Tue, 02 Jul 2013 20:42:12 -0400
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
Summary: add intrinsics using new instruction to interpreter, C1, C2, for suitable x86; add test
Reviewed-by: kvn, twisti
1 /*
2 * Copyright (c) 1997, 2013, 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 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/block.hpp"
28 #include "opto/cfgnode.hpp"
29 #include "opto/chaitin.hpp"
30 #include "opto/coalesce.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/indexSet.hpp"
33 #include "opto/machnode.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/regmask.hpp"
37 //=============================================================================
38 //------------------------------Dump-------------------------------------------
39 #ifndef PRODUCT
40 void PhaseCoalesce::dump(Node *n) const {
41 // Being a const function means I cannot use 'Find'
42 uint r = _phc._lrg_map.find(n);
43 tty->print("L%d/N%d ",r,n->_idx);
44 }
46 //------------------------------dump-------------------------------------------
47 void PhaseCoalesce::dump() const {
48 // I know I have a block layout now, so I can print blocks in a loop
49 for( uint i=0; i<_phc._cfg._num_blocks; i++ ) {
50 uint j;
51 Block *b = _phc._cfg._blocks[i];
52 // Print a nice block header
53 tty->print("B%d: ",b->_pre_order);
54 for( j=1; j<b->num_preds(); j++ )
55 tty->print("B%d ", _phc._cfg._bbs[b->pred(j)->_idx]->_pre_order);
56 tty->print("-> ");
57 for( j=0; j<b->_num_succs; j++ )
58 tty->print("B%d ",b->_succs[j]->_pre_order);
59 tty->print(" IDom: B%d/#%d\n", b->_idom ? b->_idom->_pre_order : 0, b->_dom_depth);
60 uint cnt = b->_nodes.size();
61 for( j=0; j<cnt; j++ ) {
62 Node *n = b->_nodes[j];
63 dump( n );
64 tty->print("\t%s\t",n->Name());
66 // Dump the inputs
67 uint k; // Exit value of loop
68 for( k=0; k<n->req(); k++ ) // For all required inputs
69 if( n->in(k) ) dump( n->in(k) );
70 else tty->print("_ ");
71 int any_prec = 0;
72 for( ; k<n->len(); k++ ) // For all precedence inputs
73 if( n->in(k) ) {
74 if( !any_prec++ ) tty->print(" |");
75 dump( n->in(k) );
76 }
78 // Dump node-specific info
79 n->dump_spec(tty);
80 tty->print("\n");
82 }
83 tty->print("\n");
84 }
85 }
86 #endif
88 //------------------------------combine_these_two------------------------------
89 // Combine the live ranges def'd by these 2 Nodes. N2 is an input to N1.
90 void PhaseCoalesce::combine_these_two(Node *n1, Node *n2) {
91 uint lr1 = _phc._lrg_map.find(n1);
92 uint lr2 = _phc._lrg_map.find(n2);
93 if( lr1 != lr2 && // Different live ranges already AND
94 !_phc._ifg->test_edge_sq( lr1, lr2 ) ) { // Do not interfere
95 LRG *lrg1 = &_phc.lrgs(lr1);
96 LRG *lrg2 = &_phc.lrgs(lr2);
97 // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK.
99 // Now, why is int->oop OK? We end up declaring a raw-pointer as an oop
100 // and in general that's a bad thing. However, int->oop conversions only
101 // happen at GC points, so the lifetime of the misclassified raw-pointer
102 // is from the CheckCastPP (that converts it to an oop) backwards up
103 // through a merge point and into the slow-path call, and around the
104 // diamond up to the heap-top check and back down into the slow-path call.
105 // The misclassified raw pointer is NOT live across the slow-path call,
106 // and so does not appear in any GC info, so the fact that it is
107 // misclassified is OK.
109 if( (lrg1->_is_oop || !lrg2->_is_oop) && // not an oop->int cast AND
110 // Compatible final mask
111 lrg1->mask().overlap( lrg2->mask() ) ) {
112 // Merge larger into smaller.
