1 | /* |
2 | * Copyright (c) 1997, 2017, 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 "memory/allocation.inline.hpp" |
27 | #include "opto/block.hpp" |
28 | #include "opto/c2compiler.hpp" |
29 | #include "opto/cfgnode.hpp" |
30 | #include "opto/chaitin.hpp" |
31 | #include "opto/coalesce.hpp" |
32 | #include "opto/connode.hpp" |
33 | #include "opto/indexSet.hpp" |
34 | #include "opto/machnode.hpp" |
35 | #include "opto/matcher.hpp" |
36 | #include "opto/regmask.hpp" |
37 | |
38 | #ifndef PRODUCT |
39 | void PhaseCoalesce::dump(Node *n) const { |
40 | // Being a const function means I cannot use 'Find' |
41 | uint r = _phc._lrg_map.find(n); |
42 | tty->print("L%d/N%d " ,r,n->_idx); |
43 | } |
44 | |
45 | void PhaseCoalesce::dump() const { |
46 | // I know I have a block layout now, so I can print blocks in a loop |
47 | for( uint i=0; i<_phc._cfg.number_of_blocks(); i++ ) { |
48 | uint j; |
49 | Block* b = _phc._cfg.get_block(i); |
50 | // Print a nice block header |
51 | tty->print("B%d: " ,b->_pre_order); |
52 | for( j=1; j<b->num_preds(); j++ ) |
53 | tty->print("B%d " , _phc._cfg.get_block_for_node(b->pred(j))->_pre_order); |
54 | tty->print("-> " ); |
55 | for( j=0; j<b->_num_succs; j++ ) |
56 | tty->print("B%d " ,b->_succs[j]->_pre_order); |
57 | tty->print(" IDom: B%d/#%d\n" , b->_idom ? b->_idom->_pre_order : 0, b->_dom_depth); |
58 | uint cnt = b->number_of_nodes(); |
59 | for( j=0; j<cnt; j++ ) { |
60 | Node *n = b->get_node(j); |
61 | dump( n ); |
62 | tty->print("\t%s\t" ,n->Name()); |
63 | |
64 | // Dump the inputs |
65 | uint k; // Exit value of loop |
66 | for( k=0; k<n->req(); k++ ) // For all required inputs |
67 | if( n->in(k) ) dump( n->in(k) ); |
68 | else tty->print("_ " ); |
69 | int any_prec = 0; |
70 | for( ; k<n->len(); k++ ) // For all precedence inputs |
71 | if( n->in(k) ) { |
72 | if( !any_prec++ ) tty->print(" |" ); |
73 | dump( n->in(k) ); |
74 | } |
75 | |
76 | // Dump node-specific info |
77 | n->dump_spec(tty); |
78 | tty->print("\n" ); |
79 | |
80 | } |
81 | tty->print("\n" ); |
82 | } |
83 | } |
84 | #endif |
85 | |
86 | // Combine the live ranges def'd by these 2 Nodes. N2 is an input to N1. |
87 | void PhaseCoalesce::combine_these_two(Node *n1, Node *n2) { |
88 | uint lr1 = _phc._lrg_map.find(n1); |
89 | uint lr2 = _phc._lrg_map.find(n2); |
90 | if( lr1 != lr2 && // Different live ranges already AND |
91 | !_phc._ifg->test_edge_sq( lr1, lr2 ) ) { // Do not interfere |
92 | LRG *lrg1 = &_phc.lrgs(lr1); |
93 | LRG *lrg2 = &_phc.lrgs(lr2); |
94 | // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK. |
95 | |
96 | // Now, why is int->oop OK? We end up declaring a raw-pointer as an oop |
97 | // and in general that's a bad thing. However, int->oop conversions only |
98 | // happen at GC points, so the lifetime of the misclassified raw-pointer |
99 | // is from the CheckCastPP (that converts it to an oop) backwards up |
100 | // through a merge point and into the slow-path call, and around the |
101 | // diamond up to the heap-top check and back down into the slow-path call. |
102 | // The misclassified raw pointer is NOT live across the slow-path call, |
103 | // and so does not appear in any GC info, so the fact that it is |
104 | // misclassified is OK. |
105 | |
106 | if( (lrg1->_is_oop || !lrg2->_is_oop) && // not an oop->int cast AND |
107 | // Compatible final mask |
108 | lrg1->mask().overlap( lrg2->mask() ) ) { |
109 | // Merge larger into smaller. |
110 | if( lr1 > lr2 ) { |
111 | uint tmp = lr1; lr1 = lr2; lr2 = tmp; |
112 | Node *n = n1; n1 = n2; n2 = n; |
113 | LRG *ltmp = lrg1; lrg1 = lrg2; lrg2 = ltmp; |
114 | } |
115 | // Union lr2 into lr1 |
116 | _phc.Union( n1, n2 ); |
117 | if (lrg1->_maxfreq < lrg2->_maxfreq) |
118 | lrg1->_maxfreq = lrg2->_maxfreq; |
119 | // Merge in the IFG |
120 | _phc._ifg->Union( lr1, lr2 ); |
121 | // Combine register restrictions |
122 | lrg1->AND(lrg2->mask()); |
123 | } |
124 | } |
125 | } |
126 | |
127 | // Copy coalescing |
128 | void PhaseCoalesce::coalesce_driver() { |
129 | verify(); |
130 | // Coalesce from high frequency to low |
131 | for (uint i = 0; i < _phc._cfg.number_of_blocks(); i++) { |
132 | coalesce(_phc._blks[i]); |
133 | } |
134 | } |
