Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/callnode.hpp"
28 #include "opto/chaitin.hpp"
29 #include "opto/live.hpp"
30 #include "opto/machnode.hpp"
34 //=============================================================================
35 //------------------------------PhaseLive--------------------------------------
36 // Compute live-in/live-out. We use a totally incremental algorithm. The LIVE
37 // problem is monotonic. The steady-state solution looks like this: pull a
38 // block from the worklist. It has a set of delta's - values which are newly
39 // live-in from the block. Push these to the live-out sets of all predecessor
40 // blocks. At each predecessor, the new live-out values are ANDed with what is
41 // already live-out (extra stuff is added to the live-out sets). Then the
42 // remaining new live-out values are ANDed with what is locally defined.
43 // Leftover bits become the new live-in for the predecessor block, and the pred
44 // block is put on the worklist.
45 // The locally live-in stuff is computed once and added to predecessor
46 // live-out sets. This separate compilation is done in the outer loop below.
47 PhaseLive::PhaseLive( const PhaseCFG &cfg, LRG_List &names, Arena *arena ) : Phase(LIVE), _cfg(cfg), _names(names), _arena(arena), _live(0) {
48 }
50 void PhaseLive::compute(uint maxlrg) {
51 _maxlrg = maxlrg;
52 _worklist = new (_arena) Block_List();
54 // Init the sparse live arrays. This data is live on exit from here!
55 // The _live info is the live-out info.
56 _live = (IndexSet*)_arena->Amalloc(sizeof(IndexSet)*_cfg._num_blocks);
57 uint i;
58 for( i=0; i<_cfg._num_blocks; i++ ) {
59 _live[i].initialize(_maxlrg);
60 }
62 // Init the sparse arrays for delta-sets.
63 ResourceMark rm; // Nuke temp storage on exit
65 // Does the memory used by _defs and _deltas get reclaimed? Does it matter? TT
67 // Array of values defined locally in blocks
68 _defs = NEW_RESOURCE_ARRAY(IndexSet,_cfg._num_blocks);
69 for( i=0; i<_cfg._num_blocks; i++ ) {
70 _defs[i].initialize(_maxlrg);
71 }
73 // Array of delta-set pointers, indexed by block pre_order-1.
74 _deltas = NEW_RESOURCE_ARRAY(IndexSet*,_cfg._num_blocks);
75 memset( _deltas, 0, sizeof(IndexSet*)* _cfg._num_blocks);
77 _free_IndexSet = NULL;
79 // Blocks having done pass-1
80 VectorSet first_pass(Thread::current()->resource_area());
82 // Outer loop: must compute local live-in sets and push into predecessors.
83 uint iters = _cfg._num_blocks; // stat counters
84 for( uint j=_cfg._num_blocks; j>0; j-- ) {
85 Block *b = _cfg._blocks[j-1];
87 // Compute the local live-in set. Start with any new live-out bits.
88 IndexSet *use = getset( b );
89 IndexSet *def = &_defs[b->_pre_order-1];
90 DEBUG_ONLY(IndexSet *def_outside = getfreeset();)
91 uint i;
92 for( i=b->_nodes.size(); i>1; i-- ) {
93 Node *n = b->_nodes[i-1];
94 if( n->is_Phi() ) break;
96 uint r = _names[n->_idx];
97 assert(!def_outside->member(r), "Use of external LRG overlaps the same LRG defined in this block");
98 def->insert( r );
99 use->remove( r );
100 uint cnt = n->req();
101 for( uint k=1; k<cnt; k++ ) {
102 Node *nk = n->in(k);
103 uint nkidx = nk->_idx;
104 if( _cfg._bbs[nkidx] != b ) {
105 uint u = _names[nkidx];
106 use->insert( u );
107 DEBUG_ONLY(def_outside->insert( u );)
108 }
109 }
110 }
111 #ifdef ASSERT
112 def_outside->set_next(_free_IndexSet);
113 _free_IndexSet = def_outside; // Drop onto free list
114 #endif
115 // Remove anything defined by Phis and the block start instruction
116 for( uint k=i; k>0; k-- ) {
117 uint r = _names[b->_nodes[k-1]->_idx];
118 def->insert( r );
119 use->remove( r );
120 }
122 // Push these live-in things to predecessors
123 for( uint l=1; l<b->num_preds(); l++ ) {
124 Block *p = _cfg._bbs[b->pred(l)->_idx];
125 add_liveout( p, use, first_pass );
127 // PhiNode uses go in the live-out set of prior blocks.
128 for( uint k=i; k>0; k-- )
129 add_liveout( p, _names[b->_nodes[k-1]->in(l)->_idx], first_pass );
130 }
131 freeset( b );
132 first_pass.set(b->_pre_order);
134 // Inner loop: blocks that picked up new live-out values to be propagated
135 while( _worklist->size() ) {
136 // !!!!!