113 if( lr1 > lr2 ) {
114 uint tmp = lr1; lr1 = lr2; lr2 = tmp;
115 Node *n = n1; n1 = n2; n2 = n;
116 LRG *ltmp = lrg1; lrg1 = lrg2; lrg2 = ltmp;
117 }
118 // Union lr2 into lr1
119 _phc.Union( n1, n2 );
120 if (lrg1->_maxfreq < lrg2->_maxfreq)
121 lrg1->_maxfreq = lrg2->_maxfreq;
122 // Merge in the IFG
123 _phc._ifg->Union( lr1, lr2 );
124 // Combine register restrictions
125 lrg1->AND(lrg2->mask());
126 }
127 }
128 }
130 //------------------------------coalesce_driver--------------------------------
131 // Copy coalescing
132 void PhaseCoalesce::coalesce_driver( ) {
134 verify();
135 // Coalesce from high frequency to low
136 for( uint i=0; i<_phc._cfg._num_blocks; i++ )
137 coalesce( _phc._blks[i] );
139 }
141 //------------------------------insert_copy_with_overlap-----------------------
142 // I am inserting copies to come out of SSA form. In the general case, I am
143 // doing a parallel renaming. I'm in the Named world now, so I can't do a
144 // general parallel renaming. All the copies now use "names" (live-ranges)
145 // to carry values instead of the explicit use-def chains. Suppose I need to
146 // insert 2 copies into the same block. They copy L161->L128 and L128->L132.
147 // If I insert them in the wrong order then L128 will get clobbered before it
148 // can get used by the second copy. This cannot happen in the SSA model;
149 // direct use-def chains get me the right value. It DOES happen in the named
150 // model so I have to handle the reordering of copies.
151 //
152 // In general, I need to topo-sort the placed copies to avoid conflicts.
153 // Its possible to have a closed cycle of copies (e.g., recirculating the same
154 // values around a loop). In this case I need a temp to break the cycle.
155 void PhaseAggressiveCoalesce::insert_copy_with_overlap( Block *b, Node *copy, uint dst_name, uint src_name ) {
157 // Scan backwards for the locations of the last use of the dst_name.
158 // I am about to clobber the dst_name, so the copy must be inserted
159 // after the last use. Last use is really first-use on a backwards scan.
160 uint i = b->end_idx()-1;
161 while(1) {
162 Node *n = b->_nodes[i];
163 // Check for end of virtual copies; this is also the end of the
164 // parallel renaming effort.
165 if (n->_idx < _unique) {
166 break;
167 }
168 uint idx = n->is_Copy();
169 assert( idx || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" );
170 if (idx && _phc._lrg_map.find(n->in(idx)) == dst_name) {
171 break;
172 }
173 i--;
174 }
175 uint last_use_idx = i;
177 // Also search for any kill of src_name that exits the block.
178 // Since the copy uses src_name, I have to come before any kill.
179 uint kill_src_idx = b->end_idx();
180 // There can be only 1 kill that exits any block and that is
181 // the last kill. Thus it is the first kill on a backwards scan.
182 i = b->end_idx()-1;
183 while (1) {
184 Node *n = b->_nodes[i];
185 // Check for end of virtual copies; this is also the end of the
186 // parallel renaming effort.
187 if (n->_idx < _unique) {
188 break;
189 }
190 assert( n->is_Copy() || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" );
191 if (_phc._lrg_map.find(n) == src_name) {
192 kill_src_idx = i;
193 break;
194 }
195 i--;
196 }
197 // Need a temp? Last use of dst comes after the kill of src?
198 if (last_use_idx >= kill_src_idx) {
199 // Need to break a cycle with a temp
200 uint idx = copy->is_Copy();
201 Node *tmp = copy->clone();
202 uint max_lrg_id = _phc._lrg_map.max_lrg_id();
203 _phc.new_lrg(tmp, max_lrg_id);
204 _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1);
206 // Insert new temp between copy and source
207 tmp ->set_req(idx,copy->in(idx));
208 copy->set_req(idx,tmp);
209 // Save source in temp early, before source is killed
210 b->_nodes.insert(kill_src_idx,tmp);
211 _phc._cfg._bbs.map( tmp->_idx, b );
212 last_use_idx++;
213 }
215 // Insert just after last use
216 b->_nodes.insert(last_use_idx+1,copy);
217 }
219 //------------------------------insert_copies----------------------------------
220 void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) {
221 // We do LRGs compressing and fix a liveout data only here since the other
222 // place in Split() is guarded by the assert which we never hit.
