1 /*
2 * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/ad.hpp"
28 #include "opto/block.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/runtime.hpp"
34 #include "runtime/sharedRuntime.hpp"
35
36 // Optimization - Graph Style
37
38 // Check whether val is not-null-decoded compressed oop,
39 // i.e. will grab into the base of the heap if it represents NULL.
40 static bool accesses_heap_base_zone(Node *val) {
41 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops.
42 if (val && val->is_Mach()) {
43 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
44 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
45 // decode NULL to point to the heap base (Decode_NN).
46 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
47 return true;
48 }
49 }
50 // Must recognize load operation with Decode matched in memory operand.
51 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected()
52 // returns true everywhere else. On PPC, no such memory operands
53 // exist, therefore we did not yet implement a check for such operands.
54 NOT_AIX(Unimplemented());
55 }
56 }
57 return false;
58 }
59
60 static bool needs_explicit_null_check_for_read(Node *val) {
61 // On some OSes (AIX) the page at address 0 is only write protected.
62 // If so, only Store operations will trap.
63 if (os::zero_page_read_protected()) {
64 return false; // Implicit null check will work.
65 }
66 // Also a read accessing the base of a heap-based compressed heap will trap.
67 if (accesses_heap_base_zone(val) && // Hits the base zone page.
68 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected.
69 return false;
70 }
71
72 return true;
73 }
74
75 //------------------------------implicit_null_check----------------------------
76 // Detect implicit-null-check opportunities. Basically, find NULL checks
77 // with suitable memory ops nearby. Use the memory op to do the NULL check.
78 // I can generate a memory op if there is not one nearby.
79 // The proj is the control projection for the not-null case.
80 // The val is the pointer being checked for nullness or
81 // decodeHeapOop_not_null node if it did not fold into address.
82 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
83 // Assume if null check need for 0 offset then always needed
84 // Intel solaris doesn't support any null checks yet and no
85 // mechanism exists (yet) to set the switches at an os_cpu level
86 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
87
88 // Make sure the ptr-is-null path appears to be uncommon!
89 float f = block->end()->as_MachIf()->_prob;
90 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
91 if( f > PROB_UNLIKELY_MAG(4) ) return;
92
93 uint bidx = 0; // Capture index of value into memop
94 bool was_store; // Memory op is a store op
95
96 // Get the successor block for if the test ptr is non-null
97 Block* not_null_block; // this one goes with the proj
98 Block* null_block;
99 if (block->get_node(block->number_of_nodes()-1) == proj) {
100 null_block = block->_succs[0];
101 not_null_block = block->_succs[1];
102 } else {
103 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
104 not_null_block = block->_succs[0];
105 null_block = block->_succs[1];
106 }
107 while (null_block->is_Empty() == Block::empty_with_goto) {
108 null_block = null_block->_succs[0];
109 }
110
111 // Search the exception block for an uncommon trap.
112 // (See Parse::do_if and Parse::do_ifnull for the reason
113 // we need an uncommon trap. Briefly, we need a way to
114 // detect failure of this optimization, as in 6366351.)
115 {
116 bool found_trap = false;
117 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
118 Node* nn = null_block->get_node(i1);
119 if (nn->is_MachCall() &&
120 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
121 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
122 if (trtype->isa_int() && trtype->is_int()->is_con()) {
123 jint tr_con = trtype->is_int()->get_con();
124 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
125 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
126 assert((int)reason < (int)BitsPerInt, "recode bit map");
127 if (is_set_nth_bit(allowed_reasons, (int) reason)
128 && action != Deoptimization::Action_none) {
129 // This uncommon trap is sure to recompile, eventually.
130 // When that happens, C->too_many_traps will prevent
131 // this transformation from happening again.
132 found_trap = true;
133 }
134 }
135 break;
136 }
137 }
138 if (!found_trap) {
139 // We did not find an uncommon trap.
140 return;
141 }
142 }
143
144 // Check for decodeHeapOop_not_null node which did not fold into address
145 bool is_decoden = ((intptr_t)val) & 1;
146 val = (Node*)(((intptr_t)val) & ~1);
147
148 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
149 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
150
151 // Search the successor block for a load or store who's base value is also
152 // the tested value. There may be several.
