Mon, 28 Apr 2008 08:08:12 -0700
Merge
1 /*
2 * Copyright 1998-2007 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 */
25 // Optimization - Graph Style
27 #include "incls/_precompiled.incl"
28 #include "incls/_lcm.cpp.incl"
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;
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;
47 uint bidx = 0; // Capture index of value into memop
48 bool was_store; // Memory op is a store op
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 }
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 }
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_LoadRange:
117 case Op_LoadD_unaligned:
118 case Op_LoadL_unaligned:
119 break;
120 case Op_StoreB:
121 case Op_StoreC:
122 case Op_StoreCM:
123 case Op_StoreD:
124 case Op_StoreF:
125 case Op_StoreI:
126 case Op_StoreL:
127 case Op_StoreP:
128 case Op_StoreN:
129 was_store = true; // Memory op is a store op
130 // Stores will have their address in slot 2 (memory in slot 1).
131 // If the value being nul-checked is in another slot, it means we
132 // are storing the checked value, which does NOT check the value!
133 if( mach->in(2) != val ) continue;
134 break; // Found a memory op?
135 case Op_StrComp:
136 // Not a legit memory op for implicit null check regardless of
137 // embedded loads
138 continue;
139 default: // Also check for embedded loads
140 if( !mach->needs_anti_dependence_check() )
141 continue; // Not an memory op; skip it
142 break;
143 }
144 // check if the offset is not too high for implicit exception
145 {
146 intptr_t offset = 0;
147 const TypePtr *adr_type = NULL; // Do not need this return value here
148 const Node* base = mach->get_base_and_disp(offset, adr_type);
149 if (base == NULL || base == NodeSentinel) {
150 // cannot reason about it; is probably not implicit null exception
151 } else {
152 const TypePtr* tptr = base->bottom_type()->is_ptr();
153 // Give up if offset is not a compile-time constant
154 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
155 continue;
156 offset += tptr->_offset; // correct if base is offseted
157 if( MacroAssembler::needs_explicit_null_check(offset) )
158 continue; // Give up is reference is beyond 4K page size
159 }
160 }
162 // Check ctrl input to see if the null-check dominates the memory op
163 Block *cb = cfg->_bbs[mach->_idx];
164 cb = cb->_idom; // Always hoist at least 1 block
165 if( !was_store ) { // Stores can be hoisted only one block
166 while( cb->_dom_depth > (_dom_depth + 1))
167 cb = cb->_idom; // Hoist loads as far as we want
168 // The non-null-block should dominate the memory op, too. Live
169 // range spilling will insert a spill in the non-null-block if it is
170 // needs to spill the memory op for an implicit null check.
171 if (cb->_dom_depth == (_dom_depth + 1)) {
172 if (cb != not_null_block) continue;
173 cb = cb->_idom;
174 }
175 }
176 if( cb != this ) continue;
178 // Found a memory user; see if it can be hoisted to check-block
179 uint vidx = 0; // Capture index of value into memop
180 uint j;
181 for( j = mach->req()-1; j > 0; j-- ) {
182 if( mach->in(j) == val ) vidx = j;
183 // Block of memory-op input
184 Block *inb = cfg->_bbs[mach->in(j)->_idx];
185 Block *b = this; // Start from nul check
186 while( b != inb && b->_dom_depth > inb->_dom_depth )
187 b = b->_idom; // search upwards for input
188 // See if input dominates null check
189 if( b != inb )
190 break;
191 }
192 if( j > 0 )
193 continue;
194 Block *mb = cfg->_bbs[mach->_idx];
195 // Hoisting stores requires more checks for the anti-dependence case.
196 // Give up hoisting if we have to move the store past any load.
