Sat, 24 Oct 2020 16:43:47 +0800
Merge
1 /*
2 * Copyright (c) 1998, 2016, 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 /*
26 * This file has been modified by Loongson Technology in 2015. These
27 * modifications are Copyright (c) 2015 Loongson Technology, and are made
28 * available on the same license terms set forth above.
29 */
31 #include "precompiled.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/block.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/runtime.hpp"
39 #if defined AD_MD_HPP
40 # include AD_MD_HPP
41 #elif defined TARGET_ARCH_MODEL_x86_32
42 # include "adfiles/ad_x86_32.hpp"
43 #elif defined TARGET_ARCH_MODEL_x86_64
44 # include "adfiles/ad_x86_64.hpp"
45 #elif defined TARGET_ARCH_MODEL_sparc
46 # include "adfiles/ad_sparc.hpp"
47 #elif defined TARGET_ARCH_MODEL_zero
48 # include "adfiles/ad_zero.hpp"
49 #elif defined TARGET_ARCH_MODEL_ppc_64
50 # include "adfiles/ad_ppc_64.hpp"
51 #elif defined TARGET_ARCH_MODEL_mips_64
52 # include "adfiles/ad_mips_64.hpp"
53 #endif
55 // Optimization - Graph Style
57 // Check whether val is not-null-decoded compressed oop,
58 // i.e. will grab into the base of the heap if it represents NULL.
59 static bool accesses_heap_base_zone(Node *val) {
60 if (Universe::narrow_oop_base() != NULL) { // Implies UseCompressedOops.
61 if (val && val->is_Mach()) {
62 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
63 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
64 // decode NULL to point to the heap base (Decode_NN).
65 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
66 return true;
67 }
68 }
69 // Must recognize load operation with Decode matched in memory operand.
70 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected()
71 // returns true everywhere else. On PPC, no such memory operands
72 // exist, therefore we did not yet implement a check for such operands.
73 NOT_AIX(Unimplemented());
74 }
75 }
76 return false;
77 }
79 static bool needs_explicit_null_check_for_read(Node *val) {
80 // On some OSes (AIX) the page at address 0 is only write protected.
81 // If so, only Store operations will trap.
82 if (os::zero_page_read_protected()) {
83 return false; // Implicit null check will work.
84 }
85 // Also a read accessing the base of a heap-based compressed heap will trap.
86 if (accesses_heap_base_zone(val) && // Hits the base zone page.
87 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected.
88 return false;
89 }
91 return true;
92 }
94 //------------------------------implicit_null_check----------------------------
95 // Detect implicit-null-check opportunities. Basically, find NULL checks
96 // with suitable memory ops nearby. Use the memory op to do the NULL check.
97 // I can generate a memory op if there is not one nearby.
98 // The proj is the control projection for the not-null case.
99 // The val is the pointer being checked for nullness or
100 // decodeHeapOop_not_null node if it did not fold into address.
101 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
102 // Assume if null check need for 0 offset then always needed
103 // Intel solaris doesn't support any null checks yet and no
104 // mechanism exists (yet) to set the switches at an os_cpu level
105 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
107 // Make sure the ptr-is-null path appears to be uncommon!
108 float f = block->end()->as_MachIf()->_prob;
109 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
110 if( f > PROB_UNLIKELY_MAG(4) ) return;
112 uint bidx = 0; // Capture index of value into memop
113 bool was_store; // Memory op is a store op
115 // Get the successor block for if the test ptr is non-null
116 Block* not_null_block; // this one goes with the proj
117 Block* null_block;
118 if (block->get_node(block->number_of_nodes()-1) == proj) {
119 null_block = block->_succs[0];
120 not_null_block = block->_succs[1];
121 } else {
122 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
123 not_null_block = block->_succs[0];
124 null_block = block->_succs[1];
125 }
126 while (null_block->is_Empty() == Block::empty_with_goto) {
127 null_block = null_block->_succs[0];
128 }
130 // Search the exception block for an uncommon trap.
131 // (See Parse::do_if and Parse::do_ifnull for the reason
132 // we need an uncommon trap. Briefly, we need a way to
133 // detect failure of this optimization, as in 6366351.)
134 {
135 bool found_trap = false;
136 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
137 Node* nn = null_block->get_node(i1);
138 if (nn->is_MachCall() &&
139 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
140 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
141 if (trtype->isa_int() && trtype->is_int()->is_con()) {
142 jint tr_con = trtype->is_int()->get_con();
143 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
144 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
145 assert((int)reason < (int)BitsPerInt, "recode bit map");
146 if (is_set_nth_bit(allowed_reasons, (int) reason)
147 && action != Deoptimization::Action_none) {
148 // This uncommon trap is sure to recompile, eventually.
149 // When that happens, C->too_many_traps will prevent
150 // this transformation from happening again.
151 found_trap = true;
152 }
153 }
154 break;
155 }
156 }
157 if (!found_trap) {
158 // We did not find an uncommon trap.
