1 /*
2 * Copyright (c) 1998, 2017, 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() != NULL) { // 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 case Op_ShenandoahReadBarrier:
179 case Op_ShenandoahWriteBarrier:
180 assert(mach->in(2) == val, "should be address");
181 break;
182 case Op_StoreB:
183 case Op_StoreC:
184 case Op_StoreCM:
185 case Op_StoreD:
186 case Op_StoreF:
187 case Op_StoreI:
188 case Op_StoreL:
189 case Op_StoreP:
190 case Op_StoreN:
191 case Op_StoreNKlass:
192 was_store = true; // Memory op is a store op
193 // Stores will have their address in slot 2 (memory in slot 1).
194 // If the value being nul-checked is in another slot, it means we
195 // are storing the checked value, which does NOT check the value!
196 if( mach->in(2) != val ) continue;
197 break; // Found a memory op?
198 case Op_StrComp:
199 case Op_StrEquals:
200 case Op_StrIndexOf:
201 case Op_StrIndexOfChar:
202 case Op_AryEq:
203 case Op_StrInflatedCopy:
204 case Op_StrCompressedCopy:
205 case Op_EncodeISOArray:
206 case Op_HasNegatives:
207 // Not a legit memory op for implicit null check regardless of
208 // embedded loads
209 continue;
210 default: // Also check for embedded loads
211 if( !mach->needs_anti_dependence_check() )
212 continue; // Not an memory op; skip it
213 if( must_clone[iop] ) {
214 // Do not move nodes which produce flags because
215 // RA will try to clone it to place near branch and
216 // it will cause recompilation, see clone_node().
217 continue;
218 }
219 {
220 // Check that value is used in memory address in
221 // instructions with embedded load (CmpP val1,(val2+off)).
222 Node* base;
223 Node* index;
224 const MachOper* oper = mach->memory_inputs(base, index);
225 if (oper == NULL || oper == (MachOper*)-1) {
226 continue; // Not an memory op; skip it
227 }
228 if (val == base ||
229 (val == index && val->bottom_type()->isa_narrowoop())) {
230 break; // Found it
231 } else {
232 continue; // Skip it
233 }
234 }
235 break;
236 }
237
238 // On some OSes (AIX) the page at address 0 is only write protected.
239 // If so, only Store operations will trap.
240 // But a read accessing the base of a heap-based compressed heap will trap.
241 if (!was_store && needs_explicit_null_check_for_read(val)) {
242 continue;
243 }
244
245 // Check that node's control edge is not-null block's head or dominates it,
246 // otherwise we can't hoist it because there are other control dependencies.
247 Node* ctrl = mach->in(0);
248 if (ctrl != NULL && !(ctrl == not_null_block->head() ||
249 get_block_for_node(ctrl)->dominates(not_null_block))) {
250 continue;
251 }
252
253 // check if the offset is not too high for implicit exception
254 {
255 intptr_t offset = 0;
256 const TypePtr *adr_type = NULL; // Do not need this return value here
257 const Node* base = mach->get_base_and_disp(offset, adr_type);
258 if (base == NULL || base == NodeSentinel) {
259 // Narrow oop address doesn't have base, only index.
260 // Give up if offset is beyond page size or if heap base is not protected.
261 if (val->bottom_type()->isa_narrowoop() &&
262 (MacroAssembler::needs_explicit_null_check(offset) ||
263 !Universe::narrow_oop_use_implicit_null_checks()))
264 continue;
265 // cannot reason about it; is probably not implicit null exception
266 } else {
267 const TypePtr* tptr;
268 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
269 Universe::narrow_klass_shift() == 0)) {
270 // 32-bits narrow oop can be the base of address expressions
271 tptr = base->get_ptr_type();
272 } else {
273 // only regular oops are expected here
274 tptr = base->bottom_type()->is_ptr();
275 }
276 // Give up if offset is not a compile-time constant.
277 if (offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot)
278 continue;
279 offset += tptr->_offset; // correct if base is offseted
280 // Give up if reference is beyond page size.
281 if (MacroAssembler::needs_explicit_null_check(offset))
282 continue;
283 // Give up if base is a decode node and the heap base is not protected.
284 if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN &&
285 !Universe::narrow_oop_use_implicit_null_checks())
286 continue;
287 }
288 }
289
290 // Check ctrl input to see if the null-check dominates the memory op
291 Block *cb = get_block_for_node(mach);
292 cb = cb->_idom; // Always hoist at least 1 block
293 if( !was_store ) { // Stores can be hoisted only one block
294 while( cb->_dom_depth > (block->_dom_depth + 1))
295 cb = cb->_idom; // Hoist loads as far as we want
296 // The non-null-block should dominate the memory op, too. Live
297 // range spilling will insert a spill in the non-null-block if it is
298 // needs to spill the memory op for an implicit null check.
299 if (cb->_dom_depth == (block->_dom_depth + 1)) {
300 if (cb != not_null_block) continue;
301 cb = cb->_idom;
302 }
303 }
304 if( cb != block ) continue;
305
306 // Found a memory user; see if it can be hoisted to check-block
307 uint vidx = 0; // Capture index of value into memop
308 uint j;
309 for( j = mach->req()-1; j > 0; j-- ) {
310 if( mach->in(j) == val ) {
311 vidx = j;
312 // Ignore DecodeN val which could be hoisted to where needed.
