1 /*
  2  * Copyright (c) 1998, 2020, Oracle and/or its affiliates. All rights reserved.
  3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  4  *
  5  * This code is free software; you can redistribute it and/or modify it
  6  * under the terms of the GNU General Public License version 2 only, as
  7  * published by the Free Software Foundation.
  8  *
  9  * This code is distributed in the hope that it will be useful, but WITHOUT
 10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 12  * version 2 for more details (a copy is included in the LICENSE file that
 13  * accompanied this code).
 14  *
 15  * You should have received a copy of the GNU General Public License version
 16  * 2 along with this work; if not, write to the Free Software Foundation,
 17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 18  *
 19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 20  * or visit www.oracle.com if you need additional information or have any
 21  * questions.
 22  *
 23  */
 24 
 25 #include "precompiled.hpp"
 26 #include "memory/allocation.inline.hpp"
 27 #include "memory/resourceArea.hpp"
 28 #include "opto/chaitin.hpp"
 29 #include "opto/machnode.hpp"
 30 
 31 // See if this register (or pairs, or vector) already contains the value.
 32 static bool register_contains_value(Node* val, OptoReg::Name reg, int n_regs,
 33                                     Node_List& value) {
 34   for (int i = 0; i < n_regs; i++) {
 35     OptoReg::Name nreg = OptoReg::add(reg,-i);
 36     if (value[nreg] != val)
 37       return false;
 38   }
 39   return true;
 40 }
 41 
 42 //---------------------------may_be_copy_of_callee-----------------------------
 43 // Check to see if we can possibly be a copy of a callee-save value.
 44 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const {
 45   // Short circuit if there are no callee save registers
 46   if (_matcher.number_of_saved_registers() == 0) return false;
 47 
 48   // Expect only a spill-down and reload on exit for callee-save spills.
 49   // Chains of copies cannot be deep.
 50   // 5008997 - This is wishful thinking. Register allocator seems to
 51   // be splitting live ranges for callee save registers to such
 52   // an extent that in large methods the chains can be very long
 53   // (50+). The conservative answer is to return true if we don't
 54   // know as this prevents optimizations from occurring.
 55 
 56   const int limit = 60;
 57   int i;
 58   for( i=0; i < limit; i++ ) {
 59     if( def->is_Proj() && def->in(0)->is_Start() &&
 60         _matcher.is_save_on_entry(lrgs(_lrg_map.live_range_id(def)).reg()))
 61       return true;              // Direct use of callee-save proj
 62     if( def->is_Copy() )        // Copies carry value through
 63       def = def->in(def->is_Copy());
 64     else if( def->is_Phi() )    // Phis can merge it from any direction
 65       def = def->in(1);
 66     else
 67       break;
 68     guarantee(def != NULL, "must not resurrect dead copy");
 69   }
 70   // If we reached the end and didn't find a callee save proj
 71   // then this may be a callee save proj so we return true
 72   // as the conservative answer. If we didn't reach then end
 73   // we must have discovered that it was not a callee save
 74   // else we would have returned.
 75   return i == limit;
 76 }
 77 
 78 //------------------------------yank-----------------------------------
 79 // Helper function for yank_if_dead
 80 int PhaseChaitin::yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) {
 81   int blk_adjust=0;
 82   Block *oldb = _cfg.get_block_for_node(old);
 83   oldb->find_remove(old);
 84   // Count 1 if deleting an instruction from the current block
 85   if (oldb == current_block) {
 86     blk_adjust++;
 87   }
 88   _cfg.unmap_node_from_block(old);
 89   OptoReg::Name old_reg = lrgs(_lrg_map.live_range_id(old)).reg();
 90   if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available?
