Wed, 10 Apr 2019 11:38:47 +0200
8221355: Performance regression after JDK-8155635 backport into 8u
Reviewed-by: shade, roland
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 */
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_sparc
40 # include "adfiles/ad_sparc.hpp"
41 #elif defined TARGET_ARCH_MODEL_zero
42 # include "adfiles/ad_zero.hpp"
43 #elif defined TARGET_ARCH_MODEL_ppc_64
44 # include "adfiles/ad_ppc_64.hpp"
45 #endif
47 // Optimization - Graph Style
49 // Check whether val is not-null-decoded compressed oop,
50 // i.e. will grab into the base of the heap if it represents NULL.
51 static bool accesses_heap_base_zone(Node *val) {
52 if (Universe::narrow_oop_base() != NULL) { // Implies UseCompressedOops.
53 if (val && val->is_Mach()) {
54 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
55 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
56 // decode NULL to point to the heap base (Decode_NN).
57 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
58 return true;
59 }
60 }
61 // Must recognize load operation with Decode matched in memory operand.
62 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected()
63 // returns true everywhere else. On PPC, no such memory operands
64 // exist, therefore we did not yet implement a check for such operands.
65 NOT_AIX(Unimplemented());
66 }
67 }
68 return false;
69 }
71 static bool needs_explicit_null_check_for_read(Node *val) {
72 // On some OSes (AIX) the page at address 0 is only write protected.
73 // If so, only Store operations will trap.
74 if (os::zero_page_read_protected()) {
75 return false; // Implicit null check will work.
76 }
77 // Also a read accessing the base of a heap-based compressed heap will trap.
78 if (accesses_heap_base_zone(val) && // Hits the base zone page.
79 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected.
80 return false;
81 }
83 return true;
84 }
86 //------------------------------implicit_null_check----------------------------
87 // Detect implicit-null-check opportunities. Basically, find NULL checks
88 // with suitable memory ops nearby. Use the memory op to do the NULL check.
89 // I can generate a memory op if there is not one nearby.
90 // The proj is the control projection for the not-null case.
91 // The val is the pointer being checked for nullness or
92 // decodeHeapOop_not_null node if it did not fold into address.
93 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
94 // Assume if null check need for 0 offset then always needed
95 // Intel solaris doesn't support any null checks yet and no
96 // mechanism exists (yet) to set the switches at an os_cpu level
97 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
99 // Make sure the ptr-is-null path appears to be uncommon!
100 float f = block->end()->as_MachIf()->_prob;
101 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
102 if( f > PROB_UNLIKELY_MAG(4) ) return;
104 uint bidx = 0; // Capture index of value into memop
105 bool was_store; // Memory op is a store op
107 // Get the successor block for if the test ptr is non-null
108 Block* not_null_block; // this one goes with the proj
109 Block* null_block;
110 if (block->get_node(block->number_of_nodes()-1) == proj) {
111 null_block = block->_succs[0];
112 not_null_block = block->_succs[1];
113 } else {
114 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
115 not_null_block = block->_succs[0];
116 null_block = block->_succs[1];
117 }
118 while (null_block->is_Empty() == Block::empty_with_goto) {
119 null_block = null_block->_succs[0];
120 }
122 // Search the exception block for an uncommon trap.
123 // (See Parse::do_if and Parse::do_ifnull for the reason
124 // we need an uncommon trap. Briefly, we need a way to
125 // detect failure of this optimization, as in 6366351.)
126 {
127 bool found_trap = false;
128 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
129 Node* nn = null_block->get_node(i1);
130 if (nn->is_MachCall() &&
131 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
132 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
133 if (trtype->isa_int() && trtype->is_int()->is_con()) {
134 jint tr_con = trtype->is_int()->get_con();
135 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
136 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
137 assert((int)reason < (int)BitsPerInt, "recode bit map");
138 if (is_set_nth_bit(allowed_reasons, (int) reason)
139 && action != Deoptimization::Action_none) {
140 // This uncommon trap is sure to recompile, eventually.
141 // When that happens, C->too_many_traps will prevent
142 // this transformation from happening again.
143 found_trap = true;
144 }
145 }
146 break;
147 }
148 }
149 if (!found_trap) {
150 // We did not find an uncommon trap.
151 return;
152 }
153 }
155 // Check for decodeHeapOop_not_null node which did not fold into address
156 bool is_decoden = ((intptr_t)val) & 1;
157 val = (Node*)(((intptr_t)val) & ~1);
159 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
160 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
162 // Search the successor block for a load or store who's base value is also
163 // the tested value. There may be several.