197 if( was_store ) {
198 Block *b = mb; // Start searching here for a local load
199 // mach use (faulting) trying to hoist
200 // n might be blocker to hoisting
201 while( b != this ) {
202 uint k;
203 for( k = 1; k < b->_nodes.size(); k++ ) {
204 Node *n = b->_nodes[k];
205 if( n->needs_anti_dependence_check() &&
206 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
207 break; // Found anti-dependent load
208 }
209 if( k < b->_nodes.size() )
210 break; // Found anti-dependent load
211 // Make sure control does not do a merge (would have to check allpaths)
212 if( b->num_preds() != 2 ) break;
213 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
214 }
215 if( b != this ) continue;
216 }
218 // Make sure this memory op is not already being used for a NullCheck
219 Node *e = mb->end();
220 if( e->is_MachNullCheck() && e->in(1) == mach )
221 continue; // Already being used as a NULL check
223 // Found a candidate! Pick one with least dom depth - the highest
224 // in the dom tree should be closest to the null check.
225 if( !best ||
226 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
227 best = mach;
228 bidx = vidx;
230 }
231 }
232 // No candidate!
233 if( !best ) return;
235 // ---- Found an implicit null check
236 extern int implicit_null_checks;
237 implicit_null_checks++;
239 // Hoist the memory candidate up to the end of the test block.
240 Block *old_block = cfg->_bbs[best->_idx];
241 old_block->find_remove(best);
242 add_inst(best);
243 cfg->_bbs.map(best->_idx,this);
245 // Move the control dependence
246 if (best->in(0) && best->in(0) == old_block->_nodes[0])
247 best->set_req(0, _nodes[0]);
249 // Check for flag-killing projections that also need to be hoisted
250 // Should be DU safe because no edge updates.
251 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
252 Node* n = best->fast_out(j);
253 if( n->Opcode() == Op_MachProj ) {
254 cfg->_bbs[n->_idx]->find_remove(n);
255 add_inst(n);
256 cfg->_bbs.map(n->_idx,this);
257 }
258 }
260 Compile *C = cfg->C;
261 // proj==Op_True --> ne test; proj==Op_False --> eq test.
262 // One of two graph shapes got matched:
263 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
264 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
265 // NULL checks are always branch-if-eq. If we see a IfTrue projection
266 // then we are replacing a 'ne' test with a 'eq' NULL check test.
267 // We need to flip the projections to keep the same semantics.
268 if( proj->Opcode() == Op_IfTrue ) {
269 // Swap order of projections in basic block to swap branch targets
270 Node *tmp1 = _nodes[end_idx()+1];
271 Node *tmp2 = _nodes[end_idx()+2];
272 _nodes.map(end_idx()+1, tmp2);
273 _nodes.map(end_idx()+2, tmp1);
274 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
275 tmp1->replace_by(tmp);
276 tmp2->replace_by(tmp1);
277 tmp->replace_by(tmp2);
278 tmp->destruct();
279 }
281 // Remove the existing null check; use a new implicit null check instead.
282 // Since schedule-local needs precise def-use info, we need to correct
283 // it as well.
284 Node *old_tst = proj->in(0);
285 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
286 _nodes.map(end_idx(),nul_chk);
287 cfg->_bbs.map(nul_chk->_idx,this);
288 // Redirect users of old_test to nul_chk
289 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
290 old_tst->last_out(i2)->set_req(0, nul_chk);
291 // Clean-up any dead code
292 for (uint i3 = 0; i3 < old_tst->req(); i3++)
293 old_tst->set_req(i3, NULL);
295 cfg->latency_from_uses(nul_chk);
296 cfg->latency_from_uses(best);
297 }
300 //------------------------------select-----------------------------------------
301 // Select a nice fellow from the worklist to schedule next. If there is only
302 // one choice, then use it. Projections take top priority for correctness
303 // reasons - if I see a projection, then it is next. There are a number of
304 // other special cases, for instructions that consume condition codes, et al.
305 // These are chosen immediately. Some instructions are required to immediately
306 // precede the last instruction in the block, and these are taken last. Of the
307 // remaining cases (most), choose the instruction with the greatest latency
308 // (that is, the most number of pseudo-cycles required to the end of the
309 // routine). If there is a tie, choose the instruction with the most inputs.
