the base's def reaches here.
2067 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or
2068 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND
2069 (_lrg_map.live_range_id(base) > 0) && // not a constant
2070 _cfg.get_block_for_node(base) != block) { // base not def'd in blk)
2071 // Base pointer is not currently live. Since I stretched
2072 // the base pointer to here and it crosses basic-block
2073 // boundaries, the global live info is now incorrect.
2074 // Recompute live.
2075 must_recompute_live = true;
2076 } // End of if base pointer is not live to debug info
2077 }
2078 } // End of scan all live data for derived ptrs crossing GC point
2079 } // End of if found a GC point
2080
2081 // Make all inputs live
2082 if (!n->is_Phi()) { // Phi function uses come from prior block
2083 for (uint k = 1; k < n->req(); k++) {
2084 uint lidx = _lrg_map.live_range_id(n->in(k));
2085 if (lidx < _lrg_map.max_lrg_id()) {
2086 liveout.insert(lidx);
2087 }
2088 }
2089 }
2090
2091 } // End of forall instructions in block
2092 liveout.clear(); // Free the memory used by liveout.
2093
2094 } // End of forall blocks
2095 _lrg_map.set_max_lrg_id(maxlrg);
2096
2097 // If I created a new live range I need to recompute live
2098 if (maxlrg != _ifg->_maxlrg) {
2099 must_recompute_live = true;
2100 }
2101
2102 return must_recompute_live != 0;
2103 }
2104
2105 // Extend the node to LRG mapping
2106
2107 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) {
2108 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node));
2109 }
2110
2111 #ifndef PRODUCT
2112 void PhaseChaitin::dump(const Node* n) const {
2113 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0;
2114 tty->print("L%d",r);
2115 if (r && n->Opcode() != Op_Phi) {
2116 if( _node_regs ) { // Got a post-allocation copy of allocation?
2117 tty->print("[");
2118 OptoReg::Name second = get_reg_second(n);
2119 if( OptoReg::is_valid(second) ) {
2120 if( OptoReg::is_reg(second) )
2121 tty->print("%s:",Matcher::regName[second]);
2122 else
2123 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second));
2124 }
2125 OptoReg::Name first = get_reg_first(n);
2126 if( OptoReg::is_reg(first) )
2127 tty->print("%s]",Matcher::regName[first]);
2128 else
2129 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first));
2130 } else
2131 n->out_RegMask().dump();
2132 }
2133 tty->print("/N%d\t",n->_idx);
2134 tty->print("%s === ", n->Name());
2135 uint k;
2136 for (k = 0; k < n->req(); k++) {
2137 Node *m = n->in(k);
2138 if (!m) {
2139 tty->print("_ ");
2140 }
2141 else {
2142 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2143 tty->print("L%d",r);
2144 // Data MultiNode's can have projections with no real registers.
2145 // Don't die while dumping them.
2146 int op = n->Opcode();
2147 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) {
2148 if( _node_regs ) {
2149 tty->print("[");
2150 OptoReg::Name second = get_reg_second(n->in(k));
2151 if( OptoReg::is_valid(second) ) {
2152 if( OptoReg::is_reg(second) )
2153 tty->print("%s:",Matcher::regName[second]);
2154 else
2155 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer),
2156 reg2offset_unchecked(second));
2157 }
2158 OptoReg::Name first = get_reg_first(n->in(k));
2159 if( OptoReg::is_reg(first) )
2160 tty->print("%s]",Matcher::regName[first]);
2161 else
2162 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer),
2163 reg2offset_unchecked(first));
2164 } else
2165 n->in_RegMask(k).dump();
2166 }
2167 tty->print("/N%d ",m->_idx);
2168 }
2169 }
2170 if( k < n->len() && n->in(k) ) tty->print("| ");
2171 for( ; k < n->len(); k++ ) {
