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 "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
OopFlowOopFlow91 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
compute_reach(PhaseRegAlloc * regalloc,int max_reg,Dict * safehash)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
merge(OopFlow * flow,int max_reg)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
clone(OopFlow * flow,int max_size)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
make(Arena * A,int max_size,Compile * C)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
get_live_bit(int * live,int reg)202 static int get_live_bit( int *live, int reg ) {
203 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); }
set_live_bit(int * live,int reg)204 static void set_live_bit( int *live, int reg ) {
205 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); }
clr_live_bit(int * live,int reg)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
build_oop_map(Node * n,int max_reg,PhaseRegAlloc * regalloc,int * live)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 if (false && r->is_reg() && !r->is_concrete()) {
235 continue;
236 }
237
238 // See if dead (no reaching def).
239 Node *def = _defs[reg]; // Get reaching def
240 assert( def, "since live better have reaching def" );
241
242 // Classify the reaching def as oop, derived, callee-save, dead, or other
243 const Type *t = def->bottom_type();
244 if( t->isa_oop_ptr() ) { // Oop or derived?
245 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
246 #ifdef _LP64
247 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers.
248 // Make sure both are record from the same reaching def, but do not
249 // put both into the oopmap.
250 if( (reg&1) == 1 ) { // High half of oop-pair?
251 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" );
252 continue; // Do not record high parts in oopmap
253 }
254 #endif
255
256 // Check for a legal reg name in the oopMap and bailout if it is not.
257 if (!omap->legal_vm_reg_name(r)) {
258 regalloc->C->record_method_not_compilable("illegal oopMap register name");
259 continue;
260 }
261 if( t->is_ptr()->_offset == 0 ) { // Not derived?
262 if( mcall ) {
263 // Outgoing argument GC mask responsibility belongs to the callee,
264 // not the caller. Inspect the inputs to the call, to see if
265 // this live-range is one of them.
266 uint cnt = mcall->tf()->domain()->cnt();
267 uint j;
268 for( j = TypeFunc::Parms; j < cnt; j++)
269 if( mcall->in(j) == def )
270 break; // reaching def is an argument oop
271 if( j < cnt ) // arg oops dont go in GC map
272 continue; // Continue on to the next register
273 }
274 omap->set_oop(r);
275 } else { // Else it's derived.
276 // Find the base of the derived value.
277 uint i;
278 // Fast, common case, scan
279 for( i = jvms->oopoff(); i < n->req(); i+=2 )
280 if( n->in(i) == def ) break; // Common case
281 if( i == n->req() ) { // Missed, try a more generous scan
282 // Scan again, but this time peek through copies
283 for( i = jvms->oopoff(); i < n->req(); i+=2 ) {
284 Node *m = n->in(i); // Get initial derived value
285 while( 1 ) {
286 Node *d = def; // Get initial reaching def
287 while( 1 ) { // Follow copies of reaching def to end
288 if( m == d ) goto found; // breaks 3 loops
289 int idx = d->is_Copy();
290 if( !idx ) break;
291 d = d->in(idx); // Link through copy
292 }
293 int idx = m->is_Copy();
294 if( !idx ) break;
295 m = m->in(idx);
296 }
297 }
298 guarantee( 0, "must find derived/base pair" );
299 }
300 found: ;
301 Node *base = n->in(i+1); // Base is other half of pair
302 int breg = regalloc->get_reg_first(base);
303 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot);
304
305 // I record liveness at safepoints BEFORE I make the inputs
306 // live. This is because argument oops are NOT live at a
307 // safepoint (or at least they cannot appear in the oopmap).
308 // Thus bases of base/derived pairs might not be in the
309 // liveness data but they need to appear in the oopmap.
310 if( get_live_bit(live,breg) == 0 ) {// Not live?
311 // Flag it, so next derived pointer won't re-insert into oopmap
312 set_live_bit(live,breg);
313 // Already missed our turn?
314 if( breg < reg ) {
315 if (b->is_stack() || b->is_concrete() || true ) {
316 omap->set_oop( b);
317 }
318 }
319 }
320 if (b->is_stack() || b->is_concrete() || true ) {
321 omap->set_derived_oop( r, b);
322 }
323 }
324
325 } else if( t->isa_narrowoop() ) {
326 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" );
327 // Check for a legal reg name in the oopMap and bailout if it is not.
