35 #include "opto/regmask.hpp"
36
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 }
45
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());
65
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) );
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);
205
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 }
214
215 // Insert just after last use
216 b->_nodes.insert(last_use_idx+1,copy);
217 }
218
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);
269 n->replace_by(def);
270 n->set_req(cidx,NULL);
271 b->_nodes.remove(l);
272 l--;
273 continue;
274 }
275 }
276
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
314
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 }
348
349 } // End of is two-adr
350
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);
366
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
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
399
400 } // End of if Phi
401
402 } // End of for all instructions
403 } // End of for all blocks
404 }
405
406 //=============================================================================
407 //------------------------------coalesce---------------------------------------
408 // Aggressive (but pessimistic) copy coalescing of a single block
409
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.
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).
429
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
448
449
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 }
493 lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop;
494 lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info
495
496 if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq)
497 lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq;
498
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--;
511
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 }
519
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 ) {
524
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 }
555
659 return false;
660 }
661
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 }
669
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();
676
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 }
684
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;
689
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 }
707
708 // Union the two interference sets together into '_ulr'
709 uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm );
710
711 if( reg_degree >= rm_size ) {
712 record_bias( _phc._ifg, lr1, lr2 );
713 return false;
714 }
715
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 );
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;
773
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;
778
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 }
784
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
798
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 }
|
35 #include "opto/regmask.hpp"
36
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 }
45
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.get_block_for_node(b->pred(j))->_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());
65
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) );
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);
205
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.map_node_to_block(tmp, b);
212 last_use_idx++;
213 }
214
215 // Insert just after last use
216 b->_nodes.insert(last_use_idx+1,copy);
217 }
218
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);
269 n->replace_by(def);
270 n->set_req(cidx,NULL);
271 b->_nodes.remove(l);
272 l--;
273 continue;
274 }
275 }
276
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.get_block_for_node(b->pred(j));
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.map_node_to_block(copy, 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
314
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.map_node_to_block(copy, b);
347 }
348
349 } // End of is two-adr
350
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)) {
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);
366
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
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.map_node_to_block(copy, 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
399
400 } // End of if Phi
401
402 } // End of for all instructions
403 } // End of for all blocks
404 }
405
406 //=============================================================================
407 //------------------------------coalesce---------------------------------------
408 // Aggressive (but pessimistic) copy coalescing of a single block
409
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.
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).
429
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.get_block_for_node(bs->pred(j)) != b) {
441 j++;
442 }
443
444 // Visit all the Phis in successor block
445 for( uint k = 1; k<bs->_nodes.size(); k++ ) {
446 Node *n = bs->_nodes[k];
447 if( !n->is_Phi() ) break;
448 combine_these_two( n, n->in(j) );
449 }
450 } // End of for all successor blocks
451
452
453 // Check _this_ block for 2-address instructions and copies.
454 uint cnt = b->end_idx();
455 for( i = 1; i<cnt; i++ ) {
456 Node *n = b->_nodes[i];
457 uint idx;
458 // 2-address instructions have a virtual Copy matching their input
459 // to their output
460 if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) {
461 MachNode *mach = n->as_Mach();
462 combine_these_two(mach, mach->in(idx));
463 }
496 lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop;
497 lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info
498
499 if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq)
500 lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq;
501
502 // Copy original value instead. Intermediate copies go dead, and
503 // the dst_copy becomes useless.
504 int didx = dst_copy->is_Copy();
505 dst_copy->set_req( didx, src_def );
506 // Add copy to free list
507 // _phc.free_spillcopy(b->_nodes[bindex]);
508 assert( b->_nodes[bindex] == dst_copy, "" );
509 dst_copy->replace_by( dst_copy->in(didx) );
510 dst_copy->set_req( didx, NULL);
511 b->_nodes.remove(bindex);
512 if( bindex < b->_ihrp_index ) b->_ihrp_index--;
513 if( bindex < b->_fhrp_index ) b->_fhrp_index--;
514
515 // Stretched lr1; add it to liveness of intermediate blocks
516 Block *b2 = _phc._cfg.get_block_for_node(src_copy);
517 while( b != b2 ) {
518 b = _phc._cfg.get_block_for_node(b->pred(1));
519 _phc._live->live(b)->insert(lr1);
520 }
521 }
522
523 //------------------------------compute_separating_interferences---------------
524 // Factored code from copy_copy that computes extra interferences from
525 // lengthening a live range by double-coalescing.
526 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 ) {
527
528 assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj");
529 assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj");
530 Node *prev_copy = dst_copy->in(dst_copy->is_Copy());
531 Block *b2 = b;
532 uint bindex2 = bindex;
533 while( 1 ) {
534 // Find previous instruction
535 bindex2--; // Chain backwards 1 instruction
536 while( bindex2 == 0 ) { // At block start, find prior block
537 assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" );
538 b2 = _phc._cfg.get_block_for_node(b2->pred(1));
539 bindex2 = b2->end_idx()-1;
540 }
541 // Get prior instruction
542 assert(bindex2 < b2->_nodes.size(), "index out of bounds");
543 Node *x = b2->_nodes[bindex2];
544 if( x == prev_copy ) { // Previous copy in copy chain?
