1 /*
2 * Copyright (c) 2002, 2025, 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 "code/vmreg.inline.hpp"
26 #include "compiler/oopMap.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/compile.hpp"
31 #include "opto/machnode.hpp"
32 #include "opto/matcher.hpp"
33 #include "opto/output.hpp"
34 #include "opto/phase.hpp"
35 #include "opto/regalloc.hpp"
36 #include "opto/rootnode.hpp"
37 #include "utilities/align.hpp"
38
39 // The functions in this file builds OopMaps after all scheduling is done.
40 //
41 // OopMaps contain a list of all registers and stack-slots containing oops (so
42 // they can be updated by GC). OopMaps also contain a list of derived-pointer
43 // base-pointer pairs. When the base is moved, the derived pointer moves to
44 // follow it. Finally, any registers holding callee-save values are also
45 // recorded. These might contain oops, but only the caller knows.
46 //
47 // BuildOopMaps implements a simple forward reaching-defs solution. At each
48 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are
49 // typed as pointers (no offset), then they are oops. Pointers+offsets are
50 // derived pointers, and bases can be found from them. Finally, we'll also
51 // track reaching callee-save values. Note that a copy of a callee-save value
52 // "kills" it's source, so that only 1 copy of a callee-save value is alive at
53 // a time.
54 //
55 // We run a simple bitvector liveness pass to help trim out dead oops. Due to
56 // irreducible loops, we can have a reaching def of an oop that only reaches
57 // along one path and no way to know if it's valid or not on the other path.
58 // The bitvectors are quite dense and the liveness pass is fast.
59 //
60 // At GC points, we consult this information to build OopMaps. All reaching
61 // defs typed as oops are added to the OopMap. Only 1 instance of a
62 // callee-save register can be recorded. For derived pointers, we'll have to
63 // find and record the register holding the base.
64 //
65 // The reaching def's is a simple 1-pass worklist approach. I tried a clever
66 // breadth-first approach but it was worse (showed O(n^2) in the
67 // pick-next-block code).
68 //
69 // The relevant data is kept in a struct of arrays (it could just as well be
70 // an array of structs, but the struct-of-arrays is generally a little more
71 // efficient). The arrays are indexed by register number (including
72 // stack-slots as registers) and so is bounded by 200 to 300 elements in
73 // practice. One array will map to a reaching def Node (or null for
74 // conflict/dead). The other array will map to a callee-saved register or
75 // OptoReg::Bad for not-callee-saved.
76
77
78 // Structure to pass around
79 struct OopFlow : public ArenaObj {
80 short *_callees; // Array mapping register to callee-saved
81 Node **_defs; // array mapping register to reaching def
82 // or null if dead/conflict
83 // OopFlow structs, when not being actively modified, describe the _end_ of
84 // this block.
85 Block *_b; // Block for this struct
86 OopFlow *_next; // Next free OopFlow
87 // or null if dead/conflict
88 Compile* C;
89
90 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs),
91 _b(nullptr), _next(nullptr), C(c) { }
92
93 // Given reaching-defs for this block start, compute it for this block end
94 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
95
96 // Merge these two OopFlows into the 'this' pointer.
97 void merge( OopFlow *flow, int max_reg );
98
99 // Copy a 'flow' over an existing flow
100 void clone( OopFlow *flow, int max_size);
101
102 // Make a new OopFlow from scratch
103 static OopFlow *make( Arena *A, int max_size, Compile* C );
104
105 // Build an oopmap from the current flow info
106 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
107 };
108
109 // Given reaching-defs for this block start, compute it for this block end
110 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
111
112 for( uint i=0; i<_b->number_of_nodes(); i++ ) {
113 Node *n = _b->get_node(i);
114
115 if( n->jvms() ) { // Build an OopMap here?
116 JVMState *jvms = n->jvms();
117 // no map needed for leaf calls
118 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
119 int *live = (int*) (*safehash)[n];
120 assert( live, "must find live" );
121 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
122 }
123 }
124
125 // Assign new reaching def's.
126 // Note that I padded the _defs and _callees arrays so it's legal
127 // to index at _defs[OptoReg::Bad].
