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/block.hpp"
28 #include "opto/c2compiler.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/machnode.hpp"
32 #include "opto/runtime.hpp"
33 #if defined AD_MD_HPP
34 # include AD_MD_HPP
35 #elif defined TARGET_ARCH_MODEL_x86_32
36 # include "adfiles/ad_x86_32.hpp"
37 #elif defined TARGET_ARCH_MODEL_x86_64
38 # include "adfiles/ad_x86_64.hpp"
39 #elif defined TARGET_ARCH_MODEL_aarch64
40 # include "adfiles/ad_aarch64.hpp"
41 #elif defined TARGET_ARCH_MODEL_sparc
42 # include "adfiles/ad_sparc.hpp"
43 #elif defined TARGET_ARCH_MODEL_zero
44 # include "adfiles/ad_zero.hpp"
45 #elif defined TARGET_ARCH_MODEL_ppc_64
46 # include "adfiles/ad_ppc_64.hpp"
47 #endif
48
49 // Optimization - Graph Style
50
51 // Check whether val is not-null-decoded compressed oop,
52 // i.e. will grab into the base of the heap if it represents NULL.
53 static bool accesses_heap_base_zone(Node *val) {
54 if (Universe::narrow_oop_base() != NULL) { // Implies UseCompressedOops.
55 if (val && val->is_Mach()) {
56 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
57 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
58 // decode NULL to point to the heap base (Decode_NN).
59 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
60 return true;
61 }
62 }
63 // Must recognize load operation with Decode matched in memory operand.
64 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected()
65 // returns true everywhere else. On PPC, no such memory operands
66 // exist, therefore we did not yet implement a check for such operands.
67 NOT_AIX(Unimplemented());
68 }
69 }
70 return false;
71 }
72
73 static bool needs_explicit_null_check_for_read(Node *val) {
74 // On some OSes (AIX) the page at address 0 is only write protected.
75 // If so, only Store operations will trap.
76 if (os::zero_page_read_protected()) {
77 return false; // Implicit null check will work.
78 }
79 // Also a read accessing the base of a heap-based compressed heap will trap.
80 if (accesses_heap_base_zone(val) && // Hits the base zone page.
81 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected.
82 return false;
83 }
84
85 return true;
86 }
87
88 //------------------------------implicit_null_check----------------------------
89 // Detect implicit-null-check opportunities. Basically, find NULL checks
90 // with suitable memory ops nearby. Use the memory op to do the NULL check.
91 // I can generate a memory op if there is not one nearby.
92 // The proj is the control projection for the not-null case.
93 // The val is the pointer being checked for nullness or
94 // decodeHeapOop_not_null node if it did not fold into address.
95 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
96 // Assume if null check need for 0 offset then always needed
97 // Intel solaris doesn't support any null checks yet and no
98 // mechanism exists (yet) to set the switches at an os_cpu level
99 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
100
101 // Make sure the ptr-is-null path appears to be uncommon!
102 float f = block->end()->as_MachIf()->_prob;
103 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
104 if( f > PROB_UNLIKELY_MAG(4) ) return;
105
106 uint bidx = 0; // Capture index of value into memop
107 bool was_store; // Memory op is a store op
108
109 // Get the successor block for if the test ptr is non-null
110 Block* not_null_block; // this one goes with the proj
111 Block* null_block;
112 if (block->get_node(block->number_of_nodes()-1) == proj) {
113 null_block = block->_succs[0];
114 not_null_block = block->_succs[1];
115 } else {
116 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
117 not_null_block = block->_succs[0];
118 null_block = block->_succs[1];
119 }
120 while (null_block->is_Empty() == Block::empty_with_goto) {
121 null_block = null_block->_succs[0];
122 }
123
124 // Search the exception block for an uncommon trap.
125 // (See Parse::do_if and Parse::do_ifnull for the reason
126 // we need an uncommon trap. Briefly, we need a way to
127 // detect failure of this optimization, as in 6366351.)
128 {
129 bool found_trap = false;
130 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
131 Node* nn = null_block->get_node(i1);
132 if (nn->is_MachCall() &&
133 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
134 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
135 if (trtype->isa_int() && trtype->is_int()->is_con()) {
136 jint tr_con = trtype->is_int()->get_con();
137 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
138 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
139 assert((int)reason < (int)BitsPerInt, "recode bit map");
140 if (is_set_nth_bit(allowed_reasons, (int) reason)
141 && action != Deoptimization::Action_none) {
142 // This uncommon trap is sure to recompile, eventually.
143 // When that happens, C->too_many_traps will prevent
144 // this transformation from happening again.
145 found_trap = true;
146 }
147 }
148 break;
149 }
150 }
151 if (!found_trap) {
152 // We did not find an uncommon trap.
