1 /*
2 * Copyright (c) 2002, 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 */
24
25 #include "precompiled.hpp"
26 #include "compiler/oopMap.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/callnode.hpp"
29 #include "opto/compile.hpp"
30 #include "opto/machnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/phase.hpp"
33 #include "opto/regalloc.hpp"
34 #include "opto/rootnode.hpp"
35 #ifdef TARGET_ARCH_x86
36 # include "vmreg_x86.inline.hpp"
37 #endif
38 #ifdef TARGET_ARCH_sparc
39 # include "vmreg_sparc.inline.hpp"
40 #endif
41 #ifdef TARGET_ARCH_zero
42 # include "vmreg_zero.inline.hpp"
43 #endif
44
45 // The functions in this file builds OopMaps after all scheduling is done.
46 //
47 // OopMaps contain a list of all registers and stack-slots containing oops (so
48 // they can be updated by GC). OopMaps also contain a list of derived-pointer
49 // base-pointer pairs. When the base is moved, the derived pointer moves to
50 // follow it. Finally, any registers holding callee-save values are also
51 // recorded. These might contain oops, but only the caller knows.
52 //
53 // BuildOopMaps implements a simple forward reaching-defs solution. At each
54 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are
55 // typed as pointers (no offset), then they are oops. Pointers+offsets are
56 // derived pointers, and bases can be found from them. Finally, we'll also
57 // track reaching callee-save values. Note that a copy of a callee-save value
58 // "kills" it's source, so that only 1 copy of a callee-save value is alive at
59 // a time.
60 //
61 // We run a simple bitvector liveness pass to help trim out dead oops. Due to
62 // irreducible loops, we can have a reaching def of an oop that only reaches
63 // along one path and no way to know if it's valid or not on the other path.
64 // The bitvectors are quite dense and the liveness pass is fast.
65 //
66 // At GC points, we consult this information to build OopMaps. All reaching
67 // defs typed as oops are added to the OopMap. Only 1 instance of a
68 // callee-save register can be recorded. For derived pointers, we'll have to
69 // find and record the register holding the base.
70 //
71 // The reaching def's is a simple 1-pass worklist approach. I tried a clever
72 // breadth-first approach but it was worse (showed O(n^2) in the
73 // pick-next-block code).
74 //
75 // The relevant data is kept in a struct of arrays (it could just as well be
76 // an array of structs, but the struct-of-arrays is generally a little more
77 // efficient). The arrays are indexed by register number (including
78 // stack-slots as registers) and so is bounded by 200 to 300 elements in
79 // practice. One array will map to a reaching def Node (or NULL for
80 // conflict/dead). The other array will map to a callee-saved register or
81 // OptoReg::Bad for not-callee-saved.
82
83
84 //------------------------------OopFlow----------------------------------------
85 // Structure to pass around
86 struct OopFlow : public ResourceObj {
87 short *_callees; // Array mapping register to callee-saved
88 Node **_defs; // array mapping register to reaching def
89 // or NULL if dead/conflict
90 // OopFlow structs, when not being actively modified, describe the _end_ of
91 // this block.
92 Block *_b; // Block for this struct
93 OopFlow *_next; // Next free OopFlow
94 // or NULL if dead/conflict
95 Compile* C;
96
97 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs),
98 _b(NULL), _next(NULL), C(c) { }
99
100 // Given reaching-defs for this block start, compute it for this block end
101 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
102
103 // Merge these two OopFlows into the 'this' pointer.
104 void merge( OopFlow *flow, int max_reg );
105
106 // Copy a 'flow' over an existing flow
107 void clone( OopFlow *flow, int max_size);
108
109 // Make a new OopFlow from scratch
110 static OopFlow *make( Arena *A, int max_size, Compile* C );
111
112 // Build an oopmap from the current flow info
113 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
114 };
115
116 //------------------------------compute_reach----------------------------------
117 // Given reaching-defs for this block start, compute it for this block end
118 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
119
120 for( uint i=0; i<_b->_nodes.size(); i++ ) {
121 Node *n = _b->_nodes[i];
122
123 if( n->jvms() ) { // Build an OopMap here?