164 Node_List *out = new Node_List(Thread::current()->resource_area());
165 MachNode *best = NULL; // Best found so far
166 for (DUIterator i = val->outs(); val->has_out(i); i++) {
167 Node *m = val->out(i);
168 if( !m->is_Mach() ) continue;
169 MachNode *mach = m->as_Mach();
170 was_store = false;
171 int iop = mach->ideal_Opcode();
172 switch( iop ) {
173 case Op_LoadB:
174 case Op_LoadUB:
175 case Op_LoadUS:
176 case Op_LoadD:
177 case Op_LoadF:
178 case Op_LoadI:
179 case Op_LoadL:
180 case Op_LoadP:
181 case Op_LoadN:
182 case Op_LoadS:
183 case Op_LoadKlass:
184 case Op_LoadNKlass:
185 case Op_LoadRange:
186 case Op_LoadD_unaligned:
187 case Op_LoadL_unaligned:
188 assert(mach->in(2) == val, "should be address");
189 break;
190 case Op_StoreB:
191 case Op_StoreC:
192 case Op_StoreCM:
193 case Op_StoreD:
194 case Op_StoreF:
195 case Op_StoreI:
196 case Op_StoreL:
197 case Op_StoreP:
198 case Op_StoreN:
199 case Op_StoreNKlass:
200 was_store = true; // Memory op is a store op
201 // Stores will have their address in slot 2 (memory in slot 1).
202 // If the value being nul-checked is in another slot, it means we
203 // are storing the checked value, which does NOT check the value!
204 if( mach->in(2) != val ) continue;
205 break; // Found a memory op?
206 case Op_StrComp:
207 case Op_StrEquals:
208 case Op_StrIndexOf:
209 case Op_AryEq:
210 case Op_EncodeISOArray:
211 // Not a legit memory op for implicit null check regardless of
212 // embedded loads
213 continue;
214 default: // Also check for embedded loads
215 if( !mach->needs_anti_dependence_check() )
216 continue; // Not an memory op; skip it
217 if( must_clone[iop] ) {
218 // Do not move nodes which produce flags because
219 // RA will try to clone it to place near branch and
220 // it will cause recompilation, see clone_node().
221 continue;
222 }
223 {
224 // Check that value is used in memory address in
225 // instructions with embedded load (CmpP val1,(val2+off)).
226 Node* base;
227 Node* index;
228 const MachOper* oper = mach->memory_inputs(base, index);
229 if (oper == NULL || oper == (MachOper*)-1) {
230 continue; // Not an memory op; skip it
231 }
232 if (val == base ||
233 val == index && val->bottom_type()->isa_narrowoop()) {
234 break; // Found it
235 } else {
236 continue; // Skip it
237 }
238 }
239 break;
240 }
242 // On some OSes (AIX) the page at address 0 is only write protected.
243 // If so, only Store operations will trap.
244 // But a read accessing the base of a heap-based compressed heap will trap.
245 if (!was_store && needs_explicit_null_check_for_read(val)) {
246 continue;
247 }
249 // Check that node's control edge is not-null block's head or dominates it,
250 // otherwise we can't hoist it because there are other control dependencies.
251 Node* ctrl = mach->in(0);
252 if (ctrl != NULL && !(ctrl == not_null_block->head() ||
253 get_block_for_node(ctrl)->dominates(not_null_block))) {
254 continue;
255 }
257 // check if the offset is not too high for implicit exception
258 {
259 intptr_t offset = 0;
260 const TypePtr *adr_type = NULL; // Do not need this return value here
261 const Node* base = mach->get_base_and_disp(offset, adr_type);
262 if (base == NULL || base == NodeSentinel) {
263 // Narrow oop address doesn't have base, only index
264 if( val->bottom_type()->isa_narrowoop() &&
265 MacroAssembler::needs_explicit_null_check(offset) )
266 continue; // Give up if offset is beyond page size
267 // cannot reason about it; is probably not implicit null exception
268 } else {
269 const TypePtr* tptr;
270 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
271 Universe::narrow_klass_shift() == 0)) {
272 // 32-bits narrow oop can be the base of address expressions
273 tptr = base->get_ptr_type();
274 } else {
275 // only regular oops are expected here
276 tptr = base->bottom_type()->is_ptr();
277 }
278 // Give up if offset is not a compile-time constant
279 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
280 continue;
281 offset += tptr->_offset; // correct if base is offseted
282 if( MacroAssembler::needs_explicit_null_check(offset) )
283 continue; // Give up is reference is beyond 4K page size
284 }
285 }
287 // Check ctrl input to see if the null-check dominates the memory op
288 Block *cb = get_block_for_node(mach);
289 cb = cb->_idom; // Always hoist at least 1 block
290 if( !was_store ) { // Stores can be hoisted only one block
291 while( cb->_dom_depth > (block->_dom_depth + 1))
292 cb = cb->_idom; // Hoist loads as far as we want
293 // The non-null-block should dominate the memory op, too. Live
294 // range spilling will insert a spill in the non-null-block if it is
295 // needs to spill the memory op for an implicit null check.
