1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/phaseX.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,1834 @@
1.4 +/*
1.5 + * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "precompiled.hpp"
1.29 +#include "memory/allocation.inline.hpp"
1.30 +#include "opto/block.hpp"
1.31 +#include "opto/callnode.hpp"
1.32 +#include "opto/cfgnode.hpp"
1.33 +#include "opto/connode.hpp"
1.34 +#include "opto/idealGraphPrinter.hpp"
1.35 +#include "opto/loopnode.hpp"
1.36 +#include "opto/machnode.hpp"
1.37 +#include "opto/opcodes.hpp"
1.38 +#include "opto/phaseX.hpp"
1.39 +#include "opto/regalloc.hpp"
1.40 +#include "opto/rootnode.hpp"
1.41 +
1.42 +//=============================================================================
1.43 +#define NODE_HASH_MINIMUM_SIZE 255
1.44 +//------------------------------NodeHash---------------------------------------
1.45 +NodeHash::NodeHash(uint est_max_size) :
1.46 + _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
1.47 + _a(Thread::current()->resource_area()),
1.48 + _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ), // (Node**)_a->Amalloc(_max * sizeof(Node*)) ),
1.49 + _inserts(0), _insert_limit( insert_limit() ),
1.50 + _look_probes(0), _lookup_hits(0), _lookup_misses(0),
1.51 + _total_insert_probes(0), _total_inserts(0),
1.52 + _insert_probes(0), _grows(0) {
1.53 + // _sentinel must be in the current node space
1.54 + _sentinel = new (Compile::current()) ProjNode(NULL, TypeFunc::Control);
1.55 + memset(_table,0,sizeof(Node*)*_max);
1.56 +}
1.57 +
1.58 +//------------------------------NodeHash---------------------------------------
1.59 +NodeHash::NodeHash(Arena *arena, uint est_max_size) :
1.60 + _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
1.61 + _a(arena),
1.62 + _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ),
1.63 + _inserts(0), _insert_limit( insert_limit() ),
1.64 + _look_probes(0), _lookup_hits(0), _lookup_misses(0),
1.65 + _delete_probes(0), _delete_hits(0), _delete_misses(0),
1.66 + _total_insert_probes(0), _total_inserts(0),
1.67 + _insert_probes(0), _grows(0) {
1.68 + // _sentinel must be in the current node space
1.69 + _sentinel = new (Compile::current()) ProjNode(NULL, TypeFunc::Control);
1.70 + memset(_table,0,sizeof(Node*)*_max);
1.71 +}
1.72 +
1.73 +//------------------------------NodeHash---------------------------------------
1.74 +NodeHash::NodeHash(NodeHash *nh) {
1.75 + debug_only(_table = (Node**)badAddress); // interact correctly w/ operator=
1.76 + // just copy in all the fields
1.77 + *this = *nh;
1.78 + // nh->_sentinel must be in the current node space
1.79 +}
1.80 +
1.81 +void NodeHash::replace_with(NodeHash *nh) {
1.82 + debug_only(_table = (Node**)badAddress); // interact correctly w/ operator=
1.83 + // just copy in all the fields
1.84 + *this = *nh;
1.85 + // nh->_sentinel must be in the current node space
1.86 +}
1.87 +
1.88 +//------------------------------hash_find--------------------------------------
1.89 +// Find in hash table
1.90 +Node *NodeHash::hash_find( const Node *n ) {
1.91 + // ((Node*)n)->set_hash( n->hash() );
1.92 + uint hash = n->hash();
1.93 + if (hash == Node::NO_HASH) {
1.94 + debug_only( _lookup_misses++ );
1.95 + return NULL;
1.96 + }
1.97 + uint key = hash & (_max-1);
1.98 + uint stride = key | 0x01;
1.99 + debug_only( _look_probes++ );
1.100 + Node *k = _table[key]; // Get hashed value
1.101 + if( !k ) { // ?Miss?
1.102 + debug_only( _lookup_misses++ );
1.103 + return NULL; // Miss!
1.104 + }
1.105 +
1.106 + int op = n->Opcode();
1.107 + uint req = n->req();
1.108 + while( 1 ) { // While probing hash table
1.109 + if( k->req() == req && // Same count of inputs
1.110 + k->Opcode() == op ) { // Same Opcode
1.111 + for( uint i=0; i<req; i++ )
1.112 + if( n->in(i)!=k->in(i)) // Different inputs?
1.113 + goto collision; // "goto" is a speed hack...
1.114 + if( n->cmp(*k) ) { // Check for any special bits
1.115 + debug_only( _lookup_hits++ );
1.116 + return k; // Hit!
1.117 + }
1.118 + }
1.119 + collision:
1.120 + debug_only( _look_probes++ );
1.121 + key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
1.122 + k = _table[key]; // Get hashed value
1.123 + if( !k ) { // ?Miss?
1.124 + debug_only( _lookup_misses++ );
1.125 + return NULL; // Miss!
1.126 + }
1.127 + }
1.128 + ShouldNotReachHere();
1.129 + return NULL;
1.130 +}
1.131 +
1.132 +//------------------------------hash_find_insert-------------------------------
1.133 +// Find in hash table, insert if not already present
1.134 +// Used to preserve unique entries in hash table
1.135 +Node *NodeHash::hash_find_insert( Node *n ) {
1.136 + // n->set_hash( );
1.137 + uint hash = n->hash();
1.138 + if (hash == Node::NO_HASH) {
1.139 + debug_only( _lookup_misses++ );
1.140 + return NULL;
1.141 + }
1.142 + uint key = hash & (_max-1);
1.143 + uint stride = key | 0x01; // stride must be relatively prime to table siz
1.144 + uint first_sentinel = 0; // replace a sentinel if seen.
1.145 + debug_only( _look_probes++ );
1.146 + Node *k = _table[key]; // Get hashed value
1.147 + if( !k ) { // ?Miss?
1.148 + debug_only( _lookup_misses++ );
1.149 + _table[key] = n; // Insert into table!
1.150 + debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
1.151 + check_grow(); // Grow table if insert hit limit
1.152 + return NULL; // Miss!
1.153 + }
1.154 + else if( k == _sentinel ) {
1.155 + first_sentinel = key; // Can insert here
1.156 + }
1.157 +
1.158 + int op = n->Opcode();
1.159 + uint req = n->req();
1.160 + while( 1 ) { // While probing hash table
1.161 + if( k->req() == req && // Same count of inputs
1.162 + k->Opcode() == op ) { // Same Opcode
1.163 + for( uint i=0; i<req; i++ )
1.164 + if( n->in(i)!=k->in(i)) // Different inputs?
1.165 + goto collision; // "goto" is a speed hack...
1.166 + if( n->cmp(*k) ) { // Check for any special bits
1.167 + debug_only( _lookup_hits++ );
1.168 + return k; // Hit!
1.169 + }
1.170 + }
1.171 + collision:
1.172 + debug_only( _look_probes++ );
1.173 + key = (key + stride) & (_max-1); // Stride through table w/ relative prime
1.174 + k = _table[key]; // Get hashed value
1.175 + if( !k ) { // ?Miss?
1.176 + debug_only( _lookup_misses++ );
1.177 + key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
1.178 + _table[key] = n; // Insert into table!
1.179 + debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
1.180 + check_grow(); // Grow table if insert hit limit
1.181 + return NULL; // Miss!
1.182 + }
1.183 + else if( first_sentinel == 0 && k == _sentinel ) {
1.184 + first_sentinel = key; // Can insert here
1.185 + }
1.186 +
1.187 + }
1.188 + ShouldNotReachHere();
1.189 + return NULL;
1.190 +}
1.191 +
1.192 +//------------------------------hash_insert------------------------------------
1.193 +// Insert into hash table
1.194 +void NodeHash::hash_insert( Node *n ) {
1.195 + // // "conflict" comments -- print nodes that conflict
1.196 + // bool conflict = false;
1.197 + // n->set_hash();
1.198 + uint hash = n->hash();
1.199 + if (hash == Node::NO_HASH) {
1.200 + return;
1.201 + }
1.202 + check_grow();
1.203 + uint key = hash & (_max-1);
1.204 + uint stride = key | 0x01;
1.205 +
1.206 + while( 1 ) { // While probing hash table
1.207 + debug_only( _insert_probes++ );
1.208 + Node *k = _table[key]; // Get hashed value
1.209 + if( !k || (k == _sentinel) ) break; // Found a slot
1.210 + assert( k != n, "already inserted" );
1.211 + // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
1.212 + key = (key + stride) & (_max-1); // Stride through table w/ relative prime
1.213 + }
1.214 + _table[key] = n; // Insert into table!
1.215 + debug_only(n->enter_hash_lock()); // Lock down the node while in the table.
1.216 + // if( conflict ) { n->dump(); }
1.217 +}
1.218 +
1.219 +//------------------------------hash_delete------------------------------------
1.220 +// Replace in hash table with sentinel
1.221 +bool NodeHash::hash_delete( const Node *n ) {
1.222 + Node *k;
1.223 + uint hash = n->hash();
1.224 + if (hash == Node::NO_HASH) {
1.225 + debug_only( _delete_misses++ );
1.226 + return false;
1.227 + }
1.228 + uint key = hash & (_max-1);
1.229 + uint stride = key | 0x01;
1.230 + debug_only( uint counter = 0; );
1.231 + for( ; /* (k != NULL) && (k != _sentinel) */; ) {
1.232 + debug_only( counter++ );
1.233 + debug_only( _delete_probes++ );
1.234 + k = _table[key]; // Get hashed value
1.235 + if( !k ) { // Miss?