 91     value->map(old_reg,NULL);  // Yank from value/regnd maps
 92     regnd->map(old_reg,NULL);  // This register's value is now unknown
 93   }
 94   return blk_adjust;
 95 }
 96 
 97 #ifdef ASSERT
 98 static bool expected_yanked_node(Node *old, Node *orig_old) {
 99   // This code is expected only next original nodes:
100   // - load from constant table node which may have next data input nodes:
101   //     MachConstantBase, MachTemp, MachSpillCopy
102   // - Phi nodes that are considered Junk
103   // - load constant node which may have next data input nodes:
104   //     MachTemp, MachSpillCopy
105   // - MachSpillCopy
106   // - MachProj and Copy dead nodes
107   if (old->is_MachSpillCopy()) {
108     return true;
109   } else if (old->is_Con()) {
110     return true;
111   } else if (old->is_MachProj()) { // Dead kills projection of Con node
112     return (old == orig_old);
113   } else if (old->is_Copy()) {     // Dead copy of a callee-save value
114     return (old == orig_old);
115   } else if (old->is_MachTemp()) {
116     return orig_old->is_Con();
117   } else if (old->is_Phi()) { // Junk phi's
118     return true;
119   } else if (old->is_MachConstantBase()) {
120     return (orig_old->is_Con() && orig_old->is_MachConstant());
121   }
122   return false;
123 }
124 #endif
125 
126 //------------------------------yank_if_dead-----------------------------------
127 // Removed edges from 'old'.  Yank if dead.  Return adjustment counts to
128 // iterators in the current block.
129 int PhaseChaitin::yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block,
130                                        Node_List *value, Node_List *regnd) {
131   int blk_adjust=0;
132   if (old->outcnt() == 0 && old != C->top()) {
133 #ifdef ASSERT
134     if (!expected_yanked_node(old, orig_old)) {
135       tty->print_cr("==============================================");
136       tty->print_cr("orig_old:");
137       orig_old->dump();
138       tty->print_cr("old:");
139       old->dump();
140       assert(false, "unexpected yanked node");
141     }
142     if (old->is_Con())
143       orig_old = old; // Reset to satisfy expected nodes checks.
144 #endif
145     blk_adjust += yank(old, current_block, value, regnd);
146 
147     for (uint i = 1; i < old->req(); i++) {
148       Node* n = old->in(i);
149       if (n != NULL) {
150         old->set_req(i, NULL);
151         blk_adjust += yank_if_dead_recurse(n, orig_old, current_block, value, regnd);
152       }
153     }
154     // Disconnect control and remove precedence edges if any exist
155     old->disconnect_inputs(NULL, C);
156   }
157   return blk_adjust;
158 }
159 
160 //------------------------------use_prior_register-----------------------------
161 // Use the prior value instead of the current value, in an effort to make
162 // the current value go dead.  Return block iterator adjustment, in case
163 // we yank some instructions from this block.
164 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List &regnd ) {
165   // No effect?
166   if( def == n->in(idx) ) return 0;
167   // Def is currently dead and can be removed?  Do not resurrect
168   if( def->outcnt() == 0 ) return 0;
169 
170   // Not every pair of physical registers are assignment compatible,
171   // e.g. on sparc floating point registers are not assignable to integer
172   // registers.
173   const LRG &def_lrg = lrgs(_lrg_map.live_range_id(def));
174   OptoReg::Name def_reg = def_lrg.reg();
175   const RegMask &use_mask = n->in_RegMask(idx);
176   bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0)
177                                                    : (use_mask.is_AllStack() != 0));
178   if (!RegMask::is_vector(def->ideal_reg())) {
179     // Check for a copy to or from a misaligned pair.
180     // It is workaround for a sparc with misaligned pairs.
181     can_use = can_use && !use_mask.is_misaligned_pair() && !def_lrg.mask().is_misaligned_pair();
182   }
183   if (!can_use)
184     return 0;
185 
186   // Capture the old def in case it goes dead...
187   Node *old = n->in(idx);
188 
189   // Save-on-call copies can only be elided if the entire copy chain can go
190   // away, lest we get the same callee-save value alive in 2 locations at
191   // once.  We check for the obvious trivial case here.  Although it can
192   // sometimes be elided with cooperation outside our scope, here we will just
193   // miss the opportunity.  :-(
194   if( may_be_copy_of_callee(def) ) {
195     if( old->outcnt() > 1 ) return 0; // We're the not last user
196     int idx = old->is_Copy();
197     assert( idx, "chain of copies being removed" );
198     Node *old2 = old->in(idx);  // Chain of copies
199     if( old2->outcnt() > 1 ) return 0; // old is not the last user
200     int idx2 = old2->is_Copy();
201     if( !idx2 ) return 0;       // Not a chain of 2 copies
202     if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies
203   }
204 
205   // Use the new def
206   n->set_req(idx,def);
207   _post_alloc++;
208 
209   // Is old def now dead?  We successfully yanked a copy?
