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