1 /*
2 * Copyright (c) 2002, 2018, 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_aarch64
39 # include "vmreg_aarch64.inline.hpp"
40 #endif
41 #ifdef TARGET_ARCH_sparc
42 # include "vmreg_sparc.inline.hpp"
43 #endif
44 #ifdef TARGET_ARCH_zero
45 # include "vmreg_zero.inline.hpp"
46 #endif
47 #ifdef TARGET_ARCH_arm
48 # include "vmreg_arm.inline.hpp"
49 #endif
50 #ifdef TARGET_ARCH_ppc
51 # include "vmreg_ppc.inline.hpp"
52 #endif
53
54 // The functions in this file builds OopMaps after all scheduling is done.
55 //
56 // OopMaps contain a list of all registers and stack-slots containing oops (so
57 // they can be updated by GC). OopMaps also contain a list of derived-pointer
58 // base-pointer pairs. When the base is moved, the derived pointer moves to
59 // follow it. Finally, any registers holding callee-save values are also
60 // recorded. These might contain oops, but only the caller knows.
61 //
62 // BuildOopMaps implements a simple forward reaching-defs solution. At each
63 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are
64 // typed as pointers (no offset), then they are oops. Pointers+offsets are
65 // derived pointers, and bases can be found from them. Finally, we'll also
66 // track reaching callee-save values. Note that a copy of a callee-save value
67 // "kills" it's source, so that only 1 copy of a callee-save value is alive at
68 // a time.
69 //
70 // We run a simple bitvector liveness pass to help trim out dead oops. Due to
71 // irreducible loops, we can have a reaching def of an oop that only reaches
72 // along one path and no way to know if it's valid or not on the other path.
73 // The bitvectors are quite dense and the liveness pass is fast.
74 //
75 // At GC points, we consult this information to build OopMaps. All reaching
76 // defs typed as oops are added to the OopMap. Only 1 instance of a
77 // callee-save register can be recorded. For derived pointers, we'll have to
78 // find and record the register holding the base.
79 //
80 // The reaching def's is a simple 1-pass worklist approach. I tried a clever
81 // breadth-first approach but it was worse (showed O(n^2) in the
82 // pick-next-block code).
83 //
84 // The relevant data is kept in a struct of arrays (it could just as well be
85 // an array of structs, but the struct-of-arrays is generally a little more
86 // efficient). The arrays are indexed by register number (including
87 // stack-slots as registers) and so is bounded by 200 to 300 elements in
88 // practice. One array will map to a reaching def Node (or NULL for
89 // conflict/dead). The other array will map to a callee-saved register or
90 // OptoReg::Bad for not-callee-saved.
91
92
93 // Structure to pass around
94 struct OopFlow : public ResourceObj {
95 short *_callees; // Array mapping register to callee-saved
96 Node **_defs; // array mapping register to reaching def
97 // or NULL if dead/conflict
98 // OopFlow structs, when not being actively modified, describe the _end_ of
99 // this block.
100 Block *_b; // Block for this struct
101 OopFlow *_next; // Next free OopFlow
102 // or NULL if dead/conflict
103 Compile* C;
104
105 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs),
106 _b(NULL), _next(NULL), C(c) { }
107
108 // Given reaching-defs for this block start, compute it for this block end
109 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
110
111 // Merge these two OopFlows into the 'this' pointer.
112 void merge( OopFlow *flow, int max_reg );
113
114 // Copy a 'flow' over an existing flow
115 void clone( OopFlow *flow, int max_size);
116
117 // Make a new OopFlow from scratch
118 static OopFlow *make( Arena *A, int max_size, Compile* C );
119
120 // Build an oopmap from the current flow info
121 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
122 };
123
124 // Given reaching-defs for this block start, compute it for this block end
125 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
126
127 for( uint i=0; i<_b->number_of_nodes(); i++ ) {
128 Node *n = _b->get_node(i);
129
130 if( n->jvms() ) { // Build an OopMap here?
