1 /*
2 * Copyright 1998-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 // Optimization - Graph Style
26
27 #include "incls/_precompiled.incl"
28 #include "incls/_lcm.cpp.incl"
29
30 //------------------------------implicit_null_check----------------------------
31 // Detect implicit-null-check opportunities. Basically, find NULL checks
32 // with suitable memory ops nearby. Use the memory op to do the NULL check.
33 // I can generate a memory op if there is not one nearby.
34 // The proj is the control projection for the not-null case.
35 // The val is the pointer being checked for nullness.
36 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
37 // Assume if null check need for 0 offset then always needed
38 // Intel solaris doesn't support any null checks yet and no
39 // mechanism exists (yet) to set the switches at an os_cpu level
40 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
41
42 // Make sure the ptr-is-null path appears to be uncommon!
43 float f = end()->as_MachIf()->_prob;
44 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
45 if( f > PROB_UNLIKELY_MAG(4) ) return;
46
47 uint bidx = 0; // Capture index of value into memop
48 bool was_store; // Memory op is a store op
49
50 // Get the successor block for if the test ptr is non-null
51 Block* not_null_block; // this one goes with the proj
52 Block* null_block;
53 if (_nodes[_nodes.size()-1] == proj) {
54 null_block = _succs[0];
55 not_null_block = _succs[1];
56 } else {
57 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
58 not_null_block = _succs[0];
59 null_block = _succs[1];
60 }
61
62 // Search the exception block for an uncommon trap.
63 // (See Parse::do_if and Parse::do_ifnull for the reason
64 // we need an uncommon trap. Briefly, we need a way to
65 // detect failure of this optimization, as in 6366351.)
66 {
67 bool found_trap = false;
68 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
69 Node* nn = null_block->_nodes[i1];
70 if (nn->is_MachCall() &&
71 nn->as_MachCall()->entry_point() ==
72 SharedRuntime::uncommon_trap_blob()->instructions_begin()) {
73 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
74 if (trtype->isa_int() && trtype->is_int()->is_con()) {
75 jint tr_con = trtype->is_int()->get_con();
76 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
77 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
78 assert((int)reason < (int)BitsPerInt, "recode bit map");
79 if (is_set_nth_bit(allowed_reasons, (int) reason)
80 && action != Deoptimization::Action_none) {
81 // This uncommon trap is sure to recompile, eventually.
82 // When that happens, C->too_many_traps will prevent
83 // this transformation from happening again.
84 found_trap = true;
85 }
86 }
87 break;
88 }
89 }
90 if (!found_trap) {
91 // We did not find an uncommon trap.
92 return;
93 }
94 }
95
96 // Search the successor block for a load or store who's base value is also
97 // the tested value. There may be several.
98 Node_List *out = new Node_List(Thread::current()->resource_area());
99 MachNode *best = NULL; // Best found so far
100 for (DUIterator i = val->outs(); val->has_out(i); i++) {
101 Node *m = val->out(i);
102 if( !m->is_Mach() ) continue;
103 MachNode *mach = m->as_Mach();
104 was_store = false;
105 switch( mach->ideal_Opcode() ) {
106 case Op_LoadB:
107 case Op_LoadC:
108 case Op_LoadD:
109 case Op_LoadF:
110 case Op_LoadI:
111 case Op_LoadL:
112 case Op_LoadP:
113 case Op_LoadN:
114 case Op_LoadS:
115 case Op_LoadKlass:
116 case Op_LoadNKlass:
117 case Op_LoadRange:
118 case Op_LoadD_unaligned:
119 case Op_LoadL_unaligned:
120 break;
121 case Op_StoreB:
122 case Op_StoreC:
123 case Op_StoreCM:
124 case Op_StoreD:
125 case Op_StoreF:
126 case Op_StoreI:
127 case Op_StoreL:
128 case Op_StoreP:
129 case Op_StoreN:
130 was_store = true; // Memory op is a store op
131 // Stores will have their address in slot 2 (memory in slot 1).
132 // If the value being nul-checked is in another slot, it means we
133 // are storing the checked value, which does NOT check the value!
