1 | /* |
2 | * Copyright (c) 2000, 2019, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "compiler/compileLog.hpp" |
27 | #include "compiler/oopMap.hpp" |
28 | #include "memory/allocation.inline.hpp" |
29 | #include "memory/resourceArea.hpp" |
30 | #include "opto/addnode.hpp" |
31 | #include "opto/block.hpp" |
32 | #include "opto/callnode.hpp" |
33 | #include "opto/cfgnode.hpp" |
34 | #include "opto/chaitin.hpp" |
35 | #include "opto/coalesce.hpp" |
36 | #include "opto/connode.hpp" |
37 | #include "opto/idealGraphPrinter.hpp" |
38 | #include "opto/indexSet.hpp" |
39 | #include "opto/machnode.hpp" |
40 | #include "opto/memnode.hpp" |
41 | #include "opto/movenode.hpp" |
42 | #include "opto/opcodes.hpp" |
43 | #include "opto/rootnode.hpp" |
44 | #include "utilities/align.hpp" |
45 | |
46 | #ifndef PRODUCT |
47 | void LRG::dump() const { |
48 | ttyLocker ttyl; |
49 | tty->print("%d " ,num_regs()); |
50 | _mask.dump(); |
51 | if( _msize_valid ) { |
52 | if( mask_size() == compute_mask_size() ) tty->print(", #%d " ,_mask_size); |
53 | else tty->print(", #!!!_%d_vs_%d " ,_mask_size,_mask.Size()); |
54 | } else { |
55 | tty->print(", #?(%d) " ,_mask.Size()); |
56 | } |
57 | |
58 | tty->print("EffDeg: " ); |
59 | if( _degree_valid ) tty->print( "%d " , _eff_degree ); |
60 | else tty->print("? " ); |
61 | |
62 | if( is_multidef() ) { |
63 | tty->print("MultiDef " ); |
64 | if (_defs != 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( _was_spilled1 ) tty->print("Spilled " ); |
81 | if( _was_spilled2 ) tty->print("Spilled2 " ); |
82 | if( _direct_conflict ) tty->print("Direct_conflict " ); |
83 | if( _fat_proj ) tty->print("Fat " ); |
84 | if( _was_lo ) tty->print("Lo " ); |
85 | if( _has_copy ) tty->print("Copy " ); |
86 | if( _at_risk ) tty->print("Risk " ); |
87 | |
88 | if( _must_spill ) tty->print("Must_spill " ); |
89 | if( _is_bound ) tty->print("Bound " ); |
90 | if( _msize_valid ) { |
91 | if( _degree_valid && lo_degree() ) tty->print("Trivial " ); |
92 | } |
93 | |
94 | tty->cr(); |
95 | } |
96 | #endif |
97 | |
98 | // Compute score from cost and area. Low score is best to spill. |
99 | static double raw_score( double cost, double area ) { |
100 | return cost - (area*RegisterCostAreaRatio) * 1.52588e-5; |
101 | } |
102 | |
103 | double LRG::score() const { |
104 | // Scale _area by RegisterCostAreaRatio/64K then subtract from cost. |
105 | // Bigger area lowers score, encourages spilling this live range. |
106 | // Bigger cost raise score, prevents spilling this live range. |
107 | // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer |
108 | // to turn a divide by a constant into a multiply by the reciprical). |
109 | double score = raw_score( _cost, _area); |
110 | |
111 | // Account for area. Basically, LRGs covering large areas are better |
112 | // to spill because more other LRGs get freed up. |
113 | if( _area == 0.0 ) // No area? Then no progress to spill |
114 | return 1e35; |
115 | |
116 | if( _was_spilled2 ) // If spilled once before, we are unlikely |
117 | return score + 1e30; // to make progress again. |
118 | |
119 | if( _cost >= _area*3.0 ) // Tiny area relative to cost |
120 | return score + 1e17; // Probably no progress to spill |
121 | |
122 | if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost |
123 | return score + 1e10; // Likely no progress to spill |
124 | |
125 | return score; |
126 | } |
127 | |
128 | #define NUMBUCKS 3 |
129 | |
130 | // Straight out of Tarjan's union-find algorithm |
131 | uint LiveRangeMap::find_compress(uint lrg) { |
132 | uint cur = lrg; |
133 | uint next = _uf_map.at(cur); |
134 | while (next != cur) { // Scan chain of equivalences |
135 | assert( next < cur, "always union smaller" ); |
136 | cur = next; // until find a fixed-point |
137 | next = _uf_map.at(cur); |
138 | } |
139 | |
140 | // Core of union-find algorithm: update chain of |
141 | // equivalences to be equal to the root. |
142 | while (lrg != next) { |
143 | uint tmp = _uf_map.at(lrg); |
144 | _uf_map.at_put(lrg, next); |
145 | lrg = tmp; |
146 | } |
147 | return lrg; |
148 | } |
149 | |
150 | // Reset the Union-Find map to identity |
151 | void LiveRangeMap::reset_uf_map(uint max_lrg_id) { |
152 | _max_lrg_id= max_lrg_id; |
153 | // Force the Union-Find mapping to be at least this large |
154 | _uf_map.at_put_grow(_max_lrg_id, 0); |
155 | // Initialize it to be the ID mapping. |
156 | for (uint i = 0; i < _max_lrg_id; ++i) { |
157 | _uf_map.at_put(i, i); |
158 | } |
159 | } |
160 | |
161 | // Make all Nodes map directly to their final live range; no need for |
162 | // the Union-Find mapping after this call. |
163 | void LiveRangeMap::compress_uf_map_for_nodes() { |
164 | // For all Nodes, compress mapping |
165 | uint unique = _names.length(); |
166 | for (uint i = 0; i < unique; ++i) { |
167 | uint lrg = _names.at(i); |
168 | uint compressed_lrg = find(lrg); |
169 | if (lrg != compressed_lrg) { |
170 | _names.at_put(i, compressed_lrg); |
171 | } |
172 | } |
173 | } |
174 | |
175 | // Like Find above, but no path compress, so bad asymptotic behavior |
176 | uint LiveRangeMap::find_const(uint lrg) const { |
177 | if (!lrg) { |
178 | return lrg; // Ignore the zero LRG |
179 | } |
180 | |
181 | // Off the end? This happens during debugging dumps when you got |
182 | // brand new live ranges but have not told the allocator yet. |
183 | if (lrg >= _max_lrg_id) { |
184 | return lrg; |
185 | } |
186 | |
187 | uint next = _uf_map.at(lrg); |
188 | while (next != lrg) { // Scan chain of equivalences |
189 | assert(next < lrg, "always union smaller" ); |
190 | lrg = next; // until find a fixed-point |
191 | next = _uf_map.at(lrg); |
192 | } |
193 | return next; |
194 | } |
195 | |
196 | PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated) |
197 | : PhaseRegAlloc(unique, cfg, matcher, |
198 | #ifndef PRODUCT |
199 | print_chaitin_statistics |
200 | #else |
201 | NULL |
202 | #endif |
203 | ) |
204 | , _live(0) |
205 | , _spilled_once(Thread::current()->resource_area()) |
206 | , _spilled_twice(Thread::current()->resource_area()) |
207 | , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0) |
208 | , _oldphi(unique) |
209 | #ifndef PRODUCT |
210 | , _trace_spilling(C->directive()->TraceSpillingOption) |
211 | #endif |
212 | , _lrg_map(Thread::current()->resource_area(), unique) |
213 | , _scheduling_info_generated(scheduling_info_generated) |
214 | , _sched_int_pressure(0, INTPRESSURE) |
215 | , _sched_float_pressure(0, FLOATPRESSURE) |
216 | , _scratch_int_pressure(0, INTPRESSURE) |
217 | , _scratch_float_pressure(0, FLOATPRESSURE) |
218 | { |
219 | Compile::TracePhase tp("ctorChaitin" , &timers[_t_ctorChaitin]); |
220 | |
221 | _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency()); |
222 | |
223 | // Build a list of basic blocks, sorted by frequency |
224 | _blks = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); |
225 | // Experiment with sorting strategies to speed compilation |
226 | double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket |
227 | Block **buckets[NUMBUCKS]; // Array of buckets |
228 | uint buckcnt[NUMBUCKS]; // Array of bucket counters |
229 | double buckval[NUMBUCKS]; // Array of bucket value cutoffs |
230 | for (uint i = 0; i < NUMBUCKS; i++) { |
231 | buckets[i] = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); |
232 | buckcnt[i] = 0; |
233 | // Bump by three orders of magnitude each time |
234 | cutoff *= 0.001; |
235 | buckval[i] = cutoff; |
236 | for (uint j = 0; j < _cfg.number_of_blocks(); j++) { |
237 | buckets[i][j] = NULL; |
238 | } |
239 | } |
240 | // Sort blocks into buckets |
241 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
242 | for (uint j = 0; j < NUMBUCKS; j++) { |
243 | if ((j == NUMBUCKS - 1) || (_cfg.get_block(i)->_freq > buckval[j])) { |
244 | // Assign block to end of list for appropriate bucket |
245 | buckets[j][buckcnt[j]++] = _cfg.get_block(i); |
246 | break; // kick out of inner loop |
247 | } |
248 | } |
249 | } |
250 | // Dump buckets into final block array |
251 | uint blkcnt = 0; |
252 | for (uint i = 0; i < NUMBUCKS; i++) { |
253 | for (uint j = 0; j < buckcnt[i]; j++) { |
254 | _blks[blkcnt++] = buckets[i][j]; |
255 | } |
256 | } |
257 | |
258 | assert(blkcnt == _cfg.number_of_blocks(), "Block array not totally filled" ); |
259 | } |
260 | |
261 | // union 2 sets together. |
262 | void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) { |
263 | uint src = _lrg_map.find(src_n); |
264 | uint dst = _lrg_map.find(dst_n); |
265 | assert(src, "" ); |
266 | assert(dst, "" ); |
267 | assert(src < _lrg_map.max_lrg_id(), "oob" ); |
268 | assert(dst < _lrg_map.max_lrg_id(), "oob" ); |
269 | assert(src < dst, "always union smaller" ); |
270 | _lrg_map.uf_map(dst, src); |
271 | } |
272 | |
273 | void PhaseChaitin::new_lrg(const Node *x, uint lrg) { |
274 | // Make the Node->LRG mapping |
275 | _lrg_map.extend(x->_idx,lrg); |
276 | // Make the Union-Find mapping an identity function |
277 | _lrg_map.uf_extend(lrg, lrg); |
278 | } |
279 | |
280 | |
281 | int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) { |
282 | assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections" ); |
283 | DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); ) |
284 | int found_projs = 0; |
285 | uint cnt = orig->outcnt(); |
286 | for (uint i = 0; i < cnt; i++) { |
287 | Node* proj = orig->raw_out(i); |
288 | if (proj->is_MachProj()) { |
289 | assert(proj->outcnt() == 0, "only kill projections are expected here" ); |
290 | assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections" ); |
291 | found_projs++; |
292 | // Copy kill projections after the cloned node |
293 | Node* kills = proj->clone(); |
294 | kills->set_req(0, copy); |
295 | b->insert_node(kills, idx++); |
296 | _cfg.map_node_to_block(kills, b); |
297 | new_lrg(kills, max_lrg_id++); |
298 | } |
299 | } |
300 | return found_projs; |
301 | } |
302 | |
303 | // Renumber the live ranges to compact them. Makes the IFG smaller. |
304 | void PhaseChaitin::compact() { |
305 | Compile::TracePhase tp("chaitinCompact" , &timers[_t_chaitinCompact]); |
306 | |
307 | // Current the _uf_map contains a series of short chains which are headed |
308 | // by a self-cycle. All the chains run from big numbers to little numbers. |
309 | // The Find() call chases the chains & shortens them for the next Find call. |
310 | // We are going to change this structure slightly. Numbers above a moving |
311 | // wave 'i' are unchanged. Numbers below 'j' point directly to their |
312 | // compacted live range with no further chaining. There are no chains or |
313 | // cycles below 'i', so the Find call no longer works. |
314 | uint j=1; |
315 | uint i; |
316 | for (i = 1; i < _lrg_map.max_lrg_id(); i++) { |
317 | uint lr = _lrg_map.uf_live_range_id(i); |
318 | // Ignore unallocated live ranges |
319 | if (!lr) { |
320 | continue; |
321 | } |
322 | assert(lr <= i, "" ); |
323 | _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr)); |
324 | } |
325 | // Now change the Node->LR mapping to reflect the compacted names |
326 | uint unique = _lrg_map.size(); |
327 | for (i = 0; i < unique; i++) { |
328 | uint lrg_id = _lrg_map.live_range_id(i); |
329 | _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id)); |
330 | } |
331 | |
332 | // Reset the Union-Find mapping |
333 | _lrg_map.reset_uf_map(j); |
334 | } |
335 | |
336 | void PhaseChaitin::Register_Allocate() { |
337 | |
338 | // Above the OLD FP (and in registers) are the incoming arguments. Stack |
339 | // slots in this area are called "arg_slots". Above the NEW FP (and in |