128 OptoReg::Name first = regalloc->get_reg_first(n);
129 OptoReg::Name second = regalloc->get_reg_second(n);
130 _defs[first] = n;
131 _defs[second] = n;
132
133 // Pass callee-save info around copies
134 int idx = n->is_Copy();
135 if( idx ) { // Copies move callee-save info
136 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
137 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
138 int tmp_first = _callees[old_first];
139 int tmp_second = _callees[old_second];
140 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
141 _callees[old_second] = OptoReg::Bad;
142 _callees[first] = tmp_first;
143 _callees[second] = tmp_second;
144 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
145 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
146 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
147 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
148 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
149 } else {
150 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
151 _callees[second] = OptoReg::Bad;
152
153 // Find base case for callee saves
154 if( n->is_Proj() && n->in(0)->is_Start() ) {
155 if( OptoReg::is_reg(first) &&
156 regalloc->_matcher.is_save_on_entry(first) )
157 _callees[first] = first;
158 if( OptoReg::is_reg(second) &&
159 regalloc->_matcher.is_save_on_entry(second) )
160 _callees[second] = second;
161 }
162 }
163 }
164 }
165
166 // Merge the given flow into the 'this' flow
167 void OopFlow::merge( OopFlow *flow, int max_reg ) {
168 assert( _b == nullptr, "merging into a happy flow" );
169 assert( flow->_b, "this flow is still alive" );
170 assert( flow != this, "no self flow" );
171
172 // Do the merge. If there are any differences, drop to 'bottom' which
173 // is OptoReg::Bad or null depending.
174 for( int i=0; i<max_reg; i++ ) {
175 // Merge the callee-save's
176 if( _callees[i] != flow->_callees[i] )
177 _callees[i] = OptoReg::Bad;
178 // Merge the reaching defs
179 if( _defs[i] != flow->_defs[i] )
180 _defs[i] = nullptr;
181 }
182
183 }
184
185 void OopFlow::clone( OopFlow *flow, int max_size ) {
186 _b = flow->_b;
187 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
188 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
189 }
190
191 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) {
192 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
193 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
194 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
195 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C);
196 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
197 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
198 return flow;
199 }
200
201 static int get_live_bit( int *live, int reg ) {
202 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
203 static void set_live_bit( int *live, int reg ) {
204 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
205 static void clr_live_bit( int *live, int reg ) {
206 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
207
208 // Build an oopmap from the current flow info
209 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
210 int framesize = regalloc->_framesize;
211 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
212 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
213 memset(dup_check,0,OptoReg::stack0()) );
214
215 OopMap *omap = new OopMap( framesize, max_inarg_slot );
216 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : nullptr;
217 JVMState* jvms = n->jvms();
218
219 // For all registers do...
220 for( int reg=0; reg<max_reg; reg++ ) {
221 if( get_live_bit(live,reg) == 0 )
222 continue; // Ignore if not live
223
224 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
225 // register in that case we'll get an non-concrete register for the second
226 // half. We only need to tell the map the register once!
227 //
228 // However for the moment we disable this change and leave things as they
229 // were.
230
231 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
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 if (def->is_MachTemp()) {
238 assert(!def->bottom_type()->isa_oop_ptr(),
239 "ADLC only assigns OOP types to MachTemp defs corresponding to xRegN operands");
240 // Exclude MachTemp definitions even if they are typed as oops.
241 continue;
242 }
243
244 // Classify the reaching def as oop, derived, callee-save, dead, or other
245 const Type *t = def->bottom_type();
246 if( t->isa_oop_ptr() ) { // Oop or derived?
247 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
248 #ifdef _LP64
249 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
250 // Make sure both are record from the same reaching def, but do not
251 // put both into the oopmap.
252 if( (reg&1) == 1 ) { // High half of oop-pair?
253 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
254 continue; // Do not record high parts in oopmap
255 }
256 #endif
257
258 // Check for a legal reg name in the oopMap and bailout if it is not.
259 if (!omap->legal_vm_reg_name(r)) {
260 stringStream ss;
261 ss.print("illegal oopMap register name: ");
262 r->print_on(&ss);
263 assert(false, "%s", ss.as_string());
264 regalloc->C->record_method_not_compilable(ss.as_string());
265 continue;
266 }
267 if( t->is_ptr()->_offset == 0 ) { // Not derived?
268 if( mcall ) {
269 // Outgoing argument GC mask responsibility belongs to the callee,
270 // not the caller. Inspect the inputs to the call, to see if
271 // this live-range is one of them.
