1 #ifdef USE_PRAGMA_IDENT_SRC
2 #pragma ident "@(#)buildOopMap.cpp 1.37 07/05/05 17:06:11 JVM"
3 #endif
4 /*
5 * Copyright 2002-2008 Sun Microsystems, Inc. All Rights Reserved.
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 *
8 * This code is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License version 2 only, as
10 * published by the Free Software Foundation.
11 *
12 * This code is distributed in the hope that it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 * version 2 for more details (a copy is included in the LICENSE file that
16 * accompanied this code).
17 *
18 * You should have received a copy of the GNU General Public License version
19 * 2 along with this work; if not, write to the Free Software Foundation,
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
21 *
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
24 * have any questions.
25 *
26 */
27
28 #include "incls/_precompiled.incl"
29 #include "incls/_buildOopMap.cpp.incl"
30
31 // The functions in this file builds OopMaps after all scheduling is done.
32 //
33 // OopMaps contain a list of all registers and stack-slots containing oops (so
34 // they can be updated by GC). OopMaps also contain a list of derived-pointer
35 // base-pointer pairs. When the base is moved, the derived pointer moves to
36 // follow it. Finally, any registers holding callee-save values are also
37 // recorded. These might contain oops, but only the caller knows.
38 //
39 // BuildOopMaps implements a simple forward reaching-defs solution. At each
40 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are
41 // typed as pointers (no offset), then they are oops. Pointers+offsets are
42 // derived pointers, and bases can be found from them. Finally, we'll also
43 // track reaching callee-save values. Note that a copy of a callee-save value
44 // "kills" it's source, so that only 1 copy of a callee-save value is alive at
45 // a time.
46 //
47 // We run a simple bitvector liveness pass to help trim out dead oops. Due to
48 // irreducible loops, we can have a reaching def of an oop that only reaches
49 // along one path and no way to know if it's valid or not on the other path.
50 // The bitvectors are quite dense and the liveness pass is fast.
51 //
52 // At GC points, we consult this information to build OopMaps. All reaching
53 // defs typed as oops are added to the OopMap. Only 1 instance of a
54 // callee-save register can be recorded. For derived pointers, we'll have to
55 // find and record the register holding the base.
56 //
57 // The reaching def's is a simple 1-pass worklist approach. I tried a clever
58 // breadth-first approach but it was worse (showed O(n^2) in the
59 // pick-next-block code).
60 //
61 // The relevent data is kept in a struct of arrays (it could just as well be
62 // an array of structs, but the struct-of-arrays is generally a little more
63 // efficient). The arrays are indexed by register number (including
64 // stack-slots as registers) and so is bounded by 200 to 300 elements in
65 // practice. One array will map to a reaching def Node (or NULL for
66 // conflict/dead). The other array will map to a callee-saved register or
67 // OptoReg::Bad for not-callee-saved.
68
69
70 //------------------------------OopFlow----------------------------------------
71 // Structure to pass around
72 struct OopFlow : public ResourceObj {
73 short *_callees; // Array mapping register to callee-saved
74 Node **_defs; // array mapping register to reaching def
75 // or NULL if dead/conflict
76 // OopFlow structs, when not being actively modified, describe the _end_ of
77 // this block.
78 Block *_b; // Block for this struct
79 OopFlow *_next; // Next free OopFlow
80
81 OopFlow( short *callees, Node **defs ) : _callees(callees), _defs(defs),
82 _b(NULL), _next(NULL) { }
83
84 // Given reaching-defs for this block start, compute it for this block end
85 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
86
87 // Merge these two OopFlows into the 'this' pointer.
88 void merge( OopFlow *flow, int max_reg );
89
90 // Copy a 'flow' over an existing flow
91 void clone( OopFlow *flow, int max_size);
92
93 // Make a new OopFlow from scratch
94 static OopFlow *make( Arena *A, int max_size );
95
96 // Build an oopmap from the current flow info
97 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
98 };
99
100 //------------------------------compute_reach----------------------------------
101 // Given reaching-defs for this block start, compute it for this block end
102 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
103
104 for( uint i=0; i<_b->_nodes.size(); i++ ) {
105 Node *n = _b->_nodes[i];
106
107 if( n->jvms() ) { // Build an OopMap here?