340 | // registers) is the outgoing argument area; above that is the spill/temp |
341 | // area. These are all "frame_slots". Arg_slots start at the zero |
342 | // stack_slots and count up to the known arg_size. Frame_slots start at |
343 | // the stack_slot #arg_size and go up. After allocation I map stack |
344 | // slots to actual offsets. Stack-slots in the arg_slot area are biased |
345 | // by the frame_size; stack-slots in the frame_slot area are biased by 0. |
346 | |
347 | _trip_cnt = 0; |
348 | _alternate = 0; |
349 | _matcher._allocation_started = true; |
350 | |
351 | ResourceArea split_arena(mtCompiler); // Arena for Split local resources |
352 | ResourceArea live_arena(mtCompiler); // Arena for liveness & IFG info |
353 | ResourceMark rm(&live_arena); |
354 | |
355 | // Need live-ness for the IFG; need the IFG for coalescing. If the |
356 | // liveness is JUST for coalescing, then I can get some mileage by renaming |
357 | // all copy-related live ranges low and then using the max copy-related |
358 | // live range as a cut-off for LIVE and the IFG. In other words, I can |
359 | // build a subset of LIVE and IFG just for copies. |
360 | PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false); |
361 | |
362 | // Need IFG for coalescing and coloring |
363 | PhaseIFG ifg(&live_arena); |
364 | _ifg = &ifg; |
365 | |
366 | // Come out of SSA world to the Named world. Assign (virtual) registers to |
367 | // Nodes. Use the same register for all inputs and the output of PhiNodes |
368 | // - effectively ending SSA form. This requires either coalescing live |
369 | // ranges or inserting copies. For the moment, we insert "virtual copies" |
370 | // - we pretend there is a copy prior to each Phi in predecessor blocks. |
371 | // We will attempt to coalesce such "virtual copies" before we manifest |
372 | // them for real. |
373 | de_ssa(); |
374 | |
375 | #ifdef ASSERT |
376 | // Veify the graph before RA. |
377 | verify(&live_arena); |
378 | #endif |
379 | |
380 | { |
381 | Compile::TracePhase tp("computeLive" , &timers[_t_computeLive]); |
382 | _live = NULL; // Mark live as being not available |
383 | rm.reset_to_mark(); // Reclaim working storage |
384 | IndexSet::reset_memory(C, &live_arena); |
385 | ifg.init(_lrg_map.max_lrg_id()); // Empty IFG |
386 | gather_lrg_masks( false ); // Collect LRG masks |
387 | live.compute(_lrg_map.max_lrg_id()); // Compute liveness |
388 | _live = &live; // Mark LIVE as being available |
389 | } |
390 | |
391 | // Base pointers are currently "used" by instructions which define new |
392 | // derived pointers. This makes base pointers live up to the where the |
393 | // derived pointer is made, but not beyond. Really, they need to be live |
394 | // across any GC point where the derived value is live. So this code looks |
395 | // at all the GC points, and "stretches" the live range of any base pointer |
396 | // to the GC point. |
397 | if (stretch_base_pointer_live_ranges(&live_arena)) { |
398 | Compile::TracePhase tp("computeLive (sbplr)" , &timers[_t_computeLive]); |
399 | // Since some live range stretched, I need to recompute live |
400 | _live = NULL; |
401 | rm.reset_to_mark(); // Reclaim working storage |
402 | IndexSet::reset_memory(C, &live_arena); |
403 | ifg.init(_lrg_map.max_lrg_id()); |
404 | gather_lrg_masks(false); |
405 | live.compute(_lrg_map.max_lrg_id()); |
406 | _live = &live; |
407 | } |
408 | // Create the interference graph using virtual copies |
409 | build_ifg_virtual(); // Include stack slots this time |
410 | |
411 | // The IFG is/was triangular. I am 'squaring it up' so Union can run |
412 | // faster. Union requires a 'for all' operation which is slow on the |
413 | // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - |
414 | // meaning I can visit all the Nodes neighbors less than a Node in time |
415 | // O(# of neighbors), but I have to visit all the Nodes greater than a |
416 | // given Node and search them for an instance, i.e., time O(#MaxLRG)). |
417 | _ifg->SquareUp(); |
418 | |
419 | // Aggressive (but pessimistic) copy coalescing. |
420 | // This pass works on virtual copies. Any virtual copies which are not |
421 | // coalesced get manifested as actual copies |
422 | { |
423 | Compile::TracePhase tp("chaitinCoalesce1" , &timers[_t_chaitinCoalesce1]); |
424 | |
425 | PhaseAggressiveCoalesce coalesce(*this); |
426 | coalesce.coalesce_driver(); |
427 | // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do |
428 | // not match the Phi itself, insert a copy. |
429 | coalesce.insert_copies(_matcher); |
430 | if (C->failing()) { |
431 | return; |
432 | } |
433 | } |
434 | |
435 | // After aggressive coalesce, attempt a first cut at coloring. |
436 | // To color, we need the IFG and for that we need LIVE. |
437 | { |
438 | Compile::TracePhase tp("computeLive" , &timers[_t_computeLive]); |
439 | _live = NULL; |
440 | rm.reset_to_mark(); // Reclaim working storage |
441 | IndexSet::reset_memory(C, &live_arena); |
442 | ifg.init(_lrg_map.max_lrg_id()); |
443 | gather_lrg_masks( true ); |
444 | live.compute(_lrg_map.max_lrg_id()); |
445 | _live = &live; |
446 | } |
447 | |
448 | // Build physical interference graph |
449 | uint must_spill = 0; |
450 | must_spill = build_ifg_physical(&live_arena); |
451 | // If we have a guaranteed spill, might as well spill now |
452 | if (must_spill) { |
453 | if(!_lrg_map.max_lrg_id()) { |
454 | return; |
455 | } |
456 | // Bail out if unique gets too large (ie - unique > MaxNodeLimit) |
457 | C->check_node_count(10*must_spill, "out of nodes before split" ); |
458 | if (C->failing()) { |
459 | return; |
460 | } |
461 | |
462 | uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere |
463 | _lrg_map.set_max_lrg_id(new_max_lrg_id); |
464 | // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) |
465 | // or we failed to split |
466 | C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split" ); |
467 | if (C->failing()) { |
468 | return; |
469 | } |
470 | |
471 | NOT_PRODUCT(C->verify_graph_edges();) |
472 | |
473 | compact(); // Compact LRGs; return new lower max lrg |
474 | |
475 | { |
476 | Compile::TracePhase tp("computeLive" , &timers[_t_computeLive]); |
477 | _live = NULL; |
478 | rm.reset_to_mark(); // Reclaim working storage |
479 | IndexSet::reset_memory(C, &live_arena); |
480 | ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph |
481 | gather_lrg_masks( true ); // Collect intersect mask |
482 | live.compute(_lrg_map.max_lrg_id()); // Compute LIVE |
483 | _live = &live; |
484 | } |
485 | build_ifg_physical(&live_arena); |
486 | _ifg->SquareUp(); |
487 | _ifg->Compute_Effective_Degree(); |
488 | // Only do conservative coalescing if requested |
489 | if (OptoCoalesce) { |
490 | Compile::TracePhase tp("chaitinCoalesce2" , &timers[_t_chaitinCoalesce2]); |
491 | // Conservative (and pessimistic) copy coalescing of those spills |
492 | PhaseConservativeCoalesce coalesce(*this); |
493 | // If max live ranges greater than cutoff, don't color the stack. |
494 | // This cutoff can be larger than below since it is only done once. |
495 | coalesce.coalesce_driver(); |
496 | } |
497 | _lrg_map.compress_uf_map_for_nodes(); |
498 | |
499 | #ifdef ASSERT |
500 | verify(&live_arena, true); |
501 | #endif |
502 | } else { |
503 | ifg.SquareUp(); |
504 | ifg.Compute_Effective_Degree(); |
505 | #ifdef ASSERT |
506 | set_was_low(); |
507 | #endif |
508 | } |
509 | |
510 | // Prepare for Simplify & Select |
511 | cache_lrg_info(); // Count degree of LRGs |
512 | |
513 | // Simplify the InterFerence Graph by removing LRGs of low degree. |
514 | // LRGs of low degree are trivially colorable. |
515 | Simplify(); |
516 | |
517 | // Select colors by re-inserting LRGs back into the IFG in reverse order. |
518 | // Return whether or not something spills. |
519 | uint spills = Select( ); |
520 | |
521 | // If we spill, split and recycle the entire thing |
522 | while( spills ) { |
523 | if( _trip_cnt++ > 24 ) { |
524 | DEBUG_ONLY( dump_for_spill_split_recycle(); ) |
525 | if( _trip_cnt > 27 ) { |
526 | C->record_method_not_compilable("failed spill-split-recycle sanity check" ); |
527 | return; |
528 | } |
529 | } |
530 | |
531 | if (!_lrg_map.max_lrg_id()) { |
532 | return; |
533 | } |
534 | uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere |
535 | _lrg_map.set_max_lrg_id(new_max_lrg_id); |
536 | // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) |
537 | C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split" ); |
538 | if (C->failing()) { |
539 | return; |
540 | } |
541 | |
542 | compact(); // Compact LRGs; return new lower max lrg |
543 | |
544 | // Nuke the live-ness and interference graph and LiveRanGe info |
545 | { |
546 | Compile::TracePhase tp("computeLive" , &timers[_t_computeLive]); |
547 | _live = NULL; |
548 | rm.reset_to_mark(); // Reclaim working storage |
549 | IndexSet::reset_memory(C, &live_arena); |
550 | ifg.init(_lrg_map.max_lrg_id()); |
551 | |
552 | // Create LiveRanGe array. |
553 | // Intersect register masks for all USEs and DEFs |
554 | gather_lrg_masks(true); |
555 | live.compute(_lrg_map.max_lrg_id()); |
556 | _live = &live; |
557 | } |
558 | must_spill = build_ifg_physical(&live_arena); |
559 | _ifg->SquareUp(); |
560 | _ifg->Compute_Effective_Degree(); |
561 | |
562 | // Only do conservative coalescing if requested |
563 | if (OptoCoalesce) { |
564 | Compile::TracePhase tp("chaitinCoalesce3" , &timers[_t_chaitinCoalesce3]); |
565 | // Conservative (and pessimistic) copy coalescing |
566 | PhaseConservativeCoalesce coalesce(*this); |
567 | // Check for few live ranges determines how aggressive coalesce is. |
568 | coalesce.coalesce_driver(); |
569 | } |
570 | _lrg_map.compress_uf_map_for_nodes(); |
571 | #ifdef ASSERT |
572 | verify(&live_arena, true); |
573 | #endif |
574 | cache_lrg_info(); // Count degree of LRGs |
575 | |
576 | // Simplify the InterFerence Graph by removing LRGs of low degree. |
577 | // LRGs of low degree are trivially colorable. |
578 | Simplify(); |
579 | |
580 | // Select colors by re-inserting LRGs back into the IFG in reverse order. |
581 | // Return whether or not something spills. |
582 | spills = Select(); |
583 | } |
584 | |
585 | // Count number of Simplify-Select trips per coloring success. |
586 | _allocator_attempts += _trip_cnt + 1; |
587 | _allocator_successes += 1; |
588 | |
589 | // Peephole remove copies |
590 | post_allocate_copy_removal(); |
591 | |
592 | // Merge multidefs if multiple defs representing the same value are used in a single block. |
593 | merge_multidefs(); |
594 | |
595 | #ifdef ASSERT |
596 | // Veify the graph after RA. |
597 | verify(&live_arena); |
598 | #endif |
599 | |
600 | // max_reg is past the largest *register* used. |
601 | // Convert that to a frame_slot number. |
602 | if (_max_reg <= _matcher._new_SP) { |
603 | _framesize = C->out_preserve_stack_slots(); |
604 | } |
605 | else { |
606 | _framesize = _max_reg -_matcher._new_SP; |
607 | } |
608 | assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough" ); |
609 | |
610 | // This frame must preserve the required fp alignment |
611 | _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots()); |
612 | assert(_framesize <= 1000000, "sanity check" ); |
613 | #ifndef PRODUCT |
614 | _total_framesize += _framesize; |
615 | if ((int)_framesize > _max_framesize) { |
616 | _max_framesize = _framesize; |
617 | } |
618 | #endif |
619 | |
620 | // Convert CISC spills |
621 | fixup_spills(); |
622 | |
623 | // Log regalloc results |