134 if( mach->in(2) != val ) continue;
135 break; // Found a memory op?
136 case Op_StrComp:
137 case Op_AryEq:
138 // Not a legit memory op for implicit null check regardless of
139 // embedded loads
140 continue;
141 default: // Also check for embedded loads
142 if( !mach->needs_anti_dependence_check() )
143 continue; // Not an memory op; skip it
144 break;
145 }
146 // check if the offset is not too high for implicit exception
147 {
148 intptr_t offset = 0;
149 const TypePtr *adr_type = NULL; // Do not need this return value here
150 const Node* base = mach->get_base_and_disp(offset, adr_type);
151 if (base == NULL || base == NodeSentinel) {
152 // cannot reason about it; is probably not implicit null exception
153 } else {
154 const TypePtr* tptr = base->bottom_type()->is_ptr();
155 // Give up if offset is not a compile-time constant
156 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
157 continue;
158 offset += tptr->_offset; // correct if base is offseted
159 if( MacroAssembler::needs_explicit_null_check(offset) )
160 continue; // Give up is reference is beyond 4K page size
161 }
162 }
163
164 // Check ctrl input to see if the null-check dominates the memory op
165 Block *cb = cfg->_bbs[mach->_idx];
166 cb = cb->_idom; // Always hoist at least 1 block
167 if( !was_store ) { // Stores can be hoisted only one block
168 while( cb->_dom_depth > (_dom_depth + 1))
169 cb = cb->_idom; // Hoist loads as far as we want
170 // The non-null-block should dominate the memory op, too. Live
171 // range spilling will insert a spill in the non-null-block if it is
172 // needs to spill the memory op for an implicit null check.
173 if (cb->_dom_depth == (_dom_depth + 1)) {
174 if (cb != not_null_block) continue;
175 cb = cb->_idom;
176 }
177 }
178 if( cb != this ) continue;
179
180 // Found a memory user; see if it can be hoisted to check-block
181 uint vidx = 0; // Capture index of value into memop
182 uint j;
183 for( j = mach->req()-1; j > 0; j-- ) {
184 if( mach->in(j) == val ) vidx = j;
185 // Block of memory-op input
186 Block *inb = cfg->_bbs[mach->in(j)->_idx];
187 Block *b = this; // Start from nul check
188 while( b != inb && b->_dom_depth > inb->_dom_depth )
189 b = b->_idom; // search upwards for input
190 // See if input dominates null check
191 if( b != inb )
192 break;
193 }
194 if( j > 0 )
195 continue;
196 Block *mb = cfg->_bbs[mach->_idx];
197 // Hoisting stores requires more checks for the anti-dependence case.
198 // Give up hoisting if we have to move the store past any load.
199 if( was_store ) {
200 Block *b = mb; // Start searching here for a local load
201 // mach use (faulting) trying to hoist
202 // n might be blocker to hoisting
203 while( b != this ) {
204 uint k;
205 for( k = 1; k < b->_nodes.size(); k++ ) {
206 Node *n = b->_nodes[k];
207 if( n->needs_anti_dependence_check() &&
208 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
209 break; // Found anti-dependent load
210 }
211 if( k < b->_nodes.size() )
212 break; // Found anti-dependent load
213 // Make sure control does not do a merge (would have to check allpaths)
214 if( b->num_preds() != 2 ) break;
215 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
216 }
217 if( b != this ) continue;
218 }
219
220 // Make sure this memory op is not already being used for a NullCheck
221 Node *e = mb->end();
222 if( e->is_MachNullCheck() && e->in(1) == mach )
223 continue; // Already being used as a NULL check
224
225 // Found a candidate! Pick one with least dom depth - the highest
226 // in the dom tree should be closest to the null check.
227 if( !best ||
228 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
229 best = mach;
230 bidx = vidx;
231
232 }
233 }
234 // No candidate!
235 if( !best ) return;
236
237 // ---- Found an implicit null check
238 extern int implicit_null_checks;
239 implicit_null_checks++;
240
241 // Hoist the memory candidate up to the end of the test block.