153 return;
154 }
155 }
156
157 // Check for decodeHeapOop_not_null node which did not fold into address
158 bool is_decoden = ((intptr_t)val) & 1;
159 val = (Node*)(((intptr_t)val) & ~1);
160
161 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
162 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
163
164 // Search the successor block for a load or store who's base value is also
165 // the tested value. There may be several.
166 Node_List *out = new Node_List(Thread::current()->resource_area());
167 MachNode *best = NULL; // Best found so far
168 for (DUIterator i = val->outs(); val->has_out(i); i++) {
169 Node *m = val->out(i);
170 if( !m->is_Mach() ) continue;
171 MachNode *mach = m->as_Mach();
172 was_store = false;
173 int iop = mach->ideal_Opcode();
174 switch( iop ) {
175 case Op_LoadB:
176 case Op_LoadUB:
177 case Op_LoadUS:
178 case Op_LoadD:
179 case Op_LoadF:
180 case Op_LoadI:
181 case Op_LoadL:
182 case Op_LoadP:
183 case Op_LoadN:
184 case Op_LoadS:
185 case Op_LoadKlass:
186 case Op_LoadNKlass:
187 case Op_LoadRange:
188 case Op_LoadD_unaligned:
189 case Op_LoadL_unaligned:
190 case Op_ShenandoahReadBarrier:
191 assert(mach->in(2) == val, "should be address");
192 break;
193 case Op_StoreB:
194 case Op_StoreC:
195 case Op_StoreCM:
196 case Op_StoreD:
197 case Op_StoreF:
198 case Op_StoreI:
199 case Op_StoreL:
200 case Op_StoreP:
201 case Op_StoreN:
202 case Op_StoreNKlass:
203 was_store = true; // Memory op is a store op
204 // Stores will have their address in slot 2 (memory in slot 1).
205 // If the value being nul-checked is in another slot, it means we
206 // are storing the checked value, which does NOT check the value!
207 if( mach->in(2) != val ) continue;
208 break; // Found a memory op?
209 case Op_StrComp:
210 case Op_StrEquals:
211 case Op_StrIndexOf:
212 case Op_AryEq:
213 case Op_EncodeISOArray:
214 // Not a legit memory op for implicit null check regardless of
215 // embedded loads
216 continue;
217 default: // Also check for embedded loads
218 if( !mach->needs_anti_dependence_check() )
219 continue; // Not an memory op; skip it
220 if( must_clone[iop] ) {
221 // Do not move nodes which produce flags because
222 // RA will try to clone it to place near branch and
223 // it will cause recompilation, see clone_node().
224 continue;
225 }
226 {
227 // Check that value is used in memory address in
228 // instructions with embedded load (CmpP val1,(val2+off)).
229 Node* base;
230 Node* index;
231 const MachOper* oper = mach->memory_inputs(base, index);
232 if (oper == NULL || oper == (MachOper*)-1) {
233 continue; // Not an memory op; skip it
234 }
235 if (val == base ||
236 val == index && val->bottom_type()->isa_narrowoop()) {
237 break; // Found it
238 } else {
239 continue; // Skip it
240 }
241 }
242 break;
243 }
244
245 // On some OSes (AIX) the page at address 0 is only write protected.
246 // If so, only Store operations will trap.
247 // But a read accessing the base of a heap-based compressed heap will trap.
248 if (!was_store && needs_explicit_null_check_for_read(val)) {
249 continue;
250 }
251
252 // Check that node's control edge is not-null block's head or dominates it,
253 // otherwise we can't hoist it because there are other control dependencies.
254 Node* ctrl = mach->in(0);
255 if (ctrl != NULL && !(ctrl == not_null_block->head() ||
256 get_block_for_node(ctrl)->dominates(not_null_block))) {
257 continue;
258 }
259
260 // check if the offset is not too high for implicit exception
261 {
262 intptr_t offset = 0;
263 const TypePtr *adr_type = NULL; // Do not need this return value here
264 const Node* base = mach->get_base_and_disp(offset, adr_type);
265 if (base == NULL || base == NodeSentinel) {
266 // Narrow oop address doesn't have base, only index
267 if( val->bottom_type()->isa_narrowoop() &&
268 MacroAssembler::needs_explicit_null_check(offset) )
269 continue; // Give up if offset is beyond page size
270 // cannot reason about it; is probably not implicit null exception
271 } else {
272 const TypePtr* tptr;
273 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
274 Universe::narrow_klass_shift() == 0)) {
275 // 32-bits narrow oop can be the base of address expressions
276 tptr = base->get_ptr_type();
277 } else {
278 // only regular oops are expected here
279 tptr = base->bottom_type()->is_ptr();
280 }
281 // Give up if offset is not a compile-time constant
282 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
283 continue;
284 offset += tptr->_offset; // correct if base is offseted
285 if( MacroAssembler::needs_explicit_null_check(offset) )
286 continue; // Give up is reference is beyond 4K page size
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 extern int implicit_null_checks;
369 implicit_null_checks++;
370
371 if( is_decoden ) {
372 // Check if we need to hoist decodeHeapOop_not_null first.