313 if( is_decoden ) continue;
314 }
315 // Block of memory-op input
316 Block *inb = get_block_for_node(mach->in(j));
317 Block *b = block; // Start from nul check
318 while( b != inb && b->_dom_depth > inb->_dom_depth )
319 b = b->_idom; // search upwards for input
320 // See if input dominates null check
321 if( b != inb )
322 break;
323 }
324 if( j > 0 )
325 continue;
326 Block *mb = get_block_for_node(mach);
327 // Hoisting stores requires more checks for the anti-dependence case.
328 // Give up hoisting if we have to move the store past any load.
329 if( was_store ) {
330 Block *b = mb; // Start searching here for a local load
331 // mach use (faulting) trying to hoist
332 // n might be blocker to hoisting
333 while( b != block ) {
334 uint k;
335 for( k = 1; k < b->number_of_nodes(); k++ ) {
336 Node *n = b->get_node(k);
337 if( n->needs_anti_dependence_check() &&
338 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
339 break; // Found anti-dependent load
340 }
341 if( k < b->number_of_nodes() )
342 break; // Found anti-dependent load
343 // Make sure control does not do a merge (would have to check allpaths)
344 if( b->num_preds() != 2 ) break;
345 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
346 }
347 if( b != block ) continue;
348 }
349
350 // Make sure this memory op is not already being used for a NullCheck
351 Node *e = mb->end();
352 if( e->is_MachNullCheck() && e->in(1) == mach )
353 continue; // Already being used as a NULL check
354
355 // Found a candidate! Pick one with least dom depth - the highest
356 // in the dom tree should be closest to the null check.
357 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
358 best = mach;
359 bidx = vidx;
360 }
361 }
362 // No candidate!
363 if (best == NULL) {
364 return;
365 }
366
367 // ---- Found an implicit null check
368 #ifndef PRODUCT
369 extern int implicit_null_checks;
370 implicit_null_checks++;
371 #endif
372
373 if( is_decoden ) {
374 // Check if we need to hoist decodeHeapOop_not_null first.
375 Block *valb = get_block_for_node(val);
376 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
377 // Hoist it up to the end of the test block.
378 valb->find_remove(val);
379 block->add_inst(val);
380 map_node_to_block(val, block);
381 // DecodeN on x86 may kill flags. Check for flag-killing projections
382 // that also need to be hoisted.
383 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
384 Node* n = val->fast_out(j);
385 if( n->is_MachProj() ) {
386 get_block_for_node(n)->find_remove(n);
387 block->add_inst(n);
388 map_node_to_block(n, block);
389 }
390 }
391 }
392 }
393 // Hoist the memory candidate up to the end of the test block.
394 Block *old_block = get_block_for_node(best);
395 old_block->find_remove(best);
396 block->add_inst(best);
397 map_node_to_block(best, block);
398
399 // Move the control dependence if it is pinned to not-null block.
400 // Don't change it in other cases: NULL or dominating control.
401 if (best->in(0) == not_null_block->head()) {
402 // Set it to control edge of null check.
403 best->set_req(0, proj->in(0)->in(0));
404 }
405
406 // Check for flag-killing projections that also need to be hoisted
407 // Should be DU safe because no edge updates.
408 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
409 Node* n = best->fast_out(j);
410 if( n->is_MachProj() || n->Opcode() == Op_ShenandoahWBMemProj) {
411 get_block_for_node(n)->find_remove(n);
412 block->add_inst(n);
413 map_node_to_block(n, block);
414 }
415 }
416
417 // proj==Op_True --> ne test; proj==Op_False --> eq test.
418 // One of two graph shapes got matched:
419 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
420 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
421 // NULL checks are always branch-if-eq. If we see a IfTrue projection
422 // then we are replacing a 'ne' test with a 'eq' NULL check test.
423 // We need to flip the projections to keep the same semantics.
424 if( proj->Opcode() == Op_IfTrue ) {
425 // Swap order of projections in basic block to swap branch targets
426 Node *tmp1 = block->get_node(block->end_idx()+1);
427 Node *tmp2 = block->get_node(block->end_idx()+2);
428 block->map_node(tmp2, block->end_idx()+1);
429 block->map_node(tmp1, block->end_idx()+2);
430 Node *tmp = new Node(C->top()); // Use not NULL input
431 tmp1->replace_by(tmp);
432 tmp2->replace_by(tmp1);
433 tmp->replace_by(tmp2);
434 tmp->destruct();
435 }
436
437 // Remove the existing null check; use a new implicit null check instead.
438 // Since schedule-local needs precise def-use info, we need to correct
439 // it as well.
440 Node *old_tst = proj->in(0);
441 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx);
442 block->map_node(nul_chk, block->end_idx());
443 map_node_to_block(nul_chk, block);
444 // Redirect users of old_test to nul_chk
445 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
446 old_tst->last_out(i2)->set_req(0, nul_chk);
447 // Clean-up any dead code
448 for (uint i3 = 0; i3 < old_tst->req(); i3++) {
449 Node* in = old_tst->in(i3);
450 old_tst->set_req(i3, NULL);
451 if (in->outcnt() == 0) {
452 // Remove dead input node
453 in->disconnect_inputs(NULL, C);
454 block->find_remove(in);
455 }
456 }
457
458 latency_from_uses(nul_chk);
459 latency_from_uses(best);
460
461 // insert anti-dependences to defs in this block
462 if (! best->needs_anti_dependence_check()) {
463 for (uint k = 1; k < block->number_of_nodes(); k++) {
464 Node *n = block->get_node(k);
465 if (n->needs_anti_dependence_check() &&
466 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) {
467 // Found anti-dependent load
468 insert_anti_dependences(block, n);
469 }
470 }
471 }
472 }
473
474
475 //------------------------------select-----------------------------------------
476 // Select a nice fellow from the worklist to schedule next. If there is only
477 // one choice, then use it. Projections take top priority for correctness
478 // reasons - if I see a projection, then it is next. There are a number of
479 // other special cases, for instructions that consume condition codes, et al.
