Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 2002, 2011, 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 */
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
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.
90 //------------------------------OopFlow----------------------------------------
91 // Structure to pass around
92 struct OopFlow : public ResourceObj {
93 short *_callees; // Array mapping register to callee-saved
94 Node **_defs; // array mapping register to reaching def
95 // or NULL if dead/conflict
96 // OopFlow structs, when not being actively modified, describe the _end_ of
97 // this block.
98 Block *_b; // Block for this struct
99 OopFlow *_next; // Next free OopFlow
100 // or NULL if dead/conflict
101 Compile* C;
103 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs),
104 _b(NULL), _next(NULL), C(c) { }
106 // Given reaching-defs for this block start, compute it for this block end
107 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash );
109 // Merge these two OopFlows into the 'this' pointer.
110 void merge( OopFlow *flow, int max_reg );
112 // Copy a 'flow' over an existing flow
113 void clone( OopFlow *flow, int max_size);
115 // Make a new OopFlow from scratch
116 static OopFlow *make( Arena *A, int max_size, Compile* C );
118 // Build an oopmap from the current flow info
119 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live );
120 };
122 //------------------------------compute_reach----------------------------------
123 // Given reaching-defs for this block start, compute it for this block end
124 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) {
126 for( uint i=0; i<_b->_nodes.size(); i++ ) {
127 Node *n = _b->_nodes[i];
129 if( n->jvms() ) { // Build an OopMap here?
130 JVMState *jvms = n->jvms();
131 // no map needed for leaf calls
132 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) {
133 int *live = (int*) (*safehash)[n];
134 assert( live, "must find live" );
135 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) );
136 }
137 }
139 // Assign new reaching def's.
140 // Note that I padded the _defs and _callees arrays so it's legal
141 // to index at _defs[OptoReg::Bad].
142 OptoReg::Name first = regalloc->get_reg_first(n);
143 OptoReg::Name second = regalloc->get_reg_second(n);
144 _defs[first] = n;
145 _defs[second] = n;
147 // Pass callee-save info around copies
148 int idx = n->is_Copy();
149 if( idx ) { // Copies move callee-save info
150 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx));
151 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx));
152 int tmp_first = _callees[old_first];
153 int tmp_second = _callees[old_second];
154 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location
155 _callees[old_second] = OptoReg::Bad;
156 _callees[first] = tmp_first;
157 _callees[second] = tmp_second;
158 } else if( n->is_Phi() ) { // Phis do not mod callee-saves
159 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" );
160 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" );
161 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" );
162 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" );
163 } else {
164 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value
165 _callees[second] = OptoReg::Bad;
167 // Find base case for callee saves
168 if( n->is_Proj() && n->in(0)->is_Start() ) {
169 if( OptoReg::is_reg(first) &&
170 regalloc->_matcher.is_save_on_entry(first) )
171 _callees[first] = first;
172 if( OptoReg::is_reg(second) &&
173 regalloc->_matcher.is_save_on_entry(second) )
174 _callees[second] = second;
175 }
176 }
177 }
178 }
180 //------------------------------merge------------------------------------------
181 // Merge the given flow into the 'this' flow
182 void OopFlow::merge( OopFlow *flow, int max_reg ) {
183 assert( _b == NULL, "merging into a happy flow" );
184 assert( flow->_b, "this flow is still alive" );
185 assert( flow != this, "no self flow" );
187 // Do the merge. If there are any differences, drop to 'bottom' which
188 // is OptoReg::Bad or NULL depending.
