Wed, 10 Aug 2016 14:59:21 +0200
8152172: PPC64: Support AES intrinsics
Summary: Add support for AES intrinsics on PPC64.
Reviewed-by: kvn, mdoerr, simonis, zmajo
Contributed-by: Hiroshi H Horii <horii@jp.ibm.com>
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 #ifndef PRODUCT
38 void PhaseCoalesce::dump(Node *n) const {
39 // Being a const function means I cannot use 'Find'
40 uint r = _phc._lrg_map.find(n);
41 tty->print("L%d/N%d ",r,n->_idx);
42 }
44 void PhaseCoalesce::dump() const {
45 // I know I have a block layout now, so I can print blocks in a loop
46 for( uint i=0; i<_phc._cfg.number_of_blocks(); i++ ) {
47 uint j;
48 Block* b = _phc._cfg.get_block(i);
49 // Print a nice block header
50 tty->print("B%d: ",b->_pre_order);
51 for( j=1; j<b->num_preds(); j++ )
52 tty->print("B%d ", _phc._cfg.get_block_for_node(b->pred(j))->_pre_order);
53 tty->print("-> ");
54 for( j=0; j<b->_num_succs; j++ )
55 tty->print("B%d ",b->_succs[j]->_pre_order);
56 tty->print(" IDom: B%d/#%d\n", b->_idom ? b->_idom->_pre_order : 0, b->_dom_depth);
57 uint cnt = b->number_of_nodes();
58 for( j=0; j<cnt; j++ ) {
59 Node *n = b->get_node(j);
60 dump( n );
61 tty->print("\t%s\t",n->Name());
63 // Dump the inputs
64 uint k; // Exit value of loop
65 for( k=0; k<n->req(); k++ ) // For all required inputs
66 if( n->in(k) ) dump( n->in(k) );
67 else tty->print("_ ");
68 int any_prec = 0;
69 for( ; k<n->len(); k++ ) // For all precedence inputs
70 if( n->in(k) ) {
71 if( !any_prec++ ) tty->print(" |");
72 dump( n->in(k) );
73 }
75 // Dump node-specific info
76 n->dump_spec(tty);
77 tty->print("\n");
79 }
80 tty->print("\n");
81 }
82 }
83 #endif
85 // Combine the live ranges def'd by these 2 Nodes. N2 is an input to N1.
86 void PhaseCoalesce::combine_these_two(Node *n1, Node *n2) {
87 uint lr1 = _phc._lrg_map.find(n1);
88 uint lr2 = _phc._lrg_map.find(n2);
89 if( lr1 != lr2 && // Different live ranges already AND
90 !_phc._ifg->test_edge_sq( lr1, lr2 ) ) { // Do not interfere
91 LRG *lrg1 = &_phc.lrgs(lr1);
92 LRG *lrg2 = &_phc.lrgs(lr2);
93 // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK.
95 // Now, why is int->oop OK? We end up declaring a raw-pointer as an oop
96 // and in general that's a bad thing. However, int->oop conversions only
97 // happen at GC points, so the lifetime of the misclassified raw-pointer
98 // is from the CheckCastPP (that converts it to an oop) backwards up
99 // through a merge point and into the slow-path call, and around the
100 // diamond up to the heap-top check and back down into the slow-path call.
101 // The misclassified raw pointer is NOT live across the slow-path call,
102 // and so does not appear in any GC info, so the fact that it is
103 // misclassified is OK.
105 if( (lrg1->_is_oop || !lrg2->_is_oop) && // not an oop->int cast AND
106 // Compatible final mask
107 lrg1->mask().overlap( lrg2->mask() ) ) {
108 // Merge larger into smaller.