272 uint cnt = mcall->tf()->domain()->cnt();
273 uint j;
274 for( j = TypeFunc::Parms; j < cnt; j++)
275 if( mcall->in(j) == def )
276 break; // reaching def is an argument oop
277 if( j < cnt ) // arg oops dont go in GC map
278 continue; // Continue on to the next register
279 }
280 omap->set_oop(r);
281 } else { // Else it's derived.
282 // Find the base of the derived value.
283 uint i;
284 // Fast, common case, scan
285 for( i = jvms->oopoff(); i < n->req(); i+=2 )
286 if( n->in(i) == def ) break; // Common case
287 if( i == n->req() ) { // Missed, try a more generous scan
288 // Scan again, but this time peek through copies
289 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
290 Node *m = n->in(i); // Get initial derived value
291 while( 1 ) {
292 Node *d = def; // Get initial reaching def
293 while( 1 ) { // Follow copies of reaching def to end
294 if( m == d ) goto found; // breaks 3 loops
295 int idx = d->is_Copy();
296 if( !idx ) break;
297 d = d->in(idx); // Link through copy
298 }
299 int idx = m->is_Copy();
300 if( !idx ) break;
301 m = m->in(idx);
302 }
303 }
304 guarantee( 0, "must find derived/base pair" );
305 }
306 found: ;
307 Node *base = n->in(i+1); // Base is other half of pair
308 int breg = regalloc->get_reg_first(base);
309 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
310
311 // I record liveness at safepoints BEFORE I make the inputs
312 // live. This is because argument oops are NOT live at a
313 // safepoint (or at least they cannot appear in the oopmap).
314 // Thus bases of base/derived pairs might not be in the
315 // liveness data but they need to appear in the oopmap.
316 if( get_live_bit(live,breg) == 0 ) {// Not live?
317 // Flag it, so next derived pointer won't re-insert into oopmap
318 set_live_bit(live,breg);
319 // Already missed our turn?
320 if( breg < reg ) {
321 omap->set_oop(b);
322 }
323 }
324 omap->set_derived_oop(r, b);
325 }
326
327 } else if( t->isa_narrowoop() ) {
328 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
329 // Check for a legal reg name in the oopMap and bailout if it is not.
330 if (!omap->legal_vm_reg_name(r)) {
331 stringStream ss;
332 ss.print("illegal oopMap register name: ");
333 r->print_on(&ss);
334 assert(false, "%s", ss.as_string());
335 regalloc->C->record_method_not_compilable(ss.as_string());
336 continue;
337 }
338 if( mcall ) {
339 // Outgoing argument GC mask responsibility belongs to the callee,
340 // not the caller. Inspect the inputs to the call, to see if
341 // this live-range is one of them.
342 uint cnt = mcall->tf()->domain()->cnt();
343 uint j;
344 for( j = TypeFunc::Parms; j < cnt; j++)
345 if( mcall->in(j) == def )
346 break; // reaching def is an argument oop
347 if( j < cnt ) // arg oops dont go in GC map
348 continue; // Continue on to the next register
349 }
350 omap->set_narrowoop(r);
351 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
352 // It's a callee-save value
353 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
354 debug_only( dup_check[_callees[reg]]=1; )
355 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
356 omap->set_callee_saved(r, callee);
357
358 } else {
359 // Other - some reaching non-oop value
360 #ifdef ASSERT
361 if (t->isa_rawptr()) {
362 ResourceMark rm;
363 Unique_Node_List worklist;
364 worklist.push(def);
365 for (uint i = 0; i < worklist.size(); i++) {
366 Node* m = worklist.at(i);
367 if (C->cfg()->_raw_oops.member(m)) {
368 def->dump();
369 m->dump();
370 n->dump();
371 assert(false, "there should be an oop in OopMap instead of a live raw oop at safepoint");
372 }
373 // Check users as well because def might be spilled
374 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
375 Node* u = m->fast_out(j);
376 if ((u->is_SpillCopy() && u->in(1) == m) || u->is_Phi()) {
377 worklist.push(u);
378 }
379 }
380 }
381 }
382 #endif
383 }
384
385 }
386
387 #ifdef ASSERT
388 /* Nice, Intel-only assert
389 int cnt_callee_saves=0;
390 int reg2 = 0;
391 while (OptoReg::is_reg(reg2)) {
392 if( dup_check[reg2] != 0) cnt_callee_saves++;
393 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