159 return;
160 }
161 }
163 // Check for decodeHeapOop_not_null node which did not fold into address
164 bool is_decoden = ((intptr_t)val) & 1;
165 val = (Node*)(((intptr_t)val) & ~1);
167 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
168 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
170 // Search the successor block for a load or store who's base value is also
171 // the tested value. There may be several.
172 Node_List *out = new Node_List(Thread::current()->resource_area());
173 MachNode *best = NULL; // Best found so far
174 for (DUIterator i = val->outs(); val->has_out(i); i++) {
175 Node *m = val->out(i);
176 if( !m->is_Mach() ) continue;
177 MachNode *mach = m->as_Mach();
178 was_store = false;
179 int iop = mach->ideal_Opcode();
180 switch( iop ) {
181 case Op_LoadB:
182 case Op_LoadUB:
183 case Op_LoadUS:
184 case Op_LoadD:
185 case Op_LoadF:
186 case Op_LoadI:
187 case Op_LoadL:
188 case Op_LoadP:
189 case Op_LoadN:
190 case Op_LoadS:
191 case Op_LoadKlass:
192 case Op_LoadNKlass:
193 case Op_LoadRange:
194 case Op_LoadD_unaligned:
195 case Op_LoadL_unaligned:
196 assert(mach->in(2) == val, "should be address");
197 break;
198 case Op_StoreB:
199 case Op_StoreC:
200 case Op_StoreCM:
201 case Op_StoreD:
202 case Op_StoreF:
203 case Op_StoreI:
204 case Op_StoreL:
205 case Op_StoreP:
206 case Op_StoreN:
207 case Op_StoreNKlass:
208 was_store = true; // Memory op is a store op
209 // Stores will have their address in slot 2 (memory in slot 1).
210 // If the value being nul-checked is in another slot, it means we
211 // are storing the checked value, which does NOT check the value!
212 if( mach->in(2) != val ) continue;
213 break; // Found a memory op?
214 case Op_StrComp:
215 case Op_StrEquals:
216 case Op_StrIndexOf:
217 case Op_AryEq:
218 case Op_EncodeISOArray:
219 // Not a legit memory op for implicit null check regardless of
220 // embedded loads
221 continue;
222 default: // Also check for embedded loads
223 if( !mach->needs_anti_dependence_check() )
224 continue; // Not an memory op; skip it
225 if( must_clone[iop] ) {
226 // Do not move nodes which produce flags because
227 // RA will try to clone it to place near branch and
228 // it will cause recompilation, see clone_node().
229 continue;
230 }
231 {
232 // Check that value is used in memory address in
233 // instructions with embedded load (CmpP val1,(val2+off)).
234 Node* base;
235 Node* index;
236 const MachOper* oper = mach->memory_inputs(base, index);
237 if (oper == NULL || oper == (MachOper*)-1) {
238 continue; // Not an memory op; skip it
239 }
240 if (val == base ||
241 val == index && val->bottom_type()->isa_narrowoop()) {
242 break; // Found it
243 } else {
244 continue; // Skip it
245 }
246 }
247 break;
248 }
250 // On some OSes (AIX) the page at address 0 is only write protected.
251 // If so, only Store operations will trap.
252 // But a read accessing the base of a heap-based compressed heap will trap.
253 if (!was_store && needs_explicit_null_check_for_read(val)) {
254 continue;
255 }
257 // Check that node's control edge is not-null block's head or dominates it,
258 // otherwise we can't hoist it because there are other control dependencies.
259 Node* ctrl = mach->in(0);
260 if (ctrl != NULL && !(ctrl == not_null_block->head() ||
261 get_block_for_node(ctrl)->dominates(not_null_block))) {
262 continue;
263 }
265 // check if the offset is not too high for implicit exception
266 {
267 intptr_t offset = 0;
268 const TypePtr *adr_type = NULL; // Do not need this return value here
269 const Node* base = mach->get_base_and_disp(offset, adr_type);
270 if (base == NULL || base == NodeSentinel) {
271 // Narrow oop address doesn't have base, only index
272 if( val->bottom_type()->isa_narrowoop() &&
273 MacroAssembler::needs_explicit_null_check(offset) )
274 continue; // Give up if offset is beyond page size
275 // cannot reason about it; is probably not implicit null exception
276 } else {
277 const TypePtr* tptr;
278 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
279 Universe::narrow_klass_shift() == 0)) {
280 // 32-bits narrow oop can be the base of address expressions
281 tptr = base->get_ptr_type();
282 } else {
283 // only regular oops are expected here
284 tptr = base->bottom_type()->is_ptr();
285 }
286 // Give up if offset is not a compile-time constant
287 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
288 continue;
289 offset += tptr->_offset; // correct if base is offseted
290 if( MacroAssembler::needs_explicit_null_check(offset) )
291 continue; // Give up is reference is beyond 4K page size
292 }
293 }
295 // Check ctrl input to see if the null-check dominates the memory op
296 Block *cb = get_block_for_node(mach);
297 cb = cb->_idom; // Always hoist at least 1 block
298 if( !was_store ) { // Stores can be hoisted only one block
299 while( cb->_dom_depth > (block->_dom_depth + 1))
300 cb = cb->_idom; // Hoist loads as far as we want
301 // The non-null-block should dominate the memory op, too. Live
302 // range spilling will insert a spill in the non-null-block if it is
303 // needs to spill the memory op for an implicit null check.
