1 /*
2 * Copyright (c) 2002, 2024, 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 "code/vmreg.inline.hpp"
27 #include "compiler/oopMap.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/compile.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/output.hpp"
35 #include "opto/phase.hpp"
36 #include "opto/regalloc.hpp"
37 #include "opto/rootnode.hpp"
38 #include "utilities/align.hpp"
39
40 // The functions in this file builds OopMaps after all scheduling is done.
41 //
42 // OopMaps contain a list of all registers and stack-slots containing oops (so
43 // they can be updated by GC). OopMaps also contain a list of derived-pointer
44 // base-pointer pairs. When the base is moved, the derived pointer moves to
45 // follow it. Finally, any registers holding callee-save values are also
46 // recorded. These might contain oops, but only the caller knows.
47 //
48 // BuildOopMaps implements a simple forward reaching-defs solution. At each
49 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are
50 // typed as pointers (no offset), then they are oops. Pointers+offsets are
51 // derived pointers, and bases can be found from them. Finally, we'll also
52 // track reaching callee-save values. Note that a copy of a callee-save value
53 // "kills" it's source, so that only 1 copy of a callee-save value is alive at
54 // a time.
55 //
56 // We run a simple bitvector liveness pass to help trim out dead oops. Due to
57 // irreducible loops, we can have a reaching def of an oop that only reaches
58 // along one path and no way to know if it's valid or not on the other path.
59 // The bitvectors are quite dense and the liveness pass is fast.
60 //
61 // At GC points, we consult this information to build OopMaps. All reaching
62 // defs typed as oops are added to the OopMap. Only 1 instance of a
63 // callee-save register can be recorded. For derived pointers, we'll have to
64 // find and record the register holding the base.
65 //
66 // The reaching def's is a simple 1-pass worklist approach. I tried a clever
67 // breadth-first approach but it was worse (showed O(n^2) in the
68 // pick-next-block code).
69 //
70 // The relevant data is kept in a struct of arrays (it could just as well be
71 // an array of structs, but the struct-of-arrays is generally a little more
72 // efficient). The arrays are indexed by register number (including
73 // stack-slots as registers) and so is bounded by 200 to 300 elements in
74 // practice. One array will map to a reaching def Node (or null for
75 // conflict/dead). The other array will map to a callee-saved register or
76 // OptoReg::Bad for not-callee-saved.
77
78
79 // Structure to pass around
80 struct OopFlow : public ArenaObj {
81 short *_callees; // Array mapping register to callee-saved
82 Node **_defs; // array mapping register to reaching def
83 // or null if dead/conflict
84 // OopFlow structs, when not being actively modified, describe the _end_ of
85 // this block.
86 Block *_b; // Block for this struct
87 OopFlow *_next; // Next free OopFlow
88 // or null if dead/conflict
89 Compile* C;
90
91 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs),
92 _b(nullptr), _next(nullptr), C(c) { }
93
94 // Given reaching-defs for this block start, compute it for this block end
95 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
96
97 // Merge these two OopFlows into the 'this' pointer.
98 void merge( OopFlow *flow, int max_reg );
99
100 // Copy a 'flow' over an existing flow
101 void clone( OopFlow *flow, int max_size);
102
103 // Make a new OopFlow from scratch
104 static OopFlow *make( Arena *A, int max_size, Compile* C );
105
106 // Build an oopmap from the current flow info
107 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
108 };
109
110 // Given reaching-defs for this block start, compute it for this block end
111 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
112
113 for( uint i=0; i<_b->number_of_nodes(); i++ ) {
114 Node *n = _b->get_node(i);
115
116 if( n->jvms() ) { // Build an OopMap here?
117 JVMState *jvms = n->jvms();
118 // no map needed for leaf calls
119 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
120 int *live = (int*) (*safehash)[n];
121 assert( live, "must find live" );
122 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
123 }
124 }
125
126 // Assign new reaching def's.
127 // Note that I padded the _defs and _callees arrays so it's legal
128 // to index at _defs[OptoReg::Bad].