108 JVMState *jvms = n->jvms();
109 // no map needed for leaf calls
110 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
111 int *live = (int*) (*safehash)[n];
112 assert( live, "must find live" );
113 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
114 }
115 }
116
117 // Assign new reaching def's.
118 // Note that I padded the _defs and _callees arrays so it's legal
119 // to index at _defs[OptoReg::Bad].
120 OptoReg::Name first = regalloc->get_reg_first(n);
121 OptoReg::Name second = regalloc->get_reg_second(n);
122 _defs[first] = n;
123 _defs[second] = n;
124
125 // Pass callee-save info around copies
126 int idx = n->is_Copy();
127 if( idx ) { // Copies move callee-save info
128 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
129 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
130 int tmp_first = _callees[old_first];
131 int tmp_second = _callees[old_second];
132 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
133 _callees[old_second] = OptoReg::Bad;
134 _callees[first] = tmp_first;
135 _callees[second] = tmp_second;
136 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
137 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
138 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
139 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
140 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
141 } else {
142 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
143 _callees[second] = OptoReg::Bad;
144
145 // Find base case for callee saves
146 if( n->is_Proj() && n->in(0)->is_Start() ) {
147 if( OptoReg::is_reg(first) &&
148 regalloc->_matcher.is_save_on_entry(first) )
149 _callees[first] = first;
150 if( OptoReg::is_reg(second) &&
151 regalloc->_matcher.is_save_on_entry(second) )
152 _callees[second] = second;
153 }
154 }
155 }
156 }
157
158 //------------------------------merge------------------------------------------
159 // Merge the given flow into the 'this' flow
160 void OopFlow::merge( OopFlow *flow, int max_reg ) {
161 assert( _b == NULL, "merging into a happy flow" );
162 assert( flow->_b, "this flow is still alive" );
163 assert( flow != this, "no self flow" );
164
165 // Do the merge. If there are any differences, drop to 'bottom' which
166 // is OptoReg::Bad or NULL depending.
167 for( int i=0; i<max_reg; i++ ) {
168 // Merge the callee-save's
169 if( _callees[i] != flow->_callees[i] )
170 _callees[i] = OptoReg::Bad;
171 // Merge the reaching defs
172 if( _defs[i] != flow->_defs[i] )
173 _defs[i] = NULL;
174 }
175
176 }
177
178 //------------------------------clone------------------------------------------
179 void OopFlow::clone( OopFlow *flow, int max_size ) {
180 _b = flow->_b;
181 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
182 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
183 }
184
185 //------------------------------make-------------------------------------------
186 OopFlow *OopFlow::make( Arena *A, int max_size ) {
187 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
188 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
189 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
190 OopFlow *flow = new (A) OopFlow(callees+1, defs+1);
191 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
192 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
193 return flow;
194 }
195
196 //------------------------------bit twiddlers----------------------------------
197 static int get_live_bit( int *live, int reg ) {
198 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
199 static void set_live_bit( int *live, int reg ) {
200 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
201 static void clr_live_bit( int *live, int reg ) {
202 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
203
204 //------------------------------build_oop_map----------------------------------
205 // Build an oopmap from the current flow info
206 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
207 int framesize = regalloc->_framesize;
208 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
209 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
210 memset(dup_check,0,OptoReg::stack0()) );
211
212 OopMap *omap = new OopMap( framesize, max_inarg_slot );
213 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL;
214 JVMState* jvms = n->jvms();
215
216 // For all registers do...
217 for( int reg=0; reg<max_reg; reg++ ) {
218 if( get_live_bit(live,reg) == 0 )
219 continue; // Ignore if not live
220
221 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
222 // register in that case we'll get an non-concrete register for the second
223 // half. We only need to tell the map the register once!
224 //
225 // However for the moment we disable this change and leave things as they
226 // were.
227
228 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
229
230 if (false && r->is_reg() && !r->is_concrete()) {
231 continue;
232 }
233
234 // See if dead (no reaching def).
