1 /*
2 * Copyright (c) 2000, 2024, 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 "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "compiler/oopMap.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/block.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/chaitin.hpp"
35 #include "opto/coalesce.hpp"
36 #include "opto/connode.hpp"
37 #include "opto/idealGraphPrinter.hpp"
38 #include "opto/indexSet.hpp"
39 #include "opto/machnode.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/opcodes.hpp"
43 #include "opto/rootnode.hpp"
44 #include "utilities/align.hpp"
45
46 #ifndef PRODUCT
47 void LRG::dump() const {
48 ttyLocker ttyl;
49 tty->print("%d ",num_regs());
50 _mask.dump();
51 if( _msize_valid ) {
52 if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size);
53 else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size());
54 } else {
55 tty->print(", #?(%d) ",_mask.Size());
56 }
57
58 tty->print("EffDeg: ");
59 if( _degree_valid ) tty->print( "%d ", _eff_degree );
60 else tty->print("? ");
61
62 if( is_multidef() ) {
63 tty->print("MultiDef ");
64 if (_defs != nullptr) {
65 tty->print("(");
66 for (int i = 0; i < _defs->length(); i++) {
67 tty->print("N%d ", _defs->at(i)->_idx);
68 }
69 tty->print(") ");
70 }
71 }
72 else if( _def == nullptr ) tty->print("Dead ");
73 else tty->print("Def: N%d ",_def->_idx);
74
75 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score());
76 // Flags
77 if( _is_oop ) tty->print("Oop ");
78 if( _is_float ) tty->print("Float ");
79 if( _is_vector ) tty->print("Vector ");
80 if( _is_predicate ) tty->print("Predicate ");
81 if( _is_scalable ) tty->print("Scalable ");
82 if( _was_spilled1 ) tty->print("Spilled ");
83 if( _was_spilled2 ) tty->print("Spilled2 ");
84 if( _direct_conflict ) tty->print("Direct_conflict ");
85 if( _fat_proj ) tty->print("Fat ");
86 if( _was_lo ) tty->print("Lo ");
87 if( _has_copy ) tty->print("Copy ");
88 if( _at_risk ) tty->print("Risk ");
89
90 if( _must_spill ) tty->print("Must_spill ");
91 if( _is_bound ) tty->print("Bound ");
92 if( _msize_valid ) {
93 if( _degree_valid && lo_degree() ) tty->print("Trivial ");
94 }
95
96 tty->cr();
97 }
98 #endif
99
100 // Compute score from cost and area. Low score is best to spill.
101 static double raw_score( double cost, double area ) {
102 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5;
103 }
104
105 double LRG::score() const {
106 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost.
107 // Bigger area lowers score, encourages spilling this live range.
108 // Bigger cost raise score, prevents spilling this live range.
109 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer
110 // to turn a divide by a constant into a multiply by the reciprical).
111 double score = raw_score( _cost, _area);
112
113 // Account for area. Basically, LRGs covering large areas are better
114 // to spill because more other LRGs get freed up.
115 if( _area == 0.0 ) // No area? Then no progress to spill
116 return 1e35;
117
118 if( _was_spilled2 ) // If spilled once before, we are unlikely
119 return score + 1e30; // to make progress again.
120
121 if( _cost >= _area*3.0 ) // Tiny area relative to cost
122 return score + 1e17; // Probably no progress to spill
123
124 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost
125 return score + 1e10; // Likely no progress to spill
126
127 return score;
128 }
129
130 #define NUMBUCKS 3
131
132 // Straight out of Tarjan's union-find algorithm
133 uint LiveRangeMap::find_compress(uint lrg) {
134 uint cur = lrg;
135 uint next = _uf_map.at(cur);
136 while (next != cur) { // Scan chain of equivalences
137 assert( next < cur, "always union smaller");
138 cur = next; // until find a fixed-point
139 next = _uf_map.at(cur);
140 }
141
142 // Core of union-find algorithm: update chain of
143 // equivalences to be equal to the root.
144 while (lrg != next) {
145 uint tmp = _uf_map.at(lrg);
146 _uf_map.at_put(lrg, next);
147 lrg = tmp;
148 }
149 return lrg;
150 }
151
152 // Reset the Union-Find map to identity
153 void LiveRangeMap::reset_uf_map(uint max_lrg_id) {
154 _max_lrg_id= max_lrg_id;
155 // Force the Union-Find mapping to be at least this large
156 _uf_map.at_put_grow(_max_lrg_id, 0);
157 // Initialize it to be the ID mapping.
158 for (uint i = 0; i < _max_lrg_id; ++i) {
159 _uf_map.at_put(i, i);
160 }
161 }
162
163 // Make all Nodes map directly to their final live range; no need for
164 // the Union-Find mapping after this call.
165 void LiveRangeMap::compress_uf_map_for_nodes() {
166 // For all Nodes, compress mapping
167 uint unique = _names.length();
168 for (uint i = 0; i < unique; ++i) {
169 uint lrg = _names.at(i);
170 uint compressed_lrg = find(lrg);
171 if (lrg != compressed_lrg) {
172 _names.at_put(i, compressed_lrg);
173 }
174 }
175 }
176
177 // Like Find above, but no path compress, so bad asymptotic behavior
178 uint LiveRangeMap::find_const(uint lrg) const {
179 if (!lrg) {
180 return lrg; // Ignore the zero LRG
181 }
182
183 // Off the end? This happens during debugging dumps when you got
184 // brand new live ranges but have not told the allocator yet.
185 if (lrg >= _max_lrg_id) {
186 return lrg;
187 }
188
189 uint next = _uf_map.at(lrg);
190 while (next != lrg) { // Scan chain of equivalences
191 assert(next < lrg, "always union smaller");
192 lrg = next; // until find a fixed-point
193 next = _uf_map.at(lrg);
194 }
195 return next;
196 }
197
198 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated)
199 : PhaseRegAlloc(unique, cfg, matcher,
200 #ifndef PRODUCT
201 print_chaitin_statistics
202 #else
203 nullptr
204 #endif
205 )
206 , _live(nullptr)
207 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0)
208 , _oldphi(unique)
209 #ifndef PRODUCT
210 , _trace_spilling(C->directive()->TraceSpillingOption)
211 #endif
212 , _lrg_map(Thread::current()->resource_area(), unique)
213 , _scheduling_info_generated(scheduling_info_generated)
214 , _sched_int_pressure(0, Matcher::int_pressure_limit())
215 , _sched_float_pressure(0, Matcher::float_pressure_limit())
216 , _scratch_int_pressure(0, Matcher::int_pressure_limit())
217 , _scratch_float_pressure(0, Matcher::float_pressure_limit())
218 {
219 Compile::TracePhase tp("ctorChaitin", &timers[_t_ctorChaitin]);
220
221 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency());
222
223 // Build a list of basic blocks, sorted by frequency
224 // Experiment with sorting strategies to speed compilation
225 uint nr_blocks = _cfg.number_of_blocks();
226 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket
227 Block **buckets[NUMBUCKS]; // Array of buckets
228 uint buckcnt[NUMBUCKS]; // Array of bucket counters
229 double buckval[NUMBUCKS]; // Array of bucket value cutoffs
230
231 // The space which our buckets point into.
232 Block** start = NEW_RESOURCE_ARRAY(Block *, nr_blocks*NUMBUCKS);
233
234 for (uint i = 0; i < NUMBUCKS; i++) {
235 buckets[i] = &start[i*nr_blocks];
236 buckcnt[i] = 0;
237 // Bump by three orders of magnitude each time
238 cutoff *= 0.001;
239 buckval[i] = cutoff;
240 }
241
242 // Sort blocks into buckets
243 for (uint i = 0; i < nr_blocks; i++) {
244 for (uint j = 0; j < NUMBUCKS; j++) {
245 double bval = buckval[j];
246 Block* blk = _cfg.get_block(i);
247 if (j == NUMBUCKS - 1 || blk->_freq > bval) {
248 uint cnt = buckcnt[j];
249 // Assign block to end of list for appropriate bucket
250 buckets[j][cnt] = blk;
251 buckcnt[j] = cnt+1;
252 break; // kick out of inner loop
253 }
254 }
255 }
256
257 // Squash the partially filled buckets together into the first one.
258 static_assert(NUMBUCKS >= 2, "must"); // If this isn't true then it'll mess up the squashing.
259 Block** offset = &buckets[0][buckcnt[0]];
260 for (int i = 1; i < NUMBUCKS; i++) {
261 ::memmove(offset, buckets[i], buckcnt[i]*sizeof(Block*));
262 offset += buckcnt[i];
263 }
264 assert((&buckets[0][0] + nr_blocks) == offset, "should be");
265
266 // Free the now unused memory
267 FREE_RESOURCE_ARRAY(Block*, buckets[1], (NUMBUCKS-1)*nr_blocks);
268 // Finally, point the _blks to our memory
269 _blks = buckets[0];
270
271 #ifdef ASSERT
272 uint blkcnt = 0;
273 for (uint i = 0; i < NUMBUCKS; i++) {
274 blkcnt += buckcnt[i];
275 }
276 assert(blkcnt == nr_blocks, "Block array not totally filled");
277 #endif
278 }
279
280 // union 2 sets together.
281 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) {
282 uint src = _lrg_map.find(src_n);
283 uint dst = _lrg_map.find(dst_n);
284 assert(src, "");
285 assert(dst, "");
286 assert(src < _lrg_map.max_lrg_id(), "oob");
287 assert(dst < _lrg_map.max_lrg_id(), "oob");
288 assert(src < dst, "always union smaller");
289 _lrg_map.uf_map(dst, src);
290 }
291
292 void PhaseChaitin::new_lrg(const Node *x, uint lrg) {
293 // Make the Node->LRG mapping
294 _lrg_map.extend(x->_idx,lrg);
295 // Make the Union-Find mapping an identity function
296 _lrg_map.uf_extend(lrg, lrg);
297 }
298
299
300 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) {
301 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections");
302 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); )
303 int found_projs = 0;
304 uint cnt = orig->outcnt();
305 for (uint i = 0; i < cnt; i++) {
306 Node* proj = orig->raw_out(i);
307 if (proj->is_MachProj()) {
308 assert(proj->outcnt() == 0, "only kill projections are expected here");
309 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections");
310 found_projs++;
311 // Copy kill projections after the cloned node
312 Node* kills = proj->clone();
313 kills->set_req(0, copy);
314 b->insert_node(kills, idx++);
315 _cfg.map_node_to_block(kills, b);
316 new_lrg(kills, max_lrg_id++);
317 }
318 }
319 return found_projs;
320 }
321
322 // Renumber the live ranges to compact them. Makes the IFG smaller.
323 void PhaseChaitin::compact() {
324 Compile::TracePhase tp("chaitinCompact", &timers[_t_chaitinCompact]);
325
326 // Current the _uf_map contains a series of short chains which are headed
327 // by a self-cycle. All the chains run from big numbers to little numbers.
328 // The Find() call chases the chains & shortens them for the next Find call.
329 // We are going to change this structure slightly. Numbers above a moving
330 // wave 'i' are unchanged. Numbers below 'j' point directly to their
331 // compacted live range with no further chaining. There are no chains or
332 // cycles below 'i', so the Find call no longer works.
333 uint j=1;
334 uint i;
335 for (i = 1; i < _lrg_map.max_lrg_id(); i++) {
336 uint lr = _lrg_map.uf_live_range_id(i);
337 // Ignore unallocated live ranges
338 if (!lr) {
339 continue;
340 }
341 assert(lr <= i, "");
342 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr));
343 }
344 // Now change the Node->LR mapping to reflect the compacted names
345 uint unique = _lrg_map.size();
346 for (i = 0; i < unique; i++) {
347 uint lrg_id = _lrg_map.live_range_id(i);
348 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id));
349 }
350
351 // Reset the Union-Find mapping
352 _lrg_map.reset_uf_map(j);
353 }
354
355 void PhaseChaitin::Register_Allocate() {
356
357 // Above the OLD FP (and in registers) are the incoming arguments. Stack
358 // slots in this area are called "arg_slots". Above the NEW FP (and in
359 // registers) is the outgoing argument area; above that is the spill/temp
360 // area. These are all "frame_slots". Arg_slots start at the zero
361 // stack_slots and count up to the known arg_size. Frame_slots start at
362 // the stack_slot #arg_size and go up. After allocation I map stack
363 // slots to actual offsets. Stack-slots in the arg_slot area are biased
364 // by the frame_size; stack-slots in the frame_slot area are biased by 0.