210   return yank_if_dead(old,current_block,&value,&regnd);
211 }
212 
213 
214 //------------------------------skip_copies------------------------------------
215 // Skip through any number of copies (that don't mod oop-i-ness)
216 Node *PhaseChaitin::skip_copies( Node *c ) {
217   int idx = c->is_Copy();
218   uint is_oop = lrgs(_lrg_map.live_range_id(c))._is_oop;
219   while (idx != 0) {
220     guarantee(c->in(idx) != NULL, "must not resurrect dead copy");
221     if (lrgs(_lrg_map.live_range_id(c->in(idx)))._is_oop != is_oop) {
222       break;  // casting copy, not the same value
223     }
224     c = c->in(idx);
225     idx = c->is_Copy();
226   }
227   return c;
228 }
229 
230 //------------------------------elide_copy-------------------------------------
231 // Remove (bypass) copies along Node n, edge k.
232 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List &regnd, bool can_change_regs ) {
233   int blk_adjust = 0;
234 
235   uint nk_idx = _lrg_map.live_range_id(n->in(k));
236   OptoReg::Name nk_reg = lrgs(nk_idx).reg();
237 
238   // Remove obvious same-register copies
239   Node *x = n->in(k);
240   int idx;
241   while( (idx=x->is_Copy()) != 0 ) {
242     Node *copy = x->in(idx);
243     guarantee(copy != NULL, "must not resurrect dead copy");
244     if(lrgs(_lrg_map.live_range_id(copy)).reg() != nk_reg) {
245       break;
246     }
247     blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd);
248     if (n->in(k) != copy) {
249       break; // Failed for some cutout?
250     }
251     x = copy;                   // Progress, try again
252   }
253 
254   // Phis and 2-address instructions cannot change registers so easily - their
255   // outputs must match their input.
256   if( !can_change_regs )
257     return blk_adjust;          // Only check stupid copies!
258 
259   // Loop backedges won't have a value-mapping yet
260   if( &value == NULL ) return blk_adjust;
261 
262   // Skip through all copies to the _value_ being used.  Do not change from
263   // int to pointer.  This attempts to jump through a chain of copies, where
264   // intermediate copies might be illegal, i.e., value is stored down to stack
265   // then reloaded BUT survives in a register the whole way.
266   Node *val = skip_copies(n->in(k));
267   if (val == x) return blk_adjust; // No progress?
268 
269   uint val_idx = _lrg_map.live_range_id(val);
270   OptoReg::Name val_reg = lrgs(val_idx).reg();
271   int n_regs = RegMask::num_registers(val->ideal_reg(), lrgs(val_idx));
272 
273   // See if it happens to already be in the correct register!
274   // (either Phi's direct register, or the common case of the name
275   // never-clobbered original-def register)
276   if (register_contains_value(val, val_reg, n_regs, value)) {
277     blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd);
278     if( n->in(k) == regnd[val_reg] ) // Success!  Quit trying
279       return blk_adjust;
280   }
281 
282   // See if we can skip the copy by changing registers.  Don't change from
283   // using a register to using the stack unless we know we can remove a
284   // copy-load.  Otherwise we might end up making a pile of Intel cisc-spill
285   // ops reading from memory instead of just loading once and using the
286   // register.
287 
288   // Also handle duplicate copies here.
289   const Type *t = val->is_Con() ? val->bottom_type() : NULL;
290 
291   // Scan all registers to see if this value is around already
292   for( uint reg = 0; reg < (uint)_max_reg; reg++ ) {
293     if (reg == (uint)nk_reg) {
294       // Found ourselves so check if there is only one user of this
295       // copy and keep on searching for a better copy if so.
296       bool ignore_self = true;
297       x = n->in(k);
298       DUIterator_Fast imax, i = x->fast_outs(imax);
299       Node* first = x->fast_out(i); i++;
300       while (i < imax && ignore_self) {
301         Node* use = x->fast_out(i); i++;
302         if (use != first) ignore_self = false;
303       }
304       if (ignore_self) continue;
305     }
306 
307     Node *vv = value[reg];
308     // For scalable register, number of registers may be inconsistent between
309     // "val_reg" and "reg". For example, when "val" resides in register
310     // but "reg" is located in stack.