242 Block *old_block = cfg->_bbs[best->_idx];
243 old_block->find_remove(best);
244 add_inst(best);
245 cfg->_bbs.map(best->_idx,this);
246
247 // Move the control dependence
248 if (best->in(0) && best->in(0) == old_block->_nodes[0])
249 best->set_req(0, _nodes[0]);
250
251 // Check for flag-killing projections that also need to be hoisted
252 // Should be DU safe because no edge updates.
253 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
254 Node* n = best->fast_out(j);
255 if( n->Opcode() == Op_MachProj ) {
256 cfg->_bbs[n->_idx]->find_remove(n);
257 add_inst(n);
258 cfg->_bbs.map(n->_idx,this);
259 }
260 }
261
262 Compile *C = cfg->C;
263 // proj==Op_True --> ne test; proj==Op_False --> eq test.
264 // One of two graph shapes got matched:
265 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
266 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
267 // NULL checks are always branch-if-eq. If we see a IfTrue projection
268 // then we are replacing a 'ne' test with a 'eq' NULL check test.
269 // We need to flip the projections to keep the same semantics.
270 if( proj->Opcode() == Op_IfTrue ) {
271 // Swap order of projections in basic block to swap branch targets
272 Node *tmp1 = _nodes[end_idx()+1];
273 Node *tmp2 = _nodes[end_idx()+2];
274 _nodes.map(end_idx()+1, tmp2);
275 _nodes.map(end_idx()+2, tmp1);
276 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
277 tmp1->replace_by(tmp);
278 tmp2->replace_by(tmp1);
279 tmp->replace_by(tmp2);
280 tmp->destruct();
281 }
282
283 // Remove the existing null check; use a new implicit null check instead.
284 // Since schedule-local needs precise def-use info, we need to correct
285 // it as well.
286 Node *old_tst = proj->in(0);
287 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
288 _nodes.map(end_idx(),nul_chk);
289 cfg->_bbs.map(nul_chk->_idx,this);
290 // Redirect users of old_test to nul_chk
291 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
292 old_tst->last_out(i2)->set_req(0, nul_chk);
293 // Clean-up any dead code
294 for (uint i3 = 0; i3 < old_tst->req(); i3++)
295 old_tst->set_req(i3, NULL);
296
297 cfg->latency_from_uses(nul_chk);
298 cfg->latency_from_uses(best);
299 }
300
301
302 //------------------------------select-----------------------------------------
303 // Select a nice fellow from the worklist to schedule next. If there is only
304 // one choice, then use it. Projections take top priority for correctness
305 // reasons - if I see a projection, then it is next. There are a number of
306 // other special cases, for instructions that consume condition codes, et al.
307 // These are chosen immediately. Some instructions are required to immediately
308 // precede the last instruction in the block, and these are taken last. Of the
309 // remaining cases (most), choose the instruction with the greatest latency
310 // (that is, the most number of pseudo-cycles required to the end of the
311 // routine). If there is a tie, choose the instruction with the most inputs.
312 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) {
313
314 // If only a single entry on the stack, use it
315 uint cnt = worklist.size();
316 if (cnt == 1) {
317 Node *n = worklist[0];
318 worklist.map(0,worklist.pop());
319 return n;
320 }
321
322 uint choice = 0; // Bigger is most important
323 uint latency = 0; // Bigger is scheduled first
324 uint score = 0; // Bigger is better
325 int idx = -1; // Index in worklist
326
327 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
328 // Order in worklist is used to break ties.
329 // See caller for how this is used to delay scheduling
330 // of induction variable increments to after the other
331 // uses of the phi are scheduled.