394 reg2++;
395 }
396 */
397 #endif
398
399 #ifdef ASSERT
400 bool has_derived_oops = false;
401 for( OopMapStream oms1(omap); !oms1.is_done(); oms1.next()) {
402 OopMapValue omv1 = oms1.current();
403 if (omv1.type() != OopMapValue::derived_oop_value) {
404 continue;
405 }
406 has_derived_oops = true;
407 bool found = false;
408 for( OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) {
409 OopMapValue omv2 = oms2.current();
410 if (omv2.type() != OopMapValue::oop_value) {
411 continue;
412 }
413 if( omv1.content_reg() == omv2.reg() ) {
414 found = true;
415 break;
416 }
417 }
418 assert(has_derived_oops == omap->has_derived_oops(), "");
419 assert( found, "derived with no base in oopmap" );
420 }
421
422 int num_oops = 0;
423 for (OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) {
424 OopMapValue omv = oms2.current();
425 if (omv.type() == OopMapValue::oop_value || omv.type() == OopMapValue::narrowoop_value) {
426 num_oops++;
427 }
428 }
429 assert(num_oops == omap->num_oops(), "num_oops: %d omap->num_oops(): %d", num_oops, omap->num_oops());
430 #endif
431
432 return omap;
433 }
434
435 // Compute backwards liveness on registers
436 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) {
437 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints);
438 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints];
439 Node* root = cfg->get_root_node();
440 // On CISC platforms, get the node representing the stack pointer that regalloc
441 // used for spills
442 Node *fp = NodeSentinel;
443 if (UseCISCSpill && root->req() > 1) {
444 fp = root->in(1)->in(TypeFunc::FramePtr);
445 }
446 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt));
447 // Push preds onto worklist
448 for (uint i = 1; i < root->req(); i++) {
449 Block* block = cfg->get_block_for_node(root->in(i));
450 worklist->push(block);
451 }
452
453 // ZKM.jar includes tiny infinite loops which are unreached from below.
454 // If we missed any blocks, we'll retry here after pushing all missed
455 // blocks on the worklist. Normally this outer loop never trips more
456 // than once.
457 while (1) {
458
459 while( worklist->size() ) { // Standard worklist algorithm
460 Block *b = worklist->rpop();
461
462 // Copy first successor into my tmp_live space
463 int s0num = b->_succs[0]->_pre_order;
464 int *t = &live[s0num*max_reg_ints];
465 for( int i=0; i<max_reg_ints; i++ )
466 tmp_live[i] = t[i];
467
468 // OR in the remaining live registers
469 for( uint j=1; j<b->_num_succs; j++ ) {
470 uint sjnum = b->_succs[j]->_pre_order;
471 int *t = &live[sjnum*max_reg_ints];
472 for( int i=0; i<max_reg_ints; i++ )
473 tmp_live[i] |= t[i];
474 }
475
476 // Now walk tmp_live up the block backwards, computing live
477 for( int k=b->number_of_nodes()-1; k>=0; k-- ) {
478 Node *n = b->get_node(k);
479 // KILL def'd bits
480 int first = regalloc->get_reg_first(n);
481 int second = regalloc->get_reg_second(n);
482 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
483 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
484
485 MachNode *m = n->is_Mach() ? n->as_Mach() : nullptr;
486
487 // Check if m is potentially a CISC alternate instruction (i.e, possibly
488 // synthesized by RegAlloc from a conventional instruction and a
489 // spilled input)
490 bool is_cisc_alternate = false;
491 if (UseCISCSpill && m) {
492 is_cisc_alternate = m->is_cisc_alternate();
493 }
494
495 // GEN use'd bits
496 for( uint l=1; l<n->req(); l++ ) {
497 Node *def = n->in(l);
498 assert(def != nullptr, "input edge required");
499 int first = regalloc->get_reg_first(def);
500 int second = regalloc->get_reg_second(def);
501 //If peephole had removed the node,do not set live bit for it.
502 if (!(def->is_Mach() && def->as_Mach()->get_removed())) {
503 if (OptoReg::is_valid(first)) set_live_bit(tmp_live,first);
504 if (OptoReg::is_valid(second)) set_live_bit(tmp_live,second);
505 }
506 // If we use the stack pointer in a cisc-alternative instruction,
507 // check for use as a memory operand. Then reconstruct the RegName
508 // for this stack location, and set the appropriate bit in the
509 // live vector 4987749.