135 | |
136 | // I am inserting copies to come out of SSA form. In the general case, I am |
137 | // doing a parallel renaming. I'm in the Named world now, so I can't do a |
138 | // general parallel renaming. All the copies now use "names" (live-ranges) |
139 | // to carry values instead of the explicit use-def chains. Suppose I need to |
140 | // insert 2 copies into the same block. They copy L161->L128 and L128->L132. |
141 | // If I insert them in the wrong order then L128 will get clobbered before it |
142 | // can get used by the second copy. This cannot happen in the SSA model; |
143 | // direct use-def chains get me the right value. It DOES happen in the named |
144 | // model so I have to handle the reordering of copies. |
145 | // |
146 | // In general, I need to topo-sort the placed copies to avoid conflicts. |
147 | // Its possible to have a closed cycle of copies (e.g., recirculating the same |
148 | // values around a loop). In this case I need a temp to break the cycle. |
149 | void PhaseAggressiveCoalesce::insert_copy_with_overlap( Block *b, Node *copy, uint dst_name, uint src_name ) { |
150 | |
151 | // Scan backwards for the locations of the last use of the dst_name. |
152 | // I am about to clobber the dst_name, so the copy must be inserted |
153 | // after the last use. Last use is really first-use on a backwards scan. |
154 | uint i = b->end_idx()-1; |
155 | while(1) { |
156 | Node *n = b->get_node(i); |
157 | // Check for end of virtual copies; this is also the end of the |
158 | // parallel renaming effort. |
159 | if (n->_idx < _unique) { |
160 | break; |
161 | } |
162 | uint idx = n->is_Copy(); |
163 | assert( idx || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" ); |
164 | if (idx && _phc._lrg_map.find(n->in(idx)) == dst_name) { |
165 | break; |
166 | } |
167 | i--; |
168 | } |
169 | uint last_use_idx = i; |
170 | |
171 | // Also search for any kill of src_name that exits the block. |
172 | // Since the copy uses src_name, I have to come before any kill. |
173 | uint kill_src_idx = b->end_idx(); |
174 | // There can be only 1 kill that exits any block and that is |
175 | // the last kill. Thus it is the first kill on a backwards scan. |
176 | i = b->end_idx()-1; |
177 | while (1) { |
178 | Node *n = b->get_node(i); |
179 | // Check for end of virtual copies; this is also the end of the |
180 | // parallel renaming effort. |
181 | if (n->_idx < _unique) { |
182 | break; |
183 | } |
184 | assert( n->is_Copy() || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" ); |
185 | if (_phc._lrg_map.find(n) == src_name) { |
186 | kill_src_idx = i; |
187 | break; |
188 | } |
189 | i--; |
190 | } |
191 | // Need a temp? Last use of dst comes after the kill of src? |
192 | if (last_use_idx >= kill_src_idx) { |
193 | // Need to break a cycle with a temp |
194 | uint idx = copy->is_Copy(); |
195 | Node *tmp = copy->clone(); |
196 | uint max_lrg_id = _phc._lrg_map.max_lrg_id(); |
197 | _phc.new_lrg(tmp, max_lrg_id); |
198 | _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1); |
199 | |
200 | // Insert new temp between copy and source |
201 | tmp ->set_req(idx,copy->in(idx)); |
202 | copy->set_req(idx,tmp); |
203 | // Save source in temp early, before source is killed |
204 | b->insert_node(tmp, kill_src_idx); |
205 | _phc._cfg.map_node_to_block(tmp, b); |
206 | last_use_idx++; |
207 | } |
208 | |
209 | // Insert just after last use |
210 | b->insert_node(copy, last_use_idx + 1); |
211 | } |
212 | |
213 | void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) { |
214 | // We do LRGs compressing and fix a liveout data only here since the other |
215 | // place in Split() is guarded by the assert which we never hit. |
216 | _phc._lrg_map.compress_uf_map_for_nodes(); |
217 | // Fix block's liveout data for compressed live ranges. |
218 | for (uint lrg = 1; lrg < _phc._lrg_map.max_lrg_id(); lrg++) { |
219 | uint compressed_lrg = _phc._lrg_map.find(lrg); |
220 | if (lrg != compressed_lrg) { |
221 | for (uint bidx = 0; bidx < _phc._cfg.number_of_blocks(); bidx++) { |
222 | IndexSet *liveout = _phc._live->live(_phc._cfg.get_block(bidx)); |
223 | if (liveout->member(lrg)) { |
224 | liveout->remove(lrg); |
225 | liveout->insert(compressed_lrg); |