109 if( lr1 > lr2 ) {
110 uint tmp = lr1; lr1 = lr2; lr2 = tmp;
111 Node *n = n1; n1 = n2; n2 = n;
112 LRG *ltmp = lrg1; lrg1 = lrg2; lrg2 = ltmp;
113 }
114 // Union lr2 into lr1
115 _phc.Union( n1, n2 );
116 if (lrg1->_maxfreq < lrg2->_maxfreq)
117 lrg1->_maxfreq = lrg2->_maxfreq;
118 // Merge in the IFG
119 _phc._ifg->Union( lr1, lr2 );
120 // Combine register restrictions
121 lrg1->AND(lrg2->mask());
122 }
123 }
124 }
126 // Copy coalescing
127 void PhaseCoalesce::coalesce_driver() {
128 verify();
129 // Coalesce from high frequency to low
130 for (uint i = 0; i < _phc._cfg.number_of_blocks(); i++) {
131 coalesce(_phc._blks[i]);
132 }
133 }
135 // I am inserting copies to come out of SSA form. In the general case, I am
136 // doing a parallel renaming. I'm in the Named world now, so I can't do a
137 // general parallel renaming. All the copies now use "names" (live-ranges)
138 // to carry values instead of the explicit use-def chains. Suppose I need to
139 // insert 2 copies into the same block. They copy L161->L128 and L128->L132.
140 // If I insert them in the wrong order then L128 will get clobbered before it
141 // can get used by the second copy. This cannot happen in the SSA model;
142 // direct use-def chains get me the right value. It DOES happen in the named
143 // model so I have to handle the reordering of copies.
144 //
145 // In general, I need to topo-sort the placed copies to avoid conflicts.
146 // Its possible to have a closed cycle of copies (e.g., recirculating the same
147 // values around a loop). In this case I need a temp to break the cycle.
148 void PhaseAggressiveCoalesce::insert_copy_with_overlap( Block *b, Node *copy, uint dst_name, uint src_name ) {
150 // Scan backwards for the locations of the last use of the dst_name.
151 // I am about to clobber the dst_name, so the copy must be inserted
152 // after the last use. Last use is really first-use on a backwards scan.
153 uint i = b->end_idx()-1;
154 while(1) {
155 Node *n = b->get_node(i);
156 // Check for end of virtual copies; this is also the end of the
157 // parallel renaming effort.
158 if (n->_idx < _unique) {
159 break;
160 }
161 uint idx = n->is_Copy();
162 assert( idx || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" );
163 if (idx && _phc._lrg_map.find(n->in(idx)) == dst_name) {
164 break;
165 }
166 i--;
167 }
168 uint last_use_idx = i;
170 // Also search for any kill of src_name that exits the block.
171 // Since the copy uses src_name, I have to come before any kill.
172 uint kill_src_idx = b->end_idx();
173 // There can be only 1 kill that exits any block and that is
174 // the last kill. Thus it is the first kill on a backwards scan.
175 i = b->end_idx()-1;
176 while (1) {
177 Node *n = b->get_node(i);
178 // Check for end of virtual copies; this is also the end of the
179 // parallel renaming effort.
180 if (n->_idx < _unique) {
181 break;
182 }
183 assert( n->is_Copy() || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" );
184 if (_phc._lrg_map.find(n) == src_name) {
185 kill_src_idx = i;
186 break;
187 }
188 i--;
189 }
190 // Need a temp? Last use of dst comes after the kill of src?
191 if (last_use_idx >= kill_src_idx) {
192 // Need to break a cycle with a temp
193 uint idx = copy->is_Copy();
194 Node *tmp = copy->clone();
195 uint max_lrg_id = _phc._lrg_map.max_lrg_id();
196 _phc.new_lrg(tmp, max_lrg_id);
197 _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1);
199 // Insert new temp between copy and source
200 tmp ->set_req(idx,copy->in(idx));
201 copy->set_req(idx,tmp);
202 // Save source in temp early, before source is killed
203 b->insert_node(tmp, kill_src_idx);
204 _phc._cfg.map_node_to_block(tmp, b);
205 last_use_idx++;
206 }
208 // Insert just after last use
209 b->insert_node(copy, last_use_idx + 1);
210 }
212 void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) {
213 // We do LRGs compressing and fix a liveout data only here since the other
214 // place in Split() is guarded by the assert which we never hit.
215 _phc._lrg_map.compress_uf_map_for_nodes();
216 // Fix block's liveout data for compressed live ranges.