311     if (lrgs(val_idx)._is_scalable) {
312       if (lrgs(val_idx)._is_vector) {
313         assert(val->ideal_reg() == Op_VecA, "scalable vector register");
314         if (OptoReg::is_stack(reg)) {
315           n_regs = lrgs(val_idx).scalable_reg_slots();
316         } else {
317           n_regs = RegMask::SlotsPerVecA;
318         }
319       }
320     }
321     if (n_regs > 1) { // Doubles and vectors check for aligned-adjacent set
322       uint last;
323       if (lrgs(val_idx)._is_scalable && lrgs(val_idx)._is_vector) {
324         // For scalable vector register, regmask is always SlotsPerVecA bits aligned
325         last = RegMask::SlotsPerVecA - 1;
326       } else {
327         last = (n_regs-1); // Looking for the last part of a set
328       }
329       if ((reg&last) != last) continue; // Wrong part of a set
330       if (!register_contains_value(vv, reg, n_regs, value)) continue; // Different value
331     }
332     if( vv == val ||            // Got a direct hit?
333         (t && vv && vv->bottom_type() == t && vv->is_Mach() &&
334          vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant?
335       assert( !n->is_Phi(), "cannot change registers at a Phi so easily" );
336       if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR
337           OptoReg::is_reg(reg) || // turning into a register use OR
338           regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use
339         blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd);
340         if( n->in(k) == regnd[reg] ) // Success!  Quit trying
341           return blk_adjust;
342       } // End of if not degrading to a stack
343     } // End of if found value in another register
344   } // End of scan all machine registers
345   return blk_adjust;
346 }
347 
348 
349 //
350 // Check if nreg already contains the constant value val.  Normal copy
351 // elimination doesn't doesn't work on constants because multiple
352 // nodes can represent the same constant so the type and rule of the
353 // MachNode must be checked to ensure equivalence.
354 //
355 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n,
356                                               Block *current_block,
357                                               Node_List& value, Node_List& regnd,
358                                               OptoReg::Name nreg, OptoReg::Name nreg2) {
359   if (value[nreg] != val && val->is_Con() &&
360       value[nreg] != NULL && value[nreg]->is_Con() &&
361       (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) &&
362       value[nreg]->bottom_type() == val->bottom_type() &&
363       value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) {
364     // This code assumes that two MachNodes representing constants
365     // which have the same rule and the same bottom type will produce
366     // identical effects into a register.  This seems like it must be
367     // objectively true unless there are hidden inputs to the nodes
368     // but if that were to change this code would need to updated.
369     // Since they are equivalent the second one if redundant and can
370     // be removed.
371     //
372     // n will be replaced with the old value but n might have
373     // kills projections associated with it so remove them now so that
374     // yank_if_dead will be able to eliminate the copy once the uses
375     // have been transferred to the old[value].
376     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
377       Node* use = n->fast_out(i);
378       if (use->is_Proj() && use->outcnt() == 0) {
379         // Kill projections have no users and one input
380         use->set_req(0, C->top());
381         yank_if_dead(use, current_block, &value, &regnd);
382         --i; --imax;
383       }
384     }
385     _post_alloc++;
386     return true;
387   }
388   return false;
389 }
390 
391 // The algorithms works as follows:
392 // We traverse the block top to bottom. possibly_merge_multidef() is invoked for every input edge k
393 // of the instruction n. We check to see if the input is a multidef lrg. If it is, we record the fact that we've
394 // seen a definition (coming as an input) and add that fact to the reg2defuse array. The array maps registers to their
395 // current reaching definitions (we track only multidefs though). With each definition we also associate the first
396 // instruction we saw use it. If we encounter the situation when we observe an def (an input) that is a part of the
397 // same lrg but is different from the previous seen def we merge the two with a MachMerge node and substitute
398 // all the uses that we've seen so far to use the merge. After that we keep replacing the new defs in the same lrg
399 // as they get encountered with the merge node and keep adding these defs to the merge inputs.
400 void PhaseChaitin::merge_multidefs() {
401   Compile::TracePhase tp("mergeMultidefs", &timers[_t_mergeMultidefs]);
402   ResourceMark rm;
403   // Keep track of the defs seen in registers and collect their uses in the block.
