duke@435: /*
mikael@6198:  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
duke@435:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435:  *
duke@435:  * This code is free software; you can redistribute it and/or modify it
duke@435:  * under the terms of the GNU General Public License version 2 only, as
duke@435:  * published by the Free Software Foundation.
duke@435:  *
duke@435:  * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
duke@435:  * version 2 for more details (a copy is included in the LICENSE file that
duke@435:  * accompanied this code).
duke@435:  *
duke@435:  * You should have received a copy of the GNU General Public License version
duke@435:  * 2 along with this work; if not, write to the Free Software Foundation,
duke@435:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435:  *
trims@1907:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907:  * or visit www.oracle.com if you need additional information or have any
trims@1907:  * questions.
duke@435:  *
duke@435:  */
duke@435: 
stefank@2314: #include "precompiled.hpp"
stefank@2314: #include "memory/allocation.inline.hpp"
stefank@2314: #include "opto/callnode.hpp"
stefank@2314: #include "opto/chaitin.hpp"
stefank@2314: #include "opto/live.hpp"
stefank@2314: #include "opto/machnode.hpp"
duke@435: 
duke@435: 
duke@435: // Compute live-in/live-out.  We use a totally incremental algorithm.  The LIVE
duke@435: // problem is monotonic.  The steady-state solution looks like this: pull a
duke@435: // block from the worklist.  It has a set of delta's - values which are newly
duke@435: // live-in from the block.  Push these to the live-out sets of all predecessor
duke@435: // blocks.  At each predecessor, the new live-out values are ANDed with what is
duke@435: // already live-out (extra stuff is added to the live-out sets).  Then the
duke@435: // remaining new live-out values are ANDed with what is locally defined.
duke@435: // Leftover bits become the new live-in for the predecessor block, and the pred
duke@435: // block is put on the worklist.
duke@435: //   The locally live-in stuff is computed once and added to predecessor
twisti@1040: // live-out sets.  This separate compilation is done in the outer loop below.
neliasso@4949: PhaseLive::PhaseLive( const PhaseCFG &cfg, const LRG_List &names, Arena *arena ) : Phase(LIVE), _cfg(cfg), _names(names), _arena(arena), _live(0) {
duke@435: }
duke@435: 
duke@435: void PhaseLive::compute(uint maxlrg) {
duke@435:   _maxlrg   = maxlrg;
duke@435:   _worklist = new (_arena) Block_List();
duke@435: 
duke@435:   // Init the sparse live arrays.  This data is live on exit from here!
duke@435:   // The _live info is the live-out info.
adlertz@5539:   _live = (IndexSet*)_arena->Amalloc(sizeof(IndexSet) * _cfg.number_of_blocks());
duke@435:   uint i;
adlertz@5539:   for (i = 0; i < _cfg.number_of_blocks(); i++) {
duke@435:     _live[i].initialize(_maxlrg);
duke@435:   }
duke@435: 
duke@435:   // Init the sparse arrays for delta-sets.
duke@435:   ResourceMark rm;              // Nuke temp storage on exit
duke@435: 
duke@435:   // Does the memory used by _defs and _deltas get reclaimed?  Does it matter?  TT
duke@435: 
duke@435:   // Array of values defined locally in blocks
adlertz@5539:   _defs = NEW_RESOURCE_ARRAY(IndexSet,_cfg.number_of_blocks());
adlertz@5539:   for (i = 0; i < _cfg.number_of_blocks(); i++) {
duke@435:     _defs[i].initialize(_maxlrg);
duke@435:   }
duke@435: 
duke@435:   // Array of delta-set pointers, indexed by block pre_order-1.
adlertz@5539:   _deltas = NEW_RESOURCE_ARRAY(IndexSet*,_cfg.number_of_blocks());
adlertz@5539:   memset( _deltas, 0, sizeof(IndexSet*)* _cfg.number_of_blocks());
duke@435: 
duke@435:   _free_IndexSet = NULL;
duke@435: 
duke@435:   // Blocks having done pass-1
duke@435:   VectorSet first_pass(Thread::current()->resource_area());
duke@435: 
duke@435:   // Outer loop: must compute local live-in sets and push into predecessors.
adlertz@5539:   for (uint j = _cfg.number_of_blocks(); j > 0; j--) {
adlertz@5539:     Block* block = _cfg.get_block(j - 1);
duke@435: 
duke@435:     // Compute the local live-in set.  Start with any new live-out bits.