332 Node *n = worklist[i]; // Get Node on worklist
333
334 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
335 if( n->is_Proj() || // Projections always win
336 n->Opcode()== Op_Con || // So does constant 'Top'
337 iop == Op_CreateEx || // Create-exception must start block
338 iop == Op_CheckCastPP
339 ) {
340 worklist.map(i,worklist.pop());
341 return n;
342 }
343
344 // Final call in a block must be adjacent to 'catch'
345 Node *e = end();
346 if( e->is_Catch() && e->in(0)->in(0) == n )
347 continue;
348
349 // Memory op for an implicit null check has to be at the end of the block
350 if( e->is_MachNullCheck() && e->in(1) == n )
351 continue;
352
353 uint n_choice = 2;
354
355 // See if this instruction is consumed by a branch. If so, then (as the
356 // branch is the last instruction in the basic block) force it to the
357 // end of the basic block
358 if ( must_clone[iop] ) {
359 // See if any use is a branch
360 bool found_machif = false;
361
362 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
363 Node* use = n->fast_out(j);
364
365 // The use is a conditional branch, make them adjacent
366 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
367 found_machif = true;
368 break;
369 }
370
371 // More than this instruction pending for successor to be ready,
372 // don't choose this if other opportunities are ready
373 if (ready_cnt[use->_idx] > 1)
374 n_choice = 1;
375 }
376
377 // loop terminated, prefer not to use this instruction
378 if (found_machif)
379 continue;
380 }
381
382 // See if this has a predecessor that is "must_clone", i.e. sets the
383 // condition code. If so, choose this first
384 for (uint j = 0; j < n->req() ; j++) {
385 Node *inn = n->in(j);
386 if (inn) {
387 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
388 n_choice = 3;
389 break;
390 }
391 }
392 }
393
394 // MachTemps should be scheduled last so they are near their uses
395 if (n->is_MachTemp()) {
396 n_choice = 1;
397 }
398
399 uint n_latency = cfg->_node_latency.at_grow(n->_idx);
400 uint n_score = n->req(); // Many inputs get high score to break ties
401
402 // Keep best latency found
403 if( choice < n_choice ||
404 ( choice == n_choice &&
405 ( latency < n_latency ||
406 ( latency == n_latency &&
407 ( score < n_score ))))) {
408 choice = n_choice;
409 latency = n_latency;
410 score = n_score;
411 idx = i; // Also keep index in worklist
412 }
413 } // End of for all ready nodes in worklist
414
415 assert(idx >= 0, "index should be set");
416 Node *n = worklist[(uint)idx]; // Get the winner
417
418 worklist.map((uint)idx, worklist.pop()); // Compress worklist
419 return n;
420 }
421
422
423 //------------------------------set_next_call----------------------------------
424 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
425 if( next_call.test_set(n->_idx) ) return;
426 for( uint i=0; i<n->len(); i++ ) {
427 Node *m = n->in(i);
428 if( !m ) continue; // must see all nodes in block that precede call
429 if( bbs[m->_idx] == this )
430 set_next_call( m, next_call, bbs );
431 }
432 }
433
434 //------------------------------needed_for_next_call---------------------------
435 // Set the flag 'next_call' for each Node that is needed for the next call to
436 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
437 // next subroutine call get priority - basically it moves things NOT needed
438 // for the next call till after the call. This prevents me from trying to
439 // carry lots of stuff live across a call.
440 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
441 // Find the next control-defining Node in this block
442 Node* call = NULL;
443 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
444 Node* m = this_call->fast_out(i);
445 if( bbs[m->_idx] == this && // Local-block user
446 m != this_call && // Not self-start node
447 m->is_Call() )
448 call = m;
449 break;
450 }
451 if (call == NULL) return; // No next call (e.g., block end is near)
452 // Set next-call for all inputs to this call
453 set_next_call(call, next_call, bbs);
454 }
455
456 //------------------------------sched_call-------------------------------------
457 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
458 RegMask regs;
459
460 // Schedule all the users of the call right now. All the users are
461 // projection Nodes, so they must be scheduled next to the call.
462 // Collect all the defined registers.