153 Node_List *out = new Node_List(Thread::current()->resource_area());
154 MachNode *best = NULL; // Best found so far
155 for (DUIterator i = val->outs(); val->has_out(i); i++) {
156 Node *m = val->out(i);
157 if( !m->is_Mach() ) continue;
158 MachNode *mach = m->as_Mach();
159 was_store = false;
160 int iop = mach->ideal_Opcode();
161 switch( iop ) {
162 case Op_LoadB:
163 case Op_LoadUB:
164 case Op_LoadUS:
165 case Op_LoadD:
166 case Op_LoadF:
167 case Op_LoadI:
168 case Op_LoadL:
169 case Op_LoadP:
170 case Op_LoadN:
171 case Op_LoadS:
172 case Op_LoadKlass:
173 case Op_LoadNKlass:
174 case Op_LoadRange:
175 case Op_LoadD_unaligned:
176 case Op_LoadL_unaligned:
177 assert(mach->in(2) == val, "should be address");
178 break;
179 case Op_StoreB:
180 case Op_StoreC:
181 case Op_StoreCM:
182 case Op_StoreD:
183 case Op_StoreF:
184 case Op_StoreI:
185 case Op_StoreL:
186 case Op_StoreP:
187 case Op_StoreN:
188 case Op_StoreNKlass:
189 was_store = true; // Memory op is a store op
190 // Stores will have their address in slot 2 (memory in slot 1).
191 // If the value being nul-checked is in another slot, it means we
192 // are storing the checked value, which does NOT check the value!
193 if( mach->in(2) != val ) continue;
194 break; // Found a memory op?
195 case Op_StrComp:
196 case Op_StrEquals:
197 case Op_StrIndexOf:
198 case Op_AryEq:
199 case Op_EncodeISOArray:
200 // Not a legit memory op for implicit null check regardless of
201 // embedded loads
202 continue;
203 default: // Also check for embedded loads
204 if( !mach->needs_anti_dependence_check() )
205 continue; // Not an memory op; skip it
206 if( must_clone[iop] ) {
207 // Do not move nodes which produce flags because
208 // RA will try to clone it to place near branch and
209 // it will cause recompilation, see clone_node().
210 continue;
211 }
212 {
213 // Check that value is used in memory address in
214 // instructions with embedded load (CmpP val1,(val2+off)).
215 Node* base;
216 Node* index;
217 const MachOper* oper = mach->memory_inputs(base, index);
218 if (oper == NULL || oper == (MachOper*)-1) {
219 continue; // Not an memory op; skip it
220 }
221 if (val == base ||
222 val == index && val->bottom_type()->isa_narrowoop()) {
223 break; // Found it
224 } else {
225 continue; // Skip it
226 }
227 }
228 break;
229 }
230
231 // On some OSes (AIX) the page at address 0 is only write protected.
232 // If so, only Store operations will trap.
233 // But a read accessing the base of a heap-based compressed heap will trap.
234 if (!was_store && needs_explicit_null_check_for_read(val)) {
235 continue;
236 }
237
238 // check if the offset is not too high for implicit exception
239 {
240 intptr_t offset = 0;
241 const TypePtr *adr_type = NULL; // Do not need this return value here
242 const Node* base = mach->get_base_and_disp(offset, adr_type);
243 if (base == NULL || base == NodeSentinel) {
244 // Narrow oop address doesn't have base, only index
245 if( val->bottom_type()->isa_narrowoop() &&
246 MacroAssembler::needs_explicit_null_check(offset) )
247 continue; // Give up if offset is beyond page size
248 // cannot reason about it; is probably not implicit null exception
249 } else {
250 const TypePtr* tptr;
251 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
252 Universe::narrow_klass_shift() == 0)) {
253 // 32-bits narrow oop can be the base of address expressions
254 tptr = base->get_ptr_type();
255 } else {
256 // only regular oops are expected here
257 tptr = base->bottom_type()->is_ptr();
258 }
259 // Give up if offset is not a compile-time constant
260 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
261 continue;
262 offset += tptr->_offset; // correct if base is offseted
263 if( MacroAssembler::needs_explicit_null_check(offset) )
264 continue; // Give up is reference is beyond 4K page size
265 }
266 }
267
268 // Check ctrl input to see if the null-check dominates the memory op
269 Block *cb = get_block_for_node(mach);
270 cb = cb->_idom; // Always hoist at least 1 block
271 if( !was_store ) { // Stores can be hoisted only one block
272 while( cb->_dom_depth > (block->_dom_depth + 1))
273 cb = cb->_idom; // Hoist loads as far as we want
274 // The non-null-block should dominate the memory op, too. Live
275 // range spilling will insert a spill in the non-null-block if it is
276 // needs to spill the memory op for an implicit null check.
277 if (cb->_dom_depth == (block->_dom_depth + 1)) {
278 if (cb != not_null_block) continue;
279 cb = cb->_idom;
280 }
281 }
282 if( cb != block ) continue;
283
284 // Found a memory user; see if it can be hoisted to check-block
285 uint vidx = 0; // Capture index of value into memop
286 uint j;
287 for( j = mach->req()-1; j > 0; j-- ) {
288 if( mach->in(j) == val ) {
289 vidx = j;
290 // Ignore DecodeN val which could be hoisted to where needed.