480 // These are chosen immediately. Some instructions are required to immediately
481 // precede the last instruction in the block, and these are taken last. Of the
482 // remaining cases (most), choose the instruction with the greatest latency
483 // (that is, the most number of pseudo-cycles required to the end of the
484 // routine). If there is a tie, choose the instruction with the most inputs.
485 Node* PhaseCFG::select(
486 Block* block,
487 Node_List &worklist,
488 GrowableArray<int> &ready_cnt,
489 VectorSet &next_call,
490 uint sched_slot,
491 intptr_t* recalc_pressure_nodes) {
492
493 // If only a single entry on the stack, use it
494 uint cnt = worklist.size();
495 if (cnt == 1) {
496 Node *n = worklist[0];
497 worklist.map(0,worklist.pop());
498 return n;
499 }
500
501 uint choice = 0; // Bigger is most important
502 uint latency = 0; // Bigger is scheduled first
503 uint score = 0; // Bigger is better
504 int idx = -1; // Index in worklist
505 int cand_cnt = 0; // Candidate count
506 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false;
507
508 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
509 // Order in worklist is used to break ties.
510 // See caller for how this is used to delay scheduling
511 // of induction variable increments to after the other
512 // uses of the phi are scheduled.
513 Node *n = worklist[i]; // Get Node on worklist
514
515 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
516 if( n->is_Proj() || // Projections always win
517 n->Opcode()== Op_Con || // So does constant 'Top'
518 iop == Op_CreateEx || // Create-exception must start block
519 iop == Op_CheckCastPP
520 ) {
521 worklist.map(i,worklist.pop());
522 return n;
523 }
524
525 // Final call in a block must be adjacent to 'catch'
526 Node *e = block->end();
527 if( e->is_Catch() && e->in(0)->in(0) == n )
528 continue;
529
530 // Memory op for an implicit null check has to be at the end of the block
531 if( e->is_MachNullCheck() && e->in(1) == n )
532 continue;
533
534 // Schedule IV increment last.
535 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) {
536 // Cmp might be matched into CountedLoopEnd node.
537 Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e;
538 if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) {
539 continue;
540 }
541 }
542
543 uint n_choice = 2;
544
545 // See if this instruction is consumed by a branch. If so, then (as the
546 // branch is the last instruction in the basic block) force it to the
547 // end of the basic block
548 if ( must_clone[iop] ) {
549 // See if any use is a branch
550 bool found_machif = false;
551
552 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
553 Node* use = n->fast_out(j);
554
555 // The use is a conditional branch, make them adjacent
556 if (use->is_MachIf() && get_block_for_node(use) == block) {
557 found_machif = true;
558 break;
559 }
560
561 // More than this instruction pending for successor to be ready,
562 // don't choose this if other opportunities are ready
563 if (ready_cnt.at(use->_idx) > 1)
564 n_choice = 1;
565 }
566
567 // loop terminated, prefer not to use this instruction
568 if (found_machif)
569 continue;
570 }
571
572 // See if this has a predecessor that is "must_clone", i.e. sets the
573 // condition code. If so, choose this first
574 for (uint j = 0; j < n->req() ; j++) {
575 Node *inn = n->in(j);
576 if (inn) {
577 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
578 n_choice = 3;
579 break;
580 }
581 }
582 }
583
584 // MachTemps should be scheduled last so they are near their uses
585 if (n->is_MachTemp()) {
586 n_choice = 1;
587 }
588
589 uint n_latency = get_latency_for_node(n);
590 uint n_score = n->req(); // Many inputs get high score to break ties
591
592 if (OptoRegScheduling && block_size_threshold_ok) {
593 if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) {
594 _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit());
595 _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit());
596 // simulate the notion that we just picked this node to schedule
597 n->add_flag(Node::Flag_is_scheduled);
598 // now caculate its effect upon the graph if we did
599 adjust_register_pressure(n, block, recalc_pressure_nodes, false);
600 // return its state for finalize in case somebody else wins
601 n->remove_flag(Node::Flag_is_scheduled);
602 // now save the two final pressure components of register pressure, limiting pressure calcs to short size
603 short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure();
604 short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure();
605 recalc_pressure_nodes[n->_idx] = int_pressure;
606 recalc_pressure_nodes[n->_idx] |= (float_pressure << 16);
607 }
608
609 if (_scheduling_for_pressure) {
610 latency = n_latency;
611 if (n_choice != 3) {
612 // Now evaluate each register pressure component based on threshold in the score.
613 // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks
614 // on a single instruction, but we might see it shrink on both banks.
615 // For each use of register that has a register class that is over the high pressure limit, we build n_score up for
616 // live ranges that terminate on this instruction.