226 | } |
227 | } |
228 | } |
229 | } |
230 | |
231 | // All new nodes added are actual copies to replace virtual copies. |
232 | // Nodes with index less than '_unique' are original, non-virtual Nodes. |
233 | _unique = C->unique(); |
234 | |
235 | for (uint i = 0; i < _phc._cfg.number_of_blocks(); i++) { |
236 | C->check_node_count(NodeLimitFudgeFactor, "out of nodes in coalesce" ); |
237 | if (C->failing()) return; |
238 | Block *b = _phc._cfg.get_block(i); |
239 | uint cnt = b->num_preds(); // Number of inputs to the Phi |
240 | |
241 | for( uint l = 1; l<b->number_of_nodes(); l++ ) { |
242 | Node *n = b->get_node(l); |
243 | |
244 | // Do not use removed-copies, use copied value instead |
245 | uint ncnt = n->req(); |
246 | for( uint k = 1; k<ncnt; k++ ) { |
247 | Node *copy = n->in(k); |
248 | uint cidx = copy->is_Copy(); |
249 | if( cidx ) { |
250 | Node *def = copy->in(cidx); |
251 | if (_phc._lrg_map.find(copy) == _phc._lrg_map.find(def)) { |
252 | n->set_req(k, def); |
253 | } |
254 | } |
255 | } |
256 | |
257 | // Remove any explicit copies that get coalesced. |
258 | uint cidx = n->is_Copy(); |
259 | if( cidx ) { |
260 | Node *def = n->in(cidx); |
261 | if (_phc._lrg_map.find(n) == _phc._lrg_map.find(def)) { |
262 | n->replace_by(def); |
263 | n->set_req(cidx,NULL); |
264 | b->remove_node(l); |
265 | l--; |
266 | continue; |
267 | } |
268 | } |
269 | |
270 | if (n->is_Phi()) { |
271 | // Get the chosen name for the Phi |
272 | uint phi_name = _phc._lrg_map.find(n); |
273 | // Ignore the pre-allocated specials |
274 | if (!phi_name) { |
275 | continue; |
276 | } |
277 | // Check for mismatch inputs to Phi |
278 | for (uint j = 1; j < cnt; j++) { |
279 | Node *m = n->in(j); |
280 | uint src_name = _phc._lrg_map.find(m); |
281 | if (src_name != phi_name) { |
282 | Block *pred = _phc._cfg.get_block_for_node(b->pred(j)); |
283 | Node *copy; |
284 | assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach" ); |
285 | // Rematerialize constants instead of copying them. |
286 | // We do this only for immediate constants, we avoid constant table loads |
287 | // because that will unsafely extend the live range of the constant table base. |
288 | if (m->is_Mach() && m->as_Mach()->is_Con() && !m->as_Mach()->is_MachConstant() && |
289 | m->as_Mach()->rematerialize()) { |
290 | copy = m->clone(); |
291 | // Insert the copy in the predecessor basic block |
292 | pred->add_inst(copy); |
293 | // Copy any flags as well |
294 | _phc.clone_projs(pred, pred->end_idx(), m, copy, _phc._lrg_map); |
295 | } else { |
296 | uint ireg = m->ideal_reg(); |
297 | if (ireg == 0 || ireg == Op_RegFlags) { |
298 | if (C->subsume_loads()) { |
299 | C->record_failure(C2Compiler::retry_no_subsuming_loads()); |
300 | } else { |
301 | assert(false, "attempted to spill a non-spillable item: %d: %s, ireg = %u, spill_type: %s" , |
302 | m->_idx, m->Name(), ireg, MachSpillCopyNode::spill_type(MachSpillCopyNode::PhiInput)); |
303 | C->record_method_not_compilable("attempted to spill a non-spillable item" ); |
304 | } |
305 | return; |
306 | } |
307 | const RegMask *rm = C->matcher()->idealreg2spillmask[ireg]; |
308 | copy = new MachSpillCopyNode(MachSpillCopyNode::PhiInput, m, *rm, *rm); |
309 | // Find a good place to insert. Kinda tricky, use a subroutine |
310 | insert_copy_with_overlap(pred,copy,phi_name,src_name); |
311 | } |
312 | // Insert the copy in the use-def chain |
313 | n->set_req(j, copy); |
314 | _phc._cfg.map_node_to_block(copy, pred); |
315 | // Extend ("register allocate") the names array for the copy. |
316 | _phc._lrg_map.extend(copy->_idx, phi_name); |
317 | } // End of if Phi names do not match |
318 | } // End of for all inputs to Phi |
319 | } else { // End of if Phi |
320 | |
321 | // Now check for 2-address instructions |
322 | uint idx; |
323 | if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) { |
324 | // Get the chosen name for the Node |
325 | uint name = _phc._lrg_map.find(n); |
326 | assert (name, "no 2-address specials" ); |
327 | // Check for name mis-match on the 2-address input |
328 | Node *m = n->in(idx); |
329 | if (_phc._lrg_map.find(m) != name) { |
330 | Node *copy; |
331 | assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach" ); |