365
366 _trip_cnt = 0;
367 _alternate = 0;
368 _matcher._allocation_started = true;
369
370 ResourceArea split_arena(mtCompiler); // Arena for Split local resources
371 ResourceArea live_arena(mtCompiler); // Arena for liveness & IFG info
372 ResourceMark rm(&live_arena);
373
374 // Need live-ness for the IFG; need the IFG for coalescing. If the
375 // liveness is JUST for coalescing, then I can get some mileage by renaming
376 // all copy-related live ranges low and then using the max copy-related
377 // live range as a cut-off for LIVE and the IFG. In other words, I can
378 // build a subset of LIVE and IFG just for copies.
379 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false);
380
381 // Need IFG for coalescing and coloring
382 PhaseIFG ifg(&live_arena);
383 _ifg = &ifg;
384
385 // Come out of SSA world to the Named world. Assign (virtual) registers to
386 // Nodes. Use the same register for all inputs and the output of PhiNodes
387 // - effectively ending SSA form. This requires either coalescing live
388 // ranges or inserting copies. For the moment, we insert "virtual copies"
389 // - we pretend there is a copy prior to each Phi in predecessor blocks.
390 // We will attempt to coalesce such "virtual copies" before we manifest
391 // them for real.
392 de_ssa();
393
394 #ifdef ASSERT
395 // Verify the graph before RA.
396 verify(&live_arena);
397 #endif
398
399 {
400 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
401 _live = nullptr; // Mark live as being not available
402 rm.reset_to_mark(); // Reclaim working storage
403 IndexSet::reset_memory(C, &live_arena);
404 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG
405 gather_lrg_masks( false ); // Collect LRG masks
406 live.compute(_lrg_map.max_lrg_id()); // Compute liveness
407 _live = &live; // Mark LIVE as being available
408 }
409
410 // Base pointers are currently "used" by instructions which define new
411 // derived pointers. This makes base pointers live up to the where the
412 // derived pointer is made, but not beyond. Really, they need to be live
413 // across any GC point where the derived value is live. So this code looks
414 // at all the GC points, and "stretches" the live range of any base pointer
415 // to the GC point.
416 if (stretch_base_pointer_live_ranges(&live_arena)) {
417 Compile::TracePhase tp("computeLive (sbplr)", &timers[_t_computeLive]);
418 // Since some live range stretched, I need to recompute live
419 _live = nullptr;
420 rm.reset_to_mark(); // Reclaim working storage
421 IndexSet::reset_memory(C, &live_arena);
422 ifg.init(_lrg_map.max_lrg_id());
423 gather_lrg_masks(false);
424 live.compute(_lrg_map.max_lrg_id());
425 _live = &live;
426 }
427 // Create the interference graph using virtual copies
428 build_ifg_virtual(); // Include stack slots this time
429
430 // The IFG is/was triangular. I am 'squaring it up' so Union can run
431 // faster. Union requires a 'for all' operation which is slow on the
432 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' -
433 // meaning I can visit all the Nodes neighbors less than a Node in time
434 // O(# of neighbors), but I have to visit all the Nodes greater than a
435 // given Node and search them for an instance, i.e., time O(#MaxLRG)).
436 _ifg->SquareUp();
437
438 // Aggressive (but pessimistic) copy coalescing.
439 // This pass works on virtual copies. Any virtual copies which are not
440 // coalesced get manifested as actual copies
441 {
442 Compile::TracePhase tp("chaitinCoalesce1", &timers[_t_chaitinCoalesce1]);
443
444 PhaseAggressiveCoalesce coalesce(*this);
445 coalesce.coalesce_driver();
446 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do
447 // not match the Phi itself, insert a copy.
448 coalesce.insert_copies(_matcher);
449 if (C->failing()) {
450 return;
451 }
452 }
453
454 // After aggressive coalesce, attempt a first cut at coloring.
455 // To color, we need the IFG and for that we need LIVE.
456 {
457 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
458 _live = nullptr;
459 rm.reset_to_mark(); // Reclaim working storage
460 IndexSet::reset_memory(C, &live_arena);
461 ifg.init(_lrg_map.max_lrg_id());
462 gather_lrg_masks( true );
463 live.compute(_lrg_map.max_lrg_id());
464 _live = &live;
465 }
466
467 // Build physical interference graph
468 uint must_spill = 0;
469 must_spill = build_ifg_physical(&live_arena);
470 // If we have a guaranteed spill, might as well spill now
471 if (must_spill) {
472 if(!_lrg_map.max_lrg_id()) {
473 return;
474 }
475 // Bail out if unique gets too large (ie - unique > MaxNodeLimit)
476 C->check_node_count(10*must_spill, "out of nodes before split");
477 if (C->failing()) {
478 return;
479 }
480
481 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
482 if (C->failing()) {
483 return;
484 }
485 _lrg_map.set_max_lrg_id(new_max_lrg_id);
486 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
487 // or we failed to split
488 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split");
489 if (C->failing()) {
490 return;
491 }
492
493 NOT_PRODUCT(C->verify_graph_edges();)
494
495 compact(); // Compact LRGs; return new lower max lrg
496
497 {
498 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
499 _live = nullptr;
500 rm.reset_to_mark(); // Reclaim working storage
501 IndexSet::reset_memory(C, &live_arena);
502 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph
503 gather_lrg_masks( true ); // Collect intersect mask
504 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE
505 _live = &live;
506 }
507 build_ifg_physical(&live_arena);
508 _ifg->SquareUp();
509 _ifg->Compute_Effective_Degree();
510 // Only do conservative coalescing if requested
511 if (OptoCoalesce) {
512 Compile::TracePhase tp("chaitinCoalesce2", &timers[_t_chaitinCoalesce2]);
513 // Conservative (and pessimistic) copy coalescing of those spills
514 PhaseConservativeCoalesce coalesce(*this);
515 // If max live ranges greater than cutoff, don't color the stack.
516 // This cutoff can be larger than below since it is only done once.
517 coalesce.coalesce_driver();
518 }
519 _lrg_map.compress_uf_map_for_nodes();
520
521 #ifdef ASSERT
522 verify(&live_arena, true);
523 #endif
524 } else {
525 ifg.SquareUp();
526 ifg.Compute_Effective_Degree();
527 #ifdef ASSERT
528 set_was_low();
529 #endif
530 }
531
532 // Prepare for Simplify & Select
533 cache_lrg_info(); // Count degree of LRGs
534
535 // Simplify the InterFerence Graph by removing LRGs of low degree.
536 // LRGs of low degree are trivially colorable.
537 Simplify();
538
539 // Select colors by re-inserting LRGs back into the IFG in reverse order.
540 // Return whether or not something spills.
541 uint spills = Select( );
542
543 // If we spill, split and recycle the entire thing
544 while( spills ) {
545 if( _trip_cnt++ > 24 ) {
546 DEBUG_ONLY( dump_for_spill_split_recycle(); )
547 if( _trip_cnt > 27 ) {
548 C->record_method_not_compilable("failed spill-split-recycle sanity check");
549 return;
550 }
551 }
552
553 if (!_lrg_map.max_lrg_id()) {
554 return;
555 }
556 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
557 if (C->failing()) {
558 return;
559 }
560 _lrg_map.set_max_lrg_id(new_max_lrg_id);
561 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
562 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split");
563 if (C->failing()) {
564 return;
565 }
566
567 compact(); // Compact LRGs; return new lower max lrg
568
569 // Nuke the live-ness and interference graph and LiveRanGe info
570 {
571 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
572 _live = nullptr;
573 rm.reset_to_mark(); // Reclaim working storage
574 IndexSet::reset_memory(C, &live_arena);
575 ifg.init(_lrg_map.max_lrg_id());
576
577 // Create LiveRanGe array.
578 // Intersect register masks for all USEs and DEFs
579 gather_lrg_masks(true);
580 live.compute(_lrg_map.max_lrg_id());
581 _live = &live;
582 }
583 must_spill = build_ifg_physical(&live_arena);
584 _ifg->SquareUp();
585 _ifg->Compute_Effective_Degree();
586
587 // Only do conservative coalescing if requested
588 if (OptoCoalesce) {
589 Compile::TracePhase tp("chaitinCoalesce3", &timers[_t_chaitinCoalesce3]);
590 // Conservative (and pessimistic) copy coalescing
591 PhaseConservativeCoalesce coalesce(*this);
592 // Check for few live ranges determines how aggressive coalesce is.
593 coalesce.coalesce_driver();
594 }
595 _lrg_map.compress_uf_map_for_nodes();
596 #ifdef ASSERT
597 verify(&live_arena, true);
598 #endif
599 cache_lrg_info(); // Count degree of LRGs
600
601 // Simplify the InterFerence Graph by removing LRGs of low degree.
602 // LRGs of low degree are trivially colorable.
603 Simplify();
604
605 // Select colors by re-inserting LRGs back into the IFG in reverse order.
606 // Return whether or not something spills.
607 spills = Select();
608 }
609
610 // Count number of Simplify-Select trips per coloring success.
611 _allocator_attempts += _trip_cnt + 1;
612 _allocator_successes += 1;
613
614 // Peephole remove copies
615 post_allocate_copy_removal();
616
617 // Merge multidefs if multiple defs representing the same value are used in a single block.
618 merge_multidefs();
619
620 #ifdef ASSERT
621 // Verify the graph after RA.
622 verify(&live_arena);
623 #endif
624
625 // max_reg is past the largest *register* used.
626 // Convert that to a frame_slot number.
627 if (_max_reg <= _matcher._new_SP) {
628 _framesize = C->out_preserve_stack_slots();
629 }
630 else {
631 _framesize = _max_reg -_matcher._new_SP;
632 }
633 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough");
634
635 // This frame must preserve the required fp alignment
636 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots());
637 assert(_framesize <= 1000000, "sanity check");
638 #ifndef PRODUCT
639 _total_framesize += _framesize;
640 if ((int)_framesize > _max_framesize) {
641 _max_framesize = _framesize;
642 }
643 #endif
644
645 // Convert CISC spills
646 fixup_spills();
647
648 // Log regalloc results
649 CompileLog* log = Compile::current()->log();
650 if (log != nullptr) {
651 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing());
652 }
653
654 if (C->failing()) {
655 return;
656 }
657
658 NOT_PRODUCT(C->verify_graph_edges();)
659
660 // Move important info out of the live_arena to longer lasting storage.
661 alloc_node_regs(_lrg_map.size());
662 for (uint i=0; i < _lrg_map.size(); i++) {
663 if (_lrg_map.live_range_id(i)) { // Live range associated with Node?
664 LRG &lrg = lrgs(_lrg_map.live_range_id(i));
665 if (!lrg.alive()) {
666 set_bad(i);
667 } else if ((lrg.num_regs() == 1 && !lrg.is_scalable()) ||
668 (lrg.is_scalable() && lrg.scalable_reg_slots() == 1)) {
669 set1(i, lrg.reg());
670 } else { // Must be a register-set
671 if (!lrg._fat_proj) { // Must be aligned adjacent register set
672 // Live ranges record the highest register in their mask.
673 // We want the low register for the AD file writer's convenience.
674 OptoReg::Name hi = lrg.reg(); // Get hi register
675 int num_regs = lrg.num_regs();
676 if (lrg.is_scalable() && OptoReg::is_stack(hi)) {
677 // For scalable vector registers, when they are allocated in physical
678 // registers, num_regs is RegMask::SlotsPerVecA for reg mask of scalable
679 // vector. If they are allocated on stack, we need to get the actual
680 // num_regs, which reflects the physical length of scalable registers.
681 num_regs = lrg.scalable_reg_slots();
682 }
683 if (num_regs == 1) {
684 set1(i, hi);
685 } else {
686 OptoReg::Name lo = OptoReg::add(hi, (1 - num_regs)); // Find lo
687 // We have to use pair [lo,lo+1] even for wide vectors/vmasks because
688 // the rest of code generation works only with pairs. It is safe
689 // since for registers encoding only 'lo' is used.