235 Node *def = _defs[reg]; // Get reaching def
236 assert( def, "since live better have reaching def" );
237
238 // Classify the reaching def as oop, derived, callee-save, dead, or other
239 const Type *t = def->bottom_type();
240 if( t->isa_oop_ptr() ) { // Oop or derived?
241 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
242 #ifdef _LP64
243 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
244 // Make sure both are record from the same reaching def, but do not
245 // put both into the oopmap.
246 if( (reg&1) == 1 ) { // High half of oop-pair?
247 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
248 continue; // Do not record high parts in oopmap
249 }
250 #endif
251
252 // Check for a legal reg name in the oopMap and bailout if it is not.
253 if (!omap->legal_vm_reg_name(r)) {
254 regalloc->C->record_method_not_compilable("illegal oopMap register name");
255 continue;
256 }
257 if( t->is_ptr()->_offset == 0 ) { // Not derived?
258 if( mcall ) {
259 // Outgoing argument GC mask responsibility belongs to the callee,
260 // not the caller. Inspect the inputs to the call, to see if
261 // this live-range is one of them.
262 uint cnt = mcall->tf()->domain()->cnt();
263 uint j;
264 for( j = TypeFunc::Parms; j < cnt; j++)
265 if( mcall->in(j) == def )
266 break; // reaching def is an argument oop
267 if( j < cnt ) // arg oops dont go in GC map
268 continue; // Continue on to the next register
269 }
270 omap->set_oop(r);
271 } else { // Else it's derived.
272 // Find the base of the derived value.
273 uint i;
274 // Fast, common case, scan
275 for( i = jvms->oopoff(); i < n->req(); i+=2 )
276 if( n->in(i) == def ) break; // Common case
277 if( i == n->req() ) { // Missed, try a more generous scan
278 // Scan again, but this time peek through copies
279 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
280 Node *m = n->in(i); // Get initial derived value
281 while( 1 ) {
282 Node *d = def; // Get initial reaching def
283 while( 1 ) { // Follow copies of reaching def to end
284 if( m == d ) goto found; // breaks 3 loops
285 int idx = d->is_Copy();
286 if( !idx ) break;
287 d = d->in(idx); // Link through copy
288 }
289 int idx = m->is_Copy();
290 if( !idx ) break;
291 m = m->in(idx);
292 }
293 }
294 guarantee( 0, "must find derived/base pair" );
295 }
296 found: ;
297 Node *base = n->in(i+1); // Base is other half of pair
298 int breg = regalloc->get_reg_first(base);
299 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
300
301 // I record liveness at safepoints BEFORE I make the inputs
302 // live. This is because argument oops are NOT live at a
303 // safepoint (or at least they cannot appear in the oopmap).
304 // Thus bases of base/derived pairs might not be in the
305 // liveness data but they need to appear in the oopmap.
306 if( get_live_bit(live,breg) == 0 ) {// Not live?
307 // Flag it, so next derived pointer won't re-insert into oopmap
308 set_live_bit(live,breg);
309 // Already missed our turn?
310 if( breg < reg ) {
311 if (b->is_stack() || b->is_concrete() || true ) {
312 omap->set_oop( b);
313 }
314 }
315 }
316 if (b->is_stack() || b->is_concrete() || true ) {
317 omap->set_derived_oop( r, b);
318 }
319 }
320
321 } else if( t->isa_narrowoop() ) {
322 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
323 // Check for a legal reg name in the oopMap and bailout if it is not.
324 if (!omap->legal_vm_reg_name(r)) {
325 regalloc->C->record_method_not_compilable("illegal oopMap register name");
326 continue;
327 }
328 if( mcall ) {
329 // Outgoing argument GC mask responsibility belongs to the callee,
330 // not the caller. Inspect the inputs to the call, to see if
331 // this live-range is one of them.