129 OptoReg::Name first = regalloc->get_reg_first(n);
130 OptoReg::Name second = regalloc->get_reg_second(n);
131 _defs[first] = n;
132 _defs[second] = n;
133
134 // Pass callee-save info around copies
135 int idx = n->is_Copy();
136 if( idx ) { // Copies move callee-save info
137 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
138 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
139 int tmp_first = _callees[old_first];
140 int tmp_second = _callees[old_second];
141 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
142 _callees[old_second] = OptoReg::Bad;
143 _callees[first] = tmp_first;
144 _callees[second] = tmp_second;
145 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
146 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
147 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
148 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
149 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
150 } else {
151 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
152 _callees[second] = OptoReg::Bad;
153
154 // Find base case for callee saves
155 if( n->is_Proj() && n->in(0)->is_Start() ) {
156 if( OptoReg::is_reg(first) &&
157 regalloc->_matcher.is_save_on_entry(first) )
158 _callees[first] = first;
159 if( OptoReg::is_reg(second) &&
160 regalloc->_matcher.is_save_on_entry(second) )
161 _callees[second] = second;
162 }
163 }
164 }
165 }
166
167 // Merge the given flow into the 'this' flow
168 void OopFlow::merge( OopFlow *flow, int max_reg ) {
169 assert( _b == nullptr, "merging into a happy flow" );
170 assert( flow->_b, "this flow is still alive" );
171 assert( flow != this, "no self flow" );
172
173 // Do the merge. If there are any differences, drop to 'bottom' which
174 // is OptoReg::Bad or null depending.
175 for( int i=0; i<max_reg; i++ ) {
176 // Merge the callee-save's
177 if( _callees[i] != flow->_callees[i] )
178 _callees[i] = OptoReg::Bad;
179 // Merge the reaching defs
180 if( _defs[i] != flow->_defs[i] )
181 _defs[i] = nullptr;
182 }
183
184 }
185
186 void OopFlow::clone( OopFlow *flow, int max_size ) {
187 _b = flow->_b;
188 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
189 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
190 }
191
192 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) {
193 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
194 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
195 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
196 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C);
197 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
198 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
199 return flow;
200 }
201
202 static int get_live_bit( int *live, int reg ) {
203 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
204 static void set_live_bit( int *live, int reg ) {
205 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
206 static void clr_live_bit( int *live, int reg ) {
207 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
208
209 // Build an oopmap from the current flow info
210 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
211 int framesize = regalloc->_framesize;
212 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
213 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
214 memset(dup_check,0,OptoReg::stack0()) );
215
216 OopMap *omap = new OopMap( framesize, max_inarg_slot );
217 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : nullptr;
218 JVMState* jvms = n->jvms();
219
220 // For all registers do...
221 for( int reg=0; reg<max_reg; reg++ ) {
222 if( get_live_bit(live,reg) == 0 )
223 continue; // Ignore if not live
224
225 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
226 // register in that case we'll get an non-concrete register for the second
227 // half. We only need to tell the map the register once!
228 //
229 // However for the moment we disable this change and leave things as they
230 // were.
231
232 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
233
234 // See if dead (no reaching def).
235 Node *def = _defs[reg]; // Get reaching def
236 assert( def, "since live better have reaching def" );
237
238 if (def->is_MachTemp()) {
239 assert(!def->bottom_type()->isa_oop_ptr(),
240 "ADLC only assigns OOP types to MachTemp defs corresponding to xRegN operands");
241 // Exclude MachTemp definitions even if they are typed as oops.
242 continue;
243 }
244
245 // Classify the reaching def as oop, derived, callee-save, dead, or other
246 const Type *t = def->bottom_type();
247 if( t->isa_oop_ptr() ) { // Oop or derived?
248 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
249 #ifdef _LP64
250 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
251 // Make sure both are record from the same reaching def, but do not
252 // put both into the oopmap.
253 if( (reg&1) == 1 ) { // High half of oop-pair?
254 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
255 continue; // Do not record high parts in oopmap
256 }
257 #endif
258
259 // Check for a legal reg name in the oopMap and bailout if it is not.
260 if (!omap->legal_vm_reg_name(r)) {
261 stringStream ss;
262 ss.print("illegal oopMap register name: ");
263 r->print_on(&ss);
264 assert(false, "%s", ss.as_string());
265 regalloc->C->record_method_not_compilable(ss.as_string());
266 continue;
267 }
268 if (t->is_ptr()->offset() == 0) { // Not derived?
269 if( mcall ) {
270 // Outgoing argument GC mask responsibility belongs to the callee,
271 // not the caller. Inspect the inputs to the call, to see if
272 // this live-range is one of them.