617 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
618 short int_pressure = (short)recalc_pressure_nodes[n->_idx];
619 n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score;
620 }
621 if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
622 short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16);
623 n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score;
624 }
625 } else {
626 // make sure we choose these candidates
627 score = 0;
628 }
629 }
630 }
631
632 // Keep best latency found
633 cand_cnt++;
634 if (choice < n_choice ||
635 (choice == n_choice &&
636 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
637 (!StressLCM &&
638 (latency < n_latency ||
639 (latency == n_latency &&
640 (score < n_score))))))) {
641 choice = n_choice;
642 latency = n_latency;
643 score = n_score;
644 idx = i; // Also keep index in worklist
645 }
646 } // End of for all ready nodes in worklist
647
648 assert(idx >= 0, "index should be set");
649 Node *n = worklist[(uint)idx]; // Get the winner
650
651 worklist.map((uint)idx, worklist.pop()); // Compress worklist
652 return n;
653 }
654
655 //-------------------------adjust_register_pressure----------------------------
656 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) {
657 PhaseLive* liveinfo = _regalloc->get_live();
658 IndexSet* liveout = liveinfo->live(block);
659 // first adjust the register pressure for the sources
660 for (uint i = 1; i < n->req(); i++) {
661 bool lrg_ends = false;
662 Node *src_n = n->in(i);
663 if (src_n == NULL) continue;
664 if (!src_n->is_Mach()) continue;
665 uint src = _regalloc->_lrg_map.find(src_n);
666 if (src == 0) continue;
667 LRG& lrg_src = _regalloc->lrgs(src);
668 // detect if the live range ends or not
669 if (liveout->member(src) == false) {
670 lrg_ends = true;
671 for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) {
672 Node* m = src_n->fast_out(j); // Get user
673 if (m == n) continue;
674 if (!m->is_Mach()) continue;
675 MachNode *mach = m->as_Mach();
676 bool src_matches = false;
677 int iop = mach->ideal_Opcode();
678
679 switch (iop) {
680 case Op_StoreB:
681 case Op_StoreC:
682 case Op_StoreCM:
683 case Op_StoreD:
684 case Op_StoreF:
685 case Op_StoreI:
686 case Op_StoreL:
687 case Op_StoreP:
688 case Op_StoreN:
689 case Op_StoreVector:
690 case Op_StoreNKlass:
691 for (uint k = 1; k < m->req(); k++) {
692 Node *in = m->in(k);
693 if (in == src_n) {
694 src_matches = true;
695 break;
696 }
697 }
698 break;
699
700 default:
701 src_matches = true;
702 break;
703 }
704
705 // If we have a store as our use, ignore the non source operands
706 if (src_matches == false) continue;
707
708 // Mark every unscheduled use which is not n with a recalculation
709 if ((get_block_for_node(m) == block) && (!m->is_scheduled())) {
710 if (finalize_mode && !m->is_Phi()) {
711 recalc_pressure_nodes[m->_idx] = 0x7fff7fff;
712 }
713 lrg_ends = false;
714 }
715 }
716 }
717 // if none, this live range ends and we can adjust register pressure
718 if (lrg_ends) {
719 if (finalize_mode) {
720 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
721 } else {
722 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
723 }
724 }
725 }
726
727 // now add the register pressure from the dest and evaluate which heuristic we should use:
728 // 1.) The default, latency scheduling
729 // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks
730 uint dst = _regalloc->_lrg_map.find(n);
731 if (dst != 0) {
732 LRG& lrg_dst = _regalloc->lrgs(dst);
733 if (finalize_mode) {
734 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
735 // check to see if we fall over the register pressure cliff here
736 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
737 _scheduling_for_pressure = true;
738 } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
739 _scheduling_for_pressure = true;
740 } else {
741 // restore latency scheduling mode
742 _scheduling_for_pressure = false;
743 }
744 } else {
745 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
746 }
747 }
748 }
749
750 //------------------------------set_next_call----------------------------------
751 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
752 if( next_call.test_set(n->_idx) ) return;
753 for( uint i=0; i<n->len(); i++ ) {
754 Node *m = n->in(i);
755 if( !m ) continue; // must see all nodes in block that precede call
756 if (get_block_for_node(m) == block) {
757 set_next_call(block, m, next_call);
758 }
759 }
760 }
761
762 //------------------------------needed_for_next_call---------------------------
763 // Set the flag 'next_call' for each Node that is needed for the next call to
764 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
765 // next subroutine call get priority - basically it moves things NOT needed
766 // for the next call till after the call. This prevents me from trying to
767 // carry lots of stuff live across a call.
768 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
769 // Find the next control-defining Node in this block
770 Node* call = NULL;
771 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
772 Node* m = this_call->fast_out(i);
773 if (get_block_for_node(m) == block && // Local-block user
774 m != this_call && // Not self-start node
775 m->is_MachCall()) {
776 call = m;
777 break;
778 }
779 }
780 if (call == NULL) return; // No next call (e.g., block end is near)
781 // Set next-call for all inputs to this call
782 set_next_call(block, call, next_call);
783 }
784
785 //------------------------------add_call_kills-------------------------------------
786 // helper function that adds caller save registers to MachProjNode
787 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe, bool exclude_fp) {
788 // Fill in the kill mask for the call
789 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
790 if (exclude_fp && (register_save_type[r] == Op_RegF || register_save_type[r] == Op_RegD)) {
791 continue;
792 }
793 if( !regs.Member(r) ) { // Not already defined by the call
794 // Save-on-call register?
795 if ((save_policy[r] == 'C') ||
796 (save_policy[r] == 'A') ||
797 ((save_policy[r] == 'E') && exclude_soe)) {
798 proj->_rout.Insert(r);
799 }
800 }
801 }
802 }
803
804
805 //------------------------------sched_call-------------------------------------
806 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
807 RegMask regs;
808
809 // Schedule all the users of the call right now. All the users are
810 // projection Nodes, so they must be scheduled next to the call.