310 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) {
312 // If only a single entry on the stack, use it
313 uint cnt = worklist.size();
314 if (cnt == 1) {
315 Node *n = worklist[0];
316 worklist.map(0,worklist.pop());
317 return n;
318 }
320 uint choice = 0; // Bigger is most important
321 uint latency = 0; // Bigger is scheduled first
322 uint score = 0; // Bigger is better
323 uint idx; // Index in worklist
325 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
326 // Order in worklist is used to break ties.
327 // See caller for how this is used to delay scheduling
328 // of induction variable increments to after the other
329 // uses of the phi are scheduled.
330 Node *n = worklist[i]; // Get Node on worklist
332 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
333 if( n->is_Proj() || // Projections always win
334 n->Opcode()== Op_Con || // So does constant 'Top'
335 iop == Op_CreateEx || // Create-exception must start block
336 iop == Op_CheckCastPP
337 ) {
338 worklist.map(i,worklist.pop());
339 return n;
340 }
342 // Final call in a block must be adjacent to 'catch'
343 Node *e = end();
344 if( e->is_Catch() && e->in(0)->in(0) == n )
345 continue;
347 // Memory op for an implicit null check has to be at the end of the block
348 if( e->is_MachNullCheck() && e->in(1) == n )
349 continue;
351 uint n_choice = 2;
353 // See if this instruction is consumed by a branch. If so, then (as the
354 // branch is the last instruction in the basic block) force it to the
355 // end of the basic block
356 if ( must_clone[iop] ) {
357 // See if any use is a branch
358 bool found_machif = false;
360 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
361 Node* use = n->fast_out(j);
363 // The use is a conditional branch, make them adjacent
364 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
365 found_machif = true;
366 break;
367 }
369 // More than this instruction pending for successor to be ready,
370 // don't choose this if other opportunities are ready
371 if (ready_cnt[use->_idx] > 1)
372 n_choice = 1;
373 }
375 // loop terminated, prefer not to use this instruction
376 if (found_machif)
377 continue;
378 }
380 // See if this has a predecessor that is "must_clone", i.e. sets the
381 // condition code. If so, choose this first
382 for (uint j = 0; j < n->req() ; j++) {
383 Node *inn = n->in(j);
384 if (inn) {
385 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
386 n_choice = 3;
387 break;
388 }
389 }
390 }
392 // MachTemps should be scheduled last so they are near their uses
393 if (n->is_MachTemp()) {
394 n_choice = 1;
395 }
397 uint n_latency = cfg->_node_latency.at_grow(n->_idx);
398 uint n_score = n->req(); // Many inputs get high score to break ties
400 // Keep best latency found
401 if( choice < n_choice ||
402 ( choice == n_choice &&
403 ( latency < n_latency ||
404 ( latency == n_latency &&
405 ( score < n_score ))))) {
406 choice = n_choice;
407 latency = n_latency;
408 score = n_score;
409 idx = i; // Also keep index in worklist
410 }
411 } // End of for all ready nodes in worklist
413 Node *n = worklist[idx]; // Get the winner
415 worklist.map(idx,worklist.pop()); // Compress worklist
416 return n;
417 }
420 //------------------------------set_next_call----------------------------------
421 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
422 if( next_call.test_set(n->_idx) ) return;
423 for( uint i=0; i<n->len(); i++ ) {
424 Node *m = n->in(i);
425 if( !m ) continue; // must see all nodes in block that precede call
426 if( bbs[m->_idx] == this )
427 set_next_call( m, next_call, bbs );
428 }
429 }
431 //------------------------------needed_for_next_call---------------------------
432 // Set the flag 'next_call' for each Node that is needed for the next call to
433 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
434 // next subroutine call get priority - basically it moves things NOT needed
435 // for the next call till after the call. This prevents me from trying to
436 // carry lots of stuff live across a call.
437 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
438 // Find the next control-defining Node in this block
439 Node* call = NULL;
440 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
441 Node* m = this_call->fast_out(i);
442 if( bbs[m->_idx] == this && // Local-block user
443 m != this_call && // Not self-start node
444 m->is_Call() )
445 call = m;
446 break;
447 }
448 if (call == NULL) return; // No next call (e.g., block end is near)
449 // Set next-call for all inputs to this call
450 set_next_call(call, next_call, bbs);
451 }
453 //------------------------------sched_call-------------------------------------
454 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
455 RegMask regs;
457 // Schedule all the users of the call right now. All the users are
458 // projection Nodes, so they must be scheduled next to the call.
