1 /*
2 * Copyright (c) 2000, 2020, 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
dump() const47 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 != NULL) {
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 == 0 ) 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_scalable ) tty->print("Scalable ");
81 if( _was_spilled1 ) tty->print("Spilled ");
82 if( _was_spilled2 ) tty->print("Spilled2 ");
83 if( _direct_conflict ) tty->print("Direct_conflict ");
84 if( _fat_proj ) tty->print("Fat ");
85 if( _was_lo ) tty->print("Lo ");
86 if( _has_copy ) tty->print("Copy ");
87 if( _at_risk ) tty->print("Risk ");
88
89 if( _must_spill ) tty->print("Must_spill ");
90 if( _is_bound ) tty->print("Bound ");
91 if( _msize_valid ) {
92 if( _degree_valid && lo_degree() ) tty->print("Trivial ");
93 }
94
95 tty->cr();
96 }
97 #endif
98
99 // Compute score from cost and area. Low score is best to spill.
raw_score(double cost,double area)100 static double raw_score( double cost, double area ) {
101 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5;
102 }
103
score() const104 double LRG::score() const {
105 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost.
106 // Bigger area lowers score, encourages spilling this live range.
107 // Bigger cost raise score, prevents spilling this live range.
108 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer
109 // to turn a divide by a constant into a multiply by the reciprical).
110 double score = raw_score( _cost, _area);
111
112 // Account for area. Basically, LRGs covering large areas are better
113 // to spill because more other LRGs get freed up.
114 if( _area == 0.0 ) // No area? Then no progress to spill
115 return 1e35;
116
117 if( _was_spilled2 ) // If spilled once before, we are unlikely
118 return score + 1e30; // to make progress again.
119
120 if( _cost >= _area*3.0 ) // Tiny area relative to cost
121 return score + 1e17; // Probably no progress to spill
122
123 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost
124 return score + 1e10; // Likely no progress to spill
125
126 return score;
127 }
128
129 #define NUMBUCKS 3
130
131 // Straight out of Tarjan's union-find algorithm
find_compress(uint lrg)132 uint LiveRangeMap::find_compress(uint lrg) {
133 uint cur = lrg;
134 uint next = _uf_map.at(cur);
135 while (next != cur) { // Scan chain of equivalences
136 assert( next < cur, "always union smaller");
137 cur = next; // until find a fixed-point
138 next = _uf_map.at(cur);
139 }
140
141 // Core of union-find algorithm: update chain of
142 // equivalences to be equal to the root.
143 while (lrg != next) {
144 uint tmp = _uf_map.at(lrg);
145 _uf_map.at_put(lrg, next);
146 lrg = tmp;
147 }
148 return lrg;
149 }
150
151 // Reset the Union-Find map to identity
reset_uf_map(uint max_lrg_id)152 void LiveRangeMap::reset_uf_map(uint max_lrg_id) {
153 _max_lrg_id= max_lrg_id;
154 // Force the Union-Find mapping to be at least this large
155 _uf_map.at_put_grow(_max_lrg_id, 0);
156 // Initialize it to be the ID mapping.
157 for (uint i = 0; i < _max_lrg_id; ++i) {
158 _uf_map.at_put(i, i);
159 }
160 }
161
162 // Make all Nodes map directly to their final live range; no need for
163 // the Union-Find mapping after this call.
compress_uf_map_for_nodes()164 void LiveRangeMap::compress_uf_map_for_nodes() {
165 // For all Nodes, compress mapping
166 uint unique = _names.length();
167 for (uint i = 0; i < unique; ++i) {
168 uint lrg = _names.at(i);
169 uint compressed_lrg = find(lrg);
170 if (lrg != compressed_lrg) {
171 _names.at_put(i, compressed_lrg);
172 }
173 }
174 }
175
176 // Like Find above, but no path compress, so bad asymptotic behavior
find_const(uint lrg) const177 uint LiveRangeMap::find_const(uint lrg) const {
178 if (!lrg) {
179 return lrg; // Ignore the zero LRG
180 }
181
182 // Off the end? This happens during debugging dumps when you got
183 // brand new live ranges but have not told the allocator yet.
184 if (lrg >= _max_lrg_id) {
185 return lrg;
186 }
187
188 uint next = _uf_map.at(lrg);
189 while (next != lrg) { // Scan chain of equivalences
190 assert(next < lrg, "always union smaller");
191 lrg = next; // until find a fixed-point
192 next = _uf_map.at(lrg);
193 }
194 return next;
195 }
196
PhaseChaitin(uint unique,PhaseCFG & cfg,Matcher & matcher,bool scheduling_info_generated)197 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated)
198 : PhaseRegAlloc(unique, cfg, matcher,
199 #ifndef PRODUCT
200 print_chaitin_statistics
201 #else
202 NULL
203 #endif
204 )
205 , _live(0)
206 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0)
207 , _oldphi(unique)
208 #ifndef PRODUCT
209 , _trace_spilling(C->directive()->TraceSpillingOption)
210 #endif
211 , _lrg_map(Thread::current()->resource_area(), unique)
212 , _scheduling_info_generated(scheduling_info_generated)
213 , _sched_int_pressure(0, INTPRESSURE)
214 , _sched_float_pressure(0, FLOATPRESSURE)
215 , _scratch_int_pressure(0, INTPRESSURE)
216 , _scratch_float_pressure(0, FLOATPRESSURE)
217 {
218 Compile::TracePhase tp("ctorChaitin", &timers[_t_ctorChaitin]);
219
220 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency());
221
222 // Build a list of basic blocks, sorted by frequency
223 _blks = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks());
224 // Experiment with sorting strategies to speed compilation
225 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket
226 Block **buckets[NUMBUCKS]; // Array of buckets
227 uint buckcnt[NUMBUCKS]; // Array of bucket counters
228 double buckval[NUMBUCKS]; // Array of bucket value cutoffs
229 for (uint i = 0; i < NUMBUCKS; i++) {
230 buckets[i] = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks());
231 buckcnt[i] = 0;
232 // Bump by three orders of magnitude each time
233 cutoff *= 0.001;
234 buckval[i] = cutoff;
235 for (uint j = 0; j < _cfg.number_of_blocks(); j++) {
236 buckets[i][j] = NULL;
237 }
238 }
239 // Sort blocks into buckets
240 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
241 for (uint j = 0; j < NUMBUCKS; j++) {
242 if ((j == NUMBUCKS - 1) || (_cfg.get_block(i)->_freq > buckval[j])) {
243 // Assign block to end of list for appropriate bucket
244 buckets[j][buckcnt[j]++] = _cfg.get_block(i);
245 break; // kick out of inner loop
246 }
247 }
248 }
249 // Dump buckets into final block array
250 uint blkcnt = 0;
251 for (uint i = 0; i < NUMBUCKS; i++) {
252 for (uint j = 0; j < buckcnt[i]; j++) {
253 _blks[blkcnt++] = buckets[i][j];
254 }
255 }
256
257 assert(blkcnt == _cfg.number_of_blocks(), "Block array not totally filled");
258 }
259
260 // union 2 sets together.
Union(const Node * src_n,const Node * dst_n)261 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) {
262 uint src = _lrg_map.find(src_n);
263 uint dst = _lrg_map.find(dst_n);
264 assert(src, "");
265 assert(dst, "");
266 assert(src < _lrg_map.max_lrg_id(), "oob");
267 assert(dst < _lrg_map.max_lrg_id(), "oob");
268 assert(src < dst, "always union smaller");
269 _lrg_map.uf_map(dst, src);
270 }
271
new_lrg(const Node * x,uint lrg)272 void PhaseChaitin::new_lrg(const Node *x, uint lrg) {
273 // Make the Node->LRG mapping
274 _lrg_map.extend(x->_idx,lrg);
275 // Make the Union-Find mapping an identity function
276 _lrg_map.uf_extend(lrg, lrg);
277 }
278
279
clone_projs(Block * b,uint idx,Node * orig,Node * copy,uint & max_lrg_id)280 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) {
281 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections");
282 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); )
283 int found_projs = 0;
284 uint cnt = orig->outcnt();
285 for (uint i = 0; i < cnt; i++) {
286 Node* proj = orig->raw_out(i);
287 if (proj->is_MachProj()) {
288 assert(proj->outcnt() == 0, "only kill projections are expected here");
289 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections");
290 found_projs++;
291 // Copy kill projections after the cloned node
292 Node* kills = proj->clone();
293 kills->set_req(0, copy);
294 b->insert_node(kills, idx++);
295 _cfg.map_node_to_block(kills, b);
296 new_lrg(kills, max_lrg_id++);
297 }
298 }
299 return found_projs;
300 }
301
302 // Renumber the live ranges to compact them. Makes the IFG smaller.
compact()303 void PhaseChaitin::compact() {
304 Compile::TracePhase tp("chaitinCompact", &timers[_t_chaitinCompact]);
305
306 // Current the _uf_map contains a series of short chains which are headed
307 // by a self-cycle. All the chains run from big numbers to little numbers.
308 // The Find() call chases the chains & shortens them for the next Find call.
309 // We are going to change this structure slightly. Numbers above a moving
310 // wave 'i' are unchanged. Numbers below 'j' point directly to their
311 // compacted live range with no further chaining. There are no chains or
312 // cycles below 'i', so the Find call no longer works.
313 uint j=1;
314 uint i;
315 for (i = 1; i < _lrg_map.max_lrg_id(); i++) {
316 uint lr = _lrg_map.uf_live_range_id(i);
317 // Ignore unallocated live ranges
318 if (!lr) {
319 continue;
320 }
321 assert(lr <= i, "");
322 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr));
323 }
324 // Now change the Node->LR mapping to reflect the compacted names
325 uint unique = _lrg_map.size();
326 for (i = 0; i < unique; i++) {
327 uint lrg_id = _lrg_map.live_range_id(i);
328 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id));
329 }
330
331 // Reset the Union-Find mapping
332 _lrg_map.reset_uf_map(j);
333 }
334
Register_Allocate()335 void PhaseChaitin::Register_Allocate() {
336
337 // Above the OLD FP (and in registers) are the incoming arguments. Stack
338 // slots in this area are called "arg_slots". Above the NEW FP (and in
339 // registers) is the outgoing argument area; above that is the spill/temp
340 // area. These are all "frame_slots". Arg_slots start at the zero
341 // stack_slots and count up to the known arg_size. Frame_slots start at
342 // the stack_slot #arg_size and go up. After allocation I map stack
343 // slots to actual offsets. Stack-slots in the arg_slot area are biased
344 // by the frame_size; stack-slots in the frame_slot area are biased by 0.
345
346 _trip_cnt = 0;
347 _alternate = 0;
348 _matcher._allocation_started = true;
349
350 ResourceArea split_arena(mtCompiler); // Arena for Split local resources
351 ResourceArea live_arena(mtCompiler); // Arena for liveness & IFG info
352 ResourceMark rm(&live_arena);
353
354 // Need live-ness for the IFG; need the IFG for coalescing. If the
355 // liveness is JUST for coalescing, then I can get some mileage by renaming
356 // all copy-related live ranges low and then using the max copy-related
357 // live range as a cut-off for LIVE and the IFG. In other words, I can
358 // build a subset of LIVE and IFG just for copies.
359 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false);
360
361 // Need IFG for coalescing and coloring
362 PhaseIFG ifg(&live_arena);
363 _ifg = &ifg;
364
365 // Come out of SSA world to the Named world. Assign (virtual) registers to
366 // Nodes. Use the same register for all inputs and the output of PhiNodes
367 // - effectively ending SSA form. This requires either coalescing live
368 // ranges or inserting copies. For the moment, we insert "virtual copies"
369 // - we pretend there is a copy prior to each Phi in predecessor blocks.