463 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
464 Node* n = mcall->fast_out(i);
465 assert( n->Opcode()==Op_MachProj, "" );
466 --ready_cnt[n->_idx];
467 assert( !ready_cnt[n->_idx], "" );
468 // Schedule next to call
469 _nodes.map(node_cnt++, n);
470 // Collect defined registers
471 regs.OR(n->out_RegMask());
472 // Check for scheduling the next control-definer
473 if( n->bottom_type() == Type::CONTROL )
474 // Warm up next pile of heuristic bits
475 needed_for_next_call(n, next_call, bbs);
476
477 // Children of projections are now all ready
478 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
479 Node* m = n->fast_out(j); // Get user
480 if( bbs[m->_idx] != this ) continue;
481 if( m->is_Phi() ) continue;
482 if( !--ready_cnt[m->_idx] )
483 worklist.push(m);
484 }
485
486 }
487
488 // Act as if the call defines the Frame Pointer.
489 // Certainly the FP is alive and well after the call.
490 regs.Insert(matcher.c_frame_pointer());
491
492 // Set all registers killed and not already defined by the call.
493 uint r_cnt = mcall->tf()->range()->cnt();
494 int op = mcall->ideal_Opcode();
495 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
496 bbs.map(proj->_idx,this);
497 _nodes.insert(node_cnt++, proj);
498
499 // Select the right register save policy.
500 const char * save_policy;
501 switch (op) {
502 case Op_CallRuntime:
503 case Op_CallLeaf:
504 case Op_CallLeafNoFP:
505 // Calling C code so use C calling convention
506 save_policy = matcher._c_reg_save_policy;
507 break;
508
509 case Op_CallStaticJava:
510 case Op_CallDynamicJava:
511 // Calling Java code so use Java calling convention
512 save_policy = matcher._register_save_policy;
513 break;
514
515 default:
516 ShouldNotReachHere();
517 }
518
519 // When using CallRuntime mark SOE registers as killed by the call
520 // so values that could show up in the RegisterMap aren't live in a
521 // callee saved register since the register wouldn't know where to
522 // find them. CallLeaf and CallLeafNoFP are ok because they can't
523 // have debug info on them. Strictly speaking this only needs to be
524 // done for oops since idealreg2debugmask takes care of debug info
525 // references but there no way to handle oops differently than other
526 // pointers as far as the kill mask goes.
527 bool exclude_soe = op == Op_CallRuntime;
528
529 // Fill in the kill mask for the call
530 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
531 if( !regs.Member(r) ) { // Not already defined by the call
532 // Save-on-call register?
533 if ((save_policy[r] == 'C') ||
534 (save_policy[r] == 'A') ||
535 ((save_policy[r] == 'E') && exclude_soe)) {
536 proj->_rout.Insert(r);
537 }
538 }
539 }
540
541 return node_cnt;
542 }
543
544
545 //------------------------------schedule_local---------------------------------
546 // Topological sort within a block. Someday become a real scheduler.
547 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) {
548 // Already "sorted" are the block start Node (as the first entry), and
549 // the block-ending Node and any trailing control projections. We leave
550 // these alone. PhiNodes and ParmNodes are made to follow the block start
551 // Node. Everything else gets topo-sorted.
552
553 #ifndef PRODUCT
554 if (cfg->trace_opto_pipelining()) {
555 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
556 for (uint i = 0;i < _nodes.size();i++) {
557 tty->print("# ");
558 _nodes[i]->fast_dump();
559 }
560 tty->print_cr("#");
561 }
562 #endif
563
564 // RootNode is already sorted
565 if( _nodes.size() == 1 ) return true;
566
567 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
568 uint node_cnt = end_idx();
569 uint phi_cnt = 1;
570 uint i;
571 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
572 Node *n = _nodes[i];
573 if( n->is_Phi() || // Found a PhiNode or ParmNode
574 (n->is_Proj() && n->in(0) == head()) ) {
575 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
576 _nodes.map(i,_nodes[phi_cnt]);
577 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
578 } else { // All others
579 // Count block-local inputs to 'n'
580 uint cnt = n->len(); // Input count
581 uint local = 0;
582 for( uint j=0; j<cnt; j++ ) {
583 Node *m = n->in(j);
584 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
585 local++; // One more block-local input
586 }
587 ready_cnt[n->_idx] = local; // Count em up
588
589 // A few node types require changing a required edge to a precedence edge
590 // before allocation.