328 if (!omap->legal_vm_reg_name(r)) {
329 regalloc->C->record_method_not_compilable("illegal oopMap register name");
330 continue;
331 }
332 if( mcall ) {
333 // Outgoing argument GC mask responsibility belongs to the callee,
334 // not the caller. Inspect the inputs to the call, to see if
335 // this live-range is one of them.
336 uint cnt = mcall->tf()->domain()->cnt();
337 uint j;
338 for( j = TypeFunc::Parms; j < cnt; j++)
339 if( mcall->in(j) == def )
340 break; // reaching def is an argument oop
341 if( j < cnt ) // arg oops dont go in GC map
342 continue; // Continue on to the next register
343 }
344 omap->set_narrowoop(r);
345 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save?
346 // It's a callee-save value
347 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" );
348 debug_only( dup_check[_callees[reg]]=1; )
349 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg]));
350 if ( callee->is_concrete() || true ) {
351 omap->set_callee_saved( r, callee);
352 }
353
354 } else {
355 // Other - some reaching non-oop value
356 #ifdef ASSERT
357 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) {
358 def->dump();
359 n->dump();
360 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint");
361 }
362 #endif
363 }
364
365 }
366
367 #ifdef ASSERT
368 /* Nice, Intel-only assert
369 int cnt_callee_saves=0;
370 int reg2 = 0;
371 while (OptoReg::is_reg(reg2)) {
372 if( dup_check[reg2] != 0) cnt_callee_saves++;
373 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" );
374 reg2++;
375 }
376 */
377 #endif
378
379 #ifdef ASSERT
380 for( OopMapStream oms1(omap); !oms1.is_done(); oms1.next()) {
381 OopMapValue omv1 = oms1.current();
382 if (omv1.type() != OopMapValue::derived_oop_value) {
383 continue;
384 }
385 bool found = false;
386 for( OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) {
387 OopMapValue omv2 = oms2.current();
388 if (omv2.type() != OopMapValue::oop_value) {
389 continue;
390 }
391 if( omv1.content_reg() == omv2.reg() ) {
392 found = true;
393 break;
394 }
395 }
396 assert( found, "derived with no base in oopmap" );
397 }
398 #endif
399
400 return omap;
401 }
402
403 // Compute backwards liveness on registers
do_liveness(PhaseRegAlloc * regalloc,PhaseCFG * cfg,Block_List * worklist,int max_reg_ints,Arena * A,Dict * safehash)404 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) {
405 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints);
406 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints];
407 Node* root = cfg->get_root_node();
408 // On CISC platforms, get the node representing the stack pointer that regalloc
409 // used for spills
410 Node *fp = NodeSentinel;
411 if (UseCISCSpill && root->req() > 1) {
412 fp = root->in(1)->in(TypeFunc::FramePtr);
413 }
414 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt));
415 // Push preds onto worklist
416 for (uint i = 1; i < root->req(); i++) {
417 Block* block = cfg->get_block_for_node(root->in(i));
418 worklist->push(block);
419 }
420
421 // ZKM.jar includes tiny infinite loops which are unreached from below.
422 // If we missed any blocks, we'll retry here after pushing all missed
423 // blocks on the worklist. Normally this outer loop never trips more
424 // than once.
425 while (1) {
426
427 while( worklist->size() ) { // Standard worklist algorithm
428 Block *b = worklist->rpop();
429
430 // Copy first successor into my tmp_live space
431 int s0num = b->_succs[0]->_pre_order;
432 int *t = &live[s0num*max_reg_ints];
433 for( int i=0; i<max_reg_ints; i++ )
434 tmp_live[i] = t[i];
435
436 // OR in the remaining live registers
437 for( uint j=1; j<b->_num_succs; j++ ) {
438 uint sjnum = b->_succs[j]->_pre_order;
439 int *t = &live[sjnum*max_reg_ints];
440 for( int i=0; i<max_reg_ints; i++ )
441 tmp_live[i] |= t[i];
442 }
443
444 // Now walk tmp_live up the block backwards, computing live
445 for( int k=b->number_of_nodes()-1; k>=0; k-- ) {
446 Node *n = b->get_node(k);
447 // KILL def'd bits
448 int first = regalloc->get_reg_first(n);
449 int second = regalloc->get_reg_second(n);
450 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first);
451 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second);
452
453 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL;
454
455 // Check if m is potentially a CISC alternate instruction (i.e, possibly
456 // synthesized by RegAlloc from a conventional instruction and a
457 // spilled input)
458 bool is_cisc_alternate = false;
459 if (UseCISCSpill && m) {
460 is_cisc_alternate = m->is_cisc_alternate();
461 }
462
463 // GEN use'd bits
464 for( uint l=1; l<n->req(); l++ ) {
465 Node *def = n->in(l);
466 assert(def != 0, "input edge required");
467 int first = regalloc->get_reg_first(def);
468 int second = regalloc->get_reg_second(def);
469 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first);
470 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second);
471 // If we use the stack pointer in a cisc-alternative instruction,
472 // check for use as a memory operand. Then reconstruct the RegName
473 // for this stack location, and set the appropriate bit in the
474 // live vector 4987749.