811 // Collect all the defined registers.
812 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
813 Node* n = mcall->fast_out(i);
814 assert( n->is_MachProj(), "" );
815 int n_cnt = ready_cnt.at(n->_idx)-1;
816 ready_cnt.at_put(n->_idx, n_cnt);
817 assert( n_cnt == 0, "" );
818 // Schedule next to call
819 block->map_node(n, node_cnt++);
820 // Collect defined registers
821 regs.OR(n->out_RegMask());
822 // Check for scheduling the next control-definer
823 if( n->bottom_type() == Type::CONTROL )
824 // Warm up next pile of heuristic bits
825 needed_for_next_call(block, n, next_call);
826
827 // Children of projections are now all ready
828 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
829 Node* m = n->fast_out(j); // Get user
830 if(get_block_for_node(m) != block) {
831 continue;
832 }
833 if( m->is_Phi() ) continue;
834 int m_cnt = ready_cnt.at(m->_idx) - 1;
835 ready_cnt.at_put(m->_idx, m_cnt);
836 if( m_cnt == 0 )
837 worklist.push(m);
838 }
839
840 }
841
842 // Act as if the call defines the Frame Pointer.
843 // Certainly the FP is alive and well after the call.
844 regs.Insert(_matcher.c_frame_pointer());
845
846 // Set all registers killed and not already defined by the call.
847 uint r_cnt = mcall->tf()->range()->cnt();
848 int op = mcall->ideal_Opcode();
849 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
850 map_node_to_block(proj, block);
851 block->insert_node(proj, node_cnt++);
852
853 // Select the right register save policy.
854 const char *save_policy = NULL;
855 switch (op) {
856 case Op_CallRuntime:
857 case Op_CallLeaf:
858 case Op_CallLeafNoFP:
859 // Calling C code so use C calling convention
860 save_policy = _matcher._c_reg_save_policy;
861 break;
862
863 case Op_CallStaticJava:
864 case Op_CallDynamicJava:
865 // Calling Java code so use Java calling convention
866 save_policy = _matcher._register_save_policy;
867 break;
868
869 default:
870 ShouldNotReachHere();
871 }
872
873 // When using CallRuntime mark SOE registers as killed by the call
874 // so values that could show up in the RegisterMap aren't live in a
875 // callee saved register since the register wouldn't know where to
876 // find them. CallLeaf and CallLeafNoFP are ok because they can't
877 // have debug info on them. Strictly speaking this only needs to be
878 // done for oops since idealreg2debugmask takes care of debug info
879 // references but there no way to handle oops differently than other
880 // pointers as far as the kill mask goes.
881 bool exclude_soe = op == Op_CallRuntime;
882
883 // If the call is a MethodHandle invoke, we need to exclude the
884 // register which is used to save the SP value over MH invokes from
885 // the mask. Otherwise this register could be used for
886 // deoptimization information.
887 if (op == Op_CallStaticJava) {
888 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
889 if (mcallstaticjava->_method_handle_invoke)
890 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
891 }
892
893 if (UseShenandoahGC && mcall->entry_point() == StubRoutines::shenandoah_wb_C()) {
894 assert(op == Op_CallLeafNoFP, "shenandoah_wb_C should be called with Op_CallLeafNoFP");
895 add_call_kills(proj, regs, save_policy, exclude_soe, true);
896 } else {
897 add_call_kills(proj, regs, save_policy, exclude_soe, false);
898 }
899
900 return node_cnt;
901 }
902
903 void PhaseCFG::push_ready_nodes(Node* n, Node* m, Block* block, GrowableArray<int>& ready_cnt, Node_List& worklist, uint max_idx, int c) {
904 if (get_block_for_node(m) != block) {
905 return;
906 }
907 if (m->is_Phi()) {
908 return;
909 }
910 if (m->_idx >= max_idx) { // new node, skip it
911 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
912 return;
913 }
914 int m_cnt = ready_cnt.at(m->_idx) - c;
915 ready_cnt.at_put(m->_idx, m_cnt);
916 if (m_cnt == 0) {
917 worklist.push(m);
918 }
919 }
920
921 //------------------------------schedule_local---------------------------------
922 // Topological sort within a block. Someday become a real scheduler.
923 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) {
924 // Already "sorted" are the block start Node (as the first entry), and
925 // the block-ending Node and any trailing control projections. We leave
926 // these alone. PhiNodes and ParmNodes are made to follow the block start
927 // Node. Everything else gets topo-sorted.
928
929 #ifndef PRODUCT
930 if (trace_opto_pipelining()) {
931 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
932 for (uint i = 0;i < block->number_of_nodes(); i++) {
933 tty->print("# ");
934 block->get_node(i)->fast_dump();
935 }
936 tty->print_cr("#");
937 }
938 #endif
939
940 // RootNode is already sorted
941 if (block->number_of_nodes() == 1) {
942 return true;
943 }
944
945 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false;
946
947 // We track the uses of local definitions as input dependences so that
948 // we know when a given instruction is avialable to be scheduled.