332 | // At this point it is unsafe to extend live ranges (6550579). |
333 | // Rematerialize only constants as we do for Phi above. |
334 | if (m->is_Mach() && m->as_Mach()->is_Con() && !m->as_Mach()->is_MachConstant() && |
335 | m->as_Mach()->rematerialize()) { |
336 | copy = m->clone(); |
337 | // Insert the copy in the basic block, just before us |
338 | b->insert_node(copy, l++); |
339 | l += _phc.clone_projs(b, l, m, copy, _phc._lrg_map); |
340 | } else { |
341 | uint ireg = m->ideal_reg(); |
342 | if (ireg == 0 || ireg == Op_RegFlags) { |
343 | assert(false, "attempted to spill a non-spillable item: %d: %s, ireg = %u, spill_type: %s" , |
344 | m->_idx, m->Name(), ireg, MachSpillCopyNode::spill_type(MachSpillCopyNode::TwoAddress)); |
345 | C->record_method_not_compilable("attempted to spill a non-spillable item" ); |
346 | return; |
347 | } |
348 | const RegMask *rm = C->matcher()->idealreg2spillmask[ireg]; |
349 | copy = new MachSpillCopyNode(MachSpillCopyNode::TwoAddress, m, *rm, *rm); |
350 | // Insert the copy in the basic block, just before us |
351 | b->insert_node(copy, l++); |
352 | } |
353 | // Insert the copy in the use-def chain |
354 | n->set_req(idx, copy); |
355 | // Extend ("register allocate") the names array for the copy. |
356 | _phc._lrg_map.extend(copy->_idx, name); |
357 | _phc._cfg.map_node_to_block(copy, b); |
358 | } |
359 | |
360 | } // End of is two-adr |
361 | |
362 | // Insert a copy at a debug use for a lrg which has high frequency |
363 | if (b->_freq < OPTO_DEBUG_SPLIT_FREQ || _phc._cfg.is_uncommon(b)) { |
364 | // Walk the debug inputs to the node and check for lrg freq |
365 | JVMState* jvms = n->jvms(); |
366 | uint debug_start = jvms ? jvms->debug_start() : 999999; |
367 | uint debug_end = jvms ? jvms->debug_end() : 999999; |
368 | for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) { |
369 | // Do not split monitors; they are only needed for debug table |
370 | // entries and need no code. |
371 | if (jvms->is_monitor_use(inpidx)) { |
372 | continue; |
373 | } |
374 | Node *inp = n->in(inpidx); |
375 | uint nidx = _phc._lrg_map.live_range_id(inp); |
376 | LRG &lrg = lrgs(nidx); |
377 | |
378 | // If this lrg has a high frequency use/def |
379 | if( lrg._maxfreq >= _phc.high_frequency_lrg() ) { |
380 | // If the live range is also live out of this block (like it |
381 | // would be for a fast/slow idiom), the normal spill mechanism |
382 | // does an excellent job. If it is not live out of this block |
383 | // (like it would be for debug info to uncommon trap) splitting |
384 | // the live range now allows a better allocation in the high |
385 | // frequency blocks. |
386 | // Build_IFG_virtual has converted the live sets to |
387 | // live-IN info, not live-OUT info. |
388 | uint k; |
389 | for( k=0; k < b->_num_succs; k++ ) |
390 | if( _phc._live->live(b->_succs[k])->member( nidx ) ) |
391 | break; // Live in to some successor block? |
392 | if( k < b->_num_succs ) |
393 | continue; // Live out; do not pre-split |
394 | // Split the lrg at this use |
395 | uint ireg = inp->ideal_reg(); |
396 | if (ireg == 0 || ireg == Op_RegFlags) { |
397 | assert(false, "attempted to spill a non-spillable item: %d: %s, ireg = %u, spill_type: %s" , |
398 | inp->_idx, inp->Name(), ireg, MachSpillCopyNode::spill_type(MachSpillCopyNode::DebugUse)); |
399 | C->record_method_not_compilable("attempted to spill a non-spillable item" ); |
400 | return; |
401 | } |
402 | const RegMask *rm = C->matcher()->idealreg2spillmask[ireg]; |
403 | Node* copy = new MachSpillCopyNode(MachSpillCopyNode::DebugUse, inp, *rm, *rm); |
404 | // Insert the copy in the use-def chain |
405 | n->set_req(inpidx, copy ); |
406 | // Insert the copy in the basic block, just before us |
407 | b->insert_node(copy, l++); |
408 | // Extend ("register allocate") the names array for the copy. |
409 | uint max_lrg_id = _phc._lrg_map.max_lrg_id(); |
410 | _phc.new_lrg(copy, max_lrg_id); |
411 | _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1); |
412 | _phc._cfg.map_node_to_block(copy, b); |
413 | //tty->print_cr("Split a debug use in Aggressive Coalesce"); |
414 | } // End of if high frequency use/def |
