1 /*
2 * Copyright (c) 2002, 2009, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_buildOopMap.cpp.incl"
27
28 // The functions in this file builds OopMaps after all scheduling is done.
29 //
30 // OopMaps contain a list of all registers and stack-slots containing oops (so
31 // they can be updated by GC). OopMaps also contain a list of derived-pointer
32 // base-pointer pairs. When the base is moved, the derived pointer moves to
33 // follow it. Finally, any registers holding callee-save values are also
34 // recorded. These might contain oops, but only the caller knows.
35 //
36 // BuildOopMaps implements a simple forward reaching-defs solution. At each
37 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are
38 // typed as pointers (no offset), then they are oops. Pointers+offsets are
39 // derived pointers, and bases can be found from them. Finally, we'll also
40 // track reaching callee-save values. Note that a copy of a callee-save value
41 // "kills" it's source, so that only 1 copy of a callee-save value is alive at
42 // a time.
43 //
44 // We run a simple bitvector liveness pass to help trim out dead oops. Due to
45 // irreducible loops, we can have a reaching def of an oop that only reaches
46 // along one path and no way to know if it's valid or not on the other path.
47 // The bitvectors are quite dense and the liveness pass is fast.
48 //
49 // At GC points, we consult this information to build OopMaps. All reaching
50 // defs typed as oops are added to the OopMap. Only 1 instance of a
51 // callee-save register can be recorded. For derived pointers, we'll have to
52 // find and record the register holding the base.
53 //
54 // The reaching def's is a simple 1-pass worklist approach. I tried a clever
55 // breadth-first approach but it was worse (showed O(n^2) in the
56 // pick-next-block code).
57 //
58 // The relevant data is kept in a struct of arrays (it could just as well be
59 // an array of structs, but the struct-of-arrays is generally a little more
60 // efficient). The arrays are indexed by register number (including
61 // stack-slots as registers) and so is bounded by 200 to 300 elements in
62 // practice. One array will map to a reaching def Node (or NULL for
63 // conflict/dead). The other array will map to a callee-saved register or
64 // OptoReg::Bad for not-callee-saved.
65
66
67 //------------------------------OopFlow----------------------------------------
68 // Structure to pass around
69 struct OopFlow : public ResourceObj {
70 short *_callees; // Array mapping register to callee-saved
71 Node **_defs; // array mapping register to reaching def
72 // or NULL if dead/conflict
73 // OopFlow structs, when not being actively modified, describe the _end_ of
74 // this block.
75 Block *_b; // Block for this struct
76 OopFlow *_next; // Next free OopFlow
77 // or NULL if dead/conflict
78 Compile* C;
79
80 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs),
81 _b(NULL), _next(NULL), C(c) { }
82
83 // Given reaching-defs for this block start, compute it for this block end
84 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
85
86 // Merge these two OopFlows into the 'this' pointer.
87 void merge( OopFlow *flow, int max_reg );
88
89 // Copy a 'flow' over an existing flow
90 void clone( OopFlow *flow, int max_size);
91
92 // Make a new OopFlow from scratch
93 static OopFlow *make( Arena *A, int max_size, Compile* C );
94
95 // Build an oopmap from the current flow info
96 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
97 };
98
99 //------------------------------compute_reach----------------------------------
100 // Given reaching-defs for this block start, compute it for this block end
101 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
102
103 for( uint i=0; i<_b->_nodes.size(); i++ ) {
104 Node *n = _b->_nodes[i];
105
106 if( n->jvms() ) { // Build an OopMap here?
107 JVMState *jvms = n->jvms();
108 // no map needed for leaf calls
109 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
110 int *live = (int*) (*safehash)[n];
111 assert( live, "must find live" );
112 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
113 }
114 }
115
116 // Assign new reaching def's.
117 // Note that I padded the _defs and _callees arrays so it's legal
118 // to index at _defs[OptoReg::Bad].