373 Block *valb = get_block_for_node(val);
374 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
375 // Hoist it up to the end of the test block.
376 valb->find_remove(val);
377 block->add_inst(val);
378 map_node_to_block(val, block);
379 // DecodeN on x86 may kill flags. Check for flag-killing projections
380 // that also need to be hoisted.
381 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
382 Node* n = val->fast_out(j);
383 if( n->is_MachProj() ) {
384 get_block_for_node(n)->find_remove(n);
385 block->add_inst(n);
386 map_node_to_block(n, block);
387 }
388 }
389 }
390 }
391 // Hoist the memory candidate up to the end of the test block.
392 Block *old_block = get_block_for_node(best);
393 old_block->find_remove(best);
394 block->add_inst(best);
395 map_node_to_block(best, block);
396
397 // Move the control dependence if it is pinned to not-null block.
398 // Don't change it in other cases: NULL or dominating control.
399 if (best->in(0) == not_null_block->head()) {
400 // Set it to control edge of null check.
401 best->set_req(0, proj->in(0)->in(0));
402 }
403
404 // Check for flag-killing projections that also need to be hoisted
405 // Should be DU safe because no edge updates.
406 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
407 Node* n = best->fast_out(j);
408 if( n->is_MachProj() ) {
409 get_block_for_node(n)->find_remove(n);
410 block->add_inst(n);
411 map_node_to_block(n, block);
412 }
413 }
414
415 // proj==Op_True --> ne test; proj==Op_False --> eq test.
416 // One of two graph shapes got matched:
417 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
418 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
419 // NULL checks are always branch-if-eq. If we see a IfTrue projection
420 // then we are replacing a 'ne' test with a 'eq' NULL check test.
421 // We need to flip the projections to keep the same semantics.
422 if( proj->Opcode() == Op_IfTrue ) {
423 // Swap order of projections in basic block to swap branch targets
424 Node *tmp1 = block->get_node(block->end_idx()+1);
425 Node *tmp2 = block->get_node(block->end_idx()+2);
426 block->map_node(tmp2, block->end_idx()+1);
427 block->map_node(tmp1, block->end_idx()+2);
428 Node *tmp = new (C) Node(C->top()); // Use not NULL input
429 tmp1->replace_by(tmp);
430 tmp2->replace_by(tmp1);
431 tmp->replace_by(tmp2);
432 tmp->destruct();
433 }
434
435 // Remove the existing null check; use a new implicit null check instead.
436 // Since schedule-local needs precise def-use info, we need to correct
437 // it as well.
438 Node *old_tst = proj->in(0);
439 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
440 block->map_node(nul_chk, block->end_idx());
441 map_node_to_block(nul_chk, block);
442 // Redirect users of old_test to nul_chk
443 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
444 old_tst->last_out(i2)->set_req(0, nul_chk);
445 // Clean-up any dead code
446 for (uint i3 = 0; i3 < old_tst->req(); i3++) {
447 Node* in = old_tst->in(i3);
448 old_tst->set_req(i3, NULL);
449 if (in->outcnt() == 0) {
450 // Remove dead input node
451 in->disconnect_inputs(NULL, C);
452 block->find_remove(in);
453 }
454 }
455
456 latency_from_uses(nul_chk);
457 latency_from_uses(best);
458
459 // insert anti-dependences to defs in this block
460 if (! best->needs_anti_dependence_check()) {
461 for (uint k = 1; k < block->number_of_nodes(); k++) {
462 Node *n = block->get_node(k);
463 if (n->needs_anti_dependence_check() &&
464 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) {
465 // Found anti-dependent load
466 insert_anti_dependences(block, n);
467 }
468 }
469 }
470 }
471
472
473 //------------------------------select-----------------------------------------
474 // Select a nice fellow from the worklist to schedule next. If there is only
475 // one choice, then use it. Projections take top priority for correctness
476 // reasons - if I see a projection, then it is next. There are a number of
477 // other special cases, for instructions that consume condition codes, et al.
478 // These are chosen immediately. Some instructions are required to immediately
479 // precede the last instruction in the block, and these are taken last. Of the
480 // remaining cases (most), choose the instruction with the greatest latency
481 // (that is, the most number of pseudo-cycles required to the end of the
482 // routine). If there is a tie, choose the instruction with the most inputs.