370 // We will attempt to coalesce such "virtual copies" before we manifest
371 // them for real.
372 de_ssa();
373
374 #ifdef ASSERT
375 // Veify the graph before RA.
376 verify(&live_arena);
377 #endif
378
379 {
380 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
381 _live = NULL; // Mark live as being not available
382 rm.reset_to_mark(); // Reclaim working storage
383 IndexSet::reset_memory(C, &live_arena);
384 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG
385 gather_lrg_masks( false ); // Collect LRG masks
386 live.compute(_lrg_map.max_lrg_id()); // Compute liveness
387 _live = &live; // Mark LIVE as being available
388 }
389
390 // Base pointers are currently "used" by instructions which define new
391 // derived pointers. This makes base pointers live up to the where the
392 // derived pointer is made, but not beyond. Really, they need to be live
393 // across any GC point where the derived value is live. So this code looks
394 // at all the GC points, and "stretches" the live range of any base pointer
395 // to the GC point.
396 if (stretch_base_pointer_live_ranges(&live_arena)) {
397 Compile::TracePhase tp("computeLive (sbplr)", &timers[_t_computeLive]);
398 // Since some live range stretched, I need to recompute live
399 _live = NULL;
400 rm.reset_to_mark(); // Reclaim working storage
401 IndexSet::reset_memory(C, &live_arena);
402 ifg.init(_lrg_map.max_lrg_id());
403 gather_lrg_masks(false);
404 live.compute(_lrg_map.max_lrg_id());
405 _live = &live;
406 }
407 // Create the interference graph using virtual copies
408 build_ifg_virtual(); // Include stack slots this time
409
410 // The IFG is/was triangular. I am 'squaring it up' so Union can run
411 // faster. Union requires a 'for all' operation which is slow on the
412 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' -
413 // meaning I can visit all the Nodes neighbors less than a Node in time
414 // O(# of neighbors), but I have to visit all the Nodes greater than a
415 // given Node and search them for an instance, i.e., time O(#MaxLRG)).
416 _ifg->SquareUp();
417
418 // Aggressive (but pessimistic) copy coalescing.
419 // This pass works on virtual copies. Any virtual copies which are not
420 // coalesced get manifested as actual copies
421 {
422 Compile::TracePhase tp("chaitinCoalesce1", &timers[_t_chaitinCoalesce1]);
423
424 PhaseAggressiveCoalesce coalesce(*this);
425 coalesce.coalesce_driver();
426 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do
427 // not match the Phi itself, insert a copy.
428 coalesce.insert_copies(_matcher);
429 if (C->failing()) {
430 return;
431 }
432 }
433
434 // After aggressive coalesce, attempt a first cut at coloring.
435 // To color, we need the IFG and for that we need LIVE.
436 {
437 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
438 _live = NULL;
439 rm.reset_to_mark(); // Reclaim working storage
440 IndexSet::reset_memory(C, &live_arena);
441 ifg.init(_lrg_map.max_lrg_id());
442 gather_lrg_masks( true );
443 live.compute(_lrg_map.max_lrg_id());
444 _live = &live;
445 }
446
447 // Build physical interference graph
448 uint must_spill = 0;
449 must_spill = build_ifg_physical(&live_arena);
450 // If we have a guaranteed spill, might as well spill now
451 if (must_spill) {
452 if(!_lrg_map.max_lrg_id()) {
453 return;
454 }
455 // Bail out if unique gets too large (ie - unique > MaxNodeLimit)
456 C->check_node_count(10*must_spill, "out of nodes before split");
457 if (C->failing()) {
458 return;
459 }
460
461 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
462 _lrg_map.set_max_lrg_id(new_max_lrg_id);
463 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
464 // or we failed to split
465 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split");
466 if (C->failing()) {
467 return;
468 }
469
470 NOT_PRODUCT(C->verify_graph_edges();)
471
472 compact(); // Compact LRGs; return new lower max lrg
473
474 {
475 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
476 _live = NULL;
477 rm.reset_to_mark(); // Reclaim working storage
478 IndexSet::reset_memory(C, &live_arena);
479 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph
480 gather_lrg_masks( true ); // Collect intersect mask
481 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE
482 _live = &live;
483 }
484 build_ifg_physical(&live_arena);
485 _ifg->SquareUp();
486 _ifg->Compute_Effective_Degree();
487 // Only do conservative coalescing if requested
488 if (OptoCoalesce) {
489 Compile::TracePhase tp("chaitinCoalesce2", &timers[_t_chaitinCoalesce2]);
490 // Conservative (and pessimistic) copy coalescing of those spills
491 PhaseConservativeCoalesce coalesce(*this);
492 // If max live ranges greater than cutoff, don't color the stack.
493 // This cutoff can be larger than below since it is only done once.
494 coalesce.coalesce_driver();
495 }
496 _lrg_map.compress_uf_map_for_nodes();
497
498 #ifdef ASSERT
499 verify(&live_arena, true);
500 #endif
501 } else {
502 ifg.SquareUp();
503 ifg.Compute_Effective_Degree();
504 #ifdef ASSERT
505 set_was_low();
506 #endif
507 }
508
509 // Prepare for Simplify & Select
510 cache_lrg_info(); // Count degree of LRGs
511
512 // Simplify the InterFerence Graph by removing LRGs of low degree.
513 // LRGs of low degree are trivially colorable.
514 Simplify();
515
516 // Select colors by re-inserting LRGs back into the IFG in reverse order.
517 // Return whether or not something spills.
518 uint spills = Select( );
519
520 // If we spill, split and recycle the entire thing
521 while( spills ) {
522 if( _trip_cnt++ > 24 ) {
523 DEBUG_ONLY( dump_for_spill_split_recycle(); )
524 if( _trip_cnt > 27 ) {
525 C->record_method_not_compilable("failed spill-split-recycle sanity check");
526 return;
527 }
528 }
529
530 if (!_lrg_map.max_lrg_id()) {
531 return;
532 }
533 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere
534 _lrg_map.set_max_lrg_id(new_max_lrg_id);
535 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
536 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split");
537 if (C->failing()) {
538 return;
539 }
540
541 compact(); // Compact LRGs; return new lower max lrg
542
543 // Nuke the live-ness and interference graph and LiveRanGe info
544 {
545 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
546 _live = NULL;
547 rm.reset_to_mark(); // Reclaim working storage
548 IndexSet::reset_memory(C, &live_arena);
549 ifg.init(_lrg_map.max_lrg_id());
550
551 // Create LiveRanGe array.
552 // Intersect register masks for all USEs and DEFs
553 gather_lrg_masks(true);
554 live.compute(_lrg_map.max_lrg_id());
555 _live = &live;
556 }
557 must_spill = build_ifg_physical(&live_arena);
558 _ifg->SquareUp();
559 _ifg->Compute_Effective_Degree();
560
561 // Only do conservative coalescing if requested
562 if (OptoCoalesce) {
563 Compile::TracePhase tp("chaitinCoalesce3", &timers[_t_chaitinCoalesce3]);
564 // Conservative (and pessimistic) copy coalescing
565 PhaseConservativeCoalesce coalesce(*this);
566 // Check for few live ranges determines how aggressive coalesce is.
567 coalesce.coalesce_driver();
568 }
569 _lrg_map.compress_uf_map_for_nodes();
570 #ifdef ASSERT
571 verify(&live_arena, true);
572 #endif
573 cache_lrg_info(); // Count degree of LRGs
574
575 // Simplify the InterFerence Graph by removing LRGs of low degree.
576 // LRGs of low degree are trivially colorable.
577 Simplify();
578
579 // Select colors by re-inserting LRGs back into the IFG in reverse order.
580 // Return whether or not something spills.
581 spills = Select();
582 }
583
584 // Count number of Simplify-Select trips per coloring success.
585 _allocator_attempts += _trip_cnt + 1;
586 _allocator_successes += 1;
587
588 // Peephole remove copies
589 post_allocate_copy_removal();
590
591 // Merge multidefs if multiple defs representing the same value are used in a single block.
592 merge_multidefs();
593
594 #ifdef ASSERT
595 // Veify the graph after RA.
596 verify(&live_arena);
597 #endif
598
599 // max_reg is past the largest *register* used.
600 // Convert that to a frame_slot number.
601 if (_max_reg <= _matcher._new_SP) {
602 _framesize = C->out_preserve_stack_slots();
603 }
604 else {
605 _framesize = _max_reg -_matcher._new_SP;
606 }
607 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough");
608
609 // This frame must preserve the required fp alignment
610 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots());
611 assert(_framesize <= 1000000, "sanity check");
612 #ifndef PRODUCT
613 _total_framesize += _framesize;
614 if ((int)_framesize > _max_framesize) {
615 _max_framesize = _framesize;
616 }
617 #endif
618
619 // Convert CISC spills
620 fixup_spills();
621
622 // Log regalloc results
623 CompileLog* log = Compile::current()->log();
624 if (log != NULL) {
625 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing());
626 }
627
628 if (C->failing()) {
629 return;
630 }
631
632 NOT_PRODUCT(C->verify_graph_edges();)
633
634 // Move important info out of the live_arena to longer lasting storage.
635 alloc_node_regs(_lrg_map.size());
636 for (uint i=0; i < _lrg_map.size(); i++) {
637 if (_lrg_map.live_range_id(i)) { // Live range associated with Node?
638 LRG &lrg = lrgs(_lrg_map.live_range_id(i));
639 if (!lrg.alive()) {
640 set_bad(i);
641 } else if (lrg.num_regs() == 1) {
642 set1(i, lrg.reg());
643 } else { // Must be a register-set
644 if (!lrg._fat_proj) { // Must be aligned adjacent register set
645 // Live ranges record the highest register in their mask.
646 // We want the low register for the AD file writer's convenience.
647 OptoReg::Name hi = lrg.reg(); // Get hi register
648 int num_regs = lrg.num_regs();
649 if (lrg.is_scalable() && OptoReg::is_stack(hi)) {
650 // For scalable vector registers, when they are allocated in physical
651 // registers, num_regs is RegMask::SlotsPerVecA for reg mask of scalable
652 // vector. If they are allocated on stack, we need to get the actual
653 // num_regs, which reflects the physical length of scalable registers.
654 num_regs = lrg.scalable_reg_slots();
655 }
656 OptoReg::Name lo = OptoReg::add(hi, (1-num_regs)); // Find lo
657 // We have to use pair [lo,lo+1] even for wide vectors because
658 // the rest of code generation works only with pairs. It is safe
659 // since for registers encoding only 'lo' is used.
660 // Second reg from pair is used in ScheduleAndBundle on SPARC where
661 // vector max size is 8 which corresponds to registers pair.
662 // It is also used in BuildOopMaps but oop operations are not
663 // vectorized.
664 set2(i, lo);
665 } else { // Misaligned; extract 2 bits
666 OptoReg::Name hi = lrg.reg(); // Get hi register
667 lrg.Remove(hi); // Yank from mask
668 int lo = lrg.mask().find_first_elem(); // Find lo
669 set_pair(i, hi, lo);
670 }
671 }
672 if( lrg._is_oop ) _node_oops.set(i);
673 } else {
674 set_bad(i);
675 }
676 }
677
678 // Done!
679 _live = NULL;
680 _ifg = NULL;
681 C->set_indexSet_arena(NULL); // ResourceArea is at end of scope
682 }
683
de_ssa()684 void PhaseChaitin::de_ssa() {
685 // Set initial Names for all Nodes. Most Nodes get the virtual register
686 // number. A few get the ZERO live range number. These do not
687 // get allocated, but instead rely on correct scheduling to ensure that
688 // only one instance is simultaneously live at a time.