223 _phc._lrg_map.compress_uf_map_for_nodes();
224 // Fix block's liveout data for compressed live ranges.
225 for (uint lrg = 1; lrg < _phc._lrg_map.max_lrg_id(); lrg++) {
226 uint compressed_lrg = _phc._lrg_map.find(lrg);
227 if (lrg != compressed_lrg) {
228 for (uint bidx = 0; bidx < _phc._cfg._num_blocks; bidx++) {
229 IndexSet *liveout = _phc._live->live(_phc._cfg._blocks[bidx]);
230 if (liveout->member(lrg)) {
231 liveout->remove(lrg);
232 liveout->insert(compressed_lrg);
233 }
234 }
235 }
236 }
238 // All new nodes added are actual copies to replace virtual copies.
239 // Nodes with index less than '_unique' are original, non-virtual Nodes.
240 _unique = C->unique();
242 for( uint i=0; i<_phc._cfg._num_blocks; i++ ) {
243 C->check_node_count(NodeLimitFudgeFactor, "out of nodes in coalesce");
244 if (C->failing()) return;
245 Block *b = _phc._cfg._blocks[i];
246 uint cnt = b->num_preds(); // Number of inputs to the Phi
248 for( uint l = 1; l<b->_nodes.size(); l++ ) {
249 Node *n = b->_nodes[l];
251 // Do not use removed-copies, use copied value instead
252 uint ncnt = n->req();
253 for( uint k = 1; k<ncnt; k++ ) {
254 Node *copy = n->in(k);
255 uint cidx = copy->is_Copy();
256 if( cidx ) {
257 Node *def = copy->in(cidx);
258 if (_phc._lrg_map.find(copy) == _phc._lrg_map.find(def)) {
259 n->set_req(k, def);
260 }
261 }
262 }
264 // Remove any explicit copies that get coalesced.
265 uint cidx = n->is_Copy();
266 if( cidx ) {
267 Node *def = n->in(cidx);
268 if (_phc._lrg_map.find(n) == _phc._lrg_map.find(def)) {
269 n->replace_by(def);
270 n->set_req(cidx,NULL);
271 b->_nodes.remove(l);
272 l--;
273 continue;
274 }
275 }
277 if (n->is_Phi()) {
278 // Get the chosen name for the Phi
279 uint phi_name = _phc._lrg_map.find(n);
280 // Ignore the pre-allocated specials
281 if (!phi_name) {
282 continue;
283 }
284 // Check for mismatch inputs to Phi
285 for (uint j = 1; j < cnt; j++) {
286 Node *m = n->in(j);
287 uint src_name = _phc._lrg_map.find(m);
288 if (src_name != phi_name) {
289 Block *pred = _phc._cfg._bbs[b->pred(j)->_idx];
290 Node *copy;
291 assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach");
292 // Rematerialize constants instead of copying them
293 if( m->is_Mach() && m->as_Mach()->is_Con() &&
294 m->as_Mach()->rematerialize() ) {
295 copy = m->clone();
296 // Insert the copy in the predecessor basic block
297 pred->add_inst(copy);
298 // Copy any flags as well
299 _phc.clone_projs(pred, pred->end_idx(), m, copy, _phc._lrg_map);
300 } else {
301 const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()];
302 copy = new (C) MachSpillCopyNode(m, *rm, *rm);
303 // Find a good place to insert. Kinda tricky, use a subroutine
304 insert_copy_with_overlap(pred,copy,phi_name,src_name);
305 }
306 // Insert the copy in the use-def chain
307 n->set_req(j, copy);
308 _phc._cfg._bbs.map( copy->_idx, pred );
309 // Extend ("register allocate") the names array for the copy.