291 if( is_decoden ) continue;
292 }
293 // Block of memory-op input
294 Block *inb = get_block_for_node(mach->in(j));
295 Block *b = block; // Start from nul check
296 while( b != inb && b->_dom_depth > inb->_dom_depth )
297 b = b->_idom; // search upwards for input
298 // See if input dominates null check
299 if( b != inb )
300 break;
301 }
302 if( j > 0 )
303 continue;
304 Block *mb = get_block_for_node(mach);
305 // Hoisting stores requires more checks for the anti-dependence case.
306 // Give up hoisting if we have to move the store past any load.
307 if( was_store ) {
308 Block *b = mb; // Start searching here for a local load
309 // mach use (faulting) trying to hoist
310 // n might be blocker to hoisting
311 while( b != block ) {
312 uint k;
313 for( k = 1; k < b->number_of_nodes(); k++ ) {
314 Node *n = b->get_node(k);
315 if( n->needs_anti_dependence_check() &&
316 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
317 break; // Found anti-dependent load
318 }
319 if( k < b->number_of_nodes() )
320 break; // Found anti-dependent load
321 // Make sure control does not do a merge (would have to check allpaths)
322 if( b->num_preds() != 2 ) break;
323 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
324 }
325 if( b != block ) continue;
326 }
327
328 // Make sure this memory op is not already being used for a NullCheck
329 Node *e = mb->end();
330 if( e->is_MachNullCheck() && e->in(1) == mach )
331 continue; // Already being used as a NULL check
332
333 // Found a candidate! Pick one with least dom depth - the highest
334 // in the dom tree should be closest to the null check.
335 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
336 best = mach;
337 bidx = vidx;
338 }
339 }
340 // No candidate!
341 if (best == NULL) {
342 return;
343 }
344
345 // ---- Found an implicit null check
346 extern int implicit_null_checks;
347 implicit_null_checks++;
348
349 if( is_decoden ) {
350 // Check if we need to hoist decodeHeapOop_not_null first.
351 Block *valb = get_block_for_node(val);
352 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
353 // Hoist it up to the end of the test block.
354 valb->find_remove(val);
355 block->add_inst(val);
356 map_node_to_block(val, block);
357 // DecodeN on x86 may kill flags. Check for flag-killing projections
358 // that also need to be hoisted.
359 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
360 Node* n = val->fast_out(j);
361 if( n->is_MachProj() ) {
362 get_block_for_node(n)->find_remove(n);
363 block->add_inst(n);
364 map_node_to_block(n, block);
365 }
366 }
367 }
368 }
369 // Hoist the memory candidate up to the end of the test block.
370 Block *old_block = get_block_for_node(best);
371 old_block->find_remove(best);
372 block->add_inst(best);
373 map_node_to_block(best, block);
374
375 // Move the control dependence
376 if (best->in(0) && best->in(0) == old_block->head())
377 best->set_req(0, block->head());
378
379 // Check for flag-killing projections that also need to be hoisted
380 // Should be DU safe because no edge updates.
381 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
382 Node* n = best->fast_out(j);
383 if( n->is_MachProj() ) {
384 get_block_for_node(n)->find_remove(n);
385 block->add_inst(n);
386 map_node_to_block(n, block);
387 }
388 }
389
390 // proj==Op_True --> ne test; proj==Op_False --> eq test.
391 // One of two graph shapes got matched:
392 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
393 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
394 // NULL checks are always branch-if-eq. If we see a IfTrue projection
395 // then we are replacing a 'ne' test with a 'eq' NULL check test.
396 // We need to flip the projections to keep the same semantics.
397 if( proj->Opcode() == Op_IfTrue ) {
398 // Swap order of projections in basic block to swap branch targets
399 Node *tmp1 = block->get_node(block->end_idx()+1);
400 Node *tmp2 = block->get_node(block->end_idx()+2);
401 block->map_node(tmp2, block->end_idx()+1);
402 block->map_node(tmp1, block->end_idx()+2);
403 Node *tmp = new Node(C->top()); // Use not NULL input
404 tmp1->replace_by(tmp);
405 tmp2->replace_by(tmp1);
406 tmp->replace_by(tmp2);
407 tmp->destruct();
408 }
409
410 // Remove the existing null check; use a new implicit null check instead.
411 // Since schedule-local needs precise def-use info, we need to correct
412 // it as well.
413 Node *old_tst = proj->in(0);
414 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx);
415 block->map_node(nul_chk, block->end_idx());
416 map_node_to_block(nul_chk, block);
417 // Redirect users of old_test to nul_chk
418 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
419 old_tst->last_out(i2)->set_req(0, nul_chk);
420 // Clean-up any dead code
421 for (uint i3 = 0; i3 < old_tst->req(); i3++)
422 old_tst->set_req(i3, NULL);
423
424 latency_from_uses(nul_chk);
425 latency_from_uses(best);
426 }
427
428
429 //------------------------------select-----------------------------------------
430 // Select a nice fellow from the worklist to schedule next. If there is only
431 // one choice, then use it. Projections take top priority for correctness
432 // reasons - if I see a projection, then it is next. There are a number of
433 // other special cases, for instructions that consume condition codes, et al.
