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