296 if (cb->_dom_depth == (block->_dom_depth + 1)) {
297 if (cb != not_null_block) continue;
298 cb = cb->_idom;
299 }
300 }
301 if( cb != block ) continue;
303 // Found a memory user; see if it can be hoisted to check-block
304 uint vidx = 0; // Capture index of value into memop
305 uint j;
306 for( j = mach->req()-1; j > 0; j-- ) {
307 if( mach->in(j) == val ) {
308 vidx = j;
309 // Ignore DecodeN val which could be hoisted to where needed.
310 if( is_decoden ) continue;
311 }
312 // Block of memory-op input
313 Block *inb = get_block_for_node(mach->in(j));
314 Block *b = block; // Start from nul check
315 while( b != inb && b->_dom_depth > inb->_dom_depth )
316 b = b->_idom; // search upwards for input
317 // See if input dominates null check
318 if( b != inb )
319 break;
320 }
321 if( j > 0 )
322 continue;
323 Block *mb = get_block_for_node(mach);
324 // Hoisting stores requires more checks for the anti-dependence case.
325 // Give up hoisting if we have to move the store past any load.
326 if( was_store ) {
327 Block *b = mb; // Start searching here for a local load
328 // mach use (faulting) trying to hoist
329 // n might be blocker to hoisting
330 while( b != block ) {
331 uint k;
332 for( k = 1; k < b->number_of_nodes(); k++ ) {
333 Node *n = b->get_node(k);
334 if( n->needs_anti_dependence_check() &&
335 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
336 break; // Found anti-dependent load
337 }
338 if( k < b->number_of_nodes() )
339 break; // Found anti-dependent load
340 // Make sure control does not do a merge (would have to check allpaths)
341 if( b->num_preds() != 2 ) break;
342 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
343 }
344 if( b != block ) continue;
345 }
347 // Make sure this memory op is not already being used for a NullCheck
348 Node *e = mb->end();
349 if( e->is_MachNullCheck() && e->in(1) == mach )
350 continue; // Already being used as a NULL check
352 // Found a candidate! Pick one with least dom depth - the highest
353 // in the dom tree should be closest to the null check.
354 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
355 best = mach;
356 bidx = vidx;
357 }
358 }
359 // No candidate!
360 if (best == NULL) {
361 return;
362 }
364 // ---- Found an implicit null check
365 extern int implicit_null_checks;
366 implicit_null_checks++;
368 if( is_decoden ) {
369 // Check if we need to hoist decodeHeapOop_not_null first.
370 Block *valb = get_block_for_node(val);
371 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
372 // Hoist it up to the end of the test block.
373 valb->find_remove(val);
374 block->add_inst(val);
375 map_node_to_block(val, block);
376 // DecodeN on x86 may kill flags. Check for flag-killing projections
377 // that also need to be hoisted.
378 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
379 Node* n = val->fast_out(j);
380 if( n->is_MachProj() ) {
381 get_block_for_node(n)->find_remove(n);
382 block->add_inst(n);
383 map_node_to_block(n, block);
384 }
385 }
386 }
387 }
388 // Hoist the memory candidate up to the end of the test block.
389 Block *old_block = get_block_for_node(best);
390 old_block->find_remove(best);
391 block->add_inst(best);
392 map_node_to_block(best, block);
394 // Move the control dependence if it is pinned to not-null block.
395 // Don't change it in other cases: NULL or dominating control.
396 if (best->in(0) == not_null_block->head()) {
397 // Set it to control edge of null check.
398 best->set_req(0, proj->in(0)->in(0));
399 }
401 // Check for flag-killing projections that also need to be hoisted
402 // Should be DU safe because no edge updates.
403 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
404 Node* n = best->fast_out(j);
405 if( n->is_MachProj() ) {
406 get_block_for_node(n)->find_remove(n);
407 block->add_inst(n);
408 map_node_to_block(n, block);
409 }
410 }
412 // proj==Op_True --> ne test; proj==Op_False --> eq test.
413 // One of two graph shapes got matched:
414 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
415 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
416 // NULL checks are always branch-if-eq. If we see a IfTrue projection
417 // then we are replacing a 'ne' test with a 'eq' NULL check test.
418 // We need to flip the projections to keep the same semantics.
419 if( proj->Opcode() == Op_IfTrue ) {
420 // Swap order of projections in basic block to swap branch targets
421 Node *tmp1 = block->get_node(block->end_idx()+1);
422 Node *tmp2 = block->get_node(block->end_idx()+2);
423 block->map_node(tmp2, block->end_idx()+1);
424 block->map_node(tmp1, block->end_idx()+2);
425 Node *tmp = new (C) Node(C->top()); // Use not NULL input
426 tmp1->replace_by(tmp);
427 tmp2->replace_by(tmp1);
428 tmp->replace_by(tmp2);
429 tmp->destruct();
430 }
432 // Remove the existing null check; use a new implicit null check instead.
433 // Since schedule-local needs precise def-use info, we need to correct
434 // it as well.