137 // #ifdef ASSERT
138 iters++;
139 // #endif
140 Block *b = _worklist->pop();
141 IndexSet *delta = getset(b);
142 assert( delta->count(), "missing delta set" );
144 // Add new-live-in to predecessors live-out sets
145 for( uint l=1; l<b->num_preds(); l++ )
146 add_liveout( _cfg._bbs[b->pred(l)->_idx], delta, first_pass );
148 freeset(b);
149 } // End of while-worklist-not-empty
151 } // End of for-all-blocks-outer-loop
153 // We explicitly clear all of the IndexSets which we are about to release.
154 // This allows us to recycle their internal memory into IndexSet's free list.
156 for( i=0; i<_cfg._num_blocks; i++ ) {
157 _defs[i].clear();
158 if (_deltas[i]) {
159 // Is this always true?
160 _deltas[i]->clear();
161 }
162 }
163 IndexSet *free = _free_IndexSet;
164 while (free != NULL) {
165 IndexSet *temp = free;
166 free = free->next();
167 temp->clear();
168 }
170 }
172 //------------------------------stats------------------------------------------
173 #ifndef PRODUCT
174 void PhaseLive::stats(uint iters) const {
175 }
176 #endif
178 //------------------------------getset-----------------------------------------
179 // Get an IndexSet for a block. Return existing one, if any. Make a new
180 // empty one if a prior one does not exist.
181 IndexSet *PhaseLive::getset( Block *p ) {
182 IndexSet *delta = _deltas[p->_pre_order-1];
183 if( !delta ) // Not on worklist?
184 // Get a free set; flag as being on worklist
185 delta = _deltas[p->_pre_order-1] = getfreeset();
186 return delta; // Return set of new live-out items
187 }
189 //------------------------------getfreeset-------------------------------------
190 // Pull from free list, or allocate. Internal allocation on the returned set
191 // is always from thread local storage.
192 IndexSet *PhaseLive::getfreeset( ) {
193 IndexSet *f = _free_IndexSet;
194 if( !f ) {
195 f = new IndexSet;
196 // f->set_arena(Thread::current()->resource_area());
197 f->initialize(_maxlrg, Thread::current()->resource_area());
198 } else {
199 // Pull from free list
200 _free_IndexSet = f->next();
201 //f->_cnt = 0; // Reset to empty
202 // f->set_arena(Thread::current()->resource_area());
203 f->initialize(_maxlrg, Thread::current()->resource_area());
204 }
205 return f;
206 }
208 //------------------------------freeset----------------------------------------
209 // Free an IndexSet from a block.
210 void PhaseLive::freeset( const Block *p ) {
211 IndexSet *f = _deltas[p->_pre_order-1];
212 f->set_next(_free_IndexSet);
213 _free_IndexSet = f; // Drop onto free list
214 _deltas[p->_pre_order-1] = NULL;
215 }
217 //------------------------------add_liveout------------------------------------
218 // Add a live-out value to a given blocks live-out set. If it is new, then
219 // also add it to the delta set and stick the block on the worklist.
220 void PhaseLive::add_liveout( Block *p, uint r, VectorSet &first_pass ) {
221 IndexSet *live = &_live[p->_pre_order-1];
222 if( live->insert(r) ) { // If actually inserted...
223 // We extended the live-out set. See if the value is generated locally.
224 // If it is not, then we must extend the live-in set.
225 if( !_defs[p->_pre_order-1].member( r ) ) {
226 if( !_deltas[p->_pre_order-1] && // Not on worklist?
227 first_pass.test(p->_pre_order) )
228 _worklist->push(p); // Actually go on worklist if already 1st pass
229 getset(p)->insert(r);
230 }
231 }
232 }
235 //------------------------------add_liveout------------------------------------
236 // Add a vector of live-out values to a given blocks live-out set.
237 void PhaseLive::add_liveout( Block *p, IndexSet *lo, VectorSet &first_pass ) {
238 IndexSet *live = &_live[p->_pre_order-1];
239 IndexSet *defs = &_defs[p->_pre_order-1];
240 IndexSet *on_worklist = _deltas[p->_pre_order-1];
241 IndexSet *delta = on_worklist ? on_worklist : getfreeset();
243 IndexSetIterator elements(lo);
244 uint r;
245 while ((r = elements.next()) != 0) {
246 if( live->insert(r) && // If actually inserted...
247 !defs->member( r ) ) // and not defined locally
248 delta->insert(r); // Then add to live-in set
249 }
251 if( delta->count() ) { // If actually added things
252 _deltas[p->_pre_order-1] = delta; // Flag as on worklist now
253 if( !on_worklist && // Not on worklist?