690 // Second reg from pair is used in ScheduleAndBundle with vector max
691 // size 8 which corresponds to registers pair.
692 // It is also used in BuildOopMaps but oop operations are not
693 // vectorized.
694 set2(i, lo);
695 }
696 } else { // Misaligned; extract 2 bits
697 OptoReg::Name hi = lrg.reg(); // Get hi register
698 lrg.Remove(hi); // Yank from mask
699 int lo = lrg.mask().find_first_elem(); // Find lo
700 set_pair(i, hi, lo);
701 }
702 }
703 if( lrg._is_oop ) _node_oops.set(i);
704 } else {
705 set_bad(i);
706 }
707 }
708
709 // Done!
710 _live = nullptr;
711 _ifg = nullptr;
712 C->set_indexSet_arena(nullptr); // ResourceArea is at end of scope
713 }
714
715 void PhaseChaitin::de_ssa() {
716 // Set initial Names for all Nodes. Most Nodes get the virtual register
717 // number. A few get the ZERO live range number. These do not
718 // get allocated, but instead rely on correct scheduling to ensure that
719 // only one instance is simultaneously live at a time.
720 uint lr_counter = 1;
721 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
722 Block* block = _cfg.get_block(i);
723 uint cnt = block->number_of_nodes();
724
725 // Handle all the normal Nodes in the block
726 for( uint j = 0; j < cnt; j++ ) {
727 Node *n = block->get_node(j);
728 // Pre-color to the zero live range, or pick virtual register
729 const RegMask &rm = n->out_RegMask();
730 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0);
731 }
732 }
733
734 // Reset the Union-Find mapping to be identity
735 _lrg_map.reset_uf_map(lr_counter);
736 }
737
738 void PhaseChaitin::mark_ssa() {
739 // Use ssa names to populate the live range maps or if no mask
740 // is available, use the 0 entry.
741 uint max_idx = 0;
742 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
743 Block* block = _cfg.get_block(i);
744 uint cnt = block->number_of_nodes();
745
746 // Handle all the normal Nodes in the block
747 for ( uint j = 0; j < cnt; j++ ) {
748 Node *n = block->get_node(j);
749 // Pre-color to the zero live range, or pick virtual register
750 const RegMask &rm = n->out_RegMask();
751 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0);
752 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx;
753 }
754 }
755 _lrg_map.set_max_lrg_id(max_idx+1);
756
757 // Reset the Union-Find mapping to be identity
758 _lrg_map.reset_uf_map(max_idx+1);
759 }
760
761
762 // Gather LiveRanGe information, including register masks. Modification of
763 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce.
764 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) {
765
766 // Nail down the frame pointer live range
767 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr));
768 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite
769
770 // For all blocks
771 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
772 Block* block = _cfg.get_block(i);
773
774 // For all instructions
775 for (uint j = 1; j < block->number_of_nodes(); j++) {
776 Node* n = block->get_node(j);
777 uint input_edge_start =1; // Skip control most nodes
778 bool is_machine_node = false;
779 if (n->is_Mach()) {
780 is_machine_node = true;
781 input_edge_start = n->as_Mach()->oper_input_base();
782 }
783 uint idx = n->is_Copy();
784
785 // Get virtual register number, same as LiveRanGe index
786 uint vreg = _lrg_map.live_range_id(n);
787 LRG& lrg = lrgs(vreg);
788 if (vreg) { // No vreg means un-allocable (e.g. memory)
789
790 // Check for float-vs-int live range (used in register-pressure
791 // calculations)
792 const Type *n_type = n->bottom_type();
793 if (n_type->is_floatingpoint()) {
794 lrg._is_float = 1;
795 }
796
797 // Check for twice prior spilling. Once prior spilling might have
798 // spilled 'soft', 2nd prior spill should have spilled 'hard' and
799 // further spilling is unlikely to make progress.
800 if (_spilled_once.test(n->_idx)) {
801 lrg._was_spilled1 = 1;
802 if (_spilled_twice.test(n->_idx)) {
803 lrg._was_spilled2 = 1;
804 }
805 }
806
807 #ifndef PRODUCT
808 // Collect bits not used by product code, but which may be useful for
809 // debugging.
810
811 // Collect has-copy bit
812 if (idx) {
813 lrg._has_copy = 1;
814 uint clidx = _lrg_map.live_range_id(n->in(idx));
815 LRG& copy_src = lrgs(clidx);
816 copy_src._has_copy = 1;
817 }
818
819 if (trace_spilling() && lrg._def != nullptr) {
820 // collect defs for MultiDef printing
821 if (lrg._defs == nullptr) {
822 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, nullptr);
823 lrg._defs->append(lrg._def);
824 }
825 lrg._defs->append(n);
826 }
827 #endif
828
829 // Check for a single def LRG; these can spill nicely
830 // via rematerialization. Flag as null for no def found
831 // yet, or 'n' for single def or -1 for many defs.
832 lrg._def = lrg._def ? NodeSentinel : n;
833
834 // Limit result register mask to acceptable registers
835 const RegMask &rm = n->out_RegMask();
836 lrg.AND( rm );
837
838 uint ireg = n->ideal_reg();
839 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP,
840 "oops must be in Op_RegP's" );
841
842 // Check for vector live range (only if vector register is used).
843 // On SPARC vector uses RegD which could be misaligned so it is not
844 // processes as vector in RA.
845 if (RegMask::is_vector(ireg)) {
846 lrg._is_vector = 1;
847 if (Matcher::implements_scalable_vector && ireg == Op_VecA) {
848 assert(Matcher::supports_scalable_vector(), "scalable vector should be supported");
849 lrg._is_scalable = 1;
850 // For scalable vector, when it is allocated in physical register,
851 // num_regs is RegMask::SlotsPerVecA for reg mask,
852 // which may not be the actual physical register size.
853 // If it is allocated in stack, we need to get the actual
854 // physical length of scalable vector register.
855 lrg.set_scalable_reg_slots(Matcher::scalable_vector_reg_size(T_FLOAT));
856 }
857 }
858
859 if (ireg == Op_RegVectMask) {
860 assert(Matcher::has_predicated_vectors(), "predicated vector should be supported");
861 lrg._is_predicate = 1;
862 if (Matcher::supports_scalable_vector()) {
863 lrg._is_scalable = 1;
864 // For scalable predicate, when it is allocated in physical register,
865 // num_regs is RegMask::SlotsPerRegVectMask for reg mask,
866 // which may not be the actual physical register size.
867 // If it is allocated in stack, we need to get the actual
868 // physical length of scalable predicate register.
869 lrg.set_scalable_reg_slots(Matcher::scalable_predicate_reg_slots());
870 }
871 }
872 assert(n_type->isa_vect() == nullptr || lrg._is_vector ||
873 ireg == Op_RegD || ireg == Op_RegL || ireg == Op_RegVectMask,
874 "vector must be in vector registers");
875
876 // Check for bound register masks
877 const RegMask &lrgmask = lrg.mask();
878 if (lrgmask.is_bound(ireg)) {
879 lrg._is_bound = 1;
880 }
881
882 // Check for maximum frequency value
883 if (lrg._maxfreq < block->_freq) {
884 lrg._maxfreq = block->_freq;
885 }
886
887 // Check for oop-iness, or long/double
888 // Check for multi-kill projection
889 switch (ireg) {
890 case MachProjNode::fat_proj:
891 // Fat projections have size equal to number of registers killed
892 lrg.set_num_regs(rm.Size());
893 lrg.set_reg_pressure(lrg.num_regs());
894 lrg._fat_proj = 1;
895 lrg._is_bound = 1;
896 break;
897 case Op_RegP:
898 #ifdef _LP64
899 lrg.set_num_regs(2); // Size is 2 stack words
900 #else
901 lrg.set_num_regs(1); // Size is 1 stack word
902 #endif
903 // Register pressure is tracked relative to the maximum values
904 // suggested for that platform, INTPRESSURE and FLOATPRESSURE,
905 // and relative to other types which compete for the same regs.
906 //
907 // The following table contains suggested values based on the
908 // architectures as defined in each .ad file.
909 // INTPRESSURE and FLOATPRESSURE may be tuned differently for
910 // compile-speed or performance.
911 // Note1:
912 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1
913 // since .ad registers are defined as high and low halves.
914 // These reg_pressure values remain compatible with the code
915 // in is_high_pressure() which relates get_invalid_mask_size(),
916 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE.
917 // Note2:
918 // SPARC -d32 has 24 registers available for integral values,
919 // but only 10 of these are safe for 64-bit longs.
920 // Using set_reg_pressure(2) for both int and long means
921 // the allocator will believe it can fit 26 longs into
922 // registers. Using 2 for longs and 1 for ints means the
923 // allocator will attempt to put 52 integers into registers.
924 // The settings below limit this problem to methods with
925 // many long values which are being run on 32-bit SPARC.
926 //
927 // ------------------- reg_pressure --------------------
928 // Each entry is reg_pressure_per_value,number_of_regs
929 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE
930 // IA32 2 1 1 1 1 6 6
931 // IA64 1 1 1 1 1 50 41
932 // SPARC 2 2 2 2 2 48 (24) 52 (26)
933 // SPARCV9 2 2 2 2 2 48 (24) 52 (26)
934 // AMD64 1 1 1 1 1 14 15
935 // -----------------------------------------------------
936 lrg.set_reg_pressure(1); // normally one value per register
937 if( n_type->isa_oop_ptr() ) {
938 lrg._is_oop = 1;
939 }
940 break;
941 case Op_RegL: // Check for long or double
942 case Op_RegD:
943 lrg.set_num_regs(2);
944 // Define platform specific register pressure
945 #if defined(ARM32)
946 lrg.set_reg_pressure(2);
947 #elif defined(IA32)
948 if( ireg == Op_RegL ) {
949 lrg.set_reg_pressure(2);
950 } else {
951 lrg.set_reg_pressure(1);
952 }
953 #else
954 lrg.set_reg_pressure(1); // normally one value per register
955 #endif
956 // If this def of a double forces a mis-aligned double,
957 // flag as '_fat_proj' - really flag as allowing misalignment
958 // AND changes how we count interferences. A mis-aligned
959 // double can interfere with TWO aligned pairs, or effectively
960 // FOUR registers!
961 if (rm.is_misaligned_pair()) {
962 lrg._fat_proj = 1;
963 lrg._is_bound = 1;
964 }
965 break;
966 case Op_RegVectMask:
967 assert(Matcher::has_predicated_vectors(), "sanity");
968 assert(RegMask::num_registers(Op_RegVectMask) == RegMask::SlotsPerRegVectMask, "sanity");
969 lrg.set_num_regs(RegMask::SlotsPerRegVectMask);
970 lrg.set_reg_pressure(1);
971 break;
972 case Op_RegF:
973 case Op_RegI:
974 case Op_RegN:
975 case Op_RegFlags:
976 case 0: // not an ideal register
977 lrg.set_num_regs(1);
978 lrg.set_reg_pressure(1);
979 break;
980 case Op_VecA:
981 assert(Matcher::supports_scalable_vector(), "does not support scalable vector");
982 assert(RegMask::num_registers(Op_VecA) == RegMask::SlotsPerVecA, "sanity");
983 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecA), "vector should be aligned");
984 lrg.set_num_regs(RegMask::SlotsPerVecA);
985 lrg.set_reg_pressure(1);
986 break;
987 case Op_VecS:
988 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity");
989 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity");
990 lrg.set_num_regs(RegMask::SlotsPerVecS);
991 lrg.set_reg_pressure(1);
992 break;
993 case Op_VecD:
994 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity");
995 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity");
996 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned");
997 lrg.set_num_regs(RegMask::SlotsPerVecD);
998 lrg.set_reg_pressure(1);
999 break;
1000 case Op_VecX:
1001 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity");
1002 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity");
1003 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned");
1004 lrg.set_num_regs(RegMask::SlotsPerVecX);
1005 lrg.set_reg_pressure(1);
1006 break;
1007 case Op_VecY:
1008 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity");
1009 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity");
1010 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned");
1011 lrg.set_num_regs(RegMask::SlotsPerVecY);
1012 lrg.set_reg_pressure(1);
1013 break;
1014 case Op_VecZ:
1015 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity");
1016 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity");
1017 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned");
1018 lrg.set_num_regs(RegMask::SlotsPerVecZ);
1019 lrg.set_reg_pressure(1);
1020 break;
1021 default:
1022 ShouldNotReachHere();
1023 }
1024 }
1025
1026 // Now do the same for inputs
1027 uint cnt = n->req();
1028 // Setup for CISC SPILLING
1029 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable;
1030 if( UseCISCSpill && after_aggressive ) {
1031 inp = n->cisc_operand();
1032 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable )
1033 // Convert operand number to edge index number
1034 inp = n->as_Mach()->operand_index(inp);
1035 }
1036
1037 // Prepare register mask for each input
1038 for( uint k = input_edge_start; k < cnt; k++ ) {
1039 uint vreg = _lrg_map.live_range_id(n->in(k));
1040 if (!vreg) {
1041 continue;
1042 }
1043
1044 // If this instruction is CISC Spillable, add the flags
1045 // bit to its appropriate input
1046 if( UseCISCSpill && after_aggressive && inp == k ) {
1047 #ifndef PRODUCT
1048 if( TraceCISCSpill ) {
1049 tty->print(" use_cisc_RegMask: ");
1050 n->dump();
1051 }
1052 #endif
1053 n->as_Mach()->use_cisc_RegMask();
1054 }
1055
1056 if (is_machine_node && _scheduling_info_generated) {
1057 MachNode* cur_node = n->as_Mach();
1058 // this is cleaned up by register allocation
1059 if (k >= cur_node->num_opnds()) continue;
1060 }
1061
1062 LRG &lrg = lrgs(vreg);
1063 // // Testing for floating point code shape
1064 // Node *test = n->in(k);
1065 // if( test->is_Mach() ) {
1066 // MachNode *m = test->as_Mach();
1067 // int op = m->ideal_Opcode();
1068 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) {
1069 // int zzz = 1;
1070 // }
1071 // }
1072
1073 // Limit result register mask to acceptable registers.