624 | CompileLog* log = Compile::current()->log(); |
625 | if (log != NULL) { |
626 | log->elem("regalloc attempts='%d' success='%d'" , _trip_cnt, !C->failing()); |
627 | } |
628 | |
629 | if (C->failing()) { |
630 | return; |
631 | } |
632 | |
633 | NOT_PRODUCT(C->verify_graph_edges();) |
634 | |
635 | // Move important info out of the live_arena to longer lasting storage. |
636 | alloc_node_regs(_lrg_map.size()); |
637 | for (uint i=0; i < _lrg_map.size(); i++) { |
638 | if (_lrg_map.live_range_id(i)) { // Live range associated with Node? |
639 | LRG &lrg = lrgs(_lrg_map.live_range_id(i)); |
640 | if (!lrg.alive()) { |
641 | set_bad(i); |
642 | } else if (lrg.num_regs() == 1) { |
643 | set1(i, lrg.reg()); |
644 | } else { // Must be a register-set |
645 | if (!lrg._fat_proj) { // Must be aligned adjacent register set |
646 | // Live ranges record the highest register in their mask. |
647 | // We want the low register for the AD file writer's convenience. |
648 | OptoReg::Name hi = lrg.reg(); // Get hi register |
649 | OptoReg::Name lo = OptoReg::add(hi, (1-lrg.num_regs())); // Find lo |
650 | // We have to use pair [lo,lo+1] even for wide vectors because |
651 | // the rest of code generation works only with pairs. It is safe |
652 | // since for registers encoding only 'lo' is used. |
653 | // Second reg from pair is used in ScheduleAndBundle on SPARC where |
654 | // vector max size is 8 which corresponds to registers pair. |
655 | // It is also used in BuildOopMaps but oop operations are not |
656 | // vectorized. |
657 | set2(i, lo); |
658 | } else { // Misaligned; extract 2 bits |
659 | OptoReg::Name hi = lrg.reg(); // Get hi register |
660 | lrg.Remove(hi); // Yank from mask |
661 | int lo = lrg.mask().find_first_elem(); // Find lo |
662 | set_pair(i, hi, lo); |
663 | } |
664 | } |
665 | if( lrg._is_oop ) _node_oops.set(i); |
666 | } else { |
667 | set_bad(i); |
668 | } |
669 | } |
670 | |
671 | // Done! |
672 | _live = NULL; |
673 | _ifg = NULL; |
674 | C->set_indexSet_arena(NULL); // ResourceArea is at end of scope |
675 | } |
676 | |
677 | void PhaseChaitin::de_ssa() { |
678 | // Set initial Names for all Nodes. Most Nodes get the virtual register |
679 | // number. A few get the ZERO live range number. These do not |
680 | // get allocated, but instead rely on correct scheduling to ensure that |
681 | // only one instance is simultaneously live at a time. |
682 | uint lr_counter = 1; |
683 | for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { |
684 | Block* block = _cfg.get_block(i); |
685 | uint cnt = block->number_of_nodes(); |
686 | |
687 | // Handle all the normal Nodes in the block |
688 | for( uint j = 0; j < cnt; j++ ) { |
689 | Node *n = block->get_node(j); |
690 | // Pre-color to the zero live range, or pick virtual register |
691 | const RegMask &rm = n->out_RegMask(); |
692 | _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0); |
693 | } |
694 | } |
695 | |
696 | // Reset the Union-Find mapping to be identity |
697 | _lrg_map.reset_uf_map(lr_counter); |
698 | } |
699 | |
700 | void PhaseChaitin::mark_ssa() { |
701 | // Use ssa names to populate the live range maps or if no mask |
702 | // is available, use the 0 entry. |
703 | uint max_idx = 0; |
704 | for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { |
705 | Block* block = _cfg.get_block(i); |
706 | uint cnt = block->number_of_nodes(); |
707 | |
708 | // Handle all the normal Nodes in the block |
709 | for ( uint j = 0; j < cnt; j++ ) { |
710 | Node *n = block->get_node(j); |
711 | // Pre-color to the zero live range, or pick virtual register |
712 | const RegMask &rm = n->out_RegMask(); |
713 | _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0); |
714 | max_idx = (n->_idx > max_idx) ? n->_idx : max_idx; |
715 | } |
716 | } |
717 | _lrg_map.set_max_lrg_id(max_idx+1); |
718 | |
719 | // Reset the Union-Find mapping to be identity |
720 | _lrg_map.reset_uf_map(max_idx+1); |
721 | } |
722 | |
723 | |
724 | // Gather LiveRanGe information, including register masks. Modification of |
725 | // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. |
726 | void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { |
727 | |
728 | // Nail down the frame pointer live range |
729 | uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr)); |
730 | lrgs(fp_lrg)._cost += 1e12; // Cost is infinite |
731 | |
732 | // For all blocks |
733 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
734 | Block* block = _cfg.get_block(i); |
735 | |
736 | // For all instructions |
737 | for (uint j = 1; j < block->number_of_nodes(); j++) { |
738 | Node* n = block->get_node(j); |
739 | uint input_edge_start =1; // Skip control most nodes |
740 | bool is_machine_node = false; |
741 | if (n->is_Mach()) { |
742 | is_machine_node = true; |
743 | input_edge_start = n->as_Mach()->oper_input_base(); |
744 | } |
745 | uint idx = n->is_Copy(); |
746 | |
747 | // Get virtual register number, same as LiveRanGe index |
748 | uint vreg = _lrg_map.live_range_id(n); |
749 | LRG& lrg = lrgs(vreg); |
750 | if (vreg) { // No vreg means un-allocable (e.g. memory) |
751 | |
752 | // Check for float-vs-int live range (used in register-pressure |
753 | // calculations) |
754 | const Type *n_type = n->bottom_type(); |
755 | if (n_type->is_floatingpoint()) { |
756 | lrg._is_float = 1; |
757 | } |
758 | |
759 | // Check for twice prior spilling. Once prior spilling might have |
760 | // spilled 'soft', 2nd prior spill should have spilled 'hard' and |
761 | // further spilling is unlikely to make progress. |
762 | if (_spilled_once.test(n->_idx)) { |
763 | lrg._was_spilled1 = 1; |
764 | if (_spilled_twice.test(n->_idx)) { |
765 | lrg._was_spilled2 = 1; |
766 | } |
767 | } |
768 | |
769 | #ifndef PRODUCT |
770 | // Collect bits not used by product code, but which may be useful for |
771 | // debugging. |
772 | |
773 | // Collect has-copy bit |
774 | if (idx) { |
775 | lrg._has_copy = 1; |
776 | uint clidx = _lrg_map.live_range_id(n->in(idx)); |
777 | LRG& copy_src = lrgs(clidx); |
778 | copy_src._has_copy = 1; |
779 | } |
780 | |
781 | if (trace_spilling() && lrg._def != NULL) { |
782 | // collect defs for MultiDef printing |
783 | if (lrg._defs == NULL) { |
784 | lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, NULL); |
785 | lrg._defs->append(lrg._def); |
786 | } |
787 | lrg._defs->append(n); |
788 | } |
789 | #endif |
790 | |
791 | // Check for a single def LRG; these can spill nicely |
792 | // via rematerialization. Flag as NULL for no def found |
793 | // yet, or 'n' for single def or -1 for many defs. |
794 | lrg._def = lrg._def ? NodeSentinel : n; |
795 | |
796 | // Limit result register mask to acceptable registers |
797 | const RegMask &rm = n->out_RegMask(); |
798 | lrg.AND( rm ); |
799 | |
800 | uint ireg = n->ideal_reg(); |
801 | assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, |
802 | "oops must be in Op_RegP's" ); |
803 | |
804 | // Check for vector live range (only if vector register is used). |
805 | // On SPARC vector uses RegD which could be misaligned so it is not |
806 | // processes as vector in RA. |
807 | if (RegMask::is_vector(ireg)) |
808 | lrg._is_vector = 1; |
809 | assert(n_type->isa_vect() == NULL || lrg._is_vector || ireg == Op_RegD || ireg == Op_RegL, |
810 | "vector must be in vector registers" ); |
811 | |
812 | // Check for bound register masks |
813 | const RegMask &lrgmask = lrg.mask(); |
814 | if (lrgmask.is_bound(ireg)) { |
815 | lrg._is_bound = 1; |
816 | } |
817 | |
818 | // Check for maximum frequency value |
819 | if (lrg._maxfreq < block->_freq) { |
820 | lrg._maxfreq = block->_freq; |
821 | } |
822 | |
823 | // Check for oop-iness, or long/double |
824 | // Check for multi-kill projection |
825 | switch (ireg) { |
826 | case MachProjNode::fat_proj: |
827 | // Fat projections have size equal to number of registers killed |
828 | lrg.set_num_regs(rm.Size()); |
829 | lrg.set_reg_pressure(lrg.num_regs()); |
830 | lrg._fat_proj = 1; |
831 | lrg._is_bound = 1; |
832 | break; |
833 | case Op_RegP: |
834 | #ifdef _LP64 |
835 | lrg.set_num_regs(2); // Size is 2 stack words |
836 | #else |
837 | lrg.set_num_regs(1); // Size is 1 stack word |
838 | #endif |
839 | // Register pressure is tracked relative to the maximum values |
840 | // suggested for that platform, INTPRESSURE and FLOATPRESSURE, |
841 | // and relative to other types which compete for the same regs. |
842 | // |
843 | // The following table contains suggested values based on the |
844 | // architectures as defined in each .ad file. |
845 | // INTPRESSURE and FLOATPRESSURE may be tuned differently for |
846 | // compile-speed or performance. |
847 | // Note1: |
848 | // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 |
849 | // since .ad registers are defined as high and low halves. |
850 | // These reg_pressure values remain compatible with the code |
851 | // in is_high_pressure() which relates get_invalid_mask_size(), |
852 | // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. |
853 | // Note2: |
854 | // SPARC -d32 has 24 registers available for integral values, |
855 | // but only 10 of these are safe for 64-bit longs. |
856 | // Using set_reg_pressure(2) for both int and long means |
857 | // the allocator will believe it can fit 26 longs into |
858 | // registers. Using 2 for longs and 1 for ints means the |
859 | // allocator will attempt to put 52 integers into registers. |
860 | // The settings below limit this problem to methods with |
861 | // many long values which are being run on 32-bit SPARC. |
862 | // |
863 | // ------------------- reg_pressure -------------------- |
864 | // Each entry is reg_pressure_per_value,number_of_regs |
865 | // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE |
866 | // IA32 2 1 1 1 1 6 6 |
867 | // IA64 1 1 1 1 1 50 41 |
868 | // SPARC 2 2 2 2 2 48 (24) 52 (26) |
869 | // SPARCV9 2 2 2 2 2 48 (24) 52 (26) |
870 | // AMD64 1 1 1 1 1 14 15 |
871 | // ----------------------------------------------------- |
872 | #if defined(SPARC) |
873 | lrg.set_reg_pressure(2); // use for v9 as well |
874 | #else |
875 | lrg.set_reg_pressure(1); // normally one value per register |
876 | #endif |
877 | if( n_type->isa_oop_ptr() ) { |
878 | lrg._is_oop = 1; |
879 | } |
880 | break; |
881 | case Op_RegL: // Check for long or double |
882 | case Op_RegD: |
883 | lrg.set_num_regs(2); |
884 | // Define platform specific register pressure |
885 | #if defined(SPARC) || defined(ARM32) |
886 | lrg.set_reg_pressure(2); |
887 | #elif defined(IA32) |
888 | if( ireg == Op_RegL ) { |
889 | lrg.set_reg_pressure(2); |
890 | } else { |
891 | lrg.set_reg_pressure(1); |
892 | } |
893 | #else |
894 | lrg.set_reg_pressure(1); // normally one value per register |
895 | #endif |
896 | // If this def of a double forces a mis-aligned double, |
897 | // flag as '_fat_proj' - really flag as allowing misalignment |
898 | // AND changes how we count interferences. A mis-aligned |
899 | // double can interfere with TWO aligned pairs, or effectively |
900 | // FOUR registers! |
901 | if (rm.is_misaligned_pair()) { |
902 | lrg._fat_proj = 1; |
903 | lrg._is_bound = 1; |
904 | } |
905 | break; |
906 | case Op_RegF: |
907 | case Op_RegI: |
908 | case Op_RegN: |
909 | case Op_RegFlags: |
910 | case 0: // not an ideal register |
911 | lrg.set_num_regs(1); |
912 | #ifdef SPARC |
913 | lrg.set_reg_pressure(2); |
914 | #else |
915 | lrg.set_reg_pressure(1); |
916 | #endif |
917 | break; |
918 | case Op_VecS: |
919 | assert(Matcher::vector_size_supported(T_BYTE,4), "sanity" ); |