591 if( UseConcMarkSweepGC ) {
592 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
593 // Note: Required edges with an index greater than oper_input_base
594 // are not supported by the allocator.
595 // Note2: Can only depend on unmatched edge being last,
596 // can not depend on its absolute position.
597 Node *oop_store = n->in(n->req() - 1);
598 n->del_req(n->req() - 1);
599 n->add_prec(oop_store);
600 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
601 }
602 }
603 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire &&
604 n->req() > TypeFunc::Parms ) {
605 // MemBarAcquire could be created without Precedent edge.
606 // del_req() replaces the specified edge with the last input edge
607 // and then removes the last edge. If the specified edge > number of
608 // edges the last edge will be moved outside of the input edges array
609 // and the edge will be lost. This is why this code should be
610 // executed only when Precedent (== TypeFunc::Parms) edge is present.
611 Node *x = n->in(TypeFunc::Parms);
612 n->del_req(TypeFunc::Parms);
613 n->add_prec(x);
614 }
615 }
616 }
617 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
618 ready_cnt[_nodes[i2]->_idx] = 0;
619
620 // All the prescheduled guys do not hold back internal nodes
621 uint i3;
622 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
623 Node *n = _nodes[i3]; // Get pre-scheduled
624 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
625 Node* m = n->fast_out(j);
626 if( cfg->_bbs[m->_idx] ==this ) // Local-block user
627 ready_cnt[m->_idx]--; // Fix ready count
628 }
629 }
630
631 Node_List delay;
632 // Make a worklist
633 Node_List worklist;
634 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
635 Node *m = _nodes[i4];
636 if( !ready_cnt[m->_idx] ) { // Zero ready count?
637 if (m->is_iteratively_computed()) {
638 // Push induction variable increments last to allow other uses
639 // of the phi to be scheduled first. The select() method breaks
640 // ties in scheduling by worklist order.
641 delay.push(m);
642 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
643 // Force the CreateEx to the top of the list so it's processed
644 // first and ends up at the start of the block.
645 worklist.insert(0, m);
646 } else {
647 worklist.push(m); // Then on to worklist!
648 }
649 }
650 }
651 while (delay.size()) {
652 Node* d = delay.pop();
653 worklist.push(d);
654 }
655
656 // Warm up the 'next_call' heuristic bits
657 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
658
659 #ifndef PRODUCT
660 if (cfg->trace_opto_pipelining()) {
661 for (uint j=0; j<_nodes.size(); j++) {
662 Node *n = _nodes[j];
663 int idx = n->_idx;
664 tty->print("# ready cnt:%3d ", ready_cnt[idx]);
665 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx));
666 tty->print("%4d: %s\n", idx, n->Name());
667 }
668 }
669 #endif
670
671 // Pull from worklist and schedule
672 while( worklist.size() ) { // Worklist is not ready
673
674 #ifndef PRODUCT
675 if (cfg->trace_opto_pipelining()) {
676 tty->print("# ready list:");
677 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
678 Node *n = worklist[i]; // Get Node on worklist
679 tty->print(" %d", n->_idx);
680 }
681 tty->cr();
682 }
683 #endif
684
685 // Select and pop a ready guy from worklist
686 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
687 _nodes.map(phi_cnt++,n); // Schedule him next
688
689 #ifndef PRODUCT
690 if (cfg->trace_opto_pipelining()) {
691 tty->print("# select %d: %s", n->_idx, n->Name());
692 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx));
693 n->dump();
694 if (Verbose) {
695 tty->print("# ready list:");
696 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
697 Node *n = worklist[i]; // Get Node on worklist
698 tty->print(" %d", n->_idx);
699 }
700 tty->cr();
701 }
702 }
703
704 #endif
705 if( n->is_MachCall() ) {
706 MachCallNode *mcall = n->as_MachCall();
707 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
708 continue;
709 }
710 // Children are now all ready
711 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
712 Node* m = n->fast_out(i5); // Get user
713 if( cfg->_bbs[m->_idx] != this ) continue;
714 if( m->is_Phi() ) continue;
715 if( !--ready_cnt[m->_idx] )
716 worklist.push(m);
717 }
718 }
719
720 if( phi_cnt != end_idx() ) {
721 // did not schedule all. Retry, Bailout, or Die
722 Compile* C = matcher.C;
723 if (C->subsume_loads() == true && !C->failing()) {
724 // Retry with subsume_loads == false
725 // If this is the first failure, the sentinel string will "stick"
726 // to the Compile object, and the C2Compiler will see it and retry.