2172 Node *m = n->in(k);
2173 if(!m) {
2174 break;
2175 }
2176 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2177 tty->print("L%d",r);
2178 tty->print("/N%d ",m->_idx);
2179 }
2180 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty);
2181 else n->dump_spec(tty);
2182 if( _spilled_once.test(n->_idx ) ) {
2183 tty->print(" Spill_1");
2184 if( _spilled_twice.test(n->_idx ) )
2185 tty->print(" Spill_2");
2186 }
2187 tty->print("\n");
2188 }
2189
2190 void PhaseChaitin::dump(const Block* b) const {
2191 b->dump_head(&_cfg);
2192
2193 // For all instructions
2194 for( uint j = 0; j < b->number_of_nodes(); j++ )
2195 dump(b->get_node(j));
2196 // Print live-out info at end of block
2197 if( _live ) {
2198 tty->print("Liveout: ");
2199 IndexSet *live = _live->live(b);
2200 IndexSetIterator elements(live);
2201 tty->print("{");
2202 uint i;
2203 while ((i = elements.next()) != 0) {
2204 tty->print("L%d ", _lrg_map.find_const(i));
2205 }
2206 tty->print_cr("}");
2207 }
2208 tty->print("\n");
2209 }
2210
2211 void PhaseChaitin::dump() const {
2212 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n",
2213 _matcher._new_SP, _framesize );
2214
2215 // For all blocks
2216 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2217 dump(_cfg.get_block(i));
2218 }
2219 // End of per-block dump
2220 tty->print("\n");
2221
2222 if (!_ifg) {
2223 tty->print("(No IFG.)\n");
2224 return;
2225 }
2226
2227 // Dump LRG array
2228 tty->print("--- Live RanGe Array ---\n");
2229 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) {
2230 tty->print("L%d: ",i2);
2231 if (i2 < _ifg->_maxlrg) {
2232 lrgs(i2).dump();
2233 }
2234 else {
2235 tty->print_cr("new LRG");
2236 }
2237 }
2238 tty->cr();
2239
2240 // Dump lo-degree list
2241 tty->print("Lo degree: ");
2242 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next )
2243 tty->print("L%d ",i3);
2244 tty->cr();
2245
2246 // Dump lo-stk-degree list
2247 tty->print("Lo stk degree: ");
2248 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next )
2249 tty->print("L%d ",i4);
2250 tty->cr();
2251
2252 // Dump lo-degree list
2253 tty->print("Hi degree: ");
2254 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next )
2255 tty->print("L%d ",i5);
2256 tty->cr();
2257 }
2258
2259 void PhaseChaitin::dump_degree_lists() const {
2260 // Dump lo-degree list
2261 tty->print("Lo degree: ");
2262 for( uint i = _lo_degree; i; i = lrgs(i)._next )
2263 tty->print("L%d ",i);
2264 tty->cr();
2265
2266 // Dump lo-stk-degree list
2267 tty->print("Lo stk degree: ");
2268 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next )
2269 tty->print("L%d ",i2);
2270 tty->cr();
2271
2272 // Dump lo-degree list
2273 tty->print("Hi degree: ");
2274 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next )
2275 tty->print("L%d ",i3);
2276 tty->cr();
2277 }
2278
2279 void PhaseChaitin::dump_simplified() const {
2280 tty->print("Simplified: ");
2281 for( uint i = _simplified; i; i = lrgs(i)._next )
2282 tty->print("L%d ",i);
2283 tty->cr();
2284 }
2285
2286 static char *print_reg(OptoReg::Name reg, const PhaseChaitin* pc, char* buf, size_t buf_size) {
2287 if ((int)reg < 0)
2288 os::snprintf_checked(buf, buf_size, "<OptoReg::%d>", (int)reg);
2289 else if (OptoReg::is_reg(reg))
2290 strcpy(buf, Matcher::regName[reg]);
2291 else
2292 os::snprintf_checked(buf, buf_size, "%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer),
2293 pc->reg2offset(reg));
2294 return buf+strlen(buf);
2295 }
2296
2297 // Dump a register name into a buffer. Be intelligent if we get called
2298 // before allocation is complete.
