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 {
53 tty->print(", #!!!_%d_vs_%d ", _mask_size, _mask.size());
54 }
55 } else {
56 tty->print(", #?(%d) ", _mask.size());
57 }
58
59 tty->print("EffDeg: ");
60 if( _degree_valid ) tty->print( "%d ", _eff_degree );
61 else tty->print("? ");
62
63 if( is_multidef() ) {
64 tty->print("MultiDef ");
65 if (_defs != nullptr) {
66 tty->print("(");
67 for (int i = 0; i < _defs->length(); i++) {
68 tty->print("N%d ", _defs->at(i)->_idx);
69 }
70 tty->print(") ");
71 }
72 }
73 else if( _def == nullptr ) tty->print("Dead ");
74 else tty->print("Def: N%d ",_def->_idx);
75
76 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score());
77 // Flags
78 if( _is_oop ) tty->print("Oop ");
79 if( _is_float ) tty->print("Float ");
80 if( _is_vector ) tty->print("Vector ");
81 if( _is_predicate ) tty->print("Predicate ");
82 if( _is_scalable ) tty->print("Scalable ");
83 if( _was_spilled1 ) tty->print("Spilled ");
84 if( _was_spilled2 ) tty->print("Spilled2 ");
85 if( _direct_conflict ) tty->print("Direct_conflict ");
86 if( _fat_proj ) tty->print("Fat ");
87 if( _was_lo ) tty->print("Lo ");
88 if( _has_copy ) tty->print("Copy ");
89 if( _at_risk ) tty->print("Risk ");
90
91 if( _must_spill ) tty->print("Must_spill ");
92 if( _is_bound ) tty->print("Bound ");
93 if( _msize_valid ) {
94 if( _degree_valid && lo_degree() ) tty->print("Trivial ");
95 }
96
97 tty->cr();
98 }
99 #endif
100
101 // Compute score from cost and area. Low score is best to spill.
102 static double raw_score( double cost, double area ) {
103 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5;
104 }
105
106 double LRG::score() const {
107 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost.
108 // Bigger area lowers score, encourages spilling this live range.
109 // Bigger cost raise score, prevents spilling this live range.
110 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer
111 // to turn a divide by a constant into a multiply by the reciprical).
112 double score = raw_score( _cost, _area);
113
114 // Account for area. Basically, LRGs covering large areas are better
115 // to spill because more other LRGs get freed up.
116 if( _area == 0.0 ) // No area? Then no progress to spill
117 return 1e35;
118
119 if( _was_spilled2 ) // If spilled once before, we are unlikely
120 return score + 1e30; // to make progress again.
121
122 if( _cost >= _area*3.0 ) // Tiny area relative to cost
123 return score + 1e17; // Probably no progress to spill
124
125 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost
126 return score + 1e10; // Likely no progress to spill
127
128 return score;
129 }
130
131 #define NUMBUCKS 3
132
133 // Straight out of Tarjan's union-find algorithm
134 uint LiveRangeMap::find_compress(uint lrg) {
135 uint cur = lrg;
136 uint next = _uf_map.at(cur);
137 while (next != cur) { // Scan chain of equivalences
138 assert( next < cur, "always union smaller");
139 cur = next; // until find a fixed-point
140 next = _uf_map.at(cur);
141 }
142
143 // Core of union-find algorithm: update chain of
144 // equivalences to be equal to the root.
145 while (lrg != next) {
146 uint tmp = _uf_map.at(lrg);
147 _uf_map.at_put(lrg, next);
148 lrg = tmp;
149 }
150 return lrg;
151 }
152
153 // Reset the Union-Find map to identity
154 void LiveRangeMap::reset_uf_map(uint max_lrg_id) {
155 _max_lrg_id= max_lrg_id;
156 // Force the Union-Find mapping to be at least this large
157 _uf_map.at_put_grow(_max_lrg_id, 0);
158 // Initialize it to be the ID mapping.
159 for (uint i = 0; i < _max_lrg_id; ++i) {
160 _uf_map.at_put(i, i);
161 }
162 }
163
164 // Make all Nodes map directly to their final live range; no need for
165 // the Union-Find mapping after this call.
166 void LiveRangeMap::compress_uf_map_for_nodes() {
167 // For all Nodes, compress mapping
168 uint unique = _names.length();
169 for (uint i = 0; i < unique; ++i) {
170 uint lrg = _names.at(i);
171 uint compressed_lrg = find(lrg);
172 if (lrg != compressed_lrg) {
173 _names.at_put(i, compressed_lrg);
174 }
175 }
176 }
177
178 // Like Find above, but no path compress, so bad asymptotic behavior
179 uint LiveRangeMap::find_const(uint lrg) const {
180 if (!lrg) {
181 return lrg; // Ignore the zero LRG
182 }
183
184 // Off the end? This happens during debugging dumps when you got
185 // brand new live ranges but have not told the allocator yet.
186 if (lrg >= _max_lrg_id) {
187 return lrg;
188 }
189
190 uint next = _uf_map.at(lrg);
191 while (next != lrg) { // Scan chain of equivalences
192 assert(next < lrg, "always union smaller");
193 lrg = next; // until find a fixed-point
194 next = _uf_map.at(lrg);
195 }
196 return next;
197 }
198
199 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated)
200 : PhaseRegAlloc(unique, cfg, matcher,
201 #ifndef PRODUCT
202 print_chaitin_statistics
203 #else
204 nullptr
205 #endif
206 )
207 , _live(nullptr)
208 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0)
209 , _oldphi(unique)
210 #ifndef PRODUCT
211 , _trace_spilling(C->directive()->TraceSpillingOption)
212 #endif
213 , _lrg_map(Thread::current()->resource_area(), unique)
214 , _scheduling_info_generated(scheduling_info_generated)
215 , _sched_int_pressure(0, Matcher::int_pressure_limit())
216 , _sched_float_pressure(0, Matcher::float_pressure_limit())
217 , _scratch_int_pressure(0, Matcher::int_pressure_limit())
218 , _scratch_float_pressure(0, Matcher::float_pressure_limit())
219 {
220 Compile::TracePhase tp(_t_ctorChaitin);
221
222 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency());
223
224 // Build a list of basic blocks, sorted by frequency
225 // Experiment with sorting strategies to speed compilation
226 uint nr_blocks = _cfg.number_of_blocks();
227 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket
228 Block **buckets[NUMBUCKS]; // Array of buckets
229 uint buckcnt[NUMBUCKS]; // Array of bucket counters
230 double buckval[NUMBUCKS]; // Array of bucket value cutoffs
231
232 // The space which our buckets point into.
233 Block** start = NEW_RESOURCE_ARRAY(Block *, nr_blocks*NUMBUCKS);
234
235 for (uint i = 0; i < NUMBUCKS; i++) {
236 buckets[i] = &start[i*nr_blocks];
237 buckcnt[i] = 0;
238 // Bump by three orders of magnitude each time
239 cutoff *= 0.001;
240 buckval[i] = cutoff;
241 }
242
243 // Sort blocks into buckets
244 for (uint i = 0; i < nr_blocks; i++) {
245 for (uint j = 0; j < NUMBUCKS; j++) {
246 double bval = buckval[j];
247 Block* blk = _cfg.get_block(i);
248 if (j == NUMBUCKS - 1 || blk->_freq > bval) {
249 uint cnt = buckcnt[j];
250 // Assign block to end of list for appropriate bucket
251 buckets[j][cnt] = blk;
252 buckcnt[j] = cnt+1;
253 break; // kick out of inner loop
254 }
255 }
256 }
257
258 // Squash the partially filled buckets together into the first one.
259 static_assert(NUMBUCKS >= 2, "must"); // If this isn't true then it'll mess up the squashing.
260 Block** offset = &buckets[0][buckcnt[0]];
261 for (int i = 1; i < NUMBUCKS; i++) {
262 ::memmove(offset, buckets[i], buckcnt[i]*sizeof(Block*));
263 offset += buckcnt[i];
264 }
265 assert((&buckets[0][0] + nr_blocks) == offset, "should be");
266
267 // Free the now unused memory
268 FREE_RESOURCE_ARRAY(Block*, buckets[1], (NUMBUCKS-1)*nr_blocks);
269 // Finally, point the _blks to our memory
270 _blks = buckets[0];
271
272 #ifdef ASSERT
273 uint blkcnt = 0;
274 for (uint i = 0; i < NUMBUCKS; i++) {
275 blkcnt += buckcnt[i];
276 }
277 assert(blkcnt == nr_blocks, "Block array not totally filled");
278 #endif
279 }
280
281 // union 2 sets together.
282 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) {
283 uint src = _lrg_map.find(src_n);
284 uint dst = _lrg_map.find(dst_n);
285 assert(src, "");
286 assert(dst, "");
287 assert(src < _lrg_map.max_lrg_id(), "oob");
288 assert(dst < _lrg_map.max_lrg_id(), "oob");
289 assert(src < dst, "always union smaller");
290 _lrg_map.uf_map(dst, src);
291 }
292
293 void PhaseChaitin::new_lrg(const Node *x, uint lrg) {
294 // Make the Node->LRG mapping
295 _lrg_map.extend(x->_idx,lrg);
296 // Make the Union-Find mapping an identity function
297 _lrg_map.uf_extend(lrg, lrg);
298 }
299
300
301 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) {
302 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections");
303 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); )
304 int found_projs = 0;
305 uint cnt = orig->outcnt();
306 for (uint i = 0; i < cnt; i++) {
307 Node* proj = orig->raw_out(i);
308 if (proj->is_MachProj()) {
309 assert(proj->outcnt() == 0, "only kill projections are expected here");
310 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections");
311 found_projs++;
312 // Copy kill projections after the cloned node
313 Node* kills = proj->clone();
314 kills->set_req(0, copy);
315 b->insert_node(kills, idx++);
316 _cfg.map_node_to_block(kills, b);
317 new_lrg(kills, max_lrg_id++);
318 }
319 }
320 return found_projs;
321 }
322
323 // Renumber the live ranges to compact them. Makes the IFG smaller.
324 void PhaseChaitin::compact() {
325 Compile::TracePhase tp(_t_chaitinCompact);
326
327 // Current the _uf_map contains a series of short chains which are headed
328 // by a self-cycle. All the chains run from big numbers to little numbers.
329 // The Find() call chases the chains & shortens them for the next Find call.
330 // We are going to change this structure slightly. Numbers above a moving
331 // wave 'i' are unchanged. Numbers below 'j' point directly to their
332 // compacted live range with no further chaining. There are no chains or
333 // cycles below 'i', so the Find call no longer works.
334 uint j=1;
335 uint i;
336 for (i = 1; i < _lrg_map.max_lrg_id(); i++) {
337 uint lr = _lrg_map.uf_live_range_id(i);
338 // Ignore unallocated live ranges
339 if (!lr) {
340 continue;
341 }
342 assert(lr <= i, "");
343 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr));
344 }
345 // Now change the Node->LR mapping to reflect the compacted names
346 uint unique = _lrg_map.size();
347 for (i = 0; i < unique; i++) {
348 uint lrg_id = _lrg_map.live_range_id(i);
349 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id));
350 }
351
352 // Reset the Union-Find mapping
353 _lrg_map.reset_uf_map(j);
354 }
355
356 void PhaseChaitin::Register_Allocate() {
357
358 // Above the OLD FP (and in registers) are the incoming arguments. Stack
359 // slots in this area are called "arg_slots". Above the NEW FP (and in
360 // registers) is the outgoing argument area; above that is the spill/temp
361 // area. These are all "frame_slots". Arg_slots start at the zero
362 // stack_slots and count up to the known arg_size. Frame_slots start at
363 // the stack_slot #arg_size and go up. After allocation I map stack
364 // slots to actual offsets. Stack-slots in the arg_slot area are biased
365 // by the frame_size; stack-slots in the frame_slot area are biased by 0.
366
367 _trip_cnt = 0;
368 _alternate = 0;
369 _matcher._allocation_started = true;
370
371 ResourceArea split_arena(mtCompiler, Arena::Tag::tag_regsplit); // Arena for Split local resources
372 ResourceArea live_arena(mtCompiler, Arena::Tag::tag_reglive); // Arena for liveness & IFG info
373 ResourceMark rm(&live_arena);
374
375 // Need live-ness for the IFG; need the IFG for coalescing. If the
376 // liveness is JUST for coalescing, then I can get some mileage by renaming
377 // all copy-related live ranges low and then using the max copy-related
378 // live range as a cut-off for LIVE and the IFG. In other words, I can
379 // build a subset of LIVE and IFG just for copies.
380 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false);
381
382 // Need IFG for coalescing and coloring
383 PhaseIFG ifg(&live_arena);
384 _ifg = &ifg;
385
386 // Come out of SSA world to the Named world. Assign (virtual) registers to
387 // Nodes. Use the same register for all inputs and the output of PhiNodes
388 // - effectively ending SSA form. This requires either coalescing live
389 // ranges or inserting copies. For the moment, we insert "virtual copies"
390 // - we pretend there is a copy prior to each Phi in predecessor blocks.