949 uint i;
950 if (OptoRegScheduling && block_size_threshold_ok) {
951 for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc
952 Node *n = block->get_node(i);
953 n->remove_flag(Node::Flag_is_scheduled);
954 if (!n->is_Phi()) {
955 recalc_pressure_nodes[n->_idx] = 0x7fff7fff;
956 }
957 }
958 } else {
959 #ifdef ASSERT
960 for (i = 1; i < block->number_of_nodes(); i++) {
961 Node *n = block->get_node(i);
962 assert(!n->is_scheduled(), "shouldn't be scheduled yet");
963 }
964 #endif
965 }
966
967 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
968 uint node_cnt = block->end_idx();
969 uint phi_cnt = 1;
970 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
971 Node *n = block->get_node(i);
972 if( n->is_Phi() || // Found a PhiNode or ParmNode
973 (n->is_Proj() && n->in(0) == block->head()) ) {
974 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
975 block->map_node(block->get_node(phi_cnt), i);
976 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
977 // mark n as scheduled
978 n->add_flag(Node::Flag_is_scheduled);
979 } else { // All others
980 // Count block-local inputs to 'n'
981 uint cnt = n->len(); // Input count
982 uint local = 0;
983 for( uint j=0; j<cnt; j++ ) {
984 Node *m = n->in(j);
985 if( m && get_block_for_node(m) == block && !m->is_top() )
986 local++; // One more block-local input
987 }
988 ready_cnt.at_put(n->_idx, local); // Count em up
989
990 #ifdef ASSERT
991 if( UseConcMarkSweepGC || UseG1GC ) {
992 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
993 // Check the precedence edges
994 for (uint prec = n->req(); prec < n->len(); prec++) {
995 Node* oop_store = n->in(prec);
996 if (oop_store != NULL) {
997 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
998 }
999 }
1000 }
1001 }
1002 #endif
1003
1004 // A few node types require changing a required edge to a precedence edge
1005 // before allocation.
1006 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
1007 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
1008 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
1009 // MemBarAcquire could be created without Precedent edge.
1010 // del_req() replaces the specified edge with the last input edge
1011 // and then removes the last edge. If the specified edge > number of
1012 // edges the last edge will be moved outside of the input edges array
1013 // and the edge will be lost. This is why this code should be
1014 // executed only when Precedent (== TypeFunc::Parms) edge is present.
1015 Node *x = n->in(TypeFunc::Parms);
1016 if (x != NULL && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) {
1017 // Old edge to node within same block will get removed, but no precedence
1018 // edge will get added because it already exists. Update ready count.
1019 int cnt = ready_cnt.at(n->_idx);
1020 assert(cnt > 1, "MemBar node %d must not get ready here", n->_idx);
1021 ready_cnt.at_put(n->_idx, cnt-1);
1022 }
1023 n->del_req(TypeFunc::Parms);
1024 n->add_prec(x);
1025 }
1026 }
1027 }
1028 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
1029 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
1030
1031 // All the prescheduled guys do not hold back internal nodes
1032 uint i3;
1033 for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled
1034 Node *n = block->get_node(i3); // Get pre-scheduled
1035 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1036 Node* m = n->fast_out(j);
1037 if (get_block_for_node(m) == block) { // Local-block user
1038 int m_cnt = ready_cnt.at(m->_idx)-1;
1039 // mark m as scheduled
1040 if (m_cnt < 0) {
1041 m->add_flag(Node::Flag_is_scheduled);
1042 }
1043 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
1044 }
1045 }
1046 }
1047
1048 Node_List delay;
1049 // Make a worklist
1050 Node_List worklist;
1051 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
1052 Node *m = block->get_node(i4);
1053 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
1054 if (m->is_iteratively_computed()) {
1055 // Push induction variable increments last to allow other uses
1056 // of the phi to be scheduled first. The select() method breaks
1057 // ties in scheduling by worklist order.
1058 delay.push(m);
1059 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1060 // Force the CreateEx to the top of the list so it's processed
1061 // first and ends up at the start of the block.
1062 worklist.insert(0, m);
1063 } else {
1064 worklist.push(m); // Then on to worklist!
1065 }
1066 }
1067 }
1068 while (delay.size()) {
1069 Node* d = delay.pop();
1070 worklist.push(d);
1071 }
1072
1073 if (OptoRegScheduling && block_size_threshold_ok) {
1074 // To stage register pressure calculations we need to examine the live set variables
1075 // breaking them up by register class to compartmentalize the calculations.