459 // Collect all the defined registers.
460 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
461 Node* n = mcall->fast_out(i);
462 assert( n->Opcode()==Op_MachProj, "" );
463 --ready_cnt[n->_idx];
464 assert( !ready_cnt[n->_idx], "" );
465 // Schedule next to call
466 _nodes.map(node_cnt++, n);
467 // Collect defined registers
468 regs.OR(n->out_RegMask());
469 // Check for scheduling the next control-definer
470 if( n->bottom_type() == Type::CONTROL )
471 // Warm up next pile of heuristic bits
472 needed_for_next_call(n, next_call, bbs);
474 // Children of projections are now all ready
475 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
476 Node* m = n->fast_out(j); // Get user
477 if( bbs[m->_idx] != this ) continue;
478 if( m->is_Phi() ) continue;
479 if( !--ready_cnt[m->_idx] )
480 worklist.push(m);
481 }
483 }
485 // Act as if the call defines the Frame Pointer.
486 // Certainly the FP is alive and well after the call.
487 regs.Insert(matcher.c_frame_pointer());
489 // Set all registers killed and not already defined by the call.
490 uint r_cnt = mcall->tf()->range()->cnt();
491 int op = mcall->ideal_Opcode();
492 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
493 bbs.map(proj->_idx,this);
494 _nodes.insert(node_cnt++, proj);
496 // Select the right register save policy.
497 const char * save_policy;
498 switch (op) {
499 case Op_CallRuntime:
500 case Op_CallLeaf:
501 case Op_CallLeafNoFP:
502 // Calling C code so use C calling convention
503 save_policy = matcher._c_reg_save_policy;
504 break;
506 case Op_CallStaticJava:
507 case Op_CallDynamicJava:
508 // Calling Java code so use Java calling convention
509 save_policy = matcher._register_save_policy;
510 break;
512 default:
513 ShouldNotReachHere();
514 }
516 // When using CallRuntime mark SOE registers as killed by the call
517 // so values that could show up in the RegisterMap aren't live in a
518 // callee saved register since the register wouldn't know where to
519 // find them. CallLeaf and CallLeafNoFP are ok because they can't
520 // have debug info on them. Strictly speaking this only needs to be
521 // done for oops since idealreg2debugmask takes care of debug info
522 // references but there no way to handle oops differently than other
523 // pointers as far as the kill mask goes.
524 bool exclude_soe = op == Op_CallRuntime;
526 // Fill in the kill mask for the call
527 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
528 if( !regs.Member(r) ) { // Not already defined by the call
529 // Save-on-call register?
530 if ((save_policy[r] == 'C') ||
531 (save_policy[r] == 'A') ||
532 ((save_policy[r] == 'E') && exclude_soe)) {
533 proj->_rout.Insert(r);
534 }
535 }
536 }
538 return node_cnt;
539 }
542 //------------------------------schedule_local---------------------------------
543 // Topological sort within a block. Someday become a real scheduler.
544 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) {
545 // Already "sorted" are the block start Node (as the first entry), and
546 // the block-ending Node and any trailing control projections. We leave
547 // these alone. PhiNodes and ParmNodes are made to follow the block start
548 // Node. Everything else gets topo-sorted.