304 if (cb->_dom_depth == (block->_dom_depth + 1)) {
305 if (cb != not_null_block) continue;
306 cb = cb->_idom;
307 }
308 }
309 if( cb != block ) continue;
311 // Found a memory user; see if it can be hoisted to check-block
312 uint vidx = 0; // Capture index of value into memop
313 uint j;
314 for( j = mach->req()-1; j > 0; j-- ) {
315 if( mach->in(j) == val ) {
316 vidx = j;
317 // Ignore DecodeN val which could be hoisted to where needed.
318 if( is_decoden ) continue;
319 }
320 // Block of memory-op input
321 Block *inb = get_block_for_node(mach->in(j));
322 Block *b = block; // Start from nul check
323 while( b != inb && b->_dom_depth > inb->_dom_depth )
324 b = b->_idom; // search upwards for input
325 // See if input dominates null check
326 if( b != inb )
327 break;
328 }
329 if( j > 0 )
330 continue;
331 Block *mb = get_block_for_node(mach);
332 // Hoisting stores requires more checks for the anti-dependence case.
333 // Give up hoisting if we have to move the store past any load.
334 if( was_store ) {
335 Block *b = mb; // Start searching here for a local load
336 // mach use (faulting) trying to hoist
337 // n might be blocker to hoisting
338 while( b != block ) {
339 uint k;
340 for( k = 1; k < b->number_of_nodes(); k++ ) {
341 Node *n = b->get_node(k);
342 if( n->needs_anti_dependence_check() &&
343 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
344 break; // Found anti-dependent load
345 }
346 if( k < b->number_of_nodes() )
347 break; // Found anti-dependent load
348 // Make sure control does not do a merge (would have to check allpaths)
349 if( b->num_preds() != 2 ) break;
350 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
351 }
352 if( b != block ) continue;
353 }
355 // Make sure this memory op is not already being used for a NullCheck
356 Node *e = mb->end();
357 if( e->is_MachNullCheck() && e->in(1) == mach )
358 continue; // Already being used as a NULL check
360 // Found a candidate! Pick one with least dom depth - the highest
361 // in the dom tree should be closest to the null check.
362 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
363 best = mach;
364 bidx = vidx;
365 }
366 }
367 // No candidate!
368 if (best == NULL) {
369 return;
370 }
372 // ---- Found an implicit null check
373 extern int implicit_null_checks;
374 implicit_null_checks++;
376 if( is_decoden ) {
377 // Check if we need to hoist decodeHeapOop_not_null first.
378 Block *valb = get_block_for_node(val);
379 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
380 // Hoist it up to the end of the test block.
381 valb->find_remove(val);
382 block->add_inst(val);
383 map_node_to_block(val, block);
384 // DecodeN on x86 may kill flags. Check for flag-killing projections
385 // that also need to be hoisted.
386 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
387 Node* n = val->fast_out(j);
388 if( n->is_MachProj() ) {
389 get_block_for_node(n)->find_remove(n);
390 block->add_inst(n);
391 map_node_to_block(n, block);
392 }
393 }
394 }
395 }
396 // Hoist the memory candidate up to the end of the test block.
397 Block *old_block = get_block_for_node(best);
398 old_block->find_remove(best);
399 block->add_inst(best);
400 map_node_to_block(best, block);
402 // Move the control dependence if it is pinned to not-null block.
403 // Don't change it in other cases: NULL or dominating control.
404 if (best->in(0) == not_null_block->head()) {
405 // Set it to control edge of null check.
406 best->set_req(0, proj->in(0)->in(0));
407 }
409 // Check for flag-killing projections that also need to be hoisted
410 // Should be DU safe because no edge updates.
411 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
412 Node* n = best->fast_out(j);
413 if( n->is_MachProj() ) {
414 get_block_for_node(n)->find_remove(n);
415 block->add_inst(n);
416 map_node_to_block(n, block);
417 }
418 }
420 // proj==Op_True --> ne test; proj==Op_False --> eq test.
421 // One of two graph shapes got matched:
422 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
423 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
424 // NULL checks are always branch-if-eq. If we see a IfTrue projection
425 // then we are replacing a 'ne' test with a 'eq' NULL check test.
426 // We need to flip the projections to keep the same semantics.
427 if( proj->Opcode() == Op_IfTrue ) {
428 // Swap order of projections in basic block to swap branch targets
429 Node *tmp1 = block->get_node(block->end_idx()+1);
430 Node *tmp2 = block->get_node(block->end_idx()+2);
431 block->map_node(tmp2, block->end_idx()+1);
432 block->map_node(tmp1, block->end_idx()+2);
433 Node *tmp = new (C) Node(C->top()); // Use not NULL input
434 tmp1->replace_by(tmp);
435 tmp2->replace_by(tmp1);
436 tmp->replace_by(tmp2);
437 tmp->destruct();
438 }
440 // Remove the existing null check; use a new implicit null check instead.
