1 /*
2 * Copyright (c) 1998, 2015, 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 "memory/allocation.inline.hpp"
27 #include "opto/ad.hpp"
28 #include "opto/block.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/runtime.hpp"
34 #include "opto/chaitin.hpp"
35 #include "runtime/sharedRuntime.hpp"
36
37 // Optimization - Graph Style
38
39 // Check whether val is not-null-decoded compressed oop,
40 // i.e. will grab into the base of the heap if it represents NULL.
41 static bool accesses_heap_base_zone(Node *val) {
42 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops.
43 if (val && val->is_Mach()) {
44 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
45 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
46 // decode NULL to point to the heap base (Decode_NN).
47 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
48 return true;
49 }
50 }
51 // Must recognize load operation with Decode matched in memory operand.
52 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected()
53 // returns true everywhere else. On PPC, no such memory operands
54 // exist, therefore we did not yet implement a check for such operands.
55 NOT_AIX(Unimplemented());
56 }
57 }
58 return false;
59 }
60
61 static bool needs_explicit_null_check_for_read(Node *val) {
62 // On some OSes (AIX) the page at address 0 is only write protected.
63 // If so, only Store operations will trap.
64 if (os::zero_page_read_protected()) {
65 return false; // Implicit null check will work.
66 }
67 // Also a read accessing the base of a heap-based compressed heap will trap.
68 if (accesses_heap_base_zone(val) && // Hits the base zone page.
69 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected.
70 return false;
71 }
72
73 return true;
74 }
75
76 //------------------------------implicit_null_check----------------------------
77 // Detect implicit-null-check opportunities. Basically, find NULL checks
78 // with suitable memory ops nearby. Use the memory op to do the NULL check.
79 // I can generate a memory op if there is not one nearby.
80 // The proj is the control projection for the not-null case.
81 // The val is the pointer being checked for nullness or
82 // decodeHeapOop_not_null node if it did not fold into address.
83 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
84 // Assume if null check need for 0 offset then always needed
85 // Intel solaris doesn't support any null checks yet and no
86 // mechanism exists (yet) to set the switches at an os_cpu level
87 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
88
89 // Make sure the ptr-is-null path appears to be uncommon!
90 float f = block->end()->as_MachIf()->_prob;
91 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
92 if( f > PROB_UNLIKELY_MAG(4) ) return;
93
94 uint bidx = 0; // Capture index of value into memop
95 bool was_store; // Memory op is a store op
96
97 // Get the successor block for if the test ptr is non-null
98 Block* not_null_block; // this one goes with the proj
99 Block* null_block;
100 if (block->get_node(block->number_of_nodes()-1) == proj) {
101 null_block = block->_succs[0];
102 not_null_block = block->_succs[1];
103 } else {
104 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
105 not_null_block = block->_succs[0];
106 null_block = block->_succs[1];
107 }
108 while (null_block->is_Empty() == Block::empty_with_goto) {
109 null_block = null_block->_succs[0];
110 }
111
112 // Search the exception block for an uncommon trap.
113 // (See Parse::do_if and Parse::do_ifnull for the reason
114 // we need an uncommon trap. Briefly, we need a way to
115 // detect failure of this optimization, as in 6366351.)
116 {
117 bool found_trap = false;
118 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
119 Node* nn = null_block->get_node(i1);
120 if (nn->is_MachCall() &&
121 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
122 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
123 if (trtype->isa_int() && trtype->is_int()->is_con()) {
124 jint tr_con = trtype->is_int()->get_con();
125 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
126 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
127 assert((int)reason < (int)BitsPerInt, "recode bit map");
128 if (is_set_nth_bit(allowed_reasons, (int) reason)
129 && action != Deoptimization::Action_none) {
130 // This uncommon trap is sure to recompile, eventually.
131 // When that happens, C->too_many_traps will prevent
132 // this transformation from happening again.
133 found_trap = true;
134 }
135 }
136 break;
137 }
138 }
139 if (!found_trap) {
140 // We did not find an uncommon trap.
141 return;
142 }
143 }
144
145 // Check for decodeHeapOop_not_null node which did not fold into address
146 bool is_decoden = ((intptr_t)val) & 1;
147 val = (Node*)(((intptr_t)val) & ~1);
148
149 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
150 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
151
152 // Search the successor block for a load or store who's base value is also
153 // the tested value. There may be several.
154 Node_List *out = new Node_List(Thread::current()->resource_area());
155 MachNode *best = NULL; // Best found so far
156 for (DUIterator i = val->outs(); val->has_out(i); i++) {
157 Node *m = val->out(i);
158 if( !m->is_Mach() ) continue;
159 MachNode *mach = m->as_Mach();
160 was_store = false;
161 int iop = mach->ideal_Opcode();
162 switch( iop ) {
163 case Op_LoadB:
164 case Op_LoadUB:
165 case Op_LoadUS:
166 case Op_LoadD:
167 case Op_LoadF:
168 case Op_LoadI:
169 case Op_LoadL:
170 case Op_LoadP:
171 case Op_LoadN:
172 case Op_LoadS:
173 case Op_LoadKlass:
174 case Op_LoadNKlass:
175 case Op_LoadRange:
176 case Op_LoadD_unaligned:
177 case Op_LoadL_unaligned:
178 assert(mach->in(2) == val, "should be address");
179 break;
180 case Op_StoreB:
181 case Op_StoreC:
182 case Op_StoreCM:
183 case Op_StoreD:
184 case Op_StoreF:
185 case Op_StoreI:
186 case Op_StoreL:
187 case Op_StoreP:
188 case Op_StoreN:
189 case Op_StoreNKlass:
190 was_store = true; // Memory op is a store op
191 // Stores will have their address in slot 2 (memory in slot 1).