435 Node *old_tst = proj->in(0);
436 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
437 block->map_node(nul_chk, block->end_idx());
438 map_node_to_block(nul_chk, block);
439 // Redirect users of old_test to nul_chk
440 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
441 old_tst->last_out(i2)->set_req(0, nul_chk);
442 // Clean-up any dead code
443 for (uint i3 = 0; i3 < old_tst->req(); i3++) {
444 Node* in = old_tst->in(i3);
445 old_tst->set_req(i3, NULL);
446 if (in->outcnt() == 0) {
447 // Remove dead input node
448 in->disconnect_inputs(NULL, C);
449 block->find_remove(in);
450 }
451 }
453 latency_from_uses(nul_chk);
454 latency_from_uses(best);
456 // insert anti-dependences to defs in this block
457 if (! best->needs_anti_dependence_check()) {
458 for (uint k = 1; k < block->number_of_nodes(); k++) {
459 Node *n = block->get_node(k);
460 if (n->needs_anti_dependence_check() &&
461 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) {
462 // Found anti-dependent load
463 insert_anti_dependences(block, n);
464 }
465 }
466 }
467 }
470 //------------------------------select-----------------------------------------
471 // Select a nice fellow from the worklist to schedule next. If there is only
472 // one choice, then use it. Projections take top priority for correctness
473 // reasons - if I see a projection, then it is next. There are a number of
474 // other special cases, for instructions that consume condition codes, et al.
475 // These are chosen immediately. Some instructions are required to immediately
476 // precede the last instruction in the block, and these are taken last. Of the
477 // remaining cases (most), choose the instruction with the greatest latency
478 // (that is, the most number of pseudo-cycles required to the end of the
479 // routine). If there is a tie, choose the instruction with the most inputs.
480 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
482 // If only a single entry on the stack, use it
483 uint cnt = worklist.size();
484 if (cnt == 1) {
485 Node *n = worklist[0];
486 worklist.map(0,worklist.pop());
487 return n;
488 }
490 uint choice = 0; // Bigger is most important
491 uint latency = 0; // Bigger is scheduled first
492 uint score = 0; // Bigger is better
493 int idx = -1; // Index in worklist
494 int cand_cnt = 0; // Candidate count
496 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
497 // Order in worklist is used to break ties.
498 // See caller for how this is used to delay scheduling
499 // of induction variable increments to after the other
500 // uses of the phi are scheduled.
501 Node *n = worklist[i]; // Get Node on worklist
503 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
504 if( n->is_Proj() || // Projections always win
505 n->Opcode()== Op_Con || // So does constant 'Top'
506 iop == Op_CreateEx || // Create-exception must start block
507 iop == Op_CheckCastPP
508 ) {
509 worklist.map(i,worklist.pop());
510 return n;
511 }
513 // Final call in a block must be adjacent to 'catch'
514 Node *e = block->end();
515 if( e->is_Catch() && e->in(0)->in(0) == n )
516 continue;
518 // Memory op for an implicit null check has to be at the end of the block
519 if( e->is_MachNullCheck() && e->in(1) == n )
520 continue;
522 // Schedule IV increment last.
523 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
524 e->in(1)->in(1) == n && n->is_iteratively_computed())
525 continue;
527 uint n_choice = 2;
529 // See if this instruction is consumed by a branch. If so, then (as the
530 // branch is the last instruction in the basic block) force it to the
531 // end of the basic block
532 if ( must_clone[iop] ) {
533 // See if any use is a branch
534 bool found_machif = false;
536 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
537 Node* use = n->fast_out(j);
539 // The use is a conditional branch, make them adjacent
540 if (use->is_MachIf() && get_block_for_node(use) == block) {
541 found_machif = true;
542 break;
543 }
545 // More than this instruction pending for successor to be ready,
546 // don't choose this if other opportunities are ready
547 if (ready_cnt.at(use->_idx) > 1)
548 n_choice = 1;
549 }
551 // loop terminated, prefer not to use this instruction
552 if (found_machif)
553 continue;
554 }
556 // See if this has a predecessor that is "must_clone", i.e. sets the
557 // condition code. If so, choose this first
558 for (uint j = 0; j < n->req() ; j++) {
559 Node *inn = n->in(j);
560 if (inn) {
561 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
562 n_choice = 3;
563 break;
564 }
565 }
566 }
568 // MachTemps should be scheduled last so they are near their uses
569 if (n->is_MachTemp()) {
570 n_choice = 1;
571 }
573 uint n_latency = get_latency_for_node(n);
574 uint n_score = n->req(); // Many inputs get high score to break ties
576 // Keep best latency found
577 cand_cnt++;
578 if (choice < n_choice ||
579 (choice == n_choice &&
580 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
581 (!StressLCM &&
582 (latency < n_latency ||
583 (latency == n_latency &&
584 (score < n_score))))))) {
585 choice = n_choice;
586 latency = n_latency;
587 score = n_score;
588 idx = i; // Also keep index in worklist
589 }
590 } // End of for all ready nodes in worklist
592 assert(idx >= 0, "index should be set");
593 Node *n = worklist[(uint)idx]; // Get the winner
595 worklist.map((uint)idx, worklist.pop()); // Compress worklist
596 return n;
597 }
600 //------------------------------set_next_call----------------------------------
601 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
602 if( next_call.test_set(n->_idx) ) return;
603 for( uint i=0; i<n->len(); i++ ) {
604 Node *m = n->in(i);
605 if( !m ) continue; // must see all nodes in block that precede call
606 if (get_block_for_node(m) == block) {
607 set_next_call(block, m, next_call);
608 }
609 }
610 }
612 //------------------------------needed_for_next_call---------------------------
613 // Set the flag 'next_call' for each Node that is needed for the next call to
614 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
615 // next subroutine call get priority - basically it moves things NOT needed
616 // for the next call till after the call. This prevents me from trying to
617 // carry lots of stuff live across a call.