483 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
484
485 // If only a single entry on the stack, use it
486 uint cnt = worklist.size();
487 if (cnt == 1) {
488 Node *n = worklist[0];
489 worklist.map(0,worklist.pop());
490 return n;
491 }
492
493 uint choice = 0; // Bigger is most important
494 uint latency = 0; // Bigger is scheduled first
495 uint score = 0; // Bigger is better
496 int idx = -1; // Index in worklist
497 int cand_cnt = 0; // Candidate count
498
499 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
500 // Order in worklist is used to break ties.
501 // See caller for how this is used to delay scheduling
502 // of induction variable increments to after the other
503 // uses of the phi are scheduled.
504 Node *n = worklist[i]; // Get Node on worklist
505
506 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
507 if( n->is_Proj() || // Projections always win
508 n->Opcode()== Op_Con || // So does constant 'Top'
509 iop == Op_CreateEx || // Create-exception must start block
510 iop == Op_CheckCastPP
511 ) {
512 worklist.map(i,worklist.pop());
513 return n;
514 }
515
516 // Final call in a block must be adjacent to 'catch'
517 Node *e = block->end();
518 if( e->is_Catch() && e->in(0)->in(0) == n )
519 continue;
520
521 // Memory op for an implicit null check has to be at the end of the block
522 if( e->is_MachNullCheck() && e->in(1) == n )
523 continue;
524
525 // Schedule IV increment last.
526 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
527 e->in(1)->in(1) == n && n->is_iteratively_computed())
528 continue;
529
530 uint n_choice = 2;
531
532 // See if this instruction is consumed by a branch. If so, then (as the
533 // branch is the last instruction in the basic block) force it to the
534 // end of the basic block
535 if ( must_clone[iop] ) {
536 // See if any use is a branch
537 bool found_machif = false;
538
539 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
540 Node* use = n->fast_out(j);
541
542 // The use is a conditional branch, make them adjacent
543 if (use->is_MachIf() && get_block_for_node(use) == block) {
544 found_machif = true;
545 break;
546 }
547
548 // More than this instruction pending for successor to be ready,
549 // don't choose this if other opportunities are ready
550 if (ready_cnt.at(use->_idx) > 1)
551 n_choice = 1;
552 }
553
554 // loop terminated, prefer not to use this instruction
555 if (found_machif)
556 continue;
557 }
558
559 // See if this has a predecessor that is "must_clone", i.e. sets the
560 // condition code. If so, choose this first
561 for (uint j = 0; j < n->req() ; j++) {
562 Node *inn = n->in(j);
563 if (inn) {
564 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
565 n_choice = 3;
566 break;
567 }
568 }
569 }
570
571 // MachTemps should be scheduled last so they are near their uses
572 if (n->is_MachTemp()) {
573 n_choice = 1;
574 }
575
576 uint n_latency = get_latency_for_node(n);
577 uint n_score = n->req(); // Many inputs get high score to break ties
578
579 // Keep best latency found
580 cand_cnt++;
581 if (choice < n_choice ||
582 (choice == n_choice &&
583 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
584 (!StressLCM &&
585 (latency < n_latency ||
586 (latency == n_latency &&
587 (score < n_score))))))) {
588 choice = n_choice;
589 latency = n_latency;
590 score = n_score;
591 idx = i; // Also keep index in worklist
592 }
593 } // End of for all ready nodes in worklist
594
595 assert(idx >= 0, "index should be set");
596 Node *n = worklist[(uint)idx]; // Get the winner
597
598 worklist.map((uint)idx, worklist.pop()); // Compress worklist
599 return n;
600 }
601
602
603 //------------------------------set_next_call----------------------------------
604 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
605 if( next_call.test_set(n->_idx) ) return;
606 for( uint i=0; i<n->len(); i++ ) {
607 Node *m = n->in(i);
608 if( !m ) continue; // must see all nodes in block that precede call
609 if (get_block_for_node(m) == block) {
610 set_next_call(block, m, next_call);
611 }
612 }
613 }
614
615 //------------------------------needed_for_next_call---------------------------
616 // Set the flag 'next_call' for each Node that is needed for the next call to
617 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
618 // next subroutine call get priority - basically it moves things NOT needed
619 // for the next call till after the call. This prevents me from trying to
620 // carry lots of stuff live across a call.
621 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
622 // Find the next control-defining Node in this block
623 Node* call = NULL;
624 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
625 Node* m = this_call->fast_out(i);
626 if (get_block_for_node(m) == block && // Local-block user
627 m != this_call && // Not self-start node
628 m->is_MachCall()) {
629 call = m;
630 break;
631 }
632 }
633 if (call == NULL) return; // No next call (e.g., block end is near)
634 // Set next-call for all inputs to this call
635 set_next_call(block, call, next_call);
636 }
637
638 //------------------------------add_call_kills-------------------------------------
639 // helper function that adds caller save registers to MachProjNode
640 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe, bool exclude_fp) {
641 // Fill in the kill mask for the call
642 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
643 if (exclude_fp && (register_save_type[r] == Op_RegF || register_save_type[r] == Op_RegD)) {
644 continue;
645 }
646 if( !regs.Member(r) ) { // Not already defined by the call
647 // Save-on-call register?