119 OptoReg::Name first = regalloc->get_reg_first(n);
120 OptoReg::Name second = regalloc->get_reg_second(n);
121 _defs[first] = n;
122 _defs[second] = n;
123
124 // Pass callee-save info around copies
125 int idx = n->is_Copy();
126 if( idx ) { // Copies move callee-save info
127 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
128 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
129 int tmp_first = _callees[old_first];
130 int tmp_second = _callees[old_second];
131 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
132 _callees[old_second] = OptoReg::Bad;
133 _callees[first] = tmp_first;
134 _callees[second] = tmp_second;
135 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
136 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
137 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
138 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
139 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
140 } else {
141 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
142 _callees[second] = OptoReg::Bad;
143
144 // Find base case for callee saves
145 if( n->is_Proj() && n->in(0)->is_Start() ) {
146 if( OptoReg::is_reg(first) &&
147 regalloc->_matcher.is_save_on_entry(first) )
148 _callees[first] = first;
149 if( OptoReg::is_reg(second) &&
150 regalloc->_matcher.is_save_on_entry(second) )
151 _callees[second] = second;
152 }
153 }
154 }
155 }
156
157 //------------------------------merge------------------------------------------
158 // Merge the given flow into the 'this' flow
159 void OopFlow::merge( OopFlow *flow, int max_reg ) {
160 assert( _b == NULL, "merging into a happy flow" );
161 assert( flow->_b, "this flow is still alive" );
162 assert( flow != this, "no self flow" );
163
164 // Do the merge. If there are any differences, drop to 'bottom' which
165 // is OptoReg::Bad or NULL depending.
166 for( int i=0; i<max_reg; i++ ) {
167 // Merge the callee-save's
168 if( _callees[i] != flow->_callees[i] )
169 _callees[i] = OptoReg::Bad;
170 // Merge the reaching defs
171 if( _defs[i] != flow->_defs[i] )
172 _defs[i] = NULL;
173 }
174
175 }
176
177 //------------------------------clone------------------------------------------
178 void OopFlow::clone( OopFlow *flow, int max_size ) {
179 _b = flow->_b;
180 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
181 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
182 }
183
184 //------------------------------make-------------------------------------------
185 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) {
186 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
187 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
188 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
189 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C);
190 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
191 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
192 return flow;
193 }
194
195 //------------------------------bit twiddlers----------------------------------
196 static int get_live_bit( int *live, int reg ) {
197 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
198 static void set_live_bit( int *live, int reg ) {
199 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
200 static void clr_live_bit( int *live, int reg ) {
201 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
202
203 //------------------------------build_oop_map----------------------------------
204 // Build an oopmap from the current flow info
205 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
206 int framesize = regalloc->_framesize;
207 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
208 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
209 memset(dup_check,0,OptoReg::stack0()) );
210
211 OopMap *omap = new OopMap( framesize, max_inarg_slot );
212 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL;
213 JVMState* jvms = n->jvms();
214
215 // For all registers do...
216 for( int reg=0; reg<max_reg; reg++ ) {
217 if( get_live_bit(live,reg) == 0 )
218 continue; // Ignore if not live
219
220 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
221 // register in that case we'll get an non-concrete register for the second
222 // half. We only need to tell the map the register once!
223 //
224 // However for the moment we disable this change and leave things as they
225 // were.
226
227 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
228
229 if (false && r->is_reg() && !r->is_concrete()) {
230 continue;
231 }
232
233 // See if dead (no reaching def).
234 Node *def = _defs[reg]; // Get reaching def
235 assert( def, "since live better have reaching def" );
236
237 // Classify the reaching def as oop, derived, callee-save, dead, or other
238 const Type *t = def->bottom_type();
239 if( t->isa_oop_ptr() ) { // Oop or derived?
240 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
241 #ifdef _LP64
242 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
243 // Make sure both are record from the same reaching def, but do not
244 // put both into the oopmap.
245 if( (reg&1) == 1 ) { // High half of oop-pair?
246 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
247 continue; // Do not record high parts in oopmap
248 }
249 #endif
250
251 // Check for a legal reg name in the oopMap and bailout if it is not.