415 | } // End of for all debug inputs |
416 | } // End of if low frequency safepoint |
417 | |
418 | } // End of if Phi |
419 | |
420 | } // End of for all instructions |
421 | } // End of for all blocks |
422 | } |
423 | |
424 | |
425 | // Aggressive (but pessimistic) copy coalescing of a single block |
426 | |
427 | // The following coalesce pass represents a single round of aggressive |
428 | // pessimistic coalesce. "Aggressive" means no attempt to preserve |
429 | // colorability when coalescing. This occasionally means more spills, but |
430 | // it also means fewer rounds of coalescing for better code - and that means |
431 | // faster compiles. |
432 | |
433 | // "Pessimistic" means we do not hit the fixed point in one pass (and we are |
434 | // reaching for the least fixed point to boot). This is typically solved |
435 | // with a few more rounds of coalescing, but the compiler must run fast. We |
436 | // could optimistically coalescing everything touching PhiNodes together |
437 | // into one big live range, then check for self-interference. Everywhere |
438 | // the live range interferes with self it would have to be split. Finding |
439 | // the right split points can be done with some heuristics (based on |
440 | // expected frequency of edges in the live range). In short, it's a real |
441 | // research problem and the timeline is too short to allow such research. |
442 | // Further thoughts: (1) build the LR in a pass, (2) find self-interference |
443 | // in another pass, (3) per each self-conflict, split, (4) split by finding |
444 | // the low-cost cut (min-cut) of the LR, (5) edges in the LR are weighted |
445 | // according to the GCM algorithm (or just exec freq on CFG edges). |
446 | |
447 | void PhaseAggressiveCoalesce::coalesce( Block *b ) { |
448 | // Copies are still "virtual" - meaning we have not made them explicitly |
449 | // copies. Instead, Phi functions of successor blocks have mis-matched |
450 | // live-ranges. If I fail to coalesce, I'll have to insert a copy to line |
451 | // up the live-ranges. Check for Phis in successor blocks. |
452 | uint i; |
453 | for( i=0; i<b->_num_succs; i++ ) { |
454 | Block *bs = b->_succs[i]; |
455 | // Find index of 'b' in 'bs' predecessors |
456 | uint j=1; |
457 | while (_phc._cfg.get_block_for_node(bs->pred(j)) != b) { |
458 | j++; |
459 | } |
460 | |
461 | // Visit all the Phis in successor block |
462 | for( uint k = 1; k<bs->number_of_nodes(); k++ ) { |
463 | Node *n = bs->get_node(k); |
464 | if( !n->is_Phi() ) break; |
465 | combine_these_two( n, n->in(j) ); |
466 | } |
467 | } // End of for all successor blocks |
468 | |
469 | |
470 | // Check _this_ block for 2-address instructions and copies. |
471 | uint cnt = b->end_idx(); |
472 | for( i = 1; i<cnt; i++ ) { |
473 | Node *n = b->get_node(i); |
474 | uint idx; |
475 | // 2-address instructions have a virtual Copy matching their input |
476 | // to their output |
477 | if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) { |
478 | MachNode *mach = n->as_Mach(); |
479 | combine_these_two(mach, mach->in(idx)); |
480 | } |
481 | } // End of for all instructions in block |
482 | } |
483 | |
484 | PhaseConservativeCoalesce::PhaseConservativeCoalesce(PhaseChaitin &chaitin) : PhaseCoalesce(chaitin) { |
485 | _ulr.initialize(_phc._lrg_map.max_lrg_id()); |
486 | } |
487 | |
488 | void PhaseConservativeCoalesce::verify() { |
489 | #ifdef ASSERT |
490 | _phc.set_was_low(); |
491 | #endif |
492 | } |
493 | |
494 | void PhaseConservativeCoalesce::union_helper( Node *lr1_node, Node *lr2_node, uint lr1, uint lr2, Node *src_def, Node *dst_copy, Node *src_copy, Block *b, uint bindex ) { |
495 | // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the |
496 | // union-find tree |
497 | _phc.Union( lr1_node, lr2_node ); |
498 | |
499 | // Single-def live range ONLY if both live ranges are single-def. |
500 | // If both are single def, then src_def powers one live range |
501 | // and def_copy powers the other. After merging, src_def powers |
502 | // the combined live range. |
503 | lrgs(lr1)._def = (lrgs(lr1).is_multidef() || |
504 | lrgs(lr2).is_multidef() ) |
505 | ? NodeSentinel : src_def; |
506 | lrgs(lr2)._def = NULL; // No def for lrg 2 |