441 // Since schedule-local needs precise def-use info, we need to correct
442 // it as well.
443 Node *old_tst = proj->in(0);
444 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
445 block->map_node(nul_chk, block->end_idx());
446 map_node_to_block(nul_chk, block);
447 // Redirect users of old_test to nul_chk
448 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
449 old_tst->last_out(i2)->set_req(0, nul_chk);
450 // Clean-up any dead code
451 for (uint i3 = 0; i3 < old_tst->req(); i3++) {
452 Node* in = old_tst->in(i3);
453 old_tst->set_req(i3, NULL);
454 if (in->outcnt() == 0) {
455 // Remove dead input node
456 in->disconnect_inputs(NULL, C);
457 block->find_remove(in);
458 }
459 }
461 latency_from_uses(nul_chk);
462 latency_from_uses(best);
464 // insert anti-dependences to defs in this block
465 if (! best->needs_anti_dependence_check()) {
466 for (uint k = 1; k < block->number_of_nodes(); k++) {
467 Node *n = block->get_node(k);
468 if (n->needs_anti_dependence_check() &&
469 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) {
470 // Found anti-dependent load
471 insert_anti_dependences(block, n);
472 }
473 }
474 }
475 }
478 //------------------------------select-----------------------------------------
479 // Select a nice fellow from the worklist to schedule next. If there is only
480 // one choice, then use it. Projections take top priority for correctness
481 // reasons - if I see a projection, then it is next. There are a number of
482 // other special cases, for instructions that consume condition codes, et al.
483 // These are chosen immediately. Some instructions are required to immediately
484 // precede the last instruction in the block, and these are taken last. Of the
485 // remaining cases (most), choose the instruction with the greatest latency
486 // (that is, the most number of pseudo-cycles required to the end of the
487 // routine). If there is a tie, choose the instruction with the most inputs.
488 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
490 // If only a single entry on the stack, use it
491 uint cnt = worklist.size();
492 if (cnt == 1) {
493 Node *n = worklist[0];
494 worklist.map(0,worklist.pop());
495 return n;
496 }
498 uint choice = 0; // Bigger is most important
499 uint latency = 0; // Bigger is scheduled first
500 uint score = 0; // Bigger is better
501 int idx = -1; // Index in worklist
502 int cand_cnt = 0; // Candidate count
504 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
505 // Order in worklist is used to break ties.
506 // See caller for how this is used to delay scheduling
507 // of induction variable increments to after the other
508 // uses of the phi are scheduled.
509 Node *n = worklist[i]; // Get Node on worklist
511 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
512 if( n->is_Proj() || // Projections always win
513 n->Opcode()== Op_Con || // So does constant 'Top'
514 iop == Op_CreateEx || // Create-exception must start block
515 iop == Op_CheckCastPP
516 ) {
517 worklist.map(i,worklist.pop());
518 return n;
519 }
521 // Final call in a block must be adjacent to 'catch'
522 Node *e = block->end();
523 if( e->is_Catch() && e->in(0)->in(0) == n )
524 continue;
526 // Memory op for an implicit null check has to be at the end of the block
527 if( e->is_MachNullCheck() && e->in(1) == n )
528 continue;
530 // Schedule IV increment last.
531 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
532 e->in(1)->in(1) == n && n->is_iteratively_computed())
533 continue;
535 uint n_choice = 2;
537 // See if this instruction is consumed by a branch. If so, then (as the
538 // branch is the last instruction in the basic block) force it to the
539 // end of the basic block
540 if ( must_clone[iop] ) {
541 // See if any use is a branch
542 bool found_machif = false;
544 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
545 Node* use = n->fast_out(j);
547 // The use is a conditional branch, make them adjacent
548 if (use->is_MachIf() && get_block_for_node(use) == block) {
549 found_machif = true;
550 break;
551 }
553 // More than this instruction pending for successor to be ready,
554 // don't choose this if other opportunities are ready
555 if (ready_cnt.at(use->_idx) > 1)
556 n_choice = 1;
557 }
559 // loop terminated, prefer not to use this instruction
560 if (found_machif)
561 continue;
562 }
564 // See if this has a predecessor that is "must_clone", i.e. sets the
565 // condition code. If so, choose this first
566 for (uint j = 0; j < n->req() ; j++) {
567 Node *inn = n->in(j);
568 if (inn) {
569 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
570 n_choice = 3;
571 break;
572 }
573 }
574 }
576 // MachTemps should be scheduled last so they are near their uses
577 if (n->is_MachTemp()) {
578 n_choice = 1;
579 }
581 uint n_latency = get_latency_for_node(n);
582 uint n_score = n->req(); // Many inputs get high score to break ties
584 // Keep best latency found
585 cand_cnt++;
586 if (choice < n_choice ||
587 (choice == n_choice &&
588 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
589 (!StressLCM &&
590 (latency < n_latency ||
591 (latency == n_latency &&
592 (score < n_score))))))) {
593 choice = n_choice;
594 latency = n_latency;
595 score = n_score;
596 idx = i; // Also keep index in worklist
597 }
598 } // End of for all ready nodes in worklist
600 assert(idx >= 0, "index should be set");
601 Node *n = worklist[(uint)idx]; // Get the winner
603 worklist.map((uint)idx, worklist.pop()); // Compress worklist
604 return n;
605 }
608 //------------------------------set_next_call----------------------------------
609 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
610 if( next_call.test_set(n->_idx) ) return;
611 for( uint i=0; i<n->len(); i++ ) {
612 Node *m = n->in(i);
613 if( !m ) continue; // must see all nodes in block that precede call
614 if (get_block_for_node(m) == block) {
615 set_next_call(block, m, next_call);
616 }
617 }
618 }
620 //------------------------------needed_for_next_call---------------------------
621 // Set the flag 'next_call' for each Node that is needed for the next call to
622 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
623 // next subroutine call get priority - basically it moves things NOT needed
624 // for the next call till after the call. This prevents me from trying to
625 // carry lots of stuff live across a call.
