1 /*
2 * Copyright (c) 2002, 2021, 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 ResourceObj {
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(NULL), _next(NULL), 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 == NULL, "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] = NULL;
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() : NULL;
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 // Classify the reaching def as oop, derived, callee-save, dead, or other
239 const Type *t = def->bottom_type();
240 if( t->isa_oop_ptr() ) { // Oop or derived?
241 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
242 #ifdef _LP64
243 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
244 // Make sure both are record from the same reaching def, but do not
245 // put both into the oopmap.
246 if( (reg&1) == 1 ) { // High half of oop-pair?
247 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
248 continue; // Do not record high parts in oopmap
249 }
250 #endif
251
252 // Check for a legal reg name in the oopMap and bailout if it is not.
253 if (!omap->legal_vm_reg_name(r)) {
254 regalloc->C->record_method_not_compilable("illegal oopMap register name");
255 continue;
256 }
257 if (t->is_ptr()->offset() == 0) { // Not derived?
258 if( mcall ) {
259 // Outgoing argument GC mask responsibility belongs to the callee,
260 // not the caller. Inspect the inputs to the call, to see if
261 // this live-range is one of them.
262 uint cnt = mcall->tf()->domain_cc()->cnt();
263 uint j;
264 for( j = TypeFunc::Parms; j < cnt; j++)
265 if( mcall->in(j) == def )
266 break; // reaching def is an argument oop
267 if( j < cnt ) // arg oops dont go in GC map
268 continue; // Continue on to the next register
269 }
270 omap->set_oop(r);
271 } else { // Else it's derived.
272 // Find the base of the derived value.
273 uint i;
274 // Fast, common case, scan
275 for( i = jvms->oopoff(); i < n->req(); i+=2 )
276 if( n->in(i) == def ) break; // Common case
277 if( i == n->req() ) { // Missed, try a more generous scan
278 // Scan again, but this time peek through copies
279 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
280 Node *m = n->in(i); // Get initial derived value
281 while( 1 ) {
282 Node *d = def; // Get initial reaching def
283 while( 1 ) { // Follow copies of reaching def to end
284 if( m == d ) goto found; // breaks 3 loops
285 int idx = d->is_Copy();
286 if( !idx ) break;
287 d = d->in(idx); // Link through copy
288 }
289 int idx = m->is_Copy();
290 if( !idx ) break;
291 m = m->in(idx);
292 }
293 }
294 guarantee( 0, "must find derived/base pair" );
295 }
296 found: ;
297 Node *base = n->in(i+1); // Base is other half of pair
298 int breg = regalloc->get_reg_first(base);
299 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
300
301 // I record liveness at safepoints BEFORE I make the inputs
302 // live. This is because argument oops are NOT live at a
303 // safepoint (or at least they cannot appear in the oopmap).
304 // Thus bases of base/derived pairs might not be in the
305 // liveness data but they need to appear in the oopmap.
306 if( get_live_bit(live,breg) == 0 ) {// Not live?
307 // Flag it, so next derived pointer won't re-insert into oopmap
308 set_live_bit(live,breg);
309 // Already missed our turn?
310 if( breg < reg ) {
311 omap->set_oop(b);
312 }
313 }
314 omap->set_derived_oop(r, b);
315 }
316
317 } else if( t->isa_narrowoop() ) {
318 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
319 // Check for a legal reg name in the oopMap and bailout if it is not.
320 if (!omap->legal_vm_reg_name(r)) {
321 regalloc->C->record_method_not_compilable("illegal oopMap register name");
322 continue;
323 }
324 if( mcall ) {
325 // Outgoing argument GC mask responsibility belongs to the callee,
326 // not the caller. Inspect the inputs to the call, to see if
327 // this live-range is one of them.