507 | lrgs(lr2).Clear(); // Force empty mask for LRG 2 |
508 | //lrgs(lr2)._size = 0; // Live-range 2 goes dead |
509 | lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop; |
510 | lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info |
511 | |
512 | if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq) |
513 | lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq; |
514 | |
515 | // Copy original value instead. Intermediate copies go dead, and |
516 | // the dst_copy becomes useless. |
517 | int didx = dst_copy->is_Copy(); |
518 | dst_copy->set_req( didx, src_def ); |
519 | // Add copy to free list |
520 | // _phc.free_spillcopy(b->_nodes[bindex]); |
521 | assert( b->get_node(bindex) == dst_copy, "" ); |
522 | dst_copy->replace_by( dst_copy->in(didx) ); |
523 | dst_copy->set_req( didx, NULL); |
524 | b->remove_node(bindex); |
525 | if( bindex < b->_ihrp_index ) b->_ihrp_index--; |
526 | if( bindex < b->_fhrp_index ) b->_fhrp_index--; |
527 | |
528 | // Stretched lr1; add it to liveness of intermediate blocks |
529 | Block *b2 = _phc._cfg.get_block_for_node(src_copy); |
530 | while( b != b2 ) { |
531 | b = _phc._cfg.get_block_for_node(b->pred(1)); |
532 | _phc._live->live(b)->insert(lr1); |
533 | } |
534 | } |
535 | |
536 | // Factored code from copy_copy that computes extra interferences from |
537 | // lengthening a live range by double-coalescing. |
538 | uint PhaseConservativeCoalesce::compute_separating_interferences(Node *dst_copy, Node *src_copy, Block *b, uint bindex, RegMask &rm, uint reg_degree, uint rm_size, uint lr1, uint lr2 ) { |
539 | |
540 | assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj" ); |
541 | assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj" ); |
542 | Node *prev_copy = dst_copy->in(dst_copy->is_Copy()); |
543 | Block *b2 = b; |
544 | uint bindex2 = bindex; |
545 | while( 1 ) { |
546 | // Find previous instruction |
547 | bindex2--; // Chain backwards 1 instruction |
548 | while( bindex2 == 0 ) { // At block start, find prior block |
549 | assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" ); |
550 | b2 = _phc._cfg.get_block_for_node(b2->pred(1)); |
551 | bindex2 = b2->end_idx()-1; |
552 | } |
553 | // Get prior instruction |
554 | assert(bindex2 < b2->number_of_nodes(), "index out of bounds" ); |
555 | Node *x = b2->get_node(bindex2); |
556 | if( x == prev_copy ) { // Previous copy in copy chain? |
557 | if( prev_copy == src_copy)// Found end of chain and all interferences |
558 | break; // So break out of loop |
559 | // Else work back one in copy chain |
560 | prev_copy = prev_copy->in(prev_copy->is_Copy()); |
561 | } else { // Else collect interferences |
562 | uint lidx = _phc._lrg_map.find(x); |
563 | // Found another def of live-range being stretched? |
564 | if(lidx == lr1) { |
565 | return max_juint; |
566 | } |
567 | if(lidx == lr2) { |
568 | return max_juint; |
569 | } |
570 | |
571 | // If we attempt to coalesce across a bound def |
572 | if( lrgs(lidx).is_bound() ) { |
573 | // Do not let the coalesced LRG expect to get the bound color |
574 | rm.SUBTRACT( lrgs(lidx).mask() ); |
575 | // Recompute rm_size |
576 | rm_size = rm.Size(); |
577 | //if( rm._flags ) rm_size += 1000000; |
578 | if( reg_degree >= rm_size ) return max_juint; |
579 | } |
580 | if( rm.overlap(lrgs(lidx).mask()) ) { |
581 | // Insert lidx into union LRG; returns TRUE if actually inserted |
582 | if( _ulr.insert(lidx) ) { |
583 | // Infinite-stack neighbors do not alter colorability, as they |
584 | // can always color to some other color. |
585 | if( !lrgs(lidx).mask().is_AllStack() ) { |
586 | // If this coalesce will make any new neighbor uncolorable, |
587 | // do not coalesce. |
588 | if( lrgs(lidx).just_lo_degree() ) |
589 | return max_juint; |
590 | // Bump our degree |
591 | if( ++reg_degree >= rm_size ) |
592 | return max_juint; |
593 | } // End of if not infinite-stack neighbor |
594 | } // End of if actually inserted |
595 | } // End of if live range overlaps |
596 | } // End of else collect interferences for 1 node |
597 | } // End of while forever, scan back for interferences |
598 | return reg_degree; |
599 | } |
600 | |
601 | void PhaseConservativeCoalesce::update_ifg(uint lr1, uint lr2, IndexSet *n_lr1, IndexSet *n_lr2) { |