189 for( int i=0; i<max_reg; i++ ) {
190 // Merge the callee-save's
191 if( _callees[i] != flow->_callees[i] )
192 _callees[i] = OptoReg::Bad;
193 // Merge the reaching defs
194 if( _defs[i] != flow->_defs[i] )
195 _defs[i] = NULL;
196 }
198 }
200 //------------------------------clone------------------------------------------
201 void OopFlow::clone( OopFlow *flow, int max_size ) {
202 _b = flow->_b;
203 memcpy( _callees, flow->_callees, sizeof(short)*max_size);
204 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size);
205 }
207 //------------------------------make-------------------------------------------
208 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) {
209 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1);
210 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1);
211 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) );
212 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C);
213 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" );
214 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" );
215 return flow;
216 }
218 //------------------------------bit twiddlers----------------------------------
219 static int get_live_bit( int *live, int reg ) {
220 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
221 static void set_live_bit( int *live, int reg ) {
222 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
223 static void clr_live_bit( int *live, int reg ) {
224 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); }
226 //------------------------------build_oop_map----------------------------------
227 // Build an oopmap from the current flow info
228 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) {
229 int framesize = regalloc->_framesize;
230 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP);
231 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0());
232 memset(dup_check,0,OptoReg::stack0()) );
234 OopMap *omap = new OopMap( framesize, max_inarg_slot );
235 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL;
236 JVMState* jvms = n->jvms();
238 // For all registers do...
239 for( int reg=0; reg<max_reg; reg++ ) {
240 if( get_live_bit(live,reg) == 0 )
241 continue; // Ignore if not live
243 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit
244 // register in that case we'll get an non-concrete register for the second
245 // half. We only need to tell the map the register once!
246 //
247 // However for the moment we disable this change and leave things as they
248 // were.
250 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot);
252 if (false && r->is_reg() && !r->is_concrete()) {
253 continue;
254 }
256 // See if dead (no reaching def).
257 Node *def = _defs[reg]; // Get reaching def
258 assert( def, "since live better have reaching def" );
260 // Classify the reaching def as oop, derived, callee-save, dead, or other
261 const Type *t = def->bottom_type();
262 if( t->isa_oop_ptr() ) { // Oop or derived?
263 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
264 #ifdef _LP64
265 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
266 // Make sure both are record from the same reaching def, but do not
267 // put both into the oopmap.
268 if( (reg&1) == 1 ) { // High half of oop-pair?
269 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
270 continue; // Do not record high parts in oopmap
271 }
272 #endif
274 // Check for a legal reg name in the oopMap and bailout if it is not.
275 if (!omap->legal_vm_reg_name(r)) {
276 regalloc->C->record_method_not_compilable("illegal oopMap register name");
277 continue;
278 }
279 if( t->is_ptr()->_offset == 0 ) { // Not derived?
280 if( mcall ) {
281 // Outgoing argument GC mask responsibility belongs to the callee,
282 // not the caller. Inspect the inputs to the call, to see if
283 // this live-range is one of them.
284 uint cnt = mcall->tf()->domain()->cnt();
285 uint j;
286 for( j = TypeFunc::Parms; j < cnt; j++)
287 if( mcall->in(j) == def )
288 break; // reaching def is an argument oop
289 if( j < cnt ) // arg oops dont go in GC map
290 continue; // Continue on to the next register
291 }
292 omap->set_oop(r);
293 } else { // Else it's derived.
294 // Find the base of the derived value.
295 uint i;
296 // Fast, common case, scan
297 for( i = jvms->oopoff(); i < n->req(); i+=2 )
298 if( n->in(i) == def ) break; // Common case
299 if( i == n->req() ) { // Missed, try a more generous scan
300 // Scan again, but this time peek through copies
301 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
302 Node *m = n->in(i); // Get initial derived value
303 while( 1 ) {
304 Node *d = def; // Get initial reaching def
305 while( 1 ) { // Follow copies of reaching def to end
306 if( m == d ) goto found; // breaks 3 loops
307 int idx = d->is_Copy();
308 if( !idx ) break;
309 d = d->in(idx); // Link through copy
310 }
311 int idx = m->is_Copy();
312 if( !idx ) break;
313 m = m->in(idx);
314 }
315 }
316 guarantee( 0, "must find derived/base pair" );
317 }
318 found: ;
319 Node *base = n->in(i+1); // Base is other half of pair
320 int breg = regalloc->get_reg_first(base);
321 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
323 // I record liveness at safepoints BEFORE I make the inputs
324 // live. This is because argument oops are NOT live at a
325 // safepoint (or at least they cannot appear in the oopmap).