626 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
627 // Find the next control-defining Node in this block
628 Node* call = NULL;
629 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
630 Node* m = this_call->fast_out(i);
631 if (get_block_for_node(m) == block && // Local-block user
632 m != this_call && // Not self-start node
633 m->is_MachCall()) {
634 call = m;
635 break;
636 }
637 }
638 if (call == NULL) return; // No next call (e.g., block end is near)
639 // Set next-call for all inputs to this call
640 set_next_call(block, call, next_call);
641 }
643 //------------------------------add_call_kills-------------------------------------
644 // helper function that adds caller save registers to MachProjNode
645 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
646 // Fill in the kill mask for the call
647 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
648 if( !regs.Member(r) ) { // Not already defined by the call
649 // Save-on-call register?
650 if ((save_policy[r] == 'C') ||
651 (save_policy[r] == 'A') ||
652 ((save_policy[r] == 'E') && exclude_soe)) {
653 proj->_rout.Insert(r);
654 }
655 }
656 }
657 }
660 //------------------------------sched_call-------------------------------------
661 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
662 RegMask regs;
664 // Schedule all the users of the call right now. All the users are
665 // projection Nodes, so they must be scheduled next to the call.
666 // Collect all the defined registers.
667 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
668 Node* n = mcall->fast_out(i);
669 assert( n->is_MachProj(), "" );
670 int n_cnt = ready_cnt.at(n->_idx)-1;
671 ready_cnt.at_put(n->_idx, n_cnt);
672 assert( n_cnt == 0, "" );
673 // Schedule next to call
674 block->map_node(n, node_cnt++);
675 // Collect defined registers
676 regs.OR(n->out_RegMask());
677 // Check for scheduling the next control-definer
678 if( n->bottom_type() == Type::CONTROL )
679 // Warm up next pile of heuristic bits
680 needed_for_next_call(block, n, next_call);
682 // Children of projections are now all ready
683 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
684 Node* m = n->fast_out(j); // Get user
685 if(get_block_for_node(m) != block) {
686 continue;
687 }
688 if( m->is_Phi() ) continue;
689 int m_cnt = ready_cnt.at(m->_idx)-1;
690 ready_cnt.at_put(m->_idx, m_cnt);
691 if( m_cnt == 0 )
692 worklist.push(m);
693 }
695 }
697 // Act as if the call defines the Frame Pointer.
698 // Certainly the FP is alive and well after the call.
699 regs.Insert(_matcher.c_frame_pointer());
701 // Set all registers killed and not already defined by the call.
702 uint r_cnt = mcall->tf()->range()->cnt();
703 int op = mcall->ideal_Opcode();
704 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
705 map_node_to_block(proj, block);
706 block->insert_node(proj, node_cnt++);
708 // Select the right register save policy.
709 const char *save_policy = NULL;
710 switch (op) {
711 case Op_CallRuntime:
712 case Op_CallLeaf:
713 case Op_CallLeafNoFP:
714 // Calling C code so use C calling convention
715 save_policy = _matcher._c_reg_save_policy;
716 break;
718 case Op_CallStaticJava:
719 case Op_CallDynamicJava:
720 // Calling Java code so use Java calling convention
721 save_policy = _matcher._register_save_policy;
722 break;
724 default:
725 ShouldNotReachHere();
726 }
728 // When using CallRuntime mark SOE registers as killed by the call
729 // so values that could show up in the RegisterMap aren't live in a
730 // callee saved register since the register wouldn't know where to
731 // find them. CallLeaf and CallLeafNoFP are ok because they can't
732 // have debug info on them. Strictly speaking this only needs to be
733 // done for oops since idealreg2debugmask takes care of debug info
734 // references but there no way to handle oops differently than other
735 // pointers as far as the kill mask goes.
736 bool exclude_soe = op == Op_CallRuntime;
738 // If the call is a MethodHandle invoke, we need to exclude the
739 // register which is used to save the SP value over MH invokes from
740 // the mask. Otherwise this register could be used for
741 // deoptimization information.