391 // We will attempt to coalesce such "virtual copies" before we manifest
392 // them for real.
393 de_ssa();
394
395 #ifdef ASSERT
396 // Verify the graph before RA.
397 verify(&live_arena);
398 #endif
399
400 {
401 Compile::TracePhase tp(_t_computeLive);
402 _live = nullptr; // Mark live as being not available
403 rm.reset_to_mark(); // Reclaim working storage
404 IndexSet::reset_memory(C, &live_arena);
405 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG
406 gather_lrg_masks( false ); // Collect LRG masks
407 live.compute(_lrg_map.max_lrg_id()); // Compute liveness
408 _live = &live; // Mark LIVE as being available
409 }
410
411 C->print_method(PHASE_INITIAL_LIVENESS, 4);
412
413 // Base pointers are currently "used" by instructions which define new
414 // derived pointers. This makes base pointers live up to the where the
415 // derived pointer is made, but not beyond. Really, they need to be live
416 // across any GC point where the derived value is live. So this code looks
417 // at all the GC points, and "stretches" the live range of any base pointer
418 // to the GC point.
419 if (stretch_base_pointer_live_ranges(&live_arena)) {
420 Compile::TracePhase tp("computeLive (sbplr)", _t_computeLive);
421 // Since some live range stretched, I need to recompute live
422 _live = nullptr;
423 rm.reset_to_mark(); // Reclaim working storage
424 IndexSet::reset_memory(C, &live_arena);
425 ifg.init(_lrg_map.max_lrg_id());
426 gather_lrg_masks(false);
427 live.compute(_lrg_map.max_lrg_id());
428 _live = &live;
429 C->print_method(PHASE_LIVE_RANGE_STRETCHING, 4);
430 }
431
432 // Create the interference graph using virtual copies
433 build_ifg_virtual(); // Include stack slots this time
434 if (C->failing()) {
435 return;
436 }
437
438 // The IFG is/was triangular. I am 'squaring it up' so Union can run
439 // faster. Union requires a 'for all' operation which is slow on the
440 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' -
441 // meaning I can visit all the Nodes neighbors less than a Node in time
442 // O(# of neighbors), but I have to visit all the Nodes greater than a
443 // given Node and search them for an instance, i.e., time O(#MaxLRG)).
444 _ifg->SquareUp();
445
446 // Aggressive (but pessimistic) copy coalescing.
447 // This pass works on virtual copies. Any virtual copies which are not
448 // coalesced get manifested as actual copies
449 {
450 Compile::TracePhase tp(_t_chaitinCoalesce1);
451
452 PhaseAggressiveCoalesce coalesce(*this);
453 coalesce.coalesce_driver();
454 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do
455 // not match the Phi itself, insert a copy.
456 coalesce.insert_copies(_matcher);
457 if (C->failing()) {
458 return;
459 }
460 }
461
462 // After aggressive coalesce, attempt a first cut at coloring.
463 // To color, we need the IFG and for that we need LIVE.
464 {
465 Compile::TracePhase tp(_t_computeLive);
466 _live = nullptr;
467 rm.reset_to_mark(); // Reclaim working storage
468 IndexSet::reset_memory(C, &live_arena);
469 ifg.init(_lrg_map.max_lrg_id());
470 gather_lrg_masks( true );
471 live.compute(_lrg_map.max_lrg_id());
472 _live = &live;
473 }
474
475 C->print_method(PHASE_AGGRESSIVE_COALESCING, 4);
476
477 // Build physical interference graph
478 uint must_spill = 0;
479 must_spill = build_ifg_physical(&live_arena);
480 if (C->failing()) {
481 return;
482 }
483 // If we have a guaranteed spill, might as well spill now
484 if (must_spill) {
485 if(!_lrg_map.max_lrg_id()) {
486 return;
487 }
488 // Bail out if unique gets too large (ie - unique > MaxNodeLimit)
489 C->check_node_count(10*must_spill, "out of nodes before split");
490 if (C->failing()) {
491 return;
492 }
493
494 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
495 if (C->failing()) {
496 return;
497 }
498 _lrg_map.set_max_lrg_id(new_max_lrg_id);
499 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
500 // or we failed to split
501 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split");
502 if (C->failing()) {
503 return;
504 }
505
506 NOT_PRODUCT(C->verify_graph_edges();)
507
508 compact(); // Compact LRGs; return new lower max lrg
509
510 {
511 Compile::TracePhase tp(_t_computeLive);
512 _live = nullptr;
513 rm.reset_to_mark(); // Reclaim working storage
514 IndexSet::reset_memory(C, &live_arena);
515 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph
516 gather_lrg_masks( true ); // Collect intersect mask
517 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE
518 _live = &live;
519 }
520
521 C->print_method(PHASE_INITIAL_SPILLING, 4);
522
523 build_ifg_physical(&live_arena);
524 if (C->failing()) {
525 return;
526 }
527 _ifg->SquareUp();
528 _ifg->Compute_Effective_Degree();
529 // Only do conservative coalescing if requested
530 if (OptoCoalesce) {
531 Compile::TracePhase tp(_t_chaitinCoalesce2);
532 // Conservative (and pessimistic) copy coalescing of those spills
533 PhaseConservativeCoalesce coalesce(*this);
534 // If max live ranges greater than cutoff, don't color the stack.
535 // This cutoff can be larger than below since it is only done once.
536 coalesce.coalesce_driver();
537 }
538 _lrg_map.compress_uf_map_for_nodes();
539
540 if (OptoCoalesce) {
541 C->print_method(PHASE_CONSERVATIVE_COALESCING, 4);
542 }
543
544 #ifdef ASSERT
545 verify(&live_arena, true);
546 #endif
547 } else {
548 ifg.SquareUp();
549 ifg.Compute_Effective_Degree();
550 #ifdef ASSERT
551 set_was_low();
552 #endif
553 }
554
555 // Prepare for Simplify & Select
556 cache_lrg_info(); // Count degree of LRGs
557
558 // Simplify the InterFerence Graph by removing LRGs of low degree.
559 // LRGs of low degree are trivially colorable.
560 Simplify();
561
562 // Select colors by re-inserting LRGs back into the IFG in reverse order.
563 // Return whether or not something spills.
564 uint spills = Select( );
565 if (C->failing()) {
566 return;
567 }
568
569 // If we spill, split and recycle the entire thing
570 while( spills ) {
571 if( _trip_cnt++ > 24 ) {
572 DEBUG_ONLY( dump_for_spill_split_recycle(); )
573 if( _trip_cnt > 27 ) {
574 C->record_method_not_compilable("failed spill-split-recycle sanity check");
575 return;
576 }
577 }
578
579 if (!_lrg_map.max_lrg_id()) {
580 return;
581 }
582 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
583 if (C->failing()) {
584 return;
585 }
586 _lrg_map.set_max_lrg_id(new_max_lrg_id);
587 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
588 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split");
589 if (C->failing()) {
590 return;
591 }
592
593 compact(); // Compact LRGs; return new lower max lrg
594
595 // Nuke the live-ness and interference graph and LiveRanGe info
596 {
597 Compile::TracePhase tp(_t_computeLive);
598 _live = nullptr;
599 rm.reset_to_mark(); // Reclaim working storage
600 IndexSet::reset_memory(C, &live_arena);
601 ifg.init(_lrg_map.max_lrg_id());
602
603 // Create LiveRanGe array.
604 // Intersect register masks for all USEs and DEFs
605 gather_lrg_masks(true);
606 live.compute(_lrg_map.max_lrg_id());
607 _live = &live;
608 }
609
610 C->print_method(PHASE_ITERATIVE_SPILLING, 4);
611
612 must_spill = build_ifg_physical(&live_arena);
613 if (C->failing()) {
614 return;
615 }
616 _ifg->SquareUp();
617 _ifg->Compute_Effective_Degree();
618
619 // Only do conservative coalescing if requested
620 if (OptoCoalesce) {
621 Compile::TracePhase tp(_t_chaitinCoalesce3);
622 // Conservative (and pessimistic) copy coalescing
623 PhaseConservativeCoalesce coalesce(*this);
624 // Check for few live ranges determines how aggressive coalesce is.
625 coalesce.coalesce_driver();
626 }
627 _lrg_map.compress_uf_map_for_nodes();
628
629 if (OptoCoalesce) {
630 C->print_method(PHASE_CONSERVATIVE_COALESCING, 4);
631 }
632
633 #ifdef ASSERT
634 verify(&live_arena, true);
635 #endif
636 cache_lrg_info(); // Count degree of LRGs
637
638 // Simplify the InterFerence Graph by removing LRGs of low degree.
639 // LRGs of low degree are trivially colorable.
640 Simplify();
641
642 // Select colors by re-inserting LRGs back into the IFG in reverse order.
643 // Return whether or not something spills.
644 spills = Select();
645 if (C->failing()) {
646 return;
647 }
648 }
649
650 C->print_method(PHASE_AFTER_ITERATIVE_SPILLING, 4);
651
652 // Count number of Simplify-Select trips per coloring success.
653 _allocator_attempts += _trip_cnt + 1;
654 _allocator_successes += 1;
655
656 // Peephole remove copies
657 post_allocate_copy_removal();
658
659 C->print_method(PHASE_POST_ALLOCATION_COPY_REMOVAL, 4);
660
661 // Merge multidefs if multiple defs representing the same value are used in a single block.
662 merge_multidefs();
663
664 C->print_method(PHASE_MERGE_MULTI_DEFS, 4);
665
666 #ifdef ASSERT
667 // Verify the graph after RA.
668 verify(&live_arena);
669 #endif
670
671 // max_reg is past the largest *register* used.
672 // Convert that to a frame_slot number.
673 if (_max_reg <= _matcher._new_SP) {
674 _framesize = C->out_preserve_stack_slots();
675 }
676 else {
677 _framesize = _max_reg -_matcher._new_SP;
678 }
679 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough");
680
681 // This frame must preserve the required fp alignment
682 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots());
683 assert(_framesize <= 1000000, "sanity check");
684 #ifndef PRODUCT
685 _total_framesize += _framesize;
686 if ((int)_framesize > _max_framesize) {
687 _max_framesize = _framesize;
688 }
689 #endif
690
691 // Convert CISC spills
692 fixup_spills();
693
694 C->print_method(PHASE_FIX_UP_SPILLS, 4);
695
696 // Log regalloc results
697 CompileLog* log = Compile::current()->log();
698 if (log != nullptr) {
699 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing());
700 }
701
702 if (C->failing()) {
703 return;
704 }
705
706 NOT_PRODUCT(C->verify_graph_edges();)
707
708 // Move important info out of the live_arena to longer lasting storage.
709 alloc_node_regs(_lrg_map.size());
710 for (uint i=0; i < _lrg_map.size(); i++) {
711 if (_lrg_map.live_range_id(i)) { // Live range associated with Node?
712 LRG &lrg = lrgs(_lrg_map.live_range_id(i));
713 if (!lrg.alive()) {
714 set_bad(i);
715 } else if ((lrg.num_regs() == 1 && !lrg.is_scalable()) ||
716 (lrg.is_scalable() && lrg.scalable_reg_slots() == 1)) {
717 set1(i, lrg.reg());
718 } else { // Must be a register-set
719 if (!lrg._fat_proj) { // Must be aligned adjacent register set
720 // Live ranges record the highest register in their mask.
721 // We want the low register for the AD file writer's convenience.
722 OptoReg::Name hi = lrg.reg(); // Get hi register
723 int num_regs = lrg.num_regs();
724 if (lrg.is_scalable() && OptoReg::is_stack(hi)) {
725 // For scalable vector registers, when they are allocated in physical
726 // registers, num_regs is RegMask::SlotsPerVecA for reg mask of scalable
727 // vector. If they are allocated on stack, we need to get the actual
728 // num_regs, which reflects the physical length of scalable registers.
729 num_regs = lrg.scalable_reg_slots();
730 }
731 if (num_regs == 1) {
732 set1(i, hi);
733 } else {
734 OptoReg::Name lo = OptoReg::add(hi, (1 - num_regs)); // Find lo
735 // We have to use pair [lo,lo+1] even for wide vectors/vmasks because
736 // the rest of code generation works only with pairs. It is safe
737 // since for registers encoding only 'lo' is used.
738 // Second reg from pair is used in ScheduleAndBundle with vector max
739 // size 8 which corresponds to registers pair.
740 // It is also used in BuildOopMaps but oop operations are not
741 // vectorized.
742 set2(i, lo);
743 }
744 } else { // Misaligned; extract 2 bits
745 OptoReg::Name hi = lrg.reg(); // Get hi register
746 lrg.remove(hi); // Yank from mask
747 int lo = lrg.mask().find_first_elem(); // Find lo
748 set_pair(i, hi, lo);
749 }
750 }
751 if( lrg._is_oop ) _node_oops.set(i);
752 } else {
753 set_bad(i);
754 }
755 }
756
757 // Done!