510 if (is_cisc_alternate && def == fp) {
511 const TypePtr *adr_type = nullptr;
512 intptr_t offset;
513 const Node* base = m->get_base_and_disp(offset, adr_type);
514 if (base == NodeSentinel) {
515 // Machnode has multiple memory inputs. We are unable to reason
516 // with these, but are presuming (with trepidation) that not any of
517 // them are oops. This can be fixed by making get_base_and_disp()
518 // look at a specific input instead of all inputs.
519 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
520 } else if (base != fp || offset == Type::OffsetBot) {
521 // Do nothing: the fp operand is either not from a memory use
522 // (base == nullptr) OR the fp is used in a non-memory context
523 // (base is some other register) OR the offset is not constant,
524 // so it is not a stack slot.
525 } else {
526 assert(offset >= 0, "unexpected negative offset");
527 offset -= (offset % jintSize); // count the whole word
528 int stack_reg = regalloc->offset2reg(offset);
529 if (OptoReg::is_stack(stack_reg)) {
530 set_live_bit(tmp_live, stack_reg);
531 } else {
532 assert(false, "stack_reg not on stack?");
533 }
534 }
535 }
536 }
537
538 if( n->jvms() ) { // Record liveness at safepoint
539
540 // This placement of this stanza means inputs to calls are
541 // considered live at the callsite's OopMap. Argument oops are
542 // hence live, but NOT included in the oopmap. See cutout in
543 // build_oop_map. Debug oops are live (and in OopMap).
544 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
545 for( int l=0; l<max_reg_ints; l++ )
546 n_live[l] = tmp_live[l];
547 safehash->Insert(n,n_live);
548 }
549
550 }
551
552 // Now at block top, see if we have any changes. If so, propagate
553 // to prior blocks.
554 int *old_live = &live[b->_pre_order*max_reg_ints];
555 int l;
556 for( l=0; l<max_reg_ints; l++ )
557 if( tmp_live[l] != old_live[l] )
558 break;
559 if( l<max_reg_ints ) { // Change!
560 // Copy in new value
561 for( l=0; l<max_reg_ints; l++ )
562 old_live[l] = tmp_live[l];
563 // Push preds onto worklist
564 for (l = 1; l < (int)b->num_preds(); l++) {
565 Block* block = cfg->get_block_for_node(b->pred(l));
566 worklist->push(block);
567 }
568 }
569 }
570
571 // Scan for any missing safepoints. Happens to infinite loops
572 // ala ZKM.jar
573 uint i;
574 for (i = 1; i < cfg->number_of_blocks(); i++) {
575 Block* block = cfg->get_block(i);
576 uint j;
577 for (j = 1; j < block->number_of_nodes(); j++) {
578 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == nullptr) {
579 break;
580 }
581 }
582 if (j < block->number_of_nodes()) {
583 break;
584 }
585 }
586 if (i == cfg->number_of_blocks()) {
587 break; // Got 'em all
588 }
589
590 if (PrintOpto && Verbose) {
591 tty->print_cr("retripping live calc");
592 }
593
594 // Force the issue (expensively): recheck everybody
595 for (i = 1; i < cfg->number_of_blocks(); i++) {
596 worklist->push(cfg->get_block(i));
597 }
598 }
599 }
600
601 // Collect GC mask info - where are all the OOPs?
602 void PhaseOutput::BuildOopMaps() {
603 Compile::TracePhase tp(_t_buildOopMaps);
604 // Can't resource-mark because I need to leave all those OopMaps around,
605 // or else I need to resource-mark some arena other than the default.