310 _phc._lrg_map.extend(copy->_idx, phi_name);
311 } // End of if Phi names do not match
312 } // End of for all inputs to Phi
313 } else { // End of if Phi
315 // Now check for 2-address instructions
316 uint idx;
317 if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) {
318 // Get the chosen name for the Node
319 uint name = _phc._lrg_map.find(n);
320 assert (name, "no 2-address specials");
321 // Check for name mis-match on the 2-address input
322 Node *m = n->in(idx);
323 if (_phc._lrg_map.find(m) != name) {
324 Node *copy;
325 assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach");
326 // At this point it is unsafe to extend live ranges (6550579).
327 // Rematerialize only constants as we do for Phi above.
328 if(m->is_Mach() && m->as_Mach()->is_Con() &&
329 m->as_Mach()->rematerialize()) {
330 copy = m->clone();
331 // Insert the copy in the basic block, just before us
332 b->_nodes.insert(l++, copy);
333 if(_phc.clone_projs(b, l, m, copy, _phc._lrg_map)) {
334 l++;
335 }
336 } else {
337 const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()];
338 copy = new (C) MachSpillCopyNode(m, *rm, *rm);
339 // Insert the copy in the basic block, just before us
340 b->_nodes.insert(l++, copy);
341 }
342 // Insert the copy in the use-def chain
343 n->set_req(idx, copy);
344 // Extend ("register allocate") the names array for the copy.
345 _phc._lrg_map.extend(copy->_idx, name);
346 _phc._cfg._bbs.map( copy->_idx, b );
347 }
349 } // End of is two-adr
351 // Insert a copy at a debug use for a lrg which has high frequency
352 if (b->_freq < OPTO_DEBUG_SPLIT_FREQ || b->is_uncommon(_phc._cfg._bbs)) {
353 // Walk the debug inputs to the node and check for lrg freq
354 JVMState* jvms = n->jvms();
355 uint debug_start = jvms ? jvms->debug_start() : 999999;
356 uint debug_end = jvms ? jvms->debug_end() : 999999;
357 for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) {
358 // Do not split monitors; they are only needed for debug table
359 // entries and need no code.
360 if (jvms->is_monitor_use(inpidx)) {
361 continue;
362 }
363 Node *inp = n->in(inpidx);
364 uint nidx = _phc._lrg_map.live_range_id(inp);
365 LRG &lrg = lrgs(nidx);
367 // If this lrg has a high frequency use/def
368 if( lrg._maxfreq >= _phc.high_frequency_lrg() ) {
369 // If the live range is also live out of this block (like it
370 // would be for a fast/slow idiom), the normal spill mechanism
371 // does an excellent job. If it is not live out of this block
372 // (like it would be for debug info to uncommon trap) splitting
373 // the live range now allows a better allocation in the high
374 // frequency blocks.
375 // Build_IFG_virtual has converted the live sets to
376 // live-IN info, not live-OUT info.
377 uint k;
378 for( k=0; k < b->_num_succs; k++ )
379 if( _phc._live->live(b->_succs[k])->member( nidx ) )
380 break; // Live in to some successor block?
381 if( k < b->_num_succs )
382 continue; // Live out; do not pre-split
383 // Split the lrg at this use
384 const RegMask *rm = C->matcher()->idealreg2spillmask[inp->ideal_reg()];
385 Node *copy = new (C) MachSpillCopyNode( inp, *rm, *rm );
386 // Insert the copy in the use-def chain
387 n->set_req(inpidx, copy );
388 // Insert the copy in the basic block, just before us
389 b->_nodes.insert( l++, copy );
390 // Extend ("register allocate") the names array for the copy.
391 uint max_lrg_id = _phc._lrg_map.max_lrg_id();
392 _phc.new_lrg(copy, max_lrg_id);
393 _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1);
394 _phc._cfg._bbs.map(copy->_idx, b);
395 //tty->print_cr("Split a debug use in Aggressive Coalesce");
396 } // End of if high frequency use/def
397 } // End of for all debug inputs
398 } // End of if low frequency safepoint
400 } // End of if Phi
402 } // End of for all instructions
403 } // End of for all blocks
404 }
406 //=============================================================================
407 //------------------------------coalesce---------------------------------------
408 // Aggressive (but pessimistic) copy coalescing of a single block
410 // The following coalesce pass represents a single round of aggressive
411 // pessimistic coalesce. "Aggressive" means no attempt to preserve
412 // colorability when coalescing. This occasionally means more spills, but
413 // it also means fewer rounds of coalescing for better code - and that means
414 // faster compiles.