131 JVMState *jvms = n->jvms();
132 // no map needed for leaf calls
133 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
134 int *live = (int*) (*safehash)[n];
135 assert( live, "must find live" );
136 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
137 }
138 }
139
140 // Assign new reaching def's.
141 // Note that I padded the _defs and _callees arrays so it's legal
142 // to index at _defs[OptoReg::Bad].
143 OptoReg::Name first = regalloc->get_reg_first(n);
144 OptoReg::Name second = regalloc->get_reg_second(n);
145 _defs[first] = n;
146 _defs[second] = n;
147
148 // Pass callee-save info around copies
149 int idx = n->is_Copy();
150 if( idx ) { // Copies move callee-save info
151 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
152 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
153 int tmp_first = _callees[old_first];
154 int tmp_second = _callees[old_second];
155 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
156 _callees[old_second] = OptoReg::Bad;
157 _callees[first] = tmp_first;
158 _callees[second] = tmp_second;
159 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
160 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
161 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
162 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
163 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
164 } else {
165 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
166 _callees[second] = OptoReg::Bad;
167
168 // Find base case for callee saves
169 if( n->is_Proj() && n->in(0)->is_Start() ) {
170 if( OptoReg::is_reg(first) &&
171 regalloc->_matcher.is_save_on_entry(first) )
172 _callees[first] = first;
173 if( OptoReg::is_reg(second) &&
174 regalloc->_matcher.is_save_on_entry(second) )
175 _callees[second] = second;
176 }
177 }
178 }
179 }
180
181 // Merge the given flow into the 'this' flow
182 void OopFlow::merge( OopFlow *flow, int max_reg ) {
183 assert( _b == NULL, "merging into a happy flow" );
184 assert( flow->_b, "this flow is still alive" );
185 assert( flow != this, "no self flow" );
186
187 // Do the merge. If there are any differences, drop to 'bottom' which
188 // is OptoReg::Bad or NULL depending.
189 for( int i=0; i<max_reg; i++ ) {
190 // Merge the callee-save's
191 if( _callees[i] != flow->_callees[i] )
192 _callees[i] = OptoReg::Bad;
193 // Merge the reaching defs
194 if( _defs[i] != flow->_defs[i] )
195 _defs[i] = NULL;
196 }
197
198 }
199
200 void OopFlow::clone( OopFlow *flow, int max_size ) {
201 _b = flow->_b;
202 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
203 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
204 }
205
206 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) {
207 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
208 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
209 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
210 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C);
211 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
212 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
213 return flow;
214 }
215
216 static int get_live_bit( int *live, int reg ) {
217 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
218 static void set_live_bit( int *live, int reg ) {
219 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
220 static void clr_live_bit( int *live, int reg ) {
221 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
222
223 // Build an oopmap from the current flow info
224 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
225 int framesize = regalloc->_framesize;
226 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
227 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
228 memset(dup_check,0,OptoReg::stack0()) );
229
230 OopMap *omap = new OopMap( framesize, max_inarg_slot );
231 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL;
232 JVMState* jvms = n->jvms();
233
234 // For all registers do...
235 for( int reg=0; reg<max_reg; reg++ ) {
236 if( get_live_bit(live,reg) == 0 )
237 continue; // Ignore if not live
238
239 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
240 // register in that case we'll get an non-concrete register for the second
241 // half. We only need to tell the map the register once!
242 //
243 // However for the moment we disable this change and leave things as they
244 // were.
245
246 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
247
248 if (false && r->is_reg() && !r->is_concrete()) {
249 continue;
250 }
251
252 // See if dead (no reaching def).
253 Node *def = _defs[reg]; // Get reaching def
254 assert( def, "since live better have reaching def" );
255
256 // Classify the reaching def as oop, derived, callee-save, dead, or other
257 const Type *t = def->bottom_type();
258 if( t->isa_oop_ptr() ) { // Oop or derived?
259 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
260 #ifdef _LP64
261 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
262 // Make sure both are record from the same reaching def, but do not
263 // put both into the oopmap.