328 uint cnt = mcall->tf()->domain_cc()->cnt();
329 uint j;
330 for( j = TypeFunc::Parms; j < cnt; j++)
331 if( mcall->in(j) == def )
332 break; // reaching def is an argument oop
333 if( j < cnt ) // arg oops dont go in GC map
334 continue; // Continue on to the next register
335 }
336 omap->set_narrowoop(r);
337 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
338 // It's a callee-save value
339 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
340 debug_only( dup_check[_callees[reg]]=1; )
341 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
342 omap->set_callee_saved(r, callee);
343
344 } else {
345 // Other - some reaching non-oop value
346 #ifdef ASSERT
347 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) {
348 def->dump();
349 n->dump();
350 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint");
351 }
352 #endif
353 }
354
355 }
356
357 #ifdef ASSERT
358 /* Nice, Intel-only assert
359 int cnt_callee_saves=0;
360 int reg2 = 0;
361 while (OptoReg::is_reg(reg2)) {
362 if( dup_check[reg2] != 0) cnt_callee_saves++;
363 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
364 reg2++;
365 }
366 */
367 #endif
368
369 #ifdef ASSERT
370 for( OopMapStream oms1(omap); !oms1.is_done(); oms1.next()) {
371 OopMapValue omv1 = oms1.current();
372 if (omv1.type() != OopMapValue::derived_oop_value) {
373 continue;
374 }
375 bool found = false;
376 for( OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) {
377 OopMapValue omv2 = oms2.current();
378 if (omv2.type() != OopMapValue::oop_value) {
379 continue;
380 }
381 if( omv1.content_reg() == omv2.reg() ) {
382 found = true;
383 break;
384 }
385 }
386 assert( found, "derived with no base in oopmap" );
387 }
388 #endif
389
390 return omap;
391 }
392
393 // Compute backwards liveness on registers
394 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) {
395 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints);
396 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints];
397 Node* root = cfg->get_root_node();
398 // On CISC platforms, get the node representing the stack pointer that regalloc
399 // used for spills
400 Node *fp = NodeSentinel;
401 if (UseCISCSpill && root->req() > 1) {
402 fp = root->in(1)->in(TypeFunc::FramePtr);
403 }
404 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt));
405 // Push preds onto worklist
406 for (uint i = 1; i < root->req(); i++) {
407 Block* block = cfg->get_block_for_node(root->in(i));
408 worklist->push(block);
409 }
410
411 // ZKM.jar includes tiny infinite loops which are unreached from below.
412 // If we missed any blocks, we'll retry here after pushing all missed
413 // blocks on the worklist. Normally this outer loop never trips more
414 // than once.
415 while (1) {
416
417 while( worklist->size() ) { // Standard worklist algorithm
418 Block *b = worklist->rpop();
419
420 // Copy first successor into my tmp_live space
421 int s0num = b->_succs[0]->_pre_order;
422 int *t = &live[s0num*max_reg_ints];
423 for( int i=0; i<max_reg_ints; i++ )
424 tmp_live[i] = t[i];
425
426 // OR in the remaining live registers
427 for( uint j=1; j<b->_num_succs; j++ ) {
428 uint sjnum = b->_succs[j]->_pre_order;
429 int *t = &live[sjnum*max_reg_ints];
430 for( int i=0; i<max_reg_ints; i++ )
431 tmp_live[i] |= t[i];
432 }
433
434 // Now walk tmp_live up the block backwards, computing live
435 for( int k=b->number_of_nodes()-1; k>=0; k-- ) {
436 Node *n = b->get_node(k);
437 // KILL def'd bits
438 int first = regalloc->get_reg_first(n);
439 int second = regalloc->get_reg_second(n);
440 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
441 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
442
443 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
444
445 // Check if m is potentially a CISC alternate instruction (i.e, possibly
446 // synthesized by RegAlloc from a conventional instruction and a
447 // spilled input)
448 bool is_cisc_alternate = false;
449 if (UseCISCSpill && m) {
450 is_cisc_alternate = m->is_cisc_alternate();
451 }
452
453 // GEN use'd bits
454 for( uint l=1; l<n->req(); l++ ) {
455 Node *def = n->in(l);
456 assert(def != 0, "input edge required");
457 int first = regalloc->get_reg_first(def);
458 int second = regalloc->get_reg_second(def);
459 //If peephole had removed the node,do not set live bit for it.
460 if (!(def->is_Mach() && def->as_Mach()->get_removed())) {
461 if (OptoReg::is_valid(first)) set_live_bit(tmp_live,first);
462 if (OptoReg::is_valid(second)) set_live_bit(tmp_live,second);
463 }
464 // If we use the stack pointer in a cisc-alternative instruction,
465 // check for use as a memory operand. Then reconstruct the RegName
466 // for this stack location, and set the appropriate bit in the
467 // live vector 4987749.