1.236 + debug_only( _delete_misses++ );
1.237 +#ifdef ASSERT
1.238 + if( VerifyOpto ) {
1.239 + for( uint i=0; i < _max; i++ )
1.240 + assert( _table[i] != n, "changed edges with rehashing" );
1.241 + }
1.242 +#endif
1.243 + return false; // Miss! Not in chain
1.244 + }
1.245 + else if( n == k ) {
1.246 + debug_only( _delete_hits++ );
1.247 + _table[key] = _sentinel; // Hit! Label as deleted entry
1.248 + debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
1.249 + return true;
1.250 + }
1.251 + else {
1.252 + // collision: move through table with prime offset
1.253 + key = (key + stride/*7*/) & (_max-1);
1.254 + assert( counter <= _insert_limit, "Cycle in hash-table");
1.255 + }
1.256 + }
1.257 + ShouldNotReachHere();
1.258 + return false;
1.259 +}
1.260 +
1.261 +//------------------------------round_up---------------------------------------
1.262 +// Round up to nearest power of 2
1.263 +uint NodeHash::round_up( uint x ) {
1.264 + x += (x>>2); // Add 25% slop
1.265 + if( x <16 ) return 16; // Small stuff
1.266 + uint i=16;
1.267 + while( i < x ) i <<= 1; // Double to fit
1.268 + return i; // Return hash table size
1.269 +}
1.270 +
1.271 +//------------------------------grow-------------------------------------------
1.272 +// Grow _table to next power of 2 and insert old entries
1.273 +void NodeHash::grow() {
1.274 + // Record old state
1.275 + uint old_max = _max;
1.276 + Node **old_table = _table;
1.277 + // Construct new table with twice the space
1.278 + _grows++;
1.279 + _total_inserts += _inserts;
1.280 + _total_insert_probes += _insert_probes;
1.281 + _inserts = 0;
1.282 + _insert_probes = 0;
1.283 + _max = _max << 1;
1.284 + _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
1.285 + memset(_table,0,sizeof(Node*)*_max);
1.286 + _insert_limit = insert_limit();
1.287 + // Insert old entries into the new table
1.288 + for( uint i = 0; i < old_max; i++ ) {
1.289 + Node *m = *old_table++;
1.290 + if( !m || m == _sentinel ) continue;
1.291 + debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table.
1.292 + hash_insert(m);
1.293 + }
1.294 +}
1.295 +
1.296 +//------------------------------clear------------------------------------------
1.297 +// Clear all entries in _table to NULL but keep storage
1.298 +void NodeHash::clear() {
1.299 +#ifdef ASSERT
1.300 + // Unlock all nodes upon removal from table.
1.301 + for (uint i = 0; i < _max; i++) {
1.302 + Node* n = _table[i];
1.303 + if (!n || n == _sentinel) continue;
1.304 + n->exit_hash_lock();
1.305 + }
1.306 +#endif
1.307 +
1.308 + memset( _table, 0, _max * sizeof(Node*) );
1.309 +}
1.310 +
1.311 +//-----------------------remove_useless_nodes----------------------------------
1.312 +// Remove useless nodes from value table,
1.313 +// implementation does not depend on hash function
1.314 +void NodeHash::remove_useless_nodes(VectorSet &useful) {
1.315 +
1.316 + // Dead nodes in the hash table inherited from GVN should not replace
1.317 + // existing nodes, remove dead nodes.
1.318 + uint max = size();
1.319 + Node *sentinel_node = sentinel();
1.320 + for( uint i = 0; i < max; ++i ) {
1.321 + Node *n = at(i);
1.322 + if(n != NULL && n != sentinel_node && !useful.test(n->_idx)) {
1.323 + debug_only(n->exit_hash_lock()); // Unlock the node when removed
1.324 + _table[i] = sentinel_node; // Replace with placeholder
1.325 + }
1.326 + }
1.327 +}
1.328 +
1.329 +
1.330 +void NodeHash::check_no_speculative_types() {
1.331 +#ifdef ASSERT
1.332 + uint max = size();
1.333 + Node *sentinel_node = sentinel();
1.334 + for (uint i = 0; i < max; ++i) {
1.335 + Node *n = at(i);
1.336 + if(n != NULL && n != sentinel_node && n->is_Type()) {
1.337 + TypeNode* tn = n->as_Type();
1.338 + const Type* t = tn->type();
1.339 + const Type* t_no_spec = t->remove_speculative();
1.340 + assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
1.341 + }
1.342 + }
1.343 +#endif
1.344 +}
1.345 +
1.346 +#ifndef PRODUCT
1.347 +//------------------------------dump-------------------------------------------
1.348 +// Dump statistics for the hash table
1.349 +void NodeHash::dump() {
1.350 + _total_inserts += _inserts;
1.351 + _total_insert_probes += _insert_probes;
1.352 + if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
1.353 + if (WizardMode) {
1.354 + for (uint i=0; i<_max; i++) {
1.355 + if (_table[i])
1.356 + tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
1.357 + }
1.358 + }
1.359 + tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
1.360 + tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
1.361 + tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
1.362 + tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
1.363 + // sentinels increase lookup cost, but not insert cost
1.364 + assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
1.365 + assert( _inserts+(_inserts>>3) < _max, "table too full" );
1.366 + assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
1.367 + }
1.368 +}
1.369 +
1.370 +Node *NodeHash::find_index(uint idx) { // For debugging
1.371 + // Find an entry by its index value
1.372 + for( uint i = 0; i < _max; i++ ) {
1.373 + Node *m = _table[i];
1.374 + if( !m || m == _sentinel ) continue;
1.375 + if( m->_idx == (uint)idx ) return m;
1.376 + }
1.377 + return NULL;
1.378 +}
1.379 +#endif
1.380 +
1.381 +#ifdef ASSERT
1.382 +NodeHash::~NodeHash() {
1.383 + // Unlock all nodes upon destruction of table.
1.384 + if (_table != (Node**)badAddress) clear();
1.385 +}
1.386 +
1.387 +void NodeHash::operator=(const NodeHash& nh) {
1.388 + // Unlock all nodes upon replacement of table.
1.389 + if (&nh == this) return;
1.390 + if (_table != (Node**)badAddress) clear();
1.391 + memcpy(this, &nh, sizeof(*this));
1.392 + // Do not increment hash_lock counts again.
1.393 + // Instead, be sure we never again use the source table.
1.394 + ((NodeHash*)&nh)->_table = (Node**)badAddress;
1.395 +}
1.396 +
1.397 +
1.398 +#endif
1.399 +
1.400 +
1.401 +//=============================================================================
1.402 +//------------------------------PhaseRemoveUseless-----------------------------
1.403 +// 1) Use a breadthfirst walk to collect useful nodes reachable from root.
1.404 +PhaseRemoveUseless::PhaseRemoveUseless( PhaseGVN *gvn, Unique_Node_List *worklist ) : Phase(Remove_Useless),
1.405 + _useful(Thread::current()->resource_area()) {
1.406 +
1.407 + // Implementation requires 'UseLoopSafepoints == true' and an edge from root
1.408 + // to each SafePointNode at a backward branch. Inserted in add_safepoint().
1.409 + if( !UseLoopSafepoints || !OptoRemoveUseless ) return;
1.410 +
1.411 + // Identify nodes that are reachable from below, useful.
1.412 + C->identify_useful_nodes(_useful);
1.413 + // Update dead node list
1.414 + C->update_dead_node_list(_useful);
1.415 +
1.416 + // Remove all useless nodes from PhaseValues' recorded types
1.417 + // Must be done before disconnecting nodes to preserve hash-table-invariant
1.418 + gvn->remove_useless_nodes(_useful.member_set());
1.419 +
1.420 + // Remove all useless nodes from future worklist
1.421 + worklist->remove_useless_nodes(_useful.member_set());
1.422 +
1.423 + // Disconnect 'useless' nodes that are adjacent to useful nodes
1.424 + C->remove_useless_nodes(_useful);
1.425 +
1.426 + // Remove edges from "root" to each SafePoint at a backward branch.
1.427 + // They were inserted during parsing (see add_safepoint()) to make infinite
1.428 + // loops without calls or exceptions visible to root, i.e., useful.
1.429 + Node *root = C->root();
1.430 + if( root != NULL ) {
1.431 + for( uint i = root->req(); i < root->len(); ++i ) {
1.432 + Node *n = root->in(i);
1.433 + if( n != NULL && n->is_SafePoint() ) {
1.434 + root->rm_prec(i);
1.435 + --i;
1.436 + }
1.437 + }
1.438 + }
1.439 +}
1.440 +
1.441 +
1.442 +//=============================================================================
1.443 +//------------------------------PhaseTransform---------------------------------
1.444 +PhaseTransform::PhaseTransform( PhaseNumber pnum ) : Phase(pnum),
1.445 + _arena(Thread::current()->resource_area()),
1.446 + _nodes(_arena),
1.447 + _types(_arena)
1.448 +{
1.449 + init_con_caches();
1.450 +#ifndef PRODUCT
1.451 + clear_progress();
1.452 + clear_transforms();
1.453 + set_allow_progress(true);
1.454 +#endif
1.455 + // Force allocation for currently existing nodes
1.456 + _types.map(C->unique(), NULL);
1.457 +}
1.458 +
1.459 +//------------------------------PhaseTransform---------------------------------
1.460 +PhaseTransform::PhaseTransform( Arena *arena, PhaseNumber pnum ) : Phase(pnum),
1.461 + _arena(arena),
1.462 + _nodes(arena),
1.463 + _types(arena)
1.464 +{
1.465 + init_con_caches();
1.466 +#ifndef PRODUCT
1.467 + clear_progress();
1.468 + clear_transforms();
1.469 + set_allow_progress(true);
1.470 +#endif
1.471 + // Force allocation for currently existing nodes
1.472 + _types.map(C->unique(), NULL);
1.473 +}
1.474 +
1.475 +//------------------------------PhaseTransform---------------------------------
1.476 +// Initialize with previously generated type information
1.477 +PhaseTransform::PhaseTransform( PhaseTransform *pt, PhaseNumber pnum ) : Phase(pnum),
1.478 + _arena(pt->_arena),
1.479 + _nodes(pt->_nodes),
1.480 + _types(pt->_types)
1.481 +{
1.482 + init_con_caches();
1.483 +#ifndef PRODUCT
1.484 + clear_progress();
1.485 + clear_transforms();
1.486 + set_allow_progress(true);
1.487 +#endif
1.488 +}
1.489 +
1.490 +void PhaseTransform::init_con_caches() {
1.491 + memset(_icons,0,sizeof(_icons));
1.492 + memset(_lcons,0,sizeof(_lcons));
1.493 + memset(_zcons,0,sizeof(_zcons));
1.494 +}
1.495 +
1.496 +
1.497 +//--------------------------------find_int_type--------------------------------
1.498 +const TypeInt* PhaseTransform::find_int_type(Node* n) {
1.499 + if (n == NULL) return NULL;
1.500 + // Call type_or_null(n) to determine node's type since we might be in
1.501 + // parse phase and call n->Value() may return wrong type.