689 uint lr_counter = 1;
690 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
691 Block* block = _cfg.get_block(i);
692 uint cnt = block->number_of_nodes();
693
694 // Handle all the normal Nodes in the block
695 for( uint j = 0; j < cnt; j++ ) {
696 Node *n = block->get_node(j);
697 // Pre-color to the zero live range, or pick virtual register
698 const RegMask &rm = n->out_RegMask();
699 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0);
700 }
701 }
702
703 // Reset the Union-Find mapping to be identity
704 _lrg_map.reset_uf_map(lr_counter);
705 }
706
mark_ssa()707 void PhaseChaitin::mark_ssa() {
708 // Use ssa names to populate the live range maps or if no mask
709 // is available, use the 0 entry.
710 uint max_idx = 0;
711 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) {
712 Block* block = _cfg.get_block(i);
713 uint cnt = block->number_of_nodes();
714
715 // Handle all the normal Nodes in the block
716 for ( uint j = 0; j < cnt; j++ ) {
717 Node *n = block->get_node(j);
718 // Pre-color to the zero live range, or pick virtual register
719 const RegMask &rm = n->out_RegMask();
720 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0);
721 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx;
722 }
723 }
724 _lrg_map.set_max_lrg_id(max_idx+1);
725
726 // Reset the Union-Find mapping to be identity
727 _lrg_map.reset_uf_map(max_idx+1);
728 }
729
730
731 // Gather LiveRanGe information, including register masks. Modification of
732 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce.
gather_lrg_masks(bool after_aggressive)733 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) {
734
735 // Nail down the frame pointer live range
736 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr));
737 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite
738
739 // For all blocks
740 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
741 Block* block = _cfg.get_block(i);
742
743 // For all instructions
744 for (uint j = 1; j < block->number_of_nodes(); j++) {
745 Node* n = block->get_node(j);
746 uint input_edge_start =1; // Skip control most nodes
747 bool is_machine_node = false;
748 if (n->is_Mach()) {
749 is_machine_node = true;
750 input_edge_start = n->as_Mach()->oper_input_base();
751 }
752 uint idx = n->is_Copy();
753
754 // Get virtual register number, same as LiveRanGe index
755 uint vreg = _lrg_map.live_range_id(n);
756 LRG& lrg = lrgs(vreg);
757 if (vreg) { // No vreg means un-allocable (e.g. memory)
758
759 // Check for float-vs-int live range (used in register-pressure
760 // calculations)
761 const Type *n_type = n->bottom_type();
762 if (n_type->is_floatingpoint()) {
763 lrg._is_float = 1;
764 }
765
766 // Check for twice prior spilling. Once prior spilling might have
767 // spilled 'soft', 2nd prior spill should have spilled 'hard' and
768 // further spilling is unlikely to make progress.
769 if (_spilled_once.test(n->_idx)) {
770 lrg._was_spilled1 = 1;
771 if (_spilled_twice.test(n->_idx)) {
772 lrg._was_spilled2 = 1;
773 }
774 }
775
776 #ifndef PRODUCT
777 // Collect bits not used by product code, but which may be useful for
778 // debugging.
779
780 // Collect has-copy bit
781 if (idx) {
782 lrg._has_copy = 1;
783 uint clidx = _lrg_map.live_range_id(n->in(idx));
784 LRG& copy_src = lrgs(clidx);
785 copy_src._has_copy = 1;
786 }
787
788 if (trace_spilling() && lrg._def != NULL) {
789 // collect defs for MultiDef printing
790 if (lrg._defs == NULL) {
791 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, NULL);
792 lrg._defs->append(lrg._def);
793 }
794 lrg._defs->append(n);
795 }
796 #endif
797
798 // Check for a single def LRG; these can spill nicely
799 // via rematerialization. Flag as NULL for no def found
800 // yet, or 'n' for single def or -1 for many defs.
801 lrg._def = lrg._def ? NodeSentinel : n;
802
803 // Limit result register mask to acceptable registers
804 const RegMask &rm = n->out_RegMask();
805 lrg.AND( rm );
806
807 uint ireg = n->ideal_reg();
808 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP,
809 "oops must be in Op_RegP's" );
810
811 // Check for vector live range (only if vector register is used).
812 // On SPARC vector uses RegD which could be misaligned so it is not
813 // processes as vector in RA.
814 if (RegMask::is_vector(ireg)) {
815 lrg._is_vector = 1;
816 if (ireg == Op_VecA) {
817 assert(Matcher::supports_scalable_vector(), "scalable vector should be supported");
818 lrg._is_scalable = 1;
819 // For scalable vector, when it is allocated in physical register,
820 // num_regs is RegMask::SlotsPerVecA for reg mask,
821 // which may not be the actual physical register size.
822 // If it is allocated in stack, we need to get the actual
823 // physical length of scalable vector register.
824 lrg.set_scalable_reg_slots(Matcher::scalable_vector_reg_size(T_FLOAT));
825 }
826 }
827 assert(n_type->isa_vect() == NULL || lrg._is_vector || ireg == Op_RegD || ireg == Op_RegL,
828 "vector must be in vector registers");
829
830 // Check for bound register masks
831 const RegMask &lrgmask = lrg.mask();
832 if (lrgmask.is_bound(ireg)) {
833 lrg._is_bound = 1;
834 }
835
836 // Check for maximum frequency value
837 if (lrg._maxfreq < block->_freq) {
838 lrg._maxfreq = block->_freq;
839 }
840
841 // Check for oop-iness, or long/double
842 // Check for multi-kill projection
843 switch (ireg) {
844 case MachProjNode::fat_proj:
845 // Fat projections have size equal to number of registers killed
846 lrg.set_num_regs(rm.Size());
847 lrg.set_reg_pressure(lrg.num_regs());
848 lrg._fat_proj = 1;
849 lrg._is_bound = 1;
850 break;
851 case Op_RegP:
852 #ifdef _LP64
853 lrg.set_num_regs(2); // Size is 2 stack words
854 #else
855 lrg.set_num_regs(1); // Size is 1 stack word
856 #endif
857 // Register pressure is tracked relative to the maximum values
858 // suggested for that platform, INTPRESSURE and FLOATPRESSURE,
859 // and relative to other types which compete for the same regs.
860 //
861 // The following table contains suggested values based on the
862 // architectures as defined in each .ad file.
863 // INTPRESSURE and FLOATPRESSURE may be tuned differently for
864 // compile-speed or performance.
865 // Note1:
866 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1
867 // since .ad registers are defined as high and low halves.
868 // These reg_pressure values remain compatible with the code
869 // in is_high_pressure() which relates get_invalid_mask_size(),
870 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE.
871 // Note2:
872 // SPARC -d32 has 24 registers available for integral values,
873 // but only 10 of these are safe for 64-bit longs.
874 // Using set_reg_pressure(2) for both int and long means
875 // the allocator will believe it can fit 26 longs into
876 // registers. Using 2 for longs and 1 for ints means the
877 // allocator will attempt to put 52 integers into registers.
878 // The settings below limit this problem to methods with
879 // many long values which are being run on 32-bit SPARC.
880 //
881 // ------------------- reg_pressure --------------------
882 // Each entry is reg_pressure_per_value,number_of_regs
883 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE
884 // IA32 2 1 1 1 1 6 6
885 // IA64 1 1 1 1 1 50 41
886 // SPARC 2 2 2 2 2 48 (24) 52 (26)
887 // SPARCV9 2 2 2 2 2 48 (24) 52 (26)
888 // AMD64 1 1 1 1 1 14 15
889 // -----------------------------------------------------
890 lrg.set_reg_pressure(1); // normally one value per register
891 if( n_type->isa_oop_ptr() ) {
892 lrg._is_oop = 1;
893 }
894 break;
895 case Op_RegL: // Check for long or double
896 case Op_RegD:
897 lrg.set_num_regs(2);
898 // Define platform specific register pressure
899 #if defined(ARM32)
900 lrg.set_reg_pressure(2);
901 #elif defined(IA32)
902 if( ireg == Op_RegL ) {
903 lrg.set_reg_pressure(2);
904 } else {
905 lrg.set_reg_pressure(1);
906 }
907 #else
908 lrg.set_reg_pressure(1); // normally one value per register
909 #endif
910 // If this def of a double forces a mis-aligned double,
911 // flag as '_fat_proj' - really flag as allowing misalignment
912 // AND changes how we count interferences. A mis-aligned
913 // double can interfere with TWO aligned pairs, or effectively
914 // FOUR registers!
915 if (rm.is_misaligned_pair()) {
916 lrg._fat_proj = 1;
917 lrg._is_bound = 1;
918 }
919 break;
920 case Op_RegF:
921 case Op_RegI:
922 case Op_RegN:
923 case Op_RegFlags:
924 case 0: // not an ideal register
925 lrg.set_num_regs(1);
926 lrg.set_reg_pressure(1);
927 break;
928 case Op_VecA:
929 assert(Matcher::supports_scalable_vector(), "does not support scalable vector");
930 assert(RegMask::num_registers(Op_VecA) == RegMask::SlotsPerVecA, "sanity");
931 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecA), "vector should be aligned");
932 lrg.set_num_regs(RegMask::SlotsPerVecA);
933 lrg.set_reg_pressure(1);
934 break;
935 case Op_VecS:
936 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity");
937 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity");
938 lrg.set_num_regs(RegMask::SlotsPerVecS);
939 lrg.set_reg_pressure(1);
940 break;
941 case Op_VecD:
942 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity");
943 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity");
944 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned");
945 lrg.set_num_regs(RegMask::SlotsPerVecD);
946 lrg.set_reg_pressure(1);
947 break;
948 case Op_VecX:
949 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity");
950 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity");
951 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned");
952 lrg.set_num_regs(RegMask::SlotsPerVecX);
953 lrg.set_reg_pressure(1);
954 break;
955 case Op_VecY:
956 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity");
957 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity");
958 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned");
959 lrg.set_num_regs(RegMask::SlotsPerVecY);
960 lrg.set_reg_pressure(1);
961 break;
962 case Op_VecZ:
963 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity");
964 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity");
965 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned");
966 lrg.set_num_regs(RegMask::SlotsPerVecZ);
967 lrg.set_reg_pressure(1);
968 break;
969 default:
970 ShouldNotReachHere();
971 }
972 }
973
974 // Now do the same for inputs
975 uint cnt = n->req();
976 // Setup for CISC SPILLING
977 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable;
978 if( UseCISCSpill && after_aggressive ) {
979 inp = n->cisc_operand();
980 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable )
981 // Convert operand number to edge index number
982 inp = n->as_Mach()->operand_index(inp);
983 }
984
985 // Prepare register mask for each input
986 for( uint k = input_edge_start; k < cnt; k++ ) {
987 uint vreg = _lrg_map.live_range_id(n->in(k));
988 if (!vreg) {
989 continue;
990 }
991
992 // If this instruction is CISC Spillable, add the flags
993 // bit to its appropriate input
994 if( UseCISCSpill && after_aggressive && inp == k ) {
995 #ifndef PRODUCT
996 if( TraceCISCSpill ) {
997 tty->print(" use_cisc_RegMask: ");
998 n->dump();
999 }
1000 #endif
1001 n->as_Mach()->use_cisc_RegMask();
1002 }
1003
1004 if (is_machine_node && _scheduling_info_generated) {
1005 MachNode* cur_node = n->as_Mach();
1006 // this is cleaned up by register allocation
1007 if (k >= cur_node->num_opnds()) continue;
1008 }
1009
1010 LRG &lrg = lrgs(vreg);
1011 // // Testing for floating point code shape
1012 // Node *test = n->in(k);
1013 // if( test->is_Mach() ) {
1014 // MachNode *m = test->as_Mach();
1015 // int op = m->ideal_Opcode();
1016 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) {
1017 // int zzz = 1;
1018 // }
1019 // }
1020
1021 // Limit result register mask to acceptable registers.
