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