1074 // Do not limit registers from uncommon uses before
1075 // AggressiveCoalesce. This effectively pre-virtual-splits
1076 // around uncommon uses of common defs.
1077 const RegMask &rm = n->in_RegMask(k);
1078 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) {
1079 // Since we are BEFORE aggressive coalesce, leave the register
1080 // mask untrimmed by the call. This encourages more coalescing.
1081 // Later, AFTER aggressive, this live range will have to spill
1082 // but the spiller handles slow-path calls very nicely.
1083 } else {
1084 lrg.AND( rm );
1085 }
1086
1087 // Check for bound register masks
1088 const RegMask &lrgmask = lrg.mask();
1089 uint kreg = n->in(k)->ideal_reg();
1090 bool is_vect = RegMask::is_vector(kreg);
1091 assert(n->in(k)->bottom_type()->isa_vect() == nullptr || is_vect ||
1092 kreg == Op_RegD || kreg == Op_RegL || kreg == Op_RegVectMask,
1093 "vector must be in vector registers");
1094 if (lrgmask.is_bound(kreg))
1095 lrg._is_bound = 1;
1096
1097 // If this use of a double forces a mis-aligned double,
1098 // flag as '_fat_proj' - really flag as allowing misalignment
1099 // AND changes how we count interferences. A mis-aligned
1100 // double can interfere with TWO aligned pairs, or effectively
1101 // FOUR registers!
1102 #ifdef ASSERT
1103 if (is_vect && !_scheduling_info_generated) {
1104 if (lrg.num_regs() != 0) {
1105 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned");
1106 assert(!lrg._fat_proj, "sanity");
1107 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity");
1108 } else {
1109 assert(n->is_Phi(), "not all inputs processed only if Phi");
1110 }
1111 }
1112 #endif
1113 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) {
1114 lrg._fat_proj = 1;
1115 lrg._is_bound = 1;
1116 }
1117 // if the LRG is an unaligned pair, we will have to spill
1118 // so clear the LRG's register mask if it is not already spilled
1119 if (!is_vect && !n->is_SpillCopy() &&
1120 (lrg._def == nullptr || lrg.is_multidef() || !lrg._def->is_SpillCopy()) &&
1121 lrgmask.is_misaligned_pair()) {
1122 lrg.Clear();
1123 }
1124
1125 // Check for maximum frequency value
1126 if (lrg._maxfreq < block->_freq) {
1127 lrg._maxfreq = block->_freq;
1128 }
1129
1130 } // End for all allocated inputs
1131 } // end for all instructions
1132 } // end for all blocks
1133
1134 // Final per-liverange setup
1135 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) {
1136 LRG &lrg = lrgs(i2);
1137 assert(!lrg._is_vector || !lrg._fat_proj, "sanity");
1138 if (lrg.num_regs() > 1 && !lrg._fat_proj) {
1139 lrg.clear_to_sets();
1140 }
1141 lrg.compute_set_mask_size();
1142 if (lrg.not_free()) { // Handle case where we lose from the start
1143 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG));
1144 lrg._direct_conflict = 1;
1145 }
1146 lrg.set_degree(0); // no neighbors in IFG yet
1147 }
1148 }
1149
1150 // Set the was-lo-degree bit. Conservative coalescing should not change the
1151 // colorability of the graph. If any live range was of low-degree before
1152 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
1153 // The bit is checked in Simplify.
1154 void PhaseChaitin::set_was_low() {
1155 #ifdef ASSERT
1156 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1157 int size = lrgs(i).num_regs();
1158 uint old_was_lo = lrgs(i)._was_lo;
1159 lrgs(i)._was_lo = 0;
1160 if( lrgs(i).lo_degree() ) {
1161 lrgs(i)._was_lo = 1; // Trivially of low degree
1162 } else { // Else check the Brigg's assertion
1163 // Brigg's observation is that the lo-degree neighbors of a
1164 // hi-degree live range will not interfere with the color choices
1165 // of said hi-degree live range. The Simplify reverse-stack-coloring
1166 // order takes care of the details. Hence you do not have to count
1167 // low-degree neighbors when determining if this guy colors.
1168 int briggs_degree = 0;
1169 IndexSet *s = _ifg->neighbors(i);
1170 IndexSetIterator elements(s);
1171 uint lidx;
1172 while((lidx = elements.next()) != 0) {
1173 if( !lrgs(lidx).lo_degree() )
1174 briggs_degree += MAX2(size,lrgs(lidx).num_regs());
1175 }
1176 if( briggs_degree < lrgs(i).degrees_of_freedom() )
1177 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion
1178 }
1179 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease");
1180 }
1181 #endif
1182 }
1183
1184 // Compute cost/area ratio, in case we spill. Build the lo-degree list.
1185 void PhaseChaitin::cache_lrg_info( ) {
1186 Compile::TracePhase tp("chaitinCacheLRG", &timers[_t_chaitinCacheLRG]);
1187
1188 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1189 LRG &lrg = lrgs(i);
1190
1191 // Check for being of low degree: means we can be trivially colored.
1192 // Low degree, dead or must-spill guys just get to simplify right away
1193 if( lrg.lo_degree() ||
1194 !lrg.alive() ||
1195 lrg._must_spill ) {
1196 // Split low degree list into those guys that must get a
1197 // register and those that can go to register or stack.
1198 // The idea is LRGs that can go register or stack color first when
1199 // they have a good chance of getting a register. The register-only
1200 // lo-degree live ranges always get a register.
1201 OptoReg::Name hi_reg = lrg.mask().find_last_elem();
1202 if( OptoReg::is_stack(hi_reg)) { // Can go to stack?
1203 lrg._next = _lo_stk_degree;
1204 _lo_stk_degree = i;
1205 } else {
1206 lrg._next = _lo_degree;
1207 _lo_degree = i;
1208 }
1209 } else { // Else high degree
1210 lrgs(_hi_degree)._prev = i;
1211 lrg._next = _hi_degree;
1212 lrg._prev = 0;
1213 _hi_degree = i;
1214 }
1215 }
1216 }
1217
1218 // Simplify the IFG by removing LRGs of low degree.
1219 void PhaseChaitin::Simplify( ) {
1220 Compile::TracePhase tp("chaitinSimplify", &timers[_t_chaitinSimplify]);
1221
1222 while( 1 ) { // Repeat till simplified it all
1223 // May want to explore simplifying lo_degree before _lo_stk_degree.
1224 // This might result in more spills coloring into registers during
1225 // Select().
1226 while( _lo_degree || _lo_stk_degree ) {
1227 // If possible, pull from lo_stk first
1228 uint lo;
1229 if( _lo_degree ) {
1230 lo = _lo_degree;
1231 _lo_degree = lrgs(lo)._next;
1232 } else {
1233 lo = _lo_stk_degree;
1234 _lo_stk_degree = lrgs(lo)._next;
1235 }
1236
1237 // Put the simplified guy on the simplified list.
1238 lrgs(lo)._next = _simplified;
1239 _simplified = lo;
1240 // If this guy is "at risk" then mark his current neighbors
1241 if (lrgs(lo)._at_risk && !_ifg->neighbors(lo)->is_empty()) {
1242 IndexSetIterator elements(_ifg->neighbors(lo));
1243 uint datum;
1244 while ((datum = elements.next()) != 0) {
1245 lrgs(datum)._risk_bias = lo;
1246 }
1247 }
1248
1249 // Yank this guy from the IFG.
1250 IndexSet *adj = _ifg->remove_node(lo);
1251 if (adj->is_empty()) {
1252 continue;
1253 }
1254
1255 // If any neighbors' degrees fall below their number of
1256 // allowed registers, then put that neighbor on the low degree
1257 // list. Note that 'degree' can only fall and 'numregs' is
1258 // unchanged by this action. Thus the two are equal at most once,
1259 // so LRGs hit the lo-degree worklist at most once.
1260 IndexSetIterator elements(adj);
1261 uint neighbor;
1262 while ((neighbor = elements.next()) != 0) {
1263 LRG *n = &lrgs(neighbor);
1264 #ifdef ASSERT
1265 if (VerifyRegisterAllocator) {
1266 assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
1267 }
1268 #endif
1269
1270 // Check for just becoming of-low-degree just counting registers.
1271 // _must_spill live ranges are already on the low degree list.
1272 if (n->just_lo_degree() && !n->_must_spill) {
1273 assert(!_ifg->_yanked->test(neighbor), "Cannot move to lo degree twice");
1274 // Pull from hi-degree list
1275 uint prev = n->_prev;
1276 uint next = n->_next;
1277 if (prev) {
1278 lrgs(prev)._next = next;
1279 } else {
1280 _hi_degree = next;
1281 }
1282 lrgs(next)._prev = prev;
1283 n->_next = _lo_degree;
1284 _lo_degree = neighbor;
1285 }
1286 }
1287 } // End of while lo-degree/lo_stk_degree worklist not empty
1288
1289 // Check for got everything: is hi-degree list empty?
1290 if (!_hi_degree) break;
1291
1292 // Time to pick a potential spill guy
1293 uint lo_score = _hi_degree;
1294 double score = lrgs(lo_score).score();
1295 double area = lrgs(lo_score)._area;
1296 double cost = lrgs(lo_score)._cost;
1297 bool bound = lrgs(lo_score)._is_bound;
1298
1299 // Find cheapest guy
1300 debug_only( int lo_no_simplify=0; );
1301 for (uint i = _hi_degree; i; i = lrgs(i)._next) {
1302 assert(!_ifg->_yanked->test(i), "");
1303 // It's just vaguely possible to move hi-degree to lo-degree without
1304 // going through a just-lo-degree stage: If you remove a double from
1305 // a float live range it's degree will drop by 2 and you can skip the
1306 // just-lo-degree stage. It's very rare (shows up after 5000+ methods
1307 // in -Xcomp of Java2Demo). So just choose this guy to simplify next.
1308 if( lrgs(i).lo_degree() ) {
1309 lo_score = i;
1310 break;
1311 }
1312 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; );
1313 double iscore = lrgs(i).score();
1314 double iarea = lrgs(i)._area;
1315 double icost = lrgs(i)._cost;
1316 bool ibound = lrgs(i)._is_bound;
1317
1318 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area
1319 // wins. Ties happen because all live ranges in question have spilled
1320 // a few times before and the spill-score adds a huge number which
1321 // washes out the low order bits. We are choosing the lesser of 2
1322 // evils; in this case pick largest area to spill.