727 C->record_failure(C2Compiler::retry_no_subsuming_loads());
728 }
729 // assert( phi_cnt == end_idx(), "did not schedule all" );
730 return false;
731 }
732
733 #ifndef PRODUCT
734 if (cfg->trace_opto_pipelining()) {
735 tty->print_cr("#");
736 tty->print_cr("# after schedule_local");
737 for (uint i = 0;i < _nodes.size();i++) {
738 tty->print("# ");
739 _nodes[i]->fast_dump();
740 }
741 tty->cr();
742 }
743 #endif
744
745
746 return true;
747 }
748
749 //--------------------------catch_cleanup_fix_all_inputs-----------------------
750 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
751 for (uint l = 0; l < use->len(); l++) {
752 if (use->in(l) == old_def) {
753 if (l < use->req()) {
754 use->set_req(l, new_def);
755 } else {
756 use->rm_prec(l);
757 use->add_prec(new_def);
758 l--;
759 }
760 }
761 }
762 }
763
764 //------------------------------catch_cleanup_find_cloned_def------------------
765 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
766 assert( use_blk != def_blk, "Inter-block cleanup only");
767
768 // The use is some block below the Catch. Find and return the clone of the def
769 // that dominates the use. If there is no clone in a dominating block, then
770 // create a phi for the def in a dominating block.
771
772 // Find which successor block dominates this use. The successor
773 // blocks must all be single-entry (from the Catch only; I will have
774 // split blocks to make this so), hence they all dominate.
775 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
776 use_blk = use_blk->_idom;
777
778 // Find the successor
779 Node *fixup = NULL;
780
781 uint j;
782 for( j = 0; j < def_blk->_num_succs; j++ )
783 if( use_blk == def_blk->_succs[j] )
784 break;
785
786 if( j == def_blk->_num_succs ) {
787 // Block at same level in dom-tree is not a successor. It needs a
788 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
789 Node_Array inputs = new Node_List(Thread::current()->resource_area());
790 for(uint k = 1; k < use_blk->num_preds(); k++) {
791 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
792 }
793
794 // Check to see if the use_blk already has an identical phi inserted.
795 // If it exists, it will be at the first position since all uses of a
796 // def are processed together.
797 Node *phi = use_blk->_nodes[1];
798 if( phi->is_Phi() ) {
799 fixup = phi;
800 for (uint k = 1; k < use_blk->num_preds(); k++) {
801 if (phi->in(k) != inputs[k]) {
802 // Not a match
803 fixup = NULL;
804 break;
805 }
806 }
807 }
808
809 // If an existing PhiNode was not found, make a new one.
810 if (fixup == NULL) {
811 Node *new_phi = PhiNode::make(use_blk->head(), def);
812 use_blk->_nodes.insert(1, new_phi);
813 bbs.map(new_phi->_idx, use_blk);
814 for (uint k = 1; k < use_blk->num_preds(); k++) {
815 new_phi->set_req(k, inputs[k]);
816 }
817 fixup = new_phi;
818 }
819
820 } else {
821 // Found the use just below the Catch. Make it use the clone.
822 fixup = use_blk->_nodes[n_clone_idx];
823 }
824
825 return fixup;
826 }
827
828 //--------------------------catch_cleanup_intra_block--------------------------
829 // Fix all input edges in use that reference "def". The use is in the same
830 // block as the def and both have been cloned in each successor block.
831 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
832
833 // Both the use and def have been cloned. For each successor block,
834 // get the clone of the use, and make its input the clone of the def
835 // found in that block.