618 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
619 // Find the next control-defining Node in this block
620 Node* call = NULL;
621 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
622 Node* m = this_call->fast_out(i);
623 if (get_block_for_node(m) == block && // Local-block user
624 m != this_call && // Not self-start node
625 m->is_MachCall()) {
626 call = m;
627 break;
628 }
629 }
630 if (call == NULL) return; // No next call (e.g., block end is near)
631 // Set next-call for all inputs to this call
632 set_next_call(block, call, next_call);
633 }
635 //------------------------------add_call_kills-------------------------------------
636 // helper function that adds caller save registers to MachProjNode
637 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
638 // Fill in the kill mask for the call
639 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
640 if( !regs.Member(r) ) { // Not already defined by the call
641 // Save-on-call register?
642 if ((save_policy[r] == 'C') ||
643 (save_policy[r] == 'A') ||
644 ((save_policy[r] == 'E') && exclude_soe)) {
645 proj->_rout.Insert(r);
646 }
647 }
648 }
649 }
652 //------------------------------sched_call-------------------------------------
653 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
654 RegMask regs;
656 // Schedule all the users of the call right now. All the users are
657 // projection Nodes, so they must be scheduled next to the call.
658 // Collect all the defined registers.
659 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
660 Node* n = mcall->fast_out(i);
661 assert( n->is_MachProj(), "" );
662 int n_cnt = ready_cnt.at(n->_idx)-1;
663 ready_cnt.at_put(n->_idx, n_cnt);
664 assert( n_cnt == 0, "" );
665 // Schedule next to call
666 block->map_node(n, node_cnt++);
667 // Collect defined registers
668 regs.OR(n->out_RegMask());
669 // Check for scheduling the next control-definer
670 if( n->bottom_type() == Type::CONTROL )
671 // Warm up next pile of heuristic bits
672 needed_for_next_call(block, n, next_call);
674 // Children of projections are now all ready
675 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
676 Node* m = n->fast_out(j); // Get user
677 if(get_block_for_node(m) != block) {
678 continue;
679 }
680 if( m->is_Phi() ) continue;
681 int m_cnt = ready_cnt.at(m->_idx)-1;
682 ready_cnt.at_put(m->_idx, m_cnt);
683 if( m_cnt == 0 )
684 worklist.push(m);
685 }
687 }
689 // Act as if the call defines the Frame Pointer.
690 // Certainly the FP is alive and well after the call.
691 regs.Insert(_matcher.c_frame_pointer());
693 // Set all registers killed and not already defined by the call.
694 uint r_cnt = mcall->tf()->range()->cnt();
695 int op = mcall->ideal_Opcode();
696 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
697 map_node_to_block(proj, block);
698 block->insert_node(proj, node_cnt++);
700 // Select the right register save policy.
701 const char *save_policy = NULL;
702 switch (op) {
703 case Op_CallRuntime:
704 case Op_CallLeaf:
705 case Op_CallLeafNoFP:
706 // Calling C code so use C calling convention
707 save_policy = _matcher._c_reg_save_policy;
708 break;
710 case Op_CallStaticJava:
711 case Op_CallDynamicJava:
712 // Calling Java code so use Java calling convention
713 save_policy = _matcher._register_save_policy;
714 break;
716 default:
717 ShouldNotReachHere();
718 }
720 // When using CallRuntime mark SOE registers as killed by the call
721 // so values that could show up in the RegisterMap aren't live in a
722 // callee saved register since the register wouldn't know where to
723 // find them. CallLeaf and CallLeafNoFP are ok because they can't
724 // have debug info on them. Strictly speaking this only needs to be
725 // done for oops since idealreg2debugmask takes care of debug info
726 // references but there no way to handle oops differently than other
727 // pointers as far as the kill mask goes.
728 bool exclude_soe = op == Op_CallRuntime;
730 // If the call is a MethodHandle invoke, we need to exclude the
731 // register which is used to save the SP value over MH invokes from
732 // the mask. Otherwise this register could be used for
733 // deoptimization information.