404   RegToDefUseMap reg2defuse(_max_reg, _max_reg, RegDefUse());
405   for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
406     Block* block = _cfg.get_block(i);
407     for (uint j = 1; j < block->number_of_nodes(); j++) {
408       Node* n = block->get_node(j);
409       if (n->is_Phi()) continue;
410       for (uint k = 1; k < n->req(); k++) {
411         j += possibly_merge_multidef(n, k, block, reg2defuse);
412       }
413       // Null out the value produced by the instruction itself, since we're only interested in defs
414       // implicitly defined by the uses. We are actually interested in tracking only redefinitions
415       // of the multidef lrgs in the same register. For that matter it's enough to track changes in
416       // the base register only and ignore other effects of multi-register lrgs and fat projections.
417       // It is also ok to ignore defs coming from singledefs. After an implicit overwrite by one of
418       // those our register is guaranteed to be used by another lrg and we won't attempt to merge it.
419       uint lrg = _lrg_map.live_range_id(n);
420       if (lrg > 0 && lrgs(lrg).is_multidef()) {
421         OptoReg::Name reg = lrgs(lrg).reg();
422         reg2defuse.at(reg).clear();
423       }
424     }
425     // Clear reg->def->use tracking for the next block
426     for (int j = 0; j < reg2defuse.length(); j++) {
427       reg2defuse.at(j).clear();
428     }
429   }
430 }
431 
432 int PhaseChaitin::possibly_merge_multidef(Node *n, uint k, Block *block, RegToDefUseMap& reg2defuse) {
433   int blk_adjust = 0;
434 
435   uint lrg = _lrg_map.live_range_id(n->in(k));
436   if (lrg > 0 && lrgs(lrg).is_multidef()) {
437     OptoReg::Name reg = lrgs(lrg).reg();
438 
439     Node* def = reg2defuse.at(reg).def();
440     if (def != NULL && lrg == _lrg_map.live_range_id(def) && def != n->in(k)) {
441       // Same lrg but different node, we have to merge.
442       MachMergeNode* merge;
443       if (def->is_MachMerge()) { // is it already a merge?
444         merge = def->as_MachMerge();
445       } else {
446         merge = new MachMergeNode(def);
447 
448         // Insert the merge node into the block before the first use.
449         uint use_index = block->find_node(reg2defuse.at(reg).first_use());
450         block->insert_node(merge, use_index++);
451         _cfg.map_node_to_block(merge, block);
452 
453         // Let the allocator know about the new node, use the same lrg
454         _lrg_map.extend(merge->_idx, lrg);
455         blk_adjust++;
456 
457         // Fixup all the uses (there is at least one) that happened between the first
458         // use and before the current one.
459         for (; use_index < block->number_of_nodes(); use_index++) {
460           Node* use = block->get_node(use_index);
461           if (use == n) {
462             break;
463           }
464           use->replace_edge(def, merge);
465         }
466       }
467       if (merge->find_edge(n->in(k)) == -1) {
468         merge->add_req(n->in(k));
469       }
470       n->set_req(k, merge);
471     }
472 
473     // update the uses
474     reg2defuse.at(reg).update(n->in(k), n);
475   }
476 
477   return blk_adjust;
478 }
479 
480 
481 //------------------------------post_allocate_copy_removal---------------------
482 // Post-Allocation peephole copy removal.  We do this in 1 pass over the
483 // basic blocks.  We maintain a mapping of registers to Nodes (an  array of
484 // Nodes indexed by machine register or stack slot number).  NULL means that a
485 // register is not mapped to any Node.  We can (want to have!) have several
486 // registers map to the same Node.  We walk forward over the instructions
487 // updating the mapping as we go.  At merge points we force a NULL if we have
488 // to merge 2 different Nodes into the same register.  Phi functions will give
489 // us a new Node if there is a proper value merging.  Since the blocks are
490 // arranged in some RPO, we will visit all parent blocks before visiting any
491 // successor blocks (except at loops).
492 //
493 // If we find a Copy we look to see if the Copy's source register is a stack
494 // slot and that value has already been loaded into some machine register; if
495 // so we use machine register directly.  This turns a Load into a reg-reg
496 // Move.  We also look for reloads of identical constants.
497 //
498 // When we see a use from a reg-reg Copy, we will attempt to use the copy's
499 // source directly and make the copy go dead.