192 // If the value being nul-checked is in another slot, it means we
193 // are storing the checked value, which does NOT check the value!
194 if( mach->in(2) != val ) continue;
195 break; // Found a memory op?
196 case Op_StrComp:
197 case Op_StrEquals:
198 case Op_StrIndexOf:
199 case Op_StrIndexOfChar:
200 case Op_AryEq:
201 case Op_StrInflatedCopy:
202 case Op_StrCompressedCopy:
203 case Op_EncodeISOArray:
204 case Op_HasNegatives:
205 // Not a legit memory op for implicit null check regardless of
206 // embedded loads
207 continue;
208 default: // Also check for embedded loads
209 if( !mach->needs_anti_dependence_check() )
210 continue; // Not an memory op; skip it
211 if( must_clone[iop] ) {
212 // Do not move nodes which produce flags because
213 // RA will try to clone it to place near branch and
214 // it will cause recompilation, see clone_node().
215 continue;
216 }
217 {
218 // Check that value is used in memory address in
219 // instructions with embedded load (CmpP val1,(val2+off)).
220 Node* base;
221 Node* index;
222 const MachOper* oper = mach->memory_inputs(base, index);
223 if (oper == NULL || oper == (MachOper*)-1) {
224 continue; // Not an memory op; skip it
225 }
226 if (val == base ||
227 val == index && val->bottom_type()->isa_narrowoop()) {
228 break; // Found it
229 } else {
230 continue; // Skip it
231 }
232 }
233 break;
234 }
235
236 // On some OSes (AIX) the page at address 0 is only write protected.
237 // If so, only Store operations will trap.
238 // But a read accessing the base of a heap-based compressed heap will trap.
239 if (!was_store && needs_explicit_null_check_for_read(val)) {
240 continue;
241 }
242
243 // check if the offset is not too high for implicit exception
244 {
245 intptr_t offset = 0;
246 const TypePtr *adr_type = NULL; // Do not need this return value here
247 const Node* base = mach->get_base_and_disp(offset, adr_type);
248 if (base == NULL || base == NodeSentinel) {
249 // Narrow oop address doesn't have base, only index
250 if( val->bottom_type()->isa_narrowoop() &&
251 MacroAssembler::needs_explicit_null_check(offset) )
252 continue; // Give up if offset is beyond page size
253 // cannot reason about it; is probably not implicit null exception
254 } else {
255 const TypePtr* tptr;
256 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
257 Universe::narrow_klass_shift() == 0)) {
258 // 32-bits narrow oop can be the base of address expressions
259 tptr = base->get_ptr_type();
260 } else {
261 // only regular oops are expected here
262 tptr = base->bottom_type()->is_ptr();
263 }
264 // Give up if offset is not a compile-time constant
265 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
266 continue;
267 offset += tptr->_offset; // correct if base is offseted
268 if( MacroAssembler::needs_explicit_null_check(offset) )
269 continue; // Give up is reference is beyond 4K page size
270 }
271 }
272
273 // Check ctrl input to see if the null-check dominates the memory op
274 Block *cb = get_block_for_node(mach);
275 cb = cb->_idom; // Always hoist at least 1 block
276 if( !was_store ) { // Stores can be hoisted only one block
277 while( cb->_dom_depth > (block->_dom_depth + 1))
278 cb = cb->_idom; // Hoist loads as far as we want
279 // The non-null-block should dominate the memory op, too. Live
280 // range spilling will insert a spill in the non-null-block if it is
281 // needs to spill the memory op for an implicit null check.
282 if (cb->_dom_depth == (block->_dom_depth + 1)) {
283 if (cb != not_null_block) continue;
284 cb = cb->_idom;
285 }
286 }
287 if( cb != block ) continue;
288
289 // Found a memory user; see if it can be hoisted to check-block
290 uint vidx = 0; // Capture index of value into memop
291 uint j;
292 for( j = mach->req()-1; j > 0; j-- ) {
293 if( mach->in(j) == val ) {
294 vidx = j;
295 // Ignore DecodeN val which could be hoisted to where needed.
296 if( is_decoden ) continue;
297 }
298 // Block of memory-op input
299 Block *inb = get_block_for_node(mach->in(j));
300 Block *b = block; // Start from nul check
301 while( b != inb && b->_dom_depth > inb->_dom_depth )
302 b = b->_idom; // search upwards for input
303 // See if input dominates null check
304 if( b != inb )
305 break;
306 }
307 if( j > 0 )
308 continue;
309 Block *mb = get_block_for_node(mach);
310 // Hoisting stores requires more checks for the anti-dependence case.
311 // Give up hoisting if we have to move the store past any load.
312 if( was_store ) {
313 Block *b = mb; // Start searching here for a local load
314 // mach use (faulting) trying to hoist
315 // n might be blocker to hoisting
316 while( b != block ) {
317 uint k;
318 for( k = 1; k < b->number_of_nodes(); k++ ) {
319 Node *n = b->get_node(k);
320 if( n->needs_anti_dependence_check() &&
321 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
322 break; // Found anti-dependent load
323 }
324 if( k < b->number_of_nodes() )
325 break; // Found anti-dependent load
326 // Make sure control does not do a merge (would have to check allpaths)
327 if( b->num_preds() != 2 ) break;
328 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
329 }
330 if( b != block ) continue;
331 }
332
333 // Make sure this memory op is not already being used for a NullCheck
334 Node *e = mb->end();
335 if( e->is_MachNullCheck() && e->in(1) == mach )
336 continue; // Already being used as a NULL check
337
338 // Found a candidate! Pick one with least dom depth - the highest
339 // in the dom tree should be closest to the null check.
340 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
341 best = mach;
342 bidx = vidx;
343 }
344 }
345 // No candidate!