1.502 + // (For example, a phi node at the beginning of loop parsing is not ready.)
1.503 + const Type* t = type_or_null(n);
1.504 + if (t == NULL) return NULL;
1.505 + return t->isa_int();
1.506 +}
1.507 +
1.508 +
1.509 +//-------------------------------find_long_type--------------------------------
1.510 +const TypeLong* PhaseTransform::find_long_type(Node* n) {
1.511 + if (n == NULL) return NULL;
1.512 + // (See comment above on type_or_null.)
1.513 + const Type* t = type_or_null(n);
1.514 + if (t == NULL) return NULL;
1.515 + return t->isa_long();
1.516 +}
1.517 +
1.518 +
1.519 +#ifndef PRODUCT
1.520 +void PhaseTransform::dump_old2new_map() const {
1.521 + _nodes.dump();
1.522 +}
1.523 +
1.524 +void PhaseTransform::dump_new( uint nidx ) const {
1.525 + for( uint i=0; i<_nodes.Size(); i++ )
1.526 + if( _nodes[i] && _nodes[i]->_idx == nidx ) {
1.527 + _nodes[i]->dump();
1.528 + tty->cr();
1.529 + tty->print_cr("Old index= %d",i);
1.530 + return;
1.531 + }
1.532 + tty->print_cr("Node %d not found in the new indices", nidx);
1.533 +}
1.534 +
1.535 +//------------------------------dump_types-------------------------------------
1.536 +void PhaseTransform::dump_types( ) const {
1.537 + _types.dump();
1.538 +}
1.539 +
1.540 +//------------------------------dump_nodes_and_types---------------------------
1.541 +void PhaseTransform::dump_nodes_and_types(const Node *root, uint depth, bool only_ctrl) {
1.542 + VectorSet visited(Thread::current()->resource_area());
1.543 + dump_nodes_and_types_recur( root, depth, only_ctrl, visited );
1.544 +}
1.545 +
1.546 +//------------------------------dump_nodes_and_types_recur---------------------
1.547 +void PhaseTransform::dump_nodes_and_types_recur( const Node *n, uint depth, bool only_ctrl, VectorSet &visited) {
1.548 + if( !n ) return;
1.549 + if( depth == 0 ) return;
1.550 + if( visited.test_set(n->_idx) ) return;
1.551 + for( uint i=0; i<n->len(); i++ ) {
1.552 + if( only_ctrl && !(n->is_Region()) && i != TypeFunc::Control ) continue;
1.553 + dump_nodes_and_types_recur( n->in(i), depth-1, only_ctrl, visited );
1.554 + }
1.555 + n->dump();
1.556 + if (type_or_null(n) != NULL) {
1.557 + tty->print(" "); type(n)->dump(); tty->cr();
1.558 + }
1.559 +}
1.560 +
1.561 +#endif
1.562 +
1.563 +
1.564 +//=============================================================================
1.565 +//------------------------------PhaseValues------------------------------------
1.566 +// Set minimum table size to "255"
1.567 +PhaseValues::PhaseValues( Arena *arena, uint est_max_size ) : PhaseTransform(arena, GVN), _table(arena, est_max_size) {
1.568 + NOT_PRODUCT( clear_new_values(); )
1.569 +}
1.570 +
1.571 +//------------------------------PhaseValues------------------------------------
1.572 +// Set minimum table size to "255"
1.573 +PhaseValues::PhaseValues( PhaseValues *ptv ) : PhaseTransform( ptv, GVN ),
1.574 + _table(&ptv->_table) {
1.575 + NOT_PRODUCT( clear_new_values(); )
1.576 +}
1.577 +
1.578 +//------------------------------PhaseValues------------------------------------
1.579 +// Used by +VerifyOpto. Clear out hash table but copy _types array.
1.580 +PhaseValues::PhaseValues( PhaseValues *ptv, const char *dummy ) : PhaseTransform( ptv, GVN ),
1.581 + _table(ptv->arena(),ptv->_table.size()) {
1.582 + NOT_PRODUCT( clear_new_values(); )
1.583 +}
1.584 +
1.585 +//------------------------------~PhaseValues-----------------------------------
1.586 +#ifndef PRODUCT
1.587 +PhaseValues::~PhaseValues() {
1.588 + _table.dump();
1.589 +
1.590 + // Statistics for value progress and efficiency
1.591 + if( PrintCompilation && Verbose && WizardMode ) {
1.592 + tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
1.593 + is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
1.594 + if( made_transforms() != 0 ) {
1.595 + tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
1.596 + } else {
1.597 + tty->cr();
1.598 + }
1.599 + }
1.600 +}
1.601 +#endif
1.602 +
1.603 +//------------------------------makecon----------------------------------------
1.604 +ConNode* PhaseTransform::makecon(const Type *t) {
1.605 + assert(t->singleton(), "must be a constant");
1.606 + assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
1.607 + switch (t->base()) { // fast paths
1.608 + case Type::Half:
1.609 + case Type::Top: return (ConNode*) C->top();
1.610 + case Type::Int: return intcon( t->is_int()->get_con() );
1.611 + case Type::Long: return longcon( t->is_long()->get_con() );
1.612 + }
1.613 + if (t->is_zero_type())
1.614 + return zerocon(t->basic_type());
1.615 + return uncached_makecon(t);
1.616 +}
1.617 +
1.618 +//--------------------------uncached_makecon-----------------------------------
1.619 +// Make an idealized constant - one of ConINode, ConPNode, etc.
1.620 +ConNode* PhaseValues::uncached_makecon(const Type *t) {
1.621 + assert(t->singleton(), "must be a constant");
1.622 + ConNode* x = ConNode::make(C, t);
1.623 + ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
1.624 + if (k == NULL) {
1.625 + set_type(x, t); // Missed, provide type mapping
1.626 + GrowableArray<Node_Notes*>* nna = C->node_note_array();
1.627 + if (nna != NULL) {
1.628 + Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
1.629 + loc->clear(); // do not put debug info on constants
1.630 + }
1.631 + } else {
1.632 + x->destruct(); // Hit, destroy duplicate constant
1.633 + x = k; // use existing constant
1.634 + }
1.635 + return x;
1.636 +}
1.637 +
1.638 +//------------------------------intcon-----------------------------------------
1.639 +// Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
1.640 +ConINode* PhaseTransform::intcon(int i) {
1.641 + // Small integer? Check cache! Check that cached node is not dead
1.642 + if (i >= _icon_min && i <= _icon_max) {
1.643 + ConINode* icon = _icons[i-_icon_min];
1.644 + if (icon != NULL && icon->in(TypeFunc::Control) != NULL)
1.645 + return icon;
1.646 + }
1.647 + ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
1.648 + assert(icon->is_Con(), "");
1.649 + if (i >= _icon_min && i <= _icon_max)
1.650 + _icons[i-_icon_min] = icon; // Cache small integers
1.651 + return icon;
1.652 +}
1.653 +
1.654 +//------------------------------longcon----------------------------------------
1.655 +// Fast long constant.
1.656 +ConLNode* PhaseTransform::longcon(jlong l) {
1.657 + // Small integer? Check cache! Check that cached node is not dead
1.658 + if (l >= _lcon_min && l <= _lcon_max) {
1.659 + ConLNode* lcon = _lcons[l-_lcon_min];
1.660 + if (lcon != NULL && lcon->in(TypeFunc::Control) != NULL)
1.661 + return lcon;
1.662 + }
1.663 + ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
1.664 + assert(lcon->is_Con(), "");
1.665 + if (l >= _lcon_min && l <= _lcon_max)
1.666 + _lcons[l-_lcon_min] = lcon; // Cache small integers
1.667 + return lcon;
1.668 +}
1.669 +
1.670 +//------------------------------zerocon-----------------------------------------
1.671 +// Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
1.672 +ConNode* PhaseTransform::zerocon(BasicType bt) {
1.673 + assert((uint)bt <= _zcon_max, "domain check");
1.674 + ConNode* zcon = _zcons[bt];
1.675 + if (zcon != NULL && zcon->in(TypeFunc::Control) != NULL)
1.676 + return zcon;
1.677 + zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
1.678 + _zcons[bt] = zcon;
1.679 + return zcon;
1.680 +}
1.681 +
1.682 +
1.683 +
1.684 +//=============================================================================
1.685 +//------------------------------transform--------------------------------------
1.686 +// Return a node which computes the same function as this node, but in a
1.687 +// faster or cheaper fashion.
1.688 +Node *PhaseGVN::transform( Node *n ) {
1.689 + return transform_no_reclaim(n);
1.690 +}
1.691 +
1.692 +//------------------------------transform--------------------------------------
1.693 +// Return a node which computes the same function as this node, but
1.694 +// in a faster or cheaper fashion.
1.695 +Node *PhaseGVN::transform_no_reclaim( Node *n ) {
1.696 + NOT_PRODUCT( set_transforms(); )
1.697 +
1.698 + // Apply the Ideal call in a loop until it no longer applies
1.699 + Node *k = n;
1.700 + NOT_PRODUCT( uint loop_count = 0; )
1.701 + while( 1 ) {
1.702 + Node *i = k->Ideal(this, /*can_reshape=*/false);
1.703 + if( !i ) break;
1.704 + assert( i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
1.705 + k = i;
1.706 + assert(loop_count++ < K, "infinite loop in PhaseGVN::transform");
1.707 + }
1.708 + NOT_PRODUCT( if( loop_count != 0 ) { set_progress(); } )
1.709 +
1.710 +
1.711 + // If brand new node, make space in type array.