416 // "Pessimistic" means we do not hit the fixed point in one pass (and we are
417 // reaching for the least fixed point to boot). This is typically solved
418 // with a few more rounds of coalescing, but the compiler must run fast. We
419 // could optimistically coalescing everything touching PhiNodes together
420 // into one big live range, then check for self-interference. Everywhere
421 // the live range interferes with self it would have to be split. Finding
422 // the right split points can be done with some heuristics (based on
423 // expected frequency of edges in the live range). In short, it's a real
424 // research problem and the timeline is too short to allow such research.
425 // Further thoughts: (1) build the LR in a pass, (2) find self-interference
426 // in another pass, (3) per each self-conflict, split, (4) split by finding
427 // the low-cost cut (min-cut) of the LR, (5) edges in the LR are weighted
428 // according to the GCM algorithm (or just exec freq on CFG edges).
430 void PhaseAggressiveCoalesce::coalesce( Block *b ) {
431 // Copies are still "virtual" - meaning we have not made them explicitly
432 // copies. Instead, Phi functions of successor blocks have mis-matched
433 // live-ranges. If I fail to coalesce, I'll have to insert a copy to line
434 // up the live-ranges. Check for Phis in successor blocks.
435 uint i;
436 for( i=0; i<b->_num_succs; i++ ) {
437 Block *bs = b->_succs[i];
438 // Find index of 'b' in 'bs' predecessors
439 uint j=1;
440 while( _phc._cfg._bbs[bs->pred(j)->_idx] != b ) j++;
441 // Visit all the Phis in successor block
442 for( uint k = 1; k<bs->_nodes.size(); k++ ) {
443 Node *n = bs->_nodes[k];
444 if( !n->is_Phi() ) break;
445 combine_these_two( n, n->in(j) );
446 }
447 } // End of for all successor blocks
450 // Check _this_ block for 2-address instructions and copies.
451 uint cnt = b->end_idx();
452 for( i = 1; i<cnt; i++ ) {
453 Node *n = b->_nodes[i];
454 uint idx;
455 // 2-address instructions have a virtual Copy matching their input
456 // to their output
457 if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) {
458 MachNode *mach = n->as_Mach();
459 combine_these_two(mach, mach->in(idx));
460 }
461 } // End of for all instructions in block
462 }
464 //=============================================================================
465 //------------------------------PhaseConservativeCoalesce----------------------
466 PhaseConservativeCoalesce::PhaseConservativeCoalesce(PhaseChaitin &chaitin) : PhaseCoalesce(chaitin) {
467 _ulr.initialize(_phc._lrg_map.max_lrg_id());
468 }
470 //------------------------------verify-----------------------------------------
471 void PhaseConservativeCoalesce::verify() {
472 #ifdef ASSERT
473 _phc.set_was_low();
474 #endif
475 }
477 //------------------------------union_helper-----------------------------------
478 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 ) {
479 // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the
480 // union-find tree
481 _phc.Union( lr1_node, lr2_node );
483 // Single-def live range ONLY if both live ranges are single-def.
484 // If both are single def, then src_def powers one live range
485 // and def_copy powers the other. After merging, src_def powers
486 // the combined live range.
487 lrgs(lr1)._def = (lrgs(lr1).is_multidef() ||
488 lrgs(lr2).is_multidef() )
489 ? NodeSentinel : src_def;
490 lrgs(lr2)._def = NULL; // No def for lrg 2
491 lrgs(lr2).Clear(); // Force empty mask for LRG 2
492 //lrgs(lr2)._size = 0; // Live-range 2 goes dead
493 lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop;
494 lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info
496 if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq)
497 lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq;
499 // Copy original value instead. Intermediate copies go dead, and
500 // the dst_copy becomes useless.