742 if (op == Op_CallStaticJava) {
743 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
744 if (mcallstaticjava->_method_handle_invoke)
745 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
746 }
748 add_call_kills(proj, regs, save_policy, exclude_soe);
750 return node_cnt;
751 }
754 //------------------------------schedule_local---------------------------------
755 // Topological sort within a block. Someday become a real scheduler.
756 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
757 // Already "sorted" are the block start Node (as the first entry), and
758 // the block-ending Node and any trailing control projections. We leave
759 // these alone. PhiNodes and ParmNodes are made to follow the block start
760 // Node. Everything else gets topo-sorted.
762 #ifndef PRODUCT
763 if (trace_opto_pipelining()) {
764 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
765 for (uint i = 0;i < block->number_of_nodes(); i++) {
766 tty->print("# ");
767 block->get_node(i)->fast_dump();
768 }
769 tty->print_cr("#");
770 }
771 #endif
773 // RootNode is already sorted
774 if (block->number_of_nodes() == 1) {
775 return true;
776 }
778 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
779 uint node_cnt = block->end_idx();
780 uint phi_cnt = 1;
781 uint i;
782 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
783 Node *n = block->get_node(i);
784 if( n->is_Phi() || // Found a PhiNode or ParmNode
785 (n->is_Proj() && n->in(0) == block->head()) ) {
786 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
787 block->map_node(block->get_node(phi_cnt), i);
788 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
789 } else { // All others
790 // Count block-local inputs to 'n'
791 uint cnt = n->len(); // Input count
792 uint local = 0;
793 for( uint j=0; j<cnt; j++ ) {
794 Node *m = n->in(j);
795 if( m && get_block_for_node(m) == block && !m->is_top() )
796 local++; // One more block-local input
797 }
798 ready_cnt.at_put(n->_idx, local); // Count em up
800 #ifdef ASSERT
801 if( UseConcMarkSweepGC || UseG1GC ) {
802 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
803 // Check the precedence edges
804 for (uint prec = n->req(); prec < n->len(); prec++) {
805 Node* oop_store = n->in(prec);
806 if (oop_store != NULL) {
807 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
808 }
809 }
810 }
811 }
812 #endif
814 // A few node types require changing a required edge to a precedence edge
815 // before allocation.
816 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
817 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
818 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
819 // MemBarAcquire could be created without Precedent edge.
820 // del_req() replaces the specified edge with the last input edge
821 // and then removes the last edge. If the specified edge > number of
822 // edges the last edge will be moved outside of the input edges array
823 // and the edge will be lost. This is why this code should be
824 // executed only when Precedent (== TypeFunc::Parms) edge is present.
825 Node *x = n->in(TypeFunc::Parms);
826 if (x != NULL && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) {
827 // Old edge to node within same block will get removed, but no precedence
828 // edge will get added because it already exists. Update ready count.
829 int cnt = ready_cnt.at(n->_idx);
830 assert(cnt > 1, err_msg("MemBar node %d must not get ready here", n->_idx));
831 ready_cnt.at_put(n->_idx, cnt-1);
832 }
833 n->del_req(TypeFunc::Parms);
834 n->add_prec(x);
835 }
836 }
837 }
838 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
839 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
841 // All the prescheduled guys do not hold back internal nodes
842 uint i3;
843 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
844 Node *n = block->get_node(i3); // Get pre-scheduled
845 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
846 Node* m = n->fast_out(j);
847 if (get_block_for_node(m) == block) { // Local-block user
848 int m_cnt = ready_cnt.at(m->_idx)-1;
849 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
850 }
851 }
852 }
854 Node_List delay;
855 // Make a worklist
856 Node_List worklist;
857 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
858 Node *m = block->get_node(i4);
859 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
860 if (m->is_iteratively_computed()) {
861 // Push induction variable increments last to allow other uses
862 // of the phi to be scheduled first. The select() method breaks
863 // ties in scheduling by worklist order.
864 delay.push(m);
865 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
866 // Force the CreateEx to the top of the list so it's processed
867 // first and ends up at the start of the block.
868 worklist.insert(0, m);
869 } else {
870 worklist.push(m); // Then on to worklist!