758 _live = nullptr;
759 _ifg = nullptr;
760 C->set_indexSet_arena(nullptr); // ResourceArea is at end of scope
761 }
762
763 void PhaseChaitin::de_ssa() {
764 // Set initial Names for all Nodes. Most Nodes get the virtual register
765 // number. A few get the ZERO live range number. These do not
766 // get allocated, but instead rely on correct scheduling to ensure that
767 // only one instance is simultaneously live at a time.
768 uint lr_counter = 1;
769 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
770 Block* block = _cfg.get_block(i);
771 uint cnt = block->number_of_nodes();
772
773 // Handle all the normal Nodes in the block
774 for( uint j = 0; j < cnt; j++ ) {
775 Node *n = block->get_node(j);
776 // Pre-color to the zero live range, or pick virtual register
777 const RegMask &rm = n->out_RegMask();
778 _lrg_map.map(n->_idx, !rm.is_empty() ? lr_counter++ : 0);
779 }
780 }
781
782 // Reset the Union-Find mapping to be identity
783 _lrg_map.reset_uf_map(lr_counter);
784 }
785
786 void PhaseChaitin::mark_ssa() {
787 // Use ssa names to populate the live range maps or if no mask
788 // is available, use the 0 entry.
789 uint max_idx = 0;
790 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
791 Block* block = _cfg.get_block(i);
792 uint cnt = block->number_of_nodes();
793
794 // Handle all the normal Nodes in the block
795 for ( uint j = 0; j < cnt; j++ ) {
796 Node *n = block->get_node(j);
797 // Pre-color to the zero live range, or pick virtual register
798 const RegMask &rm = n->out_RegMask();
799 _lrg_map.map(n->_idx, !rm.is_empty() ? n->_idx : 0);
800 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx;
801 }
802 }
803 _lrg_map.set_max_lrg_id(max_idx+1);
804
805 // Reset the Union-Find mapping to be identity
806 _lrg_map.reset_uf_map(max_idx+1);
807 }
808
809
810 // Gather LiveRanGe information, including register masks. Modification of
811 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce.
812 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) {
813
814 // Nail down the frame pointer live range
815 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr));
816 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite
817
818 // For all blocks
819 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
820 Block* block = _cfg.get_block(i);
821
822 // For all instructions
823 for (uint j = 1; j < block->number_of_nodes(); j++) {
824 Node* n = block->get_node(j);
825 uint input_edge_start =1; // Skip control most nodes
826 bool is_machine_node = false;
827 if (n->is_Mach()) {
828 is_machine_node = true;
829 input_edge_start = n->as_Mach()->oper_input_base();
830 }
831 uint idx = n->is_Copy();
832
833 // Get virtual register number, same as LiveRanGe index
834 uint vreg = _lrg_map.live_range_id(n);
835 LRG& lrg = lrgs(vreg);
836 if (vreg) { // No vreg means un-allocable (e.g. memory)
837
838 // Check for float-vs-int live range (used in register-pressure
839 // calculations)
840 const Type *n_type = n->bottom_type();
841 if (n_type->is_floatingpoint()) {
842 lrg._is_float = 1;
843 }
844
845 // Check for twice prior spilling. Once prior spilling might have
846 // spilled 'soft', 2nd prior spill should have spilled 'hard' and
847 // further spilling is unlikely to make progress.
848 if (_spilled_once.test(n->_idx)) {
849 lrg._was_spilled1 = 1;
850 if (_spilled_twice.test(n->_idx)) {
851 lrg._was_spilled2 = 1;
852 }
853 }
854
855 #ifndef PRODUCT
856 // Collect bits not used by product code, but which may be useful for
857 // debugging.
858
859 // Collect has-copy bit
860 if (idx) {
861 lrg._has_copy = 1;
862 uint clidx = _lrg_map.live_range_id(n->in(idx));
863 LRG& copy_src = lrgs(clidx);
864 copy_src._has_copy = 1;
865 }
866
867 if (trace_spilling() && lrg._def != nullptr) {
868 // collect defs for MultiDef printing
869 if (lrg._defs == nullptr) {
870 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, nullptr);
871 lrg._defs->append(lrg._def);
872 }
873 lrg._defs->append(n);
874 }
875 #endif
876
877 // Check for a single def LRG; these can spill nicely
878 // via rematerialization. Flag as null for no def found
879 // yet, or 'n' for single def or -1 for many defs.
880 lrg._def = lrg._def ? NodeSentinel : n;
881
882 // Limit result register mask to acceptable registers
883 const RegMask &rm = n->out_RegMask();
884 lrg.and_with(rm);
885
886 uint ireg = n->ideal_reg();
887 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP,
888 "oops must be in Op_RegP's" );
889
890 // Check for vector live range (only if vector register is used).
891 // On SPARC vector uses RegD which could be misaligned so it is not
892 // processes as vector in RA.
893 if (RegMask::is_vector(ireg)) {
894 lrg._is_vector = 1;
895 if (Matcher::implements_scalable_vector && ireg == Op_VecA) {
896 assert(Matcher::supports_scalable_vector(), "scalable vector should be supported");
897 lrg._is_scalable = 1;
898 // For scalable vector, when it is allocated in physical register,
899 // num_regs is RegMask::SlotsPerVecA for reg mask,
900 // which may not be the actual physical register size.
901 // If it is allocated in stack, we need to get the actual
902 // physical length of scalable vector register.
903 lrg.set_scalable_reg_slots(Matcher::scalable_vector_reg_size(T_FLOAT));
904 }
905 }
906
907 if (ireg == Op_RegVectMask) {
908 assert(Matcher::has_predicated_vectors(), "predicated vector should be supported");
909 lrg._is_predicate = 1;
910 if (Matcher::supports_scalable_vector()) {
911 lrg._is_scalable = 1;
912 // For scalable predicate, when it is allocated in physical register,
913 // num_regs is RegMask::SlotsPerRegVectMask for reg mask,
914 // which may not be the actual physical register size.
915 // If it is allocated in stack, we need to get the actual
916 // physical length of scalable predicate register.
917 lrg.set_scalable_reg_slots(Matcher::scalable_predicate_reg_slots());
918 }
919 }
920 assert(n_type->isa_vect() == nullptr || lrg._is_vector ||
921 ireg == Op_RegD || ireg == Op_RegL || ireg == Op_RegVectMask,
922 "vector must be in vector registers");
923
924 // Check for bound register masks
925 const RegMask &lrgmask = lrg.mask();
926 if (lrgmask.is_bound(ireg)) {
927 lrg._is_bound = 1;
928 }
929
930 // Check for maximum frequency value
931 if (lrg._maxfreq < block->_freq) {
932 lrg._maxfreq = block->_freq;
933 }
934
935 // Check for oop-iness, or long/double
936 // Check for multi-kill projection
937 switch (ireg) {
938 case MachProjNode::fat_proj:
939 // Fat projections have size equal to number of registers killed
940 lrg.set_num_regs(rm.size());
941 lrg.set_reg_pressure(lrg.num_regs());
942 lrg._fat_proj = 1;
943 lrg._is_bound = 1;
944 break;
945 case Op_RegP:
946 #ifdef _LP64
947 lrg.set_num_regs(2); // Size is 2 stack words
948 #else
949 lrg.set_num_regs(1); // Size is 1 stack word
950 #endif
951 // Register pressure is tracked relative to the maximum values
952 // suggested for that platform, INTPRESSURE and FLOATPRESSURE,
953 // and relative to other types which compete for the same regs.
954 //
955 // The following table contains suggested values based on the
956 // architectures as defined in each .ad file.
957 // INTPRESSURE and FLOATPRESSURE may be tuned differently for
958 // compile-speed or performance.
959 // Note1:
960 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1
961 // since .ad registers are defined as high and low halves.
962 // These reg_pressure values remain compatible with the code
963 // in is_high_pressure() which relates get_invalid_mask_size(),
964 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE.
965 // Note2:
966 // SPARC -d32 has 24 registers available for integral values,
967 // but only 10 of these are safe for 64-bit longs.
968 // Using set_reg_pressure(2) for both int and long means
969 // the allocator will believe it can fit 26 longs into
970 // registers. Using 2 for longs and 1 for ints means the
971 // allocator will attempt to put 52 integers into registers.
972 // The settings below limit this problem to methods with
973 // many long values which are being run on 32-bit SPARC.
974 //
975 // ------------------- reg_pressure --------------------
976 // Each entry is reg_pressure_per_value,number_of_regs
977 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE
978 // SPARC 2 2 2 2 2 48 (24) 52 (26)
979 // SPARCV9 2 2 2 2 2 48 (24) 52 (26)
980 // AMD64 1 1 1 1 1 14 15
981 // -----------------------------------------------------
982 lrg.set_reg_pressure(1); // normally one value per register
983 if( n_type->isa_oop_ptr() ) {
984 lrg._is_oop = 1;
985 }
986 break;
987 case Op_RegL: // Check for long or double
988 case Op_RegD:
989 lrg.set_num_regs(2);
990 // Define platform specific register pressure
991 #if defined(ARM32)
992 lrg.set_reg_pressure(2);
993 #else
994 lrg.set_reg_pressure(1); // normally one value per register
995 #endif
996 // If this def of a double forces a mis-aligned double,
997 // flag as '_fat_proj' - really flag as allowing misalignment
998 // AND changes how we count interferences. A mis-aligned
999 // double can interfere with TWO aligned pairs, or effectively
1000 // FOUR registers!
1001 if (rm.is_misaligned_pair()) {
1002 lrg._fat_proj = 1;
1003 lrg._is_bound = 1;
1004 }
1005 break;
1006 case Op_RegVectMask:
1007 assert(Matcher::has_predicated_vectors(), "sanity");
1008 assert(RegMask::num_registers(Op_RegVectMask) == RegMask::SlotsPerRegVectMask, "sanity");
1009 lrg.set_num_regs(RegMask::SlotsPerRegVectMask);
1010 lrg.set_reg_pressure(1);
1011 break;
1012 case Op_RegF:
1013 case Op_RegI:
1014 case Op_RegN:
1015 case Op_RegFlags:
1016 case 0: // not an ideal register
1017 lrg.set_num_regs(1);
1018 lrg.set_reg_pressure(1);
1019 break;
1020 case Op_VecA:
1021 assert(Matcher::supports_scalable_vector(), "does not support scalable vector");
1022 assert(RegMask::num_registers(Op_VecA) == RegMask::SlotsPerVecA, "sanity");
1023 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecA), "vector should be aligned");
1024 lrg.set_num_regs(RegMask::SlotsPerVecA);
1025 lrg.set_reg_pressure(1);
1026 break;
1027 case Op_VecS:
1028 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity");
1029 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity");
1030 lrg.set_num_regs(RegMask::SlotsPerVecS);
1031 lrg.set_reg_pressure(1);
1032 break;
1033 case Op_VecD:
1034 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity");
1035 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity");
1036 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned");
1037 lrg.set_num_regs(RegMask::SlotsPerVecD);
1038 lrg.set_reg_pressure(1);
1039 break;
1040 case Op_VecX:
1041 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity");
1042 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity");
1043 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned");
1044 lrg.set_num_regs(RegMask::SlotsPerVecX);
1045 lrg.set_reg_pressure(1);
1046 break;
1047 case Op_VecY:
1048 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity");
1049 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity");
1050 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned");
1051 lrg.set_num_regs(RegMask::SlotsPerVecY);
1052 lrg.set_reg_pressure(1);
1053 break;
1054 case Op_VecZ:
1055 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity");
1056 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity");
1057 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned");
1058 lrg.set_num_regs(RegMask::SlotsPerVecZ);
1059 lrg.set_reg_pressure(1);
1060 break;
1061 default:
1062 ShouldNotReachHere();
1063 }
1064 }
1065
1066 // Now do the same for inputs
1067 uint cnt = n->req();
1068 // Setup for CISC SPILLING
1069 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable;
1070 if( UseCISCSpill && after_aggressive ) {
1071 inp = n->cisc_operand();
1072 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable )
1073 // Convert operand number to edge index number
1074 inp = n->as_Mach()->operand_index(inp);
1075 }
1076
1077 // Prepare register mask for each input
1078 for( uint k = input_edge_start; k < cnt; k++ ) {
1079 uint vreg_in = _lrg_map.live_range_id(n->in(k));
1080 if (!vreg_in) {
1081 continue;
1082 }
1083
1084 // If this instruction is CISC Spillable, add the flags
1085 // bit to its appropriate input
1086 if( UseCISCSpill && after_aggressive && inp == k ) {
1087 #ifndef PRODUCT
1088 if( TraceCISCSpill ) {
1089 tty->print(" use_cisc_RegMask: ");
1090 n->dump();
1091 }
1092 #endif
1093 n->as_Mach()->use_cisc_RegMask();
1094 }
1095
1096 if (is_machine_node && _scheduling_info_generated) {
1097 MachNode* cur_node = n->as_Mach();
1098 // this is cleaned up by register allocation
1099 if (k >= cur_node->num_opnds()) continue;
1100 }
1101
1102 LRG &lrg_in = lrgs(vreg_in);
1103 // // Testing for floating point code shape
1104 // Node *test = n->in(k);
1105 // if( test->is_Mach() ) {
1106 // MachNode *m = test->as_Mach();
1107 // int op = m->ideal_Opcode();
1108 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) {
1109 // int zzz = 1;
1110 // }
1111 // }
1112
1113 // Limit result register mask to acceptable registers.