346 if (best == NULL) {
347 return;
348 }
349
350 // ---- Found an implicit null check
351 extern int implicit_null_checks;
352 implicit_null_checks++;
353
354 if( is_decoden ) {
355 // Check if we need to hoist decodeHeapOop_not_null first.
356 Block *valb = get_block_for_node(val);
357 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
358 // Hoist it up to the end of the test block.
359 valb->find_remove(val);
360 block->add_inst(val);
361 map_node_to_block(val, block);
362 // DecodeN on x86 may kill flags. Check for flag-killing projections
363 // that also need to be hoisted.
364 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
365 Node* n = val->fast_out(j);
366 if( n->is_MachProj() ) {
367 get_block_for_node(n)->find_remove(n);
368 block->add_inst(n);
369 map_node_to_block(n, block);
370 }
371 }
372 }
373 }
374 // Hoist the memory candidate up to the end of the test block.
375 Block *old_block = get_block_for_node(best);
376 old_block->find_remove(best);
377 block->add_inst(best);
378 map_node_to_block(best, block);
379
380 // Move the control dependence
381 if (best->in(0) && best->in(0) == old_block->head())
382 best->set_req(0, block->head());
383
384 // Check for flag-killing projections that also need to be hoisted
385 // Should be DU safe because no edge updates.
386 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
387 Node* n = best->fast_out(j);
388 if( n->is_MachProj() ) {
389 get_block_for_node(n)->find_remove(n);
390 block->add_inst(n);
391 map_node_to_block(n, block);
392 }
393 }
394
395 // proj==Op_True --> ne test; proj==Op_False --> eq test.
396 // One of two graph shapes got matched:
397 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
398 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
399 // NULL checks are always branch-if-eq. If we see a IfTrue projection
400 // then we are replacing a 'ne' test with a 'eq' NULL check test.
401 // We need to flip the projections to keep the same semantics.
402 if( proj->Opcode() == Op_IfTrue ) {
403 // Swap order of projections in basic block to swap branch targets
404 Node *tmp1 = block->get_node(block->end_idx()+1);
405 Node *tmp2 = block->get_node(block->end_idx()+2);
406 block->map_node(tmp2, block->end_idx()+1);
407 block->map_node(tmp1, block->end_idx()+2);
408 Node *tmp = new Node(C->top()); // Use not NULL input
409 tmp1->replace_by(tmp);
410 tmp2->replace_by(tmp1);
411 tmp->replace_by(tmp2);
412 tmp->destruct();
413 }
414
415 // Remove the existing null check; use a new implicit null check instead.
416 // Since schedule-local needs precise def-use info, we need to correct
417 // it as well.
418 Node *old_tst = proj->in(0);
419 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx);
420 block->map_node(nul_chk, block->end_idx());
421 map_node_to_block(nul_chk, block);
422 // Redirect users of old_test to nul_chk
423 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
424 old_tst->last_out(i2)->set_req(0, nul_chk);
425 // Clean-up any dead code
426 for (uint i3 = 0; i3 < old_tst->req(); i3++) {
427 Node* in = old_tst->in(i3);
428 old_tst->set_req(i3, NULL);
429 if (in->outcnt() == 0) {
430 // Remove dead input node
431 in->disconnect_inputs(NULL, C);
432 block->find_remove(in);
433 }
434 }
435
436 latency_from_uses(nul_chk);
437 latency_from_uses(best);
438 }
439
440
441 //------------------------------select-----------------------------------------
442 // Select a nice fellow from the worklist to schedule next. If there is only
443 // one choice, then use it. Projections take top priority for correctness
444 // reasons - if I see a projection, then it is next. There are a number of
445 // other special cases, for instructions that consume condition codes, et al.
446 // These are chosen immediately. Some instructions are required to immediately
447 // precede the last instruction in the block, and these are taken last. Of the
448 // remaining cases (most), choose the instruction with the greatest latency
449 // (that is, the most number of pseudo-cycles required to the end of the
450 // routine). If there is a tie, choose the instruction with the most inputs.
451 Node* PhaseCFG::select(
452 Block* block,
453 Node_List &worklist,
454 GrowableArray<int> &ready_cnt,
455 VectorSet &next_call,
456 uint sched_slot,
457 intptr_t* recalc_pressure_nodes) {
458
459 // If only a single entry on the stack, use it
460 uint cnt = worklist.size();
461 if (cnt == 1) {
462 Node *n = worklist[0];
463 worklist.map(0,worklist.pop());
464 return n;
465 }
466
467 uint choice = 0; // Bigger is most important
468 uint latency = 0; // Bigger is scheduled first
469 uint score = 0; // Bigger is better
470 int idx = -1; // Index in worklist
471 int cand_cnt = 0; // Candidate count
472 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false;
473
474 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
475 // Order in worklist is used to break ties.
476 // See caller for how this is used to delay scheduling
477 // of induction variable increments to after the other
478 // uses of the phi are scheduled.
479 Node *n = worklist[i]; // Get Node on worklist
480
481 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
482 if( n->is_Proj() || // Projections always win
483 n->Opcode()== Op_Con || // So does constant 'Top'
484 iop == Op_CreateEx || // Create-exception must start block
485 iop == Op_CheckCastPP
486 ) {
487 worklist.map(i,worklist.pop());
488 return n;
489 }
490
491 // Final call in a block must be adjacent to 'catch'
492 Node *e = block->end();
493 if( e->is_Catch() && e->in(0)->in(0) == n )
494 continue;
495
496 // Memory op for an implicit null check has to be at the end of the block
497 if( e->is_MachNullCheck() && e->in(1) == n )
498 continue;
499
500 // Schedule IV increment last.