920 | assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity" ); |
921 | lrg.set_num_regs(RegMask::SlotsPerVecS); |
922 | lrg.set_reg_pressure(1); |
923 | break; |
924 | case Op_VecD: |
925 | assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity" ); |
926 | assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity" ); |
927 | assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned" ); |
928 | lrg.set_num_regs(RegMask::SlotsPerVecD); |
929 | lrg.set_reg_pressure(1); |
930 | break; |
931 | case Op_VecX: |
932 | assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity" ); |
933 | assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity" ); |
934 | assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned" ); |
935 | lrg.set_num_regs(RegMask::SlotsPerVecX); |
936 | lrg.set_reg_pressure(1); |
937 | break; |
938 | case Op_VecY: |
939 | assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity" ); |
940 | assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity" ); |
941 | assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned" ); |
942 | lrg.set_num_regs(RegMask::SlotsPerVecY); |
943 | lrg.set_reg_pressure(1); |
944 | break; |
945 | case Op_VecZ: |
946 | assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity" ); |
947 | assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity" ); |
948 | assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned" ); |
949 | lrg.set_num_regs(RegMask::SlotsPerVecZ); |
950 | lrg.set_reg_pressure(1); |
951 | break; |
952 | default: |
953 | ShouldNotReachHere(); |
954 | } |
955 | } |
956 | |
957 | // Now do the same for inputs |
958 | uint cnt = n->req(); |
959 | // Setup for CISC SPILLING |
960 | uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; |
961 | if( UseCISCSpill && after_aggressive ) { |
962 | inp = n->cisc_operand(); |
963 | if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) |
964 | // Convert operand number to edge index number |
965 | inp = n->as_Mach()->operand_index(inp); |
966 | } |
967 | |
968 | // Prepare register mask for each input |
969 | for( uint k = input_edge_start; k < cnt; k++ ) { |
970 | uint vreg = _lrg_map.live_range_id(n->in(k)); |
971 | if (!vreg) { |
972 | continue; |
973 | } |
974 | |
975 | // If this instruction is CISC Spillable, add the flags |
976 | // bit to its appropriate input |
977 | if( UseCISCSpill && after_aggressive && inp == k ) { |
978 | #ifndef PRODUCT |
979 | if( TraceCISCSpill ) { |
980 | tty->print(" use_cisc_RegMask: " ); |
981 | n->dump(); |
982 | } |
983 | #endif |
984 | n->as_Mach()->use_cisc_RegMask(); |
985 | } |
986 | |
987 | if (is_machine_node && _scheduling_info_generated) { |
988 | MachNode* cur_node = n->as_Mach(); |
989 | // this is cleaned up by register allocation |
990 | if (k >= cur_node->num_opnds()) continue; |
991 | } |
992 | |
993 | LRG &lrg = lrgs(vreg); |
994 | // // Testing for floating point code shape |
995 | // Node *test = n->in(k); |
996 | // if( test->is_Mach() ) { |
997 | // MachNode *m = test->as_Mach(); |
998 | // int op = m->ideal_Opcode(); |
999 | // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { |
1000 | // int zzz = 1; |
1001 | // } |
1002 | // } |
1003 | |
1004 | // Limit result register mask to acceptable registers. |
1005 | // Do not limit registers from uncommon uses before |
1006 | // AggressiveCoalesce. This effectively pre-virtual-splits |
1007 | // around uncommon uses of common defs. |
1008 | const RegMask &rm = n->in_RegMask(k); |
1009 | if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) { |
1010 | // Since we are BEFORE aggressive coalesce, leave the register |
1011 | // mask untrimmed by the call. This encourages more coalescing. |
1012 | // Later, AFTER aggressive, this live range will have to spill |
1013 | // but the spiller handles slow-path calls very nicely. |
1014 | } else { |
1015 | lrg.AND( rm ); |
1016 | } |
1017 | |
1018 | // Check for bound register masks |
1019 | const RegMask &lrgmask = lrg.mask(); |
1020 | uint kreg = n->in(k)->ideal_reg(); |
1021 | bool is_vect = RegMask::is_vector(kreg); |
1022 | assert(n->in(k)->bottom_type()->isa_vect() == NULL || |
1023 | is_vect || kreg == Op_RegD || kreg == Op_RegL, |
1024 | "vector must be in vector registers" ); |
1025 | if (lrgmask.is_bound(kreg)) |
1026 | lrg._is_bound = 1; |
1027 | |
1028 | // If this use of a double forces a mis-aligned double, |
1029 | // flag as '_fat_proj' - really flag as allowing misalignment |
1030 | // AND changes how we count interferences. A mis-aligned |
1031 | // double can interfere with TWO aligned pairs, or effectively |
1032 | // FOUR registers! |
1033 | #ifdef ASSERT |
1034 | if (is_vect && !_scheduling_info_generated) { |
1035 | if (lrg.num_regs() != 0) { |
1036 | assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned" ); |
1037 | assert(!lrg._fat_proj, "sanity" ); |
1038 | assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity" ); |
1039 | } else { |
1040 | assert(n->is_Phi(), "not all inputs processed only if Phi" ); |
1041 | } |
1042 | } |
1043 | #endif |
1044 | if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) { |
1045 | lrg._fat_proj = 1; |
1046 | lrg._is_bound = 1; |
1047 | } |
1048 | // if the LRG is an unaligned pair, we will have to spill |
1049 | // so clear the LRG's register mask if it is not already spilled |
1050 | if (!is_vect && !n->is_SpillCopy() && |
1051 | (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && |
1052 | lrgmask.is_misaligned_pair()) { |
1053 | lrg.Clear(); |
1054 | } |
1055 | |
1056 | // Check for maximum frequency value |
1057 | if (lrg._maxfreq < block->_freq) { |
1058 | lrg._maxfreq = block->_freq; |
1059 | } |
1060 | |
1061 | } // End for all allocated inputs |
1062 | } // end for all instructions |
1063 | } // end for all blocks |
1064 | |
1065 | // Final per-liverange setup |
1066 | for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) { |
1067 | LRG &lrg = lrgs(i2); |
1068 | assert(!lrg._is_vector || !lrg._fat_proj, "sanity" ); |
1069 | if (lrg.num_regs() > 1 && !lrg._fat_proj) { |
1070 | lrg.clear_to_sets(); |
1071 | } |
1072 | lrg.compute_set_mask_size(); |
1073 | if (lrg.not_free()) { // Handle case where we lose from the start |
1074 | lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); |
1075 | lrg._direct_conflict = 1; |
1076 | } |
1077 | lrg.set_degree(0); // no neighbors in IFG yet |
1078 | } |
1079 | } |
1080 | |
1081 | // Set the was-lo-degree bit. Conservative coalescing should not change the |
1082 | // colorability of the graph. If any live range was of low-degree before |
1083 | // coalescing, it should Simplify. This call sets the was-lo-degree bit. |
1084 | // The bit is checked in Simplify. |
1085 | void PhaseChaitin::set_was_low() { |
1086 | #ifdef ASSERT |
1087 | for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { |
1088 | int size = lrgs(i).num_regs(); |
1089 | uint old_was_lo = lrgs(i)._was_lo; |
1090 | lrgs(i)._was_lo = 0; |
1091 | if( lrgs(i).lo_degree() ) { |
1092 | lrgs(i)._was_lo = 1; // Trivially of low degree |
1093 | } else { // Else check the Brigg's assertion |
1094 | // Brigg's observation is that the lo-degree neighbors of a |
1095 | // hi-degree live range will not interfere with the color choices |
1096 | // of said hi-degree live range. The Simplify reverse-stack-coloring |
1097 | // order takes care of the details. Hence you do not have to count |
1098 | // low-degree neighbors when determining if this guy colors. |
1099 | int briggs_degree = 0; |
1100 | IndexSet *s = _ifg->neighbors(i); |
1101 | IndexSetIterator elements(s); |
1102 | uint lidx; |
1103 | while((lidx = elements.next()) != 0) { |
1104 | if( !lrgs(lidx).lo_degree() ) |
1105 | briggs_degree += MAX2(size,lrgs(lidx).num_regs()); |
1106 | } |
1107 | if( briggs_degree < lrgs(i).degrees_of_freedom() ) |
1108 | lrgs(i)._was_lo = 1; // Low degree via the briggs assertion |
1109 | } |
1110 | assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease" ); |
1111 | } |
1112 | #endif |
1113 | } |
1114 | |
1115 | // Compute cost/area ratio, in case we spill. Build the lo-degree list. |
1116 | void PhaseChaitin::cache_lrg_info( ) { |
1117 | Compile::TracePhase tp("chaitinCacheLRG" , &timers[_t_chaitinCacheLRG]); |
1118 | |
1119 | for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { |
1120 | LRG &lrg = lrgs(i); |
1121 | |
1122 | // Check for being of low degree: means we can be trivially colored. |
1123 | // Low degree, dead or must-spill guys just get to simplify right away |
1124 | if( lrg.lo_degree() || |
1125 | !lrg.alive() || |
1126 | lrg._must_spill ) { |
1127 | // Split low degree list into those guys that must get a |
1128 | // register and those that can go to register or stack. |
1129 | // The idea is LRGs that can go register or stack color first when |
1130 | // they have a good chance of getting a register. The register-only |
1131 | // lo-degree live ranges always get a register. |
1132 | OptoReg::Name hi_reg = lrg.mask().find_last_elem(); |
1133 | if( OptoReg::is_stack(hi_reg)) { // Can go to stack? |
1134 | lrg._next = _lo_stk_degree; |
1135 | _lo_stk_degree = i; |
1136 | } else { |
1137 | lrg._next = _lo_degree; |
1138 | _lo_degree = i; |
1139 | } |
1140 | } else { // Else high degree |
1141 | lrgs(_hi_degree)._prev = i; |
1142 | lrg._next = _hi_degree; |
1143 | lrg._prev = 0; |
1144 | _hi_degree = i; |
1145 | } |
1146 | } |
1147 | } |
1148 | |
1149 | // Simplify the IFG by removing LRGs of low degree. |
1150 | void PhaseChaitin::Simplify( ) { |
1151 | Compile::TracePhase tp("chaitinSimplify" , &timers[_t_chaitinSimplify]); |
1152 | |
1153 | while( 1 ) { // Repeat till simplified it all |
1154 | // May want to explore simplifying lo_degree before _lo_stk_degree. |
1155 | // This might result in more spills coloring into registers during |
1156 | // Select(). |
1157 | while( _lo_degree || _lo_stk_degree ) { |
1158 | // If possible, pull from lo_stk first |
1159 | uint lo; |
1160 | if( _lo_degree ) { |
1161 | lo = _lo_degree; |
1162 | _lo_degree = lrgs(lo)._next; |
1163 | } else { |
1164 | lo = _lo_stk_degree; |
1165 | _lo_stk_degree = lrgs(lo)._next; |
1166 | } |
1167 | |
1168 | // Put the simplified guy on the simplified list. |
1169 | lrgs(lo)._next = _simplified; |
1170 | _simplified = lo; |
1171 | // If this guy is "at risk" then mark his current neighbors |
1172 | if( lrgs(lo)._at_risk ) { |
1173 | IndexSetIterator elements(_ifg->neighbors(lo)); |
1174 | uint datum; |
1175 | while ((datum = elements.next()) != 0) { |
1176 | lrgs(datum)._risk_bias = lo; |
1177 | } |
1178 | } |
1179 | |
1180 | // Yank this guy from the IFG. |
1181 | IndexSet *adj = _ifg->remove_node( lo ); |
1182 | |
1183 | // If any neighbors' degrees fall below their number of |
1184 | // allowed registers, then put that neighbor on the low degree |
1185 | // list. Note that 'degree' can only fall and 'numregs' is |
1186 | // unchanged by this action. Thus the two are equal at most once, |
1187 | // so LRGs hit the lo-degree worklist at most once. |
1188 | IndexSetIterator elements(adj); |
1189 | uint neighbor; |
1190 | while ((neighbor = elements.next()) != 0) { |
1191 | LRG *n = &lrgs(neighbor); |
1192 | #ifdef ASSERT |
1193 | if( VerifyOpto || VerifyRegisterAllocator ) { |
1194 | assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); |
1195 | } |
1196 | #endif |
1197 | |
1198 | // Check for just becoming of-low-degree just counting registers. |
1199 | // _must_spill live ranges are already on the low degree list. |
1200 | if( n->just_lo_degree() && !n->_must_spill ) { |