1022 // Do not limit registers from uncommon uses before
1023 // AggressiveCoalesce. This effectively pre-virtual-splits
1024 // around uncommon uses of common defs.
1025 const RegMask &rm = n->in_RegMask(k);
1026 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) {
1027 // Since we are BEFORE aggressive coalesce, leave the register
1028 // mask untrimmed by the call. This encourages more coalescing.
1029 // Later, AFTER aggressive, this live range will have to spill
1030 // but the spiller handles slow-path calls very nicely.
1031 } else {
1032 lrg.AND( rm );
1033 }
1034
1035 // Check for bound register masks
1036 const RegMask &lrgmask = lrg.mask();
1037 uint kreg = n->in(k)->ideal_reg();
1038 bool is_vect = RegMask::is_vector(kreg);
1039 assert(n->in(k)->bottom_type()->isa_vect() == NULL ||
1040 is_vect || kreg == Op_RegD || kreg == Op_RegL,
1041 "vector must be in vector registers");
1042 if (lrgmask.is_bound(kreg))
1043 lrg._is_bound = 1;
1044
1045 // If this use of a double forces a mis-aligned double,
1046 // flag as '_fat_proj' - really flag as allowing misalignment
1047 // AND changes how we count interferences. A mis-aligned
1048 // double can interfere with TWO aligned pairs, or effectively
1049 // FOUR registers!
1050 #ifdef ASSERT
1051 if (is_vect && !_scheduling_info_generated) {
1052 if (lrg.num_regs() != 0) {
1053 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned");
1054 assert(!lrg._fat_proj, "sanity");
1055 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity");
1056 } else {
1057 assert(n->is_Phi(), "not all inputs processed only if Phi");
1058 }
1059 }
1060 #endif
1061 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) {
1062 lrg._fat_proj = 1;
1063 lrg._is_bound = 1;
1064 }
1065 // if the LRG is an unaligned pair, we will have to spill
1066 // so clear the LRG's register mask if it is not already spilled
1067 if (!is_vect && !n->is_SpillCopy() &&
1068 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) &&
1069 lrgmask.is_misaligned_pair()) {
1070 lrg.Clear();
1071 }
1072
1073 // Check for maximum frequency value
1074 if (lrg._maxfreq < block->_freq) {
1075 lrg._maxfreq = block->_freq;
1076 }
1077
1078 } // End for all allocated inputs
1079 } // end for all instructions
1080 } // end for all blocks
1081
1082 // Final per-liverange setup
1083 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) {
1084 LRG &lrg = lrgs(i2);
1085 assert(!lrg._is_vector || !lrg._fat_proj, "sanity");
1086 if (lrg.num_regs() > 1 && !lrg._fat_proj) {
1087 lrg.clear_to_sets();
1088 }
1089 lrg.compute_set_mask_size();
1090 if (lrg.not_free()) { // Handle case where we lose from the start
1091 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG));
1092 lrg._direct_conflict = 1;
1093 }
1094 lrg.set_degree(0); // no neighbors in IFG yet
1095 }
1096 }
1097
1098 // Set the was-lo-degree bit. Conservative coalescing should not change the
1099 // colorability of the graph. If any live range was of low-degree before
1100 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
1101 // The bit is checked in Simplify.
set_was_low()1102 void PhaseChaitin::set_was_low() {
1103 #ifdef ASSERT
1104 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1105 int size = lrgs(i).num_regs();
1106 uint old_was_lo = lrgs(i)._was_lo;
1107 lrgs(i)._was_lo = 0;
1108 if( lrgs(i).lo_degree() ) {
1109 lrgs(i)._was_lo = 1; // Trivially of low degree
1110 } else { // Else check the Brigg's assertion
1111 // Brigg's observation is that the lo-degree neighbors of a
1112 // hi-degree live range will not interfere with the color choices
1113 // of said hi-degree live range. The Simplify reverse-stack-coloring
1114 // order takes care of the details. Hence you do not have to count
1115 // low-degree neighbors when determining if this guy colors.
1116 int briggs_degree = 0;
1117 IndexSet *s = _ifg->neighbors(i);
1118 IndexSetIterator elements(s);
1119 uint lidx;
1120 while((lidx = elements.next()) != 0) {
1121 if( !lrgs(lidx).lo_degree() )
1122 briggs_degree += MAX2(size,lrgs(lidx).num_regs());
1123 }
1124 if( briggs_degree < lrgs(i).degrees_of_freedom() )
1125 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion
1126 }
1127 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease");
1128 }
1129 #endif
1130 }
1131
1132 // Compute cost/area ratio, in case we spill. Build the lo-degree list.
cache_lrg_info()1133 void PhaseChaitin::cache_lrg_info( ) {
1134 Compile::TracePhase tp("chaitinCacheLRG", &timers[_t_chaitinCacheLRG]);
1135
1136 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) {
1137 LRG &lrg = lrgs(i);
1138
1139 // Check for being of low degree: means we can be trivially colored.
1140 // Low degree, dead or must-spill guys just get to simplify right away
1141 if( lrg.lo_degree() ||
1142 !lrg.alive() ||
1143 lrg._must_spill ) {
1144 // Split low degree list into those guys that must get a
1145 // register and those that can go to register or stack.
1146 // The idea is LRGs that can go register or stack color first when
1147 // they have a good chance of getting a register. The register-only
1148 // lo-degree live ranges always get a register.
1149 OptoReg::Name hi_reg = lrg.mask().find_last_elem();
1150 if( OptoReg::is_stack(hi_reg)) { // Can go to stack?
1151 lrg._next = _lo_stk_degree;
1152 _lo_stk_degree = i;
1153 } else {
1154 lrg._next = _lo_degree;
1155 _lo_degree = i;
1156 }
1157 } else { // Else high degree
1158 lrgs(_hi_degree)._prev = i;
1159 lrg._next = _hi_degree;
1160 lrg._prev = 0;
1161 _hi_degree = i;
1162 }
1163 }
1164 }
1165
1166 // Simplify the IFG by removing LRGs of low degree.
Simplify()1167 void PhaseChaitin::Simplify( ) {
1168 Compile::TracePhase tp("chaitinSimplify", &timers[_t_chaitinSimplify]);
1169
1170 while( 1 ) { // Repeat till simplified it all
1171 // May want to explore simplifying lo_degree before _lo_stk_degree.
1172 // This might result in more spills coloring into registers during
1173 // Select().
1174 while( _lo_degree || _lo_stk_degree ) {
1175 // If possible, pull from lo_stk first
1176 uint lo;
1177 if( _lo_degree ) {
1178 lo = _lo_degree;
1179 _lo_degree = lrgs(lo)._next;
1180 } else {
1181 lo = _lo_stk_degree;
1182 _lo_stk_degree = lrgs(lo)._next;
1183 }
1184
1185 // Put the simplified guy on the simplified list.
1186 lrgs(lo)._next = _simplified;
1187 _simplified = lo;
1188 // If this guy is "at risk" then mark his current neighbors
1189 if (lrgs(lo)._at_risk && !_ifg->neighbors(lo)->is_empty()) {
1190 IndexSetIterator elements(_ifg->neighbors(lo));
1191 uint datum;
1192 while ((datum = elements.next()) != 0) {
1193 lrgs(datum)._risk_bias = lo;
1194 }
1195 }
1196
1197 // Yank this guy from the IFG.
1198 IndexSet *adj = _ifg->remove_node(lo);
1199 if (adj->is_empty()) {
1200 continue;
1201 }
1202
1203 // If any neighbors' degrees fall below their number of
1204 // allowed registers, then put that neighbor on the low degree
1205 // list. Note that 'degree' can only fall and 'numregs' is
1206 // unchanged by this action. Thus the two are equal at most once,
1207 // so LRGs hit the lo-degree worklist at most once.
1208 IndexSetIterator elements(adj);
1209 uint neighbor;
1210 while ((neighbor = elements.next()) != 0) {
1211 LRG *n = &lrgs(neighbor);
1212 #ifdef ASSERT
1213 if (VerifyRegisterAllocator) {
1214 assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
1215 }
1216 #endif
1217
1218 // Check for just becoming of-low-degree just counting registers.
1219 // _must_spill live ranges are already on the low degree list.
1220 if (n->just_lo_degree() && !n->_must_spill) {
1221 assert(!_ifg->_yanked->test(neighbor), "Cannot move to lo degree twice");
1222 // Pull from hi-degree list
1223 uint prev = n->_prev;
1224 uint next = n->_next;
1225 if (prev) {
1226 lrgs(prev)._next = next;
1227 } else {
1228 _hi_degree = next;
1229 }
1230 lrgs(next)._prev = prev;
1231 n->_next = _lo_degree;
1232 _lo_degree = neighbor;
1233 }
1234 }
1235 } // End of while lo-degree/lo_stk_degree worklist not empty
1236
1237 // Check for got everything: is hi-degree list empty?
1238 if (!_hi_degree) break;
1239
1240 // Time to pick a potential spill guy
1241 uint lo_score = _hi_degree;
1242 double score = lrgs(lo_score).score();
1243 double area = lrgs(lo_score)._area;
1244 double cost = lrgs(lo_score)._cost;
1245 bool bound = lrgs(lo_score)._is_bound;
1246
1247 // Find cheapest guy
1248 debug_only( int lo_no_simplify=0; );
1249 for (uint i = _hi_degree; i; i = lrgs(i)._next) {
1250 assert(!_ifg->_yanked->test(i), "");
1251 // It's just vaguely possible to move hi-degree to lo-degree without
1252 // going through a just-lo-degree stage: If you remove a double from
1253 // a float live range it's degree will drop by 2 and you can skip the
1254 // just-lo-degree stage. It's very rare (shows up after 5000+ methods
1255 // in -Xcomp of Java2Demo). So just choose this guy to simplify next.
1256 if( lrgs(i).lo_degree() ) {
1257 lo_score = i;
1258 break;
1259 }
1260 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; );
1261 double iscore = lrgs(i).score();
1262 double iarea = lrgs(i)._area;
1263 double icost = lrgs(i)._cost;
1264 bool ibound = lrgs(i)._is_bound;
1265
1266 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area
1267 // wins. Ties happen because all live ranges in question have spilled
1268 // a few times before and the spill-score adds a huge number which
1269 // washes out the low order bits. We are choosing the lesser of 2
1270 // evils; in this case pick largest area to spill.
1271 // Ties also happen when live ranges are defined and used only inside
1272 // one block. In which case their area is 0 and score set to max.
1273 // In such case choose bound live range over unbound to free registers
1274 // or with smaller cost to spill.
1275 if ( iscore < score ||
1276 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ||
1277 (iscore == score && iarea == area &&
1278 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) {
1279 lo_score = i;
1280 score = iscore;
1281 area = iarea;
1282 cost = icost;
1283 bound = ibound;
1284 }
1285 }
1286 LRG *lo_lrg = &lrgs(lo_score);
1287 // The live range we choose for spilling is either hi-degree, or very
1288 // rarely it can be low-degree. If we choose a hi-degree live range
1289 // there better not be any lo-degree choices.
1290 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" );
1291
1292 // Pull from hi-degree list
1293 uint prev = lo_lrg->_prev;
1294 uint next = lo_lrg->_next;
1295 if( prev ) lrgs(prev)._next = next;
1296 else _hi_degree = next;
1297 lrgs(next)._prev = prev;
1298 // Jam him on the lo-degree list, despite his high degree.
1299 // Maybe he'll get a color, and maybe he'll spill.
1300 // Only Select() will know.