545 if( prev_copy == src_copy)// Found end of chain and all interferences
546 break; // So break out of loop
547 // Else work back one in copy chain
548 prev_copy = prev_copy->in(prev_copy->is_Copy());
549 } else { // Else collect interferences
550 uint lidx = _phc._lrg_map.find(x);
551 // Found another def of live-range being stretched?
552 if(lidx == lr1) {
553 return max_juint;
554 }
555 if(lidx == lr2) {
556 return max_juint;
557 }
558
662 return false;
663 }
664
665 // Sort; use smaller live-range number
666 Node *lr1_node = dst_copy;
667 Node *lr2_node = src_def;
668 if (lr1 > lr2) {
669 uint tmp = lr1; lr1 = lr2; lr2 = tmp;
670 lr1_node = src_def; lr2_node = dst_copy;
671 }
672
673 // Check for compatibility of the 2 live ranges by
674 // intersecting their allowed register sets.
675 RegMask rm = lrgs(lr1).mask();
676 rm.AND(lrgs(lr2).mask());
677 // Number of bits free
678 uint rm_size = rm.Size();
679
680 if (UseFPUForSpilling && rm.is_AllStack() ) {
681 // Don't coalesce when frequency difference is large
682 Block *dst_b = _phc._cfg.get_block_for_node(dst_copy);
683 Block *src_def_b = _phc._cfg.get_block_for_node(src_def);
684 if (src_def_b->_freq > 10*dst_b->_freq )
685 return false;
686 }
687
688 // If we can use any stack slot, then effective size is infinite
689 if( rm.is_AllStack() ) rm_size += 1000000;
690 // Incompatible masks, no way to coalesce
691 if( rm_size == 0 ) return false;
692
693 // Another early bail-out test is when we are double-coalescing and the
694 // 2 copies are separated by some control flow.
695 if( dst_copy != src_copy ) {
696 Block *src_b = _phc._cfg.get_block_for_node(src_copy);
697 Block *b2 = b;
698 while( b2 != src_b ) {
699 if( b2->num_preds() > 2 ){// Found merge-point
700 _phc._lost_opp_cflow_coalesce++;
701 // extra record_bias commented out because Chris believes it is not
702 // productive. Since we can record only 1 bias, we want to choose one
703 // that stands a chance of working and this one probably does not.
704 //record_bias( _phc._lrgs, lr1, lr2 );
705 return false; // To hard to find all interferences
706 }
707 b2 = _phc._cfg.get_block_for_node(b2->pred(1));
708 }
709 }
710
711 // Union the two interference sets together into '_ulr'
712 uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm );
713
714 if( reg_degree >= rm_size ) {
715 record_bias( _phc._ifg, lr1, lr2 );
716 return false;
717 }
718
719 // Now I need to compute all the interferences between dst_copy and
720 // src_copy. I'm not willing visit the entire interference graph, so
721 // I limit my search to things in dst_copy's block or in a straight
722 // line of previous blocks. I give up at merge points or when I get
723 // more interferences than my degree. I can stop when I find src_copy.
724 if( dst_copy != src_copy ) {
725 reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 );
726 if( reg_degree == max_juint ) {
727 record_bias( _phc._ifg, lr1, lr2 );
772 lrgs(lr1).set_mask(rm);
773 lrgs(lr1).compute_set_mask_size();
774 lrgs(lr1)._cost += lrgs(lr2)._cost;
775 lrgs(lr1)._area += lrgs(lr2)._area;
776
777 // While its uncommon to successfully coalesce live ranges that started out
778 // being not-lo-degree, it can happen. In any case the combined coalesced
779 // live range better Simplify nicely.
780 lrgs(lr1)._was_lo = 1;
781
782 // kinda expensive to do all the time
783 //tty->print_cr("warning: slow verify happening");
784 //_phc._ifg->verify( &_phc );
785 return true;
786 }
787
788 //------------------------------coalesce---------------------------------------
789 // Conservative (but pessimistic) copy coalescing of a single block
790 void PhaseConservativeCoalesce::coalesce( Block *b ) {
791 // Bail out on infrequent blocks
792 if (b->is_uncommon(&_phc._cfg)) {
793 return;
794 }
795 // Check this block for copies.
796 for( uint i = 1; i<b->end_idx(); i++ ) {
797 // Check for actual copies on inputs. Coalesce a copy into its
798 // input if use and copy's input are compatible.
799 Node *copy1 = b->_nodes[i];
800 uint idx1 = copy1->is_Copy();
801 if( !idx1 ) continue; // Not a copy
802
803 if( copy_copy(copy1,copy1,b,i) ) {
804 i--; // Retry, same location in block
805 PhaseChaitin::_conserv_coalesce++; // Collect stats on success
806 continue;
807 }
808 }
809 }
|