434 // These are chosen immediately. Some instructions are required to immediately
435 // precede the last instruction in the block, and these are taken last. Of the
436 // remaining cases (most), choose the instruction with the greatest latency
437 // (that is, the most number of pseudo-cycles required to the end of the
438 // routine). If there is a tie, choose the instruction with the most inputs.
439 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
440
441 // If only a single entry on the stack, use it
442 uint cnt = worklist.size();
443 if (cnt == 1) {
444 Node *n = worklist[0];
445 worklist.map(0,worklist.pop());
446 return n;
447 }
448
449 uint choice = 0; // Bigger is most important
450 uint latency = 0; // Bigger is scheduled first
451 uint score = 0; // Bigger is better
452 int idx = -1; // Index in worklist
453 int cand_cnt = 0; // Candidate count
454
455 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
456 // Order in worklist is used to break ties.
457 // See caller for how this is used to delay scheduling
458 // of induction variable increments to after the other
459 // uses of the phi are scheduled.
460 Node *n = worklist[i]; // Get Node on worklist
461
462 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
463 if( n->is_Proj() || // Projections always win
464 n->Opcode()== Op_Con || // So does constant 'Top'
465 iop == Op_CreateEx || // Create-exception must start block
466 iop == Op_CheckCastPP
467 ) {
468 worklist.map(i,worklist.pop());
469 return n;
470 }
471
472 // Final call in a block must be adjacent to 'catch'
473 Node *e = block->end();
474 if( e->is_Catch() && e->in(0)->in(0) == n )
475 continue;
476
477 // Memory op for an implicit null check has to be at the end of the block
478 if( e->is_MachNullCheck() && e->in(1) == n )
479 continue;
480
481 // Schedule IV increment last.
482 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
483 e->in(1)->in(1) == n && n->is_iteratively_computed())
484 continue;
485
486 uint n_choice = 2;
487
488 // See if this instruction is consumed by a branch. If so, then (as the
489 // branch is the last instruction in the basic block) force it to the
490 // end of the basic block
491 if ( must_clone[iop] ) {
492 // See if any use is a branch
493 bool found_machif = false;
494
495 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
496 Node* use = n->fast_out(j);
497
498 // The use is a conditional branch, make them adjacent
499 if (use->is_MachIf() && get_block_for_node(use) == block) {
500 found_machif = true;
501 break;
502 }
503
504 // More than this instruction pending for successor to be ready,
505 // don't choose this if other opportunities are ready
506 if (ready_cnt.at(use->_idx) > 1)
507 n_choice = 1;
508 }
509
510 // loop terminated, prefer not to use this instruction
511 if (found_machif)
512 continue;
513 }
514
515 // See if this has a predecessor that is "must_clone", i.e. sets the
516 // condition code. If so, choose this first
517 for (uint j = 0; j < n->req() ; j++) {
518 Node *inn = n->in(j);
519 if (inn) {
520 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
521 n_choice = 3;
522 break;
523 }
524 }
525 }
526
527 // MachTemps should be scheduled last so they are near their uses
528 if (n->is_MachTemp()) {
529 n_choice = 1;
530 }
531
532 uint n_latency = get_latency_for_node(n);
533 uint n_score = n->req(); // Many inputs get high score to break ties
534
535 // Keep best latency found
536 cand_cnt++;
537 if (choice < n_choice ||
538 (choice == n_choice &&
539 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
540 (!StressLCM &&
541 (latency < n_latency ||
542 (latency == n_latency &&
543 (score < n_score))))))) {
544 choice = n_choice;
545 latency = n_latency;
546 score = n_score;
547 idx = i; // Also keep index in worklist
548 }
549 } // End of for all ready nodes in worklist
550
551 assert(idx >= 0, "index should be set");
552 Node *n = worklist[(uint)idx]; // Get the winner
553
554 worklist.map((uint)idx, worklist.pop()); // Compress worklist
555 return n;
556 }
557
558
559 //------------------------------set_next_call----------------------------------
560 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
561 if( next_call.test_set(n->_idx) ) return;
562 for( uint i=0; i<n->len(); i++ ) {
563 Node *m = n->in(i);
564 if( !m ) continue; // must see all nodes in block that precede call
565 if (get_block_for_node(m) == block) {
566 set_next_call(block, m, next_call);
567 }
568 }
569 }
570
571 //------------------------------needed_for_next_call---------------------------
572 // Set the flag 'next_call' for each Node that is needed for the next call to
573 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
574 // next subroutine call get priority - basically it moves things NOT needed
575 // for the next call till after the call. This prevents me from trying to
576 // carry lots of stuff live across a call.