264 if( (reg&1) == 1 ) { // High half of oop-pair?
265 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
266 continue; // Do not record high parts in oopmap
267 }
268 #endif
269
270 // Check for a legal reg name in the oopMap and bailout if it is not.
271 if (!omap->legal_vm_reg_name(r)) {
272 regalloc->C->record_method_not_compilable("illegal oopMap register name");
273 continue;
274 }
275 if( t->is_ptr()->_offset == 0 ) { // Not derived?
276 if( mcall ) {
277 // Outgoing argument GC mask responsibility belongs to the callee,
278 // not the caller. Inspect the inputs to the call, to see if
279 // this live-range is one of them.
280 uint cnt = mcall->tf()->domain()->cnt();
281 uint j;
282 for( j = TypeFunc::Parms; j < cnt; j++)
283 if( mcall->in(j) == def )
284 break; // reaching def is an argument oop
285 if( j < cnt ) // arg oops dont go in GC map
286 continue; // Continue on to the next register
287 }
288 omap->set_oop(r);
289 } else { // Else it's derived.
290 // Find the base of the derived value.
291 uint i;
292 // Fast, common case, scan
293 for( i = jvms->oopoff(); i < n->req(); i+=2 )
294 if( n->in(i) == def ) break; // Common case
295 if( i == n->req() ) { // Missed, try a more generous scan
296 // Scan again, but this time peek through copies
297 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
298 Node *m = n->in(i); // Get initial derived value
299 while( 1 ) {
300 Node *d = def; // Get initial reaching def
301 while( 1 ) { // Follow copies of reaching def to end
302 if( m == d ) goto found; // breaks 3 loops
303 int idx = d->is_Copy();
304 if( !idx ) break;
305 d = d->in(idx); // Link through copy
306 }
307 int idx = m->is_Copy();
308 if( !idx ) break;
309 m = m->in(idx);
310 }
311 }
312 guarantee( 0, "must find derived/base pair" );
313 }
314 found: ;
315 Node *base = n->in(i+1); // Base is other half of pair
316 int breg = regalloc->get_reg_first(base);
317 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
318
319 // I record liveness at safepoints BEFORE I make the inputs
320 // live. This is because argument oops are NOT live at a
321 // safepoint (or at least they cannot appear in the oopmap).
322 // Thus bases of base/derived pairs might not be in the
323 // liveness data but they need to appear in the oopmap.
324 if( get_live_bit(live,breg) == 0 ) {// Not live?
325 // Flag it, so next derived pointer won't re-insert into oopmap
326 set_live_bit(live,breg);
327 // Already missed our turn?
328 if( breg < reg ) {
329 if (b->is_stack() || b->is_concrete() || true ) {
330 omap->set_oop( b);
331 }
332 }
333 }
334 if (b->is_stack() || b->is_concrete() || true ) {
335 omap->set_derived_oop( r, b);
336 }
337 }
338
339 } else if( t->isa_narrowoop() ) {
340 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
341 // Check for a legal reg name in the oopMap and bailout if it is not.
342 if (!omap->legal_vm_reg_name(r)) {
343 regalloc->C->record_method_not_compilable("illegal oopMap register name");
344 continue;
345 }
346 if( mcall ) {
347 // Outgoing argument GC mask responsibility belongs to the callee,
348 // not the caller. Inspect the inputs to the call, to see if
349 // this live-range is one of them.