606 // ResourceMark rm; // Reclaim all OopFlows when done
607 int max_reg = C->regalloc()->_max_reg; // Current array extent
608
609 Arena *A = Thread::current()->resource_area();
610 Block_List worklist; // Worklist of pending blocks
611
612 int max_reg_ints = align_up(max_reg, BitsPerInt)>>LogBitsPerInt;
613 Dict *safehash = nullptr; // Used for assert only
614 // Compute a backwards liveness per register. Needs a bitarray of
615 // #blocks x (#registers, rounded up to ints)
616 safehash = new Dict(cmpkey,hashkey,A);
617 do_liveness( C->regalloc(), C->cfg(), &worklist, max_reg_ints, A, safehash );
618 OopFlow *free_list = nullptr; // Free, unused
619
620 // Array mapping blocks to completed oopflows
621 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, C->cfg()->number_of_blocks());
622 memset( flows, 0, C->cfg()->number_of_blocks() * sizeof(OopFlow*) );
623
624
625 // Do the first block 'by hand' to prime the worklist
626 Block *entry = C->cfg()->get_block(1);
627 OopFlow *rootflow = OopFlow::make(A,max_reg,C);
628 // Initialize to 'bottom' (not 'top')
629 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
630 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
631 flows[entry->_pre_order] = rootflow;
632
633 // Do the first block 'by hand' to prime the worklist
634 rootflow->_b = entry;
635 rootflow->compute_reach( C->regalloc(), max_reg, safehash );
636 for( uint i=0; i<entry->_num_succs; i++ )
637 worklist.push(entry->_succs[i]);
638
639 // Now worklist contains blocks which have some, but perhaps not all,
640 // predecessors visited.
641 while( worklist.size() ) {
642 // Scan for a block with all predecessors visited, or any randoms slob
643 // otherwise. All-preds-visited order allows me to recycle OopFlow
644 // structures rapidly and cut down on the memory footprint.
645 // Note: not all predecessors might be visited yet (must happen for
646 // irreducible loops). This is OK, since every live value must have the
647 // SAME reaching def for the block, so any reaching def is OK.
648 uint i;
649
650 Block *b = worklist.pop();
651 // Ignore root block
652 if (b == C->cfg()->get_root_block()) {
653 continue;
654 }
655 // Block is already done? Happens if block has several predecessors,
656 // he can get on the worklist more than once.
657 if( flows[b->_pre_order] ) continue;
658
659 // If this block has a visited predecessor AND that predecessor has this
660 // last block as his only undone child, we can move the OopFlow from the
661 // pred to this block. Otherwise we have to grab a new OopFlow.
662 OopFlow *flow = nullptr; // Flag for finding optimized flow
663 Block *pred = (Block*)((intptr_t)0xdeadbeef);
664 // Scan this block's preds to find a done predecessor
665 for (uint j = 1; j < b->num_preds(); j++) {
666 Block* p = C->cfg()->get_block_for_node(b->pred(j));
667 OopFlow *p_flow = flows[p->_pre_order];
668 if( p_flow ) { // Predecessor is done
669 assert( p_flow->_b == p, "cross check" );
670 pred = p; // Record some predecessor
671 // If all successors of p are done except for 'b', then we can carry
672 // p_flow forward to 'b' without copying, otherwise we have to draw
673 // from the free_list and clone data.
674 uint k;
675 for( k=0; k<p->_num_succs; k++ )
676 if( !flows[p->_succs[k]->_pre_order] &&
677 p->_succs[k] != b )
678 break;
679
680 // Either carry-forward the now-unused OopFlow for b's use
681 // or draw a new one from the free list
682 if( k==p->_num_succs ) {
683 flow = p_flow;
684 break; // Found an ideal pred, use him
685 }
686 }
687 }
688
689 if( flow ) {
690 // We have an OopFlow that's the last-use of a predecessor.
691 // Carry it forward.
692 } else { // Draw a new OopFlow from the freelist
693 if( !free_list )
694 free_list = OopFlow::make(A,max_reg,C);
695 flow = free_list;
696 assert( flow->_b == nullptr, "oopFlow is not free" );
697 free_list = flow->_next;
698 flow->_next = nullptr;
699
700 // Copy/clone over the data
701 flow->clone(flows[pred->_pre_order], max_reg);
702 }
703
704 // Mark flow for block. Blocks can only be flowed over once,
705 // because after the first time they are guarded from entering
706 // this code again.
707 assert( flow->_b == pred, "have some prior flow" );
708 flow->_b = nullptr;
709
710 // Now push flow forward
711 flows[b->_pre_order] = flow;// Mark flow for this block
712 flow->_b = b;
713 flow->compute_reach( C->regalloc(), max_reg, safehash );
714
715 // Now push children onto worklist
716 for( i=0; i<b->_num_succs; i++ )
717 worklist.push(b->_succs[i]);
718
719 }
720 }