602 | // Some original neighbors of lr1 might have gone away |
603 | // because the constrained register mask prevented them. |
604 | // Remove lr1 from such neighbors. |
605 | IndexSetIterator one(n_lr1); |
606 | uint neighbor; |
607 | LRG &lrg1 = lrgs(lr1); |
608 | while ((neighbor = one.next()) != 0) |
609 | if( !_ulr.member(neighbor) ) |
610 | if( _phc._ifg->neighbors(neighbor)->remove(lr1) ) |
611 | lrgs(neighbor).inc_degree( -lrg1.compute_degree(lrgs(neighbor)) ); |
612 | |
613 | |
614 | // lr2 is now called (coalesced into) lr1. |
615 | // Remove lr2 from the IFG. |
616 | IndexSetIterator two(n_lr2); |
617 | LRG &lrg2 = lrgs(lr2); |
618 | while ((neighbor = two.next()) != 0) |
619 | if( _phc._ifg->neighbors(neighbor)->remove(lr2) ) |
620 | lrgs(neighbor).inc_degree( -lrg2.compute_degree(lrgs(neighbor)) ); |
621 | |
622 | // Some neighbors of intermediate copies now interfere with the |
623 | // combined live range. |
624 | IndexSetIterator three(&_ulr); |
625 | while ((neighbor = three.next()) != 0) |
626 | if( _phc._ifg->neighbors(neighbor)->insert(lr1) ) |
627 | lrgs(neighbor).inc_degree( lrg1.compute_degree(lrgs(neighbor)) ); |
628 | } |
629 | |
630 | static void record_bias( const PhaseIFG *ifg, int lr1, int lr2 ) { |
631 | // Tag copy bias here |
632 | if( !ifg->lrgs(lr1)._copy_bias ) |
633 | ifg->lrgs(lr1)._copy_bias = lr2; |
634 | if( !ifg->lrgs(lr2)._copy_bias ) |
635 | ifg->lrgs(lr2)._copy_bias = lr1; |
636 | } |
637 | |
638 | // See if I can coalesce a series of multiple copies together. I need the |
639 | // final dest copy and the original src copy. They can be the same Node. |
640 | // Compute the compatible register masks. |
641 | bool PhaseConservativeCoalesce::copy_copy(Node *dst_copy, Node *src_copy, Block *b, uint bindex) { |
642 | |
643 | if (!dst_copy->is_SpillCopy()) { |
644 | return false; |
645 | } |
646 | if (!src_copy->is_SpillCopy()) { |
647 | return false; |
648 | } |
649 | Node *src_def = src_copy->in(src_copy->is_Copy()); |
650 | uint lr1 = _phc._lrg_map.find(dst_copy); |
651 | uint lr2 = _phc._lrg_map.find(src_def); |
652 | |
653 | // Same live ranges already? |
654 | if (lr1 == lr2) { |
655 | return false; |
656 | } |
657 | |
658 | // Interfere? |
659 | if (_phc._ifg->test_edge_sq(lr1, lr2)) { |
660 | return false; |
661 | } |
662 | |
663 | // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK. |
664 | if (!lrgs(lr1)._is_oop && lrgs(lr2)._is_oop) { // not an oop->int cast |
665 | return false; |
666 | } |
667 | |
668 | // Coalescing between an aligned live range and a mis-aligned live range? |
669 | // No, no! Alignment changes how we count degree. |
670 | if (lrgs(lr1)._fat_proj != lrgs(lr2)._fat_proj) { |
671 | return false; |
672 | } |
673 | |
674 | // Sort; use smaller live-range number |
675 | Node *lr1_node = dst_copy; |
676 | Node *lr2_node = src_def; |
677 | if (lr1 > lr2) { |
678 | uint tmp = lr1; lr1 = lr2; lr2 = tmp; |
679 | lr1_node = src_def; lr2_node = dst_copy; |
680 | } |
681 | |
682 | // Check for compatibility of the 2 live ranges by |
683 | // intersecting their allowed register sets. |
684 | RegMask rm = lrgs(lr1).mask(); |
685 | rm.AND(lrgs(lr2).mask()); |
686 | // Number of bits free |
687 | uint rm_size = rm.Size(); |
688 | |
689 | if (UseFPUForSpilling && rm.is_AllStack() ) { |
690 | // Don't coalesce when frequency difference is large |
691 | Block *dst_b = _phc._cfg.get_block_for_node(dst_copy); |
692 | Block *src_def_b = _phc._cfg.get_block_for_node(src_def); |
693 | if (src_def_b->_freq > 10*dst_b->_freq ) |
694 | return false; |
695 | } |
696 | |
697 | // If we can use any stack slot, then effective size is infinite |
698 | if( rm.is_AllStack() ) rm_size += 1000000; |
699 | // Incompatible masks, no way to coalesce |
700 | if( rm_size == 0 ) return false; |
701 | |
702 | // Another early bail-out test is when we are double-coalescing and the |
703 | // 2 copies are separated by some control flow. |
704 | if( dst_copy != src_copy ) { |
705 | Block *src_b = _phc._cfg.get_block_for_node(src_copy); |
706 | Block *b2 = b; |
707 | while( b2 != src_b ) { |
708 | if( b2->num_preds() > 2 ){// Found merge-point |
709 | _phc._lost_opp_cflow_coalesce++; |