350 uint cnt = mcall->tf()->domain()->cnt();
351 uint j;
352 for( j = TypeFunc::Parms; j < cnt; j++)
353 if( mcall->in(j) == def )
354 break; // reaching def is an argument oop
355 if( j < cnt ) // arg oops dont go in GC map
356 continue; // Continue on to the next register
357 }
358 omap->set_narrowoop(r);
359 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
360 // It's a callee-save value
361 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
362 debug_only( dup_check[_callees[reg]]=1; )
363 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
364 if ( callee->is_concrete() || true ) {
365 omap->set_callee_saved( r, callee);
366 }
367
368 } else {
369 // Other - some reaching non-oop value
370 omap->set_value( r);
371 #ifdef ASSERT
372 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) {
373 def->dump();
374 n->dump();
375 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint");
376 }
377 #endif
378 }
379
380 }
381
382 #ifdef ASSERT
383 /* Nice, Intel-only assert
384 int cnt_callee_saves=0;
385 int reg2 = 0;
386 while (OptoReg::is_reg(reg2)) {
387 if( dup_check[reg2] != 0) cnt_callee_saves++;
388 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
389 reg2++;
390 }
391 */
392 #endif
393
394 #ifdef ASSERT
395 for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) {
396 OopMapValue omv1 = oms1.current();
397 bool found = false;
398 for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) {
399 if( omv1.content_reg() == oms2.current().reg() ) {
400 found = true;
401 break;
402 }
403 }
404 assert( found, "derived with no base in oopmap" );
405 }
406 #endif
407
408 return omap;
409 }
410
411 // Compute backwards liveness on registers
412 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) {
413 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints);
414 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints];
415 Node* root = cfg->get_root_node();
416 // On CISC platforms, get the node representing the stack pointer that regalloc
417 // used for spills
418 Node *fp = NodeSentinel;
419 if (UseCISCSpill && root->req() > 1) {
420 fp = root->in(1)->in(TypeFunc::FramePtr);
421 }
422 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt));
423 // Push preds onto worklist
424 for (uint i = 1; i < root->req(); i++) {
425 Block* block = cfg->get_block_for_node(root->in(i));
426 worklist->push(block);
427 }
428
429 // ZKM.jar includes tiny infinite loops which are unreached from below.
430 // If we missed any blocks, we'll retry here after pushing all missed
431 // blocks on the worklist. Normally this outer loop never trips more
432 // than once.
433 while (1) {
434
435 while( worklist->size() ) { // Standard worklist algorithm
436 Block *b = worklist->rpop();
437
438 // Copy first successor into my tmp_live space
439 int s0num = b->_succs[0]->_pre_order;
440 int *t = &live[s0num*max_reg_ints];
441 for( int i=0; i<max_reg_ints; i++ )
442 tmp_live[i] = t[i];
443
444 // OR in the remaining live registers
445 for( uint j=1; j<b->_num_succs; j++ ) {
446 uint sjnum = b->_succs[j]->_pre_order;
447 int *t = &live[sjnum*max_reg_ints];
448 for( int i=0; i<max_reg_ints; i++ )
449 tmp_live[i] |= t[i];
450 }
451
452 // Now walk tmp_live up the block backwards, computing live
453 for( int k=b->number_of_nodes()-1; k>=0; k-- ) {
454 Node *n = b->get_node(k);
455 // KILL def'd bits
456 int first = regalloc->get_reg_first(n);
457 int second = regalloc->get_reg_second(n);
458 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
459 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
460
461 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
462
463 // Check if m is potentially a CISC alternate instruction (i.e, possibly
464 // synthesized by RegAlloc from a conventional instruction and a
465 // spilled input)
466 bool is_cisc_alternate = false;
467 if (UseCISCSpill && m) {
468 is_cisc_alternate = m->is_cisc_alternate();
469 }
470
471 // GEN use'd bits
472 for( uint l=1; l<n->req(); l++ ) {
473 Node *def = n->in(l);
474 assert(def != 0, "input edge required");
475 int first = regalloc->get_reg_first(def);
476 int second = regalloc->get_reg_second(def);
477 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first);
478 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second);
479 // If we use the stack pointer in a cisc-alternative instruction,
480 // check for use as a memory operand. Then reconstruct the RegName
481 // for this stack location, and set the appropriate bit in the
482 // live vector 4987749.