501 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
502 e->in(1)->req() > 1 && e->in(1)->in(1) == n &&
503 n->is_iteratively_computed()) {
504 continue;
505 }
506
507 uint n_choice = 2;
508
509 // See if this instruction is consumed by a branch. If so, then (as the
510 // branch is the last instruction in the basic block) force it to the
511 // end of the basic block
512 if ( must_clone[iop] ) {
513 // See if any use is a branch
514 bool found_machif = false;
515
516 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
517 Node* use = n->fast_out(j);
518
519 // The use is a conditional branch, make them adjacent
520 if (use->is_MachIf() && get_block_for_node(use) == block) {
521 found_machif = true;
522 break;
523 }
524
525 // More than this instruction pending for successor to be ready,
526 // don't choose this if other opportunities are ready
527 if (ready_cnt.at(use->_idx) > 1)
528 n_choice = 1;
529 }
530
531 // loop terminated, prefer not to use this instruction
532 if (found_machif)
533 continue;
534 }
535
536 // See if this has a predecessor that is "must_clone", i.e. sets the
537 // condition code. If so, choose this first
538 for (uint j = 0; j < n->req() ; j++) {
539 Node *inn = n->in(j);
540 if (inn) {
541 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
542 n_choice = 3;
543 break;
544 }
545 }
546 }
547
548 // MachTemps should be scheduled last so they are near their uses
549 if (n->is_MachTemp()) {
550 n_choice = 1;
551 }
552
553 uint n_latency = get_latency_for_node(n);
554 uint n_score = n->req(); // Many inputs get high score to break ties
555
556 if (OptoRegScheduling && block_size_threshold_ok) {
557 if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) {
558 _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit());
559 _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit());
560 // simulate the notion that we just picked this node to schedule
561 n->add_flag(Node::Flag_is_scheduled);
562 // now caculate its effect upon the graph if we did
563 adjust_register_pressure(n, block, recalc_pressure_nodes, false);
564 // return its state for finalize in case somebody else wins
565 n->remove_flag(Node::Flag_is_scheduled);
566 // now save the two final pressure components of register pressure, limiting pressure calcs to short size
567 short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure();
568 short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure();
569 recalc_pressure_nodes[n->_idx] = int_pressure;
570 recalc_pressure_nodes[n->_idx] |= (float_pressure << 16);
571 }
572
573 if (_scheduling_for_pressure) {
574 latency = n_latency;
575 if (n_choice != 3) {
576 // Now evaluate each register pressure component based on threshold in the score.
577 // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks
578 // on a single instruction, but we might see it shrink on both banks.
579 // For each use of register that has a register class that is over the high pressure limit, we build n_score up for
580 // live ranges that terminate on this instruction.
581 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
582 short int_pressure = (short)recalc_pressure_nodes[n->_idx];
583 n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score;
584 }
585 if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
586 short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16);
587 n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score;
588 }
589 } else {
590 // make sure we choose these candidates
591 score = 0;
592 }
593 }
594 }
595
596 // Keep best latency found
597 cand_cnt++;
598 if (choice < n_choice ||
599 (choice == n_choice &&
600 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
601 (!StressLCM &&
602 (latency < n_latency ||
603 (latency == n_latency &&
604 (score < n_score))))))) {
605 choice = n_choice;
606 latency = n_latency;
607 score = n_score;
608 idx = i; // Also keep index in worklist
609 }
610 } // End of for all ready nodes in worklist
611
612 assert(idx >= 0, "index should be set");
613 Node *n = worklist[(uint)idx]; // Get the winner
614
615 worklist.map((uint)idx, worklist.pop()); // Compress worklist
616 return n;
617 }
618
619 //-------------------------adjust_register_pressure----------------------------
620 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) {
621 PhaseLive* liveinfo = _regalloc->get_live();
622 IndexSet* liveout = liveinfo->live(block);
623 // first adjust the register pressure for the sources
624 for (uint i = 1; i < n->req(); i++) {
625 bool lrg_ends = false;
626 Node *src_n = n->in(i);
627 if (src_n == NULL) continue;
628 if (!src_n->is_Mach()) continue;
629 uint src = _regalloc->_lrg_map.find(src_n);
630 if (src == 0) continue;
631 LRG& lrg_src = _regalloc->lrgs(src);
632 // detect if the live range ends or not
633 if (liveout->member(src) == false) {
634 lrg_ends = true;
635 for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) {
636 Node* m = src_n->fast_out(j); // Get user
637 if (m == n) continue;
638 if (!m->is_Mach()) continue;
639 MachNode *mach = m->as_Mach();
640 bool src_matches = false;
641 int iop = mach->ideal_Opcode();
642
643 switch (iop) {
644 case Op_StoreB:
645 case Op_StoreC:
646 case Op_StoreCM:
647 case Op_StoreD:
648 case Op_StoreF:
649 case Op_StoreI:
650 case Op_StoreL:
651 case Op_StoreP:
652 case Op_StoreN:
653 case Op_StoreVector:
654 case Op_StoreNKlass:
655 for (uint k = 1; k < m->req(); k++) {
656 Node *in = m->in(k);
657 if (in == src_n) {
658 src_matches = true;
659 break;
660 }
661 }
662 break;
663
664 default:
665 src_matches = true;
666 break;
667 }
668
669 // If we have a store as our use, ignore the non source operands
670 if (src_matches == false) continue;
671
672 // Mark every unscheduled use which is not n with a recalculation
673 if ((get_block_for_node(m) == block) && (!m->is_scheduled())) {
674 if (finalize_mode && !m->is_Phi()) {
675 recalc_pressure_nodes[m->_idx] = 0x7fff7fff;
676 }
677 lrg_ends = false;
678 }
679 }
680 }
681 // if none, this live range ends and we can adjust register pressure
682 if (lrg_ends) {
683 if (finalize_mode) {
684 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
685 } else {
686 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
687 }
688 }
689 }
690
691 // now add the register pressure from the dest and evaluate which heuristic we should use:
692 // 1.) The default, latency scheduling
693 // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks
694 uint dst = _regalloc->_lrg_map.find(n);
695 if (dst != 0) {
696 LRG& lrg_dst = _regalloc->lrgs(dst);
697 if (finalize_mode) {
698 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
699 // check to see if we fall over the register pressure cliff here
700 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
701 _scheduling_for_pressure = true;
702 } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
703 _scheduling_for_pressure = true;
704 } else {
705 // restore latency scheduling mode
706 _scheduling_for_pressure = false;
707 }
708 } else {
709 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
710 }
711 }
712 }
713
714 //------------------------------set_next_call----------------------------------
715 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
716 if( next_call.test_set(n->_idx) ) return;
717 for( uint i=0; i<n->len(); i++ ) {
718 Node *m = n->in(i);
719 if( !m ) continue; // must see all nodes in block that precede call
720 if (get_block_for_node(m) == block) {
721 set_next_call(block, m, next_call);
722 }
723 }
724 }
725
726 //------------------------------needed_for_next_call---------------------------
727 // Set the flag 'next_call' for each Node that is needed for the next call to
728 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
729 // next subroutine call get priority - basically it moves things NOT needed
730 // for the next call till after the call. This prevents me from trying to
731 // carry lots of stuff live across a call.