252 if (!omap->legal_vm_reg_name(r)) {
253 regalloc->C->record_method_not_compilable("illegal oopMap register name");
254 continue;
255 }
256 if( t->is_ptr()->_offset == 0 ) { // Not derived?
257 if( mcall ) {
258 // Outgoing argument GC mask responsibility belongs to the callee,
259 // not the caller. Inspect the inputs to the call, to see if
260 // this live-range is one of them.
261 uint cnt = mcall->tf()->domain()->cnt();
262 uint j;
263 for( j = TypeFunc::Parms; j < cnt; j++)
264 if( mcall->in(j) == def )
265 break; // reaching def is an argument oop
266 if( j < cnt ) // arg oops dont go in GC map
267 continue; // Continue on to the next register
268 }
269 omap->set_oop(r);
270 } else { // Else it's derived.
271 // Find the base of the derived value.
272 uint i;
273 // Fast, common case, scan
274 for( i = jvms->oopoff(); i < n->req(); i+=2 )
275 if( n->in(i) == def ) break; // Common case
276 if( i == n->req() ) { // Missed, try a more generous scan
277 // Scan again, but this time peek through copies
278 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
279 Node *m = n->in(i); // Get initial derived value
280 while( 1 ) {
281 Node *d = def; // Get initial reaching def
282 while( 1 ) { // Follow copies of reaching def to end
283 if( m == d ) goto found; // breaks 3 loops
284 int idx = d->is_Copy();
285 if( !idx ) break;
286 d = d->in(idx); // Link through copy
287 }
288 int idx = m->is_Copy();
289 if( !idx ) break;
290 m = m->in(idx);
291 }
292 }
293 guarantee( 0, "must find derived/base pair" );
294 }
295 found: ;
296 Node *base = n->in(i+1); // Base is other half of pair
297 int breg = regalloc->get_reg_first(base);
298 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
299
300 // I record liveness at safepoints BEFORE I make the inputs
301 // live. This is because argument oops are NOT live at a
302 // safepoint (or at least they cannot appear in the oopmap).
303 // Thus bases of base/derived pairs might not be in the
304 // liveness data but they need to appear in the oopmap.
305 if( get_live_bit(live,breg) == 0 ) {// Not live?
306 // Flag it, so next derived pointer won't re-insert into oopmap
307 set_live_bit(live,breg);
308 // Already missed our turn?
309 if( breg < reg ) {
310 if (b->is_stack() || b->is_concrete() || true ) {
311 omap->set_oop( b);
312 }
313 }
314 }
315 if (b->is_stack() || b->is_concrete() || true ) {
316 omap->set_derived_oop( r, b);
317 }
318 }
319
320 } else if( t->isa_narrowoop() ) {
321 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
322 // Check for a legal reg name in the oopMap and bailout if it is not.
323 if (!omap->legal_vm_reg_name(r)) {
324 regalloc->C->record_method_not_compilable("illegal oopMap register name");
325 continue;
326 }
327 if( mcall ) {
328 // Outgoing argument GC mask responsibility belongs to the callee,
329 // not the caller. Inspect the inputs to the call, to see if
330 // this live-range is one of them.