1201 | assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice" ); |
1202 | // Pull from hi-degree list |
1203 | uint prev = n->_prev; |
1204 | uint next = n->_next; |
1205 | if( prev ) lrgs(prev)._next = next; |
1206 | else _hi_degree = next; |
1207 | lrgs(next)._prev = prev; |
1208 | n->_next = _lo_degree; |
1209 | _lo_degree = neighbor; |
1210 | } |
1211 | } |
1212 | } // End of while lo-degree/lo_stk_degree worklist not empty |
1213 | |
1214 | // Check for got everything: is hi-degree list empty? |
1215 | if( !_hi_degree ) break; |
1216 | |
1217 | // Time to pick a potential spill guy |
1218 | uint lo_score = _hi_degree; |
1219 | double score = lrgs(lo_score).score(); |
1220 | double area = lrgs(lo_score)._area; |
1221 | double cost = lrgs(lo_score)._cost; |
1222 | bool bound = lrgs(lo_score)._is_bound; |
1223 | |
1224 | // Find cheapest guy |
1225 | debug_only( int lo_no_simplify=0; ); |
1226 | for( uint i = _hi_degree; i; i = lrgs(i)._next ) { |
1227 | assert( !(*_ifg->_yanked)[i], "" ); |
1228 | // It's just vaguely possible to move hi-degree to lo-degree without |
1229 | // going through a just-lo-degree stage: If you remove a double from |
1230 | // a float live range it's degree will drop by 2 and you can skip the |
1231 | // just-lo-degree stage. It's very rare (shows up after 5000+ methods |
1232 | // in -Xcomp of Java2Demo). So just choose this guy to simplify next. |
1233 | if( lrgs(i).lo_degree() ) { |
1234 | lo_score = i; |
1235 | break; |
1236 | } |
1237 | debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); |
1238 | double iscore = lrgs(i).score(); |
1239 | double iarea = lrgs(i)._area; |
1240 | double icost = lrgs(i)._cost; |
1241 | bool ibound = lrgs(i)._is_bound; |
1242 | |
1243 | // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area |
1244 | // wins. Ties happen because all live ranges in question have spilled |
1245 | // a few times before and the spill-score adds a huge number which |
1246 | // washes out the low order bits. We are choosing the lesser of 2 |
1247 | // evils; in this case pick largest area to spill. |
1248 | // Ties also happen when live ranges are defined and used only inside |
1249 | // one block. In which case their area is 0 and score set to max. |
1250 | // In such case choose bound live range over unbound to free registers |
1251 | // or with smaller cost to spill. |
1252 | if( iscore < score || |
1253 | (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || |
1254 | (iscore == score && iarea == area && |
1255 | ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) { |
1256 | lo_score = i; |
1257 | score = iscore; |
1258 | area = iarea; |
1259 | cost = icost; |
1260 | bound = ibound; |
1261 | } |
1262 | } |
1263 | LRG *lo_lrg = &lrgs(lo_score); |
1264 | // The live range we choose for spilling is either hi-degree, or very |
1265 | // rarely it can be low-degree. If we choose a hi-degree live range |
1266 | // there better not be any lo-degree choices. |
1267 | assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); |
1268 | |
1269 | // Pull from hi-degree list |
1270 | uint prev = lo_lrg->_prev; |
1271 | uint next = lo_lrg->_next; |
1272 | if( prev ) lrgs(prev)._next = next; |
1273 | else _hi_degree = next; |
1274 | lrgs(next)._prev = prev; |
1275 | // Jam him on the lo-degree list, despite his high degree. |
1276 | // Maybe he'll get a color, and maybe he'll spill. |
1277 | // Only Select() will know. |
1278 | lrgs(lo_score)._at_risk = true; |
1279 | _lo_degree = lo_score; |
1280 | lo_lrg->_next = 0; |
1281 | |
1282 | } // End of while not simplified everything |
1283 | |
1284 | } |
1285 | |
1286 | // Is 'reg' register legal for 'lrg'? |
1287 | static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) { |
1288 | if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) && |
1289 | lrg.mask().Member(OptoReg::add(reg,-chunk))) { |
1290 | // RA uses OptoReg which represent the highest element of a registers set. |
1291 | // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set |
1292 | // in which XMMd is used by RA to represent such vectors. A double value |
1293 | // uses [XMM,XMMb] pairs and XMMb is used by RA for it. |
1294 | // The register mask uses largest bits set of overlapping register sets. |
1295 | // On x86 with AVX it uses 8 bits for each XMM registers set. |
1296 | // |
1297 | // The 'lrg' already has cleared-to-set register mask (done in Select() |
1298 | // before calling choose_color()). Passing mask.Member(reg) check above |
1299 | // indicates that the size (num_regs) of 'reg' set is less or equal to |
1300 | // 'lrg' set size. |
1301 | // For set size 1 any register which is member of 'lrg' mask is legal. |
1302 | if (lrg.num_regs()==1) |
1303 | return true; |
1304 | // For larger sets only an aligned register with the same set size is legal. |
1305 | int mask = lrg.num_regs()-1; |
1306 | if ((reg&mask) == mask) |
1307 | return true; |
1308 | } |
1309 | return false; |
1310 | } |
1311 | |
1312 | // Choose a color using the biasing heuristic |
1313 | OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { |
1314 | |
1315 | // Check for "at_risk" LRG's |
1316 | uint risk_lrg = _lrg_map.find(lrg._risk_bias); |
1317 | if( risk_lrg != 0 ) { |
1318 | // Walk the colored neighbors of the "at_risk" candidate |
1319 | // Choose a color which is both legal and already taken by a neighbor |
1320 | // of the "at_risk" candidate in order to improve the chances of the |
1321 | // "at_risk" candidate of coloring |
1322 | IndexSetIterator elements(_ifg->neighbors(risk_lrg)); |
1323 | uint datum; |
1324 | while ((datum = elements.next()) != 0) { |
1325 | OptoReg::Name reg = lrgs(datum).reg(); |
1326 | // If this LRG's register is legal for us, choose it |
1327 | if (is_legal_reg(lrg, reg, chunk)) |
1328 | return reg; |
1329 | } |
1330 | } |
1331 | |
1332 | uint copy_lrg = _lrg_map.find(lrg._copy_bias); |
1333 | if( copy_lrg != 0 ) { |
1334 | // If he has a color, |
1335 | if( !(*(_ifg->_yanked))[copy_lrg] ) { |
1336 | OptoReg::Name reg = lrgs(copy_lrg).reg(); |
1337 | // And it is legal for you, |
1338 | if (is_legal_reg(lrg, reg, chunk)) |
1339 | return reg; |
1340 | } else if( chunk == 0 ) { |
1341 | // Choose a color which is legal for him |
1342 | RegMask tempmask = lrg.mask(); |
1343 | tempmask.AND(lrgs(copy_lrg).mask()); |
1344 | tempmask.clear_to_sets(lrg.num_regs()); |
1345 | OptoReg::Name reg = tempmask.find_first_set(lrg.num_regs()); |
1346 | if (OptoReg::is_valid(reg)) |
1347 | return reg; |
1348 | } |
1349 | } |
1350 | |
1351 | // If no bias info exists, just go with the register selection ordering |
1352 | if (lrg._is_vector || lrg.num_regs() == 2) { |
1353 | // Find an aligned set |
1354 | return OptoReg::add(lrg.mask().find_first_set(lrg.num_regs()),chunk); |
1355 | } |
1356 | |
1357 | // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate |
1358 | // copy removal to remove many more copies, by preventing a just-assigned |
1359 | // register from being repeatedly assigned. |
1360 | OptoReg::Name reg = lrg.mask().find_first_elem(); |
1361 | if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { |
1362 | // This 'Remove; find; Insert' idiom is an expensive way to find the |
1363 | // SECOND element in the mask. |
1364 | lrg.Remove(reg); |
1365 | OptoReg::Name reg2 = lrg.mask().find_first_elem(); |
1366 | lrg.Insert(reg); |
1367 | if( OptoReg::is_reg(reg2)) |
1368 | reg = reg2; |
1369 | } |
1370 | return OptoReg::add( reg, chunk ); |
1371 | } |
1372 | |
1373 | // Choose a color in the current chunk |
1374 | OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { |
1375 | 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)" ); |
1376 | 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)" ); |
1377 | |
1378 | if( lrg.num_regs() == 1 || // Common Case |
1379 | !lrg._fat_proj ) // Aligned+adjacent pairs ok |
1380 | // Use a heuristic to "bias" the color choice |
1381 | return bias_color(lrg, chunk); |
1382 | |
1383 | assert(!lrg._is_vector, "should be not vector here" ); |
1384 | assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); |
1385 | |
1386 | // Fat-proj case or misaligned double argument. |
1387 | assert(lrg.compute_mask_size() == lrg.num_regs() || |
1388 | lrg.num_regs() == 2,"fat projs exactly color" ); |
1389 | assert( !chunk, "always color in 1st chunk" ); |
1390 | // Return the highest element in the set. |
1391 | return lrg.mask().find_last_elem(); |
1392 | } |
1393 | |
1394 | // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted |
1395 | // in reverse order of removal. As long as nothing of hi-degree was yanked, |
1396 | // everything going back is guaranteed a color. Select that color. If some |
1397 | // hi-degree LRG cannot get a color then we record that we must spill. |
1398 | uint PhaseChaitin::Select( ) { |
1399 | Compile::TracePhase tp("chaitinSelect" , &timers[_t_chaitinSelect]); |
1400 | |
1401 | uint spill_reg = LRG::SPILL_REG; |
1402 | _max_reg = OptoReg::Name(0); // Past max register used |
1403 | while( _simplified ) { |
1404 | // Pull next LRG from the simplified list - in reverse order of removal |
1405 | uint lidx = _simplified; |
1406 | LRG *lrg = &lrgs(lidx); |
1407 | _simplified = lrg->_next; |
1408 | |
1409 | |
1410 | #ifndef PRODUCT |
1411 | if (trace_spilling()) { |
1412 | ttyLocker ttyl; |
1413 | tty->print_cr("L%d selecting degree %d degrees_of_freedom %d" , lidx, lrg->degree(), |
1414 | lrg->degrees_of_freedom()); |
1415 | lrg->dump(); |
1416 | } |
1417 | #endif |
1418 | |
1419 | // Re-insert into the IFG |
1420 | _ifg->re_insert(lidx); |
1421 | if( !lrg->alive() ) continue; |
1422 | // capture allstackedness flag before mask is hacked |
1423 | const int is_allstack = lrg->mask().is_AllStack(); |
1424 | |
1425 | // Yeah, yeah, yeah, I know, I know. I can refactor this |
1426 | // to avoid the GOTO, although the refactored code will not |
1427 | // be much clearer. We arrive here IFF we have a stack-based |
1428 | // live range that cannot color in the current chunk, and it |
1429 | // has to move into the next free stack chunk. |
1430 | int chunk = 0; // Current chunk is first chunk |
1431 | retry_next_chunk: |
1432 | |
1433 | // Remove neighbor colors |
1434 | IndexSet *s = _ifg->neighbors(lidx); |
1435 | |
1436 | debug_only(RegMask orig_mask = lrg->mask();) |
1437 | IndexSetIterator elements(s); |
1438 | uint neighbor; |
1439 | while ((neighbor = elements.next()) != 0) { |
1440 | // Note that neighbor might be a spill_reg. In this case, exclusion |
1441 | // of its color will be a no-op, since the spill_reg chunk is in outer |
1442 | // space. Also, if neighbor is in a different chunk, this exclusion |
1443 | // will be a no-op. (Later on, if lrg runs out of possible colors in |
1444 | // its chunk, a new chunk of color may be tried, in which case |
1445 | // examination of neighbors is started again, at retry_next_chunk.) |
1446 | LRG &nlrg = lrgs(neighbor); |
1447 | OptoReg::Name nreg = nlrg.reg(); |
1448 | // Only subtract masks in the same chunk |
1449 | if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { |
1450 | #ifndef PRODUCT |
1451 | uint size = lrg->mask().Size(); |
1452 | RegMask rm = lrg->mask(); |
1453 | #endif |
1454 | lrg->SUBTRACT(nlrg.mask()); |
1455 | #ifndef PRODUCT |
1456 | if (trace_spilling() && lrg->mask().Size() != size) { |
1457 | ttyLocker ttyl; |
1458 | tty->print("L%d " , lidx); |
1459 | rm.dump(); |
1460 | tty->print(" intersected L%d " , neighbor); |
1461 | nlrg.mask().dump(); |
1462 | tty->print(" removed " ); |
1463 | rm.SUBTRACT(lrg->mask()); |
1464 | rm.dump(); |
1465 | tty->print(" leaving " ); |