550 #ifndef PRODUCT
551 if (cfg->trace_opto_pipelining()) {
552 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
553 for (uint i = 0;i < _nodes.size();i++) {
554 tty->print("# ");
555 _nodes[i]->fast_dump();
556 }
557 tty->print_cr("#");
558 }
559 #endif
561 // RootNode is already sorted
562 if( _nodes.size() == 1 ) return true;
564 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
565 uint node_cnt = end_idx();
566 uint phi_cnt = 1;
567 uint i;
568 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
569 Node *n = _nodes[i];
570 if( n->is_Phi() || // Found a PhiNode or ParmNode
571 (n->is_Proj() && n->in(0) == head()) ) {
572 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
573 _nodes.map(i,_nodes[phi_cnt]);
574 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
575 } else { // All others
576 // Count block-local inputs to 'n'
577 uint cnt = n->len(); // Input count
578 uint local = 0;
579 for( uint j=0; j<cnt; j++ ) {
580 Node *m = n->in(j);
581 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
582 local++; // One more block-local input
583 }
584 ready_cnt[n->_idx] = local; // Count em up
586 // A few node types require changing a required edge to a precedence edge
587 // before allocation.
588 if( UseConcMarkSweepGC ) {
589 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
590 // Note: Required edges with an index greater than oper_input_base
591 // are not supported by the allocator.
592 // Note2: Can only depend on unmatched edge being last,
593 // can not depend on its absolute position.
594 Node *oop_store = n->in(n->req() - 1);
595 n->del_req(n->req() - 1);
596 n->add_prec(oop_store);
597 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
598 }
599 }
600 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ) {
601 Node *x = n->in(TypeFunc::Parms);
602 n->del_req(TypeFunc::Parms);
603 n->add_prec(x);
604 }
605 }
606 }
607 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
608 ready_cnt[_nodes[i2]->_idx] = 0;
610 // All the prescheduled guys do not hold back internal nodes
611 uint i3;
612 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
613 Node *n = _nodes[i3]; // Get pre-scheduled
614 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
615 Node* m = n->fast_out(j);
616 if( cfg->_bbs[m->_idx] ==this ) // Local-block user
617 ready_cnt[m->_idx]--; // Fix ready count
618 }
619 }
621 Node_List delay;
622 // Make a worklist
623 Node_List worklist;
624 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
625 Node *m = _nodes[i4];
626 if( !ready_cnt[m->_idx] ) { // Zero ready count?
627 if (m->is_iteratively_computed()) {
628 // Push induction variable increments last to allow other uses
629 // of the phi to be scheduled first. The select() method breaks
630 // ties in scheduling by worklist order.
631 delay.push(m);
632 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
633 // Force the CreateEx to the top of the list so it's processed
634 // first and ends up at the start of the block.
635 worklist.insert(0, m);
636 } else {
637 worklist.push(m); // Then on to worklist!
638 }
639 }
640 }
641 while (delay.size()) {
642 Node* d = delay.pop();
643 worklist.push(d);
644 }
646 // Warm up the 'next_call' heuristic bits
647 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
649 #ifndef PRODUCT
650 if (cfg->trace_opto_pipelining()) {
651 for (uint j=0; j<_nodes.size(); j++) {
652 Node *n = _nodes[j];
653 int idx = n->_idx;
654 tty->print("# ready cnt:%3d ", ready_cnt[idx]);
655 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx));
656 tty->print("%4d: %s\n", idx, n->Name());
657 }
658 }
659 #endif
661 // Pull from worklist and schedule
662 while( worklist.size() ) { // Worklist is not ready
664 #ifndef PRODUCT
665 if (cfg->trace_opto_pipelining()) {
666 tty->print("# ready list:");
667 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
668 Node *n = worklist[i]; // Get Node on worklist
669 tty->print(" %d", n->_idx);
670 }
671 tty->cr();
672 }
673 #endif
675 // Select and pop a ready guy from worklist
676 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
677 _nodes.map(phi_cnt++,n); // Schedule him next
679 #ifndef PRODUCT
680 if (cfg->trace_opto_pipelining()) {
681 tty->print("# select %d: %s", n->_idx, n->Name());
682 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx));
683 n->dump();
684 if (Verbose) {
685 tty->print("# ready list:");
686 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
687 Node *n = worklist[i]; // Get Node on worklist
688 tty->print(" %d", n->_idx);
689 }
690 tty->cr();
691 }
692 }
694 #endif
695 if( n->is_MachCall() ) {
696 MachCallNode *mcall = n->as_MachCall();
697 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
698 continue;
699 }
700 // Children are now all ready
701 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
702 Node* m = n->fast_out(i5); // Get user
703 if( cfg->_bbs[m->_idx] != this ) continue;
704 if( m->is_Phi() ) continue;
705 if( !--ready_cnt[m->_idx] )
706 worklist.push(m);
707 }
708 }
710 if( phi_cnt != end_idx() ) {
711 // did not schedule all. Retry, Bailout, or Die
712 Compile* C = matcher.C;
713 if (C->subsume_loads() == true && !C->failing()) {
714 // Retry with subsume_loads == false
715 // If this is the first failure, the sentinel string will "stick"
716 // to the Compile object, and the C2Compiler will see it and retry.