1301 lrgs(lo_score)._at_risk = true;
1302 _lo_degree = lo_score;
1303 lo_lrg->_next = 0;
1304
1305 } // End of while not simplified everything
1306
1307 }
1308
1309 // Is 'reg' register legal for 'lrg'?
is_legal_reg(LRG & lrg,OptoReg::Name reg,int chunk)1310 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) {
1311 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) &&
1312 lrg.mask().Member(OptoReg::add(reg,-chunk))) {
1313 // RA uses OptoReg which represent the highest element of a registers set.
1314 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set
1315 // in which XMMd is used by RA to represent such vectors. A double value
1316 // uses [XMM,XMMb] pairs and XMMb is used by RA for it.
1317 // The register mask uses largest bits set of overlapping register sets.
1318 // On x86 with AVX it uses 8 bits for each XMM registers set.
1319 //
1320 // The 'lrg' already has cleared-to-set register mask (done in Select()
1321 // before calling choose_color()). Passing mask.Member(reg) check above
1322 // indicates that the size (num_regs) of 'reg' set is less or equal to
1323 // 'lrg' set size.
1324 // For set size 1 any register which is member of 'lrg' mask is legal.
1325 if (lrg.num_regs()==1)
1326 return true;
1327 // For larger sets only an aligned register with the same set size is legal.
1328 int mask = lrg.num_regs()-1;
1329 if ((reg&mask) == mask)
1330 return true;
1331 }
1332 return false;
1333 }
1334
find_first_set(LRG & lrg,RegMask mask,int chunk)1335 static OptoReg::Name find_first_set(LRG &lrg, RegMask mask, int chunk) {
1336 int num_regs = lrg.num_regs();
1337 OptoReg::Name assigned = mask.find_first_set(lrg, num_regs);
1338
1339 if (lrg.is_scalable()) {
1340 // a physical register is found
1341 if (chunk == 0 && OptoReg::is_reg(assigned)) {
1342 return assigned;
1343 }
1344
1345 // find available stack slots for scalable register
1346 if (lrg._is_vector) {
1347 num_regs = lrg.scalable_reg_slots();
1348 // if actual scalable vector register is exactly SlotsPerVecA * 32 bits
1349 if (num_regs == RegMask::SlotsPerVecA) {
1350 return assigned;
1351 }
1352
1353 // mask has been cleared out by clear_to_sets(SlotsPerVecA) before choose_color, but it
1354 // does not work for scalable size. We have to find adjacent scalable_reg_slots() bits
1355 // instead of SlotsPerVecA bits.
1356 assigned = mask.find_first_set(lrg, num_regs); // find highest valid reg
1357 while (OptoReg::is_valid(assigned) && RegMask::can_represent(assigned)) {
1358 // Verify the found reg has scalable_reg_slots() bits set.
1359 if (mask.is_valid_reg(assigned, num_regs)) {
1360 return assigned;
1361 } else {
1362 // Remove more for each iteration
1363 mask.Remove(assigned - num_regs + 1); // Unmask the lowest reg
1364 mask.clear_to_sets(RegMask::SlotsPerVecA); // Align by SlotsPerVecA bits
1365 assigned = mask.find_first_set(lrg, num_regs);
1366 }
1367 }
1368 return OptoReg::Bad; // will cause chunk change, and retry next chunk
1369 }
1370 }
1371
1372 return assigned;
1373 }
1374
1375 // Choose a color using the biasing heuristic
bias_color(LRG & lrg,int chunk)1376 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) {
1377
1378 // Check for "at_risk" LRG's
1379 uint risk_lrg = _lrg_map.find(lrg._risk_bias);
1380 if (risk_lrg != 0 && !_ifg->neighbors(risk_lrg)->is_empty()) {
1381 // Walk the colored neighbors of the "at_risk" candidate
1382 // Choose a color which is both legal and already taken by a neighbor
1383 // of the "at_risk" candidate in order to improve the chances of the
1384 // "at_risk" candidate of coloring
1385 IndexSetIterator elements(_ifg->neighbors(risk_lrg));
1386 uint datum;
1387 while ((datum = elements.next()) != 0) {
1388 OptoReg::Name reg = lrgs(datum).reg();
1389 // If this LRG's register is legal for us, choose it
1390 if (is_legal_reg(lrg, reg, chunk))
1391 return reg;
1392 }
1393 }
1394
1395 uint copy_lrg = _lrg_map.find(lrg._copy_bias);
1396 if (copy_lrg != 0) {
1397 // If he has a color,
1398 if(!_ifg->_yanked->test(copy_lrg)) {
1399 OptoReg::Name reg = lrgs(copy_lrg).reg();
1400 // And it is legal for you,
1401 if (is_legal_reg(lrg, reg, chunk))
1402 return reg;
1403 } else if( chunk == 0 ) {
1404 // Choose a color which is legal for him
1405 RegMask tempmask = lrg.mask();
1406 tempmask.AND(lrgs(copy_lrg).mask());
1407 tempmask.clear_to_sets(lrg.num_regs());
1408 OptoReg::Name reg = find_first_set(lrg, tempmask, chunk);
1409 if (OptoReg::is_valid(reg))
1410 return reg;
1411 }
1412 }
1413
1414 // If no bias info exists, just go with the register selection ordering
1415 if (lrg._is_vector || lrg.num_regs() == 2) {
1416 // Find an aligned set
1417 return OptoReg::add(find_first_set(lrg, lrg.mask(), chunk), chunk);
1418 }
1419
1420 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate
1421 // copy removal to remove many more copies, by preventing a just-assigned
1422 // register from being repeatedly assigned.
1423 OptoReg::Name reg = lrg.mask().find_first_elem();
1424 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) {
1425 // This 'Remove; find; Insert' idiom is an expensive way to find the
1426 // SECOND element in the mask.
1427 lrg.Remove(reg);
1428 OptoReg::Name reg2 = lrg.mask().find_first_elem();
1429 lrg.Insert(reg);
1430 if( OptoReg::is_reg(reg2))
1431 reg = reg2;
1432 }
1433 return OptoReg::add( reg, chunk );
1434 }
1435
1436 // Choose a color in the current chunk
choose_color(LRG & lrg,int chunk)1437 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) {
1438 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)");
1439 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)");
1440
1441 if( lrg.num_regs() == 1 || // Common Case
1442 !lrg._fat_proj ) // Aligned+adjacent pairs ok
1443 // Use a heuristic to "bias" the color choice
1444 return bias_color(lrg, chunk);
1445
1446 assert(!lrg._is_vector, "should be not vector here" );
1447 assert( lrg.num_regs() >= 2, "dead live ranges do not color" );
1448
1449 // Fat-proj case or misaligned double argument.
1450 assert(lrg.compute_mask_size() == lrg.num_regs() ||
1451 lrg.num_regs() == 2,"fat projs exactly color" );
1452 assert( !chunk, "always color in 1st chunk" );
1453 // Return the highest element in the set.
1454 return lrg.mask().find_last_elem();
1455 }
1456
1457 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted
1458 // in reverse order of removal. As long as nothing of hi-degree was yanked,
1459 // everything going back is guaranteed a color. Select that color. If some
1460 // hi-degree LRG cannot get a color then we record that we must spill.
Select()1461 uint PhaseChaitin::Select( ) {
1462 Compile::TracePhase tp("chaitinSelect", &timers[_t_chaitinSelect]);
1463
1464 uint spill_reg = LRG::SPILL_REG;
1465 _max_reg = OptoReg::Name(0); // Past max register used
1466 while( _simplified ) {
1467 // Pull next LRG from the simplified list - in reverse order of removal
1468 uint lidx = _simplified;
1469 LRG *lrg = &lrgs(lidx);
1470 _simplified = lrg->_next;
1471
1472 #ifndef PRODUCT
1473 if (trace_spilling()) {
1474 ttyLocker ttyl;
1475 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(),
1476 lrg->degrees_of_freedom());
1477 lrg->dump();
1478 }
1479 #endif
1480
1481 // Re-insert into the IFG
1482 _ifg->re_insert(lidx);
1483 if( !lrg->alive() ) continue;
1484 // capture allstackedness flag before mask is hacked
1485 const int is_allstack = lrg->mask().is_AllStack();
1486
1487 // Yeah, yeah, yeah, I know, I know. I can refactor this
1488 // to avoid the GOTO, although the refactored code will not
1489 // be much clearer. We arrive here IFF we have a stack-based
1490 // live range that cannot color in the current chunk, and it
1491 // has to move into the next free stack chunk.
1492 int chunk = 0; // Current chunk is first chunk
1493 retry_next_chunk:
1494
1495 // Remove neighbor colors
1496 IndexSet *s = _ifg->neighbors(lidx);
1497 debug_only(RegMask orig_mask = lrg->mask();)
1498
1499 if (!s->is_empty()) {
1500 IndexSetIterator elements(s);
1501 uint neighbor;
1502 while ((neighbor = elements.next()) != 0) {
1503 // Note that neighbor might be a spill_reg. In this case, exclusion
1504 // of its color will be a no-op, since the spill_reg chunk is in outer
1505 // space. Also, if neighbor is in a different chunk, this exclusion
1506 // will be a no-op. (Later on, if lrg runs out of possible colors in
1507 // its chunk, a new chunk of color may be tried, in which case
1508 // examination of neighbors is started again, at retry_next_chunk.)
1509 LRG &nlrg = lrgs(neighbor);
1510 OptoReg::Name nreg = nlrg.reg();
1511 // Only subtract masks in the same chunk
1512 if (nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE) {
1513 #ifndef PRODUCT
1514 uint size = lrg->mask().Size();
1515 RegMask rm = lrg->mask();
1516 #endif
1517 lrg->SUBTRACT(nlrg.mask());
1518 #ifndef PRODUCT
1519 if (trace_spilling() && lrg->mask().Size() != size) {
1520 ttyLocker ttyl;
1521 tty->print("L%d ", lidx);
1522 rm.dump();
1523 tty->print(" intersected L%d ", neighbor);
1524 nlrg.mask().dump();
1525 tty->print(" removed ");
1526 rm.SUBTRACT(lrg->mask());
1527 rm.dump();
1528 tty->print(" leaving ");
1529 lrg->mask().dump();
1530 tty->cr();
1531 }
1532 #endif
1533 }
1534 }
1535 }
1536 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness");
1537 // Aligned pairs need aligned masks
1538 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1539 if (lrg->num_regs() > 1 && !lrg->_fat_proj) {
1540 lrg->clear_to_sets();
1541 }
1542
1543 // Check if a color is available and if so pick the color
1544 OptoReg::Name reg = choose_color( *lrg, chunk );
1545
1546 //---------------
1547 // If we fail to color and the AllStack flag is set, trigger
1548 // a chunk-rollover event
1549 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) {
1550 // Bump register mask up to next stack chunk
1551 chunk += RegMask::CHUNK_SIZE;
1552 lrg->Set_All();
1553 goto retry_next_chunk;
1554 }
1555
1556 //---------------
1557 // Did we get a color?
1558 else if( OptoReg::is_valid(reg)) {
1559 #ifndef PRODUCT
1560 RegMask avail_rm = lrg->mask();
1561 #endif
1562
1563 // Record selected register
1564 lrg->set_reg(reg);
1565
1566 if( reg >= _max_reg ) // Compute max register limit
1567 _max_reg = OptoReg::add(reg,1);
1568 // Fold reg back into normal space
1569 reg = OptoReg::add(reg,-chunk);
1570
1571 // If the live range is not bound, then we actually had some choices
1572 // to make. In this case, the mask has more bits in it than the colors
1573 // chosen. Restrict the mask to just what was picked.