734 if (op == Op_CallStaticJava) {
735 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
736 if (mcallstaticjava->_method_handle_invoke)
737 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
738 }
740 add_call_kills(proj, regs, save_policy, exclude_soe);
742 return node_cnt;
743 }
746 //------------------------------schedule_local---------------------------------
747 // Topological sort within a block. Someday become a real scheduler.
748 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
749 // Already "sorted" are the block start Node (as the first entry), and
750 // the block-ending Node and any trailing control projections. We leave
751 // these alone. PhiNodes and ParmNodes are made to follow the block start
752 // Node. Everything else gets topo-sorted.
754 #ifndef PRODUCT
755 if (trace_opto_pipelining()) {
756 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
757 for (uint i = 0;i < block->number_of_nodes(); i++) {
758 tty->print("# ");
759 block->get_node(i)->fast_dump();
760 }
761 tty->print_cr("#");
762 }
763 #endif
765 // RootNode is already sorted
766 if (block->number_of_nodes() == 1) {
767 return true;
768 }
770 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
771 uint node_cnt = block->end_idx();
772 uint phi_cnt = 1;
773 uint i;
774 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
775 Node *n = block->get_node(i);
776 if( n->is_Phi() || // Found a PhiNode or ParmNode
777 (n->is_Proj() && n->in(0) == block->head()) ) {
778 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
779 block->map_node(block->get_node(phi_cnt), i);
780 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
781 } else { // All others
782 // Count block-local inputs to 'n'
783 uint cnt = n->len(); // Input count
784 uint local = 0;
785 for( uint j=0; j<cnt; j++ ) {
786 Node *m = n->in(j);
787 if( m && get_block_for_node(m) == block && !m->is_top() )
788 local++; // One more block-local input
789 }
790 ready_cnt.at_put(n->_idx, local); // Count em up
792 #ifdef ASSERT
793 if( UseConcMarkSweepGC || UseG1GC ) {
794 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
795 // Check the precedence edges
796 for (uint prec = n->req(); prec < n->len(); prec++) {
797 Node* oop_store = n->in(prec);
798 if (oop_store != NULL) {
799 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
800 }
801 }
802 }
803 }
804 #endif
806 // A few node types require changing a required edge to a precedence edge
807 // before allocation.
808 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
809 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
810 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
811 // MemBarAcquire could be created without Precedent edge.
812 // del_req() replaces the specified edge with the last input edge
813 // and then removes the last edge. If the specified edge > number of
814 // edges the last edge will be moved outside of the input edges array
815 // and the edge will be lost. This is why this code should be
816 // executed only when Precedent (== TypeFunc::Parms) edge is present.
817 Node *x = n->in(TypeFunc::Parms);
818 n->del_req(TypeFunc::Parms);
819 n->add_prec(x);
820 }
821 }
822 }
823 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
824 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
826 // All the prescheduled guys do not hold back internal nodes
827 uint i3;
828 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
829 Node *n = block->get_node(i3); // Get pre-scheduled
830 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
831 Node* m = n->fast_out(j);
832 if (get_block_for_node(m) == block) { // Local-block user
833 int m_cnt = ready_cnt.at(m->_idx)-1;
834 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
835 }
836 }
837 }
839 Node_List delay;
840 // Make a worklist
841 Node_List worklist;
842 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
843 Node *m = block->get_node(i4);
844 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
845 if (m->is_iteratively_computed()) {
846 // Push induction variable increments last to allow other uses
847 // of the phi to be scheduled first. The select() method breaks
848 // ties in scheduling by worklist order.
849 delay.push(m);
850 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
851 // Force the CreateEx to the top of the list so it's processed
852 // first and ends up at the start of the block.
853 worklist.insert(0, m);
854 } else {
855 worklist.push(m); // Then on to worklist!