577 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
578 // Find the next control-defining Node in this block
579 Node* call = NULL;
580 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
581 Node* m = this_call->fast_out(i);
582 if (get_block_for_node(m) == block && // Local-block user
583 m != this_call && // Not self-start node
584 m->is_MachCall()) {
585 call = m;
586 break;
587 }
588 }
589 if (call == NULL) return; // No next call (e.g., block end is near)
590 // Set next-call for all inputs to this call
591 set_next_call(block, call, next_call);
592 }
593
594 //------------------------------add_call_kills-------------------------------------
595 // helper function that adds caller save registers to MachProjNode
596 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
597 // Fill in the kill mask for the call
598 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
599 if( !regs.Member(r) ) { // Not already defined by the call
600 // Save-on-call register?
601 if ((save_policy[r] == 'C') ||
602 (save_policy[r] == 'A') ||
603 ((save_policy[r] == 'E') && exclude_soe)) {
604 proj->_rout.Insert(r);
605 }
606 }
607 }
608 }
609
610
611 //------------------------------sched_call-------------------------------------
612 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
613 RegMask regs;
614
615 // Schedule all the users of the call right now. All the users are
616 // projection Nodes, so they must be scheduled next to the call.
617 // Collect all the defined registers.
618 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
619 Node* n = mcall->fast_out(i);
620 assert( n->is_MachProj(), "" );
621 int n_cnt = ready_cnt.at(n->_idx)-1;
622 ready_cnt.at_put(n->_idx, n_cnt);
623 assert( n_cnt == 0, "" );
624 // Schedule next to call
625 block->map_node(n, node_cnt++);
626 // Collect defined registers
627 regs.OR(n->out_RegMask());
628 // Check for scheduling the next control-definer
629 if( n->bottom_type() == Type::CONTROL )
630 // Warm up next pile of heuristic bits
631 needed_for_next_call(block, n, next_call);
632
633 // Children of projections are now all ready
634 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
635 Node* m = n->fast_out(j); // Get user
636 if(get_block_for_node(m) != block) {
637 continue;
638 }
639 if( m->is_Phi() ) continue;
640 int m_cnt = ready_cnt.at(m->_idx)-1;
641 ready_cnt.at_put(m->_idx, m_cnt);
642 if( m_cnt == 0 )
643 worklist.push(m);
644 }
645
646 }
647
648 // Act as if the call defines the Frame Pointer.
649 // Certainly the FP is alive and well after the call.
650 regs.Insert(_matcher.c_frame_pointer());
651
652 // Set all registers killed and not already defined by the call.
653 uint r_cnt = mcall->tf()->range()->cnt();
654 int op = mcall->ideal_Opcode();
655 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
656 map_node_to_block(proj, block);
657 block->insert_node(proj, node_cnt++);
658
659 // Select the right register save policy.
660 const char * save_policy;
661 switch (op) {
662 case Op_CallRuntime:
663 case Op_CallLeaf:
664 case Op_CallLeafNoFP:
665 // Calling C code so use C calling convention
666 save_policy = _matcher._c_reg_save_policy;
667 break;
668
669 case Op_CallStaticJava:
670 case Op_CallDynamicJava:
671 // Calling Java code so use Java calling convention
672 save_policy = _matcher._register_save_policy;
673 break;
674
675 default:
676 ShouldNotReachHere();
677 }
678
679 // When using CallRuntime mark SOE registers as killed by the call
680 // so values that could show up in the RegisterMap aren't live in a
681 // callee saved register since the register wouldn't know where to
682 // find them. CallLeaf and CallLeafNoFP are ok because they can't
683 // have debug info on them. Strictly speaking this only needs to be
684 // done for oops since idealreg2debugmask takes care of debug info
685 // references but there no way to handle oops differently than other
686 // pointers as far as the kill mask goes.
687 bool exclude_soe = op == Op_CallRuntime;
688
689 // If the call is a MethodHandle invoke, we need to exclude the
690 // register which is used to save the SP value over MH invokes from
691 // the mask. Otherwise this register could be used for
692 // deoptimization information.
693 if (op == Op_CallStaticJava) {
694 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
695 if (mcallstaticjava->_method_handle_invoke)
696 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
697 }
698
699 add_call_kills(proj, regs, save_policy, exclude_soe);
700
701 return node_cnt;
702 }
703
704
705 //------------------------------schedule_local---------------------------------
706 // Topological sort within a block. Someday become a real scheduler.