717 C->record_failure(C2Compiler::retry_no_subsuming_loads());
718 }
719 // assert( phi_cnt == end_idx(), "did not schedule all" );
720 return false;
721 }
723 #ifndef PRODUCT
724 if (cfg->trace_opto_pipelining()) {
725 tty->print_cr("#");
726 tty->print_cr("# after schedule_local");
727 for (uint i = 0;i < _nodes.size();i++) {
728 tty->print("# ");
729 _nodes[i]->fast_dump();
730 }
731 tty->cr();
732 }
733 #endif
736 return true;
737 }
739 //--------------------------catch_cleanup_fix_all_inputs-----------------------
740 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
741 for (uint l = 0; l < use->len(); l++) {
742 if (use->in(l) == old_def) {
743 if (l < use->req()) {
744 use->set_req(l, new_def);
745 } else {
746 use->rm_prec(l);
747 use->add_prec(new_def);
748 l--;
749 }
750 }
751 }
752 }
754 //------------------------------catch_cleanup_find_cloned_def------------------
755 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
756 assert( use_blk != def_blk, "Inter-block cleanup only");
758 // The use is some block below the Catch. Find and return the clone of the def
759 // that dominates the use. If there is no clone in a dominating block, then
760 // create a phi for the def in a dominating block.
762 // Find which successor block dominates this use. The successor
763 // blocks must all be single-entry (from the Catch only; I will have
764 // split blocks to make this so), hence they all dominate.
765 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
766 use_blk = use_blk->_idom;
768 // Find the successor
769 Node *fixup = NULL;
771 uint j;
772 for( j = 0; j < def_blk->_num_succs; j++ )
773 if( use_blk == def_blk->_succs[j] )
774 break;
776 if( j == def_blk->_num_succs ) {
777 // Block at same level in dom-tree is not a successor. It needs a
778 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
779 Node_Array inputs = new Node_List(Thread::current()->resource_area());
780 for(uint k = 1; k < use_blk->num_preds(); k++) {
781 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
782 }
784 // Check to see if the use_blk already has an identical phi inserted.
785 // If it exists, it will be at the first position since all uses of a
786 // def are processed together.
787 Node *phi = use_blk->_nodes[1];
788 if( phi->is_Phi() ) {
789 fixup = phi;
790 for (uint k = 1; k < use_blk->num_preds(); k++) {
791 if (phi->in(k) != inputs[k]) {
792 // Not a match
793 fixup = NULL;
794 break;
795 }
796 }
797 }
799 // If an existing PhiNode was not found, make a new one.
800 if (fixup == NULL) {
801 Node *new_phi = PhiNode::make(use_blk->head(), def);
802 use_blk->_nodes.insert(1, new_phi);
803 bbs.map(new_phi->_idx, use_blk);
804 for (uint k = 1; k < use_blk->num_preds(); k++) {
805 new_phi->set_req(k, inputs[k]);
806 }
807 fixup = new_phi;
808 }
810 } else {
811 // Found the use just below the Catch. Make it use the clone.
812 fixup = use_blk->_nodes[n_clone_idx];
813 }
815 return fixup;
816 }
818 //--------------------------catch_cleanup_intra_block--------------------------
819 // Fix all input edges in use that reference "def". The use is in the same
820 // block as the def and both have been cloned in each successor block.