1114 // Do not limit registers from uncommon uses before
1115 // AggressiveCoalesce. This effectively pre-virtual-splits
1116 // around uncommon uses of common defs.
1117 const RegMask &rm_in = n->in_RegMask(k);
1118 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) {
1119 // Since we are BEFORE aggressive coalesce, leave the register
1120 // mask untrimmed by the call. This encourages more coalescing.
1121 // Later, AFTER aggressive, this live range will have to spill
1122 // but the spiller handles slow-path calls very nicely.
1123 } else {
1124 lrg_in.and_with(rm_in);
1125 }
1126
1127 // Check for bound register masks
1128 const RegMask &lrgmask_in = lrg_in.mask();
1129 uint kreg = n->in(k)->ideal_reg();
1130 bool is_vect = RegMask::is_vector(kreg);
1131 assert(n->in(k)->bottom_type()->isa_vect() == nullptr || is_vect ||
1132 kreg == Op_RegD || kreg == Op_RegL || kreg == Op_RegVectMask,
1133 "vector must be in vector registers");
1134 if (lrgmask_in.is_bound(kreg))
1135 lrg_in._is_bound = 1;
1136
1137 // If this use of a double forces a mis-aligned double,
1138 // flag as '_fat_proj' - really flag as allowing misalignment
1139 // AND changes how we count interferences. A mis-aligned
1140 // double can interfere with TWO aligned pairs, or effectively
1141 // FOUR registers!
1142 #ifdef ASSERT
1143 if (is_vect && !_scheduling_info_generated) {
1144 if (lrg_in.num_regs() != 0) {
1145 assert(lrgmask_in.is_aligned_sets(lrg_in.num_regs()), "vector should be aligned");
1146 assert(!lrg_in._fat_proj, "sanity");
1147 assert(RegMask::num_registers(kreg) == lrg_in.num_regs(), "sanity");
1148 } else {
1149 assert(n->is_Phi(), "not all inputs processed only if Phi");
1150 }
1151 }
1152 #endif
1153 if (!is_vect && lrg_in.num_regs() == 2 && !lrg_in._fat_proj && rm_in.is_misaligned_pair()) {
1154 lrg_in._fat_proj = 1;
1155 lrg_in._is_bound = 1;
1156 }
1157 // if the LRG is an unaligned pair, we will have to spill
1158 // so clear the LRG's register mask if it is not already spilled
1159 if (!is_vect && !n->is_SpillCopy() &&
1160 (lrg_in._def == nullptr || lrg_in.is_multidef() || !lrg_in._def->is_SpillCopy()) &&
1161 lrgmask_in.is_misaligned_pair()) {
1162 lrg_in.clear();
1163 }
1164
1165 // Check for maximum frequency value
1166 if (lrg_in._maxfreq < block->_freq) {
1167 lrg_in._maxfreq = block->_freq;
1168 }
1169
1170 } // End for all allocated inputs
1171 } // end for all instructions
1172 } // end for all blocks
1173
1174 // Final per-liverange setup
1175 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) {
1176 LRG &lrg = lrgs(i2);
1177 assert(!lrg._is_vector || !lrg._fat_proj, "sanity");
1178 if (lrg.num_regs() > 1 && !lrg._fat_proj) {
1179 lrg.clear_to_sets();
1180 }
1181 lrg.compute_set_mask_size();
1182 if (lrg.not_free()) { // Handle case where we lose from the start
1183 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG));
1184 lrg._direct_conflict = 1;
1185 }
1186 lrg.set_degree(0); // no neighbors in IFG yet
1187 }
1188 }
1189
1190 // Set the was-lo-degree bit. Conservative coalescing should not change the
1191 // colorability of the graph. If any live range was of low-degree before
1192 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
1193 // The bit is checked in Simplify.
1194 void PhaseChaitin::set_was_low() {
1195 #ifdef ASSERT
1196 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1197 int size = lrgs(i).num_regs();
1198 uint old_was_lo = lrgs(i)._was_lo;
1199 lrgs(i)._was_lo = 0;
1200 if( lrgs(i).lo_degree() ) {
1201 lrgs(i)._was_lo = 1; // Trivially of low degree
1202 } else { // Else check the Brigg's assertion
1203 // Brigg's observation is that the lo-degree neighbors of a
1204 // hi-degree live range will not interfere with the color choices
1205 // of said hi-degree live range. The Simplify reverse-stack-coloring
1206 // order takes care of the details. Hence you do not have to count
1207 // low-degree neighbors when determining if this guy colors.
1208 int briggs_degree = 0;
1209 IndexSet *s = _ifg->neighbors(i);
1210 IndexSetIterator elements(s);
1211 uint lidx;
1212 while((lidx = elements.next()) != 0) {
1213 if( !lrgs(lidx).lo_degree() )
1214 briggs_degree += MAX2(size,lrgs(lidx).num_regs());
1215 }
1216 if( briggs_degree < lrgs(i).degrees_of_freedom() )
1217 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion
1218 }
1219 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease");
1220 }
1221 #endif
1222 }
1223
1224 // Compute cost/area ratio, in case we spill. Build the lo-degree list.
1225 void PhaseChaitin::cache_lrg_info( ) {
1226 Compile::TracePhase tp(_t_chaitinCacheLRG);
1227
1228 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1229 LRG &lrg = lrgs(i);
1230
1231 // Check for being of low degree: means we can be trivially colored.
1232 // Low degree, dead or must-spill guys just get to simplify right away
1233 if( lrg.lo_degree() ||
1234 !lrg.alive() ||
1235 lrg._must_spill ) {
1236 // Split low degree list into those guys that must get a
1237 // register and those that can go to register or stack.
1238 // The idea is LRGs that can go register or stack color first when
1239 // they have a good chance of getting a register. The register-only
1240 // lo-degree live ranges always get a register.
1241 OptoReg::Name hi_reg = lrg.mask().find_last_elem();
1242 if( OptoReg::is_stack(hi_reg)) { // Can go to stack?
1243 lrg._next = _lo_stk_degree;
1244 _lo_stk_degree = i;
1245 } else {
1246 lrg._next = _lo_degree;
1247 _lo_degree = i;
1248 }
1249 } else { // Else high degree
1250 lrgs(_hi_degree)._prev = i;
1251 lrg._next = _hi_degree;
1252 lrg._prev = 0;
1253 _hi_degree = i;
1254 }
1255 }
1256 }
1257
1258 // Simplify the IFG by removing LRGs of low degree.
1259 void PhaseChaitin::Simplify( ) {
1260 Compile::TracePhase tp(_t_chaitinSimplify);
1261
1262 while( 1 ) { // Repeat till simplified it all
1263 // May want to explore simplifying lo_degree before _lo_stk_degree.
1264 // This might result in more spills coloring into registers during
1265 // Select().
1266 while( _lo_degree || _lo_stk_degree ) {
1267 // If possible, pull from lo_stk first
1268 uint lo;
1269 if( _lo_degree ) {
1270 lo = _lo_degree;
1271 _lo_degree = lrgs(lo)._next;
1272 } else {
1273 lo = _lo_stk_degree;
1274 _lo_stk_degree = lrgs(lo)._next;
1275 }
1276
1277 // Put the simplified guy on the simplified list.
1278 lrgs(lo)._next = _simplified;
1279 _simplified = lo;
1280 // If this guy is "at risk" then mark his current neighbors
1281 if (lrgs(lo)._at_risk && !_ifg->neighbors(lo)->is_empty()) {
1282 IndexSetIterator elements(_ifg->neighbors(lo));
1283 uint datum;
1284 while ((datum = elements.next()) != 0) {
1285 lrgs(datum)._risk_bias = lo;
1286 }
1287 }
1288
1289 // Yank this guy from the IFG.
1290 IndexSet *adj = _ifg->remove_node(lo);
1291 if (adj->is_empty()) {
1292 continue;
1293 }
1294
1295 // If any neighbors' degrees fall below their number of
1296 // allowed registers, then put that neighbor on the low degree
1297 // list. Note that 'degree' can only fall and 'numregs' is
1298 // unchanged by this action. Thus the two are equal at most once,
1299 // so LRGs hit the lo-degree worklist at most once.
1300 IndexSetIterator elements(adj);
1301 uint neighbor;
1302 while ((neighbor = elements.next()) != 0) {
1303 LRG *n = &lrgs(neighbor);
1304 #ifdef ASSERT
1305 if (VerifyRegisterAllocator) {
1306 assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
1307 }
1308 #endif
1309
1310 // Check for just becoming of-low-degree just counting registers.
1311 // _must_spill live ranges are already on the low degree list.
1312 if (n->just_lo_degree() && !n->_must_spill) {
1313 assert(!_ifg->_yanked->test(neighbor), "Cannot move to lo degree twice");
1314 // Pull from hi-degree list
1315 uint prev = n->_prev;
1316 uint next = n->_next;
1317 if (prev) {
1318 lrgs(prev)._next = next;
1319 } else {
1320 _hi_degree = next;
1321 }
1322 lrgs(next)._prev = prev;
1323 n->_next = _lo_degree;
1324 _lo_degree = neighbor;
1325 }
1326 }
1327 } // End of while lo-degree/lo_stk_degree worklist not empty
1328
1329 // Check for got everything: is hi-degree list empty?
1330 if (!_hi_degree) break;
1331
1332 // Time to pick a potential spill guy
1333 uint lo_score = _hi_degree;
1334 double score = lrgs(lo_score).score();
1335 double area = lrgs(lo_score)._area;
1336 double cost = lrgs(lo_score)._cost;
1337 bool bound = lrgs(lo_score)._is_bound;
1338
1339 // Find cheapest guy
1340 DEBUG_ONLY( int lo_no_simplify=0; );
1341 for (uint i = _hi_degree; i; i = lrgs(i)._next) {
1342 assert(!_ifg->_yanked->test(i), "");
1343 // It's just vaguely possible to move hi-degree to lo-degree without
1344 // going through a just-lo-degree stage: If you remove a double from
1345 // a float live range it's degree will drop by 2 and you can skip the
1346 // just-lo-degree stage. It's very rare (shows up after 5000+ methods
1347 // in -Xcomp of Java2Demo). So just choose this guy to simplify next.
1348 if( lrgs(i).lo_degree() ) {
1349 lo_score = i;
1350 break;
1351 }
1352 DEBUG_ONLY( if( lrgs(i)._was_lo ) lo_no_simplify=i; );
1353 double iscore = lrgs(i).score();
1354 double iarea = lrgs(i)._area;
1355 double icost = lrgs(i)._cost;
1356 bool ibound = lrgs(i)._is_bound;
1357
1358 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area
1359 // wins. Ties happen because all live ranges in question have spilled
1360 // a few times before and the spill-score adds a huge number which
1361 // washes out the low order bits. We are choosing the lesser of 2
1362 // evils; in this case pick largest area to spill.
1363 // Ties also happen when live ranges are defined and used only inside
1364 // one block. In which case their area is 0 and score set to max.
1365 // In such case choose bound live range over unbound to free registers
1366 // or with smaller cost to spill.
1367 if ( iscore < score ||
1368 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ||
1369 (iscore == score && iarea == area &&
1370 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) {
1371 lo_score = i;
1372 score = iscore;
1373 area = iarea;
1374 cost = icost;
1375 bound = ibound;
1376 }
1377 }
1378 LRG *lo_lrg = &lrgs(lo_score);
1379 // The live range we choose for spilling is either hi-degree, or very
1380 // rarely it can be low-degree. If we choose a hi-degree live range
1381 // there better not be any lo-degree choices.
1382 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" );
1383
1384 // Pull from hi-degree list
1385 uint prev = lo_lrg->_prev;
1386 uint next = lo_lrg->_next;
1387 if( prev ) lrgs(prev)._next = next;
1388 else _hi_degree = next;
1389 lrgs(next)._prev = prev;
1390 // Jam him on the lo-degree list, despite his high degree.
1391 // Maybe he'll get a color, and maybe he'll spill.
1392 // Only Select() will know.
1393 lrgs(lo_score)._at_risk = true;
1394 _lo_degree = lo_score;
1395 lo_lrg->_next = 0;
1396
1397 } // End of while not simplified everything
1398
1399 }
1400
1401 // Is 'reg' register legal for 'lrg'?