707 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
708 // Already "sorted" are the block start Node (as the first entry), and
709 // the block-ending Node and any trailing control projections. We leave
710 // these alone. PhiNodes and ParmNodes are made to follow the block start
711 // Node. Everything else gets topo-sorted.
712
713 #ifndef PRODUCT
714 if (trace_opto_pipelining()) {
715 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
716 for (uint i = 0;i < block->number_of_nodes(); i++) {
717 tty->print("# ");
718 block->get_node(i)->fast_dump();
719 }
720 tty->print_cr("#");
721 }
722 #endif
723
724 // RootNode is already sorted
725 if (block->number_of_nodes() == 1) {
726 return true;
727 }
728
729 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
730 uint node_cnt = block->end_idx();
731 uint phi_cnt = 1;
732 uint i;
733 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
734 Node *n = block->get_node(i);
735 if( n->is_Phi() || // Found a PhiNode or ParmNode
736 (n->is_Proj() && n->in(0) == block->head()) ) {
737 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
738 block->map_node(block->get_node(phi_cnt), i);
739 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
740 } else { // All others
741 // Count block-local inputs to 'n'
742 uint cnt = n->len(); // Input count
743 uint local = 0;
744 for( uint j=0; j<cnt; j++ ) {
745 Node *m = n->in(j);
746 if( m && get_block_for_node(m) == block && !m->is_top() )
747 local++; // One more block-local input
748 }
749 ready_cnt.at_put(n->_idx, local); // Count em up
750
751 #ifdef ASSERT
752 if( UseConcMarkSweepGC || UseG1GC ) {
753 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
754 // Check the precedence edges
755 for (uint prec = n->req(); prec < n->len(); prec++) {
756 Node* oop_store = n->in(prec);
757 if (oop_store != NULL) {
758 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
759 }
760 }
761 }
762 }
763 #endif
764
765 // A few node types require changing a required edge to a precedence edge
766 // before allocation.
767 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
768 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
769 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
770 // MemBarAcquire could be created without Precedent edge.
771 // del_req() replaces the specified edge with the last input edge
772 // and then removes the last edge. If the specified edge > number of
773 // edges the last edge will be moved outside of the input edges array
774 // and the edge will be lost. This is why this code should be
775 // executed only when Precedent (== TypeFunc::Parms) edge is present.
776 Node *x = n->in(TypeFunc::Parms);
777 n->del_req(TypeFunc::Parms);
778 n->add_prec(x);
779 }
780 }
781 }
782 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
783 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
784
785 // All the prescheduled guys do not hold back internal nodes
786 uint i3;
787 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
788 Node *n = block->get_node(i3); // Get pre-scheduled
789 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
790 Node* m = n->fast_out(j);
791 if (get_block_for_node(m) == block) { // Local-block user
792 int m_cnt = ready_cnt.at(m->_idx)-1;
793 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
794 }
795 }
796 }
797
798 Node_List delay;
799 // Make a worklist
800 Node_List worklist;
801 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
802 Node *m = block->get_node(i4);
803 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
804 if (m->is_iteratively_computed()) {
805 // Push induction variable increments last to allow other uses
806 // of the phi to be scheduled first. The select() method breaks
807 // ties in scheduling by worklist order.
808 delay.push(m);
809 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
810 // Force the CreateEx to the top of the list so it's processed
811 // first and ends up at the start of the block.
812 worklist.insert(0, m);
813 } else {
814 worklist.push(m); // Then on to worklist!