1466 | lrg->mask().dump(); |
1467 | tty->cr(); |
1468 | } |
1469 | #endif |
1470 | } |
1471 | } |
1472 | //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); |
1473 | // Aligned pairs need aligned masks |
1474 | assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity" ); |
1475 | if (lrg->num_regs() > 1 && !lrg->_fat_proj) { |
1476 | lrg->clear_to_sets(); |
1477 | } |
1478 | |
1479 | // Check if a color is available and if so pick the color |
1480 | OptoReg::Name reg = choose_color( *lrg, chunk ); |
1481 | #ifdef SPARC |
1482 | debug_only(lrg->compute_set_mask_size()); |
1483 | assert(lrg->num_regs() < 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned" ); |
1484 | #endif |
1485 | |
1486 | //--------------- |
1487 | // If we fail to color and the AllStack flag is set, trigger |
1488 | // a chunk-rollover event |
1489 | if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { |
1490 | // Bump register mask up to next stack chunk |
1491 | chunk += RegMask::CHUNK_SIZE; |
1492 | lrg->Set_All(); |
1493 | |
1494 | goto retry_next_chunk; |
1495 | } |
1496 | |
1497 | //--------------- |
1498 | // Did we get a color? |
1499 | else if( OptoReg::is_valid(reg)) { |
1500 | #ifndef PRODUCT |
1501 | RegMask avail_rm = lrg->mask(); |
1502 | #endif |
1503 | |
1504 | // Record selected register |
1505 | lrg->set_reg(reg); |
1506 | |
1507 | if( reg >= _max_reg ) // Compute max register limit |
1508 | _max_reg = OptoReg::add(reg,1); |
1509 | // Fold reg back into normal space |
1510 | reg = OptoReg::add(reg,-chunk); |
1511 | |
1512 | // If the live range is not bound, then we actually had some choices |
1513 | // to make. In this case, the mask has more bits in it than the colors |
1514 | // chosen. Restrict the mask to just what was picked. |
1515 | int n_regs = lrg->num_regs(); |
1516 | assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity" ); |
1517 | if (n_regs == 1 || !lrg->_fat_proj) { |
1518 | assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity" ); |
1519 | lrg->Clear(); // Clear the mask |
1520 | lrg->Insert(reg); // Set regmask to match selected reg |
1521 | // For vectors and pairs, also insert the low bit of the pair |
1522 | for (int i = 1; i < n_regs; i++) |
1523 | lrg->Insert(OptoReg::add(reg,-i)); |
1524 | lrg->set_mask_size(n_regs); |
1525 | } else { // Else fatproj |
1526 | // mask must be equal to fatproj bits, by definition |
1527 | } |
1528 | #ifndef PRODUCT |
1529 | if (trace_spilling()) { |
1530 | ttyLocker ttyl; |
1531 | tty->print("L%d selected " , lidx); |
1532 | lrg->mask().dump(); |
1533 | tty->print(" from " ); |
1534 | avail_rm.dump(); |
1535 | tty->cr(); |
1536 | } |
1537 | #endif |
1538 | // Note that reg is the highest-numbered register in the newly-bound mask. |
1539 | } // end color available case |
1540 | |
1541 | //--------------- |
1542 | // Live range is live and no colors available |
1543 | else { |
1544 | assert( lrg->alive(), "" ); |
1545 | assert( !lrg->_fat_proj || lrg->is_multidef() || |
1546 | lrg->_def->outcnt() > 0, "fat_proj cannot spill" ); |
1547 | assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); |
1548 | |
1549 | // Assign the special spillreg register |
1550 | lrg->set_reg(OptoReg::Name(spill_reg++)); |
1551 | // Do not empty the regmask; leave mask_size lying around |
1552 | // for use during Spilling |
1553 | #ifndef PRODUCT |
1554 | if( trace_spilling() ) { |
1555 | ttyLocker ttyl; |
1556 | tty->print("L%d spilling with neighbors: " , lidx); |
1557 | s->dump(); |
1558 | debug_only(tty->print(" original mask: " )); |
1559 | debug_only(orig_mask.dump()); |
1560 | dump_lrg(lidx); |
1561 | } |
1562 | #endif |
1563 | } // end spill case |
1564 | |
1565 | } |
1566 | |
1567 | return spill_reg-LRG::SPILL_REG; // Return number of spills |
1568 | } |
1569 | |
1570 | // Set the 'spilled_once' or 'spilled_twice' flag on a node. |
1571 | void PhaseChaitin::set_was_spilled( Node *n ) { |
1572 | if( _spilled_once.test_set(n->_idx) ) |
1573 | _spilled_twice.set(n->_idx); |
1574 | } |
1575 | |
1576 | // Convert Ideal spill instructions into proper FramePtr + offset Loads and |
1577 | // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. |
1578 | void PhaseChaitin::fixup_spills() { |
1579 | // This function does only cisc spill work. |
1580 | if( !UseCISCSpill ) return; |
1581 | |
1582 | Compile::TracePhase tp("fixupSpills" , &timers[_t_fixupSpills]); |
1583 | |
1584 | // Grab the Frame Pointer |
1585 | Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr); |
1586 | |
1587 | // For all blocks |
1588 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
1589 | Block* block = _cfg.get_block(i); |
1590 | |
1591 | // For all instructions in block |
1592 | uint last_inst = block->end_idx(); |
1593 | for (uint j = 1; j <= last_inst; j++) { |
1594 | Node* n = block->get_node(j); |
1595 | |
1596 | // Dead instruction??? |
1597 | assert( n->outcnt() != 0 ||// Nothing dead after post alloc |
1598 | C->top() == n || // Or the random TOP node |
1599 | n->is_Proj(), // Or a fat-proj kill node |
1600 | "No dead instructions after post-alloc" ); |
1601 | |
1602 | int inp = n->cisc_operand(); |
1603 | if( inp != AdlcVMDeps::Not_cisc_spillable ) { |
1604 | // Convert operand number to edge index number |
1605 | MachNode *mach = n->as_Mach(); |
1606 | inp = mach->operand_index(inp); |
1607 | Node *src = n->in(inp); // Value to load or store |
1608 | LRG &lrg_cisc = lrgs(_lrg_map.find_const(src)); |
1609 | OptoReg::Name src_reg = lrg_cisc.reg(); |
1610 | // Doubles record the HIGH register of an adjacent pair. |
1611 | src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); |
1612 | if( OptoReg::is_stack(src_reg) ) { // If input is on stack |
1613 | // This is a CISC Spill, get stack offset and construct new node |
1614 | #ifndef PRODUCT |
1615 | if( TraceCISCSpill ) { |
1616 | tty->print(" reg-instr: " ); |
1617 | n->dump(); |
1618 | } |
1619 | #endif |
1620 | int stk_offset = reg2offset(src_reg); |
1621 | // Bailout if we might exceed node limit when spilling this instruction |
1622 | C->check_node_count(0, "out of nodes fixing spills" ); |
1623 | if (C->failing()) return; |
1624 | // Transform node |
1625 | MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach(); |
1626 | cisc->set_req(inp,fp); // Base register is frame pointer |
1627 | if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { |
1628 | assert( cisc->oper_input_base() == 2, "Only adding one edge" ); |
1629 | cisc->ins_req(1,src); // Requires a memory edge |
1630 | } |
1631 | block->map_node(cisc, j); // Insert into basic block |
1632 | n->subsume_by(cisc, C); // Correct graph |
1633 | // |
1634 | ++_used_cisc_instructions; |
1635 | #ifndef PRODUCT |
1636 | if( TraceCISCSpill ) { |
1637 | tty->print(" cisc-instr: " ); |
1638 | cisc->dump(); |
1639 | } |
1640 | #endif |
1641 | } else { |
1642 | #ifndef PRODUCT |
1643 | if( TraceCISCSpill ) { |
1644 | tty->print(" using reg-instr: " ); |
1645 | n->dump(); |
1646 | } |
1647 | #endif |
1648 | ++_unused_cisc_instructions; // input can be on stack |
1649 | } |
1650 | } |
1651 | |
1652 | } // End of for all instructions |
1653 | |
1654 | } // End of for all blocks |
1655 | } |
1656 | |
1657 | // Helper to stretch above; recursively discover the base Node for a |
1658 | // given derived Node. Easy for AddP-related machine nodes, but needs |
1659 | // to be recursive for derived Phis. |
1660 | Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { |
1661 | // See if already computed; if so return it |
1662 | if( derived_base_map[derived->_idx] ) |
1663 | return derived_base_map[derived->_idx]; |
1664 | |
1665 | // See if this happens to be a base. |
1666 | // NOTE: we use TypePtr instead of TypeOopPtr because we can have |
1667 | // pointers derived from NULL! These are always along paths that |
1668 | // can't happen at run-time but the optimizer cannot deduce it so |
1669 | // we have to handle it gracefully. |
1670 | assert(!derived->bottom_type()->isa_narrowoop() || |
1671 | derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity" ); |
1672 | const TypePtr *tj = derived->bottom_type()->isa_ptr(); |
1673 | // If its an OOP with a non-zero offset, then it is derived. |
1674 | if( tj == NULL || tj->_offset == 0 ) { |
1675 | derived_base_map[derived->_idx] = derived; |
1676 | return derived; |
1677 | } |
1678 | // Derived is NULL+offset? Base is NULL! |
1679 | if( derived->is_Con() ) { |
1680 | Node *base = _matcher.mach_null(); |
1681 | assert(base != NULL, "sanity" ); |
1682 | if (base->in(0) == NULL) { |
1683 | // Initialize it once and make it shared: |
1684 | // set control to _root and place it into Start block |
1685 | // (where top() node is placed). |
1686 | base->init_req(0, _cfg.get_root_node()); |
1687 | Block *startb = _cfg.get_block_for_node(C->top()); |
1688 | uint node_pos = startb->find_node(C->top()); |
1689 | startb->insert_node(base, node_pos); |
1690 | _cfg.map_node_to_block(base, startb); |
1691 | assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet" ); |
1692 | |
1693 | // The loadConP0 might have projection nodes depending on architecture |
1694 | // Add the projection nodes to the CFG |
1695 | for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) { |
1696 | Node* use = base->fast_out(i); |
1697 | if (use->is_MachProj()) { |
1698 | startb->insert_node(use, ++node_pos); |
1699 | _cfg.map_node_to_block(use, startb); |
1700 | new_lrg(use, maxlrg++); |
1701 | } |
1702 | } |
1703 | } |
1704 | if (_lrg_map.live_range_id(base) == 0) { |
1705 | new_lrg(base, maxlrg++); |
1706 | } |
1707 | 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" ); |
1708 | derived_base_map[derived->_idx] = base; |
1709 | return base; |
1710 | } |
1711 | |
1712 | // Check for AddP-related opcodes |
1713 | if (!derived->is_Phi()) { |
1714 | assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s" , derived->Name()); |
1715 | Node *base = derived->in(AddPNode::Base); |
1716 | derived_base_map[derived->_idx] = base; |
1717 | return base; |
1718 | } |
1719 | |
1720 | // Recursively find bases for Phis. |
1721 | // First check to see if we can avoid a base Phi here. |
1722 | Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); |
1723 | uint i; |
1724 | for( i = 2; i < derived->req(); i++ ) |
1725 | if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) |
1726 | break; |
1727 | // Went to the end without finding any different bases? |
1728 | if( i == derived->req() ) { // No need for a base Phi here |
1729 | derived_base_map[derived->_idx] = base; |
1730 | return base; |
1731 | } |
1732 | |
1733 | // Now we see we need a base-Phi here to merge the bases |
1734 | const Type *t = base->bottom_type(); |
1735 | base = new PhiNode( derived->in(0), t ); |
1736 | for( i = 1; i < derived->req(); i++ ) { |
1737 | base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); |
1738 | t = t->meet(base->in(i)->bottom_type()); |
1739 | } |
1740 | base->as_Phi()->set_type(t); |
1741 | |
1742 | // Search the current block for an existing base-Phi |
1743 | Block *b = _cfg.get_block_for_node(derived); |
1744 | for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi |
1745 | Node *phi = b->get_node(i); |
1746 | if( !phi->is_Phi() ) { // Found end of Phis with no match? |
1747 | b->insert_node(base, i); // Must insert created Phi here as base |
1748 | _cfg.map_node_to_block(base, b); |
1749 | new_lrg(base,maxlrg++); |
1750 | break; |
1751 | } |
1752 | // See if Phi matches. |
1753 | uint j; |
1754 | for( j = 1; j < base->req(); j++ ) |
1755 | if( phi->in(j) != base->in(j) && |