821 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
823 // Both the use and def have been cloned. For each successor block,
824 // get the clone of the use, and make its input the clone of the def
825 // found in that block.
827 uint use_idx = blk->find_node(use);
828 uint offset_idx = use_idx - beg;
829 for( uint k = 0; k < blk->_num_succs; k++ ) {
830 // Get clone in each successor block
831 Block *sb = blk->_succs[k];
832 Node *clone = sb->_nodes[offset_idx+1];
833 assert( clone->Opcode() == use->Opcode(), "" );
835 // Make use-clone reference the def-clone
836 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
837 }
838 }
840 //------------------------------catch_cleanup_inter_block---------------------
841 // Fix all input edges in use that reference "def". The use is in a different
842 // block than the def.
843 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
844 if( !use_blk ) return; // Can happen if the use is a precedence edge
846 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
847 catch_cleanup_fix_all_inputs(use, def, new_def);
848 }
850 //------------------------------call_catch_cleanup-----------------------------
851 // If we inserted any instructions between a Call and his CatchNode,
852 // clone the instructions on all paths below the Catch.
853 void Block::call_catch_cleanup(Block_Array &bbs) {
855 // End of region to clone
856 uint end = end_idx();
857 if( !_nodes[end]->is_Catch() ) return;
858 // Start of region to clone
859 uint beg = end;
860 while( _nodes[beg-1]->Opcode() != Op_MachProj ||
861 !_nodes[beg-1]->in(0)->is_Call() ) {
862 beg--;
863 assert(beg > 0,"Catch cleanup walking beyond block boundary");
864 }
865 // Range of inserted instructions is [beg, end)
866 if( beg == end ) return;
868 // Clone along all Catch output paths. Clone area between the 'beg' and
869 // 'end' indices.
870 for( uint i = 0; i < _num_succs; i++ ) {
871 Block *sb = _succs[i];
872 // Clone the entire area; ignoring the edge fixup for now.
873 for( uint j = end; j > beg; j-- ) {
874 Node *clone = _nodes[j-1]->clone();
875 sb->_nodes.insert( 1, clone );
876 bbs.map(clone->_idx,sb);
877 }
878 }
881 // Fixup edges. Check the def-use info per cloned Node
882 for(uint i2 = beg; i2 < end; i2++ ) {
883 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
884 Node *n = _nodes[i2]; // Node that got cloned
885 // Need DU safe iterator because of edge manipulation in calls.
886 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
887 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
888 out->push(n->fast_out(j1));
889 }
890 uint max = out->size();
891 for (uint j = 0; j < max; j++) {// For all users
892 Node *use = out->pop();
893 Block *buse = bbs[use->_idx];
894 if( use->is_Phi() ) {
895 for( uint k = 1; k < use->req(); k++ )
896 if( use->in(k) == n ) {
897 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
898 use->set_req(k, fixup);
899 }
900 } else {
901 if (this == buse) {
902 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
903 } else {
904 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
905 }
906 }
907 } // End for all users
909 } // End of for all Nodes in cloned area
911 // Remove the now-dead cloned ops
912 for(uint i3 = beg; i3 < end; i3++ ) {
913 _nodes[beg]->disconnect_inputs(NULL);
914 _nodes.remove(beg);
915 }
917 // If the successor blocks have a CreateEx node, move it back to the top
918 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
919 Block *sb = _succs[i4];
920 uint new_cnt = end - beg;
921 // Remove any newly created, but dead, nodes.
922 for( uint j = new_cnt; j > 0; j-- ) {
923 Node *n = sb->_nodes[j];
924 if (n->outcnt() == 0 &&
925 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
926 n->disconnect_inputs(NULL);
927 sb->_nodes.remove(j);
928 new_cnt--;
929 }
930 }
931 // If any newly created nodes remain, move the CreateEx node to the top
932 if (new_cnt > 0) {
933 Node *cex = sb->_nodes[1+new_cnt];
934 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
935 sb->_nodes.remove(1+new_cnt);
936 sb->_nodes.insert(1,cex);
937 }
938 }
939 }
940 }