332 uint cnt = mcall->tf()->domain()->cnt();
333 uint j;
334 for( j = TypeFunc::Parms; j < cnt; j++)
335 if( mcall->in(j) == def )
336 break; // reaching def is an argument oop
337 if( j < cnt ) // arg oops dont go in GC map
338 continue; // Continue on to the next register
339 }
340 omap->set_narrowoop(r);
341 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
342 // It's a callee-save value
343 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
344 debug_only( dup_check[_callees[reg]]=1; )
345 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
346 if ( callee->is_concrete() || true ) {
347 omap->set_callee_saved( r, callee);
348 }
349
350 } else {
351 // Other - some reaching non-oop value
352 omap->set_value( r);
353 }
354
355 }
356
357 #ifdef ASSERT
358 /* Nice, Intel-only assert
359 int cnt_callee_saves=0;
360 int reg2 = 0;
361 while (OptoReg::is_reg(reg2)) {
362 if( dup_check[reg2] != 0) cnt_callee_saves++;
363 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
364 reg2++;
365 }
366 */
367 #endif
368
369 return omap;
370 }
371
372 //------------------------------do_liveness------------------------------------
373 // Compute backwards liveness on registers
374 static void do_liveness( PhaseRegAlloc *regalloc, PhaseCFG *cfg, Block_List *worklist, int max_reg_ints, Arena *A, Dict *safehash ) {
375 int *live = NEW_ARENA_ARRAY(A, int, (cfg->_num_blocks+1) * max_reg_ints);
376 int *tmp_live = &live[cfg->_num_blocks * max_reg_ints];
377 Node *root = cfg->C->root();
378 // On CISC platforms, get the node representing the stack pointer that regalloc
379 // used for spills
380 Node *fp = NodeSentinel;
381 if (UseCISCSpill && root->req() > 1) {
382 fp = root->in(1)->in(TypeFunc::FramePtr);
383 }
384 memset( live, 0, cfg->_num_blocks * (max_reg_ints<<LogBytesPerInt) );
385 // Push preds onto worklist
386 for( uint i=1; i<root->req(); i++ )
387 worklist->push(cfg->_bbs[root->in(i)->_idx]);
388
389 // ZKM.jar includes tiny infinite loops which are unreached from below.
390 // If we missed any blocks, we'll retry here after pushing all missed
391 // blocks on the worklist. Normally this outer loop never trips more
392 // than once.
393 while( 1 ) {
394
395 while( worklist->size() ) { // Standard worklist algorithm
396 Block *b = worklist->rpop();
397
398 // Copy first successor into my tmp_live space
399 int s0num = b->_succs[0]->_pre_order;
400 int *t = &live[s0num*max_reg_ints];
401 for( int i=0; i<max_reg_ints; i++ )
402 tmp_live[i] = t[i];
403
404 // OR in the remaining live registers
405 for( uint j=1; j<b->_num_succs; j++ ) {
406 uint sjnum = b->_succs[j]->_pre_order;
407 int *t = &live[sjnum*max_reg_ints];
408 for( int i=0; i<max_reg_ints; i++ )
409 tmp_live[i] |= t[i];
410 }
411
412 // Now walk tmp_live up the block backwards, computing live
413 for( int k=b->_nodes.size()-1; k>=0; k-- ) {
414 Node *n = b->_nodes[k];
415 // KILL def'd bits
416 int first = regalloc->get_reg_first(n);
417 int second = regalloc->get_reg_second(n);
418 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
419 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
420
421 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
422
423 // Check if m is potentially a CISC alternate instruction (i.e, possibly
424 // synthesized by RegAlloc from a conventional instruction and a
425 // spilled input)
426 bool is_cisc_alternate = false;
427 if (UseCISCSpill && m) {
428 is_cisc_alternate = m->is_cisc_alternate();
429 }
430
431 // GEN use'd bits
432 for( uint l=1; l<n->req(); l++ ) {
433 Node *def = n->in(l);
434 assert(def != 0, "input edge required");
435 int first = regalloc->get_reg_first(def);
436 int second = regalloc->get_reg_second(def);
437 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first);
438 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second);
439 // If we use the stack pointer in a cisc-alternative instruction,
440 // check for use as a memory operand. Then reconstruct the RegName
441 // for this stack location, and set the appropriate bit in the
442 // live vector 4987749.