856 }
857 }
858 }
859 while (delay.size()) {
860 Node* d = delay.pop();
861 worklist.push(d);
862 }
864 // Warm up the 'next_call' heuristic bits
865 needed_for_next_call(block, block->head(), next_call);
867 #ifndef PRODUCT
868 if (trace_opto_pipelining()) {
869 for (uint j=0; j< block->number_of_nodes(); j++) {
870 Node *n = block->get_node(j);
871 int idx = n->_idx;
872 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
873 tty->print("latency:%3d ", get_latency_for_node(n));
874 tty->print("%4d: %s\n", idx, n->Name());
875 }
876 }
877 #endif
879 uint max_idx = (uint)ready_cnt.length();
880 // Pull from worklist and schedule
881 while( worklist.size() ) { // Worklist is not ready
883 #ifndef PRODUCT
884 if (trace_opto_pipelining()) {
885 tty->print("# ready list:");
886 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
887 Node *n = worklist[i]; // Get Node on worklist
888 tty->print(" %d", n->_idx);
889 }
890 tty->cr();
891 }
892 #endif
894 // Select and pop a ready guy from worklist
895 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
896 block->map_node(n, phi_cnt++); // Schedule him next
898 #ifndef PRODUCT
899 if (trace_opto_pipelining()) {
900 tty->print("# select %d: %s", n->_idx, n->Name());
901 tty->print(", latency:%d", get_latency_for_node(n));
902 n->dump();
903 if (Verbose) {
904 tty->print("# ready list:");
905 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
906 Node *n = worklist[i]; // Get Node on worklist
907 tty->print(" %d", n->_idx);
908 }
909 tty->cr();
910 }
911 }
913 #endif
914 if( n->is_MachCall() ) {
915 MachCallNode *mcall = n->as_MachCall();
916 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
917 continue;
918 }
920 if (n->is_Mach() && n->as_Mach()->has_call()) {
921 RegMask regs;
922 regs.Insert(_matcher.c_frame_pointer());
923 regs.OR(n->out_RegMask());
925 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
926 map_node_to_block(proj, block);
927 block->insert_node(proj, phi_cnt++);
929 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
930 }
932 // Children are now all ready
933 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
934 Node* m = n->fast_out(i5); // Get user
935 if (get_block_for_node(m) != block) {
936 continue;
937 }
938 if( m->is_Phi() ) continue;
939 if (m->_idx >= max_idx) { // new node, skip it
940 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
941 continue;
942 }
943 int m_cnt = ready_cnt.at(m->_idx)-1;
944 ready_cnt.at_put(m->_idx, m_cnt);
945 if( m_cnt == 0 )
946 worklist.push(m);
947 }
948 }
950 if( phi_cnt != block->end_idx() ) {
951 // did not schedule all. Retry, Bailout, or Die
952 if (C->subsume_loads() == true && !C->failing()) {
953 // Retry with subsume_loads == false
954 // If this is the first failure, the sentinel string will "stick"
955 // to the Compile object, and the C2Compiler will see it and retry.
956 C->record_failure(C2Compiler::retry_no_subsuming_loads());
957 }
958 // assert( phi_cnt == end_idx(), "did not schedule all" );
959 return false;
960 }
962 #ifndef PRODUCT
963 if (trace_opto_pipelining()) {
964 tty->print_cr("#");
965 tty->print_cr("# after schedule_local");
966 for (uint i = 0;i < block->number_of_nodes();i++) {
967 tty->print("# ");
968 block->get_node(i)->fast_dump();
969 }
970 tty->cr();
971 }
972 #endif
975 return true;
976 }
978 //--------------------------catch_cleanup_fix_all_inputs-----------------------
979 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
980 for (uint l = 0; l < use->len(); l++) {
981 if (use->in(l) == old_def) {
982 if (l < use->req()) {
983 use->set_req(l, new_def);
984 } else {
985 use->rm_prec(l);
986 use->add_prec(new_def);
987 l--;
988 }
989 }
990 }
991 }
993 //------------------------------catch_cleanup_find_cloned_def------------------
994 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
995 assert( use_blk != def_blk, "Inter-block cleanup only");
997 // The use is some block below the Catch. Find and return the clone of the def
998 // that dominates the use. If there is no clone in a dominating block, then
999 // create a phi for the def in a dominating block.
1001 // Find which successor block dominates this use. The successor
1002 // blocks must all be single-entry (from the Catch only; I will have
1003 // split blocks to make this so), hence they all dominate.
1004 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1005 use_blk = use_blk->_idom;
1007 // Find the successor
1008 Node *fixup = NULL;
1010 uint j;
1011 for( j = 0; j < def_blk->_num_succs; j++ )
1012 if( use_blk == def_blk->_succs[j] )
1013 break;
1015 if( j == def_blk->_num_succs ) {
1016 // Block at same level in dom-tree is not a successor. It needs a
1017 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1018 Node_Array inputs = new Node_List(Thread::current()->resource_area());
1019 for(uint k = 1; k < use_blk->num_preds(); k++) {
1020 Block* block = get_block_for_node(use_blk->pred(k));
1021 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1022 }
1024 // Check to see if the use_blk already has an identical phi inserted.
1025 // If it exists, it will be at the first position since all uses of a
1026 // def are processed together.
1027 Node *phi = use_blk->get_node(1);
1028 if( phi->is_Phi() ) {
1029 fixup = phi;
1030 for (uint k = 1; k < use_blk->num_preds(); k++) {
1031 if (phi->in(k) != inputs[k]) {
1032 // Not a match
1033 fixup = NULL;
1034 break;
1035 }
1036 }
1037 }
1039 // If an existing PhiNode was not found, make a new one.