1574 int n_regs = lrg->num_regs();
1575 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity");
1576 if (n_regs == 1 || !lrg->_fat_proj) {
1577 if (Matcher::supports_scalable_vector()) {
1578 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecA, "sanity");
1579 } else {
1580 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity");
1581 }
1582 lrg->Clear(); // Clear the mask
1583 lrg->Insert(reg); // Set regmask to match selected reg
1584 // For vectors and pairs, also insert the low bit of the pair
1585 // We always choose the high bit, then mask the low bits by register size
1586 if (lrg->is_scalable() && OptoReg::is_stack(lrg->reg())) { // stack
1587 n_regs = lrg->scalable_reg_slots();
1588 }
1589 for (int i = 1; i < n_regs; i++) {
1590 lrg->Insert(OptoReg::add(reg,-i));
1591 }
1592 lrg->set_mask_size(n_regs);
1593 } else { // Else fatproj
1594 // mask must be equal to fatproj bits, by definition
1595 }
1596 #ifndef PRODUCT
1597 if (trace_spilling()) {
1598 ttyLocker ttyl;
1599 tty->print("L%d selected ", lidx);
1600 lrg->mask().dump();
1601 tty->print(" from ");
1602 avail_rm.dump();
1603 tty->cr();
1604 }
1605 #endif
1606 // Note that reg is the highest-numbered register in the newly-bound mask.
1607 } // end color available case
1608
1609 //---------------
1610 // Live range is live and no colors available
1611 else {
1612 assert( lrg->alive(), "" );
1613 assert( !lrg->_fat_proj || lrg->is_multidef() ||
1614 lrg->_def->outcnt() > 0, "fat_proj cannot spill");
1615 assert( !orig_mask.is_AllStack(), "All Stack does not spill" );
1616
1617 // Assign the special spillreg register
1618 lrg->set_reg(OptoReg::Name(spill_reg++));
1619 // Do not empty the regmask; leave mask_size lying around
1620 // for use during Spilling
1621 #ifndef PRODUCT
1622 if( trace_spilling() ) {
1623 ttyLocker ttyl;
1624 tty->print("L%d spilling with neighbors: ", lidx);
1625 s->dump();
1626 debug_only(tty->print(" original mask: "));
1627 debug_only(orig_mask.dump());
1628 dump_lrg(lidx);
1629 }
1630 #endif
1631 } // end spill case
1632
1633 }
1634
1635 return spill_reg-LRG::SPILL_REG; // Return number of spills
1636 }
1637
1638 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
set_was_spilled(Node * n)1639 void PhaseChaitin::set_was_spilled( Node *n ) {
1640 if( _spilled_once.test_set(n->_idx) )
1641 _spilled_twice.set(n->_idx);
1642 }
1643
1644 // Convert Ideal spill instructions into proper FramePtr + offset Loads and
1645 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are.
fixup_spills()1646 void PhaseChaitin::fixup_spills() {
1647 // This function does only cisc spill work.
1648 if( !UseCISCSpill ) return;
1649
1650 Compile::TracePhase tp("fixupSpills", &timers[_t_fixupSpills]);
1651
1652 // Grab the Frame Pointer
1653 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr);
1654
1655 // For all blocks
1656 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1657 Block* block = _cfg.get_block(i);
1658
1659 // For all instructions in block
1660 uint last_inst = block->end_idx();
1661 for (uint j = 1; j <= last_inst; j++) {
1662 Node* n = block->get_node(j);
1663
1664 // Dead instruction???
1665 assert( n->outcnt() != 0 ||// Nothing dead after post alloc
1666 C->top() == n || // Or the random TOP node
1667 n->is_Proj(), // Or a fat-proj kill node
1668 "No dead instructions after post-alloc" );
1669
1670 int inp = n->cisc_operand();
1671 if( inp != AdlcVMDeps::Not_cisc_spillable ) {
1672 // Convert operand number to edge index number
1673 MachNode *mach = n->as_Mach();
1674 inp = mach->operand_index(inp);
1675 Node *src = n->in(inp); // Value to load or store
1676 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src));
1677 OptoReg::Name src_reg = lrg_cisc.reg();
1678 // Doubles record the HIGH register of an adjacent pair.
1679 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs());
1680 if( OptoReg::is_stack(src_reg) ) { // If input is on stack
1681 // This is a CISC Spill, get stack offset and construct new node
1682 #ifndef PRODUCT
1683 if( TraceCISCSpill ) {
1684 tty->print(" reg-instr: ");
1685 n->dump();
1686 }
1687 #endif
1688 int stk_offset = reg2offset(src_reg);
1689 // Bailout if we might exceed node limit when spilling this instruction
1690 C->check_node_count(0, "out of nodes fixing spills");
1691 if (C->failing()) return;
1692 // Transform node
1693 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach();
1694 cisc->set_req(inp,fp); // Base register is frame pointer
1695 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) {
1696 assert( cisc->oper_input_base() == 2, "Only adding one edge");
1697 cisc->ins_req(1,src); // Requires a memory edge
1698 }
1699 block->map_node(cisc, j); // Insert into basic block
1700 n->subsume_by(cisc, C); // Correct graph
1701 //
1702 ++_used_cisc_instructions;
1703 #ifndef PRODUCT
1704 if( TraceCISCSpill ) {
1705 tty->print(" cisc-instr: ");
1706 cisc->dump();
1707 }
1708 #endif
1709 } else {
1710 #ifndef PRODUCT
1711 if( TraceCISCSpill ) {
1712 tty->print(" using reg-instr: ");
1713 n->dump();
1714 }
1715 #endif
1716 ++_unused_cisc_instructions; // input can be on stack
1717 }
1718 }
1719
1720 } // End of for all instructions
1721
1722 } // End of for all blocks
1723 }
1724
1725 // Helper to stretch above; recursively discover the base Node for a
1726 // given derived Node. Easy for AddP-related machine nodes, but needs
1727 // to be recursive for derived Phis.
find_base_for_derived(Node ** derived_base_map,Node * derived,uint & maxlrg)1728 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) {
1729 // See if already computed; if so return it
1730 if( derived_base_map[derived->_idx] )
1731 return derived_base_map[derived->_idx];
1732
1733 // See if this happens to be a base.
1734 // NOTE: we use TypePtr instead of TypeOopPtr because we can have
1735 // pointers derived from NULL! These are always along paths that
1736 // can't happen at run-time but the optimizer cannot deduce it so
1737 // we have to handle it gracefully.
1738 assert(!derived->bottom_type()->isa_narrowoop() ||
1739 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity");
1740 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1741 // If its an OOP with a non-zero offset, then it is derived.
1742 if( tj == NULL || tj->_offset == 0 ) {
1743 derived_base_map[derived->_idx] = derived;
1744 return derived;
1745 }
1746 // Derived is NULL+offset? Base is NULL!
1747 if( derived->is_Con() ) {
1748 Node *base = _matcher.mach_null();
1749 assert(base != NULL, "sanity");
1750 if (base->in(0) == NULL) {
1751 // Initialize it once and make it shared:
1752 // set control to _root and place it into Start block
1753 // (where top() node is placed).
1754 base->init_req(0, _cfg.get_root_node());
1755 Block *startb = _cfg.get_block_for_node(C->top());
1756 uint node_pos = startb->find_node(C->top());
1757 startb->insert_node(base, node_pos);
1758 _cfg.map_node_to_block(base, startb);
1759 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet");
1760
1761 // The loadConP0 might have projection nodes depending on architecture
1762 // Add the projection nodes to the CFG
1763 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) {
1764 Node* use = base->fast_out(i);
1765 if (use->is_MachProj()) {
1766 startb->insert_node(use, ++node_pos);
1767 _cfg.map_node_to_block(use, startb);
1768 new_lrg(use, maxlrg++);
1769 }
1770 }
1771 }
1772 if (_lrg_map.live_range_id(base) == 0) {
1773 new_lrg(base, maxlrg++);
1774 }
1775 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");
1776 derived_base_map[derived->_idx] = base;
1777 return base;
1778 }
1779
1780 // Check for AddP-related opcodes
1781 if (!derived->is_Phi()) {
1782 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name());
1783 Node *base = derived->in(AddPNode::Base);
1784 derived_base_map[derived->_idx] = base;
1785 return base;
1786 }
1787
1788 // Recursively find bases for Phis.
1789 // First check to see if we can avoid a base Phi here.
1790 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg);
1791 uint i;
1792 for( i = 2; i < derived->req(); i++ )
1793 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg))
1794 break;
1795 // Went to the end without finding any different bases?
1796 if( i == derived->req() ) { // No need for a base Phi here
1797 derived_base_map[derived->_idx] = base;
1798 return base;
1799 }
1800
1801 // Now we see we need a base-Phi here to merge the bases
1802 const Type *t = base->bottom_type();
1803 base = new PhiNode( derived->in(0), t );
1804 for( i = 1; i < derived->req(); i++ ) {
1805 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg));
1806 t = t->meet(base->in(i)->bottom_type());
1807 }
1808 base->as_Phi()->set_type(t);
1809
1810 // Search the current block for an existing base-Phi
1811 Block *b = _cfg.get_block_for_node(derived);
1812 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi
1813 Node *phi = b->get_node(i);
1814 if( !phi->is_Phi() ) { // Found end of Phis with no match?
1815 b->insert_node(base, i); // Must insert created Phi here as base
1816 _cfg.map_node_to_block(base, b);
1817 new_lrg(base,maxlrg++);
1818 break;
1819 }
1820 // See if Phi matches.
1821 uint j;
1822 for( j = 1; j < base->req(); j++ )
1823 if( phi->in(j) != base->in(j) &&
1824 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs
1825 break;
1826 if( j == base->req() ) { // All inputs match?
1827 base = phi; // Then use existing 'phi' and drop 'base'
1828 break;
1829 }
1830 }
1831
1832
1833 // Cache info for later passes
1834 derived_base_map[derived->_idx] = base;
1835 return base;
1836 }
1837
1838 // At each Safepoint, insert extra debug edges for each pair of derived value/
1839 // base pointer that is live across the Safepoint for oopmap building. The
1840 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the
1841 // required edge set.
stretch_base_pointer_live_ranges(ResourceArea * a)1842 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) {
1843 int must_recompute_live = false;
1844 uint maxlrg = _lrg_map.max_lrg_id();
1845 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique());
1846 memset( derived_base_map, 0, sizeof(Node*)*C->unique() );
1847
1848 // For all blocks in RPO do...
1849 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
1850 Block* block = _cfg.get_block(i);
1851 // Note use of deep-copy constructor. I cannot hammer the original
1852 // liveout bits, because they are needed by the following coalesce pass.
1853 IndexSet liveout(_live->live(block));
1854
1855 for (uint j = block->end_idx() + 1; j > 1; j--) {
1856 Node* n = block->get_node(j - 1);
1857
1858 // Pre-split compares of loop-phis. Loop-phis form a cycle we would
1859 // like to see in the same register. Compare uses the loop-phi and so
1860 // extends its live range BUT cannot be part of the cycle. If this
1861 // extended live range overlaps with the update of the loop-phi value
1862 // we need both alive at the same time -- which requires at least 1
1863 // copy. But because Intel has only 2-address registers we end up with
1864 // at least 2 copies, one before the loop-phi update instruction and
1865 // one after. Instead we split the input to the compare just after the
1866 // phi.
1867 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) {
1868 Node *phi = n->in(1);
1869 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) {
1870 Block *phi_block = _cfg.get_block_for_node(phi);
1871 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) {
1872 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI];
1873 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask);
1874 insert_proj( phi_block, 1, spill, maxlrg++ );
1875 n->set_req(1,spill);
1876 must_recompute_live = true;
1877 }
1878 }
1879 }
1880
1881 // Get value being defined
1882 uint lidx = _lrg_map.live_range_id(n);
1883 // Ignore the occasional brand-new live range
1884 if (lidx && lidx < _lrg_map.max_lrg_id()) {
1885 // Remove from live-out set
1886 liveout.remove(lidx);
1887
1888 // Copies do not define a new value and so do not interfere.