1756 | !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs |
1757 | break; |
1758 | if( j == base->req() ) { // All inputs match? |
1759 | base = phi; // Then use existing 'phi' and drop 'base' |
1760 | break; |
1761 | } |
1762 | } |
1763 | |
1764 | |
1765 | // Cache info for later passes |
1766 | derived_base_map[derived->_idx] = base; |
1767 | return base; |
1768 | } |
1769 | |
1770 | // At each Safepoint, insert extra debug edges for each pair of derived value/ |
1771 | // base pointer that is live across the Safepoint for oopmap building. The |
1772 | // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the |
1773 | // required edge set. |
1774 | bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) { |
1775 | int must_recompute_live = false; |
1776 | uint maxlrg = _lrg_map.max_lrg_id(); |
1777 | Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); |
1778 | memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); |
1779 | |
1780 | // For all blocks in RPO do... |
1781 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
1782 | Block* block = _cfg.get_block(i); |
1783 | // Note use of deep-copy constructor. I cannot hammer the original |
1784 | // liveout bits, because they are needed by the following coalesce pass. |
1785 | IndexSet liveout(_live->live(block)); |
1786 | |
1787 | for (uint j = block->end_idx() + 1; j > 1; j--) { |
1788 | Node* n = block->get_node(j - 1); |
1789 | |
1790 | // Pre-split compares of loop-phis. Loop-phis form a cycle we would |
1791 | // like to see in the same register. Compare uses the loop-phi and so |
1792 | // extends its live range BUT cannot be part of the cycle. If this |
1793 | // extended live range overlaps with the update of the loop-phi value |
1794 | // we need both alive at the same time -- which requires at least 1 |
1795 | // copy. But because Intel has only 2-address registers we end up with |
1796 | // at least 2 copies, one before the loop-phi update instruction and |
1797 | // one after. Instead we split the input to the compare just after the |
1798 | // phi. |
1799 | if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { |
1800 | Node *phi = n->in(1); |
1801 | if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { |
1802 | Block *phi_block = _cfg.get_block_for_node(phi); |
1803 | if (_cfg.get_block_for_node(phi_block->pred(2)) == block) { |
1804 | const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; |
1805 | Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask); |
1806 | insert_proj( phi_block, 1, spill, maxlrg++ ); |
1807 | n->set_req(1,spill); |
1808 | must_recompute_live = true; |
1809 | } |
1810 | } |
1811 | } |
1812 | |
1813 | // Get value being defined |
1814 | uint lidx = _lrg_map.live_range_id(n); |
1815 | // Ignore the occasional brand-new live range |
1816 | if (lidx && lidx < _lrg_map.max_lrg_id()) { |
1817 | // Remove from live-out set |
1818 | liveout.remove(lidx); |
1819 | |
1820 | // Copies do not define a new value and so do not interfere. |
1821 | // Remove the copies source from the liveout set before interfering. |
1822 | uint idx = n->is_Copy(); |
1823 | if (idx) { |
1824 | liveout.remove(_lrg_map.live_range_id(n->in(idx))); |
1825 | } |
1826 | } |
1827 | |
1828 | // Found a safepoint? |
1829 | JVMState *jvms = n->jvms(); |
1830 | if( jvms ) { |
1831 | // Now scan for a live derived pointer |
1832 | IndexSetIterator elements(&liveout); |
1833 | uint neighbor; |
1834 | while ((neighbor = elements.next()) != 0) { |
1835 | // Find reaching DEF for base and derived values |
1836 | // This works because we are still in SSA during this call. |
1837 | Node *derived = lrgs(neighbor)._def; |
1838 | const TypePtr *tj = derived->bottom_type()->isa_ptr(); |
1839 | assert(!derived->bottom_type()->isa_narrowoop() || |
1840 | derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity" ); |
1841 | // If its an OOP with a non-zero offset, then it is derived. |
1842 | if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { |
1843 | Node *base = find_base_for_derived(derived_base_map, derived, maxlrg); |
1844 | assert(base->_idx < _lrg_map.size(), "" ); |
1845 | // Add reaching DEFs of derived pointer and base pointer as a |
1846 | // pair of inputs |
1847 | n->add_req(derived); |
1848 | n->add_req(base); |
1849 | |
1850 | // See if the base pointer is already live to this point. |
1851 | // Since I'm working on the SSA form, live-ness amounts to |
1852 | // reaching def's. So if I find the base's live range then |
1853 | // I know the base's def reaches here. |
1854 | if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or |
1855 | !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND |
1856 | (_lrg_map.live_range_id(base) > 0) && // not a constant |
1857 | _cfg.get_block_for_node(base) != block) { // base not def'd in blk) |
1858 | // Base pointer is not currently live. Since I stretched |
1859 | // the base pointer to here and it crosses basic-block |
1860 | // boundaries, the global live info is now incorrect. |
1861 | // Recompute live. |
1862 | must_recompute_live = true; |
1863 | } // End of if base pointer is not live to debug info |
1864 | } |
1865 | } // End of scan all live data for derived ptrs crossing GC point |
1866 | } // End of if found a GC point |
1867 | |
1868 | // Make all inputs live |
1869 | if (!n->is_Phi()) { // Phi function uses come from prior block |
1870 | for (uint k = 1; k < n->req(); k++) { |
1871 | uint lidx = _lrg_map.live_range_id(n->in(k)); |
1872 | if (lidx < _lrg_map.max_lrg_id()) { |
1873 | liveout.insert(lidx); |
1874 | } |
1875 | } |
1876 | } |
1877 | |
1878 | } // End of forall instructions in block |
1879 | liveout.clear(); // Free the memory used by liveout. |
1880 | |
1881 | } // End of forall blocks |
1882 | _lrg_map.set_max_lrg_id(maxlrg); |
1883 | |
1884 | // If I created a new live range I need to recompute live |
1885 | if (maxlrg != _ifg->_maxlrg) { |
1886 | must_recompute_live = true; |
1887 | } |
1888 | |
1889 | return must_recompute_live != 0; |
1890 | } |
1891 | |
1892 | // Extend the node to LRG mapping |
1893 | |
1894 | void PhaseChaitin::add_reference(const Node *node, const Node *old_node) { |
1895 | _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node)); |
1896 | } |
1897 | |
1898 | #ifndef PRODUCT |
1899 | void PhaseChaitin::dump(const Node *n) const { |
1900 | uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0; |
1901 | tty->print("L%d" ,r); |
1902 | if (r && n->Opcode() != Op_Phi) { |
1903 | if( _node_regs ) { // Got a post-allocation copy of allocation? |
1904 | tty->print("[" ); |
1905 | OptoReg::Name second = get_reg_second(n); |
1906 | if( OptoReg::is_valid(second) ) { |
1907 | if( OptoReg::is_reg(second) ) |
1908 | tty->print("%s:" ,Matcher::regName[second]); |
1909 | else |
1910 | tty->print("%s+%d:" ,OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); |
1911 | } |
1912 | OptoReg::Name first = get_reg_first(n); |
1913 | if( OptoReg::is_reg(first) ) |
1914 | tty->print("%s]" ,Matcher::regName[first]); |
1915 | else |
1916 | tty->print("%s+%d]" ,OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); |
1917 | } else |
1918 | n->out_RegMask().dump(); |
1919 | } |
1920 | tty->print("/N%d\t" ,n->_idx); |
1921 | tty->print("%s === " , n->Name()); |
1922 | uint k; |
1923 | for (k = 0; k < n->req(); k++) { |
1924 | Node *m = n->in(k); |
1925 | if (!m) { |
1926 | tty->print("_ " ); |
1927 | } |
1928 | else { |
1929 | uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; |
1930 | tty->print("L%d" ,r); |
1931 | // Data MultiNode's can have projections with no real registers. |
1932 | // Don't die while dumping them. |
1933 | int op = n->Opcode(); |
1934 | if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { |
1935 | if( _node_regs ) { |
1936 | tty->print("[" ); |
1937 | OptoReg::Name second = get_reg_second(n->in(k)); |
1938 | if( OptoReg::is_valid(second) ) { |
1939 | if( OptoReg::is_reg(second) ) |
1940 | tty->print("%s:" ,Matcher::regName[second]); |
1941 | else |
1942 | tty->print("%s+%d:" ,OptoReg::regname(OptoReg::c_frame_pointer), |
1943 | reg2offset_unchecked(second)); |
1944 | } |
1945 | OptoReg::Name first = get_reg_first(n->in(k)); |
1946 | if( OptoReg::is_reg(first) ) |
1947 | tty->print("%s]" ,Matcher::regName[first]); |
1948 | else |
1949 | tty->print("%s+%d]" ,OptoReg::regname(OptoReg::c_frame_pointer), |
1950 | reg2offset_unchecked(first)); |
1951 | } else |
1952 | n->in_RegMask(k).dump(); |
1953 | } |
1954 | tty->print("/N%d " ,m->_idx); |
1955 | } |
1956 | } |
1957 | if( k < n->len() && n->in(k) ) tty->print("| " ); |
1958 | for( ; k < n->len(); k++ ) { |
1959 | Node *m = n->in(k); |
1960 | if(!m) { |
1961 | break; |
1962 | } |
1963 | uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; |
1964 | tty->print("L%d" ,r); |
1965 | tty->print("/N%d " ,m->_idx); |
1966 | } |
1967 | if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); |
1968 | else n->dump_spec(tty); |
1969 | if( _spilled_once.test(n->_idx ) ) { |
1970 | tty->print(" Spill_1" ); |
1971 | if( _spilled_twice.test(n->_idx ) ) |
1972 | tty->print(" Spill_2" ); |
1973 | } |
1974 | tty->print("\n" ); |
1975 | } |
1976 | |
1977 | void PhaseChaitin::dump(const Block *b) const { |
1978 | b->dump_head(&_cfg); |
1979 | |
1980 | // For all instructions |
1981 | for( uint j = 0; j < b->number_of_nodes(); j++ ) |
1982 | dump(b->get_node(j)); |
1983 | // Print live-out info at end of block |
1984 | if( _live ) { |
1985 | tty->print("Liveout: " ); |
1986 | IndexSet *live = _live->live(b); |
1987 | IndexSetIterator elements(live); |
1988 | tty->print("{" ); |
1989 | uint i; |
1990 | while ((i = elements.next()) != 0) { |
1991 | tty->print("L%d " , _lrg_map.find_const(i)); |
1992 | } |
1993 | tty->print_cr("}" ); |
1994 | } |
1995 | tty->print("\n" ); |
1996 | } |
1997 | |
1998 | void PhaseChaitin::dump() const { |
1999 | tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n" , |
2000 | _matcher._new_SP, _framesize ); |
2001 | |
2002 | // For all blocks |
2003 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
2004 | dump(_cfg.get_block(i)); |
2005 | } |
2006 | // End of per-block dump |
2007 | tty->print("\n" ); |
2008 | |
2009 | if (!_ifg) { |
2010 | tty->print("(No IFG.)\n" ); |
2011 | return; |
2012 | } |
2013 | |
2014 | // Dump LRG array |
2015 | tty->print("--- Live RanGe Array ---\n" ); |
2016 | for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) { |
2017 | tty->print("L%d: " ,i2); |
2018 | if (i2 < _ifg->_maxlrg) { |
2019 | lrgs(i2).dump(); |
2020 | } |
2021 | else { |
2022 | tty->print_cr("new LRG" ); |
2023 | } |
2024 | } |
2025 | tty->cr(); |
2026 | |
2027 | // Dump lo-degree list |
2028 | tty->print("Lo degree: " ); |
2029 | for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) |
2030 | tty->print("L%d " ,i3); |
2031 | tty->cr(); |
2032 | |
2033 | // Dump lo-stk-degree list |
2034 | tty->print("Lo stk degree: " ); |
2035 | for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) |
2036 | tty->print("L%d " ,i4); |
2037 | tty->cr(); |
2038 | |
2039 | // Dump lo-degree list |
2040 | tty->print("Hi degree: " ); |
2041 | for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) |
2042 | tty->print("L%d " ,i5); |
2043 | tty->cr(); |
2044 | } |
2045 | |
2046 | void PhaseChaitin::dump_degree_lists() const { |
2047 | // Dump lo-degree list |
2048 | tty->print("Lo degree: " ); |
2049 | for( uint i = _lo_degree; i; i = lrgs(i)._next ) |
2050 | tty->print("L%d " ,i); |
2051 | tty->cr(); |
2052 | |
2053 | // Dump lo-stk-degree list |
2054 | tty->print("Lo stk degree: " ); |