732 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
733 // Find the next control-defining Node in this block
734 Node* call = NULL;
735 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
736 Node* m = this_call->fast_out(i);
737 if (get_block_for_node(m) == block && // Local-block user
738 m != this_call && // Not self-start node
739 m->is_MachCall()) {
740 call = m;
741 break;
742 }
743 }
744 if (call == NULL) return; // No next call (e.g., block end is near)
745 // Set next-call for all inputs to this call
746 set_next_call(block, call, next_call);
747 }
748
749 //------------------------------add_call_kills-------------------------------------
750 // helper function that adds caller save registers to MachProjNode
751 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
752 // Fill in the kill mask for the call
753 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
754 if( !regs.Member(r) ) { // Not already defined by the call
755 // Save-on-call register?
756 if ((save_policy[r] == 'C') ||
757 (save_policy[r] == 'A') ||
758 ((save_policy[r] == 'E') && exclude_soe)) {
759 proj->_rout.Insert(r);
760 }
761 }
762 }
763 }
764
765
766 //------------------------------sched_call-------------------------------------
767 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
768 RegMask regs;
769
770 // Schedule all the users of the call right now. All the users are
771 // projection Nodes, so they must be scheduled next to the call.
772 // Collect all the defined registers.
773 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
774 Node* n = mcall->fast_out(i);
775 assert( n->is_MachProj(), "" );
776 int n_cnt = ready_cnt.at(n->_idx)-1;
777 ready_cnt.at_put(n->_idx, n_cnt);
778 assert( n_cnt == 0, "" );
779 // Schedule next to call
780 block->map_node(n, node_cnt++);
781 // Collect defined registers
782 regs.OR(n->out_RegMask());
783 // Check for scheduling the next control-definer
784 if( n->bottom_type() == Type::CONTROL )
785 // Warm up next pile of heuristic bits
786 needed_for_next_call(block, n, next_call);
787
788 // Children of projections are now all ready
789 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
790 Node* m = n->fast_out(j); // Get user
791 if(get_block_for_node(m) != block) {
792 continue;
793 }
794 if( m->is_Phi() ) continue;
795 int m_cnt = ready_cnt.at(m->_idx) - 1;
796 ready_cnt.at_put(m->_idx, m_cnt);
797 if( m_cnt == 0 )
798 worklist.push(m);
799 }
800
801 }
802
803 // Act as if the call defines the Frame Pointer.
804 // Certainly the FP is alive and well after the call.
805 regs.Insert(_matcher.c_frame_pointer());
806
807 // Set all registers killed and not already defined by the call.
808 uint r_cnt = mcall->tf()->range()->cnt();
809 int op = mcall->ideal_Opcode();
810 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
811 map_node_to_block(proj, block);
812 block->insert_node(proj, node_cnt++);
813
814 // Select the right register save policy.
815 const char *save_policy = NULL;
816 switch (op) {
817 case Op_CallRuntime:
818 case Op_CallLeaf:
819 case Op_CallLeafNoFP:
820 // Calling C code so use C calling convention
821 save_policy = _matcher._c_reg_save_policy;
822 break;
823
824 case Op_CallStaticJava:
825 case Op_CallDynamicJava:
826 // Calling Java code so use Java calling convention
827 save_policy = _matcher._register_save_policy;
828 break;
829
830 default:
831 ShouldNotReachHere();
832 }
833
834 // When using CallRuntime mark SOE registers as killed by the call
835 // so values that could show up in the RegisterMap aren't live in a
836 // callee saved register since the register wouldn't know where to
837 // find them. CallLeaf and CallLeafNoFP are ok because they can't
838 // have debug info on them. Strictly speaking this only needs to be
839 // done for oops since idealreg2debugmask takes care of debug info
840 // references but there no way to handle oops differently than other
841 // pointers as far as the kill mask goes.
842 bool exclude_soe = op == Op_CallRuntime;
843
844 // If the call is a MethodHandle invoke, we need to exclude the
845 // register which is used to save the SP value over MH invokes from
846 // the mask. Otherwise this register could be used for
847 // deoptimization information.
848 if (op == Op_CallStaticJava) {
849 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
850 if (mcallstaticjava->_method_handle_invoke)
851 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
852 }
853
854 add_call_kills(proj, regs, save_policy, exclude_soe);
855
856 return node_cnt;
857 }
858
859
860 //------------------------------schedule_local---------------------------------
861 // Topological sort within a block. Someday become a real scheduler.