124 JVMState *jvms = n->jvms();
125 // no map needed for leaf calls
126 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
127 int *live = (int*) (*safehash)[n];
128 assert( live, "must find live" );
129 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
130 }
131 }
132
133 // Assign new reaching def's.
134 // Note that I padded the _defs and _callees arrays so it's legal
135 // to index at _defs[OptoReg::Bad].
136 OptoReg::Name first = regalloc->get_reg_first(n);
137 OptoReg::Name second = regalloc->get_reg_second(n);
138 _defs[first] = n;
139 _defs[second] = n;
140
141 // Pass callee-save info around copies
142 int idx = n->is_Copy();
143 if( idx ) { // Copies move callee-save info
144 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
145 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
146 int tmp_first = _callees[old_first];
147 int tmp_second = _callees[old_second];
148 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
149 _callees[old_second] = OptoReg::Bad;
150 _callees[first] = tmp_first;
151 _callees[second] = tmp_second;
152 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
153 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
154 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
155 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
156 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
157 } else {
158 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
159 _callees[second] = OptoReg::Bad;
160
161 // Find base case for callee saves
162 if( n->is_Proj() && n->in(0)->is_Start() ) {
163 if( OptoReg::is_reg(first) &&
164 regalloc->_matcher.is_save_on_entry(first) )
165 _callees[first] = first;
166 if( OptoReg::is_reg(second) &&
167 regalloc->_matcher.is_save_on_entry(second) )
168 _callees[second] = second;
169 }
170 }
171 }
172 }
173
174 //------------------------------merge------------------------------------------
175 // Merge the given flow into the 'this' flow
176 void OopFlow::merge( OopFlow *flow, int max_reg ) {
177 assert( _b == NULL, "merging into a happy flow" );
178 assert( flow->_b, "this flow is still alive" );
179 assert( flow != this, "no self flow" );
180
181 // Do the merge. If there are any differences, drop to 'bottom' which
182 // is OptoReg::Bad or NULL depending.
183 for( int i=0; i<max_reg; i++ ) {
184 // Merge the callee-save's
185 if( _callees[i] != flow->_callees[i] )
186 _callees[i] = OptoReg::Bad;
187 // Merge the reaching defs
188 if( _defs[i] != flow->_defs[i] )
189 _defs[i] = NULL;
190 }
191
192 }
193
194 //------------------------------clone------------------------------------------
195 void OopFlow::clone( OopFlow *flow, int max_size ) {
196 _b = flow->_b;
197 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
198 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
199 }
200
201 //------------------------------make-------------------------------------------
202 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) {
203 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
204 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
205 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
206 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C);
207 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
208 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
209 return flow;
210 }
211
212 //------------------------------bit twiddlers----------------------------------
213 static int get_live_bit( int *live, int reg ) {
214 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
215 static void set_live_bit( int *live, int reg ) {
216 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
217 static void clr_live_bit( int *live, int reg ) {
218 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
219
220 //------------------------------build_oop_map----------------------------------
221 // Build an oopmap from the current flow info
222 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
223 int framesize = regalloc->_framesize;
224 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
225 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
226 memset(dup_check,0,OptoReg::stack0()) );
227
228 OopMap *omap = new OopMap( framesize, max_inarg_slot );
229 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL;
230 JVMState* jvms = n->jvms();
231
232 // For all registers do...
233 for( int reg=0; reg<max_reg; reg++ ) {
234 if( get_live_bit(live,reg) == 0 )
235 continue; // Ignore if not live
236
237 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
238 // register in that case we'll get an non-concrete register for the second
239 // half. We only need to tell the map the register once!
240 //
241 // However for the moment we disable this change and leave things as they
242 // were.
243
244 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
245
246 if (false && r->is_reg() && !r->is_concrete()) {
247 continue;
248 }
249
250 // See if dead (no reaching def).
251 Node *def = _defs[reg]; // Get reaching def
252 assert( def, "since live better have reaching def" );
253
254 // Classify the reaching def as oop, derived, callee-save, dead, or other
255 const Type *t = def->bottom_type();
256 if( t->isa_oop_ptr() ) { // Oop or derived?