648 if ((save_policy[r] == 'C') ||
649 (save_policy[r] == 'A') ||
650 ((save_policy[r] == 'E') && exclude_soe)) {
651 proj->_rout.Insert(r);
652 }
653 }
654 }
655 }
656
657
658 //------------------------------sched_call-------------------------------------
659 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
660 RegMask regs;
661
662 // Schedule all the users of the call right now. All the users are
663 // projection Nodes, so they must be scheduled next to the call.
664 // Collect all the defined registers.
665 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
666 Node* n = mcall->fast_out(i);
667 assert( n->is_MachProj(), "" );
668 int n_cnt = ready_cnt.at(n->_idx)-1;
669 ready_cnt.at_put(n->_idx, n_cnt);
670 assert( n_cnt == 0, "" );
671 // Schedule next to call
672 block->map_node(n, node_cnt++);
673 // Collect defined registers
674 regs.OR(n->out_RegMask());
675 // Check for scheduling the next control-definer
676 if( n->bottom_type() == Type::CONTROL )
677 // Warm up next pile of heuristic bits
678 needed_for_next_call(block, n, next_call);
679
680 // Children of projections are now all ready
681 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
682 Node* m = n->fast_out(j); // Get user
683 if(get_block_for_node(m) != block) {
684 continue;
685 }
686 if( m->is_Phi() ) continue;
687 int m_cnt = ready_cnt.at(m->_idx)-1;
688 ready_cnt.at_put(m->_idx, m_cnt);
689 if( m_cnt == 0 )
690 worklist.push(m);
691 }
692
693 }
694
695 // Act as if the call defines the Frame Pointer.
696 // Certainly the FP is alive and well after the call.
697 regs.Insert(_matcher.c_frame_pointer());
698
699 // Set all registers killed and not already defined by the call.
700 uint r_cnt = mcall->tf()->range()->cnt();
701 int op = mcall->ideal_Opcode();
702 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
703 map_node_to_block(proj, block);
704 block->insert_node(proj, node_cnt++);
705
706 // Select the right register save policy.
707 const char *save_policy = NULL;
708 switch (op) {
709 case Op_CallRuntime:
710 case Op_CallLeaf:
711 case Op_CallLeafNoFP:
712 // Calling C code so use C calling convention
713 save_policy = _matcher._c_reg_save_policy;
714 break;
715
716 case Op_CallStaticJava:
717 case Op_CallDynamicJava:
718 // Calling Java code so use Java calling convention
719 save_policy = _matcher._register_save_policy;
720 break;
721
722 default:
723 ShouldNotReachHere();
724 }
725
726 // When using CallRuntime mark SOE registers as killed by the call
727 // so values that could show up in the RegisterMap aren't live in a
728 // callee saved register since the register wouldn't know where to
729 // find them. CallLeaf and CallLeafNoFP are ok because they can't
730 // have debug info on them. Strictly speaking this only needs to be
731 // done for oops since idealreg2debugmask takes care of debug info
732 // references but there no way to handle oops differently than other
733 // pointers as far as the kill mask goes.
734 bool exclude_soe = op == Op_CallRuntime;
735
736 // If the call is a MethodHandle invoke, we need to exclude the
737 // register which is used to save the SP value over MH invokes from
738 // the mask. Otherwise this register could be used for
739 // deoptimization information.
740 if (op == Op_CallStaticJava) {
741 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
742 if (mcallstaticjava->_method_handle_invoke)
743 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
744 }
745
746 if (UseShenandoahGC && mcall->entry_point() == StubRoutines::shenandoah_wb_C()) {
747 assert(op == Op_CallLeafNoFP, "shenandoah_wb_C should be called with Op_CallLeafNoFP");
748 add_call_kills(proj, regs, save_policy, exclude_soe, true);
749 } else {
750 add_call_kills(proj, regs, save_policy, exclude_soe, false);
751 }
752
753 return node_cnt;
754 }
755
756
757 //------------------------------schedule_local---------------------------------
758 // Topological sort within a block. Someday become a real scheduler.
759 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
760 // Already "sorted" are the block start Node (as the first entry), and
761 // the block-ending Node and any trailing control projections. We leave
762 // these alone. PhiNodes and ParmNodes are made to follow the block start
763 // Node. Everything else gets topo-sorted.