326 // Thus bases of base/derived pairs might not be in the
327 // liveness data but they need to appear in the oopmap.
328 if( get_live_bit(live,breg) == 0 ) {// Not live?
329 // Flag it, so next derived pointer won't re-insert into oopmap
330 set_live_bit(live,breg);
331 // Already missed our turn?
332 if( breg < reg ) {
333 if (b->is_stack() || b->is_concrete() || true ) {
334 omap->set_oop( b);
335 }
336 }
337 }
338 if (b->is_stack() || b->is_concrete() || true ) {
339 omap->set_derived_oop( r, b);
340 }
341 }
343 } else if( t->isa_narrowoop() ) {
344 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
345 // Check for a legal reg name in the oopMap and bailout if it is not.
346 if (!omap->legal_vm_reg_name(r)) {
347 regalloc->C->record_method_not_compilable("illegal oopMap register name");
348 continue;
349 }
350 if( mcall ) {
351 // Outgoing argument GC mask responsibility belongs to the callee,
352 // not the caller. Inspect the inputs to the call, to see if
353 // this live-range is one of them.
354 uint cnt = mcall->tf()->domain()->cnt();
355 uint j;
356 for( j = TypeFunc::Parms; j < cnt; j++)
357 if( mcall->in(j) == def )
358 break; // reaching def is an argument oop
359 if( j < cnt ) // arg oops dont go in GC map
360 continue; // Continue on to the next register
361 }
362 omap->set_narrowoop(r);
363 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
364 // It's a callee-save value
365 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
366 debug_only( dup_check[_callees[reg]]=1; )
367 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
368 if ( callee->is_concrete() || true ) {
369 omap->set_callee_saved( r, callee);
370 }
372 } else {
373 // Other - some reaching non-oop value
374 omap->set_value( r);
375 #ifdef ASSERT
376 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) {
377 def->dump();
378 n->dump();
379 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint");
380 }
381 #endif
382 }
384 }
386 #ifdef ASSERT
387 /* Nice, Intel-only assert
388 int cnt_callee_saves=0;
389 int reg2 = 0;
390 while (OptoReg::is_reg(reg2)) {
391 if( dup_check[reg2] != 0) cnt_callee_saves++;
392 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
393 reg2++;
394 }
395 */
396 #endif
398 #ifdef ASSERT
399 for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) {
400 OopMapValue omv1 = oms1.current();
401 bool found = false;
402 for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) {
403 if( omv1.content_reg() == oms2.current().reg() ) {
404 found = true;
405 break;
406 }
407 }
408 assert( found, "derived with no base in oopmap" );
409 }
410 #endif
412 return omap;
413 }
415 //------------------------------do_liveness------------------------------------
416 // Compute backwards liveness on registers
417 static void do_liveness( PhaseRegAlloc *regalloc, PhaseCFG *cfg, Block_List *worklist, int max_reg_ints, Arena *A, Dict *safehash ) {
418 int *live = NEW_ARENA_ARRAY(A, int, (cfg->_num_blocks+1) * max_reg_ints);
419 int *tmp_live = &live[cfg->_num_blocks * max_reg_ints];
420 Node *root = cfg->C->root();
421 // On CISC platforms, get the node representing the stack pointer that regalloc
422 // used for spills
423 Node *fp = NodeSentinel;
424 if (UseCISCSpill && root->req() > 1) {
425 fp = root->in(1)->in(TypeFunc::FramePtr);
426 }
427 memset( live, 0, cfg->_num_blocks * (max_reg_ints<<LogBytesPerInt) );
428 // Push preds onto worklist
429 for( uint i=1; i<root->req(); i++ )
430 worklist->push(cfg->_bbs[root->in(i)->_idx]);
432 // ZKM.jar includes tiny infinite loops which are unreached from below.
433 // If we missed any blocks, we'll retry here after pushing all missed
434 // blocks on the worklist. Normally this outer loop never trips more
435 // than once.