871 }
872 }
873 }
874 while (delay.size()) {
875 Node* d = delay.pop();
876 worklist.push(d);
877 }
879 // Warm up the 'next_call' heuristic bits
880 needed_for_next_call(block, block->head(), next_call);
882 #ifndef PRODUCT
883 if (trace_opto_pipelining()) {
884 for (uint j=0; j< block->number_of_nodes(); j++) {
885 Node *n = block->get_node(j);
886 int idx = n->_idx;
887 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
888 tty->print("latency:%3d ", get_latency_for_node(n));
889 tty->print("%4d: %s\n", idx, n->Name());
890 }
891 }
892 #endif
894 uint max_idx = (uint)ready_cnt.length();
895 // Pull from worklist and schedule
896 while( worklist.size() ) { // Worklist is not ready
898 #ifndef PRODUCT
899 if (trace_opto_pipelining()) {
900 tty->print("# ready list:");
901 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
902 Node *n = worklist[i]; // Get Node on worklist
903 tty->print(" %d", n->_idx);
904 }
905 tty->cr();
906 }
907 #endif
909 // Select and pop a ready guy from worklist
910 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
911 block->map_node(n, phi_cnt++); // Schedule him next
913 #ifndef PRODUCT
914 if (trace_opto_pipelining()) {
915 tty->print("# select %d: %s", n->_idx, n->Name());
916 tty->print(", latency:%d", get_latency_for_node(n));
917 n->dump();
918 if (Verbose) {
919 tty->print("# ready list:");
920 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
921 Node *n = worklist[i]; // Get Node on worklist
922 tty->print(" %d", n->_idx);
923 }
924 tty->cr();
925 }
926 }
928 #endif
929 if( n->is_MachCall() ) {
930 MachCallNode *mcall = n->as_MachCall();
931 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
932 continue;
933 }
935 if (n->is_Mach() && n->as_Mach()->has_call()) {
936 RegMask regs;
937 regs.Insert(_matcher.c_frame_pointer());
938 regs.OR(n->out_RegMask());
940 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
941 map_node_to_block(proj, block);
942 block->insert_node(proj, phi_cnt++);
944 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
945 }
947 // Children are now all ready
948 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
949 Node* m = n->fast_out(i5); // Get user
950 if (get_block_for_node(m) != block) {
951 continue;
952 }
953 if( m->is_Phi() ) continue;
954 if (m->_idx >= max_idx) { // new node, skip it
955 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
956 continue;
957 }
958 int m_cnt = ready_cnt.at(m->_idx)-1;
959 ready_cnt.at_put(m->_idx, m_cnt);
960 if( m_cnt == 0 )
961 worklist.push(m);
962 }
963 }
965 if( phi_cnt != block->end_idx() ) {
966 // did not schedule all. Retry, Bailout, or Die
967 if (C->subsume_loads() == true && !C->failing()) {
968 // Retry with subsume_loads == false
969 // If this is the first failure, the sentinel string will "stick"
970 // to the Compile object, and the C2Compiler will see it and retry.
971 C->record_failure(C2Compiler::retry_no_subsuming_loads());
972 } else {
973 assert(false, "graph should be schedulable");
974 }
975 // assert( phi_cnt == end_idx(), "did not schedule all" );
976 return false;
977 }
979 #ifndef PRODUCT
980 if (trace_opto_pipelining()) {
981 tty->print_cr("#");
982 tty->print_cr("# after schedule_local");
983 for (uint i = 0;i < block->number_of_nodes();i++) {
984 tty->print("# ");
985 block->get_node(i)->fast_dump();
986 }
987 tty->cr();
988 }
989 #endif
992 return true;
993 }
995 //--------------------------catch_cleanup_fix_all_inputs-----------------------
996 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
997 for (uint l = 0; l < use->len(); l++) {
998 if (use->in(l) == old_def) {
999 if (l < use->req()) {
1000 use->set_req(l, new_def);
1001 } else {
1002 use->rm_prec(l);
1003 use->add_prec(new_def);
1004 l--;
1005 }
1006 }
1007 }
1008 }
1010 //------------------------------catch_cleanup_find_cloned_def------------------
1011 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1012 assert( use_blk != def_blk, "Inter-block cleanup only");
1014 // The use is some block below the Catch. Find and return the clone of the def
1015 // that dominates the use. If there is no clone in a dominating block, then
1016 // create a phi for the def in a dominating block.
1018 // Find which successor block dominates this use. The successor
1019 // blocks must all be single-entry (from the Catch only; I will have
1020 // split blocks to make this so), hence they all dominate.
1021 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1022 use_blk = use_blk->_idom;
1024 // Find the successor
1025 Node *fixup = NULL;
1027 uint j;
1028 for( j = 0; j < def_blk->_num_succs; j++ )
1029 if( use_blk == def_blk->_succs[j] )
1030 break;
1032 if( j == def_blk->_num_succs ) {
1033 // Block at same level in dom-tree is not a successor. It needs a
1034 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1035 Node_Array inputs = new Node_List(Thread::current()->resource_area());
1036 for(uint k = 1; k < use_blk->num_preds(); k++) {
1037 Block* block = get_block_for_node(use_blk->pred(k));
1038 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1039 }
1041 // Check to see if the use_blk already has an identical phi inserted.
1042 // If it exists, it will be at the first position since all uses of a
1043 // def are processed together.
1044 Node *phi = use_blk->get_node(1);
1045 if( phi->is_Phi() ) {
1046 fixup = phi;
1047 for (uint k = 1; k < use_blk->num_preds(); k++) {
1048 if (phi->in(k) != inputs[k]) {
1049 // Not a match
1050 fixup = NULL;
1051 break;
1052 }
1053 }
1054 }
1056 // If an existing PhiNode was not found, make a new one.