500 void PhaseChaitin::post_allocate_copy_removal() {
501   Compile::TracePhase tp("postAllocCopyRemoval", &timers[_t_postAllocCopyRemoval]);
502   ResourceMark rm;
503 
504   // Need a mapping from basic block Node_Lists.  We need a Node_List to
505   // map from register number to value-producing Node.
506   Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg.number_of_blocks() + 1);
507   memset(blk2value, 0, sizeof(Node_List*) * (_cfg.number_of_blocks() + 1));
508   // Need a mapping from basic block Node_Lists.  We need a Node_List to
509   // map from register number to register-defining Node.
510   Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg.number_of_blocks() + 1);
511   memset(blk2regnd, 0, sizeof(Node_List*) * (_cfg.number_of_blocks() + 1));
512 
513   // We keep unused Node_Lists on a free_list to avoid wasting
514   // memory.
515   GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16);
516 
517   // For all blocks
518   for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
519     uint j;
520     Block* block = _cfg.get_block(i);
521 
522     // Count of Phis in block
523     uint phi_dex;
524     for (phi_dex = 1; phi_dex < block->number_of_nodes(); phi_dex++) {
525       Node* phi = block->get_node(phi_dex);
526       if (!phi->is_Phi()) {
527         break;
528       }
529     }
530 
531     // If any predecessor has not been visited, we do not know the state
532     // of registers at the start.  Check for this, while updating copies
533     // along Phi input edges
534     bool missing_some_inputs = false;
535     Block *freed = NULL;
536     for (j = 1; j < block->num_preds(); j++) {
537       Block* pb = _cfg.get_block_for_node(block->pred(j));
538       // Remove copies along phi edges
539       for (uint k = 1; k < phi_dex; k++) {
540         elide_copy(block->get_node(k), j, block, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false);
541       }
542       if (blk2value[pb->_pre_order]) { // Have a mapping on this edge?
543         // See if this predecessor's mappings have been used by everybody
544         // who wants them.  If so, free 'em.
545         uint k;
546         for (k = 0; k < pb->_num_succs; k++) {
547           Block* pbsucc = pb->_succs[k];
548           if (!blk2value[pbsucc->_pre_order] && pbsucc != block) {
549             break;              // Found a future user
550           }
551         }
552         if (k >= pb->_num_succs) { // No more uses, free!
553           freed = pb;           // Record last block freed
554           free_list.push(blk2value[pb->_pre_order]);
555           free_list.push(blk2regnd[pb->_pre_order]);
556         }
557       } else {                  // This block has unvisited (loopback) inputs
558         missing_some_inputs = true;
559       }
560     }
561 
562 
563     // Extract Node_List mappings.  If 'freed' is non-zero, we just popped
564     // 'freed's blocks off the list
565     Node_List &regnd = *(free_list.is_empty() ? new Node_List() : free_list.pop());
566     Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop());
567     assert( !freed || blk2value[freed->_pre_order] == &value, "" );
568     value.map(_max_reg,NULL);
569     regnd.map(_max_reg,NULL);
570     // Set mappings as OUR mappings
571     blk2value[block->_pre_order] = &value;
572     blk2regnd[block->_pre_order] = &regnd;
573 
574     // Initialize value & regnd for this block
575     if (missing_some_inputs) {
576       // Some predecessor has not yet been visited; zap map to empty
577       for (uint k = 0; k < (uint)_max_reg; k++) {
578         value.map(k,NULL);
579         regnd.map(k,NULL);
580       }
581     } else {
582       if( !freed ) {            // Didn't get a freebie prior block
583         // Must clone some data
584         freed = _cfg.get_block_for_node(block->pred(1));
585         Node_List &f_value = *blk2value[freed->_pre_order];
586         Node_List &f_regnd = *blk2regnd[freed->_pre_order];
587         for( uint k = 0; k < (uint)_max_reg; k++ ) {
588           value.map(k,f_value[k]);
589           regnd.map(k,f_regnd[k]);
590         }
591       }
592       // Merge all inputs together, setting to NULL any conflicts.
593       for (j = 1; j < block->num_preds(); j++) {
594         Block* pb = _cfg.get_block_for_node(block->pred(j));
595         if (pb == freed) {
596           continue; // Did self already via freelist
597         }
598         Node_List &p_regnd = *blk2regnd[pb->_pre_order];
599         for( uint k = 0; k < (uint)_max_reg; k++ ) {
600           if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs?