273 uint cnt = mcall->tf()->domain_cc()->cnt();
274 uint j;
275 for( j = TypeFunc::Parms; j < cnt; j++)
276 if( mcall->in(j) == def )
277 break; // reaching def is an argument oop
278 if( j < cnt ) // arg oops dont go in GC map
279 continue; // Continue on to the next register
280 }
281 omap->set_oop(r);
282 } else { // Else it's derived.
283 // Find the base of the derived value.
284 uint i;
285 // Fast, common case, scan
286 for( i = jvms->oopoff(); i < n->req(); i+=2 )
287 if( n->in(i) == def ) break; // Common case
288 if( i == n->req() ) { // Missed, try a more generous scan
289 // Scan again, but this time peek through copies
290 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
291 Node *m = n->in(i); // Get initial derived value
292 while( 1 ) {
293 Node *d = def; // Get initial reaching def
294 while( 1 ) { // Follow copies of reaching def to end
295 if( m == d ) goto found; // breaks 3 loops
296 int idx = d->is_Copy();
297 if( !idx ) break;
298 d = d->in(idx); // Link through copy
299 }
300 int idx = m->is_Copy();
301 if( !idx ) break;
302 m = m->in(idx);
303 }
304 }
305 guarantee( 0, "must find derived/base pair" );
306 }
307 found: ;
308 Node *base = n->in(i+1); // Base is other half of pair
309 int breg = regalloc->get_reg_first(base);
310 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
311
312 // I record liveness at safepoints BEFORE I make the inputs
313 // live. This is because argument oops are NOT live at a
314 // safepoint (or at least they cannot appear in the oopmap).
315 // Thus bases of base/derived pairs might not be in the
316 // liveness data but they need to appear in the oopmap.
317 if( get_live_bit(live,breg) == 0 ) {// Not live?
318 // Flag it, so next derived pointer won't re-insert into oopmap
319 set_live_bit(live,breg);
320 // Already missed our turn?
321 if( breg < reg ) {
322 omap->set_oop(b);
323 }
324 }
325 omap->set_derived_oop(r, b);
326 }
327
328 } else if( t->isa_narrowoop() ) {
329 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
330 // Check for a legal reg name in the oopMap and bailout if it is not.
331 if (!omap->legal_vm_reg_name(r)) {
332 stringStream ss;
333 ss.print("illegal oopMap register name: ");
334 r->print_on(&ss);
335 assert(false, "%s", ss.as_string());
336 regalloc->C->record_method_not_compilable(ss.as_string());
337 continue;
338 }
339 if( mcall ) {
340 // Outgoing argument GC mask responsibility belongs to the callee,
341 // not the caller. Inspect the inputs to the call, to see if
342 // this live-range is one of them.
343 uint cnt = mcall->tf()->domain_cc()->cnt();
344 uint j;
345 for( j = TypeFunc::Parms; j < cnt; j++)
346 if( mcall->in(j) == def )
347 break; // reaching def is an argument oop
348 if( j < cnt ) // arg oops dont go in GC map
349 continue; // Continue on to the next register
350 }
351 omap->set_narrowoop(r);
352 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
353 // It's a callee-save value
354 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
355 debug_only( dup_check[_callees[reg]]=1; )
356 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
357 omap->set_callee_saved(r, callee);
358
359 } else {
360 // Other - some reaching non-oop value
361 #ifdef ASSERT
362 if (t->isa_rawptr()) {
363 ResourceMark rm;
364 Unique_Node_List worklist;
365 worklist.push(def);
366 for (uint i = 0; i < worklist.size(); i++) {
367 Node* m = worklist.at(i);
368 if (C->cfg()->_raw_oops.member(m)) {
369 def->dump();
370 m->dump();
371 n->dump();
372 assert(false, "there should be an oop in OopMap instead of a live raw oop at safepoint");
373 }
374 // Check users as well because def might be spilled
375 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
376 Node* u = m->fast_out(j);
377 if ((u->is_SpillCopy() && u->in(1) == m) || u->is_Phi()) {
378 worklist.push(u);
379 }
380 }
381 }
382 }
383 #endif
384 }
385
386 }
387
388 #ifdef ASSERT
389 /* Nice, Intel-only assert
390 int cnt_callee_saves=0;
391 int reg2 = 0;
392 while (OptoReg::is_reg(reg2)) {
393 if( dup_check[reg2] != 0) cnt_callee_saves++;
394 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