1402 static bool is_legal_reg(LRG& lrg, OptoReg::Name reg) {
1403 if (lrg.mask().can_represent(reg) && lrg.mask().member(reg)) {
1404 // RA uses OptoReg which represent the highest element of a registers set.
1405 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set
1406 // in which XMMd is used by RA to represent such vectors. A double value
1407 // uses [XMM,XMMb] pairs and XMMb is used by RA for it.
1408 // The register mask uses largest bits set of overlapping register sets.
1409 // On x86 with AVX it uses 8 bits for each XMM registers set.
1410 //
1411 // The 'lrg' already has cleared-to-set register mask (done in Select()
1412 // before calling choose_color()). Passing mask.Member(reg) check above
1413 // indicates that the size (num_regs) of 'reg' set is less or equal to
1414 // 'lrg' set size.
1415 // For set size 1 any register which is member of 'lrg' mask is legal.
1416 if (lrg.num_regs()==1)
1417 return true;
1418 // For larger sets only an aligned register with the same set size is legal.
1419 int mask = lrg.num_regs()-1;
1420 if ((reg&mask) == mask)
1421 return true;
1422 }
1423 return false;
1424 }
1425
1426 static OptoReg::Name find_first_set(LRG& lrg, RegMask& mask) {
1427 int num_regs = lrg.num_regs();
1428 OptoReg::Name assigned = mask.find_first_set(lrg, num_regs);
1429
1430 if (lrg.is_scalable()) {
1431 // a physical register is found
1432 if (OptoReg::is_reg(assigned)) {
1433 assert(!lrg.mask().is_offset(),
1434 "offset register masks can only contain stack slots");
1435 return assigned;
1436 }
1437
1438 // find available stack slots for scalable register
1439 if (lrg._is_vector) {
1440 num_regs = lrg.scalable_reg_slots();
1441 // if actual scalable vector register is exactly SlotsPerVecA * 32 bits
1442 if (num_regs == RegMask::SlotsPerVecA) {
1443 return assigned;
1444 }
1445
1446 // mask has been cleared out by clear_to_sets(SlotsPerVecA) before choose_color, but it
1447 // does not work for scalable size. We have to find adjacent scalable_reg_slots() bits
1448 // instead of SlotsPerVecA bits.
1449 assigned = mask.find_first_set(lrg, num_regs); // find highest valid reg
1450 while (OptoReg::is_valid(assigned)) {
1451 assert(mask.can_represent(assigned), "sanity");
1452 // Verify the found reg has scalable_reg_slots() bits set.
1453 if (mask.is_valid_reg(assigned, num_regs)) {
1454 return assigned;
1455 } else {
1456 // Remove more for each iteration
1457 mask.remove(assigned - num_regs + 1); // Unmask the lowest reg
1458 mask.clear_to_sets(RegMask::SlotsPerVecA); // Align by SlotsPerVecA bits
1459 assigned = mask.find_first_set(lrg, num_regs);
1460 }
1461 }
1462 return OptoReg::Bad; // will cause chunk change, and retry next chunk
1463 } else if (lrg._is_predicate) {
1464 assert(num_regs == RegMask::SlotsPerRegVectMask, "scalable predicate register");
1465 num_regs = lrg.scalable_reg_slots();
1466 mask.clear_to_sets(num_regs);
1467 return mask.find_first_set(lrg, num_regs);
1468 }
1469 }
1470
1471 return assigned;
1472 }
1473
1474 OptoReg::Name PhaseChaitin::select_bias_lrg_color(LRG& lrg) {
1475 uint bias_lrg1_idx = _lrg_map.find(lrg._copy_bias);
1476 uint bias_lrg2_idx = _lrg_map.find(lrg._copy_bias2);
1477
1478 // If bias_lrg1 has a color
1479 if (bias_lrg1_idx != 0 && !_ifg->_yanked->test(bias_lrg1_idx)) {
1480 OptoReg::Name reg = lrgs(bias_lrg1_idx).reg();
1481 // and it is legal for lrg
1482 if (is_legal_reg(lrg, reg)) {
1483 return reg;
1484 }
1485 }
1486
1487 // If bias_lrg2 has a color
1488 if (bias_lrg2_idx != 0 && !_ifg->_yanked->test(bias_lrg2_idx)) {
1489 OptoReg::Name reg = lrgs(bias_lrg2_idx).reg();
1490 // and it is legal for lrg
1491 if (is_legal_reg(lrg, reg)) {
1492 return reg;
1493 }
1494 }
1495
1496 uint bias_lrg_idx = 0;
1497 if (bias_lrg1_idx != 0 && bias_lrg2_idx != 0) {
1498 // Since none of the bias live ranges are part of the IFG yet, constrain the
1499 // definition mask with the bias live range with the least degrees of
1500 // freedom. This will increase the chances of register sharing once the bias
1501 // live range becomes part of the IFG.
1502 lrgs(bias_lrg1_idx).compute_set_mask_size();
1503 lrgs(bias_lrg2_idx).compute_set_mask_size();
1504 bias_lrg_idx = lrgs(bias_lrg1_idx).degrees_of_freedom() >
1505 lrgs(bias_lrg2_idx).degrees_of_freedom()
1506 ? bias_lrg2_idx
1507 : bias_lrg1_idx;
1508 } else if (bias_lrg1_idx != 0) {
1509 bias_lrg_idx = bias_lrg1_idx;
1510 } else if (bias_lrg2_idx != 0) {
1511 bias_lrg_idx = bias_lrg2_idx;
1512 }
1513
1514 // Register masks with offset excludes all mask bits before the offset.
1515 // Such masks are mainly used for allocation from stack slots. Constrain the
1516 // register mask of definition live range using bias mask only if
1517 // both masks have zero offset.
1518 if (bias_lrg_idx != 0 && !lrg.mask().is_offset() &&
1519 !lrgs(bias_lrg_idx).mask().is_offset()) {
1520 // Choose a color which is legal for bias_lrg
1521 ResourceMark rm(C->regmask_arena());
1522 RegMask tempmask(lrg.mask(), C->regmask_arena());
1523 tempmask.and_with(lrgs(bias_lrg_idx).mask());
1524 tempmask.clear_to_sets(lrg.num_regs());
1525 OptoReg::Name reg = find_first_set(lrg, tempmask);
1526 if (OptoReg::is_valid(reg)) {
1527 return reg;
1528 }
1529 }
1530 return OptoReg::Bad;
1531 }
1532
1533 // Choose a color using the biasing heuristic
1534 OptoReg::Name PhaseChaitin::bias_color(LRG& lrg) {
1535
1536 // Check for "at_risk" LRG's
1537 uint risk_lrg = _lrg_map.find(lrg._risk_bias);
1538 if (risk_lrg != 0 && !_ifg->neighbors(risk_lrg)->is_empty()) {
1539 // Walk the colored neighbors of the "at_risk" candidate
1540 // Choose a color which is both legal and already taken by a neighbor
1541 // of the "at_risk" candidate in order to improve the chances of the
1542 // "at_risk" candidate of coloring
1543 IndexSetIterator elements(_ifg->neighbors(risk_lrg));
1544 uint datum;
1545 while ((datum = elements.next()) != 0) {
1546 OptoReg::Name reg = lrgs(datum).reg();
1547 // If this LRG's register is legal for us, choose it
1548 if (is_legal_reg(lrg, reg)) {
1549 return reg;
1550 }
1551 }
1552 }
1553
1554 // Try biasing the color with non-interfering bias live range[s].
1555 OptoReg::Name reg = select_bias_lrg_color(lrg);
1556 if (OptoReg::is_valid(reg)) {
1557 return reg;
1558 }
1559
1560 // If no bias info exists, just go with the register selection ordering
1561 if (lrg._is_vector || lrg.num_regs() == 2 || lrg.is_scalable()) {
1562 // Find an aligned set
1563 ResourceMark rm(C->regmask_arena());
1564 RegMask tempmask(lrg.mask(), C->regmask_arena());
1565 return find_first_set(lrg, tempmask);
1566 }
1567
1568 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate
1569 // copy removal to remove many more copies, by preventing a just-assigned
1570 // register from being repeatedly assigned.
1571 reg = lrg.mask().find_first_elem();
1572 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) {
1573 // This 'Remove; find; Insert' idiom is an expensive way to find the
1574 // SECOND element in the mask.
1575 lrg.remove(reg);
1576 OptoReg::Name reg2 = lrg.mask().find_first_elem();
1577 lrg.insert(reg);
1578 if (OptoReg::is_reg(reg2)) {
1579 reg = reg2;
1580 }
1581 }
1582 return reg;
1583 }
1584
1585 // Choose a color in the current chunk
1586 OptoReg::Name PhaseChaitin::choose_color(LRG& lrg) {
1587 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)");
1588 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)");
1589
1590 if( lrg.num_regs() == 1 || // Common Case
1591 !lrg._fat_proj ) // Aligned+adjacent pairs ok
1592 // Use a heuristic to "bias" the color choice
1593 return bias_color(lrg);
1594
1595 assert(!lrg._is_vector, "should be not vector here" );
1596 assert( lrg.num_regs() >= 2, "dead live ranges do not color" );
1597
1598 // Fat-proj case or misaligned double argument.
1599 assert(lrg.compute_mask_size() == lrg.num_regs() ||
1600 lrg.num_regs() == 2,"fat projs exactly color" );
1601 assert(!lrg.mask().is_offset(), "always color in 1st chunk");
1602 // Return the highest element in the set.
1603 return lrg.mask().find_last_elem();
1604 }
1605
1606 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted
1607 // in reverse order of removal. As long as nothing of hi-degree was yanked,
1608 // everything going back is guaranteed a color. Select that color. If some
1609 // hi-degree LRG cannot get a color then we record that we must spill.
1610 uint PhaseChaitin::Select( ) {
1611 Compile::TracePhase tp(_t_chaitinSelect);
1612
1613 uint spill_reg = LRG::SPILL_REG;
1614 _max_reg = OptoReg::Name(0); // Past max register used
1615 while( _simplified ) {
1616 // Pull next LRG from the simplified list - in reverse order of removal
1617 uint lidx = _simplified;
1618 LRG *lrg = &lrgs(lidx);
1619 _simplified = lrg->_next;
1620
1621 #ifndef PRODUCT
1622 if (trace_spilling()) {
1623 ttyLocker ttyl;
1624 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(),
1625 lrg->degrees_of_freedom());
1626 lrg->dump();
1627 }
1628 #endif
1629
1630 // Re-insert into the IFG
1631 _ifg->re_insert(lidx);
1632 if( !lrg->alive() ) continue;
1633 // capture infinitestackedness flag before mask is hacked
1634 const int is_infinite_stack = lrg->mask().is_infinite_stack();
1635
1636 // Yeah, yeah, yeah, I know, I know. I can refactor this
1637 // to avoid the GOTO, although the refactored code will not
1638 // be much clearer. We arrive here IFF we have a stack-based
1639 // live range that cannot color in the current chunk, and it
1640 // has to move into the next free stack chunk.
1641 retry_next_chunk:
1642
1643 // Remove neighbor colors
1644 IndexSet *s = _ifg->neighbors(lidx);
1645 #ifndef PRODUCT
1646 ResourceMark rm(C->regmask_arena());
1647 RegMask orig_mask(lrg->mask(), C->regmask_arena());
1648 #endif
1649
1650 if (!s->is_empty()) {
1651 IndexSetIterator elements(s);
1652 uint neighbor;
1653 while ((neighbor = elements.next()) != 0) {
1654 LRG &nlrg = lrgs(neighbor);
1655 OptoReg::Name nreg = nlrg.reg();
1656 // The neighbor might be a spill_reg. In this case, exclusion of its
1657 // color will be a no-op, since the spill_reg is in outer space. In
1658 // this case, do not exclude the corresponding mask. Later on, if lrg
1659 // runs out of possible colors in its chunk, a new chunk of color may
1660 // be tried, in which case examination of neighbors is started again,
1661 // at retry_next_chunk.
1662 if (nreg < LRG::SPILL_REG) {
1663 #ifndef PRODUCT
1664 uint size = lrg->mask().size();
1665 ResourceMark rm(C->regmask_arena());
1666 RegMask trace_mask(lrg->mask(), C->regmask_arena());
1667 #endif
1668 lrg->subtract_inner(nlrg.mask());
1669 #ifndef PRODUCT
1670 if (trace_spilling() && lrg->mask().size() != size) {
1671 ttyLocker ttyl;
1672 tty->print("L%d ", lidx);
1673 trace_mask.dump();
1674 tty->print(" intersected L%d ", neighbor);
1675 nlrg.mask().dump();
1676 tty->print(" removed ");
1677 trace_mask.subtract(lrg->mask());
1678 trace_mask.dump();
1679 tty->print(" leaving ");
1680 lrg->mask().dump();
1681 tty->cr();
1682 }
1683 #endif
1684 }
1685 }
1686 }
1687
1688 Node* def = lrg->_def;
1689 if (lrg->is_singledef() && !lrg->_is_bound && def->is_Mach()) {
1690 MachNode* mdef = def->as_Mach();
1691 if (Matcher::is_register_biasing_candidate(mdef, 1)) {
1692 Node* in1 = mdef->in(mdef->operand_index(1));
1693 if (in1 != nullptr && lrg->_copy_bias == 0) {
1694 lrg->_copy_bias = _lrg_map.find(in1);
1695 }
1696 }
1697
1698 // For commutative operations, def allocation can also be
1699 // biased towards LRG of second input's def.