764
765 #ifndef PRODUCT
766 if (trace_opto_pipelining()) {
767 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
768 for (uint i = 0;i < block->number_of_nodes(); i++) {
769 tty->print("# ");
770 block->get_node(i)->fast_dump();
771 }
772 tty->print_cr("#");
773 }
774 #endif
775
776 // RootNode is already sorted
777 if (block->number_of_nodes() == 1) {
778 return true;
779 }
780
781 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
782 uint node_cnt = block->end_idx();
783 uint phi_cnt = 1;
784 uint i;
785 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
786 Node *n = block->get_node(i);
787 if( n->is_Phi() || // Found a PhiNode or ParmNode
788 (n->is_Proj() && n->in(0) == block->head()) ) {
789 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
790 block->map_node(block->get_node(phi_cnt), i);
791 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
792 } else { // All others
793 // Count block-local inputs to 'n'
794 uint cnt = n->len(); // Input count
795 uint local = 0;
796 for( uint j=0; j<cnt; j++ ) {
797 Node *m = n->in(j);
798 if( m && get_block_for_node(m) == block && !m->is_top() )
799 local++; // One more block-local input
800 }
801 ready_cnt.at_put(n->_idx, local); // Count em up
802
803 #ifdef ASSERT
804 if( UseConcMarkSweepGC || UseG1GC ) {
805 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
806 // Check the precedence edges
807 for (uint prec = n->req(); prec < n->len(); prec++) {
808 Node* oop_store = n->in(prec);
809 if (oop_store != NULL) {
810 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
811 }
812 }
813 }
814 }
815 #endif
816
817 // A few node types require changing a required edge to a precedence edge
818 // before allocation.
819 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
820 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
821 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
822 // MemBarAcquire could be created without Precedent edge.
823 // del_req() replaces the specified edge with the last input edge
824 // and then removes the last edge. If the specified edge > number of
825 // edges the last edge will be moved outside of the input edges array
826 // and the edge will be lost. This is why this code should be
827 // executed only when Precedent (== TypeFunc::Parms) edge is present.
828 Node *x = n->in(TypeFunc::Parms);
829 n->del_req(TypeFunc::Parms);
830 n->add_prec(x);
831 }
832 }
833 }
834 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
835 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
836
837 // All the prescheduled guys do not hold back internal nodes
838 uint i3;
839 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
840 Node *n = block->get_node(i3); // Get pre-scheduled
841 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
842 Node* m = n->fast_out(j);
843 if (get_block_for_node(m) == block) { // Local-block user
844 int m_cnt = ready_cnt.at(m->_idx)-1;
845 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
846 }
847 }
848 }
849
850 Node_List delay;
851 // Make a worklist
852 Node_List worklist;
853 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
854 Node *m = block->get_node(i4);
855 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
856 if (m->is_iteratively_computed()) {
857 // Push induction variable increments last to allow other uses
858 // of the phi to be scheduled first. The select() method breaks
859 // ties in scheduling by worklist order.
860 delay.push(m);
861 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
862 // Force the CreateEx to the top of the list so it's processed
863 // first and ends up at the start of the block.
864 worklist.insert(0, m);
865 } else {
866 worklist.push(m); // Then on to worklist!
867 }
868 }
869 }
870 while (delay.size()) {
871 Node* d = delay.pop();
872 worklist.push(d);
873 }
874
875 // Warm up the 'next_call' heuristic bits
876 needed_for_next_call(block, block->head(), next_call);
877
878 #ifndef PRODUCT
879 if (trace_opto_pipelining()) {
880 for (uint j=0; j< block->number_of_nodes(); j++) {
881 Node *n = block->get_node(j);
882 int idx = n->_idx;
883 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
884 tty->print("latency:%3d ", get_latency_for_node(n));
885 tty->print("%4d: %s\n", idx, n->Name());
886 }
887 }
888 #endif
889
890 uint max_idx = (uint)ready_cnt.length();
891 // Pull from worklist and schedule
892 while( worklist.size() ) { // Worklist is not ready
893
894 #ifndef PRODUCT
895 if (trace_opto_pipelining()) {
896 tty->print("# ready list:");
897 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
898 Node *n = worklist[i]; // Get Node on worklist
899 tty->print(" %d", n->_idx);
900 }
901 tty->cr();
902 }
903 #endif
904
905 // Select and pop a ready guy from worklist
906 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
907 block->map_node(n, phi_cnt++); // Schedule him next
908
909 #ifndef PRODUCT
910 if (trace_opto_pipelining()) {
911 tty->print("# select %d: %s", n->_idx, n->Name());
912 tty->print(", latency:%d", get_latency_for_node(n));
913 n->dump();
914 if (Verbose) {
915 tty->print("# ready list:");
916 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
917 Node *n = worklist[i]; // Get Node on worklist
918 tty->print(" %d", n->_idx);
919 }
920 tty->cr();
921 }
922 }
923
924 #endif
925 if( n->is_MachCall() ) {
926 MachCallNode *mcall = n->as_MachCall();
927 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
928 continue;
929 }
930
931 if (n->is_Mach() && n->as_Mach()->has_call()) {
932 RegMask regs;
933 regs.Insert(_matcher.c_frame_pointer());
934 regs.OR(n->out_RegMask());
935
936 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
937 map_node_to_block(proj, block);
938 block->insert_node(proj, phi_cnt++);
939
940 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false, false);
941 }
942
943 // Children are now all ready
944 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
945 Node* m = n->fast_out(i5); // Get user
946 if (get_block_for_node(m) != block) {
947 continue;
948 }
949 if( m->is_Phi() ) continue;
950 if (m->_idx >= max_idx) { // new node, skip it
951 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
952 continue;
953 }
954 int m_cnt = ready_cnt.at(m->_idx)-1;
955 ready_cnt.at_put(m->_idx, m_cnt);
956 if( m_cnt == 0 )
957 worklist.push(m);
958 }
959 }
960
961 if( phi_cnt != block->end_idx() ) {
962 // did not schedule all. Retry, Bailout, or Die
963 if (C->subsume_loads() == true && !C->failing()) {
964 // Retry with subsume_loads == false
965 // If this is the first failure, the sentinel string will "stick"
966 // to the Compile object, and the C2Compiler will see it and retry.