217 for (uint lrg = 1; lrg < _phc._lrg_map.max_lrg_id(); lrg++) {
218 uint compressed_lrg = _phc._lrg_map.find(lrg);
219 if (lrg != compressed_lrg) {
220 for (uint bidx = 0; bidx < _phc._cfg.number_of_blocks(); bidx++) {
221 IndexSet *liveout = _phc._live->live(_phc._cfg.get_block(bidx));
222 if (liveout->member(lrg)) {
223 liveout->remove(lrg);
224 liveout->insert(compressed_lrg);
225 }
226 }
227 }
228 }
230 // All new nodes added are actual copies to replace virtual copies.
231 // Nodes with index less than '_unique' are original, non-virtual Nodes.
232 _unique = C->unique();
234 for (uint i = 0; i < _phc._cfg.number_of_blocks(); i++) {
235 C->check_node_count(NodeLimitFudgeFactor, "out of nodes in coalesce");
236 if (C->failing()) return;
237 Block *b = _phc._cfg.get_block(i);
238 uint cnt = b->num_preds(); // Number of inputs to the Phi
240 for( uint l = 1; l<b->number_of_nodes(); l++ ) {
241 Node *n = b->get_node(l);
243 // Do not use removed-copies, use copied value instead
244 uint ncnt = n->req();
245 for( uint k = 1; k<ncnt; k++ ) {
246 Node *copy = n->in(k);
247 uint cidx = copy->is_Copy();
248 if( cidx ) {
249 Node *def = copy->in(cidx);
250 if (_phc._lrg_map.find(copy) == _phc._lrg_map.find(def)) {
251 n->set_req(k, def);
252 }
253 }
254 }
256 // Remove any explicit copies that get coalesced.
257 uint cidx = n->is_Copy();
258 if( cidx ) {
259 Node *def = n->in(cidx);
260 if (_phc._lrg_map.find(n) == _phc._lrg_map.find(def)) {
261 n->replace_by(def);
262 n->set_req(cidx,NULL);
263 b->remove_node(l);
264 l--;
265 continue;
266 }
267 }
269 if (n->is_Phi()) {
270 // Get the chosen name for the Phi
271 uint phi_name = _phc._lrg_map.find(n);
272 // Ignore the pre-allocated specials
273 if (!phi_name) {
274 continue;
275 }
276 // Check for mismatch inputs to Phi
277 for (uint j = 1; j < cnt; j++) {
278 Node *m = n->in(j);
279 uint src_name = _phc._lrg_map.find(m);
280 if (src_name != phi_name) {
281 Block *pred = _phc._cfg.get_block_for_node(b->pred(j));
282 Node *copy;
283 assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach");
284 // Rematerialize constants instead of copying them.
285 // We do this only for immediate constants, we avoid constant table loads
286 // because that will unsafely extend the live range of the constant table base.
287 if (m->is_Mach() && m->as_Mach()->is_Con() && !m->as_Mach()->is_MachConstant() &&
288 m->as_Mach()->rematerialize()) {
289 copy = m->clone();
290 // Insert the copy in the predecessor basic block
291 pred->add_inst(copy);
292 // Copy any flags as well
293 _phc.clone_projs(pred, pred->end_idx(), m, copy, _phc._lrg_map);
294 } else {
295 const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()];
296 copy = new (C) MachSpillCopyNode(m, *rm, *rm);
297 // Find a good place to insert. Kinda tricky, use a subroutine
298 insert_copy_with_overlap(pred,copy,phi_name,src_name);
299 }
300 // Insert the copy in the use-def chain
301 n->set_req(j, copy);
302 _phc._cfg.map_node_to_block(copy, pred);
303 // Extend ("register allocate") the names array for the copy.