601             value.map(k,NULL); // Then no value handy
602             regnd.map(k,NULL);
603           }
604         }
605       }
606     }
607 
608     // For all Phi's
609     for (j = 1; j < phi_dex; j++) {
610       uint k;
611       Node *phi = block->get_node(j);
612       uint pidx = _lrg_map.live_range_id(phi);
613       OptoReg::Name preg = lrgs(pidx).reg();
614 
615       // Remove copies remaining on edges.  Check for junk phi.
616       Node *u = NULL;
617       for (k = 1; k < phi->req(); k++) {
618         Node *x = phi->in(k);
619         if( phi != x && u != x ) // Found a different input
620           u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input
621       }
622       if (u != NodeSentinel) {    // Junk Phi.  Remove
623         phi->replace_by(u);
624         j -= yank_if_dead(phi, block, &value, &regnd);
625         phi_dex--;
626         continue;
627       }
628       // Note that if value[pidx] exists, then we merged no new values here
629       // and the phi is useless.  This can happen even with the above phi
630       // removal for complex flows.  I cannot keep the better known value here
631       // because locally the phi appears to define a new merged value.  If I
632       // keep the better value then a copy of the phi, being unable to use the
633       // global flow analysis, can't "peek through" the phi to the original
634       // reaching value and so will act like it's defining a new value.  This
635       // can lead to situations where some uses are from the old and some from
636       // the new values.  Not illegal by itself but throws the over-strong
637       // assert in scheduling.
638       if( pidx ) {
639         value.map(preg,phi);
640         regnd.map(preg,phi);
641         int n_regs = RegMask::num_registers(phi->ideal_reg(), lrgs(pidx));
642         for (int l = 1; l < n_regs; l++) {
643           OptoReg::Name preg_lo = OptoReg::add(preg,-l);
644           value.map(preg_lo,phi);
645           regnd.map(preg_lo,phi);
646         }
647       }
648     }
649 
650     // For all remaining instructions
651     for (j = phi_dex; j < block->number_of_nodes(); j++) {
652       Node* n = block->get_node(j);
653 
654       if(n->outcnt() == 0 &&   // Dead?
655          n != C->top() &&      // (ignore TOP, it has no du info)
656          !n->is_Proj() ) {     // fat-proj kills
657         j -= yank_if_dead(n, block, &value, &regnd);
658         continue;
659       }
660 
661       // Improve reaching-def info.  Occasionally post-alloc's liveness gives
662       // up (at loop backedges, because we aren't doing a full flow pass).
663       // The presence of a live use essentially asserts that the use's def is
664       // alive and well at the use (or else the allocator fubar'd).  Take
665       // advantage of this info to set a reaching def for the use-reg.
666       uint k;
667       for (k = 1; k < n->req(); k++) {
668         Node *def = n->in(k);   // n->in(k) is a USE; def is the DEF for this USE
669         guarantee(def != NULL, "no disconnected nodes at this point");
670         uint useidx = _lrg_map.live_range_id(def); // useidx is the live range index for this USE
671 
672         if( useidx ) {
673           OptoReg::Name ureg = lrgs(useidx).reg();
674           if( !value[ureg] ) {
675             int idx;            // Skip occasional useless copy
676             while( (idx=def->is_Copy()) != 0 &&
677                    def->in(idx) != NULL &&  // NULL should not happen
678                    ureg == lrgs(_lrg_map.live_range_id(def->in(idx))).reg())
679               def = def->in(idx);
680             Node *valdef = skip_copies(def); // tighten up val through non-useless copies
681             value.map(ureg,valdef); // record improved reaching-def info
682             regnd.map(ureg,   def);
683             // Record other half of doubles
684             uint def_ideal_reg = def->ideal_reg();
685             int n_regs = RegMask::num_registers(def_ideal_reg, lrgs(_lrg_map.live_range_id(def)));
686             for (int l = 1; l < n_regs; l++) {
687               OptoReg::Name ureg_lo = OptoReg::add(ureg,-l);
688               if (!value[ureg_lo] &&
689                   (!RegMask::can_represent(ureg_lo) ||
690                    lrgs(useidx).mask().Member(ureg_lo))) { // Nearly always adjacent
691                 value.map(ureg_lo,valdef); // record improved reaching-def info
692                 regnd.map(ureg_lo,   def);
693               }
694             }
695           }
696         }
697       }
698 
699       const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0;
700 
701       // Remove copies along input edges
702       for (k = 1; k < n->req(); k++) {
703         j -= elide_copy(n, k, block, value, regnd, two_adr != k);
704       }
705 
706       // Unallocated Nodes define no registers
707       uint lidx = _lrg_map.live_range_id(n);
708       if (!lidx) {
709         continue;
710       }
711 
712       // Update the register defined by this instruction
713       OptoReg::Name nreg = lrgs(lidx).reg();
714       // Skip through all copies to the _value_ being defined.