1.712 + ensure_type_or_null(k);
1.713 +
1.714 + // Since I just called 'Value' to compute the set of run-time values
1.715 + // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1.716 + // cache Value. Later requests for the local phase->type of this Node can
1.717 + // use the cached Value instead of suffering with 'bottom_type'.
1.718 + const Type *t = k->Value(this); // Get runtime Value set
1.719 + assert(t != NULL, "value sanity");
1.720 + if (type_or_null(k) != t) {
1.721 +#ifndef PRODUCT
1.722 + // Do not count initial visit to node as a transformation
1.723 + if (type_or_null(k) == NULL) {
1.724 + inc_new_values();
1.725 + set_progress();
1.726 + }
1.727 +#endif
1.728 + set_type(k, t);
1.729 + // If k is a TypeNode, capture any more-precise type permanently into Node
1.730 + k->raise_bottom_type(t);
1.731 + }
1.732 +
1.733 + if( t->singleton() && !k->is_Con() ) {
1.734 + NOT_PRODUCT( set_progress(); )
1.735 + return makecon(t); // Turn into a constant
1.736 + }
1.737 +
1.738 + // Now check for Identities
1.739 + Node *i = k->Identity(this); // Look for a nearby replacement
1.740 + if( i != k ) { // Found? Return replacement!
1.741 + NOT_PRODUCT( set_progress(); )
1.742 + return i;
1.743 + }
1.744 +
1.745 + // Global Value Numbering
1.746 + i = hash_find_insert(k); // Insert if new
1.747 + if( i && (i != k) ) {
1.748 + // Return the pre-existing node
1.749 + NOT_PRODUCT( set_progress(); )
1.750 + return i;
1.751 + }
1.752 +
1.753 + // Return Idealized original
1.754 + return k;
1.755 +}
1.756 +
1.757 +#ifdef ASSERT
1.758 +//------------------------------dead_loop_check--------------------------------
1.759 +// Check for a simple dead loop when a data node references itself directly
1.760 +// or through an other data node excluding cons and phis.
1.761 +void PhaseGVN::dead_loop_check( Node *n ) {
1.762 + // Phi may reference itself in a loop
1.763 + if (n != NULL && !n->is_dead_loop_safe() && !n->is_CFG()) {
1.764 + // Do 2 levels check and only data inputs.
1.765 + bool no_dead_loop = true;
1.766 + uint cnt = n->req();
1.767 + for (uint i = 1; i < cnt && no_dead_loop; i++) {
1.768 + Node *in = n->in(i);
1.769 + if (in == n) {
1.770 + no_dead_loop = false;
1.771 + } else if (in != NULL && !in->is_dead_loop_safe()) {
1.772 + uint icnt = in->req();
1.773 + for (uint j = 1; j < icnt && no_dead_loop; j++) {
1.774 + if (in->in(j) == n || in->in(j) == in)
1.775 + no_dead_loop = false;
1.776 + }
1.777 + }
1.778 + }
1.779 + if (!no_dead_loop) n->dump(3);
1.780 + assert(no_dead_loop, "dead loop detected");
1.781 + }
1.782 +}
1.783 +#endif
1.784 +
1.785 +//=============================================================================
1.786 +//------------------------------PhaseIterGVN-----------------------------------
1.787 +// Initialize hash table to fresh and clean for +VerifyOpto
1.788 +PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn, const char *dummy ) : PhaseGVN(igvn,dummy), _worklist( ),
1.789 + _stack(C->unique() >> 1),
1.790 + _delay_transform(false) {
1.791 +}
1.792 +
1.793 +//------------------------------PhaseIterGVN-----------------------------------
1.794 +// Initialize with previous PhaseIterGVN info; used by PhaseCCP
1.795 +PhaseIterGVN::PhaseIterGVN( PhaseIterGVN *igvn ) : PhaseGVN(igvn),
1.796 + _worklist( igvn->_worklist ),
1.797 + _stack( igvn->_stack ),
1.798 + _delay_transform(igvn->_delay_transform)
1.799 +{
1.800 +}
1.801 +
1.802 +//------------------------------PhaseIterGVN-----------------------------------
1.803 +// Initialize with previous PhaseGVN info from Parser
1.804 +PhaseIterGVN::PhaseIterGVN( PhaseGVN *gvn ) : PhaseGVN(gvn),
1.805 + _worklist(*C->for_igvn()),
1.806 + _stack(C->unique() >> 1),
1.807 + _delay_transform(false)
1.808 +{
1.809 + uint max;
1.810 +
1.811 + // Dead nodes in the hash table inherited from GVN were not treated as
1.812 + // roots during def-use info creation; hence they represent an invisible
1.813 + // use. Clear them out.
1.814 + max = _table.size();
1.815 + for( uint i = 0; i < max; ++i ) {
1.816 + Node *n = _table.at(i);
1.817 + if(n != NULL && n != _table.sentinel() && n->outcnt() == 0) {
1.818 + if( n->is_top() ) continue;
1.819 + assert( false, "Parse::remove_useless_nodes missed this node");
1.820 + hash_delete(n);
1.821 + }
1.822 + }
1.823 +
1.824 + // Any Phis or Regions on the worklist probably had uses that could not
1.825 + // make more progress because the uses were made while the Phis and Regions
1.826 + // were in half-built states. Put all uses of Phis and Regions on worklist.
1.827 + max = _worklist.size();
1.828 + for( uint j = 0; j < max; j++ ) {
1.829 + Node *n = _worklist.at(j);
1.830 + uint uop = n->Opcode();
1.831 + if( uop == Op_Phi || uop == Op_Region ||
1.832 + n->is_Type() ||
1.833 + n->is_Mem() )
1.834 + add_users_to_worklist(n);
1.835 + }
1.836 +}
1.837 +
1.838 +
1.839 +#ifndef PRODUCT
1.840 +void PhaseIterGVN::verify_step(Node* n) {
1.841 + _verify_window[_verify_counter % _verify_window_size] = n;
1.842 + ++_verify_counter;
1.843 + ResourceMark rm;
1.844 + ResourceArea *area = Thread::current()->resource_area();
1.845 + VectorSet old_space(area), new_space(area);
1.846 + if (C->unique() < 1000 ||
1.847 + 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
1.848 + ++_verify_full_passes;
1.849 + Node::verify_recur(C->root(), -1, old_space, new_space);
1.850 + }
1.851 + const int verify_depth = 4;
1.852 + for ( int i = 0; i < _verify_window_size; i++ ) {
1.853 + Node* n = _verify_window[i];
1.854 + if ( n == NULL ) continue;
1.855 + if( n->in(0) == NodeSentinel ) { // xform_idom
1.856 + _verify_window[i] = n->in(1);
1.857 + --i; continue;
1.858 + }
1.859 + // Typical fanout is 1-2, so this call visits about 6 nodes.
1.860 + Node::verify_recur(n, verify_depth, old_space, new_space);
1.861 + }
1.862 +}
1.863 +#endif
1.864 +
1.865 +
1.866 +//------------------------------init_worklist----------------------------------
1.867 +// Initialize worklist for each node.
1.868 +void PhaseIterGVN::init_worklist( Node *n ) {
1.869 + if( _worklist.member(n) ) return;
1.870 + _worklist.push(n);
1.871 + uint cnt = n->req();
1.872 + for( uint i =0 ; i < cnt; i++ ) {
1.873 + Node *m = n->in(i);
1.874 + if( m ) init_worklist(m);
1.875 + }
1.876 +}
1.877 +
1.878 +//------------------------------optimize---------------------------------------
1.879 +void PhaseIterGVN::optimize() {
1.880 + debug_only(uint num_processed = 0;);
1.881 +#ifndef PRODUCT
1.882 + {
1.883 + _verify_counter = 0;
1.884 + _verify_full_passes = 0;
1.885 + for ( int i = 0; i < _verify_window_size; i++ ) {
1.886 + _verify_window[i] = NULL;
1.887 + }
1.888 + }
1.889 +#endif
1.890 +
1.891 +#ifdef ASSERT
1.892 + Node* prev = NULL;
1.893 + uint rep_cnt = 0;
1.894 +#endif
1.895 + uint loop_count = 0;
1.896 +
1.897 + // Pull from worklist; transform node;
1.898 + // If node has changed: update edge info and put uses on worklist.