adlertz@5539:     IndexSet* use = getset(block);
adlertz@5539:     IndexSet* def = &_defs[block->_pre_order-1];
duke@435:     DEBUG_ONLY(IndexSet *def_outside = getfreeset();)
duke@435:     uint i;
adlertz@5635:     for (i = block->number_of_nodes(); i > 1; i--) {
adlertz@5635:       Node* n = block->get_node(i-1);
adlertz@5539:       if (n->is_Phi()) {
adlertz@5539:         break;
adlertz@5539:       }
duke@435: 
adlertz@5722:       uint r = _names.at(n->_idx);
duke@435:       assert(!def_outside->member(r), "Use of external LRG overlaps the same LRG defined in this block");
duke@435:       def->insert( r );
duke@435:       use->remove( r );
duke@435:       uint cnt = n->req();
adlertz@5539:       for (uint k = 1; k < cnt; k++) {
duke@435:         Node *nk = n->in(k);
duke@435:         uint nkidx = nk->_idx;
adlertz@5539:         if (_cfg.get_block_for_node(nk) != block) {
adlertz@5722:           uint u = _names.at(nkidx);
adlertz@5539:           use->insert(u);
adlertz@5539:           DEBUG_ONLY(def_outside->insert(u);)
duke@435:         }
duke@435:       }
duke@435:     }
duke@435: #ifdef ASSERT
duke@435:     def_outside->set_next(_free_IndexSet);
duke@435:     _free_IndexSet = def_outside;     // Drop onto free list
duke@435: #endif
duke@435:     // Remove anything defined by Phis and the block start instruction
adlertz@5539:     for (uint k = i; k > 0; k--) {
adlertz@5722:       uint r = _names.at(block->get_node(k - 1)->_idx);
adlertz@5539:       def->insert(r);
adlertz@5539:       use->remove(r);
duke@435:     }
duke@435: 
duke@435:     // Push these live-in things to predecessors
adlertz@5539:     for (uint l = 1; l < block->num_preds(); l++) {
adlertz@5539:       Block* p = _cfg.get_block_for_node(block->pred(l));
adlertz@5539:       add_liveout(p, use, first_pass);
duke@435: 
duke@435:       // PhiNode uses go in the live-out set of prior blocks.
adlertz@5539:       for (uint k = i; k > 0; k--) {
adlertz@5722:         add_liveout(p, _names.at(block->get_node(k-1)->in(l)->_idx), first_pass);
adlertz@5539:       }
duke@435:     }
adlertz@5539:     freeset(block);
adlertz@5539:     first_pass.set(block->_pre_order);
duke@435: 
duke@435:     // Inner loop: blocks that picked up new live-out values to be propagated
adlertz@5539:     while (_worklist->size()) {
adlertz@5539:       Block* block = _worklist->pop();
adlertz@5539:       IndexSet *delta = getset(block);
duke@435:       assert( delta->count(), "missing delta set" );
duke@435: 
duke@435:       // Add new-live-in to predecessors live-out sets
adlertz@5539:       for (uint l = 1; l < block->num_preds(); l++) {
adlertz@5539:         Block* predecessor = _cfg.get_block_for_node(block->pred(l));
adlertz@5539:         add_liveout(predecessor, delta, first_pass);
adlertz@5509:       }
duke@435: 
adlertz@5539:       freeset(block);
duke@435:     } // End of while-worklist-not-empty
duke@435: 
duke@435:   } // End of for-all-blocks-outer-loop
duke@435: 
duke@435:   // We explicitly clear all of the IndexSets which we are about to release.
duke@435:   // This allows us to recycle their internal memory into IndexSet's free list.
duke@435: 
adlertz@5539:   for (i = 0; i < _cfg.number_of_blocks(); i++) {
duke@435:     _defs[i].clear();
duke@435:     if (_deltas[i]) {
duke@435:       // Is this always true?
duke@435:       _deltas[i]->clear();
duke@435:     }
duke@435:   }
duke@435:   IndexSet *free = _free_IndexSet;
duke@435:   while (free != NULL) {
duke@435:     IndexSet *temp = free;
duke@435:     free = free->next();
duke@435:     temp->clear();
duke@435:   }
duke@435: 
duke@435: }
duke@435: 
duke@435: #ifndef PRODUCT
duke@435: void PhaseLive::stats(uint iters) const {
duke@435: }
duke@435: #endif
duke@435: 
duke@435: // Get an IndexSet for a block.  Return existing one, if any.  Make a new
duke@435: // empty one if a prior one does not exist.
duke@435: IndexSet *PhaseLive::getset( Block *p ) {
duke@435:   IndexSet *delta = _deltas[p->_pre_order-1];
duke@435:   if( !delta )                  // Not on worklist?
duke@435:     // Get a free set; flag as being on worklist
duke@435:     delta = _deltas[p->_pre_order-1] = getfreeset();
duke@435:   return delta;                 // Return set of new live-out items
duke@435: }
duke@435: 
duke@435: // Pull from free list, or allocate.  Internal allocation on the returned set
duke@435: // is always from thread local storage.