443 if (is_cisc_alternate && def == fp) {
444 const TypePtr *adr_type = NULL;
445 intptr_t offset;
446 const Node* base = m->get_base_and_disp(offset, adr_type);
447 if (base == NodeSentinel) {
448 // Machnode has multiple memory inputs. We are unable to reason
449 // with these, but are presuming (with trepidation) that not any of
450 // them are oops. This can be fixed by making get_base_and_disp()
451 // look at a specific input instead of all inputs.
452 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
453 } else if (base != fp || offset == Type::OffsetBot) {
454 // Do nothing: the fp operand is either not from a memory use
455 // (base == NULL) OR the fp is used in a non-memory context
456 // (base is some other register) OR the offset is not constant,
457 // so it is not a stack slot.
458 } else {
459 assert(offset >= 0, "unexpected negative offset");
460 offset -= (offset % jintSize); // count the whole word
461 int stack_reg = regalloc->offset2reg(offset);
462 if (OptoReg::is_stack(stack_reg)) {
463 set_live_bit(tmp_live, stack_reg);
464 } else {
465 assert(false, "stack_reg not on stack?");
466 }
467 }
468 }
469 }
470
471 if( n->jvms() ) { // Record liveness at safepoint
472
473 // This placement of this stanza means inputs to calls are
474 // considered live at the callsite's OopMap. Argument oops are
475 // hence live, but NOT included in the oopmap. See cutout in
476 // build_oop_map. Debug oops are live (and in OopMap).
477 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
478 for( int l=0; l<max_reg_ints; l++ )
479 n_live[l] = tmp_live[l];
480 safehash->Insert(n,n_live);
481 }
482
483 }
484
485 // Now at block top, see if we have any changes. If so, propagate
486 // to prior blocks.
487 int *old_live = &live[b->_pre_order*max_reg_ints];
488 int l;
489 for( l=0; l<max_reg_ints; l++ )
490 if( tmp_live[l] != old_live[l] )
491 break;
492 if( l<max_reg_ints ) { // Change!
493 // Copy in new value
494 for( l=0; l<max_reg_ints; l++ )
495 old_live[l] = tmp_live[l];
496 // Push preds onto worklist
497 for( l=1; l<(int)b->num_preds(); l++ )
498 worklist->push(cfg->_bbs[b->pred(l)->_idx]);
499 }
500 }
501
502 // Scan for any missing safepoints. Happens to infinite loops
503 // ala ZKM.jar
504 uint i;
505 for( i=1; i<cfg->_num_blocks; i++ ) {
506 Block *b = cfg->_blocks[i];
507 uint j;
508 for( j=1; j<b->_nodes.size(); j++ )
509 if( b->_nodes[j]->jvms() &&
510 (*safehash)[b->_nodes[j]] == NULL )
511 break;
512 if( j<b->_nodes.size() ) break;
513 }
514 if( i == cfg->_num_blocks )
515 break; // Got 'em all
516 #ifndef PRODUCT
517 if( PrintOpto && Verbose )
518 tty->print_cr("retripping live calc");
519 #endif
520 // Force the issue (expensively): recheck everybody
521 for( i=1; i<cfg->_num_blocks; i++ )
522 worklist->push(cfg->_blocks[i]);
523 }
524
525 }
526
527 //------------------------------BuildOopMaps-----------------------------------
528 // Collect GC mask info - where are all the OOPs?
529 void Compile::BuildOopMaps() {
530 NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); )
531 // Can't resource-mark because I need to leave all those OopMaps around,
532 // or else I need to resource-mark some arena other than the default.