1076 uint float_pressure = Matcher::float_pressure(FLOATPRESSURE);
1077 _regalloc->_sched_int_pressure.init(INTPRESSURE);
1078 _regalloc->_sched_float_pressure.init(float_pressure);
1079 _regalloc->_scratch_int_pressure.init(INTPRESSURE);
1080 _regalloc->_scratch_float_pressure.init(float_pressure);
1081
1082 _regalloc->compute_entry_block_pressure(block);
1083 }
1084
1085 // Warm up the 'next_call' heuristic bits
1086 needed_for_next_call(block, block->head(), next_call);
1087
1088 #ifndef PRODUCT
1089 if (trace_opto_pipelining()) {
1090 for (uint j=0; j< block->number_of_nodes(); j++) {
1091 Node *n = block->get_node(j);
1092 int idx = n->_idx;
1093 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
1094 tty->print("latency:%3d ", get_latency_for_node(n));
1095 tty->print("%4d: %s\n", idx, n->Name());
1096 }
1097 }
1098 #endif
1099
1100 uint max_idx = (uint)ready_cnt.length();
1101 // Pull from worklist and schedule
1102 while( worklist.size() ) { // Worklist is not ready
1103
1104 #ifndef PRODUCT
1105 if (trace_opto_pipelining()) {
1106 tty->print("# ready list:");
1107 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1108 Node *n = worklist[i]; // Get Node on worklist
1109 tty->print(" %d", n->_idx);
1110 }
1111 tty->cr();
1112 }
1113 #endif
1114
1115 // Select and pop a ready guy from worklist
1116 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes);
1117 block->map_node(n, phi_cnt++); // Schedule him next
1118
1119 n->add_flag(Node::Flag_is_scheduled);
1120
1121 if (OptoRegScheduling && block_size_threshold_ok) {
1122 // Now adjust the resister pressure with the node we selected
1123 if (!n->is_Phi()) {
1124 adjust_register_pressure(n, block, recalc_pressure_nodes, true);
1125 }
1126 }
1127
1128 #ifndef PRODUCT
1129 if (trace_opto_pipelining()) {
1130 tty->print("# select %d: %s", n->_idx, n->Name());
1131 tty->print(", latency:%d", get_latency_for_node(n));
1132 n->dump();
1133 if (Verbose) {
1134 tty->print("# ready list:");
1135 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1136 Node *n = worklist[i]; // Get Node on worklist
1137 tty->print(" %d", n->_idx);
1138 }
1139 tty->cr();
1140 }
1141 }
1142
1143 #endif
1144 if( n->is_MachCall() ) {
1145 MachCallNode *mcall = n->as_MachCall();
1146 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
1147 continue;
1148 }
1149
1150 if (n->is_Mach() && n->as_Mach()->has_call()) {
1151 RegMask regs;
1152 regs.Insert(_matcher.c_frame_pointer());
1153 regs.OR(n->out_RegMask());
1154
1155 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
1156 map_node_to_block(proj, block);
1157 block->insert_node(proj, phi_cnt++);
1158
1159 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false, false);
1160 }
1161
1162 // Children are now all ready
1163 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
1164 Node* m = n->fast_out(i5); // Get user
1165 push_ready_nodes(n, m, block, ready_cnt, worklist, max_idx, 1);
1166 }
1167
1168 replace_uses_with_shenandoah_barrier(n, block, worklist, ready_cnt, max_idx, phi_cnt);
1169 }
1170
1171 if( phi_cnt != block->end_idx() ) {
1172 // did not schedule all. Retry, Bailout, or Die
1173 if (C->subsume_loads() == true && !C->failing()) {
1174 // Retry with subsume_loads == false
1175 // If this is the first failure, the sentinel string will "stick"
1176 // to the Compile object, and the C2Compiler will see it and retry.
1177 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1178 } else {
1179 assert(false, "graph should be schedulable");
1180 }
1181 // assert( phi_cnt == end_idx(), "did not schedule all" );
1182 return false;
1183 }
1184
1185 if (OptoRegScheduling && block_size_threshold_ok) {
1186 _regalloc->compute_exit_block_pressure(block);
1187 block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure();
1188 block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure();
1189 }
1190
1191 #ifndef PRODUCT
1192 if (trace_opto_pipelining()) {
1193 tty->print_cr("#");
1194 tty->print_cr("# after schedule_local");
1195 for (uint i = 0;i < block->number_of_nodes();i++) {
1196 tty->print("# ");
1197 block->get_node(i)->fast_dump();
1198 }
1199 tty->print_cr("# ");
1200
1201 if (OptoRegScheduling && block_size_threshold_ok) {
1202 tty->print_cr("# pressure info : %d", block->_pre_order);
1203 _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info");
1204 _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info");
1205 }
1206 tty->cr();
1207 }
1208 #endif
1209
1210 return true;
1211 }
1212
1213 //--------------------------catch_cleanup_fix_all_inputs-----------------------
1214 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
1215 for (uint l = 0; l < use->len(); l++) {
1216 if (use->in(l) == old_def) {
1217 if (l < use->req()) {
1218 use->set_req(l, new_def);
1219 } else {
1220 use->rm_prec(l);
1221 use->add_prec(new_def);
1222 l--;
1223 }
1224 }
1225 }
1226 }
1227
1228 //------------------------------catch_cleanup_find_cloned_def------------------
1229 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1230 assert( use_blk != def_blk, "Inter-block cleanup only");
1231
1232 // The use is some block below the Catch. Find and return the clone of the def
1233 // that dominates the use. If there is no clone in a dominating block, then
1234 // create a phi for the def in a dominating block.
1235
1236 // Find which successor block dominates this use. The successor
1237 // blocks must all be single-entry (from the Catch only; I will have
1238 // split blocks to make this so), hence they all dominate.
1239 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1240 use_blk = use_blk->_idom;
1241
1242 // Find the successor
1243 Node *fixup = NULL;
1244
1245 uint j;
1246 for( j = 0; j < def_blk->_num_succs; j++ )
1247 if( use_blk == def_blk->_succs[j] )
1248 break;
1249
1250 if( j == def_blk->_num_succs ) {
1251 // Block at same level in dom-tree is not a successor. It needs a
1252 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1253 Node_Array inputs = new Node_List(Thread::current()->resource_area());
1254 for(uint k = 1; k < use_blk->num_preds(); k++) {
1255 Block* block = get_block_for_node(use_blk->pred(k));
1256 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1257 }
1258
1259 // Check to see if the use_blk already has an identical phi inserted.
1260 // If it exists, it will be at the first position since all uses of a
1261 // def are processed together.