304 _phc._lrg_map.extend(copy->_idx, phi_name);
305 } // End of if Phi names do not match
306 } // End of for all inputs to Phi
307 } else { // End of if Phi
309 // Now check for 2-address instructions
310 uint idx;
311 if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) {
312 // Get the chosen name for the Node
313 uint name = _phc._lrg_map.find(n);
314 assert (name, "no 2-address specials");
315 // Check for name mis-match on the 2-address input
316 Node *m = n->in(idx);
317 if (_phc._lrg_map.find(m) != name) {
318 Node *copy;
319 assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach");
320 // At this point it is unsafe to extend live ranges (6550579).
321 // Rematerialize only constants as we do for Phi above.
322 if (m->is_Mach() && m->as_Mach()->is_Con() && !m->as_Mach()->is_MachConstant() &&
323 m->as_Mach()->rematerialize()) {
324 copy = m->clone();
325 // Insert the copy in the basic block, just before us
326 b->insert_node(copy, l++);
327 l += _phc.clone_projs(b, l, m, copy, _phc._lrg_map);
328 } else {
329 const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()];
330 copy = new (C) MachSpillCopyNode(m, *rm, *rm);
331 // Insert the copy in the basic block, just before us
332 b->insert_node(copy, l++);
333 }
334 // Insert the copy in the use-def chain
335 n->set_req(idx, copy);
336 // Extend ("register allocate") the names array for the copy.
337 _phc._lrg_map.extend(copy->_idx, name);
338 _phc._cfg.map_node_to_block(copy, b);
339 }
341 } // End of is two-adr
343 // Insert a copy at a debug use for a lrg which has high frequency
344 if (b->_freq < OPTO_DEBUG_SPLIT_FREQ || _phc._cfg.is_uncommon(b)) {
345 // Walk the debug inputs to the node and check for lrg freq
346 JVMState* jvms = n->jvms();
347 uint debug_start = jvms ? jvms->debug_start() : 999999;
348 uint debug_end = jvms ? jvms->debug_end() : 999999;
349 for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) {
350 // Do not split monitors; they are only needed for debug table
351 // entries and need no code.
352 if (jvms->is_monitor_use(inpidx)) {
353 continue;
354 }
355 Node *inp = n->in(inpidx);
356 uint nidx = _phc._lrg_map.live_range_id(inp);
357 LRG &lrg = lrgs(nidx);
359 // If this lrg has a high frequency use/def
360 if( lrg._maxfreq >= _phc.high_frequency_lrg() ) {
361 // If the live range is also live out of this block (like it
362 // would be for a fast/slow idiom), the normal spill mechanism
363 // does an excellent job. If it is not live out of this block
364 // (like it would be for debug info to uncommon trap) splitting
365 // the live range now allows a better allocation in the high
366 // frequency blocks.
367 // Build_IFG_virtual has converted the live sets to
368 // live-IN info, not live-OUT info.
369 uint k;
370 for( k=0; k < b->_num_succs; k++ )
371 if( _phc._live->live(b->_succs[k])->member( nidx ) )
372 break; // Live in to some successor block?
373 if( k < b->_num_succs )
374 continue; // Live out; do not pre-split
375 // Split the lrg at this use
376 const RegMask *rm = C->matcher()->idealreg2spillmask[inp->ideal_reg()];
377 Node *copy = new (C) MachSpillCopyNode( inp, *rm, *rm );
378 // Insert the copy in the use-def chain
379 n->set_req(inpidx, copy );
380 // Insert the copy in the basic block, just before us
381 b->insert_node(copy, l++);
382 // Extend ("register allocate") the names array for the copy.
383 uint max_lrg_id = _phc._lrg_map.max_lrg_id();
384 _phc.new_lrg(copy, max_lrg_id);
385 _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1);
386 _phc._cfg.map_node_to_block(copy, b);
387 //tty->print_cr("Split a debug use in Aggressive Coalesce");
388 } // End of if high frequency use/def
389 } // End of for all debug inputs
390 } // End of if low frequency safepoint
392 } // End of if Phi
394 } // End of for all instructions
395 } // End of for all blocks
396 }
399 // Aggressive (but pessimistic) copy coalescing of a single block
401 // The following coalesce pass represents a single round of aggressive
402 // pessimistic coalesce. "Aggressive" means no attempt to preserve
403 // colorability when coalescing. This occasionally means more spills, but
404 // it also means fewer rounds of coalescing for better code - and that means
405 // faster compiles.