395 reg2++;
396 }
397 */
398 #endif
399
400 #ifdef ASSERT
401 bool has_derived_oops = false;
402 for( OopMapStream oms1(omap); !oms1.is_done(); oms1.next()) {
403 OopMapValue omv1 = oms1.current();
404 if (omv1.type() != OopMapValue::derived_oop_value) {
405 continue;
406 }
407 has_derived_oops = true;
408 bool found = false;
409 for( OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) {
410 OopMapValue omv2 = oms2.current();
411 if (omv2.type() != OopMapValue::oop_value) {
412 continue;
413 }
414 if( omv1.content_reg() == omv2.reg() ) {
415 found = true;
416 break;
417 }
418 }
419 assert(has_derived_oops == omap->has_derived_oops(), "");
420 assert( found, "derived with no base in oopmap" );
421 }
422
423 int num_oops = 0;
424 for (OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) {
425 OopMapValue omv = oms2.current();
426 if (omv.type() == OopMapValue::oop_value || omv.type() == OopMapValue::narrowoop_value) {
427 num_oops++;
428 }
429 }
430 assert(num_oops == omap->num_oops(), "num_oops: %d omap->num_oops(): %d", num_oops, omap->num_oops());
431 #endif
432
433 return omap;
434 }
435
436 // Compute backwards liveness on registers
437 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) {
438 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints);
439 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints];
440 Node* root = cfg->get_root_node();
441 // On CISC platforms, get the node representing the stack pointer that regalloc
442 // used for spills
443 Node *fp = NodeSentinel;
444 if (UseCISCSpill && root->req() > 1) {
445 fp = root->in(1)->in(TypeFunc::FramePtr);
446 }
447 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt));
448 // Push preds onto worklist
449 for (uint i = 1; i < root->req(); i++) {
450 Block* block = cfg->get_block_for_node(root->in(i));
451 worklist->push(block);
452 }
453
454 // ZKM.jar includes tiny infinite loops which are unreached from below.
455 // If we missed any blocks, we'll retry here after pushing all missed
456 // blocks on the worklist. Normally this outer loop never trips more
457 // than once.
458 while (1) {
459
460 while( worklist->size() ) { // Standard worklist algorithm
461 Block *b = worklist->rpop();
462
463 // Copy first successor into my tmp_live space
464 int s0num = b->_succs[0]->_pre_order;
465 int *t = &live[s0num*max_reg_ints];
466 for( int i=0; i<max_reg_ints; i++ )
467 tmp_live[i] = t[i];
468
469 // OR in the remaining live registers
470 for( uint j=1; j<b->_num_succs; j++ ) {
471 uint sjnum = b->_succs[j]->_pre_order;
472 int *t = &live[sjnum*max_reg_ints];
473 for( int i=0; i<max_reg_ints; i++ )
474 tmp_live[i] |= t[i];
475 }
476
477 // Now walk tmp_live up the block backwards, computing live
478 for( int k=b->number_of_nodes()-1; k>=0; k-- ) {
479 Node *n = b->get_node(k);
480 // KILL def'd bits
481 int first = regalloc->get_reg_first(n);
482 int second = regalloc->get_reg_second(n);
483 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
484 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
485
486 MachNode *m = n->is_Mach() ? n->as_Mach() : nullptr;
487
488 // Check if m is potentially a CISC alternate instruction (i.e, possibly
489 // synthesized by RegAlloc from a conventional instruction and a
490 // spilled input)
491 bool is_cisc_alternate = false;
492 if (UseCISCSpill && m) {
493 is_cisc_alternate = m->is_cisc_alternate();
494 }
495
496 // GEN use'd bits
497 for( uint l=1; l<n->req(); l++ ) {
498 Node *def = n->in(l);
499 assert(def != nullptr, "input edge required");
500 int first = regalloc->get_reg_first(def);
501 int second = regalloc->get_reg_second(def);
502 //If peephole had removed the node,do not set live bit for it.
503 if (!(def->is_Mach() && def->as_Mach()->get_removed())) {
504 if (OptoReg::is_valid(first)) set_live_bit(tmp_live,first);
505 if (OptoReg::is_valid(second)) set_live_bit(tmp_live,second);
506 }
507 // If we use the stack pointer in a cisc-alternative instruction,
508 // check for use as a memory operand. Then reconstruct the RegName
509 // for this stack location, and set the appropriate bit in the
510 // live vector 4987749.