475 if (is_cisc_alternate && def == fp) {
476 const TypePtr *adr_type = NULL;
477 intptr_t offset;
478 const Node* base = m->get_base_and_disp(offset, adr_type);
479 if (base == NodeSentinel) {
480 // Machnode has multiple memory inputs. We are unable to reason
481 // with these, but are presuming (with trepidation) that not any of
482 // them are oops. This can be fixed by making get_base_and_disp()
483 // look at a specific input instead of all inputs.
484 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input");
485 } else if (base != fp || offset == Type::OffsetBot) {
486 // Do nothing: the fp operand is either not from a memory use
487 // (base == NULL) OR the fp is used in a non-memory context
488 // (base is some other register) OR the offset is not constant,
489 // so it is not a stack slot.
490 } else {
491 assert(offset >= 0, "unexpected negative offset");
492 offset -= (offset % jintSize); // count the whole word
493 int stack_reg = regalloc->offset2reg(offset);
494 if (OptoReg::is_stack(stack_reg)) {
495 set_live_bit(tmp_live, stack_reg);
496 } else {
497 assert(false, "stack_reg not on stack?");
498 }
499 }
500 }
501 }
502
503 if( n->jvms() ) { // Record liveness at safepoint
504
505 // This placement of this stanza means inputs to calls are
506 // considered live at the callsite's OopMap. Argument oops are
507 // hence live, but NOT included in the oopmap. See cutout in
508 // build_oop_map. Debug oops are live (and in OopMap).
509 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints);
510 for( int l=0; l<max_reg_ints; l++ )
511 n_live[l] = tmp_live[l];
512 safehash->Insert(n,n_live);
513 }
514
515 }
516
517 // Now at block top, see if we have any changes. If so, propagate
518 // to prior blocks.
519 int *old_live = &live[b->_pre_order*max_reg_ints];
520 int l;
521 for( l=0; l<max_reg_ints; l++ )
522 if( tmp_live[l] != old_live[l] )
523 break;
524 if( l<max_reg_ints ) { // Change!
525 // Copy in new value
526 for( l=0; l<max_reg_ints; l++ )
527 old_live[l] = tmp_live[l];
528 // Push preds onto worklist
529 for (l = 1; l < (int)b->num_preds(); l++) {
530 Block* block = cfg->get_block_for_node(b->pred(l));
531 worklist->push(block);
532 }
533 }
534 }
535
536 // Scan for any missing safepoints. Happens to infinite loops
537 // ala ZKM.jar
538 uint i;
539 for (i = 1; i < cfg->number_of_blocks(); i++) {
540 Block* block = cfg->get_block(i);
541 uint j;
542 for (j = 1; j < block->number_of_nodes(); j++) {
543 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == NULL) {
544 break;
545 }
546 }
547 if (j < block->number_of_nodes()) {
548 break;
549 }
550 }
551 if (i == cfg->number_of_blocks()) {
552 break; // Got 'em all
553 }
554
555 if (PrintOpto && Verbose) {
556 tty->print_cr("retripping live calc");
557 }
558
559 // Force the issue (expensively): recheck everybody
560 for (i = 1; i < cfg->number_of_blocks(); i++) {
561 worklist->push(cfg->get_block(i));
562 }
563 }
564 }
565
566 // Collect GC mask info - where are all the OOPs?
BuildOopMaps()567 void PhaseOutput::BuildOopMaps() {
568 Compile::TracePhase tp("bldOopMaps", &timers[_t_buildOopMaps]);
569 // Can't resource-mark because I need to leave all those OopMaps around,
570 // or else I need to resource-mark some arena other than the default.