468 if (is_cisc_alternate && def == fp) {
469 const TypePtr *adr_type = NULL;
470 intptr_t offset;
471 const Node* base = m->get_base_and_disp(offset, adr_type);
472 if (base == NodeSentinel) {
473 // Machnode has multiple memory inputs. We are unable to reason
474 // with these, but are presuming (with trepidation) that not any of
475 // them are oops. This can be fixed by making get_base_and_disp()
476 // look at a specific input instead of all inputs.
477 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
478 } else if (base != fp || offset == Type::OffsetBot) {
479 // Do nothing: the fp operand is either not from a memory use
480 // (base == NULL) OR the fp is used in a non-memory context
481 // (base is some other register) OR the offset is not constant,
482 // so it is not a stack slot.
483 } else {
484 assert(offset >= 0, "unexpected negative offset");
485 offset -= (offset % jintSize); // count the whole word
486 int stack_reg = regalloc->offset2reg(offset);
487 if (OptoReg::is_stack(stack_reg)) {
488 set_live_bit(tmp_live, stack_reg);
489 } else {
490 assert(false, "stack_reg not on stack?");
491 }
492 }
493 }
494 }
495
496 if( n->jvms() ) { // Record liveness at safepoint
497
498 // This placement of this stanza means inputs to calls are
499 // considered live at the callsite's OopMap. Argument oops are
500 // hence live, but NOT included in the oopmap. See cutout in
501 // build_oop_map. Debug oops are live (and in OopMap).
502 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
503 for( int l=0; l<max_reg_ints; l++ )
504 n_live[l] = tmp_live[l];
505 safehash->Insert(n,n_live);
506 }
507
508 }
509
510 // Now at block top, see if we have any changes. If so, propagate
511 // to prior blocks.
512 int *old_live = &live[b->_pre_order*max_reg_ints];
513 int l;
514 for( l=0; l<max_reg_ints; l++ )
515 if( tmp_live[l] != old_live[l] )
516 break;
517 if( l<max_reg_ints ) { // Change!
518 // Copy in new value
519 for( l=0; l<max_reg_ints; l++ )
520 old_live[l] = tmp_live[l];
521 // Push preds onto worklist
522 for (l = 1; l < (int)b->num_preds(); l++) {
523 Block* block = cfg->get_block_for_node(b->pred(l));
524 worklist->push(block);
525 }
526 }
527 }
528
529 // Scan for any missing safepoints. Happens to infinite loops
530 // ala ZKM.jar
531 uint i;
532 for (i = 1; i < cfg->number_of_blocks(); i++) {
533 Block* block = cfg->get_block(i);
534 uint j;
535 for (j = 1; j < block->number_of_nodes(); j++) {
536 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == NULL) {
537 break;
538 }
539 }
540 if (j < block->number_of_nodes()) {
541 break;
542 }
543 }
544 if (i == cfg->number_of_blocks()) {
545 break; // Got 'em all
546 }
547
548 if (PrintOpto && Verbose) {
549 tty->print_cr("retripping live calc");
550 }
551
552 // Force the issue (expensively): recheck everybody
553 for (i = 1; i < cfg->number_of_blocks(); i++) {
554 worklist->push(cfg->get_block(i));
555 }
556 }
557 }
558
559 // Collect GC mask info - where are all the OOPs?
560 void PhaseOutput::BuildOopMaps() {
561 Compile::TracePhase tp("bldOopMaps", &timers[_t_buildOopMaps]);
562 // Can't resource-mark because I need to leave all those OopMaps around,
563 // or else I need to resource-mark some arena other than the default.