511 if (is_cisc_alternate && def == fp) {
512 const TypePtr *adr_type = nullptr;
513 intptr_t offset;
514 const Node* base = m->get_base_and_disp(offset, adr_type);
515 if (base == NodeSentinel) {
516 // Machnode has multiple memory inputs. We are unable to reason
517 // with these, but are presuming (with trepidation) that not any of
518 // them are oops. This can be fixed by making get_base_and_disp()
519 // look at a specific input instead of all inputs.
520 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
521 } else if (base != fp || offset == Type::OffsetBot) {
522 // Do nothing: the fp operand is either not from a memory use
523 // (base == nullptr) OR the fp is used in a non-memory context
524 // (base is some other register) OR the offset is not constant,
525 // so it is not a stack slot.
526 } else {
527 assert(offset >= 0, "unexpected negative offset");
528 offset -= (offset % jintSize); // count the whole word
529 int stack_reg = regalloc->offset2reg(offset);
530 if (OptoReg::is_stack(stack_reg)) {
531 set_live_bit(tmp_live, stack_reg);
532 } else {
533 assert(false, "stack_reg not on stack?");
534 }
535 }
536 }
537 }
538
539 if( n->jvms() ) { // Record liveness at safepoint
540
541 // This placement of this stanza means inputs to calls are
542 // considered live at the callsite's OopMap. Argument oops are
543 // hence live, but NOT included in the oopmap. See cutout in
544 // build_oop_map. Debug oops are live (and in OopMap).
545 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
546 for( int l=0; l<max_reg_ints; l++ )
547 n_live[l] = tmp_live[l];
548 safehash->Insert(n,n_live);
549 }
550
551 }
552
553 // Now at block top, see if we have any changes. If so, propagate
554 // to prior blocks.
555 int *old_live = &live[b->_pre_order*max_reg_ints];
556 int l;
557 for( l=0; l<max_reg_ints; l++ )
558 if( tmp_live[l] != old_live[l] )
559 break;
560 if( l<max_reg_ints ) { // Change!
561 // Copy in new value
562 for( l=0; l<max_reg_ints; l++ )
563 old_live[l] = tmp_live[l];
564 // Push preds onto worklist
565 for (l = 1; l < (int)b->num_preds(); l++) {
566 Block* block = cfg->get_block_for_node(b->pred(l));
567 worklist->push(block);
568 }
569 }
570 }
571
572 // Scan for any missing safepoints. Happens to infinite loops
573 // ala ZKM.jar
574 uint i;
575 for (i = 1; i < cfg->number_of_blocks(); i++) {
576 Block* block = cfg->get_block(i);
577 uint j;
578 for (j = 1; j < block->number_of_nodes(); j++) {
579 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == nullptr) {
580 break;
581 }
582 }
583 if (j < block->number_of_nodes()) {
584 break;
585 }
586 }
587 if (i == cfg->number_of_blocks()) {
588 break; // Got 'em all
589 }
590
591 if (PrintOpto && Verbose) {
592 tty->print_cr("retripping live calc");
593 }
594
595 // Force the issue (expensively): recheck everybody
596 for (i = 1; i < cfg->number_of_blocks(); i++) {
597 worklist->push(cfg->get_block(i));
598 }
599 }
600 }
601
602 // Collect GC mask info - where are all the OOPs?
603 void PhaseOutput::BuildOopMaps() {
604 Compile::TracePhase tp("bldOopMaps", &timers[_t_buildOopMaps]);
605 // Can't resource-mark because I need to leave all those OopMaps around,
606 // or else I need to resource-mark some arena other than the default.