436 while( 1 ) {
438 while( worklist->size() ) { // Standard worklist algorithm
439 Block *b = worklist->rpop();
441 // Copy first successor into my tmp_live space
442 int s0num = b->_succs[0]->_pre_order;
443 int *t = &live[s0num*max_reg_ints];
444 for( int i=0; i<max_reg_ints; i++ )
445 tmp_live[i] = t[i];
447 // OR in the remaining live registers
448 for( uint j=1; j<b->_num_succs; j++ ) {
449 uint sjnum = b->_succs[j]->_pre_order;
450 int *t = &live[sjnum*max_reg_ints];
451 for( int i=0; i<max_reg_ints; i++ )
452 tmp_live[i] |= t[i];
453 }
455 // Now walk tmp_live up the block backwards, computing live
456 for( int k=b->_nodes.size()-1; k>=0; k-- ) {
457 Node *n = b->_nodes[k];
458 // KILL def'd bits
459 int first = regalloc->get_reg_first(n);
460 int second = regalloc->get_reg_second(n);
461 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
462 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
464 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
466 // Check if m is potentially a CISC alternate instruction (i.e, possibly
467 // synthesized by RegAlloc from a conventional instruction and a
468 // spilled input)
469 bool is_cisc_alternate = false;
470 if (UseCISCSpill && m) {
471 is_cisc_alternate = m->is_cisc_alternate();
472 }
474 // GEN use'd bits
475 for( uint l=1; l<n->req(); l++ ) {
476 Node *def = n->in(l);
477 assert(def != 0, "input edge required");
478 int first = regalloc->get_reg_first(def);
479 int second = regalloc->get_reg_second(def);
480 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first);
481 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second);
482 // If we use the stack pointer in a cisc-alternative instruction,
483 // check for use as a memory operand. Then reconstruct the RegName
484 // for this stack location, and set the appropriate bit in the
485 // live vector 4987749.
486 if (is_cisc_alternate && def == fp) {
487 const TypePtr *adr_type = NULL;
488 intptr_t offset;
489 const Node* base = m->get_base_and_disp(offset, adr_type);
490 if (base == NodeSentinel) {
491 // Machnode has multiple memory inputs. We are unable to reason
492 // with these, but are presuming (with trepidation) that not any of
493 // them are oops. This can be fixed by making get_base_and_disp()
494 // look at a specific input instead of all inputs.
495 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
496 } else if (base != fp || offset == Type::OffsetBot) {
497 // Do nothing: the fp operand is either not from a memory use
498 // (base == NULL) OR the fp is used in a non-memory context
499 // (base is some other register) OR the offset is not constant,
500 // so it is not a stack slot.
501 } else {
502 assert(offset >= 0, "unexpected negative offset");
503 offset -= (offset % jintSize); // count the whole word
504 int stack_reg = regalloc->offset2reg(offset);
505 if (OptoReg::is_stack(stack_reg)) {
506 set_live_bit(tmp_live, stack_reg);
507 } else {
508 assert(false, "stack_reg not on stack?");
509 }
510 }
511 }
512 }
514 if( n->jvms() ) { // Record liveness at safepoint
516 // This placement of this stanza means inputs to calls are
517 // considered live at the callsite's OopMap. Argument oops are
518 // hence live, but NOT included in the oopmap. See cutout in
519 // build_oop_map. Debug oops are live (and in OopMap).
520 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
521 for( int l=0; l<max_reg_ints; l++ )
522 n_live[l] = tmp_live[l];
523 safehash->Insert(n,n_live);
524 }
526 }
528 // Now at block top, see if we have any changes. If so, propagate
529 // to prior blocks.
530 int *old_live = &live[b->_pre_order*max_reg_ints];
531 int l;
532 for( l=0; l<max_reg_ints; l++ )
533 if( tmp_live[l] != old_live[l] )
534 break;
535 if( l<max_reg_ints ) { // Change!