1.899 + while( _worklist.size() ) {
1.900 + if (C->check_node_count(NodeLimitFudgeFactor * 2,
1.901 + "out of nodes optimizing method")) {
1.902 + return;
1.903 + }
1.904 + Node *n = _worklist.pop();
1.905 + if (++loop_count >= K * C->live_nodes()) {
1.906 + debug_only(n->dump(4);)
1.907 + assert(false, "infinite loop in PhaseIterGVN::optimize");
1.908 + C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1.909 + return;
1.910 + }
1.911 +#ifdef ASSERT
1.912 + if (n == prev) {
1.913 + if (++rep_cnt > 3) {
1.914 + n->dump(4);
1.915 + assert(false, "loop in Ideal transformation");
1.916 + }
1.917 + } else {
1.918 + rep_cnt = 0;
1.919 + }
1.920 + prev = n;
1.921 +#endif
1.922 + if (TraceIterativeGVN && Verbose) {
1.923 + tty->print(" Pop ");
1.924 + NOT_PRODUCT( n->dump(); )
1.925 + debug_only(if( (num_processed++ % 100) == 0 ) _worklist.print_set();)
1.926 + }
1.927 +
1.928 + if (n->outcnt() != 0) {
1.929 +
1.930 +#ifndef PRODUCT
1.931 + uint wlsize = _worklist.size();
1.932 + const Type* oldtype = type_or_null(n);
1.933 +#endif //PRODUCT
1.934 +
1.935 + Node *nn = transform_old(n);
1.936 +
1.937 +#ifndef PRODUCT
1.938 + if (TraceIterativeGVN) {
1.939 + const Type* newtype = type_or_null(n);
1.940 + if (nn != n) {
1.941 + // print old node
1.942 + tty->print("< ");
1.943 + if (oldtype != newtype && oldtype != NULL) {
1.944 + oldtype->dump();
1.945 + }
1.946 + do { tty->print("\t"); } while (tty->position() < 16);
1.947 + tty->print("<");
1.948 + n->dump();
1.949 + }
1.950 + if (oldtype != newtype || nn != n) {
1.951 + // print new node and/or new type
1.952 + if (oldtype == NULL) {
1.953 + tty->print("* ");
1.954 + } else if (nn != n) {
1.955 + tty->print("> ");
1.956 + } else {
1.957 + tty->print("= ");
1.958 + }
1.959 + if (newtype == NULL) {
1.960 + tty->print("null");
1.961 + } else {
1.962 + newtype->dump();
1.963 + }
1.964 + do { tty->print("\t"); } while (tty->position() < 16);
1.965 + nn->dump();
1.966 + }
1.967 + if (Verbose && wlsize < _worklist.size()) {
1.968 + tty->print(" Push {");
1.969 + while (wlsize != _worklist.size()) {
1.970 + Node* pushed = _worklist.at(wlsize++);
1.971 + tty->print(" %d", pushed->_idx);
1.972 + }
1.973 + tty->print_cr(" }");
1.974 + }
1.975 + }
1.976 + if( VerifyIterativeGVN && nn != n ) {
1.977 + verify_step((Node*) NULL); // ignore n, it might be subsumed
1.978 + }
1.979 +#endif
1.980 + } else if (!n->is_top()) {
1.981 + remove_dead_node(n);
1.982 + }
1.983 + }
1.984 +
1.985 +#ifndef PRODUCT
1.986 + C->verify_graph_edges();
1.987 + if( VerifyOpto && allow_progress() ) {
1.988 + // Must turn off allow_progress to enable assert and break recursion
1.989 + C->root()->verify();
1.990 + { // Check if any progress was missed using IterGVN
1.991 + // Def-Use info enables transformations not attempted in wash-pass
1.992 + // e.g. Region/Phi cleanup, ...
1.993 + // Null-check elision -- may not have reached fixpoint
1.994 + // do not propagate to dominated nodes
1.995 + ResourceMark rm;
1.996 + PhaseIterGVN igvn2(this,"Verify"); // Fresh and clean!
1.997 + // Fill worklist completely
1.998 + igvn2.init_worklist(C->root());
1.999 +
1.1000 + igvn2.set_allow_progress(false);
1.1001 + igvn2.optimize();
1.1002 + igvn2.set_allow_progress(true);
1.1003 + }
1.1004 + }
1.1005 + if ( VerifyIterativeGVN && PrintOpto ) {
1.1006 + if ( _verify_counter == _verify_full_passes )
1.1007 + tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
1.1008 + (int) _verify_full_passes);
1.1009 + else
1.1010 + tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1.1011 + (int) _verify_counter, (int) _verify_full_passes);
1.1012 + }
1.1013 +#endif
1.1014 +}
1.1015 +
1.1016 +
1.1017 +//------------------register_new_node_with_optimizer---------------------------
1.1018 +// Register a new node with the optimizer. Update the types array, the def-use
1.1019 +// info. Put on worklist.
1.1020 +Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
1.1021 + set_type_bottom(n);
1.1022 + _worklist.push(n);
1.1023 + if (orig != NULL) C->copy_node_notes_to(n, orig);
1.1024 + return n;
1.1025 +}
1.1026 +
1.1027 +//------------------------------transform--------------------------------------
1.1028 +// Non-recursive: idealize Node 'n' with respect to its inputs and its value
1.1029 +Node *PhaseIterGVN::transform( Node *n ) {
1.1030 + if (_delay_transform) {
1.1031 + // Register the node but don't optimize for now
1.1032 + register_new_node_with_optimizer(n);
1.1033 + return n;
1.1034 + }
1.1035 +
1.1036 + // If brand new node, make space in type array, and give it a type.
1.1037 + ensure_type_or_null(n);
1.1038 + if (type_or_null(n) == NULL) {
1.1039 + set_type_bottom(n);
1.1040 + }
1.1041 +
1.1042 + return transform_old(n);
1.1043 +}
1.1044 +
1.1045 +//------------------------------transform_old----------------------------------
1.1046 +Node *PhaseIterGVN::transform_old( Node *n ) {
1.1047 +#ifndef PRODUCT
1.1048 + debug_only(uint loop_count = 0;);
1.1049 + set_transforms();
1.1050 +#endif
1.1051 + // Remove 'n' from hash table in case it gets modified
1.1052 + _table.hash_delete(n);
1.1053 + if( VerifyIterativeGVN ) {
1.1054 + assert( !_table.find_index(n->_idx), "found duplicate entry in table");
1.1055 + }
1.1056 +
1.1057 + // Apply the Ideal call in a loop until it no longer applies
1.1058 + Node *k = n;
1.1059 + DEBUG_ONLY(dead_loop_check(k);)
1.1060 + DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
1.1061 + Node *i = k->Ideal(this, /*can_reshape=*/true);
1.1062 + assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1.1063 +#ifndef PRODUCT
1.1064 + if( VerifyIterativeGVN )
1.1065 + verify_step(k);
1.1066 + if( i && VerifyOpto ) {
1.1067 + if( !allow_progress() ) {
1.1068 + if (i->is_Add() && i->outcnt() == 1) {
1.1069 + // Switched input to left side because this is the only use
1.1070 + } else if( i->is_If() && (i->in(0) == NULL) ) {
1.1071 + // This IF is dead because it is dominated by an equivalent IF When
1.1072 + // dominating if changed, info is not propagated sparsely to 'this'
1.1073 + // Propagating this info further will spuriously identify other
1.1074 + // progress.
1.1075 + return i;
1.1076 + } else
1.1077 + set_progress();
1.1078 + } else
1.1079 + set_progress();
1.1080 + }
1.1081 +#endif
1.1082 +
1.1083 + while( i ) {
1.1084 +#ifndef PRODUCT
1.1085 + debug_only( if( loop_count >= K ) i->dump(4); )
1.1086 + assert(loop_count < K, "infinite loop in PhaseIterGVN::transform");
1.1087 + debug_only( loop_count++; )
1.1088 +#endif
1.1089 + assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
1.1090 + // Made a change; put users of original Node on worklist
1.1091 + add_users_to_worklist( k );
1.1092 + // Replacing root of transform tree?
1.1093 + if( k != i ) {
1.1094 + // Make users of old Node now use new.
1.1095 + subsume_node( k, i );
1.1096 + k = i;
1.1097 + }
1.1098 + DEBUG_ONLY(dead_loop_check(k);)
1.1099 + // Try idealizing again
1.1100 + DEBUG_ONLY(is_new = (k->outcnt() == 0);)
1.1101 + i = k->Ideal(this, /*can_reshape=*/true);
1.1102 + assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
1.1103 +#ifndef PRODUCT
1.1104 + if( VerifyIterativeGVN )
1.1105 + verify_step(k);
1.1106 + if( i && VerifyOpto ) set_progress();
1.1107 +#endif
1.1108 + }
1.1109 +
1.1110 + // If brand new node, make space in type array.
1.1111 + ensure_type_or_null(k);
1.1112 +
1.1113 + // See what kind of values 'k' takes on at runtime
1.1114 + const Type *t = k->Value(this);
1.1115 + assert(t != NULL, "value sanity");
1.1116 +
1.1117 + // Since I just called 'Value' to compute the set of run-time values
1.1118 + // for this Node, and 'Value' is non-local (and therefore expensive) I'll
1.1119 + // cache Value. Later requests for the local phase->type of this Node can
1.1120 + // use the cached Value instead of suffering with 'bottom_type'.
1.1121 + if (t != type_or_null(k)) {
1.1122 + NOT_PRODUCT( set_progress(); )
1.1123 + NOT_PRODUCT( inc_new_values();)
1.1124 + set_type(k, t);
1.1125 + // If k is a TypeNode, capture any more-precise type permanently into Node
1.1126 + k->raise_bottom_type(t);
1.1127 + // Move users of node to worklist
1.1128 + add_users_to_worklist( k );
1.1129 + }
1.1130 +
1.1131 + // If 'k' computes a constant, replace it with a constant
1.1132 + if( t->singleton() && !k->is_Con() ) {
1.1133 + NOT_PRODUCT( set_progress(); )
1.1134 + Node *con = makecon(t); // Make a constant
1.1135 + add_users_to_worklist( k );
1.1136 + subsume_node( k, con ); // Everybody using k now uses con
1.1137 + return con;
1.1138 + }
1.1139 +
1.1140 + // Now check for Identities
1.1141 + i = k->Identity(this); // Look for a nearby replacement
1.1142 + if( i != k ) { // Found? Return replacement!
1.1143 + NOT_PRODUCT( set_progress(); )
1.1144 + add_users_to_worklist( k );
1.1145 + subsume_node( k, i ); // Everybody using k now uses i
1.1146 + return i;
1.1147 + }
1.1148 +
1.1149 + // Global Value Numbering
1.1150 + i = hash_find_insert(k); // Check for pre-existing node
1.1151 + if( i && (i != k) ) {
1.1152 + // Return the pre-existing node if it isn't dead
1.1153 + NOT_PRODUCT( set_progress(); )
1.1154 + add_users_to_worklist( k );
1.1155 + subsume_node( k, i ); // Everybody using k now uses i
1.1156 + return i;
1.1157 + }
1.1158 +
1.1159 + // Return Idealized original
1.1160 + return k;
1.1161 +}
1.1162 +
1.1163 +//---------------------------------saturate------------------------------------
1.1164 +const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
1.1165 + const Type* limit_type) const {
1.1166 + return new_type->narrow(old_type);
1.1167 +}
1.1168 +
1.1169 +//------------------------------remove_globally_dead_node----------------------
1.1170 +// Kill a globally dead Node. All uses are also globally dead and are
1.1171 +// aggressively trimmed.