331 uint cnt = mcall->tf()->domain()->cnt();
332 uint j;
333 for( j = TypeFunc::Parms; j < cnt; j++)
334 if( mcall->in(j) == def )
335 break; // reaching def is an argument oop
336 if( j < cnt ) // arg oops dont go in GC map
337 continue; // Continue on to the next register
338 }
339 omap->set_narrowoop(r);
340 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
341 // It's a callee-save value
342 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
343 debug_only( dup_check[_callees[reg]]=1; )
344 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
345 if ( callee->is_concrete() || true ) {
346 omap->set_callee_saved( r, callee);
347 }
348
349 } else {
350 // Other - some reaching non-oop value
351 omap->set_value( r);
352 #ifdef ASSERT
353 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) {
354 def->dump();
355 n->dump();
356 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint");
357 }
358 #endif
359 }
360
361 }
362
363 #ifdef ASSERT
364 /* Nice, Intel-only assert
365 int cnt_callee_saves=0;
366 int reg2 = 0;
367 while (OptoReg::is_reg(reg2)) {
368 if( dup_check[reg2] != 0) cnt_callee_saves++;
369 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
370 reg2++;
371 }
372 */
373 #endif
374
375 #ifdef ASSERT
376 for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) {
377 OopMapValue omv1 = oms1.current();
378 bool found = false;
379 for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) {
380 if( omv1.content_reg() == oms2.current().reg() ) {
381 found = true;
382 break;
383 }
384 }
385 assert( found, "derived with no base in oopmap" );
386 }
387 #endif
388
389 return omap;
390 }
391
392 //------------------------------do_liveness------------------------------------
393 // Compute backwards liveness on registers
394 static void do_liveness( PhaseRegAlloc *regalloc, PhaseCFG *cfg, Block_List *worklist, int max_reg_ints, Arena *A, Dict *safehash ) {
395 int *live = NEW_ARENA_ARRAY(A, int, (cfg->_num_blocks+1) * max_reg_ints);
396 int *tmp_live = &live[cfg->_num_blocks * max_reg_ints];
397 Node *root = cfg->C->root();
398 // On CISC platforms, get the node representing the stack pointer that regalloc
399 // used for spills
400 Node *fp = NodeSentinel;
401 if (UseCISCSpill && root->req() > 1) {
402 fp = root->in(1)->in(TypeFunc::FramePtr);
403 }
404 memset( live, 0, cfg->_num_blocks * (max_reg_ints<<LogBytesPerInt) );
405 // Push preds onto worklist
406 for( uint i=1; i<root->req(); i++ )
407 worklist->push(cfg->_bbs[root->in(i)->_idx]);
408
409 // ZKM.jar includes tiny infinite loops which are unreached from below.
410 // If we missed any blocks, we'll retry here after pushing all missed
411 // blocks on the worklist. Normally this outer loop never trips more
412 // than once.
413 while( 1 ) {
414
415 while( worklist->size() ) { // Standard worklist algorithm
416 Block *b = worklist->rpop();
417
418 // Copy first successor into my tmp_live space
419 int s0num = b->_succs[0]->_pre_order;
420 int *t = &live[s0num*max_reg_ints];
421 for( int i=0; i<max_reg_ints; i++ )
422 tmp_live[i] = t[i];
423
424 // OR in the remaining live registers
425 for( uint j=1; j<b->_num_succs; j++ ) {
426 uint sjnum = b->_succs[j]->_pre_order;
427 int *t = &live[sjnum*max_reg_ints];
428 for( int i=0; i<max_reg_ints; i++ )
429 tmp_live[i] |= t[i];
430 }
431
432 // Now walk tmp_live up the block backwards, computing live
433 for( int k=b->_nodes.size()-1; k>=0; k-- ) {
434 Node *n = b->_nodes[k];
435 // KILL def'd bits
436 int first = regalloc->get_reg_first(n);
437 int second = regalloc->get_reg_second(n);
438 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
439 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
440
441 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
442
443 // Check if m is potentially a CISC alternate instruction (i.e, possibly
444 // synthesized by RegAlloc from a conventional instruction and a
445 // spilled input)
446 bool is_cisc_alternate = false;
447 if (UseCISCSpill && m) {
448 is_cisc_alternate = m->is_cisc_alternate();
449 }
450
451 // GEN use'd bits
452 for( uint l=1; l<n->req(); l++ ) {
453 Node *def = n->in(l);
454 assert(def != 0, "input edge required");
455 int first = regalloc->get_reg_first(def);
456 int second = regalloc->get_reg_second(def);
457 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first);
458 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second);
459 // If we use the stack pointer in a cisc-alternative instruction,
460 // check for use as a memory operand. Then reconstruct the RegName
461 // for this stack location, and set the appropriate bit in the
462 // live vector 4987749.