715       // Do not change from int to pointer
716       Node *val = skip_copies(n);
717 
718       // Clear out a dead definition before starting so that the
719       // elimination code doesn't have to guard against it.  The
720       // definition could in fact be a kill projection with a count of
721       // 0 which is safe but since those are uninteresting for copy
722       // elimination just delete them as well.
723       if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) {
724         regnd.map(nreg, NULL);
725         value.map(nreg, NULL);
726       }
727 
728       uint n_ideal_reg = n->ideal_reg();
729       int n_regs = RegMask::num_registers(n_ideal_reg, lrgs(lidx));
730       if (n_regs == 1) {
731         // If Node 'n' does not change the value mapped by the register,
732         // then 'n' is a useless copy.  Do not update the register->node
733         // mapping so 'n' will go dead.
734         if( value[nreg] != val ) {
735           if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, OptoReg::Bad)) {
736             j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
737           } else {
738             // Update the mapping: record new Node defined by the register
739             regnd.map(nreg,n);
740             // Update mapping for defined *value*, which is the defined
741             // Node after skipping all copies.
742             value.map(nreg,val);
743           }
744         } else if( !may_be_copy_of_callee(n) ) {
745           assert(n->is_Copy(), "");
746           j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
747         }
748       } else if (RegMask::is_vector(n_ideal_reg)) {
749         // If Node 'n' does not change the value mapped by the register,
750         // then 'n' is a useless copy.  Do not update the register->node
751         // mapping so 'n' will go dead.
752         if (!register_contains_value(val, nreg, n_regs, value)) {
753           // Update the mapping: record new Node defined by the register
754           regnd.map(nreg,n);
755           // Update mapping for defined *value*, which is the defined
756           // Node after skipping all copies.
757           value.map(nreg,val);
758           for (int l = 1; l < n_regs; l++) {
759             OptoReg::Name nreg_lo = OptoReg::add(nreg,-l);
760             regnd.map(nreg_lo, n );
761             value.map(nreg_lo,val);
762           }
763         } else if (n->is_Copy()) {
764           // Note: vector can't be constant and can't be copy of calee.
765           j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
766         }
767       } else {
768         // If the value occupies a register pair, record same info
769         // in both registers.
770         OptoReg::Name nreg_lo = OptoReg::add(nreg,-1);
771         if( RegMask::can_represent(nreg_lo) &&     // Either a spill slot, or
772             !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent
773           // Sparc occasionally has non-adjacent pairs.
774           // Find the actual other value
775           RegMask tmp = lrgs(lidx).mask();
776           tmp.Remove(nreg);
777           nreg_lo = tmp.find_first_elem();
778         }
779         if (value[nreg] != val || value[nreg_lo] != val) {
780           if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, nreg_lo)) {
781             j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
782           } else {
783             regnd.map(nreg   , n );
784             regnd.map(nreg_lo, n );
785             value.map(nreg   ,val);
786             value.map(nreg_lo,val);
787           }
788         } else if (!may_be_copy_of_callee(n)) {
789           assert(n->is_Copy(), "");
790           j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
791         }
792       }
793 
794       // Fat projections kill many registers
795       if( n_ideal_reg == MachProjNode::fat_proj ) {
796         RegMask rm = n->out_RegMask();
797         // wow, what an expensive iterator...
798         nreg = rm.find_first_elem();
799         while( OptoReg::is_valid(nreg)) {
800           rm.Remove(nreg);
801           value.map(nreg,n);
802           regnd.map(nreg,n);
803           nreg = rm.find_first_elem();
804         }
805       }
806 
807     } // End of for all instructions in the block
808 
809   } // End for all blocks
810 }