407 // "Pessimistic" means we do not hit the fixed point in one pass (and we are
408 // reaching for the least fixed point to boot). This is typically solved
409 // with a few more rounds of coalescing, but the compiler must run fast. We
410 // could optimistically coalescing everything touching PhiNodes together
411 // into one big live range, then check for self-interference. Everywhere
412 // the live range interferes with self it would have to be split. Finding
413 // the right split points can be done with some heuristics (based on
414 // expected frequency of edges in the live range). In short, it's a real
415 // research problem and the timeline is too short to allow such research.
416 // Further thoughts: (1) build the LR in a pass, (2) find self-interference
417 // in another pass, (3) per each self-conflict, split, (4) split by finding
418 // the low-cost cut (min-cut) of the LR, (5) edges in the LR are weighted
419 // according to the GCM algorithm (or just exec freq on CFG edges).
421 void PhaseAggressiveCoalesce::coalesce( Block *b ) {
422 // Copies are still "virtual" - meaning we have not made them explicitly
423 // copies. Instead, Phi functions of successor blocks have mis-matched
424 // live-ranges. If I fail to coalesce, I'll have to insert a copy to line
425 // up the live-ranges. Check for Phis in successor blocks.
426 uint i;
427 for( i=0; i<b->_num_succs; i++ ) {
428 Block *bs = b->_succs[i];
429 // Find index of 'b' in 'bs' predecessors
430 uint j=1;
431 while (_phc._cfg.get_block_for_node(bs->pred(j)) != b) {
432 j++;
433 }
435 // Visit all the Phis in successor block
436 for( uint k = 1; k<bs->number_of_nodes(); k++ ) {
437 Node *n = bs->get_node(k);
438 if( !n->is_Phi() ) break;
439 combine_these_two( n, n->in(j) );
440 }
441 } // End of for all successor blocks
444 // Check _this_ block for 2-address instructions and copies.
445 uint cnt = b->end_idx();
446 for( i = 1; i<cnt; i++ ) {
447 Node *n = b->get_node(i);
448 uint idx;
449 // 2-address instructions have a virtual Copy matching their input
450 // to their output
451 if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) {
452 MachNode *mach = n->as_Mach();
453 combine_these_two(mach, mach->in(idx));
454 }
455 } // End of for all instructions in block
456 }
458 PhaseConservativeCoalesce::PhaseConservativeCoalesce(PhaseChaitin &chaitin) : PhaseCoalesce(chaitin) {
459 _ulr.initialize(_phc._lrg_map.max_lrg_id());
460 }
462 void PhaseConservativeCoalesce::verify() {
463 #ifdef ASSERT
464 _phc.set_was_low();
465 #endif
466 }
468 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 ) {
469 // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the
470 // union-find tree
471 _phc.Union( lr1_node, lr2_node );
473 // Single-def live range ONLY if both live ranges are single-def.
474 // If both are single def, then src_def powers one live range
475 // and def_copy powers the other. After merging, src_def powers
476 // the combined live range.
477 lrgs(lr1)._def = (lrgs(lr1).is_multidef() ||
478 lrgs(lr2).is_multidef() )
479 ? NodeSentinel : src_def;
480 lrgs(lr2)._def = NULL; // No def for lrg 2
481 lrgs(lr2).Clear(); // Force empty mask for LRG 2
482 //lrgs(lr2)._size = 0; // Live-range 2 goes dead
483 lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop;
484 lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info
486 if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq)
487 lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq;
489 // Copy original value instead. Intermediate copies go dead, and
490 // the dst_copy becomes useless.