1057 if (fixup == NULL) {
1058 Node *new_phi = PhiNode::make(use_blk->head(), def);
1059 use_blk->insert_node(new_phi, 1);
1060 map_node_to_block(new_phi, use_blk);
1061 for (uint k = 1; k < use_blk->num_preds(); k++) {
1062 new_phi->set_req(k, inputs[k]);
1063 }
1064 fixup = new_phi;
1065 }
1067 } else {
1068 // Found the use just below the Catch. Make it use the clone.
1069 fixup = use_blk->get_node(n_clone_idx);
1070 }
1072 return fixup;
1073 }
1075 //--------------------------catch_cleanup_intra_block--------------------------
1076 // Fix all input edges in use that reference "def". The use is in the same
1077 // block as the def and both have been cloned in each successor block.
1078 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1080 // Both the use and def have been cloned. For each successor block,
1081 // get the clone of the use, and make its input the clone of the def
1082 // found in that block.
1084 uint use_idx = blk->find_node(use);
1085 uint offset_idx = use_idx - beg;
1086 for( uint k = 0; k < blk->_num_succs; k++ ) {
1087 // Get clone in each successor block
1088 Block *sb = blk->_succs[k];
1089 Node *clone = sb->get_node(offset_idx+1);
1090 assert( clone->Opcode() == use->Opcode(), "" );
1092 // Make use-clone reference the def-clone
1093 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1094 }
1095 }
1097 //------------------------------catch_cleanup_inter_block---------------------
1098 // Fix all input edges in use that reference "def". The use is in a different
1099 // block than the def.
1100 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1101 if( !use_blk ) return; // Can happen if the use is a precedence edge
1103 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1104 catch_cleanup_fix_all_inputs(use, def, new_def);
1105 }
1107 //------------------------------call_catch_cleanup-----------------------------
1108 // If we inserted any instructions between a Call and his CatchNode,
1109 // clone the instructions on all paths below the Catch.
1110 void PhaseCFG::call_catch_cleanup(Block* block) {
1112 // End of region to clone
1113 uint end = block->end_idx();
1114 if( !block->get_node(end)->is_Catch() ) return;
1115 // Start of region to clone
1116 uint beg = end;
1117 while(!block->get_node(beg-1)->is_MachProj() ||
1118 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1119 beg--;
1120 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1121 }
1122 // Range of inserted instructions is [beg, end)
1123 if( beg == end ) return;
1125 // Clone along all Catch output paths. Clone area between the 'beg' and
1126 // 'end' indices.
1127 for( uint i = 0; i < block->_num_succs; i++ ) {
1128 Block *sb = block->_succs[i];
1129 // Clone the entire area; ignoring the edge fixup for now.
1130 for( uint j = end; j > beg; j-- ) {
1131 Node *clone = block->get_node(j-1)->clone();
1132 sb->insert_node(clone, 1);
1133 map_node_to_block(clone, sb);
1134 if (clone->needs_anti_dependence_check()) {
1135 insert_anti_dependences(sb, clone);
1136 }
1137 }
1138 }
1141 // Fixup edges. Check the def-use info per cloned Node
1142 for(uint i2 = beg; i2 < end; i2++ ) {
1143 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1144 Node *n = block->get_node(i2); // Node that got cloned
1145 // Need DU safe iterator because of edge manipulation in calls.
1146 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1147 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1148 out->push(n->fast_out(j1));
1149 }
1150 uint max = out->size();
1151 for (uint j = 0; j < max; j++) {// For all users
1152 Node *use = out->pop();
1153 Block *buse = get_block_for_node(use);
1154 if( use->is_Phi() ) {
1155 for( uint k = 1; k < use->req(); k++ )
1156 if( use->in(k) == n ) {
1157 Block* b = get_block_for_node(buse->pred(k));
1158 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1159 use->set_req(k, fixup);
1160 }
1161 } else {
1162 if (block == buse) {
1163 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1164 } else {
1165 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1166 }
1167 }
1168 } // End for all users
1170 } // End of for all Nodes in cloned area
1172 // Remove the now-dead cloned ops
1173 for(uint i3 = beg; i3 < end; i3++ ) {
1174 block->get_node(beg)->disconnect_inputs(NULL, C);
1175 block->remove_node(beg);
1176 }
1178 // If the successor blocks have a CreateEx node, move it back to the top
1179 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1180 Block *sb = block->_succs[i4];
1181 uint new_cnt = end - beg;
1182 // Remove any newly created, but dead, nodes.
1183 for( uint j = new_cnt; j > 0; j-- ) {
1184 Node *n = sb->get_node(j);
1185 if (n->outcnt() == 0 &&
1186 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1187 n->disconnect_inputs(NULL, C);
1188 sb->remove_node(j);
1189 new_cnt--;
1190 }
1191 }
1192 // If any newly created nodes remain, move the CreateEx node to the top
1193 if (new_cnt > 0) {
1194 Node *cex = sb->get_node(1+new_cnt);
1195 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1196 sb->remove_node(1+new_cnt);
1197 sb->insert_node(cex, 1);
1198 }
1199 }
1200 }
1201 }