1.1172 +void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
1.1173 + enum DeleteProgress {
1.1174 + PROCESS_INPUTS,
1.1175 + PROCESS_OUTPUTS
1.1176 + };
1.1177 + assert(_stack.is_empty(), "not empty");
1.1178 + _stack.push(dead, PROCESS_INPUTS);
1.1179 +
1.1180 + while (_stack.is_nonempty()) {
1.1181 + dead = _stack.node();
1.1182 + uint progress_state = _stack.index();
1.1183 + assert(dead != C->root(), "killing root, eh?");
1.1184 + assert(!dead->is_top(), "add check for top when pushing");
1.1185 + NOT_PRODUCT( set_progress(); )
1.1186 + if (progress_state == PROCESS_INPUTS) {
1.1187 + // After following inputs, continue to outputs
1.1188 + _stack.set_index(PROCESS_OUTPUTS);
1.1189 + if (!dead->is_Con()) { // Don't kill cons but uses
1.1190 + bool recurse = false;
1.1191 + // Remove from hash table
1.1192 + _table.hash_delete( dead );
1.1193 + // Smash all inputs to 'dead', isolating him completely
1.1194 + for (uint i = 0; i < dead->req(); i++) {
1.1195 + Node *in = dead->in(i);
1.1196 + if (in != NULL && in != C->top()) { // Points to something?
1.1197 + int nrep = dead->replace_edge(in, NULL); // Kill edges
1.1198 + assert((nrep > 0), "sanity");
1.1199 + if (in->outcnt() == 0) { // Made input go dead?
1.1200 + _stack.push(in, PROCESS_INPUTS); // Recursively remove
1.1201 + recurse = true;
1.1202 + } else if (in->outcnt() == 1 &&
1.1203 + in->has_special_unique_user()) {
1.1204 + _worklist.push(in->unique_out());
1.1205 + } else if (in->outcnt() <= 2 && dead->is_Phi()) {
1.1206 + if (in->Opcode() == Op_Region) {
1.1207 + _worklist.push(in);
1.1208 + } else if (in->is_Store()) {
1.1209 + DUIterator_Fast imax, i = in->fast_outs(imax);
1.1210 + _worklist.push(in->fast_out(i));
1.1211 + i++;
1.1212 + if (in->outcnt() == 2) {
1.1213 + _worklist.push(in->fast_out(i));
1.1214 + i++;
1.1215 + }
1.1216 + assert(!(i < imax), "sanity");
1.1217 + }
1.1218 + }
1.1219 + if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
1.1220 + in->is_Proj() && in->in(0) != NULL && in->in(0)->is_Initialize()) {
1.1221 + // A Load that directly follows an InitializeNode is
1.1222 + // going away. The Stores that follow are candidates
1.1223 + // again to be captured by the InitializeNode.
1.1224 + for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
1.1225 + Node *n = in->fast_out(j);
1.1226 + if (n->is_Store()) {
1.1227 + _worklist.push(n);
1.1228 + }
1.1229 + }
1.1230 + }
1.1231 + } // if (in != NULL && in != C->top())
1.1232 + } // for (uint i = 0; i < dead->req(); i++)
1.1233 + if (recurse) {
1.1234 + continue;
1.1235 + }
1.1236 + } // if (!dead->is_Con())
1.1237 + } // if (progress_state == PROCESS_INPUTS)
1.1238 +
1.1239 + // Aggressively kill globally dead uses
1.1240 + // (Rather than pushing all the outs at once, we push one at a time,
1.1241 + // plus the parent to resume later, because of the indefinite number
1.1242 + // of edge deletions per loop trip.)
1.1243 + if (dead->outcnt() > 0) {
1.1244 + // Recursively remove output edges
1.1245 + _stack.push(dead->raw_out(0), PROCESS_INPUTS);
1.1246 + } else {
1.1247 + // Finished disconnecting all input and output edges.
1.1248 + _stack.pop();
1.1249 + // Remove dead node from iterative worklist
1.1250 + _worklist.remove(dead);
1.1251 + // Constant node that has no out-edges and has only one in-edge from
1.1252 + // root is usually dead. However, sometimes reshaping walk makes
1.1253 + // it reachable by adding use edges. So, we will NOT count Con nodes
1.1254 + // as dead to be conservative about the dead node count at any
1.1255 + // given time.
1.1256 + if (!dead->is_Con()) {
1.1257 + C->record_dead_node(dead->_idx);
1.1258 + }
1.1259 + if (dead->is_macro()) {
1.1260 + C->remove_macro_node(dead);
1.1261 + }
1.1262 + if (dead->is_expensive()) {
1.1263 + C->remove_expensive_node(dead);
1.1264 + }
1.1265 + }
1.1266 + } // while (_stack.is_nonempty())
1.1267 +}
1.1268 +
1.1269 +//------------------------------subsume_node-----------------------------------
1.1270 +// Remove users from node 'old' and add them to node 'nn'.
1.1271 +void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
1.1272 + assert( old != hash_find(old), "should already been removed" );
1.1273 + assert( old != C->top(), "cannot subsume top node");
1.1274 + // Copy debug or profile information to the new version:
1.1275 + C->copy_node_notes_to(nn, old);
1.1276 + // Move users of node 'old' to node 'nn'
1.1277 + for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
1.1278 + Node* use = old->last_out(i); // for each use...
1.1279 + // use might need re-hashing (but it won't if it's a new node)
1.1280 + bool is_in_table = _table.hash_delete( use );
1.1281 + // Update use-def info as well
1.1282 + // We remove all occurrences of old within use->in,
1.1283 + // so as to avoid rehashing any node more than once.
1.1284 + // The hash table probe swamps any outer loop overhead.
1.1285 + uint num_edges = 0;
1.1286 + for (uint jmax = use->len(), j = 0; j < jmax; j++) {
1.1287 + if (use->in(j) == old) {
1.1288 + use->set_req(j, nn);
1.1289 + ++num_edges;
1.1290 + }
1.1291 + }
1.1292 + // Insert into GVN hash table if unique
1.1293 + // If a duplicate, 'use' will be cleaned up when pulled off worklist
1.1294 + if( is_in_table ) {
1.1295 + hash_find_insert(use);
1.1296 + }
1.1297 + i -= num_edges; // we deleted 1 or more copies of this edge
1.1298 + }
1.1299 +
1.1300 + // Smash all inputs to 'old', isolating him completely
1.1301 + Node *temp = new (C) Node(1);
1.1302 + temp->init_req(0,nn); // Add a use to nn to prevent him from dying
1.1303 + remove_dead_node( old );
1.1304 + temp->del_req(0); // Yank bogus edge
1.1305 +#ifndef PRODUCT
1.1306 + if( VerifyIterativeGVN ) {
1.1307 + for ( int i = 0; i < _verify_window_size; i++ ) {
1.1308 + if ( _verify_window[i] == old )
1.1309 + _verify_window[i] = nn;
1.1310 + }
1.1311 + }
1.1312 +#endif
1.1313 + _worklist.remove(temp); // this can be necessary
1.1314 + temp->destruct(); // reuse the _idx of this little guy
1.1315 +}
1.1316 +
1.1317 +//------------------------------add_users_to_worklist--------------------------
1.1318 +void PhaseIterGVN::add_users_to_worklist0( Node *n ) {
1.1319 + for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1.1320 + _worklist.push(n->fast_out(i)); // Push on worklist
1.1321 + }
1.1322 +}
1.1323 +
1.1324 +void PhaseIterGVN::add_users_to_worklist( Node *n ) {
1.1325 + add_users_to_worklist0(n);
1.1326 +
1.1327 + // Move users of node to worklist
1.1328 + for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1.1329 + Node* use = n->fast_out(i); // Get use
1.1330 +
1.1331 + if( use->is_Multi() || // Multi-definer? Push projs on worklist
1.1332 + use->is_Store() ) // Enable store/load same address
1.1333 + add_users_to_worklist0(use);
1.1334 +
1.1335 + // If we changed the receiver type to a call, we need to revisit
1.1336 + // the Catch following the call. It's looking for a non-NULL
1.1337 + // receiver to know when to enable the regular fall-through path
1.1338 + // in addition to the NullPtrException path.
1.1339 + if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
1.1340 + Node* p = use->as_CallDynamicJava()->proj_out(TypeFunc::Control);
1.1341 + if (p != NULL) {
1.1342 + add_users_to_worklist0(p);
1.1343 + }
1.1344 + }
1.1345 +
1.1346 + if( use->is_Cmp() ) { // Enable CMP/BOOL optimization
1.1347 + add_users_to_worklist(use); // Put Bool on worklist
1.1348 + // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
1.1349 + // phi merging either 0 or 1 onto the worklist
1.1350 + if (use->outcnt() > 0) {
1.1351 + Node* bol = use->raw_out(0);
1.1352 + if (bol->outcnt() > 0) {
1.1353 + Node* iff = bol->raw_out(0);
1.1354 + if (iff->outcnt() == 2) {
1.1355 + Node* ifproj0 = iff->raw_out(0);
1.1356 + Node* ifproj1 = iff->raw_out(1);
1.1357 + if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
1.1358 + Node* region0 = ifproj0->raw_out(0);
1.1359 + Node* region1 = ifproj1->raw_out(0);
1.1360 + if( region0 == region1 )
1.1361 + add_users_to_worklist0(region0);
1.1362 + }
1.1363 + }
1.1364 + }
1.1365 + }
1.1366 + }
1.1367 +
1.1368 + uint use_op = use->Opcode();
1.1369 + // If changed Cast input, check Phi users for simple cycles
1.1370 + if( use->is_ConstraintCast() || use->is_CheckCastPP() ) {
1.1371 + for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1.1372 + Node* u = use->fast_out(i2);
1.1373 + if (u->is_Phi())
1.1374 + _worklist.push(u);
1.1375 + }
1.1376 + }
1.1377 + // If changed LShift inputs, check RShift users for useless sign-ext
1.1378 + if( use_op == Op_LShiftI ) {
1.1379 + for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1.1380 + Node* u = use->fast_out(i2);
1.1381 + if (u->Opcode() == Op_RShiftI)
1.1382 + _worklist.push(u);
1.1383 + }
1.1384 + }
1.1385 + // If changed AddI/SubI inputs, check CmpU for range check optimization.