491 int didx = dst_copy->is_Copy();
492 dst_copy->set_req( didx, src_def );
493 // Add copy to free list
494 // _phc.free_spillcopy(b->_nodes[bindex]);
495 assert( b->get_node(bindex) == dst_copy, "" );
496 dst_copy->replace_by( dst_copy->in(didx) );
497 dst_copy->set_req( didx, NULL);
498 b->remove_node(bindex);
499 if( bindex < b->_ihrp_index ) b->_ihrp_index--;
500 if( bindex < b->_fhrp_index ) b->_fhrp_index--;
502 // Stretched lr1; add it to liveness of intermediate blocks
503 Block *b2 = _phc._cfg.get_block_for_node(src_copy);
504 while( b != b2 ) {
505 b = _phc._cfg.get_block_for_node(b->pred(1));
506 _phc._live->live(b)->insert(lr1);
507 }
508 }
510 // Factored code from copy_copy that computes extra interferences from
511 // lengthening a live range by double-coalescing.
512 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 ) {
514 assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj");
515 assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj");
516 Node *prev_copy = dst_copy->in(dst_copy->is_Copy());
517 Block *b2 = b;
518 uint bindex2 = bindex;
519 while( 1 ) {
520 // Find previous instruction
521 bindex2--; // Chain backwards 1 instruction
522 while( bindex2 == 0 ) { // At block start, find prior block
523 assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" );
524 b2 = _phc._cfg.get_block_for_node(b2->pred(1));
525 bindex2 = b2->end_idx()-1;
526 }
527 // Get prior instruction
528 assert(bindex2 < b2->number_of_nodes(), "index out of bounds");
529 Node *x = b2->get_node(bindex2);
530 if( x == prev_copy ) { // Previous copy in copy chain?
531 if( prev_copy == src_copy)// Found end of chain and all interferences
532 break; // So break out of loop
533 // Else work back one in copy chain
534 prev_copy = prev_copy->in(prev_copy->is_Copy());
535 } else { // Else collect interferences
536 uint lidx = _phc._lrg_map.find(x);
537 // Found another def of live-range being stretched?
538 if(lidx == lr1) {
539 return max_juint;
540 }
541 if(lidx == lr2) {
542 return max_juint;
543 }
545 // If we attempt to coalesce across a bound def
546 if( lrgs(lidx).is_bound() ) {
547 // Do not let the coalesced LRG expect to get the bound color
548 rm.SUBTRACT( lrgs(lidx).mask() );
549 // Recompute rm_size
550 rm_size = rm.Size();
551 //if( rm._flags ) rm_size += 1000000;
552 if( reg_degree >= rm_size ) return max_juint;
553 }
554 if( rm.overlap(lrgs(lidx).mask()) ) {
555 // Insert lidx into union LRG; returns TRUE if actually inserted
556 if( _ulr.insert(lidx) ) {
557 // Infinite-stack neighbors do not alter colorability, as they
558 // can always color to some other color.
559 if( !lrgs(lidx).mask().is_AllStack() ) {
560 // If this coalesce will make any new neighbor uncolorable,
561 // do not coalesce.
562 if( lrgs(lidx).just_lo_degree() )
563 return max_juint;
564 // Bump our degree
565 if( ++reg_degree >= rm_size )
566 return max_juint;
567 } // End of if not infinite-stack neighbor
568 } // End of if actually inserted
569 } // End of if live range overlaps
570 } // End of else collect interferences for 1 node
571 } // End of while forever, scan back for interferences
572 return reg_degree;
573 }
575 void PhaseConservativeCoalesce::update_ifg(uint lr1, uint lr2, IndexSet *n_lr1, IndexSet *n_lr2) {
576 // Some original neighbors of lr1 might have gone away
577 // because the constrained register mask prevented them.
578 // Remove lr1 from such neighbors.
579 IndexSetIterator one(n_lr1);
580 uint neighbor;
581 LRG &lrg1 = lrgs(lr1);
582 while ((neighbor = one.next()) != 0)
583 if( !_ulr.member(neighbor) )
584 if( _phc._ifg->neighbors(neighbor)->remove(lr1) )
585 lrgs(neighbor).inc_degree( -lrg1.compute_degree(lrgs(neighbor)) );
588 // lr2 is now called (coalesced into) lr1.
589 // Remove lr2 from the IFG.