1323 // Ties also happen when live ranges are defined and used only inside
1324 // one block. In which case their area is 0 and score set to max.
1325 // In such case choose bound live range over unbound to free registers
1326 // or with smaller cost to spill.
1327 if ( iscore < score ||
1328 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ||
1329 (iscore == score && iarea == area &&
1330 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) {
1331 lo_score = i;
1332 score = iscore;
1333 area = iarea;
1334 cost = icost;
1335 bound = ibound;
1336 }
1337 }
1338 LRG *lo_lrg = &lrgs(lo_score);
1339 // The live range we choose for spilling is either hi-degree, or very
1340 // rarely it can be low-degree. If we choose a hi-degree live range
1341 // there better not be any lo-degree choices.
1342 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" );
1343
1344 // Pull from hi-degree list
1345 uint prev = lo_lrg->_prev;
1346 uint next = lo_lrg->_next;
1347 if( prev ) lrgs(prev)._next = next;
1348 else _hi_degree = next;
1349 lrgs(next)._prev = prev;
1350 // Jam him on the lo-degree list, despite his high degree.
1351 // Maybe he'll get a color, and maybe he'll spill.
1352 // Only Select() will know.
1353 lrgs(lo_score)._at_risk = true;
1354 _lo_degree = lo_score;
1355 lo_lrg->_next = 0;
1356
1357 } // End of while not simplified everything
1358
1359 }
1360
1361 // Is 'reg' register legal for 'lrg'?
1362 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) {
1363 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) &&
1364 lrg.mask().Member(OptoReg::add(reg,-chunk))) {
1365 // RA uses OptoReg which represent the highest element of a registers set.
1366 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set
1367 // in which XMMd is used by RA to represent such vectors. A double value
1368 // uses [XMM,XMMb] pairs and XMMb is used by RA for it.
1369 // The register mask uses largest bits set of overlapping register sets.
1370 // On x86 with AVX it uses 8 bits for each XMM registers set.
1371 //
1372 // The 'lrg' already has cleared-to-set register mask (done in Select()
1373 // before calling choose_color()). Passing mask.Member(reg) check above
1374 // indicates that the size (num_regs) of 'reg' set is less or equal to
1375 // 'lrg' set size.
1376 // For set size 1 any register which is member of 'lrg' mask is legal.
1377 if (lrg.num_regs()==1)
1378 return true;
1379 // For larger sets only an aligned register with the same set size is legal.
1380 int mask = lrg.num_regs()-1;
1381 if ((reg&mask) == mask)
1382 return true;
1383 }
1384 return false;
1385 }
1386
1387 static OptoReg::Name find_first_set(LRG &lrg, RegMask mask, int chunk) {
1388 int num_regs = lrg.num_regs();
1389 OptoReg::Name assigned = mask.find_first_set(lrg, num_regs);
1390
1391 if (lrg.is_scalable()) {
1392 // a physical register is found
1393 if (chunk == 0 && OptoReg::is_reg(assigned)) {
1394 return assigned;
1395 }
1396
1397 // find available stack slots for scalable register
1398 if (lrg._is_vector) {
1399 num_regs = lrg.scalable_reg_slots();
1400 // if actual scalable vector register is exactly SlotsPerVecA * 32 bits
1401 if (num_regs == RegMask::SlotsPerVecA) {
1402 return assigned;
1403 }
1404
1405 // mask has been cleared out by clear_to_sets(SlotsPerVecA) before choose_color, but it
1406 // does not work for scalable size. We have to find adjacent scalable_reg_slots() bits
1407 // instead of SlotsPerVecA bits.
1408 assigned = mask.find_first_set(lrg, num_regs); // find highest valid reg
1409 while (OptoReg::is_valid(assigned) && RegMask::can_represent(assigned)) {
1410 // Verify the found reg has scalable_reg_slots() bits set.
1411 if (mask.is_valid_reg(assigned, num_regs)) {
1412 return assigned;
1413 } else {
1414 // Remove more for each iteration
1415 mask.Remove(assigned - num_regs + 1); // Unmask the lowest reg
1416 mask.clear_to_sets(RegMask::SlotsPerVecA); // Align by SlotsPerVecA bits
1417 assigned = mask.find_first_set(lrg, num_regs);
1418 }
1419 }
1420 return OptoReg::Bad; // will cause chunk change, and retry next chunk
1421 } else if (lrg._is_predicate) {
1422 assert(num_regs == RegMask::SlotsPerRegVectMask, "scalable predicate register");
1423 num_regs = lrg.scalable_reg_slots();
1424 mask.clear_to_sets(num_regs);
1425 return mask.find_first_set(lrg, num_regs);
1426 }
1427 }
1428
1429 return assigned;
1430 }
1431
1432 // Choose a color using the biasing heuristic
1433 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) {
1434
1435 // Check for "at_risk" LRG's
1436 uint risk_lrg = _lrg_map.find(lrg._risk_bias);
1437 if (risk_lrg != 0 && !_ifg->neighbors(risk_lrg)->is_empty()) {
1438 // Walk the colored neighbors of the "at_risk" candidate
1439 // Choose a color which is both legal and already taken by a neighbor
1440 // of the "at_risk" candidate in order to improve the chances of the
1441 // "at_risk" candidate of coloring
1442 IndexSetIterator elements(_ifg->neighbors(risk_lrg));
1443 uint datum;
1444 while ((datum = elements.next()) != 0) {
1445 OptoReg::Name reg = lrgs(datum).reg();
1446 // If this LRG's register is legal for us, choose it
1447 if (is_legal_reg(lrg, reg, chunk))
1448 return reg;
1449 }
1450 }
1451
1452 uint copy_lrg = _lrg_map.find(lrg._copy_bias);
1453 if (copy_lrg != 0) {
1454 // If he has a color,
1455 if(!_ifg->_yanked->test(copy_lrg)) {
1456 OptoReg::Name reg = lrgs(copy_lrg).reg();
1457 // And it is legal for you,
1458 if (is_legal_reg(lrg, reg, chunk))
1459 return reg;
1460 } else if( chunk == 0 ) {
1461 // Choose a color which is legal for him
1462 RegMask tempmask = lrg.mask();
1463 tempmask.AND(lrgs(copy_lrg).mask());
1464 tempmask.clear_to_sets(lrg.num_regs());
1465 OptoReg::Name reg = find_first_set(lrg, tempmask, chunk);
1466 if (OptoReg::is_valid(reg))
1467 return reg;
1468 }
1469 }
1470
1471 // If no bias info exists, just go with the register selection ordering
1472 if (lrg._is_vector || lrg.num_regs() == 2 || lrg.is_scalable()) {
1473 // Find an aligned set
1474 return OptoReg::add(find_first_set(lrg, lrg.mask(), chunk), chunk);
1475 }
1476
1477 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate
1478 // copy removal to remove many more copies, by preventing a just-assigned
1479 // register from being repeatedly assigned.
1480 OptoReg::Name reg = lrg.mask().find_first_elem();
1481 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) {
1482 // This 'Remove; find; Insert' idiom is an expensive way to find the
1483 // SECOND element in the mask.
1484 lrg.Remove(reg);
1485 OptoReg::Name reg2 = lrg.mask().find_first_elem();
1486 lrg.Insert(reg);
1487 if( OptoReg::is_reg(reg2))
1488 reg = reg2;
1489 }
1490 return OptoReg::add( reg, chunk );
1491 }
1492
1493 // Choose a color in the current chunk
1494 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) {
1495 assert( C->in_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP-1)), "must not allocate stack0 (inside preserve area)");
1496 assert(C->out_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP+0)), "must not allocate stack0 (inside preserve area)");
1497
1498 if( lrg.num_regs() == 1 || // Common Case
1499 !lrg._fat_proj ) // Aligned+adjacent pairs ok
1500 // Use a heuristic to "bias" the color choice
1501 return bias_color(lrg, chunk);
1502
1503 assert(!lrg._is_vector, "should be not vector here" );
1504 assert( lrg.num_regs() >= 2, "dead live ranges do not color" );
1505
1506 // Fat-proj case or misaligned double argument.
1507 assert(lrg.compute_mask_size() == lrg.num_regs() ||
1508 lrg.num_regs() == 2,"fat projs exactly color" );
1509 assert( !chunk, "always color in 1st chunk" );
1510 // Return the highest element in the set.
1511 return lrg.mask().find_last_elem();
1512 }
1513
1514 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted
1515 // in reverse order of removal. As long as nothing of hi-degree was yanked,
1516 // everything going back is guaranteed a color. Select that color. If some
1517 // hi-degree LRG cannot get a color then we record that we must spill.
1518 uint PhaseChaitin::Select( ) {
1519 Compile::TracePhase tp("chaitinSelect", &timers[_t_chaitinSelect]);
1520
1521 uint spill_reg = LRG::SPILL_REG;
1522 _max_reg = OptoReg::Name(0); // Past max register used
1523 while( _simplified ) {
1524 // Pull next LRG from the simplified list - in reverse order of removal
1525 uint lidx = _simplified;
1526 LRG *lrg = &lrgs(lidx);
1527 _simplified = lrg->_next;
1528
1529 #ifndef PRODUCT
1530 if (trace_spilling()) {
1531 ttyLocker ttyl;
1532 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(),
1533 lrg->degrees_of_freedom());
1534 lrg->dump();
1535 }
1536 #endif
1537
1538 // Re-insert into the IFG
1539 _ifg->re_insert(lidx);
1540 if( !lrg->alive() ) continue;
1541 // capture allstackedness flag before mask is hacked
1542 const int is_allstack = lrg->mask().is_AllStack();
1543
1544 // Yeah, yeah, yeah, I know, I know. I can refactor this
1545 // to avoid the GOTO, although the refactored code will not
1546 // be much clearer. We arrive here IFF we have a stack-based
1547 // live range that cannot color in the current chunk, and it
1548 // has to move into the next free stack chunk.
1549 int chunk = 0; // Current chunk is first chunk
1550 retry_next_chunk:
1551
1552 // Remove neighbor colors
1553 IndexSet *s = _ifg->neighbors(lidx);
1554 debug_only(RegMask orig_mask = lrg->mask();)
1555
1556 if (!s->is_empty()) {
1557 IndexSetIterator elements(s);
1558 uint neighbor;
1559 while ((neighbor = elements.next()) != 0) {
1560 // Note that neighbor might be a spill_reg. In this case, exclusion
1561 // of its color will be a no-op, since the spill_reg chunk is in outer
1562 // space. Also, if neighbor is in a different chunk, this exclusion
1563 // will be a no-op. (Later on, if lrg runs out of possible colors in
1564 // its chunk, a new chunk of color may be tried, in which case
1565 // examination of neighbors is started again, at retry_next_chunk.)
1566 LRG &nlrg = lrgs(neighbor);
1567 OptoReg::Name nreg = nlrg.reg();
1568 // Only subtract masks in the same chunk
1569 if (nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE) {
1570 #ifndef PRODUCT
1571 uint size = lrg->mask().Size();
1572 RegMask rm = lrg->mask();
1573 #endif
1574 lrg->SUBTRACT(nlrg.mask());
1575 #ifndef PRODUCT
1576 if (trace_spilling() && lrg->mask().Size() != size) {
1577 ttyLocker ttyl;
1578 tty->print("L%d ", lidx);
1579 rm.dump();
1580 tty->print(" intersected L%d ", neighbor);
1581 nlrg.mask().dump();
1582 tty->print(" removed ");
1583 rm.SUBTRACT(lrg->mask());
1584 rm.dump();
1585 tty->print(" leaving ");
1586 lrg->mask().dump();
1587 tty->cr();
1588 }
1589 #endif
1590 }
1591 }
1592 }
1593 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness");
1594 // Aligned pairs need aligned masks
1595 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1596 if (lrg->num_regs() > 1 && !lrg->_fat_proj) {
1597 lrg->clear_to_sets();
1598 }
1599
1600 // Check if a color is available and if so pick the color
1601 OptoReg::Name reg = choose_color( *lrg, chunk );
1602
1603 //---------------
1604 // If we fail to color and the AllStack flag is set, trigger
1605 // a chunk-rollover event
1606 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) {
1607 // Bump register mask up to next stack chunk
1608 chunk += RegMask::CHUNK_SIZE;
1609 lrg->Set_All();
1610 goto retry_next_chunk;
1611 }
1612
1613 //---------------
1614 // Did we get a color?