257 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
258 #ifdef _LP64
259 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
260 // Make sure both are record from the same reaching def, but do not
261 // put both into the oopmap.
262 if( (reg&1) == 1 ) { // High half of oop-pair?
263 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
264 continue; // Do not record high parts in oopmap
265 }
266 #endif
267
268 // Check for a legal reg name in the oopMap and bailout if it is not.
269 if (!omap->legal_vm_reg_name(r)) {
270 regalloc->C->record_method_not_compilable("illegal oopMap register name");
271 continue;
272 }
273 if( t->is_ptr()->_offset == 0 ) { // Not derived?
274 if( mcall ) {
275 // Outgoing argument GC mask responsibility belongs to the callee,
276 // not the caller. Inspect the inputs to the call, to see if
277 // this live-range is one of them.
278 uint cnt = mcall->tf()->domain()->cnt();
279 uint j;
280 for( j = TypeFunc::Parms; j < cnt; j++)
281 if( mcall->in(j) == def )
282 break; // reaching def is an argument oop
283 if( j < cnt ) // arg oops dont go in GC map
284 continue; // Continue on to the next register
285 }
286 omap->set_oop(r);
287 } else { // Else it's derived.
288 // Find the base of the derived value.
289 uint i;
290 // Fast, common case, scan
291 for( i = jvms->oopoff(); i < n->req(); i+=2 )
292 if( n->in(i) == def ) break; // Common case
293 if( i == n->req() ) { // Missed, try a more generous scan
294 // Scan again, but this time peek through copies
295 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
296 Node *m = n->in(i); // Get initial derived value
297 while( 1 ) {
298 Node *d = def; // Get initial reaching def
299 while( 1 ) { // Follow copies of reaching def to end
300 if( m == d ) goto found; // breaks 3 loops
301 int idx = d->is_Copy();
302 if( !idx ) break;
303 d = d->in(idx); // Link through copy
304 }
305 int idx = m->is_Copy();
306 if( !idx ) break;
307 m = m->in(idx);
308 }
309 }
310 guarantee( 0, "must find derived/base pair" );
311 }
312 found: ;
313 Node *base = n->in(i+1); // Base is other half of pair
314 int breg = regalloc->get_reg_first(base);
315 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
316
317 // I record liveness at safepoints BEFORE I make the inputs
318 // live. This is because argument oops are NOT live at a
319 // safepoint (or at least they cannot appear in the oopmap).
320 // Thus bases of base/derived pairs might not be in the
321 // liveness data but they need to appear in the oopmap.
322 if( get_live_bit(live,breg) == 0 ) {// Not live?
323 // Flag it, so next derived pointer won't re-insert into oopmap
324 set_live_bit(live,breg);
325 // Already missed our turn?
326 if( breg < reg ) {
327 if (b->is_stack() || b->is_concrete() || true ) {
328 omap->set_oop( b);
329 }
330 }
331 }
332 if (b->is_stack() || b->is_concrete() || true ) {
333 omap->set_derived_oop( r, b);
334 }
335 }
336
337 } else if( t->isa_narrowoop() ) {
338 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
339 // Check for a legal reg name in the oopMap and bailout if it is not.
340 if (!omap->legal_vm_reg_name(r)) {
341 regalloc->C->record_method_not_compilable("illegal oopMap register name");
342 continue;
343 }
344 if( mcall ) {
345 // Outgoing argument GC mask responsibility belongs to the callee,
346 // not the caller. Inspect the inputs to the call, to see if
347 // this live-range is one of them.