710 | // extra record_bias commented out because Chris believes it is not |
711 | // productive. Since we can record only 1 bias, we want to choose one |
712 | // that stands a chance of working and this one probably does not. |
713 | //record_bias( _phc._lrgs, lr1, lr2 ); |
714 | return false; // To hard to find all interferences |
715 | } |
716 | b2 = _phc._cfg.get_block_for_node(b2->pred(1)); |
717 | } |
718 | } |
719 | |
720 | // Union the two interference sets together into '_ulr' |
721 | uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm ); |
722 | |
723 | if( reg_degree >= rm_size ) { |
724 | record_bias( _phc._ifg, lr1, lr2 ); |
725 | return false; |
726 | } |
727 | |
728 | // Now I need to compute all the interferences between dst_copy and |
729 | // src_copy. I'm not willing visit the entire interference graph, so |
730 | // I limit my search to things in dst_copy's block or in a straight |
731 | // line of previous blocks. I give up at merge points or when I get |
732 | // more interferences than my degree. I can stop when I find src_copy. |
733 | if( dst_copy != src_copy ) { |
734 | reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 ); |
735 | if( reg_degree == max_juint ) { |
736 | record_bias( _phc._ifg, lr1, lr2 ); |
737 | return false; |
738 | } |
739 | } // End of if dst_copy & src_copy are different |
740 | |
741 | |
742 | // ---- THE COMBINED LRG IS COLORABLE ---- |
743 | |
744 | // YEAH - Now coalesce this copy away |
745 | assert( lrgs(lr1).num_regs() == lrgs(lr2).num_regs(), "" ); |
746 | |
747 | IndexSet *n_lr1 = _phc._ifg->neighbors(lr1); |
748 | IndexSet *n_lr2 = _phc._ifg->neighbors(lr2); |
749 | |
750 | // Update the interference graph |
751 | update_ifg(lr1, lr2, n_lr1, n_lr2); |
752 | |
753 | _ulr.remove(lr1); |
754 | |
755 | // Uncomment the following code to trace Coalescing in great detail. |
756 | // |
757 | //if (false) { |
758 | // tty->cr(); |
759 | // tty->print_cr("#######################################"); |
760 | // tty->print_cr("union %d and %d", lr1, lr2); |
761 | // n_lr1->dump(); |
762 | // n_lr2->dump(); |
763 | // tty->print_cr("resulting set is"); |
764 | // _ulr.dump(); |
765 | //} |
766 | |
767 | // Replace n_lr1 with the new combined live range. _ulr will use |
768 | // n_lr1's old memory on the next iteration. n_lr2 is cleared to |
769 | // send its internal memory to the free list. |
770 | _ulr.swap(n_lr1); |
771 | _ulr.clear(); |
772 | n_lr2->clear(); |
773 | |
774 | lrgs(lr1).set_degree( _phc._ifg->effective_degree(lr1) ); |
775 | lrgs(lr2).set_degree( 0 ); |
776 | |
777 | // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the |
778 | // union-find tree |
779 | union_helper( lr1_node, lr2_node, lr1, lr2, src_def, dst_copy, src_copy, b, bindex ); |
780 | // Combine register restrictions |
781 | lrgs(lr1).set_mask(rm); |
782 | lrgs(lr1).compute_set_mask_size(); |
783 | lrgs(lr1)._cost += lrgs(lr2)._cost; |
784 | lrgs(lr1)._area += lrgs(lr2)._area; |
785 | |
786 | // While its uncommon to successfully coalesce live ranges that started out |
787 | // being not-lo-degree, it can happen. In any case the combined coalesced |
788 | // live range better Simplify nicely. |
789 | lrgs(lr1)._was_lo = 1; |
790 | |
791 | // kinda expensive to do all the time |
792 | //tty->print_cr("warning: slow verify happening"); |
793 | //_phc._ifg->verify( &_phc ); |
794 | return true; |
795 | } |
796 | |
797 | // Conservative (but pessimistic) copy coalescing of a single block |
798 | void PhaseConservativeCoalesce::coalesce( Block *b ) { |
799 | // Bail out on infrequent blocks |
800 | if (_phc._cfg.is_uncommon(b)) { |
801 | return; |
802 | } |
803 | // Check this block for copies. |
804 | for( uint i = 1; i<b->end_idx(); i++ ) { |
805 | // Check for actual copies on inputs. Coalesce a copy into its |
806 | // input if use and copy's input are compatible. |
807 | Node *copy1 = b->get_node(i); |
808 | uint idx1 = copy1->is_Copy(); |
809 | if( !idx1 ) continue; // Not a copy |
810 | |
811 | if( copy_copy(copy1,copy1,b,i) ) { |
812 | i--; // Retry, same location in block |
813 | PhaseChaitin::_conserv_coalesce++; // Collect stats on success |
814 | continue; |
815 | } |
816 | } |
817 | } |
818 | |