1889 // Remove the copies source from the liveout set before interfering.
1890 uint idx = n->is_Copy();
1891 if (idx) {
1892 liveout.remove(_lrg_map.live_range_id(n->in(idx)));
1893 }
1894 }
1895
1896 // Found a safepoint?
1897 JVMState *jvms = n->jvms();
1898 if (jvms && !liveout.is_empty()) {
1899 // Now scan for a live derived pointer
1900 IndexSetIterator elements(&liveout);
1901 uint neighbor;
1902 while ((neighbor = elements.next()) != 0) {
1903 // Find reaching DEF for base and derived values
1904 // This works because we are still in SSA during this call.
1905 Node *derived = lrgs(neighbor)._def;
1906 const TypePtr *tj = derived->bottom_type()->isa_ptr();
1907 assert(!derived->bottom_type()->isa_narrowoop() ||
1908 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity");
1909 // If its an OOP with a non-zero offset, then it is derived.
1910 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) {
1911 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg);
1912 assert(base->_idx < _lrg_map.size(), "");
1913 // Add reaching DEFs of derived pointer and base pointer as a
1914 // pair of inputs
1915 n->add_req(derived);
1916 n->add_req(base);
1917
1918 // See if the base pointer is already live to this point.
1919 // Since I'm working on the SSA form, live-ness amounts to
1920 // reaching def's. So if I find the base's live range then
1921 // I know the base's def reaches here.
1922 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or
1923 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND
1924 (_lrg_map.live_range_id(base) > 0) && // not a constant
1925 _cfg.get_block_for_node(base) != block) { // base not def'd in blk)
1926 // Base pointer is not currently live. Since I stretched
1927 // the base pointer to here and it crosses basic-block
1928 // boundaries, the global live info is now incorrect.
1929 // Recompute live.
1930 must_recompute_live = true;
1931 } // End of if base pointer is not live to debug info
1932 }
1933 } // End of scan all live data for derived ptrs crossing GC point
1934 } // End of if found a GC point
1935
1936 // Make all inputs live
1937 if (!n->is_Phi()) { // Phi function uses come from prior block
1938 for (uint k = 1; k < n->req(); k++) {
1939 uint lidx = _lrg_map.live_range_id(n->in(k));
1940 if (lidx < _lrg_map.max_lrg_id()) {
1941 liveout.insert(lidx);
1942 }
1943 }
1944 }
1945
1946 } // End of forall instructions in block
1947 liveout.clear(); // Free the memory used by liveout.
1948
1949 } // End of forall blocks
1950 _lrg_map.set_max_lrg_id(maxlrg);
1951
1952 // If I created a new live range I need to recompute live
1953 if (maxlrg != _ifg->_maxlrg) {
1954 must_recompute_live = true;
1955 }
1956
1957 return must_recompute_live != 0;
1958 }
1959
1960 // Extend the node to LRG mapping
1961
add_reference(const Node * node,const Node * old_node)1962 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) {
1963 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node));
1964 }
1965
1966 #ifndef PRODUCT
dump(const Node * n) const1967 void PhaseChaitin::dump(const Node* n) const {
1968 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0;
1969 tty->print("L%d",r);
1970 if (r && n->Opcode() != Op_Phi) {
1971 if( _node_regs ) { // Got a post-allocation copy of allocation?
1972 tty->print("[");
1973 OptoReg::Name second = get_reg_second(n);
1974 if( OptoReg::is_valid(second) ) {
1975 if( OptoReg::is_reg(second) )
1976 tty->print("%s:",Matcher::regName[second]);
1977 else
1978 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second));
1979 }
1980 OptoReg::Name first = get_reg_first(n);
1981 if( OptoReg::is_reg(first) )
1982 tty->print("%s]",Matcher::regName[first]);
1983 else
1984 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first));
1985 } else
1986 n->out_RegMask().dump();
1987 }
1988 tty->print("/N%d\t",n->_idx);
1989 tty->print("%s === ", n->Name());
1990 uint k;
1991 for (k = 0; k < n->req(); k++) {
1992 Node *m = n->in(k);
1993 if (!m) {
1994 tty->print("_ ");
1995 }
1996 else {
1997 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
1998 tty->print("L%d",r);
1999 // Data MultiNode's can have projections with no real registers.
2000 // Don't die while dumping them.
2001 int op = n->Opcode();
2002 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) {
2003 if( _node_regs ) {
2004 tty->print("[");
2005 OptoReg::Name second = get_reg_second(n->in(k));
2006 if( OptoReg::is_valid(second) ) {
2007 if( OptoReg::is_reg(second) )
2008 tty->print("%s:",Matcher::regName[second]);
2009 else
2010 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer),
2011 reg2offset_unchecked(second));
2012 }
2013 OptoReg::Name first = get_reg_first(n->in(k));
2014 if( OptoReg::is_reg(first) )
2015 tty->print("%s]",Matcher::regName[first]);
2016 else
2017 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer),
2018 reg2offset_unchecked(first));
2019 } else
2020 n->in_RegMask(k).dump();
2021 }
2022 tty->print("/N%d ",m->_idx);
2023 }
2024 }
2025 if( k < n->len() && n->in(k) ) tty->print("| ");
2026 for( ; k < n->len(); k++ ) {
2027 Node *m = n->in(k);
2028 if(!m) {
2029 break;
2030 }
2031 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0;
2032 tty->print("L%d",r);
2033 tty->print("/N%d ",m->_idx);
2034 }
2035 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty);
2036 else n->dump_spec(tty);
2037 if( _spilled_once.test(n->_idx ) ) {
2038 tty->print(" Spill_1");
2039 if( _spilled_twice.test(n->_idx ) )
2040 tty->print(" Spill_2");
2041 }
2042 tty->print("\n");
2043 }
2044
dump(const Block * b) const2045 void PhaseChaitin::dump(const Block* b) const {
2046 b->dump_head(&_cfg);
2047
2048 // For all instructions
2049 for( uint j = 0; j < b->number_of_nodes(); j++ )
2050 dump(b->get_node(j));
2051 // Print live-out info at end of block
2052 if( _live ) {
2053 tty->print("Liveout: ");
2054 IndexSet *live = _live->live(b);
2055 IndexSetIterator elements(live);
2056 tty->print("{");
2057 uint i;
2058 while ((i = elements.next()) != 0) {
2059 tty->print("L%d ", _lrg_map.find_const(i));
2060 }
2061 tty->print_cr("}");
2062 }
2063 tty->print("\n");
2064 }
2065
dump() const2066 void PhaseChaitin::dump() const {
2067 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n",
2068 _matcher._new_SP, _framesize );
2069
2070 // For all blocks
2071 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2072 dump(_cfg.get_block(i));
2073 }
2074 // End of per-block dump
2075 tty->print("\n");
2076
2077 if (!_ifg) {
2078 tty->print("(No IFG.)\n");
2079 return;
2080 }
2081
2082 // Dump LRG array
2083 tty->print("--- Live RanGe Array ---\n");
2084 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) {
2085 tty->print("L%d: ",i2);
2086 if (i2 < _ifg->_maxlrg) {
2087 lrgs(i2).dump();
2088 }
2089 else {
2090 tty->print_cr("new LRG");
2091 }
2092 }
2093 tty->cr();
2094
2095 // Dump lo-degree list
2096 tty->print("Lo degree: ");
2097 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next )
2098 tty->print("L%d ",i3);
2099 tty->cr();
2100
2101 // Dump lo-stk-degree list
2102 tty->print("Lo stk degree: ");
2103 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next )
2104 tty->print("L%d ",i4);
2105 tty->cr();
2106
2107 // Dump lo-degree list
2108 tty->print("Hi degree: ");
2109 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next )
2110 tty->print("L%d ",i5);
2111 tty->cr();
2112 }
2113
dump_degree_lists() const2114 void PhaseChaitin::dump_degree_lists() const {
2115 // Dump lo-degree list
2116 tty->print("Lo degree: ");
2117 for( uint i = _lo_degree; i; i = lrgs(i)._next )
2118 tty->print("L%d ",i);
2119 tty->cr();
2120
2121 // Dump lo-stk-degree list
2122 tty->print("Lo stk degree: ");
2123 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next )
2124 tty->print("L%d ",i2);
2125 tty->cr();
2126
2127 // Dump lo-degree list
2128 tty->print("Hi degree: ");
2129 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next )
2130 tty->print("L%d ",i3);
2131 tty->cr();
2132 }
2133
dump_simplified() const2134 void PhaseChaitin::dump_simplified() const {
2135 tty->print("Simplified: ");
2136 for( uint i = _simplified; i; i = lrgs(i)._next )
2137 tty->print("L%d ",i);
2138 tty->cr();
2139 }
2140
print_reg(OptoReg::Name reg,const PhaseChaitin * pc,char * buf)2141 static char *print_reg(OptoReg::Name reg, const PhaseChaitin* pc, char* buf) {
2142 if ((int)reg < 0)
2143 sprintf(buf, "<OptoReg::%d>", (int)reg);
2144 else if (OptoReg::is_reg(reg))
2145 strcpy(buf, Matcher::regName[reg]);
2146 else
2147 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer),
2148 pc->reg2offset(reg));
2149 return buf+strlen(buf);
2150 }
2151
2152 // Dump a register name into a buffer. Be intelligent if we get called
2153 // before allocation is complete.