1.1386 + if (use_op == Op_AddI || use_op == Op_SubI) {
1.1387 + for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1.1388 + Node* u = use->fast_out(i2);
1.1389 + if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
1.1390 + _worklist.push(u);
1.1391 + }
1.1392 + }
1.1393 + }
1.1394 + // If changed AddP inputs, check Stores for loop invariant
1.1395 + if( use_op == Op_AddP ) {
1.1396 + for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1.1397 + Node* u = use->fast_out(i2);
1.1398 + if (u->is_Mem())
1.1399 + _worklist.push(u);
1.1400 + }
1.1401 + }
1.1402 + // If changed initialization activity, check dependent Stores
1.1403 + if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
1.1404 + InitializeNode* init = use->as_Allocate()->initialization();
1.1405 + if (init != NULL) {
1.1406 + Node* imem = init->proj_out(TypeFunc::Memory);
1.1407 + if (imem != NULL) add_users_to_worklist0(imem);
1.1408 + }
1.1409 + }
1.1410 + if (use_op == Op_Initialize) {
1.1411 + Node* imem = use->as_Initialize()->proj_out(TypeFunc::Memory);
1.1412 + if (imem != NULL) add_users_to_worklist0(imem);
1.1413 + }
1.1414 + }
1.1415 +}
1.1416 +
1.1417 +/**
1.1418 + * Remove the speculative part of all types that we know of
1.1419 + */
1.1420 +void PhaseIterGVN::remove_speculative_types() {
1.1421 + assert(UseTypeSpeculation, "speculation is off");
1.1422 + for (uint i = 0; i < _types.Size(); i++) {
1.1423 + const Type* t = _types.fast_lookup(i);
1.1424 + if (t != NULL) {
1.1425 + _types.map(i, t->remove_speculative());
1.1426 + }
1.1427 + }
1.1428 + _table.check_no_speculative_types();
1.1429 +}
1.1430 +
1.1431 +//=============================================================================
1.1432 +#ifndef PRODUCT
1.1433 +uint PhaseCCP::_total_invokes = 0;
1.1434 +uint PhaseCCP::_total_constants = 0;
1.1435 +#endif
1.1436 +//------------------------------PhaseCCP---------------------------------------
1.1437 +// Conditional Constant Propagation, ala Wegman & Zadeck
1.1438 +PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
1.1439 + NOT_PRODUCT( clear_constants(); )
1.1440 + assert( _worklist.size() == 0, "" );
1.1441 + // Clear out _nodes from IterGVN. Must be clear to transform call.
1.1442 + _nodes.clear(); // Clear out from IterGVN
1.1443 + analyze();
1.1444 +}
1.1445 +
1.1446 +#ifndef PRODUCT
1.1447 +//------------------------------~PhaseCCP--------------------------------------
1.1448 +PhaseCCP::~PhaseCCP() {
1.1449 + inc_invokes();
1.1450 + _total_constants += count_constants();
1.1451 +}
1.1452 +#endif
1.1453 +
1.1454 +
1.1455 +#ifdef ASSERT
1.1456 +static bool ccp_type_widens(const Type* t, const Type* t0) {
1.1457 + assert(t->meet(t0) == t, "Not monotonic");
1.1458 + switch (t->base() == t0->base() ? t->base() : Type::Top) {
1.1459 + case Type::Int:
1.1460 + assert(t0->isa_int()->_widen <= t->isa_int()->_widen, "widen increases");
1.1461 + break;
1.1462 + case Type::Long:
1.1463 + assert(t0->isa_long()->_widen <= t->isa_long()->_widen, "widen increases");
1.1464 + break;
1.1465 + }
1.1466 + return true;
1.1467 +}
1.1468 +#endif //ASSERT
1.1469 +
1.1470 +//------------------------------analyze----------------------------------------
1.1471 +void PhaseCCP::analyze() {
1.1472 + // Initialize all types to TOP, optimistic analysis
1.1473 + for (int i = C->unique() - 1; i >= 0; i--) {
1.1474 + _types.map(i,Type::TOP);
1.1475 + }
1.1476 +
1.1477 + // Push root onto worklist
1.1478 + Unique_Node_List worklist;
1.1479 + worklist.push(C->root());
1.1480 +
1.1481 + // Pull from worklist; compute new value; push changes out.
1.1482 + // This loop is the meat of CCP.
1.1483 + while( worklist.size() ) {
1.1484 + Node *n = worklist.pop();
1.1485 + const Type *t = n->Value(this);
1.1486 + if (t != type(n)) {
1.1487 + assert(ccp_type_widens(t, type(n)), "ccp type must widen");
1.1488 +#ifndef PRODUCT
1.1489 + if( TracePhaseCCP ) {
1.1490 + t->dump();
1.1491 + do { tty->print("\t"); } while (tty->position() < 16);
1.1492 + n->dump();
1.1493 + }
1.1494 +#endif
1.1495 + set_type(n, t);
1.1496 + for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1.1497 + Node* m = n->fast_out(i); // Get user
1.1498 + if( m->is_Region() ) { // New path to Region? Must recheck Phis too
1.1499 + for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1.1500 + Node* p = m->fast_out(i2); // Propagate changes to uses
1.1501 + if( p->bottom_type() != type(p) ) // If not already bottomed out
1.1502 + worklist.push(p); // Propagate change to user
1.1503 + }
1.1504 + }
1.1505 + // If we changed the receiver type to a call, we need to revisit
1.1506 + // the Catch following the call. It's looking for a non-NULL
1.1507 + // receiver to know when to enable the regular fall-through path
1.1508 + // in addition to the NullPtrException path
1.1509 + if (m->is_Call()) {
1.1510 + for (DUIterator_Fast i2max, i2 = m->fast_outs(i2max); i2 < i2max; i2++) {
1.1511 + Node* p = m->fast_out(i2); // Propagate changes to uses
1.1512 + if (p->is_Proj() && p->as_Proj()->_con == TypeFunc::Control && p->outcnt() == 1)
1.1513 + worklist.push(p->unique_out());
1.1514 + }
1.1515 + }
1.1516 + if( m->bottom_type() != type(m) ) // If not already bottomed out
1.1517 + worklist.push(m); // Propagate change to user
1.1518 + }
1.1519 + }
1.1520 + }
1.1521 +}
1.1522 +
1.1523 +//------------------------------do_transform-----------------------------------
1.1524 +// Top level driver for the recursive transformer
1.1525 +void PhaseCCP::do_transform() {
1.1526 + // Correct leaves of new-space Nodes; they point to old-space.
1.1527 + C->set_root( transform(C->root())->as_Root() );
1.1528 + assert( C->top(), "missing TOP node" );
1.1529 + assert( C->root(), "missing root" );
1.1530 +}
1.1531 +
1.1532 +//------------------------------transform--------------------------------------
1.1533 +// Given a Node in old-space, clone him into new-space.
1.1534 +// Convert any of his old-space children into new-space children.
1.1535 +Node *PhaseCCP::transform( Node *n ) {
1.1536 + Node *new_node = _nodes[n->_idx]; // Check for transformed node
1.1537 + if( new_node != NULL )
1.1538 + return new_node; // Been there, done that, return old answer
1.1539 + new_node = transform_once(n); // Check for constant
1.1540 + _nodes.map( n->_idx, new_node ); // Flag as having been cloned
1.1541 +
1.1542 + // Allocate stack of size _nodes.Size()/2 to avoid frequent realloc
1.1543 + GrowableArray <Node *> trstack(C->unique() >> 1);
1.1544 +
1.1545 + trstack.push(new_node); // Process children of cloned node
1.1546 + while ( trstack.is_nonempty() ) {
1.1547 + Node *clone = trstack.pop();
1.1548 + uint cnt = clone->req();
1.1549 + for( uint i = 0; i < cnt; i++ ) { // For all inputs do
1.1550 + Node *input = clone->in(i);
1.1551 + if( input != NULL ) { // Ignore NULLs
1.1552 + Node *new_input = _nodes[input->_idx]; // Check for cloned input node
1.1553 + if( new_input == NULL ) {
1.1554 + new_input = transform_once(input); // Check for constant
1.1555 + _nodes.map( input->_idx, new_input );// Flag as having been cloned
1.1556 + trstack.push(new_input);
1.1557 + }
1.1558 + assert( new_input == clone->in(i), "insanity check");
1.1559 + }
1.1560 + }
1.1561 + }
1.1562 + return new_node;
1.1563 +}
1.1564 +
1.1565 +
1.1566 +//------------------------------transform_once---------------------------------
1.1567 +// For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
1.1568 +Node *PhaseCCP::transform_once( Node *n ) {
1.1569 + const Type *t = type(n);
1.1570 + // Constant? Use constant Node instead
1.1571 + if( t->singleton() ) {
1.1572 + Node *nn = n; // Default is to return the original constant
1.1573 + if( t == Type::TOP ) {
1.1574 + // cache my top node on the Compile instance
1.1575 + if( C->cached_top_node() == NULL || C->cached_top_node()->in(0) == NULL ) {
1.1576 + C->set_cached_top_node( ConNode::make(C, Type::TOP) );
1.1577 + set_type(C->top(), Type::TOP);
1.1578 + }
1.1579 + nn = C->top();
1.1580 + }
1.1581 + if( !n->is_Con() ) {
1.1582 + if( t != Type::TOP ) {
1.1583 + nn = makecon(t); // ConNode::make(t);
1.1584 + NOT_PRODUCT( inc_constants(); )
1.1585 + } else if( n->is_Region() ) { // Unreachable region
1.1586 + // Note: nn == C->top()
1.1587 + n->set_req(0, NULL); // Cut selfreference
1.1588 + // Eagerly remove dead phis to avoid phis copies creation.