483 if (is_cisc_alternate && def == fp) {
484 const TypePtr *adr_type = NULL;
485 intptr_t offset;
486 const Node* base = m->get_base_and_disp(offset, adr_type);
487 if (base == NodeSentinel) {
488 // Machnode has multiple memory inputs. We are unable to reason
489 // with these, but are presuming (with trepidation) that not any of
490 // them are oops. This can be fixed by making get_base_and_disp()
491 // look at a specific input instead of all inputs.
492 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
493 } else if (base != fp || offset == Type::OffsetBot) {
494 // Do nothing: the fp operand is either not from a memory use
495 // (base == NULL) OR the fp is used in a non-memory context
496 // (base is some other register) OR the offset is not constant,
497 // so it is not a stack slot.
498 } else {
499 assert(offset >= 0, "unexpected negative offset");
500 offset -= (offset % jintSize); // count the whole word
501 int stack_reg = regalloc->offset2reg(offset);
502 if (OptoReg::is_stack(stack_reg)) {
503 set_live_bit(tmp_live, stack_reg);
504 } else {
505 assert(false, "stack_reg not on stack?");
506 }
507 }
508 }
509 }
510
511 if( n->jvms() ) { // Record liveness at safepoint
512
513 // This placement of this stanza means inputs to calls are
514 // considered live at the callsite's OopMap. Argument oops are
515 // hence live, but NOT included in the oopmap. See cutout in
516 // build_oop_map. Debug oops are live (and in OopMap).
517 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
518 for( int l=0; l<max_reg_ints; l++ )
519 n_live[l] = tmp_live[l];
520 safehash->Insert(n,n_live);
521 }
522
523 }
524
525 // Now at block top, see if we have any changes. If so, propagate
526 // to prior blocks.
527 int *old_live = &live[b->_pre_order*max_reg_ints];
528 int l;
529 for( l=0; l<max_reg_ints; l++ )
530 if( tmp_live[l] != old_live[l] )
531 break;
532 if( l<max_reg_ints ) { // Change!
533 // Copy in new value
534 for( l=0; l<max_reg_ints; l++ )
535 old_live[l] = tmp_live[l];
536 // Push preds onto worklist
537 for (l = 1; l < (int)b->num_preds(); l++) {
538 Block* block = cfg->get_block_for_node(b->pred(l));
539 worklist->push(block);
540 }
541 }
542 }
543
544 // Scan for any missing safepoints. Happens to infinite loops
545 // ala ZKM.jar
546 uint i;
547 for (i = 1; i < cfg->number_of_blocks(); i++) {
548 Block* block = cfg->get_block(i);
549 uint j;
550 for (j = 1; j < block->number_of_nodes(); j++) {
551 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == NULL) {
552 break;
553 }
554 }
555 if (j < block->number_of_nodes()) {
556 break;
557 }
558 }
559 if (i == cfg->number_of_blocks()) {
560 break; // Got 'em all
561 }
562 #ifndef PRODUCT
563 if( PrintOpto && Verbose )
564 tty->print_cr("retripping live calc");
565 #endif
566 // Force the issue (expensively): recheck everybody
567 for (i = 1; i < cfg->number_of_blocks(); i++) {
568 worklist->push(cfg->get_block(i));
569 }
570 }
571 }
572
573 // Collect GC mask info - where are all the OOPs?
574 void Compile::BuildOopMaps() {
575 NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); )
576 // Can't resource-mark because I need to leave all those OopMaps around,
577 // or else I need to resource-mark some arena other than the default.