2299 char *PhaseChaitin::dump_register(const Node* n, char* buf, size_t buf_size) const {
2300 if( _node_regs ) {
2301 // Post allocation, use direct mappings, no LRG info available
2302 print_reg( get_reg_first(n), this, buf, buf_size);
2303 } else {
2304 uint lidx = _lrg_map.find_const(n); // Grab LRG number
2305 if( !_ifg ) {
2306 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2307 } else if( !lidx ) { // Special, not allocated value
2308 strcpy(buf,"Special");
2309 } else {
2310 if (lrgs(lidx)._is_vector) {
2311 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs()))
2312 print_reg( lrgs(lidx).reg(), this, buf, buf_size); // a bound machine register
2313 else
2314 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2315 } else if( (lrgs(lidx).num_regs() == 1)
2316 ? lrgs(lidx).mask().is_bound1()
2317 : lrgs(lidx).mask().is_bound_pair() ) {
2318 // Hah! We have a bound machine register
2319 print_reg( lrgs(lidx).reg(), this, buf, buf_size);
2320 } else {
2321 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2322 }
2323 }
2324 }
2325 return buf+strlen(buf);
2326 }
2327
2328 void PhaseChaitin::dump_for_spill_split_recycle() const {
2329 if( WizardMode && (PrintCompilation || PrintOpto) ) {
2330 // Display which live ranges need to be split and the allocator's state
2331 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt);
2332 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) {
2333 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
2334 tty->print("L%d: ", bidx);
2335 lrgs(bidx).dump();
2336 }
2337 }
2338 tty->cr();
2339 dump();
2340 }
2341 }
2342
2343 void PhaseChaitin::dump_frame() const {
2344 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer);
2345 const TypeTuple *domain = C->tf()->domain_cc();
2346 const int argcnt = domain->cnt() - TypeFunc::Parms;
2347
2348 // Incoming arguments in registers dump
2349 for( int k = 0; k < argcnt; k++ ) {
2350 OptoReg::Name parmreg = _matcher._parm_regs[k].first();
2351 if( OptoReg::is_reg(parmreg)) {
2352 const char *reg_name = OptoReg::regname(parmreg);
2353 tty->print("#r%3.3d %s", parmreg, reg_name);
2354 parmreg = _matcher._parm_regs[k].second();
2355 if( OptoReg::is_reg(parmreg)) {
2356 tty->print(":%s", OptoReg::regname(parmreg));
2357 }
2358 tty->print(" : parm %d: ", k);
2359 domain->field_at(k + TypeFunc::Parms)->dump();
2360 tty->cr();
2361 }
2362 }
2363
2364 // Check for un-owned padding above incoming args
2365 OptoReg::Name reg = _matcher._new_SP;
2366 if( reg > _matcher._in_arg_limit ) {
2367 reg = OptoReg::add(reg, -1);
2368 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg));
2369 }
2370
2371 // Incoming argument area dump
2372 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots());
2373 while( reg > begin_in_arg ) {
2374 reg = OptoReg::add(reg, -1);
2375 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2376 int j;
2377 for( j = 0; j < argcnt; j++) {
2378 if( _matcher._parm_regs[j].first() == reg ||
2379 _matcher._parm_regs[j].second() == reg ) {
2380 tty->print("parm %d: ",j);
2381 domain->field_at(j + TypeFunc::Parms)->dump();
2382 tty->cr();
2383 break;
2384 }
2385 }
2386 if( j >= argcnt )
2387 tty->print_cr("HOLE, owned by SELF");
2388 }
2389
2390 // Old outgoing preserve area
2391 while( reg > _matcher._old_SP ) {
2392 reg = OptoReg::add(reg, -1);
2393 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg));
2394 }
2395
2396 // Old SP
2397 tty->print_cr("# -- Old %s -- Framesize: %d --",fp,
2398 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize);
2399
2400 // Preserve area dump
2401 int fixed_slots = C->fixed_slots();
2402 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots());
2403 OptoReg::Name return_addr = _matcher.return_addr();
2404
2405 reg = OptoReg::add(reg, -1);
2406 while (OptoReg::is_stack(reg)) {
2407 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2408 if (return_addr == reg) {
2409 tty->print_cr("return address");
2410 } else if (reg >= begin_in_preserve) {
2411 // Preserved slots are present on x86
2412 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word))
2413 tty->print_cr("saved fp register");
2414 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) &&
2415 VerifyStackAtCalls)
2416 tty->print_cr("0xBADB100D +VerifyStackAtCalls");