1615 else if( OptoReg::is_valid(reg)) {
1616 #ifndef PRODUCT
1617 RegMask avail_rm = lrg->mask();
1618 #endif
1619
1620 // Record selected register
1621 lrg->set_reg(reg);
1622
1623 if( reg >= _max_reg ) // Compute max register limit
1624 _max_reg = OptoReg::add(reg,1);
1625 // Fold reg back into normal space
1626 reg = OptoReg::add(reg,-chunk);
1627
1628 // If the live range is not bound, then we actually had some choices
1629 // to make. In this case, the mask has more bits in it than the colors
1630 // chosen. Restrict the mask to just what was picked.
1631 int n_regs = lrg->num_regs();
1632 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1633 if (n_regs == 1 || !lrg->_fat_proj) {
1634 if (Matcher::supports_scalable_vector()) {
1635 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecA, "sanity");
1636 } else {
1637 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity");
1638 }
1639 lrg->Clear(); // Clear the mask
1640 lrg->Insert(reg); // Set regmask to match selected reg
1641 // For vectors and pairs, also insert the low bit of the pair
1642 // We always choose the high bit, then mask the low bits by register size
1643 if (lrg->is_scalable() && OptoReg::is_stack(lrg->reg())) { // stack
1644 n_regs = lrg->scalable_reg_slots();
1645 }
1646 for (int i = 1; i < n_regs; i++) {
1647 lrg->Insert(OptoReg::add(reg,-i));
1648 }
1649 lrg->set_mask_size(n_regs);
1650 } else { // Else fatproj
1651 // mask must be equal to fatproj bits, by definition
1652 }
1653 #ifndef PRODUCT
1654 if (trace_spilling()) {
1655 ttyLocker ttyl;
1656 tty->print("L%d selected ", lidx);
1657 lrg->mask().dump();
1658 tty->print(" from ");
1659 avail_rm.dump();
1660 tty->cr();
1661 }
1662 #endif
1663 // Note that reg is the highest-numbered register in the newly-bound mask.
1664 } // end color available case
1665
1666 //---------------
1667 // Live range is live and no colors available
1668 else {
1669 assert( lrg->alive(), "" );
1670 assert( !lrg->_fat_proj || lrg->is_multidef() ||
1671 lrg->_def->outcnt() > 0, "fat_proj cannot spill");
1672 assert( !orig_mask.is_AllStack(), "All Stack does not spill" );
1673
1674 // Assign the special spillreg register
1675 lrg->set_reg(OptoReg::Name(spill_reg++));
1676 // Do not empty the regmask; leave mask_size lying around
1677 // for use during Spilling
1678 #ifndef PRODUCT
1679 if( trace_spilling() ) {
1680 ttyLocker ttyl;
1681 tty->print("L%d spilling with neighbors: ", lidx);
1682 s->dump();
1683 debug_only(tty->print(" original mask: "));
1684 debug_only(orig_mask.dump());
1685 dump_lrg(lidx);
1686 }
1687 #endif
1688 } // end spill case
1689
1690 }
1691
1692 return spill_reg-LRG::SPILL_REG; // Return number of spills
1693 }
1694
1695 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
1696 void PhaseChaitin::set_was_spilled( Node *n ) {
1697 if( _spilled_once.test_set(n->_idx) )
1698 _spilled_twice.set(n->_idx);
1699 }
1700
1701 // Convert Ideal spill instructions into proper FramePtr + offset Loads and
1702 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are.
1703 void PhaseChaitin::fixup_spills() {
1704 // This function does only cisc spill work.
1705 if( !UseCISCSpill ) return;
1706
1707 Compile::TracePhase tp("fixupSpills", &timers[_t_fixupSpills]);
1708
1709 // Grab the Frame Pointer
1710 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr);
1711
1712 // For all blocks
1713 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1714 Block* block = _cfg.get_block(i);
1715
1716 // For all instructions in block
1717 uint last_inst = block->end_idx();
1718 for (uint j = 1; j <= last_inst; j++) {
1719 Node* n = block->get_node(j);
1720
1721 // Dead instruction???
1722 assert( n->outcnt() != 0 ||// Nothing dead after post alloc
1723 C->top() == n || // Or the random TOP node
1724 n->is_Proj(), // Or a fat-proj kill node
1725 "No dead instructions after post-alloc" );
1726
1727 int inp = n->cisc_operand();
1728 if( inp != AdlcVMDeps::Not_cisc_spillable ) {
1729 // Convert operand number to edge index number
1730 MachNode *mach = n->as_Mach();
1731 inp = mach->operand_index(inp);
1732 Node *src = n->in(inp); // Value to load or store
1733 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src));
1734 OptoReg::Name src_reg = lrg_cisc.reg();
1735 // Doubles record the HIGH register of an adjacent pair.
1736 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs());
1737 if( OptoReg::is_stack(src_reg) ) { // If input is on stack
1738 // This is a CISC Spill, get stack offset and construct new node
1739 #ifndef PRODUCT
1740 if( TraceCISCSpill ) {
1741 tty->print(" reg-instr: ");
1742 n->dump();
1743 }
1744 #endif
1745 int stk_offset = reg2offset(src_reg);
1746 // Bailout if we might exceed node limit when spilling this instruction
1747 C->check_node_count(0, "out of nodes fixing spills");
1748 if (C->failing()) return;
1749 // Transform node
1750 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach();
1751 cisc->set_req(inp,fp); // Base register is frame pointer
1752 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) {
1753 assert( cisc->oper_input_base() == 2, "Only adding one edge");
1754 cisc->ins_req(1,src); // Requires a memory edge
1755 } else {
1756 // There is no space reserved for a memory edge before the inputs for
1757 // instructions which have "stackSlotX" parameter instead of "memory".
1758 // For example, "MoveF2I_stack_reg". We always need a memory edge from
1759 // src to cisc, else we might schedule cisc before src, loading from a
1760 // spill location before storing the spill. On some platforms, we land
1761 // in this else case because mach->oper_input_base() > 1, i.e. we have
1762 // multiple inputs. In some rare cases there are even multiple memory
1763 // operands, before and after spilling.
1764 // (e.g. spilling "addFPR24_reg_mem" to "addFPR24_mem_cisc")
1765 // In either case, there is no space in the inputs for the memory edge
1766 // so we add an additional precedence / memory edge.
1767 cisc->add_prec(src);
1768 }
1769 block->map_node(cisc, j); // Insert into basic block
1770 n->subsume_by(cisc, C); // Correct graph
1771 //
1772 ++_used_cisc_instructions;
1773 #ifndef PRODUCT
1774 if( TraceCISCSpill ) {
1775 tty->print(" cisc-instr: ");
1776 cisc->dump();
1777 }
1778 #endif
1779 } else {
1780 #ifndef PRODUCT
1781 if( TraceCISCSpill ) {
1782 tty->print(" using reg-instr: ");
1783 n->dump();
1784 }
1785 #endif
1786 ++_unused_cisc_instructions; // input can be on stack
1787 }
1788 }
1789
1790 } // End of for all instructions
1791
1792 } // End of for all blocks
1793 }
1794
1795 // Helper to stretch above; recursively discover the base Node for a
1796 // given derived Node. Easy for AddP-related machine nodes, but needs
1797 // to be recursive for derived Phis.
1798 Node* PhaseChaitin::find_base_for_derived(Node** derived_base_map, Node* derived, uint& maxlrg) {
1799 // See if already computed; if so return it
1800 if (derived_base_map[derived->_idx]) {
1801 return derived_base_map[derived->_idx];
1802 }
1803
1804 #ifdef ASSERT
1805 if (derived->is_Mach() && derived->as_Mach()->ideal_Opcode() == Op_VerifyVectorAlignment) {
1806 // Bypass the verification node
1807 Node* base = find_base_for_derived(derived_base_map, derived->in(1), maxlrg);
1808 derived_base_map[derived->_idx] = base;
1809 return base;
1810 }
1811 #endif
1812
1813 // See if this happens to be a base.
1814 // NOTE: we use TypePtr instead of TypeOopPtr because we can have
1815 // pointers derived from null! These are always along paths that
1816 // can't happen at run-time but the optimizer cannot deduce it so
1817 // we have to handle it gracefully.
1818 assert(!derived->bottom_type()->isa_narrowoop() ||
1819 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity");
1820 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1821 // If its an OOP with a non-zero offset, then it is derived.
1822 if( tj == nullptr || tj->_offset == 0 ) {
1823 derived_base_map[derived->_idx] = derived;
1824 return derived;
1825 }
1826 // Derived is null+offset? Base is null!
1827 if( derived->is_Con() ) {
1828 Node *base = _matcher.mach_null();
1829 assert(base != nullptr, "sanity");
1830 if (base->in(0) == nullptr) {
1831 // Initialize it once and make it shared:
1832 // set control to _root and place it into Start block
1833 // (where top() node is placed).
1834 base->init_req(0, _cfg.get_root_node());
1835 Block *startb = _cfg.get_block_for_node(C->top());
1836 uint node_pos = startb->find_node(C->top());
1837 startb->insert_node(base, node_pos);
1838 _cfg.map_node_to_block(base, startb);
1839 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet");
1840
1841 // The loadConP0 might have projection nodes depending on architecture
1842 // Add the projection nodes to the CFG
1843 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) {
1844 Node* use = base->fast_out(i);
1845 if (use->is_MachProj()) {
1846 startb->insert_node(use, ++node_pos);
1847 _cfg.map_node_to_block(use, startb);
1848 new_lrg(use, maxlrg++);
1849 }
1850 }
1851 }
1852 if (_lrg_map.live_range_id(base) == 0) {
1853 new_lrg(base, maxlrg++);
1854 }
1855 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");
1856 derived_base_map[derived->_idx] = base;
1857 return base;
1858 }
1859
1860 // Check for AddP-related opcodes
1861 if (!derived->is_Phi()) {
1862 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name());
1863 Node *base = derived->in(AddPNode::Base);
1864 derived_base_map[derived->_idx] = base;
1865 return base;
1866 }
1867
1868 // Recursively find bases for Phis.
1869 // First check to see if we can avoid a base Phi here.
1870 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg);
1871 uint i;
1872 for( i = 2; i < derived->req(); i++ )
1873 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg))
1874 break;
1875 // Went to the end without finding any different bases?
1876 if( i == derived->req() ) { // No need for a base Phi here
1877 derived_base_map[derived->_idx] = base;
1878 return base;
1879 }
1880
1881 // Now we see we need a base-Phi here to merge the bases
1882 const Type *t = base->bottom_type();
1883 base = new PhiNode( derived->in(0), t );
1884 for( i = 1; i < derived->req(); i++ ) {
1885 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg));
1886 t = t->meet(base->in(i)->bottom_type());
1887 }
1888 base->as_Phi()->set_type(t);
1889
1890 // Search the current block for an existing base-Phi
1891 Block *b = _cfg.get_block_for_node(derived);
1892 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi
1893 Node *phi = b->get_node(i);
1894 if( !phi->is_Phi() ) { // Found end of Phis with no match?
1895 b->insert_node(base, i); // Must insert created Phi here as base
1896 _cfg.map_node_to_block(base, b);
1897 new_lrg(base,maxlrg++);
1898 break;
1899 }
1900 // See if Phi matches.
1901 uint j;
1902 for( j = 1; j < base->req(); j++ )
1903 if( phi->in(j) != base->in(j) &&
1904 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different nulls
1905 break;
1906 if( j == base->req() ) { // All inputs match?
1907 base = phi; // Then use existing 'phi' and drop 'base'
1908 break;
1909 }
1910 }
1911
1912
1913 // Cache info for later passes
1914 derived_base_map[derived->_idx] = base;
1915 return base;
1916 }
1917
1918 // At each Safepoint, insert extra debug edges for each pair of derived value/
1919 // base pointer that is live across the Safepoint for oopmap building. The
1920 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the
1921 // required edge set.