duke@435: IndexSet *PhaseLive::getfreeset( ) {
duke@435:   IndexSet *f = _free_IndexSet;
duke@435:   if( !f ) {
duke@435:     f = new IndexSet;
duke@435: //    f->set_arena(Thread::current()->resource_area());
duke@435:     f->initialize(_maxlrg, Thread::current()->resource_area());
duke@435:   } else {
duke@435:     // Pull from free list
duke@435:     _free_IndexSet = f->next();
duke@435:   //f->_cnt = 0;                        // Reset to empty
duke@435: //    f->set_arena(Thread::current()->resource_area());
duke@435:     f->initialize(_maxlrg, Thread::current()->resource_area());
duke@435:   }
duke@435:   return f;
duke@435: }
duke@435: 
duke@435: // Free an IndexSet from a block.
duke@435: void PhaseLive::freeset( const Block *p ) {
duke@435:   IndexSet *f = _deltas[p->_pre_order-1];
duke@435:   f->set_next(_free_IndexSet);
duke@435:   _free_IndexSet = f;           // Drop onto free list
duke@435:   _deltas[p->_pre_order-1] = NULL;
duke@435: }
duke@435: 
duke@435: // Add a live-out value to a given blocks live-out set.  If it is new, then
duke@435: // also add it to the delta set and stick the block on the worklist.
duke@435: void PhaseLive::add_liveout( Block *p, uint r, VectorSet &first_pass ) {
duke@435:   IndexSet *live = &_live[p->_pre_order-1];
duke@435:   if( live->insert(r) ) {       // If actually inserted...
duke@435:     // We extended the live-out set.  See if the value is generated locally.
duke@435:     // If it is not, then we must extend the live-in set.
duke@435:     if( !_defs[p->_pre_order-1].member( r ) ) {
duke@435:       if( !_deltas[p->_pre_order-1] && // Not on worklist?
duke@435:           first_pass.test(p->_pre_order) )
duke@435:         _worklist->push(p);     // Actually go on worklist if already 1st pass
duke@435:       getset(p)->insert(r);
duke@435:     }
duke@435:   }
duke@435: }
duke@435: 
duke@435: // Add a vector of live-out values to a given blocks live-out set.
duke@435: void PhaseLive::add_liveout( Block *p, IndexSet *lo, VectorSet &first_pass ) {
duke@435:   IndexSet *live = &_live[p->_pre_order-1];
duke@435:   IndexSet *defs = &_defs[p->_pre_order-1];
duke@435:   IndexSet *on_worklist = _deltas[p->_pre_order-1];
duke@435:   IndexSet *delta = on_worklist ? on_worklist : getfreeset();
duke@435: 
duke@435:   IndexSetIterator elements(lo);
duke@435:   uint r;
duke@435:   while ((r = elements.next()) != 0) {
duke@435:     if( live->insert(r) &&      // If actually inserted...
duke@435:         !defs->member( r ) )    // and not defined locally
duke@435:       delta->insert(r);         // Then add to live-in set
duke@435:   }
duke@435: 
duke@435:   if( delta->count() ) {                // If actually added things
duke@435:     _deltas[p->_pre_order-1] = delta; // Flag as on worklist now
duke@435:     if( !on_worklist &&         // Not on worklist?
duke@435:         first_pass.test(p->_pre_order) )
duke@435:       _worklist->push(p);       // Actually go on worklist if already 1st pass
duke@435:   } else {                      // Nothing there; just free it
duke@435:     delta->set_next(_free_IndexSet);
duke@435:     _free_IndexSet = delta;     // Drop onto free list
duke@435:   }
duke@435: }
duke@435: 
duke@435: #ifndef PRODUCT
duke@435: // Dump the live-out set for a block
duke@435: void PhaseLive::dump( const Block *b ) const {
duke@435:   tty->print("Block %d: ",b->_pre_order);
duke@435:   tty->print("LiveOut: ");  _live[b->_pre_order-1].dump();
adlertz@5635:   uint cnt = b->number_of_nodes();
duke@435:   for( uint i=0; i<cnt; i++ ) {
adlertz@5722:     tty->print("L%d/", _names.at(b->get_node(i)->_idx));
adlertz@5635:     b->get_node(i)->dump();
duke@435:   }
duke@435:   tty->print("\n");
duke@435: }
duke@435: 
duke@435: // Verify that base pointers and derived pointers are still sane.