2417 else
2418 tty->print_cr("in_preserve");
2419 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) {
2420 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg));
2421 } else {
2422 tty->print_cr("pad2, stack alignment");
2423 }
2424 reg = OptoReg::add(reg, -1);
2425 }
2426
2427 // Spill area dump
2428 reg = OptoReg::add(_matcher._new_SP, _framesize );
2429 while( reg > _matcher._out_arg_limit ) {
2430 reg = OptoReg::add(reg, -1);
2431 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg));
2432 }
2433
2434 // Outgoing argument area dump
2435 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) {
2436 reg = OptoReg::add(reg, -1);
2437 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg));
2438 }
2439
2440 // Outgoing new preserve area
2441 while( reg > _matcher._new_SP ) {
2442 reg = OptoReg::add(reg, -1);
2443 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg));
2444 }
2445 tty->print_cr("#");
2446 }
2447
2448 void PhaseChaitin::dump_bb(uint pre_order) const {
2449 tty->print_cr("---dump of B%d---",pre_order);
2450 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2451 Block* block = _cfg.get_block(i);
2452 if (block->_pre_order == pre_order) {
2453 dump(block);
2454 }
2455 }
2456 }
2457
2458 void PhaseChaitin::dump_lrg(uint lidx, bool defs_only) const {
2459 tty->print_cr("---dump of L%d---",lidx);
2460
2461 if (_ifg) {
2462 if (lidx >= _lrg_map.max_lrg_id()) {
2463 tty->print("Attempt to print live range index beyond max live range.\n");
2464 return;
2465 }
2466 tty->print("L%d: ",lidx);
2467 if (lidx < _ifg->_maxlrg) {
2468 lrgs(lidx).dump();
2469 } else {
2470 tty->print_cr("new LRG");
2471 }
2472 }
2473 if( _ifg && lidx < _ifg->_maxlrg) {
2474 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx));
2475 _ifg->neighbors(lidx)->dump();
2476 tty->cr();
2477 }
2478 // For all blocks
2479 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2480 Block* block = _cfg.get_block(i);
2481 int dump_once = 0;
2482
2483 // For all instructions
2484 for( uint j = 0; j < block->number_of_nodes(); j++ ) {
2485 Node *n = block->get_node(j);
2486 if (_lrg_map.find_const(n) == lidx) {
2487 if (!dump_once++) {
2488 tty->cr();
2489 block->dump_head(&_cfg);
2490 }
2491 dump(n);
2492 continue;
2493 }
2494 if (!defs_only) {
2495 uint cnt = n->req();
2496 for( uint k = 1; k < cnt; k++ ) {
2497 Node *m = n->in(k);
2498 if (!m) {
2499 continue; // be robust in the dumper
2500 }
2501 if (_lrg_map.find_const(m) == lidx) {
2502 if (!dump_once++) {
2503 tty->cr();
2504 block->dump_head(&_cfg);
2505 }
2506 dump(n);
2507 }
2508 }
2509 }
2510 }
2511 } // End of per-block dump
2512 tty->cr();
2513 }
2514 #endif // not PRODUCT
2515
2516 #ifdef ASSERT
2517 // Verify that base pointers and derived pointers are still sane.
2518 void PhaseChaitin::verify_base_ptrs(ResourceArea* a) const {
2519 Unique_Node_List worklist(a);
2520 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2521 Block* block = _cfg.get_block(i);
2522 for (uint j = block->end_idx() + 1; j > 1; j--) {
2523 Node* n = block->get_node(j-1);
2524 if (n->is_Phi()) {
2525 break;
2526 }
2527 // Found a safepoint?
2528 if (n->is_MachSafePoint()) {
2529 MachSafePointNode* sfpt = n->as_MachSafePoint();
2530 JVMState* jvms = sfpt->jvms();
2531 if (jvms != nullptr) {
2532 // Now scan for a live derived pointer
2533 if (jvms->oopoff() < sfpt->req()) {
2534 // Check each derived/base pair
2535 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
2536 Node* check = sfpt->in(idx);
2537 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
2538 // search upwards through spills and spill phis for AddP
2539 worklist.clear();
2540 worklist.push(check);
2541 uint k = 0;
2542 while (k < worklist.size()) {
2543 check = worklist.at(k);
2544 assert(check, "Bad base or derived pointer");
2545 // See PhaseChaitin::find_base_for_derived() for all cases.
2546 int isc = check->is_Copy();
2547 if (isc) {
2548 worklist.push(check->in(isc));
2549 } else if (check->is_Phi()) {
2550 for (uint m = 1; m < check->req(); m++) {
2551 worklist.push(check->in(m));
2552 }
2553 } else if (check->is_Con()) {
2554 if (is_derived && check->bottom_type()->is_ptr()->offset() != 0) {
2555 // Derived is null+non-zero offset, base must be null.