254 first_pass.test(p->_pre_order) )
255 _worklist->push(p); // Actually go on worklist if already 1st pass
256 } else { // Nothing there; just free it
257 delta->set_next(_free_IndexSet);
258 _free_IndexSet = delta; // Drop onto free list
259 }
260 }
262 #ifndef PRODUCT
263 //------------------------------dump-------------------------------------------
264 // Dump the live-out set for a block
265 void PhaseLive::dump( const Block *b ) const {
266 tty->print("Block %d: ",b->_pre_order);
267 tty->print("LiveOut: "); _live[b->_pre_order-1].dump();
268 uint cnt = b->_nodes.size();
269 for( uint i=0; i<cnt; i++ ) {
270 tty->print("L%d/", _names[b->_nodes[i]->_idx] );
271 b->_nodes[i]->dump();
272 }
273 tty->print("\n");
274 }
276 //------------------------------verify_base_ptrs-------------------------------
277 // Verify that base pointers and derived pointers are still sane.
278 void PhaseChaitin::verify_base_ptrs( ResourceArea *a ) const {
279 #ifdef ASSERT
280 Unique_Node_List worklist(a);
281 for( uint i = 0; i < _cfg._num_blocks; i++ ) {
282 Block *b = _cfg._blocks[i];
283 for( uint j = b->end_idx() + 1; j > 1; j-- ) {
284 Node *n = b->_nodes[j-1];
285 if( n->is_Phi() ) break;
286 // Found a safepoint?
287 if( n->is_MachSafePoint() ) {
288 MachSafePointNode *sfpt = n->as_MachSafePoint();
289 JVMState* jvms = sfpt->jvms();
290 if (jvms != NULL) {
291 // Now scan for a live derived pointer
292 if (jvms->oopoff() < sfpt->req()) {
293 // Check each derived/base pair
294 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
295 Node *check = sfpt->in(idx);
296 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
297 // search upwards through spills and spill phis for AddP
298 worklist.clear();
299 worklist.push(check);
300 uint k = 0;
301 while( k < worklist.size() ) {
302 check = worklist.at(k);
303 assert(check,"Bad base or derived pointer");
304 // See PhaseChaitin::find_base_for_derived() for all cases.
305 int isc = check->is_Copy();
306 if( isc ) {
307 worklist.push(check->in(isc));
308 } else if( check->is_Phi() ) {
309 for (uint m = 1; m < check->req(); m++)
310 worklist.push(check->in(m));
311 } else if( check->is_Con() ) {
312 if (is_derived) {
313 // Derived is NULL+offset
314 assert(!is_derived || check->bottom_type()->is_ptr()->ptr() == TypePtr::Null,"Bad derived pointer");
315 } else {
316 assert(check->bottom_type()->is_ptr()->_offset == 0,"Bad base pointer");
317 // Base either ConP(NULL) or loadConP
318 if (check->is_Mach()) {
319 assert(check->as_Mach()->ideal_Opcode() == Op_ConP,"Bad base pointer");
320 } else {
321 assert(check->Opcode() == Op_ConP &&
322 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null,"Bad base pointer");
323 }
324 }
325 } else if( check->bottom_type()->is_ptr()->_offset == 0 ) {
326 if(check->is_Proj() || check->is_Mach() &&
327 (check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
328 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
329 check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
330 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
331 #ifdef _LP64
332 UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP ||
333 UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN ||
334 #endif
335 check->as_Mach()->ideal_Opcode() == Op_LoadP ||
336 check->as_Mach()->ideal_Opcode() == Op_LoadKlass)) {
337 // Valid nodes
338 } else {
339 check->dump();
340 assert(false,"Bad base or derived pointer");
341 }
342 } else {
343 assert(is_derived,"Bad base pointer");
344 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP,"Bad derived pointer");
345 }
346 k++;
347 assert(k < 100000,"Derived pointer checking in infinite loop");
348 } // End while
349 }
350 } // End of check for derived pointers
351 } // End of Kcheck for debug info
352 } // End of if found a safepoint
353 } // End of forall instructions in block
354 } // End of forall blocks
355 #endif
356 }
358 //------------------------------verify-------------------------------------
359 // Verify that graphs and base pointers are still sane.
360 void PhaseChaitin::verify( ResourceArea *a, bool verify_ifg ) const {
361 #ifdef ASSERT
362 if( VerifyOpto || VerifyRegisterAllocator ) {
363 _cfg.verify();
364 verify_base_ptrs(a);
365 if(verify_ifg)
366 _ifg->verify(this);
367 }
368 #endif
369 }
371 #endif