836
837 uint use_idx = blk->find_node(use);
838 uint offset_idx = use_idx - beg;
839 for( uint k = 0; k < blk->_num_succs; k++ ) {
840 // Get clone in each successor block
841 Block *sb = blk->_succs[k];
842 Node *clone = sb->_nodes[offset_idx+1];
843 assert( clone->Opcode() == use->Opcode(), "" );
844
845 // Make use-clone reference the def-clone
846 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
847 }
848 }
849
850 //------------------------------catch_cleanup_inter_block---------------------
851 // Fix all input edges in use that reference "def". The use is in a different
852 // block than the def.
853 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
854 if( !use_blk ) return; // Can happen if the use is a precedence edge
855
856 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
857 catch_cleanup_fix_all_inputs(use, def, new_def);
858 }
859
860 //------------------------------call_catch_cleanup-----------------------------
861 // If we inserted any instructions between a Call and his CatchNode,
862 // clone the instructions on all paths below the Catch.
863 void Block::call_catch_cleanup(Block_Array &bbs) {
864
865 // End of region to clone
866 uint end = end_idx();
867 if( !_nodes[end]->is_Catch() ) return;
868 // Start of region to clone
869 uint beg = end;
870 while( _nodes[beg-1]->Opcode() != Op_MachProj ||
871 !_nodes[beg-1]->in(0)->is_Call() ) {
872 beg--;
873 assert(beg > 0,"Catch cleanup walking beyond block boundary");
874 }
875 // Range of inserted instructions is [beg, end)
876 if( beg == end ) return;
877
878 // Clone along all Catch output paths. Clone area between the 'beg' and
879 // 'end' indices.
880 for( uint i = 0; i < _num_succs; i++ ) {
881 Block *sb = _succs[i];
882 // Clone the entire area; ignoring the edge fixup for now.
883 for( uint j = end; j > beg; j-- ) {
884 Node *clone = _nodes[j-1]->clone();
885 sb->_nodes.insert( 1, clone );
886 bbs.map(clone->_idx,sb);
887 }
888 }
889
890
891 // Fixup edges. Check the def-use info per cloned Node
892 for(uint i2 = beg; i2 < end; i2++ ) {
893 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
894 Node *n = _nodes[i2]; // Node that got cloned
895 // Need DU safe iterator because of edge manipulation in calls.
896 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
897 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
898 out->push(n->fast_out(j1));
899 }
900 uint max = out->size();
901 for (uint j = 0; j < max; j++) {// For all users
902 Node *use = out->pop();
903 Block *buse = bbs[use->_idx];
904 if( use->is_Phi() ) {
905 for( uint k = 1; k < use->req(); k++ )
906 if( use->in(k) == n ) {
907 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
908 use->set_req(k, fixup);
909 }
910 } else {
911 if (this == buse) {
912 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
913 } else {
914 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
915 }
916 }
917 } // End for all users
918
919 } // End of for all Nodes in cloned area
920
921 // Remove the now-dead cloned ops
922 for(uint i3 = beg; i3 < end; i3++ ) {
923 _nodes[beg]->disconnect_inputs(NULL);
924 _nodes.remove(beg);
925 }
926
927 // If the successor blocks have a CreateEx node, move it back to the top
928 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
929 Block *sb = _succs[i4];
930 uint new_cnt = end - beg;
931 // Remove any newly created, but dead, nodes.
932 for( uint j = new_cnt; j > 0; j-- ) {
933 Node *n = sb->_nodes[j];
934 if (n->outcnt() == 0 &&
935 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
936 n->disconnect_inputs(NULL);
937 sb->_nodes.remove(j);
938 new_cnt--;
939 }
940 }
941 // If any newly created nodes remain, move the CreateEx node to the top
942 if (new_cnt > 0) {
943 Node *cex = sb->_nodes[1+new_cnt];
944 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
945 sb->_nodes.remove(1+new_cnt);
946 sb->_nodes.insert(1,cex);
947 }
948 }
949 }
950 }