501 int didx = dst_copy->is_Copy();
502 dst_copy->set_req( didx, src_def );
503 // Add copy to free list
504 // _phc.free_spillcopy(b->_nodes[bindex]);
505 assert( b->_nodes[bindex] == dst_copy, "" );
506 dst_copy->replace_by( dst_copy->in(didx) );
507 dst_copy->set_req( didx, NULL);
508 b->_nodes.remove(bindex);
509 if( bindex < b->_ihrp_index ) b->_ihrp_index--;
510 if( bindex < b->_fhrp_index ) b->_fhrp_index--;
512 // Stretched lr1; add it to liveness of intermediate blocks
513 Block *b2 = _phc._cfg._bbs[src_copy->_idx];
514 while( b != b2 ) {
515 b = _phc._cfg._bbs[b->pred(1)->_idx];
516 _phc._live->live(b)->insert(lr1);
517 }
518 }
520 //------------------------------compute_separating_interferences---------------
521 // Factored code from copy_copy that computes extra interferences from
522 // lengthening a live range by double-coalescing.
523 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 ) {
525 assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj");
526 assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj");
527 Node *prev_copy = dst_copy->in(dst_copy->is_Copy());
528 Block *b2 = b;
529 uint bindex2 = bindex;
530 while( 1 ) {
531 // Find previous instruction
532 bindex2--; // Chain backwards 1 instruction
533 while( bindex2 == 0 ) { // At block start, find prior block
534 assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" );
535 b2 = _phc._cfg._bbs[b2->pred(1)->_idx];
536 bindex2 = b2->end_idx()-1;
537 }
538 // Get prior instruction
539 assert(bindex2 < b2->_nodes.size(), "index out of bounds");
540 Node *x = b2->_nodes[bindex2];
541 if( x == prev_copy ) { // Previous copy in copy chain?
542 if( prev_copy == src_copy)// Found end of chain and all interferences
543 break; // So break out of loop
544 // Else work back one in copy chain
545 prev_copy = prev_copy->in(prev_copy->is_Copy());
546 } else { // Else collect interferences
547 uint lidx = _phc._lrg_map.find(x);
548 // Found another def of live-range being stretched?
549 if(lidx == lr1) {
550 return max_juint;
551 }
552 if(lidx == lr2) {
553 return max_juint;
554 }
556 // If we attempt to coalesce across a bound def
557 if( lrgs(lidx).is_bound() ) {
558 // Do not let the coalesced LRG expect to get the bound color
559 rm.SUBTRACT( lrgs(lidx).mask() );
560 // Recompute rm_size
561 rm_size = rm.Size();
562 //if( rm._flags ) rm_size += 1000000;
563 if( reg_degree >= rm_size ) return max_juint;
564 }
565 if( rm.overlap(lrgs(lidx).mask()) ) {
566 // Insert lidx into union LRG; returns TRUE if actually inserted
567 if( _ulr.insert(lidx) ) {
568 // Infinite-stack neighbors do not alter colorability, as they
569 // can always color to some other color.
570 if( !lrgs(lidx).mask().is_AllStack() ) {
571 // If this coalesce will make any new neighbor uncolorable,
572 // do not coalesce.
573 if( lrgs(lidx).just_lo_degree() )
574 return max_juint;
575 // Bump our degree
576 if( ++reg_degree >= rm_size )
577 return max_juint;
578 } // End of if not infinite-stack neighbor
579 } // End of if actually inserted
580 } // End of if live range overlaps
581 } // End of else collect interferences for 1 node
582 } // End of while forever, scan back for interferences
583 return reg_degree;
584 }
586 //------------------------------update_ifg-------------------------------------
587 void PhaseConservativeCoalesce::update_ifg(uint lr1, uint lr2, IndexSet *n_lr1, IndexSet *n_lr2) {
588 // Some original neighbors of lr1 might have gone away
589 // because the constrained register mask prevented them.
590 // Remove lr1 from such neighbors.
591 IndexSetIterator one(n_lr1);
592 uint neighbor;
593 LRG &lrg1 = lrgs(lr1);
594 while ((neighbor = one.next()) != 0)
595 if( !_ulr.member(neighbor) )
596 if( _phc._ifg->neighbors(neighbor)->remove(lr1) )
597 lrgs(neighbor).inc_degree( -lrg1.compute_degree(lrgs(neighbor)) );
600 // lr2 is now called (coalesced into) lr1.