463 if (is_cisc_alternate && def == fp) {
464 const TypePtr *adr_type = NULL;
465 intptr_t offset;
466 const Node* base = m->get_base_and_disp(offset, adr_type);
467 if (base == NodeSentinel) {
468 // Machnode has multiple memory inputs. We are unable to reason
469 // with these, but are presuming (with trepidation) that not any of
470 // them are oops. This can be fixed by making get_base_and_disp()
471 // look at a specific input instead of all inputs.
472 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
473 } else if (base != fp || offset == Type::OffsetBot) {
474 // Do nothing: the fp operand is either not from a memory use
475 // (base == NULL) OR the fp is used in a non-memory context
476 // (base is some other register) OR the offset is not constant,
477 // so it is not a stack slot.
478 } else {
479 assert(offset >= 0, "unexpected negative offset");
480 offset -= (offset % jintSize); // count the whole word
481 int stack_reg = regalloc->offset2reg(offset);
482 if (OptoReg::is_stack(stack_reg)) {
483 set_live_bit(tmp_live, stack_reg);
484 } else {
485 assert(false, "stack_reg not on stack?");
486 }
487 }
488 }
489 }
490
491 if( n->jvms() ) { // Record liveness at safepoint
492
493 // This placement of this stanza means inputs to calls are
494 // considered live at the callsite's OopMap. Argument oops are
495 // hence live, but NOT included in the oopmap. See cutout in
496 // build_oop_map. Debug oops are live (and in OopMap).
497 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
498 for( int l=0; l<max_reg_ints; l++ )
499 n_live[l] = tmp_live[l];
500 safehash->Insert(n,n_live);
501 }
502
503 }
504
505 // Now at block top, see if we have any changes. If so, propagate
506 // to prior blocks.
507 int *old_live = &live[b->_pre_order*max_reg_ints];
508 int l;
509 for( l=0; l<max_reg_ints; l++ )
510 if( tmp_live[l] != old_live[l] )
511 break;
512 if( l<max_reg_ints ) { // Change!
513 // Copy in new value
514 for( l=0; l<max_reg_ints; l++ )
515 old_live[l] = tmp_live[l];
516 // Push preds onto worklist
517 for( l=1; l<(int)b->num_preds(); l++ )
518 worklist->push(cfg->_bbs[b->pred(l)->_idx]);
519 }
520 }
521
522 // Scan for any missing safepoints. Happens to infinite loops
523 // ala ZKM.jar
524 uint i;
525 for( i=1; i<cfg->_num_blocks; i++ ) {
526 Block *b = cfg->_blocks[i];
527 uint j;
528 for( j=1; j<b->_nodes.size(); j++ )
529 if( b->_nodes[j]->jvms() &&
530 (*safehash)[b->_nodes[j]] == NULL )
531 break;
532 if( j<b->_nodes.size() ) break;
533 }
534 if( i == cfg->_num_blocks )
535 break; // Got 'em all
536 #ifndef PRODUCT
537 if( PrintOpto && Verbose )
538 tty->print_cr("retripping live calc");
539 #endif
540 // Force the issue (expensively): recheck everybody
541 for( i=1; i<cfg->_num_blocks; i++ )
542 worklist->push(cfg->_blocks[i]);
543 }
544
545 }
546
547 //------------------------------BuildOopMaps-----------------------------------
548 // Collect GC mask info - where are all the OOPs?
549 void Compile::BuildOopMaps() {
550 NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); )
551 // Can't resource-mark because I need to leave all those OopMaps around,
552 // or else I need to resource-mark some arena other than the default.