duke@435: void PhaseChaitin::verify_base_ptrs( ResourceArea *a ) const {
kvn@985: #ifdef ASSERT
kvn@985:   Unique_Node_List worklist(a);
adlertz@5539:   for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
adlertz@5539:     Block* block = _cfg.get_block(i);
adlertz@5539:     for (uint j = block->end_idx() + 1; j > 1; j--) {
adlertz@5635:       Node* n = block->get_node(j-1);
adlertz@5539:       if (n->is_Phi()) {
adlertz@5539:         break;
adlertz@5539:       }
duke@435:       // Found a safepoint?
adlertz@5539:       if (n->is_MachSafePoint()) {
duke@435:         MachSafePointNode *sfpt = n->as_MachSafePoint();
duke@435:         JVMState* jvms = sfpt->jvms();
duke@435:         if (jvms != NULL) {
duke@435:           // Now scan for a live derived pointer
duke@435:           if (jvms->oopoff() < sfpt->req()) {
duke@435:             // Check each derived/base pair
kvn@985:             for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
duke@435:               Node *check = sfpt->in(idx);
kvn@985:               bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
duke@435:               // search upwards through spills and spill phis for AddP
kvn@985:               worklist.clear();
kvn@985:               worklist.push(check);
kvn@985:               uint k = 0;
kvn@985:               while( k < worklist.size() ) {
kvn@985:                 check = worklist.at(k);
kvn@985:                 assert(check,"Bad base or derived pointer");
kvn@985:                 // See PhaseChaitin::find_base_for_derived() for all cases.
kvn@985:                 int isc = check->is_Copy();
kvn@985:                 if( isc ) {
kvn@985:                   worklist.push(check->in(isc));
kvn@985:                 } else if( check->is_Phi() ) {
kvn@985:                   for (uint m = 1; m < check->req(); m++)
kvn@985:                     worklist.push(check->in(m));
kvn@985:                 } else if( check->is_Con() ) {
kvn@985:                   if (is_derived) {
kvn@985:                     // Derived is NULL+offset
kvn@985:                     assert(!is_derived || check->bottom_type()->is_ptr()->ptr() == TypePtr::Null,"Bad derived pointer");
kvn@985:                   } else {
kvn@985:                     assert(check->bottom_type()->is_ptr()->_offset == 0,"Bad base pointer");
kvn@985:                     // Base either ConP(NULL) or loadConP
kvn@985:                     if (check->is_Mach()) {
kvn@985:                       assert(check->as_Mach()->ideal_Opcode() == Op_ConP,"Bad base pointer");
kvn@985:                     } else {
kvn@985:                       assert(check->Opcode() == Op_ConP &&
kvn@985:                              check->bottom_type()->is_ptr()->ptr() == TypePtr::Null,"Bad base pointer");
kvn@985:                     }
kvn@985:                   }
kvn@985:                 } else if( check->bottom_type()->is_ptr()->_offset == 0 ) {
kvn@985:                   if(check->is_Proj() || check->is_Mach() &&
kvn@985:                      (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
kvn@985:                       check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
kvn@985:                       check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
kvn@985:                       check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
kvn@985: #ifdef _LP64
kvn@985:                       UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP ||
kvn@985:                       UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN ||
ehelin@5694:                       UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass ||
kvn@985: #endif
kvn@985:                       check->as_Mach()->ideal_Opcode() == Op_LoadP ||
kvn@1001:                       check->as_Mach()->ideal_Opcode() == Op_LoadKlass)) {
kvn@1001:                     // Valid nodes
kvn@1001:                   } else {
kvn@1001:                     check->dump();
kvn@985:                     assert(false,"Bad base or derived pointer");
kvn@1001:                   }
kvn@985:                 } else {
kvn@985:                   assert(is_derived,"Bad base pointer");
kvn@985:                   assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP,"Bad derived pointer");
kvn@985:                 }
kvn@985:                 k++;
kvn@985:                 assert(k < 100000,"Derived pointer checking in infinite loop");
duke@435:               } // End while
duke@435:             }
duke@435:           } // End of check for derived pointers
duke@435:         } // End of Kcheck for debug info
duke@435:       } // End of if found a safepoint
duke@435:     } // End of forall instructions in block
duke@435:   } // End of forall blocks
kvn@985: #endif
duke@435: }
kvn@1001: 
kvn@1001: // Verify that graphs and base pointers are still sane.
kvn@1001: void PhaseChaitin::verify( ResourceArea *a, bool verify_ifg ) const {
kvn@1001: #ifdef ASSERT
kvn@1001:   if( VerifyOpto || VerifyRegisterAllocator ) {
kvn@1001:     _cfg.verify();
kvn@1001:     verify_base_ptrs(a);
kvn@1001:     if(verify_ifg)
kvn@1001:       _ifg->verify(this);
kvn@1001:   }
duke@435: #endif
kvn@1001: }
kvn@1001: 
kvn@1001: #endif