967 C->record_failure(C2Compiler::retry_no_subsuming_loads());
968 } else {
969 assert(false, "graph should be schedulable");
970 }
971 // assert( phi_cnt == end_idx(), "did not schedule all" );
972 return false;
973 }
974
975 #ifndef PRODUCT
976 if (trace_opto_pipelining()) {
977 tty->print_cr("#");
978 tty->print_cr("# after schedule_local");
979 for (uint i = 0;i < block->number_of_nodes();i++) {
980 tty->print("# ");
981 block->get_node(i)->fast_dump();
982 }
983 tty->cr();
984 }
985 #endif
986
987
988 return true;
989 }
990
991 //--------------------------catch_cleanup_fix_all_inputs-----------------------
992 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
993 for (uint l = 0; l < use->len(); l++) {
994 if (use->in(l) == old_def) {
995 if (l < use->req()) {
996 use->set_req(l, new_def);
997 } else {
998 use->rm_prec(l);
999 use->add_prec(new_def);
1000 l--;
1001 }
1002 }
1003 }
1004 }
1005
1006 //------------------------------catch_cleanup_find_cloned_def------------------
1007 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1008 assert( use_blk != def_blk, "Inter-block cleanup only");
1009
1010 // The use is some block below the Catch. Find and return the clone of the def
1011 // that dominates the use. If there is no clone in a dominating block, then
1012 // create a phi for the def in a dominating block.
1013
1014 // Find which successor block dominates this use. The successor
1015 // blocks must all be single-entry (from the Catch only; I will have
1016 // split blocks to make this so), hence they all dominate.
1017 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1018 use_blk = use_blk->_idom;
1019
1020 // Find the successor
1021 Node *fixup = NULL;
1022
1023 uint j;
1024 for( j = 0; j < def_blk->_num_succs; j++ )
1025 if( use_blk == def_blk->_succs[j] )
1026 break;
1027
1028 if( j == def_blk->_num_succs ) {
1029 // Block at same level in dom-tree is not a successor. It needs a
1030 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1031 Node_Array inputs = new Node_List(Thread::current()->resource_area());
1032 for(uint k = 1; k < use_blk->num_preds(); k++) {
1033 Block* block = get_block_for_node(use_blk->pred(k));
1034 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1035 }
1036
1037 // Check to see if the use_blk already has an identical phi inserted.
1038 // If it exists, it will be at the first position since all uses of a
1039 // def are processed together.
1040 Node *phi = use_blk->get_node(1);
1041 if( phi->is_Phi() ) {
1042 fixup = phi;
1043 for (uint k = 1; k < use_blk->num_preds(); k++) {
1044 if (phi->in(k) != inputs[k]) {
1045 // Not a match
1046 fixup = NULL;
1047 break;
1048 }
1049 }
1050 }
1051
1052 // If an existing PhiNode was not found, make a new one.
1053 if (fixup == NULL) {
1054 Node *new_phi = PhiNode::make(use_blk->head(), def);
1055 use_blk->insert_node(new_phi, 1);
1056 map_node_to_block(new_phi, use_blk);
1057 for (uint k = 1; k < use_blk->num_preds(); k++) {
1058 new_phi->set_req(k, inputs[k]);
1059 }
1060 fixup = new_phi;
1061 }
1062
1063 } else {
1064 // Found the use just below the Catch. Make it use the clone.