578 // ResourceMark rm; // Reclaim all OopFlows when done
579 int max_reg = _regalloc->_max_reg; // Current array extent
580
581 Arena *A = Thread::current()->resource_area();
582 Block_List worklist; // Worklist of pending blocks
583
584 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt;
585 Dict *safehash = NULL; // Used for assert only
586 // Compute a backwards liveness per register. Needs a bitarray of
587 // #blocks x (#registers, rounded up to ints)
588 safehash = new Dict(cmpkey,hashkey,A);
589 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash );
590 OopFlow *free_list = NULL; // Free, unused
591
592 // Array mapping blocks to completed oopflows
593 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->number_of_blocks());
594 memset( flows, 0, _cfg->number_of_blocks() * sizeof(OopFlow*) );
595
596
597 // Do the first block 'by hand' to prime the worklist
598 Block *entry = _cfg->get_block(1);
599 OopFlow *rootflow = OopFlow::make(A,max_reg,this);
600 // Initialize to 'bottom' (not 'top')
601 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
602 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
603 flows[entry->_pre_order] = rootflow;
604
605 // Do the first block 'by hand' to prime the worklist
606 rootflow->_b = entry;
607 rootflow->compute_reach( _regalloc, max_reg, safehash );
608 for( uint i=0; i<entry->_num_succs; i++ )
609 worklist.push(entry->_succs[i]);
610
611 // Now worklist contains blocks which have some, but perhaps not all,
612 // predecessors visited.
613 while( worklist.size() ) {
614 // Scan for a block with all predecessors visited, or any randoms slob
615 // otherwise. All-preds-visited order allows me to recycle OopFlow
616 // structures rapidly and cut down on the memory footprint.
617 // Note: not all predecessors might be visited yet (must happen for
618 // irreducible loops). This is OK, since every live value must have the
619 // SAME reaching def for the block, so any reaching def is OK.
620 uint i;
621
622 Block *b = worklist.pop();
623 // Ignore root block
624 if (b == _cfg->get_root_block()) {
625 continue;
626 }
627 // Block is already done? Happens if block has several predecessors,
628 // he can get on the worklist more than once.
629 if( flows[b->_pre_order] ) continue;
630
631 // If this block has a visited predecessor AND that predecessor has this
632 // last block as his only undone child, we can move the OopFlow from the
633 // pred to this block. Otherwise we have to grab a new OopFlow.
634 OopFlow *flow = NULL; // Flag for finding optimized flow
635 Block *pred = (Block*)((intptr_t)0xdeadbeef);
636 // Scan this block's preds to find a done predecessor
637 for (uint j = 1; j < b->num_preds(); j++) {
638 Block* p = _cfg->get_block_for_node(b->pred(j));
639 OopFlow *p_flow = flows[p->_pre_order];
640 if( p_flow ) { // Predecessor is done
641 assert( p_flow->_b == p, "cross check" );
642 pred = p; // Record some predecessor
643 // If all successors of p are done except for 'b', then we can carry
644 // p_flow forward to 'b' without copying, otherwise we have to draw
645 // from the free_list and clone data.
646 uint k;
647 for( k=0; k<p->_num_succs; k++ )
648 if( !flows[p->_succs[k]->_pre_order] &&
649 p->_succs[k] != b )
650 break;
651
652 // Either carry-forward the now-unused OopFlow for b's use
653 // or draw a new one from the free list
654 if( k==p->_num_succs ) {
655 flow = p_flow;
656 break; // Found an ideal pred, use him
657 }
658 }
659 }
660
661 if( flow ) {
662 // We have an OopFlow that's the last-use of a predecessor.
663 // Carry it forward.
664 } else { // Draw a new OopFlow from the freelist
665 if( !free_list )
666 free_list = OopFlow::make(A,max_reg,C);
667 flow = free_list;
668 assert( flow->_b == NULL, "oopFlow is not free" );
669 free_list = flow->_next;
670 flow->_next = NULL;
671
672 // Copy/clone over the data
673 flow->clone(flows[pred->_pre_order], max_reg);
674 }
675
676 // Mark flow for block. Blocks can only be flowed over once,
677 // because after the first time they are guarded from entering
678 // this code again.
679 assert( flow->_b == pred, "have some prior flow" );
680 flow->_b = NULL;
681
682 // Now push flow forward
683 flows[b->_pre_order] = flow;// Mark flow for this block
684 flow->_b = b;
685 flow->compute_reach( _regalloc, max_reg, safehash );
686
687 // Now push children onto worklist
688 for( i=0; i<b->_num_succs; i++ )
689 worklist.push(b->_succs[i]);
690
691 }
692 }