1.1589 + for (DUIterator i = n->outs(); n->has_out(i); i++) {
1.1590 + Node* m = n->out(i);
1.1591 + if( m->is_Phi() ) {
1.1592 + assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
1.1593 + replace_node(m, nn);
1.1594 + --i; // deleted this phi; rescan starting with next position
1.1595 + }
1.1596 + }
1.1597 + }
1.1598 + replace_node(n,nn); // Update DefUse edges for new constant
1.1599 + }
1.1600 + return nn;
1.1601 + }
1.1602 +
1.1603 + // If x is a TypeNode, capture any more-precise type permanently into Node
1.1604 + if (t != n->bottom_type()) {
1.1605 + hash_delete(n); // changing bottom type may force a rehash
1.1606 + n->raise_bottom_type(t);
1.1607 + _worklist.push(n); // n re-enters the hash table via the worklist
1.1608 + }
1.1609 +
1.1610 + // Idealize graph using DU info. Must clone() into new-space.
1.1611 + // DU info is generally used to show profitability, progress or safety
1.1612 + // (but generally not needed for correctness).
1.1613 + Node *nn = n->Ideal_DU_postCCP(this);
1.1614 +
1.1615 + // TEMPORARY fix to ensure that 2nd GVN pass eliminates NULL checks
1.1616 + switch( n->Opcode() ) {
1.1617 + case Op_FastLock: // Revisit FastLocks for lock coarsening
1.1618 + case Op_If:
1.1619 + case Op_CountedLoopEnd:
1.1620 + case Op_Region:
1.1621 + case Op_Loop:
1.1622 + case Op_CountedLoop:
1.1623 + case Op_Conv2B:
1.1624 + case Op_Opaque1:
1.1625 + case Op_Opaque2:
1.1626 + _worklist.push(n);
1.1627 + break;
1.1628 + default:
1.1629 + break;
1.1630 + }
1.1631 + if( nn ) {
1.1632 + _worklist.push(n);
1.1633 + // Put users of 'n' onto worklist for second igvn transform
1.1634 + add_users_to_worklist(n);
1.1635 + return nn;
1.1636 + }
1.1637 +
1.1638 + return n;
1.1639 +}
1.1640 +
1.1641 +//---------------------------------saturate------------------------------------
1.1642 +const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
1.1643 + const Type* limit_type) const {
1.1644 + const Type* wide_type = new_type->widen(old_type, limit_type);
1.1645 + if (wide_type != new_type) { // did we widen?
1.1646 + // If so, we may have widened beyond the limit type. Clip it back down.
1.1647 + new_type = wide_type->filter(limit_type);
1.1648 + }
1.1649 + return new_type;
1.1650 +}
1.1651 +
1.1652 +//------------------------------print_statistics-------------------------------
1.1653 +#ifndef PRODUCT
1.1654 +void PhaseCCP::print_statistics() {
1.1655 + tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
1.1656 +}
1.1657 +#endif
1.1658 +
1.1659 +
1.1660 +//=============================================================================
1.1661 +#ifndef PRODUCT
1.1662 +uint PhasePeephole::_total_peepholes = 0;
1.1663 +#endif
1.1664 +//------------------------------PhasePeephole----------------------------------
1.1665 +// Conditional Constant Propagation, ala Wegman & Zadeck
1.1666 +PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
1.1667 + : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
1.1668 + NOT_PRODUCT( clear_peepholes(); )
1.1669 +}
1.1670 +
1.1671 +#ifndef PRODUCT
1.1672 +//------------------------------~PhasePeephole---------------------------------
1.1673 +PhasePeephole::~PhasePeephole() {
1.1674 + _total_peepholes += count_peepholes();
1.1675 +}
1.1676 +#endif
1.1677 +
1.1678 +//------------------------------transform--------------------------------------
1.1679 +Node *PhasePeephole::transform( Node *n ) {
1.1680 + ShouldNotCallThis();
1.1681 + return NULL;
1.1682 +}
1.1683 +
1.1684 +//------------------------------do_transform-----------------------------------
1.1685 +void PhasePeephole::do_transform() {
1.1686 + bool method_name_not_printed = true;
1.1687 +
1.1688 + // Examine each basic block
1.1689 + for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
1.1690 + Block* block = _cfg.get_block(block_number);
1.1691 + bool block_not_printed = true;
1.1692 +
1.1693 + // and each instruction within a block
1.1694 + uint end_index = block->number_of_nodes();
1.1695 + // block->end_idx() not valid after PhaseRegAlloc
1.1696 + for( uint instruction_index = 1; instruction_index < end_index; ++instruction_index ) {
1.1697 + Node *n = block->get_node(instruction_index);
1.1698 + if( n->is_Mach() ) {
1.1699 + MachNode *m = n->as_Mach();
1.1700 + int deleted_count = 0;
1.1701 + // check for peephole opportunities
1.1702 + MachNode *m2 = m->peephole( block, instruction_index, _regalloc, deleted_count, C );
1.1703 + if( m2 != NULL ) {
1.1704 +#ifndef PRODUCT
1.1705 + if( PrintOptoPeephole ) {
1.1706 + // Print method, first time only
1.1707 + if( C->method() && method_name_not_printed ) {
1.1708 + C->method()->print_short_name(); tty->cr();
1.1709 + method_name_not_printed = false;
1.1710 + }
1.1711 + // Print this block
1.1712 + if( Verbose && block_not_printed) {
1.1713 + tty->print_cr("in block");
1.1714 + block->dump();
1.1715 + block_not_printed = false;
1.1716 + }
1.1717 + // Print instructions being deleted
1.1718 + for( int i = (deleted_count - 1); i >= 0; --i ) {
1.1719 + block->get_node(instruction_index-i)->as_Mach()->format(_regalloc); tty->cr();
1.1720 + }
1.1721 + tty->print_cr("replaced with");
1.1722 + // Print new instruction
1.1723 + m2->format(_regalloc);
1.1724 + tty->print("\n\n");
1.1725 + }
1.1726 +#endif
1.1727 + // Remove old nodes from basic block and update instruction_index
1.1728 + // (old nodes still exist and may have edges pointing to them
1.1729 + // as register allocation info is stored in the allocator using
1.1730 + // the node index to live range mappings.)
1.1731 + uint safe_instruction_index = (instruction_index - deleted_count);
1.1732 + for( ; (instruction_index > safe_instruction_index); --instruction_index ) {
1.1733 + block->remove_node( instruction_index );
1.1734 + }
1.1735 + // install new node after safe_instruction_index
1.1736 + block->insert_node(m2, safe_instruction_index + 1);
1.1737 + end_index = block->number_of_nodes() - 1; // Recompute new block size
1.1738 + NOT_PRODUCT( inc_peepholes(); )
1.1739 + }
1.1740 + }
1.1741 + }
1.1742 + }
1.1743 +}
1.1744 +
1.1745 +//------------------------------print_statistics-------------------------------
1.1746 +#ifndef PRODUCT
1.1747 +void PhasePeephole::print_statistics() {
1.1748 + tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
1.1749 +}
1.1750 +#endif
1.1751 +
1.1752 +
1.1753 +//=============================================================================
1.1754 +//------------------------------set_req_X--------------------------------------
1.1755 +void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
1.1756 + assert( is_not_dead(n), "can not use dead node");
1.1757 + assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" );
1.1758 + Node *old = in(i);
1.1759 + set_req(i, n);
1.1760 +
1.1761 + // old goes dead?
1.1762 + if( old ) {
1.1763 + switch (old->outcnt()) {
1.1764 + case 0:
1.1765 + // Put into the worklist to kill later. We do not kill it now because the
1.1766 + // recursive kill will delete the current node (this) if dead-loop exists
1.1767 + if (!old->is_top())
1.1768 + igvn->_worklist.push( old );
1.1769 + break;
1.1770 + case 1:
1.1771 + if( old->is_Store() || old->has_special_unique_user() )
1.1772 + igvn->add_users_to_worklist( old );
1.1773 + break;
1.1774 + case 2:
1.1775 + if( old->is_Store() )
1.1776 + igvn->add_users_to_worklist( old );
1.1777 + if( old->Opcode() == Op_Region )
1.1778 + igvn->_worklist.push(old);
1.1779 + break;
1.1780 + case 3:
1.1781 + if( old->Opcode() == Op_Region ) {
1.1782 + igvn->_worklist.push(old);
1.1783 + igvn->add_users_to_worklist( old );
1.1784 + }
1.1785 + break;
1.1786 + default:
1.1787 + break;
1.1788 + }
1.1789 + }
1.1790 +
1.1791 +}
1.1792 +
1.1793 +//-------------------------------replace_by-----------------------------------
1.1794 +// Using def-use info, replace one node for another. Follow the def-use info
1.1795 +// to all users of the OLD node. Then make all uses point to the NEW node.
1.1796 +void Node::replace_by(Node *new_node) {
1.1797 + assert(!is_top(), "top node has no DU info");
1.1798 + for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
1.1799 + Node* use = last_out(i);
1.1800 + uint uses_found = 0;
1.1801 + for (uint j = 0; j < use->len(); j++) {
1.1802 + if (use->in(j) == this) {
1.1803 + if (j < use->req())
1.1804 + use->set_req(j, new_node);
1.1805 + else use->set_prec(j, new_node);
1.1806 + uses_found++;
1.1807 + }
1.1808 + }
1.1809 + i -= uses_found; // we deleted 1 or more copies of this edge
1.1810 + }
1.1811 +}
1.1812 +
1.1813 +//=============================================================================
1.1814 +//-----------------------------------------------------------------------------
1.1815 +void Type_Array::grow( uint i ) {
1.1816 + if( !_max ) {
1.1817 + _max = 1;
1.1818 + _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
1.1819 + _types[0] = NULL;
1.1820 + }
1.1821 + uint old = _max;
1.1822 + while( i >= _max ) _max <<= 1; // Double to fit
1.1823 + _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
1.1824 + memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
1.1825 +}
1.1826 +
1.1827 +//------------------------------dump-------------------------------------------
1.1828 +#ifndef PRODUCT
1.1829 +void Type_Array::dump() const {
1.1830 + uint max = Size();
1.1831 + for( uint i = 0; i < max; i++ ) {
1.1832 + if( _types[i] != NULL ) {
1.1833 + tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
1.1834 + }
1.1835 + }
1.1836 +}
1.1837 +#endif