1065 fixup = use_blk->get_node(n_clone_idx);
1066 }
1067
1068 return fixup;
1069 }
1070
1071 //--------------------------catch_cleanup_intra_block--------------------------
1072 // Fix all input edges in use that reference "def". The use is in the same
1073 // block as the def and both have been cloned in each successor block.
1074 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1075
1076 // Both the use and def have been cloned. For each successor block,
1077 // get the clone of the use, and make its input the clone of the def
1078 // found in that block.
1079
1080 uint use_idx = blk->find_node(use);
1081 uint offset_idx = use_idx - beg;
1082 for( uint k = 0; k < blk->_num_succs; k++ ) {
1083 // Get clone in each successor block
1084 Block *sb = blk->_succs[k];
1085 Node *clone = sb->get_node(offset_idx+1);
1086 assert( clone->Opcode() == use->Opcode(), "" );
1087
1088 // Make use-clone reference the def-clone
1089 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1090 }
1091 }
1092
1093 //------------------------------catch_cleanup_inter_block---------------------
1094 // Fix all input edges in use that reference "def". The use is in a different
1095 // block than the def.
1096 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1097 if( !use_blk ) return; // Can happen if the use is a precedence edge
1098
1099 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1100 catch_cleanup_fix_all_inputs(use, def, new_def);
1101 }
1102
1103 //------------------------------call_catch_cleanup-----------------------------
1104 // If we inserted any instructions between a Call and his CatchNode,
1105 // clone the instructions on all paths below the Catch.
1106 void PhaseCFG::call_catch_cleanup(Block* block) {
1107
1108 // End of region to clone
1109 uint end = block->end_idx();
1110 if( !block->get_node(end)->is_Catch() ) return;
1111 // Start of region to clone
1112 uint beg = end;
1113 while(!block->get_node(beg-1)->is_MachProj() ||
1114 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1115 beg--;
1116 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1117 }
1118 // Range of inserted instructions is [beg, end)
1119 if( beg == end ) return;
1120
1121 // Clone along all Catch output paths. Clone area between the 'beg' and
1122 // 'end' indices.
1123 for( uint i = 0; i < block->_num_succs; i++ ) {
1124 Block *sb = block->_succs[i];
1125 // Clone the entire area; ignoring the edge fixup for now.
1126 for( uint j = end; j > beg; j-- ) {
1127 Node *clone = block->get_node(j-1)->clone();
1128 sb->insert_node(clone, 1);
1129 map_node_to_block(clone, sb);
1130 if (clone->needs_anti_dependence_check()) {
1131 insert_anti_dependences(sb, clone);
1132 }
1133 }
1134 }
1135
1136
1137 // Fixup edges. Check the def-use info per cloned Node
1138 for(uint i2 = beg; i2 < end; i2++ ) {
1139 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1140 Node *n = block->get_node(i2); // Node that got cloned
1141 // Need DU safe iterator because of edge manipulation in calls.
1142 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1143 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1144 out->push(n->fast_out(j1));
1145 }
1146 uint max = out->size();
1147 for (uint j = 0; j < max; j++) {// For all users
1148 Node *use = out->pop();
1149 Block *buse = get_block_for_node(use);
1150 if( use->is_Phi() ) {
1151 for( uint k = 1; k < use->req(); k++ )
1152 if( use->in(k) == n ) {
1153 Block* b = get_block_for_node(buse->pred(k));
1154 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1155 use->set_req(k, fixup);
1156 }
1157 } else {
1158 if (block == buse) {
1159 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1160 } else {
1161 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1162 }
1163 }
1164 } // End for all users
1165
1166 } // End of for all Nodes in cloned area
1167
1168 // Remove the now-dead cloned ops
1169 for(uint i3 = beg; i3 < end; i3++ ) {
1170 block->get_node(beg)->disconnect_inputs(NULL, C);
1171 block->remove_node(beg);
1172 }
1173
1174 // If the successor blocks have a CreateEx node, move it back to the top
1175 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1176 Block *sb = block->_succs[i4];
1177 uint new_cnt = end - beg;
1178 // Remove any newly created, but dead, nodes.
1179 for( uint j = new_cnt; j > 0; j-- ) {
1180 Node *n = sb->get_node(j);
1181 if (n->outcnt() == 0 &&
1182 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1183 n->disconnect_inputs(NULL, C);
1184 sb->remove_node(j);
1185 new_cnt--;
1186 }
1187 }
1188 // If any newly created nodes remain, move the CreateEx node to the top
1189 if (new_cnt > 0) {
1190 Node *cex = sb->get_node(1+new_cnt);
1191 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1192 sb->remove_node(1+new_cnt);
1193 sb->insert_node(cex, 1);
1194 }
1195 }
1196 }
1197 }