348 uint cnt = mcall->tf()->domain()->cnt();
349 uint j;
350 for( j = TypeFunc::Parms; j < cnt; j++)
351 if( mcall->in(j) == def )
352 break; // reaching def is an argument oop
353 if( j < cnt ) // arg oops dont go in GC map
354 continue; // Continue on to the next register
355 }
356 omap->set_narrowoop(r);
357 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
358 // It's a callee-save value
359 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
360 debug_only( dup_check[_callees[reg]]=1; )
361 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
362 if ( callee->is_concrete() || true ) {
363 omap->set_callee_saved( r, callee);
364 }
365
366 } else {
367 // Other - some reaching non-oop value
368 omap->set_value( r);
369 #ifdef ASSERT
370 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) {
371 def->dump();
372 n->dump();
373 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint");
374 }
375 #endif
376 }
377
378 }
379
380 #ifdef ASSERT
381 /* Nice, Intel-only assert
382 int cnt_callee_saves=0;
383 int reg2 = 0;
384 while (OptoReg::is_reg(reg2)) {
385 if( dup_check[reg2] != 0) cnt_callee_saves++;
386 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
387 reg2++;
388 }
389 */
390 #endif
391
392 #ifdef ASSERT
393 for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) {
394 OopMapValue omv1 = oms1.current();
395 bool found = false;
396 for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) {
397 if( omv1.content_reg() == oms2.current().reg() ) {
398 found = true;
399 break;
400 }
401 }
402 assert( found, "derived with no base in oopmap" );
403 }
404 #endif
405
406 return omap;
407 }
408
409 //------------------------------do_liveness------------------------------------
410 // Compute backwards liveness on registers
411 static void do_liveness( PhaseRegAlloc *regalloc, PhaseCFG *cfg, Block_List *worklist, int max_reg_ints, Arena *A, Dict *safehash ) {
412 int *live = NEW_ARENA_ARRAY(A, int, (cfg->_num_blocks+1) * max_reg_ints);
413 int *tmp_live = &live[cfg->_num_blocks * max_reg_ints];
414 Node *root = cfg->C->root();
415 // On CISC platforms, get the node representing the stack pointer that regalloc
416 // used for spills
417 Node *fp = NodeSentinel;
418 if (UseCISCSpill && root->req() > 1) {
419 fp = root->in(1)->in(TypeFunc::FramePtr);
420 }
421 memset( live, 0, cfg->_num_blocks * (max_reg_ints<<LogBytesPerInt) );
422 // Push preds onto worklist
423 for( uint i=1; i<root->req(); i++ )
424 worklist->push(cfg->_bbs[root->in(i)->_idx]);
425
426 // ZKM.jar includes tiny infinite loops which are unreached from below.
427 // If we missed any blocks, we'll retry here after pushing all missed
428 // blocks on the worklist. Normally this outer loop never trips more
429 // than once.
430 while( 1 ) {
431
432 while( worklist->size() ) { // Standard worklist algorithm
433 Block *b = worklist->rpop();
434
435 // Copy first successor into my tmp_live space
436 int s0num = b->_succs[0]->_pre_order;
437 int *t = &live[s0num*max_reg_ints];
438 for( int i=0; i<max_reg_ints; i++ )
439 tmp_live[i] = t[i];
440
441 // OR in the remaining live registers
442 for( uint j=1; j<b->_num_succs; j++ ) {
443 uint sjnum = b->_succs[j]->_pre_order;
444 int *t = &live[sjnum*max_reg_ints];
445 for( int i=0; i<max_reg_ints; i++ )
446 tmp_live[i] |= t[i];
447 }
448
449 // Now walk tmp_live up the block backwards, computing live
450 for( int k=b->_nodes.size()-1; k>=0; k-- ) {
451 Node *n = b->_nodes[k];
452 // KILL def'd bits
453 int first = regalloc->get_reg_first(n);
454 int second = regalloc->get_reg_second(n);
455 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
456 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
457
458 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
459
460 // Check if m is potentially a CISC alternate instruction (i.e, possibly
461 // synthesized by RegAlloc from a conventional instruction and a
462 // spilled input)
463 bool is_cisc_alternate = false;
464 if (UseCISCSpill && m) {
465 is_cisc_alternate = m->is_cisc_alternate();
466 }
467
468 // GEN use'd bits
469 for( uint l=1; l<n->req(); l++ ) {
470 Node *def = n->in(l);
471 assert(def != 0, "input edge required");
472 int first = regalloc->get_reg_first(def);
473 int second = regalloc->get_reg_second(def);
474 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first);
475 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second);
476 // If we use the stack pointer in a cisc-alternative instruction,
477 // check for use as a memory operand. Then reconstruct the RegName
478 // for this stack location, and set the appropriate bit in the
479 // live vector 4987749.