536 // Copy in new value
537 for( l=0; l<max_reg_ints; l++ )
538 old_live[l] = tmp_live[l];
539 // Push preds onto worklist
540 for( l=1; l<(int)b->num_preds(); l++ )
541 worklist->push(cfg->_bbs[b->pred(l)->_idx]);
542 }
543 }
545 // Scan for any missing safepoints. Happens to infinite loops
546 // ala ZKM.jar
547 uint i;
548 for( i=1; i<cfg->_num_blocks; i++ ) {
549 Block *b = cfg->_blocks[i];
550 uint j;
551 for( j=1; j<b->_nodes.size(); j++ )
552 if( b->_nodes[j]->jvms() &&
553 (*safehash)[b->_nodes[j]] == NULL )
554 break;
555 if( j<b->_nodes.size() ) break;
556 }
557 if( i == cfg->_num_blocks )
558 break; // Got 'em all
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->_num_blocks; i++ )
565 worklist->push(cfg->_blocks[i]);
566 }
568 }
570 //------------------------------BuildOopMaps-----------------------------------
571 // Collect GC mask info - where are all the OOPs?
572 void Compile::BuildOopMaps() {
573 NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); )
574 // Can't resource-mark because I need to leave all those OopMaps around,
575 // or else I need to resource-mark some arena other than the default.
576 // ResourceMark rm; // Reclaim all OopFlows when done
577 int max_reg = _regalloc->_max_reg; // Current array extent
579 Arena *A = Thread::current()->resource_area();
580 Block_List worklist; // Worklist of pending blocks
582 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt;
583 Dict *safehash = NULL; // Used for assert only
584 // Compute a backwards liveness per register. Needs a bitarray of
585 // #blocks x (#registers, rounded up to ints)
586 safehash = new Dict(cmpkey,hashkey,A);
587 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash );
588 OopFlow *free_list = NULL; // Free, unused
590 // Array mapping blocks to completed oopflows
591 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->_num_blocks);
592 memset( flows, 0, _cfg->_num_blocks*sizeof(OopFlow*) );
595 // Do the first block 'by hand' to prime the worklist
596 Block *entry = _cfg->_blocks[1];
597 OopFlow *rootflow = OopFlow::make(A,max_reg,this);
598 // Initialize to 'bottom' (not 'top')
599 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
600 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
601 flows[entry->_pre_order] = rootflow;
603 // Do the first block 'by hand' to prime the worklist
604 rootflow->_b = entry;
605 rootflow->compute_reach( _regalloc, max_reg, safehash );
606 for( uint i=0; i<entry->_num_succs; i++ )
607 worklist.push(entry->_succs[i]);
609 // Now worklist contains blocks which have some, but perhaps not all,
610 // predecessors visited.
611 while( worklist.size() ) {
612 // Scan for a block with all predecessors visited, or any randoms slob
613 // otherwise. All-preds-visited order allows me to recycle OopFlow
614 // structures rapidly and cut down on the memory footprint.
615 // Note: not all predecessors might be visited yet (must happen for
616 // irreducible loops). This is OK, since every live value must have the
617 // SAME reaching def for the block, so any reaching def is OK.
618 uint i;
620 Block *b = worklist.pop();
621 // Ignore root block
622 if( b == _cfg->_broot ) continue;
623 // Block is already done? Happens if block has several predecessors,
624 // he can get on the worklist more than once.
625 if( flows[b->_pre_order] ) continue;
627 // If this block has a visited predecessor AND that predecessor has this
628 // last block as his only undone child, we can move the OopFlow from the
629 // pred to this block. Otherwise we have to grab a new OopFlow.
630 OopFlow *flow = NULL; // Flag for finding optimized flow
631 Block *pred = (Block*)0xdeadbeef;
632 uint j;
633 // Scan this block's preds to find a done predecessor
634 for( j=1; j<b->num_preds(); j++ ) {
635 Block *p = _cfg->_bbs[b->pred(j)->_idx];
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;
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 }
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;
669 // Copy/clone over the data
670 flow->clone(flows[pred->_pre_order], max_reg);
671 }
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;
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 );
684 // Now push children onto worklist
685 for( i=0; i<b->_num_succs; i++ )
686 worklist.push(b->_succs[i]);
688 }
689 }