862 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) {
863 // Already "sorted" are the block start Node (as the first entry), and
864 // the block-ending Node and any trailing control projections. We leave
865 // these alone. PhiNodes and ParmNodes are made to follow the block start
866 // Node. Everything else gets topo-sorted.
867
868 #ifndef PRODUCT
869 if (trace_opto_pipelining()) {
870 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
871 for (uint i = 0;i < block->number_of_nodes(); i++) {
872 tty->print("# ");
873 block->get_node(i)->fast_dump();
874 }
875 tty->print_cr("#");
876 }
877 #endif
878
879 // RootNode is already sorted
880 if (block->number_of_nodes() == 1) {
881 return true;
882 }
883
884 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false;
885
886 // We track the uses of local definitions as input dependences so that
887 // we know when a given instruction is avialable to be scheduled.
888 uint i;
889 if (OptoRegScheduling && block_size_threshold_ok) {
890 for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc
891 Node *n = block->get_node(i);
892 n->remove_flag(Node::Flag_is_scheduled);
893 if (!n->is_Phi()) {
894 recalc_pressure_nodes[n->_idx] = 0x7fff7fff;
895 }
896 }
897 }
898
899 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
900 uint node_cnt = block->end_idx();
901 uint phi_cnt = 1;
902 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
903 Node *n = block->get_node(i);
904 if( n->is_Phi() || // Found a PhiNode or ParmNode
905 (n->is_Proj() && n->in(0) == block->head()) ) {
906 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
907 block->map_node(block->get_node(phi_cnt), i);
908 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
909 if (OptoRegScheduling && block_size_threshold_ok) {
910 // mark n as scheduled
911 n->add_flag(Node::Flag_is_scheduled);
912 }
913 } else { // All others
914 // Count block-local inputs to 'n'
915 uint cnt = n->len(); // Input count
916 uint local = 0;
917 for( uint j=0; j<cnt; j++ ) {
918 Node *m = n->in(j);
919 if( m && get_block_for_node(m) == block && !m->is_top() )
920 local++; // One more block-local input
921 }
922 ready_cnt.at_put(n->_idx, local); // Count em up
923
924 #ifdef ASSERT
925 if( UseConcMarkSweepGC || UseG1GC ) {
926 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
927 // Check the precedence edges
928 for (uint prec = n->req(); prec < n->len(); prec++) {
929 Node* oop_store = n->in(prec);
930 if (oop_store != NULL) {
931 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
932 }
933 }
934 }
935 }
936 #endif
937
938 // A few node types require changing a required edge to a precedence edge
939 // before allocation.
940 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
941 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
942 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
943 // MemBarAcquire could be created without Precedent edge.
944 // del_req() replaces the specified edge with the last input edge
945 // and then removes the last edge. If the specified edge > number of
946 // edges the last edge will be moved outside of the input edges array
947 // and the edge will be lost. This is why this code should be
948 // executed only when Precedent (== TypeFunc::Parms) edge is present.
949 Node *x = n->in(TypeFunc::Parms);
950 n->del_req(TypeFunc::Parms);
951 n->add_prec(x);
952 }
953 }
954 }
955 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
956 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
957
958 // All the prescheduled guys do not hold back internal nodes
959 uint i3;
960 for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled
961 Node *n = block->get_node(i3); // Get pre-scheduled
962 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
963 Node* m = n->fast_out(j);
964 if (get_block_for_node(m) == block) { // Local-block user
965 int m_cnt = ready_cnt.at(m->_idx)-1;
966 if (OptoRegScheduling && block_size_threshold_ok) {
967 // mark m as scheduled
968 if (m_cnt < 0) {
969 m->add_flag(Node::Flag_is_scheduled);
970 }
971 }
972 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
973 }
974 }
975 }
976
977 Node_List delay;
978 // Make a worklist
979 Node_List worklist;
980 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
981 Node *m = block->get_node(i4);
982 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
983 if (m->is_iteratively_computed()) {
984 // Push induction variable increments last to allow other uses
985 // of the phi to be scheduled first. The select() method breaks
986 // ties in scheduling by worklist order.
987 delay.push(m);
988 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
989 // Force the CreateEx to the top of the list so it's processed
990 // first and ends up at the start of the block.
991 worklist.insert(0, m);
992 } else {
993 worklist.push(m); // Then on to worklist!
994 }
995 }
996 }
997 while (delay.size()) {
998 Node* d = delay.pop();
999 worklist.push(d);
1000 }
1001
1002 if (OptoRegScheduling && block_size_threshold_ok) {
1003 // To stage register pressure calculations we need to examine the live set variables
1004 // breaking them up by register class to compartmentalize the calculations.