590 IndexSetIterator two(n_lr2);
591 LRG &lrg2 = lrgs(lr2);
592 while ((neighbor = two.next()) != 0)
593 if( _phc._ifg->neighbors(neighbor)->remove(lr2) )
594 lrgs(neighbor).inc_degree( -lrg2.compute_degree(lrgs(neighbor)) );
596 // Some neighbors of intermediate copies now interfere with the
597 // combined live range.
598 IndexSetIterator three(&_ulr);
599 while ((neighbor = three.next()) != 0)
600 if( _phc._ifg->neighbors(neighbor)->insert(lr1) )
601 lrgs(neighbor).inc_degree( lrg1.compute_degree(lrgs(neighbor)) );
602 }
604 static void record_bias( const PhaseIFG *ifg, int lr1, int lr2 ) {
605 // Tag copy bias here
606 if( !ifg->lrgs(lr1)._copy_bias )
607 ifg->lrgs(lr1)._copy_bias = lr2;
608 if( !ifg->lrgs(lr2)._copy_bias )
609 ifg->lrgs(lr2)._copy_bias = lr1;
610 }
612 // See if I can coalesce a series of multiple copies together. I need the
613 // final dest copy and the original src copy. They can be the same Node.
614 // Compute the compatible register masks.
615 bool PhaseConservativeCoalesce::copy_copy(Node *dst_copy, Node *src_copy, Block *b, uint bindex) {
617 if (!dst_copy->is_SpillCopy()) {
618 return false;
619 }
620 if (!src_copy->is_SpillCopy()) {
621 return false;
622 }
623 Node *src_def = src_copy->in(src_copy->is_Copy());
624 uint lr1 = _phc._lrg_map.find(dst_copy);
625 uint lr2 = _phc._lrg_map.find(src_def);
627 // Same live ranges already?
628 if (lr1 == lr2) {
629 return false;
630 }
632 // Interfere?
633 if (_phc._ifg->test_edge_sq(lr1, lr2)) {
634 return false;
635 }
637 // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK.
638 if (!lrgs(lr1)._is_oop && lrgs(lr2)._is_oop) { // not an oop->int cast
639 return false;
640 }
642 // Coalescing between an aligned live range and a mis-aligned live range?
643 // No, no! Alignment changes how we count degree.
644 if (lrgs(lr1)._fat_proj != lrgs(lr2)._fat_proj) {
645 return false;
646 }
648 // Sort; use smaller live-range number
649 Node *lr1_node = dst_copy;
650 Node *lr2_node = src_def;
651 if (lr1 > lr2) {
652 uint tmp = lr1; lr1 = lr2; lr2 = tmp;
653 lr1_node = src_def; lr2_node = dst_copy;
654 }
656 // Check for compatibility of the 2 live ranges by
657 // intersecting their allowed register sets.
658 RegMask rm = lrgs(lr1).mask();
659 rm.AND(lrgs(lr2).mask());
660 // Number of bits free
661 uint rm_size = rm.Size();
663 if (UseFPUForSpilling && rm.is_AllStack() ) {
664 // Don't coalesce when frequency difference is large
665 Block *dst_b = _phc._cfg.get_block_for_node(dst_copy);
666 Block *src_def_b = _phc._cfg.get_block_for_node(src_def);
667 if (src_def_b->_freq > 10*dst_b->_freq )
668 return false;
669 }
671 // If we can use any stack slot, then effective size is infinite
672 if( rm.is_AllStack() ) rm_size += 1000000;
673 // Incompatible masks, no way to coalesce
674 if( rm_size == 0 ) return false;
676 // Another early bail-out test is when we are double-coalescing and the
677 // 2 copies are separated by some control flow.
678 if( dst_copy != src_copy ) {
679 Block *src_b = _phc._cfg.get_block_for_node(src_copy);
680 Block *b2 = b;
681 while( b2 != src_b ) {
682 if( b2->num_preds() > 2 ){// Found merge-point
683 _phc._lost_opp_cflow_coalesce++;
684 // extra record_bias commented out because Chris believes it is not
685 // productive. Since we can record only 1 bias, we want to choose one
686 // that stands a chance of working and this one probably does not.