564 // ResourceMark rm; // Reclaim all OopFlows when done
565 int max_reg = C->regalloc()->_max_reg; // Current array extent
566
567 Arena *A = Thread::current()->resource_area();
568 Block_List worklist; // Worklist of pending blocks
569
570 int max_reg_ints = align_up(max_reg, BitsPerInt)>>LogBitsPerInt;
571 Dict *safehash = NULL; // Used for assert only
572 // Compute a backwards liveness per register. Needs a bitarray of
573 // #blocks x (#registers, rounded up to ints)
574 safehash = new Dict(cmpkey,hashkey,A);
575 do_liveness( C->regalloc(), C->cfg(), &worklist, max_reg_ints, A, safehash );
576 OopFlow *free_list = NULL; // Free, unused
577
578 // Array mapping blocks to completed oopflows
579 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, C->cfg()->number_of_blocks());
580 memset( flows, 0, C->cfg()->number_of_blocks() * sizeof(OopFlow*) );
581
582
583 // Do the first block 'by hand' to prime the worklist
584 Block *entry = C->cfg()->get_block(1);
585 OopFlow *rootflow = OopFlow::make(A,max_reg,C);
586 // Initialize to 'bottom' (not 'top')
587 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
588 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
589 flows[entry->_pre_order] = rootflow;
590
591 // Do the first block 'by hand' to prime the worklist
592 rootflow->_b = entry;
593 rootflow->compute_reach( C->regalloc(), max_reg, safehash );
594 for( uint i=0; i<entry->_num_succs; i++ )
595 worklist.push(entry->_succs[i]);
596
597 // Now worklist contains blocks which have some, but perhaps not all,
598 // predecessors visited.
599 while( worklist.size() ) {
600 // Scan for a block with all predecessors visited, or any randoms slob
601 // otherwise. All-preds-visited order allows me to recycle OopFlow
602 // structures rapidly and cut down on the memory footprint.
603 // Note: not all predecessors might be visited yet (must happen for
604 // irreducible loops). This is OK, since every live value must have the
605 // SAME reaching def for the block, so any reaching def is OK.
606 uint i;
607
608 Block *b = worklist.pop();
609 // Ignore root block
610 if (b == C->cfg()->get_root_block()) {
611 continue;
612 }
613 // Block is already done? Happens if block has several predecessors,
614 // he can get on the worklist more than once.
615 if( flows[b->_pre_order] ) continue;
616
617 // If this block has a visited predecessor AND that predecessor has this
618 // last block as his only undone child, we can move the OopFlow from the
619 // pred to this block. Otherwise we have to grab a new OopFlow.
620 OopFlow *flow = NULL; // Flag for finding optimized flow
621 Block *pred = (Block*)((intptr_t)0xdeadbeef);
622 // Scan this block's preds to find a done predecessor
623 for (uint j = 1; j < b->num_preds(); j++) {
624 Block* p = C->cfg()->get_block_for_node(b->pred(j));
625 OopFlow *p_flow = flows[p->_pre_order];
626 if( p_flow ) { // Predecessor is done
627 assert( p_flow->_b == p, "cross check" );
628 pred = p; // Record some predecessor
629 // If all successors of p are done except for 'b', then we can carry
630 // p_flow forward to 'b' without copying, otherwise we have to draw
631 // from the free_list and clone data.
632 uint k;
633 for( k=0; k<p->_num_succs; k++ )
634 if( !flows[p->_succs[k]->_pre_order] &&
635 p->_succs[k] != b )
636 break;
637
638 // Either carry-forward the now-unused OopFlow for b's use
639 // or draw a new one from the free list
640 if( k==p->_num_succs ) {
641 flow = p_flow;
642 break; // Found an ideal pred, use him
643 }
644 }
645 }
646
647 if( flow ) {
648 // We have an OopFlow that's the last-use of a predecessor.
649 // Carry it forward.
650 } else { // Draw a new OopFlow from the freelist
651 if( !free_list )
652 free_list = OopFlow::make(A,max_reg,C);
653 flow = free_list;
654 assert( flow->_b == NULL, "oopFlow is not free" );
655 free_list = flow->_next;
656 flow->_next = NULL;
657
658 // Copy/clone over the data
659 flow->clone(flows[pred->_pre_order], max_reg);
660 }
661
662 // Mark flow for block. Blocks can only be flowed over once,
663 // because after the first time they are guarded from entering
664 // this code again.
665 assert( flow->_b == pred, "have some prior flow" );
666 flow->_b = NULL;
667
668 // Now push flow forward
669 flows[b->_pre_order] = flow;// Mark flow for this block
670 flow->_b = b;
671 flow->compute_reach( C->regalloc(), max_reg, safehash );
672
673 // Now push children onto worklist
674 for( i=0; i<b->_num_succs; i++ )
675 worklist.push(b->_succs[i]);
676
677 }
678 }