815 }
816 }
817 }
818 while (delay.size()) {
819 Node* d = delay.pop();
820 worklist.push(d);
821 }
822
823 // Warm up the 'next_call' heuristic bits
824 needed_for_next_call(block, block->head(), next_call);
825
826 #ifndef PRODUCT
827 if (trace_opto_pipelining()) {
828 for (uint j=0; j< block->number_of_nodes(); j++) {
829 Node *n = block->get_node(j);
830 int idx = n->_idx;
831 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
832 tty->print("latency:%3d ", get_latency_for_node(n));
833 tty->print("%4d: %s\n", idx, n->Name());
834 }
835 }
836 #endif
837
838 uint max_idx = (uint)ready_cnt.length();
839 // Pull from worklist and schedule
840 while( worklist.size() ) { // Worklist is not ready
841
842 #ifndef PRODUCT
843 if (trace_opto_pipelining()) {
844 tty->print("# ready list:");
845 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
846 Node *n = worklist[i]; // Get Node on worklist
847 tty->print(" %d", n->_idx);
848 }
849 tty->cr();
850 }
851 #endif
852
853 // Select and pop a ready guy from worklist
854 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
855 block->map_node(n, phi_cnt++); // Schedule him next
856
857 #ifndef PRODUCT
858 if (trace_opto_pipelining()) {
859 tty->print("# select %d: %s", n->_idx, n->Name());
860 tty->print(", latency:%d", get_latency_for_node(n));
861 n->dump();
862 if (Verbose) {
863 tty->print("# ready list:");
864 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
865 Node *n = worklist[i]; // Get Node on worklist
866 tty->print(" %d", n->_idx);
867 }
868 tty->cr();
869 }
870 }
871
872 #endif
873 if( n->is_MachCall() ) {
874 MachCallNode *mcall = n->as_MachCall();
875 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
876 continue;
877 }
878
879 if (n->is_Mach() && n->as_Mach()->has_call()) {
880 RegMask regs;
881 regs.Insert(_matcher.c_frame_pointer());
882 regs.OR(n->out_RegMask());
883
884 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
885 map_node_to_block(proj, block);
886 block->insert_node(proj, phi_cnt++);
887
888 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
889 }
890
891 // Children are now all ready
892 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
893 Node* m = n->fast_out(i5); // Get user
894 if (get_block_for_node(m) != block) {
895 continue;
896 }
897 if( m->is_Phi() ) continue;
898 if (m->_idx >= max_idx) { // new node, skip it
899 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
900 continue;
901 }
902 int m_cnt = ready_cnt.at(m->_idx)-1;
903 ready_cnt.at_put(m->_idx, m_cnt);
904 if( m_cnt == 0 )
905 worklist.push(m);
906 }
907 }
908
909 if( phi_cnt != block->end_idx() ) {
910 // did not schedule all. Retry, Bailout, or Die
911 if (C->subsume_loads() == true && !C->failing()) {
912 // Retry with subsume_loads == false
913 // If this is the first failure, the sentinel string will "stick"
914 // to the Compile object, and the C2Compiler will see it and retry.
915 C->record_failure(C2Compiler::retry_no_subsuming_loads());
916 }
917 // assert( phi_cnt == end_idx(), "did not schedule all" );
918 return false;
919 }
920
921 #ifndef PRODUCT
922 if (trace_opto_pipelining()) {
923 tty->print_cr("#");
924 tty->print_cr("# after schedule_local");
925 for (uint i = 0;i < block->number_of_nodes();i++) {
926 tty->print("# ");
927 block->get_node(i)->fast_dump();
928 }
929 tty->cr();
930 }
931 #endif
932
933
934 return true;
935 }
936
937 //--------------------------catch_cleanup_fix_all_inputs-----------------------
938 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
939 for (uint l = 0; l < use->len(); l++) {
940 if (use->in(l) == old_def) {
941 if (l < use->req()) {
942 use->set_req(l, new_def);
943 } else {
944 use->rm_prec(l);
945 use->add_prec(new_def);
946 l--;
947 }
948 }
949 }
950 }
951
952 //------------------------------catch_cleanup_find_cloned_def------------------
953 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
954 assert( use_blk != def_blk, "Inter-block cleanup only");
955
956 // The use is some block below the Catch. Find and return the clone of the def
957 // that dominates the use. If there is no clone in a dominating block, then
958 // create a phi for the def in a dominating block.
959
960 // Find which successor block dominates this use. The successor
961 // blocks must all be single-entry (from the Catch only; I will have
962 // split blocks to make this so), hence they all dominate.
963 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
964 use_blk = use_blk->_idom;
965
966 // Find the successor
967 Node *fixup = NULL;
968
969 uint j;
970 for( j = 0; j < def_blk->_num_succs; j++ )
971 if( use_blk == def_blk->_succs[j] )
972 break;
973
974 if( j == def_blk->_num_succs ) {
975 // Block at same level in dom-tree is not a successor. It needs a
976 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
977 Node_Array inputs = new Node_List(Thread::current()->resource_area());
978 for(uint k = 1; k < use_blk->num_preds(); k++) {
979 Block* block = get_block_for_node(use_blk->pred(k));
980 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
981 }
982
983 // Check to see if the use_blk already has an identical phi inserted.
984 // If it exists, it will be at the first position since all uses of a
985 // def are processed together.
986 Node *phi = use_blk->get_node(1);
987 if( phi->is_Phi() ) {
988 fixup = phi;
989 for (uint k = 1; k < use_blk->num_preds(); k++) {
990 if (phi->in(k) != inputs[k]) {
991 // Not a match
992 fixup = NULL;
993 break;
994 }
995 }
996 }
997
998 // If an existing PhiNode was not found, make a new one.
999 if (fixup == NULL) {
1000 Node *new_phi = PhiNode::make(use_blk->head(), def);
1001 use_blk->insert_node(new_phi, 1);
1002 map_node_to_block(new_phi, use_blk);
1003 for (uint k = 1; k < use_blk->num_preds(); k++) {
1004 new_phi->set_req(k, inputs[k]);
1005 }
1006 fixup = new_phi;
1007 }
1008
1009 } else {
1010 // Found the use just below the Catch. Make it use the clone.