1005 uint float_pressure = Matcher::float_pressure(FLOATPRESSURE);
1006 _regalloc->_sched_int_pressure.init(INTPRESSURE);
1007 _regalloc->_sched_float_pressure.init(float_pressure);
1008 _regalloc->_scratch_int_pressure.init(INTPRESSURE);
1009 _regalloc->_scratch_float_pressure.init(float_pressure);
1010
1011 _regalloc->compute_entry_block_pressure(block);
1012 }
1013
1014 // Warm up the 'next_call' heuristic bits
1015 needed_for_next_call(block, block->head(), next_call);
1016
1017 #ifndef PRODUCT
1018 if (trace_opto_pipelining()) {
1019 for (uint j=0; j< block->number_of_nodes(); j++) {
1020 Node *n = block->get_node(j);
1021 int idx = n->_idx;
1022 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
1023 tty->print("latency:%3d ", get_latency_for_node(n));
1024 tty->print("%4d: %s\n", idx, n->Name());
1025 }
1026 }
1027 #endif
1028
1029 uint max_idx = (uint)ready_cnt.length();
1030 // Pull from worklist and schedule
1031 while( worklist.size() ) { // Worklist is not ready
1032
1033 #ifndef PRODUCT
1034 if (trace_opto_pipelining()) {
1035 tty->print("# ready list:");
1036 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1037 Node *n = worklist[i]; // Get Node on worklist
1038 tty->print(" %d", n->_idx);
1039 }
1040 tty->cr();
1041 }
1042 #endif
1043
1044 // Select and pop a ready guy from worklist
1045 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes);
1046 block->map_node(n, phi_cnt++); // Schedule him next
1047
1048 if (OptoRegScheduling && block_size_threshold_ok) {
1049 n->add_flag(Node::Flag_is_scheduled);
1050
1051 // Now adjust the resister pressure with the node we selected
1052 if (!n->is_Phi()) {
1053 adjust_register_pressure(n, block, recalc_pressure_nodes, true);
1054 }
1055 }
1056
1057 #ifndef PRODUCT
1058 if (trace_opto_pipelining()) {
1059 tty->print("# select %d: %s", n->_idx, n->Name());
1060 tty->print(", latency:%d", get_latency_for_node(n));
1061 n->dump();
1062 if (Verbose) {
1063 tty->print("# ready list:");
1064 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1065 Node *n = worklist[i]; // Get Node on worklist
1066 tty->print(" %d", n->_idx);
1067 }
1068 tty->cr();
1069 }
1070 }
1071
1072 #endif
1073 if( n->is_MachCall() ) {
1074 MachCallNode *mcall = n->as_MachCall();
1075 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
1076 continue;
1077 }
1078
1079 if (n->is_Mach() && n->as_Mach()->has_call()) {
1080 RegMask regs;
1081 regs.Insert(_matcher.c_frame_pointer());
1082 regs.OR(n->out_RegMask());
1083
1084 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
1085 map_node_to_block(proj, block);
1086 block->insert_node(proj, phi_cnt++);
1087
1088 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
1089 }
1090
1091 // Children are now all ready
1092 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
1093 Node* m = n->fast_out(i5); // Get user
1094 if (get_block_for_node(m) != block) {
1095 continue;
1096 }
1097 if( m->is_Phi() ) continue;
1098 if (m->_idx >= max_idx) { // new node, skip it
1099 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
1100 continue;
1101 }
1102 int m_cnt = ready_cnt.at(m->_idx) - 1;
1103 ready_cnt.at_put(m->_idx, m_cnt);
1104 if( m_cnt == 0 )
1105 worklist.push(m);
1106 }
1107 }
1108
1109 if( phi_cnt != block->end_idx() ) {
1110 // did not schedule all. Retry, Bailout, or Die
1111 if (C->subsume_loads() == true && !C->failing()) {
1112 // Retry with subsume_loads == false
1113 // If this is the first failure, the sentinel string will "stick"
1114 // to the Compile object, and the C2Compiler will see it and retry.
1115 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1116 }
1117 // assert( phi_cnt == end_idx(), "did not schedule all" );
1118 return false;
1119 }
1120
1121 if (OptoRegScheduling && block_size_threshold_ok) {
1122 _regalloc->compute_exit_block_pressure(block);
1123 block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure();
1124 block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure();
1125 }
1126
1127 #ifndef PRODUCT
1128 if (trace_opto_pipelining()) {
1129 tty->print_cr("#");
1130 tty->print_cr("# after schedule_local");
1131 for (uint i = 0;i < block->number_of_nodes();i++) {
1132 tty->print("# ");
1133 block->get_node(i)->fast_dump();
1134 }
1135 tty->print_cr("# ");
1136
1137 if (OptoRegScheduling && block_size_threshold_ok) {
1138 tty->print_cr("# pressure info : %d", block->_pre_order);
1139 _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info");
1140 _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info");
1141 }
1142 tty->cr();
1143 }
1144 #endif
1145
1146 return true;
1147 }
1148
1149 //--------------------------catch_cleanup_fix_all_inputs-----------------------
1150 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
1151 for (uint l = 0; l < use->len(); l++) {
1152 if (use->in(l) == old_def) {
1153 if (l < use->req()) {
1154 use->set_req(l, new_def);
1155 } else {
1156 use->rm_prec(l);
1157 use->add_prec(new_def);
1158 l--;
1159 }
1160 }
1161 }
1162 }
1163
1164 //------------------------------catch_cleanup_find_cloned_def------------------
1165 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1166 assert( use_blk != def_blk, "Inter-block cleanup only");
1167
1168 // The use is some block below the Catch. Find and return the clone of the def
1169 // that dominates the use. If there is no clone in a dominating block, then
1170 // create a phi for the def in a dominating block.
1171
1172 // Find which successor block dominates this use. The successor
1173 // blocks must all be single-entry (from the Catch only; I will have
1174 // split blocks to make this so), hence they all dominate.
1175 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1176 use_blk = use_blk->_idom;
1177
1178 // Find the successor
1179 Node *fixup = NULL;
1180
1181 uint j;
1182 for( j = 0; j < def_blk->_num_succs; j++ )
1183 if( use_blk == def_blk->_succs[j] )
1184 break;
1185
1186 if( j == def_blk->_num_succs ) {
1187 // Block at same level in dom-tree is not a successor. It needs a
1188 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1189 Node_Array inputs = new Node_List(Thread::current()->resource_area());
1190 for(uint k = 1; k < use_blk->num_preds(); k++) {
1191 Block* block = get_block_for_node(use_blk->pred(k));
1192 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1193 }
1194
1195 // Check to see if the use_blk already has an identical phi inserted.