1700 if (Matcher::is_register_biasing_candidate(mdef, 2)) {
1701 Node* in2 = mdef->in(mdef->operand_index(2));
1702 if (in2 != nullptr && lrg->_copy_bias2 == 0) {
1703 lrg->_copy_bias2 = _lrg_map.find(in2);
1704 }
1705 }
1706 }
1707
1708 //assert(is_infinite_stack == lrg->mask().is_infinite_stack(), "nbrs must not change InfiniteStackedness");
1709 // Aligned pairs need aligned masks
1710 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1711 if (lrg->num_regs() > 1 && !lrg->_fat_proj) {
1712 lrg->clear_to_sets();
1713 }
1714
1715 // Check if a color is available and if so pick the color
1716 OptoReg::Name reg = choose_color(*lrg);
1717
1718 //---------------
1719 // If we fail to color and the infinite flag is set, we must trigger
1720 // a chunk-rollover event and continue searching for a color in the next set
1721 // of slots (which are all necessarily stack slots, as registers are only in
1722 // the initial chunk)
1723 if (!OptoReg::is_valid(reg) && is_infinite_stack) {
1724 // Bump register mask up to next stack chunk
1725 bool success = lrg->rollover();
1726 if (!success) {
1727 // We should never get here in practice. Bail out in product,
1728 // assert in debug.
1729 assert(false, "the next available stack slots should be within the "
1730 "OptoRegPair range");
1731 C->record_method_not_compilable(
1732 "chunk-rollover outside of OptoRegPair range");
1733 return -1;
1734 }
1735 goto retry_next_chunk;
1736 }
1737
1738 //---------------
1739 // Did we get a color?
1740 else if (OptoReg::is_valid(reg)) {
1741 #ifndef PRODUCT
1742 ResourceMark rm(C->regmask_arena());
1743 RegMask avail_rm(lrg->mask(), C->regmask_arena());
1744 #endif
1745
1746 // Record selected register
1747 lrg->set_reg(reg);
1748
1749 if (reg >= _max_reg) { // Compute max register limit
1750 _max_reg = OptoReg::add(reg, 1);
1751 }
1752
1753 // If the live range is not bound, then we actually had some choices
1754 // to make. In this case, the mask has more bits in it than the colors
1755 // chosen. Restrict the mask to just what was picked.
1756 int n_regs = lrg->num_regs();
1757 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1758 if (n_regs == 1 || !lrg->_fat_proj) {
1759 if (Matcher::supports_scalable_vector()) {
1760 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecA, "sanity");
1761 } else {
1762 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity");
1763 }
1764 lrg->clear(); // Clear the mask
1765 lrg->insert(reg); // Set regmask to match selected reg
1766 // For vectors and pairs, also insert the low bit of the pair
1767 // We always choose the high bit, then mask the low bits by register size
1768 if (lrg->is_scalable() && OptoReg::is_stack(lrg->reg())) { // stack
1769 n_regs = lrg->scalable_reg_slots();
1770 }
1771 for (int i = 1; i < n_regs; i++) {
1772 lrg->insert(OptoReg::add(reg, -i));
1773 }
1774 lrg->set_mask_size(n_regs);
1775 } else { // Else fatproj
1776 // mask must be equal to fatproj bits, by definition
1777 }
1778 #ifndef PRODUCT
1779 if (trace_spilling()) {
1780 ttyLocker ttyl;
1781 tty->print("L%d selected ", lidx);
1782 lrg->mask().dump();
1783 tty->print(" from ");
1784 avail_rm.dump();
1785 tty->cr();
1786 }
1787 #endif
1788 // Note that reg is the highest-numbered register in the newly-bound mask.
1789 } // end color available case
1790
1791 //---------------
1792 // Live range is live and no colors available
1793 else {
1794 assert( lrg->alive(), "" );
1795 assert( !lrg->_fat_proj || lrg->is_multidef() ||
1796 lrg->_def->outcnt() > 0, "fat_proj cannot spill");
1797 assert( !orig_mask.is_infinite_stack(), "infinite stack does not spill" );
1798
1799 // Assign the special spillreg register
1800 lrg->set_reg(OptoReg::Name(spill_reg++));
1801 // Do not empty the regmask; leave mask_size lying around
1802 // for use during Spilling
1803 #ifndef PRODUCT
1804 if( trace_spilling() ) {
1805 ttyLocker ttyl;
1806 tty->print("L%d spilling with neighbors: ", lidx);
1807 s->dump();
1808 DEBUG_ONLY(tty->print(" original mask: "));
1809 DEBUG_ONLY(orig_mask.dump());
1810 dump_lrg(lidx);
1811 }
1812 #endif
1813 } // end spill case
1814
1815 }
1816
1817 return spill_reg-LRG::SPILL_REG; // Return number of spills
1818 }
1819
1820 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
1821 void PhaseChaitin::set_was_spilled( Node *n ) {
1822 if( _spilled_once.test_set(n->_idx) )
1823 _spilled_twice.set(n->_idx);
1824 }
1825
1826 // Convert Ideal spill instructions into proper FramePtr + offset Loads and
1827 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are.
1828 void PhaseChaitin::fixup_spills() {
1829 // This function does only cisc spill work.
1830 if( !UseCISCSpill ) return;
1831
1832 Compile::TracePhase tp(_t_fixupSpills);
1833
1834 // Grab the Frame Pointer
1835 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr);
1836
1837 // For all blocks
1838 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1839 Block* block = _cfg.get_block(i);
1840
1841 // For all instructions in block
1842 uint last_inst = block->end_idx();
1843 for (uint j = 1; j <= last_inst; j++) {
1844 Node* n = block->get_node(j);
1845
1846 // Dead instruction???
1847 assert( n->outcnt() != 0 ||// Nothing dead after post alloc
1848 C->top() == n || // Or the random TOP node
1849 n->is_Proj(), // Or a fat-proj kill node
1850 "No dead instructions after post-alloc" );
1851
1852 int inp = n->cisc_operand();
1853 if( inp != AdlcVMDeps::Not_cisc_spillable ) {
1854 // Convert operand number to edge index number
1855 MachNode *mach = n->as_Mach();
1856 inp = mach->operand_index(inp);
1857 Node *src = n->in(inp); // Value to load or store
1858 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src));
1859 OptoReg::Name src_reg = lrg_cisc.reg();
1860 // Doubles record the HIGH register of an adjacent pair.
1861 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs());
1862 if( OptoReg::is_stack(src_reg) ) { // If input is on stack
1863 // This is a CISC Spill, get stack offset and construct new node
1864 #ifndef PRODUCT
1865 if( TraceCISCSpill ) {
1866 tty->print(" reg-instr: ");
1867 n->dump();
1868 }
1869 #endif
1870 int stk_offset = reg2offset(src_reg);
1871 // Bailout if we might exceed node limit when spilling this instruction
1872 C->check_node_count(0, "out of nodes fixing spills");
1873 if (C->failing()) return;
1874 // Transform node
1875 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach();
1876 cisc->set_req(inp,fp); // Base register is frame pointer
1877 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) {
1878 assert( cisc->oper_input_base() == 2, "Only adding one edge");
1879 cisc->ins_req(1,src); // Requires a memory edge
1880 } else {
1881 // There is no space reserved for a memory edge before the inputs for
1882 // instructions which have "stackSlotX" parameter instead of "memory".
1883 // For example, "MoveF2I_stack_reg". We always need a memory edge from
1884 // src to cisc, else we might schedule cisc before src, loading from a
1885 // spill location before storing the spill. On some platforms, we land
1886 // in this else case because mach->oper_input_base() > 1, i.e. we have
1887 // multiple inputs. In some rare cases there are even multiple memory
1888 // operands, before and after spilling.
1889 // (e.g. spilling "addFPR24_reg_mem" to "addFPR24_mem_cisc")
1890 // In either case, there is no space in the inputs for the memory edge
1891 // so we add an additional precedence / memory edge.
1892 cisc->add_prec(src);
1893 }
1894 block->map_node(cisc, j); // Insert into basic block
1895 n->subsume_by(cisc, C); // Correct graph
1896 //
1897 ++_used_cisc_instructions;
1898 #ifndef PRODUCT
1899 if( TraceCISCSpill ) {
1900 tty->print(" cisc-instr: ");
1901 cisc->dump();
1902 }
1903 #endif
1904 } else {
1905 #ifndef PRODUCT
1906 if( TraceCISCSpill ) {
1907 tty->print(" using reg-instr: ");
1908 n->dump();
1909 }
1910 #endif
1911 ++_unused_cisc_instructions; // input can be on stack
1912 }
1913 }
1914
1915 } // End of for all instructions
1916
1917 } // End of for all blocks
1918 }
1919
1920 // Helper to stretch above; recursively discover the base Node for a
1921 // given derived Node. Easy for AddP-related machine nodes, but needs
1922 // to be recursive for derived Phis.
1923 Node* PhaseChaitin::find_base_for_derived(Node** derived_base_map, Node* derived, uint& maxlrg) {
1924 // See if already computed; if so return it
1925 if (derived_base_map[derived->_idx]) {
1926 return derived_base_map[derived->_idx];
1927 }
1928
1929 #ifdef ASSERT
1930 if (derived->is_Mach() && derived->as_Mach()->ideal_Opcode() == Op_VerifyVectorAlignment) {
1931 // Bypass the verification node
1932 Node* base = find_base_for_derived(derived_base_map, derived->in(1), maxlrg);
1933 derived_base_map[derived->_idx] = base;
1934 return base;
1935 }
1936 #endif
1937
1938 // See if this happens to be a base.
1939 // NOTE: we use TypePtr instead of TypeOopPtr because we can have
1940 // pointers derived from null! These are always along paths that
1941 // can't happen at run-time but the optimizer cannot deduce it so
1942 // we have to handle it gracefully.
1943 assert(!derived->bottom_type()->isa_narrowoop() ||
1944 derived->bottom_type()->make_ptr()->is_ptr()->offset() == 0, "sanity");
1945 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1946 // If its an OOP with a non-zero offset, then it is derived.
1947 if (tj == nullptr || tj->offset() == 0) {
1948 derived_base_map[derived->_idx] = derived;
1949 return derived;
1950 }
1951 // Derived is null+offset? Base is null!
1952 if( derived->is_Con() ) {
1953 Node *base = _matcher.mach_null();
1954 assert(base != nullptr, "sanity");
1955 if (base->in(0) == nullptr) {
1956 // Initialize it once and make it shared:
1957 // set control to _root and place it into Start block
1958 // (where top() node is placed).
1959 base->init_req(0, _cfg.get_root_node());
1960 Block *startb = _cfg.get_block_for_node(C->top());
1961 uint node_pos = startb->find_node(C->top());
1962 startb->insert_node(base, node_pos);
1963 _cfg.map_node_to_block(base, startb);
1964 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet");
1965
1966 // The loadConP0 might have projection nodes depending on architecture
1967 // Add the projection nodes to the CFG
1968 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) {
1969 Node* use = base->fast_out(i);
1970 if (use->is_MachProj()) {
1971 startb->insert_node(use, ++node_pos);
1972 _cfg.map_node_to_block(use, startb);
1973 new_lrg(use, maxlrg++);
1974 }
1975 }
1976 }
1977 if (_lrg_map.live_range_id(base) == 0) {
1978 new_lrg(base, maxlrg++);
1979 }
1980 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");
1981 derived_base_map[derived->_idx] = base;
1982 return base;
1983 }
1984
1985 // Check for AddP-related opcodes
1986 if (!derived->is_Phi()) {
1987 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name());
1988 Node *base = derived->in(AddPNode::Base);
1989 derived_base_map[derived->_idx] = base;
1990 return base;
1991 }
1992
1993 // Recursively find bases for Phis.
1994 // First check to see if we can avoid a base Phi here.
1995 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg);
1996 uint i;
1997 for( i = 2; i < derived->req(); i++ )
1998 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg))
1999 break;
2000 // Went to the end without finding any different bases?
2001 if( i == derived->req() ) { // No need for a base Phi here
2002 derived_base_map[derived->_idx] = base;
2003 return base;
2004 }
2005
2006 // Now we see we need a base-Phi here to merge the bases
2007 const Type *t = base->bottom_type();
2008 base = new PhiNode( derived->in(0), t );
2009 for( i = 1; i < derived->req(); i++ ) {
2010 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg));
2011 t = t->meet(base->in(i)->bottom_type());
2012 }
2013 base->as_Phi()->set_type(t);
2014
2015 // Search the current block for an existing base-Phi
2016 Block *b = _cfg.get_block_for_node(derived);
2017 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi
2018 Node *phi = b->get_node(i);
2019 if( !phi->is_Phi() ) { // Found end of Phis with no match?