533 // ResourceMark rm; // Reclaim all OopFlows when done
534 int max_reg = _regalloc->_max_reg; // Current array extent
535
536 Arena *A = Thread::current()->resource_area();
537 Block_List worklist; // Worklist of pending blocks
538
539 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt;
540 Dict *safehash = NULL; // Used for assert only
541 // Compute a backwards liveness per register. Needs a bitarray of
542 // #blocks x (#registers, rounded up to ints)
543 safehash = new Dict(cmpkey,hashkey,A);
544 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash );
545 OopFlow *free_list = NULL; // Free, unused
546
547 // Array mapping blocks to completed oopflows
548 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->_num_blocks);
549 memset( flows, 0, _cfg->_num_blocks*sizeof(OopFlow*) );
550
551
552 // Do the first block 'by hand' to prime the worklist
553 Block *entry = _cfg->_blocks[1];
554 OopFlow *rootflow = OopFlow::make(A,max_reg);
555 // Initialize to 'bottom' (not 'top')
556 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
557 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
558 flows[entry->_pre_order] = rootflow;
559
560 // Do the first block 'by hand' to prime the worklist
561 rootflow->_b = entry;
562 rootflow->compute_reach( _regalloc, max_reg, safehash );
563 for( uint i=0; i<entry->_num_succs; i++ )
564 worklist.push(entry->_succs[i]);
565
566 // Now worklist contains blocks which have some, but perhaps not all,
567 // predecessors visited.
568 while( worklist.size() ) {
569 // Scan for a block with all predecessors visited, or any randoms slob
570 // otherwise. All-preds-visited order allows me to recycle OopFlow
571 // structures rapidly and cut down on the memory footprint.
572 // Note: not all predecessors might be visited yet (must happen for
573 // irreducible loops). This is OK, since every live value must have the
574 // SAME reaching def for the block, so any reaching def is OK.
575 uint i;
576
577 Block *b = worklist.pop();
578 // Ignore root block
579 if( b == _cfg->_broot ) continue;
580 // Block is already done? Happens if block has several predecessors,
581 // he can get on the worklist more than once.
582 if( flows[b->_pre_order] ) continue;
583
584 // If this block has a visited predecessor AND that predecessor has this
585 // last block as his only undone child, we can move the OopFlow from the
586 // pred to this block. Otherwise we have to grab a new OopFlow.
587 OopFlow *flow = NULL; // Flag for finding optimized flow
588 Block *pred = (Block*)0xdeadbeef;
589 uint j;
590 // Scan this block's preds to find a done predecessor
591 for( j=1; j<b->num_preds(); j++ ) {
592 Block *p = _cfg->_bbs[b->pred(j)->_idx];
593 OopFlow *p_flow = flows[p->_pre_order];
594 if( p_flow ) { // Predecessor is done
595 assert( p_flow->_b == p, "cross check" );
596 pred = p; // Record some predecessor
597 // If all successors of p are done except for 'b', then we can carry
598 // p_flow forward to 'b' without copying, otherwise we have to draw
599 // from the free_list and clone data.
600 uint k;
601 for( k=0; k<p->_num_succs; k++ )
602 if( !flows[p->_succs[k]->_pre_order] &&
603 p->_succs[k] != b )
604 break;
605
606 // Either carry-forward the now-unused OopFlow for b's use
607 // or draw a new one from the free list
608 if( k==p->_num_succs ) {
609 flow = p_flow;
610 break; // Found an ideal pred, use him
611 }
612 }
613 }
614
615 if( flow ) {
616 // We have an OopFlow that's the last-use of a predecessor.
617 // Carry it forward.
618 } else { // Draw a new OopFlow from the freelist
619 if( !free_list )
620 free_list = OopFlow::make(A,max_reg);
621 flow = free_list;
622 assert( flow->_b == NULL, "oopFlow is not free" );
623 free_list = flow->_next;
624 flow->_next = NULL;
625
626 // Copy/clone over the data
627 flow->clone(flows[pred->_pre_order], max_reg);
628 }
629
630 // Mark flow for block. Blocks can only be flowed over once,
631 // because after the first time they are guarded from entering
632 // this code again.
633 assert( flow->_b == pred, "have some prior flow" );
634 flow->_b = NULL;
635
636 // Now push flow forward
637 flows[b->_pre_order] = flow;// Mark flow for this block
638 flow->_b = b;
639 flow->compute_reach( _regalloc, max_reg, safehash );
640
641 // Now push children onto worklist
642 for( i=0; i<b->_num_succs; i++ )
643 worklist.push(b->_succs[i]);
644
645 }
646 }