1922 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) {
1923 int must_recompute_live = false;
1924 uint maxlrg = _lrg_map.max_lrg_id();
1925 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique());
1926 memset( derived_base_map, 0, sizeof(Node*)*C->unique() );
1927
1928 // For all blocks in RPO do...
1929 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1930 Block* block = _cfg.get_block(i);
1931 // Note use of deep-copy constructor. I cannot hammer the original
1932 // liveout bits, because they are needed by the following coalesce pass.
1933 IndexSet liveout(_live->live(block));
1934
1935 for (uint j = block->end_idx() + 1; j > 1; j--) {
1936 Node* n = block->get_node(j - 1);
1937
1938 // Pre-split compares of loop-phis. Loop-phis form a cycle we would
1939 // like to see in the same register. Compare uses the loop-phi and so
1940 // extends its live range BUT cannot be part of the cycle. If this
1941 // extended live range overlaps with the update of the loop-phi value
1942 // we need both alive at the same time -- which requires at least 1
1943 // copy. But because Intel has only 2-address registers we end up with
1944 // at least 2 copies, one before the loop-phi update instruction and
1945 // one after. Instead we split the input to the compare just after the
1946 // phi.
1947 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) {
1948 Node *phi = n->in(1);
1949 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) {
1950 Block *phi_block = _cfg.get_block_for_node(phi);
1951 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) {
1952 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI];
1953 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask);
1954 insert_proj( phi_block, 1, spill, maxlrg++ );
1955 n->set_req(1,spill);
1956 must_recompute_live = true;
1957 }
1958 }
1959 }
1960
1961 // Get value being defined
1962 uint lidx = _lrg_map.live_range_id(n);
1963 // Ignore the occasional brand-new live range
1964 if (lidx && lidx < _lrg_map.max_lrg_id()) {
1965 // Remove from live-out set
1966 liveout.remove(lidx);
1967
1968 // Copies do not define a new value and so do not interfere.
1969 // Remove the copies source from the liveout set before interfering.
1970 uint idx = n->is_Copy();
1971 if (idx) {
1972 liveout.remove(_lrg_map.live_range_id(n->in(idx)));
1973 }
1974 }
1975
1976 // Found a safepoint?
1977 JVMState *jvms = n->jvms();
1978 if (jvms && !liveout.is_empty()) {
1979 // Now scan for a live derived pointer
1980 IndexSetIterator elements(&liveout);
1981 uint neighbor;
1982 while ((neighbor = elements.next()) != 0) {
1983 // Find reaching DEF for base and derived values
1984 // This works because we are still in SSA during this call.
1985 Node *derived = lrgs(neighbor)._def;
1986 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1987 assert(!derived->bottom_type()->isa_narrowoop() ||
1988 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity");
1989 // If its an OOP with a non-zero offset, then it is derived.
1990 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) {
1991 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg);
1992 assert(base->_idx < _lrg_map.size(), "");
1993 // Add reaching DEFs of derived pointer and base pointer as a
1994 // pair of inputs
1995 n->add_req(derived);
1996 n->add_req(base);
1997
1998 // See if the base pointer is already live to this point.
1999 // Since I'm working on the SSA form, live-ness amounts to
2000 // reaching def's. So if I find the base's live range then
2001 // I know the base's def reaches here.
2002 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or
2003 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND
2004 (_lrg_map.live_range_id(base) > 0) && // not a constant
2005 _cfg.get_block_for_node(base) != block) { // base not def'd in blk)
2006 // Base pointer is not currently live. Since I stretched
2007 // the base pointer to here and it crosses basic-block
2008 // boundaries, the global live info is now incorrect.
2009 // Recompute live.
2010 must_recompute_live = true;
2011 } // End of if base pointer is not live to debug info
2012 }
2013 } // End of scan all live data for derived ptrs crossing GC point
2014 } // End of if found a GC point
2015
2016 // Make all inputs live
2017 if (!n->is_Phi()) { // Phi function uses come from prior block
2018 for (uint k = 1; k < n->req(); k++) {
2019 uint lidx = _lrg_map.live_range_id(n->in(k));
2020 if (lidx < _lrg_map.max_lrg_id()) {
2021 liveout.insert(lidx);
2022 }
2023 }
2024 }
2025
2026 } // End of forall instructions in block
2027 liveout.clear(); // Free the memory used by liveout.
2028
2029 } // End of forall blocks
2030 _lrg_map.set_max_lrg_id(maxlrg);
2031
2032 // If I created a new live range I need to recompute live
2033 if (maxlrg != _ifg->_maxlrg) {
2034 must_recompute_live = true;
2035 }
2036
2037 return must_recompute_live != 0;
2038 }
2039
2040 // Extend the node to LRG mapping
2041
2042 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) {
2043 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node));
2044 }
2045
2046 #ifndef PRODUCT
2047 void PhaseChaitin::dump(const Node* n) const {
2048 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0;
2049 tty->print("L%d",r);
2050 if (r && n->Opcode() != Op_Phi) {
2051 if( _node_regs ) { // Got a post-allocation copy of allocation?
2052 tty->print("[");
2053 OptoReg::Name second = get_reg_second(n);
2054 if( OptoReg::is_valid(second) ) {
2055 if( OptoReg::is_reg(second) )
2056 tty->print("%s:",Matcher::regName[second]);
2057 else
2058 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second));
2059 }
2060 OptoReg::Name first = get_reg_first(n);
2061 if( OptoReg::is_reg(first) )
2062 tty->print("%s]",Matcher::regName[first]);
2063 else
2064 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first));
2065 } else
2066 n->out_RegMask().dump();
2067 }
2068 tty->print("/N%d\t",n->_idx);
2069 tty->print("%s === ", n->Name());
2070 uint k;
2071 for (k = 0; k < n->req(); k++) {
2072 Node *m = n->in(k);
2073 if (!m) {
2074 tty->print("_ ");
2075 }
2076 else {
2077 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2078 tty->print("L%d",r);
2079 // Data MultiNode's can have projections with no real registers.
2080 // Don't die while dumping them.
2081 int op = n->Opcode();
2082 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) {
2083 if( _node_regs ) {
2084 tty->print("[");
2085 OptoReg::Name second = get_reg_second(n->in(k));
2086 if( OptoReg::is_valid(second) ) {
2087 if( OptoReg::is_reg(second) )
2088 tty->print("%s:",Matcher::regName[second]);
2089 else
2090 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer),
2091 reg2offset_unchecked(second));
2092 }
2093 OptoReg::Name first = get_reg_first(n->in(k));
2094 if( OptoReg::is_reg(first) )
2095 tty->print("%s]",Matcher::regName[first]);
2096 else
2097 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer),
2098 reg2offset_unchecked(first));
2099 } else
2100 n->in_RegMask(k).dump();
2101 }
2102 tty->print("/N%d ",m->_idx);
2103 }
2104 }
2105 if( k < n->len() && n->in(k) ) tty->print("| ");
2106 for( ; k < n->len(); k++ ) {
2107 Node *m = n->in(k);
2108 if(!m) {
2109 break;
2110 }
2111 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2112 tty->print("L%d",r);
2113 tty->print("/N%d ",m->_idx);
2114 }
2115 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty);
2116 else n->dump_spec(tty);
2117 if( _spilled_once.test(n->_idx ) ) {
2118 tty->print(" Spill_1");
2119 if( _spilled_twice.test(n->_idx ) )
2120 tty->print(" Spill_2");
2121 }
2122 tty->print("\n");
2123 }
2124
2125 void PhaseChaitin::dump(const Block* b) const {
2126 b->dump_head(&_cfg);
2127
2128 // For all instructions
2129 for( uint j = 0; j < b->number_of_nodes(); j++ )
2130 dump(b->get_node(j));
2131 // Print live-out info at end of block
2132 if( _live ) {
2133 tty->print("Liveout: ");
2134 IndexSet *live = _live->live(b);
2135 IndexSetIterator elements(live);
2136 tty->print("{");
2137 uint i;
2138 while ((i = elements.next()) != 0) {
2139 tty->print("L%d ", _lrg_map.find_const(i));
2140 }
2141 tty->print_cr("}");
2142 }
2143 tty->print("\n");
2144 }
2145
2146 void PhaseChaitin::dump() const {
2147 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n",
2148 _matcher._new_SP, _framesize );
2149
2150 // For all blocks
2151 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2152 dump(_cfg.get_block(i));
2153 }
2154 // End of per-block dump
2155 tty->print("\n");
2156
2157 if (!_ifg) {
2158 tty->print("(No IFG.)\n");
2159 return;
2160 }
2161
2162 // Dump LRG array
2163 tty->print("--- Live RanGe Array ---\n");
2164 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) {
2165 tty->print("L%d: ",i2);
2166 if (i2 < _ifg->_maxlrg) {
2167 lrgs(i2).dump();
2168 }
2169 else {
2170 tty->print_cr("new LRG");
2171 }
2172 }
2173 tty->cr();
2174
2175 // Dump lo-degree list
2176 tty->print("Lo degree: ");
2177 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next )
2178 tty->print("L%d ",i3);
2179 tty->cr();
2180
2181 // Dump lo-stk-degree list
2182 tty->print("Lo stk degree: ");
2183 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next )
2184 tty->print("L%d ",i4);
2185 tty->cr();
2186
2187 // Dump lo-degree list
2188 tty->print("Hi degree: ");
2189 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next )
2190 tty->print("L%d ",i5);
2191 tty->cr();
2192 }
2193
2194 void PhaseChaitin::dump_degree_lists() const {
2195 // Dump lo-degree list
2196 tty->print("Lo degree: ");
2197 for( uint i = _lo_degree; i; i = lrgs(i)._next )
2198 tty->print("L%d ",i);
2199 tty->cr();
2200
2201 // Dump lo-stk-degree list
2202 tty->print("Lo stk degree: ");
2203 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next )
2204 tty->print("L%d ",i2);
2205 tty->cr();
2206
2207 // Dump lo-degree list
2208 tty->print("Hi degree: ");
2209 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next )
2210 tty->print("L%d ",i3);
2211 tty->cr();
2212 }
2213
2214 void PhaseChaitin::dump_simplified() const {
2215 tty->print("Simplified: ");
2216 for( uint i = _simplified; i; i = lrgs(i)._next )
2217 tty->print("L%d ",i);
2218 tty->cr();
2219 }
2220
2221 static char *print_reg(OptoReg::Name reg, const PhaseChaitin* pc, char* buf, size_t buf_size) {
2222 if ((int)reg < 0)
2223 os::snprintf_checked(buf, buf_size, "<OptoReg::%d>", (int)reg);
2224 else if (OptoReg::is_reg(reg))
2225 strcpy(buf, Matcher::regName[reg]);
2226 else
2227 os::snprintf_checked(buf, buf_size, "%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer),
2228 pc->reg2offset(reg));
2229 return buf+strlen(buf);
2230 }
2231
2232 // Dump a register name into a buffer. Be intelligent if we get called
2233 // before allocation is complete.