2020 b->insert_node(base, i); // Must insert created Phi here as base
2021 _cfg.map_node_to_block(base, b);
2022 new_lrg(base,maxlrg++);
2023 break;
2024 }
2025 // See if Phi matches.
2026 uint j;
2027 for( j = 1; j < base->req(); j++ )
2028 if( phi->in(j) != base->in(j) &&
2029 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different nulls
2030 break;
2031 if( j == base->req() ) { // All inputs match?
2032 base = phi; // Then use existing 'phi' and drop 'base'
2033 break;
2034 }
2035 }
2036
2037
2038 // Cache info for later passes
2039 derived_base_map[derived->_idx] = base;
2040 return base;
2041 }
2042
2043 // At each Safepoint, insert extra debug edges for each pair of derived value/
2044 // base pointer that is live across the Safepoint for oopmap building. The
2045 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the
2046 // required edge set.
2047 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) {
2048 int must_recompute_live = false;
2049 uint maxlrg = _lrg_map.max_lrg_id();
2050 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique());
2051 memset( derived_base_map, 0, sizeof(Node*)*C->unique() );
2052
2053 // For all blocks in RPO do...
2054 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2055 Block* block = _cfg.get_block(i);
2056 // Note use of deep-copy constructor. I cannot hammer the original
2057 // liveout bits, because they are needed by the following coalesce pass.
2058 IndexSet liveout(_live->live(block));
2059
2060 for (uint j = block->end_idx() + 1; j > 1; j--) {
2061 Node* n = block->get_node(j - 1);
2062
2063 // Pre-split compares of loop-phis. Loop-phis form a cycle we would
2064 // like to see in the same register. Compare uses the loop-phi and so
2065 // extends its live range BUT cannot be part of the cycle. If this
2066 // extended live range overlaps with the update of the loop-phi value
2067 // we need both alive at the same time -- which requires at least 1
2068 // copy. But because Intel has only 2-address registers we end up with
2069 // at least 2 copies, one before the loop-phi update instruction and
2070 // one after. Instead we split the input to the compare just after the
2071 // phi.
2072 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) {
2073 Node *phi = n->in(1);
2074 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) {
2075 Block *phi_block = _cfg.get_block_for_node(phi);
2076 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) {
2077 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI];
2078 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask);
2079 insert_proj( phi_block, 1, spill, maxlrg++ );
2080 n->set_req(1,spill);
2081 must_recompute_live = true;
2082 }
2083 }
2084 }
2085
2086 // Get value being defined
2087 uint lidx = _lrg_map.live_range_id(n);
2088 // Ignore the occasional brand-new live range
2089 if (lidx && lidx < _lrg_map.max_lrg_id()) {
2090 // Remove from live-out set
2091 liveout.remove(lidx);
2092
2093 // Copies do not define a new value and so do not interfere.
2094 // Remove the copies source from the liveout set before interfering.
2095 uint idx = n->is_Copy();
2096 if (idx) {
2097 liveout.remove(_lrg_map.live_range_id(n->in(idx)));
2098 }
2099 }
2100
2101 // Found a safepoint?
2102 JVMState *jvms = n->jvms();
2103 if (jvms && !liveout.is_empty()) {
2104 // Now scan for a live derived pointer
2105 IndexSetIterator elements(&liveout);
2106 uint neighbor;
2107 while ((neighbor = elements.next()) != 0) {
2108 // Find reaching DEF for base and derived values
2109 // This works because we are still in SSA during this call.
2110 Node *derived = lrgs(neighbor)._def;
2111 const TypePtr *tj = derived->bottom_type()->isa_ptr();
2112 assert(!derived->bottom_type()->isa_narrowoop() ||
2113 derived->bottom_type()->make_ptr()->is_ptr()->offset() == 0, "sanity");
2114 // If its an OOP with a non-zero offset, then it is derived.
2115 if (tj && tj->offset() != 0 && tj->isa_oop_ptr()) {
2116 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg);
2117 assert(base->_idx < _lrg_map.size(), "");
2118 // Add reaching DEFs of derived pointer and base pointer as a
2119 // pair of inputs
2120 n->add_req(derived);
2121 n->add_req(base);
2122
2123 // See if the base pointer is already live to this point.
2124 // Since I'm working on the SSA form, live-ness amounts to
2125 // reaching def's. So if I find the base's live range then
2126 // I know the base's def reaches here.
2127 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or
2128 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND
2129 (_lrg_map.live_range_id(base) > 0) && // not a constant
2130 _cfg.get_block_for_node(base) != block) { // base not def'd in blk)
2131 // Base pointer is not currently live. Since I stretched
2132 // the base pointer to here and it crosses basic-block
2133 // boundaries, the global live info is now incorrect.
2134 // Recompute live.
2135 must_recompute_live = true;
2136 } // End of if base pointer is not live to debug info
2137 }
2138 } // End of scan all live data for derived ptrs crossing GC point
2139 } // End of if found a GC point
2140
2141 // Make all inputs live
2142 if (!n->is_Phi()) { // Phi function uses come from prior block
2143 for (uint k = 1; k < n->req(); k++) {
2144 uint lidx = _lrg_map.live_range_id(n->in(k));
2145 if (lidx < _lrg_map.max_lrg_id()) {
2146 liveout.insert(lidx);
2147 }
2148 }
2149 }
2150
2151 } // End of forall instructions in block
2152 liveout.clear(); // Free the memory used by liveout.
2153
2154 } // End of forall blocks
2155 _lrg_map.set_max_lrg_id(maxlrg);
2156
2157 // If I created a new live range I need to recompute live
2158 if (maxlrg != _ifg->_maxlrg) {
2159 must_recompute_live = true;
2160 }
2161
2162 return must_recompute_live != 0;
2163 }
2164
2165 // Extend the node to LRG mapping
2166
2167 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) {
2168 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node));
2169 }
2170
2171 #ifndef PRODUCT
2172 void PhaseChaitin::dump(const Node* n) const {
2173 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0;
2174 tty->print("L%d",r);
2175 if (r && n->Opcode() != Op_Phi) {
2176 if( _node_regs ) { // Got a post-allocation copy of allocation?
2177 tty->print("[");
2178 OptoReg::Name second = get_reg_second(n);
2179 if( OptoReg::is_valid(second) ) {
2180 if( OptoReg::is_reg(second) )
2181 tty->print("%s:",Matcher::regName[second]);
2182 else
2183 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second));
2184 }
2185 OptoReg::Name first = get_reg_first(n);
2186 if( OptoReg::is_reg(first) )
2187 tty->print("%s]",Matcher::regName[first]);
2188 else
2189 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first));
2190 } else
2191 n->out_RegMask().dump();
2192 }
2193 tty->print("/N%d\t",n->_idx);
2194 tty->print("%s === ", n->Name());
2195 uint k;
2196 for (k = 0; k < n->req(); k++) {
2197 Node *m = n->in(k);
2198 if (!m) {
2199 tty->print("_ ");
2200 }
2201 else {
2202 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2203 tty->print("L%d",r);
2204 // Data MultiNode's can have projections with no real registers.
2205 // Don't die while dumping them.
2206 int op = n->Opcode();
2207 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) {
2208 if( _node_regs ) {
2209 tty->print("[");
2210 OptoReg::Name second = get_reg_second(n->in(k));
2211 if( OptoReg::is_valid(second) ) {
2212 if( OptoReg::is_reg(second) )
2213 tty->print("%s:",Matcher::regName[second]);
2214 else
2215 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer),
2216 reg2offset_unchecked(second));
2217 }
2218 OptoReg::Name first = get_reg_first(n->in(k));
2219 if( OptoReg::is_reg(first) )
2220 tty->print("%s]",Matcher::regName[first]);
2221 else
2222 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer),
2223 reg2offset_unchecked(first));
2224 } else
2225 n->in_RegMask(k).dump();
2226 }
2227 tty->print("/N%d ",m->_idx);
2228 }
2229 }
2230 if( k < n->len() && n->in(k) ) tty->print("| ");
2231 for( ; k < n->len(); k++ ) {
2232 Node *m = n->in(k);
2233 if(!m) {
2234 break;
2235 }
2236 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2237 tty->print("L%d",r);
2238 tty->print("/N%d ",m->_idx);
2239 }
2240 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty);
2241 else n->dump_spec(tty);
2242 if( _spilled_once.test(n->_idx ) ) {
2243 tty->print(" Spill_1");
2244 if( _spilled_twice.test(n->_idx ) )
2245 tty->print(" Spill_2");
2246 }
2247 tty->print("\n");
2248 }
2249
2250 void PhaseChaitin::dump(const Block* b) const {
2251 b->dump_head(&_cfg);
2252
2253 // For all instructions
2254 for( uint j = 0; j < b->number_of_nodes(); j++ )
2255 dump(b->get_node(j));
2256 // Print live-out info at end of block
2257 if( _live ) {
2258 tty->print("Liveout: ");
2259 IndexSet *live = _live->live(b);
2260 IndexSetIterator elements(live);
2261 tty->print("{");
2262 uint i;
2263 while ((i = elements.next()) != 0) {
2264 tty->print("L%d ", _lrg_map.find_const(i));
2265 }
2266 tty->print_cr("}");
2267 }
2268 tty->print("\n");
2269 }
2270
2271 void PhaseChaitin::dump() const {
2272 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n",
2273 _matcher._new_SP, _framesize );
2274
2275 // For all blocks
2276 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2277 dump(_cfg.get_block(i));
2278 }
2279 // End of per-block dump
2280 tty->print("\n");
2281
2282 if (!_ifg) {
2283 tty->print("(No IFG.)\n");
2284 return;
2285 }
2286
2287 // Dump LRG array
2288 tty->print("--- Live RanGe Array ---\n");
2289 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) {
2290 tty->print("L%d: ",i2);
2291 if (i2 < _ifg->_maxlrg) {
2292 lrgs(i2).dump();
2293 }
2294 else {
2295 tty->print_cr("new LRG");
2296 }
2297 }
2298 tty->cr();
2299
2300 // Dump lo-degree list
2301 tty->print("Lo degree: ");
2302 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next )
2303 tty->print("L%d ",i3);
2304 tty->cr();
2305
2306 // Dump lo-stk-degree list
2307 tty->print("Lo stk degree: ");
2308 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next )
2309 tty->print("L%d ",i4);
2310 tty->cr();
2311
2312 // Dump lo-degree list
2313 tty->print("Hi degree: ");
2314 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next )
2315 tty->print("L%d ",i5);
2316 tty->cr();
2317 }
2318
2319 void PhaseChaitin::dump_degree_lists() const {
2320 // Dump lo-degree list
2321 tty->print("Lo degree: ");
2322 for( uint i = _lo_degree; i; i = lrgs(i)._next )
2323 tty->print("L%d ",i);
2324 tty->cr();
2325
2326 // Dump lo-stk-degree list
2327 tty->print("Lo stk degree: ");
2328 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next )
2329 tty->print("L%d ",i2);
2330 tty->cr();
2331
2332 // Dump lo-degree list
2333 tty->print("Hi degree: ");
2334 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next )
2335 tty->print("L%d ",i3);
2336 tty->cr();
2337 }
2338
2339 void PhaseChaitin::dump_simplified() const {
2340 tty->print("Simplified: ");
2341 for( uint i = _simplified; i; i = lrgs(i)._next )
2342 tty->print("L%d ",i);
2343 tty->cr();
2344 }
2345
2346 static char *print_reg(OptoReg::Name reg, const PhaseChaitin* pc, char* buf, size_t buf_size) {
2347 if ((int)reg < 0)
2348 os::snprintf_checked(buf, buf_size, "<OptoReg::%d>", (int)reg);
2349 else if (OptoReg::is_reg(reg))
2350 strcpy(buf, Matcher::regName[reg]);
2351 else
2352 os::snprintf_checked(buf, buf_size, "%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer),
2353 pc->reg2offset(reg));
2354 return buf+strlen(buf);
2355 }
2356
2357 // Dump a register name into a buffer. Be intelligent if we get called
2358 // before allocation is complete.