2556 assert(check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad derived pointer");
2557 } else {
2558 assert(check->bottom_type()->is_ptr()->offset() == 0, "Bad base pointer");
2559 // Base either ConP(nullptr) or loadConP
2560 if (check->is_Mach()) {
2561 assert(check->as_Mach()->ideal_Opcode() == Op_ConP, "Bad base pointer");
2562 } else {
2563 assert(check->Opcode() == Op_ConP &&
2564 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad base pointer");
2565 }
2566 }
2567 } else if (check->bottom_type()->is_ptr()->offset() == 0) {
2568 if (check->is_Proj() || (check->is_Mach() &&
2569 (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
2570 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
2571 check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
2572 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
2573 #ifdef _LP64
2574 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP) ||
2575 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN) ||
2576 (UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass) ||
2577 #endif // _LP64
2578 check->as_Mach()->ideal_Opcode() == Op_LoadP ||
2579 check->as_Mach()->ideal_Opcode() == Op_LoadKlass))) {
2580 // Valid nodes
2581 } else {
2582 check->dump();
2583 assert(false, "Bad base or derived pointer");
2584 }
2585 } else {
2586 assert(is_derived, "Bad base pointer");
2587 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP, "Bad derived pointer");
2588 }
2589 k++;
2590 assert(k < 100000, "Derived pointer checking in infinite loop");
2591 } // End while
2592 }
2593 } // End of check for derived pointers
2594 } // End of Kcheck for debug info
2595 } // End of if found a safepoint
2596 } // End of forall instructions in block
2597 } // End of forall blocks
2598 }
2599
2600 // Verify that graphs and base pointers are still sane.
2601 void PhaseChaitin::verify(ResourceArea* a, bool verify_ifg) const {
2602 if (VerifyRegisterAllocator) {
2603 _cfg.verify();
2604 if (C->failing()) {
2605 return;
2606 }
2607 verify_base_ptrs(a);
2608 if (verify_ifg) {
2609 _ifg->verify(this);
2610 }
2611 }
2612 }
2613 #endif // ASSERT
2614
2615 int PhaseChaitin::_final_loads = 0;
2616 int PhaseChaitin::_final_stores = 0;
2617 int PhaseChaitin::_final_memoves= 0;
2618 int PhaseChaitin::_final_copies = 0;
2619 double PhaseChaitin::_final_load_cost = 0;
2620 double PhaseChaitin::_final_store_cost = 0;
2621 double PhaseChaitin::_final_memove_cost= 0;
2622 double PhaseChaitin::_final_copy_cost = 0;
2623 int PhaseChaitin::_conserv_coalesce = 0;
2624 int PhaseChaitin::_conserv_coalesce_pair = 0;
2625 int PhaseChaitin::_conserv_coalesce_trie = 0;
2626 int PhaseChaitin::_conserv_coalesce_quad = 0;
2627 int PhaseChaitin::_post_alloc = 0;
2628 int PhaseChaitin::_lost_opp_pp_coalesce = 0;
2629 int PhaseChaitin::_lost_opp_cflow_coalesce = 0;
2630 int PhaseChaitin::_used_cisc_instructions = 0;
2631 int PhaseChaitin::_unused_cisc_instructions = 0;
2632 int PhaseChaitin::_allocator_attempts = 0;
2633 int PhaseChaitin::_allocator_successes = 0;
2634
2635 #ifndef PRODUCT
2636 uint PhaseChaitin::_high_pressure = 0;
2637 uint PhaseChaitin::_low_pressure = 0;
2638
2639 void PhaseChaitin::print_chaitin_statistics() {
2640 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);
2641 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);
2642 tty->print_cr("Adjusted spill cost = %7.0f.",
2643 _final_load_cost*4.0 + _final_store_cost * 2.0 +
2644 _final_copy_cost*1.0 + _final_memove_cost*12.0);
2645 tty->print("Conservatively coalesced %d copies, %d pairs",
2646 _conserv_coalesce, _conserv_coalesce_pair);
2647 if( _conserv_coalesce_trie || _conserv_coalesce_quad )
2648 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad);
2649 tty->print_cr(", %d post alloc.", _post_alloc);
2650 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce )
2651 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.",
2652 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce );
2653 if( _used_cisc_instructions || _unused_cisc_instructions )
2654 tty->print_cr("Used cisc instruction %d, remained in register %d",
2655 _used_cisc_instructions, _unused_cisc_instructions);
2656 if( _allocator_successes != 0 )
2657 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes);
2658 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure);
2659 }
2660 #endif // not PRODUCT