1040 if (fixup == NULL) {
1041 Node *new_phi = PhiNode::make(use_blk->head(), def);
1042 use_blk->insert_node(new_phi, 1);
1043 map_node_to_block(new_phi, use_blk);
1044 for (uint k = 1; k < use_blk->num_preds(); k++) {
1045 new_phi->set_req(k, inputs[k]);
1046 }
1047 fixup = new_phi;
1048 }
1050 } else {
1051 // Found the use just below the Catch. Make it use the clone.
1052 fixup = use_blk->get_node(n_clone_idx);
1053 }
1055 return fixup;
1056 }
1058 //--------------------------catch_cleanup_intra_block--------------------------
1059 // Fix all input edges in use that reference "def". The use is in the same
1060 // block as the def and both have been cloned in each successor block.
1061 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1063 // Both the use and def have been cloned. For each successor block,
1064 // get the clone of the use, and make its input the clone of the def
1065 // found in that block.
1067 uint use_idx = blk->find_node(use);
1068 uint offset_idx = use_idx - beg;
1069 for( uint k = 0; k < blk->_num_succs; k++ ) {
1070 // Get clone in each successor block
1071 Block *sb = blk->_succs[k];
1072 Node *clone = sb->get_node(offset_idx+1);
1073 assert( clone->Opcode() == use->Opcode(), "" );
1075 // Make use-clone reference the def-clone
1076 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1077 }
1078 }
1080 //------------------------------catch_cleanup_inter_block---------------------
1081 // Fix all input edges in use that reference "def". The use is in a different
1082 // block than the def.
1083 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1084 if( !use_blk ) return; // Can happen if the use is a precedence edge
1086 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1087 catch_cleanup_fix_all_inputs(use, def, new_def);
1088 }
1090 //------------------------------call_catch_cleanup-----------------------------
1091 // If we inserted any instructions between a Call and his CatchNode,
1092 // clone the instructions on all paths below the Catch.
1093 void PhaseCFG::call_catch_cleanup(Block* block) {
1095 // End of region to clone
1096 uint end = block->end_idx();
1097 if( !block->get_node(end)->is_Catch() ) return;
1098 // Start of region to clone
1099 uint beg = end;
1100 while(!block->get_node(beg-1)->is_MachProj() ||
1101 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1102 beg--;
1103 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1104 }
1105 // Range of inserted instructions is [beg, end)
1106 if( beg == end ) return;
1108 // Clone along all Catch output paths. Clone area between the 'beg' and
1109 // 'end' indices.
1110 for( uint i = 0; i < block->_num_succs; i++ ) {
1111 Block *sb = block->_succs[i];
1112 // Clone the entire area; ignoring the edge fixup for now.
1113 for( uint j = end; j > beg; j-- ) {
1114 Node *clone = block->get_node(j-1)->clone();
1115 sb->insert_node(clone, 1);
1116 map_node_to_block(clone, sb);
1117 if (clone->needs_anti_dependence_check()) {
1118 insert_anti_dependences(sb, clone);
1119 }
1120 }
1121 }
1124 // Fixup edges. Check the def-use info per cloned Node
1125 for(uint i2 = beg; i2 < end; i2++ ) {
1126 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1127 Node *n = block->get_node(i2); // Node that got cloned
1128 // Need DU safe iterator because of edge manipulation in calls.
1129 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1130 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1131 out->push(n->fast_out(j1));
1132 }
1133 uint max = out->size();
1134 for (uint j = 0; j < max; j++) {// For all users
1135 Node *use = out->pop();
1136 Block *buse = get_block_for_node(use);
1137 if( use->is_Phi() ) {
1138 for( uint k = 1; k < use->req(); k++ )
1139 if( use->in(k) == n ) {
1140 Block* b = get_block_for_node(buse->pred(k));
1141 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1142 use->set_req(k, fixup);
1143 }
1144 } else {
1145 if (block == buse) {
1146 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1147 } else {
1148 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1149 }
1150 }
1151 } // End for all users
1153 } // End of for all Nodes in cloned area
1155 // Remove the now-dead cloned ops
1156 for(uint i3 = beg; i3 < end; i3++ ) {
1157 block->get_node(beg)->disconnect_inputs(NULL, C);
1158 block->remove_node(beg);
1159 }
1161 // If the successor blocks have a CreateEx node, move it back to the top
1162 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1163 Block *sb = block->_succs[i4];
1164 uint new_cnt = end - beg;
1165 // Remove any newly created, but dead, nodes.
1166 for( uint j = new_cnt; j > 0; j-- ) {
1167 Node *n = sb->get_node(j);
1168 if (n->outcnt() == 0 &&
1169 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1170 n->disconnect_inputs(NULL, C);
1171 sb->remove_node(j);
1172 new_cnt--;
1173 }
1174 }
1175 // If any newly created nodes remain, move the CreateEx node to the top
1176 if (new_cnt > 0) {
1177 Node *cex = sb->get_node(1+new_cnt);
1178 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1179 sb->remove_node(1+new_cnt);
1180 sb->insert_node(cex, 1);
1181 }
1182 }
1183 }
1184 }