1011 fixup = use_blk->get_node(n_clone_idx);
1012 }
1013
1014 return fixup;
1015 }
1016
1017 //--------------------------catch_cleanup_intra_block--------------------------
1018 // Fix all input edges in use that reference "def". The use is in the same
1019 // block as the def and both have been cloned in each successor block.
1020 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1021
1022 // Both the use and def have been cloned. For each successor block,
1023 // get the clone of the use, and make its input the clone of the def
1024 // found in that block.
1025
1026 uint use_idx = blk->find_node(use);
1027 uint offset_idx = use_idx - beg;
1028 for( uint k = 0; k < blk->_num_succs; k++ ) {
1029 // Get clone in each successor block
1030 Block *sb = blk->_succs[k];
1031 Node *clone = sb->get_node(offset_idx+1);
1032 assert( clone->Opcode() == use->Opcode(), "" );
1033
1034 // Make use-clone reference the def-clone
1035 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1036 }
1037 }
1038
1039 //------------------------------catch_cleanup_inter_block---------------------
1040 // Fix all input edges in use that reference "def". The use is in a different
1041 // block than the def.
1042 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1043 if( !use_blk ) return; // Can happen if the use is a precedence edge
1044
1045 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1046 catch_cleanup_fix_all_inputs(use, def, new_def);
1047 }
1048
1049 //------------------------------call_catch_cleanup-----------------------------
1050 // If we inserted any instructions between a Call and his CatchNode,
1051 // clone the instructions on all paths below the Catch.
1052 void PhaseCFG::call_catch_cleanup(Block* block) {
1053
1054 // End of region to clone
1055 uint end = block->end_idx();
1056 if( !block->get_node(end)->is_Catch() ) return;
1057 // Start of region to clone
1058 uint beg = end;
1059 while(!block->get_node(beg-1)->is_MachProj() ||
1060 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1061 beg--;
1062 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1063 }
1064 // Range of inserted instructions is [beg, end)
1065 if( beg == end ) return;
1066
1067 // Clone along all Catch output paths. Clone area between the 'beg' and
1068 // 'end' indices.
1069 for( uint i = 0; i < block->_num_succs; i++ ) {
1070 Block *sb = block->_succs[i];
1071 // Clone the entire area; ignoring the edge fixup for now.
1072 for( uint j = end; j > beg; j-- ) {
1073 // It is safe here to clone a node with anti_dependence
1074 // since clones dominate on each path.
1075 Node *clone = block->get_node(j-1)->clone();
1076 sb->insert_node(clone, 1);
1077 map_node_to_block(clone, sb);
1078 }
1079 }
1080
1081
1082 // Fixup edges. Check the def-use info per cloned Node
1083 for(uint i2 = beg; i2 < end; i2++ ) {
1084 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1085 Node *n = block->get_node(i2); // Node that got cloned
1086 // Need DU safe iterator because of edge manipulation in calls.
1087 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1088 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1089 out->push(n->fast_out(j1));
1090 }
1091 uint max = out->size();
1092 for (uint j = 0; j < max; j++) {// For all users
1093 Node *use = out->pop();
1094 Block *buse = get_block_for_node(use);
1095 if( use->is_Phi() ) {
1096 for( uint k = 1; k < use->req(); k++ )
1097 if( use->in(k) == n ) {
1098 Block* b = get_block_for_node(buse->pred(k));
1099 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1100 use->set_req(k, fixup);
1101 }
1102 } else {
1103 if (block == buse) {
1104 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1105 } else {
1106 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1107 }
1108 }
1109 } // End for all users
1110
1111 } // End of for all Nodes in cloned area
1112
1113 // Remove the now-dead cloned ops
1114 for(uint i3 = beg; i3 < end; i3++ ) {
1115 block->get_node(beg)->disconnect_inputs(NULL, C);
1116 block->remove_node(beg);
1117 }
1118
1119 // If the successor blocks have a CreateEx node, move it back to the top
1120 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1121 Block *sb = block->_succs[i4];
1122 uint new_cnt = end - beg;
1123 // Remove any newly created, but dead, nodes.
1124 for( uint j = new_cnt; j > 0; j-- ) {
1125 Node *n = sb->get_node(j);
1126 if (n->outcnt() == 0 &&
1127 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1128 n->disconnect_inputs(NULL, C);
1129 sb->remove_node(j);
1130 new_cnt--;
1131 }
1132 }
1133 // If any newly created nodes remain, move the CreateEx node to the top
1134 if (new_cnt > 0) {
1135 Node *cex = sb->get_node(1+new_cnt);
1136 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1137 sb->remove_node(1+new_cnt);
1138 sb->insert_node(cex, 1);
1139 }
1140 }
1141 }
1142 }