480 if (is_cisc_alternate && def == fp) {
481 const TypePtr *adr_type = NULL;
482 intptr_t offset;
483 const Node* base = m->get_base_and_disp(offset, adr_type);
484 if (base == NodeSentinel) {
485 // Machnode has multiple memory inputs. We are unable to reason
486 // with these, but are presuming (with trepidation) that not any of
487 // them are oops. This can be fixed by making get_base_and_disp()
488 // look at a specific input instead of all inputs.
489 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
490 } else if (base != fp || offset == Type::OffsetBot) {
491 // Do nothing: the fp operand is either not from a memory use
492 // (base == NULL) OR the fp is used in a non-memory context
493 // (base is some other register) OR the offset is not constant,
494 // so it is not a stack slot.
495 } else {
496 assert(offset >= 0, "unexpected negative offset");
497 offset -= (offset % jintSize); // count the whole word
498 int stack_reg = regalloc->offset2reg(offset);
499 if (OptoReg::is_stack(stack_reg)) {
500 set_live_bit(tmp_live, stack_reg);
501 } else {
502 assert(false, "stack_reg not on stack?");
503 }
504 }
505 }
506 }
507
508 if( n->jvms() ) { // Record liveness at safepoint
509
510 // This placement of this stanza means inputs to calls are
511 // considered live at the callsite's OopMap. Argument oops are
512 // hence live, but NOT included in the oopmap. See cutout in
513 // build_oop_map. Debug oops are live (and in OopMap).
514 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
515 for( int l=0; l<max_reg_ints; l++ )
516 n_live[l] = tmp_live[l];
517 safehash->Insert(n,n_live);
518 }
519
520 }
521
522 // Now at block top, see if we have any changes. If so, propagate
523 // to prior blocks.
524 int *old_live = &live[b->_pre_order*max_reg_ints];
525 int l;
526 for( l=0; l<max_reg_ints; l++ )
527 if( tmp_live[l] != old_live[l] )
528 break;
529 if( l<max_reg_ints ) { // Change!
530 // Copy in new value
531 for( l=0; l<max_reg_ints; l++ )
532 old_live[l] = tmp_live[l];
533 // Push preds onto worklist
534 for( l=1; l<(int)b->num_preds(); l++ )
535 worklist->push(cfg->_bbs[b->pred(l)->_idx]);
536 }
537 }
538
539 // Scan for any missing safepoints. Happens to infinite loops
540 // ala ZKM.jar
541 uint i;
542 for( i=1; i<cfg->_num_blocks; i++ ) {
543 Block *b = cfg->_blocks[i];
544 uint j;
545 for( j=1; j<b->_nodes.size(); j++ )
546 if( b->_nodes[j]->jvms() &&
547 (*safehash)[b->_nodes[j]] == NULL )
548 break;
549 if( j<b->_nodes.size() ) break;
550 }
551 if( i == cfg->_num_blocks )
552 break; // Got 'em all
553 #ifndef PRODUCT
554 if( PrintOpto && Verbose )
555 tty->print_cr("retripping live calc");
556 #endif
557 // Force the issue (expensively): recheck everybody
558 for( i=1; i<cfg->_num_blocks; i++ )
559 worklist->push(cfg->_blocks[i]);
560 }
561
562 }
563
564 //------------------------------BuildOopMaps-----------------------------------
565 // Collect GC mask info - where are all the OOPs?
566 void Compile::BuildOopMaps() {
567 NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); )
568 // Can't resource-mark because I need to leave all those OopMaps around,
569 // or else I need to resource-mark some arena other than the default.