607 // ResourceMark rm; // Reclaim all OopFlows when done
608 int max_reg = C->regalloc()->_max_reg; // Current array extent
609
610 Arena *A = Thread::current()->resource_area();
611 Block_List worklist; // Worklist of pending blocks
612
613 int max_reg_ints = align_up(max_reg, BitsPerInt)>>LogBitsPerInt;
614 Dict *safehash = nullptr; // Used for assert only
615 // Compute a backwards liveness per register. Needs a bitarray of
616 // #blocks x (#registers, rounded up to ints)
617 safehash = new Dict(cmpkey,hashkey,A);
618 do_liveness( C->regalloc(), C->cfg(), &worklist, max_reg_ints, A, safehash );
619 OopFlow *free_list = nullptr; // Free, unused
620
621 // Array mapping blocks to completed oopflows
622 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, C->cfg()->number_of_blocks());
623 memset( flows, 0, C->cfg()->number_of_blocks() * sizeof(OopFlow*) );
624
625
626 // Do the first block 'by hand' to prime the worklist
627 Block *entry = C->cfg()->get_block(1);
628 OopFlow *rootflow = OopFlow::make(A,max_reg,C);
629 // Initialize to 'bottom' (not 'top')
630 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
631 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
632 flows[entry->_pre_order] = rootflow;
633
634 // Do the first block 'by hand' to prime the worklist
635 rootflow->_b = entry;
636 rootflow->compute_reach( C->regalloc(), max_reg, safehash );
637 for( uint i=0; i<entry->_num_succs; i++ )
638 worklist.push(entry->_succs[i]);
639
640 // Now worklist contains blocks which have some, but perhaps not all,
641 // predecessors visited.
642 while( worklist.size() ) {
643 // Scan for a block with all predecessors visited, or any randoms slob
644 // otherwise. All-preds-visited order allows me to recycle OopFlow
645 // structures rapidly and cut down on the memory footprint.
646 // Note: not all predecessors might be visited yet (must happen for
647 // irreducible loops). This is OK, since every live value must have the
648 // SAME reaching def for the block, so any reaching def is OK.
649 uint i;
650
651 Block *b = worklist.pop();
652 // Ignore root block
653 if (b == C->cfg()->get_root_block()) {
654 continue;
655 }
656 // Block is already done? Happens if block has several predecessors,
657 // he can get on the worklist more than once.
658 if( flows[b->_pre_order] ) continue;
659
660 // If this block has a visited predecessor AND that predecessor has this
661 // last block as his only undone child, we can move the OopFlow from the
662 // pred to this block. Otherwise we have to grab a new OopFlow.
663 OopFlow *flow = nullptr; // Flag for finding optimized flow
664 Block *pred = (Block*)((intptr_t)0xdeadbeef);
665 // Scan this block's preds to find a done predecessor
666 for (uint j = 1; j < b->num_preds(); j++) {
667 Block* p = C->cfg()->get_block_for_node(b->pred(j));
668 OopFlow *p_flow = flows[p->_pre_order];
669 if( p_flow ) { // Predecessor is done
670 assert( p_flow->_b == p, "cross check" );
671 pred = p; // Record some predecessor
672 // If all successors of p are done except for 'b', then we can carry
673 // p_flow forward to 'b' without copying, otherwise we have to draw
674 // from the free_list and clone data.
675 uint k;
676 for( k=0; k<p->_num_succs; k++ )
677 if( !flows[p->_succs[k]->_pre_order] &&
678 p->_succs[k] != b )
679 break;
680
681 // Either carry-forward the now-unused OopFlow for b's use
682 // or draw a new one from the free list
683 if( k==p->_num_succs ) {
684 flow = p_flow;
685 break; // Found an ideal pred, use him
686 }
687 }
688 }
689
690 if( flow ) {
691 // We have an OopFlow that's the last-use of a predecessor.
692 // Carry it forward.
693 } else { // Draw a new OopFlow from the freelist
694 if( !free_list )
695 free_list = OopFlow::make(A,max_reg,C);
696 flow = free_list;
697 assert( flow->_b == nullptr, "oopFlow is not free" );
698 free_list = flow->_next;
699 flow->_next = nullptr;
700
701 // Copy/clone over the data
702 flow->clone(flows[pred->_pre_order], max_reg);
703 }
704
705 // Mark flow for block. Blocks can only be flowed over once,
706 // because after the first time they are guarded from entering
707 // this code again.
708 assert( flow->_b == pred, "have some prior flow" );
709 flow->_b = nullptr;
710
711 // Now push flow forward
712 flows[b->_pre_order] = flow;// Mark flow for this block
713 flow->_b = b;
714 flow->compute_reach( C->regalloc(), max_reg, safehash );
715
716 // Now push children onto worklist
717 for( i=0; i<b->_num_succs; i++ )
718 worklist.push(b->_succs[i]);
719
720 }
721 }