553 // ResourceMark rm; // Reclaim all OopFlows when done
554 int max_reg = _regalloc->_max_reg; // Current array extent
555
556 Arena *A = Thread::current()->resource_area();
557 Block_List worklist; // Worklist of pending blocks
558
559 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt;
560 Dict *safehash = NULL; // Used for assert only
561 // Compute a backwards liveness per register. Needs a bitarray of
562 // #blocks x (#registers, rounded up to ints)
563 safehash = new Dict(cmpkey,hashkey,A);
564 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash );
565 OopFlow *free_list = NULL; // Free, unused
566
567 // Array mapping blocks to completed oopflows
568 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->_num_blocks);
569 memset( flows, 0, _cfg->_num_blocks*sizeof(OopFlow*) );
570
571
572 // Do the first block 'by hand' to prime the worklist
573 Block *entry = _cfg->_blocks[1];
574 OopFlow *rootflow = OopFlow::make(A,max_reg,this);
575 // Initialize to 'bottom' (not 'top')
576 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
577 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
578 flows[entry->_pre_order] = rootflow;
579
580 // Do the first block 'by hand' to prime the worklist
581 rootflow->_b = entry;
582 rootflow->compute_reach( _regalloc, max_reg, safehash );
583 for( uint i=0; i<entry->_num_succs; i++ )
584 worklist.push(entry->_succs[i]);
585
586 // Now worklist contains blocks which have some, but perhaps not all,
587 // predecessors visited.
588 while( worklist.size() ) {
589 // Scan for a block with all predecessors visited, or any randoms slob
590 // otherwise. All-preds-visited order allows me to recycle OopFlow
591 // structures rapidly and cut down on the memory footprint.
592 // Note: not all predecessors might be visited yet (must happen for
593 // irreducible loops). This is OK, since every live value must have the
594 // SAME reaching def for the block, so any reaching def is OK.
595 uint i;
596
597 Block *b = worklist.pop();
598 // Ignore root block
599 if( b == _cfg->_broot ) continue;
600 // Block is already done? Happens if block has several predecessors,
601 // he can get on the worklist more than once.
602 if( flows[b->_pre_order] ) continue;
603
604 // If this block has a visited predecessor AND that predecessor has this
605 // last block as his only undone child, we can move the OopFlow from the
606 // pred to this block. Otherwise we have to grab a new OopFlow.
607 OopFlow *flow = NULL; // Flag for finding optimized flow
608 Block *pred = (Block*)0xdeadbeef;
609 uint j;
610 // Scan this block's preds to find a done predecessor
611 for( j=1; j<b->num_preds(); j++ ) {
612 Block *p = _cfg->_bbs[b->pred(j)->_idx];
613 OopFlow *p_flow = flows[p->_pre_order];
614 if( p_flow ) { // Predecessor is done
615 assert( p_flow->_b == p, "cross check" );
616 pred = p; // Record some predecessor
617 // If all successors of p are done except for 'b', then we can carry
618 // p_flow forward to 'b' without copying, otherwise we have to draw
619 // from the free_list and clone data.
620 uint k;
621 for( k=0; k<p->_num_succs; k++ )
622 if( !flows[p->_succs[k]->_pre_order] &&
623 p->_succs[k] != b )
624 break;
625
626 // Either carry-forward the now-unused OopFlow for b's use
627 // or draw a new one from the free list
628 if( k==p->_num_succs ) {
629 flow = p_flow;
630 break; // Found an ideal pred, use him
631 }
632 }
633 }
634
635 if( flow ) {
636 // We have an OopFlow that's the last-use of a predecessor.
637 // Carry it forward.
638 } else { // Draw a new OopFlow from the freelist
639 if( !free_list )
640 free_list = OopFlow::make(A,max_reg,C);
641 flow = free_list;
642 assert( flow->_b == NULL, "oopFlow is not free" );
643 free_list = flow->_next;
644 flow->_next = NULL;
645
646 // Copy/clone over the data
647 flow->clone(flows[pred->_pre_order], max_reg);
648 }
649
650 // Mark flow for block. Blocks can only be flowed over once,
651 // because after the first time they are guarded from entering
652 // this code again.
653 assert( flow->_b == pred, "have some prior flow" );
654 flow->_b = NULL;
655
656 // Now push flow forward
657 flows[b->_pre_order] = flow;// Mark flow for this block
658 flow->_b = b;
659 flow->compute_reach( _regalloc, max_reg, safehash );
660
661 // Now push children onto worklist
662 for( i=0; i<b->_num_succs; i++ )
663 worklist.push(b->_succs[i]);
664
665 }
666 }