1196 // If it exists, it will be at the first position since all uses of a
1197 // def are processed together.
1198 Node *phi = use_blk->get_node(1);
1199 if( phi->is_Phi() ) {
1200 fixup = phi;
1201 for (uint k = 1; k < use_blk->num_preds(); k++) {
1202 if (phi->in(k) != inputs[k]) {
1203 // Not a match
1204 fixup = NULL;
1205 break;
1206 }
1207 }
1208 }
1209
1210 // If an existing PhiNode was not found, make a new one.
1211 if (fixup == NULL) {
1212 Node *new_phi = PhiNode::make(use_blk->head(), def);
1213 use_blk->insert_node(new_phi, 1);
1214 map_node_to_block(new_phi, use_blk);
1215 for (uint k = 1; k < use_blk->num_preds(); k++) {
1216 new_phi->set_req(k, inputs[k]);
1217 }
1218 fixup = new_phi;
1219 }
1220
1221 } else {
1222 // Found the use just below the Catch. Make it use the clone.
1223 fixup = use_blk->get_node(n_clone_idx);
1224 }
1225
1226 return fixup;
1227 }
1228
1229 //--------------------------catch_cleanup_intra_block--------------------------
1230 // Fix all input edges in use that reference "def". The use is in the same
1231 // block as the def and both have been cloned in each successor block.
1232 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1233
1234 // Both the use and def have been cloned. For each successor block,
1235 // get the clone of the use, and make its input the clone of the def
1236 // found in that block.
1237
1238 uint use_idx = blk->find_node(use);
1239 uint offset_idx = use_idx - beg;
1240 for( uint k = 0; k < blk->_num_succs; k++ ) {
1241 // Get clone in each successor block
1242 Block *sb = blk->_succs[k];
1243 Node *clone = sb->get_node(offset_idx+1);
1244 assert( clone->Opcode() == use->Opcode(), "" );
1245
1246 // Make use-clone reference the def-clone
1247 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1248 }
1249 }
1250
1251 //------------------------------catch_cleanup_inter_block---------------------
1252 // Fix all input edges in use that reference "def". The use is in a different
1253 // block than the def.
1254 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1255 if( !use_blk ) return; // Can happen if the use is a precedence edge
1256
1257 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1258 catch_cleanup_fix_all_inputs(use, def, new_def);
1259 }
1260
1261 //------------------------------call_catch_cleanup-----------------------------
1262 // If we inserted any instructions between a Call and his CatchNode,
1263 // clone the instructions on all paths below the Catch.
1264 void PhaseCFG::call_catch_cleanup(Block* block) {
1265
1266 // End of region to clone
1267 uint end = block->end_idx();
1268 if( !block->get_node(end)->is_Catch() ) return;
1269 // Start of region to clone
1270 uint beg = end;
1271 while(!block->get_node(beg-1)->is_MachProj() ||
1272 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1273 beg--;
1274 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1275 }
1276 // Range of inserted instructions is [beg, end)
1277 if( beg == end ) return;
1278
1279 // Clone along all Catch output paths. Clone area between the 'beg' and
1280 // 'end' indices.
1281 for( uint i = 0; i < block->_num_succs; i++ ) {
1282 Block *sb = block->_succs[i];
1283 // Clone the entire area; ignoring the edge fixup for now.
1284 for( uint j = end; j > beg; j-- ) {
1285 // It is safe here to clone a node with anti_dependence
1286 // since clones dominate on each path.
1287 Node *clone = block->get_node(j-1)->clone();
1288 sb->insert_node(clone, 1);
1289 map_node_to_block(clone, sb);
1290 }
1291 }
1292
1293
1294 // Fixup edges. Check the def-use info per cloned Node
1295 for(uint i2 = beg; i2 < end; i2++ ) {
1296 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1297 Node *n = block->get_node(i2); // Node that got cloned
1298 // Need DU safe iterator because of edge manipulation in calls.
1299 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1300 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1301 out->push(n->fast_out(j1));
1302 }
1303 uint max = out->size();
1304 for (uint j = 0; j < max; j++) {// For all users
1305 Node *use = out->pop();
1306 Block *buse = get_block_for_node(use);
1307 if( use->is_Phi() ) {
1308 for( uint k = 1; k < use->req(); k++ )
1309 if( use->in(k) == n ) {
1310 Block* b = get_block_for_node(buse->pred(k));
1311 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1312 use->set_req(k, fixup);
1313 }
1314 } else {
1315 if (block == buse) {
1316 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1317 } else {
1318 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1319 }
1320 }
1321 } // End for all users
1322
1323 } // End of for all Nodes in cloned area
1324
1325 // Remove the now-dead cloned ops
1326 for(uint i3 = beg; i3 < end; i3++ ) {
1327 block->get_node(beg)->disconnect_inputs(NULL, C);
1328 block->remove_node(beg);
1329 }
1330
1331 // If the successor blocks have a CreateEx node, move it back to the top
1332 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1333 Block *sb = block->_succs[i4];
1334 uint new_cnt = end - beg;
1335 // Remove any newly created, but dead, nodes.
1336 for( uint j = new_cnt; j > 0; j-- ) {
1337 Node *n = sb->get_node(j);
1338 if (n->outcnt() == 0 &&
1339 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1340 n->disconnect_inputs(NULL, C);
1341 sb->remove_node(j);
1342 new_cnt--;
1343 }
1344 }
1345 // If any newly created nodes remain, move the CreateEx node to the top
1346 if (new_cnt > 0) {
1347 Node *cex = sb->get_node(1+new_cnt);
1348 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1349 sb->remove_node(1+new_cnt);
1350 sb->insert_node(cex, 1);
1351 }
1352 }
1353 }
1354 }