601 // Remove lr2 from the IFG.
602 IndexSetIterator two(n_lr2);
603 LRG &lrg2 = lrgs(lr2);
604 while ((neighbor = two.next()) != 0)
605 if( _phc._ifg->neighbors(neighbor)->remove(lr2) )
606 lrgs(neighbor).inc_degree( -lrg2.compute_degree(lrgs(neighbor)) );
608 // Some neighbors of intermediate copies now interfere with the
609 // combined live range.
610 IndexSetIterator three(&_ulr);
611 while ((neighbor = three.next()) != 0)
612 if( _phc._ifg->neighbors(neighbor)->insert(lr1) )
613 lrgs(neighbor).inc_degree( lrg1.compute_degree(lrgs(neighbor)) );
614 }
616 //------------------------------record_bias------------------------------------
617 static void record_bias( const PhaseIFG *ifg, int lr1, int lr2 ) {
618 // Tag copy bias here
619 if( !ifg->lrgs(lr1)._copy_bias )
620 ifg->lrgs(lr1)._copy_bias = lr2;
621 if( !ifg->lrgs(lr2)._copy_bias )
622 ifg->lrgs(lr2)._copy_bias = lr1;
623 }
625 //------------------------------copy_copy--------------------------------------
626 // See if I can coalesce a series of multiple copies together. I need the
627 // final dest copy and the original src copy. They can be the same Node.
628 // Compute the compatible register masks.
629 bool PhaseConservativeCoalesce::copy_copy(Node *dst_copy, Node *src_copy, Block *b, uint bindex) {
631 if (!dst_copy->is_SpillCopy()) {
632 return false;
633 }
634 if (!src_copy->is_SpillCopy()) {
635 return false;
636 }
637 Node *src_def = src_copy->in(src_copy->is_Copy());
638 uint lr1 = _phc._lrg_map.find(dst_copy);
639 uint lr2 = _phc._lrg_map.find(src_def);
641 // Same live ranges already?
642 if (lr1 == lr2) {
643 return false;
644 }
646 // Interfere?
647 if (_phc._ifg->test_edge_sq(lr1, lr2)) {
648 return false;
649 }
651 // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK.
652 if (!lrgs(lr1)._is_oop && lrgs(lr2)._is_oop) { // not an oop->int cast
653 return false;
654 }
656 // Coalescing between an aligned live range and a mis-aligned live range?
657 // No, no! Alignment changes how we count degree.
658 if (lrgs(lr1)._fat_proj != lrgs(lr2)._fat_proj) {
659 return false;
660 }
662 // Sort; use smaller live-range number
663 Node *lr1_node = dst_copy;
664 Node *lr2_node = src_def;
665 if (lr1 > lr2) {
666 uint tmp = lr1; lr1 = lr2; lr2 = tmp;
667 lr1_node = src_def; lr2_node = dst_copy;
668 }
670 // Check for compatibility of the 2 live ranges by
671 // intersecting their allowed register sets.
672 RegMask rm = lrgs(lr1).mask();
673 rm.AND(lrgs(lr2).mask());
674 // Number of bits free
675 uint rm_size = rm.Size();
677 if (UseFPUForSpilling && rm.is_AllStack() ) {
678 // Don't coalesce when frequency difference is large
679 Block *dst_b = _phc._cfg._bbs[dst_copy->_idx];
680 Block *src_def_b = _phc._cfg._bbs[src_def->_idx];
681 if (src_def_b->_freq > 10*dst_b->_freq )
682 return false;
683 }
685 // If we can use any stack slot, then effective size is infinite
686 if( rm.is_AllStack() ) rm_size += 1000000;
687 // Incompatible masks, no way to coalesce
688 if( rm_size == 0 ) return false;
690 // Another early bail-out test is when we are double-coalescing and the
691 // 2 copies are separated by some control flow.
692 if( dst_copy != src_copy ) {
693 Block *src_b = _phc._cfg._bbs[src_copy->_idx];
694 Block *b2 = b;
695 while( b2 != src_b ) {
696 if( b2->num_preds() > 2 ){// Found merge-point
697 _phc._lost_opp_cflow_coalesce++;
698 // extra record_bias commented out because Chris believes it is not
699 // productive. Since we can record only 1 bias, we want to choose one
700 // that stands a chance of working and this one probably does not.