570 // ResourceMark rm; // Reclaim all OopFlows when done
571 int max_reg = _regalloc->_max_reg; // Current array extent
572
573 Arena *A = Thread::current()->resource_area();
574 Block_List worklist; // Worklist of pending blocks
575
576 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt;
577 Dict *safehash = NULL; // Used for assert only
578 // Compute a backwards liveness per register. Needs a bitarray of
579 // #blocks x (#registers, rounded up to ints)
580 safehash = new Dict(cmpkey,hashkey,A);
581 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash );
582 OopFlow *free_list = NULL; // Free, unused
583
584 // Array mapping blocks to completed oopflows
585 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->_num_blocks);
586 memset( flows, 0, _cfg->_num_blocks*sizeof(OopFlow*) );
587
588
589 // Do the first block 'by hand' to prime the worklist
590 Block *entry = _cfg->_blocks[1];
591 OopFlow *rootflow = OopFlow::make(A,max_reg,this);
592 // Initialize to 'bottom' (not 'top')
593 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
594 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
595 flows[entry->_pre_order] = rootflow;
596
597 // Do the first block 'by hand' to prime the worklist
598 rootflow->_b = entry;
599 rootflow->compute_reach( _regalloc, max_reg, safehash );
600 for( uint i=0; i<entry->_num_succs; i++ )
601 worklist.push(entry->_succs[i]);
602
603 // Now worklist contains blocks which have some, but perhaps not all,
604 // predecessors visited.
605 while( worklist.size() ) {
606 // Scan for a block with all predecessors visited, or any randoms slob
607 // otherwise. All-preds-visited order allows me to recycle OopFlow
608 // structures rapidly and cut down on the memory footprint.
609 // Note: not all predecessors might be visited yet (must happen for
610 // irreducible loops). This is OK, since every live value must have the
611 // SAME reaching def for the block, so any reaching def is OK.
612 uint i;
613
614 Block *b = worklist.pop();
615 // Ignore root block
616 if( b == _cfg->_broot ) continue;
617 // Block is already done? Happens if block has several predecessors,
618 // he can get on the worklist more than once.
619 if( flows[b->_pre_order] ) continue;
620
621 // If this block has a visited predecessor AND that predecessor has this
622 // last block as his only undone child, we can move the OopFlow from the
623 // pred to this block. Otherwise we have to grab a new OopFlow.
624 OopFlow *flow = NULL; // Flag for finding optimized flow
625 Block *pred = (Block*)0xdeadbeef;
626 uint j;
627 // Scan this block's preds to find a done predecessor
628 for( j=1; j<b->num_preds(); j++ ) {
629 Block *p = _cfg->_bbs[b->pred(j)->_idx];
630 OopFlow *p_flow = flows[p->_pre_order];
631 if( p_flow ) { // Predecessor is done
632 assert( p_flow->_b == p, "cross check" );
633 pred = p; // Record some predecessor
634 // If all successors of p are done except for 'b', then we can carry
635 // p_flow forward to 'b' without copying, otherwise we have to draw
636 // from the free_list and clone data.
637 uint k;
638 for( k=0; k<p->_num_succs; k++ )
639 if( !flows[p->_succs[k]->_pre_order] &&
640 p->_succs[k] != b )
641 break;
642
643 // Either carry-forward the now-unused OopFlow for b's use
644 // or draw a new one from the free list
645 if( k==p->_num_succs ) {
646 flow = p_flow;
647 break; // Found an ideal pred, use him
648 }
649 }
650 }
651
652 if( flow ) {
653 // We have an OopFlow that's the last-use of a predecessor.
654 // Carry it forward.
655 } else { // Draw a new OopFlow from the freelist
656 if( !free_list )
657 free_list = OopFlow::make(A,max_reg,C);
658 flow = free_list;
659 assert( flow->_b == NULL, "oopFlow is not free" );
660 free_list = flow->_next;
661 flow->_next = NULL;
662
663 // Copy/clone over the data
664 flow->clone(flows[pred->_pre_order], max_reg);
665 }
666
667 // Mark flow for block. Blocks can only be flowed over once,
668 // because after the first time they are guarded from entering
669 // this code again.
670 assert( flow->_b == pred, "have some prior flow" );
671 flow->_b = NULL;
672
673 // Now push flow forward
674 flows[b->_pre_order] = flow;// Mark flow for this block
675 flow->_b = b;
676 flow->compute_reach( _regalloc, max_reg, safehash );
677
678 // Now push children onto worklist
679 for( i=0; i<b->_num_succs; i++ )
680 worklist.push(b->_succs[i]);
681
682 }
683 }