2359 char *PhaseChaitin::dump_register(const Node* n, char* buf, size_t buf_size) const {
2360 if( _node_regs ) {
2361 // Post allocation, use direct mappings, no LRG info available
2362 print_reg( get_reg_first(n), this, buf, buf_size);
2363 } else {
2364 uint lidx = _lrg_map.find_const(n); // Grab LRG number
2365 if( !_ifg ) {
2366 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2367 } else if( !lidx ) { // Special, not allocated value
2368 strcpy(buf,"Special");
2369 } else {
2370 if (lrgs(lidx)._is_vector) {
2371 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs()))
2372 print_reg( lrgs(lidx).reg(), this, buf, buf_size); // a bound machine register
2373 else
2374 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2375 } else if( (lrgs(lidx).num_regs() == 1)
2376 ? lrgs(lidx).mask().is_bound1()
2377 : lrgs(lidx).mask().is_bound_pair() ) {
2378 // Hah! We have a bound machine register
2379 print_reg( lrgs(lidx).reg(), this, buf, buf_size);
2380 } else {
2381 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2382 }
2383 }
2384 }
2385 return buf+strlen(buf);
2386 }
2387
2388 void PhaseChaitin::dump_for_spill_split_recycle() const {
2389 if( WizardMode && (PrintCompilation || PrintOpto) ) {
2390 // Display which live ranges need to be split and the allocator's state
2391 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt);
2392 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) {
2393 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
2394 tty->print("L%d: ", bidx);
2395 lrgs(bidx).dump();
2396 }
2397 }
2398 tty->cr();
2399 dump();
2400 }
2401 }
2402
2403 void PhaseChaitin::dump_frame() const {
2404 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer);
2405 const TypeTuple *domain = C->tf()->domain_cc();
2406 const int argcnt = domain->cnt() - TypeFunc::Parms;
2407
2408 // Incoming arguments in registers dump
2409 for( int k = 0; k < argcnt; k++ ) {
2410 OptoReg::Name parmreg = _matcher._parm_regs[k].first();
2411 if( OptoReg::is_reg(parmreg)) {
2412 const char *reg_name = OptoReg::regname(parmreg);
2413 tty->print("#r%3.3d %s", parmreg, reg_name);
2414 parmreg = _matcher._parm_regs[k].second();
2415 if( OptoReg::is_reg(parmreg)) {
2416 tty->print(":%s", OptoReg::regname(parmreg));
2417 }
2418 tty->print(" : parm %d: ", k);
2419 domain->field_at(k + TypeFunc::Parms)->dump();
2420 tty->cr();
2421 }
2422 }
2423
2424 // Check for un-owned padding above incoming args
2425 OptoReg::Name reg = _matcher._new_SP;
2426 if( reg > _matcher._in_arg_limit ) {
2427 reg = OptoReg::add(reg, -1);
2428 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg));
2429 }
2430
2431 // Incoming argument area dump
2432 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots());
2433 while( reg > begin_in_arg ) {
2434 reg = OptoReg::add(reg, -1);
2435 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2436 int j;
2437 for( j = 0; j < argcnt; j++) {
2438 if( _matcher._parm_regs[j].first() == reg ||
2439 _matcher._parm_regs[j].second() == reg ) {
2440 tty->print("parm %d: ",j);
2441 domain->field_at(j + TypeFunc::Parms)->dump();
2442 tty->cr();
2443 break;
2444 }
2445 }
2446 if( j >= argcnt )
2447 tty->print_cr("HOLE, owned by SELF");
2448 }
2449
2450 // Old outgoing preserve area
2451 while( reg > _matcher._old_SP ) {
2452 reg = OptoReg::add(reg, -1);
2453 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg));
2454 }
2455
2456 // Old SP
2457 tty->print_cr("# -- Old %s -- Framesize: %d --",fp,
2458 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize);
2459
2460 // Preserve area dump
2461 int fixed_slots = C->fixed_slots();
2462 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots());
2463 OptoReg::Name return_addr = _matcher.return_addr();
2464
2465 reg = OptoReg::add(reg, -1);
2466 while (OptoReg::is_stack(reg)) {
2467 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2468 if (return_addr == reg) {
2469 tty->print_cr("return address");
2470 } else if (reg >= begin_in_preserve) {
2471 // Preserved slots are present on x86
2472 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word))
2473 tty->print_cr("saved fp register");
2474 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) &&
2475 VerifyStackAtCalls)
2476 tty->print_cr("<Majik cookie> +VerifyStackAtCalls");
2477 else
2478 tty->print_cr("in_preserve");
2479 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) {
2480 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg));
2481 } else {
2482 tty->print_cr("pad2, stack alignment");
2483 }
2484 reg = OptoReg::add(reg, -1);
2485 }
2486
2487 // Spill area dump
2488 reg = OptoReg::add(_matcher._new_SP, _framesize );
2489 while( reg > _matcher._out_arg_limit ) {
2490 reg = OptoReg::add(reg, -1);
2491 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg));
2492 }
2493
2494 // Outgoing argument area dump
2495 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) {
2496 reg = OptoReg::add(reg, -1);
2497 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg));
2498 }
2499
2500 // Outgoing new preserve area
2501 while( reg > _matcher._new_SP ) {
2502 reg = OptoReg::add(reg, -1);
2503 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg));
2504 }
2505 tty->print_cr("#");
2506 }
2507
2508 void PhaseChaitin::dump_bb(uint pre_order) const {
2509 tty->print_cr("---dump of B%d---",pre_order);
2510 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2511 Block* block = _cfg.get_block(i);
2512 if (block->_pre_order == pre_order) {
2513 dump(block);
2514 }
2515 }
2516 }
2517
2518 void PhaseChaitin::dump_lrg(uint lidx, bool defs_only) const {
2519 tty->print_cr("---dump of L%d---",lidx);
2520
2521 if (_ifg) {
2522 if (lidx >= _lrg_map.max_lrg_id()) {
2523 tty->print("Attempt to print live range index beyond max live range.\n");
2524 return;
2525 }
2526 tty->print("L%d: ",lidx);
2527 if (lidx < _ifg->_maxlrg) {
2528 lrgs(lidx).dump();
2529 } else {
2530 tty->print_cr("new LRG");
2531 }
2532 }
2533 if( _ifg && lidx < _ifg->_maxlrg) {
2534 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx));
2535 _ifg->neighbors(lidx)->dump();
2536 tty->cr();
2537 }
2538 // For all blocks
2539 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2540 Block* block = _cfg.get_block(i);
2541 int dump_once = 0;
2542
2543 // For all instructions
2544 for( uint j = 0; j < block->number_of_nodes(); j++ ) {
2545 Node *n = block->get_node(j);
2546 if (_lrg_map.find_const(n) == lidx) {
2547 if (!dump_once++) {
2548 tty->cr();
2549 block->dump_head(&_cfg);
2550 }
2551 dump(n);
2552 continue;
2553 }
2554 if (!defs_only) {
2555 uint cnt = n->req();
2556 for( uint k = 1; k < cnt; k++ ) {
2557 Node *m = n->in(k);
2558 if (!m) {
2559 continue; // be robust in the dumper
2560 }
2561 if (_lrg_map.find_const(m) == lidx) {
2562 if (!dump_once++) {
2563 tty->cr();
2564 block->dump_head(&_cfg);
2565 }
2566 dump(n);
2567 }
2568 }
2569 }
2570 }
2571 } // End of per-block dump
2572 tty->cr();
2573 }
2574 #endif // not PRODUCT
2575
2576 #ifdef ASSERT
2577 // Verify that base pointers and derived pointers are still sane.
2578 void PhaseChaitin::verify_base_ptrs(ResourceArea* a) const {
2579 Unique_Node_List worklist(a);
2580 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2581 Block* block = _cfg.get_block(i);
2582 for (uint j = block->end_idx() + 1; j > 1; j--) {
2583 Node* n = block->get_node(j-1);
2584 if (n->is_Phi()) {
2585 break;
2586 }
2587 // Found a safepoint?
2588 if (n->is_MachSafePoint()) {
2589 MachSafePointNode* sfpt = n->as_MachSafePoint();
2590 JVMState* jvms = sfpt->jvms();
2591 if (jvms != nullptr) {
2592 // Now scan for a live derived pointer
2593 if (jvms->oopoff() < sfpt->req()) {
2594 // Check each derived/base pair
2595 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
2596 Node* check = sfpt->in(idx);
2597 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
2598 // search upwards through spills and spill phis for AddP
2599 worklist.clear();
2600 worklist.push(check);
2601 uint k = 0;
2602 while (k < worklist.size()) {
2603 check = worklist.at(k);
2604 assert(check, "Bad base or derived pointer");
2605 // See PhaseChaitin::find_base_for_derived() for all cases.
2606 int isc = check->is_Copy();
2607 if (isc) {
2608 worklist.push(check->in(isc));
2609 } else if (check->is_Phi()) {
2610 for (uint m = 1; m < check->req(); m++) {
2611 worklist.push(check->in(m));
2612 }
2613 } else if (check->is_Con()) {
2614 if (is_derived && check->bottom_type()->is_ptr()->offset() != 0) {
2615 // Derived is null+non-zero offset, base must be null.
2616 assert(check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad derived pointer");
2617 } else {
2618 assert(check->bottom_type()->is_ptr()->offset() == 0, "Bad base pointer");
2619 // Base either ConP(nullptr) or loadConP
2620 if (check->is_Mach()) {
2621 assert(check->as_Mach()->ideal_Opcode() == Op_ConP, "Bad base pointer");
2622 } else {
2623 assert(check->Opcode() == Op_ConP &&
2624 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad base pointer");
2625 }
2626 }
2627 } else if (check->bottom_type()->is_ptr()->offset() == 0) {
2628 if (check->is_Proj() || (check->is_Mach() &&
2629 (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
2630 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
2631 check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
2632 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
2633 #ifdef _LP64
2634 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP) ||
2635 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN) ||
2636 (UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass) ||
2637 #endif // _LP64
2638 check->as_Mach()->ideal_Opcode() == Op_LoadP ||
2639 check->as_Mach()->ideal_Opcode() == Op_LoadKlass))) {
2640 // Valid nodes
2641 } else {
2642 check->dump();
2643 assert(false, "Bad base or derived pointer");
2644 }
2645 } else {
2646 assert(is_derived, "Bad base pointer");
2647 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP, "Bad derived pointer");
2648 }
2649 k++;
2650 assert(k < 100000, "Derived pointer checking in infinite loop");
2651 } // End while
2652 }
2653 } // End of check for derived pointers
2654 } // End of Kcheck for debug info
2655 } // End of if found a safepoint
2656 } // End of forall instructions in block
2657 } // End of forall blocks
2658 }
2659
2660 // Verify that graphs and base pointers are still sane.
2661 void PhaseChaitin::verify(ResourceArea* a, bool verify_ifg) const {
2662 if (VerifyRegisterAllocator) {
2663 _cfg.verify();
2664 if (C->failing()) {
2665 return;
2666 }
2667 verify_base_ptrs(a);
2668 if (verify_ifg) {
2669 _ifg->verify(this);
2670 }
2671 }
2672 }
2673 #endif // ASSERT
2674
2675 int PhaseChaitin::_final_loads = 0;
2676 int PhaseChaitin::_final_stores = 0;
2677 int PhaseChaitin::_final_memoves= 0;
2678 int PhaseChaitin::_final_copies = 0;
2679 double PhaseChaitin::_final_load_cost = 0;
2680 double PhaseChaitin::_final_store_cost = 0;
2681 double PhaseChaitin::_final_memove_cost= 0;
2682 double PhaseChaitin::_final_copy_cost = 0;
2683 int PhaseChaitin::_conserv_coalesce = 0;
2684 int PhaseChaitin::_conserv_coalesce_pair = 0;
2685 int PhaseChaitin::_conserv_coalesce_trie = 0;
2686 int PhaseChaitin::_conserv_coalesce_quad = 0;
2687 int PhaseChaitin::_post_alloc = 0;
2688 int PhaseChaitin::_lost_opp_pp_coalesce = 0;
2689 int PhaseChaitin::_lost_opp_cflow_coalesce = 0;
2690 int PhaseChaitin::_used_cisc_instructions = 0;
2691 int PhaseChaitin::_unused_cisc_instructions = 0;
2692 int PhaseChaitin::_allocator_attempts = 0;
2693 int PhaseChaitin::_allocator_successes = 0;
2694
2695 #ifndef PRODUCT
2696 uint PhaseChaitin::_high_pressure = 0;
2697 uint PhaseChaitin::_low_pressure = 0;
2698
2699 void PhaseChaitin::print_chaitin_statistics() {
2700 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);
2701 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);
2702 tty->print_cr("Adjusted spill cost = %7.0f.",
2703 _final_load_cost*4.0 + _final_store_cost * 2.0 +
2704 _final_copy_cost*1.0 + _final_memove_cost*12.0);
2705 tty->print("Conservatively coalesced %d copies, %d pairs",
2706 _conserv_coalesce, _conserv_coalesce_pair);
2707 if( _conserv_coalesce_trie || _conserv_coalesce_quad )
2708 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad);
2709 tty->print_cr(", %d post alloc.", _post_alloc);
2710 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce )
2711 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.",
2712 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce );
2713 if( _used_cisc_instructions || _unused_cisc_instructions )
2714 tty->print_cr("Used cisc instruction %d, remained in register %d",
2715 _used_cisc_instructions, _unused_cisc_instructions);
2716 if( _allocator_successes != 0 )
2717 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes);
2718 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure);
2719 }
2720 #endif // not PRODUCT