1262 Node *phi = use_blk->get_node(1);
1263 if( phi->is_Phi() ) {
1264 fixup = phi;
1265 for (uint k = 1; k < use_blk->num_preds(); k++) {
1266 if (phi->in(k) != inputs[k]) {
1267 // Not a match
1268 fixup = NULL;
1269 break;
1270 }
1271 }
1272 }
1273
1274 // If an existing PhiNode was not found, make a new one.
1275 if (fixup == NULL) {
1276 Node *new_phi = PhiNode::make(use_blk->head(), def);
1277 use_blk->insert_node(new_phi, 1);
1278 map_node_to_block(new_phi, use_blk);
1279 for (uint k = 1; k < use_blk->num_preds(); k++) {
1280 new_phi->set_req(k, inputs[k]);
1281 }
1282 fixup = new_phi;
1283 }
1284
1285 } else {
1286 // Found the use just below the Catch. Make it use the clone.
1287 fixup = use_blk->get_node(n_clone_idx);
1288 }
1289
1290 return fixup;
1291 }
1292
1293 //--------------------------catch_cleanup_intra_block--------------------------
1294 // Fix all input edges in use that reference "def". The use is in the same
1295 // block as the def and both have been cloned in each successor block.
1296 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1297
1298 // Both the use and def have been cloned. For each successor block,
1299 // get the clone of the use, and make its input the clone of the def
1300 // found in that block.
1301
1302 uint use_idx = blk->find_node(use);
1303 uint offset_idx = use_idx - beg;
1304 for( uint k = 0; k < blk->_num_succs; k++ ) {
1305 // Get clone in each successor block
1306 Block *sb = blk->_succs[k];
1307 Node *clone = sb->get_node(offset_idx+1);
1308 assert( clone->Opcode() == use->Opcode(), "" );
1309
1310 // Make use-clone reference the def-clone
1311 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1312 }
1313 }
1314
1315 //------------------------------catch_cleanup_inter_block---------------------
1316 // Fix all input edges in use that reference "def". The use is in a different
1317 // block than the def.
1318 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1319 if( !use_blk ) return; // Can happen if the use is a precedence edge
1320
1321 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1322 catch_cleanup_fix_all_inputs(use, def, new_def);
1323 }
1324
1325 //------------------------------call_catch_cleanup-----------------------------
1326 // If we inserted any instructions between a Call and his CatchNode,
1327 // clone the instructions on all paths below the Catch.
1328 void PhaseCFG::call_catch_cleanup(Block* block) {
1329
1330 // End of region to clone
1331 uint end = block->end_idx();
1332 if( !block->get_node(end)->is_Catch() ) return;
1333 // Start of region to clone
1334 uint beg = end;
1335 while(!block->get_node(beg-1)->is_MachProj() ||
1336 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1337 beg--;
1338 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1339 }
1340 // Range of inserted instructions is [beg, end)
1341 if( beg == end ) return;
1342
1343 // Clone along all Catch output paths. Clone area between the 'beg' and
1344 // 'end' indices.
1345 for( uint i = 0; i < block->_num_succs; i++ ) {
1346 Block *sb = block->_succs[i];
1347 // Clone the entire area; ignoring the edge fixup for now.
1348 for( uint j = end; j > beg; j-- ) {
1349 Node *clone = block->get_node(j-1)->clone();
1350 sb->insert_node(clone, 1);
1351 map_node_to_block(clone, sb);
1352 if (clone->needs_anti_dependence_check()) {
1353 insert_anti_dependences(sb, clone);
1354 }
1355 }
1356 }
1357
1358
1359 // Fixup edges. Check the def-use info per cloned Node
1360 for(uint i2 = beg; i2 < end; i2++ ) {
1361 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1362 Node *n = block->get_node(i2); // Node that got cloned
1363 // Need DU safe iterator because of edge manipulation in calls.
1364 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1365 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1366 out->push(n->fast_out(j1));
1367 }
1368 uint max = out->size();
1369 for (uint j = 0; j < max; j++) {// For all users
1370 Node *use = out->pop();
1371 Block *buse = get_block_for_node(use);
1372 if( use->is_Phi() ) {
1373 for( uint k = 1; k < use->req(); k++ )
1374 if( use->in(k) == n ) {
1375 Block* b = get_block_for_node(buse->pred(k));
1376 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1377 use->set_req(k, fixup);
1378 }
1379 } else {
1380 if (block == buse) {
1381 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1382 } else {
1383 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1384 }
1385 }
1386 } // End for all users
1387
1388 } // End of for all Nodes in cloned area
1389
1390 // Remove the now-dead cloned ops
1391 for(uint i3 = beg; i3 < end; i3++ ) {
1392 block->get_node(beg)->disconnect_inputs(NULL, C);
1393 block->remove_node(beg);
1394 }
1395
1396 // If the successor blocks have a CreateEx node, move it back to the top
1397 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1398 Block *sb = block->_succs[i4];
1399 uint new_cnt = end - beg;
1400 // Remove any newly created, but dead, nodes.
1401 for( uint j = new_cnt; j > 0; j-- ) {
1402 Node *n = sb->get_node(j);
1403 if (n->outcnt() == 0 &&
1404 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1405 n->disconnect_inputs(NULL, C);
1406 sb->remove_node(j);
1407 new_cnt--;
1408 }
1409 }
1410 // If any newly created nodes remain, move the CreateEx node to the top
1411 if (new_cnt > 0) {
1412 Node *cex = sb->get_node(1+new_cnt);
1413 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1414 sb->remove_node(1+new_cnt);
1415 sb->insert_node(cex, 1);
1416 }
1417 }
1418 }
1419 }