571 // ResourceMark rm; // Reclaim all OopFlows when done
572 int max_reg = C->regalloc()->_max_reg; // Current array extent
573
574 Arena *A = Thread::current()->resource_area();
575 Block_List worklist; // Worklist of pending blocks
576
577 int max_reg_ints = align_up(max_reg, BitsPerInt)>>LogBitsPerInt;
578 Dict *safehash = NULL; // Used for assert only
579 // Compute a backwards liveness per register. Needs a bitarray of
580 // #blocks x (#registers, rounded up to ints)
581 safehash = new Dict(cmpkey,hashkey,A);
582 do_liveness( C->regalloc(), C->cfg(), &worklist, max_reg_ints, A, safehash );
583 OopFlow *free_list = NULL; // Free, unused
584
585 // Array mapping blocks to completed oopflows
586 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, C->cfg()->number_of_blocks());
587 memset( flows, 0, C->cfg()->number_of_blocks() * sizeof(OopFlow*) );
588
589
590 // Do the first block 'by hand' to prime the worklist
591 Block *entry = C->cfg()->get_block(1);
592 OopFlow *rootflow = OopFlow::make(A,max_reg,C);
593 // Initialize to 'bottom' (not 'top')
594 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) );
595 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) );
596 flows[entry->_pre_order] = rootflow;
597
598 // Do the first block 'by hand' to prime the worklist
599 rootflow->_b = entry;
600 rootflow->compute_reach( C->regalloc(), max_reg, safehash );
601 for( uint i=0; i<entry->_num_succs; i++ )
602 worklist.push(entry->_succs[i]);
603
604 // Now worklist contains blocks which have some, but perhaps not all,
605 // predecessors visited.
606 while( worklist.size() ) {
607 // Scan for a block with all predecessors visited, or any randoms slob
608 // otherwise. All-preds-visited order allows me to recycle OopFlow
609 // structures rapidly and cut down on the memory footprint.
610 // Note: not all predecessors might be visited yet (must happen for
611 // irreducible loops). This is OK, since every live value must have the
612 // SAME reaching def for the block, so any reaching def is OK.
613 uint i;
614
615 Block *b = worklist.pop();
616 // Ignore root block
617 if (b == C->cfg()->get_root_block()) {
618 continue;
619 }
620 // Block is already done? Happens if block has several predecessors,
621 // he can get on the worklist more than once.
622 if( flows[b->_pre_order] ) continue;
623
624 // If this block has a visited predecessor AND that predecessor has this
625 // last block as his only undone child, we can move the OopFlow from the
626 // pred to this block. Otherwise we have to grab a new OopFlow.
627 OopFlow *flow = NULL; // Flag for finding optimized flow
628 Block *pred = (Block*)((intptr_t)0xdeadbeef);
629 // Scan this block's preds to find a done predecessor
630 for (uint j = 1; j < b->num_preds(); j++) {
631 Block* p = C->cfg()->get_block_for_node(b->pred(j));
632 OopFlow *p_flow = flows[p->_pre_order];
633 if( p_flow ) { // Predecessor is done
634 assert( p_flow->_b == p, "cross check" );
635 pred = p; // Record some predecessor
636 // If all successors of p are done except for 'b', then we can carry
637 // p_flow forward to 'b' without copying, otherwise we have to draw
638 // from the free_list and clone data.
639 uint k;
640 for( k=0; k<p->_num_succs; k++ )
641 if( !flows[p->_succs[k]->_pre_order] &&
642 p->_succs[k] != b )
643 break;
644
645 // Either carry-forward the now-unused OopFlow for b's use
646 // or draw a new one from the free list
647 if( k==p->_num_succs ) {
648 flow = p_flow;
649 break; // Found an ideal pred, use him
650 }
651 }
652 }
653
654 if( flow ) {
655 // We have an OopFlow that's the last-use of a predecessor.
656 // Carry it forward.
657 } else { // Draw a new OopFlow from the freelist
658 if( !free_list )
659 free_list = OopFlow::make(A,max_reg,C);
660 flow = free_list;
661 assert( flow->_b == NULL, "oopFlow is not free" );
662 free_list = flow->_next;
663 flow->_next = NULL;
664
665 // Copy/clone over the data
666 flow->clone(flows[pred->_pre_order], max_reg);
667 }
668
669 // Mark flow for block. Blocks can only be flowed over once,
670 // because after the first time they are guarded from entering
671 // this code again.
672 assert( flow->_b == pred, "have some prior flow" );
673 flow->_b = NULL;
674
675 // Now push flow forward
676 flows[b->_pre_order] = flow;// Mark flow for this block
677 flow->_b = b;
678 flow->compute_reach( C->regalloc(), max_reg, safehash );
679
680 // Now push children onto worklist
681 for( i=0; i<b->_num_succs; i++ )
682 worklist.push(b->_succs[i]);
683
684 }
685 }
686