687 //record_bias( _phc._lrgs, lr1, lr2 );
688 return false; // To hard to find all interferences
689 }
690 b2 = _phc._cfg.get_block_for_node(b2->pred(1));
691 }
692 }
694 // Union the two interference sets together into '_ulr'
695 uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm );
697 if( reg_degree >= rm_size ) {
698 record_bias( _phc._ifg, lr1, lr2 );
699 return false;
700 }
702 // Now I need to compute all the interferences between dst_copy and
703 // src_copy. I'm not willing visit the entire interference graph, so
704 // I limit my search to things in dst_copy's block or in a straight
705 // line of previous blocks. I give up at merge points or when I get
706 // more interferences than my degree. I can stop when I find src_copy.
707 if( dst_copy != src_copy ) {
708 reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 );
709 if( reg_degree == max_juint ) {
710 record_bias( _phc._ifg, lr1, lr2 );
711 return false;
712 }
713 } // End of if dst_copy & src_copy are different
716 // ---- THE COMBINED LRG IS COLORABLE ----
718 // YEAH - Now coalesce this copy away
719 assert( lrgs(lr1).num_regs() == lrgs(lr2).num_regs(), "" );
721 IndexSet *n_lr1 = _phc._ifg->neighbors(lr1);
722 IndexSet *n_lr2 = _phc._ifg->neighbors(lr2);
724 // Update the interference graph
725 update_ifg(lr1, lr2, n_lr1, n_lr2);
727 _ulr.remove(lr1);
729 // Uncomment the following code to trace Coalescing in great detail.
730 //
731 //if (false) {
732 // tty->cr();
733 // tty->print_cr("#######################################");
734 // tty->print_cr("union %d and %d", lr1, lr2);
735 // n_lr1->dump();
736 // n_lr2->dump();
737 // tty->print_cr("resulting set is");
738 // _ulr.dump();
739 //}
741 // Replace n_lr1 with the new combined live range. _ulr will use
742 // n_lr1's old memory on the next iteration. n_lr2 is cleared to
743 // send its internal memory to the free list.
744 _ulr.swap(n_lr1);
745 _ulr.clear();
746 n_lr2->clear();
748 lrgs(lr1).set_degree( _phc._ifg->effective_degree(lr1) );
749 lrgs(lr2).set_degree( 0 );
751 // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the
752 // union-find tree
753 union_helper( lr1_node, lr2_node, lr1, lr2, src_def, dst_copy, src_copy, b, bindex );
754 // Combine register restrictions
755 lrgs(lr1).set_mask(rm);
756 lrgs(lr1).compute_set_mask_size();
757 lrgs(lr1)._cost += lrgs(lr2)._cost;
758 lrgs(lr1)._area += lrgs(lr2)._area;
760 // While its uncommon to successfully coalesce live ranges that started out
761 // being not-lo-degree, it can happen. In any case the combined coalesced
762 // live range better Simplify nicely.
763 lrgs(lr1)._was_lo = 1;
765 // kinda expensive to do all the time
766 //tty->print_cr("warning: slow verify happening");
767 //_phc._ifg->verify( &_phc );
768 return true;
769 }
771 // Conservative (but pessimistic) copy coalescing of a single block
772 void PhaseConservativeCoalesce::coalesce( Block *b ) {
773 // Bail out on infrequent blocks
774 if (_phc._cfg.is_uncommon(b)) {
775 return;
776 }
777 // Check this block for copies.
778 for( uint i = 1; i<b->end_idx(); i++ ) {
779 // Check for actual copies on inputs. Coalesce a copy into its
780 // input if use and copy's input are compatible.
781 Node *copy1 = b->get_node(i);
782 uint idx1 = copy1->is_Copy();
783 if( !idx1 ) continue; // Not a copy
785 if( copy_copy(copy1,copy1,b,i) ) {
786 i--; // Retry, same location in block
787 PhaseChaitin::_conserv_coalesce++; // Collect stats on success
788 continue;
789 }
790 }
791 }