2234 char *PhaseChaitin::dump_register(const Node* n, char* buf, size_t buf_size) const {
2235 if( _node_regs ) {
2236 // Post allocation, use direct mappings, no LRG info available
2237 print_reg( get_reg_first(n), this, buf, buf_size);
2238 } else {
2239 uint lidx = _lrg_map.find_const(n); // Grab LRG number
2240 if( !_ifg ) {
2241 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2242 } else if( !lidx ) { // Special, not allocated value
2243 strcpy(buf,"Special");
2244 } else {
2245 if (lrgs(lidx)._is_vector) {
2246 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs()))
2247 print_reg( lrgs(lidx).reg(), this, buf, buf_size); // a bound machine register
2248 else
2249 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2250 } else if( (lrgs(lidx).num_regs() == 1)
2251 ? lrgs(lidx).mask().is_bound1()
2252 : lrgs(lidx).mask().is_bound_pair() ) {
2253 // Hah! We have a bound machine register
2254 print_reg( lrgs(lidx).reg(), this, buf, buf_size);
2255 } else {
2256 os::snprintf_checked(buf, buf_size, "L%d",lidx); // No register binding yet
2257 }
2258 }
2259 }
2260 return buf+strlen(buf);
2261 }
2262
2263 void PhaseChaitin::dump_for_spill_split_recycle() const {
2264 if( WizardMode && (PrintCompilation || PrintOpto) ) {
2265 // Display which live ranges need to be split and the allocator's state
2266 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt);
2267 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) {
2268 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
2269 tty->print("L%d: ", bidx);
2270 lrgs(bidx).dump();
2271 }
2272 }
2273 tty->cr();
2274 dump();
2275 }
2276 }
2277
2278 void PhaseChaitin::dump_frame() const {
2279 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer);
2280 const TypeTuple *domain = C->tf()->domain();
2281 const int argcnt = domain->cnt() - TypeFunc::Parms;
2282
2283 // Incoming arguments in registers dump
2284 for( int k = 0; k < argcnt; k++ ) {
2285 OptoReg::Name parmreg = _matcher._parm_regs[k].first();
2286 if( OptoReg::is_reg(parmreg)) {
2287 const char *reg_name = OptoReg::regname(parmreg);
2288 tty->print("#r%3.3d %s", parmreg, reg_name);
2289 parmreg = _matcher._parm_regs[k].second();
2290 if( OptoReg::is_reg(parmreg)) {
2291 tty->print(":%s", OptoReg::regname(parmreg));
2292 }
2293 tty->print(" : parm %d: ", k);
2294 domain->field_at(k + TypeFunc::Parms)->dump();
2295 tty->cr();
2296 }
2297 }
2298
2299 // Check for un-owned padding above incoming args
2300 OptoReg::Name reg = _matcher._new_SP;
2301 if( reg > _matcher._in_arg_limit ) {
2302 reg = OptoReg::add(reg, -1);
2303 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg));
2304 }
2305
2306 // Incoming argument area dump
2307 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots());
2308 while( reg > begin_in_arg ) {
2309 reg = OptoReg::add(reg, -1);
2310 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2311 int j;
2312 for( j = 0; j < argcnt; j++) {
2313 if( _matcher._parm_regs[j].first() == reg ||
2314 _matcher._parm_regs[j].second() == reg ) {
2315 tty->print("parm %d: ",j);
2316 domain->field_at(j + TypeFunc::Parms)->dump();
2317 tty->cr();
2318 break;
2319 }
2320 }
2321 if( j >= argcnt )
2322 tty->print_cr("HOLE, owned by SELF");
2323 }
2324
2325 // Old outgoing preserve area
2326 while( reg > _matcher._old_SP ) {
2327 reg = OptoReg::add(reg, -1);
2328 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg));
2329 }
2330
2331 // Old SP
2332 tty->print_cr("# -- Old %s -- Framesize: %d --",fp,
2333 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize);
2334
2335 // Preserve area dump
2336 int fixed_slots = C->fixed_slots();
2337 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots());
2338 OptoReg::Name return_addr = _matcher.return_addr();
2339
2340 reg = OptoReg::add(reg, -1);
2341 while (OptoReg::is_stack(reg)) {
2342 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2343 if (return_addr == reg) {
2344 tty->print_cr("return address");
2345 } else if (reg >= begin_in_preserve) {
2346 // Preserved slots are present on x86
2347 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word))
2348 tty->print_cr("saved fp register");
2349 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) &&
2350 VerifyStackAtCalls)
2351 tty->print_cr("0xBADB100D +VerifyStackAtCalls");
2352 else
2353 tty->print_cr("in_preserve");
2354 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) {
2355 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg));
2356 } else {
2357 tty->print_cr("pad2, stack alignment");
2358 }
2359 reg = OptoReg::add(reg, -1);
2360 }
2361
2362 // Spill area dump
2363 reg = OptoReg::add(_matcher._new_SP, _framesize );
2364 while( reg > _matcher._out_arg_limit ) {
2365 reg = OptoReg::add(reg, -1);
2366 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg));
2367 }
2368
2369 // Outgoing argument area dump
2370 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) {
2371 reg = OptoReg::add(reg, -1);
2372 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg));
2373 }
2374
2375 // Outgoing new preserve area
2376 while( reg > _matcher._new_SP ) {
2377 reg = OptoReg::add(reg, -1);
2378 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg));
2379 }
2380 tty->print_cr("#");
2381 }
2382
2383 void PhaseChaitin::dump_bb(uint pre_order) const {
2384 tty->print_cr("---dump of B%d---",pre_order);
2385 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2386 Block* block = _cfg.get_block(i);
2387 if (block->_pre_order == pre_order) {
2388 dump(block);
2389 }
2390 }
2391 }
2392
2393 void PhaseChaitin::dump_lrg(uint lidx, bool defs_only) const {
2394 tty->print_cr("---dump of L%d---",lidx);
2395
2396 if (_ifg) {
2397 if (lidx >= _lrg_map.max_lrg_id()) {
2398 tty->print("Attempt to print live range index beyond max live range.\n");
2399 return;
2400 }
2401 tty->print("L%d: ",lidx);
2402 if (lidx < _ifg->_maxlrg) {
2403 lrgs(lidx).dump();
2404 } else {
2405 tty->print_cr("new LRG");
2406 }
2407 }
2408 if( _ifg && lidx < _ifg->_maxlrg) {
2409 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx));
2410 _ifg->neighbors(lidx)->dump();
2411 tty->cr();
2412 }
2413 // For all blocks
2414 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2415 Block* block = _cfg.get_block(i);
2416 int dump_once = 0;
2417
2418 // For all instructions
2419 for( uint j = 0; j < block->number_of_nodes(); j++ ) {
2420 Node *n = block->get_node(j);
2421 if (_lrg_map.find_const(n) == lidx) {
2422 if (!dump_once++) {
2423 tty->cr();
2424 block->dump_head(&_cfg);
2425 }
2426 dump(n);
2427 continue;
2428 }
2429 if (!defs_only) {
2430 uint cnt = n->req();
2431 for( uint k = 1; k < cnt; k++ ) {
2432 Node *m = n->in(k);
2433 if (!m) {
2434 continue; // be robust in the dumper
2435 }
2436 if (_lrg_map.find_const(m) == lidx) {
2437 if (!dump_once++) {
2438 tty->cr();
2439 block->dump_head(&_cfg);
2440 }
2441 dump(n);
2442 }
2443 }
2444 }
2445 }
2446 } // End of per-block dump
2447 tty->cr();
2448 }
2449 #endif // not PRODUCT
2450
2451 #ifdef ASSERT
2452 // Verify that base pointers and derived pointers are still sane.
2453 void PhaseChaitin::verify_base_ptrs(ResourceArea* a) const {
2454 Unique_Node_List worklist(a);
2455 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2456 Block* block = _cfg.get_block(i);
2457 for (uint j = block->end_idx() + 1; j > 1; j--) {
2458 Node* n = block->get_node(j-1);
2459 if (n->is_Phi()) {
2460 break;
2461 }
2462 // Found a safepoint?
2463 if (n->is_MachSafePoint()) {
2464 MachSafePointNode* sfpt = n->as_MachSafePoint();
2465 JVMState* jvms = sfpt->jvms();
2466 if (jvms != nullptr) {
2467 // Now scan for a live derived pointer
2468 if (jvms->oopoff() < sfpt->req()) {
2469 // Check each derived/base pair
2470 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
2471 Node* check = sfpt->in(idx);
2472 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
2473 // search upwards through spills and spill phis for AddP
2474 worklist.clear();
2475 worklist.push(check);
2476 uint k = 0;
2477 while (k < worklist.size()) {
2478 check = worklist.at(k);
2479 assert(check, "Bad base or derived pointer");
2480 // See PhaseChaitin::find_base_for_derived() for all cases.
2481 int isc = check->is_Copy();
2482 if (isc) {
2483 worklist.push(check->in(isc));
2484 } else if (check->is_Phi()) {
2485 for (uint m = 1; m < check->req(); m++) {
2486 worklist.push(check->in(m));
2487 }
2488 } else if (check->is_Con()) {
2489 if (is_derived && check->bottom_type()->is_ptr()->_offset != 0) {
2490 // Derived is null+non-zero offset, base must be null.
2491 assert(check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad derived pointer");
2492 } else {
2493 assert(check->bottom_type()->is_ptr()->_offset == 0, "Bad base pointer");
2494 // Base either ConP(nullptr) or loadConP
2495 if (check->is_Mach()) {
2496 assert(check->as_Mach()->ideal_Opcode() == Op_ConP, "Bad base pointer");
2497 } else {
2498 assert(check->Opcode() == Op_ConP &&
2499 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad base pointer");
2500 }
2501 }
2502 } else if (check->bottom_type()->is_ptr()->_offset == 0) {
2503 if (check->is_Proj() || (check->is_Mach() &&
2504 (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
2505 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
2506 check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
2507 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
2508 #ifdef _LP64
2509 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP) ||
2510 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN) ||
2511 (UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass) ||
2512 #endif // _LP64
2513 check->as_Mach()->ideal_Opcode() == Op_LoadP ||
2514 check->as_Mach()->ideal_Opcode() == Op_LoadKlass))) {
2515 // Valid nodes
2516 } else {
2517 check->dump();
2518 assert(false, "Bad base or derived pointer");
2519 }
2520 } else {
2521 assert(is_derived, "Bad base pointer");
2522 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP, "Bad derived pointer");
2523 }
2524 k++;
2525 assert(k < 100000, "Derived pointer checking in infinite loop");
2526 } // End while
2527 }
2528 } // End of check for derived pointers
2529 } // End of Kcheck for debug info
2530 } // End of if found a safepoint
2531 } // End of forall instructions in block
2532 } // End of forall blocks
2533 }
2534
2535 // Verify that graphs and base pointers are still sane.
2536 void PhaseChaitin::verify(ResourceArea* a, bool verify_ifg) const {
2537 if (VerifyRegisterAllocator) {
2538 _cfg.verify();
2539 verify_base_ptrs(a);
2540 if (verify_ifg) {
2541 _ifg->verify(this);
2542 }
2543 }
2544 }
2545 #endif // ASSERT
2546
2547 int PhaseChaitin::_final_loads = 0;
2548 int PhaseChaitin::_final_stores = 0;
2549 int PhaseChaitin::_final_memoves= 0;
2550 int PhaseChaitin::_final_copies = 0;
2551 double PhaseChaitin::_final_load_cost = 0;
2552 double PhaseChaitin::_final_store_cost = 0;
2553 double PhaseChaitin::_final_memove_cost= 0;
2554 double PhaseChaitin::_final_copy_cost = 0;
2555 int PhaseChaitin::_conserv_coalesce = 0;
2556 int PhaseChaitin::_conserv_coalesce_pair = 0;
2557 int PhaseChaitin::_conserv_coalesce_trie = 0;
2558 int PhaseChaitin::_conserv_coalesce_quad = 0;
2559 int PhaseChaitin::_post_alloc = 0;
2560 int PhaseChaitin::_lost_opp_pp_coalesce = 0;
2561 int PhaseChaitin::_lost_opp_cflow_coalesce = 0;
2562 int PhaseChaitin::_used_cisc_instructions = 0;
2563 int PhaseChaitin::_unused_cisc_instructions = 0;
2564 int PhaseChaitin::_allocator_attempts = 0;
2565 int PhaseChaitin::_allocator_successes = 0;
2566
2567 #ifndef PRODUCT
2568 uint PhaseChaitin::_high_pressure = 0;
2569 uint PhaseChaitin::_low_pressure = 0;
2570
2571 void PhaseChaitin::print_chaitin_statistics() {
2572 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);
2573 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);
2574 tty->print_cr("Adjusted spill cost = %7.0f.",
2575 _final_load_cost*4.0 + _final_store_cost * 2.0 +
2576 _final_copy_cost*1.0 + _final_memove_cost*12.0);
2577 tty->print("Conservatively coalesced %d copies, %d pairs",
2578 _conserv_coalesce, _conserv_coalesce_pair);
2579 if( _conserv_coalesce_trie || _conserv_coalesce_quad )
2580 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad);
2581 tty->print_cr(", %d post alloc.", _post_alloc);
2582 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce )
2583 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.",
2584 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce );
2585 if( _used_cisc_instructions || _unused_cisc_instructions )
2586 tty->print_cr("Used cisc instruction %d, remained in register %d",
2587 _used_cisc_instructions, _unused_cisc_instructions);
2588 if( _allocator_successes != 0 )
2589 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes);
2590 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure);
2591 }
2592 #endif // not PRODUCT