701 //record_bias( _phc._lrgs, lr1, lr2 );
702 return false; // To hard to find all interferences
703 }
704 b2 = _phc._cfg._bbs[b2->pred(1)->_idx];
705 }
706 }
708 // Union the two interference sets together into '_ulr'
709 uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm );
711 if( reg_degree >= rm_size ) {
712 record_bias( _phc._ifg, lr1, lr2 );
713 return false;
714 }
716 // Now I need to compute all the interferences between dst_copy and
717 // src_copy. I'm not willing visit the entire interference graph, so
718 // I limit my search to things in dst_copy's block or in a straight
719 // line of previous blocks. I give up at merge points or when I get
720 // more interferences than my degree. I can stop when I find src_copy.
721 if( dst_copy != src_copy ) {
722 reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 );
723 if( reg_degree == max_juint ) {
724 record_bias( _phc._ifg, lr1, lr2 );
725 return false;
726 }
727 } // End of if dst_copy & src_copy are different
730 // ---- THE COMBINED LRG IS COLORABLE ----
732 // YEAH - Now coalesce this copy away
733 assert( lrgs(lr1).num_regs() == lrgs(lr2).num_regs(), "" );
735 IndexSet *n_lr1 = _phc._ifg->neighbors(lr1);
736 IndexSet *n_lr2 = _phc._ifg->neighbors(lr2);
738 // Update the interference graph
739 update_ifg(lr1, lr2, n_lr1, n_lr2);
741 _ulr.remove(lr1);
743 // Uncomment the following code to trace Coalescing in great detail.
744 //
745 //if (false) {
746 // tty->cr();
747 // tty->print_cr("#######################################");
748 // tty->print_cr("union %d and %d", lr1, lr2);
749 // n_lr1->dump();
750 // n_lr2->dump();
751 // tty->print_cr("resulting set is");
752 // _ulr.dump();
753 //}
755 // Replace n_lr1 with the new combined live range. _ulr will use
756 // n_lr1's old memory on the next iteration. n_lr2 is cleared to
757 // send its internal memory to the free list.
758 _ulr.swap(n_lr1);
759 _ulr.clear();
760 n_lr2->clear();
762 lrgs(lr1).set_degree( _phc._ifg->effective_degree(lr1) );
763 lrgs(lr2).set_degree( 0 );
765 // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the
766 // union-find tree
767 union_helper( lr1_node, lr2_node, lr1, lr2, src_def, dst_copy, src_copy, b, bindex );
768 // Combine register restrictions
769 lrgs(lr1).set_mask(rm);
770 lrgs(lr1).compute_set_mask_size();
771 lrgs(lr1)._cost += lrgs(lr2)._cost;
772 lrgs(lr1)._area += lrgs(lr2)._area;
774 // While its uncommon to successfully coalesce live ranges that started out
775 // being not-lo-degree, it can happen. In any case the combined coalesced
776 // live range better Simplify nicely.
777 lrgs(lr1)._was_lo = 1;
779 // kinda expensive to do all the time
780 //tty->print_cr("warning: slow verify happening");
781 //_phc._ifg->verify( &_phc );
782 return true;
783 }
785 //------------------------------coalesce---------------------------------------
786 // Conservative (but pessimistic) copy coalescing of a single block
787 void PhaseConservativeCoalesce::coalesce( Block *b ) {
788 // Bail out on infrequent blocks
789 if( b->is_uncommon(_phc._cfg._bbs) )
790 return;
791 // Check this block for copies.
792 for( uint i = 1; i<b->end_idx(); i++ ) {
793 // Check for actual copies on inputs. Coalesce a copy into its
794 // input if use and copy's input are compatible.
795 Node *copy1 = b->_nodes[i];
796 uint idx1 = copy1->is_Copy();
797 if( !idx1 ) continue; // Not a copy
799 if( copy_copy(copy1,copy1,b,i) ) {
800 i--; // Retry, same location in block
801 PhaseChaitin::_conserv_coalesce++; // Collect stats on success
802 continue;
803 }
804 }
805 }