2055 | for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) |
2056 | tty->print("L%d " ,i2); |
2057 | tty->cr(); |
2058 | |
2059 | // Dump lo-degree list |
2060 | tty->print("Hi degree: " ); |
2061 | for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) |
2062 | tty->print("L%d " ,i3); |
2063 | tty->cr(); |
2064 | } |
2065 | |
2066 | void PhaseChaitin::dump_simplified() const { |
2067 | tty->print("Simplified: " ); |
2068 | for( uint i = _simplified; i; i = lrgs(i)._next ) |
2069 | tty->print("L%d " ,i); |
2070 | tty->cr(); |
2071 | } |
2072 | |
2073 | static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { |
2074 | if ((int)reg < 0) |
2075 | sprintf(buf, "<OptoReg::%d>" , (int)reg); |
2076 | else if (OptoReg::is_reg(reg)) |
2077 | strcpy(buf, Matcher::regName[reg]); |
2078 | else |
2079 | sprintf(buf,"%s + #%d" ,OptoReg::regname(OptoReg::c_frame_pointer), |
2080 | pc->reg2offset(reg)); |
2081 | return buf+strlen(buf); |
2082 | } |
2083 | |
2084 | // Dump a register name into a buffer. Be intelligent if we get called |
2085 | // before allocation is complete. |
2086 | char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { |
2087 | if( _node_regs ) { |
2088 | // Post allocation, use direct mappings, no LRG info available |
2089 | print_reg( get_reg_first(n), this, buf ); |
2090 | } else { |
2091 | uint lidx = _lrg_map.find_const(n); // Grab LRG number |
2092 | if( !_ifg ) { |
2093 | sprintf(buf,"L%d" ,lidx); // No register binding yet |
2094 | } else if( !lidx ) { // Special, not allocated value |
2095 | strcpy(buf,"Special" ); |
2096 | } else { |
2097 | if (lrgs(lidx)._is_vector) { |
2098 | if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs())) |
2099 | print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register |
2100 | else |
2101 | sprintf(buf,"L%d" ,lidx); // No register binding yet |
2102 | } else if( (lrgs(lidx).num_regs() == 1) |
2103 | ? lrgs(lidx).mask().is_bound1() |
2104 | : lrgs(lidx).mask().is_bound_pair() ) { |
2105 | // Hah! We have a bound machine register |
2106 | print_reg( lrgs(lidx).reg(), this, buf ); |
2107 | } else { |
2108 | sprintf(buf,"L%d" ,lidx); // No register binding yet |
2109 | } |
2110 | } |
2111 | } |
2112 | return buf+strlen(buf); |
2113 | } |
2114 | |
2115 | void PhaseChaitin::dump_for_spill_split_recycle() const { |
2116 | if( WizardMode && (PrintCompilation || PrintOpto) ) { |
2117 | // Display which live ranges need to be split and the allocator's state |
2118 | tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges" , _trip_cnt); |
2119 | for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) { |
2120 | if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { |
2121 | tty->print("L%d: " , bidx); |
2122 | lrgs(bidx).dump(); |
2123 | } |
2124 | } |
2125 | tty->cr(); |
2126 | dump(); |
2127 | } |
2128 | } |
2129 | |
2130 | void PhaseChaitin::dump_frame() const { |
2131 | const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); |
2132 | const TypeTuple *domain = C->tf()->domain(); |
2133 | const int argcnt = domain->cnt() - TypeFunc::Parms; |
2134 | |
2135 | // Incoming arguments in registers dump |
2136 | for( int k = 0; k < argcnt; k++ ) { |
2137 | OptoReg::Name parmreg = _matcher._parm_regs[k].first(); |
2138 | if( OptoReg::is_reg(parmreg)) { |
2139 | const char *reg_name = OptoReg::regname(parmreg); |
2140 | tty->print("#r%3.3d %s" , parmreg, reg_name); |
2141 | parmreg = _matcher._parm_regs[k].second(); |
2142 | if( OptoReg::is_reg(parmreg)) { |
2143 | tty->print(":%s" , OptoReg::regname(parmreg)); |
2144 | } |
2145 | tty->print(" : parm %d: " , k); |
2146 | domain->field_at(k + TypeFunc::Parms)->dump(); |
2147 | tty->cr(); |
2148 | } |
2149 | } |
2150 | |
2151 | // Check for un-owned padding above incoming args |
2152 | OptoReg::Name reg = _matcher._new_SP; |
2153 | if( reg > _matcher._in_arg_limit ) { |
2154 | reg = OptoReg::add(reg, -1); |
2155 | tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER" , reg, fp, reg2offset_unchecked(reg)); |
2156 | } |
2157 | |
2158 | // Incoming argument area dump |
2159 | OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); |
2160 | while( reg > begin_in_arg ) { |
2161 | reg = OptoReg::add(reg, -1); |
2162 | tty->print("#r%3.3d %s+%2d: " ,reg,fp,reg2offset_unchecked(reg)); |
2163 | int j; |
2164 | for( j = 0; j < argcnt; j++) { |
2165 | if( _matcher._parm_regs[j].first() == reg || |
2166 | _matcher._parm_regs[j].second() == reg ) { |
2167 | tty->print("parm %d: " ,j); |
2168 | domain->field_at(j + TypeFunc::Parms)->dump(); |
2169 | tty->cr(); |
2170 | break; |
2171 | } |
2172 | } |
2173 | if( j >= argcnt ) |
2174 | tty->print_cr("HOLE, owned by SELF" ); |
2175 | } |
2176 | |
2177 | // Old outgoing preserve area |
2178 | while( reg > _matcher._old_SP ) { |
2179 | reg = OptoReg::add(reg, -1); |
2180 | tty->print_cr("#r%3.3d %s+%2d: old out preserve" ,reg,fp,reg2offset_unchecked(reg)); |
2181 | } |
2182 | |
2183 | // Old SP |
2184 | tty->print_cr("# -- Old %s -- Framesize: %d --" ,fp, |
2185 | reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); |
2186 | |
2187 | // Preserve area dump |
2188 | int fixed_slots = C->fixed_slots(); |
2189 | OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots()); |
2190 | OptoReg::Name return_addr = _matcher.return_addr(); |
2191 | |
2192 | reg = OptoReg::add(reg, -1); |
2193 | while (OptoReg::is_stack(reg)) { |
2194 | tty->print("#r%3.3d %s+%2d: " ,reg,fp,reg2offset_unchecked(reg)); |
2195 | if (return_addr == reg) { |
2196 | tty->print_cr("return address" ); |
2197 | } else if (reg >= begin_in_preserve) { |
2198 | // Preserved slots are present on x86 |
2199 | if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word)) |
2200 | tty->print_cr("saved fp register" ); |
2201 | else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) && |
2202 | VerifyStackAtCalls) |
2203 | tty->print_cr("0xBADB100D +VerifyStackAtCalls" ); |
2204 | else |
2205 | tty->print_cr("in_preserve" ); |
2206 | } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) { |
2207 | tty->print_cr("Fixed slot %d" , OptoReg::reg2stack(reg)); |
2208 | } else { |
2209 | tty->print_cr("pad2, stack alignment" ); |
2210 | } |
2211 | reg = OptoReg::add(reg, -1); |
2212 | } |
2213 | |
2214 | // Spill area dump |
2215 | reg = OptoReg::add(_matcher._new_SP, _framesize ); |
2216 | while( reg > _matcher._out_arg_limit ) { |
2217 | reg = OptoReg::add(reg, -1); |
2218 | tty->print_cr("#r%3.3d %s+%2d: spill" ,reg,fp,reg2offset_unchecked(reg)); |
2219 | } |
2220 | |
2221 | // Outgoing argument area dump |
2222 | while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { |
2223 | reg = OptoReg::add(reg, -1); |
2224 | tty->print_cr("#r%3.3d %s+%2d: outgoing argument" ,reg,fp,reg2offset_unchecked(reg)); |
2225 | } |
2226 | |
2227 | // Outgoing new preserve area |
2228 | while( reg > _matcher._new_SP ) { |
2229 | reg = OptoReg::add(reg, -1); |
2230 | tty->print_cr("#r%3.3d %s+%2d: new out preserve" ,reg,fp,reg2offset_unchecked(reg)); |
2231 | } |
2232 | tty->print_cr("#" ); |
2233 | } |
2234 | |
2235 | void PhaseChaitin::dump_bb( uint pre_order ) const { |
2236 | tty->print_cr("---dump of B%d---" ,pre_order); |
2237 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
2238 | Block* block = _cfg.get_block(i); |
2239 | if (block->_pre_order == pre_order) { |
2240 | dump(block); |
2241 | } |
2242 | } |
2243 | } |
2244 | |
2245 | void PhaseChaitin::dump_lrg( uint lidx, bool defs_only ) const { |
2246 | tty->print_cr("---dump of L%d---" ,lidx); |
2247 | |
2248 | if (_ifg) { |
2249 | if (lidx >= _lrg_map.max_lrg_id()) { |
2250 | tty->print("Attempt to print live range index beyond max live range.\n" ); |
2251 | return; |
2252 | } |
2253 | tty->print("L%d: " ,lidx); |
2254 | if (lidx < _ifg->_maxlrg) { |
2255 | lrgs(lidx).dump(); |
2256 | } else { |
2257 | tty->print_cr("new LRG" ); |
2258 | } |
2259 | } |
2260 | if( _ifg && lidx < _ifg->_maxlrg) { |
2261 | tty->print("Neighbors: %d - " , _ifg->neighbor_cnt(lidx)); |
2262 | _ifg->neighbors(lidx)->dump(); |
2263 | tty->cr(); |
2264 | } |
2265 | // For all blocks |
2266 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
2267 | Block* block = _cfg.get_block(i); |
2268 | int dump_once = 0; |
2269 | |
2270 | // For all instructions |
2271 | for( uint j = 0; j < block->number_of_nodes(); j++ ) { |
2272 | Node *n = block->get_node(j); |
2273 | if (_lrg_map.find_const(n) == lidx) { |
2274 | if (!dump_once++) { |
2275 | tty->cr(); |
2276 | block->dump_head(&_cfg); |
2277 | } |
2278 | dump(n); |
2279 | continue; |
2280 | } |
2281 | if (!defs_only) { |
2282 | uint cnt = n->req(); |
2283 | for( uint k = 1; k < cnt; k++ ) { |
2284 | Node *m = n->in(k); |
2285 | if (!m) { |
2286 | continue; // be robust in the dumper |
2287 | } |
2288 | if (_lrg_map.find_const(m) == lidx) { |
2289 | if (!dump_once++) { |
2290 | tty->cr(); |
2291 | block->dump_head(&_cfg); |
2292 | } |
2293 | dump(n); |
2294 | } |
2295 | } |
2296 | } |
2297 | } |
2298 | } // End of per-block dump |
2299 | tty->cr(); |
2300 | } |
2301 | #endif // not PRODUCT |
2302 | |
2303 | int PhaseChaitin::_final_loads = 0; |
2304 | int PhaseChaitin::_final_stores = 0; |
2305 | int PhaseChaitin::_final_memoves= 0; |
2306 | int PhaseChaitin::_final_copies = 0; |
2307 | double PhaseChaitin::_final_load_cost = 0; |
2308 | double PhaseChaitin::_final_store_cost = 0; |
2309 | double PhaseChaitin::_final_memove_cost= 0; |
2310 | double PhaseChaitin::_final_copy_cost = 0; |
2311 | int PhaseChaitin::_conserv_coalesce = 0; |
2312 | int PhaseChaitin::_conserv_coalesce_pair = 0; |
2313 | int PhaseChaitin::_conserv_coalesce_trie = 0; |
2314 | int PhaseChaitin::_conserv_coalesce_quad = 0; |
2315 | int PhaseChaitin::_post_alloc = 0; |
2316 | int PhaseChaitin::_lost_opp_pp_coalesce = 0; |
2317 | int PhaseChaitin::_lost_opp_cflow_coalesce = 0; |
2318 | int PhaseChaitin::_used_cisc_instructions = 0; |
2319 | int PhaseChaitin::_unused_cisc_instructions = 0; |
2320 | int PhaseChaitin::_allocator_attempts = 0; |
2321 | int PhaseChaitin::_allocator_successes = 0; |
2322 | |
2323 | #ifndef PRODUCT |
2324 | uint PhaseChaitin::_high_pressure = 0; |
2325 | uint PhaseChaitin::_low_pressure = 0; |
2326 | |
2327 | void PhaseChaitin::print_chaitin_statistics() { |
2328 | 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); |
2329 | 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); |
2330 | tty->print_cr("Adjusted spill cost = %7.0f." , |
2331 | _final_load_cost*4.0 + _final_store_cost * 2.0 + |
2332 | _final_copy_cost*1.0 + _final_memove_cost*12.0); |
2333 | tty->print("Conservatively coalesced %d copies, %d pairs" , |
2334 | _conserv_coalesce, _conserv_coalesce_pair); |
2335 | if( _conserv_coalesce_trie || _conserv_coalesce_quad ) |
2336 | tty->print(", %d tries, %d quads" , _conserv_coalesce_trie, _conserv_coalesce_quad); |
2337 | tty->print_cr(", %d post alloc." , _post_alloc); |
2338 | if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) |
2339 | tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered." , |
2340 | _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); |
2341 | if( _used_cisc_instructions || _unused_cisc_instructions ) |
2342 | tty->print_cr("Used cisc instruction %d, remained in register %d" , |
2343 | _used_cisc_instructions, _unused_cisc_instructions); |
2344 | if( _allocator_successes != 0 ) |
2345 | tty->print_cr("Average allocation trips %f" , (float)_allocator_attempts/(float)_allocator_successes); |
2346 | tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d" , _high_pressure, _low_pressure); |
2347 | } |
2348 | #endif // not PRODUCT |
2349 | |