dump_register(const Node * n,char * buf) const2154 char *PhaseChaitin::dump_register(const Node* n, char* buf) const {
2155 if( _node_regs ) {
2156 // Post allocation, use direct mappings, no LRG info available
2157 print_reg( get_reg_first(n), this, buf );
2158 } else {
2159 uint lidx = _lrg_map.find_const(n); // Grab LRG number
2160 if( !_ifg ) {
2161 sprintf(buf,"L%d",lidx); // No register binding yet
2162 } else if( !lidx ) { // Special, not allocated value
2163 strcpy(buf,"Special");
2164 } else {
2165 if (lrgs(lidx)._is_vector) {
2166 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs()))
2167 print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register
2168 else
2169 sprintf(buf,"L%d",lidx); // No register binding yet
2170 } else if( (lrgs(lidx).num_regs() == 1)
2171 ? lrgs(lidx).mask().is_bound1()
2172 : lrgs(lidx).mask().is_bound_pair() ) {
2173 // Hah! We have a bound machine register
2174 print_reg( lrgs(lidx).reg(), this, buf );
2175 } else {
2176 sprintf(buf,"L%d",lidx); // No register binding yet
2177 }
2178 }
2179 }
2180 return buf+strlen(buf);
2181 }
2182
dump_for_spill_split_recycle() const2183 void PhaseChaitin::dump_for_spill_split_recycle() const {
2184 if( WizardMode && (PrintCompilation || PrintOpto) ) {
2185 // Display which live ranges need to be split and the allocator's state
2186 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt);
2187 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) {
2188 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
2189 tty->print("L%d: ", bidx);
2190 lrgs(bidx).dump();
2191 }
2192 }
2193 tty->cr();
2194 dump();
2195 }
2196 }
2197
dump_frame() const2198 void PhaseChaitin::dump_frame() const {
2199 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer);
2200 const TypeTuple *domain = C->tf()->domain();
2201 const int argcnt = domain->cnt() - TypeFunc::Parms;
2202
2203 // Incoming arguments in registers dump
2204 for( int k = 0; k < argcnt; k++ ) {
2205 OptoReg::Name parmreg = _matcher._parm_regs[k].first();
2206 if( OptoReg::is_reg(parmreg)) {
2207 const char *reg_name = OptoReg::regname(parmreg);
2208 tty->print("#r%3.3d %s", parmreg, reg_name);
2209 parmreg = _matcher._parm_regs[k].second();
2210 if( OptoReg::is_reg(parmreg)) {
2211 tty->print(":%s", OptoReg::regname(parmreg));
2212 }
2213 tty->print(" : parm %d: ", k);
2214 domain->field_at(k + TypeFunc::Parms)->dump();
2215 tty->cr();
2216 }
2217 }
2218
2219 // Check for un-owned padding above incoming args
2220 OptoReg::Name reg = _matcher._new_SP;
2221 if( reg > _matcher._in_arg_limit ) {
2222 reg = OptoReg::add(reg, -1);
2223 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg));
2224 }
2225
2226 // Incoming argument area dump
2227 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots());
2228 while( reg > begin_in_arg ) {
2229 reg = OptoReg::add(reg, -1);
2230 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2231 int j;
2232 for( j = 0; j < argcnt; j++) {
2233 if( _matcher._parm_regs[j].first() == reg ||
2234 _matcher._parm_regs[j].second() == reg ) {
2235 tty->print("parm %d: ",j);
2236 domain->field_at(j + TypeFunc::Parms)->dump();
2237 tty->cr();
2238 break;
2239 }
2240 }
2241 if( j >= argcnt )
2242 tty->print_cr("HOLE, owned by SELF");
2243 }
2244
2245 // Old outgoing preserve area
2246 while( reg > _matcher._old_SP ) {
2247 reg = OptoReg::add(reg, -1);
2248 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg));
2249 }
2250
2251 // Old SP
2252 tty->print_cr("# -- Old %s -- Framesize: %d --",fp,
2253 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize);
2254
2255 // Preserve area dump
2256 int fixed_slots = C->fixed_slots();
2257 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots());
2258 OptoReg::Name return_addr = _matcher.return_addr();
2259
2260 reg = OptoReg::add(reg, -1);
2261 while (OptoReg::is_stack(reg)) {
2262 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
2263 if (return_addr == reg) {
2264 tty->print_cr("return address");
2265 } else if (reg >= begin_in_preserve) {
2266 // Preserved slots are present on x86
2267 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word))
2268 tty->print_cr("saved fp register");
2269 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) &&
2270 VerifyStackAtCalls)
2271 tty->print_cr("0xBADB100D +VerifyStackAtCalls");
2272 else
2273 tty->print_cr("in_preserve");
2274 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) {
2275 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg));
2276 } else {
2277 tty->print_cr("pad2, stack alignment");
2278 }
2279 reg = OptoReg::add(reg, -1);
2280 }
2281
2282 // Spill area dump
2283 reg = OptoReg::add(_matcher._new_SP, _framesize );
2284 while( reg > _matcher._out_arg_limit ) {
2285 reg = OptoReg::add(reg, -1);
2286 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg));
2287 }
2288
2289 // Outgoing argument area dump
2290 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) {
2291 reg = OptoReg::add(reg, -1);
2292 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg));
2293 }
2294
2295 // Outgoing new preserve area
2296 while( reg > _matcher._new_SP ) {
2297 reg = OptoReg::add(reg, -1);
2298 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg));
2299 }
2300 tty->print_cr("#");
2301 }
2302
dump_bb(uint pre_order) const2303 void PhaseChaitin::dump_bb(uint pre_order) const {
2304 tty->print_cr("---dump of B%d---",pre_order);
2305 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2306 Block* block = _cfg.get_block(i);
2307 if (block->_pre_order == pre_order) {
2308 dump(block);
2309 }
2310 }
2311 }
2312
dump_lrg(uint lidx,bool defs_only) const2313 void PhaseChaitin::dump_lrg(uint lidx, bool defs_only) const {
2314 tty->print_cr("---dump of L%d---",lidx);
2315
2316 if (_ifg) {
2317 if (lidx >= _lrg_map.max_lrg_id()) {
2318 tty->print("Attempt to print live range index beyond max live range.\n");
2319 return;
2320 }
2321 tty->print("L%d: ",lidx);
2322 if (lidx < _ifg->_maxlrg) {
2323 lrgs(lidx).dump();
2324 } else {
2325 tty->print_cr("new LRG");
2326 }
2327 }
2328 if( _ifg && lidx < _ifg->_maxlrg) {
2329 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx));
2330 _ifg->neighbors(lidx)->dump();
2331 tty->cr();
2332 }
2333 // For all blocks
2334 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2335 Block* block = _cfg.get_block(i);
2336 int dump_once = 0;
2337
2338 // For all instructions
2339 for( uint j = 0; j < block->number_of_nodes(); j++ ) {
2340 Node *n = block->get_node(j);
2341 if (_lrg_map.find_const(n) == lidx) {
2342 if (!dump_once++) {
2343 tty->cr();
2344 block->dump_head(&_cfg);
2345 }
2346 dump(n);
2347 continue;
2348 }
2349 if (!defs_only) {
2350 uint cnt = n->req();
2351 for( uint k = 1; k < cnt; k++ ) {
2352 Node *m = n->in(k);
2353 if (!m) {
2354 continue; // be robust in the dumper
2355 }
2356 if (_lrg_map.find_const(m) == lidx) {
2357 if (!dump_once++) {
2358 tty->cr();
2359 block->dump_head(&_cfg);
2360 }
2361 dump(n);
2362 }
2363 }
2364 }
2365 }
2366 } // End of per-block dump
2367 tty->cr();
2368 }
2369 #endif // not PRODUCT
2370
2371 #ifdef ASSERT
2372 // Verify that base pointers and derived pointers are still sane.
verify_base_ptrs(ResourceArea * a) const2373 void PhaseChaitin::verify_base_ptrs(ResourceArea* a) const {
2374 Unique_Node_List worklist(a);
2375 for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
2376 Block* block = _cfg.get_block(i);
2377 for (uint j = block->end_idx() + 1; j > 1; j--) {
2378 Node* n = block->get_node(j-1);
2379 if (n->is_Phi()) {
2380 break;
2381 }
2382 // Found a safepoint?
2383 if (n->is_MachSafePoint()) {
2384 MachSafePointNode* sfpt = n->as_MachSafePoint();
2385 JVMState* jvms = sfpt->jvms();
2386 if (jvms != NULL) {
2387 // Now scan for a live derived pointer
2388 if (jvms->oopoff() < sfpt->req()) {
2389 // Check each derived/base pair
2390 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
2391 Node* check = sfpt->in(idx);
2392 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
2393 // search upwards through spills and spill phis for AddP
2394 worklist.clear();
2395 worklist.push(check);
2396 uint k = 0;
2397 while (k < worklist.size()) {
2398 check = worklist.at(k);
2399 assert(check, "Bad base or derived pointer");
2400 // See PhaseChaitin::find_base_for_derived() for all cases.
2401 int isc = check->is_Copy();
2402 if (isc) {
2403 worklist.push(check->in(isc));
2404 } else if (check->is_Phi()) {
2405 for (uint m = 1; m < check->req(); m++) {
2406 worklist.push(check->in(m));
2407 }
2408 } else if (check->is_Con()) {
2409 if (is_derived && check->bottom_type()->is_ptr()->_offset != 0) {
2410 // Derived is NULL+non-zero offset, base must be NULL.
2411 assert(check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad derived pointer");
2412 } else {
2413 assert(check->bottom_type()->is_ptr()->_offset == 0, "Bad base pointer");
2414 // Base either ConP(NULL) or loadConP
2415 if (check->is_Mach()) {
2416 assert(check->as_Mach()->ideal_Opcode() == Op_ConP, "Bad base pointer");
2417 } else {
2418 assert(check->Opcode() == Op_ConP &&
2419 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad base pointer");
2420 }
2421 }
2422 } else if (check->bottom_type()->is_ptr()->_offset == 0) {
2423 if (check->is_Proj() || (check->is_Mach() &&
2424 (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
2425 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
2426 check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
2427 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
2428 #ifdef _LP64
2429 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP) ||
2430 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN) ||
2431 (UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass) ||
2432 #endif // _LP64
2433 check->as_Mach()->ideal_Opcode() == Op_LoadP ||
2434 check->as_Mach()->ideal_Opcode() == Op_LoadKlass))) {
2435 // Valid nodes
2436 } else {
2437 check->dump();
2438 assert(false, "Bad base or derived pointer");
2439 }
2440 } else {
2441 assert(is_derived, "Bad base pointer");
2442 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP, "Bad derived pointer");
2443 }
2444 k++;
2445 assert(k < 100000, "Derived pointer checking in infinite loop");
2446 } // End while
2447 }
2448 } // End of check for derived pointers
2449 } // End of Kcheck for debug info
2450 } // End of if found a safepoint
2451 } // End of forall instructions in block
2452 } // End of forall blocks
2453 }
2454
2455 // Verify that graphs and base pointers are still sane.
verify(ResourceArea * a,bool verify_ifg) const2456 void PhaseChaitin::verify(ResourceArea* a, bool verify_ifg) const {
2457 if (VerifyRegisterAllocator) {
2458 _cfg.verify();
2459 verify_base_ptrs(a);
2460 if (verify_ifg) {
2461 _ifg->verify(this);
2462 }
2463 }
2464 }
2465 #endif // ASSERT
2466
2467 int PhaseChaitin::_final_loads = 0;
2468 int PhaseChaitin::_final_stores = 0;
2469 int PhaseChaitin::_final_memoves= 0;
2470 int PhaseChaitin::_final_copies = 0;
2471 double PhaseChaitin::_final_load_cost = 0;
2472 double PhaseChaitin::_final_store_cost = 0;
2473 double PhaseChaitin::_final_memove_cost= 0;
2474 double PhaseChaitin::_final_copy_cost = 0;
2475 int PhaseChaitin::_conserv_coalesce = 0;
2476 int PhaseChaitin::_conserv_coalesce_pair = 0;
2477 int PhaseChaitin::_conserv_coalesce_trie = 0;
2478 int PhaseChaitin::_conserv_coalesce_quad = 0;
2479 int PhaseChaitin::_post_alloc = 0;
2480 int PhaseChaitin::_lost_opp_pp_coalesce = 0;
2481 int PhaseChaitin::_lost_opp_cflow_coalesce = 0;
2482 int PhaseChaitin::_used_cisc_instructions = 0;
2483 int PhaseChaitin::_unused_cisc_instructions = 0;
2484 int PhaseChaitin::_allocator_attempts = 0;
2485 int PhaseChaitin::_allocator_successes = 0;
2486
2487 #ifndef PRODUCT
2488 uint PhaseChaitin::_high_pressure = 0;
2489 uint PhaseChaitin::_low_pressure = 0;
2490
print_chaitin_statistics()2491 void PhaseChaitin::print_chaitin_statistics() {
2492 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);
2493 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);
2494 tty->print_cr("Adjusted spill cost = %7.0f.",
2495 _final_load_cost*4.0 + _final_store_cost * 2.0 +
2496 _final_copy_cost*1.0 + _final_memove_cost*12.0);
2497 tty->print("Conservatively coalesced %d copies, %d pairs",
2498 _conserv_coalesce, _conserv_coalesce_pair);
2499 if( _conserv_coalesce_trie || _conserv_coalesce_quad )
2500 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad);
2501 tty->print_cr(", %d post alloc.", _post_alloc);
2502 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce )
2503 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.",
2504 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce );
2505 if( _used_cisc_instructions || _unused_cisc_instructions )
2506 tty->print_cr("Used cisc instruction %d, remained in register %d",
2507 _used_cisc_instructions, _unused_cisc_instructions);
2508 if( _allocator_successes != 0 )
2509 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes);
2510 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure);
2511 }
2512 #endif // not PRODUCT
2513